Synthetic Defect Sample Generation for Sparse Manufacturing Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for manufacturing defect detection suffer from insufficient training datasets, leading to overfitting and inaccurate predictions due to sparse data, and current methods for augmenting datasets rely on supervised images, which are time-consuming and costly.
Innovation Solution
A system and method using an unsupervised segmentation network and generative adversarial network (GAN) to automatically generate synthetic training images with high confidence labels, enhancing the training dataset by combining original and synthetic images based on segmentation masks, and employing a defect classifier to validate the generated images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If supervised images are used to augment training datasets, then the training data quality improves, but the time and cost required for manual labeling increases
Solution Approach 1:
The patent uses generative adversarial networks (GANs) to create synthetic copies of defect images that closely resemble real defect patterns. These synthetic images serve as training data without requiring manual labeling, thus maintaining training data quality while eliminating the time-consuming manual annotation process. The generator network creates realistic defect variations by learning from a small set of labeled images and generating numerous synthetic training samples.
Solution Approach 2:
The system employs self-supervised learning where the model automatically generates its own training data and labels through the GAN framework. The discriminator network provides automatic feedback to the generator, creating a self-improving system that doesn't require external human annotators. This self-service mechanism allows the system to continuously generate and refine training data autonomously.
2Ease of manufacture
If traditional defect inspection methods are used, then the inspection process is simple, but the ability to detect surface defects in new display panel modules is insufficient
Solution Approach 1:
The patent transforms the inspection approach by changing the data representation parameters through synthetic image generation. Instead of relying on traditional inspection parameters, the system generates training data with varied parameters including different defect types, positions, sizes, and patterns. This parameter diversification enables the machine learning model to detect a broader range of defects while maintaining process simplicity through automated analysis.
Solution Approach 2:
The patent replaces traditional mechanical/optical inspection mechanisms with an AI-based system. The machine learning model, trained on synthetic data, substitutes complex manual or automated optical inspection processes. This substitution maintains ease of operation while dramatically improving defect detection capability, particularly for surface defects in new display panel technologies that traditional methods struggle to identify.
3Productivity
If sparse training data is used, then the training process is faster, but the model accuracy and generalization ability deteriorate
Solution Approach 1:
The patent performs preliminary data preparation by generating a comprehensive set of synthetic training images before model training begins. The GAN system pre-generates diverse defect patterns, variations, and edge cases that would be time-consuming to collect manually. This preliminary action ensures that when training starts, the model has access to abundant, high-quality training data, improving both accuracy and generalization while maintaining efficient training throughput.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Systems and methods for classifying products are disclosed. A first data sample having a first portion and a second portion is identified from a training dataset. A first mask is generated based on the first data sample, where the first mask is associated with the first portion of the first data sample. A second data sample is generated based on a noise input. The first mask is applied to the second data sample for outputting a third portion of the second data sample. The third portion of the second data sample is combined with the second portion of the first data sample for generating a first combined data sample. Confidence and classification of the first combined data sample are predicted. The first combined data sample is added to the training dataset in response to predicting the confidence and the classification.