PMSM fault detection using unsupervised learning methods based on conditional convolution autoencoder

The challenges of fault detection and condition monitoring in powertrain systems have become increasingly prominent, particularly with the widespread adoption of fail-operational systems. These systems are pivotal in diverse sectors, including the robotics, automotive industry, and various industrial applications. A critical attribute of such systems lies in their capability to identify non-standard behaviour of the system. This study describes a inovative conditional convolutional autoencoder-based fault detection algorithm for the permanent magnet synchronous motor. The study compares a train process of conditional convolutional autoencoder with a classical convolutional autoencoder. The presented autoencoder structure was designed to be implementable into the target microcontroller AURIX TC397 while providing sufficient recognition capabilities of the interturn short-circuit. Autoencoders are trained on data obtained during healthy motor operation and subsequently used to detect interturn short-circuit faults on the experimental dual three-phase permanent magnet synchronous motor with the possibility of emulating an interturn short-circuit fault. The paper provides insights into the achieved autoencoder inference times and the sensitivity in detecting the fault.