Parameter Model Generation for Complex Data Anomaly Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing data modeling methods, such as those using normal distribution assumptions, are limited in handling complex data distributions, and anomaly detection methods like sparse coding consolidate information into few variables, losing detailed data representation.

Innovation Solution

A model generation apparatus that selects reference data based on conformity with specific data, specifies parameters, and generates a parameter model indicating the distribution of these parameters to represent complex data distributions effectively.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If sparse coding is used to detect anomalies by consolidating information into two variables (reconstruction error and L1 norm), then anomaly detection capability is improved, but detailed information about original data is lost

Engineering Contradiction:

Improveanomaly detection capabilityVSAvoiddetailed information about original data

Core Design Contradiction:

ReliabilityVSLoss of information

Solution Approach 1:

The patent segments the consolidated anomaly detection task into multiple independent parameter models, each modeling the distribution of a specific parameter (reconstruction error, L1 norm, and other features) separately. This segmentation preserves detailed information about each parameter while maintaining anomaly detection capability.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent transitions from a two-dimensional model (reconstruction error and L1 norm only) to a multi-dimensional parameter space by introducing additional parameter models for other features. This dimensional expansion preserves more information about the original data while maintaining anomaly detection effectiveness.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

2Ease of manufacture

If normal distribution assumption is used for data modeling, then modeling simplicity is improved, but capability to handle complex data distributions deteriorates

Engineering Contradiction:

Improvemodeling simplicityVSAvoidcapability to handle complex data distributions

Core Design Contradiction:

Ease of manufactureVSAdaptability or versatility

Solution Approach 1:

The patent changes the modeling approach from assuming normal distribution to using non-parametric density estimation methods. This allows the model to adapt to complex data distributions without sacrificing simplicity, as the method automatically learns the underlying distribution from data without requiring parametric assumptions.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS10580127B2Model generation apparatus, evaluation apparatus, model generation method, evaluation method, and storage medium

Publication Date: 2020.03.03 CANON KK

AI summary

At least one model generation apparatus that generates a model representing a feature of specific data belonging to a specific category includes a selection unit configured to select, based on a degree of conformity between each of a plurality of pieces of specific data belonging to the specific category and each of a plurality of pieces of reference data included in a previously-set reference data group, at least one piece of reference data from the reference data group with respect to each piece of specific data, a parameter specifying unit configured to specify a parameter corresponding to the reference data selected by the selection unit with respect to each of the plurality of pieces of specific data, and a model generation unit configured to generate, as a model of the specific data, a parameter model indicating a distribution of at least one parameter specified by the parameter specifying unit.