Development of a comprehensive classification model for railway track geometry condition severity based on both safety and ride quality

Abstract

The condition of the railway track geometry is an important determinant of both operational safety and ride quality. According to train accident report data, track geometry defects are among the leading causes of train derailments. Traditional track condition evaluation methods are primarily based on safety focused threshold exceedances of geometry parameters or ride quality focused track quality indices (TQIs). However, traditional TQIs exhibit shortcomings in that they either characterize safety or ride quality but not both. Yet, railroads need to address both safety and ride quality to ensure overall operational efficiency, maintenance condition and safety. This dissertation aims to address these limitations by proposing a novel Combined Track Quality and Safety Index (CTQSI) which provides a comprehensive framework for railway track condition assessment to include both safety and ride quality. ☐ In order to develop this CTQSI, a detailed preprocessing and exploratory analysis of over 180 miles of geometry data from a Class 4 U.S. freight track line was performed to identify key patterns, defects, and distribution characteristics. Exploratory Data Analysis (EDA) revealed meaningful spatial patterns, defect signatures, and threshold exceedances, forming a foundational understanding of the track’s baseline condition. Using the profile signal as a base case geometry parameter, detailed feature extraction tasks were undertaken to identify features that characterize safety and ride quality aspects of the track. These tasks were performed using statistical metrics, gradient-based indicators, rainflow cycle counting, and multilevel wavelet decomposition methods. The predictive performance of each extracted feature with respect to safety and ride quality criteria was evaluated using correlation analysis and feature importance assessments. ☐ An eighth of a car vehicle dynamic model was developed in MATLAB/Simulink to simulate the response of a typical 143-ton freight car under three representative operational speeds 40 mph, 60 mph, and 80 mph corresponding to FRA track classes 3, 4, and 5 respectively. Dynamic response variables, including car body acceleration and vertical wheel-rail force, were analyzed to derive safety and ride comfort indicators which served as the “ground truth” for this research. The root-mean square (RMS) of the car body acceleration (based on ISO 2631-1) was used to define ride quality, while vertical wheel unloading (per FRA, 49 CFR §213.333) was adopted as the safety criterion. These criteria were used to develop aggregate ride quality and safety models based on features extracted from the track geometry (profile) data. These models were then used to develop the Ride Quality Score (RQS) and Safety Score (SS) respectively. ☐ Two approaches were employed to develop the Ride Quality Score (RQS) and Safety Score (SS) models. The first was a machine learning-based method, in which a Random Forest regression model was trained to predict ride quality and safety criteria using the extracted geometry features. The second approach was a Weighted-Average Aggregation method, where features with high Feature Importance Scores (FIS) and strong correlation to the ground truth were combined using their FIS values as weighting factors. This dual-method strategy allowed for both data-driven learning and interpretable feature-based scoring. The RQS was strongly associated with the change in gradient of the profile (curvature) and low-frequency wavelet energy, while SS was primarily correlated with energy of mid-frequency detail coefficients (particularly at decomposition level 2) and slope of the irregularity. ☐ The RQS and SS were combined into a unified index, CTQSI, using a polar transformation framework. Two mathematical formulations were evaluated: (1) a rotated cosine-angle model, which allows directional weighting based on the relative contribution of safety and comfort, and (2) an elliptic contour model, which uses trigonometric scaling functions to shape the amplification based on the angle θ between the RQS-SS axes. The elliptic model was selected for its smooth contour properties, mathematical tractability, and its ability to generate risk zones that reflect realistic operational trade-offs. CTQSI contour maps were developed for each speed class, dividing the RQS-SS space into four risk zones: low-risk (safe zone), moderate-risk (alert zone), high-risk (danger zone), and critical (emergency zone). These zones were established by combining regulatory standards from the ISO and FRA with data-driven thresholds derived from segment distribution percentiles. ☐ Finally, the CTQSI was used to classify about 3,400 track segments across the three speed regimes. Comparative analyses revealed that at higher speeds larger number of segments fall into high-risk and critical zones. This is consistent with ‘real life” behavior. Segment-level analysis revealed that some segments with no defects exceeding regulatory thresholds could still pose safety risks at higher speeds due to sharp, localized geometry changes that may trigger dynamic unloading. Conversely, some segments with large but gradual undulations produced lower CTQSI values, as the long-wavelength nature of the defect results in less severe dynamic responses. Similarly, moderate profile irregularities may lead to elevated RQS without violating safety thresholds, highlighting how both the magnitude and frequency content of geometry deviations influence CTQSI classification.