A Transformer-Based Bridge Structural Response Prediction Framework.

Area of Science:

• Civil Engineering
• Structural Health Monitoring
• Artificial Intelligence

Background:

• Accurate structural response prediction is crucial for bridge health monitoring.
• Complex environments and noise hinder feature extraction, leading to poor prediction accuracy.
• Existing methods like Long Short-Term Memory (LSTM) face limitations in handling complex time-series data.

Purpose of the Study:

• To propose a novel Transformer-based framework for enhanced bridge structural response prediction.
• To accurately extract history-dependent features from time-series data for improved accuracy.
• To validate the framework's effectiveness against established methods.

Main Methods:

• Development of a Transformer-based framework with multi-layer encoder and attention modules.
• Utilizing six-month strain response data from a concrete bridge for validation.
• Comparative analysis with a Long Short-Term Memory (LSTM)-based framework.

Main Results:

• The Transformer-based framework demonstrated superior performance in predicting structural responses.
• The proposed method achieved a prediction error less than 50% of the LSTM-based framework.
• Effective capture of history-dependent features in time-series data was confirmed.

Conclusions:

• The Transformer-based framework offers a significant advancement in bridge structural response prediction.
• This method holds potential for improving bridge damage diagnosis and disaster warning systems.
• The framework's ability to handle complex time-series data makes it suitable for real-world applications.