DCGAN Synthetic NDT Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The limited availability of representative data in non-destructive testing (NDT) domains hinders the effectiveness of artificial intelligence (AI) and machine learning (ML) techniques for defect detection and classification, as current data augmentation methods are computationally costly and lack variety in flaw characteristics.
Innovation Solution
A method and system utilizing a Deep Convolutional Generative Adversarial Network (DCGAN) to generate synthetic non-destructive testing datasets by performing numerical analysis on real-time experimentation data, including defective sample dimensions, defect morphologies, and instrument sensitivity, thereby creating a large volume of training datasets with varied flaw characteristics efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation is performed through experimentation and existing techniques, then the volume of training data increases, but the process becomes computationally time-consuming and costly
Solution Approach 1:
The patent uses Generative Adversarial Networks (GANs) to generate synthetic copies of NDT data that mimic real experimental data. The GAN framework creates realistic defect patterns and imaging results without requiring physical experimentation, thus providing large volumes of training data while avoiding the computational costs of traditional data augmentation methods.
Solution Approach 2:
The patent replaces physical experimentation and mechanical data collection processes with computational models and simulations. By using numerical simulation models to generate defect patterns and imaging data, the system eliminates the need for time-consuming physical tests while maintaining data realism and variety.
2Quantity of substance
If data augmentation is performed through experimentation and existing techniques, then the volume of training data increases, but the augmented data lacks variety in flaw characteristics
Solution Approach 1:
The patent introduces diverse defect patterns, geometries, and imaging conditions at the local level within the synthetic data generation process. The GAN model is trained to create variations in flaw characteristics including different defect types, sizes, orientations, and material properties, ensuring each generated dataset contains rich local diversity rather than uniform patterns.
Solution Approach 2:
The patent creates dynamic and varied flaw characteristics by allowing the GAN model to generate continuously varying defect parameters. The system can adaptively create different defect scenarios based on input specifications, enabling the training data to cover a wide range of possible real-world conditions rather than static, limited patterns.
3Productivity
If limited representative data is used for AI and ML techniques, then the implementation becomes faster and cheaper, but accurate detection and classification of defects fails
Solution Approach 1:
The patent performs preliminary data preparation by generating comprehensive synthetic training datasets before deploying AI/ML models. By pre-generating diverse defect scenarios and imaging conditions using GANs, the system ensures that models are trained on sufficient representative data upfront, enabling accurate defect detection when deployed without requiring additional time-consuming data collection during implementation.
Data Source
AI summary
Disclosed herein is a method and system for generating synthetic non-destructive testing dataset. The system receives non-destructive testing datasets related to real-time experimentation of non-destructive testing as input. The testing datasets include dimensions of defective samples, expected defect morphologies, defect probabilities, the sensitivity of instruments, observation from experimental datasets, noise from instrumentation. The system performs numerical analysis on the received one or more non-destructive testing datasets containing one or more flawed geometrical features for generating one or more non-destructive training datasets by using a numerical simulation model. The system further trains a Deep Convolutional Generative Adversarial Network (DCGAN) by using the generated one or more non-destructive training datasets with flaw geometrical features. The system receives a plurality of random number input vectors iteratively at the trained DCGAN and generates a synthetic non-destructive dataset for each of the plurality of received random number input vectors using the trained DCGAN.