TFD-former: Time-frequency domain fusion decoders for effective and robust fault diagnosis under time-varying speeds

Abstract

Research on fault diagnosis using deep learning methodologies is crucial for ensuring the safety and efficiency of industrial systems. In recent years, numerous analytical approaches have been developed for processing time-domain signals, frequency-domain signals, and time-frequency analysis. This study demonstrates that time-frequency domain fusion decoders exhibit remarkable effectiveness and robustness in fault diagnosis applications. We propose TFDFormer, a novel fault diagnosis framework built on two key designs. First, we introduce a lightweight encoder based on the CNN-Transformer structure, to extract signal features from both time and frequency domains, ensuring precise alignment across these domains. Additionally, we develop a contrastive learning loss function specifically tailored for time-frequency domain embeddings to enhance the model's performance. Second, we employ cross-attention mechanism in the decoder to facilitate efficient feature fusion, enabling seamless integration of domain-specific information. To validate the effectiveness of domain fusion and the diagnostic accuracy of TFDFormer, we evaluate it across three prominent fault diagnosis scenarios: fault diagnosis under time-varying speeds, domain generation fault diagnosis, and fault diagnosis on small-scale datasets. The experiments reveal that time-frequency domain fusion decoders significantly enhance model's capabilities in addressing complex diagnostic tasks. TFDFormer consistently outperforms state-of-the-art methods in terms of accuracy, robustness, generalization, and computational efficiency, demonstrating its superiority in fault diagnosis applications.