Data-Driven Adaptive Modeling For Nonlinear Dynamic Processes With Its Applications

Safe,efficient,and sustainable operations and control are primary objectives in industrial production processes.Traditional modeling and control technologies heavily rely on expert experience and mechanism knowledge,thereby showing apparent limitations in increasing complex industry.The burgeoning era of big data is influencing the fourth industrial revolution tremendously,providing unprecedented opportunities to achieve industrial intelligence.To attain this goal,data analytics and machine learning are indispensable.In the case that the mechanism model is difficult to build,effective use of off-line/online data to realize process prediction,modeling and control becomes an important way to achieve industrial intelligence in the era of big data.Nonlinear dynamic industrial processes have characteristics of complex mechanism model,high orders and strong nonlinearities,which makes it difficult to design real-time effective controller for control purpose.While such industrial processes can produce and store a large amount of process data and present them in the form of time series.Therefore,model constructed upon historical data that enables representing the underlying dynamics can be used for describing nonlinear dynamic processes.However,for practical process industry,factors such as working conditions changes,operation mode switching as well as external disturbance can lead to process drift.This makes process information in the form of fast data streams exhibit nonstationary characteristic.Off-line model constructed upon historical data cannot capture the changing process dynamics.Therefore,it is necessary to provide model with online adaptation capacity.The online model should not only have high adaptability to track the fast-changing process dynamics,but also maintain low computational complexity to meet real-time restrictions,so as to design real-time controller based on data model to achieve effective control of nonlinear dynamic industrial processes.This dissertation focuses on modeling problem in industrial nonlinear dynamic processes.Based on data-driven modeling methods,off-line and online modeling approaches are investigated according to the dynamic data characteristics,which ensures the data-driven model can adapt properly to the nonlinear dynamic process.Taking the tunnel microwave heating process as a practical application,data-driven model of the microwave heating process is constructed and optimal controller is designed to improve the overall operational effectiveness.The main work and contributions of this dissertation include the following aspects:1)Focusing on off-line data from dynamic process,the adaptive construction and optimization of recurrent neural network(RNN)is investigated.In order to solve the problem that the structure of RNN is difficult to determine,an adaptive construction method is proposed to gradually increase the complexity of network structure until the model can represent the nonlinear dynamic relationship between the process input and output.The cuckoo search algorithm and orthogonal least squares(OLS)are adapted to ensure the parameter optimality of recurrent neurons and orthogonal relationship of hidden neurons.The proposed adaptive RNN is applied to the modeling and temperature prediction of tunnel microwave heating process,and it can attain higher temperature prediction accuracy as well as compact network structure.2)Multiple local linear models learning based online modeling approach is investigated for process drifting data streams.A multiple local model based selective ensemble regression(SER)algorithm is first proposed.This algorithm can automatically identify newly emerged process state and fit by a local model with growing strategy.In the meantime,it selects the most up-to-data local models from the local model set to make online prediction.To reduce heavy computational burden raised by local model growing,a growing and pruning SER algorithm is further proposed.A probability based matric is adapted to measure the local model performance.Additionally,local model pruning strategy is employed to effectively remove unwanted local models,which significantly reduces online computational complexity.This algorithm can well mimic biological system,that is,update the memory by learning new knowledge and removing out-of-date knowledge so that intelligent decision can be made based on the relevant knowledge.3)In order to construct fixed-size adaptive model,global nonlinear model learning based online modeling approach is investigated for process drifting data streams.An extended gradient radial basis function(GRBF)network is first proposed for process modeling and identification.The OLS is adopted to construct a compact network structure on training dataset,which makes it well capture characteristic of nonstationary time series with changing local mean and trend.To improve the GRBF model adaptive ability in time-varying environments,an online adaptation strategy is further proposed to adjust model parameters and structure based on current modeling error.The strategy contains weight adaptation mode and node adaptation mode.The GRBF network updates weights by the recursive least square.If the current network structure cannot capture the changing process dynamics,an insignificant node is replaced by a new node.By exploiting the property of GRBF node,the new node optimization is very efficient and can automatically encode the new data state.4)Based on the developing RNN model,data-driven optimal control algorithm is investigated and applied for temperature tracking control in tunnel microwave heating process.The temperature tracking problem is first converted to an error regulation problem with stability analysis.A recurrent neuro-controller is developed based on the trained system model to acquire the desired control laws.Based on RNN,adaptive dynamic programming(ADP)is introduced to construct the optimal error regulator,which is combined with desired controller.By online iterative updating critic network and action network in ADP,the optimal cost function and control law can be gradually obtained.This method effectively integrates the nonlinear dynamic mapping ability of RNN into the design of ADP-based optimal tracking controller.In the application of microwave heating process,the control law produced by action network can automatically adjust multiple microwave powers and conveyor speed,so that the heating material can well track the desired temperature.