Identifying a consistent diagnostic approach for bearing faults in the same mechanical equipment operating under continuously varying rotational speeds remains a critical challenge. In this study, a time–frequency residual neural network method is proposed by integrating the short-time Fourier transform (STFT) with deep residual connections. First, the vibration signals are transformed into time–frequency representations using STFT, from which training and testing datasets are constructed. The training dataset is used to train the deep residual neural network, while the testing dataset is employed to evaluate its performance. Batch normalization (BN) is incorporated to enhance training stability, and Adapter Tuning is adopted for parameter fine-tuning to improve convergence efficiency. After validating the effectiveness of the proposed deep time–frequency residual network using publicly available datasets, its diagnostic accuracy and practical applicability are further verified through bearing vibration data collected from a dedicated experimental test bench. The experimental validation results show that, for fault diagnosis under changing bearing rotational speeds, the proposed method can achieve an accuracy of approximately 97.25% after only 200 iterations. Using a model structure with three residual blocks, the optimal diagnostic performance can be achieved with an input sample size of 3000.
Li et al. (Tue,) studied this question.