Process Anomaly Detection Using NMF-Based State Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection systems struggle to accurately detect anomalies in processes with multiple operational states due to significant estimation errors arising from improper data classification, whether too fine or too coarse, leading to missed detections or dominant local tendencies.
Innovation Solution
An anomaly detection device that utilizes non-negative matrix factorization (NMF) to generate optimum feature information for each process characteristic, calculating a degree of anomaly based on the difference between process data and normal state, using a coefficient matrix and hyperplane representation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If process data is estimated using the same estimation model for all operational states, then the device complexity is reduced, but the measurement precision deteriorates due to significant estimation errors in certain operational states
Solution Approach 1:
The patent segments the process data into multiple clusters based on operational states using unsupervised learning. Each cluster represents a distinct operational state with its own characteristics. This segmentation allows the system to handle different operational states separately, improving estimation precision without requiring a completely different model for each state, thus balancing complexity and precision.
Solution Approach 2:
The patent dynamically adapts the estimation model by selecting cluster-specific parameters and characteristics based on the current operational state. The system automatically adjusts its behavior according to the identified cluster, enabling accurate estimation across varying operational conditions while maintaining a single unified model structure, thereby resolving the contradiction between model complexity and precision.
2Measurement precision
If process data is finely classified according to operational state, then the measurement precision is improved, but the device complexity increases due to multiple estimation models required
Solution Approach 1:
The patent uses unsupervised learning to automatically segment process data into clusters representing different operational states. This data-driven segmentation achieves fine classification precision without manually designing multiple complex models, as the clustering algorithm automatically identifies patterns and groupings in the data.
Solution Approach 2:
The system employs unsupervised learning algorithms that automatically discover and classify operational states without human intervention or predefined models. The algorithm self-organizes the data into meaningful clusters based on inherent patterns, achieving precise classification while minimizing the complexity of manual model design and configuration.
3Device complexity
If data is roughly classified regardless of operational state, then the device complexity is reduced, but the measurement precision deteriorates due to loss of feature information
Solution Approach 1:
The patent implements a dynamic classification approach where the system automatically adapts its classification granularity based on the characteristics of the input data. The unsupervised learning algorithm dynamically identifies appropriate cluster structures, achieving rough classification when data is homogeneous and fine classification when distinct operational states are present, all within a single flexible model framework.
Solution Approach 2:
The system changes its classification parameters automatically based on the data being processed. The unsupervised learning algorithm adjusts clustering parameters, number of clusters, and feature weighting dynamically according to the operational data characteristics, enabling the same model to achieve appropriate classification precision across different scenarios without manual reconfiguration.
4Ease of operation
If conventional intuitive classification by trial and error is used, then the ease of operation is improved, but the measurement precision deteriorates due to suboptimal data classification
Solution Approach 1:
The patent employs unsupervised learning algorithms that automatically perform data classification without requiring operator intervention or expertise. The system self-service classifies process data into operational state clusters based on inherent data patterns, eliminating the need for trial-and-error classification while achieving superior precision through algorithmic optimization of cluster structures and parameters.
Solution Approach 2:
The patent replaces the manual trial-and-error classification process (mechanical operation) with an automated unsupervised learning algorithm. This substitution eliminates the need for operator expertise and iterative adjustments, achieving both ease of operation (automatic classification) and high precision (algorithm-optimized clustering) simultaneously.
Data Source
AI summary
An anomaly detection device includes: a process data obtaining unit that obtains process data of a process having a plurality of process characteristics; and an anomaly detection unit that detects an anomalous state of the process based on a degree of anomaly that indicates a difference between the obtained process data and a normal state of the process represented by feature information that represents a feature of normal process data for each of the process characteristics. The feature information that represents the normal state of the process is optimum feature information for each of the process characteristics found by evaluating a candidate of feature information based on an error between the normal process data and estimated process data calculated using the candidate of feature information, the candidate of feature information being generated for each of the process characteristics from a plurality of items of normal process data obtained when the process is in the normal state.