A PI-Dual-STGCN Fault Diagnosis Model Based on the SHAP-LLM Joint Explanation Framework.

Area of Science:

• * Artificial Intelligence
• * Machine Learning
• * Industrial Fault Diagnosis

Background:

• * Deep learning models often lack transparency and interpretability in fault diagnosis.
• * Existing methods struggle with clear explanations of diagnostic outcomes.
• * Industrial applications require verifiable and understandable decision support systems.

Purpose of the Study:

• * To develop a transparent and interpretable fault diagnosis model for industrial applications.
• * To enhance the explainability of deep learning diagnostic processes.
• * To leverage large language models (LLMs) for verifiable decision support without direct diagnosis.

Main Methods:

• * Proposed a PI-Dual-STGCN (Physics-Informed Dual-Graph Spatio-Temporal Graph Convolutional Network) model incorporating physical constraints.
• * Constructed a dual-graph architecture combining physical topology and signal similarity graphs.
• * Developed a SHAP-LLM (SHapley Additive exPlanations-Large Language Model) joint explanation framework using Explainable AI (XAI) and Retrieval-Augmented Generation (RAG).

Main Results:

• * The PI-Dual-STGCN model achieved a diagnostic accuracy of 99.22%.
• * The SHAP-LLM framework generated visual explanations and a hierarchical knowledge base.
• * LLMs were used for reasoning over fact-based textual information (performance metrics, SHAP results), avoiding direct diagnosis and hallucinations.

Conclusions:

• * The combined PI-Dual-STGCN and SHAP-LLM approach significantly enhances diagnostic accuracy and transparency.
• * XAI provides visual and quantitative explanations of model decision logic.
• * This method offers verifiable intelligent decision support for industrial fault diagnosis by grounding LLM reasoning.