AAPN-Tiny: A Compact Edge-Deployable Adaptive Attention Pyramid Architecture for Multi-Class Fault Diagnosis in Solar Photovoltaic Modules

Solar PV arrays are susceptible to various faults, such as hotspots, cracks, and Potential Induced Degradation, which can impair efficiency and longevity. Traditional fault detection methods are time-intensive and limited in accuracy, especially for large-scale installations. This paper proposes an Adaptive Attention Pyramid Network (AAPN) for accurate and efficient fault detection in PV modules. AAPN integrates depthwise separable convolutions, squeeze-and-excitation blocks, and adaptive attention mechanisms to achieve high accuracy in identifying fault types across different classification complexities. Extensive experimentation on a comprehensive dataset of infrared PV images, organized into 12 fault classes, demonstrated AAPNs high classification accuracy of up to 96% in binary and 92% in 12-class classification scenarios. The proposed model is tested using an infrared solar module dataset for 2-class, 8- class, 11-class, and 12-class fault categories. Its effectiveness is compared with 69 existing deep-learning models for various fault classes. An ablation study was conducted to evaluate the impact of different architectural components, such as depthwise separable convolutions and squeeze-and-excitation blocks, on the models performance, showing an optimal trade-off between accuracy and computational efficiency. The proposed architecture model is very lightweight, utilizing only 0.8 million parameters. Its effective balance between high accuracies and low parameter utilization makes it highly suitable for deployment on drone-based edge devices, facilitating on-site real-time PV fault monitoring, maintenance, and detection. Additionally, the model has been successfully implemented on the Google Coral Edge TPU, achieving 40.2 ms inference time per image, confirming its efficiency and suitability for real-time applications in resource-constrained environments.