Time Series Anomaly Detection Framework
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Anomaly detection in time series data is challenging due to the difficulty in defining anomalies across different domains, unequal class distribution, and the need for efficient online detection methods that minimize false positives, especially in real-world applications with limited tagged datasets and varying time series characteristics.
Innovation Solution
A framework for automatically determining anomaly detection methods by classifying time series based on characteristics such as stationarity, seasonality, and concept drift, using evaluation metrics like area under the curve analysis and Numenta Anomaly Benchmark scores to select the most suitable method for each type of time series.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If a machine learning classifier approach is used for anomaly detection, then the system can automatically detect anomalies, but the unequal class distribution causes the accuracy measure to reflect only the unequal class distribution rather than true detection performance
Solution Approach 1:
The patent changes the evaluation parameter from accuracy to F1 score, which is appropriate for imbalanced datasets. The F1 score harmonizes precision and recall, providing a reliable measure of anomaly detection performance even when anomalous data points are scarce compared to normal data points.
2Adaptability or versatility
If anomaly detection is performed on real-world streaming applications with limited tagged datasets, then the system can operate in practical scenarios, but it encounters anomalous behavior that was not present in the training set
Solution Approach 1:
The patent performs preliminary action by evaluating multiple anomaly detection methods across diverse time series characteristics before deployment. By pre-testing methods like STL decomposition, Prophet, SARIMAX, and Twitter Anomaly Detection on various dataset types (seasonality, trend, concept drift, missing time steps), the system builds a knowledge base that improves reliability when encountering unseen anomalies in production.
Solution Approach 2:
The patent applies local quality by selecting different anomaly detection methods tailored to specific time series characteristics. Instead of using a single universal method, the system chooses the most appropriate method based on the local characteristics of each time series (e.g., using STL for seasonal data, Prophet for trend data with holidays, SARIMAX for stationary data).
3Adaptability or versatility
If different types of anomalies (point, contextual, and collective) are detected, then the system can identify various anomaly patterns, but different detection methods are better at detecting certain types making selection difficult
Solution Approach 1:
The patent segments the anomaly detection problem by categorizing time series into distinct characteristics (seasonality, trend, concept drift, missing time steps) and matching specific detection methods to each segment. This segmentation approach simplifies the selection process by providing a clear mapping between time series types and optimal detection methods, reducing the complexity of choosing among multiple anomaly types.
4Productivity
If anomaly detection methods are evaluated without considering time series characteristics, then the evaluation process is simpler, but the detection accuracy does not account for specific data properties like stationarity, seasonality, and concept drift
Solution Approach 1:
The patent applies local quality by evaluating anomaly detection methods based on their performance on time series with specific local characteristics. The system separately evaluates methods on datasets with seasonality, trend, concept drift, and missing time steps, ensuring that the evaluation accurately reflects how well each method handles particular data properties rather than averaging performance across heterogeneous data types.
Data Source
AI summary
Disclosed are a framework and method for selecting an anomaly detection method for each of a plurality of class of time series based on characteristics a time series example that represents an expected form of data. The method provides classification of a given time series into one of known classes based on expected properties of the time series, filtering the set of possible detection methods based on the time series class, evaluating the remaining detection methods on the given time series using the specific evaluation metric and selecting and returning a recommended anomaly detection method based on the specific evaluation metric.