基于孪生网络的自监督太阳能电池板裂纹检测方法

太阳能电池板的裂纹缺陷检测能够避免电能转换效率低,以及短路造成起火的损失。针对现存对比学习方法中存在 细微裂纹漏检导致检测精度低,并且严重依赖构建负样本等问题,提出了一种基于孪生网络的两阶段自监督裂纹检测方法。 第1阶段提出了一种基于卷积神经网络 (convolutional neural network,CNN)和 Transformer 的预训练编码器模型,通过孪生 网络架构学习样本的精细特征表示,提高对电池板细微裂纹的特征表示能力;第2阶段基于预训练模型在少量标注样本下学 习分类器以区分缺陷样本。为进一步区分不影响电池板功能的纵向裂纹,另增加了一个分类头进行判别。在 ELPV 数据集 上的实验结果表明,方法在测试准确度方面优于其他相关检测方法,在只对数据进行少量标注的情况下准确度达到83.26%, 单张检测时间为6.1 ms, 同时在裂纹图像中检出纵向裂纹的召回率也有76.7%。     

面向GPU并行编程的线程同步综述

并行计算已成为主流趋势. 在并行计算系统中, 同步是关键设计之一, 对硬件性能的充分利用至关重要. 近年来, GPU (graphic processing unit, 图形处理器)作为应用最为广加速器得到了快速发展, 众多应用也对GPU线程同步提出更高要求. 然而, 现有GPU系统却难以高效地支持真实应用中复杂的线程同步. 研究者虽然提出了很多支持GPU线程同步的方法并取得了较大进展, 但GPU独特的体系结构及并行模式导致GPU线程同步的研究仍然面临很多挑战. 根据不同的线程同步目的和粒度对GPU并行编程中的线程同步进行分类. 在此基础上, 围绕GPU线程同步的表达和执行, 首先分析总结GPU线程同步存在的难以高效表达、错误频发、执行效率低的关键问题及挑战; 而后依据不同的GPU线程同步粒度, 从线程同步表达方法和性能优化方法两个方面入手, 介绍近年来学术界和产业界对GPU线程竞争同步及合作同步的研究, 对现有研究方法进行分析与总结. 最后, 指出GPU线程同步未来的研究趋势和发展前景, 并给出可能的研究思路, 从而为该领域的研究人员提供参考.     

40Gb/s 0.18μm CMOS甚短距离并行光接收前端放大器

研制成功一种应用于甚短距离(VSR)光传输系统的40Gb/s 并行光接收前端放大器芯片.该电路采用12路并行信道结构和0.18μm CMOS工艺,单信道传输速率达到了3.318Gb/s.电路设计采用了RGC结构和噪声优化技术,克服了CMOS光检测器大寄生电容造成的带宽不够的问题.提出了一种同时采用P 保护环(PGR)、N 保护环(NGR)和深N阱(DNW)的并行放大器隔离结构,有效地抑制了并行放大器之间的串扰,减小了放大器之间的衬底耦合噪声.测试表明,所有信道在3.318Gb/s数据速率、2mVpp输入和2pF的寄生电容下均得到了清晰的眼图.芯片采用1.8V电源供电,单路前端放大器的功耗为85mW,12路总功耗约为1W.     

Parallel‐series connected rectenna array using frequency selective surface (FSS) for power harvesting applications at 5.8 GHz

This article presents a frequency selective surface (FSS) and rectenna array in parallel‐series connection to harvest power of wireless power systems. To improve the basic antenna parameters, a frequency selective surface was mounted on the reverse side of the substrate. According to the results, the proposed method showed significant improvement in comparison to other methods in both single and rectenna array. Moreover, the conversion efficiency of the presented technique was increased to 76%. The output voltage of 4.5 and 9 V and the current of 60 and 120 mA were resulted for 5 and 10 rectenna units, respectively. The proposed arrays can be expanded to large scale integrated array without any significant degradation in conversion efficiency.     

Ensemble mating selection in evolutionary many-objective search

Traditional multi-objective evolutionary algorithms have encountered difficulties when handling many-objective problems. This is due to the loss of selection pressure incurred by the growing size of objective space. A variety of environmental selection operators have been proposed to address the issue, each has its distinct benefits and drawbacks. We develop a novel ensemble framework to enhance the effectiveness and robustness of many-objective optimization. The framework incorporates multiple environmental selection operators to guide the search, which are then viewed as voters to construct a mating pool. We design an ensemble mating selection strategy that makes decisions based on the preference information provided by the voters: individuals elected by more voters will be assigned larger possibilities to enter the mating pool. By doing so, high quality offspring can be reproduced from the elected promising candidates. To accommodate the multiple selection operators for voting, the framework maintains multiple parallel populations, where each population is updated by one of the selection operators. An instantiation of the framework with three popular operators is presented as a prime example. Extensive experiments have been conducted on a number of many-objective problems to examine the effectiveness of the proposed approach. Experimental results show that the mating selection strategy is capable of improving the quality of the obtained solution set.     

Quantitative evaluation of reheat cracking susceptibility by in situ observation and measurement using laser confocal microscope*

In the post-weld heat treatment process, the reheat cracking which might occur in the weldments of low-alloy steels has been a serious problem. So, it is considered to be important to predict the possibility of occurrence of reheat cracking in these steels. It is however recognized as a time-consuming procedure to evaluate quantitatively the susceptibility to this type of cracking. In the present study, a new quantitative evaluation method of reheat cracking susceptibility by in situ observation and measurement using a laser confocal microscope has been proposed. Through this new method, the reheat cracking susceptibility of any kind of steels can be evaluated with the same standard. Moreover, because the position of the initial crack can be focused and the critical ductility to initiate the crack is measured by in situ observation, the reheat cracking susceptibility can be evaluated using only one specimen. So the newly developed method can provide efficient quantitative assessment of the reheat cracking sensitivity with high accuracy.     

弥散燃料颗粒裂纹起源的有限元模拟分析

通过建立含多气泡的燃料颗粒模型,采用有限元方法分析了燃料颗粒在裂变气体气泡内压作用下的应力分布,统计了燃料颗粒内部气泡位置对气泡内壁处的最大拉应力的影响,并结合实验结果探寻了弥散燃料颗粒在辐照后退火时的裂纹起源。结果表明:当弥散燃料颗粒内部含有多个裂变气体气泡时,受气泡内压作用,气泡内壁径向应力为压应力,环向应力为拉应力;气泡位置距燃料颗粒心部越远,气泡内壁处的最大环向拉应力越大;表层气泡的最大环向拉应力远大于心部气泡的;燃料颗粒裂纹起源于表层气泡内壁。     

基于模糊状态观测器的核反应堆功率T-S模糊积分控制系统

为提高核反应堆运行的负荷跟随能力,设计了基于T-S模糊模型的包含模糊状态观测器的积分控制系统。该方法将非线性的点堆动力学方程在一些选定工况点进行线性化,采用并行分布补偿（PDC）方法,设计模糊积分控制器和模糊状态观测器,以相对于额定功率平衡状态的相对中子密度为前件变量,构造T-S模糊控制系统。引入模糊状态观测器,成功解决了部分状态变量不能测量的困难。使用线性矩阵不等式（LMI）方法进行稳定性分析,保证了此控制系统的大范围稳定性。仿真结果验证了这种控制系统能在大范围运行工况下工作良好。     

Simulation of the fracture behavior of Zircaloy-4 cladding under reactivity-initiated accident conditions with a damage mechanics model combined with fuel performance codes FEMAXI-7 and RANNS

A continuum damage mechanics model using FEM calculations was proposed to be applied to an analysis of the fuel failure due to pellet cladding mechanical interaction (PCMI) under reactivity-initiated accident conditions. The model expressed ductile fracture via two processes: damage nucleation related to void nucleation and damage evolution related to void growth and linkage. The boundary conditions for the simulations were input from the fuel performance codes FEMAXI-7 and RANNS. The simulation made reasonable predictions for the cladding hoop strain at failure and reproduced the typical fracture behavior of the fuel cladding under the PCMI loading, characterized by a ductile shear zone in the inner region of the cladding wall. It was shown that occurrence of a through-wall crack is determined at an early stage of crack propagation, and the rest of the through-wall penetration process is achieved with a negligible increment in strain. The effect of a local temperature rise in the cladding inner region on the failure strain was found to be less than 5% for the conditions investigated. Failure strains predicted under a plane strain loading were smaller by 20%–30% than those predicted under equibiaxial tensions between the hoop and the axial directions.     

数值堆材料辐照损伤并行动力学蒙特卡罗软件MISA-AKMC的设计与实现

材料辐照损伤模拟是数值堆软件的重要内容。原子动力学蒙特卡罗（AKMC）方法是研究核材料辐照行为的重要手段,可在保持原子级别精度下,有效地扩展模拟的时间尺度到秒甚至年量级。但在面向实际应用需求时,其仍面临内存限制和复杂的计算量等挑战。通过并行计算技术提升计算效率是解决这一挑战的有效手段。本文论述一款并行AKMC程序MISA-AKMC的设计原理与实现技术。MISA-AKMC实现了一种并行KMC模拟框架,重点包括sub-lattice并行算法的实现、加速优化方法、转发通信方法、KMC模型接口等。基于该框架,开发了空位-间隙的演化模型,可实现热老化模拟和析出模拟。通过并行性能测试获得了良好的并行性能,结合算例结果验证说明了MISA-AKMC的正确性和可靠性。