Two-Stage Lightweight CNN with Weakly Supervised Defect Localization for Multimodal Photovoltaic Fault Detection

This repository contains three distinct photovoltaic (PV) datasets representing the full spectrum of solar fault detection: internal cell degradation (EL), thermal anomalies (IR), and surface-level obstructions (RGB). Each dataset includes both the Original (raw) images and the Refined (auto-annotated) versions generated by our Two-Stage Framework. 1. Electroluminescence (EL) PV Dataset Modality: 8-bit Grayscale (Internal Inspection). Purpose: Used to detect subsurface defects invisible to the naked eye, such as microcracks and electrical disconnects. Content: The dataset comprises 2,624 images extracted from 44 distinct solar modules. It includes both monocrystalline (uniform background) and polycrystalline (complex grain boundary) cells. Classes: Binary classification (Functional vs. Defective). Challenge: The primary challenge is distinguishing actual microcracks from the jagged grain boundaries of polycrystalline cells, which often mimic defect patterns. 2. Thermal Infrared (IR) PV Dataset Modality: Thermal Infrared (Heat Anomaly Inspection). Purpose: Used to identify heat-induced faults, such as hotspots and diode failures, captured via UAVs. Content: A large-scale dataset of 20,000 thermal images. Resolution: Extremely low native resolution ($40 \times 24$ pixels), presenting a significant challenge for feature extraction. Classes: 12 Categories, including No-Anomaly, Hotspot, Hotspot-Multi, Diode, Diode-Multi, Shadowing, Cell, Cell-Multi, Cracking, Soiling, Vegetation, and Offline-Module. Challenge: Extreme class imbalance (50% of data is "No-Anomaly") and spatial sparsity. 3. Aerial RGB PV Dataset Modality: Optical RGB (Surface Inspection). Purpose: Used to classify surface-level obstructions and physical degradation visible from low-altitude drone flights. Content: 875 high-resolution images acquired under varying outdoor lighting conditions. Classes: 6 Categories: Clean, Dusty, Bird Drop, Electrical Damage, Physical Damage, and Snow-Covered. Challenge: Distinguishing between transient environmental obstructions (e.g., bird drops, snow) and permanent physical damage (e.g., shattered glass) amidst complex ground clutter. Dataset Structure in Zip Files Original Datasets: Contain the raw images exactly as collected/standardized. Refined Datasets: Contain the output of Phase 1 (Auto-Annotation). These images have been processed by the WSOL module to generate cropped or spatially focused regions of interest (ROI), removing background noise (e.g., roof tiles in RGB or grain boundaries in EL) to facilitate high-precision fine-tuning in Phase 2.本仓库包含三个不同的光伏（photovoltaic, PV）数据集，覆盖了太阳能故障检测的全场景：电池片内部退化（电致发光，Electroluminescence, EL）、热异常（红外，Infrared, IR）以及表面遮挡（RGB）。每个数据集均包含原始（raw）图像与经我们的两阶段框架生成的精炼（auto-annotated）版本。 1. 电致发光（EL）光伏数据集 模态：8位灰度图像（内部检测）。 用途：用于检测肉眼不可见的亚表面缺陷，例如微裂纹与电气断路。 内容：该数据集包含从44块不同太阳能组件中提取的2624张图像，涵盖单晶硅（背景均匀）与多晶硅（存在复杂晶界）两种电池片。 类别：二分类任务（正常 vs 缺陷）。 挑战：主要难点在于区分真实微裂纹与多晶硅电池片的锯齿状晶界——后者常与缺陷模式高度相似。 2. 热红外（IR）光伏数据集 模态：热红外成像（热异常检测）。 用途：用于识别经无人机（UAV）捕获的热诱导故障，例如热斑与二极管失效。 内容：包含20000张热图像的大规模数据集。 分辨率：原生分辨率极低（$40 imes24$像素），为特征提取带来显著挑战。 类别：共12个类别，包括无异常、热斑、多区域热斑、二极管故障、多区域二极管故障、阴影遮挡、单电池片异常、多电池片异常、裂纹、组件积灰、植被遮挡与组件离线。 挑战：存在严重的类别不平衡（50%的数据为“无异常”类别）与空间稀疏性问题。 3. 航拍RGB光伏数据集 模态：可见光RGB成像（表面检测）。 用途：用于分类低空无人机飞行中可观测到的表面遮挡与物理退化。 内容：包含875张在不同户外光照条件下采集的高分辨率图像。 类别：共6个类别：清洁、积尘、鸟粪、电气损伤、物理损伤与积雪覆盖。 挑战：在复杂地面杂波背景中，需区分瞬时环境遮挡（例如鸟粪、积雪）与永久性物理损伤（例如玻璃碎裂）。 Zip文件内的数据集结构 原始数据集：包含采集/标准化后的原始图像，未做额外处理。 精炼数据集：包含第一阶段（自动标注）的输出结果。这些图像经弱监督目标定位（Weakly Supervised Object Localization, WSOL）模块处理，生成裁剪或空间聚焦的感兴趣区域（Region of Interest, ROI），并移除背景噪声（例如RGB数据中的屋顶瓦片、EL数据中的晶界），以助力第二阶段的高精度微调。