Enhanced Solar Defect Detection via Deep Learning: A CNN-Wavelet Transform-LSTM Approach

Nowadays the efficiency of solar energy generation is compromised by the different kinds of solar defects occurred due to regular operations or environmental conditions.Such defects can be visualized by electroluminescence (EL) images.Recently various techniques introduced which are based on image processing and machine learning functions using the EL images.The conventional machine learning methods are semi-automatic which needs the handcrafted features extraction.In this paper, the automatic solar defect detection and classification proposed using deep learning.The proposed methodology consists of three phases which are Pre-processing, Feature Extraction and Reduction, and Classification.Convolution Neural Network (CNN) based features extraction and reduction, and Long-Short-Term-Memory (LSTM) for the classification of solar defects are used.In the pre-processing phase, the distortion correction algorithm introduced to remove the distortions using the special kind of Gaussian filtering and improve the contrast.The distortion correction helps to estimate the more robust and reliable features during the CNN which deliver the improved accuracy of detection.This paper enhances the existing CNN-based feature extraction process by incorporating Wavelet Transform (WT) for improved feature representation and applying Principal Component Analysis (PCA) for feature reduction.This optimization reduces the high-dimensional feature vectors into compact, unique, and smaller-sized representations, enabling more efficient and accurate defect detection.The proposed model in this paper, called D-CNN-WT-P-LSTM, is simulated and evaluated with recent deep learning and conventional machine learning methods.The proposed model, D-CNN-WT-P-LSTM, outperforms existing methods, achieving accuracy improvements of 14% and 20% for 2-class and 13% and 11% for 4-class compared to DCNN and CNN models, respectively.