Time-Series Anomaly Detection with Deviation-Point Windows
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing time-series anomaly detection methods are slow, require a sliding window approach, and struggle to accurately identify anomalies in non-stationary signals, particularly when they start before exceeding the confidence interval, leading to signal discontinuities and false positives in critical applications.
Innovation Solution
A method involving a first predictor trained on a training window to estimate a confidence envelope, moving the training window to the deviation point of an outlier, and training a second predictor on the updated window to efficiently and accurately identify and replace anomalies, using techniques like autoregressive integrated moving average (ARIMA) and convolutional neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a sliding window approach is used for anomaly detection, then the method can be applied systematically to time-series signals, but the detection process becomes slow and inefficient
Solution Approach 1:
The patent pre-calculates and stores confidence intervals for different time windows before actual anomaly detection. When detecting anomalies, the system directly compares signal values against these pre-computed intervals, eliminating the need for real-time predictive model execution and significantly speeding up detection while maintaining accuracy
2Measurement precision
If the training window is moved incrementally through the time-series signal, then all portions can be examined for anomalies, but the process requires excessive computational time and resources
Solution Approach 1:
The patent divides the time-series signal into fixed-size windows and processes each window independently using pre-computed confidence intervals. This segmentation allows parallel processing of multiple windows and eliminates the need for sequential incremental movement, reducing computational time while ensuring complete coverage of the signal
Solution Approach 2:
Confidence intervals are pre-calculated for all possible windows before anomaly detection begins. This preliminary computation stores the expected signal behavior patterns, allowing rapid comparison during detection without re-computing statistical parameters for each window position
3Ease of operation
If only the portion outside the confidence interval is considered as anomaly, then the detection rule is simple, but anomalies beginning before the first exceedance are missed and signal discontinuities are introduced
Solution Approach 1:
The patent identifies the deviation point (where the signal first deviates from the expected pattern) before determining the anomaly boundaries. By pre-identifying this critical point, the system can accurately mark the start of the anomaly even when it begins before the signal exceeds the confidence interval, preventing missed detections and discontinuities
4Adaptability or versatility
If existing predictive models are used on non-stationary signals, then the models can process varying statistical properties, but they fail to identify anomalies accurately when statistical properties change over time
Solution Approach 1:
The patent dynamically adjusts the confidence interval calculations based on the local statistical properties of each time window. Instead of using fixed global parameters, the system computes window-specific mean and standard deviation values, allowing the detection thresholds to adapt to changing signal characteristics while maintaining accurate anomaly identification
Data Source
AI summary
A first predictor trained on observations within a training window of a time-series signal is obtained. A confidence envelope for a prediction window of the time-series signal is estimated using the first predictor. An outlier portion is identified within the prediction window and a deviation point for the outlier portion is determined. The training window is moved such that the training window ends proximate the deviation point. A second predictor is trained on observations within the training window of the time-series signal that has been moved according to the update process