A fault diagnosis method for rotating machinery with variable speed based on multi-feature fusion and improved ShuffleNet V2

Using deep learning to classify the time-frequency images of bearing vibration signals has become a mainstream method in the field of fault diagnosis. Most studies, however, assume a constant rotational speed, and the accuracy and reliability of the diagnosis model diminishes once the rotational speed changes. Moreover, due to the large model size and high computational complexity, the convolutional neural networks are not suitable for industrial applications. This paper proposes a novel fault diagnosis method for rotating machinery with variable speed based on multi-feature fusion and improved ShuffleNet V2. First, complementary ensemble empirical mode decomposition is used to denoise the time-domain signal. Then the denoised time domain signal is converted into an angular domain signal using the resampling technique, while the envelope spectrums of the angular domain signals are obtained by the Hilbert transform, and the three signals are fused into an red-green-blue image form to enhance the sample features. Finally, to perform fast and accurate classification of the features, the squeeze and excitation blocks are placed after the branch splicing operation of ShuffleNet V2 to enhance the recognition accuracy, and the rectified linear unit activation function is replaced by the HardSwish activation function to avoid necrosis. The experimental results show that the fault recognition accuracy of the proposed method for rolling bearings under variable speed is 96.4%, and the size of the fault diagnosis model is 7.82 MB, indicating that the method can effectively improve the accuracy and ensures that the model size does not increase significantly.