Source-Free Black-Box Adaptation for Machine Fault Diagnosis

Abstract

Despite the impressive process of current domain adaptation-based fault diagnosis approaches, access to source data and source model parameters is a sine qua non, resulting in obvious limitations when deploying to real industry, particularly considering the data storage, transmission, and privacy issues. In light of this, an interesting and challenging diagnosis scenario is studied in this article, i.e., source-free black-box adaptation diagnosis (SBAD), where only the model output information from the source domain is available for target tasks. To address this issue, a novel diagnosis framework named knowledge transfer from distillation to adaptation (KTDA) is proposed accordingly. Without source data and source model details, KTDA first develops a decoupled self-distillation mechanism to distill source domain knowledge from the black-box model's outputs to the target model, in which the noisy knowledge is simultaneously dealt with by the global and local self-regularization. In addition, a self-adaptation strategy is presented to further adjust the model, where the unlabeled target data is treated differently to reduce the intradomain divergence for improving the fit to the target task. Note that, the target model is not restricted to be the same as the source model in KTDA, thus having more flexibility and versatility in realistic industrial applications. We conduct a variety of fault diagnosis tasks for performance verification, empirical evidence shows the effectiveness and prospect of our method.