Multiclass Image Classification Using GANs and CNN Based on Holes Drilled in Laminated Chipboard.

Area of Science:

• Machine Learning
• Predictive Maintenance
• Computer Vision

Background:

• Drill bit wear is a significant issue in manufacturing, leading to reduced quality and potential equipment damage.
• Current methods for assessing drill condition often rely on subjective expert judgment.
• Automated systems for drill wear detection can improve efficiency and prevent costly failures.

Purpose of the Study:

• To develop a multiclass prediction model for classifying drilled hole images into quality categories: "very fine," "acceptable," and "unacceptable."
• To create a system that warns of impending drill wear, thereby reducing damage from blunt tools.
• To compare the performance of custom convolutional neural networks against a benchmark service.

Main Methods:

• Gathering and normalizing real-world drilled hole images.
• Employing data augmentation techniques, including a novel transformation and generative adversarial networks (GANs), to expand the dataset.
• Training various convolutional neural networks (CNNs) for multiclass prediction and comparing them with Microsoft's Custom Vision service.

Main Results:

• The developed CNN models demonstrated effective multiclass prediction of drilled hole quality.
• Several custom-trained models outperformed the benchmark in recognizing less-represented quality classes.
• Data augmentation and GANs aided in dataset rebalancing and improving model robustness.

Conclusions:

• A robust machine learning approach can accurately classify drilled hole quality to predict drill wear.
• Custom CNN architectures offer competitive or superior performance compared to established services for this specific task.
• This predictive model has the potential to significantly reduce damage caused by worn drill bits.