Design Of Interactive Modular Neural Network And Its Application

Modular Neural Network(MNN)is a kind of brain neural network designed from the perspective of simulating the modularization of human brain.It simplifies complex problems through the idea of "divide and conquer".It uses different modules to learn,so that the learned knowledge can be retained and has strong memory ability.The design and research of MNN is of great scientific significance to the development of neural network models that are more consistent with brain-like functional modules,and to the theory of MNN structural design and research.At the same time,the application prospect of MNN in industrial process is expanded.At present,the main limitation of MNN lies in the research on the interaction of function and structure:the different subtasks of MNNs and the different subnetworks are independent of each other.How to build MNNs with modular interaction to get closer to the brain network from the structure and function is the main problem.Aiming at the structural design of modular neural network and its application in practical problems,the main construct of MNN are studied in detail.The main work contents and innovations are as follows:1 Functional task partition construction design of interactive modular neural network(1)In order to simulate the functional modular partition of human brain,a method of task decomposition using adaptive feature clustering algorithm is proposed,which realizes the feature partition of the original task.Compared with the traditional sample partition method,the MNN model established by the feature partition method clusters the samples from the feature dimension,simulating the functional characteristics of human brain network modularization.The model is applied to nonlinear system modeling.The experimental results show that the proposed model has higher prediction accuracy than the MNN based on sample partition.(2)In order to realize the functional interaction of modular neural network,an improved soft subspace clustering algorithm is proposed,which is used in the task decomposition method of MNN to realize the soft partition of the feature dimension of input samples,so that some features exist in multiple subtasks and form information interaction between subtasks.The model is applied to nonlinear system modeling.The results show that the functional interactivity construction of MNN can improve its generalization performance.2 Interactive design of subnetwork structure of modular neural network(1)Focusing on the lack of structural interactivity of subnetworks in current MNN design,a MNN model for local interaction of subnetworks is proposed.This model is based on the lateral inhibition mechanism between neurons,adding an interactive structure of inhibitory connection between excitatory nodes in different subnetworks,so that the information between subnetworks can be exchanged through the connection between modules,and the structural interaction of subnetworks can be realized.Experimental verification is carried out on the benchmark data sets in the UCI database,and the results show that the MNN with inter-module interactive connection can effectively improve the accuracy of the MNN compared with the independent sub-network structure.(2)Aiming at the problem that the fixed structure interactive subnetwork cannot be adjusted according to the actual problem and may lead to structural redundancy,a self-organizing interactive MNN model is proposed.The model first connects the outputs of different modules with the hidden nodes of other subnetworks to form an interactive subnetwork structure,and then through the sliding window,after the data enters the subnetwork,use the contribution degree to dynamically grow and prune the nodes of the subnetwork.It realizes the dynamic adjustment of the internal structure of the subnetwork after the establishment of interaction.The experimental results also prove the effectiveness of the proposed self-organization mechanism,and obtain relatively ideal prediction accuracy.(3)Aiming at the problem that structural optimization for interactive MNN,a multiobjective optimization algorithm is proposed to further optimize the structure of interactive MNN.By encoding the interactive MNN and setting three objective functions:model accuracy,inter-modular structure,and intra-modular structure between subnetworks,the NSGAII algorithm is used to optimize the structure of the interactive MNN.The growth and pruning strategy of gradient and threshold is proposed to guide the model to change towards the optimal evolution direction in the whole evolution process.The experimental results show that the generated interactive MNN has the structure of tight connection within modules and sparse connection between modules,and has higher accuracy.(4).In view of the problem that the number of subnetworks of interactive MNN needs to be specified in advance,and the simultaneous evolution of structure and parameters can not be carried out at the same time,the structure of interactive MNN is further improved,and the method of NeuroEvolution of Augmenting Topologies(NEAT)is proposed to evolve the structure of interactive MNN while adjusting parameters.In addition,the module attribute index is introduced,In this way,the modular attribute is taken into account while the structure is adjusted,and a reasonable interactive MNN structure is generated.The experimental results show that the NEAT algorithm can automatically generate an interactive MNN structure with inter-module connections without setting the number of subnetworks.3 Application of interactive modular neural network in wastewater treatment processMNN has been widely applied in numerous industrial processes,so the designed interactive MNN also is applied to industrial processes for verification.Aiming at the problem that the key effluent parameter BOD is difficult to measure accurately in the process of wastewater treatment,the prediction model of interactive MNN based on functional connection and the prediction model of interactive MNN based on structural connection are established respectively.Among them,the interactive MNN based on functional connection carries out feature soft partition by the auxiliary variables related to effluent BOD,then use the subnetwork with interactive structure to learn all partitions,and finally integrate the effluent BOD results of all subnetworks to get the results.The interactive MNN based on structural connection directly inputs auxiliary variables related to effluent BOD into the network,and generates corresponding prediction models through NEAT algorithm and modular indicators.The experimental results show that the interactive MNN can effectively predict the effluent BOD parameters.