Physics-informed deep learning for structural vibration identification and its application on a benchmark structure.

Area of Science:

• Civil Engineering
• Machine Learning
• Structural Dynamics

Background:

• Structural vibration identification is crucial for civil engineering, relying on sensor data.
• Predicting responses at unmonituted locations with limited data is challenging.
• Current deep learning methods struggle to incorporate physics laws into vibration analysis.

Purpose of the Study:

• To present a novel physics-informed deep learning (PIDL) framework for structural vibration identification.
• To combine deep generative networks with structural dynamics knowledge.
• To improve the accuracy and reliability of vibration analysis in civil structures.

Main Methods:

• Developed a PIDL framework integrating a convolutional neural network and a physics-informed variational autoencoder.
• Utilized explicit time-domain (ETD) analysis with generated unit impulse response (UIR) signals.
• Incorporated an ETD method-aided convolution function into the loss function for UIR signal alignment.

Main Results:

• The PIDL framework effectively extracts physics-related dynamics features.
• Accurate identification of excitation signals and latent physics parameters was achieved across various damage patterns.
• Demonstrated improved accuracy and reliability compared to conventional deep learning methods.

Conclusions:

• The PIDL framework successfully integrates structural dynamics knowledge into deep learning for vibration analysis.
• The study highlights the potential of PIDL in civil structure monitoring and health assessment.
• This research advances structural vibration identification by incorporating physics-informed machine learning.