Problems Associated With Statistical Pattern Recognition of Acoustic Emission Signals in a Compact Tension Fatigue Specimen

Abstract Acoustic emission (AE) data were acquired during fatigue testing of an alumi-num 2024-T4 compact tension specimen using a commercially available AE sys-tem. AE signals from crack extension were identified and separated from noisespikes, signals that reflected from the specimen edges, and signals that saturatedthe instrumentation. A commercially available software package was used to traina statistical pattern recognition system to classify the signals. The softwaretrained a network to recognize signals with a 91-percent accuracy when com-pared with the researcher's interpretation of the data. Reasons for the discrepan-cies are examined and it is postulated that additional preprocessing of the AEdata to focus on the extensional wave mode and eliminate 'other effects beforetraining the pattern recognition system will result in increased accuracy. Introduction Acoustic emission (AE) is defined as "the class ofphenomena whereby transient elastic waves are gener-ated by the rapid release of energy from localizedsources within a material (or structure) or the transientwaves so generated" (ref. 1). Acoustic emission can begenerated by a variety of sources, including cracknucleation and propagation, multiple dislocation slip,twinning, grain boundary sliding, Barkhausen effect(realignment or growth of magnetic domains), phasetransformations, and debonding and fracture of inclu-sion. Acoustic emission can also be generated bysources other than materials under stress, such as com-ponents rubbing against one another (fretting), leaks,structural vibrations, electrical transients. Spanner(ref. 2) and Williams (ref. 3) have provided discus-sions of sources of acoustic emission in a variety ofmaterials and applications. Effective use of acousticemission for monitoring damage progression in struc-tures requires interpretation of the AE signals to deter-mine the sources of the AE, their locations, and theirseverity. An experienced AE practitioner can learn torecognize signals from different sources, but alwaysuncertainty about some of the data exists. Current AEsystems, such as the one used in this study, can recordup to 200 waveforms per second. Pattern recognitionalgorithms exist for training computers to recognizeand interpret the signals. The objective of this projectwas to investigate the applicability of statistical pat-tern recognition to the identification of crack signals ina well-controlled test with limited sources of acousticemission as a prelude to a possible application to mon-itoring crack growth in aging aircraft. The initialapproach was to use a commercially available soft-ware package to extract features from the acousticemission signals and perform the pattern recognition.Pattern recognition methods require that a networkfirst be trained to recognize signals; this is also calledlearning. A set of signals representing the differentclasses of data to be learned are provided as inputs tothe network along with their classes. The network ana-lyzes the differences between the signals and deter-mines which characteristics best define each class ofdata. It compares its calculations with the knownclasses of the signals provided by the user. Wherethere is ambiguity, or disagreement with the classesprovided, there is training error. The network can con-tinue to refine its analysis to minimize the trainingerror. Once the training error is minimized, the learn-ing is complete and one or more classifiers are devel-oped. These classifiers may be developed with thesame technique used in the learning phase, or differenttechniques may be used.The second phase of pattern recognition is classi-fication. New signals are input to the network and ana-lyzed by using the classifiers developed in the learningstage. The network does not know the classes of thesesignals but determines their classes based upon theclassifiers. If several classifiers are used, they may notall agree on the classes of all the signals. If the userknows the classes of the signals, he may evaluate theresults of the classification based upon his knowledgeof the signals. Any discrepancies between the classifi-ers and the user's knowledge are classification errors.In this work, a k-nearest neighbor algorithm wasused in the learning phase, and the training error was