AD-FHA: An unsupervised fault diagnosis method with visual scanning and memory retrieval capabilities based on a human information system fusion approach

Unsupervised intelligent fault diagnosis models face several challenges under diverse operating conditions. These challenges include strong coupling interference between faults and operating conditions, insufficient single-scale feature representation, poor classification robustness, and a lack of interpretability. Therefore, we propose an unsupervised fault diagnosis method (AD-FHA) with the fusion of human information systems, and it possesses visual scanning and memory retrieval capabilities. This method simulates human cognitive mechanisms and introduces a selective attention-inspired dynamic sparse feature extraction architecture (ADSF), which uses fault saliency detection and secondary attention screening to selectively enhance fault impact information while dynamically suppressing operating condition-irrelevant interference; simultaneously, we propose the Adaptive Multiscale Sparse Perception Enhancement (MSSE) theory, which comprehensively characterizes diverse fault modes — such as weak impacts and periodic modulation — through multiscale parallel branching and adaptive weight fusion of feature energy; Finally, a memory retrieval-inspired hash-based diagnostic method (FHA) is designed. By constructing a standard fault memory database and performing Hamming distance matching, it achieves unsupervised robust diagnosis while ensuring traceability of the decision-making process. Experiments conducted on two cross-domain bearing datasets and one real-world industrial composite failure dataset demonstrate that AD-FHA outperforms mainstream methods in unsupervised domain adaptation, diagnosis with sparse samples, and noise-robust diagnosis, providing preliminary validation of its effectiveness and engineering potential across various operating conditions.