Multi-Stage Contrastive Causal Learning Framework for Rolling Bearing Fault Diagnosis Under Different Working Conditions

Deep learning methods have shown remarkable performance in intelligent fault diagnosis. However, traditional models often rely heavily on large amounts of labeled data and exhibit limited generalization capabilities across different operating conditions. To address this issue, this paper revisits the latent representations of fault data from a causal perspective and proposes a structural causal model to guide the decoupling of time-domain and frequency-domain representations in deep learning models. Based on this, a multi-stage contrastive causal learning diagnosis framework is constructed. This framework leverages self-supervised time-frequency domain contrastive learning and supervised multi-domain contrastive learning to explore general causal representations, thereby effectively decoupling the features of fault data. Finally, by fine-tuning the model and training the classification head, the fault classification task is accomplished. Experimental results demonstrate that the proposed method achieves outstanding diagnostic performance on multiple fault-bearing datasets, showcasing its potential for widespread application in complex industrial scenarios.