Predicting Power Loss for future Axle Designs using Neural Networks

Sammanfattning

This thesis investigates whether neural networks can predict axle power loss from experimental drag-loss data provided by Volvo Construction Equipment. Four approaches are compared: a black-box neural network, a physics-informed neural network, and two grey-box models based on simplified physical assumptions. The models perform well when trained and tested on randomly split data, showing that neural networks can capture patterns within known axle configurations. However, performance drops in harder tests such as block and holdout splits, meaning the models do not yet generalize reliably to unseen axle designs. The study concludes that more standardized data and more geometric configurations are needed for industrial use.