基于RGB成像的太阳能光伏故障分类深度学习及预处理技术比较

Abstract Effective fault detection in solar photovoltaic (PV) systems is essential for ensuring optimal performance and maintenance. This study explores how various image preprocessing techniques affect the accuracy of a deep learning-based classification model using RGB images of PV panels (bird droppings, dust, physical/electrical damage, snow, clean) were collected from Kaggle. Each pixel’s R, G, B values capture visual patterns, later enhanced using preprocessing. RGB thus provides the raw input for CNN classification. The work demonstrated that raw RGB images alone yielded only 85–89 % accuracy, but when combined with preprocessing (grayscale conversion + Gaussian blur), performance improved significantly up to 94 %. It mainly focusses on the variations in the balanced and unbalanced dataset where data augmentation was used to improve the class distribution. The CNN model, with 5 convolutional layers, pooling, flattening, 2 dense layers, and a softmax output, classifies six PV fault types. Using an 80:20 train-test split, the augmented balanced dataset (12,000 images) achieved higher accuracy (78.75 %) than the unbalanced one (67.24 %). The CNN achieved training accuracy of ∼ 94 % and validation accuracy of ∼ 91 %, showing competitive performance with fewer computational resources. Further, five preprocessing methods like Grayscale conversion, Gaussian Blur, Canny Edge Detection, Thresholding, and Histogram Equalization were applied to the balanced dataset to assess their impact on model performance. Among these, Grayscale conversion and Gaussian Blur produced the highest accuracies of 80.68 % and 80.42 %, respectively. These results highlight the importance of preprocessing choices in enhancing fault detection performance using RGB imagery in PV systems.