LSTM-based bi-directional urban safety network using a conditional vector of frequency domain decomposition data

Area of Science:

• Civil Engineering
• Structural Engineering
• Seismology

Background:

• Structural health monitoring (SHM) systems are vital for seismic safety but are hampered by sensor data loss.
• Sensor malfunctions, damage, or communication failures compromise the reliability of existing SHM systems.

Purpose of the Study:

• To develop a robust framework for predicting structural responses in sensor-deficient environments.
• To enhance the accuracy and reliability of structural health monitoring during seismic events.
• To improve regional seismic resilience through advanced data recovery techniques.

Main Methods:

• A bi-directional urban safety network utilizing Long Short-Term Memory (LSTM) models was proposed.
• The framework integrates time-domain displacement data and frequency-domain features, transformed into conditional vectors.
• Evaluations were conducted on linear and nonlinear structural systems under seismic loads.

Main Results:

• The proposed method significantly improved prediction accuracy, reducing Root Mean Square Error (RMSE) by up to 27.13% compared to baseline models.
• Conditional vectors enhanced the LSTM model's ability to predict dynamic structural responses.
• The framework accurately predicted maximum response amplitudes and captured time-dependent nonlinear behavior.

Conclusions:

• The bi-directional urban safety network offers a scalable solution for structural response recovery in urban environments with limited sensor data.
• This approach enhances seismic resilience at a regional scale by ensuring the continuous monitoring of structural integrity.
• The findings highlight the potential of advanced machine learning techniques in overcoming SHM limitations.