Agricultural machinery fault diagnosis with fusion of graph convolutional networks and large language models

Agricultural machinery operates long-term in harsh and complex field environments, confronting severe conditions such as high-frequency vibrations, variable loads, high humidity, and excessive dust. These adverse factors easily induce component loosening, corrosion, or aging, seriously endangering operational stability. However, existing fault diagnosis methods for agricultural machinery suffer from two critical limitations: first, they mostly rely on single-level monitoring of independent variables, failing to effectively model the spatial topological correlations and temporal coupling relationships among multi-source sensor parameters; second, they only output binary “normal/abnormal” judgments without providing semantic explanations of fault causes, propagation paths, or actionable maintenance suggestions, resulting in a “detection without interpretation” bottleneck that hinders efficient operation and maintenance. Although some graph neural network (GNN)-based methods attempt to model parameter correlations, they often construct graphs from a single dimension; meanwhile, large language models applied in industrial diagnosis lack adaptation to agricultural machinery-specific scenarios and are prone to “hallucinations.” To address these issues, this study proposes an agricultural machinery fault diagnosis method fusing Graph Convolutional Networks and LLMs, establishing a closed-loop workflow of “fault detection-localization-interpretation-decision.” The core design comprises three parts: First, a Spatial-Temporal Fusion Graph is constructed to model multi-parameter relationships. Nodes in the STFG adhere to a “system-component-parameter” three-level mapping principle, covering five core subsystems of agricultural machinery and 20 key operating parameters, ensuring each node corresponds to a unique physical component. Edge information integrates two types of correlations: Workflow Topology Graph edges and Time Sequence Graph edges. Second, a GCN-based feature learning and graph spectral anomaly extraction module is designed. After comparing network structures with 1–5 layers, a 3-layer GCN is ultimately adopted. High-dimensional node embeddings are generated through neighborhood feature aggregation; these embeddings are then projected into the graph Laplacian spectral domain, where low-frequency energy corresponds to global steady-state changes of the system (e.g., slow speed adjustments caused by engine load) and high-frequency energy characterizes local anomalies (e.g., torque mutations induced by fuel injection faults). Two types of filters are designed: a steady-state filter to retain global trends and an adaptive filter to amplify local disturbances. The anomaly score is calculated as the ratio of high-frequency abnormal energy to low-frequency steady-state energy, normalized using historical data to enhance comparability across time windows. Finally, an LLM-based maintenance decision module adapted to agricultural machinery scenarios is built. When the anomaly score exceeds a dynamic threshold, the system converts structured information (including abnormal nodes, propagation paths, and parameter change trends) into prompts, which are input to the DeepSeek-R1 model fine-tuned via Low-Rank Adaptation to reduce computational costs. Meanwhile, a mechanistic consistency verification mechanism is incorporated, cross-validating outputs with agricultural machinery’s physical topology and typical fault laws to suppress model hallucinations, ensuring generated content aligns with engineering reality and providing actionable maintenance suggestions. Experimental validation was conducted using operational data from a 35-horsepower Shifeng tractor during ridge-sowing operations, collected in Wudalianchi, Heilongjiang Province, in May 2025. The dataset includes three types of natural faults and over 24 hours of continuous data sampled at 100Hz. The dataset was divided into training, validation, and test sets at a ratio of 7:2:1 (using time-slice division to avoid data leakage). Comparative experiments show that compared with Support Vector Machine (Accuracy 80.6%, F1-score 0.798), 1D-Convolutional Neural Network (87.6%, 0.848), standalone GCN (92.8%, 0.863), and Graph Attention Network (95.7%, 0.927), the proposed method achieves an accuracy of 98.5% and an F1-score of 0.970, with an average detection delay of 0.028 seconds. Under noisy conditions (signal-to-noise ratio ≥15dB), the accuracy of the proposed method decreases by less than 3%, while that of traditional methods drops by 10%-15%. Evaluation of the LLM module by six agricultural machinery maintenance engineers yielded an average score of 4.32, with 92% consistency between fault cause explanations and actual scenarios, and a 40% reduction in fault tracing time. Edge deployment tests confirm that the total delay of the diagnostic process is ≤250ms, making it suitable for deployment on agricultural machinery embedded systems. This method effectively addresses the "detection without interpretation" drawback of traditional diagnostic approaches, enhances sensitivity to weak and coupled faults, and provides a practical technical route for intelligent agricultural machinery maintenance, holding significant value for improving operational reliability and maintenance efficiency.