This paper extends the work of Shankar et al to the classification of three types of machined defects in Inconel 600 steam-generator tubing: electrodischarge machined slots, uniform thinning, and elliptical wastage. Three different pattern-recognition techniques were used for classification: (1) an empirical Bayes procedure, (2) a nearest-neighbor algorithm, and (3) a multicategory linear discriminate function. The three types of defects were classified correctly with an overall accuracy of 96 to 98 percent depending on the technique used. Two pattern-recognition algorithms, least squares and nearest neighbor, were used to size uniform-thinning defects in steam-generator tubing. All of the defects were between 25 and 75 percent of the wall in depth. With the least-squares algorithm, we achieved a fit correlation of 0.99 with a 95 percent confidence interval of (0.98, 1.00).

Acoustic emissions are a type of nondestructive evaluation data that have received increasing attention, particularly as a method of monitoring the condition of inservice structures. Models have been developed that relate the rate of acoustic emissions to structural integrity. The implementation of these techniques in the field has been hindered by the noisy environment in which the data must be taken. Acoustic emissions from noncritical sources are recorded in addition to those produced by critical sources, such as flaws. This can lead to incorrect interpretation of the models. The purpose of this report is to discuss a technique for prescreening acoustic events and filtering out those that are produced by noncritical sources. The methodology that we investigated is "pattern recognition." We applied three different pattern recognition techniques to a data set that consisted of acoustic emissions caused by crack growth and acoustic signals caused by extraneous noise sources. Our examination of the acoustic emission data presented in this report has uncovered several features of the data that can provide a reasonable filter. Two of the most valuable features are the frequency of maximum response and the autocorrelation coefficient at Lag 13. When these two features and several others were combined with a least squares decision algorithm, we were able to correctly classify 90% of the acoustic emissions in the data set. Although these results must be verified on additional data sets, it appears possible to design filters that eliminate extraneous noise sources from flawgrowth acoustic emissions using pattern recognition techniques. iii ..

No abstract

No abstract