ADVANCED ARTIFICIAL INTELLIGENCE ALGORITHMS FOR BATTERY MANAGEMENT SYSTEMS

The thesis “Advanced Artificial Intelligence Algorithms for Battery Management Systems” addresses the development of advanced Artificial Intelligence algorithms for Battery Management Systems, with particular reference to automotive applications and electric and hybrid vehicles. The work was carried out within the context of an industrial PhD at Marelli, with the aim of bridging academic research and the application needs of the industrial sector. The central problem concerns the accurate estimation of battery states, particularly State of Charge (SOC), internal resistance (R0), and State of Health (SOH), which are fundamental quantities for ensuring the safety, efficiency, performance, and lifetime of the battery pack. The main motivation derives from the growing diffusion of electric vehicles, which makes the role of the Battery Management System increasingly important. Lithium-ion batteries are currently among the most widely used technologies in the automotive sector, but they require advanced control systems because their internal states cannot be directly measured. In this thesis, the Lithium Titanate Oxide (LTO) chemistry is considered in particular, chosen for its high safety, thermal stability, fast-charging capability, and long cycle life. The work is organized into five chapters. The first chapter introduces the industrial and scientific context, defining the research objectives. These include: developing AI algorithms capable of estimating SOC, R0, and SOH using only directly measurable signals, such as voltage, current, and temperature; verifying the robustness of the algorithms under different operating conditions; compressing and implementing the SOC algorithm on an embedded platform; and validating its operation through Hardware-in-the-Loop (HIL) simulations. The second chapter provides the technical background needed to understand the work. It describes the principles of electrochemical cells, redox reactions, the main electrical quantities such as voltage, capacity, energy, and power, and the characteristics of different lithium-ion chemistries. The thesis also explains the operational definitions of SOC and SOH, introducing the equivalent circuit model of the battery, based on an open-circuit voltage source, an internal resistance, and RC networks used to represent dynamic phenomena. The role of the BMS is then discussed: it monitors voltage, current, and temperature, protects the battery, manages cell balancing, and estimates internal states. The final part of the chapter presents an overview of traditional and data-driven methods for SOC and SOH estimation, including Coulomb Counting, OCV, Kalman filters, neural networks, LSTM, CNN, and Transformers. The third chapter describes the development of the virtual environment in Matlab/Simulink used to generate synthetic datasets. Since the work was carried out in an industrial context and refers to a cell parameterized with internal data, synthetic data were generated instead of relying exclusively on public datasets. The virtual environment includes a dynamic vehicle model, an electrochemical battery pack model, and a control block to alternate discharge and charge phases. The considered pack consists of 400 LTO cells connected in series. The cell model is based on a second-order equivalent circuit, validated by comparing the simulated voltage with experimental data. The validation errors are limited, confirming the reliability of the model. An aging model was also integrated into the environment, accounting for capacity reduction and internal resistance increase. The synthetic datasets were generated using standardized driving cycles, including WLTP, UDDS, US06, FTP, HWFET, and J1015. These cycles make it possible to simulate realistic and diversified operating conditions. Synthetic generation reduces time and costs compared with real experimental campaigns, especially for aging studies, which would require months or years of physical testing. The fourth chapter represents the core of the thesis and describes the development of the AI algorithms. The proposed architecture is a cascaded co-estimation approach: first SOC is estimated, then internal resistance R0 is estimated also using the estimated SOC, and finally SOH is estimated using SOC and R0. For SOC estimation, a Decision Tree and an LSTM network were compared. The results show that the LSTM is more robust and accurate, especially in the presence of sensor noise, temperature variations, aging, and impulsive current profiles. Under nominal conditions, the LSTM achieves very low errors, with RMSE, MAE, and MAPE below 1%. The LSTM-based SOC algorithm was also compared with traditional methods used in industry, such as OCV, OCV-Coulomb Counting Mixer, and Unscented Kalman Filter. The LSTM shows generally superior or highly competitive performance, especially in realistic scenarios with noise and dynamic operating conditions. An important element of the work is the compression of the network using PCA-based techniques, which is necessary to meet the ECU memory constraints. The compressed network reduces the number of parameters by approximately 70%. Despite the compression, performance remains good. Subsequently, the compressed SOC algorithm was converted into C code using Embedded Coder, integrated onto an automotive ECU based on the Infineon TC377 microcontroller, and validated through a dSPACE HIL platform. For R0 and SOH estimation, hybrid LSTM + Conv1D architectures were developed. For internal resistance estimation, the number of cycles was identified as a more informative feature than charging time. The networks were trained on synthetic datasets with accelerated degradation down to an SOH of 60%. The results show good accuracy, with limited errors at both 25°C and 35°C. However, the thesis highlights that generalization worsens when the temperature deviates significantly from the values used during training, for example at 20°C, 50°C, 60°C, or 70°C. This confirms the future need for broader and more thermally variable datasets. The fifth chapter concludes the work by highlighting the main contributions. From a scientific perspective, the thesis proposes a complete methodology for the co-estimation of SOC, R0, and SOH on LTO cells. From an industrial perspective, the most relevant contribution is the entire end-to-end workflow: synthetic data generation, AI development, validation, compression, code generation, ECU deployment, and HIL verification. The thesis therefore demonstrates that advanced neural algorithms can be made compatible with the embedded constraints typical of the automotive sector. Future developments include extending the datasets to more variable thermal profiles, implementing the R0 and SOH algorithms on embedded platforms as well, applying the methodology to new chemistries such as solid-state batteries, and integrating it with cloud architectures.