A Multiscale Bearing Fault Diagnosis Method Based on Dilated Residual Bottleneck Block for High Noise Conditions

Abstract

As critical components of mechanical equipment, the health condition of bearings directly determines the stable operation and operational safety of machinery. Thus, timely early fault diagnosis of bearings is of great practical significance. Nevertheless, collected vibration signals are frequently contaminated by strong ambient noise, which hinders effective extraction of discriminative fault features and degrades diagnosis accuracy. To tackle this problem, this paper proposes a multi-scale noise-robust residual network built with dilated structures to boost diagnosis accuracy under high-noise environments. Specifically, stacked dilated convolutions are embedded inside residual bottleneck blocks to enlarge the receptive field of shallow networks during fault feature extraction, and dimensionality reduction operations are adopted to suppress noise interference. Stacked such dilated residual blocks are assembled to construct a dedicated noise-robust module named DRes-Block. Furthermore, a Channel Attention Module (CAM) is deployed after the noise-robust module to strengthen the weight responses of feature channels highly correlated with faults, which further improves the anti-noise capability of the module. To fully capture comprehensive fault features, a multi-scale feature extraction module is designed, which adopts distinct dilation rates on parallel branches to mine multi-scale inherent characteristics of fault vibration signals. This module is integrated to construct an end-to-end bearing fault diagnosis model, namely DResB-MDCNN. Experimental results demonstrate that the proposed model achieves diagnosis accuracy higher than 95% under strong noise with an SNR of -8 dB on both the CWRU bearing dataset and self-simulated bearing dataset. Even under the extreme noise condition of SNR = -10 dB, the model still reaches an accuracy of 87.07%, outperforming ResNet18, VGG16, CNN-TCN and WDCNN-BiGRU by 2.53%, 29.45%, 4.95% and 11.93% respectively. Based on the data sampling and partitioning strategy proposed in this paper and the experimental platform used, each training epoch only takes 3.4 seconds. The total number of trainable parameters of the model is 2,326,438, with a storage size of approximately 8.87 MB, which renders the model lightweight. Therefore, the DResB-MDCNN model proposed in this paper possesses prominent comprehensive advantages for bearing fault diagnosis under heavy noise interference.