Investigating Data Drift in Federated Learning : A Case Study on Temperature Prediction in Electric Powertrains

Sammanfattning

Electric powertrains are critical components in the advancement of electric vehicles. Accurate temperature prediction for the drive is essential to prevent failures and ensure reliable operation. While traditional federated learning methods can be employed to monitor temperatures across different drives, the presence of data drift—both across drives and over time—can render these methods ineffective. This thesis explores the application of personalized layers and clustering strategies to detect and mitigate data drift in federated learning. Using a dataset comprising laboratory-recorded data, we compare and analyze the performance of these methods against traditional federated learning. The results demonstrate that our proposed approaches significantly reduce the adverse effects of data drift, achieving performance levels comparable to scenarios without drift.