Anomaly Detection Using Similarity Graphs and Label Noise
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current anomaly detection techniques in industrial machinery lack reliability due to insufficient utilization of past anomaly patterns and require arbitrary preprocessing, failing to effectively incorporate label information from both labeled and unlabeled samples.
Innovation Solution
The technique introduces the concept of similarity-based label information by expressing samples as multi-dimensional vectors with sensor values generated by a linear sum of latent variables and varying observation noise, using a graph Laplacian to determine an optimal linear transformation matrix and calculate anomaly scores.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If past anomaly patterns are incorporated into the detection model, then anomaly detection reliability is improved, but arbitrary preprocessing is required which complicates the system
Solution Approach 1:
The patent introduces a similarity matrix as an intermediary structure that captures relationships between samples based on their feature vectors. This similarity matrix serves as a mediator that enables the system to utilize past anomaly patterns without requiring complex arbitrary preprocessing, as the similarity computation naturally handles the data relationships
Solution Approach 2:
The patent changes the approach from traditional preprocessing methods to a parameter-based similarity computation. By computing similarity between samples using their feature vectors and constructing a similarity matrix, the system transforms the problem into a parameter optimization problem that automatically incorporates past anomaly patterns without arbitrary preprocessing steps
2Reliability
If label information from both labeled and unlabeled samples is incorporated, then anomaly detection reliability is improved, but the difficulty of detecting and measuring increases
Solution Approach 1:
The patent merges labeled and unlabeled samples into a unified framework for anomaly detection. By computing similarity between all samples (both labeled and unlabeled) and incorporating this information into the detection model, the system effectively utilizes label information from both sources without requiring separate processing methods, thereby reducing the difficulty of detection and measurement
Data Source
AI summary
A method providing an analytical technique introducing label information into an anomaly detection model. Effective utilization of label information is based on introducing the degree of similarity between samples. Assuming, for example, there is a degree of similarity between normally labeled samples and no similarity between normally labeled and abnormally labeled samples. Also each sensor value is generated by the linear sum of a latent variable and a coefficient vector specific to each sensor. However, the magnitude of observation noise is formulated to vary according to the label information for the sensor values, and set so that normal label≦unlabeled≦anomalously labeled. A graph Laplacian is created based on the degree of similarity between samples, and determines the optimal linear transformation matrix according to a gradient method. A optimal linear transformation matrix is used to calculate an anomaly score for each sensor in the test samples.