Study on an enhanced YOLOv9 algorithm for detecting stains and damage in photovoltaic panels

Abstract

In recent years, photovoltaic (PV) panels have been increasingly adopted globally. However, their surfaces are susceptible to damage and soiling, which can negatively impact the output characteristics of PV modules. While intelligent cleaning robots are now widely used for maintaining PV panels, they still have limitations, such as low recognition accuracy and poor real-time performance in identifying dirt. To address these issues, this study proposes an algorithm based on an improved YOLOv9t model for detecting stains and damage on PV panels. The improvements include adding an All - in - One Dehazing Network (AOD - Net) to reduce the effects of overexposure and blurriness in captured images, replacing the original Conv with Spatial - Depth Conversion Convolution (SPD - Conv) to enhance accuracy and reduce computational complexity, and incorporating an Inverted Residual Mobile Block - Efficient Multi - Scale Attention (iRMB - EMA) mechanism to improve the algorithm's accuracy in complex backgrounds and during camera movements. Experimental results show that the improved YOLOv9t algorithm increases mAP by 5.83 % and reduces the weight file size by 18.21 % compared to other algorithms. This makes it a promising solution for PV panel maintenance and offers a new approach for large-scale photovoltaic power station automation, with the potential to significantly lower costs and enhance solar power generation efficiency.