Inverter fault diagnosis based on Fourier transform and evolutionary neural network

Abstract

The fault diagnosis of the inverter is fundamental to energy intelligence. Due to the complex characteristics of the inverter (e.g., high-dimensional decision and poor stability), it is challenging to solve the problem using traditional fault diagnosis methods. Recently, artificial intelligence (AI)-based approaches have emerged as the most promising methods. However, they often require to set hyperparameters manually, which hinders further AI-based applications in fault diagnosis of inverters. To fill the gap, we propose an inverter fault diagnosis method using fast Fourier Transform (FFT) and evolutionary neural network. This method combines the amplitude of low-frequency harmonic component of the three-phase inverter output current which is obtained by FFT and the average value in a period of three-phase inverter output current into the fault eigenvector. This method uses an evolutionary neural network trained by combining genetic algorithm (GA), ant colony optimization (ACO) algorithm and Back-propagation (BP) algorithm to realize fault diagnosis. This method can effectively resist noise interference and reduce the number of independent variables in the part of feature extraction, so that it can simplify the network model. In addition, this method can avoid the network training from trapping in local optima in the part of fault classification, with high accuracy and fast response speed. The experimental results show that the proposed algorithm and method of fault feature extraction can effectively detect and locate the insulated-gate bipolar transistor (IGBT) with open circuit (OC) fault in three-phase inverter, and can be applied to online monitoring.