Dynamic Network Model Research Articles

Monitoring and analysis of the temperature field of a cold-region highway tunnel considering the traffic-induced thermal effect

In this paper, the tunnel temperature (air, structure, and rock) and wall heat flow were monitored in-situ for 63 days in cold and warm terms. The in-situ monitoring results reveal that the traffic-induced thermal effect (TTE) is one of the main factors inducing tunnel temperature variation. The tunnel wall temperature change rate caused by traffic-induced thermal effect is 20.53–43.18% in the cold term. A simplified analytical solution for the tunnel wall temperature is proposed based on Green's function and verified by the monitoring data in this study. The influence coefficients of traffic flow formed by different vehicle types on the tunnel wall temperature in the cold and warm terms are obtained using an optimized machine learning method. The impact ratio is LDV:HDV: HDT = 6:1:1 in the cold (that of LDT is minimal); LDV:HDV:LDT: HDT = 1:8:4:10 in the warm. The assumed tunnel heat absorption ratio (THAR) ranges from 0.19 to 2.05% in the cold term and −0.22 to 0.52% in the warm. The thermal loss degree in tunnel lining is lower than that of ignoring the traffic-induced thermal effect in the cold term. TTE increases THAR by about 1.45% and partly delays the heat dissipation of the tunnel wall. Furthermore, a dynamic network model with vehicles as moving heat sources is established to investigate the thermal effect of different vehicle types. The study results provide reference to new ideas and calculation solutions for designing cold-region highway tunnels and implementing new energy utilization systems.

Dynamic interactive multiview memory network for emotion recognition in conversation

When available, multimodal data is key for enhanced emotion recognition in conversation. Text, audio, and video in dialogues can facilitate and complement each other in analyzing speakers’ emotions. However, it is very challenging to effectively fuse multimodal features to understand the detailed contextual information in conversations. In this work, we focus on dynamic interactions during the information fusion process and propose a Dynamic Interactive Multiview Memory Network (DIMMN) model to integrate interaction information for recognizing emotions. Specifically, the information fusion within DIMMN is through multiple perspectives (combining different modalities). We designed multiview layers in attention networks to enable the model to mine the crossmodal dynamic dependencies between different groups in the process of dynamic modal interaction. In order to learn the long-term dependency information, temporal convolutional networks are introduced to synthesize contextual information of a single person. Then, the gated recurrent units and memory networks are used to model the global session to detect contextual dependencies for multi-round, multi-speaker interactive emotion information. Experimental results on IEMOCAP and MELD demonstrate that DIMMN achieves better and comparable performance to the state-of-the-art methods, with an accuracy of 64.7% and 60.6%, respectively.

Route to chaos in a branching model of neural network dynamics

Simplified models are a necessary steppingstone for understanding collective neural network dynamics, in particular the transitions between different kinds of behavior, whose universality can be captured by such models, without prejudice. One such model, the cortical branching model (CBM), has previously been used to characterize part of the universal behavior of neural network dynamics and also led to the discovery of a second, chaotic transition which has not yet been fully characterized. Here, we study the properties of this chaotic transition, that occurs in the mean-field approximation to the kin=1 CBM by focusing on the constraints the model imposes on initial conditions, parameters, and the imprint thereof on the Lyapunov spectrum. Although the model seems similar to the Hénon map, we find that the Hénon map cannot be recovered using orthogonal transformations to decouple the dynamics. Fundamental differences between the two, namely that the CBM is defined on a compact space and features a non-constant Jacobian, indicate that the CBM maps, more generally, represent a class of generalized Hénon maps which has yet to be fully understood.

Spatiotemporal dynamic network for regional maritime vessel flow prediction amid COVID-19

The COVID-19 pandemic has stifled international trade and the global maritime industry. Its impact on the routing of the regional vessel traffic flow provides supportive data to port authorities, ship owners, shippers, and consignees. This study proposes a spatiotemporal dynamic graph neural network (STDGNN) model that includes the usual primary part of the vessel flow and an auxiliary part of newly confirmed COVID-19 cases near the port. The primary part consists of a time-embedding (TE) block, two dynamic graph neural network (DGNN) blocks, and a gated recurrent unit block, to capture the spatiotemporal dependence in the regional vessel traffic flow. The auxiliary part is made of multiple blocks to exploit the dynamic temporal relationships in hours, days, and weeks. Moreover, the performance of the STDGNN model is verified by utilising real vessel traffic flow data (i.e. inflow, outflow, and volume) and the new cases of COVID-19 near the port of New York, USA, provided by the automatic identification system and the Johns Hopkins University Centre for Systems Science and Engineering. The 2-h prediction result shows a 37.7%, 17.23%, and 11.4% improvement in the mean absolute error (MAE) over the gated recurrent unit (GRU), STGCN, and TGCN models, respectively. The delicate and adaptable prediction of vessel traffic flow could help the port relieve congestion, enhance efficiency, and further assist the recovery of regional maritime industries in the post-COVID era.

Artificial Neural Network Travel Time Prediction Model for Buses Using Only GPS Data

Real-time and accurate travel time information of transit vehicles is valuable as it allows passengers to plan their trips to minimize waiting times. The objective of this research was to develop a dynamic artificial neural network (ANN) model that can provide accurate prediction of bus travel times to give real-time information at a given downstream bus stop using only global positioning system (GPS) data. The ANN model is trained off-line but can be used to provide real-time travel time information. To achieve this, care was taken in selecting a unique set of input-output combinations for prediction. The results obtained from the case study are promising to implement an Advanced Public Transportation System (APTS). The performance of the proposed ANN model was compared with a historical average model under two criteria: prediction accuracy and robustness. It was shown that the ANN outperformed the average approach in both aspects.

A Dynamic Prediction Neural Network Model of Cross-Border e-Commerce Sales for Virtual Community Knowledge Sharing.

In this paper, a neural network algorithm is used to conduct in-depth research and analysis on the sales dynamics prediction of virtual community knowledge sharing in cross-border e-commerce. Both the expected returns and the social network structure are analyzed, and both have positive effects on knowledge sharing in the actual development process, but the degree of them also possesses certain variability. A model of the factors influencing the quality of knowledge shared by users is constructed to explore the relationship between the dimensions in social capital and how they affect the community users' perceptions of knowledge quality. Exploring the strong influencing factors of product repurchase rate has key implications for promoting product sales and sales forecasting. The scale of this paper has undergone several minor revisions, and the content validity is very good. Criterion validity is generally reflected by Person correlation coefficient, and construct validity includes exploratory factor analysis and confirmatory factor analysis. First, the values were clustered, and the optimal variables were selected using stepwise regression and fitted with Poisson regression models to explore the relationship between the repurchase rate of different products and the factors that strongly influence the repurchase rate in the case. To predict the sales of goods, the advantages of the BP neural network, LSTM neural network, and Verhulst gray model are combined: BP neural network can predict sales by combining the current data corresponding to independent variables, LSTM neural network can explore the influence of historical data on sales, and Verhulst model can predict sales based on the growth trend of variables. A BP-LSTM-Verhulst nested neural network based on the AP algorithm is constructed to predict sales volume, and the accuracy of this method is proved by example. Finally, it is found that the proposed sales prediction method has higher accuracy than exponential regression and shallow neural networks. The deep learning prediction method combining unstructured data such as images proposed in the paper not only provides a more accurate sales prediction method for short life cycle products in e-commerce but also provides an effective deep learning method for management practices. The KMO value obtained after the test is less than 0.5, which is not suitable for factor analysis. Using SPSS22.0 to expand the KMO value and Bartlett's sphericity test, the KMO value is significantly greater than the minimum requirement of 0.5, the KMO value of self-efficacy is 0.826, and the KMO value of expected return is 0.870.

Reliability optimization of flexible test system based on pyro-mechanical device products production driven

The quality evaluation of aerospace pyro-mechanical device products(PDPs) is very important to the safety of aerospace engineering. Aerospace PDPs are small-lot batch, multiple batches and one-off products which rely on the analysis of ignition test with limited samples. To accommodate multiple batches and small-lot batch of aerospace PDPs, the flexible test system(FTS) switches between different batches of products by the selection of instruments and channels. The conventional test mode is driven by the completion time of batch products, which usually leads to frequent switching of instruments and channels and affects the reliability of test system. Therefore, the FTS is configured according to the completion of production missions within a certain time period in the future. This paper considers the FTS of aerospace PDPs as a phased mission system, and designs the corresponding dynamic Bayesian network(DBN) model. A test of one batch is taken as a phase of phased mission system. In order to improve efficiency and optimize reliability of the aerospace PDPs FTS, the receding horizon optimization problem is constructed. Numerical experiments show that the FTS with phased mission model can optimize the test sequence during the test process, which improves efficiency and reliability of test system.

Research on visual question answering based on dynamic memory network model of multiple attention mechanisms

Since the existing visual question answering model lacks long-term memory modules for answering complex questions, it is easy to cause the loss of effective information. In order to further improve the accuracy of the visual question answering model, this paper applies the multiple attention mechanism combining channel attention and spatial attention to memory networks for the first time and proposes a dynamic memory network model (DMN-MA) based on the multiple attention mechanism. The model uses the multiple attention mechanism in the situational memory module to obtain the most relevant visual vectors for answering questions based on continuous memory updating, storage and iterative inference of the questions, and effectively uses contextual information for answer inference. The experimental results show that the accuracy of the model in this paper reaches 64.57% and 67.18% on the large-scale public datasets COCO-QA and VQA2.0, respectively.

Dynamic state feedback controller and observer design for dynamic artificial neural network models

Artificial neural networks are black-box models that can be used to model nonlinear dynamical systems. This article presents a synthesis method for full dynamic state feedback controllers and state and output observers that have guaranteed properties for systems approximated by dynamic artificial neural networks. The resulting control designs are applicable to the practical situation in which the steady-state values for the control input are not known. Dynamic artificial neural networks are written in the standard nonlinear operator form, also known in the literature as the Luré formulation. A generalized form of the Luré formulation is adopted to allow for the representation of deep ℓ-layer networks, ℓ≥1. Sufficient conditions for controller synthesis and observer design are derived in the form of linear matrix inequalities, using a quadratic Lyapunov function. The synthesis method is demonstrated for the control of pH in two tanks in series and a numerical example.

Sequential estimation of temporally evolving latent space network models

Dynamic network data describe interactions among a fixed population through time. This data type can be modelled using the latent space framework, where the probability of a connection forming is expressed as a function of low-dimensional latent coordinates associated with the nodes, and sequential estimation of model parameters can be achieved via Sequential Monte Carlo (SMC) methods. In this setting, SMC is a natural candidate for estimation which offers greater scalability than existing approaches commonly considered in the literature, allows for estimates to be conveniently updated given additional observations and facilitates both online and offline inference. A novel approach to sequentially infer parameters of dynamic latent space network models is proposed by building on techniques from the high-dimensional SMC literature. The scalability and performance of the proposed approach is explored via simulation, and the flexibility under model variants is demonstrated. Finally, a real-world dataset describing classroom contacts is analysed using the proposed methodology.

Dynamic Availability Assessment Using Dynamic Evidential Network: Water Deluge System Case Study

Probabilistic modeling is widespread in engineering practices, mainly to evaluate the safety, risk analysis, and reliability of complex systems. However, insufficient data makes it difficult to estimate the state probability of components or the global system in dynamic complex systems. Furthermore, conventional methods for dependability analysis typically have little capacity to cope with dependence, failure behavior, epistemic uncertainty, and common cause failure simultaneously. This paper presents the application of an extended discrete-time dynamic evidential network (DEN) model to assess the availability of complex systems. The model application combines Dempster-Shafer's theory to treat epistemic uncertainty over a new state-space reconstruction of components and the dynamic Bayesian network to present multi-state system dependability. This model is demonstrated in a real case study of a water deluge system installed as a safety barrier from Algeria's oil and gas plant. The results show the significant influence of these factors on the system's availability. The goal of this modeling is to assure the high availability of a safety barrier in a volatile setting by providing a decision-making tool to prioritize maintenance tasks, preventing the failure of complicated redundant systems, and recommending alterations to the design.

Convolutional neural networks combined with Runge–Kutta methods

A convolutional neural network can be constructed using numerical methods for solving dynamical systems, since the forward pass of the network can be regarded as a trajectory of a dynamical system. However, existing models based on numerical solvers cannot avoid the iterations of implicit methods, which makes the models inefficient at inference time. In this paper, we reinterpret the pre-activation Residual Networks (ResNets) and their variants from the dynamical systems view. We consider that the iterations of implicit Runge-Kutta methods are fused into the training of these models. Moreover, we propose a novel approach to constructing network models based on high-order Runge-Kutta methods in order to achieve higher efficiency. Our proposed models are referred to as the Runge-Kutta Convolutional Neural Networks (RKCNNs). The RKCNNs are evaluated on multiple benchmark datasets. The experimental results show that RKCNNs are vastly superior to other dynamical system network models: they achieve higher accuracy with much fewer resources. They also expand the family of network models based on numerical methods for dynamical systems.

Study of in situ dynamic modulus prediction of asphalt mixture utilizing Ground penetrating radar technology

• The small-scale specimen’s applicability in dynamic modulus test was investigated. • Empirical and ANN dynamic modulus prediction models were proposed using GPR data. • The proposed models provided acceptable results at the confidence level of 95%. It is widely known that the asphalt mixture dynamic modulus is the key mechanical property that affects the quality and service life of asphalt pavement greatly. Ground penetrating radar (GPR) technology is a prospective method commonly used in determining the material information of asphalt pavement. Its utilization is supported by studying the relationship between the characterizations of GPR waveform and the parameters representing the condition of asphalt pavement. However, it is rarely applied in dynamic modulus prediction of asphalt mixture. To fill this gap, this study focuses on developing the corresponding dynamic modulus predictive method on the basis of GPR data. First, the applicability of the ϕ 38 mm × h 110 mm small-scale specimen covering wide range of air void content in the uniaxial compression test was analyzed. Then, the empirical and artificial neural network (ANN) dynamic modulus prediction models were developed by taking into account the dielectric constant of asphalt mixture. Last, the two proposed models were validated based on four asphalt mixture lanes in the field. It was found that the test errors of small-scale specimens were smaller than 15 %, and the statistical difference of dynamic moduli and phase angles measured from the full-size and small-scale specimens is insignificant at the 95 % confidence level meaning the size influence can be ignored. According to the verification experiment, the proposed empirical model and ANN model could yield acceptable results with the average prediction error of 19.3 % and 24.8 %, respectively. It is recommended that the two proposed predictive models coupled with GPR data could be used as an alternative to the currently available methods in estimating the dynamic modulus of asphalt mixture in situ.

Using machine learning to improve neutron identification in water Cherenkov detectors.

Water Cherenkov detectors like Super-Kamiokande, and the next generation Hyper-Kamiokande are adding gadolinium to their water to improve the detection of neutrons. By detecting neutrons in addition to the leptons in neutrino interactions, an improved separation between neutrino and anti-neutrinos, and reduced backgrounds for proton decay searches can be expected. The neutron signal itself is still small and can be confused with muon spallation and other background sources. In this paper, machine learning techniques are employed to optimize the neutron capture detection capability in the new intermediate water Cherenkov detector (IWCD) for Hyper-K. In particular, boosted decision tree (XGBoost), graph convolutional network (GCN), and dynamic graph convolutional neural network (DGCNN) models are developed and benchmarked against a statistical likelihood-based approach, achieving up to a 10% increase in classification accuracy. Characteristic features are also engineered from the datasets and analyzed using SHAP (SHapley Additive exPlanations) to provide insight into the pivotal factors influencing event type outcomes. The dataset used in this research consisted of roughly 1.6 million simulated particle gun events, divided nearly evenly between neutron capture and a background electron source. The current samples used for training are representative only, and more realistic samples will need to be made for the analyses of real data. The current class split is 50/50, but there is expected to be a difference between the classes in the real experiment, and one might consider using resampling techniques to address the issue of serious imbalances in the class distribution in real data if necessary.

Optimization of the Trust Propagation on Supply Chain Network Based on Blockchain Plus

The decentralization of blockchain technology greatly improves the trust relationship in the supply chain network. In view of the lack of trust, uncertainty, and asymmetry in the supply chain network, this paper integrates the blockchain technology to build a network dynamics model of trust representation, calculation, and propagation, and explores how the blockchain influences the supply chain network. The result indicates that the network scale increased by 115.89%, the network connectivity increased by 60.31%, and the average shortest path decreased by 4.95%, after the blockchain trust framework had been deployed in the agricultural supply chain. Meanwhile, the network topology performance such as degree distribution and average clustering coefficient were optimized to varying degrees. Taking agricultural supply chain as an example, the practical significance of topological change was explained. Overall, the blockchain trust mechanism improves the topology of the supply chain network by affecting the trust relationship between nodes.

Multistate time lag dynamic Bayesian networks model for reliability prediction of smart meters

Smart meters are the terminals of smart grids and have the characteristics of complex systems that units in the systems always have many tasks. The reliability prediction model of smart meters is less investigated and traditional models with the two-condition hypothesis are not applicable. In complex systems, multitasking unit has degradation characteristic. Failures of different tasks will affect other units with varying degrees and even lead to the failures of other units in advance. Hence, linear correlation analysis is unsuitable for failure correlation with time lags. Firstly, this paper proposes a two-parameter correlation estimated method for multistate units and multitasking units. The status data of units are divided into one-dimensional and two-state time series, cross-correlation coefficients and time migrations are obtained by analysing two-state time series and used to describe the time-lag correlation. Secondly, the Conditional Probability Table (CPT) of the Dynamic Bayesian Network (DBN) is updated by adding cross-correlation coefficients and time migrations, then time-lag Dynamic Bayesian Network (TDBN) is built. Additionally, the simulation analysis shows that prediction results of TDBN are more consistent of mean failure data. Lastly, the TDBN of the smart meter is built, and the effectiveness of TDBN is further proved by analysing the failure data of smart meters.

Local node feature modeling for edge computing based on network embedding in dynamic networks

In a dynamic network, the characteristics of local nodes include first and higher-order proximity among the nodes as well as different attributes attached to each node. This complexity impose significant challenge for dynamic network modeling. As a result, few dynamic network studies have considered high-order proximity among local nodes. In this paper, we adopt the network embedding method to map high-order proximity of local nodes into low-dimensional, dense and real-valued vectors. Morevoer, we incorporate it into a model-based evolutionary clustering method through regularity conditions. Such a unified framework can increase the effectiveness and robustness of dynamic community detection while pertaining a good explanatory and visualization ability. Experiments based on synthetic and real world data sets show that our model can produce better community detection results than other popular models such as DECS and Genlouvain in dense networks. This result is consistent with the advantage of network embedding method in dense networks.

A novel dynamic Bayesian network approach for data mining and survival data analysis

BackgroundCensorship is the primary challenge in survival modeling, especially in human health studies. The classical methods have been limited by applications like Kaplan–Meier or restricted assumptions like the Cox regression model. On the other hand, Machine learning algorithms commonly rely on the high dimensionality of data and ignore the censorship attribute. In addition, these algorithms are more sophisticated to understand and utilize. We propose a novel approach based on the Bayesian network to address these issues.MethodsWe proposed a two-slice temporal Bayesian network model for the survival data, introducing the survival and censorship status in each observed time as the dynamic states. A score-based algorithm learned the structure of the directed acyclic graph. The likelihood approach conducted parameter learning. We conducted a simulation study to assess the performance of our model in comparison with the Kaplan–Meier and Cox proportional hazard regression. We defined various scenarios according to the sample size, censoring rate, and shapes of survival and censoring distributions across time. Finally, we fit the model on a real-world dataset that includes 760 post gastrectomy surgery due to gastric cancer. The validation of the model was explored using the hold-out technique based on the posterior classification error. Our survival model performance results were compared using the Kaplan–Meier and Cox proportional hazard models.ResultsThe simulation study shows the superiority of DBN in bias reduction for many scenarios compared with Cox regression and Kaplan–Meier, especially in the late survival times. In the real-world data, the structure of the dynamic Bayesian network model satisfied the finding from Kaplan–Meier and Cox regression classical approaches. The posterior classification error found from the validation technique did not exceed 0.04, representing that our network predicted the state variables with more than 96% accuracy.ConclusionsOur proposed dynamic Bayesian network model could be used as a data mining technique in the context of survival data analysis. The advantages of this approach are feature selection ability, straightforward interpretation, handling of high-dimensional data, and few assumptions.

Dynamic Bayesian network model for comprehensive risk analysis of fatigue-critical structural details

Dynamic Bayesian network (DBN) models are widely used for structural risk analysis because of their powerful parameter learning ability and capability of simplifying the problem by parsing it using nodes and arcs. However, these models update the reliability after inspection and maintenance (I&M) in a manner different than that specified by the aircraft structural integrity program (ASIP). In this study, a DBN model is developed to correctly represent the ASIP method. The model updates the crack length distribution after I&M based on the nondestructive testing (NDT) reliability and repair crack length distribution. The nodes for inequality and equality data are explicitly represented in the DBN, which correspond to the crack length vs. signal amplitude and noise characteristics of the NDT system. The proposed model overcomes drawbacks of existing models—initial overestimation and final underestimation—by appropriately considering the repair crack length distribution. Specifically, a decision node is used, which records the fraction of the crack length distribution removed after inspection. Furthermore, a method for constructing conditional probability tables is presented. The proposed model is applied to the in-service fatigue problem for a J85 engine compressor rotor blade. The findings demonstrate that the proposed model can be used in a wide range of applications.

Soft sensing of SO2 emission for ultra-low emission coal-fired power plant with dynamic model and segmentation model

Static models, segmentation models and dynamic models for soft sensing of SO2 emission were developed based on the Ordinary Least Squares, Support Vector Regression, eXtreme Gradient Boosting, Random Forest and Neural Network algorithms with real operational data from a 1000 MW coal-fired power plant with ultra-low emission systems. The prediction results of test set show that static models are not suitable for modeling the desulfurization systems with complex condition and time-delay process. Thus, segmentation models and dynamic models were used to optimize the prediction accuracy. All the prediction performance of the nonlinear models was improved significantly with dynamic model, while the prediction performance of the linear model was improved with segmentation model. The dynamic Neural Network model with one hidden layer achieves the most accurate regression result (R2 = 70.4 %, RMSE = 1.95 mg/m3, MAE = 1.53 mg/m3). The results show the superiority of the dynamic Neural Network model over the other and the dynamic Support Vector Regression model perform second-best.