This paper focuses on the development of a Machine Learning (ML) approach to classify the severity of photovoltaic (PV) defects using a low-cost and low-power embedded system. The main idea is to develop a TinyML model to classify four classes of severity (i.e., healthy, low, medium, high) based on PV thermal images and to integrate this model on an embedded platform. For this work, a new dataset of a total of 429 thermography images among healthy and faulty PV modules were collected and analysed. A comprehensive comparison of different type of CNN parameters, image input size and quantization methods have been investigated, considering the hardware constraints like maximum RAM, flash size and latency. Eventually, the hardware prototype has been enclosed in a 3D printed case with an integrated thermal camera and an LCD screen to improve the handling and comfort of use for any type of user. The results indicate that this approach is feasible and accurate, reaching an average accuracy of almost 90%. The main advantage of the proposed technique is that also a non-expert user can analyse its own PV plant in order to prevent possible failures with a very low-cost system rather than expensive thermal cameras without any support from experts in the field.