Study On Soft-Sensing Technology's Application Base On ANN And LSSVM

Advanced control technology has been widely used in China petrol companiessince the 1990s, because it enhances the market competition ability of factory ofprocess industry. The control of product quality is the core of all kinds of control, sothe information of or about product quality of important process variables must be gotin time to have good control of product quality. But the online sensors to get theinformation are expensive and difficult to be maintained, and the sensors' delay ofanalysis will induce to the performance decrement of control system. To solve thisproblem, the technology of soft sensing has been researched deeply in these years andbecome one of the most important technologies in advanced control theory. The basicidea of soft sensing is selecting some variables easy to be measured (secondaryvariables) and set up the mathematical function (soft sensor) of them to the importantvariables (primary variables) that can't or difficult to be measured, then the primaryvariables can be inferred via the soft sensor. Soft sensing technique can give theprimary variable's information quickly and continuously, it's also cheap and easy to bemaintained, so it can enhance the product quality while decrease the manufacturingcost. So it's thought as one of the important problems needed to be researched deeplyin control field in the future.The complexity, nonlinearity, time variety and the on-time requirement of a petrolfactory in practice make it difficult to set up a nonlinear soft sensing model based onstructural risk minimization rule, and to modify the mode online. So this paper madeuse of artificial neural network and least squares support vector machine (LSSVM)based on the practical data of a factory to do the following research in the following.1. The theory and training algorithms of back propagation neural network's (BP)and RBF network were introduced in this paper with its advancements and backdraws. BP and RBF soft sensor for end point of gasoline of atmosphericdistillation column was established. The calculation results show thatprediction precision of the soft sensor can satisfy the requirement of processcontrol.2.The theory and training algorithms of LSSVM were introduced in this paperwith G-LSSVM algorithm, which can choose regression LSSVM Lhyperparameters. A regression benchmarking problem was used to check theeffectivity of G-LSSVM. At last G-LSSVM was used to set up the soft sensorfor the end point of gasoline, the result showsthat the precision of G-LSSVMmodel can satisfy the requirement and avoid the difficulty of selecting properstructure and the result of local minimum.3.In this paper, the increment online training algorithm was introduced, thenaccording to matrix computation theory, a new non-increment online trainingalgorithm was proposed based on the modification of kernel matrix inverse.The proposed online algorithm was applied to setup a soft sensor for end pointof gasoline too. The result shows that the model can be trained on time and theprediction precision is good.