Multivariable time sequence anomaly detection method and system based on graph neural network

Abstract

The invention provides a multivariable time series anomaly detection method and system based on a graph neural network, and the method comprises the steps: taking a sensor in a physical system as a node in a probabilistic graph model, taking the data monitored by the sensor as a time series, carrying out the modeling of a multi-dimensional time series relation, and obtaining a dynamic graph neural network model; obtaining a predicted value of each node at the next time point, and generating an adjacent matrix of each node by using a normalized time alignment measure; when the time reaches the next time point, obtaining the true value of the node, constructing a loss function introducing an adjacent matrix reconstruction error according to the predicted value and the true value so as to train and update the dynamic graph neural network model, and meanwhile, determining the dynamic graph neural network model according to the loss function value of each node, the distribution difference of the neighbor nodes and the adjacent matrix value. Obtaining an abnormal value of each node; and when the error between the node predicted value and the real value is greater than an abnormal value, generating an abnormal alarm. According to the invention, the stability of the abnormal value of the system and the accuracy of slow change anomaly detection are improved.

Description

technical field [0001] The invention relates to anomaly detection / stability detection of industrial equipment systems, in particular to anomaly detection / stability detection of equipment systems in a process control. Background technique [0002] For a complex physical system (such as power plants, data centers, smart factories, etc.), there are a large number of different types of sensors (such as current, voltage, temperature, speed and pressure, etc.) to monitor the operating status of each equipment in the system. In the process of managing such a complex physical system, a key task is to monitor system abnormalities in real time, so that measures can be taken at the appropriate time to resolve the cause of abnormal alarms. [0003] Multivariate time series anomaly detection models can be divided into three categories: supervised, unsupervised, and semi-supervised, according to whether there is a corresponding positive anomaly label. At the same time, considering that t...

AI Technical Summary

Problems solved by technology

[0009] The purpose of the present invention is to solve the problem that the outliers of the output of the existing multidimensional time series anomaly detection model based on the graph neural network vary greatly, and combine the probability graph model theory to propose a multidimensional time series prediction model based on the dynamic time series model

Images