Predicting equipment utilization in agricultural tractors using field data and machine learning

Determining the implement in use during agricultural tractor operations from onboard data alone represents a practical challenge in field equipment utilization monitoring. This study investigates whether standard tractor CAN-Bus signals are sufficient to automatically identify the active implement without additional sensing hardware or manual operator input. Field tests were conducted with a 105 HP agricultural tractor performing three distinct operations ploughing, rotary tilling, and beet harvesting under real field conditions. A dataset was recorded at 10 Hz via an IoT-based edge-to-cloud telemetry system across five SAE J1939 parameters: wheel-based vehicle speed, engine torque percentage, hitch position, traction load, and engine speed. Random Forest and XGBoost classifiers were trained on the collected field data, and both achieved perfect classification performance on stratified hold-out test data. SHAP-based sensitivity analysis was subsequently applied to quantify the contribution of each parameter to the classification decisions and to validate the physical interpretability of the learned models. Class-level analysis further revealed that each operation is governed by a distinct feature hierarchy: speed is the primary discriminator for ploughing, torque dominates rotary tillage identification, and beet harvesting exhibits a distributed multi-parametric signature. These results demonstrate that routine CAN-Bus field data contains sufficient information to reliably predict equipment utilization, offering a scalable and infrastructure-free approach to implement identification in agricultural machinery.