Research On An Intelligent Aided System For Chemical Process Safety Operation Based On Kernel And Knowledge

Characteristics of process engineering are a complex flows and colossal system. There are many unsafe factors. Abnormal situations occur from time to time. Therefore, an operation safe is a most important thing for the process engineering. Using enough information technologies, it can assure that the production process is safe, and which it also is an important technology problem for the process engineering. At the present time, with the extending application of DCS (Distributed Computer System) and computer techniques, a great amount of process data which includes process information can be sampled and collected. But, these data are not employed availably, thus it leads abundant data and poor information. In other hand, an amount of empirical knowledge, especially treating abnormal situations, are accumulated. But, these information are not handled better. Therefore, it is research contents in the paper how to transform these data into informations which can be used, and extract feature information which effects on the operation safety of the process from the process data, and further make enough use of these information and combine the empirical knowledge to support the long, steady, and safe operation for the process engineering.Based on the motivation, a conception and frame for an intelligent aided system of process safety operation is presented in the paper. The structure, function, and some key techniques for the system are described in detail. The proposed system is developed and applied to a real industrial process.Firstly, according to the characteristic of chemical engineering, the intelligent aided system for process safety operation is presented based on an intelligent algorithm for identifying an operation state and combined with process models, production objects, and intelligent techniques. A hybrid intelligent method for fault diagnosis is presented based on the discussion and analysis for some key techniques.Secondly, an improved kernel principal component analysis (KPCA) based on multi-block feature vector selection is presented in the paper. A kernel matrix K need be computed in the process of constructing a model for KPCA. But, for a large sample, it is very difficult to solve the kernel matrix K. Using a method of multi-block feature vector selection, a sub-sample is selected to describe the all sample, then a model for KPCA is trained by the sub-sample. The experiment shows that performance of the proposed method and KPCA are almost equivalent. The dimension of kernel matrix K for the proposed method reduces. Consequently, the computing complexity of K drops.Thirdly, a classification method combining a kernel-based feature extraction with a least square support vector machines (LSSVM) for an operation mode of process is presented. A process data are characteristic of high dimension, strongly nonlinearity, high noise and small fault samples. Therefore, it is very difficult to directly identifying the process operation mode from these data. Using the kernel-based feature extraction, the inputting data are decomposed, and removed from redundant or related information, and extracted from process feature information. Then, the classification model based on LSSVM is built up for identifying the process operation mode. The real data from an industrial process is used to prove the proposed method. The experiment result shows that the proposed method is effective.Fourthly, a fuzzy least squares support vector machines classification method incorporated with a prior knowledge on data is presented. To address the drawback which a least squares support vector machines is sensitive to noises or outliers, the least squares support vector machines model combining with the prior knowledge on process data is represented based on a noise distribution model and a strategy based on a sample affinity. Information of noise distribution model for process data is introduced in the training process of the model;The strategy based on the sample affinity is presented to discriminate between data and noises. A fuzzy membership is automatically generated and assigned to each corresponding data point by using the strategy and the noise model. Thus, the fuzzy membership in the LSSVM model can be adjusted. The experiment result shows that the proposed method has better performance and robustness against the noise data.Fifthly, a fault diagnosis expert system for the chemical process is presented. The following problem is considered in the process of the design for the expert system. In the process of knowledge acquisition, a problem which a knowledge acquisition has always been the bottleneck in developing expert systems is solved by using a method combined a fault tree and an empirical knowledge table with a decision tress; A knowledge verification is an import step for the knowledge acquisition. The knowledge verification is carried out based on a directed graph approach in the paper; A selection strategy of knowledge rules in the memory knowledge base is presented; Incorporating process knowledge, a dynamical strategy of knowledge conflict resolution based on a statistics and a time-series analysis method is introduced for resolving a knowledge conflict.Sixthly, based on the previous research and a requirement of real application environment, the intelligent aided system for chemical process safety operation is developed. The structure and the main function for the proposed system are described in detail. The proposed system is applied to a lubricating oil process in a petroleum plant. The validity and practicability of the proposed system is proven. Further, the frame of the proposed system as well as some key techniques used to the system can be confirmed with the availability.