Fault Diagnosis of Rotating Machinery Bearings Based on Multi-source Wavelet Transform Neural Network

A multi-source wavelet time-frequency transform convolutional neural network was proposed to address the issues of limited fault samples in rotating machinery bearing fault diagnosis, along with the vulnerability to overfitting and the poor generalization ability of traditional models when dealing with small datasets. Initially, for high-frequency data obtained from a single vibration sensor, a wavelet transform-based time-frequency convolutional layer was formulated to integrate both the real and imaginary components of wavelet coefficients. Here, the real component represented the amplitude information of vibration signals, while the imaginary component depicted phase information. Compared with a convolution layer that only considering real part, this convolutional layer may extract comprehensive time-frequency features. Subsequently, the time-frequency convolutional layer was employed to independently extract features from high-frequency data acquired by multi-sensors on a single device, and these features were then concatenated. Lastly, a dense module utilizing lightweight depth-separable convolution was developed to conduct further feature extraction from the concatenated features, facilitating fault classification. The effectiveness of the model was confirmed through experimentation using Case Western Reserve University rolling bearing dataset, achieving an accuracy of 98.5%.Additionally, the model was deployed for fault diagnosis in rotary kilns, belt conveyors, and grate coolers, demonstrating an average accuracy of 97.19%. © 2024 Chinese Mechanical Engineering Society. All rights reserved.