A dual shared space-driven domain continuous learning network for rotating machinery fault diagnosis in dynamic environments

Departing from the ideal static deployment environment, diagnostic models in real-world industrial scenarios typically require to maintain robust performance across multiple working conditions driven by production demands and external factor. At this point, conventional diagnostic models inevitably suffer from insufficient generalization or catastrophic forgetting during gradual adaptation to new working conditions, thereby failing to effectively meet diagnostic requirements. To this end, this study proposes a dual shared space-driven domain continuous learning network (DSS-DCL) to alleviate the aforementioned limitations and provide reliable and sustainable fault diagnosis for rotating machinery under dynamic environments. Specifically, DSS-DCL primarily consists of a shared parameter space constraint and a shared feature space constraint. Among them, the shared parameter space constraint aims to regulate the distribution of model parameters toward uniformity to enhance the possibility of obtaining a generalized diagnostic model. The shared feature space constraint aims to prompt the model to retain sufficient plasticity for new domains learning while exhibiting excellent stability to preserve diagnostic performance on each observed domain. Integrating the above constraints, DSS-DCL enables continuous adaptation to new working conditions while suppressing catastrophic forgetting of accumulated diagnostic knowledge, rather than allocating a separate proxy for each condition or retraining from scratch through exhaustive data aggregation. Extensive experiments and discussions conducted on two publicly available datasets and one experimental dataset encompassing multiple working conditions validate the feasibility and effectiveness of the proposed method.