基于深度学习的太阳能电池EL图像不良检测数据

本数据可以在串焊、层前、层后、终检环节通用，支持多种输入，能够处理不同规格和尺寸的光伏电池EL图像。其次是本技术的功能完整性。本技术支持17种电池片本身缺陷（断栅、黑边、黑斑、划伤、破片、单隐裂、死片、叉隐裂、虚焊、亮斑、污染、印记、脏污、炸点、条偏、黑心、氧环）的检测以及4种外部缺陷（拼接不良、明暗片、条码、间距）的检测，基本涵盖了所有可能出现的缺陷。本数据基于多工序、多片型、多片源的大量样本进行预训练，且构建了完整的负反馈体系，很大程度地提升了多场景下的算法检测精度，并搭配复杂且可配置的后处理规则，使用者可以根据不同车间的工艺要求进行个性化配置。本数据借鉴了人工智能领域的深度学习技术，对现有神经网络结构进行针对性的优化改进，并提出了适用太阳能面板生产过程中EL图像处理的特殊策略以及定制化方法，以解决上述背景技术中提出的问题，以下为算法逻辑：1) 输入待检图像及参数。2) 切除图像黑边。3) 切分电池片。4) 亮度及尺寸调整。5) 训练神经网络。6) 部署推理模型。7) 特征提取。8) 生成热力图。9) 计算偏置、计算中心点、计算目标长宽。10) 遍历不良信息。11) 规则过滤。12) 输出检出结果。将过滤后的有效缺陷信息进行组织和汇总。13) 拼接电池片图像，构造返回json串。对满足检出标准的缺陷坐标进行反推和拼接，计算得出其在原图中的具体位置，附加上算法得出的缺陷类型、置信度、单位电池片中的位置以及面积占比，构造出json信息串后，返回至前端软件。以上检测出不良后，后台单独统计各不良数量，以及单机产量，计算出此类不良比例，计算公式：不良比例=此类不良串数/总串数；例如：虚焊比例=虚焊串数/总串数；通过定时捞取此项数据，获取机台状态，针对性调整。

This dataset is universally applicable to series welding, pre-lamination, post-lamination, and final inspection stages, supports multiple input modes, and can handle photovoltaic (PV) cell Electroluminescence (EL) images of various specifications and sizes. Next, regarding the functional completeness of this technology: it supports detection of 17 types of inherent cell defects, including broken grid, black edge, black spot, scratch, broken wafer, single hidden crack, dead cell, cross hidden crack, poor soldering, bright spot, contamination, mark, dirt, burst point, strip misalignment, black core, and oxygen ring, as well as 4 types of external defects: poor string assembly, uneven cell brightness, barcode, and spacing, basically covering all possible defects. This dataset is pre-trained on a large-scale sample set from multiple production processes, diverse cell types, and various cell sources, and a complete negative feedback system has been established, which greatly improves the algorithm's detection accuracy across diverse scenarios. It is also equipped with complex and configurable post-processing rules, allowing users to perform personalized configurations based on the process requirements of different workshops. This technology leverages deep learning technologies from the field of artificial intelligence, makes targeted optimizations and improvements to existing neural network architectures, and proposes specialized strategies and customized methods applicable to EL image processing in solar panel production lines to address the issues raised in the aforementioned background art. The algorithm workflow is as follows:
1) Input the image to be inspected and corresponding parameters.
2) Remove the black borders from the image.
3) Segment individual PV cells from the image.
4) Adjust brightness and resize the image.
5) Train the neural network.
6) Deploy the inference model.
7) Extract image features.
8) Generate a heat map.
9) Calculate the offset, center point, and target length and width.
10) Traverse all detected defective information.
11) Filter results via predefined rules.
12) Output the detection results, and organize and summarize the filtered valid defect information.
13) Stitch the segmented cell images. Reverse-derive and stitch the coordinates of defects that meet the detection criteria to calculate their exact positions in the original full image, attach the defect type, confidence score, position within the individual cell, and area proportion derived by the algorithm, construct a JSON-formatted information string, and return it to the front-end software.
After detecting defective products, the background system separately counts the quantity of each type of defect and the single-machine output volume, and calculates the proportion of such defects. The calculation formula is: Defect proportion = Number of defective strings of this type / Total number of strings. For example: Poor soldering proportion = Number of poor soldering strings / Total number of strings. By periodically extracting this data, the operating status of the production machine can be obtained for targeted adjustments.

搜集汇总

数据集介绍

特点

该数据集是一个基于深度学习的太阳能电池EL图像不良检测数据集，包含大量样本和多种缺陷类型，适用于多工序和多场景的检测需求。