A Deep LSTM-Sliding Window Model for Real-Time Monitoring of Railroad Conditions Using Distributed Acoustic Sensing (DAS)

Abstract Ensuring safety remains the paramount concern within railway systems. While extensive research has been conducted on multiple facets of train safety, the ongoing challenge lies in the real-time monitoring and timely detection of defects, including their occurrence, causes, and severity. Optical fiber cable has been proven to sense long-distance condition monitoring by using optical time domain reflectometry (OTDR). Distributed Acoustic Sensing (DAS) uses fiber optic cables along the track to detect any anomaly indicator such as vibration-based defective features. DAS systems can collect data over long distances. DAS became an excellent solution for real-time condition monitoring due to their high-speed data transmission capabilities, sensing certain mechanical and operational properties, such as strain, vibration, temperature, and pressure, which made the optical cables applicable for real-time structural condition monitoring. Conventional monitoring methods need time, physical inspection, scheduling, and higher financial involvement, while a knowledge-based method contains realtime monitoring, analysis, prediction, and classification of any occurrences at any remote distance from the monitoring station. The crack, rock fail, broken fail, flat wheel, rail condition, and track bed condition estimate from the DAS data require massive data analytics with an intelligence interface. Meanwhile, using sliding windows, machine learning, and deep learning tools, a data-driven intelligent method for fault detection has become an ideal technology among researchers. Fault diagnosis methods based on data-driven algorithms can identify failure types at the inspection site and provide a predictive failure plan to the maintenance team, which leads to increased safety, reliability, and profitability, as well as an improvement in the overall implementation of the data-driven predictive algorithms in the fields. This paper explores railway tracks’ structural health and condition monitoring using DAS data extracted from a High Tonnage Loop (HTL)-fiber optic bed In MxV Rail facilities (Pueblo, CO) and applying a Deep Long-Short Term Memory-Sliding Window (DLSTM-SW) model, that achieved a condition detection accuracy higher than 97% with a swift data processing time which led the model to the application of real-time monitoring of the remote condition of railroad. The findings of this article include automatically labeling each of the railroad’s distributed points or locations, defining the condition such as Defective Location (DL), Non-Defective Location (NDL), and Train Position (TP) along the fiber cable length of the railroad. In addition, as DAS generates a huge amount of data, it takes higher time for data processing, filtering, and feature extraction with traditional machine learning or deep learning methods, while the processed DLSTM-SW model takes only a few seconds which makes the model applicable for real-time monitoring.