| The development of edge computing and machine learning technology has brought new opportunities for data intelligence of the Internet of Things.As a new computing paradigm,edge computing collects and intelligently analyzes data at the edge of the network,which solves the problem of high communication cost and large delay that is difficult for centralized cloud computing.Though combined with edge computing,data transmission is still the main source of energy consumption for energy-constrained wireless sensors,so high energy consumption caused by data collection is one of the problems that data intelligence has to settle urgently.Besides,though the computing capability of the edge device is stronger than that of the sensors,it’s still much weaker than that of the cloud data center.Machine learning models with big resource demand and high computational complexity cannot provide real-time guarantees.Thus,research on the low energy consumption,lightweight and fast data gathering and classification technologies based on edge computing has important application value.Given the above problems,we mined and utilized the spatiotemporal correlation of sensor readings to gather compressive data to reduce the data transmission energy consumption in terms of data collection.And from the perspective of data analysis,a lightweight intelligent classification method for edge devices was studied.The main contributions are as follows:(1)A low-energy consumption compressive data gathering network model based on edge computing scenario was established.The energy consumption caused by the data transmission implemented by direct and relaying transmission was modeled and derived based on the distributed spatiotemporal compressive data gathering scenarios with a cluster-based network structure.The conclusions are that it has the smallest intra-cluster energy cost when the edge devices used as cluster heads are placed at the centers of clusters,and it has the smallest network communication energy cost when using direct transmission for smaller clusters and relaying transmission for bigger ones.The simulation results show that the model can effectively reduce the energy consumption of data transmission.(2)A spatiotemporal compressive data gathering method combining block-wise compressed sensing with logical mapping was proposed.Aiming at the large volume and high spatiotemporal correlation of the consecutively sensed data generated by the densely distributed sensors,random sampling on temporal data is conducted at sensor nodes and spatial compressed sensing on the data gathered inside clusters is performed at edge nodes to reduce redundant data transmission.Besides,an improved logical mapping scheme was proposed,which adjusts the intra-cluster and inter-cluster data at edge nodes and the sink respectively to make them roughly ordered spatially to further reduce the number of transmissions required for accurate reconstruction.Simulation results show that our method can effectively reduce data transmission energy consumption while ensuring data recovery quality.(3)A deep learning fault classification algorithm based on non-reconstructed data was proposed.Aiming at the high dimension of the original sensor data and the heavy computational load of the data reconstruction process of compressed sensing,we use random sampling to obtain temporal compressed data,build a deep neural network model based on stacked sparse auto-encoders and Softmax regression,and use back-propagation algorithms to fine-tune network parameters to achieve the automatic feature extraction and classification based on compressed data.Simulation results show that the method achieves similar classification accuracy and lower running delay. |
