TinyTripleNet: A lightweight architecture for solar photo voltaic fault detection optimized for edge tensor processing unit hardware deployment
Kolisetty Himavarshini, Ranjith Raja B, Yanamala Rama Muni Reddy, Archana Pallakonda, Rayappa David Amar Raj, Edris Pouresmaeil, Jamshid Aghaei
Abstract
• TinyTripleNet is a lightweight neural network developed for detecting faults in solar photovoltaic panels on edge devices. • It accurately classifies various solar panel faults using thermographic images. • The model is optimised for Coral Edge Tensor Processing Unit, enabling fast and memory-efficient deployment. • TinyTripleNet outperforms well-known models like VGG16 and ResNet50 in both speed and resource usage. The growing adoption of solar energy underscores the need for efficient and accurate fault detection in photovoltaic (PV) systems to minimize energy loss and maintenance overhead. This study introduces TinyTripleNet, a lightweight convolutional neural network optimized for real-time solar panel fault classification on edge devices. The model integrates three convolutional layers with Residual and Squeeze-and-Excitation (SE) blocks to enhance feature extraction while maintaining a compact size of only 0.9 million parameters. TinyTripleNet is trained using binary cross-entropy loss and the Adam optimizer on a thermographic PV dataset, achieving class-wise accuracies of 96.71 % (2-class), 92.16 % (8-class), 90.25 % (11-class), and 93.07 % (12-class). It also achieved a mean Average Precision (mAP) of 92.4 % on the 12-class dataset. Compared to 29 baseline models, including VGG16, ResNet50, and MobileNet, TinyTripleNet demonstrated a 75 % reduction in inference time and over 80 % savings in memory usage, with an inference latency of 4.8 ms/image on the Coral Edge TPU and a runtime memory footprint of 1.7 MB. These results position TinyTripleNet as a robust solution for drone-based PV fault detection, offering a practical trade-off between accuracy, model size, and real-time deployment efficiency.