Online system identification and excitation for thermal monitoring of electric machines using machine learning and model predictive control

Data-driven thermal models are demonstrated to be practical for real-time temperature estimation in electric drive applications. However, the model’s performance and scope are highly influenced by the information contained in the empirical dataset. So far, thermal modeling and system excitation are separate processes where the system is excited using offline-designed, open-loop experiments that do not consider the system or model responses. This typically requires lengthy experimental measurement sessions at test benches and the missing feedback for the excitation leads to biased data distributions in the state-input space. To overcome these issues, model predictive based closed-loop system excitation is introduced, which identifies a neural network-based thermal model online using the continual data stream. Simulation results show that a significantly improved state-input dataset distribution can be achieved with the proposed solution in less time than traditional open-loop offline design of experiment strategies.