An interpretable integration fusion time-frequency prototype contrastive learning for machine fault diagnosis with limited labeled samples

The rise of Industry 4.0 and Industry 5.0, focusing on digital transformation and human-machine collaboration, has boosted the need for advanced fault diagnosis technologies. These must be interpretable to ensure industrial efficiency, reliability, and safety. However, current methods often rely on single-sensor information, require many labeled samples for training, and struggle to justify diagnostic decisions. These limitations reduce their effectiveness in real-world production environments. Aiming at these problems, this paper proposed an interpretable integration fusion time-frequency prototype contrastive learning (IIF-TFPCL) for machine fault diagnosis with limited labeled samples. First, a data-level fusion method based on integrated Gini coefficient entropy is designed to achieve credible fusion of multi-sensor signals while enhancing the fault characteristics of the fused signals. Second, an interpretable wavelet feature fusion convolutional transformer architecture is constructed to achieve interpretable fault extraction from faulty signals. Then, a dual dynamic pseudo-labeling selection strategy is devised to efficiently choose high-confidence unlabeled samples from the original imbalanced unlabeled data. In this process, a self-attention mechanism is employed to measure the correlation between unlabeled samples and initial prototypes. Finally, a time-frequency prototype contrastive loss is constructed to enhance the discriminative ability and robustness of the network in fault diagnosis tasks. The IIF-TFPCL was validated using fused multi-sensor signals and various original single-sensor signals. The experiments display that it is significantly superior to the remaining seven comparison methods. The experimental analysis demonstrates the excellent fault identification performance and interpretability of the IIF-TFPCL with limited labeled data.