Anomaly Detection via Segmented Autoencoder Feature Attribution

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional anomaly detection systems using autoencoders face challenges in handling heterogeneous data features and lack of labeled anomalous data, leading to ineffective anomaly identification and explanation.

Innovation Solution

An unsupervised anomaly detection method using a one-class classifier trained on normal data, which analyzes individual features through a compound loss function and word embedding for categorical features, providing explainability by highlighting contributing features.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If one-class classifier is used for anomaly detection with heterogeneous features, then the system can handle diverse data types, but it fails to analyze individual features and provide explanation for detected anomalies

Engineering Contradiction:
Improveability to handle heterogeneous data featuresVSAvoidlack of feature-level explanation for anomalies
Core Design Contradiction:
Adaptability or versatilityVSLoss of information

Solution Approach 1:

The patent divides the heterogeneous input data into multiple feature-specific sub-models, where each sub-model processes a particular feature type (e.g., numerical, categorical, temporal) independently. This segmentation allows the system to maintain adaptability across diverse data types while enabling individual feature analysis through dedicated processing paths.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary explanation layer that aggregates outputs from multiple sub-models and generates human-interpretable explanations for detected anomalies. This intermediary component translates the complex multi-feature analysis into feature-level insights, resolving the information loss about which specific features contribute to anomaly detection.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If multi-class classifier is used with limited anomalous training data, then it can detect different anomaly types, but the insufficient anomalous data causes undertraining and reduced detection accuracy

Engineering Contradiction:
Improveability to detect multiple anomaly typesVSAvoiddetection accuracy due to undertraining
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent segments the anomaly detection task into multiple specialized sub-models, where each sub-model is trained on a specific feature type or anomaly category. This segmentation allows the system to handle multiple anomaly types without requiring sufficient training data for each specific anomaly class, as each sub-model focuses on a narrower scope where less data is needed.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent creates a universal anomaly detection framework where a small amount of anomalous training data can be effectively utilized across multiple sub-models through feature sharing and transfer learning. The system achieves multi-functionality in detecting various anomaly types while maintaining reliability by leveraging the limited anomalous data more efficiently through the segmented architecture.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If conventional autoencoder is used for anomaly detection, then it can reconstruct input data, but it cannot provide information about which specific features contribute to the detected anomaly

Engineering Contradiction:
Improveanomaly detection capabilityVSAvoidlack of feature attribution information
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent segments the autoencoder architecture into feature-specific encoder and decoder components, where each segment handles a particular feature type. This segmentation enables the system to maintain precise anomaly detection capability while tracking which specific feature segments contribute most to the reconstruction error, thereby providing feature attribution information.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary attribution mechanism that analyzes the output from multiple feature-specific sub-models and identifies which features contribute most to anomaly detection. This intermediary layer preserves the precise detection capability while recovering the lost feature-level explanation information.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4128062B1Apparatus and method for anomaly detection
Publication Date: 2024.08.07 MITSUBISHI ELECTRIC CORP
  • EP4128062B1 patent drawingFigure 1
  • EP4128062B1 patent drawingFigure 2
  • EP4128062B1 patent drawingFigure 3

AI summary

An anomaly detector for detecting anomaly in input data comprises an auto-encoder trained to encode the input data and decode the encoded input data to reconstruct the input data. Further, the anomaly detector comprises a classifier trained to determine a reconstruction loss indicative of a difference between the accepted input data and the reconstructed input data, where the reconstruction loss includes a weighted combination of a plurality of loss functions evaluating reconstruction losses of a plurality of parts of the reconstructed input data, different types of loss functions, or both. The classifier is further configured to detect an anomaly in the reconstructed input data when the reconstruction loss is above a threshold.