Machine Learning for Force Simulation and Material Identification

Abstract

Virtual product development using numerical simulations is crucial for heavy machinery design. This research focuses on using machine learning models to increase the efficiency of granular loading simulations. In addition, the machine learning models are trained with data from a miniature wheel loader rig. The models predict interaction forces and classify materials using sound and vibration data, offering a robust alternative to vision-based methods in challenging conditions. To improve transparency, interpretable models are compared with deep neural networks. The conclusion is that force predictions made by a regression-based machine learning model in particular show the most similarity to DEM-based force predictions. In addition, nearly all models achieved a test accuracy of over 90% during particle classification, with the perceptron and linear kernel SVM providing the highest accuracy with the shortest training times. Lastly, a conclusion cannot be made as to whether a deep or interpretable model is better suited for the force prediction task. However, either feedback-based force generation is unsuitable or it has to be tweaked more to produce better results.