Synthetic Defect Dataset Generation for Weld Inspection AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning and deep learning technologies face challenges in generating sufficient and balanced datasets, particularly for rare defects in manufacturing processes, which are often time-consuming and require large amounts of annotated data, leading to inefficiencies in defect detection.
Innovation Solution
A computer system generates millions of diverse and realistic images of welded metallic plates with defects, using a process that combines base images with randomness to create training datasets, allowing for the training of deep learning models that can detect defects in real-time manufacturing scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If real-world defect images are collected for training, then the dataset reflects actual manufacturing defects, but the quantity of rare defect samples is insufficient
Solution Approach 1:
The patent uses Generative Adversarial Networks (GANs) to generate synthetic defect images that copy the characteristics of real defect patterns. The generative model creates realistic defect samples by learning from limited real defect images and producing additional synthetic variations, thereby multiplying the available training data without requiring proportional increases in real defect collection
Solution Approach 2:
The system performs preliminary defect pattern analysis by training the generative model on available real defect images before actual production defect detection. This preliminary training phase creates a robust synthetic dataset that prepares the detection model in advance, allowing it to handle rare defect types even when real examples are scarce
2Measurement precision
If manual annotation of defect images is performed, then accurate ground truth labels are obtained, but the process is time-consuming and labor-intensive
Solution Approach 1:
The patent implements self-service annotation through the GAN framework, where the generative model automatically produces synthetic defect images with embedded ground truth labels. The system annotates itself by generating images with known defect patterns and locations, eliminating the need for manual human annotation while maintaining high label accuracy
Solution Approach 2:
The generative model acts as an intermediary between real defect images and the training process. Instead of directly using limited real images or relying on slow manual annotation, the intermediary model synthesizes abundant training samples with automatic annotations, bridging the gap between scarce real data and the need for large-scale training
3Adaptability or versatility
If diverse defect types are included in the training dataset, then the model becomes more robust to various anomalies, but the complexity of data collection and balancing increases
Solution Approach 1:
The patent employs dynamic defect generation where the GAN model can adaptively create diverse defect types by modifying generation parameters. The system dynamically adjusts defect characteristics such as type, location, size, and severity during the generation process, allowing comprehensive coverage of defect variations without manually curating separate datasets for each defect category
Solution Approach 2:
The generative model serves multiple functions simultaneously: it generates diverse defect types, balances class distributions, creates realistic images, and produces ground truth labels. This multi-functional approach consolidates several complex data preparation tasks into a single unified system, reducing overall process complexity while achieving diverse and balanced training data
Data Source
AI summary
A system for generating training datasets is provided. The system uses base images to generate a large number of images that include non-defective and defective characteristics. The generated images are then used to train a model that may be used to predict defects in real world images of a manufacturing process.