Time Series Modeling for Complex System Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex physical systems generate vast amounts of time series data with nonlinear and heterogeneous characteristics, posing challenges for modeling system dynamics without prior knowledge, especially in anomaly detection and management tasks.
Innovation Solution
A method and system that profile data properties, classify dependencies, and generate models based on time series data using sliding window segmentation, linear or nonlinear subspace decomposition, and Vector-Autoregressive modeling, integrated with anomaly scoring and alarm generation for efficient system monitoring.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If domain specific techniques are used to extract knowledge from measurement data, then system understanding and predictive capabilities are improved, but extensive human involvement and domain knowledge are required which becomes difficult to obtain as system scale and complexity increase
Solution Approach 1:
The system performs self-profiling by automatically analyzing measurement data to discover data properties, dependencies, and generate models without requiring external domain experts. The analytic engine autonomously profiles thousands of attributes, classifies dependencies, and generates multiple models that adapt to the specific characteristics of the measurement data.
Solution Approach 2:
The patent replaces manual domain expert analysis with an automated analytic engine that uses computational methods including sliding window segmentation, subspace decomposition, and Vector-Autoregressive modeling to extract knowledge from measurement data, eliminating the need for extensive human involvement.
2Adaptability or versatility
If domain independent solutions are used to extract knowledge from data, then system scale and complexity can be handled, but large amount of historical data and computational resources are required
Solution Approach 1:
The analytic engine segments the analysis process into distinct modules: sliding window segmentation for time series data, subspace decomposition for feature extraction, and model generation for different data patterns. This segmented approach reduces computational complexity by processing data in manageable segments rather than requiring analysis of all historical data simultaneously.
Solution Approach 2:
The system generates multiple models (AR, periodic, constant, CUSUM) and selects the most appropriate ones based on data characteristics, rather than attempting to build a single comprehensive model using all available data. This partial action approach reduces computational burden while maintaining effectiveness.
3Measurement precision
If multiple models are generated to capture different data properties, then modeling accuracy and anomaly detection performance are improved, but computational burden increases
Solution Approach 1:
The system changes parameters dynamically by selecting different model types (AR, periodic, constant, CUSUM) based on the characteristics of the measurement data. The analytic engine profiles data properties and adjusts model parameters accordingly, generating only the necessary models for each specific data pattern rather than uniformly generating all possible models.
Data Source
AI summary
A system and method for analysis of complex systems which includes determining model parameters based on time series data, further including profiling a plurality of types of data properties to discover complex data properties and dependencies; classifying the data dependencies into predetermined categories for analysis; and generating a plurality of models based on the discovered properties and dependencies. The system and method may analyze, using a processor, the generated models based on a fitness score determined for each model to generate a status report for each model; integrate the status reports for each model to determine an anomaly score for the generated models; and generate an alarm when the anomaly score exceeds a predefined threshold.