Research On Anomaly Detection Method For Sensor Data Stream And Network Traffic In Wireless Sensor Network

With the further development of Io T technology,Wireless Sensor Networks(WSNs)play an indispensable role in environmental monitoring based on their ability to continuously monitor,sense,and acquire real-time data streams.Sensor nodes in WSNs continuously collect environmental parameters in their monitoring range to form data streams over time.However,sensor nodes in WSNs are often deployed in harsh environments that are inaccessible to humans and are subject to many interferences,which can cause the collected data streams to be inaccurate,affecting the end-users’ ability to make informed decisions.Additionally,sensor nodes are highly susceptible to external malicious network attacks,posing a significant security risk to the secure operation of sensor networks.Therefore,studying abnormal detection methods of sensor node data stream and network traffic reaching user terminals is of great theoretical significance and engineering value.In this thesis,three different anomaly detection methods are proposed for detecting anomalies in sensor data streams and network traffic in WSNs,and their performance is verified through simulation experiments.The main contributions of this thesis are as follows:(1)In order to fully extract and utilize the spatial-temporal characteristics of sensor data streams,a combination of Temporal Convolution Network(TCN)and Bi-direction Gated Recurrent Unit(BiGRU)based on the attention mechanism is proposed for detecting anomalies in sensor data streams.This method first extracts spatial features of sensor data streams by using TCN,extracts time-dependent features using BiGRU from forward and backward propagation,and then introduces the attention mechanism to assign higher weights to key features to maximize the retention of important features of sensor data streams.Finally,the model is trained to achieve abnormal data detection and abnormal class classification of sensor data streams.The experimental results show that the TCN-BiGRU-Attention method can effectively capture the spatial-temporal features of sensor data streams and has higher F1 scores,Recall,and more stable anomaly detection results.(2)The previous anomaly detection methods for high-dimensional sensor data streams are mostly "two-step" methods,i.e.,the feature extraction and dimensionality reduction stages and the anomaly detection stages are independent of each other,which are easy to fall into the local optimum.To address this problem,an end-to-end anomaly detection method SDA-OCNN based on the combination of Stack Denoising Autoencoder(SDA)and One-Class Neural Network(OCNN)is proposed for detecting anomalies in high-dimensional sensor data streams.This method uses the idea of selfencoder data reconstruction,using SDA to reconstruct the high-dimensional sensor data stream with noise reduction and OCNN to obtain a hyperplane to separate normal data from abnormal data by using the low-dimensional features of the hidden layer of SDA to train and achieve the purpose of abnormality detection.By jointly optimizing the parameters of the SDA and OCNN models,this method organically combines the two tasks of data feature extraction and anomaly detection,effectively avoiding the traditional "two-step" method that can fall into the dilemma of local optimum.It is proved that the SDA-OCNN method has higher F1 scores and Recall,and can better perform the anomaly detection task of high-dimensional sensor data streams.(3)Most of the existing traffic anomaly detection methods are oriented to networks with fixed topology,and the research focuses more on the temporal characteristics of traffic,but in practical scenarios,the topology of wireless sensor networks changes due to node energy depletion or failure,etc.,and the traditional anomaly detection methods cannot extract the dynamic features of the network space.To address this problem,an anomaly detection method based on spatial-temporal dynamic graph convolutional networks is proposed for detecting anomalies in network traffic data in WSNs with changing network topology.This method uses graph data structure to describe the spatial characteristics of wireless sensor networks,and captures the spatial-temporal evolution characteristics of wireless sensor networks through spatial-temporal dynamic graph convolutional networks.Additionally,the method introduces the generative adversarial network and the autoencoder to address the characteristics of network traffic data with unbalanced and insufficient features,using the generative adversarial network to generate a few classes of samples to improve the unbalanced distribution of traffic data and the autoencoder to extract the self-supervised features of traffic data and merge them with the spatial-temporal evolution features extracted by the spatial-temporal dynamic graph convolutional neural network to achieve feature enhancement.Experiments prove that the proposed method has stronger detection performance and higher performance measures in networks with changing network topology compared to other network traffic anomaly detection methods.