Manufacturing Product Classification Using One-Class Adversarial Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle to accurately classify manufacturing defects due to imbalanced training datasets, where one class significantly outnumber the other, leading to poor generalization and accuracy.
Innovation Solution
Employ transfer learning using a pre-trained deep convolutional neural network to extract features from manufacturing trace data, and utilize a generative adversarial network to iteratively train a one-class adversarial network to learn a decision boundary for distinguishing between 'good' and 'no good' products, addressing class imbalance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning models are trained on imbalanced manufacturing data, then the majority class classification accuracy is maintained, but the minority class (defect) detection accuracy deteriorates
Solution Approach 1:
The patent introduces a pre-trained deep convolutional neural network as an intermediary feature extractor that transforms raw manufacturing data into enriched feature representations. This intermediary layer bridges the gap between limited imbalanced data and high-quality defect detection, allowing the model to learn effective features even with scarce minority class samples.
Solution Approach 2:
The patent performs preliminary feature extraction using a pre-trained deep convolutional neural network before the actual classification task. This preliminary action of extracting meaningful features from raw data enables the subsequent classifier to achieve high accuracy with limited training data, effectively preparing the data in advance to overcome the imbalanced class distribution.
2Measurement precision
If more training data is collected to improve minority class accuracy, then defect detection improves, but data collection time and cost increase
Solution Approach 1:
The patent uses transfer learning by copying knowledge from a pre-trained deep convolutional neural network that was trained on large-scale general image data. This copied knowledge is adapted to the specific manufacturing defect detection task, enabling high classification accuracy without needing to collect and train on large amounts of domain-specific defect data.
Solution Approach 2:
The pre-trained deep convolutional neural network serves multiple functions: it acts as a feature extractor, a data augmentor through learned representations, and a regularization mechanism. This universal component handles the data scarcity problem across different manufacturing contexts without requiring task-specific data collection.
3Measurement precision
If complex deep learning models are used to handle imbalanced data, then classification accuracy improves, but model complexity and training difficulty increase
Solution Approach 1:
The patent segments the deep learning model into two distinct functional modules: a pre-trained deep convolutional neural network for feature extraction and a simpler classifier for binary classification. This segmentation allows each component to be optimized independently, reducing overall training complexity while maintaining high classification accuracy on imbalanced manufacturing data.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
A system and method for classifying products manufactured via a manufacturing process. A processor receives an input dataset, and extracts features of the input dataset at two or more levels of abstraction. The processor combines the extracted features and provides the combined extracted features to a classifier. The classifier is trained based on the combined extracted features for learning a pattern of not-faulty products. The trained classifier is configured to receive data for a product to be classified, to output a prediction for the product based on the received data.