Pruned and Low-Rank Optimized Tiny Residual Architecture for Solar Photovoltaic Fault Classification on Edge TPU

Abstract Photovoltaic (PV) power systems are essential for attaining sustainable and low-carbon energy production, providing substantial opportunities for diminishing greenhouse gas emissions and improving grid resilience. Operational inefficiencies resulting from defects in photovoltaic modules—frequently occurring during production or maintenance—can diminish total energy output and system dependability. This study introduces a small, high-performance fault classification system designed to facilitate the transition to intelligent, self-diagnostic renewable energy infrastructure, suited for energy-efficient embedded deployment. Utilizing infrared thermography and an innovative residual deep learning architecture, the model employs singular value decomposition (SVD)-based low-rank approximation and pruning methods to enhance parameter efficiency, resulting in a lightweight configuration of merely 0.65 million parameters. Benchmark assessments in 2-class, 8-class, 11-class, and 12-class configurations demonstrate enhanced performance compared to 59 current convolutional neural network (CNN) models, with an accuracy of 94.73% for anomaly detection and up to 88.44% for multiclass fault diagnosis. The model demonstrates an average inference time of 4.3 ms per image on Edge TPU, validating its application for real-time drone-assisted photovoltaic inspection and facilitating intelligent diagnostics in renewable energy conversion systems. This study advances the optimization and digitization of thermodynamically efficient energy systems.