Constrained neural classifier training method for flaw detection in industrial pipes using particle swarm optimisation

摘要

A novel method for constrained training of multi-class artificial neural network classifiers is proposed in this work. The traditional training procedure is usually based on mean square error minimisation and thus, all classes of interest are considered as having the same relevance for system performance. This is not always the case for real-world applications in which the class relevance may be unbalanced. In this paper, cost functions designed to introduce classification performance constraints for specific classes are presented and particle swarm optimisation is used as global optimisation method. The proposed method is applied to a non-destructive evaluation decision support problem using pulsed eddy currents signals. Experimental results obtained from thermally insulated industrial pipes indicate the efficiency of the proposed method in comparison to neural networks trained from the traditional back-propagation algorithm.