Parameter Prediction and Optimisation of Working Element Parameters for a Novel Tracked Multi-Axis in Situ Soil Remediation Device Based on Machine Learning Algorithms

To improve the operational efficiency of in situ soil remediation, this study investigated the operating parameters of the crushing–mixing working element of a novel tracked multi-axis in situ soil remediation device according to the soil contamination characteristics and process requirements. A DEM-based numerical simulation model was established, and response surface methodology (RSM) and machine learning algorithms were further integrated to model the response relationships, predict the evaluation indicators, and optimise the operating parameters. Single-factor experiments were conducted using the dispersion coefficient and soil fragmentation rate as the main evaluation indicators to determine the parameter range for the steepest ascent test. The steepest ascent test was used to rapidly approach the optimal parameter region, and RSM was then applied to establish the nonlinear mapping relationships between the operating parameters and response indicators. On this basis, machine learning models were introduced to further analyse and predict the experimental data, thereby improving the multi-objective optimisation process. A comparative analysis showed that, under the same dataset and evaluation metrics, the machine learning models achieved higher prediction accuracy than the RSM model. Among them, the decision tree model exhibited the best overall performance and provided a more stable optimisation result than the random forest and support vector regression models. The optimal parameter combination, identified by the decision tree model, was a rotational speed of 81 rpm, an average mixing pitch of 195 mm, a descent speed of 0.061 m/s, and an average mixing time of 1.1 s. Under this parameter combination, the dispersion coefficient was 0.171, and the residual bond count was 1511. The comparison of the RSM and machine learning models showed that the machine learning models achieved higher prediction accuracy. The relative errors between the optimal and actual simulation values were 2.56% and 4.92%, respectively. These results demonstrate that machine learning algorithms are applicable to the parameter optimisation of soil remediation working elements. The proposed DEM–RSM–machine learning framework can improve the efficiency and accuracy of equipment development and process optimisation, providing a scientific and technical basis for the development of intelligent agricultural equipment and sustainable agricultural engineering.