Anomaly Detection Method Using Mode-Specific Normal Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods in facilities, such as those using gas turbines, struggle with accurately identifying anomalies when observational data is not included in learning data, leading to reduced inspection reliability and increased user burden, and fail to associate anomaly prior-warnings with events effectively.
Innovation Solution
A method that divides operating states based on event signals, creates a normal model for each mode, and computes anomaly measurements by comparing sensor signals with the model, setting thresholds based on learning data sufficiency to detect anomalies with high sensitivity and precision, while automatically selecting features and learning data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If observational data not included in learning data is treated as anomaly data, then anomaly detection sensitivity is improved, but inspection reliability deteriorates due to false positives in normal signals
Solution Approach 1:
The patent segments the learning data into multiple clusters representing different operational states or patterns. By dividing the data space into distinct regions, the system can recognize normal variations that fall within known clusters while identifying true anomalies as points outside all clusters, thereby reducing false positives while maintaining detection sensitivity.
Solution Approach 2:
The patent performs preliminary clustering analysis on learning data before actual anomaly detection. This preliminary action establishes a reference model of normal operational patterns, allowing the system to distinguish between normal variations and true anomalies more reliably, thus preventing false positives while maintaining high detection sensitivity.
2Measurement precision
If exhaustive past various-state data is stored in database as learning data, then anomaly detection accuracy is improved, but user burden increases due to data management complexity
Solution Approach 1:
The patent implements self-service through automated clustering algorithms that automatically organize and structure learning data without requiring manual curation. The system autonomously identifies patterns, creates clusters, and maintains the learning database, eliminating the need for users to manually manage exhaustive datasets while preserving detection accuracy.
Solution Approach 2:
The patent transforms the raw learning data into clustered representations with extracted features and patterns. By changing the parameter representation from raw exhaustive data to structured cluster models, the system maintains high detection accuracy while significantly reducing the complexity of data management and user burden.
3Quantity of substance
If anomaly is mixed into learning data, then data completeness is improved, but anomaly detection accuracy deteriorates due to reduced divergence degree
Solution Approach 1:
The patent extracts and separates anomaly patterns from the mixed learning data through clustering analysis. By identifying and isolating outlier patterns during the clustering process, the system can maintain complete datasets while preventing anomalous data points from contaminating the normal operation models, thus preserving both data completeness and detection accuracy.
Solution Approach 2:
The patent introduces clustering structures as an intermediary layer between raw mixed data and anomaly detection. This intermediary organization allows the system to handle incomplete or contaminated data by routing observations through cluster-based reasoning, maintaining accuracy even when anomalies are present in the learning dataset.
Data Source
AI summary
This invention provides method for detecting advance signs of anomalies, event signals outputted from the facility are used to create a separate mode for each operating state, a normal model is created for each mode, the sufficiency of learning data for each mode is checked, a threshold is set according to the results of said check, and anomaly identification is performed using said threshold. Also, for diagnosis, a frequency matrix is created in advance, with result events on the horizontal axis and cause events on the vertical axis, and the frequency matrix is used to predict malfunctions. Malfunction events are inputted as result events, and quantized sensor signals having anomaly measures over the threshold are inputted as cause events.