Deep Neural Network and Feature Extraction Approach for Anomaly Detection and Classification in Infrared Images of Photovoltaic Systems

This work presents a solution based on a deep neural network for detecting and classifying faults in photovoltaic modules using thermographic infrared images comparing the performance of different machine learning techniques. Three scenarios were evaluated, those being the identification of the presence of faults, whether the faults are permanent or not and finally a classification of the type of fault. The core of the proposed methodology utilizes a Convolutional Neural Network (CNN) for classification. Initially, a relatively deep CNN is developed as a standalone system for fault detection and classification. Subsequently, the feature extraction layers from the proposed CNN are employed to convert dataset images into more straightforward vectors of attributes, which are then used to train other more general classifiers. This approach aims to create easier and promising learning conditions for any classifier, especially those that are not intrinsically devised to operate directly with digital images as input data. The results emphasize the strengths and weaknesses of the evaluated models and encourage further research investments on the proposed feature extraction approach.