Particle Swarm and Neural Network Approach for Fault Clearing of Multilevel Inverters

This study presents a machine learning technique for fault diagnostics in induction motor drives. A normal model and an extensive range of faulted models for the inverter-motor combination were developed and implemented using a generic commercial simulation tool to generate voltages and current signals at a broad range of operating points selected by a Particle Swarm Optimization (PSO) based machine learning algorithm. A structured Particle Swarm (PS)-neural network system has been designed, developed and trained to detect and isolate the most common types of faults: single switch open circuit faults, post-short circuits, short circuits and the unknown faults. Extensive simulation experiments were conducted to test the system with added noise and the results show that the structured neural network system which was trained by using the proposed machine learning approach gives high accuracy in detecting whether a faulty condition has occurred, thus isolating and pin-pointing to the type of faulty conditions occurring in power electronics inverter based electrical drives. Finally, the authors show that the proposed structured PS-neural network system has the capability of real-time detection of any of the faulty conditions mentioned above within 20 milliseconds or less.