Multivariate Time Series Anomaly Detection Using Matrix Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing prediction techniques for multivariate time series struggle to effectively analyze multiple events within a time interval and provide explanations for anomaly events, failing to determine whether an anomaly occurred.
Innovation Solution
A system that processes a multivariate time series by generating true and forecast matrices, calculating forecasting errors, and using neural networks and attention mechanisms to identify anomalies, incorporating techniques like dilated convolutional neural networks and convolutional long short-term memory networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional prediction techniques are used for multivariate time series, then the system is simple to implement, but it cannot effectively analyze multiple events within a time interval and cannot provide explanations for anomaly events
Solution Approach 1:
The patent segments the time series analysis into multiple components: true value matrix, true frequency matrix, true correlation matrix, forecast value matrix, forecast frequency matrix, and forecast correlation matrix. Each matrix captures different aspects of the time series data, enabling comprehensive analysis of multiple events within time intervals while maintaining systematic organization.
Solution Approach 2:
The patent transforms the time series data into multiple dimensional representations including frequency domain (Fourier transform) and correlation domain. This dimensional transformation allows the system to capture events at different scales and provide explanations for anomalies by analyzing relationships across multiple dimensions rather than just temporal sequences.
2Loss of information
If traditional prediction techniques are used, then computational resources are conserved, but the system cannot determine whether an anomaly took place or provide effective explanations
Solution Approach 1:
The patent performs preliminary computations by pre-calculating and storing the true value matrix, true frequency matrix, and true correlation matrix from historical data before anomaly detection is needed. This preliminary action allows the system to quickly compare against forecast matrices during anomaly detection without performing heavy computations in real-time, reducing energy consumption during actual detection while maintaining information completeness.
Solution Approach 2:
The patent creates forecast copies of the time series data through matrix transformations and comparisons. By generating forecast value matrices, forecast frequency matrices, and forecast correlation matrices that mirror the structure of true matrices, the system can identify anomalies through differences without needing to store and process all raw historical data, thus balancing information retention with computational efficiency.
Data Source
AI summary
Provided is a system for detecting an anomaly in a multivariate time series that includes at least one processor programmed or configured to receive a dataset of a plurality of data instances, wherein each data instance comprises a time series of data points, determine a set of target data instances based on the dataset, determine a set of historical data instances based on the dataset, generate, based on the set of target data instances, a true value matrix, a true frequency matrix, and a true correlation matrix, generate a forecast value matrix, a forecast frequency matrix, and a forecast correlation matrix based on the set of target data instances and the set of historical data instances, determine an amount of forecasting error, and determine whether the amount of forecasting error corresponds to an anomalous event associated with the dataset of data instances. Methods and computer program products are also provided.