State-of-health estimation using a neural network trained on vehicle data

The validation of battery aging models in automotive applications requires reliable aging data to compare the accuracy of each proposed model. Using a sample of 704 vehicles aged up to eight years under diverse nominal conditions two aging estimation models are proposed. By analyzing relevant automobile battery data a more relevant fit of a semi-empirical holistic model is provided with an Arrhenius temperature dependence and pseudo-Tafel voltage dependence. As a comparison, a neural network capturing the aging behavior using the most correlated variables available in the data-set was also developed. Over 110,000 measurements from seven relevant indicators are available as aging predictors, as well as, highly-accurate capacity measurements which is used as the ground truth capacity targets to train and validate the proposed models. Against these points the Semi-Empirical and Neural Network models achieved a root mean squared error of 3.4%-SOH and 3.0%-SOH, respectively.