Bearing Fault Diagnosis Based on Improved DBN Combining Attention Mechanism

Aiming at the difficulty of fault extraction and classification caused by traditional feature extraction methods due to the large noise of rolling bearings in practical work, a bearing fault diagnosis method based on improved DBN combining attention mechanism is proposed. The Gaussian Bernoulli restricted boltzmann machine model is introduced to solve the problem that the input vector of traditional restricted boltzmann machine is limited by Bernoulli binary distribution and has a poor fitting effect for non-binomial data reconstruction. The cosine loss function is used as the loss function, which retains the advantage of softmax loss function to enlarge the difference between classes, it reduces the sensitivity to different signal intensity. Combined with the attention mechanism adaptive, more "attention" is given to the effective features describing the bearing state. At the same time, multiple features in time domain and frequency domain are used to prepare for fault diagnosis. Experiments show that this method can effectively improve the adaptive feature extraction ability of the model and the accuracy of fault diagnosis, and has good generalization ability.