Abnormality Sign Detection Model for Spurious Correlation Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing abnormality detection techniques in large plants, such as nuclear and thermal power plants, suffer from erroneous detections due to spurious correlations between process amounts, leading to inaccurate classification and inefficient performance of machine learning models.
Innovation Solution
An abnormality sign detection system using a neural network-based autoencoder that classifies process amounts into decorrelation data and correlation data, suppressing spurious correlations by treating process amounts with potential erroneous changes as single monitoring signals, thereby improving detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning techniques are used to detect abnormality signs by learning correlation between process amounts, then detection capability is improved, but spurious correlation causes erroneous detection
Solution Approach 1:
The patent segments the learning process into two distinct phases: first learning physical correlations between process amounts during normal operation, then separately detecting abnormalities. This segmentation prevents spurious correlations from contaminating the abnormality detection model, as the system distinguishes between normal operational patterns and true anomalies.
Solution Approach 2:
The patent extracts and removes spurious correlations from the learning data by identifying and excluding process amounts that exhibit false correlations during the learning phase. This extraction process ensures that only genuine physical correlations are used for abnormality detection, eliminating the source of erroneous predictions.
2Measurement precision
If process amounts are classified into different groups to avoid spurious correlation, then erroneous detection is reduced, but classification work becomes huge and accuracy depends on user skill
Solution Approach 1:
The patent implements self-service by enabling the system to automatically perform the classification of process amounts into correlated and uncorrelated groups. The machine learning algorithm autonomously identifies spurious correlations and separates process amounts accordingly, eliminating the need for manual user intervention and ensuring consistent, skill-independent classification accuracy.
Solution Approach 2:
The system incorporates feedback mechanisms that automatically adjust classification based on detected spurious correlations. By continuously monitoring learning results and refining classifications accordingly, the system improves classification accuracy over time without requiring additional user effort or expertise.
3Device complexity
If all process amounts are monitored together in a single model, then system complexity is reduced, but spurious correlation causes performance degradation
Solution Approach 1:
The patent segments the monitoring system into multiple specialized models: one for learning physical correlations during normal operation and another for detecting abnormalities. This segmentation maintains manageable system complexity while significantly improving detection performance by preventing spurious correlations from degrading model accuracy.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
According to one embodiment, an abnormality sign detection system comprising one or more computers configured to perform machine learning of an abnormality-sign detection-model that detects at least one of an abnormality in a target facility to be monitored and a sign of the abnormality, wherein the one or more computers are configured to: acquire a plurality of process amounts generated at the target facility; classify each of the plurality of process amounts into either correlation data for which correlation between the plurality of process amounts is learned or decorrelation data for which correlation between the plurality of process amounts is not learned; generate learning input data depending on classification, the learning input data being data in which each of the plurality of process amounts is associated with the correlation data or the decorrelation data; and perform the machine learning by inputting the learning input data to the abnormality-sign detection-model.