This activity concerns the development of an innovative methodology for predictive maintenance of agricultural machinery through anomaly detection based on the analysis of sensor data collected from the tractor CAN-bus. The goal is to identify faults and abnormal behaviors in a timely and accurate manner, improving machine reliability and reducing downtime. The methodology relies on deep learning models (autoencoders) specialized by terrain type and activity (e.g., plowing, moving, turning, idling), capable of detecting subtle and context-specific deviations in sensor signals. To address the scarcity of real fault data, an innovative multi-agent workflow based on Large Language Models (LLMs) was developed, capable of generating realistic synthetic anomalies (such as engine overheating, torque instability, battery voltage drop). These anomalies are automatically validated and injected into the datasets to test the system’s reliability.

Results showed high accuracy in classifying operational activities and strong anomaly detection performance, confirming the robustness of the method and its effectiveness in recognizing rare faults that are difficult to observe under real operating conditions.

Our approach represents an advanced and scalable solution for predictive maintenance of agricultural machinery:

• Improves reliability and reduces downtime

• • Enables testing of models even in the absence of real fault data thanks to LLM-generated synthetic anomalies

• Supports the adoption of precision agriculture practices in complex and resource-constrained environments