Fault diagnosis in electric drives using machine learning approaches

This paper applies machine learning techniques to fault diagnosis in electric motor drives. As faults in motor drives can cause safety hazards in applications such as electric traction, propulsion, aircraft, and others, it is desired to diagnose the fault, i.e., detect it and isolate it, in order to recover or engage safe-mode operation. Machine learning methods are divided into three main steps: 1) Feature extraction using Principal Component Analysis (PCA); 2) Classification using the k-Nearest Neighbor (k-NN) or Probabilistic Neural Network (PNN) methods; and 3) Classifier performance evaluation using the Cross-Validation (CV) method. While electric machine, inverter, and sensor faults are introduced, the supervised learning algorithms are applied to four case studies where two fault modes occur in a current sensor, and two occur in the speed encoder. Classification accuracy, i.e., the ability to diagnose a fault, for all four cases is shown to exceed 98%. The paper also investigates a load profile used in automotive driving cycles which produces richer dynamic responses that are more interesting from an application perspective. The final goal is to implement these algorithms in real-time such that fault diagnosis can be used to isolate drive faults, especially sensor faults, and operate in a safe mode.