Feature extraction and diagnosis of bearing fault signals based on LE and DBN algorithms

In order to improve the operation stability of mechanical transmission system, a fault signal feature extraction method based on LE and DBN algorithm was proposed. LE algorithm was selected to extract manifold parameters of high-dimensional vibration signals. The manifold learning data was input into DBN to realize the secondary mining process of feature data and complete the classification of different faults. The results show that the le-DBN model used in this paper achieves better performance than other models. LE algorithm can significantly shorten the computation time of LE-DBN combined model. The recognition accuracy of training set samples is almost 100%, indicating that the model can play a good fitting effect on training data. Compared with PCA and KPCA, LE algorithm has better feature extraction performance and can achieve nearly 100% accuracy when appropriate parameters are set. When the number of label samples is between 60 and 120, DBN network shows better classification performance than CNN network. The LE-DBN model achieves the ideal classification accuracy and fast recognition requirements for different bearing fault diagnosis.