Multiscale Fusion Attention Convolutional Neural Network for Fault Diagnosis of Aero-Engine Rolling Bearing

Abstract

Considering the nonstationary characteristics of the vibration signal of aircraft engine rolling bearings and the insufficient ability of convolutional neural network (CNN) to extract important fault features from the signals affecting the fault diagnosis performance, a fault diagnosis method based on multiscale fusion attention CNN (MSFACNN) is proposed. First, the raw 1-D vibration signal is matrixed and converted into 2-D gray scale images as input, which preserves the relevant information between the time series data. Second, considering the multiple time scale characteristics of the vibration signal, a multiscale convolution module is constructed to extract the multiscale features of the samples. Then, the importance of the fault feature is learned by the improved attention mechanism and then weighted fusion, making the network learn to focus on more important information, and realizing a dynamic scale-selection mechanism. Finally, fault diagnosis is realized by the fully connected (FC) layer and softmax function. The short connection is introduced to avoid the problem of network degradation. The method is applied to two different rolling bearing datasets, and the results of experiments reveal that MSFACNN can achieve higher recognition accuracy with smaller training parameters than other classical diagnostic methods.