A Lightweight and Efficient Method of Structural Damage Detection Using Stochastic Configuration Network.

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Neural networks are increasingly used in structural health monitoring systems (SHMSs).
• High-performance, low-latency networks are crucial for integrating numerous neural networks in SHMSs.
• Traditional neural networks struggle with temporal data properties and computational costs in SHMSs.

Purpose of the Study:

• To enhance damage detection accuracy and efficiency in SHMSs using vibration signals.
• To address the limitations of stochastic configuration networks (SCNs) in capturing temporal data and managing computational load.
• To introduce novel methods for improving SCN performance without relying on backpropagation.

Main Methods:

• Utilized a stochastic configuration network (SCN) for damage detection.
• Developed a stochastic convolutional feature extraction approach independent of backpropagation.
• Proposed a random node deletion algorithm for automatic pruning of redundant neurons in SCNs.

Main Results:

• The stochastic convolutional feature extraction method improved damage detection accuracy by 30% compared to the original SCN.
• The random node deletion algorithm successfully removed approximately 10% of redundant neurons.
• The proposed methods enhance the speed and accuracy of SHMS operations.

Conclusions:

• The developed stochastic convolutional feature extraction and random node deletion techniques significantly improve SCN performance for SHMSs.
• These methods offer a viable alternative to traditional neural networks for vibration-based damage detection.
• The study demonstrates a pathway to more efficient and accurate structural health monitoring.