Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.795965
Title: Automatic classification of tubing defects by analysis of their eddy current signals
Author: MacFarlane, Kenneth T.
Awarding Body: University of Glasgow
Current Institution: University of Glasgow
Date of Award: 1987
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
This thesis contains the results of a study conducted upon the application of pattern recognition techniques to the signals from a rotating eddy current probe in essentially two circumstances. These were (firstly) the inspection of 316 finned austenitic stainless steel heat exchanger tubes destined for an Advanced Gas Cooled Reactor using single frequency eddy currents with a differential probe, and secondly, the inspection of Inconel 600 tubing for Pressurised Water Reactor steam generators with a multiplexed 4 frequency pancake probe. In the first case, an automatic defect detection algorithm has been developed which will isolate any discernable axially oriented crack-type defect. Various classifiers (multiclass linear, k,1 nearest neighbour, nearest mean, scan cross correlation, and adaptive learning network) have been coded and tested on a 3 class sample set and 96.9% correct classification achieved with the two least easily separable classes, whilst the harmless third class was removed with the segmentation algorithm. In the second case, a new logging system was developed, and 38 tubes containing 45 synthesised defects were logged to disk and scans from each were plotted in the impedance plane, and against time. There were six types of defect, mostly in 4 different classes of depth. Seven different segmentation methods were devised and tested on the scans and a fitted linear threshold in the impedance plane was found to be best for sensitivity and speed of operation given the background (Pilger) noise present in the tube set,though another method which uses a convex hull algorithm is able to discriminate against any definable undesirable background signal. An automatically annulable method of generating a subtractive mix of two channels was also developed for use with simple magnitude-based. thresholding segmentation. Ad hoc geometrical and spectral featuresets were tested on the 7 class defect set with a multiclass linear and a nearest neighbour classifier. The correct classification rate achieved was 61.5%, and in view of this, and in an attempt to produce a site-trainable system, some development was done on parametrically-based feature extractors (FFT and Fourier Descriptors), and on a class distribution-independant feature set selector.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.795965  DOI: Not available
Share: