Probabilistic Semantic Compression for Zero-Shot Fault Diagnosis

Wu, Yifan and Li, Chuan and Xia, Min (2025) Probabilistic Semantic Compression for Zero-Shot Fault Diagnosis. IEEE Transactions on Instrumentation and Measurement, 74: 3568111. ISSN 0018-9456

Abstract

Semantically bridging known and novel domains, zero-shot fault diagnosis (ZSFD) often faces two main challenges. First, deterministic semantic–fault associations fail to capture generalization. Second, automatically generated semantics are flat and redundant, impairing diagnostic inference. To address these issues, a probabilistic semantic compression (PSC) framework is proposed for ZSFD in this work. The approach extracts discriminative features through a contrastive projection head and constructs probabilistic semantics via a semantic projection head. The probabilistic semantics capture the uncertainty during the dynamic evolution of nonstationary fault patterns, replacing hard categorical assignments. A semantic compression module adaptively compresses the flat and redundant probabilistic semantics in a structure with both high interclass discriminability and interclass correlation. Features and semantics are mutually embedded in a shared latent space, where they are aligned based on their probabilistic distributions to enable the detection of novel faults. Extensive validation on three industrial datasets demonstrates strong generalization and effective transfer from bearings with artificially induced damage to those with naturally occurring faults.