Abstract: Fault diagnosis is critical for ensuring the safe operation of industrial systems. Traditional diagnostic models typically produce abstract outputs, such as anomaly scores or fault categories, but fail to address key questions like “why did the fault occur” or “how should it be repaired”. Although large language models have shown great potential for fault diagnosis, they face a semantic gap when processing time-series industrial signals, as continuous temporal data are difficult to encode into discrete tokens that language models can effectively handle. Unlike the conventional “signal-to-category” paradigm, this paper proposes a novel framework for fault diagnosis, termed Signals-to-Semantics Fault Detection and Diagnosis (S2S-FDD). First, we design a LLM-based conversion module based on a temporal analysis semantic bank, which transforms abstract time-series signals into natural language summaries by extracting features such as trends, periodicity, and deviations, thereby mapping the abstract signal space to a semantic space. Second, based on the extracted semantic descriptions, an iterative reasoning method is developed to align with expert diagnostic logic. This method leverages LLMs to compare the semantic descriptions with historical maintenance documents, generating preliminary diagnostic hypotheses and actively planning additional sensor data acquisition for multi-round verification when information is insufficient, ultimately producing a traceable fault diagnosis reasoning chain. Experimental results on multiphase flow systems and real thermal power plant processes demonstrate that the proposed method achieves high diagnostic accuracy even in the absence of fault samples. More importantly, instead of merely outputting abstract fault labels, it generates a human-understandable and verifiable reasoning process and diagnostic report, offering a new paradigm for realizing true human-AI collaborative diagnosis.