XAI Graphical Representation for ML Anomaly Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Machine-learning algorithms used in industrial settings, such as monitoring heavy machinery, often produce complex and 'black-box' outputs, making it difficult to understand how predictions are made and leading to issues in anomaly detection and trust in technical device control.

Innovation Solution

The implementation of explainable artificial intelligence (XAI) techniques, specifically through a graphical representation of a graph based on feature relevance vectors, allows for human-understandable explanations of ML algorithm outputs, enabling identification of common explanation patterns and anomalous data points.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If complex machine-learning algorithms are used to process sensor data and detect anomalies, then detection capability and productivity are improved, but the algorithms become black-box systems that are not human-understandable, making it difficult to configure and control technical devices

Engineering Contradiction:
Improveanomaly detection capabilityVSAvoidalgorithm complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent introduces explanation artifacts as intermediary elements between the black-box ML algorithm and the human user. These artifacts (feature importance scores, attention maps, counterfactual explanations) serve as mediators that translate the algorithm's internal reasoning into human-understandable formats, allowing users to configure and control technical devices based on algorithmic predictions without needing to understand the complex underlying mechanisms

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent replaces the need for direct human understanding of complex algorithmic mechanisms with alternative explanation mechanisms. Instead of requiring users to comprehend the black-box algorithm's internal workings, the system substitutes this with interpretable output artifacts that convey the same informational value through different, more accessible means

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If black-box machine-learning algorithms are deployed for anomaly detection, then detection accuracy is improved, but trust and acceptance from user perspective deteriorate due to lack of understandability

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoiduser trust
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent implements feedback loops where explanation artifacts are provided to users alongside algorithmic predictions. Users can review these explanations (feature importance, attention patterns, counterfactuals) and provide feedback that refines both the explanations and the algorithmic predictions, thereby building trust through transparent iterative improvement while maintaining high detection accuracy

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent uses visual representation techniques where different aspects of the explanation are highlighted through color coding and visual emphasis. Attention maps use color gradients to show which features the algorithm focuses on, and counterfactual explanations use visual differentiation to show what changes would alter predictions, making the black-box reasoning visually accessible and trustworthy

Inventive Principle:
Principle #32Color changes

3Productivity

If complex machine-learning algorithms are used to process operational data, then predictive capability is improved, but the ability to provide actionable control decisions deteriorates due to unclear explanations

Engineering Contradiction:
Improvepredictive capabilityVSAvoidactionable control decision-making
Core Design Contradiction:
ProductivityVSEase of operation

Solution Approach 1:

The patent segments the algorithmic prediction process into distinct interpretable components: input feature analysis, attention mechanism outputs, prediction results, and counterfactual scenarios. Each segment is presented separately with its own explanation artifacts, allowing operators to understand and act on specific aspects of the prediction without being overwhelmed by the complexity of the entire system

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent inverts the traditional approach by not just predicting outcomes but also explicitly modeling and presenting the inverse: counterfactual explanations that show what conditions would lead to different outcomes. This inversion provides operators with actionable guidance by showing them what changes to make rather than just what will happen if nothing changes

Inventive Principle:
Principle #13The other way round (Inversion)

Data Source

PatentEP4312162A1Graphical representation of explainable artificial intelligence
Publication Date: 2024.01.31 SIEMENS AG
  • EP4312162A1 patent drawingFigure 1
  • EP4312162A1 patent drawingFigure 2
  • EP4312162A1 patent drawingFigure 3

AI summary

A graph is determined based on an explainable artificial intelligence analysis of predictions made by a machine-learning algorithm. Global dependencies between different data points are highlighted by the graph.