Fault detection, localization and diagnosis are critical for increasing the efficiency and dependability of photovoltaic (PV) plants. In this paper, a fault diagnosis method for PV modules is developed using a Machine Learning (ML) platform (Edge Impulse) and infrared images. The idea is to develop a TinyML model to classify certain defects that can frequently occur on PV modules (e.g., dirty, short-circuit and sand deposit on PV modules) and check the feasibility of the developed TinyML model by integrating it into a low-cost and lower power microcontroller (MCU). In this regard, a database of infrared thermography images was built and used. Then, different MobileNets versions are evaluated and compared in terms of accuracy, hardware resources and inference latency. Simulation and co-simulation results clearly report the feasibility of the method based on two different MCUs, with a maximum classification accuracy of 98%. The key advantage of the used platform is that the embedded ML models can be developed quickly and run locally without internet connection. Moreover, edge processes are not affected by latency, bandwidth limitations, energy constraints, or data privacy issues, making them outstanding methods for real-time diagnostics. To verify the ability of the developed TinyML model to classify defects in real- time, the model has been experimentally evaluated and the results demonstrated its capability to diagnose PV modules with good accuracy.
