Anomaly Detection Mapping Model Using Subspace Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods, such as Deep SVDD, fail to accurately separate normal and anomalous data in control systems with multiple states due to anomalous data being mapped inside the hypersphere, leading to undetectable anomalies and increased hypersphere radius, which complicates the separation of normal and anomalous data.
Innovation Solution
A learning apparatus and method that learns first and second parameters for a mapping model to a region set based on a subspace and distance, generating feature vectors and adjusting the distance to accurately separate normal and anomalous data by mapping feature vectors to a narrow region set within a subspace, preventing anomalous data from being mapped inside the normal region.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Deep SVDD is used to map normal data inside a hypersphere, then normal data can be fitted inside the hypersphere, but anomalous data is also mapped inside the hypersphere leading to inaccurate separation
Solution Approach 1:
The invention divides the normal data region into multiple separate hyperspheres instead of using a single large hypersphere. Each hypersphere encloses a cluster of normal data points, creating distinct regions that prevent anomalous data from being misclassified. This segmentation allows for more precise separation between normal and anomalous data while maintaining the ability to fit normal data within the hyperspheres.
2Adaptability or versatility
If the hypersphere radius is increased to accommodate all normal data, then all normal data can be included, but the volume of the hypersphere increases causing anomalous data to be mapped inside
Solution Approach 1:
Instead of using one large hypersphere that must accommodate all normal data points, the invention segments the normal data into multiple clusters and creates separate smaller hyperspheres for each cluster. This reduces the volume of each individual hypersphere while collectively covering all normal data, thereby preventing anomalous data from being included in the hypersphere volume.
Solution Approach 2:
The invention transitions from a single-dimensional hypersphere expansion approach to a multi-dimensional clustering approach. By organizing normal data into multiple spatial clusters and creating hyperspheres for each cluster, the solution utilizes the dimensional space more efficiently, achieving comprehensive coverage without requiring a single large hypersphere volume.
3Device complexity
If a single hypersphere is used for normal data, then the mapping model is simple, but the system cannot accurately handle control systems with multiple states
Solution Approach 1:
The invention segments the normal data into multiple clusters corresponding to different system states, creating multiple hyperspheres instead of a single one. This segmentation allows the mapping model to accurately represent control systems with multiple states by assigning distinct hyperspheres to different state clusters, thereby improving anomaly detection accuracy while maintaining manageable model complexity through systematic organization.
Solution Approach 2:
The invention introduces dynamic adaptability to the mapping model by allowing the number, size, and position of hyperspheres to vary based on the distribution of normal data. This dynamic structure enables the model to adapt to different system states and data patterns, improving reliability for complex control systems while keeping the overall model architecture flexible and manageable.
Data Source
AI summary
A learning apparatus includes: a learning unit that learns a first parameter and a second parameter that are included in a mapping model for mapping, to a region set based on a subspace set in advance and a distance from the subspace, a feature vector generated based on normal data input as training data, the first parameter being for generating the feature vector and the second parameter being for adjusting the distance.