Multimode Process Monitoring Method Based On Local Information Preservation

With the development of computer technology and the improvement of automation production equipment, modern industrial system continues to expand the scale and complexity. Therefore, great importance is attached to the safety of production and reliability to reduce or avoid adverse effects. However, the complexity of the industry processes makes traditional monitoring methods face challenge. The widespread use of the filedbus technology and the distributed control systems(DCS) makes saving huge amount of processes’s data possible. Therefore, as an effective method to guarantee industrial process operation well and improve the quality, process monitoring technology has received the attention of many scholars and research.However, most of exsited fault detection ways often have some assumptions of processes and data during modeling period, such as unimodal processes and singal Gaussian distribution. The existence of these assumptions makes monitoring methods can play their best performance only under the specific production conditions. On the other hand, the majority of the methods focus more on the global information of the process data, while they do not give enough attentions to the local information of the data, so these methods are difficult to achieve a good monitoring when the data distribution is complex. To deal with the complex mutilmodal problems and data distribution problems of actual industrial processes, this article firstly studys the research status in this field, and then analysis data information in-depth, and finally puts forward some novally efficient process monitoring strategies, which do not depend on the data distributions. The main contributions of this dissertation are summarized as follows:(1) To handle the mutimodal problem of processes, a new technique based on local density estimation is developed. For the issue of the practical data does not meet the data distribution assumption for the traditional multivariate statistical methods, this method is firstly use manifold learning algorithm which has no single gaussian distribution assumption to map process high-dimensional data into low-dimensional subspace, and then apply the improved kernel density estimation to construct efficient monitoring statistics in the subspace. Due to the strategy adopts the variable bandwidth kernel density estimation, it has good robustness for the data distribution. (2) For the complex distribution of the process data from different modal and from the same modal, in order to overcome the inaccurate problem of description of the fault detection methods based on density of process data using the traditional Euclidean distance, puts forward a new strategy based on local weighted. This strategy study the data’s local information, and distinguishes the effects of different sample points to the current sample by using different weights. This method can describe the local information of the data more accurately and achieve better process fault detection.After the analysis and study the theoretical basis of the above methods, the typical numerical examples, non-isothermal continuous stirred tank reactor (CSTR) and famous Tennessee Eastman process are applied to validate the efficiency of proposed strategies. Finally, after summarizing the main research work of the thesis, some future research directions are discussed.