A Fault Diagnosis Method for Axial Piston Pumps Based on Temporal-Textual Feature Alignment Using Large Language Models

Axial piston pumps serve as critical power components in hydraulic systems, and their fault diagnosis is of great practical significance. However, the diversity of operating conditions and environments limits the effectiveness and generalizability of traditional diagnostic approaches. Large language models (LLMs) offer a promising solution to improve model generalization, but their application to time-series vibration signals presents significant challenges. To address this issue, this paper proposes a novel fault diagnosis method for axial piston pumps based on temporal-textual feature alignment using LLMs. First, raw time-series signals are mapped into the textual domain through a reprogramming mechanism, enabling cross-modal alignment between temporal and textual features. To further enhance the LLM’s reasoning capability over time-series data, the proposed method incorporates prompt engineering, which not only enriches contextual information but also guides the model’s interpretation of reprogrammed inputs. Finally, fault states are identified by interpreting the LLM’s outputs on the transformed time-series data. Experimental results on an axial piston pump dataset validate the superior performance and robustness of the proposed approach across varying operating conditions.