Adversarial prototype fuzzy-weighted sub-domain adaptation with singular value equilibrium for universal transfer diagnosis of bridge

In recent years, cross-domain bridge structural damage diagnosis has emerged as a burgeoning research direction in structural health monitoring. Most existing transfer learning-based structural damage diagnosis studies focus on closed-set scenarios where source and target domains share identical label spaces, significantly limiting their applicability to real-world bridges. While recent advances address more practical partial-set and open-set scenarios, obtaining prior knowledge of damage patterns in field bridge environments during testing remains challenging. Consequently, source and target domains may share a public set while maintaining distinct private sets, thereby introducing additional class discrepancies, i.e., a highly challenging universal cross-domain damage diagnosis problem. To address this, an Adversarial Prototype Fuzzy-weighted Subdomain Adaptation (APFSA) method is proposed for universal scenarios. Firstly, an adversarial fuzzy prototype joint weighting scheme is designed to quantify sample transferability through domain-wise weighting for deriving the public and private sets. This effectively separates “known damage” from “unknown damage” in the target domain while mitigating negative transfer effects caused by the private set in source domain. Secondly, a multi-channel joint subdomain adaptation strategy is developed to capture fine-grained transferable features of known damage to enhance diagnostic precision. Additionally, to address the degradation of discriminative capabilities during transfer learning, a singular value equilibrium mechanism is proposed to enhance recognition accuracy while ensuring robust transfer performance. Experimental validation across three bridge cases demonstrates the superiority and effectiveness of the proposed method in cross-domain damage diagnosis.