Time series anomaly detection method using GRU-based model

Patent metadata

Field	Value
Publication number	US-12587552-B2
Application number	US-202318098777-A
Country	US
Kind code	B2
Filing date	Jan 19, 2023
Priority date	Aug 3, 2022
Publication date	Mar 24, 2026
Grant date	Mar 24, 2026

Abstract

Provided is a time series anomaly detection method using a GRU-based model. The method includes inputting time series data from a predetermined second to 1 second before a target point in time into a stacked GRU, outputting an output of the stacked GRU after passing through a fully connected layer, adding data from the predetermined second before the target point in time to an output value that has passed through the fully connected layer via a skip connection, and learning the GRU-based model to obtain predicted data at the target time point through actual data at the target time point and mean square error (MSE) loss on the basis of the added data. A similarity calculation unit is additionally provided to the GRU-based model, and the similarity calculation unit receives the time series data, generates latent variables, calculates a sum of similarities, and reflects importance of the latent variables, thereby learning the GRU-based model. This calculates a sum of similarities between variables and reflects importance in predicting data at a target point in time, so that the performance of a time series anomaly detection system can be improved.

First claim

What is claimed is:

1 . A time series anomaly detection method using a GRU-based model, the method comprising:
inputting time series data from a predetermined second to 1 second before a target point in time into a stacked GRU; outputting an output of the stacked GRU after passing through a fully connected layer; adding data from the predetermined second before the target point in time to an output value that has passed through the fully connected layer via a skip connection; and learning the GRU-based model to obtain predicted data at the target time point through actual data at the target time point and mean square error (MSE) loss on the basis of the added data, wherein a similarity calculation unit is additionally provided to the GRU-based model, wherein the similarity calculation unit receives the time series data, generates latent variables, calculates a sum of similarities, and reflects importance of the latent variables, thereby learning the GRU-based model, and wherein a method in which the similarity calculation unit reflects the importance of the latent variables in learning of the GRU-based model comprises: inputting the time series data into an autoencoder; outputting the same data as the input time series data and generating the latent variables when the autoencoder is learned by compressing and restoring the input time series data on a time axis; obtaining a matrix value by calculating mutual similarity for all of the generated latent variables; calculating a sum of similarities of the latent variables on the basis of the matrix value; and multiplying the calculated sum of similarities by the actual data at the target point in time and the predicted data of the target point in time.

2 . The method of claim 1 , wherein the calculated mutual similarity is cosine similarity.

3 . The method of claim 2 , wherein the cosine similarity is a value obtained by dividing a product of first and second latent variables by a product of a L2 norm of the first and second latent variables.

4 . The method of claim 1 , wherein the importance of the latent variables is proportional to the sum of similarities of the latent variables.

5 . The method of claim 1 , wherein the time series data is time series data at an interval of 1 second from the predetermined second to 1 second before the target point in time.

Assignees

• Research & Business Found Sungkyunkwan Univ

Inventors

• Cheong Yungyung
• Seong Changmin

Classifications

• H04L41/16

  using machine learning or artificial intelligence · CPC title

• G06N20/00

  Machine learning · CPC title

• G06F11/2263

  using neural networks · CPC title

• G06F11/3447

  Performance evaluation by modeling · CPC title

• H04L63/1425 Primary

  Traffic logging, e.g. anomaly detection · CPC title