Estimating Temperature in a Permanent-Magnet Synchronous Motor Using Hammerstein and Nonlinear Autoregressive Models Initialized via Thermal Networks

Monitoring the temperature of permanent-magnet synchronous motors is crucial to prevent failures in sensitive components such as windings and permanent magnets. In this respect, machine learning techniques have been used to generate models to estimate the temperature of rotor and stator hotspots. However, the effective use of data-driven methods requires large datasets, extensive training time, and substantial computational power. Moreover, machine learning methods mostly operate with a black-box approach; they do not account for the physics of the system to be modeled. This paper proposes and compares Hammerstein and nonlinear autoregressive exogenous models to estimate the temperature of the permanent magnets and windings of an out-runner permanent-magnet synchronous motor. A linear time-invariant component, used for both the Hammerstein and nonlinear autoregressive exogenous models, is initialized via a previously identified fourth-order lumped parameter thermal network. This model accounts for the thermal behavior of the machine. The nonlinear component is modeled via a neuron sigmoid network. Results show that the Hammerstein model achieves a lower mean squared error for the winding temperature estimation than the nonlinear autoregressive exogenous model. The opposite is true for the magnet temperature estimation.