Development of a preliminary Anomaly Detection solution for Lithium-Ion Batteries exploiting Support Vector Machine Algorithm

This thesis presents the development of a novel application for the detection of anomalies in Lithium-Ion battery cells, utilizing the Support Vector Machine algorithm. The increasing presence of lithium-based energy storage systems, in any industry, demands innovative methods for early detection of anomalous behaviors, enhancing the safety and performance of the systems. The project has emerged in the context of Predictive Diagnosis for electric vehicle powertrains. First, research was carried out in the machine learning field to pick the most suitable algorithm for anomaly detection, then extended computation and tests have been performed to prove the effectiveness. The most challenging, creative and innovative part has been to develop techniques to compute the raw data coming from the dynamic system (battery cell) in order to have an instantaneous view of dynamic behavior. Thus, the machine learning algorithm itself, adopted in this work, does not consider the dynamic behavior of quantities, but outputs a value that considers this aspect. The methodology used starts from a very simple situation consisting on limited data training with one anomaly. After decent results has been obtained, the complexity has increased until a final test with a full drive cycle, which includes several types of anomalies. Finally, an experiment has been carried out in a Real Time (RT) environment, using a Nucleo boards that communicate over CAN bus, simulating as realistically as possible an Automotive environment, in order to see possible criticality. The overall results are satisfactory with a good level of accuracy, but the methods and approach have limitations on generalizing the detection to any type of anomaly, indeed the prior selection of the input and computation of the features is aimed to the detection of specific anomalies.