Researchers have developed an open-access dataset consisting of 6,200 images to help solar panel cleaning robots navigate more accurately under various contamination conditions.
As you know, solar panels need regular cleaning to operate efficiently. Dust, sand, or water stains accumulating on the panels can hinder sunlight from reaching the cells, negatively impacting energy production. Especially in large solar power plants, autonomous robots are used to facilitate the cleaning process. For these robots to move reliably on the panels, they need to accurately perceive surface lines and structural boundaries. Researchers, led by the Universiti Teknologi Malaysia, have developed a new dataset consisting of 6,200 images to help solve this problem. The open-access dataset aims to improve the navigation performance of AI-powered cleaning robots under various contamination conditions.
AI's Navigation on Dirty Solar Panels Becomes Easier
Solar panel cleaning robots analyze images from their cameras to determine which direction to move on the panel. However, surface dirt and light reflections can make it difficult to detect lines on the panel. Therefore, a robot that performs well on a clean panel may not achieve the same performance on a heavily soiled panel.
While traditional solar panel image datasets typically focus on detecting damage and defects on panels, this study aims directly to improve robots' navigation capabilities. Researchers imaged the panel surface with an HF868-2 model camera placed on the front of a cleaning robot. The camera, positioned at approximately 15 centimeters in height, used a wide-angle lens with a 135-degree field of view. Images recorded at 60 frames per second and 320 × 240 pixels resolution were downscaled to 160 × 120 pixels after geometric calibration. This allowed the images to be analyzed with the limited processing power available on the robots.
To create the dataset, 800 clean panel images were first manually labeled. Three types of reference lines that robots could use for navigation were identified in these images: white gap lines between solar cells, and horizontal and vertical boundary lines where cells meet the panel frame. Subsequently, five different contamination levels were artificially generated for each clean image. Effects such as bird droppings, water stains, puddles, light reflections, and dust accumulated on frames were simulated. Depending on the contamination level, the goal was to obscure 1% to 50% of the reference lines. This method allows for the examination of only the effect of contamination on the image, without altering the panel's geometry. Thus, developers can more systematically evaluate under which conditions AI models struggle to recognize reference lines.
Dataset Also Tested Under Real Contamination Conditions
Of the total 6,200 images in the dataset, 4,800 were allocated for training and validation. Among these, 3,846 were prepared for training AI models, and 954 for evaluating their performance.
During the testing phase, in addition to 1,200 images covering six different clean and artificially contaminated conditions, 200 images obtained from two physically contaminated solar panels in a laboratory environment were also used. These panels were soiled with pigment drops, sand, wastewater, and clean water. The effect of reflections caused by intense light was also tested.
The label files provided with the images contain information on the positions, orientations, start and end points, and class data of the reference lines. This information allows developers to measure how accurately AI models detect these lines. Images obtained under real contamination conditions can also be used to evaluate how successful the models are not only in simulations but also in physical environments.
The researchers' next goal is to go beyond the dataset and enable solar panel cleaning robots to move autonomously more reliably. The team is developing AI models that can estimate the robot's orientation based on reference lines on the panel and operate with low processing power. They are also working on localization methods that combine images from cameras with data from inertial measurement units and wheel encoders. This approach aims to help robots determine their positions and directions of movement more accurately.
0 Comments: