Bearing Fault Diagnosis Based on the Combination of Attention Entropy and Bayesian Network

In order to improve the accuracy of bearing fault identification, this paper proposes a bearing fault diagnosis method that combines the Improved Refined Composite Multiscale Attention Entropy (IRCMATE) with Bayesian networks. The method uses Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose the original signal, obtain multiple intrinsic modal functions (IMFs), calculate the Pearson correlation coefficient (PCC) between each IMF and the original signal, and select the IMFs with PCC greater than 0.3 to reconstruct the signal. Aiming at the problem that the multiscale attention entropy depends on the data length and is easily affected by the outliers in the sample, the improved refined composite multiscale attention entropy is proposed. The paper calculates the IRCMATE of the reconstructed signals, and select the IRCMATE under the appropriate scale as the new feature vector, so as to realize the fault feature extraction of the bearings. Improved Crown Porcupine Optimizer (ICPO) is proposed. ICPO solves the variable conflict problem of Crown Porcupine Optimizer (CPO) by realizing binary conversion through Sigmoid function. The dynamic strategy selection factor is introduced into the CPO, which improves the global exploration ability of the algorithm early in the iteration. After each iteration, the local optimization operations of increasing edge, decreasing edge and reversing edge are implemented on the current global optimal solution, so as to avoid falling into the local optimal solution. Using ICPO to get Bayesian network structure, and the bearing fault diagnosis model is obtained after training using the training set. It is shown through experiments that the fault features extracted using IRCMATE are more obvious. The comparison with other methods shows that IRCMATE-ICPO-BN has a higher fault diagnosis recognition rate.