Data-driven fault diagnosis in the converter system

Abstract

In modern power electronics industry, three-phase pulse width modulation (PWM) converter has been widely applied due to the noise resistance and economic efficiency. To guarantee the steady working state of converter, fault diagnosis method has become an essential and popular area. In this paper, a data-driven methodology based on randomized learning algorithm is proposed to diagnose insulated gate bipolar transistor (IGBT) open-circuit (OC). To exclude the disturbance of frequency variation, RELIEFF algorithm is combined with Fast Fourier Transform (FFT) to select certain frequency components as input. In order to improve diagnosis performance, the hybrid ensemble time-adaptive model is designed to balance the tradeoff between diagnostic accuracy and time. Besides, parameters in this diagnosis model are optimized by multi-objective programming. Furthermore, parameters are selected from solution set of Pareto Optimal Front (POF), which is able to satisfy multiple practical requirements in online application. Simulated and experimental data are utilized in offline model training, as well as validate online performance of the proposed fault diagnosis method.