Abstract

In industrial applications, Proportional-Integral (PI) controllers are frequently employed for controlling Permanent Magnet Synchronous Motors (PMSMs) due to their fast response rate and comprehensibility. However, their control performance may deteriorate with unforeseen environmental disturbances and uncertainties. To enhance the adaptivity of the controller, Gaussian Process Regression (GPR), a machine learning technique, is used to mitigate the impact of unknown components in system dynamics in this paper. In particular, GPR is adopted to autonomously tune the parameters of the PI controller, composing a novel GPR-based PI (GPR-PI) controller that maintains both interpretability and safety, because of its theoretical prediction error bound. Moreover, the stability of the system is guaranteed under the sufficiently small designed learning rate of the PI coefficients in the gradient descent rule, indicating the tradeoff between stability and adaptivity. Then, the GPR-based PI is compared with the NN-based PI and demonstrates the priority of enlarging the stabilized speed range. Ultimately, the experiments validate the efficacy of the GPR-PI approach, showcasing a significant reduction of over 60% in response time when compared to the standard PI controller.