Mahalanobis Anomaly Diagnosis with Adaptive Thresholds for Sparse Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Anomaly detection systems in power generation facilities face challenges in accurately detecting anomalies when the amount of data is small or the number of data points varies, particularly in equipment that operates for extended periods, such as gas turbines, where collecting sufficient data for normal distribution analysis is time-consuming.
Innovation Solution
A diagnosing device and method that calculates the Mahalanobis distance (MD value) and determines anomalies based on this value, with a higher likelihood of no anomaly when the number of samples per unit space is small, using a Mahalanobis distance calculating unit and an anomaly determination unit that adjusts thresholds and corrections for anomaly detection, considering the t-distribution for small sample sizes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the Mahalanobis distance method is used for anomaly detection, then anomaly detection accuracy is improved, but the requirement for large amounts of data increases
Solution Approach 1:
The patent changes the statistical distribution assumption from normal distribution to t-distribution, which is specifically designed for small sample sizes. This parameter change in the statistical model allows the Mahalanobis distance method to work effectively with limited data while maintaining anomaly detection accuracy.
Solution Approach 2:
The patent dynamically adjusts the anomaly determination threshold based on the number of available samples. When sample size is small, the threshold is adjusted to account for the higher uncertainty in the t-distribution, enabling the system to adapt its detection criteria according to data availability.
2Measurement precision
If data collection is extended to gather sufficient samples for normal distribution analysis, then detection accuracy is improved, but the time required increases
Solution Approach 1:
The patent prepares the t-distribution based model in advance, which is specifically designed to handle small sample sizes. This preliminary preparation of the statistical model eliminates the need for extensive data collection during operation, as the model is already optimized for scenarios with limited data availability.
Solution Approach 2:
By changing the statistical parameters from normal distribution assumptions to t-distribution parameters, the system can achieve reliable anomaly detection with fewer samples, thereby reducing the time required for data collection while maintaining detection accuracy.
3Reliability
If the anomaly determination threshold is set strictly, then false positives are reduced, but the likelihood of missing actual anomalies increases
Solution Approach 1:
The patent implements dynamic threshold adjustment based on the number of samples available. The anomaly determination threshold is not fixed but adapts according to the sample size, being more conservative when samples are limited and allowing for better detection sensitivity when sufficient data is available. This dynamic approach balances false positive reduction with anomaly detection sensitivity.
Data Source
AI summary
Provided are a diagnosing device, a diagnosing method, and a program with which it is possible to detect abnormalities accurately even if there is a small amount of data or the number of data points varies. This diagnosing device is provided with a Mahalanobis distance calculating unit which calculates the Mahalanobis distance (referred to as ‘MD value’ hereinbelow) of a detected value, and an abnormality determining unit which determines the presence or absence of an abnormality on the basis of the MD value, wherein the abnormality determining unit determines the presence or absence of an abnormality by arranging that a determination that there is no abnormality is more likely to occur if the number of samples per unit space is small than if the number of samples per unit space is large.