The Research And Application Of Neural Network Soft-sensing Algorithm Based On Adaptive Nonnegative Garrote

At present,as industrial processes become increasingly complex,some critical parameters in the process are difficult to measure with sensors directly or are costly.Soft-sensing technology provides an effective method.The technology uses easy-to-measure variables to infer difficult-to-measure parameters and has been widely used in process control.It overcomes the shortcomings of insufficient real-time performance and high cost caused by the direct use of sensor measurement.Based on the existing algorithm,this paper analyzes the problems of the current variable selection algorithm in detail.In response to these problems,the article innovatively designs a kind of soft-sensing algorithm that can adaptively select the input variables according to the relative importance of the input variables.Firstly,because of the biased estimation problem of the nonnegative garrote（NNG）algorithm for input variable compression,an adaptive variable selection algorithm based on NNG（Adaptive-NNG）is proposed.The algorithm adds the reciprocal of the absolute value of least squares regression estimation to the constraints of NNG and realizes the purpose of adaptively compressing the coefficients of input variables.Cross-validation and Bayesian information criterion（BIC）are used to determine the optimal garrote parameter of the proposed algorithm.The effectiveness of the algorithm is tested through two linear regression numerical simulation cases.Secondly,the researched Adaptive-NNG algorithm is extended to the field of nonlinear regression.An adaptive variable selection algorithm based on multi-layer perceptron（MLP）is proposed（Adaptive-NNG-MLP,ANNG-MLP）and used for complex nonlinear process modeling.The mean impact value（MIV）is designed as an adaptive operator and introduced into the constraints of NNG-MLP,which overcomes the shortcomings of the NNG-MLP algorithm’s excessive dependence on garrote parameter selection and biased estimation of variable compression.Improve the accuracy of variable selection,thereby improving the generalization performance of the model.Numerical simulation experiments show that our proposed algorithm is superior to other algorithms in model accuracy and model complexity.Finally,the ANNG-MLP algorithm is applied to the actual flue gas desulfurization system to predict the SO₂ concentration of the flue gas emitted by the system.The simulation results show that the proposed algorithm can accurately predict the dynamic changes of the target variables.The selected vital variables are also consistent with the actual chemical process reaction mechanism and consistent with the production experience of the on-site operators.Therefore,the ANNG-MLP algorithm has high reliability in predicting and modeling the SO₂ content of flue gas emissions from the flue gas desulfurization industrial process and can provide technical support for the process control system design and production process improvement.