An Autoencoder-Based Deep Learning Approach for Load Identification in Structural Dynamics.

Area of Science:

• Civil Engineering
• Data Science
• Machine Learning

Background:

• Structural health monitoring and load identification generate large volumes of multivariate time-series data from sensor networks.
• Effective dimensionality reduction is crucial for machine learning algorithms to process this big data, retaining informative content and inferring correlations.

Purpose of the Study:

• To propose a novel time series AutoEncoder (AE) model for civil engineering applications, specifically for load identification tasks.
• To develop an AE with inception modules and residual learning for efficient encoding and decoding, creating a highly reduced latent representation.
• To investigate the optimal dimensionality of the latent representation considering data variability and the AE's inverse-forward nature.

Main Methods:

• Utilized a time series AutoEncoder (AE) architecture incorporating inception modules and residual learning.
• Developed an extremely reduced latent representation tailored for load identification.
• Employed the false nearest neighbor heuristic to determine the optimal latent space dimensionality.
• Validated the model on shear building structures subjected to dynamic loadings.

Main Results:

• The proposed AE demonstrated strong signal reconstruction capabilities.
• The model successfully accomplished the load identification task.
• The chosen latent representation dimensionality proved effective for the specific application.

Conclusions:

• The developed AutoEncoder model offers an effective solution for dimensionality reduction in structural vibration analysis.
• The AE is capable of performing accurate load identification, a critical task in structural health monitoring.
• The methodology provides a framework for optimizing latent space dimensionality in similar time-series analysis problems.