Anomaly Detection Model Selection for Time Series Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current anomaly detection models for time series data fail to account for the time-ordered sequence of data and discontinuous model fitting, leading to unnecessary discarding of valid models and reduced symptom reporting, especially during system changes.
Innovation Solution
A method that determines the fitness of a model to a part of the data, specifically ensuring the fitted part includes the most recent data, and builds a reduced model for deployment, considering a threshold ratio of 50% or more for model usability, while accounting for phase changes and anomalies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a model is discarded when it does not fit the entire period of data, then model accuracy is maintained, but the number of discarded valid models increases and infrastructure protection is reduced
Solution Approach 1:
The patent segments the time series data into multiple periods and evaluates model fitness for each period separately. Instead of requiring a model to fit the entire dataset, the system identifies contiguous periods where the model achieves sufficient fit (e.g., 50% or more of the data period), allowing partial model deployment. This segmentation approach resolves the contradiction by maintaining accuracy within valid periods while extending coverage to previously discarded scenarios.
Solution Approach 2:
The patent applies local quality by allowing model fitness to vary across different time periods. A model may fit certain periods well while failing in others due to system changes or anomalies. The system deploy s the model for the specific periods where it demonstrates adequate fit, rather than uniformly accepting or rejecting it across the entire dataset. This enables accurate anomaly detection in stable periods while accommodating system evolution.
2Measurement precision
If traditional statistical distribution tests are used to verify model fit, then statistical validity is achieved, but the time-ordered sequence and contiguity of data are not considered
Solution Approach 1:
The patent introduces dynamic evaluation by considering the temporal characteristics of data periods. Instead of static statistical tests that treat all data points equally, the system evaluates model fitness across contiguous time periods, considering the sequence and continuity of data. This dynamic approach adapts to system changes over time while maintaining statistical rigor through fitness threshold evaluation.
Solution Approach 2:
The patent adds a temporal dimension to model evaluation by analyzing data in sequential periods rather than as a single aggregated dataset. The system considers the contiguity and ordering of data points, evaluating whether a model consistently fits across time periods. This dimensional addition preserves statistical validity while incorporating time-awareness that traditional tests lack.
3Reliability
If multiple anomaly detection models are deployed to cover all metrics, then comprehensive symptom detection is achieved, but the complexity of model management and validation increases
Solution Approach 1:
The patent applies discarding and recovering by temporarily setting aside models that do not fit the entire dataset, then recovering and deploying them for the specific periods where they demonstrate adequate fit. Instead of permanently discarding potentially useful models, the system identifies valid time windows for their deployment. This reduces unnecessary model management complexity while maintaining comprehensive detection coverage across different system states.
Data Source
AI summary
Deploying a model for anomaly detection in time series data. A period of data is received. A model of the period of data is received. It is determined that the model fits a part of the period of data and that the fitted part of the period of data includes the most recent data. A reduced model for the part of the period of data that fit the received model is built. The reduced model is deployed.