Results of the AI

We used supervised learning, which is where a model learns to make predictions based on labeled training data. In the context of solar panel classification, this means that a dataset of solar panel images with labeled categories (such as broken, dirty, microcracks, hot spots, and good) are collected from various sources. This dataset is then used to train the convolutional neural network (CNN), which is a type of deep learning algorithm commonly used for image classification tasks.

In this particular case, the images in the dataset were 512x512 pixels in size, and the data was split into two, 80% of which was used for training and 20% for validation sets. The training images were augmented, which means that they were modified in various ways to create more variations of the same image, such as by rotating, flipping, stretching, or skewing the image. This is a common technique used in deep learning to help the model better generalize novel, unseen images.

The model was able to accurately detect broken panels, hot spots, and microcracks, which are common types of solar panel defects. However, it struggled to accurately detect panels covered in dust, pollen, and other debris. This is likely due to the limited training data available for this category, which resulted in the model being overfitted to the other categories. Overfitting occurs when a model is trained on a limited dataset and starts to memorize the data rather than learning to generalize to new data. This can result in the model performing well on the training data but poorly on new, unseen data. In this case, the model's high accuracy for broken panel detection, hot spot, and microcrack detection may be a result of overfitting due to the lack of available images to train these categories.

To improve the model's accuracy, future work will involve implementing instance segmentation and finding ways to collect more data on solar panel defects. Instance segmentation is a technique used to identify specific objects within an image, using OpenCV and edge detection, we isolate a single panel and give that tagged image to the AI for training and validation. Then, using the same edge detection, we isolate a panel from a live image and run it through the AI for classification, this will help the model better detect and classify different types of debris on solar panels. Additionally, collecting more data on less common categories like hot spots and microcracks could help the model better generalize to new, unseen images.

The use of a supervised learning approach and a CNN to classify solar panels into different categories is a promising application of machine learning. However, the accuracy of the model is limited by the availability of training data, and future work is needed to improve its performance.

UTC IEEE Robotics Club - 2023