Product Processing Quality Classification Model for Small-Sample and Imbalanced Data Environment.

Area of Science:

• Manufacturing Technology
• Machine Learning Applications
• Data Science

Background:

• Machine learning is increasingly vital in manufacturing, yet faces challenges with small and imbalanced datasets for quality classification.
• Existing methods like GANs can exacerbate class imbalance during data expansion.

Purpose of the Study:

• To propose and evaluate MSMOTE-GAN, a data generation model designed to address small sample and imbalanced data issues in product quality classification.
• To improve the accuracy and robustness of classification models in data-scarce manufacturing environments.

Main Methods:

• Developed MSMOTE-GAN, integrating Mahalanobis Synthetic Minority Oversampling Technology (MSMOTE) with Generative Adversarial Networks (GAN).
• MSMOTE modifies SMOTE by using Mahalanobis distance for more accurate sample generation, accounting for attribute correlations.
• GAN generates synthetic data from a balanced dataset to enlarge sample size and mitigate overfitting.

Main Results:

• The MSMOTE-GAN model effectively balances datasets by addressing majority class oversampling issues inherent in GANs.
• Experiments on water heater liner datasets demonstrated that MSMOTE-GAN significantly improves classification model performance compared to Bootstrap and tableGAN.
• The generated data expanded by MSMOTE-GAN enhanced the overall ability of the classification model.

Conclusions:

• MSMOTE-GAN is a viable solution for product quality classification in manufacturing environments characterized by small and imbalanced datasets.
• The proposed method offers a significant improvement over traditional techniques, paving the way for more reliable automated quality control.