Deep learning algorithm and IoT based system for Photovoltaic (PV) defect images classification

Renewable energy had become a critical alternative to traditional power sources due to the depletion of non-renewable resources in recent years. Solar energy offered significant advantages, including sustainability, environmental friendliness, and costeffectiveness. Photovoltaic (PV) systems, which converted sunlight into electrical energy through PV modules, play a crucial role in solar power generation. However, maintaining the efficiency of PV modules was essential for long-term energy sustainability, as defects can significantly impact performance. Traditional inspection methods required extensive on-site assessments, making the process labour-intensive and inefficient. This study proposed an automated PV defect classification system using Convolutional Neural Network (CNN) with transfer learning, supported by an IoTbased monitoring platform. The total of 1,394 RGB type images were used in this research, undergoing image pre-processing that included data augmentation techniques such as resizing, rotation, and X-Y translation to enhance the dataset and improved model generalization during training and evaluation. The system employed CNN with two distinct transfer learning architectures, Series Networks which were AlexNet, VGG-16, and VGG-19 and Directed Acyclic Graph (DAG) Networks which were ResNet-18, ResNet-50, Inception-V3, and GoogLeNet. Series Networks, known for their simpler, layer-wise feature extraction, perform well in certain classification tasks, such as specificity and precision. Meanwhile, DAG Networks leveraged deeper architectures and residual connections, enhancing feature learning efficiency and classification accuracy. Experimental results revealed that ResNet-50, a DAG Network, outperforms other architectures by achieving the highest classification accuracy (98.96%) and F1-score (97.89%), largely due to its ability to retain critical features across multiple layers. This model particularly excelled in classifying the Cracks defect with perfect accuracy. In contrast, AlexNet, a Series Network, achieved the highest specificity (99.53%), also performing best in identifying Cracks. VGG-16 demonstrated superior precision (98.65%), with outstanding classification performance across Cracks, Delamination, and Encapsulation Discolouring classes. Meanwhile, VGG-19 excelled in sensitivity (98.10%), effectively identifying all true cases of Cracks and Delamination. These results highlighted the strength of Series Networks in handling structured, less complex classification tasks, while DAG architectures like ResNet-50 show superior performance in capturing deeper feature representations for high overall accuracy. The IoT integration, implemented via the ThingSpeak platform, enable remote visualization of classification performance metrics transmitted from MATLAB, with the system's effectiveness evaluated based on data transmission accuracy and the reliable display of results on the platform. This research demonstrated that while Series Networks offered efficiency in structured classification, DAG Networks provided superior deep feature extraction, making them more suitable for complex PV defect detection. The integration of deep learning with IoT enhanced result accessibility and supports more informed decision-making, potentially reducing manual inspection efforts in future implementations.