Research For Soft Sensor Modeling Based On Neural Networks And EAEA Optimization Algorithm

In the process of industrial modeling, the object usually contains complicated interplay of variables. In this case, the mechanism analysis method does not have an ability to describe the change rule of modeling object. Therefore, the artificial neural networks get fast development. During the neural networks modeling, if the number of training data is less than that of parameters, the network is likely to take place over-fitting problem. In order to avoid over-fitting problem, the expert knowledge of monotonicity is used as constraint. If the candidate solution violates the constraint, the training target function will be inflicted punishment. In this case, the candidate model with over-fitting problem will not be outputted.Because of the weakness in neural network training algorithms, an improved AEA algorithm-EAEA is proposed. EAEA is one of the intelligence optimization algorithms. In EAEA, the estimation of distribution algorithm and elite reservation are embedded into AEA to generate a comparison population for iteration needs. Then the optimal parameters are obtained by computing initial population and comparison population. The performance of EAEA is tested on22unconstrained benchmark functions and the testing results show that EAEA has satisfactory search ability and convergence speed.Due the reason that neural network cannot reflect the linear relation between variables, a high-low order mixed modeling method is proposed. It is the combination of multiple linear regression and neural networks. In the course of high-low order mixed modeling, the relationship between measurable variables and unmeasured variables is divided into linear part and nonlinear part. The different parts are modeled by multiple linear regression and neural networks respectively. Then the low order model and the high order model are obtained. The low order model is used to describe the linear relationship between measurable variables and unmeasured variables. The high order model is used to compensate the nonlinear error in low order model. The high-low order mixed modeling method can adequately exert the advantages of multiple linear regression and artificial neural networks, and provides good performances.