Deep Engineering Dives

Written by Allan Zarembski, Ph.D., P.E. FASME, Hon. Mbr. AREMA, Professor of Practice and Director, Railroad Engineering and Safety Program, Department of Civil and Environmental Engineering, University of Delaware
Figure 2: Use of AI for crossing gate status[3].

Figure 2: Use of AI for crossing gate status[3].

RAILWAY AGE, MARCH 2021 ISSUE (expanded version): The railroad industry is using data science to better understand and maintain right-of-way performance. Here's how.

As noted in the February 2021 Railway Age article[1], the University of Delaware’s annual “Big Data in Railroad Maintenance Planning” conference provides a forum for railroad and data analytics professionals to come together with academia and discuss the latest applications and research in railway-related data science. Clearly evident from these annual conferences, there is growing use of data analytics, often referred to as “Big Data,” to address maintenance and safety issues in all aspects of railroading: Engineering (Track and Structures), Equipment (Rolling Stock) and Transportation (Operations). This article will continue that discussion of the December 2020 conference, focusing on the Track and Right-of-Way-related applications of data analytics to the issues of railroad safety and maintenance. Although held in a virtual format, the conference—organized by Drs. Allan M. Zarembski, Nii Attoh-Okine and Joseph Palese—was able to attract more than 270 attendees from the full spectrum of railroad Big Data-related activities and research.

One major area of safety focus that was discussed in several of the presentations was grade crossing safety. Not surprising, grade crossing safety is an ongoing concern in the railroad industry, and a range of research efforts are making use of Artificial Intelligence (AI) and Machine Learning (ML) tools to address it. The presentation by the Federal Railroad Administration (FRA) showed a range of programs such as the use of a combination of vision systems/cameras and AI-driven recognition software to identify vehicles in grade crossings and trespassers on the right-of-way. Another FRA program, labeled i-CATTS for integrated Crossing Assessment & Traffic Sharing System, is aimed at developing an affordable and field-deployable system using AI to provide real-time traffic information and estimate grade crossing delay time due to train blockage (see Figure 1, below)[2]. This information can then be shared with motorists and first-responders. A related program uses AI processing to evaluate grade crossing status, such as illustrated in Figure 2 (above)[3] .

Figure 1: Distribution of vehicular traffic at grade crossings [2].

In the area of inspection technologies, data analytics is helping to improve the accuracy, precision and effectiveness of inspection systems. This was likewise seen in several presentations. For example, ML, in the form of Deep Neural Networks (DNNs), is being used to improve the accuracy of rail testing, specifically to improve the detection of internal rail defects as well as their location, as seen in Figure 3 (below)[4].

Figure 3: Neural Network application for improved rail testing[4].

In another application, DNNs are being used to identify changes in images from cameras and other vision systems, to allow for the identification of missing or degraded track components. Such a DNN is illustrated in Figure 4A (below), where multiple layers of the DNN allow for improved analysis capability as well as the ability of the system to “train itself” and “learn” from the input data[5].

Figure 4A: LRAIL Deep Neural Network structure [5].

Figure 4B (below) shows how this is then applied to the problem of identifying missing or failed fasteners.

Figure 4B: LRAIL Deep Neural Network application [5].

Another area that has seen significant influence of improved data analysis, as applied to large data sets (i.e., Big Data), is that of maintenance forecasting and prediction of component failure. It has been the subject of numerous presentations in both this and earlier Big Data conferences, often showing different and increasingly effective applications of data analytics tools and approaches. This is the case illustrated in Figure 5 (below), where track geometry degradation is modeled and forecast using a data analytics technique referred to as Auto Regressive Integrated Moving Average or ARIMA[6]. The 2019 Big Data conference showed the use of the same technique to forecast rail wear[7], thus showing the versatility of these data analytics tools and techniques, as well as their increased application to real problems.

Figure 5: Use of ARIMA for track geometry degradation forecasting [6].

Broken rail prediction, again using a Neural Network approach was the topic of another presentation[8]. As can be seen in Figure 6 (below), a specific form of Neural Network, the Soft-Tile Coding-Based Neural Network or STC-NN, is used to develop a model to predict, on a probabilistic basis, when and where broken rails will occur. The resulting AI model was able to “catch” more than 71% of broken rails on a 20,000-mile rail network.

Figure 6: Broken rail prediction methodology [8].

Data analytics was also used to integrate the inspection, analysis, forecasting and maintenance planning process for an application of an axle-box mounted accelerometer (ABA) inspection system, as shown in Figure 7[9]. The result was a condition-based maintenance decision system for rail surface condition as applied on a regional railway in Romania. The specific approach shown in Figure 7 is as follows:

Figure 7: inspection, analysis, forecasting and maintenance planning process for application of an axle-box mounted accelerometer (ABA) inspection system.

• Data collection from an onboard ABA.
• Signal processing of raw data.
• Use of Hilbert spectrum analysis to identify defect signatures based on the axle-box acceleration measurement. 
• Degradation model used to identify different degradation scenarios over time. 
• Use of multi-objective optimization process to find the trade-offs between the stochastic degradation scenarios and the number of interventions (replacements or maintenance). 
• Use of Pareto analysis for maintenance decisions.

The result in an integrated process which allow for the identification of where the highest increase of performance can be achieved, while keeping control of the budget for each intervention. This process and the resulting maintenance decision parameters are presented in Figure 8 for a segment of the study line with a corrugation condition. Based on this process, 70 rail segments were identified as requiring replacement.

Figure 8: Segment of the study line with a corrugation condition.

As noted, the range of presentations went from highly theoretical to quite practical. In addition to the applications and theoretical modeling, other papers addressed such issues as data quality and the potential source of error from such factors as Covariate Shift in track geometry data[10]. These issues focus on data quality measure,  which plays an important and significant role in determining the performance of an ML model. Thus, Covariate Shift, which is defined as a change in the distribution of the independent variables, can adversely affect the development and performance of such an ML model.

The University of Delaware’s Big Data in Railroad Maintenance Conference continues to provide a spotlight on and a venue for learning about the growing application of data analytics, ML and AI in the rail and transit industry. 

The next Big Data conference is scheduled for December 15-16, 2021. 

For further information, contact Professor Allan M. Zarembski at [email protected]


1. Zarembski, A. M., “Using Data Science to Better Understand and Maintain Rolling Stock Performance,” Railway Age, February 2021.

2. Jay Baillargeon and Francesco Bedini Jacobini, Federal Railroad Administration, “Leveraging Big Data to Advance Grade Crossing Safety & Trespasser Prevention,” 2020 Big Data in Railroad Maintenance Planning Conference.

3. Larry Jordon, Wi-Tronix, “IoT for Rail and Case Study of Applied Video Analytics for Maintenance in Rail,” 2020 Big Data in Railroad Maintenance Planning Conference.

4. Bobby Gilbert, Sperry Rail Service, “An Introduction to Elmer®: Artificial Intelligence and Big Data for Rail Inspection,” 2020 Big Data in Railroad Maintenance Planning Conference.

5. John Laurent, Pavemetrics, “Use of 3D Laser Scanning and Artificial Intelligence to Detect Changes in Track Condition,” 2020 Big Data in Railroad Maintenance Planning Conference.

6. Serkan Sandikcioglu, ENSCO, Jay Baillargeon, FRA, Jackie Van Der Westhuizen and Radim Bruzek, ENSCO, “Rail Geometry Predictive Analytics with Time Series Models,” 2020 Big Data in Railroad Maintenance Planning Conference.

7. Palese, J. W., “Application of Data Analytics to Rail Wear Forecasting,” 2020 Big Data in Railroad Maintenance Planning Conference.

8. Xiang Liu, Associate Professor, Rutgers University, “Artificial Intelligence-Aided Broken Rail Prediction on Freight Railroads,” 2020 Big Data in Railroad Maintenance Planning Conference.

9. Alfredo Núñez, Delft University, Netherlands, “Evolutionary Multi-objective Optimization for Maintenance of Rail in a Regional Railway Network,” 2020 Big Data in Railroad Maintenance Planning Conference.

10. Nii Attoh Okine, University of Delaware, “Covariate Shift Problems in Track Geometry Data,” 2020 Big Data in Railroad Maintenance Planning Conference. 

Categories: Analytics, Class I, Freight, M/W, News, Short Lines & Regionals, Switching & Terminal Tags: , ,