Prediction of Vehicle Component Weights using Multi-Output Regression

Abstract

With the ongoing digitalisation, many companies are exploring opportunities to improve operational efficiency through machine learning. Volvo Cars is part of this trend, and their Weight Management and Optimization team seeks an automated way of managing data for component weights of vehicles. The team performs data harmonisation as a preliminary step, prior to production, which is currently computed manually in Excel. This is inefficient and highly time-consuming, however, machine learning models are believed to facilitate the process. Hence, this thesis aims to investigate the feasibility of implementing a machine learning model for automated data harmonisation, in combination with a pipeline for managing the weight data and calculations. The problem was formulated as a multi-output regression problem, and neural network models and ensemble models were deemed appropriate for the purpose. A significant part of this thesis involved experimenting with different preprocessing techniques and model configurations to find the optimal model performance. The results of the experiments indicated that the ensemble models outperformed the neural network models, where XGBoost was the top candidate. The best-performing XGBoost model achieved a validation accuracy of 84% which was promising, however, the test accuracy was modest, reaching only 22%. This performance was not considered sufficient for the model to be implemented as a part of the final pipeline, since the effort required for deployment of the tool was believed to exceed the benefits. Nevertheless, there remains potential to optimise the model further, and this study can be considered a guideline for further work within this field. The results do not suggest that the use of ensemble models for predicting vehicle component weights is unsuitable, but rather that the optimal way of structuring this problem was not found.