Examenspresentation för Charlotte De Geer och Alex Matsson

Opponents: Hanna Bähr and Maria Svensson

Abstract
This study aims to explore possible and feasible ways to personalize the existing software within the heavy-duty vehicles at Volvo Group. This will be done using machine learning algorithms, specifically focusing on Long Short-Term Memory (LSTM) neural networks and traditional classification algorithms. The goal is to explore potential approaches for enhancing the software to improve the vehicle's drivability and optimize fuel consumption. The research methodology involved collecting relevant data from the heavy-duty vehicle, including various readings using CAN-bus and map-based data.

This data was preprocessed to train and evaluate the LSTM neural network and traditional classification algorithms. The results obtained were not completely satisfactory. The accuracy and predictive performance of the LSTM models fell short of expectations. Data quality, feature representation, and model complexity were factors that contributed to the results. However, the predictions from the LSTM models were adequate. From the training progress, it is possible to see that the model learns and was able to identify some trends. Furthermore, the classification using the traditional and the LSTM classifiers ranged from 93 \% to 99 \% accuracy.

These findings highlight the challenges and limitations of employing LSTM neural networks and traditional classification algorithms for software adaptation. Further research is necessary to explore alternative approaches, such as using sufficient and more suitable data for transfer- and deep learning. The insights gained from this study help comprehend machine learning applications in heavy-duty vehicles and suggest future research efforts to enhance software adaptation and thus improve vehicle performance.

Welcome!

Charlotte, Alex and Jonas