Multi-End Physics-Informed Deep Learning for Seismic Response Estimation | JoVE Visualize

Area of Science:

Structural Health Monitoring (SHM)
Deep Learning
Computational Mechanics

Background:

Structural health monitoring (SHM) systems often face limitations in measuring all necessary structural responses.
Estimating unmeasured responses from limited sensor data is crucial for accurate structural analysis.
Obtaining large datasets for training deep neural networks (NNs) can be challenging, particularly for rare events like earthquakes.

Purpose of the Study:

To develop a deep neural network framework for reconstructing structural responses using small datasets, especially for rare events.
To enhance the accuracy and robustness of response estimation in structural health monitoring.
To address the challenge of limited training data in applying NNs for structural response reconstruction.

Main Methods:

A novel multi-end autoencoder architecture with skip connections was proposed to compress the parameter space and estimate unmeasured responses.
A physics-based loss function, derived from the dynamic equilibrium equation, was employed to guide training and mitigate overfitting.
Convolutional neural networks (NNs) were utilized to learn the mapping between measured and target responses from limited input data.

Main Results:

The proposed physics-informed NN framework demonstrated superior performance compared to ordinary NNs when trained on small datasets, particularly with noisy data.
The multi-end autoencoder effectively extracted shared patterns and reconstructed diverse structural responses (displacement, velocity, acceleration) at all positions.
The framework was validated through numerical studies on both linear and nonlinear systems, confirming its applicability.

Conclusions:

The developed physics-informed neural network approach is effective for structural response reconstruction under rare events with limited training data.
The combination of a multi-end autoencoder and physics-based loss function significantly improves performance and reduces overfitting.
This method offers a promising solution for enhancing structural health monitoring capabilities, especially in scenarios with data scarcity.