Auto-Tuned Window Anomaly Detection for Changing Time-Series Trends
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods in time series data struggle with false positives due to changing trends and require frequent updates in data contracts, while current forecasting algorithms are inaccurate and resource-intensive, lacking support for multiple seasonal components and holiday effects.
Innovation Solution
A method that separates time series data into multiple windows, computes upper and lower bounds for each window, and uses a voting scheme to detect anomalies based on outlier status across windows, with feedback mechanisms to adjust parameters for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional anomaly detection methods are used with fixed bounds, then the system is simple to implement, but it generates false positives due to changing trends and requires frequent updates
Solution Approach 1:
The patent implements dynamic window sizing where the window size adapts based on the detected trend in time series data. When an increasing or decreasing trend is detected, the window size is adjusted accordingly. This dynamic adaptation allows the system to maintain accurate anomaly detection without requiring frequent manual updates to data contracts, resolving the contradiction between reliability and complexity by automating the adaptation mechanism.
Solution Approach 2:
The system performs preliminary trend detection and window size adjustment before anomaly detection. By pre-adapting the window size to match the underlying trend pattern, the system prepares the detection mechanism in advance, reducing false positives and eliminating the need for reactive updates to detection parameters.
2Measurement precision
If forecasting algorithms are used to improve anomaly detection accuracy, then detection precision improves, but resource consumption increases and training time is excessive
Solution Approach 1:
The patent extracts and removes the forecasting component from the anomaly detection system. Instead of using complex forecasting algorithms that require extensive training and computational resources, the system directly analyzes the trend in historical data and adapts window sizes accordingly. This extraction eliminates the resource-intensive forecasting step while maintaining the ability to detect anomalies accurately through trend-aware dynamic windowing.
Solution Approach 2:
The system replaces expensive, long-training forecasting models with simple, lightweight trend detection mechanisms that require minimal computational resources. The trend detection uses basic statistical methods on historical data, providing a cost-effective alternative to complex forecasting algorithms while achieving sufficient detection precision.
3Adaptability or versatility
If current forecasting algorithms are used, then some prediction capability is provided, but they lack support for multiple seasonal components and holiday effects
Solution Approach 1:
The patent segments the time series analysis into multiple windows of different sizes, each capturing different temporal patterns. By dividing the data into multiple segments with varying window sizes, the system can simultaneously capture short-term and long-term seasonal patterns without requiring a single complex forecasting model. This segmentation approach naturally handles multiple seasonal components by allowing different windows to focus on different periodicities.
Data Source
AI summary
Disclosed is a method of detecting anomalies in time series data. The method includes computing a first bound for a first window of the time series a second bound for a second window of the time series, wherein the second window includes more samples of the time series data. The method also includes generating a first outlier status that indicates whether a current value of the time series data exceeds the first bound, and generating a second outlier status that indicates whether the current value of the time series data exceeds the second bound. The method also includes determining, by a processing device, whether an anomaly is detected in the time series data based on values of the first outlier status and the second outlier status. The method also includes generating an alert in response to determining that the anomaly is detected and sending the alert to a notification system.