Reaserch And Implementation Of Anomaly Detection Of Smart Gateway Supporting Edge-Cloud Collaboration

With the development of the digital economy,more and more industries have embarked on the road of digital transformation.As an important research direction for data value mining,anomaly detection increasingly gains more attention.However,in the real scenario,between the central cloud platform and the device side,there often exists communication-limited difficutilty,which leads to a model with bad results in anomaly detection after training,so that the system cannot detect data timely and accurately.This thesis studies and realizes a smart gateway of anomaly detection supporting edge-cloud collaboration.When the communication is limited,this gateway can not only receive data from the device,compress and detect anomaly data at the edge,but also train model for anomaly detection in the cloud so as to alarm after discovering anomaly.In order to achieve the targets,this thesis proposes an algorithm of sampling and compression of reservoir based on kernel density estimation and Variational AutoEncoder and Bi-directional Long Short-Term Memory based on Self-Attention(VAE-SABiLSTM)network structure.The former introduces adaptive kernel density estimation in the process of reservoir sampling,and determines the sampling probability according to data density to complete data compression.Due to data being compressed,its space becomes fragmented,so the system use latter for training.It uses compressed data as input,and then undergoes a variational autoencoder and an improved bidirectional long short-term memory based on self attention mechanism.Finally,the output of anomaly detection is obtained.This network structure changes the distribution of variational autoencoder,and introduces self attention mechanism into bi-directional long short-term memory,enabling it to obtain more comprehensive contextual information,thereby learns the important features of the current local sequence and improves the accuracy of detection.In order to solve the feature drift in the latitude of time,this thesis also proposes a negative feedback mechanism,which continuously adjusts the parameters in the data compression process based on the results of anomaly detection,ensuring the feasibility and accuracy of long-term anomaly detection for the device.After a series of experiments,the above method has higher representability and better detection accuracy in the results of compressed data compared to other schemes,like VAE,LSTM,VAE-LSTM,Federated Learning using Gaussian Variation Self-Encode Nework(FL-MGVN).This thesis first introduces the relevant background and technology of smart gateway of anomaly detection supporting edge-cloud collaboration,and analyzes its requirements by the results of research on relevant products.Then,an algorithm of sampling and compression of reservoir based on kernel density estimation and Variational autoencoder and Bidirectional Long short-term memory based on Self-Attention(VAESABiLSTM)network structure are proposed and experimented in this thesis.Finally,the thesis introduces the design architecture of overall system and each module,and their implementation and evaluation.