Related Experiment Videos
Identification of Roll Defect or Damage Based on Rayleigh Waves and Deep Convolutional Neural Network Models
Biao Xiao1, Yue Zhang2, Zhiwei Liu3
1Shanghai Institute of Special Equipment Inspection and Technical Research Co., Ltd., Shanghai 200062, China.
Materials (Basel, Switzerland)
|July 28, 2026
Summary
This study introduces a deep learning method for identifying roller damages using ultrasonic Rayleigh wave signals and power spectrum images. The approach accurately classifies four common damage types, improving quality control in manufacturing.
Area of Science:
- Materials Science and Engineering
- Non-Destructive Testing
- Artificial Intelligence in Engineering
Background:
- Roller damage negatively impacts product quality and requires timely detection and repair.
- Accurate identification of damage types is crucial for effective repair strategies and maintaining production.
- Ultrasonic testing offers efficient, accurate, and safe defect detection but requires intelligent identification methods.
Purpose of the Study:
- To develop an intelligent deep learning-based method for classifying common roller damages.
- To evaluate the effectiveness of using Rayleigh wave signals and their power spectrum images for damage detection.
- To compare the performance of various deep learning models for automated damage classification.
Main Methods:
- Designed experimental setup with an organic glass inclined block and clamping device for ultrasonic testing.
- Acquired time-domain Rayleigh wave signals and generated power spectrum images via Short-Time Fourier Transform.
- Established and compared deep learning models (ResNet, GoogLeNet, DenseNet, AlexNet) with 1D/2D convolutional channels for damage classification.
Main Results:
- Deep learning models effectively classified four common roller damages: void, hole, crack, and adhesion.
- Power spectrum images of Rayleigh waves, particularly at high sampling rates (0.5 MS/s), outperformed raw time-domain signals.
- ResNet-18 demonstrated high accuracy and shorter training times when applied to power spectrum images.
Conclusions:
- Deep learning classification of roller damages using ultrasonic power spectrum images is a viable and effective approach.
- The proposed method enhances the intelligence of ultrasonic detection for improved manufacturing quality control.
- High sampling rates and power spectrum image analysis combined with ResNet-18 offer optimal performance for roller damage identification.