With the adoption of battery-powered electric vehicles, the need for accurate state of health (SOH) estimation of batteries has arisen, especially to maximize the life expectancy of the battery pack. While most SOH estimators are based on physical state observers, we propose a data-driven machine-learning approach. Typically before the production of battery packs starts, in the design phase, the characteristics of cells are measured in the lab. Once the vehicles successfully reach the market and are in use, they send telemetry data to the manufacturer. This data allows a deeper understanding of the battery pack's behavior in actual use cases. This work explores using these lab measurements to build more accurate state of health prediction models on vehicle fleet data. We describe multiple transfer learning approaches and apply them to three machine learning models: multilayer perceptron, convolutional neural network and extreme gradient boosted trees. The models are trained on two datasets, one from accelerated life cycle tests on cells in a lab and one of the telemetry data from a vehicle fleet in actual use. While training, we leverage sliding windows and features regarding the aging of batteries created together with domain experts. The use of transfer learning successfully reduces the error of the models on the field data. For the MLP and XGboost models, improvements of around 35% on the loss from start measure are achieved. We show that transfer learning can be used to transfer information from laboratory cell tests to a vehicle fleet and therefore increase the prediction accuracy on the state of health of batteries. Additionally, the amount of data needed for training is evaluated in terms of the size of a vehicle fleet and the time range of measurements of this fleet. We highlight the advantages of transfer learning models for reduced datasets and show that good results can be achieved even with little data available.
Mr. Alexander Palmisano, Graduand Process Innovation & Implementation, AVL List GmbH