Spatial Prediction of Tractor Noise and Detection of Atypical Sound Pressure Levels Using Machine Learning

Abstract

Prolonged exposure to high sound pressure levels (SPLs) in agricultural tractors, especially when the power take-off (PTO) is engaged, poses serious health risks to operators. While machine learning has been applied in agriculture, little progress has been made in predicting SPLs in agricultural tractors or detecting patterns that may indicate early mechanical issues. The main goal of this study was to spatially predict SPLs and identify observations whose SPLs departed from the model-expected level for normally operating tractors under different stationary pre-check conditions, as a preliminary basis for future condition-monitoring applications. SPLs were measured at the operator’s seat and at distances up to 15 m in three tractors at PTO speeds of 540, 750, and 1000 rpm. Statistical tests and machine learning models were used to model spatial noise distribution. The RF model performed best (R2 = 0.93, RMSE = 1.65 dB(A)), with spatial position being the key predictor of SPL. In this sample, open-station tractors exceeded 100 dB(A) at and near the operator’s position under the loudest tested conditions, far above the 85 dB(A) screening criterion, whereas no statistically significant effect of PTO speed was detected under the tested conditions. Because the reported values are short-duration source measurements rather than full-shift personal exposures, they are interpreted as screening indicators of potentially hazardous conditions rather than as a formal daily noise-dose assessment. In addition, an out-of-range analysis identified observations exceeding the 99th percentile of positive residuals as atypical, indicating SPLs substantially higher than expected for a normally operating tractor. The proposed approach demonstrates a novel way to identify screening distances and conditions, providing a practical tool to protect workers’ health, monitor equipment, and inform maintenance strategies.