Research On Soft Sensing Of Marine Enzyme Fermentation Process Based On Gaussian Process Regression

In industrial processes,the parameters of some key variables are difficult to be detected on-line,which brings great restrictions in control and optimization of the industrial production.The emergence of soft sensing technology is one of the effective ways to solve this problem.At present,common tools to build a complex nonlinear soft sensing model include neural network,least squares,radial basis function network and support vector machine.However,it can't give the accuracy parameter and uncertainty interval of soft sensing results,which restricts its application in actual production.At the same time,marine enzyme fermentation is a nonlinear biochemical reaction process with complex internal mechanism.If the soft sensing model is built directly,it will not only increase the model complexity and introduce a lot of noise,but also the reliability of the soft sensing results is difficult to guarantee.Besides,Gaussian process regression(GPR)soft sensing model,which can obtain the uncertainty information,also has the problem of low optimization efficiency and poor stability.In view of the above problems,multiple optimization algorithms are researched with the GPR soft sensing model,and the process data from a real industrial marine enzyme fermentation is adopted to verify and evaluate its rationality.The main work of this thesis is as follows:(1)To solve the difficulty that some traditional soft sensing models can't provide the confidence interval,a GPR soft sensing model is established.Compared with the traditional modeling method,the proposed method does not only realize the soft sensing of the activity of marine enzyme,but also give the uncertainty interval of the soft sensing results,and it can reasonably reflect the abnormal input.(2)There are many redundant auxiliary variables in the fermentation process of marine enzymes,which lead to difficulties in modeling and low accuracy in soft sensing result.Aiming at to this problem,a neural network(NN)based on mean impact value(MIV)variable selection method(NN-MIV)is proposed.The NN-MIV method comprehensively considers the internal contribution rate and the external contribution rate of the auxiliary variable to the dominant variable,and combines the two contribution rate indexes as the basis for screening auxiliary variables to obtaining the optimal auxiliary variable,which is used to optimize the GPR soft sensing model.This method has been applied in a real marine enzyme fermentation process and the results show that it reduces the calculation,simplifies the model and achieves better performance.(3)Due to many important variables may be discarded in the process of NN-MIV method,resulting in serious information loss,a new variable selection method based on NNMIV-PCA is proposed.Firstly,the NNMIV method investigates the characteristic weight of the auxiliary variable to the dominant variable.By setting one screening threshold dynamically,several variables with larger feature weight are obtained.Secondly,the remaining variables with smaller feature weights are selected by the principal component analysis(PCA)method.Some principal components with the largest contribution rate are obtained by setting another screening threshold.Lastly,the final auxiliary variables are determined according to the above results.Experimental results show that the NNMIV-PCA method ensures high contribution rate while the information utilization rate of variables is also improved.(4)Aiming to the problem that GPR soft sensing model has low optimization efficiency and poor stability during the solution,a GPR soft sensing model based on gravitational search algorithm(GSA-GPR)is proposed.The gravitational search algorithm(GSA),based on Newton's law of kinematics,assumes that each particle is attracted to others by gravitation.The particle with larger fitness value has bigger inertia mass and all particles will move towards the mass with largest inertia.By setting the optimization parameter to approximate the optimal solution of the problem,GSA algorithm shows better optimization efficiency and avoids the problem of falling into local extremum.The experimental results show that GSA-GPR soft sensing model has preferable generalization ability and stability by the subgroup data.