Structural Design Of Pruning Smallworld Neural Network And Its Application In Soft Measurement Of Wastewater Treatment Process

In recent years,as a mathematical model inspired by biological neural network（BNN）,artificial neural network（ANN）has become a research hotspot in the field of artificial intelligence.Compared with ANN,the topology of BNN lies somewhere between regular and random topologies and exhibits with a small-world property.It has been proved that the small-world neural network（SWNN）constructed by combining ANNs and small-world property can approach to the topology of BNN.Furthermore,it has received considerable attention of domestic and overseas scholars to design SWNN.However,the architecture of SWNN is too complex.An excessively large neural network may degenerate the generalization performance caused by an overfitting problem.Therefore,it is crucial to explore how to prune the structure of SWNN to further improve its generalization ability.The main research work of this thesis is mainly divided into the following points:（1）Investigation of small-world property in SWNN.Small-world property is the most important property of SWNN.This thesis investigates the influence of the number of rewiring connections on small-world property in the construction of SWNN to find the best small-world attribute.Firstly,an SWNN is constructed based on the rewiring rule of Watts and Strogatz.The effect of the number of rewiring connections on the small-world property in SWNN is investigated.The simulation results show that the network has better small-world property when the number of reconnected connections is set to about half of the maximal number of rewiring connections,which proves the effectiveness of the construction method for SWNN.It lays a foundation for the follow-up study of structural pruning for SWNN.（2）Design of a pruning feedforward small-world neural network based on Katz centrality.Aiming at the problems that the structure of SWNN is too large and there are redundant nodes in SWNN,this thesis proposes a pruning small-world neural network based on Katz centrality（PSWNN-Katz）.Firstly,the importance of hidden neurons is evaluated based on the Katz centrality.Then a more compact network structure is obtained by combining unimportant hidden neurons with their most correlated neurons.The convergence of PSWNN-Katz in structure fixed phase and structure pruning phase is theoretically analyzed.Finally,experimental results show that the PSWNN-Katz has better generalization ability owning to a compact structure and good small-world property.（3）Design of a pruning feedforward small-world neural network based on a dynamic regularization method with the smoothing l_1/2 norm.Aiming at the problem that the proposed PSWNN-Katz may lead to pruning some important weights,from the perspective of weight pruning,this thesis proposes a pruning small-world neural network based on a dynamic regularization method with the smoothing l_1/2 norm(PSWNN-DSRL_1/2).By minimizing the modified error function added with a smoothing l_1/2 norm,redundant weights are pruned to generate a sparse architecture.Aiming at the problem that the regularization parameter is fixed during the training process,a dynamic adjusting strategy is further designed to adaptively adjust regularization parameter.Several experiments are carried out to evaluate the performance of the proposed PSWNN-DSRL_1/2,validating the effectiveness of the proposed model on nonlinear system modeling.The results indicate the generalization ability of PSWNN-DSRL_1/2 is further improved with a sparse architecture.（4）Design of an effluent BOD soft measurement model based on PSWNN-Katz and PSWNN-DSRL_1/2.To solve the problem that it is difficult to predict the effluent biochemical oxygen demand concentration（BOD）in the wastewater treatment process accurately,this thesis established a soft measurement model based on the proposed PSWNN-Katz and PSWNN-DSRL_1/2.Firstly,a method based on mutual information is employed to select feature variables.Then a soft measurement model based on pruning feedforward small-world neural network is established.In order to verify the effectiveness of the soft measurement model,it is applied to the effluent BOD prediction in actual wastewater treatment.The results show that the soft measurement model based on PSWNN-DSRL_1/2has better generalization ability than other soft measurement models and can predict the effluent BOD accurately.（5）Design of the intelligent software for effluent BOD prediction in wastewater treatment process.To solve the problem that the traditional measurement for effluent BOD takes a long time,and is easily interfered by the environment.An online intelligent software for effluent BOD prediction based on the MATLAB GUI is designed.After analysing the software requirements,there are some modules including user registration and login module,model introduction module,data processing module,BOD prediction module and help module.The software has the advantage of low cost and it can predict effluent BOD rapidly and visualize the prediction results.The software has practical significance for ensuring stable and efficient operation of wastewater treatment plants.