Bearing Fault Diagnosis Method Based on Multi-scale Dilated Convolution Under Various Noise Conditions

With the development of artificial intelligence, the rolling bearing fault diagnosis methods based on deep learning have gained widespread adoption. However, these diagnostic approaches face the challenges in achieving high recognition accuracy, particularly in the presence of noise interference, and the burden of extensive model parameters. To address these problems, this paper proposes a lightweight fault diagnosis model. The model utilizes an innovative end-to-end architecture known as the multi-scale dilated convolutional neural network (MDCNN). This architecture effectively extracts fault features with different resolutions by combining multi-scale techniques with dilated convolution, while significantly reducing the parameter burden associated with multi-scale approach. To further strengthen the model's robustness and enhance its performance in noisy interference, residual connections and the H-swish activation function are incorporated into the MDCNN architecture. Experimental validation using publicly available datasets demonstrates that the proposed method outperforms other existing methods cross different signal-to-noise (SNR) ratios and with fewer samples.