Machine vision based fault diagnosis of photovoltaic modules using lazy learning approach

Abstract

Machine Vision is an advanced and powerful imaging based technique that has been applied in various fields like robotics, inspection and process control. Machine vision, in industrial terms, is termed as a subcategory of computer vision. The primary aim of the proposed study is to distinguish various visual fault conditions that hinder the performance and life span of photovoltaic (PV) modules using computer vision and a machine learning approach. Literatures state that thermographic and electroluminescence images were used in deep learning for identifying faults in photovoltaic modules. However, the effectiveness of normal RGB images with the fusion of deep learning and machine learning techniques is less explored. This paper deals with the classification of normal RGB images of PV modules acquired using a fusion of deep learning and machine learning techniques. Visual faults like delamination, snail trail, burn marks, glass breakage and discoloration that occur in a photovoltaic module (PVM) were considered in the study. A machine learning approach was used to handle this problem which contains three stages: (i) feature extraction, (ii) feature selection and (iii) feature classification. Initially, the features from the aerial images of PVM (acquired from unmanned aerial vehicles (UAVs) equipped with digital cameras) were extracted using convolutional neural networks (CNN). Secondly, J48 decision tree algorithm was employed to select the features of utmost significance and dominance from the extracted image features. Finally, a set of lazy classifiers like locally weighted learning (LWL), K-star algorithm (KS), nearest neighbor (NN) and k-nearest neighbor (kNN) were adopted to execute the classification task on the selected image features. The classification accuracies of all the aforementioned classifiers were compared and it was found that the k-nearest neighbor classifier achieved a maximum accuracy of 98.95% with a lesser computational time of 0.04 s.