Solar Cell Defects Classification: Evaluating CNN Architectures using Local Interpretable Model Agnostic Explanations and Feature Map Visualizations

Abstract

This thesis explores the use of Convolutional Neural Networks (CNNs) for detecting and classifying of defects in solar cell images. With the rapid expansion of solar energy systems, ensuring the reliability and efficiency of solar cells through accurate defect detection has become increasingly important. Traditional manual inspection methods are often insufficient due to their time-consuming nature and susceptibility to human error. Therefore, this study explores the use of automated, deep learning-based techniques to enhance defect detection processes. The research involves the development and evaluation of six different CNN architectures, each varying in the number of convolutional layers, maxpooling layers, and dense layers. The models were trained and tested on a dataset of electroluminescence (EL) images, containing 2,624 samples categorized into four classes: No defect, minor defect, moderate defect, and severe defect. The performance of the models was assessed using accuracy, precision, recall, and loss metrics. To interpret the decision-making processes of the CNN models, Local Interpretable Model-agnostic Explanations (LIME) and feature map visualizations were employed. These techniques provided insights into the specific regions of the images that influenced the models’ predictions, thereby enhancing the transparency and trustworthiness of the AI systems. The results indicated that while more complex CNN architectures have the potential to achieve higher accuracy, they also risk overfitting, especially with limited training data. The study found that simpler models, such as those with fewer convolutional and dense layers, tended to converge more quickly and exhibited stable performance. However, these models sometimes failed to capture intricate patterns in the data, leading to lower recall rates. Comparing these findings with existing literature, the study highlights the importance of balancing model complexity and performance. The integration of LIME explanations and feature map visualizations underscores the necessity for model interpretability in real-world applications. This research contributes to the field of defect detection in solar cells by providing a comprehensive analysis of CNN architectures and emphasizing the need for transparent AI systems.