Symbolic Decision Tree Diagnosis for Systems Without Equation Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Diagnosing complex industrial systems without knowledge of their mathematical equations is challenging, especially as systems become more complex, making traditional structural analysis methods difficult to apply.

Innovation Solution

A method is proposed to construct a decision tree for system diagnosis using symbolic classification, which does not require knowledge of the system's governing equations. This involves creating a training dataset with measured values and labels representing nominal and failure states, and using a genetic algorithm to generate and test classification functions that can discriminate between these states.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional structural analysis methods are used to diagnose complex industrial systems, then diagnostic capability can be achieved through analytical redundancy relations, but the method becomes increasingly difficult to apply as system complexity increases and mathematical equations become more complex

Engineering Contradiction:
Improvediagnostic capabilityVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent replaces traditional structural analysis methods (which rely on mathematical equations and analytical redundancy relations) with a machine learning-based approach using neural networks. The neural network learns diagnostic patterns directly from operational data without requiring explicit mathematical models of the system, thereby maintaining diagnostic capability while overcoming the limitations imposed by increasing system complexity.

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

Solution Approach 2:

The patent transforms the diagnostic approach by changing from equation-based parameters to data-driven parameters. Instead of using analytical redundancy relations derived from mathematical equations, the system uses neural network weights and activation functions that are trained on operational data, allowing the diagnostic system to adapt to complex systems without requiring explicit mathematical models.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If structural analysis methods are applied to complex industrial systems, then failure indicators can be constructed using analytical redundancy relations, but deep knowledge of the system's mathematical equations is required which is often unavailable

Engineering Contradiction:
Improvefailure detection capabilityVSAvoidlack of system equation knowledge
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

The patent creates a virtual model of the system through a neural network that copies the system's operational behavior from training data. This neural network model serves as a surrogate for the unknown mathematical equations, allowing failure detection without requiring actual knowledge of the system's governing equations. The network learns the relationships between inputs and outputs empirically.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent introduces a neural network as an intermediary between the complex industrial system and the diagnostic process. The neural network acts as a mediator that translates raw operational data into diagnostic information without requiring direct access to or understanding of the system's mathematical equations, thereby overcoming the information loss caused by unavailable system knowledge.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Measurement precision

If machine learning approaches are combined with structural analysis for ARR identification, then diagnostic performance improves, but it is still necessary to know the variables involved in the mathematical equations which makes the method unusable for many industrial systems

Engineering Contradiction:
Improvediagnostic performanceVSAvoidapplicability to industrial systems
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent creates a universal diagnostic approach using a neural network that can be applied to any industrial system regardless of its specific mathematical equations or variables. The neural network is trained on operational data and can diagnose various types of systems (process control, manufacturing, energy, etc.) without requiring system-specific knowledge, thereby achieving both high diagnostic performance and broad applicability.

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

Solution Approach 2:

The patent extracts the essential diagnostic functionality from the complex combination of structural analysis and machine learning by using a standalone neural network that operates directly on sensor data. This extraction removes the requirement for knowledge of mathematical equations and variables, while preserving the improved diagnostic performance that machine learning provides.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP4478143A1Method for constructing a decision tree for diagnosing a system, method for diagnosing the system, corresponding devices and computer programs
Publication Date: 2024.12.18 ATOS FRANCE
  • EP4478143A1 patent drawingFigure 1a~1b
  • EP4478143A1 patent drawingFigure 2
  • EP4478143A1 patent drawingFigure 3a~3b

AI summary

The invention relates to a method and device for constructing a decision tree for diagnosing a system comprising a plurality of components, comprising: - obtaining a training dataset comprising pairs, each pair comprising a vector of measured values ​​of observable variables representative of a system operation and an associated label, the label belonging to a group of labels representative of a nominal operating state of said system or a failure state of said component, - processing a current node created by dividing a previous node and associated with a subset of the training dataset, referred to as the current dataset, said processing comprising, when at least one division criterion is satisfied,the division of the current node into a first and a second child node by applying a classification function obtained from the current dataset and defined to associate with several of said observable variables, a first class, called the nominal class, representing the nominal operating state of the system or a second class, called the fault class, representing said fault states of said system, and - the provision of the decision tree for the diagnosis of the system.