单晶硅的广泛应用推动了石英坩埚市场的飞速发展，石英坩埚透明层直接与熔融物质接触，其厚度不均匀以及因侵蚀导致原有厚度变薄都会造成坩埚破裂、漏硅。目前石英坩埚透明层的厚度测量工作只能依赖人工完成，效率低下、精度不足且仅能测量坩埚口沿位置，无法确保坩埚整体质量合格。本文提出了一种基于层界边缘视觉特征分析的石英坩埚透明层层厚测量方法，设计研发了一台机械臂加持电子显微镜的自动拍摄装置，通过识别不同深度的透明层特征、确定层界关键帧并建立数学模型最终对厚度进行测量。本文主要研究内容包括：

（1）基于石英坩埚超透层不同深度的气泡特征及分布规律，提出了一种改进的霍夫圆检测算法，结合Canny算子与Prewitt算子对透明层气泡图像进行边缘检测，利用粒子群优化算法对所得边缘进行搜索得到最优圆心和半径组合，最后利用最小二乘法拟合圆检测结果，检测结果显示改进后算法的气泡识别率平均达到了93.4%。

（2）针对“海绵状”特征的判断问题，建立了基于ViT的CBAM增强模型，通过将CBAM注意力模块引入ViT模型增强了模型对气泡与“海绵状”结构等不同图像特征的关注；将ViT模型的分块策略修改为重叠移动滑窗，避免由于图像块边缘特征不足导致的信息损失问题，有助于捕获细节和边缘特征；针对帧图像特征相似的特点，在交叉熵损失函数的基础上添加相似损失函数作为综合损失，提升模型对细微差异的感知能力，使模型可以更好的区分透明层与非透明层层界区域的相似特征，结果显示改进后的ViT模型相较于原始版本在准确率上提升了5.63%，能够准确识别两层之间的交界区域。

（3）基于得到的关键帧信息建立透明层层厚测量模型，并设计搭建了自动化检测装置。考虑到不同坩埚的大小尺寸特点，选择合适组件进行组装，设计了可以旋转、加紧的转台与机械臂搭配协作，并根据拍摄需求规划了合理路径。实验最后通过设计机械臂与转台之间的信号交互，实现了短时间内全自动的坩埚检测工作。

The widespread application of monocrystalline silicon has promoted the rapid development of Quartz crucible market. The transparent layer of Quartz crucible is in direct contact with the molten material, and its uneven thickness and the thinning of the original thickness due to the erosion of solution will destroy the structure of the crucible, resulting in the rupture of the crucible and the leakage of silicon. At present, the thickness measurement of the transparent layer of Quartz crucible can only be done manually, which is inefficient, lack of precision and can only measure the position of the crucible mouth edge, and cannot ensure the overall quality of the crucible is qualified. In this thesis, a method of Quartz crucible transparent layer thickness measurement based on the analysis of visual characteristics of layer boundary edge is proposed, and a robotic arm with an electron microscope is designed and investigated to measure the thickness by identifying the characteristics of the transparent layer at different depths, determining the key frames of the layer boundary, and establishing a mathematical model. The major research content of this thesis includes:

(1) An improved Hough circle detection algorithm is proposed based on the characteristics and distribution laws of bubbles at different depths in the ultra-transparent layer of Quartz crucible, combining Canny operator and Prewitt operator to detect the edges of the bubble images in the transparent layer, and using the particle swarm optimization algorithm to search the resulting edges to get the optimal combination of the circle center and the radius, and finally, using the least-squares fitting of the circle detection results. The improved bubble recognition rate reaches 93.5% on average.

(2) For the judgement of "spongy" features, a CBAM enhancement model based on ViT model is established, which enhances the model's attention to the different features such as bubbles and "spongy" structures by introducing the CBAM attention module into the ViT model. The chunking strategy of the ViT model is modified to an Overlapping sliding window, which avoids the problem of information loss due to the lack of edge features in the image blocks, and ensures that each image block contains enough contextual information, which helps to better capture the details and edge features; For the similarity of the image features in the transition layer, we add a similarity loss function on top of the cross-entropy loss function as a comprehensive loss of the model, which improves the model's ability to perceive the subtle differences, so that the model can better distinguish the similar features in the boundary region of transparent and non-transparent layers, and the results indicate that the improved ViT model improves the accuracy by 5.63% compared to the original version, able to accurately identify the junction region between the two layers.

(3) The transparent Layer thickness measurement model is established based on the obtained key frame information, and the automatic inspection device is designed and built. In consideration of the size and dimension characteristics of different crucibles, the appropriate components are selected for assembly, the rotary table that can be rotated and tightened is designed to be paired with the robotic arm, and a sensible path is planned according to the filming requirements. Eventually, through designing the signal interaction between the robotic arm and the rotary table, the fully automatic crucible inspection work is realized in a short time.

[1]梁启超, 乔芬, 杨健, 等. 太阳能电池的研究现状与进展[J]. 中国材料进展, 2019, 38(5): 505-511.

[2]张良. 欧晶科技: 双向驱动勇做细分产业龙头[J]. 董事会, 2023(08): 38-39.

[3]Li J, Li Z, Liu L, et al. Effects of melt depth on oxygen transport in silicon crystal growth by continuous-feeding Czochralski method[J]. Journal of Crystal Growth, 2023, 610: 127180.

[4]刘宇浩. 碳化硅材料性能优异,应用迎来发展良机[J]. 中国集成电路, 2023, 32(05): 16-21.

[5]陈远新, 马英英, 陈芸清, 等. 半导体级多晶硅的低成本清洗方法[J]. 中国集成电路, 2023, 32(05): 80-85.

[6]Ogbonda C. Crystal growth method of gallium arsenide using czochralski method[J]. Asian Journal of Multidimensional Research(AJMR), 2020, 9(3): 8-14.

[7]罗晓斌, 张波, 周燕青, 等. 功率和埚转对直拉单晶硅漏硅事故率的影响及分析[J]. 山西化工, 2020, 40(03): 55-57.

[8]史珺, 宗卫峰. 铸锭炉热场对坩埚可靠性影响的机理研究——硅液对于石英坩埚的侵蚀及其与温度、坩埚材质和热场的关系分析[J]. 太阳能学报, 2014, 35(03): 502-507.

[9]JC/T 1048-2018. 单晶硅生长用石英坩埚[S]. 北京: 中国建材工业出版社, 2019.

[10]张鹏远, 万烨, 刘见华, 等. 合成石英砂工艺概况及几种石英砂的检测数据对比分析[J]. 中国非金属矿工业导刊, 2024(02): 55-59.

[11]Hansen, Sebastian Holand. Investigation of bubble distribution and evolution in solar cell quartz crucibles[D]. Trondheim: Norwegian University of Science and Technology, 2017.

[12]Ove Paulsen et al. Bubble distribution in fused quartz crucibles studied by micro X-Ray computational tomography. Comparing 2D and 3D analysis[J]. Journal of Crystal Growth, 2019, 520: 96-104.

[13]Hirsch A, Trempa M, Kupka I, et al. Investigation of gas bubble growth in fused silica crucibles for silicon Czochralski crystal growth[J]. Journal of Crystal Growth, 2020, 533: 125470.

[14]张大虎, 田青越, 汪灵, 罗大伟, 潘俊良. 单晶硅生长用石英坩埚的组成与结构特征[J]. 矿物岩石, 2016, 36(1): 12-16.

[15]Zhao Q, Li R R, Qian Q U. Research on statistic detection method of micro bubbles in transparent layer of quartz crucible based on image processing[J]. Journal of Crystal Growth, 2021, 556: 125966.

[16]赵谦, 史凌云. 融合细胞自动机的坩埚单层气泡计数方法研究[J]. 计算机仿真, 2020, 37 (05): 429- 433.

[17]赵谦, 李蓉蓉, 钱渠, 马莉, 尹怡晨. 基于ARM的坩埚气泡动态智能检测系统[J]. 仪表技术与传感器, 2021(04): 97-101.

[18]石星宇. 连续拉晶下石英坩埚内气泡变化规律及晶棒的性能研究[D]. 银川: 宁夏大学, 2022.

[19]黄维. 坩埚透明层多层气泡计量方法研究[D]. 西安: 西安科技大学, 2021.

[20]赵谦, 郑超, 尹怡晨, 马文越. 基于视频图像的坩埚气泡三维测量方法[J]. 国外电子测量技术, 2022, 41(10): 178-184.

[21]Zhao Q, Zheng C, Ma W. An Improved Crucible Spatial Bubble Detection Based on YOLOv5 Fusion Target Tracking[J]. Sensors, 2022, 22(17): 6356.

[22]刘志辉, 罗玉峰, 饶森林, 等. 坩埚厚度对多晶硅定向凝固的影响[J]. 热加工工艺, 2017, 46(11): 74-77.

[23]韩揽月, 袁平. γ射线反散射石英坩埚厚度仪样机的研制[J]. 核电子学与探测技术, 2013, 33(06): 705-707.

[24]T/CEMIA 004-2018. 光伏单晶硅生长用石英坩埚[S]. 北京: 中国标准出版社, 2018.

[25]祝立君, 于宏宇, 王建龙, 周鼎军, 左立超. 一种自动测量坩埚透明层厚度的装置及方法[P]. 内蒙古： CN108592844A, 2018-09-28.

[26]刘福鸿, 陈宝昌. 一种石英坩埚壁厚快速检测装置[P]. 江苏省： CN113390318A, 2021-09-14.

[27]焦瑾瑾. 基于Roberts边缘检测的数字媒体激光重现方法[J]. 激光杂志, 2022, 43(05): 197-200.

[28]刘源, 夏春蕾. 一种基于Sobel算子的带钢表面缺陷图像边缘检测算法[J]. 电子测量技术, 2021, 44(03): 138-143.

[29]Saeed B, Hossein B. Edge detection on noisy images using Prewitt operator and fractional order differentiation[J]. Multimedia Tools and Applications, 2022, 81(7): 9759-9770.

[30]Wang Q, Chen W, Peng H. An intelligent image detection method using improved canny edge detection operator[J]. International Journal of Information Technology and Management, 2022, 21(4): 369-381.

[31]张萍萍, 李童, 李茹, 等. 一种改进的Sobel图像边缘检测算法及其实现[J]. 电视技术, 2022, 46(05): 42-45.

[32]王晶晶, 杨俊杰. 改进Canny算法在绝缘子串红外图像边缘检测中的应用[J]. 电瓷避雷器, 2022(03): 179-185.

[33]杨勇军, 肖志文, 常明, 等. 基于Prewitt算子和Radon变换的雷达高速微弱目标检测[J]. 电光与控制, 2024, 31(06): 101-104+111.

[34]崔玉, 俞建定, 邹洪彬, 等. 改进的Canny实时边缘检测算法在FPGA上的实现[J]. 无线通信技术, 2021, 30(04): 27-31+36.

[35]赵明庆, 常月军. 基于神经网络的船舶红外图像边缘检测方法[J]. 舰船科学技术, 2023, 45(17): 154-157.

[36]Liu C, Sharan L, Adelson E H. Exploring features in a Bayesian framework for material recognition[C]// The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010. IEEE, 2010.

[37]Yang M, Thung G. Classification of trash for recyclability status[J]. CS229 project report, 2016, 2016(1): 3.

[38]L. K S, S. H, Kanagaraj N, et al. AlexNet approach for early stage Alzheimer’s disease detection from MRI brain images[J]. Materials Today: Proceedings, 2022, 51(P1): 58-65.

[39]杨文艺. 基于深度学习的岩石薄片图像智能识别与分类[D]. 焦作: 河南理工大学, 2021.

[40]于晓, 庄光耀. 基于轻量化VGG16和BCBAM的电力设备故障红外图像诊断识别[J]. 河南科技学院学报(自然科学版), 2023, 51(06): 60-69.

[41]He K, Zhang X, Ren S, et al. Deep residual learning for image recognition [C]// Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 770-778.

[42]叶冲, 杨晶东. 基于CBAM-EfficientNet的垃圾图像分类算法研究[J]. 智能计算机与应用, 2021, 11(05): 218-222.

[43]Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. Advances in neural information processing systems, 2017, 30.

[44]Thayer F, Alex L, Marco C, et al. Marginal Bayesian Posterior Inference using Recurrent Neural Networks with Application to Sequential Models[J]. Statistica Sinica, 2023, 33(SI): 1507-1532.

[45]Zhiheng Z, Xiaomei H, Naixue X, et al. A self-attention model with contrastive learning for online group recommendation in event-based social networks[J]. The Journal of Supercomputing, 2023, 80(7): 9713-9741.

[46]Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arXiv preprint arXiv: 2010.11929, 2020.

[47]Aloraini M. An effective human monkeypox classification using vision transformer[J]. International Journal of Imaging Systems and Technology, 2024, 34(1): e22944.

[48]杜显君. 基于ViT的高速公路车辆细分类研究[J]. 现代计算机, 2022, 28(12): 51-55.

[49]Chen C F R, Fan Q, Panda R. Crossvit: Cross-attention multi-scale vision transformer for image classification[C]// Proceedings of the IEEE/CVF international conference on computer vision. 2021: 357-366.

[50]Han K, Xiao A, Wu E, et al. Transformer in transformer[J]. Advances in Neural Information Processing Systems, 2021, 34: 15908-15919.

[51]孙丹, 朱立军. 基于BP神经网络和霍夫变换的非理想虹膜定位[J]. 沈阳化工大学学报, 2023, 37(01): 60-67.

[52]陈江. 基于改进支持向量机的纸张缺陷图像分类识别方法[J]. 造纸科学与技术, 2023, 42(05): 39-43+78.

[53]王丽霞, 史正涛, 喜文飞等. 基于灰度共生矩阵的基岩滑坡纹理特征提取分析研究[J]. 城市勘测, 2023, (02): 187-192.

[54]潘瑞, 曾良才, 伍世虔等. 低光照图像特征检测方法[J]. 传感器与微系统, 2021, 40(10): 110-113+117.

[55]张永强, 宋美霖, 刘天虎, 等. 混合骨骼特征的三帧间差分手势识别方法研究[J]. 智能科学与技术学报, 2022, 4(04): 592-599.

[56]张新雨, 李思雨, 李婧华, 等. 基于改进智能霍夫变换的红热圆形工件直径检测[J]. 西安理工大学学报, 2022, 38(01): 41-47.

[57]冉宇辰, 黄民, 王文胜等. 一种改进霍夫梯度变换的纱笼坐标检测方法研究[J]. 机械设计与制造, 2022, (10): 94-99.

[58]Jiawei S, Bobo W, Tong Z, et al. Classification for thyroid nodule using ViT with contrastive learning in ultrasound images[J]. Computers in Biology and Medicine, 2023, 152106444-106444.

[59]张楠楠, 张晓, 白铁成, 等. 基于CBAM-YOLO v7的自然环境下棉叶病虫害识别方法[J]. 农业机械学报, 2023, 54(S1): 239-244.

[60]王大浩, 平雪良. 基于规定性能的六自由度机械臂视觉伺服控制[J]. 传感器与微系统, 2022, 41(03): 104-108.

[61]朗杰智能. BRTIRUS0707A六轴机器人[EB/OL]. http://www.nblangjie.com/ goodsread_ 43 .html, 2024-05-29.

[62]黄梓效, 易增强, 谢清俊, 等. 激光玻璃测厚仪校准装置及校准方法的研究[J]. 中国仪器仪表, 2023(05): 63-67.

[63]罗文, 张伟, 赫丽华, 等. 碳纤维复合材料表面雷达吸波涂层超声测厚方法研究[J]. 失效分析与预防, 2021, 16(04): 233-237.