Diffusion Bridge-Based Gradual Domain Adaptation for Fault Diagnosis: Concept, Algorithms and Applications

Abstract

Gradual domain adaptation (GDA) has been recognized as an effective strategy for alleviating domain shift by decomposing large distribution discrepancies into a sequence of progressive adaptation steps, thereby enabling smoother knowledge transfer from source to target domains. However, existing GDA methods still face two major challenges: the construction of intermediate domains is often heuristic and poorly constrained, which may result in unstable or unreliable adaptation trajectories; meanwhile, semantic information can be progressively distorted during gradual adaptation, leading to distribution misalignment and performance degradation. To address these issues, a diffusion bridge-based data adaptation (DBDA) framework is proposed in this article. Specifically, a forward-reverse diffusion bridge is constructed to progressively map target-domain features toward the source domain in a scalable and flexible manner. In addition, a knowledge retention mechanism is incorporated to preserve semantic consistency throughout the adaptation trajectory. Furthermore, a history-aware boundary extension loss is developed to enhance decision robustness under out-of-distribution conditions. Extensive experiments on cross-domain pipeline fault diagnosis datasets, as well as the public sleep european data format (Sleep-EDF) dataset, demonstrate that the proposed DBDA framework consistently outperforms existing GDA approaches in terms of accuracy, generalization capability, and stability.