An Improved Convolutional Neural Network Algorithm (ICNN) With Channel Attention Mechanism and Hardware Acceleration Concept for Solar Panel Dust Detection

The efficiency of photovoltaic power generation is significantly affected by dust accumulation, yet existing image recognition methods struggle to achieve both real-time performance and accurate feature extraction under resource-constrained environments. Traditional CNN-based solutions are either too computationally intensive for embedded devices or too simplified to remain robust in complex outdoor conditions. To address this challenge, this study proposes an Improved Convolutional Neural Network (ICNN) that integrates a channel attention mechanism with hardware-aware acceleration. By combining FPGA-based parallelism, pipelining, and fixed-point operations with adaptive channel recalibration, the proposed ICNN achieves real-time inference with low power consumption while maintaining high classification accuracy. Experimental results on multi-weather photovoltaic datasets demonstrate that ICNN achieves 85.83% Accuracy, 92.82% AUC, and an 82.53% F1 score, outperforming traditional machine learning models (SVM, RF, XGBoost), classic deep networks (ResNet50), and lightweight CNNs (MobileNet, EfficientNet, ShuffleNet). Deployment in a 10 MW solar plant further validates its practical value, showing reduced annual maintenance costs (~$128,000) and improved energy efficiency (+4.6%). This work establishes the first FPGA-embedded, lightweight solution for photovoltaic dust detection, bridging theoretical innovation with practical engineering applications.