Enhanced convolutional neural networks for defect detection in fiber-reinforced composites: a hyperparameter optimization approach
Abstract Fiber-reinforced composites are widely utilized in aerospace, automotive, and structural applications due to their superior strength-to-weight ratio. Despite their advantages, they are prone to internal defects such as fiber-matrix separation, fiber breakage, fiber pullout, void formation,...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
SpringerOpen
2025-08-01
|
| Series: | Journal of Engineering and Applied Science |
| Subjects: | |
| Online Access: | https://doi.org/10.1186/s44147-025-00698-6 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Fiber-reinforced composites are widely utilized in aerospace, automotive, and structural applications due to their superior strength-to-weight ratio. Despite their advantages, they are prone to internal defects such as fiber-matrix separation, fiber breakage, fiber pullout, void formation, and micro-cracking, which can compromise structural integrity and performance. This study presents an automated defect classification framework using convolutional neural networks (CNNs) designed to identify these five defect types from scanning electron microscopy (SEM) images. A total of 3670 SEM images, evenly distributed across the five classes, were used to train and evaluate CNN models. Two hyperparameter optimization techniques—Keras Tuner and particle swarm optimization (PSO)—were applied to enhance the model’s performance by tuning parameters such as the number of filters, kernel size, pooling size, and the number of convolutional layers.The PSO-optimized CNN achieved a superior classification accuracy of 99.23% with a validation loss of 0.02, outperforming the Keras Tuner model, which reached 96.88% accuracy with a validation loss of 0.1119. Additionally, the PSO-based model demonstrated stronger evaluation metrics, including 99.8% precision, 99.75% recall, and a 99.77% F1 score. It also achieved faster inference, processing each image batch in just 25 ms compared to 35 ms for the Keras Tuner model. The novelty of this study lies in its focused application of swarm intelligence for hyperparameter tuning in the context of composite material inspection—a domain that remains relatively underexplored. The proposed method provides a powerful tool for improving the reliability and efficiency of quality assurance processes in fiber-reinforced composite manufacturing. |
|---|---|
| ISSN: | 1110-1903 2536-9512 |