Research On Deep Compressed Sensing For Industrial IoT Networks

Industrial Internet of Things(ⅡoT)is an emerging technology that is attached great significance by the national strategies of the Industrial 4.0 in Germany,the Industrial Internet in American and the Intelligent manufacturing in China.It has become the core driving force of digitalization and intelligent transformation and upgrading of modern manufacturing industry.With the help of ⅡoT networks,the intelligent sensing and value extraction of massive industrial data is the primary prerequisites and crucial task to achieve industrial intelligence.Compressed sensing(CS)breaks through the limitation of Nyquist Shannon sampling theorem and can achieve perfect reconstruction of signal with less sampling.Therefore,CS is especially suitable for the sensing and acquisition of data from the underlying devices in ⅡoT networks with limited resources.However,the existing CS schemes rely on sparse transformation,incoherent subsampling and non-linear iterative reconstruction methods,which have many shortcomings in compression ratio,reconstruction accuracy and model robustness.Along with the great success of artificial intelligence,the Deep Compressed Sensing(DCS)method,which combines compressed sensing and deep learning,provides new ideas from data perspective and achieves effective improvement in reconstruction speed and accuracy.Currently,there is still no successful solution in theory and application to achieve efficient sensing and acquisition of industrial data with DCS method,and the specific implementation process of DCS faces huge challenges:First,the data generated from industrial scenarios are not really sparse,The linear DCS method still suffers from the thorny problems of sparse representation and suboptimal linear measurement matrix faced by traditional CS,resulting in difficulty in improving the compression ratio and reconstruction accuracy.Second,the process of solving the high-dimensional inverse problem of CS reconstruction is NP hard,leading to slow convergence of many existing reconstruction algorithms,and thus very difficult to apply and deploy models in resource-constrained ⅡoT networks.Third,the sensing,acquisition and transmission environment of industrial data is very complex,and the existing compression and reconstruction models are susceptible to tiny perturbations.As a result,exploring the compressed acquisition and robust reconstruction methods for Industrial data transmissions has great theoretical and application value.Therefore,this thesis focuses on the core issue of "how to design an efficient and robust compressed acquisition and accurate reconstruction method to break through the bottleneck of limited resources in ⅡoT devices,effectively reduce the transmission latency and resource consumption,and realize the efficient sensing and acquisition of the heterogeneous data across the ⅡoT networks".Taking the combination of CS and deep learning methods as the core technical means,this thesis focuses on the deep compressed sensing(DCS)method for data acquisition and aggregation in ⅡoT networks.The purpose of this thesis is to explore the model design,deployment and model robustness improvement methods of deep compressed sensing with low latency data transmission features from the perspective of data-driven.The specific research content and innovation are briefly described as follows:1)For the scenarios where there exists complex sparsity in data with different structured sparse models,a novel deep compressed sensing method is proposed for the compressed data aggregation process in ⅡoT networks.The data compression and data reconstruction are modeled as the mapping learning process between high and low dimensional space,and a deep compressed sensing network(DCSNet)is designed to jointly train the encoding and decoding network of DCSNet by end-to-end learning.In DCSNet,a linear compression model for mining low-dimensional structures from high-dimensional data is constructed to extract the core features from structured sparse data.The residual mapping learning module is designed to realize the superimposed fusion of high-order features and low-order features of the neural network,which ensures the high-precision reconstruction of the data.Then,a mechanism for the deployment and implementation of DCSNet in ⅡoT networks is designed based on the hybrid CS model,which uses a learning-based measurement matrix instead of a random measurement matrix and replaces the traditional iterative reconstruction process with a learning-based sparse data reconstruction network,thus achieving efficient aggregation of industrial data.Experimental results show that for the data with different structured sparse models and real sensor data,the proposed DCSNet can effectively improve data reconstruction accuracy and communication efficiency in the resource-constrained scenarios.2)A novel quantized deep compressed sensing network(QDCS-Net)for edge-cloud collaborative ⅡoT scenarios is proposed to address the challenges in efficient data transmission requirements and AI model deployment.In QDCS-Net,the joint design of customized quantization layers,dual-path structures,and Swish activation function is adopted to achieve efficient linear and nonlinear low-dimensional embedding and high-precision data reconstruction at higher compression ratios.By formulating determining the deployment of QDCS-Net models as a latency optimization problem,an edge-cloud collaborative transmission strategy is proposed to solve the problem of reasonable deployment of QDCS-Net models,then to improve the delay performance of data transmission under the premise of given transmission bandwidth and computational capacity.Finally,a real vibration data acquisition system is further built to evaluate the transmission and reconstruction performance of QDCS-Net.Experimental results show that our proposed QDCS-Net can effectively improve the reconstruction accuracy as well as the communication efficiency.The MSE,SSIM and PSNR performance of the QDCS-Net outperforms that of those existing state-of-the-art baselines.The proposed QDCS-Net has better performance in reconstructing signals even at extremely low compression ratios.3)Aiming at addressing both data transmission and computing in ⅡoT,we propose an edge-cloud collaborative framework for the ⅡoT,which can help achieve the design goal.A compressed transmission mechanism with edge-cloud collaboration based on deep compressed sensing and dual prediction is proposed and validated to improve transmission efficiency.Since tiny perturbations in original data or compressed data can lead to the emergence of instabilities in deep learning models,the data reconstruction and model instability problems are jointly modeled as a min-max optimization problem,and the robustness of the DCS network is further enhanced by introducing the idea of adversarial learning.Then,a segmented mixed adversarial training(SMART)strategy is proposed to improve the robustness of the DCS models.In the strategy,the adversarial samples with different perturbation levels are generated to train the DCS models in three stages:non-adversarial training,segmented adversarial training and mixed adversarial training.The experimental results show that the reconstruction performance of the segmented adversarial training models and the mixed adversarial training models under different perturbation levels are significantly better than that of the non-adversarial training models,and the proposed SMART strategy can effectively improve the robustness of the quantized DCS model to the random perturbations in the source data and the random bit errors in the compressed data.In summary,in order to solve the challenging problems of complex data sparsity structure,NP-hard for solving high-dimension inverse problems,as well as the poor anti-disturbance capability of DCS models,this thesis focuses on three aspects of design of DCS for efficient compressed data aggregation,design of communication efficient quantized DCS for Edge-Cloud collaborative ⅡoT networks,and the robustness enhancement method of DCS for ⅡoT networks.An efficient and robust data compression and transmission solution of DCS under high compression ratio is proposed,which can significantly improve the reconstruction performance and reduce the transmission consumption.