Applications of Pattern Recognition to the Diagnosis of Equipment Failures

Abstract

The theory of a feature extraction method, called correspondence analysis, is given. It is applied to learning datas made of all available informations about a given type of equipment, as stored in a reliability and maintenance data bank.

Correspondence analysis is a variant of principal component analysis, based upon a CHI-square distributional metric wherein patterns and observations play symmetrical roles. A simultaneous graphical representation of both patterns and observations helps in analyzing the operational behaviour of the equipment for design review and maintenance control; two examples hereof are given for an airborne equipment, and for resistor components.

Real time diagnosis and fault localization have been achieved, thanks to a real time data compression into the reduced feature space generated by correspondence analysis, and by a sequential recognition procedure. This procedure is the generalized nearest neighbour rule, applied sequentially to learning sets of increasing size; the stopping rule uses a compromise between recognition time and the misclassification probability. An example of automated real time testing is given, showing a 92% true recognition rate for acceptable items produced by machine-tools, and a mean correct diagnosis rate of 81% for non-accepted items. 16 references.

This paper is part of a course in Technical Diagnosis at the Ecole Nationale Supérieure de l’Aéronautique et de l’Espace, France.