ARTIFICIAL NEURAL NETWORK BASED ALGORITHM FOR ACOUSTIC IMPACT BASED NONDESTRUCTIVE PROCESS MONITORING OF COMPOSITE PRODUCTS V Srivatsan, Krishnan Balasubramaniam* and N. V. Nair Center for Nondestructive Evaluation, ^Department of Mechanical Engineering, Indian Institute of Technology, Chennai-600036, INDIA ABSTRACT. Damages like cracks, delaminations, etc., in composite parts have traditionally been evaluated using manual methods like acoustic impact (using measurements in the audio frequencies). This technique is currently used during manufacturing for product quality testing and later for maintenance and assurance of structural integrity. The automation of this technique will significantly improve the reliability of inspection. The signals obtained from the composites are analyzed using signal-processing techniques in the time-frequency domain to build a robust algorithm for detection and identification of defects. A feature vector is constructed using these techniques and then applied to a neural network for defect identification. Comparative studies are conducted to search for the best and most comprehensive feature vector. Results using different signal processing techniques are presented. Similarly comparative results are presented between two different kinds of neural networks (namely Radial Basis functions and MLP) and various architectures in each kind. A low cost data acquisition system has also been developed for acquiring audio signals using the sound card and the microphone in a multi-media PC. INTRODUCTION Damage like cracks, voids and de-laminations in composite parts that are manufactured in high volumes must be detected during manufacture to ensure quality. This process requires the skill and experience of an operator, and thus, the reliability of the system is not certain. The concept of acoustic impact itself is also longstanding, having come to light through the potters of ancient ages and more rigorously automated in recent times for NDE of composites and aerospace components for maintenance applications. [19] In this paper, we have attempted to develop an automated process based on the similar principles that can be directly applied to quality assurance of composite parts in a mass production environment. The specimens are excited by an impact actuated by a solenoid, and the acoustic signal is recorded using a microphone and a multimedia PC. Frequency extraction is done on the time domain signal and the frequency spectral data was used as input to a neural network to detects the defects or delaminations in the composites at very high production rates (2-5 parts per second). CP657, Review of Quantitative Nondestructive Evaluation Vol. 22, ed. by D. O. Thompson and D. E. Chimenti © 2003 American Institute of Physics 0-7354-0117-9/03/$20.00 1651 PRODUCTION PROCESS PRODUCTION PROCESS Usually homogeneous composites are fabricated by compressing the powdered Usually homogeneous composites are fabricated by compressing the powdered constituents under high temperature and pressure. The composition and dimensions of the constituents under high temperature and pressure. The composition and dimensions of the lining is normally manipulated to achieve the right values of the control parameters. A lining is normally manipulated to achieve the right values of the control parameters. A typical example is brake linings in which case the usual control parameters are coefficient typical example is brake linings in which case the usual control parameters are coefficient ofoffriction and thermal conductivity. During the compression process, by product gases friction and thermal conductivity. During the compression process, by product gases are formed, these are formed,and andproper properarrangements arrangements need need to to be be made made to to vent vent them. them. Occasionally, Occasionally, these gases do not find an escape route and thus form voids within the composite part causing gases do not find an escape route and thus form voids within the composite part causing porosity porosityoror'blisters'. ‘blisters’. These As the the Thesevoids voidsand and 'blisters' ‘blisters’ adversely adversely affect affect the the performance performance of of the the parts. parts. As part wears out due to repeated use, the voids come to the surface. This has a heavily part wears out due to repeated use, the voids come to the surface. This has a heavily detrimental detect and and detrimentaleffect effecton onthe theperformance performance of of the the part. part. This This makes makes itit essential essential to to detect reject defective parts on the manufacturing line itself. Unless the voids are of significant reject defective parts on the manufacturing line itself. Unless the voids are of significant size be size(this (thisresults resultsininaaconspicuous conspicuousbulge bulgeon onthe the surface surface of of the the part), part), they they cannot cannot usually usually be detected by the naked eye. detected by the naked eye. ACOUSTIC ACOUSTICIMPACT IMPACTBASED BASEDRESONANCE RESONANCE On Onimpact, impact,any anysolid solidwould wouldemit emitsound sound due duetotoresonance resonance in in either either of of the the following following two twomanners. manners.The The first first being being the the actual actual resonance resonance of of the the structure structure and and the the other other the the formation of standing waves inside the solid. One of these frequencies usually lies in the formation of standing waves inside the solid. One of these frequencies audible audiblerange rangeofoffrequencies. frequencies.The Thepresence presence of of aa defect defect such such as as aa delamination delamination causes the object objecttotoresonate resonatewith withaadifferent differentfrequency frequencythan than aa 'perfect' ‘perfect’ or or defect defect less less composite. This isisthe resonance thebasic basicprinciple principleon onwhich whichour ourdetection detection technique technique isis based. based. The The modes of resonance are aredepicted depictedininthe theFig. Fig.11below. below. FEM were FEMbased basedmodels modelsof of aatypical typical composite composite part part of of rectangular rectangular cross-section were studied studiedusing usingcommercially commerciallyavailable available software software so so as as to to obtain obtain an an initial initial insight into the resonance study are are resonancemodes modes available. available. The The mode mode shapes shapes obtained obtained as as aa result result of this study presented presentedbelow. below.Fig Fig22depicts depictsthe thefirst firstfour fourmode modeshapes shapes liiiiiiiiiiipi^^ : fiii it: 4»i: FIGURE FIGURE1.1.Resonant Resonantmodes. modes. FIGURE2.2.FEM FEMprediction predictionofofmodes modesofofresonance. resonance. FIGURE 1652 INITIAL EXPERIMENTAL SETUP INITIAL EXPERIMENTAL SETUP Impactor Impactor The design of the impact mechanism becomes a critical issue since the interference Thedue design of the impact becomes is a critical issue since the interferencethe occurring to acoustic signalsmechanism from the impactor quite significant. Standardizing occurring due to acoustic signals from the impactor is quite significant. Standardizing the impactor, though, can relatively easily circumvent this problem. We used a solenoid-based impactor, though, can relatively easily circumvent this problem. We used a solenoid-based impactor with a ball hammerhead. This ball was introduced essentially to increase the impactor with especially a ball hammerhead. Thisimpactors ball was had introduced essentially increase weight since lightweight a tendency to liftto off and the cause weight since especially lightweight impactors had a tendency to lift off and cause secondary impacts. The solenoid has a rated frequency of 12 Hz. secondary impacts. The solenoid has a rated frequency of 12 Hz. Initial Test Stand Initial Test Stand The test stand for the experiment was designed using aluminum L sections and this The test stand for the experiment was designed using aluminum L sections and this was acoustically insulated from fromthe thecomposite compositepart partusing usingStyrofoam Styrofoam base. The height was acoustically insulated base. The height of of the stand was also chosen so as to remove air column resonance. Fig. 3 shows the the stand was also chosen so as to remove air column resonance. Fig. 3 shows the testtest stand. stand. Data Acquisition Acquisition Data The data data acquisition acquisitionmechanism mechanismwas waskept keptasassimple simpleasaspossible possible allow The to to allow forfor thethe easy adaptability adaptability into into aa production productionenvironment. environment.We Wehave haveshown shown that a simple passive easy that a simple passive computer microphone microphonecan canbe beused usedtotodetect detectthe thedifferences differencesin in frequencies.Our Our prototype computer frequencies. prototype was built with a computer microphone and a LabView based data acquisition system was built with a computer microphone and a LabView based data acquisition system forfor determining the the frequency frequencyspectra. spectra.The TheANN ANNanalysis analysiswas wascarried carried initially post determining outout initially in in thethe post processingmode. mode.Fig. Fig.44shows showsaascreen screenshot shotofofthe thedata dataacquisition acquisition software. processing software. Impactor Impactor Composite db Composite Styrofoam Styrofoam FIGURE FIGURE3.3. Test Testsetup. setup. FIGURE 4. 4. Screen Screenshot shotofofdata dataacquisition acquisitionsoftware. software. FIGURE 1653 ARTIFICIAL NEURAL NETWORKS There is a distinct difference in the frequencies across different part thickness and compositions. This difference cannot be effectively tacked with a simple rule based system based simply on the frequencies of resonance. The complete and accurate knowledge of the composite part's composition is almost impossible in real-life scenario making the accurate prediction of the resonance frequencies very difficult. An ANN based approach assumes no knowledge of the composition or thickness of the parts. Moreover an on-site train and use approach is more attractive to a production plant, which would need to conduct an in-depth study each time it changed its parameters, if it were to rely on rule based methods. Various architectures were tried out with two different types of classifiers, the Multi Layer Perceptron (MLP) and the Radial Basis Function (RBF) based neural networks. These were chosen since they are the two most commonly used classifiers and also the fastest. They were trained with various source vectors, all in the frequency domain. RESULTS The first set of training of the MLP was done with the complete frequency domain data (comprising of 517 data points). This training set was later modified using a data set with the 10 values of frequency appearing with the highest amplitude in the frequency data. Then the data was collected at 9 points on each specimen and all the 90 points were used for training. The training was successful and the time taken was found to be generally proportional to the input vector size since the rest of the architecture was kept unchanged and fully connected at all time. Testing gave best results with the 90-point data with less than 10% error (1 missed call in 13 samples). The RBF network was trained with the 10-point data and the 90-point data. The time taken for training only doubled in the case of the 90-point data, showing promise that this is less affected by the size of the input vector. The testing correlations were found to be slightly better than the MLP case, (1 error in 15 data sets, correlation of 0.932). It must also be said that though the use of the 517 point data is not time-efficient, the results of testing using that input vector with the MLP network gave extremely promising results - 0 wrong calls among 50 data sets. PROTOTYPE CONVEYOR INSPECTION SYSTEM A prototype of the entire mechanism was designed which can be inserted into a production environment. The conveyer is built using Aluminum L sections for a frame with a cotton belt. A stepper motor given using an MS shaft operates the drive. A computer microphone is used for data acquisition. The setup is designed to fit as the last end of a conveyer belt. The figure of the entire system is shown in Fig 5 below. Software was developed using the MLP ANN algorithm with lights indicating GO/ NO-GO for each component. In Fig. 6, the screenshot of the software written in Labview is illustrated. 1654 FIGURE 5. Complete system, Impact station (two views) and Drive system (in counter clockwise order). FIGURE 5. Complete system, Impact station (two views) and Drive system (in counter clockwise order). FIGURE 6. 6. Automated Automated Software. Software. FIGURE FUTURE WORK WORK FUTURE The algorithm algorithm needs needs to to be be generalized generalized across across various various thickness thickness and and compositions compositions The by training the network appropriately. Possibilities of applications to situations involving by training the network appropriately. Possibilities of applications to situations involving particulate composites can be investigated. particulate composites can be investigated. ACKNOWLEDGEMENTS ACKNOWLEDGEMENTS We register our gratitude to to Dr Dr Satish Satish Udpa Udpa and and his his team, team, atat Michigan Michigan State State We register our gratitude University for the invaluable guidance and help extended to us. We thank all the staff and University for the invaluable guidance and help extended to us. We thank all the staff and students at CNDE, IIT Madras for their assistance. students at CNDE, IIT Madras for their assistance. REFERENCES REFERENCES 1. 2. 2. P. of the the mechanical mechanical impedance impedance method method of of nondestructive nondestructive P. Cawley, Cawley, ‘The 'The sensitivity sensitivity of testing’, (August 1987) pp. 209-215. 209 – 215. testing', NDT NDT International, International, (August 1987) pp. P. “The mechanics mechanics of of coin coin tap tap method method of of nondestructive nondestructive P. Cawley Cawley and and R.D. R.D. Adams, Adams, "The testing,” 299-3 16 (1988). (1988). testing," J.Sound and Vibration, 122(2), 299-3 1655 3. 4. 5. 6. 7. 8. 9. P. Cawley and R. D. Adams, "Sensitivity of the Coin-Tap Method of Nondestructive Testing," Materials Evaluation, Vol. 47, May 1989, 558-563. J. J. Peters, D. J. Barnard, N. A. Hudelson, T. S. Simpson and D. K. Hsu, "A prototype tap test imaging system: Initial field test results," Rev. Prog. Quantitative Nondestructive Evaluation, Vol. 19, edited by D. O. Thompson and D. E. Chimenti, AIP, Melville, NY, 2000. pp. 2053-2060. Daniel J. Barnard, John J. Peters and David K. Hsu "Development of a magnetic cam for the computerAided tap test system" Review of Progress in Quantitative Nondestructive Evaluation Vol. 20, ed. by D. O. Thompson and D. E. Chimenti pg. 1966-712 (2001) G. E. Georgeson, S. Lea and J.Hansen, "Electronic Tap Hammer for Composite Damage Assessment," SPIE Proceedings, Vol. 2945, Nondestructive Evaluation of Aging Aircraft, Airports and Aerospace Hardware, edited by R. D. Rempt and A. L. Broz, 1996. pp. 328- 338. (The RD3 is available from WichiTech, Inc., 8980-L, Route 108, Columbia, MD 21045.) R. D. Adams, A. M. Alien and P. Cawley, "Nondestructive Inspection of Composite Structures by Low Velocity Impact", in Review of Prog, in Quantitative NDE, Vol. 5, edited by D. 0. Thompson and D. E. Chimenti, PlenumPress, New York, 1986. pp. 1253-1258. D. K. Hsu, J. J. Peters, D. J. Barnard, D. Fei, T. A. Simpson and V. Dayal, "Imaging Flaws and Damages Using Instrumented Tap Test", Paper Summary of ASNT Spring Conference and Research Symposium Orlando, FL, March 2226,1999. pp. 139-141. J. J. Peters, D. J. Barnard, A. Hudelson, T. S. Simpson and D. K. Hsu, "A Prototype Tap Test Imaging System: Initial Field Test Results," Rev. of Prog, in Quantitative NDE, Vol. 20, edited by D. 0. Thompson and D. E. Chimenti, 2000. 1656
© Copyright 2025 Paperzz