Research On Time Series Anomaly Detection Based On Graph Neural Network

Anomaly detection is important in industrial applications,such as power grids and transportation networks.These applications often contain multivariate time-series data,and detecting anomalies becomes challenging due to the complex interdependencies between different components in the series: most time-series anomaly detection cannot deeply mine the underlying correlations of the data,performance efficiency Unbalanced with training optimization and lack of exceptional interpretability.Facing the above problems,this thesis proposes two methods for anomaly detection of multivariate time series data with unsupervised anomaly detection of graph neural networks as the core: one is an anomaly detection method based on graph-enhanced normalized flow,and the other is an anomaly detection method based on graph neural and gated recurrent networks.The main research work of this thesis includes the following:To solve the problem of insufficient depth of dependency mining of temporal data,this paper proposes a graph-enhanced normalized stream model GCRNF.Introducing a new stream model incorporating a Bayesian network for causality modeling in the constituent sequences,decomposes the joint probabilities into easily evaluated conditional probabilities and models the conditional dependencies among the constituent sequences.In addition,a graph-dependent encoder is designed to combine a hybrid convolutional neural network and a recurrent neural network to aggregate the hidden states of nodes and capture the local features of sequences as well as the interdependencies between sequences in depth for joint estimation of DAG stream parameters based on a hybrid convolutional neural network and a recurrent neural network.The results demonstrate the superiority of GCRNF in density estimation,anomaly detection,and identification of time series distribution drift.To ensure the balance between excellent detection performance and training optimization and further improve the ability of anomaly interpretability of the detection model,this thesis proposes an anomaly detection method GRN based on a graph neural network and recurrent gated network from the time dimension and variable dimension.First,the data is transformed into vector embedding,and the dependency relationship between the embedded vectors is learned based on the graph structure learning method.Based on obtaining the dependencies between each embedded vector,the sensor vectors are predicted based on the recurrent gated network,and the root cause localization of anomalous events is achieved by combining the attention mechanism and anomaly scoring.The experimental results show that GRN has better anomaly detection performance and more reliable interpretability than the baseline while solving the problems of gradient disappearance and gradient explosion in the training process.In this thesis,extensive experiments using multiple real-world datasets are conducted to validate the effectiveness of the proposed method.The two anomaly detection methods described above provide valuable experience for multivariate time series anomaly detection in industrial control systems and may improve the security and reliability of critical infrastructures.