Vibration-Language Model for Fault Diagnosis with Numerically Reliable Evidence

Abstract

Fault diagnosis methods for mechanical equipment should not only identify fault categories but also provide verifiable diagnostic evidence. Existing deep learning models usually output only labels or confidence scores, making it difficult to estab-lish an interpretable diagnostic reasoning process. Large language models have strong capabilities in evidence organization and explanation generation. However, a modality gap exists between vibration signals and discrete language tokens. In addi-tion, key numerical evidence in generated diagnostic reports may be inaccurate or hallucinated. To address these issues, this paper proposes a Vibration-Language Model (ViLM). The proposed method encodes angle-domain waveforms and order spectra into learnable vibration tokens and maps them into the embedding space of a large language model. With signal-description alignment and diagnostic instruction tuning, this architecture enables interpretable fault diagnosis based on vibration-informed language generation. Furthermore, a numerical evidence-constrained de-coding method is designed to embed the computation and backfilling of key numeri-cal evidence into the generation process. Experiments show that ViLM improves fault classification and evidence-supported explanation while maintaining high nu-merical reliability in generated diagnostic reports.