Source-free adaptation based on information maximization and central prediction separation for fault diagnosis

Fault diagnosis plays a critical role in ensuring the normal operation of machinery. Traditional methods, relying on manual analysis and expert systems, no longer suffice for the demands of intelligent diagnosis. Deep learning provides a new way to solve this dilemma. However, data-driven neural network models require the training set(source domain) and the test set(target domain) to follow similar or identical distributions. Numerous domain adaptation methods have been proposed and applied to tackle domain shift. Nonetheless, these methods require access to source data. In practical industrial scenarios, due to data privacy concerns, enterprises often cannot provide source data but can only share the trained source model.Therefore, this article proposes a passive domain adaptive method. This method can complete model migration using only the source domain model without accessing the source domain data. Specifically, this article trains the feature extractor to maximize domain information and meet the prediction requirements of classifier certainty and diversity. In addition, this article proposes a center prediction separation loss function to separate the centers of different faults, resulting in clearer classification boundaries.Experimental results demonstrate that our method improves the diagnostic performance on the target domain, achieving an average diagnostic accuracy increase of 13.52% on six tasks. For specific transfer tasks, the classification accuracy is boosted by approximately 23%. Furthermore, feature visualization experiments validate that the extracted features exhibit intra-class compactness and inter-class separability. The proposed source-free domain adaptation method contributes to enhancing robustness for model.