Comparative study between laser vibrometer and accelerometer measurements for mechanical fault detection of electric motors

Transcription

1 Comparative study between laser vibrometer and accelerometer measurements for mechanical fault detection of electric motors R.M.Rodriguez*, C.Cristalli*, N.Paone* *1 *4jA s.r.1., Angeli di Rosora (AN), ITALY * * University of Ancona, ITALY ABSTRACT This paper presents a comparative study between accelerometer and laser vibrometer measurements aimed at on-line quality control carried out on universal motors of washing machines, which exhibit defects localized mainly in the bearings. The test is designed so to study measurement problems of an on-line station. Bearing defects include faults in the cage, in the rolling element and in the outer and inner ring. A set of no defective and defective motors were analyzed by means of the acceleration signal provided by the accelerometer, and the displacement and velocity signals given by a single point laser vibrometer. Advantages and disadvantages of both absolute and relative sensors and of contact and non-contact instrumentation will be exposed taking into account the applicability in real on line quality control measurements and bringing to evidence the related measurement problems due to the specific environmental conditions of assembly lines and sensor installation constraints. Keywords: Quality control, vibration measurements, laser vibrometer, accelerometer. 1. INTRODUCTION It is well known that vibration tests allow the discrimination between good and faulty products through the analysis of vibration signatures1'2' and therefore vibration signals can be used for quality control on production lines. In fact, a variety of manufacturing defects show normally by increased noise and vibration. The typical test stations employ advanced techniques of Pattern Recognition integrated into a complete system of data acquisition and signal processing. A fundamental step in the design of the test station is the selection of the appropriate transducer. Vibrations have been traditionally measured by means of accelerometers, although non-contact methods appear as a possible and attracting alternative, because of the potential simplification of the test station and of the equipment dedicated to place the sensors on the motors. Quality control in production processes of universal motors has become of primary importance because of the size of the market of household appliances. As a matter of fact, nowadays universal motors are widely employed in thousands of electrical household appliances all over the world, and the demands of standards of quality in the final product are very high. To reduce the number of faulty samples at the end of the production line is the scope of many manufacturers who have adopted quality control as a key part of the production process, also from an economical point of view. State of art quality control strategies based on statistical tests on a random selection of samples from the lines, cannot guarantee the quality check of 100% of the production and therefore, they cannot be considered as an appropriate method to assure a high standard of quality control. In most of the cases, if 100% on-line quality control of electric motors is performed nowadays it is done by an operator that listens to the motor and detects the existence of any malfunctions; he operates in an acoustic isolated cabin and the motor is driven during a specific test cycle. The capacity of the operator to discriminate the different types of defects cannot be easily surpassed, however, one of the disadvantages lies on the lack of reproducibility due to the subjective component of the 1 * r.rodriguez(l1occioni.com; c.cristinalchloccioni.com; ABA, sri. Via Fiume, 16; Angeli di Rosora (AN) Italy ** nico1a(i?mehp1.unian.it; Department of Mechanics, University of Ancona; Via Brecce Bianche, Ancona, Italia Fifth International Conference on Vibration Measurements by Laser Techniques: Advances and Applications, Enrico Primo Tomasini, Editor, Proceedings of SPIE Vol (2002) 2002 SPIE X/02/$

2 listener whose attention can vary along the daily working period. This is the reason why an automatic system based on the measurement of vibrations that guarantees the control of 100% of the production is more and more required. Regarding electric motors, quality control tests are not only addressed to verify the electrical functionality and efficiency, but should include also the vibrational aspect4 in order to detect and classify possible mechanical manufacturing defects, e.g. unbalance in the shaft, or faults in bearings or brushes. Thus, a complete and profound test of the final products is an essential requirement to obtain the necessary information that permits an adequate control of the production process. 2. MEASUREMENT PROBLEMS IN ON-LINE INSTALLATIONS Vibration tests are critical and complex tests to be carried out in a production line due mainly to: - intrinsiccomplexity of the vibration phenomena in assembled systems, - noisyenvironmental conditions in the production line, - short time available, usually the test must be performed in a few seconds. An automatic system should fulfill the following requirements: - good repeatability - reliability - flexibility, adaptability to different types of motors and defects - capacity of being integrated in the production line together with the performance of other tests. The existing methods are normally based on the use of accelerometers that have demonstrated a good performance in production processes, but installation may pose several problems. Microphones are often used to analyze noise, but require isolation from other sound sources. Other measurement techniques, like the laser vibrometer, offer the advantage of being a non contact method avoiding the intrusive character of the accelerometer. The single point laser vibrometer is, in fact, being employed nowadays to obtain fast, accurate on-line testing of products during the manufacturing process of an increasing range of appliances and devices. This paper analyses the information provided by the signals acquired with an accelerometer and a laser vibrometer that measure contemporary the vibrations of an electric motor. The measurement process consists basically in accelerating the motor, to keep it rotating at constant speed for several seconds, and finally to stop it. The vibration signals are acquired and analyzed in order to reveal anomalous vibrations of the motors and to classify them as good or faulty. During the measurements the motor is standing on a typical pallet for assembly lines, with its axis in the vertical position. Motors have two bearings, one in the lower part of the rotor and the other one in the upper part. In order to meet specifications for on-line operation, the tests have been performed with the motors placed on a line transport pallet. This situation poses a variety of problems to sensor installation. In fact, motor position varies, because no exact geometric reference is available. Furthermore during run-up the motor moves on its support; this will affect sensor performance. In particular, the laser displacement measurements will require a large range to avoid saturation, but this will reduce signal amplitude during measurement; signal drop-out will occur as well, due to the speckle noise caused by the laser spot motion on the rough surface. On the other side, also the accelerometer will not operate in ideal conditions. First of all, large noise from the electric brushes and the windings of the motor will decrease signal to noise ratio. Also the mechanical coupling will be loose, because the accelerometer has to be installed on a small curved surface, the bearing housing, which does not allow for fixing devices such as tapes or magnets; therefore it has been chosen a probe for hand-held acceleration measurements, whose tip is pressed on the bearing housing by a spring, through a prismatic guide. It is well known that this kind of mounting affects sensor bandwidth, which reduces to about 3kHz; nevertheless, for quality control measurements, sensor use outside its operating bandwidth is still possible if tests have a comparative character and are not absolute measurements, even if signal amplitude is distorted. This solution is a constraint imposed by the on-line installation; it offers advantages for automatic sensor mounting on the bearing seats and reduces installation shocks which may decrease operating life of the accelerometer, therefore it has been chosen. On the other side, the non-contact nature of the laser vibrometer fully eliminates these constraints, so it overcomes the related limits. Nevertheless, beam positioning depends on motor position in the same way of accelerometer mounting, but is sufficient depth of focus is available, measurements on the bearing seats can be easily taken. The main limit of the laser vibrometer is the requirement of isolation from floor vibrations, which can be easily achieved by mechanical filtering in the frequency of interest. 522 Proc. SPIE Vol. 4827

3 3. INSTRUMENTATION ANDMETHOD All measurements were taken on both bearings in radial direction. In every measurement, both the accelerometer and the vibrometer signals were acquired simultaneously. In order to focus both sensors in the same point it is necessary to locate the laser vibrometer with an angle of 30 degrees with respect to radial direction of the bearing. The transducer employed was a Precision Quartz Shear ICP accelerometer, with a sensitivity of mv/g and a frequency range between 1 and Hz, reduced by the installation on the probe tip, as discussed. The single point laser vibrometeris hold by a tripod, is isolated from the vibrations of the pallet by rubber elements and is located at about 60 cm from the motor. The controller demodulates the vibration signal from the optical head and provides an analog voltage proportional to the velocity or displacement of the surface. The sensitivity was fixed to 5 mm s'jv for the velocity and 8 mm/v for the measurements of displacement. The maximum signal bandwidth is 250 khz, and the maximum velocity is 1.6 m/s. The amplified analog signal from the accelerometer and the vibrometer is then acquired by means of a 16 bit acquisition board. Data is sampled in the time domain with a sampling frequency of Hz that provides a useful frequency range from 0 to Hz. A LabVIEW program, especially developed for this purpose, activates and controls both the functioning of the motor and the acquisition board, and processes the signals. The duration of the measurement depends on the type of motor, and in particular on the time the motor needs to reach the stationary state and on the duration of steady state. The transient phase comprises the run up of the motor from 0 to a prefixed threshold of rpm, when the motor has reached the maximum speed. This phase lasts around 4 s in a typical test. After this transient phase, the motor runs in stationary state for about 6 s, which means a total time of 10 s required to acquire data during both phases. Then the motor stops. The measurements were carried out on a total of 21 motors. 6 of these motors had no defects. The damage in the rest of the bearings was caused intentionally by means of shocks in the outer surface and introducing abrasive dust, producing the following defects: - defectin the cage - defectin the rolling element - defectin the outer ring - defectin the inner ring. The same signal processing techniques, including Fourier analysis, order analysis and Short Time Fourier Transform (STFT), were applied to the three signals. 4. RESULTS 4.1. Time signals In the time domain, the localized defects in bearings are shown by a series of impulses that occur with a certain cadence. This impulse nature of the vibration of rolling bearings generates high frequency components. Signal are recorded in the time domain simultaneously. Qualitatively they can be described as follows. The displacement signals reflected a strong influence of the large movements of the motor when it is started and during the run up state, as it can be seen in the following figures, during the first seconds. These movements are not repeatable and depend on motor position on the pallet. Laser vibrometer signal drop-out is observed as a consequence and determines noise spikes. During the tests, several acceleration ramps where employed to bring the motor to the stationary state, displacement amplitudes were more relevant using the faster ramps. Due to the large movements, in order to avoid saturation the AD converter is set to a large full scale range, therefore signal resolution is decreased; this is a limiting factor in the measurement of low amplitude steady state vibrations., which makes displacement signals poor. Proc. SPIE Vol

4 4,OE-5 3OE-5-- 2,QE-5 E 1OE-5-3,4E-21 I' -2.OE-0-3.OE-5 i:8910 ::.3.5E-18 2.OE-2- -1OE-2- -2,OE-2-1 _ ry Time (s) r ,0:.IJ1I - 25,0. -so.o ; Figure 1.- Displacement, velocity and acceleration time signals for a good motor Time Os) 4.OE-5 3.OE-5 a :' 1.OE E-21 a i Tine (s).- -1.OE-0- ' -2.OE OE-5-2.OE OE-2- C.. 3.5E-18-2.OE-2 - I Time (s) Velocity signals measured by the laser vibrometer appear more uniform in amplitude, so that an appropriate range can be selected. Some eventual peaks have been found also in some velocity measurements due to speckle noise. Acceleration signals from the accelerometer appear uniform in amplitude during most of the signal, so an appropriate measurement range can be set. Some noise peaks can be found in the accelerometer signal. In this case, it is due to an incorrect contact between the probe tip and the bearing surface, together with electric disturbances; it all results in an important source of noise. A first approach to a faulty bearing can be made through measurements of RMS values of the time signal during steady state operation, which represents a good alarm indicator. By comparing all signals of figures 1 and 2, correlation between defect and signal RMS value is easily observed for velocity and acceleration signals. Nevertheless, RMS does not provide any mean to classify the type of defect, therefore it can only be used as indicator in a pass-or-fail test. It is important to look at other features of the signal for type of defect classification. In fact, a further understanding of the defects existing in the motors requires a more profound study of the signals in the frequency domain both in the stationary and in the transient state by means of different signal processing techniques, aimed at feature extraction which correlate to the defects1''9. Once a certain number of features, that are representative of the behavior of the motor, have been selected from the different techniques, they will be the input of a neural network that automatically will be able to classify the state of the motor Feature selection in the transient phase For a fast diagnosis it would be interesting to shorten as much as possible the steady state, or even eliminate it completely, therefore the computation of the steady state RMS or of frequency spectra may be impossible. In this case, it is interesting to evaluate if the transient data alone can be used for motor diagnosis. During the transient phase the spectra vary in function of time. During the acquisition time the motor is accelerated from 0 to rpm. In order to analyze data, the short time Fourier transform is a very descriptive tool, which can provide features for diagnostics.. The following figures 3 and 4 represent the STFT computed for a good and a faulty motor. The displacement measurements by laser vibrometer reveal horizontal and vertical lines caused by signal drop out, which makes the signal noisy. Theses disturbances have not appeared in any of the measurements of velocity and acceleration OE-2-75,0-._50, i ii Figure 2.- Displacement, velocity and acceleration time signals of a faulty motor, with defects in the outer ring. Time 8s) 524 Proc. SPIE Vol. 4827

5 5TFTDspace TI-I Ac t_r E :::: E \ :i1 j.:e....5e _OE :I ; o.o o Figure 4.- Short Time Fourier Transform in displacement, velocity and acceleration of a faulty motor. Resolution in frequency domain is not sufficient to locate the single frequencies identified during the steady state, therefore the analysis of the type of defect is difficult to be done. Anyway, a motor diagnosis is still possible. From the observation of these diagrams, the process of diagnosis can be automatized calculating as significant feature the energy contained in a selected window in time-frequency domain. The time interval comprises from 0.5 to 2 s, the frequency range has been different depending of the measurement parameter: Hz for the displacement, Hz for the velocity; and for the acceleration. From the data obtained, it has been observed that the displacement measured by laser vibrometry is not a useful parameter to establish a correct classification between good and faulty motors. Instead, both the acceleration and the velocity appear to be good parameters to detect defects in the motors in this transient phase. Similar results are found for all defects. The discrimination is then automatically carried out by establishing a threshold value for every defect. Every motor that surpass that limit will be considered faulty Selection of characteristic features in frequency/order domain at steady state If defects have to be classified, it appears important to distinguish single lines in the spectra with good resolution. Therefore tests should have a steady state region, where Fourier analysis can be carried out. Observing the averaged FFT spectra obtained from the time signals (figures 5 and 6), it can be seen for all the three measured quantities that the difference between a good and a faulty motor are noticeable. It is also worth noting the different distribution of the frequencies along the spectrum. The spectra of the three parameters, displacement, velocity and acceleration are well correlated to each other. In the figures 5 and 6, the y axis has been amplified in order to be able to observe the frequency components existing in the higher ranges. In order to extract meaningful features for classification, it is important to concentrate attention on the particular frequencies which carry each defect. They are usually in a constant ratio to the angular speed of the rotor; this can be exploited to make the analysis independent of the small rotor velocity fluctuations that in any real case occur. Therefore instead of looking at frequency spectra in the frequency domain, it is better to look at the order domain, which is a frequency normalized to the rotational frequency of the motor. Proc. SPIE Vol

6 5 1.OE-16 5.o E-1 I. - II - Q.+o -- :i JJUUL1LtLL. Ii Iij 4.OE-1,..!. U 75 llee O UO U 2O Frecpiency Freciexy Figure 5.-FFF analysis of displacement, velocity and acceleration signals of a good motor. 4E E-7-. : ,7E-7,-' OE-1 20E+2 1.BE+2,.15E+2 50( LJ1k. haii l A , i o 2sa soa a Fteqiency 2, I, E+2- I ;::' 1.OE SE-S U- 5E+1 J J 11f , OE+O 3.SE16 3,5E-9 OE16 3, O 2.5E16 a: 2, E+O 1.2E+O 2.OE o+o 15E E d II.1, il,li Ii JI I.h ii, ii.ii i Frequency Figure 6.- FVF analysis of displacement, velocity and acceleration signals of a faulty motor. Within the spectrum, there are some frequencies that characterizes the vibration of the bearings and once known help to define which defect is present; therefore features must be derived by looking at those components of the signals. These frequencies depend on the particular dimension and geometry of every bearing and on the rotational speed of the motor15. The order analysis has been used to detect the presence of these frequencies, always relating them to the rotational speed. In our particular case, according to the geometrical dimension and the type of bearing existing in the motor, the orders (characteristic frequencies divided by the rotational speed, rpm) Order will be the following: (Freuuencv/RPS) The FTF frequency represents the Fundamental Train Frequency and Main Lower Upper reveals the existence of defects in the bearing cage; RPS is the rotational frequencies Bearing Bearing speed of the motor; BSF is the Ball Spin Frequency and shows the FTP defects in the rolling element; BPFO indicates the Ball Pass Frequency RPS 1 1 of the Outer Race, and finally the BPFI Orders represents the Ball Pass BSF Frequency of the Inner Race. Both BPFO and BPFI reveal defects in the BPFO outer and inner ring respectively. BPFI The orders of rotation can be considered harmonics, but unlike.. harmonics many interesting orders are non integer harmonics of the 1st Table 1.- Order frequencies according to the type... order (of the rotational speed) as it can see in the table 1. of bearmg and defect.... Not only the main frequencies can be found m the spectra, its harmomcs are also present, being the most probable of them the harmonics of the rotational speed (k*rps). In addition, in most practical cases, a faulty bearing will have several defects, and additional groups of frequencies and its harmonics will be also detected161. It can be a series of components with frequencies (BPFO+RPS) and a series of harmonics of the frequencies (k BPFO + j FTF), (k BPFO + j BSF), (k BPFO + j BPFI), being k and j integers, with sometimes j=k and other times j. k; other combinations are observed as well. TI fun 526 Proc. SPIE Vol. 4827

7 3.5E-7 FTF 3,OE-7, 25E-7-2OE-7-1.SE-7 1OE-7 5,OE-8 RPM 2*RPM U) S Freiency 22 Frequency Figure 7.- Velocity order spectrum of a good motor Figure 8.- Velocity order spectrum of a faulty motor with defect in the lower bearing A general observation of the order spectra has shown that within the range of the lower orders, the differences between a good and a faulty motor are not well evident. However, a large difference appears in the higher orders, which reveal features quite sensitive to defects. Therefore we have concentrated on those features, which are discussed hereafter for one of the defects, in the lower bearing. Figures 7 and 8 show an example of order analysis carried out in a good and a faulty motor. The most relevant frequencies that can be observed in the spectra in the good motor are FTF, the rotational speed RPM and some harmonics; the spectrum of the faulty motor reveals components in the higher orders, that also existed in the good one, but with a much lower amplitude. These components are combinations of the characteristic frequencies of table 1 and a subset of them as been chosen. They have been defined as feature 1 (Fl), feature 2 (F2), feature 3 (F3), etc. Velocity spectra measured by laser vibrometer and acceleration spectra by accelerometer have been compared versus those features in a series of good and faulty motors, in order to find correlation with the defect. The displacement measurements do not provide sufficient amplitude resolution in the range of orders of interest, being the displacement extremely small at high frequencies; therefore they are not considered for the analysis. In figure 9 the amplitudes of the velocity power spectrum proves that there is a clear difference between the amplitudes of good and faulty motors, which group around two separate sets. Therefore laser vibrometer data confirm to be well correlated to the defect. Similar results have been obtain with motors that exhibited other type of defects. Figure 10 shows the same features observed by the accelerometer during the same series of tests. The data of good and Velocity measurements 01 (U) E E C) U)) 0. U, U)) Feature 1. Feature 2 A Feature 3 0 Feature 4 Feature 5 Feature 6 + Feature 7 - Feature 8 order Figure 9.- Comparison of power spectrum amplitudes (order analysis) between good (black) and faulty motors (grey) with defect in the lower Proc. SPIE Vol

8 faulty motors do not separate in two distinct sets. This is an indicator that in this application accelerometer data do not provide signals which allow computation of reliable features to be used for defect classification. Reasons for this low degree of correlation with defects can be found by analysing the accelerometer installation that one is forced to adopt for the online application on motor bearings. The probe support which is used for the purpose in fact shows problems when the motor moves rigidly, the probe tip may loose contact, thus increasing noise. Noise is also rather high due to the strong electromagnetic interference from the motor coils and the brushes, which operate at a few centimetres from the sensor and its cable. Therefore in these operating conditions, typical of a production line, mounting conditions and electrical noise tend to compromise accelerometer performance versus the performance of a laser vibrometer, if the laser head is isolated from environment vibrations. The non-contact character of this instrument and the fact that it is an optical probe result key factors in improving the diagnostic capabilities of the test station. Acceleration measurements 1.OE+03 1.OE+02 ' /4 Featurel 1 OE+O1. 0 Feature 2 S A Feature 3 E 1 OE+OO A S A ofeature4 1 OE-O1 ; Feature 5 c 1 A0 x; = Feature6 Q 1 OE-02 * + I +Feature 7 0 Feature8 1OE-03 a 1.OE-04 - Order Figure 10 - Comparison of acceleration power spectrum amplitudes (by accelerometer) between good (black) and faulty motors (grey) with defect in the lower bearing 5. CONCLUSIONS Two vibration measurement instruments have been compared for feature extraction for on-line diagnostic purposes, namely a laser Doppler vibrometer and an accelerometer. They have been installed so to operate on a series of electric motors positioned on a pallet, so to simulate on-line operation; the accelerometer is mounted on a probe tip, with a spring to assure contact, due to the geometry of the system and the impossibility to fix it on the motor in real line conditions. The comparison has taken into account the three different output signals available: displacement and velocity from the laser vibrometer, acceleration from the accelerometer. In order to keep diagnostic time as short as possible, the transient state is of interest. This has been analysed by STFT algorithm; it is possible to distinguish between good and faulty motors and a well correlated feature can be designed as an integral within a selected window in time-frequency domain. Displacement data are poor, while velocity and acceleration data correlate well with the defects. Characteristic features have been computed in the frequency domain during the steady state, as specific non-integer orders of rotation; they allow to classify the defect. The main results show a good correlation of defects with the velocity data measured by the laser vibrometer, while both displacement and acceleration signal are affected by noise, which makes feature correlation to defects poorer. The poor results of the accelerometer can be explained by the installation constraints and by the electromagnetic noise; they both contribute to the noise. These differences caused by the line environment, affect sensor performance therefore changing the diagnostic system reliability. The non-contact and optical nature of the laser vibrometer in this application provides relevant advantages. 528 Proc. SPIE Vol. 4827

9 ACKNOWLEDGEMENTS The work has been partly supported by the EU Marie Curie Research Training Contract n. G1TR-CT , "Pattern recognition in defect evaluation for on-line quality control" and partly by an Italian Research Ministry contract MIUR- Cofin2000. The contribution of Renato Cariglia to the execution of the measurements is acknowledged. REFERENCES 1. R.H.Lyon, Machinery noise and diagnostics, ed. Butterworth, Pouliezos A.D., Stavrakakis G.S., "Real Time Fault Monitoring of Industrial Processes", Kiuwer Academic Publishers, A. Lucifredi, "Elementi di analisi del segnale e della firma, monitoraggio di condizione e diagnostica, per la manutenzione predittiva delle macchine e degli impianti ", ed. Franco Angeli, Hammer P.S., "Acceptance testing of electric motors and generators", IEEE Transaction on Industrial Application, Vol. 24 (6), pp , N.Tandon, A.Choudhury, A review of vibration and acoustic methods for the detection of defects in rolling element bearings", Tribology International, vol. 32, pp , Azovtsev Yu., Barkov A., Yudin I., "Automatic Diagnostics and Condition Prediction of Rolling Element Bearings Using Enveloping Methods", 18th Annual Meeting of the Vibration Institute, June, N.Paone, L.Scalise, G.Stavrakakis, A.Pouliezos, "Fault detection for quality control of house-hold appliances by non-invasive laser Doppler technique and likelihood classifier", Measurement, vol. 25, pp , Portnoff M. R., "Time-Frequency Representation of digital signals and systems based on Short Time Fourier Analysis". IEEE Transactions on Acoustic Speech and Signal Processing, Vol. 28 (1), pp.55-69, Cammarata S., "Reti neuronali", ETASLIBRI, Proc. SPIE Vol