Deep learning-based state-of-health estimation with battery charging data

Summary:

The scope of this work is the development of a data-driven capacity estimation model for cells under real-world working conditions with recurrent neural networks having long short-term memory capability. Voltage-time sensor data from the constant current phase charging curve is used as input, reflecting input availability in the real world. The network achieves a best-case mean absolute percentage error of 0.76% and is extremely robust while handling input noise. It also has the ability to handle variations in the length of the input time series and can generate a viable estimation even with an incomplete collection of input due to sensor errors. The model validation with several scenarios is done in a local embedded device, highlighting the use case of such models in future battery management systems.

This work aimed to create a capacity estimation model that is suitable for application in electric vehicles under real-world operation, to be associated with future battery management systems with the capability to take advantage of intelligent health prognostics algorithms. At the time of the literature review for this work, no comparable work was found in the same domain, which implements LSTM networks to use raw voltage-time sensor data from partial CC-phase charging curves to obtain remaining capacity estimation as output per cycle. The validation of the model with a market-ready machine learning capable embedded device is also done to show the viability of acceptance of deep learning-based prognostics in future battery management systems. The cloud-based training and serving to localized devices for implementation makes model operation and updating a seamless process with no change to the architecture needed even with new data, the model parameters only need to be updated online with new training and then be distributed.