Embedding Time-Frequency Transform Neural Networks for Efficient Fault Diagnosis

To address the challenges of non-stationary signals, noise corruption, and limited interpretability in intelligent fault diagnosis, we propose a Time-Frequency Dual Transformation (TFDT) framework. By embedding a trainable Time-Frequency Convolution (TFConv) layer into a convolutional neural network, TFDT integrates physics-informed time-frequency transforms—STTF, Morlet wavelet, and Chirplet—into the learning process. This allows the model to adaptively extract fault-sensitive features while maintaining physical interpretability. Experimental results on the CWRU dataset show that TFDT outperforms standard CNNs and fixed-transform pipelines in accuracy, convergence speed, and robustness under noisy conditions. Ablation studies confirm the critical role of the TFConv layer, and t-SNE visualizations reveal discriminative and compact feature clusters, supporting the model’s interpretability claims.