| The “Made in China 2025†plan has put forward the primary goal for China’s intelligent manufacturing in the decade to come, which is, on the premise of emancipating productive forces, a significant measure to promote the productivity as well as production quality. Along with the implementation of the “Made in China 2025†plan, the traditional industrial automation equipments will be inevitably upgraded massively, and the integration of machine vision with the automation control is considered as one of the important implementations. The parameters of manipulated variables acquired in the traditional automatic control systems are those indirectly related with product quality, e.g. pressure sensor, liquid level sensor, flow sensor, and temperature sensor. Visualization of the parameters, on the other hand, which are directly related with product quality have been neglected completely. A large amount of information is contained in images, and nearly 80% of human beings’ information is obtained through vision, as image information is massive. Machine vision technologies will dramatically improve the information acquisition capacity of industrial automation. It also can introduce visual information related to product quality to industrial automation, and further make the information obtained by the industrial automatic systems not just direct and straight, but also massive and multi-dimensional. At present, although machine vision technologies have been applied to online inspection, but no intelligently supervised strategy based on machine vision is formed, which does not help realize effective online control on product quality. Therefore, a study on how to realize the integration of a machine vision system with an industrial control system to deploy the control system with computer supervision has a fundamental significance in promoting the improvement and stability of product quality. It also supports the development and innovation of machine vision technology applications, whilst improving China’s intelligent manufacturing.This thesis focuses on the study of machine vision supervised control. While the control problem of large time delay systems with uncertain parameters such as hoop standard granulator is solved, the machine vision technology will be applied to inspect product quality(such as the superficial flaws and hardness of particles) online. Thus, the segmentation and detection of particles under a complex environment will be achieved, and the machine vision supervised control of hoop standard granulator based on the inspection results will be realized. The main novel contributions of the research can be summarized as follows:Firstly, in order to overcome the large time delay problem of a granulator’s temperature, a new Fuzzy Analytic Hierarchy Process(FAHP) and swarm intelligence based feature weight SVM scheme has been presented. The fuzzy analytic hierarchy process is used to evaluate the weights of input features. The aim of weight evaluation is to get better classification results when kernel functions can be mapped to a high dimensional space. Moreover, a new bacterial foraging algorithm based on Gaussian distributions is used to solve the partial convergence problem in the standard SVM training model.Secondly, for the large time delay systems such as a hoop standard granulator with uncertain parameters, three strategies have been proposed, namely, a model-free based expert control strategy, a fusion model based golden section method and a fuzzy expert control approach. Evaluations have been conducted, and the results show that SVM based fuzzy expert control strategy is better than the others.Thirdly, in order to inspect the particles’ quality, a Meanshift filter of using the standard Memory Gradient Algorithm(MGA) is proposed to solve the segmentation problem of particles. In this way, noise is reduced, and the damage to the image is reduced as well under a complex environment(such as changing illumination and severe noise). Moreover, the algorithm’s convergence is also derived and analyzed theoretically. On the other hand, considering the particle attachment, a watershed segmentation algorithm using entropy is proposed, which can reduce the iteration number, shorten the computation time of segmentation, and optimize the watershed segmentation steps to avoid over-segmentation. In addition, in response to the issue of the efficiency problem in random selection of particles, an adaptive segmentation decision strategy is presented to choose the best segmentation methods automatically.Fourthly, considering the particle classification problem, a rotation invariant raw statistics feature is proposed. The combination of Patch features and MR8 filter features will enable the new RIRS feature to be rotation invariant. Meanwhile, a sparse presentation based on texton learning algorithm is introduced to make up the shortcoming of K-means with the low differentiation of subclass. Then, a sparse random projection method is used to build histogram features with a 2l-0l architecture and the histogram features are classified using a random forest classifier. The classification approach is the key and the foundation of the supervised control based on machine vision.Finally, we have designed a supervised control system based on machine vision. By means of establishing both an intelligent control platform and a machine vision hardware platform for the granulator, the supervised control test based on machine vision is conducted in experiments. The experimental results suggest that the machine vision based supervisory system which is designed in this work is capable of effectively detecting product flaws, and making accurate intervention by means of optimizing control objectives effectively and giving proper alarms.
|
PDF Full Text Request