Exemplar-Based Anomaly Detection in Time Series Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting anomalies in time series data are inefficient due to the need to store large training datasets and slow computation, which hinders real-time applications and accuracy, especially when dealing with diverse types of data.
Innovation Solution
A divide-and-conquer approach to learn a compact set of exemplars from training time series data, where each exemplar represents a set of similar windows, allowing for efficient anomaly detection by summarizing the data into a smaller, more manageable set of features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If the entire training time series is stored as the model, then the training time is negligible, but the storage requirement becomes very large and the anomaly detection computation is slow
Solution Approach 1:
The patent extracts only the essential characteristics from the training time series by creating a set of representative exemplars. Instead of storing the entire training dataset, the method selects and stores only the most informative subsequences (exemplars) that capture the underlying patterns, thereby reducing storage requirements while maintaining detection effectiveness.
Solution Approach 2:
The patent creates simplified copies of the training data in the form of exemplars. These exemplars are condensed representations that capture the essential patterns of the original time series without requiring storage of the complete dataset. The exemplars serve as lightweight proxies that enable fast comparison during anomaly detection.
2Loss of time
If the entire training time series is stored as the model, then training is simple and fast, but the anomaly detection computation becomes slow
Solution Approach 1:
The method extracts only the necessary pattern information from the training data by identifying and storing representative exemplars. This extraction process creates a compact model that enables rapid comparison during detection, significantly improving anomaly detection speed while maintaining training efficiency.
Solution Approach 2:
The patent segments the training time series into smaller subsequences and selects representative exemplars from these segments. This segmentation approach allows the method to process training data efficiently while creating a compact set of exemplars that can be quickly compared against test data, thereby improving detection speed.
3Quantity of substance
If predictive techniques are used to model time series data, then the model is compact, but the prediction accuracy may be insufficient for some time series data
Solution Approach 1:
Instead of using predictive models that generate future values, the patent creates accurate copies of actual training patterns in the form of exemplars. These exemplars are direct representations of observed data patterns, ensuring high fidelity to the actual time series behavior while maintaining a compact model size through selective sampling.
Solution Approach 2:
The patent changes the modeling approach from predictive parameter estimation to direct pattern representation. By storing actual subsequences (exemplars) rather than predictive parameters, the method achieves both compactness and high accuracy, as the exemplars directly encode the true patterns without approximation errors inherent in predictive models.
4Device complexity
If piecewise linear paths or boxes are used to represent valid paths, then the learning complexity is O(nlogn), but the method may not accurately represent all types of time series data
Solution Approach 1:
The patent creates a universal exemplar-based model that can represent diverse time series patterns across different data types. The exemplar approach is inherently versatile because it stores actual patterns rather than assuming a specific mathematical form, allowing the same method to effectively model various types of time series data including periodic, trending, and irregular patterns.
Solution Approach 2:
The patent changes from fixed-structure representations (piecewise linear paths or boxes) to flexible pattern-based representations (exemplars). This parameter change allows the model to adapt to different data types by selecting exemplars that reflect the actual patterns in the training data, thereby improving versatility while maintaining manageable complexity through efficient exemplar selection algorithms.
Data Source
AI summary
A method detects anomalies in time series data, by first learning a final set of exemplars by summarizing training time series data using a divide-and-conquer procedure. Then, for each window of testing time series data, a distance to a nearest exemplar in the final set of exemplars is determined, wherein the distance is an anomaly score. Finally, an anomaly is signaled when the anomaly score for a window is greater than a threshold.