Predictive maintenance in agricultural machinery: Activity-aware anomaly detection with language model-generated synthetic faults

The increasing digitalization of agricultural machinery within the Agriculture 4.0 paradigm has led to widespread deployment of smart tractors equipped with high-frequency, geo-referenced sensors. These systems generate large volumes of multivariate time-series data, creating new opportunities to enhance operational reliability. However, anomaly detection in this context remains challenging due to high data dimensionality, spatiotemporal complexity, strong context dependence of faults, and limited labeled failure data. To address these challenges, this work proposes a spatiotemporal anomaly detection framework that integrates terrain- and activity-specific autoencoders with a synthetic anomaly generation pipeline powered by Large Language Models (LLMs). Tractor sensor data are partitioned by terrain and clustered to uncover recurring activity patterns, enabling fine-grained, context-aware anomaly detection. To overcome the lack of real fault data, we introduce an LLM-driven, rule-based approach to generate realistic synthetic anomalies for systematic testing. The framework is validated on multivariate time-series data collected from the Controller Area Network bus of a New Holland T7.165 S tractor operating under diverse agricultural conditions. Experimental results show that context-specific autoencoders achieve high detection accuracy across a range of operational scenarios, while LLM-generated synthetic faults enable robust evaluation. Overall, by combining context-aware anomaly detection with LLM-driven fault generation, the proposed framework enhances predictive maintenance in precision agriculture and provides an effective approach to improving the reliability of modern agricultural machinery.