Quality Control of Carbon Look Components via Surface Defect Classification with Deep Neural Networks.

Area of Science:

• Materials Science
• Computer Science
• Manufacturing Engineering

Background:

• Industry 4.0 relies on data-driven methods like Machine Learning (ML) and Deep Learning (DL) for smart factories.
• Automatic quality control is crucial for precision and standardization in manufacturing.
• Existing methods often combine DL with Nondestructive Testing (NDT) requiring specific setups.

Purpose of the Study:

• To investigate the application of deep neural networks (DNNs) and transfer learning for surface defect classification.
• To assess the feasibility of using plain images for defect detection in automotive carbon fiber components.
• To develop a reproducible methodology for automated quality control.

Main Methods:

• Collected a dataset of 1900 images of Carbon Fiber Reinforced Polymers (CFRP) components.
• Developed and tested ten DNN classifiers using ten pre-trained Convolutional Neural Networks (CNNs) as feature extractors (e.g., VGG16, ResNet, DenseNet121).
• Evaluated binary (defect/no defect) and multiclass (no defect, recoverable, non-recoverable) classification performance.

Main Results:

• The DenseNet121-based classifier achieved 97% accuracy in multiclass defect classification.
• The methodology successfully classified defects from smartphone images under varying conditions.
• No specialized settings or NDT equipment were required.

Conclusions:

• Deep neural networks and transfer learning are viable for classifying surface defects in CFRP components using standard images.
• The proposed method offers a cost-effective and reproducible solution for automated quality control in the automotive sector.
• Open-access data and code facilitate further research and industrial application.