Bearings fault detection based on AR self-correlation entropy

To solve the problems of difficultly obtaining abnormal samples in bearings fault detection application and overfitting of conventional classifications due to the abnormal data imbalanced, a novel one-class detection model based on AR model self-correlation´s entropy characteristic is presented in this paper. The AR model is employed to extract the normal samples´ parameter characteristics and consequently normal samples´ AR sub-space is established. The self-correlation of errors generated by other samples´ being projected onto the AR model space is calculated. The entropy of the previously calculated self-correlation is used as the metrics of similarity with the normal sub-space. The experiments show that the proposed approach can efficiently overcome the drawback of computational complexity and detection rate´s sensitivity to the length of samples of conventional detection methods based on AR´ parameters. The single fault detection and diagnosis schemes based on AR self-correlation´s entropy also are proposed in this paper. The model´s threshold value settings are also analyzed and the determination approach based on the Particle Swarm Optimization is investigated in this paper. The proposed detection scheme is compared against MLP and detection techniques based on AR parameters as features in the experiments. The results illustrate effectiveness of the investigated techniques with some concluding remarks.