Time & frequency domain consistency generic learning for fault diagnosis of HST bogie via self-supervised contrastive pre-training

With the outbreak of general large model technique, the development of pre-trained models has been promoted for high-end equipment maintenance (such as high-speed trains, HST). Nevertheless, industrial data often lacks rich corpus, strong logicality and regularity, which leads to potential mismatch between pre-trained models and target sub-domains. To address this challenge, pre-trained models are required to adapt to target sub-domains with various temporal dynamics without seeing any target instances. Considering that information representations in frequency-domain are more stable, embedding time-based neighborhood near its frequency-based neighborhood and performing consistency characterization is beneficial for pre-trained model learning. Hence, this paper proposes a time & frequency domain consistency pre-trained model (TF-DCNet) combined with self-supervised contrastive learning for cross-device fault diagnosis of HST bogies, whose main contributions include: 1) A decomposable pre-training model is defined, which adopts time-space contrastive estimation to bring time-based and frequency-based local neighborhoods close in latent space. 2) A self-supervised pre-training strategy is designed, which creatively utilizes consistency loss to enforce model, transfers relationship knowledge to downstream diagnostic tasks and improves the performance in data of interest with limited fine-tuning. The validity of TF-DCNet is verified in several one-to-many diagnostic tasks, which reflects the breadth of proposed method in multi-task application.