Source-free universal domain adaptation for compressor component fault diagnosis guided by hybrid clustering strategy

Universal domain adaptation has emerged as a promising approach to address fault diagnosis challenges in industrial scenarios, garnering significant attention in recent years. However, existing methods typically process data from different sources in a unified manner, neglecting the issue of data silos and their impact on domain adaptation performance. Thus, this study develops a novel source-free universal domain adaptation method. The approach aims to enable efficient transfer of shared fault patterns across domains and accurate identification of novel fault modes while preserving data privacy. Specifically, a global clustering strategy is designed to generate pseudo-labels for target domain samples, tailored for universal domain adaptation tasks. This is complemented by a suppression mechanism to mitigate the interference of private fault patterns from the source domain effectively. Additionally, a local consensus clustering strategy is introduced to fully exploit the intrinsic structural characteristics of target domain data, thereby improving the accuracy of pseudo-label assignment. Finally, a contrastive learning-based unknown category identification strategy is established, significantly enhancing the model's ability to identify novel fault modes within the target domain. Experimental results on multiple domain adaptation tasks involving compressor components demonstrate the superiority of the proposed method over other algorithms. The method exhibits higher accuracy and improved generalization capabilities when addressing diverse domain adaptation challenges, further underscoring its practical value and effectiveness.