Hi-LLM: a physics-anchored heterogeneous interaction large language model for robust health monitoring of special equipment
Аннотация
Large Language Models (LLMs) have been widely explored for time-series forecasting and classification, yet their application in fault diagnosis of special equipment remains limited by a fundamental input-paradigm dichotomy. Existing methods generally rely either on text-converted features or raw sensory signals. The former provides compact engineering semantics but loses critical information in weak fault scenarios, whereas the latter preserves signal fidelity but lacks semantic grounding. To address this tension, this study proposes the Heterogeneous Interaction Large Language Model (Hi-LLM), a physics-anchored model for vibration-based diagnosis of electromechanical subsystems. Hi-LLM combines a natural-language description of engineering diagnostic features with projected patches from the same vibration signal. The text representation supplies sample-specific semantic context, while signal patches retain local waveform details. Experiments on three rotating-component datasets evaluate the model under in-domain, cross-condition, cross-domain, and limited-target-data settings. The reported performance supports the practical value of jointly presenting engineering semantics and fine-grained signal information to the LLM under the evaluated protocols. This setting is described as direct transfer without target-condition samples. The results highlight the potential of Hi-LLM for diagnosis under varying operating conditions and limited target data.