Semantic information hybrid distribution calibration-enabled multi-modal fusion network for unsupervised health state diagnosis of manipulator

Multimodal data fusion-driven intelligent health diagnosis is integral to predictive maintenance of mechanical equipment, yet it confronts two critical practical hurdles: unverified credibility of disentangled modal-invariant features and the sacrifice of fine-grained critical information for macro-consistency. These challenges significantly hinder its performance improvement and broader adoption. Inspired by this, a novel unsupervised fusion framework — the Semantic Information Hybrid Distribution Calibration-enabled Fusion Network (SIHDC-FN) — is developed, with its core comprising two modules: the Semantic-guided Vibration-Acoustics Joint Disentanglement (SVAJD) Module and the Category-aware Fine-grained Distribution Calibration (CFDC) Module. Within each module, maintenance log information — textual data that, despite being frequently overlooked, records the real health state of manipulators — is treated as a bridge endowed with authentic semantic attributes and an ideal calibration anchor, a dual role that enables it to facilitate, on one hand, the credible disentanglement of modal-invariant information at the macroscale. On the other hand, through the incorporation of a multivariate variational distribution joint constraint strategy, it further ensures the aligned enhancement of detailed features at the class-aware fine-grained scale, a process that ultimately preserves the subtle critical information inherent in modal-invariant features. The comprehensive performance of the proposed fusion method — encompassing feasibility, superiority, robustness, and anti-interference capacity — is thoroughly verified via multi-scenario tests on a typical 3-PRR planar parallel manipulator. © 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.