An embedded solution for fault detection and diagnosis of photovoltaic modules using thermographic images and deep convolutional neural networks

In this work, an embedded system for fault detection and diagnosis of photovoltaic (PV) modules based on infrared thermographic images and deep conventional neural networks (DCNNs) is introduced. First, a binary classifier is developed for PV modules fault detection. Then, a multiclass classifier is developed to diagnose the type of defects occurred on PV modules. In this study, four common defects are examined: partial shading effect, dust deposit on PV modules surface, short-circuited PV module and bypass diode failure. The developed DCNN-based classifiers have been first optimized and then embedded into a low-cost microprocessor (Raspberry Pi 4). The models have also been compared with three main TF-Lite optimization techniques (Simple conversion, Dynamic range quantization and Float 16 quantization). Experimental results demonstrate the feasibility of the developed embedded system to operate in real-time, and can detect and diagnose anomalies with acceptable accuracy. The trade-off between accuracy and models size has been also discussed. Furthermore, the operator could be notified about the state of the PV array through SMS (phone message) using a GSM module (SIM808), as well as by email. This embedded solution can help to make real-time analyses for decision making (i.e. removing fault, cleaning PV modules, changing PV modules, replacing diodes, etc.).