Vibration-based Longitudinal Rail Stress Estimation Exploiting Acoustic Measurement and Machine Learning

The mission of this project is to minimize the risks of rail track buckling by predicting the stress state of rail, considering the rail neutral temperature (RNT), in a practical, accurate, and effective manner. The research team developed and implemented an approach that exploits contactless acoustic sensing, finite element modeling (FEM), and machine learning to estimate RNT using vibration data without disturbing track structure or using knowledge from prior baseline measurements. This study is composed of three key components: data collection at a revenue-service line to ensure the proposed sensing technology can function in the field, finite element modeling to interpret field observations and understand the behavior of track vibration characteristics, and machine learning algorithms to establish an input-output relationship between track vibration signatures and in-situ RNT. The tasks in this project were divided into two stages. Stage I collected track vibration data in a realistic field environment covering a wide range of rail temperature, and thereby thermal stress conditions. The data collected in Stage I was used to build up training data for Stage II. During Stage I, the team designed and fabricated fixtures for field data collection, and coordinated with the rail partner to establish an instrumented field site that provides continuous rail temperature, axial rail strain, axial rail load, and RNT data on a heavily used revenue-service line in Illinois. Track vibration data were collected over a wide range of temperatures throughout each day during six field trips to the test site. The acoustic vibration test data prove that consistent high-quality signals can be obtained in the frequency range of 20 to 80 kHz. The high-frequency range is desirable because preliminary results show that rail vibrations in this frequency range are not disrupted by variations in track substructure and foundation. Stage II activities focused on the development of FEM and machine learning algorithms for RNT prediction using the field-collected and FEM data. Using FEM tools, the behavior of high-frequency rail track vibration was predicted under mechanical and thermal loads, and FEM predictions of resonance frequency were within 0.1% of those collected in the field. Considering the FEM simulation, field vibration data, and ground-truth rail stress and temperature data, the team demonstrated that several specific vibration modes of the rail are affected by thermal stress changes and RNT. Using FEM frequency data from these specific modes under the influence of thermal load, a neural network was designed to predict RNT. The results from the neural network demonstrate that it is feasible to predict RNT using the identified high-frequency modes; the system performance with field data indicates that the proposed framework can support RNT prediction within ±5.5ºC (±9.9ºF) when measurement/model noise is low.

Language

  • English

Media Info

  • Media Type: Digital/other
  • Features: Figures; References; Tables;
  • Pagination: 41p
  • Serial:
  • Publication flags:

    Open Access (libre)

Subject/Index Terms

Filing Info

  • Accession Number: 01757821
  • Record Type: Publication
  • Report/Paper Numbers: Rail Safety IDEA Project 41
  • Files: TRIS, TRB, ATRI
  • Created Date: Nov 13 2020 1:41PM