Advances in Machine Vision and Manufacturing Automation

对于制造商来说，提高自动化水平是生存的必要条件，才能保持竞争力。机器视觉(MV)是自动化工具包中的关键工具之一，因为它提供了对零件和生产这些零件的机器的基于图像的检测和分析。这项工作将分为两个项目，并将导致培养2名博士，3名硕士和5名理科学生。总体目标是用新颖的算法和系统设计推动基于MV的系统的最新水平，为加拿大工业更多地接受自动化技术奠定基础。项目A是关于部件检测的MV主题。几十年来，人们已经认识到需要一种自适应机器视觉(AMV)系统，这种系统可以在不同的应用中实现，而不需要广泛的重新调整。到目前为止已经开发的AMV零件检测系统一直是伪自适应的，从这个意义上说，它们没有在不同的应用上进行测试，而是在相同零件的不同型号上进行了测试。该项目的目标是开发一种真正的自适应系统，并在3种不同的应用上进行测试：硬币、齿轮和压敏电阻。最初的申请是一个特别具有挑战性的问题，因为“部分”是一枚印度硬币。出于测试目的，硬币将被放置在移动的传送带上，以模拟制造操作。该系统将被要求实时识别并分类硬币。传统的MV系统包括两个部分：1)以图像为输入的特征选择；2)以特征为输入的分类。根据到目前为止的经验，提出的AMV系统将使用基于偶然性的特征选择器和新的基于模糊决策树的分类器。性能将以AlexNet(一种非传统的深度神经网络)为基准。项目B的主题是机器故障检测的MV。自动化装配机日以继夜地运转，以实现高生产率。连续操作会导致高机械磨损，从而可能导致机器故障。传统的故障检测方法使用多个常规传感器来检查与固定阈值的偏差。该项目的目标是开发一种基于MV的检测系统，用单个摄像头检测已知和未知的故障。一台高速工业装配机可用于这项工作。该方法将基于高斯混合模型(GMM)方法进行视频分析。分析的出发点是找出偏离正常的图像，换句话说，这是一种“故障”。然后对图像进行分析，以找出差异发生在哪里。此定位阶段将提供有关故障性质的基本信息。因此，提出的基于GMM的系统开始检测和定位机器上的故障，同时将诊断留给操作员。将使用MATLAB进行离线软件原型设计，使用OpenCV进行在线测试。

For manufacturers to stay competitive, increased levels of automation are a must for survival. Machine vision (MV) is one of the key tools in the automation toolkit as it provides for image-based inspection and analysis of parts and the machines that produce those parts. The work will be broken down into two projects, and will lead to the training of 2 PhD, 3 MSc and 5 BSc students. The overall goal is to advance the state of the art of MV-based systems with novel algorithms and system designs that will create the basis for greater acceptance of automation technology by Canadian industry.Project A is on the subject of MV for part inspection. For decades it has been recognized that there is a need for an adaptive machine vision (AMV) system, one that can be implemented in different applications without extensive retuning. AMV part inspection systems that have been developed to date have been pseudo adaptive, in the sense that they were not tested on different applications, but instead tested on different models of the same part. The goal of this project is to develop a truly adaptive system and test it on 3 different applications: coins, gears and varistors. The initial application is a particularly challenging problem where the “part” is an Indian coin. For testing purposes, coins will be placed on a moving conveyor to mimic a manufacturing operation. The system will be required to recognize the coins and sort them, in real time. A traditional MV system has two parts: 1) feature selection with images as the input and 2) classification with features as the input. Based upon experience to date, the proposed AMV system will use a Contingency-based feature selector and a novel Fuzzy Decision Tree-based classifier. Performance will be benchmarked against AlexNet (a non-traditional Deep Neural Net).Project B is on the subject of MV for machine fault detection. Automated assembly machines operate around-the-clock to achieve high production rates. Continuous operation results in high mechanical wear that can led to machine faults. Traditional fault detection methods check for deviations from fixed threshold limits with multiple conventional sensors. The goal of this project is to develop a MV-based detection system to detect known and unknown faults with a single camera. A high speed industrial assembly machine is available for this work. The proposed approach will be based upon the Gaussian Mixture Model (GMM) method for video analysis. The analysis sets out to identify images that deviate from the normal, in other words a “fault”. The images are then analyzed to find out where the difference has occurred. This localization stage will give basic information about the nature of the fault. Thus, the proposed GMM-based system sets out to detect and locate the fault on the machine while leaving the diagnosis to the operator.MATLAB will be used for off-line software prototyping and OpenCV will be used for on-line testing.