Prototype-Guided Pseudo-Label Filtering for Dynamic Distribution Adaptation in cross-condition fault diagnosis

Dynamic Distribution Adaptation (DDA) offers a solution for fault diagnosis in wastewater treatment processes. However, DDA typically relies on pseudo-labels of the target domain to achieve domain alignment. Due to domain shift, pseudo-labels generated by the source-trained classifier often contain significant label noise, which accumulates over successive training iterations and ultimately degrades model performance. To address this issue, this work proposes a Prototype-Guided Pseudo-Label Filtering (PGPLF) framework for DDA. To mitigate the introduction of pseudo-label noise, PGPLF abandons the source-trained classifier, which is potentially unreliable, and instead assigns pseudo-labels based on class prototypes. To further filter noisy pseudo-labels, a sample-to-prototype distance-based filtering strategy is proposed, which adaptively selects high-confidence samples for training. Considering the potential value of low-confidence samples, PGPLF adaptively assigns negative pseudo-labels to them and introduces a negative label-based alignment strategy. This allows the model to fully leverage low-confidence samples while avoiding introducing significant noise. The proposed PGPLF was validated on datasets generated by Benchmark Simulation Model No. 1. Experimental results demonstrate that PGPLF outperforms the comparison methods, achieving the highest fault classification accuracy across six domain adaptation tasks. • Pseudo-label noise is effectively filtered in dynamic distribution adaptation. • An adaptive prototype-guided noise filtering strategy is introduced. • Low-confidence samples are assigned negative labels to exploit class information. • A negative label-based alignment strategy enhances domain alignment.