/Structural-State-Translation

Primary LanguagePythonMIT LicenseMIT

Condition transfer between prestressed bridges using structural state translation for structural health monitoring

This repository contains the codes and the dataset used to transfer the structural condition between two prestressed bridges using Structural State Translation, published in a AI in Civil Engineering by Springer.

Publication: Condition transfer between prestressed bridges using structural state translation for structural health monitoring

The study

  • This study uses Structural State Translation (SST) methodology for condition transfer between two structurally dissimilar prestressed concrete bridges, Bridge #1 and Bridge #2, by translating the state (or condition) of Bridge #2 to a new state based on the knowledge obtained from Bridge #1.
  • A Domain-Generalized Cycle-Generative (DGCG) model is trained on two distinct data domains, State-H (healthy) and State-D (damaged), acquired from Bridge #1 in an unsupervised setting, with the cycle-consistent adversarial technique and the Domain Generalization (DG) learning approach implemented.
  • The model is used to generalize and transfer its knowledge to Bridge #2. In this sense, DGCG translates the condition of Bridge #2 to the condition that the model learned after training, which is the 50% missing strands in addition to 10% cross-section loss in the area of the middle girder
  • Specifically, in one scenario, Bridge #2’s State-H is translated to State-D; and in another scenario, Bridge #2’s State-D is translated to State-H.
  • Finally, the translated bridge states are evaluated by comparing them to the real states based on their modal parameters and Average MMSC (Mean Magnitude-Squared Coherence) values, showing that the translated states are remarkably similar to the real ones.
  • The comparison results show a max difference of 1.12% in the bridges' natural frequencies, a difference of 0.28% in their damping ratios, a minimum MAC of 0.923, and an Average MMSC value of 0.947.

Picture3

The code

  • dataset.py provides the loading the dataset
  • blocks.py provides the blocks used in generator and critic (DGCG model)
  • config.py provides the configurations used in the model training
  • critic.py provides the critic model
  • generator.py provides the generator model
  • metric.py provides the FID used in the training
  • train.py is the file for training the DGCG model and some visualization
  • utils.py is only used for gradient penalty used for the critics during the training

The dataset

The dataset used in for the SST is created from numeric a bridge deck models as modeled and analyzed in the Finite Element Analysis (FEA) program.

  • First, the bridge decks are modelled in the FEA program.
  • Then, they went through with Time History Analysis (THA) after Gaussian noise is applied.
  • Subsequently, the acceleration response signals are extracted from the virtual sensor channels placed on each bridge deck model, so as to form the respective acceleration response dataset for each bridge deck state (4 bridge deck state in total).
  • The acceleration response signals are extracted from the virtual sensor channels of each bridge deck model for 1024 seconds and 256 Hz, so as to form the respective dataset for each bridge deck state. Each dataset consists of a 15-channel acceleration response signal. The datasets are denoted as Dataset 1H, Dataset 1D, Dataset 2H, and Dataset 2D, where the numbers (1 and 2) represent the bridge sequence and the letters (H and D) represent the state of the bridges, with H meaning "healthy" and D meaning "damaged".
  • Bridge#1 is used for training, which are Dataset 1H, Dataset 1D.
  • Bridge#2 is used for test, which are Dataset 2H, Dataset 2D.

The train_4096_span1.rar folder includes the undamaged (a0) and damaged (a0) folders, where each folder has 16-second acceleration response tensors (each tensor has 4096 data points) collected from undamaged and damaged conditions of Bridge#1. This folder is used for training.

The test_4096_span2.rar folder includes the undamaged (a0) and damaged (a0) folders, where each folder has 16-second acceleration response tensors (each tensor has 4096 data points) collected from undamaged and damaged conditions of Bridge#2. This folder is used for testing.

The DGCG model

The number of learnable model parameters DGCG model have is 53.7 million. The single DGCG network architecture is shown in the figure below, as there are two of the same networks due to the cycle-consistent adversarial training nature. For instance, one network is responsible for State-H to State-D, and the other is for State-D to State-H.

Picture2