Large model for fault diagnosis of industrial equipment based on a knowledge graph construction

Abstract

To address the significant heterogeneity of multi-modal data and the challenges in capturing fault semantics for industrial equipment, a fault diagnosis framework that integrates a time-frequency knowledge graph with the large model DeepSeek-V3 is proposed. Specifically, an unsupervised knowledge graph construction method is designed based on multi-modal vibration data signals. This method mines temporal evolution relationships using dynamic time warping and quantifies the relevance between features and faults via mutual information, thereby forming a dynamic graph representation. Additionally, DeepSeek-V3 encodes the natural language descriptions of vibration features, integrating graph structure and time-frequency map features to achieve collaborative reasoning and diagnosis among text, graphs, and maps. Experimental results show that the proposed method achieves high accuracy and significantly outperforms benchmark models, surpassing traditional methods. The proposed framework, through the deep integration of data-driven knowledge graphs and large model semantic understanding, demonstrates high precision, strong robustness, and transparent decision-making capabilities, providing new insights for intelligent diagnosis of industrial equipment.