In this study, the project team developed new technology for in-place rail stress (rail neutral temperature or RNT) for continuously welded rail (CWR). Track alignment irregularities caused by excessive longitudinal rail stress can lead to rail thermal buckling, and thereby disrupt safe railway operation. The ability to monitor RNT capability without disturbing track structure or using additional knowledge from prior baseline measurements is a critical need and represents a breakthrough in rail monitoring technology. The project team developed and implemented technology that makes of acoustic vibration measurements, finite element modeling (FEM), and machine learning to address this problem. The study was composed of three key components: acoustic vibration data collection at an instrumented revenue-service rail line, finite element modeling to interpret field observations, and machine learning algorithms to establish an input-output relationship between track vibration signatures and in-situ RNT.
Rail temperature, stress, RNT, and vibration data were collected, over a six-month span, through six field trips to the instrumented test site, comprising a wide range of temperatures and stress conditions. The stress state varied throughout each day, and RNT evolved across the seasons. High-quality acoustic vibration data across a frequency range of 20 to 80 kHz were collected from the field test site. Using FEM tools, the behaviors of high-frequency rail track vibration were predicted under simulated mechanical and thermal loads; the FEM predictions of resonance frequency were within 0.1% of those collected in the field. Using FEM modal frequency data from specific vibration 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 indicated that the proposed framework can support RNT prediction within ±5.5ºC (±9.9ºF) when measurement/model noise is low.
The most important payoffs of our technology and approach include (i) it does not require modification of track structures; (ii) its sensing configuration is simple, robust, and does not involve any hazardous phenomena for humans; and (iii) it analyzes vibrational data using FEM-assist machine learning algorithms, which eliminates the need for reference measurements and reduces influences from varying tie, clip/fastener, and sub-base conditions. Our work was enabled through the close cooperation with our rail industry partner BNSF Railway. We aim to continue this in the future to increase the likelihood that the technology will eventually be implemented in the rail industry. In addition, the results and methodologies will be presented in national conferences, published in national newsletters and journals for a broader outreach.
The final report is available.
