Intelligent Dust Detection System for Solar Panels Using Computer Vision

Abstract

Algeria's National Renewable Energy Programme targets 15000 MW of installed photo￾voltaic (PV) capacity by 2035, most of it in Saharan and semi-arid environments where mineral dust routinely cuts panel output by ten to thirty percent. The conventional ways of managing soiling, namely calendar-based cleaning, isolated soiling-ratio reference stations, and manual visual inspection, were built for installations one or two orders of magnitude smaller than the utility-class plants now under construction, and none of them combines the spatial coverage, objectivity, and scalability this setting demands. This thesis investigates an automated, image-based alternative grounded in deep learning. It implements and evaluates a Convolutional Neural Network approach to binary clean￾versus-dusty classification, using a MobileNetV3-Large backbone pretrained on ImageNet and adapted to the target domain through transfer learning, with Contrast-Limited Adaptive Histogram Equalisation (CLAHE) applied to normalise illumination. The model is trained on a merged public dataset of 2029 labelled solar panel images under a production-grade pipeline using AdamW optimisation, cosine annealing with linear warmup, label smoothing, RandAugment, and Random Erasing augmentation. It reaches a best validation accuracy of 97.04 percent (96.55 percent with four-view test-time augmentation), with 97.08 percent precision and 94.86 percent recall on the Dusty class. Deployment via the ONNX Runtime achieves up to a 3.66× inference speedup relative to native PyTorch, reducing per-image latency to 1.87 milliseconds. The system also includes a statistical drift-detection module based on the Kolmogorov--Smirnov test, together with a drift-resilient fine-tuning procedure that lifts accuracy on drifted data from 67 to 85 percent.