Real-time risk assessment framework for tractors: part 1

Tractor rollovers remain a critical safety issue. We propose a risk assessment framework (Part 1) that estimates quantitative risk points (RP) from accident-factor variables available during or prior to operation. Using 1,164 de-identified tractor accident records from Japan and nine accessible variables, we trained a deep neural network regressor. The optimised model (five hidden layers) achieved r = 0.87 and R2 = 0.71, with errors decreasing as RP increased. This model outperformed the baseline models in accuracy. The framework estimates potential risk points of the tractor operational environment and generates a numerical index to inform operators of the prevailing risk level, enabling proactive safety management. Part 2 incorporates a dynamic risk index derived from vehicle-behaviour signals to provide a comprehensive assessment.