Transferring state of health estimation neural networks for different battery chemistries and charging protocols using renormalization and transfer learning

The State of Health (SOH) is a metric defined for quantifying battery ageing. It is very useful for identifying when a battery has suffered sudden degradation and when it should be replaced. However, accurate estimation of the SOH is not trivial and the use of data-driven approaches such as Neural Networks (NNs) is becoming increasingly common. While promising, these approachs are in principle limited by the fact that new models must be trained for each kind of different batteries, which limit their usability in real-world use case. To address this issue, this work explores two ways of transferring SOH estimation models to batteries with different chemistries and charging protocols than the ones they were trained for: 1. renormalization; and 2. normalization parameter training. We show that updating normalization parameters is sufficient to make models follow SOH evolution, but results usually present an offset and distortion. Optimizing these parameters yields results close to and sometimes better than the reference models, if their protocols are sufficiently similar. Our results lead us to believe that battery chemistry does not influence the model transferring process, but that differences between the training and target datasets’ charging protocols may hinder its success.