A Controller-Aware Convolutional Neural Network-Based Inverter Open-Switch Fault Diagnosis Technique for Renewable Energy Sources in Isolated Microgrids

This paper presents a novel open-switch fault diagnosis (FD) technique for three-phase voltage source inverters (VSIs) in renewable energy sources in isolated microgrids. Given the critical role that an interfacing inverter's control schemes play in a microgrid's operations, the impact of three grid-forming inverter control schemes (droop control, virtual synchronous generator (VSG) control, and VSG with a Fuzzy secondary controller) is evaluated for inverter open-switch faults diagnosis. In this paper, a dataset of inverter open-switch faults measured in our laboratory through Opal-RT real-time simulator is used, for which a 2 kW three-phase VSI under healthy and 21 single- and multi-open-switch faults conditions is tested, along with the three control schemes and various inverter loading, in an isolated microgrid. To develop robust FD techniques, several methods converting time-series signals to images are evaluated in this study, including Gramian angular summation field (GASF), recurrence plots (RPs), and continuous wavelet transform (CWT). Through these methods, three-phase currents measured at the load terminal are transformed into RGB images, the images are then analyzed using the convolutional neural network (CNN) to classify inverter open-switch faults. The best CNN models achieve 99.86% (GASF), 99.45% (RPs), and 98.84% (CWT) fault classification accuracy. The fault detection time is within 5-8 ms at 15 kHz sampling. The proposed techniques can reliably detect the type and location of inverter open-switch faults under different control schemes, achieving high accuracy and fast detection in isolated microgrids.