Complex Network And Deep Learning-based State Analysis And Identification Research Of Complex Systems

Complex systems widely exist in the real world.Exploring the evolution mechanisms and dynamic behaviors in different system states,and realizing the highprecision system state identification are the current research hotspots in complex system fields.Oil-water two-phase flow and human brain are two typical complex systems.For the former,the dynamic mechanisms of the evolution and development under different flow patterns are still unclear.The measurement of complex flow parameters has not met the application requirements in terms of accuracy and real-time performance.For the latter,the research of intelligent brain-computer interface system based on human brain still has huge research demands in the fields of brain cognitive enhancement and rehabilitation medicine.In this dissertation,these two complex systems are used as the research objects.Taking multi-source data fusion and complex system topological domain analysis as the main direction,we have developed and proposed a variety of novel multilayer complex networks and modular deep learning models,and then have effectively achieved the dynamic analysis and identification of complex system states.The innovative results obtained in this dissertation are as follows:1.In view of the requirements of complex system state analysis from lowdimensional measurements,this dissertation proposes a multilayer modality transitionbased complex network and a multilayer recurrence complex network,which echoes the multi-component,multi-dimensional nature of the complex systems.Multilayer modality transition-based complex network extracts modalities from low-dimensional measurements to characterize system units,and then uses multilayer design to achieve multi-dimensional topological mapping of complex system states.Differently,multilayer recurrence complex network uses phase space vectors as network nodes,and reconstructs complex system dynamics through the recurrence phenomena between nodes.Self-recurrence and cross-recurrence are both been considered.The practical application results show that the former network can describe the formation and evolution of oil slugs,and characterize the intermittent flow behaviors.The latter method can quantitatively indicate the fragmentation mechanism of oil phase from large to small in oil-water bubble flow.2.In view of the requirements of complex system state analysis from highdimensional measurements,this dissertation proposes a frequency-dependent multilayer complex network framework,which echoes the multi-frequency and timevarying characteristics of the majority complex systems.This framework uses channels as network nodes,and uses a multilayer structure to achieve topological mapping of complex system states in all effective frequency bands.The practical application results show that this framework can effectively identify the transient phase distribution differences in the pipeline section during the evolution of oil-water two-phase flow.In addition,after expansion,this framework can also quantitatively explore the brain topological changes during long-time driving,and mine the key brain electrodes of driving fatigue such as FP1,F3,FZ,C4,and CP4.3.In view of the complex system state identification problem based on multisource measurements,this dissertation proposes a spatial-temporal fluid soft-measuring model,a multi-harmonic linkage convolutional neural network model,and a FlashlightNet model.They are respectively applied to three complex system state identification tasks,i.e.,the water cut prediction in oil-water two-phase flow,the multi-category visual evoked potential(VEP)signal classification,and the motor imagery(MI)braincontrolled intention recognition.All three models adopt multi-path and modular design,and configure corresponding feature modules according to the system selfcharacteristics.The practical application results show that they outperform the state-ofthe-art methods in their respective tasks.4.In order to further introduce topological domain information for solving the problem of complex system state identification,this dissertation proposes a topological deep learning framework based on the complex network.This framework first topologically maps the complex system states through multilayer complex network,and then builds a deep learning model to realize the system state identification.This framework is helpful to realize efficient complementarity between complex networks(system topological features)and deep learning(nonlinear feature extraction capabilities).The practical application results show that this framework can achieve the classification accuracies of 95.34% and 91.31% for driving fatigue detection task and multi-category emotion recognition task,respectively.