Research On Dynamic Soft Sensing Modeling Of Complex Industrial Processes

In chemical industry,the real-time measurement of quality variables is an important part of closed-loop control.However,due to many factors such as economic reasons and some limitations of industrial environments,some quality variables(we called primary variables)have difficulties in measurement.Soft sensing technique can avoid these inconveniences through real-time estimation of primary variables according to measurable physical quantities(we called second variables),and has been successfully applied in many producing processes.However,current soft sensor modeling researches are generally based on static assumptions,which only focuse on modeling the instantaneous relationship between the primary variables and second variables,but ignore the dynamic features of industrial processes.As a result,large estimation errors may happen when process variables change dramatically.Dynamic soft sensing modeling takes the historical state information of industrial processes into account,which is capable in modeling dynamic system.Compared with the traditional soft sensing model,it has higher accuracy and robustness.Based on Artificial Neural Network(ANN),the dynamic soft sensor modeling has been studied from two aspects,the model parameter optimization and the model structure.The main contents are given as follows:Nonlinear moving average(NMA)models often contain too many model parameters,which may poor the efficiency of training process.In order to ease this problem,an improved parameter optimization strategy is proposed.The main idea of this strategy is to divide the parameter optimization problem into two halves,in this case,the difficulty of optimizing problem could be reduced.To be specific,we first define a new cost function to optimize the parameters contained in hidden layer.Then,the parameters of the output layer are optimized by analytic expression.Not only does this strategy improve the convergence rate of parameter optimization,but also it reduces the randomness in the optimization process.The Finite impulse response(FIR)based soft sensing modeling is studied in this paper.By fully parameterizing the FIR layer,its dynamic feature extracting ability has been boosted.As a result,model we proposed can solve several problems in industrial process,likes dynamic response,measurement noise and time delay.Its accuracy and dynamic feature extracting ability has been verified by data modeling experiments.A new nonlinear autoregressive with exogenous input(NARX)dynamic soft sensor model is built by combining two network models.When the industrial environment cannot provide the history information of primary variables in real time,the real-time prediction of primary variables can be guaranteed by multi-step prediction method.By reducing the autocorrelation of prediction sequence,the cumulative error caused by multi-step estimation is suppressed.At the cost of reducing the accuracy of single-step prediction properly,the model can obtain better prediction effect in the much rough industrial environment.