A Bearing Fault Diagnosis Method Based on TimeFrequency Enhanced Residual Network

Aiming at the recognition problem of non-stationary fault signals in rotating machinery, this paper proposes a novel bearing diagnosis method named Time-Frequency Enhanced Residual Network (TFE-ResNet). This method is composed of the Time-Scale Synchrosqueezing Transform (TSST) and an improved residual neural network. Firstly, the generalized group delay estimation and dynamic threshold clipping mechanism are used to generate high-resolution time-frequency images. Secondly, the generated time-frequency images are used as feature maps to input into the improved ResNet architecture for training. The improved ResNet architecture introduces the DropBlock regularization technique, which randomly discards spatially continuous feature blocks to break the dependence between local features. The temperature scaling technique and label smoothing technique are introduced to enhance the model’s ability to estimate uncertainty. Finally, according to the training results, the impact of time-frequency resolution on the model’s classification accuracy is analyzed, the optimal model structure is constructed, and the bearing condition diagnosis is realized. The dataset from Case Western Reserve University is used to validate the model and compare it with other bearing diagnosis methods. The classification accuracy of the proposed algorithm is better than that of the contrastive algorithms.