Automated identification of autocorrelated control chart patterns utilizing developed convolutional neural networks

Area of Science:

• Industrial Engineering
• Machine Learning
• Quality Control

Background:

• Control Chart Pattern Recognition (CCPR) is vital for product quality.
• Traditional methods require manual feature engineering, which is complex and time-consuming.
• Deep learning offers automated feature extraction but often needs large datasets.

Purpose of the Study:

• To develop and evaluate deep learning models for CCPR, focusing on autocorrelated processes.
• To investigate the effectiveness of transfer learning in enhancing CCPR model performance with limited data.
• To compare custom convolutional networks with pre-trained models (VGG19, MobileNet, LeNet) for CCPR.

Main Methods:

• Utilized custom convolutional neural networks (CNNs) and pre-trained VGG19, MobileNet, and LeNet models.
• Employed transfer learning to adapt pre-trained models for the CCPR task, addressing data scarcity.
• Compared the performance of various CNN architectures, including 1D CNN, 2D CNN, and transfer learning approaches.

Main Results:

• Pre-trained networks achieved superior recognition accuracy compared to a custom 2D CNN without transfer learning, particularly with smaller datasets.
• Pre-trained VGG19 outperformed 1D CNN and conventional machine learning techniques.
• Transfer learning simplified feature extraction and reduced hyperparameter tuning challenges.

Conclusions:

• Pre-trained deep learning models, especially VGG19, are highly effective for CCPR, offering significant advantages over traditional methods and custom CNNs.
• Transfer learning is a viable strategy to overcome data limitations in deep network training for CCPR.
• The proposed approach demonstrates strong potential for real-world CCPR applications.