For the future of energy production, large solar panels are essential without the huge carbon dioxide emissions. If left untreated, microfractures, hot spots and other defects can grow over time. This can lead to a reduction in output or even failure. Manivannan Svan’s solution to this problem revolves around computer vision and machine-learning to locate small defects on the surface, before reporting it automatically.
Sivan began by taking images of solar panels with visible cracks with an Arduino Portenta and Vision Shield, and then he created bounding boxes around each panel. He then trained a MobileNetV2 model using Edge Impulse’s new FOMO object detection algorithm. To improve the accuracy of the model, he added images from different angles and lighting conditions to the dataset to enhance its accuracy. This was to avoid mistaking white boundaries for cracks
The Edge Impulse Studio had the model tested and deployed to Sivan’s Portenta H7 board. It was able to detect cracks on a solar panel’s surface in around 80% of cases. Sivan may add more features to the model that use the onboard connectivity for faster response times with outside services. Read more about the project.
