Multiple Source-Free Domain Adaptation Network Based on Knowledge Distillation for Machinery Fault Diagnosis
Ke Yue, Jipu Li, Zhuyun Chen, Ruyi Huang, Weihua Li
Abstract
Data privacy protection is a hot-button issue in the field of intelligent fault diagnosis. For this purpose, plenty of methods are recently proposed to adapt a machine learning model to a target domain without any labeled data from the target domain or access to the source domain’s data distribution, which is called source-free domain adaptation (SFDA). However, existing methods generally focus on SFDA with a single source domain and the fault categories are often inconsistent between different working conditions. A natural idea is to derive the fault knowledge of different fault categories from multiple source domains. Therefore, a knowledge distillation based multiple source-free domain adaptation framework (KD-MSFDA) is proposed in this study. To be specific, multiple source predictors are pre-trained locally and transferred to the target domain. A KD with predictor confidence vote process is designed to filter the invalid source domains, which can extremely help extract more reliable unitive expert knowledge. Meanwhile, a knowledge contribution-based domain weight adaptation strategy is proposed to automatically assign the weight of each source domain. Extensive experiments on an automobile transmission dataset and a bearing dataset are designed to demonstrate the proposed framework. And the experimental performance verifies that the proposed framework is effective for multiple source-free domain adaptation scenarios.