A Novel Method of Bridge Deflection Prediction Using Probabilistic Deep Learning and Measured Data.

Area of Science:

• Structural Engineering
• Artificial Intelligence in Civil Engineering
• Bridge Health Monitoring

Background:

• Suspension bridges are flexible structures requiring precise deflection control for operational safety.
• Predicting vertical girder deflection is complex due to stochastic traffic loads and environmental temperature variations.

Purpose of the Study:

• To develop an integrated method for predicting vertical deflection intervals of suspension bridge girders.
• To enhance the accuracy of deflection prediction and establish methods for identifying abnormal deflections and warning thresholds.

Main Methods:

• Utilized a Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network for time-series data analysis.
• Integrated a probability density estimation layer with Gaussian distribution (GD) for interval prediction.
• Trained the model using bridge health monitoring data, including environmental temperature, vehicle load, and deflection.

Main Results:

• The CNN-LSTM-GD model significantly improved Root Mean Squared Error (RMSE) and coefficient of determination (R2) compared to LSTM and CNN-LSTM models for both short and long time scales.
• Achieved up to 54.40% RMSE improvement and 12.37% R2 increase over baseline models.
• Demonstrated effectiveness in identifying abnormal deflections and setting warning thresholds.

Conclusions:

• The proposed CNN-LSTM-GD model provides a robust and accurate approach for suspension bridge deflection prediction.
• The method is crucial for developing effective bridge deflection early-warning systems.
• Accurate deflection prediction and abnormal deflection identification enhance bridge operational safety and maintenance.