For state of the art State of Health (SoH) diagnosis methods a lot of domain-specific expertise and experimental data is needed. Utilizing parameter estimation methods like Kalman filters, the internal resistance and capacity may be estimated. Those methods require a battery model that fits to the battery and the desired application. Established Machine learning methods on the other hand exchange that need of domain knowledge against the need of extensive laboratory experimentation data. This work proposes a new machine learning method for SoH diagnosis that minimizes the amount of required laboratory data while taking no assumptions about the battery chemistry and parameters.

The method solely depends on the during operation directly measurable quantities voltage and current with no need of estimating State of Charge. This self-supervised learning approach is enabled by the concept of virtual tests. Virtual tests apply standardized laboratory methods like current pulse tests for internal resistance measurement to a machine learning model. That machine learning model imitates the behavior of the battery by estimating the terminal voltage in dependency of the current, therefore mitigating the need for labeled SoH data. This self-supervised learning method allows an online adaptation of the model to changes of behavior of the aging battery while being used in the working environment.

For example for estimating the internal resistance of a traction battery in a battery electric vehicle the current and voltage profiles may be measured while everyday driving.[Figure 1] A machine learning model maps the measured current profile to the voltage profile, therefore memorizing the battery dynamics. Because the model needs to track internal states that represent this dynamic and state of charge a recurrent neural network is a good fit for that application. When the model is trained on the data, a rectangular current profile will be applied to the model, representing a virtual pulse test. The internal resistance can be calculated out of the applied current and the voltage drop of the estimated voltage profile.[Figure 2] In the same way a constant current discharge may be applied for capacity estimation purposes. Later, when the state of health degenerated further, the existing model may be trained on more recent data in order to update the SoH. Figure 3 visualizes the described process.

This new method allows the estimation of the SoH with neither specific knowledge about the battery nor elaborate experimental aging data of many instances of the same battery. Instead it utilizes established laboratory methods as is described, for instance, in ISO 12405-4:2018. That way it also does not rely on labeled SoH values for training of the model. Also the accuracy of the method does not suffer from a drift error that increases with the age of the battery. These advantages make it particularly suitable for low quantities and prototypes where extensive laboratory experiments are too expensive or demand too much time.

This work was funded by the German Federal Ministry of Education and Research (BMBF, Funding number: 16EMO0262).