基于空间灰度重心推进的线结构光中心提取

利用线结构光对工件进行三维测量时,快速、准确提取激光光条中心是测量系统的关键技术之一。提出一种基于空间灰度重心推进的激光中心提取方法。该方法基于激光条纹空间灰度重心向前推进,结合八邻域判定法进行光条空间区域的选取,接着对选取的区域进行新的空间灰度重心坐标提取;在完成整幅图像提取后,采取3σ准则剔除异常中心点,最终获得激光中心坐标信息。实验结果表明,所提算法对不同类型条纹的适用性强,提取效果稳定。提取中心均方根误差为0.492pixel,相较于提取精度最优的Steger改进算法提升了9.8%,提取速度提高了5倍;相较于处理速度最优的内部推进算法,在保持提取速度基础上,提取精度提升了24.1%。同时,所提算法极大增强了对欠曝光光条图像的处理能力,有效降低了环境对线结构光测量的限制。

基于中心点对特征的无序堆叠工件位姿估计算法

针对无序堆叠场景中工件相互遮挡导致的位姿估计精度下降的问题,提出一种基于中心点对特征的六维位姿估计算法。首先,模拟无序堆叠场景下点云目标多姿态随机摆放的物理环境,生成特征提取网络需要的带真实标签的数据集。进而,以中心点对特征构建离线阶段的全局特征描述。然后,对在线阶段的工件无序堆叠场景,利用动态图卷积神经网络(dynamic graph convolutional neural network, DGCNN)算法提取点云中心特征分数,确定工件中心点,并以此作为改进点对特征(point pair feature, PPF)算法的参考点。最后,使用IPA数据集和自采集场景数据对算法性能进行了验证。实验结果表明:所提算法降低了参考点选择的随机性,相比原始PPF算法在场景工件数目为30时平均准确率提升19.5百分点,5种工件场景下平均运行时间缩短29.00%。

高反光物体的局部点云孔洞补全方法

针对高反光物体点云数据存在噪声明显、孔洞过大,使用传统贪婪投影三角化算法无法达到点云补全目的的问题,提出一种改进的贪婪三角化点云补全算法。利用统计滤波和高斯滤波进行离群点去除和平滑处理;采用移动最小二乘法对局部点云进行上采样增强数据,并且对点云进行进一步平滑。将贪婪三角化算法中的kd树(k-dimensional tree)算法替换为效率更高的八叉树(octree)搜索算法,将主成分分析法替换为准确率更高的移动最小二乘法进行法线估计;最后,进行点云三角化,完成点云补全。实验结果表明,改进算法能够更好地补全孔洞,补全后表面更加平滑、结构更为准确,且花费时间更短。

基于改进Mask R-CNN的饮料分拣系统实例分割算法

针对饮料瓶摆放过程中的遮挡、光线干扰等因素影响实例分割精度问题,提出了改进掩膜区域卷积神经网络(mask region convolutional neural network,Mask R-CNN)的饮料瓶实例分割算法(drink Mask R-CNN,DRK-Mask R-CNN)。在原Mask R-CNN的网络结构上,使用ResNet101作为网络的特征提取主干网络,通过引入Transformer中的自注意力模块获取全局与局部的联系;通过引入通道注意力模块以及空间注意力模块增加表现力,关注重要特征,并抑制不必要特征;并引入空洞卷积网络模块以扩大感受野,获取多尺度上下文信息,降低计算量。在此基础上完成对自建饮料瓶数据集的数据训练,完成对不同饮料瓶的识别与图像分割。实验结果表明,DRK-Mask R-CNN算法的实例分割平均精度为95.8%,相较于原算法提高4.7百分点。

基于改进YOLOv3-Tiny网络的嵌入式物流商品查验系统设计与实现

针对电商店家在发货前人工验货效率低下的问题,设计一种基于改进YOLOv3-Tiny网络图像目标检测算法的嵌入式商品查验系统,可高效完成物流商品查验工作。系统通过摄像头获得实时视频帧图片,将帧图片送入改进的YOLOv3-Tiny网络,直接在图片上识别回归分析出商品名称信息及坐标信息,与本地订单数据进行比较,判断配货是否正确。整个系统可在Linux或Windows环境下使用。系统优化了YOLOv3-Tiny的特征融合部分、卷积层、锚框数目、锚框尺寸。实验结果表明,该系统单次商品检测分析的运行时间在355~370 ms之间,准确率为96.15%,满足商品实时查验检测的使用要求。

多特征融合的抗遮挡相关滤波目标跟踪算法

针对传统相关滤波目标跟踪算法由于背景杂乱、目标尺度变化、目标遮挡等导致鲁棒性较差的问题,提出一种多特征融合抗遮挡的相关滤波目标跟踪算法。该算法对核相关滤波(kernel correlation filter,KCF)算法进行改进,通过将方向梯度直方图(histogram of oriented gradient,HOG)特征和多尺度块局部二值模式(multiscale block local binary pattern,MB-LBP)特征加权融合,并将所得特征再次与颜色(color name,CN)特征进行二次加权融合的方式来增强特征表达能力;增加尺度滤波器用于目标尺度估计;利用峰值旁瓣比(peak sidelobe ratio,PSR)和目标重定位算法对被遮挡目标进行判定和检测。选取OTB100数据集,将改进算法与几种主流算法进行了跟踪效果对比。实验结果表明,改进算法在背景杂乱、目标尺度变化、目标遮挡等情况下具有较好的鲁棒性。

工业室内环境中建立增强现实系统模型研究

针对增强现实(AR)技术在工业领域的应用问题,提出了一个在工业室内环境下应用的增强现实系统模型。该模型主要针对工业中不会在较短时间内发生较大变化的室内环境,首先对头盔式显示器(HMD)采集的环境图像通过Harris检测算法来检测特征点,然后通过程序比对图像在特征点添加相关注释,最后将包含注释的现实图像显示在头盔显示器上。主要介绍了该模型的具体步骤及实现方法,Harris角点检测算法具有应对光强变化、图像旋转等特点,故增加了模型整体的鲁棒性。最后将该模型应用于风力发电机机舱内部环境进行验证,其中特征点检测部分可根据阈值调节特征点数的多少,实验中采用的特征点数为20,在图像匹配部分匹配程度超过85%,实验结果表明该模型实现了增强现实的功能。

基于改进蚁群算法包装机器人轨迹规划研究

针对六轴包装机器人在货物抓取到放入过程,由于路径规划时间长计算出的路径不准确导致机械臂在包装过程中的平稳性和重复精度不精确的问题,提出一种利用人工势场法和蚁群算法相结合的轨迹规划方法。通过将蚁群算法状态转移函数中加入人工势场合力,增加蚁群算法的初始值。基于MATLAB用D-H参数建立机器人模型,对改进后的算法进行仿真验证,对传统蚁群算法和改进蚁群算法分别在二维环境和三维环境中进行仿真验证,最后在真实环境下进行实验验证。结果表明,相较于传统算法改进后的算法具有较高的精度和稳定性,在搜索路径耗时、迭代次数和收敛速度等方面能够提高两倍以上。

基于多特征融合的HL-S工件识别算法

为解决目前工业生产过程中,工件检测技术在复杂环境下特征不清晰且背景复杂情况下识别的不准确等问题,提出了一种多特征融合的HL-S(HOG、LBP特征融合-Softmax分类识别)工件识别算法。首先对获取的图片进行预处理操作,将光照较暗不均匀的图像执行Retinex图像增强,在此操作中使用双边滤波代替高斯滤波;然后进行工件的HOG轮廓特征和LBP纹理特征的提取操作,使用线性融合的方法对处理后的图像进行特征融合;最后,将所有样本数据通过Softmax分类器进行归类判别。实验结果表明,使用该方法进行工件识别,能够提高在光照条件不好情况下的识别率,大大提高了工业生产的效率。

复杂背景下的风力机组叶片图像分割方法研究

针对单一的图像分割方法对复杂背景下的风力机组叶片图像进行分割时,得到的结果不具备准确性的问题,提出一种将Canny边缘检测与形态学分析相结合的方法,对Canny算法处理之后的二值图像进行腐蚀膨胀,去除背景区域。在分割一种具有颜色差异的风机叶片图像时,将RGB图像转化为HSV图像,再分别对HSV图像的色度、饱和度、和明度分量进行处理,并运用开运算去除图像噪声,得到理想的分割结果。实验结果表明,提出的图像分割方法能得到完整的风机叶片图像分割结果。

基于卷积神经网络的燃气轮机故障诊断研究

针对燃气轮机故障信号为变速非平稳信号的特点,使用非线性调频分量分解方法对燃气轮机机匣振动信号进行时频分析,并利用卷积神经网络对燃气轮机转子进行故障诊断。首先使用非线性调频分量分解方法将信号分解成多个频率分量,得到各个频率成分的瞬时频率,并绘制时频图。将时频图经过灰度处理以及尺寸压缩等预处理后作为输入训练卷积神经网络,利用卷积神经网络强大的特征提取能力实现燃气轮机转子故障的有效诊断。实验结果表明,针对燃气轮机转子故障,该模型能够获得很好的诊断效果,测试准确率约为99%。

基于YOLO-GT网络的零售商品目标检测方法

提出改进的YOLO-GT深度学习网络,实现基于嵌入式平台的零售商品数据集RPC的图像目标检测。YOLO-GT网络以YOLOv3-Tiny网络为基础,采用线性扩展优化用于特征提取的卷积层,并将深度学习网络浅层信息与深层信息融合使得网络成为三尺度预测网络,提高网络对小目标商品的识别能力。YOLO-GT采用Mish函数优化了网络激活函数,并重新利用FCM聚类算法生成多个新的先验框。利用RPC零售商品数据集在嵌入式平台Jetson nano核心板上实验,YOLO-GT网络对于每张图片的平均检测速度为342.67 ms/张,平均检测准确率(m AP)为97.21%,与YOLOv3-Tiny相比,检测速度提升了57.81%,m AP提升了1.84%,并降低了小目标漏检率,能够满足使用要求。

基于BOA服务器的同步数据采集管理系统设计

针对目前风电机组管理中存在的监测数据获取和查看不便等问题,提出了一种基于B/S架构的嵌入式八通道同步数据采集配置管理系统。系统采用以ARM232为主控板的核心处理器,以Linux嵌入式系统为运行平台并搭建了Boa服务器;系统通过CGI接口技术实现了客户端与服务器端的数据交互,并采用Ajax进行异步数据传输、以JSON格式作为数据交互格式。同时在登录成功之后使用Session机制保存登录状态并实现超时退出功能,保证了系统准确、高效、安全的运行。经测试,该系统在运行时占用内存资源少,能够实现在页面整体无刷新的条件下对数据进行更新,运行稳定。

基于改进Mask RCNN的道路信息检测算法

针对目前目标检测算法应用于道路信息检测精度低、速度慢、小目标检测效果差的问题,提出一种基于改进掩膜区域卷积神经网络(mask region convolutional network, Mask RCNN)的道路信息检测算法。引入深度可分离卷积提升检测速度;引入卷积注意力模块(convolutional block attention module, CBAM)与双向特征金字塔网络(bidirectional feature pyramid network, Bi-FPN)提高模型精度;借鉴迁移学习思想,基于PASCAL-VOC2012数据集对模型预训练,提高模型学习特征的能力;基于自制道路信息数据集完成模型正式训练。实验结果表明,使用改进方法优化后的基于ResNet50的Mask RCNN算法整体性能较好,平均精度均值达到95.2%,较原算法提高了4.5%,检测帧率达到24.8帧/s,较原算法提高了8.3帧/s,且小目标漏检现象变少,证明改进方法可以提高道路信息检测算法的检测精度、检测速度与小目标检测性能。

Early faults prediction of running state of electromechanical systems and reconfigurable integration of series safety monitoring systems

Fault prediction technology of running state of electromechanical systems is one of the key technologies that ensure safe and reliable operation of electromechanical equipment in health state. For multiple types of modern, high-end and key electromechanical equipment, this paper will describe the early faults prediction method for multi-type electromechanical systems, which is favorable for predicting early faults of complex electromechanical systems in non-stationary, nonlinear, variable working conditions and long-time running state; the paper shall introduce the reconfigurable integration technology of series safety monitoring systems based on which the integrated development platform of series safety monitoring systems is built. This platform can adapt to integrated R&D of series safety monitoring systems characterized by high technology, multiple species and low volume. With the help of this platform, series safety monitoring systems were developed, and the Remote Network Security Monitoring Center for Facility Groups was built. Experimental research and engineering applications show that: this new fault prediction method has realized the development trend features extraction of typical electromechanical systems, multi-information fusion, intelligent information decision-making and so on, improving the processing accuracy, relevance and applicability of information; new reconfigurable integration technologies have improved the integration level and R&D efficiency of series safety monitoring systems as well as expanded the scope of application; the series safety monitoring systems developed based on reconfigurable integration platform has already played an important role in many aspects including ensuring safety operation of equipment, stabilizing product quality, optimizing running state, saving energy consumption, reducing environmental pollution, improving working conditions, carrying out scientific maintenance, advancing equipment utilization, saving maintenance charge and enhancing the level of information management.

Monitoring and control technology of solar photovoltaic power generation system

In view of characteristics of solar photovoltaic (PV) power station such as the decentralized layout and massive monitoring and control information, a solar PV power generation monitoring and control system has been designed. The system is designed into three layers namely the sensor and actuator layer, the PLC field monitoring and control layer and the remote network monitoring and control layer. Through ZigBee wireless network, PROFIBUS and GPRS wireless network, the system makes the three layers exchange information rapidly, and the system supervises not only various operational parameters of the power generating system but also weather changes as a way to change the solar tracking strategy of the PV power generating system and reduce the operating energy consumption of the system. Through the hardware redundant design of PLC central controller and the upper computer, the solar PV power station can be more secure and reliable when running.

Application of Noise Reduction Method Based on Birgé-Massart Threshold

Flue gas generator set is a kind of large high-speed rotating machinery in petrochemical industry.This research focuses on noise reduction algorithms basis ontheBirgé-Massartthreshold.Obtained the threshold through Penalization Strategy Provided by Birgé-Massart;constructed different modulus maximum vertex neighborhood on different wavelet transform decomposition scales to influence the search process of modulus maximum point;obtained the appropriate modulus maximum points sequence on various wavelet decomposition scales;highlighted state feature information;finally usedMallat staggered projection to reconstruct signals.In order to validate the effectiveness of the algorithm,it was compared with four kinds of threshold noise suppression methods namely Rigrsure,Sqtwolog,Heursure,Minimaxi.The results show that this algorithm has a better signal to noise ratio and mean-square error.

Research on information technology of state monitoring and fault prediction for mechatronics system

The safety and reliability of mechatronics systems,particularly the high-end,large and key mechatronics equipment in service,can strongly influence on production efficiency,personnel safety,resources and environment.Based on the demands of development of modern industries and technologies such as international industry 4.0,Made-in-China 2025 and Internet + and so on,this paper started from revealing the regularity of evolution of running state of equipment and the methods of signal processing of low signal noise ratio,proposed the key information technology of state monitoring and earlyfault-warning for equipment,put forward the typical technical line and major technical content,introduced the application of the technology to realize modern predictive maintenance of equipment and introduced the development of relevant safety monitoring instruments.The technology will play an important role in ensuring the safety of equipment in service,preventing accidents and realizing scientific maintenance.

Research on motion compensation method based on neural network of radial basis function

The machining precision not only depends on accurate mechanical structure but also depends on motion compensation method. If manufacturing precision of mechanical structure cannot be improved, the motion compensation is a reasonable way to improve motion precision. A motion compensation method based on neural network of radial basis function(RBF) was presented in this paper. It utilized the infinite approximation advantage of RBF neural network to fit the motion error curve. The best hidden neural quantity was optimized by training the motion error data and calculating the total sum of squares. The best curve coefficient matrix was got and used to calculate motion compensation values. The experiments showed that the motion errors could be reduced obviously by utilizing the method in this paper.