基於深度學習之太陽能板檢測研究

Abstract

隨著全球再生能源需求快速增長，太陽能發電已成為能源轉型之核心。然而，大規模光電場域的維運面臨人工巡檢效率受限與判讀標準不一之挑戰。為提升檢測效率與判讀一致性，本研究提出一套基於深度學習之自動化太陽能板瑕疵檢測架構，整合實例分割、幾何校正、瑕疵分類與品質篩選四大核心模組，旨在實現高信賴度之無人機巡檢應用。 本研究採用 YOLO 實例分割技術，從複雜背景中精確提取太陽能板區域（ROI）；實驗數據顯示，該模型於物件定位與輪廓分割之平均準確率 （mAP@0.50）分別達到 98.9% 與 98.4%，確保前端輸入之品質。系統接續整合透視轉換策略，修正拍攝視角造成之幾何變形以標準化輸入特徵。針對瑕疵樣本稀缺與類別不平衡特性，本研究設計優化之卷積神經網路，整合全域平均池化（GAP）、標籤平滑（Label Smoothing）與 Dropout 等正規化策略，強化模型對物理損壞與環境髒汙之辨識能力。此外，為解決開放場域非預期樣本干擾，引入基於資訊熵（Entropy）之品質篩選機制，賦予系統風險控制之能力。 實驗結果顯示，所提出之分類模型於測試集達98.1% 之準確率。品質篩選機制證實能以約 3.8% 之人工複檢率，主動篩選 84.6% 之無效影像，並將自動化分類準確率提升至 98.9%。本研究提出之系統架構在維持高準確度與覆蓋率之前提下，有效降低人工維運成本，為太陽能光電場智慧化巡檢提供具實務價值之技術解決方案。

Solar photovoltaic (PV) power is pivotal to the global energy transition, yet manual inspection of large-scale plants remains inefficient and inconsistent. To address these challenges, this study proposes an automated defect detection framework based on deep learning. It integrates instance segmentation, geometric correction, classification, and quality screening for reliable UAV inspections. Methodologically, YOLO instance segmentation is employed to extract panel regions, achieving a mean Average Precision (mAP@0.50) of 98.9% for localization and 98.4% for segmentation. Perspective transformation is then applied to standardize geometric features. To overcome data scarcity and class imbalance, an optimized CNN incorporating Global Average Pooling, Label Smoothing, and Dropout is designed for robust defect identification. Furthermore, an entropy-based screening mechanism is introduced to mitigate risks from out-of-distribution samples. Experimental results demonstrate a classification accuracy of 98.1% on independent tests. The screening mechanism intercepts 84.6% of invalid images with a mere 3.8% manual review rate, elevating accuracy to 98.9%. Effectively balancing high accuracy with reduced maintenance costs, this framework presents a practical solution for intelligent PV plant inspection.