Some Examples of Big Data in Railroad Engineering

2014 IEEE International Conference on Big Data Some Examples of Big Data in Railroad Engineering Allan M Zarembski Department of Civil and Environmental Engineering University of Delaware Newark, DE, USA Email: dramz@udel.edu Abstract The railroad industry is an infrastructure intensive industry that relies on significant amounts of information and data to operate and maintain each railroad. Using the US railroad industry as a model, this data collection encompasses the full range of railroad activities from tracking of goods shipments and car locations to managing train crews to inspecting and maintaining the infrastructure. This paper will look at this last area, inspection and maintaining the infrastructure and in particular the 330,000 km (200,000 miles) of railroad track in active use in the US.. Using a broad range of inspection vehicle to collect data and a new generation of maintenance management software systems to analyze and interpret this data, railroads represent an industry that is starting to make extensive use of its big data to optimize its capital infrastructure and safely manage its operations while keeping costs under control. This paper presents examples of collection, storage and use of big data in the railroad engineering environment. Keywords Railroad, track, track geometry, rail inspection, maintenance planning I. INTRODUCTION The railroad industry is an infrastructure intensive industry that relies on significant amounts of information and data to operate and maintain each railroad. Using the US railroad industry as a model, this data collection encompasses the full range of railroad activities from monitoring over 30,000, 000 car loads (shipments) per year, to managing the railroad fleet of over 1.3 Million rail cars and 24,000 locomotives to managing the infrastructure of over 330,000 km (200,000 miles) of track, which is owned and maintained by the railroads themselves. The US railroad industry s annual revenues are of the order of $60 Billion and their annual capital program is well over $15 Billion a year which includes replacements of such capital assets as rolling stock (cars and locomotives) and key track components such as rails, ties, ballast and bridges. Large scale data bases are used to manage everything from tracking of goods shipments and car locations to managing train crews to inspecting and maintaining the infrastructure. This paper will look at this last area, inspection and maintaining the infrastructure and in particular the 330,000 km (200,000 miles) of railroad track in active use in the US. Over the last half century, the inspection and management of the infrastructure has evolved from a subjective activity performed by a large labor force geographically distributed along the railroad lines, such as the tradition section gangs location every 30 to 50 km (20 to 30 miles) to an objective, technology active, data focused centrally managed activity. Using a broad range of inspection vehicle to collect data and a new generation of maintenance management software systems to analyze and interpret this data, railroads represent an industry that is starting to make extensive use of its big data to optimize its capital infrastructure and safely manage its operations while keeping costs under control. Noting that virtually all US railroads are private corporations, the ability to manage their infrastructure safely and cost-effectively is of great importance in maintain the financial viability and of the US railroads. That US railroad stocks are currently at all-time highs and the railroad industry is viewed by the investment community as an effective financial performer is testimony to the railroads growing ability to manage this data and use it effectively and efficiently. Noting that the six largest US railroads have over 75% of the tracks, with between 30,000 and 60,000+ km of track each, data management and analysis of big data has become of growing importance for these major railroads. This for example, railroads operate track geometry inspection vehicles, at track speeds of up to 130 kph (80 mph) for freight railroads and 200 kph (125 mph) for passenger railroads, with main line tracks being measured from one to up to 12 times a year. These measurement vehicles collect 10 to 12+ channel of data with a measurement taken as often as every foot. This represents well over 2,500,000,000 data measurements per year for just this one class of inspection vehicle. Large volumes of data are also collected from other vehicle based inspection systems such as: Ultrasonic rail test vehicles 978-1-4799-5666-1/14/$31.00 2014 IEEE 96

Rail wear inspection vehicles (laser wear measurement) Gauge restraint measurement vehicles Ballast profile and subsurface inspection vehicles (LIDAR and GPR) Tie (sleeper) inspection systems Dynamic load measurement systems As well as track based measurements of vehicle condition such as: Wheel load/impact detectors Lateral force detectors Overheated bearing detectors Dragging equipment detectors which record the data on a per car or per wheel basis. Thus on a busy mainline railroad, such a detector would see over 400,000 cars a year or over 3 Million wheels a year, taking a measurement on each. This data is then used at a number of levels: At the first level, basic threshold analyses are performed to determine if the measured value exceed a predefined threshold to include both maintenance and safety thresholds [1], [2]. At the second level, this data is entered into large data bases to allow for historical monitoring, trend analysis and first generation forecasting of rates of degradation or failure [3], [4], and [5]. At the third level, this data is used in state of the art statistical analyses such a multivariate regression analysis or Multivariate Adaptive Regressive Splines (MARS) analysis to develop higher order forecasting and trend analysis [1], [3], and [5]. At the next level, these forecasting models are combined with the large data bases in maintenance planning models for determination of maintenance requirements and scheduling of maintenance activities across these large networks [6], [7], [8], and [9]. These maintenance planning and management models often combine economic analyses with the projected failure analyses to calculate the optimum maintenance and replacement requirements for each local strength of territory on these large railroad systems [9], [10]. This data collection and analysis will be discussed in greater detail in this paper. II. EXAMPLE OF RAIL ENGINEERING DATA In order to better understand the scope of data involved, the inspection data from a major eastern US Class 1 railroad with approximately 36,000 km (22,000 miles) of track will be examined. Selecting several of the largest track based engineering data bases would include the following classes of data: Track geometry data Rail defect data Traffic and tonnage data Vertical Track Interaction (VTI) data This is an excerpt of the total set of track inspection data that is consolidated into a common data base using track location as the primary reference variable for correlation between different data bases and inspection data. Track location includes Railroad, Division and Subdivision, Mile Post (MP) 1 and track number 2. To give an idea of the scale of the data, for a railroad with approximately 36,000 km (22,000 miles) of track the following data is collected and stored in a data base. A. Track geometry data On board measurements every foot. Based on a system average frequency 1 inspection per km per year, this would represent over 100,000,000 measurements year with at least 12 channels of data collected at each measurement. Recorded exception data, stored in an active data base, represents approximately 70,000 measurements per year with at least 12 channels of data collected at each measurement B. Rail defect data On board measurements on a continuous basis. Based on a system average frequency 1 inspection per km per year, this would represent over 36,000 km of inspection data Recorded exception or defect data, stored in an active data base, represents approximately 20,000 data sets per year. C. Vertical Track Interaction (VTI) data Based on partial coverage of the network, represents approximately 200, 000 stored data records per year. Table 1 shows a breakdown of several years worth of exception/defect data (based on 5 years geometry exception data and 3 years rail defect data) for the full railroad system as well as by division, together with mileage and traffic density. Note traffic density is defined in terms of Millions of Gross Tons of traffic (MGT) where 1 MGT would be equivalent to approximately 15,000 railway cars per year. This data will be discussed in further detail below. 1 US railroads use the English system and use MilePost or MP as the primary location identifier. 2 Railroad routes can have 1,2,3 or even four tracks, so that the track number is part of the unique track ID. 97

TABLE I. SUMMARY OF GEOMETRY AND RAIL DEFECTS BY DIVISION AND FULL SYSTEM Divisions Length in Miles Annual MGT Reported Geo Defects Unique Geo Defects Rail Defects BA 1042.79 24.05 18351 11462 1543 Fig 1. Cross-Level defect AT 2042.1 25.68 27536 16044 2285 AL 1871.05 25.02 22738 14645 1774 2221.24 17.29 49616 26383 4786 CO l 2051.42 19.03 61677 39658 2374 CH 1630.73 18.65 16978 10053 1464 FL 3056.51 15.37 38785 22457 4019 Fig 2. Profile or Surface Defect GL 2326.16 32.8 18485 13202 2520 JA 2920.07 14.97 28308 15727 2272 LO 1406.79 17.24 31123 17435 1657 NA 1658.86 30.61 21340 15275 1746 Total full 22228 21.3 334937 202341 26440 Fig 3. Alignment defect III. TRACK GEOMETRY DATA Track geometry data refers to measurement of any variation in the geometry of the track, usually measured at the rail. Table II presents a brief description of the key track geometry defects. Defect Type Cross-level (XLEVEL) (Fig. 1.) Gage Profile or Surface (Fig. 2.) Warp Alignment (Fig. 3.) Elevation Cant Loaded gauge, GWP, GWR TABLE II. TRACK GEOMETRY EXCEPTIONS Defect Description A deviation in height of one rail with respect to the second rail at the same location Variation in the distance between the two rails (track gage) A variation in vertical track profile Warp is the difference between two cross-level defects in a specified interval A variation in lateral track profile over a specified interval A deviation in height of one rail with respect to the second rail at the same location on a curve. An unwanted rotation of the rail or superelevation on a curve Variation in the distance between the two rails (track gage) under a test load Noting that several of these measurements are taken per rail (thus two per track location) and several use different chord lengths (Warp 31 vs Warp62); the full set of geometry measurements included in this data set for every measurement point is shown in Table III. TABLE III. ONE COMPLETE SET OF TRACK GEOMETRY MEASUREMENTS (MEASURED EVERY FOOT) CROSSLEVEL EXECESS ELEVATION ALINGMENT LEFT ALINGMENT RIGHT WARP 31 WARP 62 CLIM GWR 2ND LEVEL PROFILE LEFT PROFILE RIGHT RIGHT RAIL CANT LEFT RAIL CANT LOADED GAGE PLG 24 2ND LEVEL WIDE GAGE 98

In addition, the following additional information is provided as shown in Table IV. Division, subdivision Location ( prefix, MP, offset in ft, track) Location ( GPS; lat-long) Date found Defect type Geometry car Class of track Reduced class of track (for defect) Length of defect Maximum value Curve/tangent Nearest event Freight speed Passenger speed Cant defect Priority Action date Action type (corrected, etc.) IV. RAIL DEFECT DATA Rail defect data, as recorded in the data base, is any defect detected, either with an inspection vehicle (usually ultrasonic test car) or visually by a track inspector (Track is inspected visually once or twice a week). Table IV provides a sample listing of the defect codes used in the data file. TABLE IV. DEFECT CODES V. TONNAGE DATA (ANNUAL MGT) This data consists of tonnage data by line segment to include all line segments system wide. The format is as shown in Table V. Div Sub Prefix TABLE V. TONNAGE DATA FORMAT Begin_ Milepost End_ Milepost Track Tonnage_ Measurement _D MGT CMG 1 2 1 01/14/2013 25.36 CMG 1 2 2 01/14/2013 20.49 CMG 2 2 1 01/14/2013 25.46 CMG 2 2 2 01/14/2013 20.68 CMG 2 2 1 01/14/2013 25.52 CMG 2 2 2 01/14/2013 20.80 CMG 2 3 1 01/14/2013 24.89 CMG 2 3 2 01/14/2013 20.70 CMG 3 5 1 01/14/2013 24.20 CMG 3 5 2 01/14/2013 20.52 CMG 5 5 1 01/14/2013 24.05 CMG 5 5 2 01/14/2013 20.59 CMG 5 7 1 01/14/2013 23.83 CMG 5 7 2 01/14/2013 20.60 CMG 7 8 1 01/14/2013 23.51 Description Transverse Fissure Compound Fissure Detail Fracture Engine Burn Fracture Defective Plant Weld In-Track Electric Flash Butt Weld Defective Field Weld Horizontal Split Head Vertical Split Head Split Web Piped Rail Head Web Separation Bolt Hole Crack Broken Base Defect Codes TDT TDC TDD EBF EFBW/OAW EFBW TW HSH VSH/VSH SW/SW PIPE/PIPE HW/HW BHB/BHB BB/BB VI. VERTICAL TRACK INTERACTION (VTI) DATA This data represents vehicle-track dynamic data as recorded by an inspection vehicle, in this case data collected using unmanned vehicle mounted Vertical Track Interaction measurement systems. The types of measurements include: Urgent Wheel Impact (AXV) Carbody Vertical (CBV) Carbody Lateral (CBL) Truck Lateral (TRL-RMS) Truck Lateral (TRL-PkPk) Mid-Chord Offset (MCO) This data is collected continuously (at a high digital sampling rate) and only values that exceed specific exception Near U 99

values set by the railroad are recorded. Table VI presents the specific data set column headers. TABLE VI. VTI EXCEPTION REPORT COLUMN HEADER Column Header Title Description A ID Unique exception identifier B SUB_DIV_NAME Subdivision of exception C LINECODE Linecode of exception within subdivision D GPS_SPEED From GPS reported in mph E DIRECTION From GPS reported in degrees F MPDECIMAL Closest MP to GPS location G CARID Locomotive number H TRAINSET Locomotive type I UNIT_ID V/TI Monitor ID J RUN_DATE Date of exception K LATITUDE GPS latitude L LONGITUDE GPS longitude M EXCEPTION_VA LUE Exception value N EXC_TYPE Exception type O SEVERITY Severity level (Priority, Near Urgent, or Urgent) VII. CROSS-TIES Cross-ties represent another area of Big Data in a railroad. Typically there are 1950 ties per km (3250 ties per mile), so that a railroad with 33,000 km of track (22,000 miles) would have over 70 million ties. These ties are usually inspected on a four to five year cycle, so that 15 to 17 Million ties per year are inspected as to their condition and whether they need replacement. The data collected by the inspection personnel or inspection vehicles is uploaded into the railroad system database and then used to determine required annual replacement ties by mile of track. Typically, a railroad of this size would replace 2 Million ties a year. VIII. DATA ANALYSIS As noted earlier, this data is entered into large data bases to allow for historical monitoring, trend analysis and forecasting of rates of degradation or failure. To illustrate this use, the rail defect data presented earlier combined with rail wear measurement data (often taken with rail profile measurement technology mounted on a geometry or other inspection car) can be used to forecast the rate of defect development and the correspond replacement requirements for rail. Information about the condition of the rail, to include the number, type, and location of internal defects, rate and degree of wear, extent and depth of surface defects, and deterioration of the rail head profile, are all key inputs, taken from the inspection processes and input into forecasting and planning models. In general, rail life is analyzed on a segment by segment basis with the definition of these segments based on the optimum length required for effective analysis and specific track and traffic conditions. These segments are assumed to be "homogeneous" in that the rail life as well as certain key parameters will be the same for the entire segment. Rail fatigue life forecasting algorithms utilize Weibull analysis techniques to predict defect growth rates and future defect levels. Studies have shown that rail develops fatigue defects as a function of the cumulative traffic that passes over the rail section as well as such factors as axle load, wheel-rail contact, and rail metallurgy and cleanliness. The rate of defect formation and accumulation with traffic has been shown to follow a Weibull distribution [3], [9], which is in the form of a logarithmic relationship as shown in Fig. 4. Thus, as the rail ages (cumulative MGT increases) the expected rate of defect occurrence increases significantly, corresponding to the logarithmic nature of the Weibull equation. 100

Fig. 4. Weibull Curve of Rail Defects Fig. 6. Consolidated Rail Requirement Forecast The output of the algorithms is an annual forecast of defect rate (defects/mile/year,) and cumulative defects for each segment, together with the forecast life of the rail. Fig. 5. presents this segment by segment rail life forecast in track chart form while Fig. 6 presents a 20 year replacement forecast (together with the rail installation history for that territory). The rail degradation cycles can be clearly seen from the relationship between installation and forecast replacement years. Fig. 7. presents a similar type of forecast for cross-ties. Again using a forecasting algorithm based on actual tie condition within the data base, the year in which a tie replacement gang is needed for each individual milepost is calculated and presented for a 200 mile long subdivision; based on a 1000 ties/mile replacement threshold [7], [8]. Fig. 7. Mile by Mile Tie replacement Forecast Fig. 5. Rail Life forecast by Segment (Track Chart) As the size and extent of the data bases continue to grow, more refined statistical analyses such a multivariate regression analysis or Multivariate Adaptive Regressive Splines (MARS) analysis are used to develop higher order forecasting and trend analysis [1]. This is illustrated in Fig. 8. and Fig. 9. which show a MARS application to geometry and rail defect data for a big data application, representing over 500,000 data records. 101

Fig. 8. MARS Analysis of Geometry and rail data to predict Rail defect development significantly more efficient method of managing the rail asset than the traditional rules based approach, because it takes into account the local differences in behavior and performance, as they effect the degradation of the rails. In addition, it allows for a more accurate planning and scheduling of rail maintenance activities, since the times and locations for the key production activities are more accurately known. In today s era of system optimization and cost sensitivity, the proper planning and management of track maintenance can be an important tool in assisting railways in controlling their capital and maintenance costs. This is particularly true for high capital cost items such as rail, ties, and ballast which represents a sizable portion of a railway's maintenance of way budget. As the cost of maintenance continues to increase, the ability to properly plan and execute track maintenance programs in an efficient and cost- effective manner becomes increasingly important. Fig 9. Sensitivity of Rail Life to geometry defects Finally, as noted previously, these forecasting models are combined with the large data bases in maintenance planning models for determination of maintenance requirements and scheduling of maintenance activities across these large networks [6], [7], [8] and [9]. The increasing use of inspection technology and the concurrent introduction of new technologies and associated analyses techniques provide railways with increasingly accurate and timely information about the condition of the track and its key components. This inspection data, which provides the railway with detailed information about the condition of the rail, forms the basis for improved planning, analysis and management based on the actual condition of the track. This use of condition based planning and management has been shown to be a REFERENCES [1] A. Zarembski, Integration of multiple inspection system data to identify potentially unsafe track rail conditions: data collection, consolidation and preparation, Report prepared for US Federal Railroad Administration, 2014. [2] A. Zarembski, Track Maintenance Planning Using Forecasting Software, Railway Technology International, 1988. [3] A. Zarembski, Forecasting of Track Component Lives and its Use in Track Maintenance Planning, in International Heavy Haul Association/Transportation Research Board Workshop, Vancouver, B.C., 1991. [4] A. Zarembski, Computerized Maintenance Planning and Reporting Systems, in Rail Transit 95: Design, Construction and Maintenance of Transit Track and Structures, New York, NY, June 1995 [5] A. Zarembski, Development and Implementation of Integrated Maintenance Planning Systems, in Transportation Research Board Annual Meeting, Washington, DC, 1998. [6] A. Zarembski and J. Palese, Application of Maintenance of Way Information Systems to Enhance Operations and Safety in Passenger Transport, American Public Transit Association, 1999. [7] A. Zarembski, L. Parker, J. Palese and C. Bonaventura, Computerized Tie Condition Inspection and Use of Tie Condition Data in Cross-Tie Maintenance Planning, in International Heavy Haul Conference, 2003. [8] A. Zarembski, L. Parker and J. Palese, Use of Comprehensive Tie Condition Data in Cross-Tie Maintenance Planning and Management on the BNSF, in American Railway Engineering Maintenance Association Annual Technical Conference, 2002. [9] A. Zarembski, T. Euston and J. Palese, Use of Track Component Life Prediction Models in Infrastructure Management, in AusRail Conference and Exhibition, Sydney, 2005. [10] A, Zarembski, Rail Maintenance Planning and Management Using State of the Art Maintenance Management Software, in International Railway Journal, 2010. 102