Predicting track geometry using machine-learning methods
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
The track is an essential and critical component of the railway system, playing a vital role in ensuring the safe and efficient transportation of people and goods. However, according to Federal Railroad Administration (FRA), between 2018 and 2021, there were 2097 train accidents ( human causes are excluded), and 781 of them were caused by track defects (37%). Maintaining and monitoring railway tracks are crucial for ensuring safe and efficient railway operations. Among various maintenance strategies, predictive maintenance is the most desirable approach due to its ability to schedule maintenance activities based on predicted defect occurrences. In this research, we aim to develop a model that accurately predicts track geometry irregularities, such as profile values, using geometry data alone, thereby enabling track engineers to anticipate maintenance needs promptly. ☐ Maintenance engineers rely on detailed information about the condition of railroad tracks to effectively plan and carry out maintenance activities like tamping and undercutting. In the past, maintenance scheduling was based on experience, manual inspections, field observations, and cyclical maintenance over a fixed period. Delaying maintenance can accelerate the rate of degradation of track geometry data, and the absence of maintenance activities may result in safety-related issues, while performing maintenance too early can be costly. To address this, data-driven models have been used to predict and track degradation rates based on geometry data. However, this proposed research takes a unique approach by combining mechanical models with data-driven models, specifically utilizing functional networks, Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM) networks. RNNs are well-suited for sequence problems, while LSTM networks excel in capturing long-term dependencies, making them suitable for predicting sequences in track degradation. ☐ The research will utilize historically collected railway track geometry data, which has been collected monthly over a period of 24 months. Exploratory Data Analysis (EDA) will be conducted to identify relationships and patterns within the data through visualizations and graphs. The data will then be processed, addressing missing values and removing unnecessary data to ensure data alignment. ☐ Once the exploratory data analysis is complete, the Track Quality Index (TQI) will identify previous maintenance interventions. This step is crucial as it accounts for the effects of maintenance activities, ensuring that the models consider the improvements made and do not seek correlations that do not exist. Subsequently, machine learning models will be generated and evaluated using the prepared data. ☐ The development of machine learning models/approaches accurately predicts localized growth of track irregularities, such as the profile value, using geometry data using functional networks and LSTM. The functional network model will incorporate domain knowledge and data to estimate neuron functions, capturing the complex dynamics associated with railway behavior. The research demonstrates the superior predictive accuracy of the functional network model compared to a neural network model. It highlights the model’s strengths in terms of accuracy, simplicity, and interpretability. The LSTM model offered several strengths. It effectively captured the underlying patterns and variations in the track geometry data, enabling accurate forecasts for maintenance planning and safe railway operations.
Description
Keywords
Track geometry, Machine-learning methods, Railway system, Maintenance needs