Bearing fault detection using SOM based on singular value spectrum

A novel fault diagnosis approach based on singular value spectrum of signals in phase-space is presented in this paper, which can effectively solve the problems of detection-time delay due to the requirement for a large number of samples to calculate those dynamic invariants, such as the largest Lyapunov exponent etc. In order to avoid the practical problems of difficultly obtaining abnormal samples in fault diagnosis applications, a diagnosis model based on SOM is proposed in this paper. The corresponding models are established according to different types of training samples to diagnose new test sample, which is determined by the samplepsilas distance to the current model. In experiments, comparison of different distance functions is conducted which indicates the Euclidean distance achieves the best performance. The performance of detectors based on singular value spectrum and singular value spectrum entropy is also compared. The results show the effectiveness of the investigated techniques. Finally, the proposed approach is compared against the detector based on the largest Lyapunov exponent and MLP techniques based on 10-order temporary signals. The results illustrate the proposed approach significantly outperforms other methods in terms of detection rate.