Development of a. waviness measurement for coated products. (Wavimeter) EUR EN. Research and Innovation

Transcription

1 Development of a waviness measurement for coated products (Wavimeter) Research and Innovation EUR EN

2 EUROPEAN COMMISSION Directorate-General for Research and Innovation Directorate G Industrial Technologies Unit G.5 Research Fund for Coal and Steel rtd-steel-coal@ec.europa.eu RTD-PUBLICATIONS@ec.europa.eu Contact: RFCS Publications European Commission B-1049 Brussels

3 European Commission Research Fund for Coal and Steel Development of a waviness measurement for coated products (Wavimeter) G. Moreas, A. Loroy, A. Vion, P. Albart, L. Bellavia, O. Lemaire CRM Avenue du Bois St-Jean 21 (P59), 4000 Liège, BELGIUM G. Fricout ArcelorMittal Research Voie Romaine, Maizières-lès-Metz, FRANCE O. Deutscher BFI Sohnstrasse 65, Düsseldorf, GERMANY E. Montagna Segal Chaussée de Ramioul 50, 4400 Flémalle, BELGIUM G. Vecino, A. Espina ArcelorMittal Spain Residencia La Grando, Asturia, SPAIN Grant Agreement RFSR-CT July 2006 to 31 January 2010 Final report Directorate-General for Research and Innovation 2013 EUR EN

4 LEGAL NOTICE Neither the European Commission nor any person acting on behalf of the Commission is responsible for the use which might be made of the following information. The views expressed in this publication are the sole responsibility of the authors and do not necessarily reflect the views of the European Commission. Europe Direct is a service to help you find answers to your questions about the European Union Freephone number (*): (*) Certain mobile telephone operators do not allow access to numbers or these calls may be billed. More information on the European Union is available on the Internet ( Cataloguing data can be found at the end of this publication. Luxembourg: Publications Office of the European Union, 2013 ISBN doi: /73039 European Union, 2013 Reproduction is authorised provided the source is acknowledged. Printed in Luxembourg Printed on white chlorine-free paper

5 TABLE OF CONTENT 1. FINAL SUMMARY SCIENTIFIC AND TECHNICAL DESCRIPTION OF THE RESULTS OBJECTIVES AND BACKGROUND COMPARISON OF INITIALLY PLANNED ACTIVITIES AND WORK ACCOMPLISHED MAJOR ACHIEVEMENTS DESCRIPTION OF ACTIVITIES AND DISCUSSION Study of the different types of waviness (all the partners) (WP1) Samples collection (Task1.1.) Laboratory topographic measurements (Task 1.2.) BFI instrument ARsa instrument CRM instrument Study of the results (Task 1.3.) Samples used to define the waviness Investigation on different kinds of waviness Investigation of possible practicable waviness characterising parameters for online measurements Comparison of roughness and waviness measured by each partner ARsa measurements CRM measurements Explanation of outliers Global considerations on measurements Conclusion Development of waviness sensor (CRM ARsa AM Spain) (WP2) Wavimeter: acquisition part (CRM) (Task 2.1, Task 2.2) General requirements Camera requirements Laser requirements Laser test Fibre coupling of laser Control software for laser and camera Synchronisation mode of camera and laser Synchronisation signal Optical mounting for laser Acquisition and control unit Wavimeter Wavimeter: image stitching (ARsa-AM Spain-CRM) (Task 2.3) Stitching requirements Description of the main algorithms Specific problems Results of stitching process C++ implementation and optimisation Real time system Sub-pixel precision Wavimeter: stitched image analysis (CRM) (Task 2.3) Wavimeter: profile analysis (CRM) (Task 2.3) Wavimeter: profile analysis (BFI) (Task 2.3) Wavimeter: time cycle (CRM) (Task 2.3) Validation on industrial samples in laboratory (CRM-ARsa-BFI) (WP3) Validation on static samples (Task 3.1) Validation of the method Mounting and conditions of test Solved problems Results Validation on dynamic samples (Task 3.2) Mounting and conditions of test Tests

6 Investigation in the principle possibilities to measure waviness with the Raumed measuring system Conclusion Integrated on-line strip surface waviness measurements on galvanising line at Segal (Segal-CRM-BFI) (WP4) Delay in industrial tests Mechanical mounting (Task 4.1) On-line pre-test (Task 4.1) Results Conclusion on pre-test On-line installation (Task 4.2) Electrical, mechanical and software preparations First position Second position Adjustment of the sensor Validation on samples (Task 4.2) First campaign of tests Sensor modifications Validation on specific coil (Task 4.3) Database constitution and study of interactions parameters at Segal (Segal-CRM) (WP5) Database constitution (Task 5.1) Database analysis (Task 5.2) Integrated on-line strip surface waviness measurement on galvanization line at ArcelorMittal Spain (AM Spain-CRM-ARsa-BFI) (WP6) Installation on-line with mechanical and electrical adaptation of plant, and communication with plant control system (AMSpain, CRM, ARsa) (Task 6.1) Validation of measurements on head and tails (AM Spain, CRM, BFI) (Task 6.2) Second campaign of tests Tail samples Validation of measurements on specific coils (AM Spain, CRM, BFI) (Task 6.3) Database constitution and study of interactions parameters at ArcelorMittal Spain (AM Spain-ARsa) (WP7) Database constitution (Task 7.1) Database analysis (Task 7.2) Conclusion Determination of the guidelines to improve the process (WP8) (All the partners) Study on results from WP5 and WP7 - Study of the common interactions on both lines - Study of parameters specific to each line (Task 8.1) Guidelines description (Task 8.2) Conclusion CONCLUSIONS EXPLOITATION AND IMPACT OF THE RESEARCH RESULTS LIST OF FIGURES LIST OF REFERENCES

7 1. FINAL SUMMARY Basically, the surface topography refers to the horizontal and vertical information on the height variation of the surface and can be divided into (micro) roughness and waviness. Generally, the roughness is defined to include only the height variations having a horizontal spacing or wavelength of 5 to 10 times the vertical peak to valley distance corresponding to wavelengths below mm. The waviness, on the other hand, refers to structural components with longer wavelengths, approximately between 0.1 and 30 mm. Waviness means, in general, the shape deviation of a surface in the wavelength range above the roughness, i.e. the shape above the usual cut-off frequency (Lc) of mostly 0.8 or 2.5 mm in case of steel. If the waviness of the steel sheet is above defined limits, the steel user does not accept the product because it causes scrap or bad quality in the further processing (such as deep drawing, painting or brushing). As waviness, except for high glossy surfaces, is mostly not eye visible before painting, a measurement is required to evaluate its level. Nowadays, this is done by the use of profilometers on static samples from the tail or the head of some coils. However, this measurement is not available for the whole coil and, moreover, is only realised on some ones (1 on 5 in average). Nevertheless, the waviness is becoming more and more important for the customers, specially the automotive ones, who require better quality to avoid scrap and to try to put less painting on the steel product without losing aspect quality. So, as nothing exists today, an on-line system, measuring the waviness and recognising its type during the production process, is of great interest and was the main objective of this project. The scientific and technological objectives of this project were: - To implement an accurate on-line measurement of the strip waviness based on microscope techniques. This will be done first in laboratory and then in industrial lines at Segal (Corus) and at Arcelor Mittal Spain. Working on two different lines producing under different conditions and also different textures will improve the knowledge of the waviness sources. - To constitute a database to evaluate the parameters influencing the strip waviness and to determine the interdependencies between these parameters in order to have a better understanding of the physical mechanisms influencing the surface, leading to the definition of guidelines to control the strip waviness and to improve process, - To define a common reference as well for steel producers and customers. A collection of samples has been done following specifications given to the industrial partners (AM Spain and Segal). BFI collected also another series of samples covering EBT, EDT and Pretex textures. Mainly all the samples have turned around partners to characterize the surface with different measuring devices and to compare the results. All the collected samples have been measured using different laboratory instruments: 2 mechanical stylus based instruments from BFI and CRM respectively, and one interferometric instrument from ARsa. During the project, the different types of steel sheet waviness like for example chatter marks, micro waviness and Lüders bands have been investigated. Because strip waviness in the type of chatter marks is one of the most monitored waviness, extensive investigation on that kind of waviness was made. The chatter marks are long wavelength components in the rolling direction, that are, for example, caused by roller vibration, and that can give rise to form deviations or thickness variations. As the amplitude is only of the order of a few microns, they are extremely difficult to visualise on the metal surface as they are buried by the shorter wavelength surface roughness components. Adequate filters may be found which allow the chatter to be visualised and to be analysed according to amplitude and wavelength. Measurement of both the upper and lower strip surfaces allows a separate analysis of form 5

8 or thickness variations. The study advices to separate shape and thickness waviness using power spectrum analysis. Beside the strip waviness kind chatter marks, other kinds of strip waviness were also investigated. Further investigations were made on the basis of the last results of the CARSTEEL project. In this project, a new uniform parameter for characterising the waviness of steel sheet surfaces (outer skin panels for cars) and the according measuring conditions were proposed in close cooperation with several steel, paint and car producers. From this consortium, a clear defined Wa(1-5)L30-value was developed on the base of the correlation between the waviness parameters of the steel sheet surface and the waviness in the top coat. So, following the CARSTEEL demand, an online waviness measuring system should normally measures a surface profile length of minimal 30mm with a point density 300 points/mm, using a low cut-off of 1mm and a high cut-off of 5mm. This means that 25 to 50 microscope fields (1200 µm, 600 µm) minimum (without any recovery) have to been stitched to get the profile for calculating a single Wa(1-5) value. Study on smaller measurement lengths and on the required number of points per mm has also been done. The conclusion is that it seems to be possible to reduce the profile length from 30 up to 20 mm and down to 1000 points with good correlation. In laboratory, interferometric measurements showed clearly limitations in the precision. This problem appeared to be mainly due to problem of interpolation required on high reflective surfaces. So, due to their possible influence on the results, interpolation for non-measured points, surface texture, have to be taken into account. Moreover, for on-line Wavimeter validation, the computation method and realistic target accuracy have to be very carefully assessed. The comparison of the laboratory measurements, when all done along rolling direction, has validated the Ra, calculated with a cut-off of 2.5mm, and Wa calculated with a low cut-off of 1mm and a high cutoff of 5mm, on a length of 45mm maximum. During this comparison of various measurements, adequate filtering of base profile has been deeply studied. To calculate waviness and separate it from the roughness, filtering the profile is required. In general, the Gaussian filter is to be preferred as this is phase correct i.e. it does not introduce phase distortion into the transmitted wave. Various implementations of this filter have been implemented and tested to optimise the time processing. The conclusion was the use of Gaussian like filter of "n" magnitude order implementation for laboratory and on-line measurements. This filtering compared with the BFI instrument filtering gave a very good correlation. Because it is extremely costly and difficult to design a vision system with a high resolution and a large field of view at the same time, the adopted method combined two techniques. The first one is the high resolution triangulation method used in Topometer project in which the profile of the surface is measured by the analysis of the deformation of the line projected onto the surface at a known angle. The field of view width is in the range of 1mm. The adding of the Wavimeter consists in the grabbing at high frequency of a series of such images with a slight recovery (defined by laboratory tests). These images, whose background is visible through the use of adequate illumination, are then stitched together to get the required length of measurement. To apply the defined method, following the general specifications for the waviness measurements, high speed camera and high speed laser have been chosen. So, a new sensor head and all the control boxes have been built and tested. 6

9 The high speed laser appears to be a critical component of the installation as it failed two times during this project leading, combined with economical conditions in 2008 and 2009, to high delay in the completion of the sensor and the industrial tests. After discussion with the European Commission, the decision was taken to suspend the project for 7 months in order to wait better industrial conditions to run the tests in plants. Anyway, the possibility to modify the laser unit should be studied for the realisation of an industrial sensor. High precision external synchronisation technique, based on the conversion of speed signal to reach continuously the defined recovery amount between images, has been implemented. It allows synchronising very precisely the grabbing frequency with the speed of the product. This results in a nearly constant amount of pixels of recovery image from image. This leads so in a higher speed of computation for the stitching as the stitching area is not changing a lot. High attention has also been given to the quality of the optical components and the stability of the used mountings to reach a configuration as stable as possible. Indeed, stable lighting of the background and stable level of line gives higher stability in the results in time. The system is multitasking by the combination of multiple processes in parallel: checking of the security position, deciding of the adequate acquisition moment, grabbing images, calculation of the Ra and Wa values. In parallel, stitching techniques have been investigated. Stitching algorithms refers to a set of techniques used to recover a large image from a set of smaller one. It is commonly used for instance to recover a full panorama based on a set of pictures taken with a camera. In this project, the trade-off between resolution and field of view is impossible to handle without using such techniques. A review of possible algorithms has been made. Finally, the used stitching method is based on a simple transformation model which consists in a translation which is estimated between two consecutive image using a correlation technique and FFT algorithms. No rotation or change in zoom is considered. This is realistic because the optical configuration of the acquisition device is very well controlled. A few problems specific to the optical and mechanical configuration under study have been highlighted. They are detailed below. An element that could limit the stitching accuracy is the occurrence of bright spots due to the specular reflexion of the laser line. The position of these spots depends on the exact direction from which the laser line is projected on the surface. So, it is important to use background sub-windows that do not contain the laser line to evaluate the translation vector. The most noticeable feature is the orientation of the laser line compared to the strip direction. Ideally, the laser line would be exactly projected in the direction of the strip movement, so that acquiring several successive images would be equivalent to project a longer laser line on a larger field-of-view. On the first experimental acquisition, a slight misalignment was almost each time visible. This phenomenon is a problem in itself, because it means that the reconstructed profile will not be fully continuous. To align perfectly the system, the camera and the projected line are adjustable. The camera is oriented by the use of thin plates placed on the adequate side between the camera and its support. The adjustment is optimised when the transverse stitching displacement (orthogonal to rolling direction) is near 0 pixels. The line direction is adjusted through the use of a high precision rotation table (on which the line projection optics is fixed combined to an FFT analysis of the resulting stitched image. Orientating the line in the rolling direction will result in the extinction of the peak at the length of stitching. The homogeneity of the source can also play a role in the stitching process. In the successive improvements of the sensor, this has been corrected and in the current version, the homogeneity of the light source is quite good. 7

10 The suitable choice of the initial parameters in the search of the displacement between matching and displaced windows is a very important factor in order to reduce the processing time. Specific algorithms have been implemented depending on the type of texture for example. A fast algorithm implementation for computation of the displacement vector between two subsequent images, and the generation of the stitched image made of a set of adjacent images, has been optimised. It is based on the FFT-based CCN algorithm. It has been implemented as a C++ class for easy integration in other software pieces. One objective has been to reduce the processing time in order to achieve a real time system, which permits to process all the images captured in each sample (around 70 images). The algorithm was optimized and several tests have been done looking for reaching the suitable values of parameters like initial horizontal and vertical displacements, this way the size of the displaced window is reduced and therefore the processing time. 3 sub-pixels methods have been studied to get a higher accuracy in stitching. The conclusion was that sub-pixel processing does not improve image stitching in an important way meanwhile it would cause some disadvantages, mainly in time processing. Having got a complete stitched image, the line present in the image has to be isolated for further analysis. The algorithms initially developed for Topometer have been adapted. Specific techniques based on linear regression and filtering on standard deviation have been implemented. Due essentially to the transfer of images from camera to PC, the acquisition cycle is around 20 seconds. Beside this, the image processing takes presently around 25 seconds. The first point could be decreased using the new GigE products having a higher transfer rate or even, a camera including processor that could potentially allow at least part of processing inside the camera. By the way, only reduced data would have to be transferred reducing considerably this operation. On the side of image processing, as it is well defined today, possibility to reduce some parts which are time consuming could be to use specifically designed electronic circuit. In an industrial version of the Wavimeter, both aspects should be deeper investigated. All the described studies and developments have been used for the validation of the method principle. The method has been so validated in laboratory using a simulator to move the samples from point to point and to take images at each. Then, the developed analysis algorithms were used to calculate Wa and Ra on the huge database of images. The analysis validated the method in laboratory, as well for the image stitching as for the image processing and for profile filtering. The precision is very good (mainly in +/-15%) and the correlation with mechanical stylus measurement also (above 0.86). The dispersion could probably be decreased by using an inclined configuration that will be less influenced by the reflectivity of the surface. But, anyway, even when comparing mechanical data, the mechanical Wa value is very sensitive and shows higher dispersion than Ra data. This has to be taken into account in the evaluation of quality of optical data. The dynamic tests validated the synchronisation and helped to improve the electrical connections and electronic circuit to reach the best grabbing frequency, the linearity relative to speed frequency and the stability with a frequency variation lower than 1%. Another instrument has been tested by BFI for its possible application on-line. The Raumed measuring system normally is used to measure the roughness profiles of steel strips online. It was adapted for a laboratory test on the collected samples described earlier. The analysis of data shows that there is good correlation between the calculated Wca values and the measured ones. Taking away outliers, there is a good correlation between the calculated Wa values and the mechanical ones. Therefore, the Raumed measurement could be tried in the future for on-line trials test to evaluate and compare its precision as online waviness measuring system, after comparison with Wavimeter on a same series of samples in laboratory. 8

11 Prior to installing the Wavimeter in Segal plant, it was desirable to evaluate the distance variations between the rotating roll and a fixed point with respect to the roll frame, to check the possibility of having a sufficiently long set of well-focussed images for Waviness calculation. Vibrations, roll eccentricity or roll finishing defects can cause such distance variations. The distance variation over one roll turn was measured by two techniques and was around 80µm. It was not a problem for the waviness measurement, while the vibrations could be. However, due to the higher weight of the actual waviness measurement system and due to the fact that it will be positioned on a shorter arm, the amplitude of the vibrations was supposed to be smaller and the resonance frequency should be modified. Some damping system in the support frame could also help in this regard. The conclusion of this test was that one can expect that working on the selected roll might be feasible, and will be tried. The Wavimeter has so been installed at the pre-test position in Segal. But, unhappily, the vibrations were too high with a too high frequency. This induced a variation of more than 200µm height on a measurement length of 50mm. These vibrations coming from the roll and its structure were transmitted and amplified by the beam support to the sensor. As the focus depth is only 20µm, these 200µm were of course inacceptable. A first test of improvement was to rigidify the base support of the sensor head itself. It decreased the amplitude of vibrations to 150µm but increase the frequency, so the situation was worse. A second test was to change completely the beam support by reducing its size to the centre of the structure of the roll structure. This resulted in a decrease of the frequency of vibrations and a slight decrease of the amplitude. However, the variation on 50mm length was still too high (around 90µm). So it was decided to change the measurement position. The second position is just after the skin-pass (around 30m). The disadvantage is that the measurable side is the bottom side of the strip but, for a validation step, it is acceptable. The big advantage was that it existed already a base beam support fixed on the floor and no more on the roll structure. So, a mechanical adaptation was realised quickly in CRM to be placed on the beam and support the sensor head. There, the vibrations were far lower (only around 15µm along 50mm of measurement). This place was so validated for some test weeks. In parallel, AM Spain was contacted to modify its support taking into account the tests done at Segal. The first thing to do when the sensor is installed is to adjust its global position. This is done through the use of the distance sensor signal and the positioning tables supporting the core of the sensor. This signal is minimal when the optical axis is orthogonal to the roll surface. Moreover, specific algorithms have been written to adjust the Wavimeter position on line. The focus is secondly checked by blurring evaluation. The camera inclination is modified based on the stitching translation values. The triangulation line is aligned along rolling direction by the analysis of the profile spectrum. A first campaign of measurements was finally started at Segal in December Samples were taken from head and tails. It appears quickly that even an angle of 4 (tested in laboratory) was too low to get a sufficient amount of the line through the camera. But, the tests to go further in tilting let appear a high loose of quality in the background image. This problem is due to the high reflectivity of the strip and did not appear on laboratory samples that were no more so reflective ( old material). The result of this first campaign was the modification of the implantation of the camera inside the core of the sensor. It implied also modifications in other support plates. The system was so modified to be able to be at 8 with the lighting of the background not coming from the microscope optics anymore but from external, and so being more in a bright field configuration than in a dark field configuration that does not fit with high reflective products. These modifications were done just before leaving to AM Spain for a second campaign of tests. 9

12 The modified Wavimeter was so installed in January 2010 in AM Spain on a support beam fixed with rubber cushion used to avoid vibrations. Samples were taken from head and tails and measured mechanically beside the line with the CRM laboratory instrument. The position of mechanical measurement was the same as the one of Wavimeter in the width of strip. It has quickly appeared that the AM Spain product is more reflective than Segal product. The result was a bad quality image even in 8 degree inclination with a very low fill factor of the line. After some study on site, a second mechanical modification was decided. This modification results in a final inclination of the optical axis at 16 related to the orthogonal to the surface. The grabbed images are far better and the fill factor too. Even with the small focus depth, the resulting image does not appear blurred. This solution improves the line quality but results in a small loss of precision in the measurement due to inclination. It is a compromise between a good fill factor and a better precision. Nevertheless, image processing had to be slightly adapted to take into account the inclination and the high product reflectivity. During the days of measurement, samples were taken at the end and in middle of coils. Some of the samples, covering a priori a large range of Wa values have been sent to BFI for the validation of mechanical measurements. The correlation between CRM and BFI Wa measurements is very good. However, it must be noted that relatively large deviations were observed on all kinds of measurements: it reaches regularly 15%! That can imply also difficulties in the validation of measurements. The variation of the Wa on product should probably be checked along more than 30mm on a sample, for example, along long lengths of product. The samples sent to BFI covered as well tail of coils as mid of coils. The Wavimeter data were in the correct range but the correlation with mechanical data was not good on these samples, except for those taken in mid of coils. This is due to the specific and very local process conditions near the welding when it is passing the skin-pass. So, to analyse correctly the results, on the whole campaign done in AM Spain, due to this specific and very local process conditions near the welding, only the samples cut in the middle of the coils have been considered. Considering the average of the last 4 measurements, the correlation is good and the variations are in +/-10% related to the mechanical Wa. This is a very good result considering, for the correlation factor of 0.71, that it is influenced by the small range of data (from0.28 to 0.42µm) and that it should be higher with a wider range of data. A series of coils have been followed during the campaign. This has been used, coupled with samples at the end of coils to check the Wavimeter measurements in comparison with mechanical data during variation of production parameters. It was shown that the Wavimeter was so able to follow variations of Wa highlighted during the tests. Most important result of the database analysis is that the variations in skin-pass forces induced Wa variations. Due to the relatively short time let to conduct industrial trials, not all operating conditions have been checked also because they were not planned in the plant schedule overloaded at that time. It should be very interesting to continue analysis of such elements combined with Wavimeter on-line measurements. The study of guidelines to be followed to reach a determined Wa was started but, during the campaign, due to the short let time and the overloaded plant planning, all the process conditions were not covered leading to difficulties to separate all the influences. It should be very interesting to continue analysis of such elements combined with on-line measurements but during long term periods to cover a large range of process conditions without imposing specific tests to plant. 10

13 In conclusion, all the tasks planned in the research program have been covered and notably the industrial trials in two sites. The developed measurement technology allows characterising on line the waviness. The technical problem encountered during the setup of the sensor as well the economical situation prevailing in 2008 and 2009 have shortened the time let for the industrial testing and have not allowed covering a large range of conditions as initially expected. The obtained results are nevertheless sufficiently convincing to envisage a further development and the industrial implementation of the technique. This further test could be managed inside a pilot project in collaboration with plants and an industrial manufacturer of sensor. 11

14 2. Scientific and technical description of the results 2.1. Objectives and background If the waviness of the steel sheet is above defined limits, the steel user does not accept the product because it causes scrap or bad quality in the further processing (such as deep drawing, painting or brushing). As waviness, except for high glossy surfaces, is mostly not eye visible before painting, a measurement is required to evaluate its level. Nowadays, this is done by the use of profilometers on static samples from the tail or the head of some coils. However, this measurement is not available for the whole coil and, moreover, is only realised on some ones (1 on 5 in average). Nevertheless, the waviness is becoming more and more important for the customers, specially the automotive ones, who require better quality to avoid scrap and to try to put less painting on the steel product without losing aspect quality. So, as nothing exists today, an on-line system, measuring the waviness and recognising its type during the production process, is of great interest and was the main objective of this project. The scientific and technological objectives of this project were: - To implement an accurate on-line measurement of the strip waviness based on microscope techniques. This will be done first in laboratory and then in industrial lines at Segal (Corus) and at ArcelorMittal Spain. Working on two different lines producing under different conditions and also different textures will improve our knowledge of the waviness sources. - To constitute a database to evaluate the parameters influencing the strip waviness and to determine the interdependencies between these parameters in order to have a better understanding of the physical mechanisms influencing the surface, leading to the definition of guidelines to control the strip waviness and to improve process, - To define a common reference as well for steel producers and customers, Basically, the surface topography refers to the horizontal and vertical information on the height variation of the surface and can be divided into (micro) roughness and waviness. Generally, the roughness is defined to include only the height variations having a horizontal spacing or wavelength of 5 to 10 times the vertical peak to valley distance corresponding to wavelengths below mm. The waviness, on the other hand, refers to structural components with longer wavelengths, approximately between 0.1 and 30 mm. Waviness means, in general, the shape deviation of a surface in the wavelength range above the roughness, i.e. the shape above the usual cut-off frequency (Lc) of mostly 0.8 or 2.5 mm in case of steel. (Fig. 1) a. b. Fig. 1 Waviness The wavelengths that can be observed in the surface correspond to roughness and contribute to the dullness effect if they are below 0.1mm. The wavelengths from 0.1 to 30 mm determine the waviness. The range between 0.1 and 0.3mm (Wa) affects to the Distinctness Of Image, whilst the range 0.3 to 1 mm (Wb) and 1.0 to 30 mm (Wc) determine the orange peel specially. Longer wavelengths contribute to create unflatness. 12

15 According to state of art, it can be stated in a first approach that the waviness amplitude is below 3 µm and the maximum wavelength 30 mm. The most important kinds of waviness are: Waviness type a: Waviness in the wavelength range from circa 0.5 or 1 mm to about 5 or 10 mm in all directions (they are like buckles in the surface with amplitudes of about 0.5 to 2.5 µm). Waviness type b: Transverse undulations / Chatter marks. Waviness type c: Edge or central waves. Waviness type d: Stretcher strains. Waviness type e: Coil breaks. Waviness type b is may be the most known waviness of steel flat products. Chatter marks are waves of unevenness in the strip in the direction of rolling. This usually extends over the entire width of the product. This can be due to shape waviness or thickness variations. Waviness type c (edge and central waves) consists in uneven undulations running longitudinal to the rolling direction. They do not extend over the entire width of the strip. Edge and central waves can be caused by profile defects (e.g. ridges) in the starting material. In hot-dip metal coating lines, they can also be the result of incorrect support roller profiles and incorrect support roller positioning. Edge waves may also be induced by bright edge (edge overgrowth), or trimming with incorrect knife settings, or by contact (tub) of the strip edge with stationary structural components (e.g. guides). If the strip is free of tension, edge and central waves can be sometimes identified with the naked eye. Waviness type d (stretcher strains) consists in lines which occur during processing and which are caused by local strain. The shape of the line is not always clearly defined. During processing, nonuniform plastic deformation occurs when the yield point is exceeded, producing strain-marked zones located at about 45 to the direction of the force (Lüders bands). With increasing stretching, these zones multiply and intersect one another, until the entire surface is covered with them. Stretcher strains occur in parts and/or in zones subjected to slight deformation. Sheet and strip products that are prone to stretcher strains exhibit a pronounced yield point elongation in the stress-strain-diagram. A sufficient degree of temper rolling should suppress these pronounced yield points and thus also the occurrence of stretcher strains. Waviness type e, i.e. coil breaks, is cross breaks oriented at right angles to the rolling direction. They may run across the width of the strip or be located at the edge of the strip, with either regular or irregular spacing. If these coil breaks occur randomly with relatively wide spacing, they are termed guide roller breaks. If they occur more closely packed at short intervals and are regularly spaced over the width and length of the strip, they are termed levelling breaks (material stretched to remove coiling kinks exhibits levelling breaks). Coil breaks occurring closely spaced in the strip-edge area are termed edge breaks. The causes of coil breaks are essentially similar to those of stretcher strains. 13

16 2.2. Comparison of initially planned activities and work accomplished WP1 Study of the different types of waviness T Collect of samples T Laboratory topographic measurements T Study of the results Deliverable: list of waviness types with dimensional characterisation WP2 Development of waviness sensor T Integration of high speed camera with high speed laser T Integration of multiprocessor system T Development of adapted image processing and calculation algorithms Deliverable: prototype sensor, high speed processing unit WP3 Validation of the measurements on industrial samples in laboratory T Validation on static samples T Validation on dynamic samples Deliverable: validation reports of laboratory measurement WP4 Integrated on-line strip surface waviness measurement on galvanisation line at Segal T Installation on-line T Validation on head/tail samples T Validation on specific coils Deliverable: validation reports of measurements campaigns WP5 Database constitution and study of interactions parameters at Segal T Database constitution T Database analysis Deliverable: list of identified parameters WP6 Integrated on-line strip surface waviness measurement on galvanisation line at AM Spain T Installation on-line T Validation on head/tail samples T Validation on specific coils Deliverable: validation reports of measurements campaigns WP7 Database constitution and study of interactions parameters at AM Spain T Database constitution T Database analysis Deliverable: list of identified parameters 14

17 WP8 Determination of the guidelines to improve the process T Study on the results from WP5 and WP T Guidelines description Deliverable: general and specific guidelines list WP9 Reporting T. 9. Co-ordination meetings and reporting Deliverable: various progress reports and final report Period of suspension (from 1/12/2008 to 30/06/2009) Initially planned Realised All the tasks planned in the research program have been covered and notably the industrial trials in two sites. The developed measurement technology allows characterising on line the waviness. The technical problem encountered during the setup of the sensor as well the economical situation prevailing in 2008 and 2009 have delayed some tasks. A suspension of the contract has been asked and granted by the European Commission from 1/12/2008 to 30/06/2009. Anyway, the mentioned conditions have shortened the time let for the industrial testing and have not allowed covering a large range of conditions as initially expected. The obtained results are nevertheless sufficiently convincing to envisage a further development and the implementation of the technique fully industrially to be tested inside a pilot project in collaboration with plants and an industrial manufacturer of sensor Major achievements In summary, the reached deliverables of the project are the following: Deliverable D1 A list of waviness types with dimensional characterization has been delivered. The parameters to be applied for the on-line waviness measurement, which are: - 30mm minimum of length of measurement - Low cut-off: 1mm, high cut-off: 5mm - At least a density of 300 points per mm Mechanical stylus measurement has been assessed as reference. Deliverable D2 A prototype sensor for on-line waviness measurement has been developed in laboratory. Careful attention must be paid to synchronisation with line speed, illumination homogeneity and calculation optimisation. Deliverable D3 Wavimeter sensor has been validated in laboratory. Tests have highlighted that careful attention must be paid to synchronisation with line speed, illumination homogeneity and calculation optimisation. All these points have been integrated in D2 deliverable. The method has been validated in term of precision, mainly in +/-15%, by comparison with mechanical stylus reference, correlation above

18 Deliverable D4 Industrial tests done in Segal have demonstrated the influence of vibrations, as well as the product reflectivity. These points have to be carefully managed to reach a correct measurement. The industrial campaigns have validated the synchronisation and the image quality. Deliverable D5 The production parameters applied during the relatively short test campaign has not shown impact on waviness values. Deliverable D6 Industrial tests done in AM Spain have validated the method industrially. Comparing optical measurement of Wavimeter sensor with mechanical measurements on samples taken from mid of coils, the correlation reaches 0.71 and the variations are kept mainly in +/-10% range, which is very good. Deliverable D7 The production parameters applied during the relatively short test campaign has not shown impact on waviness values, except for skin pass force variation. The Wavimeter sensor has correctly followed the Wa change induced by this variation. Deliverable D8 Due to the shortened period of test and the limited number of process conditions during this period, the guidelines description has been limited in the assessment of influence of level of skin-pass force on Wa. Deliverable D9 All the progress reports and the draft final reports have been delivered in time Description of activities and discussion Study of the different types of waviness (all the partners) (WP1) Samples collection (Task1.1.) The collection of samples has been done following specifications given to the industrial partners (AM Spain and Segal). BFI collected also another series of samples covering EBT, EDT and Pretex textures. The samples have 200mm along the transversal side and 250mm along the rolling direction. They are marked on the side to be measured. Mainly all the samples have turned around partners to characterize the surface with different measuring devices and to compare the results. Respectively, 10, 6 and 33 samples have been collected by AM Spain, Segal and BFI Laboratory topographic measurements (Task 1.2.) In order to obtain an accurate, cross-validated reference waviness measurement on a set of reference samples, it has been decided to perform measurements with several acquisition devices. BFI, ARsa and CRM have done these measurements BFI instrument In BFI, all 2D measurements (roughness) and 3D measurements (topography) were made with an X-Ymeasuring table, resting on an air cushion. Roughness profiles were measured with a skidless stylus (stylus tip radius 5 µm) or an autofocus laser optical sensor with a spot diameter of about 1 µm. Threedimensional topography measurements were made with the same stylus or with the autofocus laser 16

19 optical sensor. The measuring device is shown in Fig. 2. The longitudinal resolution is about 1 µm. The vertical resolution of the stylus is about 12 nm and of the autofocus laser optical sensor about 6 nm. Skidless Stylus Autofocus Laser optical Sensor X-Y- Measuring Table, resting on a air cushion ARsa instrument Fig. 2 Waviness and roughness measuring system used by BFI At ARSA facilities, an interferometric measurement has been used. Interferometry is an optical, contact less method to measure 2D surface topographies. The principle if to create an interference pattern between a reference light beam and a light beam reflected on the surface under study. This interference pattern enables to compute the optical path difference between the two beams, and therefore, the Z position of the points of the surface. The main advantages of such a technique are the following: - The measurement is contact less - The topography is measured on a 2D area - The measurement is relatively fast - Resolution and Z accuracy are interesting and can be changed with the microscope objective However, the interferometry has also some drawbacks that are important to consider: - If the light flow in the microscope camera is not high enough, it is not possible to exploit the interferogram. - When a high resolution measurement is needed, the field of view is limited, because the size of the CCD is fixed, and the resolution adjusted with optical objective. So, to acquire a large area with a high resolution, it is important to perform a stitching operation (specific software is delivered with the microscope). The configuration used to perform the measurement is the following: - Objective: x5 and Zoom x0,5 - Stitching (internal microscope algorithm): 37x1 maps with Overlap: 20% between maps. - Size of stitched image: 84,2 x 2,11 mm - Size of each tile image: 2,82x 2,11 mm - Matrix camera: 320 x 240 pixels (8.8x8.8 µm) 17

20 Fig. 3 Waviness and roughness measuring system used by ARsa Because the exact definition of the parameters computation (cut-off of filters, order of polynomial, etc ) can strongly influence the waviness value, the 9 profiles extracted from the three topographical maps for each samples have been provided to the other partners to check the coherence of waviness measurements CRM instrument All 2D measurements (profiles) have been made on an optical table. The sample was fixed on it by three sides. Profiles were measured with a Talysurf system with a skidless stylus (stylus tip radius 2 µm). The measuring device is shown in Fig. 2; it is used with a connection to a portable PC with Taylor- Hobson software. The lateral resolution is about 0.5µm and the height resolution is of 16nm. Fig. 4 Waviness and roughness measuring system used by CRM The measurements were done on 50mm length (maximum possible with this instrument). 3 lines were measured in 2 zones. The results are so an averaging of 6 profiles for each sample. All the tests were done with the support of Advanced materials, solutions and sensors department. As double filtering (low and high cut-offs on the same profile) was not implemented inside the Taylor- Hobson software, filtering has been implemented by CRM. This filter, which is an Order n filter, has been used for all the profiles and will also be used for on-line calculations with the Wavimeter. (See Wavimeter: profile analysis (CRM), p69 for details) 18

21 The profiles have been given to BFI so that BFI filter could be applied on it and compare to CRM results Study of the results (Task 1.3.) Samples used to define the waviness The samples used to define the waviness are the ones provided by BFI. The sample surfaces were measured to document the kind of surface topography three dimensionally. Furthermore, the roughness and waviness parameters of the samples were measured and calculated. After this, the samples were turned over to the partners to give them the possibility to measure them with their own measuring devices too and to be able to compare the results. Some of the three dimensionally scanned surfaces are shown in Fig. 5. EDT EDT EDT EBT EBT EBT Pretex Pretex Pretex Fig. 5 Examples of the different surface topographies of the samples investigated As usual roughness parameters, the mean roughness value Ra (cut off = 2.5 mm) and the peak count value RPc (cutting lines = ± 0.5 mm) were calculated. These values are shown in Fig. 6 and Fig

22 Ra [µm] RPc [1/cm] Samples Samples Fig. 6 Roughness values Ra (mean value ± standard deviation) of the investigated samples Fig. 7 Peak count values RPc (mean value ± standard deviation) of the investigated samples To characterise the surface waviness the following characterizing parameters were measured and calculated: - the standard deviation of gradient distribution after band pass filtering 1 to 8 mm (sst; Fig. 8), - the W motif values (Fig. 9), - the Wca(1-5) values ( CARSTEEL uniform parameter; Fig. 10), - the Ra macro values (Fig. 11) and - the mean values of the power spectrum in the wave range length from 1 to 8 mm (FFT(1-8) mean; (Fig. 12) sst [µm] Samples Fig. 8 Standard deviation of gradient distribution after band passes filtering 1 to 8 mm (mean value ± standard deviation) of the investigated samples W Motif [µm] Samples Fig. 9 W Motif values (mean value ± standard deviation) of the investigated samples Wa(1-5) [µm] Samples Fig. 10 Wa values (mean value ± standard deviation) of the investigated samples Ra macro [µm] Ra macro Samples Fig. 11 Ra macro values (mean value ± standard deviation) of the investigated samples 20

23 40 35 FFT (1-8) mean [µm 2 *10-3 ] Samples Fig. 12 Mean value of the power spectrum in the wave length range from 1 to 8 mm (mean value ± standard deviation) of the investigated samples The Wa(1-5) CARSTEEL new uniform waviness parameter is defined as follow: Measuring device: - Skid less pick-up system (or other high resolution pick-up system) - Tip radius: 5 µm - Tip angle: 90 Measuring conditions: - Measuring length: 30 mm - Point density = > 300 points/mm Calculations: - Levelling mean values of beginning 0.1 mm and ending 0.1 mm to 0 - Band pass filter (50% Gaussian filter; high pass: 5 mm, low pass 1 mm) - Average deviation of the profile to the mean line (evaluation length: 25 mm centre of the profile). This new waviness parameter accepted by the European steel, automotive and paint industries should be used to compare the results of laboratory and on-line measurements Investigation on different kinds of waviness During the project, BFI investigated on the different types of steel sheet waviness like for example chatter marks, micro waviness and Lüders bands. Because strip waviness in the type of chatter marks is one of the most monitored waviness, BFI made extensive investigation on that kind of waviness. For analysing the kind of waviness, both sides of the strip have to be measured in the same area. So, the samples were drilled at two points with small drilling holes. With this you can be sure that you are measuring the samples in the same area on both sides. After that, the samples were fixed in a sample frame to avoid different deflections during measurements. Fig. 13 shows the sample frame with a mounted sample. Drilling hole markings Fig. 13 Sample frame for measuring samples on both sides to analyse strip waviness in laboratory 21

24 The chatter marks are long wavelength components in the rolling direction, that are, for example, caused by roller vibration, and that can give rise to form deviations or thickness variations. As the amplitude is only of the order of a few microns, they are extremely difficult to visualise on the metal surface as they are buried by the shorter wavelength surface roughness components. Normally, chatter is imaged by means of a grinding operation of the surface. If, however, the surface structure is measured as a 3 dimensional image, filters may be found which allow the chatter to be visualised and to be analysed according to amplitude and wavelength. Measurement of both the upper and lower strip surfaces allows a separate analysis of form or thickness variations. As an example, Fig. 14 shows the surface of a steel sheet (upper and according lower side) with chatter marks threedimensional scanned as colour contours and Fig. 15 as three-dimensional display with virtual illumination. The waviness becomes more clearly (especially when with the three-dimensional display with virtual illumination that simulates the visual impression of the real surface) if the roughness is removed for example by low pass filtering the surface with a cut off of 1 mm. Non filtered surface (upper side) 1 mm low pass filtered surface (upper side) Non filtered surface (lower side) 1 mm low pass filtered surface (lower side) Fig. 14 Three-dimensional scanned surface of sample W01 with scatter marks (monotone rising colour contours; 100 mm x 10 mm) Non filtered surface (upper side) 1 mm low pass filtered surface (upper side) Non filtered surface (lower side) 1 mm low pass filtered surface (lower side) Fig. 15 Three-dimensional scanned surface of sample W01 with scatter marks (three-dimensional display, with virtual illumination;100 mm x 10 mm) 22

25 If parallel surface profiles are measured in rolling direction as shown in Fig. 16 (upper side) and in Fig. 17 (lower side), it appears that the parallel measured profiles have nearly the same shape (inversed for the lower profile). This is one indicator for chatter marks. Fig. 16 Surface profiles of upper side of sample W01 (100 mm, 300 points/mm) Fig. 17 Surface profiles of lower side of sample W01 (100 mm, 300 points/mm) The presence of form waviness or thickness waviness can be checked by the analysis of the difference between the upper side profile and the inverted lower side profile as it is shown in Fig. 18. The low waviness in the difference profile (thickness profile) adverts to form waviness. 23

26 Surface profiles of upper side Surface profiles of lower side Thickness profile (upper side plus lower side) Fig. 18 Analysing kind of waviness of sample W01 (=> form waviness) A second method for analysing the kind of chatter mark waviness (form or thickness waviness) is to calculate the power spectra of the upper surface profile, the lower surface profile and the thickness profile (upper side profile minus inverted lower side profile) as shown exemplary in Fig. 19. For both samples W01 and W02 the amplitudes in the range of lower frequencies are much higher for the upper and the lower side profiles than for the thickness profile. This indicates form waviness. Sample W01 shows in the power spectrum of the thickness profile higher values in the range of low frequencies than W02. This means that W01 shows also certain thickness waviness. 24

27 1.E+01 1.E+01 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 1.E-05 1.E Frequencies [1/mm] Frequencies [1/mm] Power spectrum (W01; upper side) Power spectrum (W02; upper side) 1.E+01 1.E+01 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 1.E-05 1.E Frequencies [1/mm] Frequencies [1/mm] Power spectrum (W01; lower side) Power spectrum (W02; lower side) 1.E+01 1.E+01 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 1.E-05 1.E Frequencies [1/mm] Frequencies [1/mm] Power spectrum (W01; thickness profile) Power spectrum (W02; thickness profile) Fig. 19 Analysing kind of waviness of samples W01 and W02 by comparing the power spectra How different waviness of the kind chatter marks can be, is demonstrated with two other samples W03 and W04. The scanned profiles of sample W03 are shown in Fig. 20 and in Fig. 21. This sample shows waviness on the upper and on the lower side of about ± 4 to 5 µm. 25

28 Fig. 20 Surface profiles of upper side of sample W03 (100 mm, 300 points/mm) Fig. 21 Surface profiles of lower side of sample W03 (100 mm, 300 points/mm) After adding the profiles of the upper and lower sides, thickness profiles with low waviness are obtained (Fig. 22) showing nearly only the surface roughness. 26

29 Fig. 22 Thickness profiles of sample W03 (100 mm, 300 points/mm) The surface profiles of the sample W04 having low surface roughness are shown in Fig. 23 (upper side) and Fig. 24 (lower side). Fig. 23 Surface profiles of upper side of sample W04 (100 mm, 300 points/mm) 27

30 Fig. 24 Surface profiles of lower side of sample W04 (100 mm, 300 points/mm) Different amplitudes and wavelengths in the upper and the lower side of this sample can be observed. After adding the profiles of the upper and lower side, thickness profiles with low waviness are obtained (Fig. 25) showing a clear waviness. Fig. 25 Thickness profiles of sample W04 (100 mm, 300 points/mm) Fig. 26 shows the related power spectra. While sample W03 shows for the upper and the lower side nearly the same amplitudes in the low frequency range, there are clear differences in the amplitudes of the upper and the lower side of sample W04. As a result, the power spectrum of the thickness profile of sample W03 shows no really high amplitude (form waviness) while the power spectrum of the thickness profile of sample W04 shows clearly a high amplitude at about 1,5 1/mm (thickness waviness with wavelength of about 0.6 mm). 28

31 1.E+01 1.E+01 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 1.E-05 1.E Frequencies [1/mm] Frequencies [1/mm] Power spectrum (W03; upper side) Power spectrum (W04; upper side) 1.E+01 1.E+01 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 1.E-05 1.E Frequencies [1/mm] Frequencies [1/mm] Power spectrum (W03; lower side) Power spectrum (W04; lower side) 1.E+01 1.E+01 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 1.E-05 1.E Frequencies [1/mm] Frequencies [1/mm] Power spectrum (W03; thickness profile) Power spectrum (W04; thickness profile) Fig. 26 Analysing kind of waviness of sample W03 and W04 by comparing the power spectra It was checked whether this thickness waviness influences the CARSTEEL waviness characterising parameter Wa(1-5). As shown in Fig. 27, the lower side of the sample W04 shows the highest value. 29

32 Fig. 27 CARSTEEL waviness characterizing parameter Wa(1-5) of the samples W03 and W04 (U = upper side, L = lower side) Most of strip waviness kinds, investigated by BFI during the project, were analysed as chatter marks (mostly form waviness). Beside the strip waviness kind chatter marks, other kinds of strip waviness were also investigated. As an example of the results shows Fig. 28, the three dimensional scanned and visualised surface of steel sheets with different distinct Lüders bands and the belonging surface profiles. It can be observed that the amplitudes and wavelengths are in the same ranges as the ones we can analyse in the surface of skin passed steel sheets. Fig. 28 Surface (10 mm x 10 mm) and belonging surface profiles of steel sheets with different Lüders bands 30

33 An example for typical form waviness, that is not chatter marks, is shown in Fig. 29. Upper side Inverted lower side Fig. 29 Micro waviness of a steel sheet (Area: 100 mm x 100 mm; 1mm to 60 mm band pass filtered surface) To analyse whether there exist form or thickness variation, profiles of the upper side and the lower side were measured in the same area and compared. Fig. 30 shows that there exist mainly form deviations (form waviness) and only a small thickness variation. The wavelength range goes from about 10 mm up to about 30 mm. A small thickness variation of about ± 1 µm is also observed. Top side Inverted lower side Difference of top side and inverted lower side (variation of thickness) Fig. 30 Profiles of micro waviness of a steel sheet (1 mm to 60 mm band pass filtered surface) 31

34 Also extensive investigations were made by BFI in sheet surface waviness. Fig. 31 shows exemplary the 3D scanned surface of a strip surface without any and with different filtering. As the surface roughness (short wavelength) hides the identification of waviness, the surface was low pass filtered with cut -off of 1, 2, 3, 4 and 5 mm (wave length range of CARSTEEL waviness parameter Wa(1-5). The different waviness are clearly recognised. The surface waviness is like micro waviness (Fig. 29) but with shorter wavelengths. Without low pass filter 1 mm low pass filtered 2 mm low pass filtered 3 mm low pass filtered 4 mm low pass filtered 5 mm low pass filtered Fig. 31 By low pass filtering visualized sheet surface waviness (area: 20 x 20 mm) In general, differentiation has to be done between waviness of the strip as form deviations, thickness deviations or flatness and waviness in the sheet surface roughness. Rough estimated shows Fig. 32 the different areas for different kinds of waviness compared with the areas of gloss and roughness. With regard to the waviness, differentiation has to be done between more global (large area) flatness, the strip waviness and the surface waviness. 32

35 Amplitude range [µm] Wave length range [mm] Flatness, Stripwaviness 1 Surfacewaviness 0.1 Roughness "CARSTEEL" waviness parameter 0.01 Wa(1-5) Gloss wave lengths Fig. 32 Analysed areas for the different kinds of waviness characterised by wave length and amplitude compared with gloss and roughness areas Including all investigated samples, differentiation between the steel strip roughness and the different strip waviness follows the data of Fig. 33. Surface texture Wave length range Range of amplitudes Roughness mm (20) µm Surface waviness mm 1 5 mm (Wa(1-5)) µm Flatness and strip waviness mm µm Fig. 33 Differentiation between roughness and strip waviness In Fig. 33, there are areas of overlapping of wavelength ranges and ranges of amplitudes. Analysing the different waviness especially in these overlapping areas requires measuring of upper and lower side of strip and/or measuring more than one parallel profile. Also, analysing the dominating wavelengths by calculating the power spectra gives information about the type of waviness. Accepted wavelength limits are available for characterising the roughness (Ra; high pass filter 2.5 mm) and the surface waviness (Wa(1-5); band pass filter 1 5 mm). Therefore, the further investigation in this project verifying the online waviness measuring system has been concentrated on these parameters Investigation of possible practicable waviness characterising parameters for online measurements Further investigations were made by BFI on the basis of the last results of the CARSTEEL project. In this project, a new uniform parameter for characterising the waviness of steel sheet surfaces (outer skin panels for cars) and the according measuring conditions were proposed in close cooperation with several steel, paint and car producers. From this consortium, a clear defined Wa(1-5)L30-value was developed on the base of the correlation between the waviness parameters of the steel sheet surface and the waviness in the top coat. 33

36 R² (correlation with wavescan (long)) FFTmean Wa(1-5) Wa(0.8-8) WaP5(0.8G) sst Line and kind of paintings Ra macro W Motif Round Robin Lab1 (horizontal) Lab2 (horizontal) Lab3 (horizontal) Lab4 (horizontal) L1 / horizontal L2 / horizontal L3 / horizontal P1 / horizontal P2 / horizontal Lab1 (vertical) Lab2 (vertical) Lab3 (vertical) Lab4 (vertical) L1 / vertical L2 / vertical L3 / vertical P1 / vertical P2 / vertical Fig. 34 Square of correlation factor (sheet surface top coat; all in the CARSTEEL project measured paintings!) / Results of RFS-CR CARSTEEL Fig. 34 shows as a result of the CARSTEEL project the square of correlation factor (correlation sheet surface waviness top coat waviness) of all paintings measured in this project (laboratory paintings and paint shop paintings). The coefficient of determination depends on the sheet surface characterising parameter, the surface waviness of the painted sample set, the paint system and the painting conditions. The determination of the new uniform sheet surface waviness characterising parameter resulted from the ranking of the mean values of the coefficient of determination with regard to a practical small measuring length. As shown in Fig. 35, the new Wa(1-5)-value shows nearly the same coefficients of determination as the FFTmean value but requires only 30 mm measuring length instead of 82 mm. That is the reason why the CARSTEEL consortium (research institute, steel -, paint and car producers) agreed to use the Wa(1-5) value as uniform European sheet surface waviness characterising parameter in the future. R² (correlation with wavescan (long)) FFTmean Wa(1-5) Wa(0.8-8) WaP5(0.8G) sst Ra macro W Motif mean hor mean vert mean all Kind of paintings Fig. 35 Ranking of correlation for all measured paintings! Square of correlation factor (sheet surface top coat) / Results of RFS-CR CARSTEEL The not so high coefficients of determination and the high sizes of standard deviations are caused by the different surface waviness of the different painted sample sets, the different paint systems and different painting conditions. It seems that in general the horizontal paint application (engine hood or car roof) shows higher correlation with the sheet surface waviness than the vertical paint applications. So, following the CARSTEEL demand an online waviness measuring system should normally measures a surface profile length of minimal 30mm with a point density 300 points/mm. 34

37 This means that 25 to 50 microscope fields (1200 µm, 600 µm) minimum (without any recovery) have to been stitched to get the profile for calculating a single Wa(1-5) value. To reduce this error-prone stitching procedure, investigations were made by BFI whether it is possible to reduce the surface profile length and/or the profile resolution with good correlation with the origin CARSTEEL sheet surface waviness characterising parameter Wa(1-5) which follows the usual rule that the measuring length should be minimum 6 times the highest filtering length. These investigations were made using the BFI round robin sample set. For these investigations, the samples were measured with different measuring lengths (30, 25, 20, and 17.5 mm (usual roughness measuring length)) and different point densities 300, 100, 50 and 10 points/mm. Based on these measurements, the Wa(1-5) values were calculated according to the calculating rules of the CARSTEEL Wa(1-5)L30(res.:300). As shown in Fig. 36, it seems to be possible to reduce the profile length from 30 up to 20 mm with a really high correlation of these values. 0.8 Wa(1-5)-values calculated by measurments with reduced measuring length [µm] y = 0.95x R 2 = 0.91 y = 0.98x R 2 = 0.85 y = 0.98x R 2 = 0.90 Wa (1-5) L25 Wa (1-5) L20 Wa (1-5) L17.5 Linear (Wa (1-5) L25) Linear (Wa (1-5) L20) Linear (Wa (1-5) L17.5) Wa(1-5)L30 [µm] (as proposed in "CARSTEEL" project) Fig. 36 Correlation of Wa(1-5) values with different measuring lengths with the origin CARSTEEL Wa(1-5) The effects of reducing the point density at the different measuring lengths are shown in Fig. 37 to Fig. 40. Wa(1-5)-values calculated by measurments with reduced point resolutions [µm] y = 0.93x R 2 = 0.95 y = 0.94x R 2 = 0.95 y = 1.04x R 2 = 0.34 Wa (1-5) L30 (res.:100) Wa (1-5) L30 (res.:50) Wa (1-5) L30 (res.:10) Linear (Wa (1-5) L30 (res.:100)) Linear (Wa (1-5) L30 (res.:50)) Linear (Wa (1-5) L30 (res.:10)) Wa(1-5)L30(res.:300) [µm] (as proposed in "CARSTEEL" project) Fig. 37 Correlation of Wa(1-5)L30 (resolution 300 points/mm) values with Wa(1-5)L30 (different resolutions) 35

38 Wa(1-5)-values calculated by measurments with reduced point resolutions [µm] y = 0.97x R 2 = 1.00 y = 0.99x R 2 = 1.00 Wa (1-5) L25 (res.:100) Wa (1-5) L25 (res.:50) Wa (1-5) L25 (res.:10) Linear (Wa (1-5) L25 (res.:100)) Linear (Wa (1-5) L25 (res.:50)) y = 1.11x Linear (Wa (1-5) L25 R 2 (res.:10)) = Wa(1-5)L25(res.:300) [µm] Fig. 38 Correlation of Wa(1-5)L25 (resolution 300 points/mm) values with Wa(1-5)L25 (different resolutions) Wa(1-5)-values calculated by measurments with reduced point resolutions [µm] y = 0.97x R 2 = 1.00 y = 0.99x R 2 = 1.00 y = 1.09x R 2 = 0.64 Wa (1-5) L20 (res.:100) Wa (1-5) L20 (res.:50) Wa (1-5) L20 (res.:10) Linear (Wa (1-5) L20 (res.:100)) Linear (Wa (1-5) L20 (res.:50)) Linear (Wa (1-5) L20 (res.:10)) Wa(1-5)L20(res.:300) [µm] Fig. 39 Correlation of Wa(1-5)L20 (resolution 300 points/mm) values with Wa(1-5)L20 (different resolutions) Wa(1-5)-values calculated by measurments with reduced measuring length [µm] y = 0.97x R 2 = 1.00 y = 0.99x R 2 = 1.00 y = 1.06x R 2 = 0.55 Wa (1-5) L17.5 (res.:100) Wa (1-5) L17.5 (res.:50) Wa (1-5) L17.5 (res.:10) Linear (Wa (1-5) L17.5 (res.:100)) Linear (Wa (1-5) L17.5 (res.:50)) Linear (Wa (1-5) L17.5 (res.:10)) Wa(1-5)L17.5(res.:300) [µm] Fig. 40 Correlation of Wa(1-5)L17.5 (resolution 300 points/mm) values with Wa(1-5)L17.5 (different resolutions) Summarising all results gained in these investigations, it seems that for online waviness measuring the Wa(1-5)L20(res.:50) value should be favoured. So, 20 mm length and 1000 points is an advice. 36

39 Comparison of roughness and waviness measured by each partner The comparison of the measurements, initially done by each partner, shows good correlation between roughness CRM measurements and BFI measurements (see Fig. 41). The observed differences are due, for CRM measurements, to different cut-offs. Indeed, using a lower cut-off will filters more the profiles, resulting in lower Ra values. The measurements done by ARsa are more scattered. This can be explained by the measurement technique and the resulting interpolated points. (See ARsa instrument, p17). In any case, some outliers (in particular BFI17) were observed and had to be investigated. Fig. 41 Ra measurements For the waviness values, the results are more dispersed. This can easily be explained by the different cut-off values that were used, these cut-off values having a higher impact on the waviness value than on roughness values. Wa got from ARsa was calculated with low filtering 0.8mm and high filtering based on a polynomial of degree 5. Those of CRM were first calculated with (0.8mm, 2.5mm) and (0.8mm, 8mm) instead of those finally determined by the Carsteel project. In any case, the CRM Wa(0.8-8) values correlate very well with the BFI calculated Wa(1-5) values (although other filter limits were used, but including 1-5mm limits). Fig. 42 Wa initial measurements 37

40 To compare equivalent data, the profiles measured by CRM and ARsa have been processed with the filter available on their system by BFI (BFI17 to BFI33 samples profiles) and with Order n filter implemented by CRM for on-line analysis (BFI1 to BFI33 samples, A5 to A9 samples). The cut-offs were 1mm for the low one and 5mm for the high one. These results have then been compared to the BFI data (got with mechanical skidless pick-up system). Fig. 43 Wa (1-5) As shown in Fig. 43, the ARsa and CRM measurements leads sometimes to lower Wa(1-5) values and sometimes to higher ones. Especially, there are some outliers explained later in this report. Fig. 44 Wa (1-5) without biggest outliers/mismeasurements Fig. 44 shows the results after elimination of the biggest outliers/mismeasurements (points eliminated from the data). The correlations are generally good (at least 0.85 of R²), except for ARsa measurements filtered with BFI filter. 38

41 To analyse the differences in the surface profiles measured by ARSA and BFI the power spectra of the profiles were calculated. Fig. 45 shows the power spectra of the samples 05, 07 and 16. The power spectrum of sample 05 shows in the interested frequency range from 0.2 to 1 mm-1 (wavelength range from 1 to 5 mm) nearly the same, of sample 07 lower and of sample 16 higher values in the ARSA measurements than in the BFI measurements. A separate frequency caused by the stitching could not be analysed. 1.E+01 1.E+01 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 1.E E Frequencies [1/mm] Frequencies [1/mm] BFI 05; BFI measurement BFI 05; ARSA measurement 1.E+01 1.E+01 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 1.E-05 1.E Frequencies [1/mm] Frequencies [1/mm] BFI 07; BFI measurement BFI 07; ARSA measurement 1.E+01 1.E+01 Amplitudes of Powerspectrum [µm²] 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 1.E-05 1.E Amplitudes of Powerspectrum [µm²] Frequencies [1/mm] Frequencies [1/mm] BFI 16; BFI measurement BFI 16; ARSA measurement Fig. 45 Power spectra of the waviness profiles measured by BFI and by ARSA To document the surface topographies of AM Spain samples, the surfaces were scanned 3 dimensionally and the surfaces were visualised (Fig. 46). The influence of measuring length of the profile to the calculated Wa(1-5) values was checked to verify the Investigation of possible for online measurements practicable waviness characterising parameters (Fig. 47). It appears nearly no differences between the Wa(1-5) values calculated from roughness profiles with 30, 25 and 20 mm measuring length. This confirms the results of earlier investigations. 39

42 Sample 09a Sample 09b Sample 14a Sample 14b Sample 37a Sample 37b Fig. 46 3D scanned and visualized surfaces of the investigated AM Spain samples Wa(1-5) [µm] y = 1.00x R 2 = 1.00 y = 0.99x R 2 = 1.00 Wa(1-5) L=25 Wa(1-5) L=20 Linear (Wa(1-5) L=20) Linear (Wa(1-5) L=25) Wa(1-5) L=30 [µm] Fig. 47 Comparison of the waviness parameters of the investigated AM samples measured with different measuring length 40

43 ARsa measurements Complementary measurements have been performed to understand the origin of discrepancies regarding measurements with ARSA interferometers. Because of the good results obtained while comparing mechanical measurements performed by CRM and BFI, these two methods have been retained as the validation method for Wavimeter assessment. However, it was important to understand the origin of the difference between ARSA optical measurements and the others. So, additional measurements for understanding the origin of the problem have been performed with a specific focus on the first 16 BFI samples. This study has been conducted internally in ARSA, because only the interferometric measurements performed there were showing major irregularities. The main reason suspected for the discrepancies was the necessary interpolation of the non-measured points with the interferometer. The total amount of non measured points is directly linked with the steepness of the surface slopes, and it can therefore vary from one sample to another depending on the surface texture. To have a better idea of the influence f this parameter, the first sixteen BFI samples have been measured a second time with a careful registration of the amount of non-measured points. In parallel, profile measurements have been performed with a confocal microscope which is a bit less sensitive to non-measured points (but which is much slower for measuring 2D surfaces). The Wa results have been compared for the 16 samples on the one hand, with a quite poor correlation, and for the 8 samples whose number of non-measured points is below 10%. The correlation between both values is then above 90%. Also, except one outlier, the correlation between the mechanical BFI measurements and the confocal measurements is quite good (93%), even without a very careful adjustment of computation techniques which have proved to be a very important phenomena. The main conclusion of this study is that the ratio of non measured points is very important for the quality of an optical waviness measurement. This is particularly true when the optical measurement device uses stitching for reconstructing the profile. Regarding the Wavimeter, the conclusion is relevant, because it means the number of non measured points, i.e. the number of points where the line is not visible (for instance due to specular reflection) should be limited to the minimum (less than 10% would be a good value) CRM measurements To try to understand the differences between mechanical measurements from BFI and mechanical measurements from CRM, the results got from CRM measurements with CRM filter have been compared to the results got from CRM measurements with BFI filter (Fig. 48). The correlation is very good, only the slope is lower than expected. So, the filter introduces only a difference in level of results but no dispersion. This tends to confirm that perhaps some mismeasurements were done. Comparing the algorithm of the filters is difficult as the BFI filter is included in their measurement system and so is not available. 41

44 Explanation of outliers Fig. 48 Comparison of Wa got with BFI filter and CRM filter As shown in Fig. 43, there were some outliers comparing the measurements between CRM and BFI waviness measurements. High differences were found in the waviness values of the samples P16, P17 and P33 even though the calculation of the Wa(1-5) values were made by the BFI using the BFI and the CRM measured surface profiles. After analysis, it appears that the basic difference was the measuring direction. The BFI measurements were made as usual perpendicular to the rolling direction. CRM measured in rolling directions as planned in the online measurements. To revise these differences in the waviness values check measurements were made by the BFI with all samples in rolling direction. For these, the number of measurements was increased to 10 profiles for each sample. As shown in Fig. 49, there are some high differences in the Wa(1-5) values calculated from surface profiles measured in and perpendicular to the rolling direction. Wa (BFI; in rolling direction) [µm] P16 P17 P 33 y = x R 2 = Wa (BFI; perpendicular to the rolling direction) [µm] Fig. 49 Comparison of first BFI measurements (inaccurate!) and BFI check measurements 42

45 The differences that were found in the waviness values of the samples P16 and P17 seems to be the difference of the waviness in and perpendicular to the rolling direction. The mean value of Wa(1-5) of sample P 33 was nearly the same in both measuring directions. With regard to these differences the WA(1-5) values calculated by the BFI and CRM measured profiles shows a good correlation with a coefficient of determination of 0.9 (Fig. 50). Wa (CRM measurement - BFI calculation) P33 y = x R 2 = Wa (BFI; perpendicular to the rolling direction) [µm] Fig. 50 Comparison of waviness measurements by CRM and BFI (both calculated by BFI) The difference in the Wa(1-5) value of sample P 33 cannot be defined. If the P33 profile measurement is considered as outlier, a really excellent correlation exists between the BFI and the CRM measurements with a coefficient of determination of 0.97 (Fig. 50). Wa (CRM measurement - BFI calculation) y = x R 2 = Wa (BFI; perpendicular to the rolling direction) [µm] Fig. 51 Comparison of waviness measurements by CRM and BFI without the outliner P 33 (both calculated by BFI) But it seems that the CRM stylus measuring system measures the amplitudes of the profiles about 27% higher than the BFI measuring system. This difference should be checked by measuring measuring standards (e.g. a roughness standard or the CARSTEEL waviness standard). This difference has to be taken into account when values measured by BFI and CRM will be compared. 43

46 Nevertheless it must be stressed, that the BFI and CRM measurements show a really excellent correlation. One important result of these comparisons of measuring results is that the waviness values measured in and perpendicular to the rolling direction may show significant differences Global considerations on measurements Possible draw-backs in Roughness or Waviness computation Globally, when computing the waviness and roughness, it is important to notice specific phenomena that could influence the measurement results. First, the interpolation procedure could modify the slightly the topography at small scales, and therefore influence the Ra computation, in particular if the number of non-measured point is high because of a surface with high slopes for instance. Second, depending on the exact method used, the computation of the Wa could be strongly influenced by the shape extraction procedure, which could modify strongly the behaviour of the surface at large scale. Third, working on deterministic textures such as EBT has proved to provide waviness measurement results that are highly variable with the angle between the profile and the texture orientation. In addition to these major sources of disparities between waviness measurements, it is also important to remember that each measurement system has a specific transfer function applied on the real topographical surface. For instance, a mechanical stylus will in fact measure the original profile dilated with a structuring element having the size of the measurement probe. The transfer function is variable from one system to another, and this could also induce slight variations in the waviness values obtained with different systems. All these considerations have to be kept in mind when trying to compare the parameters obtain through various measurements device and establish a reference waviness measurement on a specific set of samples. Possible solutions Firstly, it is important to assess very carefully the computation method used to extract the waviness parameters (interpolation, shape extraction, Wa computation ) when comparing. Secondly, due to intrinsic acquisition differences between the various devices, a realistic target accuracy coming has to be established to compare the Wavimeter measurements with the waviness of reference samples. The self coherence and reproducibility of Wavimeter measurements (acquisition, computation of the profile, extraction of waviness parameters) will be very important to assess independently from the comparison with other devices. One of the most important problems might be the assessment of deterministic texture waviness. A common practical way of getting rid of the measurement angle dependency is to either use 45 which has proved to give the more stable results, or to average the results on various angles. Unfortunately, it is impossible to apply such solutions with the Wavimeter, as the acquisition direction will always be the rolling direction. However, the background image provides information about the texture orientation and its angle with the laser line projection, so it might be possible to use this background information to correct the obtained measurements using specific calibration techniques. 44

47 Conclusion A list of waviness types with dimensional characterization has been delivered (Deliverable D1). This has led to define the parameters to be applied for the on-line waviness measurement, which are: - 30mm minimum of length of measurement - Low cut-off: 1mm, high cut-off: 5mm - At least a density of 300 points per mm Deeper tests have shown that it seems to be possible to reduce the profile length from 30 up to 20 mm and down to 1000 points with good correlation. Laboratory measurements have highlighted some limitations of interferometric measurements linked to high reflectivity of the surface. Mechanical stylus measurement has been assessed as reference. Laboratory tests have established that: - Due to their possible influence on the results, interpolation for non-measured points, surface texture, have to be taken into account. Moreover, for on-line Wavimeter validation, the computation method and realistic target accuracy have to be very carefully assessed. - The comparison of the laboratory measurements, when all done along rolling direction, has validated the Ra, calculated with a cut-off of 2.5mm, and Wa calculated with a low cut-off of 1mm and a high cut-off of 5mm, on a length of 45mm maximum. - To calculate waviness and separate it from the roughness, Gaussian filtering will be used as this is phase correct. 45

48 Development of waviness sensor (CRM ARsa AM Spain) (WP2) Wavimeter: acquisition part (CRM) (Task 2.1, Task 2.2) General requirements To be able to measure a large number of types of waviness, a length of at least 30mm (CarSteel definition) is required. The stitching algorithm requires an amount of recovery image from image. A recovery of at least 50µm has been determined by image analysis of the database constituted in laboratory (see WP3). Considering a field of view of 850µm, a recovery amount of 50 pixels and a maximum speed of 180m/min on a galvanising line, the frequency of grabbing should be 3.75kHz. In the initial version of on-line microscope, the maximum frequency of the camera and the laser used was 10Hz. Moreover, the images are acquired each time the system is seen as focused inducing that the collected images are not contiguous. So, material has to be changed to fulfil the waviness measurements requirements Camera requirements In the determination of the best camera to be used, not only the frequency of grabbing but also the pixel size and the sensitivity have been considered. The chosen camera has a relatively good sensitivity. It is a CMOS one giving the advantage of no blooming in comparison with CCD. Its full resolution is 1280 pixels by 1024 pixels. It can acquire full image resolution up to 2kHz, and 512 lines up to 4kHz. This last configuration has been used on line Laser requirements In the determination of the best laser to be used, not only the frequency of illuminating pulse but also the level of energy output and the width of the pulse have been considered. The amount of energy used in the current configuration of microscope has been measured, giving so the minimum required for the new laser for a camera having the same sensitivity. The width of the pulse has also been calculated taking into account field of view width and product speed. The chosen laser works at 532nm, as the Topometer one, and has a pulse width of 30ns at a frequency of 3kHz. This laser unit appears to be very critical. Two failures with very long delays of repair have been encountered during the project implying long delays in the sensor development and validation in laboratory and in on-line tests. At the moment, no explanation has been given by the manufacturer. As it is an essential part of the sensor, this unit should be improved or even changed in a final industrial version. It must be précised that, since the last repair, the laser has run satisfactory and seems to be stable. 46

49 Laser test Laser has been tested in term of pulse width, stability and delay. Pulse duration The most frequently used definition for pulse duration is Full Width Half Maximum (FWHM) and will be used here. Since the temporal shape of the pulse is non-gaussian, an 1/e² definition of the pulse has no meaning. It was computed that the energy comprised in the time window of the FWHM definition represents between 60 % and 70 % of the total energy of the pulse. YAG lasers typically exhibit an irregular temporal pulse shape that can be decomposed into a smooth component and an oscillating one. Although the shape of the oscillations can vary from pulse to pulse, its frequency has been found constant over the pulse frequency and energy ranges observed, which facilitates subsequent FWHM computation. To facilitate the computation of the FWHM over irregular pulse shapes, a moving average over one period of the oscillating component was applied to the raw signal. This allows to rather well eliminating the oscillations, and the FWHM duration can easily be computed from the averaged curve. The figure below shows a typical YAG pulse (blue curve), and its decomposition into averaged (red) and oscillating (green) components. (Fig. 52) Pulse stability Fig. 52 Typical YAG pulse To estimate the stability of the pulse in terms of duration and energy, a set of 100 consecutive pulses has been recorded for different values of pulse repetition rate (PRR) and energy (the parameter of influence for the energy is in fact the current). The mean duration and laser energy have been computed for each set, as well as the standard deviation: 47

50 Current - Duration (ns) Pulse Energy Frequency Mean Std dev Mean Std dev 36A - 1kHz A - 1kHz A - 1kHz A - 10kHz A - 10kHz A - 10kHz A - 30kHz A - 30kHz A - 30kHz Fig. 53 Pulse stability data The manufacturer of the laser gives pulse duration of 23ns at 30 khz (current unknown). The pulse energy was computed by integrating the signal of the photodetector over time. This photodetector has not been calibrated for energy or intensity measurement, thus these values represent relative values (Fig. 53). Delay A delay of ns between the trigger (rising edge of the trigger signal) and the pulse has been measured. The uncertainty in the value comes from the fact that the generator used has a pretty long rise time. This data can be used for synchronisation realisation Fibre coupling of laser As the pulse energy of the high frequency laser is much reduced compared to the slow one used in Topometer, the coupling between the laser and the fibre has to be improved in order to reach similar energy levels at the output of the line generating optics. The previous set-up simply consisted of the fibre core being illuminated by the free-space laser beam (Fig. 54.a). This very straightforward arrangement causes a lot of losses by absorption, reflection and scattering on the cladding of the fibre, its connector and its support, since the transmitting area of a single mode fibre is very small (typically 5-10 µm diameter) compared to the other dimensions (cladding 125 µm). The energy available at the fibre input is 0.28 mj/pulse, while the energy at the output of the line generating optics is only 26nJ/pulse, giving a global efficiency of only 0.1 % (-40 db) which is not explained by fibre attenuation. a. Direct illumination b. Focused beam into the core of fibre Fig. 54 Illumination of fibre By comparison, the pulse energy of the high-speed laser at a pulse repetition rate (PRR) of 1 khz is 0.3 mj when the laser is functioning at its maximum power. This means that all of the energy of the laser should be used only to generate the line, without any background, demonstrating the need for an improved coupling of the fibre. 48

51 Improvement of coupling can be reached by using a plano-convex lens to focus the laser beam into the core of the fibre, which reveals to be far more efficient than the direct illumination arrangement. The focal length of the lens has been chosen so that the convergence angle of the focused beam matches the acceptance angle of the fibre. This ensures that only little light propagates through non-transmitting modes of the fibre, and that, essentially, the core of the fibre is illuminated with the focused beam. (Fig. 54.b) This configuration greatly improves the coupling efficiency, since a throughput of more than 2 µj/pulse was obtained while the laser operated at reduced power and using only 10% of the energy of the pulse, as a 90/10 beamsplitter was used in addition. This allows the main part of the energy to be used for other purposes such as e.g. background illumination. The drawback of this system is an increased sensitivity towards the positioning of the focused beam with regard to the fibre core. The illumination angle and the focus distance are also very sensitive parameters. It is, however, relatively easy to find a good set of parameters by trial and errors in a recursive way Control software for laser and camera To be able to integrate the complete control of the laser in the high-speed application written by CRM, new software have been developed and successfully tested. The same work has been done for the camera. Fig. 55 shows the result of background illumination with the chosen high-speed camera and the chosen high speed laser Synchronisation mode of camera and laser Fig. 55 Background test of high speed camera As the laser pulse must be very short to freeze the image of a moving product, it is important to have a good synchronisation of the camera and the laser, in order to be able to correctly catch the pulse during the exposition time of the camera. Otherwise, the camera misses the pulse and the recorded image is not usable (black image). Two modes of synchronisation are available for driving the camera: internal and external. The internal synchronisation uses an internally generated signal, which is adjustable by the user through a software interface. The external synchronisation uses an external signal, which is repeated on the sync out output and can therefore be used to synchronise other devices. 49

52 The internal synchronisation was tested at various frequencies, from 30 Hz to 4 khz, and with variable exposure times ranging from the minimum (1 µs) to the maximum allowed by the acquisition frequency. In every case, the camera caught correctly the laser pulses. In external mode, the synchronisation signal is sent to the camera only and the sync out output is used to drive the laser. In every case, the camera catches correctly the pulse at any frequency and for any exposure time. This is this last mode that has been chosen. It gives higher security in the reach of correct synchronisation as the camera sends a pulse only when it is ready to take an image, i.e. to integrate the light received in the next frame period Synchronisation signal Various possibilities were possible for the synchronisation signal. The frequency of this signal can be fixed or variable in function of speed. If this is fixed, it has to be sufficiently high to keep the recovery amount for the higher speed. This will result in a very high amount of recovery for low speed, resulting so in a high number of images stocked. This implies that, for a line of contiguous images, the recovery can vary highly inducing more calculations to eliminate the extra images. The frequency can be recalculated each time the speed changes to keep the amount of recovery in the same range. So, the number of stocked images will be relatively constant and no extra calculations will be required. On the contrary, this will help the stitching algorithm by reducing the area of research. Calculating continuously in real time this frequency through the PC is not the best choice. It has been replaced by the use of a Voltage Controlled Oscillator associated with an integrated circuit that will combine the speed and two programmable factors: Gain and Offset (Fig. 56). These two last ones can be changed through the PC to modify promptly the amount of recovery. Line Speed Frequency Box Offset Gain IC VCO Grabbing Frequency Software Camera Triggers: Pre-start Start Camera Laser trigger Laser control Fig. 56 Scheme for synchronisation signal Small adaptation has been required to take into account positive offset appearing on output, even with line speed equal to 0. This offset is due to electrical noise. As it has been demonstrated that this noise was stable for every line speed and camera grabbing frequency, the circuit has been adjusted with a small negative offset. This can be eliminated by putting through the software control a positive offset signal. 50

53 AM Spain has studied a second solution. It consists in the use of PLL's, which is a very accurate method of synchronisation. This electronic device allows transforming any pulse signal into another one with a proportional frequency. For example, starting from the encoder signal of the roll where the acquisition system is, which sends a pulse when the roll moves a specific angle, it is possible to obtain a pulse each 500 µm using the relation between the angle of the roll and the movement of the strip. This signal could be used to trigger the camera. After a few milliseconds of the camera trigger and still in its integrating period the laser will be triggered in order to allow the CCD to integrate the line image and the necessary background. The block diagram of this circuit is the one of Fig. 57. fenc & osc /b Camera trigger lp /a l p a = = = = l b i delay Laser trigger li Fig. 57 Scheme for synchronisation signal with PLL's There are several commercial PLL s which permit to configure a and b parameters and could be used in this way. For example MC of MOTOROLA, or TRF3750 of TEXAS INSTRUMENTS. Finally, the version with VCO showed sufficient good precision during industrial tests and was used for all the measurement campaigns Optical mounting for laser The scheme of optical mounting for the laser is shown in Fig. 54. To compact the system, mirrors have been used. A beamsplitter divides the energy in two beams for the fibre for line projection and for the fibre for background illumination. A motorised attenuator has been added in each optical path to be able from the PC to modify independently the intensity of the line and of the background. This mounting can be used in combination with automatic evaluation of line and background levels to modify it on line in function of product surface. 51

54 Fig. 58 Scheme for optical mounting for laser This scheme has been implemented as shown on Fig. 59. Fig. 59 Optical mounting for laser coupling with background and line projection fibres To improve the resistance to vibrations, the translation tables on which fibres are connected have lockable position Acquisition and control unit The structure of the initial program (Topometer) has been modified to integrate the new camera and the new laser. A high speed PC is used to control the whole system. Considering that n images are required to cover the desired length with a defined recovery amount. The camera is used in cycle mode. In a first step, the camera is pre-triggered by the PC. It will acquire images and store them in a cycle buffer of n images size. When a complete cycle of acquisition is done, the first time the focus point is reached, the PC triggers the camera again. This second step is completed when n/2 images are acquired. In that way, the buffer of the camera consists in n/2 images acquired before focus and n/2 images acquired after focus. A priori, this is the best way to obtain a line of images greater focused. The grabbing sequence is controlled by the PC only at three times: the pre-trigger that is not critical in time, the trigger which is more critical and the completion of acquisition which is not critical too. The critical trigger signal is continuously checked by the PC through the acquisition of the distance signal. 52

55 During this step, the other operations in charge of the computer is the speed acquisition for post saving and the check of security signals to take off the system if required. Control Box Computer CPU Line Speed Offset Gain Frequency Box IC VCO Security Axis control Distance control Software Camera Triggers: Pre-trigger Trigger Grabbing Frequency Camera Laser Box Laser & Optics Laser control Laser trigger Fig. 60 General synchronisation Wavimeter Using the high speed camera which is more voluminous and heavier than the camera used in Topometer implied to make a mechanical revamping of the installation. It required to modify the implantation of the distance sensor used for focalisation and synchronisation, but also to use positioning rotation tables able to support higher load. So, most parts of the support and interfacing mechanical elements were redesigned and built in collaboration with Operational engineering department. In result, the Wavimeter head is shown in Fig. 61. Three control boxes are associated to the sensor: - the first one includes the screen and the PC (with acquisition board), all the electrical interfaces, the control unit for the distance sensor and the control unit for the positioning motors of the sensor head. This box is air conditioned as the PC and the control unit for motors are consuming resulting in heat dissipation (see Fig. 62, upper box), - the second one includes the laser power, the laser head and optical mounting for coupling laser to background and line projection fibres, and two serial modules to control the motorised attenuators of the optical mounting (see Fig. 62, lower box), - the third one includes the air control system used to control pneumatic support on which the sensor head is fixed. It will allow to position the sensor near the product when measurement is possible and to take it off when line incidents occur (heat buckles, ) (see Fig. 63) 53

56 Fig. 61 Wavimeter head Fig. 62 PC control box Fig. 63 Pneumatic control box A heating/cooling Peltier system has also been added on the head of Wavimeter taking into account the experience of Topometer during the development of this project. Indeed, for a long time stable running, it has been highlighted that a stable temperature must be maintained inside the measuring head. This is achieved using the Peltier system. The focusing unit used a distance sensor that appeared to give noisy response and even no response on high reflective surface. Study has been made to change it. A new distance sensor has so been implanted in the Wavimeter: it works as well on high reflective surfaces as on diffractive surfaces. 54

57 This change has greatly helped to manage the modification implied by the reflectivity on the strip. Indeed, on fresh (non oxidised) samples, it appeared that the reflectivity was so high that the projected line was for a big part reflected specularly and so not seen by the camera. The solution was to incline the whole sensor. The result is that the optical axis is inclined at 4, the line making still a 45 with this optical axis and the line is more reflected in direction of the camera. The new sensor was placed in such a way their measurements were not impacted by this configuration. This configuration has been used in the first industrial tests. Tests done after laser failure, have demonstrated an influence of the grabbing frequency on the distance measurement due to electrical coupling. So, the electrical circuit (power supply, synchronisation input and output) has been completely isolated from the other parts of the installation solving so this problem Wavimeter: image stitching (ARsa-AM Spain-CRM) (Task 2.3) Stitching requirements Stitching algorithms refers to a set of techniques used to recover a large image from a set of smaller one. It is commonly used for instance to recover a full panorama based on a set of pictures taken with a camera. In this project, the trade-off between resolution and field of view is impossible to handle without using such techniques. As a matter of fact, a very high resolution is necessary to capture the deviation of the laser line with an accuracy compatible with Wa measurements (the order of magnitude for Wa values is between 0.5 and 1 µm). Nevertheless, Wa is a measurement referring to quite large wavelengths (around 1mm and greater), and a long profile (30 mm) is therefore needed to average the waves amplitude with a good statistical representation. Because it is extremely costly and difficult to design a vision system with a high resolution and a large field of view at the same time, it has been decided to focus on the resolution aspect for the hardware part (camera, optical system), and to recover the large field of view property by stitching consecutive recovering images. Fig. 64 Stitching principle Stitching algorithms already exists in various image processing fields. A review of possible algorithms has been made. In the following, the explanations will be more concentrated on the algorithm selected. Indeed, this last one provided results accurate enough for the application with a speed compatible with on-line applications. In spite of the large amount of literature existing on stitching algorithms, it is always necessary to adapt a generic concept to a specific case. In this study, there are specificities that have to be taken into account to optimise the algorithms. In particular, the optical configuration is normally very well known, contrary to most panorama situations, but there could be some distortions due to the laser line that have to be taken into account. Difficulties specific to our system have been highlighted and some solutions have been proposed to finally obtain results whose quality seems to be enough to perform on-line Wa measurements. 55

58 Description of the main algorithms The algorithms have to fulfil specific conditions: - They have to be precise enough to reconstruct an accurate waviness profile - They have to be fast enough for a real time processing The basic principle of stitching algorithms involves usually several steps: - Pre-processing of the images - Estimation of a transformation from one image to another (usually, the transformation is a parametric model whose parameters have to be estimated) - Mixing of the two images based on the estimated transformation to reconstruct a stitched image - Post-processing of the images to obtain higher-level data. The role of the pre-processing step is to correct possible artefacts that would prevent the transformation computation to be accurate enough for the target application. The estimation of the transformation parameters is really the heart of the processing. Of course, the right transformation model has to be chosen depending on the specific features of the acquisition device. In the case of a usual camera to reconstruct panorama for instance, a transformation model involving variation of position, orientation and zoom of the camera has to be taken into account. Fortunately, in the considered case, the acquisition conditions are very well constrained, which makes this step easier. The mixing of two images is a kind of data reconciliation to obtain a final result as accurate as possible. The post-processing in our case is the exploitation of the stitched images to recover the topographical information (profile and Wa). The question of the pre-processing and post-processing will be addressed in specific sections, as it is usually problem dependant. In the following, the choice of a reasonable transformation model and the estimation of its parameters will be detailed. Transformation model In this application, the camera is not moving, but the strip follows a simple translation displacement in front of it. As a consequence, the translation model, which is one of the simplest transformations in the stitching paradigm, seems to be well suited. With more formalism, the following notations could be introduced: - I 0 (x) is the first image value at position x, where x denotes a 2-dimensional vector. - I 1 (y) is the second image at position y. - T(x,p) is a transformation model with parameterized by the vector p that makes the correspondence between a point in I0 referential and a point in I 1 referential. The goal is to find a vector p such as I (.) I1( T (., )) is minimal in a functional sense. 0 p The considered norm is usually a least-square integral over a recovering area. In the discrete world for instance with images of limited domain D: p 0 = arg Min( 0 x D, T ( x, p) D I ( x) I 1 ( T ( x, p)) 2 ) In this case, the transformation T ( x, p) is simply a translation of vector p: T ( x, p) = x + p It is easy to show that minimizing the previous function of p, is in fact equivalent to maximizing the 56

59 cross-product expressed as follow in the case where the transformation is a simple translation of vector p: V ( p) = I ( x). I ( x + p 0 1 ) x D, x+ p D V(p) can be interpreted as the covariance function between a patch of image I 0 and a patch of image I 1. This is easily tractable. However, previous studies have proved that the use of covariance function could be in most cases replaced with improved results by the use of the correlation function. In that case the correlation function could be expressed as follow: ( I ( x) E ).( I ( x + p) E ) W ( p) = 0 x D, x+ p D 0 σ. σ with: - E 0 and E 1 being respectively the average value of image I 0 and image I 1 over the recovering area. - σ 0 and σ 1 being respectively the average value of image I 0 and image I 1 over the recovering area. This slight transformation enables to take into account variation of global illumination from one image to another (due for instance to flash non-homogeneity). In the specific case of correlation function, it is obvious that the expression of W(p) as a convolution product provides an easy way to maximize this function over a reasonable range of value p. The translation model is therefore directly associated to the maximization of the correlation function between a patch of image I 0 and a patch of image I 1, but of course, the same methodology could be used with another more sophisticated transformation model. The function to minimize would be different, and the minimization techniques presented in the following paragraph would be different as well. However, because the simple translation model have given results that are to be good enough for our application, it seems unnecessary to investigate more complicated and more costly transformation functions. Estimation of the parameters Now that the transformation model has been fixed, it is important to provide an efficient way to compute the best parameter p for the transformation model (in our case, p is simply a translation vector). One of the ways to perform the computation is to perform a sliding average of a patch P1 of image I 1 with a kernel corresponding to a patch P0 of image I 0, with the normalizing constant carefully computed. (Fig. 65) It is important to notice that the size of the patches have to be chosen given an estimation of the recovering area (which is not exactly known), the realistic range of vector p and the time available for computing: the larger are the patches, the more precise is the p vector estimation, but the longer is the computation

60 Patch P0 Patch P1 W(p) * = Max over p Fig. 65 Sliding average of a patch P1 with a patch P0 As commonly stated, performing the convolution with a given kernel is equivalent to performing a multiplication in the Fourier space: 1 f * g( t) = F ( Ff Fg)( t) Where: Ff designs the usual Fourier transform (or FFT in the discrete case). The patch P0 of image I 0 can be normalized with average E 0 and standard deviationσ 0. However, for each position of the convolution kernel over patch P1, the value of E 1 and σ 1 have to be computed to match with the recovering area whose size is the size of patch P0. Fortunately, the computation of the local average and standard deviation of image I 1 can easily be performed by replacing the patch P0 of image I 0 by a window function whose value is 1 over the patch P0 support and 0 everywhere else. I has to be replaced by ( ) 2 To compute the standard deviation, 1 I 1. The following python code (Fig. 66 ) to compute the correlation function W(p) should provide enough details of the computation without entering into mathematical details. Once W(p) has been computed for a reasonable range of vector p, the maximum over this range can be computed to get the vector p maximising the correlation. This provides the translation p of the transformation model and gives enough information to superpose at best the two images I 0 and I 1. 58

61 Fig. 66 Source code in Python for correlation function computation 59

62 Reconstruction of a stitched image Fig. 67 Data fusion principle Once the recovering area has been computed accurately through the translation vector p, it is possible to reconstruct the information on the overlapping area using all information at disposal. In our case, the average of grey level of image I 0 and I 1 has been computed in the recovering area. In summary, the stitching method we are going to use is based on: - A simple transformation model which is a translation of vector p. No rotation or change in zoom is considered. This is realistic because the optical configuration of the acquisition device is very well controlled. - The vector p of this transformation is estimated between two consecutive image using a correlation technique and FFT algorithms. - The fusion of data on the recovering area is done by averaging the two images. - The process can be easily iterated to stitch more several images. In the following sections, a few problems specific to the optical and mechanical configuration under study will be highlighted and slight adjustment to this basic algorithm will be adopted to overtake these limitations Specific problems When applying this algorithm to real successive images coming from the microscope, several specific features can be observed. Specular caustics Another element that could limit the stitching accuracy is the occurrence of bright spots due to the specular reflexion of the laser line. The position of these spots depends on the exact direction from which the laser line is projected on the surface. These spots can therefore influence the correlation function between two consecutive images. As a consequence of the possible misalignment and also of the specular spots, it is important to use background sub-windows that do not contain the laser line to evaluate the translation vector p. Slight misalignments The most noticeable feature is the orientation of the laser line compared to the strip direction. Ideally, the laser line would be exactly projected in the direction of the strip movement, so that acquiring several successive images would be equivalent to project a longer laser line on a larger field-of-view. However, this alignment is difficult to set-up perfectly without a specific control device. On the first experimental acquisition, a slight misalignment is almost each time visible. 60

63 Error! Fig. 68 Slight misalignment in line This phenomenon is a problem in itself, because it means that the reconstructed profile will not be fully continuous. But before the profile reconstruction step, this misalignment can induce artefacts in the computation of the vector p. Because the laser line is linked to the field of view, it does not follow the same translation as the background image. The images of Fig. 69 coming from real data got in laboratory illustrate this problem. Fig. 69 Slight misalignment in line (examples) Each of these figures (Fig. 69, left and right) is the stitching of the same two consecutive images. On the left figure, the whole image has been used to compute the vector p, whereas on the right figure, only the background of the top part of the image has been used. It is clear that the computed vector is not the same in both cases: on the left, the vector p is such that the laser line (in particular the quite bright tip in the middle of the image) from the first image is superposed to the laser line of the second image. On the opposite, on the right figure, the correlation between the background of the first image and the background of the second image is maximised: the misalignment of the laser line becomes visible. In one acquisition, a systematic computation has shown to compute a misalignment of 0.3 in average. To align perfectly the system, the camera and the projected line are adjustable. For the camera, concerning its orientation in the plane of focus, no automatic system is implemented. The camera is oriented by the use of thin plates placed on the adequate side between the camera and its support. The adjustment is optimised when the transverse stitching displacement (orthogonal to rolling direction) is near 0 pixels. The line direction is adjusted through the use of the rotation table (precision: ) on which the line projection optics is fixed combined to an FFT analysis of the resulting stitched image. 61

64 Indeed, if the line is not perfectly in the rolling direction, it will result in the stitched image in a profile with steps appearing at the length of stitching (which is constant as the grabbing frequency is directly linked to the speed product). These steps will appear in the FFT profile of both Wa and Ra profiles. Orientating the line in the rolling direction will result in the extinction of the concerned peak. (Fig. 70) Light source homogeneity 1 degree tilt 0 degree tilt Fig. 70 Spectrum analysis for line position adjustment The homogeneity of the source can also play a role in the stitching process. In fact, like the laser line, the background is linked to the referential of the image and not to the strip movement, so a point of the strip that is seen two consecutive images can be lighted with a different intensity in the first and the second image. However, because the spatial frequency of the source non-homogeneity is usually low, the artefact is compensated by using the normalized correlation version of the algorithms: removing the average value and dividing by the standard deviation for the patches compensate the changes in source intensity if the lighting is homogeneous on the patches used to compute the vector p. In the first trials, the spatial flash variation as well as the intensity variation of the flash from one image to another was clearly visible on the images: averaging several consecutive images gives an idea of the flash aspect (Fig. 71). In the successive improvements of the sensor, this has been corrected and in the current version, the homogeneity of the light source is quite good. Average of 100 images with the first version of Average of 100 images with the improved version of illumination system: the lighting non-homogeneity is illumination system: the lighting non-homogeneity is clearly visible greatly reduced Fig. 71 Light source homogeneity Moreover, as explained later in this document, in the last configuration of the sensor (inclined at x ), the illumination is no more done through the optics. Indeed, when illuminating reflective surfaces with non specular light, the background is very dark resulting in high inaccuracy in stitching. It is the reason why the illumination is now realised obliquely to the optical axis. 62

65 Size of matching window The suitable choice of the initial parameters in the search of the displacement between matching and displaced windows is a very important factor in order to reduce the processing time. For example, the size of the matching window is critical, seeing that the number of Standardized Crossed Correlation and comparisons needed is directly affected. In this area, several improvements were attained. In Fig. 72, two windows sizes are shown; the left plane illustrates the window size used in previous trials whilst the right one indicates the window size used in the last software version which being much smaller provides shorter computing time. The selection of the window size is only possible after several trials and an exhaustive study working with the particular images conditions, so it may vary depending on the different implementations Fig. 72 Different windows size With this smaller window size, the system can be sure than the maximum value will fit on it, but it must take into account another important parameter: the initial displacement of the displaced window. This displacement is calculated into an iterative process looking the previous values of the displacement, this is a very interesting improvement, which permits to reduce considerably the time needed. Next values are different parameters for the EBT texture: Rows=1220 Cols=1616 Horizontal displacement matching=20 Vertical displacement matching=20 DHIni Matching=-1343 DVIni Matching=1 Origin window column=1344 Origin window row=1 Rows of the matching window=464 Columns of the matching window=100 63

66 Results of stitching process The previous algorithms in addition to the choice of the right parameters (size of patches, range of admissible translations, normalization) and specific tuning (sub-pixel interpolation) have enabled us to obtain good results in the stitching process in a reasonable amount of time C++ implementation and optimisation Fig. 73 Result of the stitching of two consecutive images A fast algorithm implementation for computation of the displacement vector between two subsequent images, and the generation of the stitched image made of a set of adjacent images, has been done by CRM. It is based on the FFT-based CCN algorithm proposed by ARsa. It has been implemented as a C++ class for easy integration in other software pieces. The developed software has the following features: - Self-contained optimised implementation for the displacement vector computation (CCN), based on a statistically pre-computed initial guess vector, to reduce the search area. - Dedicated low-level load and save functions for bitmap images allow for significantly faster stitching execution, as these steps are bottlenecks. This is important as the load function is used repeatedly during the stitching process and the save function has to deal with very large files (> 100 Mb). The latter takes advantage of some bitmap format features to even improve saving time when possible. - Use of an efficient (open-source) library: fftw3, for the computation of the many Fourier transforms computations required by the algorithm. The library guarantees an O(N log N) complexity for arbitrary transforms size, even for prime size transforms, by the selection of the most approriate FFT algorithm. This is very interesting as the size of the data (depending on parameters such as the search area size) to be transformed is always an odd number. Also, the computation of single-sided transforms yields an improvement by a factor of 2 in the execution time. - Dynamically (and repeatedly) allocating large memory blocks during execution is quite time consuming. For this reason, a static memory implementation was developed: the required (large) memory blocks are allocated at the beginning of the execution, in such a way that they are reused by subsequent calls to the functions. It has the drawback of needing to know in advance the size of the images and of the stitched image, which is not really a problem (the size of the stitched image can easily be estimated). More generally, a very careful memory management is one of the characteristics of this piece of software. Similarly, attention was paid to avoiding repeated computation of the same thing: for example, if one given transform should be computed for every stitching computation, it is made only the first time and the result is saved for subsequent uses. Many parameters are pre-computed in this way. Dynamic construction of the stitched image, with possibility of recording only a region of interest (ROI) around the line, can increase execution speed and reduce memory requirements. 64

67 The quick merge version takes advantage of the physical data arrangement in memory to allow the stitched image to grow without needing to re-copy it each time. This gives a very good execution time dependence on the number of input images since it is better than linear. The drawback of the method is that the image is stored vertically. This can be corrected when saving to hard drive or by postprocessing but doing so is quite time-consuming for large images. Nevertheless, this vertical orientation is not a problem for the subsequent steps of the waviness computation Real time system The main objective has been to reduce the processing time in order to achieve a real time system, which permits to process all the images captured in each sample (around 20 images). The algorithm was optimized and several tests have been done looking for reaching the suitable values of parameters like initial horizontal and vertical displacements, this way the size of the displaced window is reduced and therefore the processing time. The simplification of the application allowed removing unnecessary program parts, making a faster application. This code has been tested with different series of images obtaining good results and it is quick enough to process all the images in real time. The modifications made are: - Automatic recognition of the minimum window size after the processing of the first images - Approaching of the matching starting point to the minimum possible in an automatic way - Early recognition of the maximum CCN to stop the search in the whole window. - Living the images in memory, saving only stitched images. - Possibility of work with different parameters depending on the texture. The average duration is 70 msec/matching, which allows to process 70 images in 5 sec Sub-pixel precision In a first approach, the displacements were calculated in pixels, but it could be calculated with a very high accuracy using sub-pixel methods. 3 different methods were considered to calculate where the maximum correlation is within the displaced window. Cubic spline In this case, polynomial interpolation has been applied using cubic spline. 5 points around the maximum and 2 decimal precision has been used. The two more distant points are used to get the boundary conditions in the polynomial curve. More than 5 points do not provide better results but instead increases the computational cost. The Fig. 73 shows a 3D view of the correlation surfaces. 65

68 Fig. 74 Correlation surface for the first correspondence of the image group with displacement of 550 µm. Fig. 82 represents the correspondence values for the horizontal displacements with which it is obtained a maximum value of vertical displacement Fig. 75 Correspondence curve for the horizontal displacement value in which the maximum vertical displacement value is found Finally, Fig. 76 represents the interpolated values obtained in the maximum neighbourhood. Notice that to obtain the maximum with a subpixel precision of 0.01 pixels it is needed to calculate 200 points around the maximum, whilst to obtain a precision of 0,1 pixels 20 points would be enough. It is therefore necessary to study if the computational cost carried out by working with two decimal precision is worthwhile compared with the one decimal calculations. Equally, to obtain a subpixel precision in the two axes in a correct way there should be used bi-lineal or bi-cubical interpolation. 66

69 CCN DV Fig. 76 Neighbourhood to the maximum for the previous curve. In blue discrete values and in red interpolated Parabolic method using 3 points The cubic spline method achieves a very high accuracy, but it has the inconvenient of a huge computational cost, seeing that it should calculate 200 interpolations to obtain a precision of 0.01 pixels. In this system, time is critical and this method is non-viable, that is why a simpler method is more interesting. The main aim in all these methods is to obtain the position of the maximum but not its value. We are not interested in Y-value, only in the X-position. This consideration reduces a lot the processing time of the method because the position of the maximum in a parabolic approximation is known if 3 values of the contour are known. It is not necessary to calculate the value in each point and reconstruct the parabola. Fig. 77 Position of the maximum in a parabolic approximation The position of the point M (Fig. 77) is easily calculated with the expression: Where a, b, c are the values of the correlation in the maximum pixel (b) and its neighbours (a,c). This method has a good accuracy with a very low computational cost. 67

70 Linear method using 5 points Another interesting method based in looking for the x-position not for the Y-value, is this one, which considers the information of 5 points around the pixel with the maximum value and obtains the subpixel position. The linear method, illustrated in Fig. 78, calculates the straight lines that pass over the neighbours and where they cross themselves defines where the maximum is. Fig. 78 Linear approximation compared with cubic spline Where the red lines cross themselves defines where the maximum is with a very high precision. Green line is the cubic spline approximation and it s clear the similitude between the both methods. Trials made in Matlab reveal that its accuracy is similar than accuracy reached with cubic spline evaluated each 0.01 pixel. The computational cost is very low because the calculus of the slopes and the maximum are simple trigonometric relations. Conclusions of using sub-pixel precision During the trials using polynomial interpolation, it appeared that the computational cost is almost double than the other two methods since to obtain a 0.01 pixel precision are required 200 polynomial evaluations. The other 2 methods do not affect significantly the processing time. In the table of Fig. 79 are represented the comparison of all the methods in the stitching of 8 pictures. 68

71 Matching\Time(sec) No Subpixel Cubic Spline Parabolic Linear Fig. 79 Comparison of different sub-pixel techniques With all of these methods, it is possible to obtain a great precision searching the displacement, but to reconstruct the entire image, sub-pixel techniques involves another important disadvantage seeing that to join one image with the next one, is needed to resize the original images into a more larger ones. To reach a 0.01 pixel precision is needed to make the original images 100 x 100 times larger. This process would take a long time and it is possible to do not be worthwhile. Besides, the final reconstructed image would be very large, and further processing would be more expensive. In conclusion, sub-pixel processing do not improve image stitching in an important way meanwhile it would cause some disadvantages. This option would only be suitable if the list of displacements were required with a very high accuracy. It would not be suitable only to obtain the reconstruction of the sequence and the location of the laser line Wavimeter: stitched image analysis (CRM) (Task 2.3) Having got a complete stitched image, the line present in the image has to be isolated for further analysis. The algorithms initially developed for Topometer had to be adapted to the higher size of image but also to the presence of potential slope of line in the stitched image. Indeed, if the camera is not perfectly aligned with the rolling direction, the stitching process is not influenced by this inclination but the result appears as the line would be inclined in the resulting image (in reality, this is the resulting image which is inclined). Specific techniques based on linear regression and filtering on standard deviation have been implemented Wavimeter: profile analysis (CRM) (Task 2.3) When the line is found, a raw profile is available for further calculations. Having this profile, roughness and waviness parameters can be extracted. To calculate waviness and separate it from the roughness, filtering the profile is required. In general, the Gaussian filter is to be preferred as this is phase correct i.e. it does not introduce phase distortion into the transmitted wave. Various implementations of this filter have been implemented and tested to optimise the time processing. A first implementation is the Krystek one that is considered as a fast and reliable convolution algorithm to calculate the mean line of a roughness profile (see [1]). The second implementation is an Order n filter that consists in a simplified realization for the Gaussian filter in surface metrology (see [2]). 69

72 For both implementations, the data to supply is: the profile, the number of pints and the cut-off value. In both cases, at least, the result profile is shorter of a length equal to cut-off value (half of the cut-off value at both sides). To calculate waviness, the initial profile is first filtered with the low cut-off value giving so let say ProfileLow. Then, the initial profile is filtered with the high cut-off value giving so ProfileHigh. The waviness profile is then got by the subtraction of ProfileHigh from ProfileLow. Then, the waviness is got by the following formula: 1 n W a = n i= 1 Z i ( x i ). xi where: n is the number of points in the profile Z i(x i) is the difference between the height of the profile at the point i and the average of the profile x i is the step along the profile at the point i Fig. 80 shows the comparison of the result of Krystek and Order 32 filters in function of Gaussian filter on one profile. The Gaussian filter result has been got by using two times (one for each cut-off) the Gaussian filter implemented in Taylor-Hobson software and by subtracting the resulting profiles. The upper curves refer to left axis Y. The three profiles seems to be perfectly superposed. Looking closer at the difference between Krystek profile and Gaussian profile, and between Order 32 profile and Gaussian profile, the lower curves displayed are got. These are referred to the right axis Y. So, the differences are limited to +/-0.01µm that results in a variation of 2% in the Wa value which is acceptable. From the point of view of timing, the Order32 filter takes around 80ms to process points with a low cut-off of 2.5mm and a high cut-off of 8mm in comparison of ~12 seconds for Gaussian filter and 9 seconds for Krystek filter. Fig. 80 Filters comparison on one profile Fig. 81 shows the comparison between Wa values got using Krystek filter, Order8 filter and Order32 filter on the CRM mechanical profiles got on the round robin samples. The Krystek filter has been taken as reference. The variations in Wa results is only of 2%. The correlation and the slope are very good: nearly 1 for both. 70

73 Fig. 81 Filters comparison on round robin samples So, Order n filter will be used for laboratory and on-line measurements Wavimeter: profile analysis (BFI) (Task 2.3) BFI had added investigations in the development of waviness sensor by a detailed study of other various filters that can be applied on the profiles got on the industrial samples in laboratory. To meet these demands BFI studied the influence of different filter types, measuring lengths and evaluation lengths to the waviness value of steel sheet surfaces. Especially the effects of a draft of a new ISO (ISO/DTS ) were investigated. This could provide methods of treading end effects of linear profile filters. End effects means unintentional changes in the filtration response in the end portion of an open profile. This can be done by modifying the filter equation. These modified filters with an automatic correction of end effects are called new Gaussian filter and new spline filter. The conclusion was that the Order32 filter is finally better than the new ones in term of precision even if the new ones require a shorter length of measurement Wavimeter: time cycle (CRM) (Task 2.3) Due essentially to the transfer of images from camera to PC, the acquisition cycle is around 20 seconds. Beside this, the image processing takes presently around 25 seconds. The first point could be decreased using the new GigE products having a higher transfer rate or even, a camera including processor that could potentially allow at least part of processing inside the camera. By the way, only reduced data would have to be transferred reducing considerably this operation. On the side of image processing, as it is well defined today, possibility to reduce some parts which are time consuming could be to use specifically designed electronic circuit. In an industrial version of the Wavimeter, both aspects should be deeper investigated. 71

74 Validation on industrial samples in laboratory (CRM-ARsa-BFI) (WP3) Validation on static samples (Task 3.1) Validation of the method The validation principle on static samples consists in: - displacing the sample in the line direction to simulate on line moving, - taking an image after each displacement, - stitching all the grabbed images, - analysing the resulting image to find the line and get the profile, - filtering the profile with adequate cut-offs, - calculating the Ra and Wa, - comparing the got values with the measurements given by classical off line laboratory methods Mounting and conditions of test Topometer has been used for optical measurements. Each sample is mounted on a system composed by 2 translation tables and 1 rotation table (Fig. 82). This allows: - to place the sample surface perfectly in a plane orthogonal to the optical axis, - to move the sample in the direction of measurement, - to move the sample to measure on parallel lines to compare results obtained at different positions. View of Topometer and sample b. View of translation and rotation tables of sample support Fig. 82 Laboratory mounting for optical measurements Some difficulties have been encountered due to the fact that the samples are not flat at a scale higher than waviness. This induces that even when positioned orthogonal to the optical axis, the sample does not stay in focus from the beginning to the end of the travel. Fixing the sample on each border on a flat support has resulted in lower variations. This problem will not been encountered on plant as the strip will be under traction and as the observed surface will be well applied on a roll. The positioning of the sample is done first by use of vertical translation table and rotation table. The 72

75 focus is done in such a way the longest length of scanning is in focus. Then, the sample is placed at the start point and is moved along the measurement direction, an image being grabbed for each step, without any further focusing. The total length of measurement is 120mm. Each image has 600µm in the direction of measurement. Three types of step have been applied: 600µm, 550µm and 500µm, corresponding to a recovery image from image of 0µm, 50µm and 100µm. 3 to 5 lines of measurement have been done for each condition, each spaced of 5mm. Two directions of measurements have been used: one in the rolling direction (the one that will be applied on line) and the second in the transverse direction (for laboratory study). So, a large range of waviness types can be covered. The result is a huge database of images that has been used to develop and validate the image processing whose aim is to collate automatically the images grabbed on-line Solved problems First results have highlighted problems of positioning of the samples. First of all, some images were black due to a lack of synchronisation between laser and camera. This has been solved. The second problem appeared due to a misalignment of the line. To get a perfectly connected profile, it is required that the projected line is nearly perfectly in the rolling direction (simulated by a translation table in laboratory). Initially, it was not the case. This is presently corrected by precise visual inspection. An automatic method will be studied during next period. The third problem appeared due to movement of the translation table around its translation axis. This appeared in the profile as relatively low variations to higher level in the analysed profile, followed by a flat mean level, followed again by low variations to lower level, then a flat mean level and so on. (Fig. 83) Fig. 83 Level variation in profile due to translation table This has been solved by the adaptation of mechanical mounting and by the use of another table. 73

76 Results The profiles of a series of samples have then been reconstructed (stitching, line detection, filtering). The images have been acquired with a recovery of 50µm image to image. Fig. 84 shows the type of profile got on the sample with the Wavimeter method. The Wavimeter values of Ra and Wa are very near to the mechanical measurements. Fig. 84 Sample of roughness and waviness profiles The global results for Ra calculated with a cut-off of 2.5mm are shown in Fig. 85. Fig. 85 Optical Ra in function of BFI Ra and CRM Ra The correlation is good between optical Ra (measured with Wavimeter method) and CRM mechanical Ra. It is less good between optical Ra and BFI Ra, but it is due to outlier (for BFI17 sample as already shown in the round robin test). It is better when the outliers are suppressed (Fig. 86). 74

77 Fig. 86 Optical Ra in function of BFI Ra and CRM Ra without outliers The global results for Wa calculated with a low cut-off of 1mm and a high cut-off of 5mm are shown in Fig. 85. The correlation is good between optical Wa (measured with Wavimeter method) and CRM mechanical Wa, even with the outlier BFI17. It is less good between optical Wa and BFI Wa, but, again, it is due to outliers (for BFI17 sample as already shown in the round robin test). Fig. 87 Optical Wa in function of BFI Wa and CRM Wa It is better when the outliers are suppressed, the correlation is at 0.86 and the slope at 0.98 for comparison with Wa CRM and the correlation is at 0.91 and the slope at 0.92 for comparison with Wa BFI. (Fig. 88) 75

78 Fig. 88 Optical Wa in function of BFI Wa and CRM Wa (without outliers) Except for the outliers that have to be deeper inspected, these results validates the method in laboratory, as well for the image stitching as for the image processing and for profile filtering. The precision is good (mainly in +/-15%) and the correlation with mechanical stylus measurement also. The dispersion can probably be decreased by using an inclined configuration that will be less influenced by the reflectivity of the surface. After verification, as shown earlier in this report, it appeared that BFI first measurements had been done in direction perpendicular to the rolling direction. After making new measurements, the outliers disappeared Validation on dynamic samples (Task 3.2) Mounting and conditions of test The dynamic tests have been prepared through the realisation of a rotating system including a wheel on which a strip sample of wheel circumference length has been fixed. This wheel can rotate at a user defined speed given by a controlled motor. The aim was to allow the validation of stability of recovery areas and the ability of the Wavimeter to follow dynamically the sample. The mounting is shown in Fig. 89. Fig. 89 Laboratory mounting for dynamic tests 76

79 Tests After the laser came back from the second repair, first tests were done. It appeared that the height variations were far too high (more than 1mm!) and far from the one that could be encountered on line. This could not give good results as on the measurement length, the height variation will be too high resulting in blurred images (for recall, the focus depth is around 20µm). Trials to decrease these variations failed. Due to the remaining time in the project, it was decided to go on line without investigating more this point. Indeed, reproduce the line conditions appeared to be more difficult than expected, in the point of view of sample stability (traction) and speed (reach 3m/sec) in stable conditions. Anyway, these tests validated the synchronisation and helped to improve the electrical connections and electronic circuit to reach: - the best frequency grabbing frequency in term of overlapping between images and time delay to speed variation, - the linearity: frequency directly proportional to speed, - stability: frequency variation less than 1% Investigation in the principle possibilities to measure waviness with the Raumed measuring system Another instrument has been tested by BFI for its possible application on-line. The Raumed measuring system normally is used to measure the roughness profiles of steel strips online. For this a laser beam is projected to the surface and the reflected beam will be changed after passing a special optical system into an electrical signal which corresponds to the distance between the sensor and the locally surface point. (Fig. 90) Fig. 90 Principle of the Raumed measurement For the investigation whether this measuring system is appropriate to measure the surface profile in the accuracy to calculate the waviness value Wca, a measuring system was installed in a laboratory beside a tactile and a laser-optical roughness measuring system. (Fig. 91) 77

80 Fig. 91 In the laboratory installed Raumed measuring system Next, the output signal of the Raumed was adapted to the needed input signal of the usual measurement device used by the BFI. After this the surface of a cocked flat sample was measured with Raumed and a tactile measuring system. By those measured data the Raumed was gauged. (Fig. 92) Taktil Raumed Messwert [µm] Weg [mm] Fig. 92 Measured profiles of a cocked flat sample Then the surface profiles of steel sheets samples with different topographies were measured with both the tactile measurement and the Raumed measuring system. 78

81 EDT1 EDT2 EDT3 EBTstoch. EBTdet. Pretex1 Pretex2 Edelstahl Fig. 93 For the measurement comparison measured surfaces The Wca values were calculated by the tactile and Raumed measured profiles. The interrelation between the calculated data is shown in the diagram of Fig. 94 along with the according ± standard deviations. 0.9 Wa(1-5) Raumed measurement [µm] y = 1.26x R 2 = Wa(1-5) tactile measurement [µm] Fig. 94 Comparison between the Wa waviness values calculated by the profiles measured tactile and optical with the Raumed measuring system without outliers Taking away outliers, there is a good correlation between the calculated Wa values and the mechanical ones. Therefore, the Raumed measurement could be tried in the future for on-line trials test to evaluate and compare its precision as online waviness measuring system, after comparison with Wavimeter on a same series of samples in laboratory Conclusion Wavimeter sensor has been validated in laboratory (D3 deliverable). Tests have highlighted that careful attention must be paid to synchronisation with line speed, illumination homogeneity and calculation optimisation. All these points have been integrated in D2 deliverable. The method has been validated in term of precision, mainly in +/-15%, by comparison with mechanical stylus reference, correlation above Dynamic tests have, in particular, validated the synchronisation with line speed 79

82 Integrated on-line strip surface waviness measurements on galvanising line at Segal (Segal-CRM-BFI) (WP4) Delay in industrial tests The technical problem encountered during the setup of the sensor as well the economical situation partially in 2008 and 2009 have delayed some tasks. A suspension of the contract has been asked and granted by the European Commission from 1/12/2008 to 30/06/2009. Anyway, the mentioned conditions have shortened the time let for the industrial testing and have not allowed covering a large range of conditions as initially expected Mechanical mounting (Task 4.1) A roll has been selected in Segal plant for the first measurement campaigns of the Wavimeter. It is a roll located in an easily accessible area, after the skin pass beside an accumulator. The mechanical support has been realised following the principles given by CRM. This support places the optical axis of the Wavimeter in such a way it passes through the axis of the roll (Error! Reference source not found.). In that way, the optical axis is perpendicular to the surface assuring that the image is bet focussed. Moreover, the support is fixed on a beam attached to the structure supporting the roll with the aim of following the vibrations of the roll and so being always at the same distance to the roll On-line pre-test (Task 4.1) Fig. 95 Mechanical implantation principle Prior to installing the Wavimeter, it was desirable to evaluate the distance variations between the rotating roll and a fixed point with respect to the roll frame, to check the possibility of having a sufficiently long set of well-focussed images for Waviness calculation. Vibrations, roll eccentricity or roll finishing defects can cause such distance variations. In order to measure this, a distance sensor was installed in front of the roll, mounted on a specially built stand, in such a way that the sensors optical axis crossed the rotation axis of the roll. Doing so, the actual orthogonal distance between roll surface and sensor is measured. Two different distance sensors were used for the purpose of the trial, both being optical sensors but having different measurement principles: - Triangulation sensor - Chromatic confocal sensor 80

83 Fig. 96 Pre-test at Segal Results The first measured distance variations showed apparently random, large amplitude, at a much higher frequency than the expected roll rotation speed. This could let one suspect sensor malfunctioning due to sheet reflectivity. However, new measurements using the other sensor showed comparable behaviour. So, as they are identical, only the results obtained with the triangulation sensor are presented in the following. The analysis of the data revealed a permanent oscillation of the distance at a frequency close to 11 Hz (660 rpm), with very variable amplitude over time, superimposed to a slower variation, with constant amplitude and variable frequency. The latter is the consequence of the rotation of the roll, as a repeatable pattern appears that corresponds to the roll profile over one turn. The amplitude of the distance variations due to roll imperfections is the parameter of interest of these trials, and was found to be about 80 µm. This is sufficiently small to allow for waviness measurements as it was previously computed that 250µm is still acceptable (leading to 20µm of variation of height on 30cm of measurement). On the other hand, the amplitude of the 11 Hz oscillations that would be caused by structure vibrations, is extremely variable and ranges from a few tens to 350 µm, or even 600 µm in the extreme case. These excessive variations seem to be caused by the speed of the production line, that is accessible through the rate of variation of the roll profile. The relation between line speed and vibrations amplitude is in fact roughly exponential. This is illustrated by the Error! Reference source not found. and Error! Reference source not found.: in blue the raw signal as delivered by the sensor, and in red the filtered signal, free of vibrations contributions, with the roll profile clearly visible. They show the speed variations of the production line, with the corresponding amplitudes of the raw signal (NB: the scales are different from figure to figure). 81

84 Fig. 97 Distance record 1 Error! Reference source not found. shows relatively low speed, followed by a stop and a restarting. The slow and comparatively large variations probably correspond to sheet traction variations. Fig. 98 Distance record 2 Error! Reference source not found. shows speed change (the «glitches» in the red curve are caused by acquisition problems). Although there is a clear link between line speed and amplitude of the vibrations, it is probable that other (undetermined) factors are affecting it. Some amplitude variations were observed where no speed change occurred. Also, some shocks, that is large amplitude vibrations during a short time, could be seen at more or less constant intervals and could correspond to some plant events. However, the line speed seems to be the most determining factor. It must be pointed out that only the amplitude of the vibrations is modified by the line speed, while the frequency remains unchanged. This last observation lets one think to a resonance effect of the mechanical support of the sensor (or of the roll frame itself). 82

85 Conclusion on pre-test The distance variation over one roll turn (80µm) is not a problem for the waviness measurement, while the vibrations could be. However, due to the higher weight of the actual waviness measurement system and due to the fact that it will be positioned on a shorter arm, the amplitude of the vibrations was supposed to be smaller and the resonance frequency should be modified. Some damping system in the support frame could also help in this regard. The conclusion of this test was that one can expect that working on the selected roll might be feasible, and will be tried On-line installation (Task 4.2) Electrical, mechanical and software preparations In collaboration with Segal plant, electrical connections, mechanical parts and communication software have been realised at the end of 2007 and during First position The Wavimeter has been installed at the pre-test position in October But, unhappily, the vibrations were too high with a too high frequency. This induced a variation of more than 200µm height on a measurement length of 50mm. These vibrations coming from the roll and its structure were transmitted and amplified by the beam support to the sensor. As the focus depth is only 20µm, these 200µm were of course inacceptable. A first test of improvement was to rigidify the base support of the sensor head itself. It decreased the amplitude of vibrations to 150µm but increase the frequency, so the situation was worse. A second test was to change completely the beam support by reducing its size to the centre of the structure of the roll structure. This resulted in a decrease of the frequency of vibrations and a slight decrease of the amplitude. However, the variation on 50mm length was still too high (around 90µm). So it was decided to change the measurement position Second position The second position is just after the skin-pass (around 30m). The disadvantage is that the measurable side is the bottom side of the strip but for a validation step it is acceptable. The big advantage was that it existed already a base beam support fixed on the floor and no more on the roll structure. So, a mechanical adaptation was realised quickly in CRM to be placed on the beam and support the sensor head. The Fig. 99 shows this installation. There, the vibrations were far lower (only around 15µm along 50mm of measurement. This place was so validated for some test weeks. In parallel, AM Spain was contacted to modify its support taking into account the tests done at Segal. Al the electrical connections had to be re-done. This task 4.2. has taken longer time due to the vibrations encountered at the first chosen position. But the very positive result is that another place was determined and this was relatively free of vibrations. 83

86 Fig. 99 Wavimeter at Segal Adjustment of the sensor The first thing to do when the sensor is installed is to adjust its global position. This is done through the use of the distance sensor signal and the positioning tables supporting the core of the sensor. This signal is minimal when the optical axis is orthogonal to the roll surface. To facilitate the next adjustments of the sensor on line some algorithms have been written. Fig. 100 Dedicated algorithms for adjustments 84

87 The first part of this software concerns stitching. A focus parameter is calculated on 6 regions of each image: upper left corner, upper mid part, upper right corner, lower left corner, lower mid part, lower right corner. The term upper means above the projected line, and lower below the line. The average on all the base images is also calculated for each zone. This helps to evaluate if the system is well in focus. The parameters: longitudinal displacement, transverse displacement and correlation factor, concerning the stitching between 2 successive images are displayed in graphs. The longitudinal displacement is along rolling direction, the transverse one is perpendicular to rolling direction. For recall, the line is projected along rolling direction. This helps to evaluate the quality of the stitching. If variations in displacement are present, it means either a bad original search region for stitching, but it can also be a problem of focus or if quality of background image. If transverse displacement is too big, it highlights a big inclination of the camera. This should be corrected by small plates between one camera side and its support. The second part concerns profile analysis and so, also stitching. The line profile is searched by image processing in the stitched image and then analysed to get the Ra and the Wa profiles. Then, as already explained, the spectrums of these profiles are calculated to check if there is or not a peak at the length of stitching. For recall, the presence of this peak is significant of a misalignment of the projected line Validation on samples (Task 4.2) In order to validate the measurements done by the online waviness sensor some samples of each coil were taken and measured manually. These samples were taken from the head and tail of the coils, and the measurements were done in the same side and the same place as the waviness sensor. To grab a maximum of images and have so a maximum of data, as the grabbing cycle takes around 20 seconds (due to the transmission of images from the camera to the PC essentially), the image processing was not enabled on line. Indeed, the image processing itself takes around 25 seconds, so, it has been applied after grabbing First campaign of tests A first campaign of measurements was started at Segal in December It appears quickly that even an angle of 4 (tested in laboratory) was too low to get a sufficient amount of the line through the camera. The tests to go further in tilting let appear a high loose of quality in the background image. This problem is due to the high reflectivity of the strip and did not appear on laboratory samples that were no more so reflective ( old material) Sensor modifications The result of this first campaign was the modification of the implantation of the camera inside the core of the sensor. It implied also modifications in other support plates. The system was so modified to be able to be at 8 with the lighting of the background not coming from the microscope optics but from external, and so being more in a brightfield configuration than in a darkfield configuration that does not fit with high reflective products. These modifications were done just before leaving to AM Spain for a second campaign of tests. They were tested during one hour in Segal to check that the grabbed image was correct before leaving to Spain. 85

88 Validation on specific coil (Task 4.3) A series of coils have been followed during the campaign. However, due to the adjustments required on the sensor, no specific condition has been correlated to Wa data. All the experience got on Segal line has allowed to realise successful tests in AM Spain (see further for details) Database constitution and study of interactions parameters at Segal (Segal-CRM) (WP5) Database constitution (Task 5.1) A database has been constituted with the data from the line Database analysis (Task 5.2) A series of coils have been followed during the campaigns. However, the production parameters variations applied during this period has not shown impact on waviness values. These last ones remain in the same range and do not seem to follow any rule. But, as explained for the AM Spain tests, it appeared clearly during the on-line tests that samples coming from middle of coil are required for validation by mechanical measurements. (see and ) Fig. 101 Mechanical Wa during variation of production parameters Last developments in production has highlighted that the analysis should be done with specific tests related, for example, to the variation of skin-pass force and traction, change of roll type. Due to the relatively short time let to conduct industrial trials, not all operating conditions have been checked also because they were not planned in the plant schedule overloaded at that time. It should be very interesting to continue analysis of such elements combined with Wavimeter on-line measurements. 86

89 Integrated on-line strip surface waviness measurement on galvanization line at ArcelorMittal Spain (AM Spain-CRM-ARsa-BFI) (WP6) Installation on-line with mechanical and electrical adaptation of plant, and communication with plant control system (AMSpain, CRM, ARsa) (Task 6.1) Once the sensor was improved, some tests were in the HDGL #2 of ArcelorMittal Spain, during 3 weeks by people from CRM, ARSA and Arcelor Mittal Spain. The tests were finished on end of January The device was installed in front of a roll that is placed 165 m after the skin-pass. In this place, there is a roughness measurement system developed in a previous RFCS research project (TOPOMETER) and the good results obtained with that device were a decisive argument to decide the location of the trials. In order to integrate the prototype in the line properly, some specifications were defined by CRM according the device requirements and the previous trials in SEGAL: - The optical axis of the device should pass through the centre of the roll in front of which it would be placed - the sensor should be placed horizontally on a beam fixed on ground - The support should not have its own vibration - A protection of the installation should also be planned. - The maximum distance between head and control cabinets should be12m. Following this specifications, a new support was designed where it could be possible to install both gages.(see Fig. 102) Fig. 102 Wavimeter support scheme at AM Spain As below the device, there is not any steady ground the supporting beam had to be designed quite long and this fact caused a deformation of the beam when the device was installed due to its weight but it did not affect the measurements and the trials results. In the other hand, 2 lines of 220 V and 16 A were installed as power supply for the device. Additionally some other electrical signals were provided following CRM specifications (Fig. 103) 87

90 Description Type Analog signals Speed at sensor place 4-20mA Digital signals Skin opening Volt-free contact Ga / Gi Normally closed contact for Gi Welding passing Volt-free contact Fig. 103 Electrical signals used There is also a safety signal that indicates if the strip has broken and it allows to move away the device using the pneumatic cylinders of the device and to avoid a collision. In order to storage the measurement made and to receive some information, a communication protocol was implemented between the device and the process computer. It is a TCP/IP protocol with these characteristics: - Each connection is a client/server connection (TCP/IP), mode TCP (not UDP). - For a bi-directional link between 2 partners, there is one logical connection. - Each connection is built up by the client application. - The server application is capable of accepting several incoming connections; for the communication with Wavimeter, 2 process computers can connect to the Wavimeter-server. - Each message has a fixed header with the following layout: struct PamSockProtocolHeader { char headeridentification[6]; unsigned long datalength; unsigned long sequencenumber; }; HeaderIdentification:is always the string $PAM$ zero terminated. DataLength: is the length of the message data including the header. SequenceNumber:starts at one (after establishing the link) and increments by one each message sent. This number can be used to test for lost messages. - Server and client programs must be able to regroup split messages. - Every protocol-error (e.g. $PAM$ not found in header) must result in the termination of the link on both server and client-side. The client must retry (automatically) to re-establish the link. - The Wavimeter system is the server application, the process computer is the client. - The fixed header is followed by the application part of the message; this part starts with a message-number (type int16) which identifies the message, followed by the application data. Through this protocol the devices sends the roughness, peak and waviness values and it receives the production time and the coil ID number. All the information is stored in the process database with the rest of information of the line that was used to validate the measurements. The pictures of Fig. 104 show the installation at AM Spain, with the control cabinets, the Wavimeter head and the rubber cushion used to avoid vibrations in support beam, and finally the mechanical stylus measurements for validation on samples. 88

91 Control cabinet Rubber cushion to avoid vibrations Wavimeter in position at AM Spain Stylus measurement by CRM Fig. 104 Wavimeter installation and checking unit on line at AM Spain Validation of measurements on head and tails (AM Spain, CRM, BFI) (Task 6.2) In order to validate the measurements done by the online waviness sensor, some samples of each coil were taken and measured manually. These samples were taken from the head and tail of the coils, and the measurements were done in the same side and the same place as the waviness sensor Second campaign of tests The second campaign of tests for Wavimeter started in AM Spain with the modified configuration (detailed previously). It has quickly appeared that the AM Spain product is more reflective than Segal product. The result was a bad quality image even in 8 degree inclination with a very low fill factor of the line. After some study on site, a second mechanical modification was decided. This modification results in a final inclination of the optical axis at 16 related to the orthogonal to the surface. The images quality is far better and the fill factor too (Fig. 105). Even with the small focus depth, the resulting image does not appear blurred. This solution improves the line quality but results in a small loss of precision in the measurement due to inclination. It is a compromise between a good fill factor and a better precision. 89

92 a/ Before second mechanical modification b/ After second mechanical modification Fig. 105 Improvement of image quality (FOV width: 850µm) Nevertheless, image processing had to be slightly adapted to take into account the inclination and the high product reflectivity Tail samples During the days of measurement, samples were taken at the end of coils. These samples have been measured by CRM with TalySurf instrument used for the validation in laboratory. Most of the measurements have been done on plant, the rest in CRM laboratory. Some measurements have also been doubled to check the repeatability and the validity. Some of the samples, covering a priori a large range of Wa values have been sent to BFI for the validation of mechanical measurements. Fig. 106 Comparison between Wa mechanical measurements Fig. 106 shows the comparison between BFI measurements and TalySurf measurements. The X axis corresponds to TalySurf measurements (Wa TS) made by CRM. Three profiles, measured on 45mm on each sample, in the rolling direction, are then processed by Gaussian filter Order n method (detailed earlier in this report) and averaged to get the final Wa and Ra values. BFI measurements (Wa BFI) consist in 10 profiles in the rolling direction on 30mm each filtered by a method delivered with the instrument. TalySurf profiles have also been processed by BFI algorithm (Wa TS by BFI), taking the middle 30mm from the 45mm measured.. Wa BFI and Wa TS by BFI are represented on the Y axis. All the processes have used low cut-off of 1mm and high cut-off of 5mm. Correlation is very good between CRM and BFI filtering algorithms (correlation between Wa TS and Wa TS by BFI). 90