Calibrated source-free adaptation for intelligent diagnosis

Abstract

The domain adaptation-based fault diagnosis method has achieved an appealing cross-domain fault diagnosis performance. However, source domain data is often inaccessible due to data privacy concerns or the need to reduce the burden of data storage and transmission. Additionally, the credibility of diagnosis results is rarely considered, which is indispensable for decision-making, especially in critical safety scenarios. To address these challenges, this paper proposes a novel cross-domain fault diagnosis method called Calibrated Source-free Adaptation Diagnosis (CSAD). Specifically, a pseudo-label learning-based model adaptation diagnosis framework without the assistance of source data is formulated, where a determinant-based mutual information regular is developed to mitigate the adverse impact of noisy pseudo labels. Furthermore, an unsupervised target-mimic-oriented model calibration method is devised for more credible diagnosis results in source-free scenarios. Comparative experiments are conducted to validate the effectiveness and superiority of the proposed method. In terms of diagnosis accuracy, CSAD achieved average accuracies of 99.42%, 99.60%, and 98.61% on the three datasets, respectively, surpassing other methods. Regarding model calibration performance, our approach reduced ECE by 4.56%, 5.58%, and 4.62% on the three datasets compared to the uncalibrated models.