Deep Learning Algorithms in Industry 4.0; Application of Surface Defect Inspection for Quality Control

Ferromagnetic parts, such as fasteners, are commonly used in industries such as automotive, aerospace, and machinery. However, defects can occur during the production processes, which can compromise the integrity and performance of these parts. To address this issue, magnetic particle inspection (MPI) is a widely used non-destructive testing method to detect surface and near-surface defects in ferromagnetic parts.

However, manual inspection by qualified operators is time-consuming, subjective, and error-prone. Therefore, this study aims to develop an automated method for defect identification based on the magnetic particle technique using deep learning.

The proposed approach involves the application of convolutional neural networks (CNNs) for automatic defect identification in ferromagnetic parts. CNNs are effective in image recognition tasks and can learn complex features without manual preprocessing.

The system is trained using a large dataset of labeled images of fasteners, both with and without defects. By learning the patterns and features associated with defects, the CNN can generalize its learning to new images and identify the presence and locations of defects. Challenges such as overfitting, limited data, imbalanced classes, and model complexity are addressed using regularization techniques, data augmentation, synthetic images generated by GANs, and multi-task learning.

This study also focuses on the design of a reliable image acquisition system that combines frame and line scan cameras to capture high-resolution images of rotating fasteners, enabling detailed analysis of surface finish and dimensional tolerances.

A data-centric approach using data augmentation and synthetic images, along with a model-centric approach using multi-task learning, improves the defect detection model´s performance by incorporating diverse sources of information. Explainable AI techniques, including GradCAM heatmaps, are employed to enhance the interpretability and explainability of the defect detection model.

The combination of knowledge distillation, transfer learning, and fine-tuning further improves the model´s speed and accuracy. Overall, this proposed methodology aims to enhance the efficiency and effectiveness of defect detection processes in various industries.

This study also focuses on the design of a reliable image acquisition system that combines frame and line scan cameras to capture high-resolution images of rotating fasteners, enabling detailed analysis of surface finish and dimensional tolerances.

A data-centric approach using data augmentation and synthetic images, along with a model-centric approach using multi-task learning, improves the defect detection model´s performance by incorporating diverse sources of information.

Explainable AI techniques, including GradCAM heatmaps, are employed to enhance the interpretability and explainability of the defect detection model. The combination of knowledge distillation, transfer learning, and fine-tuning further improves the model´s speed and accuracy. Overall, this proposed methodology aims to enhance the efficiency and effectiveness of defect detection processes in various industries.