Study On Genetic Programming And Its Application On Data Based Soft Sensors

Genetic programming(GP),as a special case of genetic algorithm and an subclass of evolution computation,is a new and important technology in the field of computational intelligence.It simulates inheritance and evolution in the nature and gets optimal solutions by using reproduction,crossover and mutation operations.Genetic programming,which combines genetic algorithms and computer programs,can learn solutions of potentially unbound complexity,so it gets more and more interest and has been widely used in symbolic regression, function synthesis,data mining,data analysis,pattern recognition,biochemistry,and etc..As artificial neural network(ANN) and Support Vector Machine(SVM) which are popular in soft sensor,Genetic programming can model a system without any priori knowledge,solve derivative-free optimization problem,has a good non-linear mapping capability and self-learning ability.In addition to these advantages,it also has the following particular characters which are of special importance to industry.These unique features are few design parameters in GP,natural selection of the most important process inputs and the lag time of these inputs,its results are parsimonious analytical functions and several candidate solutions after a run.An additional advantage is the low implementation cost of such type of soft sensors. It can be applied directly into the existing Distributed Control Systems(DCS) avoiding additional specialized software packages which is typical for neural net-based soft sensors.At the same time there are challenges in implementing industrial soft sensors generated by genetic programming,including function generation with noisy industrial data,dealing with large input variables and the slow speed of model development due to the inherent high computational requirements of this method.In order to make the advantages of its unique characters in the soft-sensor modeling,we studies how to overcome these shortcomings.In detail,the major contributions of this thesiss are summarized as following:(1) Basic concepts,technologies,theories of genetic programming are reviewd.The development,research situation and applications of GP are summarized and the focus is the research situation of GP in soft sensor modeling.The advantages and shortcomings of different modeling methods in soft sensor thechniques are analyzed and the unique characters of GP is introduced which makes GP is of special interest to soft sensor.(2) To prevent good solutions to be eliminated because of unfit parameters in them, defferential evolution is used to find the best fit data for these parameters.An immunity defferential evolution,which intergrates immune idea,is proposed to make excellent schemata proliferate and repair these schemata destroyed by mutation and crossover operation,as a result the convergent speed is increased greatly.In order to improve the efficiency and correctness of vaccine for the parameters optimization problems,a new form of vaccine,a new autonomous obtaining method,and the method of deciding the probability of vaccination are proposed too.(3) The immune idea is integrated in GP too.Here the vaccine help to choose the location of crossover points.In this way,the probability of the good schema destructed by crossover is reduced and the learning speed increases.(4) The overfitting phenomenon in genetic programming(GP) is analyzed and a new method is proposed to improve the ability of generalization.First,a new protected approach based on interpolation for some mathematical functions is proposed.Second,a multi-objective GP is proposed to make the evaluation of these goals impersonally and maintenance the diversity of population.Finally,In order to further improve the efficiency and guid the search direction,the preference information,which is introduced according to the user's preference in the different performance criteria,is joined(5) A novel robust genetic programming based on M-estimator is proposed to enhance the ability of dealing with outliers.Moreover,these cut-off parameters in the estimator play a crucial role in degrading the effects of outliers.Usually an optimal value of the cut-off parameter exists,but without a priori knowledge of the training data,it is difficult to define.So a segregated GP using two different cut-off parameters is proposed to solve this problem.The novel feature of this approach is that the algorithm can perform multi-directional search on the whole problem space so it can get mixed information from different directional search and has more chance to find an acceptable solution.In addition,the proposed approach is less sensitive to the value of the cut-off parameters and performs almost as well as GP with an ideal cut-off parameter.(6) the methods,by which GP is applied to soft sensor are studied and a hybrid modeling method which is based on GP and PLS is proposed for offline and online estimation.A brief review of this thesis is given and Some future research directions are highlighted.