# Prediction of tractor drawbar pull under different tillage tools using machine learning and low-cost sensors

## TL;DR

This study uses machine learning and low-cost sensors to accurately predict tractor drawbar pull for different plows in various soil conditions.

## Contribution

The study introduces a cost-effective approach by comparing input variable combinations for different plows using machine learning models.

## Key findings

- RF in Model E achieved the highest performance for moldboard plow prediction (R2 = 0.977).
- ANN in Models B and C provided strong accuracy for chisel plow prediction (R2 = 0.953).
- ANN in Models B and E performed well for subsoiler plow prediction (R2 = 0.953).

## Abstract

Machine-learning models were developed to predict the drawbar pull of a 78-kW-class tractor for moldboard, chisel, and subsoiler plows. Four models were tested: random forest (RF), extreme gradient boosting (XGB), artificial neural network (ANN), and support vector machine (SVM). The training variables included engine speed (ES), engine torque (ET), travel speed (TS), tillage depth (TD), and slip ratio (SR). Unlike prior studies that focused mainly on engine parameters, this study incorporated nonlinear variables to improve both accuracy and practical applicability. Data were collected from three Korean paddy fields with different soil conditions, and the dataset was divided into 70% for training and 30% for testing. Five input variable combinations were used: Model A (ES, ET), Model B (ES, ET, TD), Model C (ES, ET, TS, SR), Model D (TD, TS), and Model E (ES, TD, TS). The results showed that, for the moldboard plow, RF in Model E achieved the highest performance (R2 = 0.977). For the chisel plow, ANN in Models B and C provided strong predictive accuracy (R2 = 0.953). The subsoiler also performed well with ANN in Models B and E (R2 = 0.953). Overall, the proposed models—particularly RF and ANN—proved effective in predicting drawbar pull and outperformed XGB and SVM. This study is distinguished by its comparison of various input variable combinations for different plows (moldboard, chisel, and subsoiler) and by its proposal of a cost-effective approach using low-cost sensors.