Defect detection in partially complete SAW and TIG welds using on-line radioscopy and image processing

Transcription

1 Defect detection in partially complete SAW and TIG welds using on-line radioscopy and image processing G. Bonser and S. W. Lawson Mechatronic Systems and Robotics Research Group, School of Mechanical and Materials Engineering, University of Surrey, Guildford, Surrey, GU 5XH, United Kingdom. ABSTRACT An application of machine vision applied to the analysis of radioscopic images of incomplete weld geometries is described. The rationale of the work is to identify weld defects as soon as they are produced, thereby reducing the costs of any subsequent repairs. Existing methods of weld and defect identification are compared, leading to the development of filtering and 'window' based variance operator for segmentation of suspect defect areas inside the weld region is described. The software and radioscopic imaging system have been benchmarked through a series of demonstration trials on both 8mm thick carbon steel submerged arc welded (SAW) testpieces, and 5mm thick carbon steel tungsten inert gas (TIG) welded testpieces. The range of intentionally implanted defects, from root s to lack of side wall fusion, were detected with an overall accuracy of 87%, and classified in terms of defect size, shape, and position within the weld region. Keywords: Radioscopy, NDT, image segmentation, incomplete welds, x-ray, defect detection, defect classification.. INTRODUCTION The introduction of digital imaging has brought about a wide range of improvements to radiographic non destructive testing (NDT). These advances include calibration of sensor wear, assisted operator interpretation using image enhancement, and cost effective digital image archiving. Image integration methods are now routinely used in practice to remove image noise, whilst simple edge enhancements and grey look up tables can also be used to aid defect detection in poor contrast images. In addition, to these image enhancement techniques a number of researchers have addressed the possibilities of semi, or even fully, automated inspection using image analysis and pattern recognition techniques. The work described here has been based on a recent European Commission (EC) funded project entitled NDT Methods for Flaw Detection During Welding which has been directed at the development and evaluation of techniques for the detection of defects during the welding process. The problems associated with undertaking NDT during welding include high temperatures (which can affect SNR and other data acquisition parameters), incomplete weld geometries and large local electro-magnetic fields. In addition any defect information must be made available almost immediately so that any errors in the welding parameters can be corrected in an on-line manner. It is for this latter reason that automated image processing methods have been developed for flaw detection and evaluation. Work has concentrated on the use of both ultrasonic Time of Flight Diffraction (TOFD) and x-ray radioscopic methods for flaw detection. The development of image processing for automatic TOFD is described elsewhere in these proceedings (paper number ).

2 Section of this paper describes a filtering approach used to detect flaws in radioscopic images, whilst Section 3 describes the approach adopted for defect characterisation. Section 4 describes the results of automatic defect detection on images generated in the final demonstration trials of the EC funded project. Overall defect detection rates are given in Section 5, along with general conclusions to the work.. IMAGE PROCESSING FOR DEFECT DETECTION Defects are characterised in a radiographic picture by local changes in x-ray collection density which translate to transitions in grey pixel intensity, although the magnitude of such changes may vary greatly from defect to defect, as is illustrated in Figure. The figure shows an example radioscopic image of a weld containing three s - none of which are immediately visible from the original grey picture. When the contrast of the image is enhanced using histogram equalisation (Sonka, Hlavac and Boyle, 993) the s become visible to the human eye - however the central has a mean grey very close to the background value. It is the goal of defect detection operators to locate such changes, whether slight or severe, and to label defect pixels accordingly. A B grey 63 A (a) (b) (c) B Figure : Variation of defect intensities in radioscopic images. (a) raw image of weld with three vertical s, varying in severity. (b) histogram equalised version of (a) with s labelled. (c) intensity profile along line AB in original image, showing s and have large deflections from the background whilst is virtually indistinguishable from background noise. Note that an automatic defect detection scheme would be required to locate all three s. One of the fastest and most powerful local area techniques is a convolution, using a one or two dimensional kernel and standard convolution 'filters', such as edge accentuators and high pass operators, are already used in practice as aids to defect detection by human operators. However, no single filter is suited to the automatic detection of all different flaw types, since variations in defect shape, grey intensity, edge attributes, orientation, and position in the component make it very difficult to produce a generic defect model. Both Inoue and Sakai (985) and Wallingford et al (99) describe the use of matched filters for flaw enhancement and detection in weld and simulated images respectively. The method is based on the optimisation of signal to noise ratios (SNR's) given the expected two dimensional signature of a particular flaw. The technique may be implemented as a convolution operator and therefore can easily be applied at real time speeds. However the results presented show that although the technique can be used to locate almost invisible flaws from noisy backgrounds, further processing is required to accurately segment the defect region. Furthermore the technique relies on the use of different filters for each flaw shape and orientation that may occur. It may be desirable, on the other hand, to have a limited number of filters, each geared to detect a particular group of defects, whose outputs may be optimally combined to produce the correct segmentation. A method has

3 been developed (Lawson, 996) which uses two filters as a pre-processing stage for defect detection in radioscopic images of welds. Two specific filter masks were chosen from a number of standard filters by experimental means rather than through formal or iterative methods. The filters used are the Laws "E5S5" (a combination of Laws' spot, E5, and edge, S5, one dimensional detector masks) and the horizontal Kirsch filter. The integer coefficients of the (5x5) variations of these filters are given in Figure. The Laws mask, was selected for its ability to accentuate 'spot' or 'blob' type defects (such as porosities and cavities) and the Kirsch filter since it was more suited to highlighting longitudinal defect types (such as s, lack of penetration and slag inclusions), and was therefore complementary in performance (a) (b) Figure : 5x5 filter masks for defect detection in radioscopic images (a) Laws E5S5 filter; (b) horizontal Kirsch filter. These two masks may be applied, separately, as local linear property extractors to every pixel f(x,y) in an image to yield a new pixel value f'(x,y) given by :- where M is the convolution mask, or N /... Equ. () f'(x,y) = M(k,l). f(x, y) k,l= N / f'(x,y) = N / {f mn}xy. {Mmn}... Equ. () m,n= N / where the set {Mmn} is the coefficients of the convolution mask M, and {f mn}xy is the set of members of grey values contained in the local population within the rectangular (Nx x Ny) box centered at position (x, y) within the image, and m = (x -(Nx/)... y+(nx/)) and n = (y-(ny/)... y +(Ny/)). Although the two filters described above will accentuate any defects present in an image they will not actually segment them from the image background. To achieve this a post-processing method for both radioscopic and ultrasonic TOFD data has been developed which is 'tuneable' to the SNR characteristics which are present in the NDT data - and therefore the characteristics which are present in a sequence of images acquired at a particular temperature and with a fixed image intensifier gain setting. The method uses the standard deviation of a local area population to determine whether or not a pixel is of interest. The standard deviation, σ, of a local area population given by :- σ = N xn NxN (i n - i) n=... Equ. (3)

4 where i is the intensity of a pixel, i is the mean intensity of the local population and N is the side length of the square population. The standard deviation, σ, may then be thresholded to detect areas of high variance (likely to be defects or component echoes) or low variance (likely to be background). The operator can be applied to images which have been pre-processed with either the Laws or Kirsch convolution filter. (a) T var =.5 T var =. (b) T var =.7 T var =.5 (c) T var =. T var =. (d) T var =. T var =.4 (e) T var =.3 T var =.5 Figure 3: Effect of varying thresholds Tvar and Tvar when applying standard deviation, σ, operator to image filtered using Laws kernel. The image contains a cavity type defect. The effect of using a dual threshold approach to the segmentation of an example radioscopic image is given in Figure 3 above. The image has been pre-processed with the Laws filter and the local grey scale standard deviation thresholded at two s, Tvar and Tvar. 3. DEFECT CHARACTERISATION Defect characterisation as such is not achieved by the automated defect detection algorithms. Instead pattern analysis is used to extract a number of features or descriptors from a detected defect. These descriptors are then used to concisely describe the defect. Ideally this interpretation would produce a complete description of each defect - including its whereabouts, severity and, most critically, its type. However, in practice this has proven very difficult to achieve given the complexities and subtleties of the subjective manual inspection procedure employed by a skilled radiographer. A computerised interpretation is however able to give an objective description of each defect including its shape, dimensions and location - often these may be combined to give an indication of the actual defect type. The procedure adopted extracts six descriptors from a defect detected image denoted d... d6, where:- (d) (d) (d3) (d4) (d5) (d6) normalised length/width. elongation (perimeter*perimeter/area). mean grey scale value of defective area (extracted from raw image). mean grey scale value of defect/mean grey scale value of weld area. normalised Euclidean distance from weld edge. extent of defect (area). Each defect in a defect detected image may analysed in such a way as to produce a feature vector of the six descriptors. Defects are labelled -N in the defect detected image using a blob colouring technique (Ballard & Brown, 98). An example result of the characterisation is shown in Figure 4.

5 DEFECT 'E' in 5% SAW weld ~~~~~~~~~~~~~~~~~~~~~~~~~~ Number of defects to be classified = ******************************************************* Defect() length: 87 width: len/wid: 7.5 area: 74 coords: 73,47 perim/area meangrey: 9 rel grey:.9766 centrity:.44 class: large, longitudinal, air defect location: at weld centre ******************************************************* Figure 4: Example of defect analysis in a SAW testpiece. Note that any defect with a density less than that of the surrounding weld (determined by descriptor d4) is labelled as an 'air' defect. The slag defect in the example shown is hence described as large and longitudinal, of low density and positioned at the centre of the weld - this description could easily also be of a root or a lack of penetration. Hence in this case the description cannot exactly classify the defect type. 4. RESULTS OF THE DEMONSTRATION TRIALS During the final demonstration trials three TIG testpieces (TIG A-R, TIG B-R, & TIG C-R) and two SAW testpieces (SAW A-R & SAW C-R) were fabricated with intentionally placed defects. Break up of the intentional defects plus additional unintended defects produced slightly different and extra defects. The radioscopic information was gathered on-line at approximately 5, 5, 85 and % weld completion for the TIG testpieces, whilst at only 8 and 3% completion of the SAW testpieces due to the small power of the available x-ray source. The testpieces were analysed using conventional NDT methods (both radiographic and ultrasonic) against which the effectiveness of the automated radioscopic defect detection algorithms have been adjudged. The ITS Gammascope system For the duration of the project the partner responsible for the radioscopic data acquisition, Isotopen Technik Dr. Sauerwein (ITS), have used a single data acquisition device to capture digital radioscopic images for subsequent processing. This is the Gammascope device which is a fully integrated real time radiography system with interfaces between the x-ray set and the component manipulator all being controlled by a DEC host computer. The system also features digital image processing capabilities, including real time image integration. The images generated by the Gammascope are of size 5x5 by 8 bits and cover approximately mm of weld. The ITS system acquires images at normal video frame-rate but performs frame averaging by an integration process to reduce noise effects. The integration operation requires that the radioscopy device must remain stationary relative to the weld for 5 seconds whilst gathering a single image.

6 After acquisition each image was processed by the automated algorithms described above to locate any defects. The effectiveness of the automatic defect identification process has been measured using the rating system shown in Table for both the visibility of the defect and the of associated false alarms. The full effectiveness of the algorithms was quantified by a comparison using full destructive testing of the testpieces, which was performed by the Institut de Soudure (the French Welding Institute). rating visibilty of defect detected well detected but distorted and/or fragmented only partially detected but visible very poorly detected/obscured by noise not detected false alarm none very few isolated or insignificant pixels significant false alarms high of false alarms obscuring defects unacceptably high Table : matic defect detection and false alarm ratings used in the assessment of the radioscopic images. For each of the radioscopic images, an indication of the success of the automatic defect detection for each defect in the image on a scale of -5 was generated, as per Table () above. In addition a further figure was given (also summarised in Table ) stating an indication of the of false alarms detected by the algorithms. Analysis of the TIG testpieces The TIG welded testpieces contained a total of 5 defects of 6 types. The defects are abbreviated as follows: LOFS - lack of side-wall fusion P - porosity W - weld inclusion RC - root LOP - lack of penetration LOIRF - lack of inter-run fusion. Three sets of m long testpieces were manufactured, denoted sample TIG A-R, TIG B-R, and TIG C-R. Defects were intentionally positioned along the length of each weld at various stages of weld completion. Ultrasonic TOFD was performed at approx.:,, 46, 85 and % weld completion. From the conventional NDT of the welds it was summarised that the defects present in the TIG testpieces were of the intended type and in the correct location. However, extra unintentional defects were also present and the conventional pulse echo ultrasonics fails to detect several defects (detected by radiography). Figure 5 shows the results of the automated signal processing software for the detection of defects in the TIG welded testpieces.

7 TIG A TIG B TIG C % % % % % % LOFS % Root % % Clustered porosity Root + LOP % LOFS + porosity LOF + porosity % - - LOFS % - Porosity - LOIRF LOP % Figure 5: The results of the radioscopic signal processing software on the TIG images. For the TIG testpieces the signal processing software identified the majority of defects except for root s and very small unintentional weld inclusions. Analysis of the SAW testpieces The SA welded testpieces were meant to contain a total of 5 defects of 6 types. The defects are abbreviated as follows: LOFS - lack of side-wall fusion P - porosity SI - slag inclusion RC - root LOI - lack of root (inter-run) fusion LOIRF - lack of inter-run fusion. Two sets of m long testpieces were manufactured, SAW A-R and SAW C-R. Defects were intentionally positioned along the length of each weld at various stages of weld completion. Radioscopy was performed at various stages of completion for each testpiece, at: 8 and 3% weld completion. From the manual conventional NDT of the welds it can be summarised that the defects present in the SAW testpieces are of the intentioned type and were at the intended location. However, extra unintentional defects were again present and conventional ultrasonic testing failed to detect several defects (detected by radiography). Figure 6 shows the results of the automated signal processing software for the detection of defects in the SAW welded testpieces.

8 SAW A SAW C 8% 3% 8% 3% LOFS Root Clustered porosity Root + LOP LOFS + porosity LOF + porosity Slag inclusions LOFS Porosity LOIRF LOP 8% % % 8% 8% 8% % % % 5% 5% Figure 6: The results of the radioscopic signal processing software on the SAW images. 5. CONCLUSIONS The detection of false alarms, i.e. the automated detection of a defect which is in fact not present was found to be quite low, typically with a few very small isolated groups of pixels which could be ignored due to their small size. The worst results were produced on the lower s of completion for the TIG welds where the pronounced rippling of the surface of the weld could produce marked contrasts on the radioscopic image. The defects which were not automatically detected were usually very small in area with root s proving to be the most difficult flaw to detect. Figures 7 shows the overall effectiveness of the radioscopic signal processing for defect detection for both TIG and SAW testpieces.

9 Percentage completion No of defects detected manually No of defects detected automatically Percentage detected automatically Percentage completion No of defects detected manually No of defects detected automatically Percentage detected automatically % % 8% % 4 87% 3% % 7 8% % 4 9 9% (a) Figure 7: Overall effectiveness on (a) TIG radioscopic images, (b) SAW radioscopic images. The high degree of success that has been achieved demonstrates that on-line weld inspection is possible for incomplete weld geometries. 6. ACKNOWLEDGEMENTS The authors wish to acknowledge the funding of this work by the Commission of the European Communities under the BRITE-EURAM II (Industrial Materials and Technologies) shared cost project scheme (project number BRE-39). In addition the authors wish to thank the partners of the project, the Institut de Soudure, Nordon and CIE, Isotopen Technik Dr Sauerwein GmbH, and Mitsui-Babcock Energy Ltd. (b) 7. REFERENCES Wallingford, R.M., Siwek, E.M. and Gray, J.N., "Application of two-dimensional matched filters to x-ray radiographic flaw detection and enhancement", Review of Progress in Quantative Non-destructive Evaluation vol. ; eds. D.O. Thomson and D.E. Chimenti, Plenum, New York, 99, pp Sonka, M., Hlavac, V., and Boyle, R., Image Processing, Analysis and Machine Vision, Pibl. By Chapman and Hall, London, 993. Lawson, S.W., matic Defect Detection in Industrial Radioscopic and Ultrasonic Images, Ph.D. Thesis, University of Surrey, April, 996. Inoue, K. and Sakai, M., mation of inspection for weld, Trans. Of Japanese Welding Research Institute, Osaka University, vol. 4(), pp , 985. Ballard, D.H. and Brown, C.M., Computer Vision, Publ. By Prentice-Hall, 98. Further author information: G. Bonser. g.bonser@surrey.ac.uk; Tel: S. Lawson. s.lawson@surrey.ac.uk; Tel: Web site: