Enhancing the reliability of permanent magnet synchronous motors through data-driven approaches: an industrial perspective

The modern Automotive sector widely adopts Permanent Magnet Synchronous Motors for the propulsion of Electric Vehicles (EVs). As the market for EVs grows, reducing costs and increasing performance in terms of reliability, range, and efficiency becomes paramount. The recent advent of data-driven methods, particularly Machine Learning, offers a great opportunity to evolve the way condition monitoring is performed. With increasingly interconnected machines and growing sensor data, effective processing is essential to extract valuable insights. Data-driven approaches enable the analysis of large, heterogeneous datasets and the automatic learning of useful representations. In particular, Deep Learning (DL) represents the last frontier of data-driven methods, and, due to its effectiveness, it is gradually permeating our daily lives. This thesis was developed in HPE Group, a company that includes engineering, production, and testing departments. In this particular context, the research and application of State of the Art data-driven techniques led to facing challenges closer to the industrial world. In this dissertation, we tackled, through DL methods, two primary applications related to HPE’s electric motors. The first implements a cost-effective, integrated fault detection system using a MEMS vibration sensor with 3.3 kHz bandwidth and an Autoencoder NN. The model achieves an accuracy of 98 % in detecting the incipience of a natural bearing fault. The second application uses Encoder-Decoder Recurrent NNs to estimate, in real-time, the thermal state of the e-motor, which control is crucial for preventing permanent magnets demagnetization. We evaluated the performance of DL models on three motors. Our estimation of permanent magnet temperature achieved a Maximum Absolute Error of 10.11 °C, which corresponds to 5.6 % of the total temperature range. Both models were deployed on the Microcontroller Unit that drives the Inverter, a STM StellarE1, utilizing 19.34 % of flash memory and 13.0 % of available RAM.