Aging Diagnostics of Lithium-ion Batteries using Machine Learning and real Vehicle Fleet Data

MOTIVATION

Most approaches for predicting the SOH under a wide range of conditions have one common drawback: They rely on bench testing, which is tedious and time-consuming, cell-specific and expensive. This rarely reflects the actual usage behavior, a crucial factor in battery aging. Instead, our approach uses real driving data.

Several classic Machine Learning methods, especially Neural Networks, have been adapted for and applied to batteries, but only a small number of publications used actual field data. We are the first to explore how the training of the network could benefit from additional information of an entire vehicle fleet.

METHODOLOGY

In the FeBaL project (“Felddatenbasierte Batteriediagnose und Lebensdauerprognose“, i.e., “Field data-based battery diagnosis and lifetime prediction”), we apply a Neural Network (NN) to real applications.

The Fleet Model consists of three parts. The first part is for data preprocessing and feature extraction. Part 2 is a joined battery model for all vehicles, where information is transferred between the devices. This enables the extrapolation to wider operation ranges. In this model part, we use a structured NN where knowledge about the specific cell (for instance the OCV-curve) can be incorporated. The third part comprises the storage and adjustment of the device-specific aging parameters. This allows for different aging rates for each vehicle, thus taking the different operation conditions into account. Furthermore, it differentiates between calendric and cyclic effects as well as between resistance increase and capacity decrease.

FINDINGS

The Fleet Model is trained and evaluated on a fleet of busses which operate the same NMC-cells. The busses differ, however, in their charging strategy as well as SOC and temperature ranges, so we can fully exploit the benefits of our fleet approach.

Using validation sections, we show that the battery behavior as well as the aging is accurately detected by the Fleet Model. The estimated resistances and capacities are within the expected value range.

To confirm the benefits of the fleet approach, we trained a model with data of only a single bus and compared the results. The Fleet Model improves the prediction accuracy, especially in regions with few data.

DISCUSSION

We accomplished the first milestone of the project: the accurate tracking of the SOH of an entire vehicle fleet. In the second part of the project, we will parametrize a heuristic aging model to enable end-of-life (EOL) prediction. To this end, information from laboratory tests will also be included to cover all operation conditions and to enhance the extrapolation capabilities. In addition, we want to establish an application-specific EOL criterion by simulating specific usage scenarios and/or usage recommendations.