Generative Adversarial Networks Based Dynamic Correction Of Operational Indices In Process Industries

In the process of operation optimization and control of the whole process of complex industrial production,it is difficult to guarantee the optimization of the whole industrial process through the local optimization of each production process.At present,most of the studies on the optimization control of industrial production remain in the stage of local optimization control of the process,while the research on the global optimization control of the whole mineral processing process is less.Therefore,in order to achieve the global optimization and dynamic correction of the whole production process under complex uncertainties,it is very important to study the decision-making process of industrial process operation indicators.In order to coordinate the operation indicators of each production process and solve the optimization problems related to the comprehensive production indicators such as product quality,output,efficiency and consumption of the whole production process,many scholars and experts have proposed many knowledge-based methods,but these methods do not model the priori knowledge well that exists in the industrial process.So they encounters "knowledge bottleneck" in varying degrees.According to the existing problems,relying on the project of National Natural Science Foundation of China(NSFC)"Closed-loop optimization decision-making method for multi-process process indices in complex industrial processes under dynamic environment"(61273031),this paper studies the optimization decision-making method for the whole process operation index of complex industrial production,and proposes a dynamic correction model based on adversarial learning to realize the global optimization,whole process optimization and dynamic adjustment under dynamic uncertainty.The main work is as follows:1)The mathematical description of dynamic correction in complex industrial process is given,and three objectives of dynamic correction for complex industrial operation indices are proposed to assess this problem;a strategy of dynamic correction for complex industrial process based on Generative Adversarial Networks is proposed;based on variational inference,the derivation of value function of this strategy is given.2)A framework of Decision making Generative adversarial network(DMGAN)is proposed,which includes an encoder,a generator(also called decision-making network)and two discriminators.The two discriminators constrain the latent variable space and the decision space respectively.In DMGAN,the learning and reasoning process is realized by two adversarial learning criteria and three cycle-consistency criteria to achieve bijective mapping between latent codes space and target decision space,and a mapping from target space to decision space.In order to match the increasing complexity of industrial processes,a RU-Net(Reinforced U-Net)is proposed.The networks improves the generalization of traditional U-Net in three ways:a more general conbinator,building block design and Drop-Level regularization.Combinator function and building block design are used to enhance the local representation while Drop-Level regularizaiton training criteria is used to prevent the networks from overfitting;KDA(Knowledge Dissimilarity Assessment)based on Parzen Window probability density estimation and Mutual Information is proposed to evaluate the difference between decision space generated by the generator and actual operational knowledge.In addition,two methods based MSE(Mean Square Error)are used to evaluate the performance of dynamic correction of operational indicices in other ways.The actual production data of beneficiation process are taken as test platform.The simulation and comparison experiments show the effectiveness and versatility of DMGAN.3)A task-driven multi-step dynamic correction model for complex industrial processes,named RAGAN(Recurrent Attention GAN),is proposed.The model includes a DA(Distributed Attention)mechanism and a RAAE(Recurrent Adversarial Auto-encoder)framework.A distributed attention DA mechanism is proposed and sensing networks is built on the basis of this mechanism.Including encoder read-in network,decision-making agent read-in network and condition read-in network,decision-making agent write-in network,and discriminator read-in network and condition read-in network.Read-in network and condition read-in network respectively achieve dynamic selection of input indices and conditional variables to produce corresponding perceptive region.Writing-in network can partly modify the matrix of the CCM(Cumulative Canvas Matrix).Reinforcement learning is used to train the sensing networks based on DA mechanism how to allocate perceptive resources reasonably.LSTM(Long Short-Term Memory)is used to construct an encoder,a decision-making agent and a discriminator,in which decision-making agent updates the intermediate state produced by MC(Memory Cell)in LSTM based on the perceived industrial operational state of th e sensing network.The intermediate state is used to guide the actions of each sensing network in the next time step,including determining the parameters needed by the sensing network based on DA mechanism to ascertain how the sensing network allocates the sensing resources and reconstruction action achieved by connecting a recurrent encoder sequentially and thus resulting in a RAE(Recurrent Auto-encoder)to reconstruct the existing decision space in multi-step way and the interaction action between agent and environment by generating correction values of operational indices based on conditional variables in current time step to update the initial decision value and then inputing to the process model.Reward value is composed loss functions of supervised learning and these of reinforcement learning.A new recurrent GAN framework RAAE is proposed,which constructs GAN model and the autoencoder in a recurrent way,and the recurrent discriminator in GAN model can achieve Nash equilibrium with the current generator to improve the performance of RAAE model.The actual operation data of mineral processing is used as test platform.The simulation and comparison experiments show the effectiveness of the proposed RAGAN.