Adaptive generalization: An active perception framework for robust intelligent fault diagnosis under dynamic environments

The distribution shift problem induced by dynamic environments significantly degrades the performance of data-driven fault diagnosis systems in real-world applications. Although domain generalization techniques have been introduced to address distribution discrepancies, static generalized models passively handle all unseen domains and neglect the value of the testing data. To bridge this gap, this study investigates a novel learning paradigm termed adaptive generalization-based fault diagnosis and develops an active perception framework designed to improve robustness under dynamic operating conditions. The developed framework initially fosters universal generalization during offline training and subsequently, during deployment, actively perceives environmental information from the incoming inference data, facilitating targeted adaptation. Perturbation effect minimization is applied to the available multi-source domain data to achieve flat minima, providing a reliable initialization for subsequent online adaptation. The model parameters are then adjusted based on knowledge transferability and structural information extracted from the stream of unseen target domain data, allowing the model to effectively adapt to the inference environment. The effectiveness and superiority of the proposed framework are validated through extensive experiments on three failure datasets. Our code is publicly available at: https://github.com/CHAOZHAO-1/SGTA.