Abstract: Agricultural machinery can operate under harsh and complex environments in fields. Adverse factors can easily induce component loosening, corrosion, or aging, seriously endangering operational stability, such as high-frequency vibrations, variable loads, high humidity, and excessive dust. However, existing fault diagnosis can suffer from two limitations: 1) Single-level monitoring of independent variables cannot effectively simulate the spatial topological correlations and temporal coupling relationships among multi-source sensor parameters; 2) Binary “normal/abnormal” judgments cannot provide the semantic explanations of fault causes, propagation paths, or available maintenance suggestions, thus resulting in the “detection without interpretation” to hinder efficient operation and maintenance. Some graph neural networks (GNNs) can be constructed to simulate the parameter correlations from a single dimension. Meanwhile, large language models (LLMs) have been applied in industrial diagnosis. However, it is still lacking in the agricultural machinery under specific scenarios, leading to hallucinations from large language models. In this study, a fault diagnosis was proposed for agricultural machinery. A closed-loop workflow of fault detection, localization, interpretation, and decision was also established using GNN and LLMs. Three procedures were included: firstly, a spatial-temporal fusion graph was constructed to determine multi-parameter relationships. Nodes in the STFG followed a "system-component-parameter" three-level mapping. Five subsystems were obtained in the agricultural machinery. 20 parameters were arranged into each node corresponding to one physical component. Edge information was then integrated with workflow topology and time sequence graph edges. Secondly, an anomaly extraction module was designed on a graph spectral sequence using GCN feature learning. Network structures were compared with 1-5 layers. A 3-layer GCN was adopted after comparison. High-dimensional node embeddings were generated and then projected into the graph Laplacian spectral domain after neighborhood feature aggregation. Low-frequency energy corresponded to the global steady-states of the system (e.g., slow speed adjustments caused by engine load), whereas the high-frequency energy was characterized by the local anomalies (e.g., torque mutations induced by fuel injection faults). A steady-state filter was designed to retain global trends, while an adaptive filter was designed to amplify local disturbances. The anomaly score was calculated as the ratio of high-frequency abnormal energy to low-frequency steady-state energy. Historical data were normalized to enhance comparability over time windows. Finally, an LLM maintenance decision module was constructed under agricultural machinery scenarios. Once the anomaly score exceeded a dynamic threshold, the structural information (including abnormal nodes, propagation paths, and parameter trends) was converted into prompts, which were input to the DeepSeek-R1 model fine-tuned via low-rank adaptation to reduce computational costs. Meanwhile, a mechanistic consistency verification was incorporated to suppress model hallucinations. Agricultural machinery’s physical topology and typical faults were output to align with engineering reality for the actionable maintenance suggestions. Experimental data were collected from a 35-horsepower Shifeng tractor during ridge sowing in Wudalianchi, Heilongjiang Province, in May 2025. Over 24 hours of continuous data were sampled on three types of natural faults at 100 Hz. The dataset was then divided into training, validation, and test sets at a ratio of 7:2:1 (using time-slice division to avoid data leakage). The results show that an accuracy of 98.5% and an F1-score of 0.970 were achieved with an average detection delay of 0.028 s, 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 accuracy decreased by less than 3% under noisy conditions (signal-to-noise ratio not less than 15 dB), compared with the conventional method by 10%-15%. The LLM module was also evaluated by six agricultural machinery maintenance engineers. An average score of 4.32 was obtained with 92% consistency between fault cause explanations and actual scenarios. The fault tracing time was reduced by 40% after evaluation. Edge deployment showed that the total delay of the diagnostic process was not more than 250 ms, suitable for the deployment on agricultural machinery embedded systems. The conventional "detection without interpretation" was avoided to effectively enhance the sensitivity to weak and coupled faults. The finding can also provide a practical technical route for intelligent maintenance in agricultural machinery, particularly with its high reliability and efficiency.