Machine learning approach for online estimation of Li-ion battery State of Health

As the world moves more towards sustainability and reducing greenhouse effects, the age of electric cars is approaching in a quick pace and, batteries, which are the electric heart of these cars, are gaining more and more attention. This includes not only developments related to batteries’ material and structure but also the ones related to charging and monitoring battery health. The analysis and monitoring of cell degradation in batteries is crucial for assuring the safety of those electric vehicles. Alongside this, the expanding advances made in machine learning techniques and their wide range of applications has made it a very powerful tool that can be utilized in understanding the complex behavior of systems like batteries. The main objective of this thesis is to develop and test a machine learning algorithm capable of monitoring and estimating the battery State of Health (SOH) in real-time, with the possibility of deploying this algorithm on a hardware, like in the case of the battery management system of an electric vehicle. This work is composed of two main stages. Firstly, a current profile is applied and features are selected. For the purpose of online and real-time estimation, a discharge train of current pulses is applied through the aging process of Li-ion battery cells. Since terminal voltage of the aged battery drops faster than that of a fresh cell, a combination of features are extracted from the terminal voltage response to the current pulse test. Correlation analysis is carried to decide on the best features that can be used in the estimation of battery SOH. Secondly, the selected set of features is used to train an artificial neural network so that it can accurately classify the health of the battery cell. In order to decide on the best level of battery State of Charge (SOC) at which the training and SOH estimation should be done, two current pulses at three different levels of SOC are utilized and then compared to better understand the effect of SOC level and choose the optimal and more consistent level of SOC. Further analysis is done on tuning of the neural network hyperparameters using metaheuristic techniques like the particle swarm optimization (PSO), in order to find the best combination of number of layers, dimension and gradient descent values. The trained neural network is deployed with the use of code generation from Simulink model on a TI development board to test the real time performance of the trained network on classifying and estimating the SOH level. Analysis of the obtained results showed that all candidate features extracted were highly correlated with the SOH of the battery. Using only five of these features; two knee points, slope, voltage difference and calculated internal resistance, the trained network is capable of classifying the SOH level with very low error rates and showed consistency in the performance when tested on other battery cells. In conclusion, applying the short-term current pulse approach which only takes up few seconds in addition to the help of machine learning algorithm, the developed technique is capable of providing a real-time estimation for battery SOH with great accuracy across all the testing cells.