Deep Learning Approach to Impact Classification in Sensorized Panels Using Self-Attention.

Area of Science:

• Engineering
• Materials Science
• Computer Science

Background:

• Structural Health Monitoring (SHM) systems are crucial for assessing infrastructure integrity.
• Traditional impact detection methods face challenges in speed and robustness.
• Deep learning, particularly neural networks, shows promise for enhancing SHM capabilities.

Purpose of the Study:

• To propose and evaluate a novel impact classification method for SHM using the Transformer neural network.
• To compare the Transformer's performance against the Convolutional Neural Network (CNN) for impact detection.
• To assess the Transformer's suitability in terms of performance, scalability, and computational time.

Main Methods:

• Utilized a Transformer neural network, incorporating Self-Attention mechanisms, for impact classification.
• Employed a Convolutional Neural Network (CNN) as a benchmark for comparison.
• Preprocessed time-series data from piezoelectric sensors using Fourier Transform for network input.

Main Results:

• The Transformer method demonstrated a significant reduction in computational complexity for impact detection.
• The Transformer achieved excellent prediction results, comparable or superior to the CNN.
• The study provided insights into the advantages and disadvantages of both Transformer and CNN architectures.

Conclusions:

• The Transformer neural network is a highly suitable and promising architecture for SHM impact classification.
• Transformer-based methods offer improved speed and robustness in impact detection.
• This approach has the potential to significantly advance the field of structural health monitoring.