A compressed sensing and CNN-based method for fault diagnosis of photovoltaic inverters in edge computing scenarios

Accurate and real-time diagnosis of the inverter is crucial for the reliability, safety and generation efficiency of the photovoltaic (PV) system. Recently, deep learning (DL) is widely used for accurate diagnosis, which automatically extracts useful features instead of relying on experts. For real-time diagnosis, in an emerging edge computing paradigm, massive data generated by end-device are processed in nearby edge nodes, which reduces energy consumption and latency. However, large amounts of data are required for model training in the cloud, which leads to tremendous burdens of transmission and computation. To tackle these issues, a data-driven diagnosis method based on compressed sensing (CS) and convolutional neural network (CNN) is proposed for open-circuit faults of PV inverters. Based on CS, raw signals are compressed in edge nodes with the optimal compression ratio determined by experiments. The proposed CS-CNN is evaluated on an edge-cloud semi-physical experiment platform. Compared with common compression methods, the accuracy of CS is improved by more than 3%, and the time consumption is only one-fifth. Compared with the CNN of the same structure, the transmission and computation time is largely reduced by an order of magnitude. The developed CS-CNN compresses 85% of data with a test accuracy of 99.18%.