Rotor temperature estimation in Induction Motors with Supervised Machine Learning

Abstract: The electrification of the automotive industry and artificial intelligence are both growing rapidly and can be greatly beneficial for a more sustainable future when combined. Induction machines exhibit many complex relationships between physical and electromagnetic properties that must be calculated in order to produce the correct quantities of torque and speed commanded by the driver. This is why calculations that depend on sensory information are often cross-monitored and supervised to prevent unsafe conditions or damage to the equipment. Safe torque estimation has a substantial role in safety which requires the fulfillment of ASIL C defined by ISO 26262. The calculation of safe torque is based on rotor temperature among other safety parameters. Traditional methods of obtaining rotor temperature include thermal models, state observers, and active parameter estimation. These methods rely on complex mathematical equations that have the risk of being incorrect and can sometimes be unfeasible in a practical environment. Naturally, we investigate whether we can embed Artificial Neural Networks in the software since we know that they can solve complex non-linear problems exceptionally well when combined with supervised machine learning. To supervise and train the network, we must first acquire the rotor temperature in an experimental setting with a temperature sensor. Then we embed the model into the software of an electrical inverter produced by Inmotion using a microcontroller framework. This way, predictions of rotor temperature can be made in a live environment without the sensor. Using the mean squared error of the output and k-fold cross-validation we can apply a corrected t-test to make a comparison and statistical evaluation of the models. The results in this research prove that a machine learning model can in fact be used to replace the current traditional state observer model that is based on stator temperature. We find that when stator and rotor temperatures are uncorrelated and different, the machine learning model is able to generalize much more accurately passing the t-test with an alpha threshold of  α  = 0.05. Results also reveal that the obtained rotor temperature can be used as reliable input for estimating safe torque by evaluating the measurements from a live motor with a realistic safety requirement.