Anomaly Diagnosis via 2D Sensor Density
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods struggle to accurately diagnose the specific sensors associated with anomalies in multi-dimensional time series data from equipment, often resulting in false alarms, especially in facilities with frequently switching operations and varying conditions.
Innovation Solution
The method involves extracting multi-dimensional feature vectors from sensor signals, calculating reference vectors, and identifying anomaly-related sensors based on 2-dimensional distribution density, rather than solely relying on contribution amounts, to enhance sensitivity and accuracy in anomaly detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If anomaly detection is performed using T2 statistic and Q statistic with contribution analysis, then anomaly can be detected as a combination of sensor signals, but false alarms increase frequently in facilities with switching operations
Solution Approach 1:
The patent applies dynamics by making the reference data adaptive rather than static. The reference data is continuously updated based on recently collected sensor data, allowing the anomaly detection system to adapt to changing operational conditions and switching operations without increasing false alarms
Solution Approach 2:
The patent changes the parameter of reference data from fixed historical data to dynamically updated reference data. By periodically refreshing the reference data with recent normal operation data, the system adjusts to varying operational states while maintaining reliable anomaly detection
2Loss of information
If contribution amount of each sensor is used to diagnose anomaly-related sensors, then sensor association can be identified, but incorrect diagnosis occurs when reference data follows the anomaly
Solution Approach 1:
The patent applies preliminary action by updating the reference data before anomaly diagnosis occurs. By periodically refreshing the reference data with recently collected normal operation data, the system ensures that the reference data reflects current operational conditions and does not contain anomaly information, enabling accurate sensor association diagnosis
3Ease of operation
If monitoring criteria are applied to each sensor individually, then simple anomaly detection can be performed, but combination anomalies of multiple sensors cannot be detected
Solution Approach 1:
The patent merges multiple sensor signals by calculating the distance between the current sensor data vector and the reference data vector in a multi-dimensional space. This combining approach detects anomalies that involve multiple sensors simultaneously while maintaining operational simplicity through unified distance-based criteria
Data Source
AI summary
To sensing an anomaly on the basis of a multi-dimensional time series sensor signal, in order to determine the next action for a countermeasure, survey, or the like, the present invention is configured such that a multi-dimensional feature vector for each time is extracted on the basis of a sensor signal, a reference feature vector for each time is extracted on the basis of a set of characteristic vectors for a predetermined learning period and the characteristic vector of each time, an anomaly measure is calculated on the basis of the difference between the feature vectors for the times and the reference feature vectors, an anomaly is detected by comparing the anomaly measure and a predetermined threshold value, and the anomaly-related sensor for the time the anomaly is detected is identified on the basis of a 2-dimensional distribution density of feature values.