Explaining 9 ML Anomaly Detection Methods Used by ThresholdIQ
This article provides a detailed explanation of the 9 machine learning anomaly detection methods used by ThresholdIQ, a platform that analyzes data for anomalies. The methods are explained in plain English with concrete examples.
Why it matters
This article provides valuable insights into the technical approaches used by a leading AI-powered anomaly detection platform, which can help users better understand the capabilities and limitations of the system.
Key Points
- 1ThresholdIQ uses 9 separate ML methods to detect different types of anomalies in data
- 2The primary severity driver is the multi-window Z-score, which compares data points to multiple time horizons
- 3Other methods include EWMA spike detection, SARIMA seasonal residuals, and correlation deviation
- 4The methods work together, with the multi-window Z-score driving the primary severity and the other methods boosting severity
- 5The article explains how each method works, what it catches, and what it misses
Details
ThresholdIQ is a platform that analyzes data and flags anomalies using 9 different machine learning methods. These methods look for different types of problems, such as sudden spikes, slow drift, sensor failures, and correlated failures. The primary severity driver is the multi-window Z-score, which compares each data point to the rolling mean and standard deviation across 4 different time windows (50 points, 100 points, 200 points, and 500 points). The more windows that agree a value is unusual, the higher the severity. Other methods include EWMA spike detection, SARIMA seasonal residuals, isolation forest for multivariate outliers, and correlation deviation to catch correlated failures. The article provides plain-English explanations and worked examples for each method, as well as an overview of how they work together to provide a comprehensive anomaly detection system.
No comments yet
Be the first to comment