Research On Space-time Data Prediction Method In The Industrial Internet Of Things Based On Deep Learning

With the popularization and application of information systems,low latency and efficient networking makes it possible to share large amounts of data in the Industrial Internet of Things(IIo T).The user experience and system performance are extremely improved through predictive analysis of large amounts of shared data.The deployment of mobile sensors has become a new paradigm to monitor critical information for IIo T applications in various internal and external environments.In IIo T deployments for intelligent monitoring,due to the large scale and the indeterminate location of nodes,the network is highly unstructured and dynamic,thus the prediction of monitoring data becomes a research hotspot and difficulty.In this thesis,a dynamic space-time data prediction algorithm based on Temporal Convolutional Network and Graph Convolutional Network(TCN-GCN)is proposed to provide accurate and reliable prediction for intelligent monitoring systems.First,an adaptive graph learning module is designed to construct a dynamic graph adjacency matrix according to the mobile nodes.Moreover,the dilation rate of the dilated temporal convolution module is increased to fully exploit the temporal features of space-time data and improve the prediction efficiency.Furthermore,the graph convolution network is developed to effectively capture hidden space relationship between mobile nodes combining the dynamic graph adjacency matrix.Simulation analysis shows that the dynamic prediction algorithm based on TCN-GCN can realize high-precision space-time data prediction,enhancing the perception and utilization of environmental information by smart devices.To compress the above space-time data prediction network so that it can be deployed on edge devices,this thesis proposes a Knowledge Distillation-Based Space Time Data Prediction(KD-ST)algorithm to improve prediction efficiency and support real-time decision-making.The Generative Adversarial Network(GAN)discriminant method and outlier elimination method are developed to transfer the output features of the teacher network to the student network,compressing the model and reducing the interference of outliers for the student network to improve the model prediction accuracy.Moreover,the student networks with different architectures are designed to meet various IIo T applications.In addition,a weight transfer strategy is adopted to improve model training efficiency.Experimental results show that the proposed KDST can provide users with millisecond-level prediction latency on the edge device NVIDIA Jetson Nano.A prediction method based on Federated Learning Combining Temporal Convolutional Network and Graph Convolutional Network(Fed TCN-GCN)is proposed to accurately predict space-time data under the constraints of privacy protection.First,we combine an emerging federated learning strategy with a TCNGCN-based prediction algorithm to provide robust data privacy protection by training models with local datasets.In addition,random sampling of customers participating in federated learning reduces the communication overhead of the algorithm,which is suitable for large-scale IIo T networks and distributed prediction.Simulation analysis shows the Fed TCN-GCN effectively prevents the privacy leakage of the original dataset with a low loss of accuracy.This thesis optimizes in terms of prediction accuracy,model complexity and data privacy protection.The graph convolutional neural network,knowledge distillation and federated learning are adopted to propose a TCN-GCN based dynamic space-time data prediction algorithm,a KD-ST based compress space-time data prediction algorithm and a Fed TCN-GCN based privacy-preserving space-time data prediction algorithm.Experimental results show that the three algorithms proposed in this thesis improve the prediction accuracy of the model,reduce the model complexity and the number of model operations,and achieve privacy protection.