PDF-Based Synthetic Data Generation for NDE Anomaly Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing NDE/NDT techniques face inefficiencies due to a lack of sufficient training data for anomaly detection, leading to inefficient implementation and reduced accuracy.

Innovation Solution

A method and system that generate simulated data using probability density functions (PDF) to extrapolate or interpolate experimental data, creating a large volume of training data for deep learning and machine learning models to detect anomalies, including new anomalies not present in the original data, and validate the models before deployment.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If feature-based classification, artificial neural networks and adaptive filtering techniques are applied for automatic radiographic inspections, then anomaly detection capability is improved, but implementation efficiency deteriorates due to lack of sufficient training data

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoidimplementation efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent creates synthetic copies of experimental data by applying probability density functions to generate simulated data. This copying approach generates additional training data that mimics real experimental data characteristics, enabling efficient training of machine learning models without requiring extensive physical experimentation.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent performs preliminary data generation and model training by creating simulated training datasets before actual anomaly detection is needed. The simulated data is generated in advance using PDF-based methods, allowing the learning model to be pre-trained and validated, thus improving implementation efficiency when deployed for actual inspections.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If more experimental data is collected to train the learning model, then model accuracy is improved, but time consumption and resource requirements increase

Engineering Contradiction:
Improvemodel accuracyVSAvoiddata collection time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

Instead of collecting additional physical experimental data, the patent creates synthetic copies through simulation. The PDF-based generation method reproduces the statistical characteristics of experimental data, providing sufficient training data without the time and resource costs of physical data collection.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent transforms the data generation approach by changing from physical experimentation to mathematical simulation. By using probability density functions and varying simulation parameters, the system generates diverse training data efficiently, avoiding the time-consuming nature of physical data collection while maintaining model accuracy.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3997574B1Method for simulation assisted data generation and deep learning intelligence creation in non-destructive evaluation systems
Publication Date: 2025.09.17 INDIAN INST OF TECH MADRAS
  • EP3997574B1 patent drawingFigure 1
  • EP3997574B1 patent drawingFigure 2
  • EP3997574B1 patent drawingFigure 3

AI summary

Method and system for detecting one or more anomalies in an object are provided. The system receives experimental data of the object and applies a probability density function (PDF) upon one or more variables associated with the experimental data to determine corresponding one or more PDF estimates. The system further generates simulated data associated with the object based on at least one of the one or more PDF estimates and priori data associated with the testing of the object. The simulated data comprises one or more new anomalies unknown in the experimental data along with the one or more anomalies of the experimental data. Furthermore, the system trains a learning model based on the one or more new anomalies and the one or more anomalies of the experimental data. The learning model is applied for detecting any anomaly in an object.