Neural-Network-Based Algorithm for Fault Detection and Identification in PMSM Drive for UAV Electric Propulsion

Together with energy density of storage devices, reliability and safety of power systems still remain critical issues on the path to electrification of aerial vehicles, due to the strict airworthiness certification requirements. Permanent Magnet Synchronous Motors (PMSMs) demonstrated to be applicable for both propulsion and actuation in terms of performances, but the susceptibility to faults of power electronics is problematic. This work deals with the development of a neural-network-based algorithm for the detection and identification of open-circuit faults in PMSM power drives. A low complexity feedforward neural network is trained on simulation-generated data for multiclass classification problem and then combined with a counter-based decision logic to ensure fast, robust, and computationally efficient detection. Feature selection analysis is conducted to reduce the size of the neural network without significant loss in terms of classification accuracy. Real-time simulation tests show that faults can be reliably detected within few electrical periods, achieving a median classification accuracy of 98.1%.