Battery Modelling, Simulation and Artificial Intelligence for Health Monitoring and Anomaly Detection

The objective of this thesis is to simulate electrochemically a set of batteries, based on the configuration of the customer, then to develop and apply machine learning and neural network algorithms to forecast the performance and behaviour of batteries. The work begins with the data acquisition phase, executed through sensors and connectors from the client, and managed through a proprietary platform capable of real-time graphical display, data analysis tools, as well as integration and management capabilities for machine learning solutions. Subsequently, the data undergoes a cleaning and preprocessing phase. Initially, the quality of the acquired data is assessed; however, the raw data obtained from the sensors are typically noisy and contain gaps due to the client's inconsistent broadband connection. Hence, the data must be cleaned and prepared for subsequent analysis. Due to limited availability of data from the customer, an electrochemical simulation of battery performance was conducted, to obtain the dataset required for the machine learning model training and test. Initially, data acquisition was planned over four months, from April to July, allowing two months for the model development and optimization of the forecasting strategy. Although potential strategies were identified in advance, the project strategy was completely changed to deliver this project; hence, a model has been developed from zero, and the algorithm strategy completely revised. These records are processed and selected, due to their different properties, to train, validate and test the model separately. Two strategies are chosen: the first relies on supervised learning, to predict precisely State of Health, the second is the anomaly detection, due to a general lack of precise information in real contexts. The results are commented and discussed, with the possibility of real-life potential implementation after retraining.