选择并融合粗细粒度特征的细粒度图像识别

目的 在细粒度图像识别任务中,类内图像姿势方差大,需要找到类内变化小的共性,决定该任务依赖于鉴别性局部的细粒度特征;类间图像局部差异小,需要找到类间更全面的不同,决定该任务还需要多样性局部的粗粒度特征。现有方法主要关注粗细粒度下的局部定位,没有考虑如何选择粗细粒度的特征及如何融合不同粒度的特征。为此,提出一种选择并融合粗细粒度特征的细粒度图像识别方法。方法 设计一个细粒度特征选择模块,通过空间选择和通道选择来突出局部的细粒度鉴别性特征;构建一个粗粒度特征选择模块,基于细粒度模块选择后的局部,挖掘各局部间的语义和位置关系,从而获得为细粒度局部提供补充信息的粗粒度多样性特征;融合这两个模块中提取到的细粒度特征和粗粒度特征,形成互补的粗细粒度表示,以提高细粒度图像识别方法的准确性。结果 在CUB-200-2011（caltech-UCSD birds-200-2011）、Stanford Cars和FGVC-Aircraft（fine-grained visual classification aircraft） 3个公开的标准数据集上进行广泛实验,结果表明,所提方法的识别准确率分别达到90.3%、95.6%和94.8%,明显优于目前主流的细粒度图像识别方法,相较于对比方法中的最好结果,准确率相对提升0.7%、0.5%和1.4%。结论 提出的方法能够提取粗粒度和细粒度两种类型的视觉特征,同时保证特征的鉴别性和多样性,使细粒度图像识别的结果更加精准。     

基于3D路径聚合高分辨率网络的胰腺分割

目的 胰腺的准确分割是胰腺癌识别和分析的重要前提。现有基于深度学习的主流胰腺分割网络大多是编码—解码结构,对特征图采用先降低再增加分辨率的方式,严重丢失了胰腺位置和细节信息,导致分割效果不佳。针对上述问题,提出了基于3D路径聚合高分辨率网络的胰腺分割方法。方法 首先,为了捕获更多3D特征上下文信息,将高分辨率网络中的2D运算拓展为3D运算;其次,提出全分辨特征路径聚合模块,利用连续非线性变换缩小全分辨率输入图像与分割头网络输出特征语义差异的同时,减少茎网络下采样丢失的位置和细节信息对分割结果的影响;最后,提出多尺度特征路径聚合模块,利用渐进自适应特征压缩融合方式,避免低分辨率特征通道过度压缩导致的信息内容损失。结果 在公开胰腺数据集上,提出方法在Dice系数（Dice similarity coefficient,DSC）、Jaccard系数（Jaccard index,JI）、精确率（precision）和召回率（recall）上相比3D高分辨率网络（3D high-resolution net,3DHRNet）分别提升了1.41%、2.09%、2.35%和0.49%,相比具有代表性编码—解码结构的胰腺分割方法,取得了更高的分割精度。结论 本文提出的3D路径聚合高分辨率网络（3D pathaggregation high-resolution network,3DPAHRNet）具有更强的特征位置和细节信息的保留能力,能够显著改善在腹部CT（computed tomography）图像中所占比例较小的胰腺器官的分割结果。开源代码可在https：//github.com/qiuchengjian/PAHRNet3D获得。     

Adaptive fault-tolerant control for a class of flexible air-breathing hypersonic vehicles

In this paper, a novel adaptive fault-tolerant control (FTC) scheme is proposed for a class of flexible air-breathing hypersonic vehicles with unknown inertial and aerodynamic parameters and even input constraints. The fault model under consideration covers the case that all actuators suffer from unknown time-varying faults. In the controller design and stability analysis, we introduce new Lyapunov functions, some differentiable auxiliary functions, a bound estimation approach, and a Nussbaum function, which help us successfully circumvent the obstacle caused by the faults, input constraints, and flexible modes. In addition to higher reliability, the proposed scheme is able to ensure that all closed-loop signals are globally uniformly bounded and to steer the tracking errors of altitude and velocity into predefined arbitrarily small residual sets. As a result, the tracking accuracy can be designated in advance. Simulation results illustrate the effectiveness of the proposed scheme.     

On Dual-Mode Driving Control Method for a Novel Unmanned Tractor With High Safety and Reliability

Due to the non-standardization and complexity of the farmland environment, it is always a huge challenge for tractors to achieve fully autonomy (work at Self-driving mode) all the time in agricultural industry. Whereas, when tractors work in the Tele-driving (or Remote driving) mode, the operators are prone to fatigue because they need to concentrate for long periods of time. In response to these, a dual-mode control strategy was proposed to integrate the advantages of both approaches, i.e., by combing Self-driving at most of the time with Tele-driving under special (complex and hazardous) conditions through switching control method. First, the state switcher was proposed, which is used for smooth switching the driving modes according to different working states of a tractor. Then, the state switching control law and the corresponding subsystem tracking controllers were designed. Finally, the effectiveness and superiority of the dual-mode control method were evaluated via actual experimental testing of a tractor whose results show that the proposed control method can switch smoothly, stably, and efficiently between the two driving modes automatically. The average control accuracy has been improved by 20% and 15% respectively, compared to the conventional Tele-driving control and Self-driving control with low-precision navigation. In conclusion, the proposed dual-mode control method can not only satisfy the operation in the complex and changeable farmland environment, but also free drivers from high-intensity and fatiguing work. This provides a perfect application solution and theoretical support for the intelligentization of unmanned farm agricultural machinery with high safety and reliability.      

Adaptive non-singular fast terminal sliding mode based on prescribed performance for robot manipulators

A fast convergent non-singular terminal sliding mode adaptive control law based on prescribed performance is formulated to solve the uncertainties and external disturbances of robot manipulators. First, the tracking error of robot manipulators is transformed by using the prescribed performance function, which improves the transient behaviors and steady-state accuracy of robot manipulators. Then, a novel fast convergent non-singular terminal sliding mode surface is brought up according to the transformed error, and the control law is derived to meet the stability requirements of robot manipulators. In practice, the upper boundary of the lumped disturbances cannot be accurately obtained. Therefore, an adaptive prescribed performance control (PPC) controller to lumped disturbances is brought up to ensure the stability and finite-time convergence of robot manipulators. Finally, the system stability of robot manipulators is proved by the Lyapunov theorem. Simulation results and comparative analysis demonstrate the superiority and robustness of the raised strategy.     

Digital twin-assisted fault diagnosis system for robot joints with insufficient data

The robot joint is an important component of the construction robot, and its fault diagnosis can ensure the exact execution of building jobs, stable operation, and timely prevention of probable safety mishaps. However, deep learning-based fault diagnosis needs a multitude of measured fault data, which is difficult to obtain for various reasons. To solve the problem of insufficient data, a digital twin-assisted fault diagnosis system for robot joints is proposed. First, a simplified dynamics model of the robot joint is developed to generate the virtual entity data which can be used as the X-domain data for the digital twin model. Second, a CycleGAN-based digital twin model is proposed to map the virtual entity (X-domain) data to the physical entity (Y-domain) utilizing only a small amount of measured data. In the end, a test-rig for the robot joint is built to simulate the robot's working conditions, and the CNN-ResNet classifier is utilized to verify the effectiveness of the simulated data generated by the digital twin model. The results show that the fault diagnosis accuracy can be increased from 32.5% to 98.86% utilizing only 400 sets of measured data.     

基于平面电容的油水多界面检测方法研究

含水量是衡量原油质量的重要指标,在原油的生产和储运过程中油水混合液分界面不断变化,因此全过程都要用到高精度传感器对其进行界面检测。基于电容边缘效应设计了一种侵入式平面电容传感器,其主要结构由基板和平面电极阵列两部分组成。运用有限元软件建立8电极阵列传感器模型,对不同电极工作时的电场分布进行研究,分析了平面电容传感器的检测灵敏度和成像精度。并且,研究了电极宽度、长度和相邻间距对传感器灵敏场分布的影响。对介电分布进行图像重建,使设计的平面电容阵列传感器可以检测3个分界面的高度,且经过尺寸参数优化,提高了传感器的成像精度。实验证明,运用平面电容阵列检测油水界面的方法具有可行性和有效性。     

大型发电机定子线棒端部防晕结构温度和电场分布仿真研究

大型发电机定子线棒端部防晕层温度和电场分布是表征防晕结构性能的重要参数,关系着发电机的长期安全稳定运行。工频电气试验过程中,防晕材料的电导损耗是线棒端部绝缘温度梯度分布的主要原因,电场分布取决于防晕层电导和主绝缘电容,而电导率又是电场强度的函数,因此发电机定子线棒端部防晕结构温度场和电场的计算是一个相互耦合的过程。本文借助COMSOL Multiphysics有限元分析软件多物理场仿真功能,以某大型水氢冷发电机定子线棒端部防晕结构作为研究对象,仿真分析了其工频状况下的电场分布情况和时域条件下的温度场分布情况,以匹配线棒端部防晕结构设计开发过程。将仿真温度场结果与真机线棒试验结果对比,两者吻合较好,为进一步优化设计防晕结构和材料参数奠定了基础。     

新型电力负荷管理系统国内标准体系研究与建议

现阶段,电力负荷管理系统标准体系化程度需要进一步加强。国内相关标准主要分布在电力负荷管理、用电信息采集两个业务分支内,各项标准内容与当前信息通信技术、自动化控制技术、大数据分析技术的发展不匹配,标准的功能及性能指标缺乏先进性,很多关键内容难以适应当下负荷精准分类、实时监测、实时控制的需要。首先阐述了国内电力负荷管理系统标准体系架构,包括系统类、终端类、接口类和管理类4个方面内容;其次对电力负荷管理系统标准体系现状进行分类梳理和解读,每个方面选取典型标准进行介绍,探讨了各个环节标准发展现状及实际应用中存在的问题;最后结合电网发展对电力负荷管理系统的新要求以及实际运行中面临的新问题,提出新型电力负荷管理系统标准体系完善建议及重点行动计划,对今后新型电力负荷管理系统相关标准的修订具有一定参考价值。