Fault detection based on evolving LVQ neural networks

This paper proposes a novel approach of employing genetic algorithms (GA) to set initial reference vectors in the Kohonen layer of learning vector quantization (LVQ) neural networks. The aim of this investigation is to improve the learning characteristics of LVQ so as to get more accurate classification results. In the proposed scheme, the reference vectors are set to the locations mostly matching the probability distribution of training vectors. Genetic algorithms are applied to optimize the locations and distribution of the reference vectors. After competitive learning of LVQ, the reference vectors are employed to be representatives of various patterns to determine the categories which testing vectors belong to a comparison study is reported based on LVQ with random initial reference vectors, LVQ with GA learning and LVQ with initial reference vectors set by GA. Experimental results of a case study have shown that the proposed method is promising for machine fault classification.