Feature-optimized deep neural networks for multiple fault classification in PMSM drives

Permanent Magnet Synchronous Motors (PMSMs) are extensively used in Electric Vehicle (EV) drives. However, the mechanical and electrical stresses result in different faults during PMSM drive operation. Among these, the Open Circuit (OC) and Short Circuit (SC) faults of the stator winding are the most frequent and severe faults. Especially, the detection and classification faults in EV applications are very crucial at various dynamic operating points (speed, load and noisy conditions). Therefore, this paper proposes a feature-optimised Deep Neural Network (DNN) fault detection and classification method during dynamic operating conditions. Considering healthy, open and short circuit scenarios, the phase currents, torque, and speed signals of the drive are measured through simulation. Further, the hybrid feature optimized framework integrating statistical time-domain, transient dynamic and Fast Fourier Transform (FFT) based spectral features are extracted from the raw signals. Thereby, the proposed DNN model has been trained using the featured dataset generated. Then, the effectiveness of the proposed DNN is analysed by comparing with SVM, k-NN and LSTM networks. The comparison is performed in terms of accuracy, precision, and Area Under Curve (AUC) of the Receiver Operating Characteristics (ROC) curve. The values of accuracy (99.97%), precision (99.96%), and AUC demonstrate the superiority of the DNN model.