Siamese Neural Network Architecture for Fault Detection in a Voltage Source Inverter

Abstract

This study presents a methodology for evaluating the operating conditions of power converters through similarity analysis using the Siamesa Neural Network technique. The time series of the inverter waveforms are collected for 3 cycles, then converted and analyzed using image processing. For the evaluation of the method by simulation, initially the structure of the H bridge three-phase inverter is used, and later the ANPC hybrid topology with asymmetric modulation, both operating in open loop. The line voltage and current waveforms for the first and the line and phase voltage for the second are the parameters for defining and comparing normal steady-state operating conditions and open circuit conditions in the switches. The work presents the promising ability of siamese neural networks to observe the inverter operating conditions and carry out the identification and diagnosis of faults, given the prior knowledge of the circuit behavior, with the advantage of calculating the similarities even without the use of specific data bank for network training.