基于深度学习的零件实例分割识别研究

针对传统零件识别方法图像特征提取鲁棒性不足,零件识别准确率较低、不能对图像进行实例分割的问题,文章提出了一种基于Mask R-CNN的零件识别方法。该方法利用卷积神经网络对零件图像进行特征提取,选取数据集中标注好的图像微调Mask R-CNN网络,以保证零件识别的准确性,并生成Mask分割掩码,对零件进行实例分割。同时,对数据集进行数据增强和划分K折交叉验证来提高模型的鲁棒性。最后通过搭建实验平台对零件进行识别,证明了该方法的有效性。     

基于CS优化深度学习卷积神经网络的目标检测算法

目前对于形状比较复杂且密集摆放的工件,传统的工业机器人视觉分拣技术已经无法有效检测和识别。因此,为了提高生产线上分拣工件检测的准确率,提出了一种基于布谷鸟搜索算法(Cuckoo Search,CS)优化深度学习卷积神经网络(Convolutional Neural Network,CNN)的目标检测算法。首先对视觉分拣系统的组成进行了分析。然后采用经典的Faster R-CNN的模型结构来实现目标检测,并将CS优化算法应用到CNN模型的参数训练中,解决了反向传播的局部最优问题,同时提高了迭代速度。工件检测实验结果表明:相比于传统的CNN模型,提出CS-CNN模型具有更好的目标检测的准确率,提高了网络的收敛速度。     

基于CS优化深度学习卷积神经网络的目标检测算法（英文）

目前对于形状比较复杂且密集摆放的工件,传统的工业机器人视觉分拣技术已经无法有效检测和识别。因此,为了提高生产线上分拣工件检测的准确率,提出了一种基于布谷鸟搜索算法(Cuckoo Search,CS)优化深度学习卷积神经网络(Convolutional Neural Network,CNN)的目标检测算法。首先对视觉分拣系统的组成进行了分析。然后采用经典的Faster R-CNN的模型结构来实现目标检测,并将CS优化算法应用到CNN模型的参数训练中,解决了反向传播的局部最优问题,同时提高了迭代速度。工件检测实验结果表明:相比于传统的CNN模型,提出CS-CNN模型具有更好的目标检测的准确率,提高了网络的收敛速度。     

基于单目视觉的工业机器人拆垛系统设计与实验

基于单目视觉提出了工业机器人拆垛系统构想。为实现工业机器人视觉拆垛系统中机器人精确运动控制与工件位姿识别,运用D-H位移矩阵法建立了机器人运动学模型,得到机器人末端相对机器人坐标系的位姿信息。基于形状模板匹配法提出目标图像识别算法,得到各个目标图像在相机坐标系下的位姿信息;将三维视觉模型与机器人运动学模型进行信息融合,建立机器人视觉拆垛控制系统数学模型。基于CCD工业相机、4-DOF工业机器人搭建视觉定位抓取实验系统。通过对空间目标进行抓取的实验,验证基于单目视觉的工业机器人拆垛系统的正确性、精确性、鲁棒性。     

基于四自由度视觉机器人的织物分拣方法

针对人工分拣织物效率低,劳动成本高,而且对工人身体产生危害的问题,提出机器视觉与机器人相结合代替人工分拣织物的方法,从而解决了人工分拣织物存在的缺陷。构造RBF神经网络的颜色分类器将不同颜色织物分类,利用阈值分割对目标织物进行分割;使用遍历中心扩散法求取机器人抓取织物的抓取点;经过摄像机标定和机器人轨迹规划实现机器人对纯色织物的自主分拣。实验结果表明,该方法分拣织物的准确率能够满足纺织业的实际应用需求。     

基于PointECA网络的无序工件点云分割算法

针对无序、采样不均匀以及存在相互遮挡的工件点云分割效果不佳的问题,提出一种多尺度自适应通道维度注意力点云分割网络(PointECA)。该算法中的多尺度特征提取模块能够较好地融合不同尺度的局部邻域特征,得到较为丰富的全局特征信息;自适应性通道注意力模块能够对不同尺度局部特征的通道维度交互学习,实现较好的语义分割效果。此外,制作了用于语义分割实验的Workpieces数据集。大量实验数据表明：PointECA在无序且有相互遮挡场景下,对工件部件分割的平均交并比达到了95.42%,能够为无序工件的快速分拣提供较好的条件。     

基于视觉的工业机器人装配演示示教研究

针对工业机器人编程效率低下、智能化程度不高和人机交互性能弱等问题，提出一种基于视觉的工业机器人装配演示示教系统，该系统包括目标检测与中心点定位模块、装配动作分类识别模块和机器人动作执行模块。在目标检测与中心点定位模块中,提出一种目标物体中心点定位和机器人抓取方法，使用实例分割算法识别物体类别，通过掩码均值化处理和坐标转换计算物体3D姿态信息；在装配动作分类识别模块中，建立基于深度学习网络的动作分类识别模型，该模型的输入为装配动作视频帧，输出为动作分类标签；最后，机器人动作执行模块根据物体类别、物体3D姿态和动作分类标签等信息规划机器人装配动作，控制机器人执行装配任务。以轴孔装配为例，验证了上述方法的有效性，实现了基于视觉演示的机器人装配模仿编程，对机器人演示示教研究具有一定的参考价值。     

基于深度学习的固结磨料研磨垫表面形态表征

固结磨料研磨垫的表面形态与其加工性能有着密切关系,为更好地了解固结磨料研磨垫表面形态,尤其是研磨垫中的金刚石、孔隙、金刚石脱落坑等的分布特征,提出一种基于深度学习的固结磨料研磨垫表面形态分析方法。首先,利用徕卡DVM6数字显微镜及其配套软件获取固结磨料研磨垫表面图像;然后,采用python3+OpenCV对图像进行预处理,并利用标注软件Labelme对图像进行标注,用于后续的训练和测试;最后,运用深度学习框架Tensorflow搭建Mask R-CNN模型。结果表明：Mask R-CNN模型能对单一固结磨料垫表面图像中的多目标进行有效分割与识别,其主要评价指标平均准确率达到78.9%,达到了图像识别的主流水平。     

基于3D视觉的机械零件位姿分析

针对工业制造领域中存在的弱纹理或无纹理的机械零件位姿估计问题，以DenseFusion网络为基础，提出了一种基于3D视觉的机械零件位姿估计方法。首先，依照工业制造场景构建了用于网络训练和测试的机械零件仿真数据集；其次，使用目标零件分割出的深度图像构建点云，对其进行曲率下采样处理，提高关键点的质量；最后，将颜色特征、点云特征和法线特征融合，使用融合特征回归目标零件的6D位姿。在构建的机械零件仿真数据集和LineMod公共数据集进行了实验和比较，其结果表明了所提出的方法相较于同类其他方法具有更高的准确率，更好的收敛性能，对于弱纹理或无纹理的机械零件有较好的位姿估计效果。     

基于改进Faster R-CNN的轴承表面缺陷检测

针对轴承表面缺陷小、几何形状多变以及低对比度的特点，提出了一种改进的Faster R-CNN算法，对轴承表面缺陷进行检测。首先，以ResNet-50结合特征金字塔网络对轴承表面缺陷进行特征提取；其次，在改进的特征提取网络中引入可变形卷积，通过卷积学习偏移量自适应调整感受野，提高了缺陷的提取能力；最后，针对ROI Pooling因二次量化而导致的区域不匹配问题，采用基于双线插值的ROI Align改进ROI Pooling。实验结果表明，在采集的轴承表面缺陷数据集上，改进的Faster R-CNN平均精度均值为97.6%,与改进前相比，提高了11.76%,可以实现对轴承表面各类缺陷更为准确的检测，具有较强的实用性。     

Effects of Post-weld Heat Treatment on Microstructure,Mechanical Properties and the Role of Weld Reinforcement in 2219 Aluminum Alloy TIG-Welded Joints

In as-welded state, each region of 2219 aluminum alloy TIG-welded joint shows diff erent microstructure and microhardness due to the diff erent welding heat cycles and the resulting evolution of second phases. After the post-weld heat treatment, both the amount and the size of the eutectic structure or θ phases decreased. Correspondingly, both the Cu content in α-Al matrix and the microhardness increased to a similar level in each region of the joint, and the tensile strength of the entire joint was greatly improved. Post-weld heat treatment played the role of solid solution strengthening and aging strengthening. After the post-weld heat treatment, the weld performance became similar to other regions, but weld reinforcements lost their reinforcing eff ect on the weld and their existence was more of an adverse eff ect. The joint without weld reinforcements after the post-weld heat treatment had the optimal tensile properties, and the specimens randomly crack in the weld zone.     

The First Phase of the SRIF Industrialized Base in ＂West Area＂ Has Been Put into Operation

After nearly two years＇ tense construction, the first phase of industrialized base of Shenyang Research Institute of Foundry （SRIF）, located at the Tiexi Casting and Forging Industrial Park in the west of Tiexi District, has now been completed and formally put into operation.     

Thermodynamics Study on the Decomposition of Chromite with KOH

Institute of Process Engineering, Chinese Academy of Sciences, China, has proposed a method for oxidative leaching of chromite with potassium hydroxide. Understanding the mechanism of chromite decomposition, especially in the potassium hydroxide fusion, is important for the optimization of the operating parameters of the oxidative leaching process. A traditional thermodynamic method is proposed and the thermal decomposition and the reaction decomposition during the oxidative leaching of chromite with KOH and oxygen is discussed, which suggests that chromite is mainly destroyed by reactions with KOH and oxygen. Meanwhile, equilibrium of the main reactions of the above process was calculated at different temperatures and oxygen partial pressures. The stable zones of productions, namely, K2CrO4 and Fe2O3, increase with the decrease of temperature, which indicates that higher temperature is not beneficial to thermodynamic reactions. In addition, a comparison of the general alkali methods is carried out, and it is concluded that the KOH leaching process is thermodynamically superior to the conventional chromate production process.     

Influence of Heat Treatment on Damping Behaviour of the Magnesium Wrought Alloy AZ61

The effect of isochronal heat treatments for 1h on variation of damping, hardness and microstructural change of the magnesium wrought alloy AZ61 was investigated. Damping and hardness behaviour could be attributed to the evolution of precipitation process. The influence of precipitation on damping behaviour was explained in the framework of the dislocation string model of Granato and Lücke.     

Effect of Rare Earth Doping on Impedance,Modulus and Conductivity Properties of Multiferroic Composites:0.5(BiLa_xFe_(1-x)O_3)–0.5(PbTiO_3)

The Lanthanum-doped bismuth ferrite–lead titanate compositions of 0.5(Bi LaxFe1-xO3)–0.5(Pb Ti O3)(x = 0.05,0.10,0.15,0.20)(BLxF1-x-PT) were prepared by mixed oxide method.Structural characterization was performed by X-ray diffraction and shows a tetragonal structure at room temperature.The lattice parameter c/a ratio decreases with increasing of La(x = 0.05–0.20) concentration of the composites.The effect of charge carrier/ion hopping mechanism,conductivity,relaxation process and impedance parameters was studied using an impedance analyzer in a wide frequency range(102–106Hz) at different temperatures.The nature of Nyquist plot confirms the presence of bulk effects only,and non-Debye type of relaxation processes occurs in the composites.The electrical modulus exhibits an important role of the hopping mechanism in the electrical transport process of the materials.The ac conductivity and dc conductivity of the materials were studied,and the activation energy found to be 0.81,0.77,0.76 and 0.74 e V for all compositions of x = 0.05–0.20 at different temperatures(200–300 °C).     

Quantitative Analysis of the Crystallographic Orientation Relationship Between the Martensite and Austenite in Quenching–Partitioning–Tempering Steels

The orientation relationships(ORs)between the martensite and the retained austenite in low-and medium-carbon steels after quenching–partitioning–tempering process were studied in this work.The ORs in the studied steels are identified by selected-area electron diffraction(SAED)as either K–S or N–W ORs.Meanwhile,the ORs were also studied based on numerical fitting of electron backscatter diffraction data method suggested by Miyamoto.The simulated K–S and N–W ORs in the low-index directions generally do not well coincide with the experimental pole figure,which may be attributed to both the orientation spread from the ideal variant orientations and high symmetry of the low-index directions.However,the simulated results coincide well with experimental pole figures in the high-index directions{123}_(bcc).A modified method with simplicity based on Miyamoto’s work was proposed.The results indicate that the ORs determined by modified method are similar to those determined by Miyamoto’method,that is,the OR is near K–S OR for the low-carbon Q–P–T steel,and with the increase of carbon content,the OR is closer to N–W OR in medium-carbon Q–P–T steel.     

Using Finite Element and Contour Method to Evaluate Residual Stress in Thick Ti-6Al-4V Alloy Welded by Electron Beam Welding

This work was to reveal the residual stress profile in electron beam welded Ti-6Al-4V alloy plates(50 mm thick) by using finite element and contour measurement methods.A three-dimensional finite element model of 50-mmthick titanium component was proposed,in which a column–cone combined heat source model was used to simulate the temperature field and a thermo-elastic–plastic model to analyze residual stress in a weld joint based on ABAQUS software.Considering the uncertainty of welding simulation,the computation was calibrated by experimental data of contour measurement method.Both test and simulated results show that residual stresses on the surface and inside the weld zone are significantly different and present a narrow and large gradient feature in the weld joint.The peak tensile stress exceeds the yield strength of base materials inside weld,which are distinctly different from residual stress of the thin Ti-6Al-4V alloy plates presented in references before.     

Microstructures and Tensile Properties of Ultrafine-Grained Ni–(1–3.5) wt% SiC_(NP) Composites Prepared by a Powder Metallurgy Route

Silicon carbide nanoparticle-reinforced nickel-based composites(Ni–Si CNP),with a Si CNPcontent ranged from1 to 3.5 wt%,were prepared using mechanical alloying and spark plasma sintering.In addition,unreinforced pure nickel samples were also prepared for comparative purposes.To characterize the microstructural properties of both the unreinforced pure nickel and the Ni–Si CNPcomposites transmission electron microscopy(TEM) was used,while their mechanical behavior was investigated using the Vickers pyramid method for hardness measurements and a universal tensile testing machine for tensile tests.TEM results showed an array of dislocation lines decorated in the sintered pure nickel sample,whereas,for the Ni–Si CNPcomposites,the presence of nano-dispersed Si CNPand twinning crystals was observed.These homogeneously distributed Si CNPwere found located either within the matrix,between twins or on grain boundaries.For the Ni–Si CNPcomposites,coerced coarsening of the Si CNPassembly occurred with increasing Si CNPcontent.Furthermore,the grain sizes of the Ni–Si CNPcomposites were much finer than that of the unreinforced pure nickel,which was considered to be due to the composite ball milling process.In all cases,the Ni–Si CNPcomposites showed higher strengths and hardness values than the unreinforced pure nickel,likely due to a combination of dispersion strengthening(Orowan effects) and particle strengthening(Hall–Petch effects).For the Ni–Si CNPcomposites,the strength increased initially and then decreased as a function of Si CNPcontent,whereas their elongation percentages decreased linearly.Compared to all materials tested,the Ni–Si CNPcomposite containing 1.5% Si C was found more superior considering both their strength and plastic properties.     

Directional Growth of Tin Crystals Controlled by Combined Solute Concentration Gradient Field and Static Magnetic Field

A new method was introduced to achieve directional growth of Sn crystals. Microstructures in liquid(Pb)/liquid(Sn) diffusion couples were investigated under various static magnetic fields. Results show that the β-Sn crystals mainly reveal an irregular dendritic morphology without or with a relatively low static magnetic field(B0.3 T). When the magnetic field is increased to 0.5 T, the β-Sn dendrites close to the final stage of growth begin to show some directional character. With a further increase in the magnetic field to a higher level(0.8–5 T), the β-Sn dendrites have an enhanced directional growth character, but the dendrites show a certain deflection. As the magnetic field is increased to 12 T, the directional growth of the β-Sn dendrites in the center of the couple is severely destroyed. The mechanism of the directional growth of the β-Sn crystals and the deflection of the β-Sn crystals with the application of static magnetic field was tentatively discussed.     

Growth Kinetics of Single Inclusion Particle in Molten Melts

On the basis of the single-particle framework, a new theory on inclusion growth in metallurgical melts is developed to study the kinetics of inclusion growth on account of reaction and collision. The studies show that the early growth of inclusion depends on reaction growth and Brawnian motion collision, and where the former is decisive, the late growth depends on turbulence collision and Stokes＇ collision, and where the former is dominant; collision growth is very quick during the smelting process, lessened in the refining process, but nearly negligible in the continuous casting process.