Correlation Model Learning Reliability for Abnormality Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Invariant relation analysis may fail to detect abnormalities or produce erroneous reports due to correlations having insufficient learning during the metric learning period, where time-series values remain unchanged or change linearly, leading to decreased detection capability.
Innovation Solution
A system analysis device generates a correlation model based on time-series metrics and calculates learning reliability by analyzing the behavior of each metric's time series, determining the degree of fitness of single time-series models to assess the reliability of correlations, thereby improving abnormality detection and reducing false reports.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If invariant relation analysis is performed using correlations learned from time-series metrics, then the system can detect abnormalities by identifying correlation destruction, but the detection capability decreases or erroneous reports occur when the time-series values remain unchanged or change linearly during the learning period
Solution Approach 1:
The patent applies preliminary action by performing preliminary actions to identify and exclude metrics with insufficient learning (unchanged or linearly changing time-series) before conducting invariant relation analysis. The system calculates learning reliability for each metric and pre-filters out metrics that do not meet the reliability threshold, thereby preventing correlation destruction detection errors before they occur.
2Productivity
If correlations are selected based on weights calculated from prediction errors, then the correlation model can be generated efficiently, but the model includes correlations that have not been sufficiently learned, leading to decreased detection capability
Solution Approach 1:
The patent implements feedback by introducing a learning reliability evaluation mechanism that provides feedback on the quality of correlation learning. The system calculates learning reliability for each metric based on time-series behavior, and this feedback is used to adjust the selection of correlations included in the invariant relation analysis, ensuring only sufficiently learned correlations are used.
3Measurement precision
If the system monitors prediction errors against a threshold to detect correlation destruction, then abnormalities can be identified, but false abnormality reports occur when the underlying correlation was not properly learned during the learning period
Solution Approach 1:
The patent introduces an intermediary element - the learning reliability calculation mechanism - that acts as a mediator between the correlation model generation and the abnormality detection processes. This intermediary evaluates the quality of learning for each metric and filters out unreliable metrics before they participate in correlation destruction detection, thereby reducing false positives without significantly increasing system complexity.
Data Source
AI summary
In an invariant relation analysis, a capability to detect abnormalities is improved and erroneous abnormality reports are reduced. A system analysis device (100) includes a correlation model generation unit (130) and a learning reliability calculation unit (140). The correlation model generation unit (130) generates, based on time series of a plurality of metrics in a system in a learning period, a correlation model that includes a correlation between metrics. The learning reliability calculation unit (140) calculates learning reliability of the correlation, based on a behavior of a time series of each of metrics relevant to the correlation included in the correlation model, in the learning period.