Application of local outlier factor method and back-propagation neural network for steel plates fault diagnosis

Abstract

Fault diagnosis, which is a task to identify the nature of the occurred fault, is of paramount importance to ensure the steadiness of industrial and domestic machinery. Essentially, fault diagnosis is a problem of classification. A method based on Local Outlier Factor (LOF) anomaly detection and BP neural network is proposed to apply to steel plates fault diagnosis. The LOF method is firstly used to find the anomaly samples and process the relevant samples detected. Then the processed samples are used to train a back-propagation neural network (BPNN) to classify steel plate faults. It is to be noted that the LOF method's effect of outlier elimination nicely overcomes the specific defect of the BP neural network model in which the training process is very sensitive to singularities in the training samples. Results of contrastive experiments indicate that the proposed method can reliably improve the classification accuracy and decrease the training time.