A Novel CNN-LSTM Algorithm for Strain Time Series Prediction of Orthotropic Steel Bridge Decks

Area of Science:

• Structural Engineering
• Computational Mechanics
• Data Science

Background:

• Orthotropic steel bridge decks (OSBDs) exhibit complex strain behaviors due to stochasticity and nonlinearity.
• Accurate strain prediction is crucial for bridge health monitoring and fatigue assessment.

Purpose of the Study:

• To develop a hybrid prediction framework for accurate strain time series forecasting in OSBDs.
• To evaluate the proposed model's performance against existing methods across various time horizons.

Main Methods:

• Wavelet decomposition (Daubechies 10) to decouple strain signals into low-frequency (temperature) and high-frequency (vehicle) components.
• A cascaded Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) architecture to process decoupled signals.
• Comparative analysis with CNN-GRU, LSTM, and GRU models using real-world monitoring data.

Main Results:

• The hybrid CNN-LSTM model achieved superior predictive accuracy, with Mean Absolute Percentage Error (MAPE) below 0.6% and R² of 0.961 across all time horizons (24h, 1h, 10min).
• Achieved a single-step inference latency of 0.63 milliseconds, suitable for real-time applications.
• The decouple-then-predict approach effectively mitigated feature interference from mixed signals.

Conclusions:

• The proposed hybrid framework provides a robust and efficient solution for OSBD strain prediction.
• The model's high-fidelity output supports online fatigue evaluation and continuous structural health monitoring.
• This method satisfies real-time computational demands for advanced bridge infrastructure management.