A Transferrable and Noise-Tolerant Data-Driven Method for Open-Circuit Fault Diagnosis of Multiple Inverters in a Microgrid
Yang Xia, Yan Xu, Nan Zhou
Abstract
A new data-driven method is developed in this article for open-circuit fault diagnosis of multiple inverters in a microgrid. The diagnosis problem is hierarchically modelled as a faulty inverter localization subproblem and a switch fault classification subproblem. First, eigen-decomposition is conducted for measurement signals (bus voltage and active power) and then certain features are extracted to represent the distribution characteristics of decomposed eigenvalues. By utilizing these features as the input, extreme learning machines (ELMs) are trained to identify the faulty inverter in the microgrid based on a decision-making scheme. Given different manufacturers of the inverter, one trained switch classifier may not work for other similar inverters. Therefore, a transfer learning algorithm, domain adaption ELM, is utilized to train the switch fault classifier at the second stage. The proposed method is advantageous for much stronger robustness under noises and generalization capability compared to existing approaches, as shown in test results.