Heteogenous Simulated Train Track Bridge Dataset for Live Load Estimation

Heterogeneous Bridge Health Monitoring Dataset for Live Load Estimation

Overview

This dataset provides simulated sensor data for bridge health monitoring with a focus on estimating live load on a bridge deck using both displacement and acceleration signals. The data was generated using the Train-Track Bridge (TTB) simulator [1], a Matlab-based Finite Element Method tool that realistically models the interactions among trains, tracks, and bridges under varying load and temperature conditions. The dataset includes temperature as an exogenous variable, which has a strong influence on both displacement and acceleration signals, notably causing shifts in the displacement signal's magnitude and acceleration signal's frequency.

Statistics:
490,398 samples
756 unique conditions under different load and temperature conditions
21 features (10 displacement + 10 acceleration + 1 exogenous variable (temperature))

Citation:
If you found this dataset helpful, please consider citing our work as follows:
Zhao, Mengjie, et al. "Graph neural networks for virtual sensing in complex systems: Addressing heterogeneous temporal dynamics." arXiv preprint arXiv:2407.18691 (2024).

Dataset Generation

• Simulation Setup:
  Following the setup in [2], the simulation models an ICE3 Velaro train with eight wagons traversing a bridge. The bridge is defined by:

  • Length: 50 m
  • Second moment of area: 51.3 m⁴
  • Mass per unit length: 69,000 kg/m
  • Modulus of elasticity: 3.5 × 10¹⁰ N/m

• Environmental and Operational Conditions:

  • Temperature: Hourly temperature data (0.1°C precision) from Zurich’s Fluntern weather station was integrated to reflect realistic temperature variations.
  • Train Loads:
    • Base load was set at 42,100 kg.
    • Load factors derived from SBB long-distance passenger hourly traffic statistics (2023) were applied. Daily factors were drawn randomly (standard deviation 0.1) and adjusted for weekdays and weekends.
    • Wagon-to-wagon variations were modeled using a normal distribution (standard deviation 500 kg), resulting in load values ranging from 42,100 kg to 53,500 kg.

• Simulation Setup:
  Nine train runs were simulated daily at two-hour intervals from 6:00 AM to 10:00 PM. To capture diverse conditions, simulations were conducted for the first week of each month over one year, yielding 756 unique train runs in total.

Data Processing and Structure

• Signal Acquisition:
  Raw sensor signals (displacement and acceleration) were originally sampled at 1000 Hz during the simulation. Displacement and acceleration values were extracted at 10 equidistant points along the bridge.
• Data Cropping:
  Based on displacement sensor magnitudes, only the period when the train was fully on the bridge (excluding the entry and exit phases) was retained.
• Noise and Filtering:
  Realistic sensor noise was added using additive white Gaussian noise (SNR = 35 dB). Subsequently, the data was downsampled to 100 Hz using an 8th-order Chebyshev Type I filter.

Dataset Partitioning
To emulate real-world data collection and model evaluation scenarios, the dataset was split temporally:

• Training Set: 48 odd-numbered days
• Test Set: 36 even-numbered days
  
  

Software and Code Availability

A data loader and the complete code for processing and utilizing this dataset are provided. Both the code and data are available at:
https://github.com/EPFL-IMOS/htgnn.

This dataset and its corresponding code are associated with the paper Graph Neural Networks for Virtual Sensing in Complex Systems: Addressing Heterogeneous Temporal Dynamics, which has been accepted for publication in the journal Mechanical Systems and Signal Processing (MSSP).

Acknowledgments

This dataset is built upon the following key publications:

[1] D. Cantero, “TTB-2D: Train–track–bridge interaction simulation tool for matlab,” SoftwareX, vol. 20, p. 101253, 2022.

[2] M. Z. Sarwar and D. Cantero, “Probabilistic autoencoder-based bridge damage assessment using train-induced responses,” Mechanical Systems and Signal Processing, vol. 208, p. 111046, 2024.

This work was supported by the Swiss National Science Foundation under Grant 200021 200461. The authors would also like to thank Muhammad Zohaib Sarwar for his valuable insights and discussions regarding the TTB simulator setup.