Drowning in data and seeking a solution
MAINTENANCE: Train-mounted measurement systems and condition monitoring equipment are generating large quantities of data, posing a challenge for railways to collate and interpret the information in a meaningful way. Software which can harness this data more effectively will help to inform infrastructure maintenance decisions.
Stuart Grassie and Martin Saxon
Railways are complex and disparate systems. The people who manage them frequently believe the popular mantra ‘what you don’t measure you can’t monitor’. Being conscientious stewards, and wishing to monitor and manage the assets for which they are responsible, they obtain what they believe to be necessary measurements of asset condition.
Unfortunately it seems that in many cases the collection of data is started without first thinking through how the measurements are to be organised and used. Sometimes it is not clear whether they are even being obtained in a format that allows the data to be manipulated and used routinely for the purpose intended.
Our experience suggests that this problem exists from almost the simplest format of data collection through to the most complicated; the accumulation of measurements becomes an end in itself.
Interestingly, the problem appears to be worse for smaller railways, such as metros. This may in part reflect the fact that larger operators have more resources to buy in the required expertise. This is certainly true when railways are purchasing a new measuring train; companies that measure track geometry increasingly have a suitable software platform that can accommodate a range of inputs, undertake trending analysis, and deliver graphic and informative displays. A typical example is the DataMap platform, a suite of software for the visualisation and assessment of all types of infrastructure information, developed by Balfour Beatty Rail Technologies.
It is also the case that pan-European research projects looking at track maintenance have understandably been directed more towards main line railways than metros, notably Ecotrack in the 1990s and the current Innotrack project. However, even ‘big’ railways can be negligent in the collection and use of data. Some track databases are directed more towards displaying information about individual work items than track condition; this is good for seeing what work needs to be done, but almost useless as a way of showing track condition concisely, informatively and graphically.
This article is based on some examples of the work undertaken by Railmeasurement Ltd, where we have helped to develop software to derive meaningful information from large volumes of data. One author’s expertise is primarily in railway engineering, whereas the other is an expert in software and systems development. In some projects we were involved from an early stage, and in others we were brought in later to help dig the hapless victims out from the mountains of data under which they had become buried.
From discussions and our own experience, we have helped to determine why something is being measured, what can be learnt from the measurements, how data can be presented in an intelligible format, and how the information gained can be used to improve the maintenance and management of the railway. This understanding of the requirements then leads on to the implementation of data management systems in a format that is not only attractive and useable but can also be undertaken to a tight budget. Feedback from the users in turn has led to modifications and further development.
We have also endeavoured to make the software as versatile as possible, so that it is more open to development for use on different railways or with different inputs, as the following examples will demonstrate.
Monitoring track condition
TrackMonitor was originally developed in a project for one particular metro operator, which was measuring rail wear in order to determine re-railing requirements and priorities. Although the metro expected a written report, we suggested that the results would be more comprehensible if presented on a schematic metro map, like the famous London Underground design now copied by almost every other metro. We also proposed a single simple equation for ‘equivalent wear’, which combined vertical and transverse measurements in a manner that was directly relevant to the metro’s wear limits.
Fig 1 gives a typical example. The red sections are those in which one or more measurements are outside the ‘must replace’ wear limit, yellow sections have at least one measurement outside the ‘plan to replace’ limit, while green sections have all measurements inside the wear limits. The summary at the left of the display shows that there is only 5·7 km of rail in the red and yellow categories out of a total of 241·5 km. So in total there is very little badly-worn rail, but unfortunately it is distributed throughout the network — with red or yellow categories showing in several sections of the map.
As well as the simple overview, more detail is available for each section. The next level of detail shows equivalent wear along a section (Fig 2), with summary graphs and figures. Even this level of detail can highlight interesting features to the inquisitive observer. There are two relatively new sections of rail towards the the bottom of the graph, shown by the blue dots indicating rail age. However, gauge wear is greater on these sections than on the significantly older, adjacent rails, whereas the crown wear is very much lower in line with the age. There is also relatively high gauge wear on both rails towards the upper end of the graph, and slightly high crown wear; might these be indicators of a tight gauge? If so, it would be better to fix the gauge than replace the rails prematurely.
A further level of detail brings up the full spreadsheet where wear measurements are recorded. It is thus possible to get an overall impression of the severity of wear on the network, see where resources should be directed, and also correct the basic geometry when the rail is replaced.
In this particular example, wear measurements were taken manually. However, this software could readily be extended to handle measurements taken from a moving vehicle, as with the RailData software described below. Other options that have been discussed but not fully implemented include display of historical data and the derivation of information such as wear rates, as well as an informative display of other permanent way data, such as switch & crossing condition, measurements of ballast condition using ground-penetrating radar, and the condition of the overhead line.
Localising the data
One of the authors was working with a metro to improve wheel/rail interface conditions, when it was mentioned that routine measurements of transverse rail profile had been made for several years, using a measurement train. This should have been a treasure of information on rail condition, but to date it had not proved as valuable as had been anticipated. For example, the measuring system produced a table of individual ‘exceptions’ to particular limits on rail wear, gauge, gauge face angle, and so on. But when the permanent way staff went to check the exceptions manually, they almost invariably found that there was nothing to see.
There had been justifiably high hopes for the profile measuring system, but confidence in it had plummeted. At least one member of the staff had been routinely making manual measurements along the line that would have been unnecessary if the system had given the information that it should have produced. Inadvertently, the manual measurements later provided a useful way of validating of measurements made from the train!
We were provided with measurements that had been taken over a period of a few years, essentially in a series of spreadsheets containing data on vertical and lateral wear, rail type, gauge face angle and so on, as functions of the distance along the line. But the measuring train seldom operated along exactly the same route, being regularly switched from one track to the other and back. This added to the challenge of sorting useful information from the pile of data.
It was clear that an early challenge would be to localise the measurements made at different times, and fix their position along the track sufficiently accurately that changes in condition could be detected from one date to the next. This is a problem experienced by anyone who has worked in this area. Our solution, which is adequate for a 40 km metro line but might require modification elsewhere, is to use the extremely repeatable characteristics of the measurements themselves to localise the data using a correlation technique. Some signal processing helps to remove spurious ‘noise’, but does not remove significant features of the measurements.
An example of the RailData analysis is shown in Fig 3, displaying vertical wear on both rails over a 200 m length of curved track, taken over a period of about 30 months. There are very repeatable characteristics in the wear pattern, and a steady increase in wear from the earliest measurements (black) to the most recent (blue and purple) taken three months apart. It was already clear from this work that staff might not find exceptions exactly where they were flagged up but perhaps to within 10 m either side; this is essentially a result of wear of the wheel on which the tachometer was mounted. And while there are inevitably spurious exceptions, these make more sense if the measured quantity is shown for some distance along the track.
Fig 3 also shows a pronounced periodic wear pattern with a wavelength of about 10 m, particularly on the left rail which is the inside rail of the curve. Although it had previously been known that periodic wear of this type was a problem, the simple graphic display showed that the problem was not only endemic over the whole line but in some sections was starting to occur within months of new rail being laid.
One objective of this project was to determine if these measurements provided information that could be used for track maintenance management — for example demonstrating areas of rapid rail wear which might arise from poor lubricator maintenance. This would give an incentive to continue taking regular measurements, and would also indicate what was required to process them into useful information.
Although in principle this may appear simple, in practice it is a considerable challenge, since the typical annual wear rates on this line are about 0·3 mm, which is within the accuracy levels quoted for all laser-based measuring systems of the type being used. It nevertheless appears that statistically-significant information can be obtained, even for a line on which the rails are wearing periodically.
Fig 4 shows vertical wear rates (in mm per year) as a function of distance along the same 200 m stretch of track as Fig 3. The wear rate is calculated as a difference between the first measurements and the three subsequent sets, divided by the appropriate time interval. There is clearly tremendous variation, some of which is ‘real’, some of which results from errors in the measuring system (now 0·6 mm for the difference in measurements), and some of which is ‘noise’.
Some sense can be made from an apparent jumble of scribbles by looking at the percentage exceedence curve, which shows the fraction of the record (on the y-axis) as a function of the wear rate (on the x-axis). Two of the three graphs have substantially identical statistical variation, with a median wear rate of about 0·45 mm per year. The other records have not only a significantly higher median wear rate (about 0·55 mm per year) but also a greater variation in wear rate through the curve. It was a pleasant surprise to find that statistically-significant differences in rail wear could be demonstrated which were a small fraction of the supposed accuracy of the measuring equipment.
What happened to reduce the vertical wear rate on this section of track? It is now too late to be certain, although it might have been possible to check if the software had been available when measurements were made. However, we do know that the curve was initially over-canted. Coincidentally, train speeds were increased between the second and third sets of measurements, so the vehicles were then curving closer to equilibrium speed. This tends to orient wheelsets more radial to a curve, which reduces wear. Although the speeds were increased for entirely different reasons, this work demonstrated that there was an unforeseen bonus entirely consistent with sound principles of vehicle/track interaction.
A possible development of this software would be to include track geometry measurements, in order to show the extent to which rail wear (particularly the periodic pattern shown in Fig 3) is exacerbated by poor geometry. This synthesis of information would contribute to understanding the root cause of a problem rather than its most obvious manifestation. In this case, for example, it might be more effective to improve the geometry than to maintain or renew rails that wear periodically as a result of vehicle dynamic behaviour excited by the poor geometry.
The addition of vehicle ride measurements would demonstrate the influence of geometry and rail wear on passenger comfort, and thereby help to prioritise maintenance on those aspects which would provide the greatest benefit to passengers (and potentially the greatest pay-off).
Aligning wheel and rail profiles
Measurements made using hand-held equipment exist at almost the opposite extreme from the never-ending flow of measurements produced by recording trains. The authors are not alone in having used excellent transverse-profile measuring equipment to take a series of measurements for rail grinding, and for routine monitoring of test sites or vehicle wheels.
Unfortunately the software provided with some instruments is oriented more towards ‘scientific’ than routine use. After some frustrations, we developed software that would automatically align up to 10 wheel or rail measurements with a reference profile, show the difference between the reference and all the measurements, and calculate the average difference. With a few clicks of the mouse, another reference profile can be selected and all measurements are automatically aligned again. What had been a time-consuming and often frustrating task became ‘a doddle’.
An example of alignment of several measured profiles of high rails, calculation of the deviation from the reference profile and of the average deviation is shown in Fig 5. This type of application is extremely useful for rail grinding in order to see average metal removal required to achieve a desired profile.
As with the other packages, this was achieved by combining an understanding of what was useful to the railway engineer (derived from both personal experience and discussions with others who had suffered similar frustrations) with an ability to write robust software that does exactly what is required.
Making the most of data
It is all too easy for those who maintain railway track and vehicles to become submerged in measurements of asset condition. In some cases, a little advance thought as to how the measurements are to be accessed, displayed intelligibly and used by those who need them, might identify ways of dealing with the data before it becomes unmanageable. It should be possible to identify those measurements from which most benefit will be gained so that the expense of gathering data of marginal value can be avoided.
Railways are not unique in this respect, and we do not claim that the approaches described here are unique solutions. But we have gained some experience in advising on data collection and designing software, in order to alleviate problems or prevent them happening.
The TrackMonitor software, which we initially developed as a readily-accessible means of viewing and using rail wear data, could readily be extended to become a general-purpose database for the permanent way. As with the Rail Data profile assessment system, an understanding of what would be valuable to the railway engineer helped to provide the means of gaining an insight into underlying problems. Although the software cost a fraction of the money spent on acquiring the data, the added value it provides has been enormous.
- Fig 1. The front end for the TrackMonitor database software displays rail condition on a simple schematic map.
- Fig 2. Detail of rail wear for a sample 200 m section of metro track.
- Fig 3. Vertical wear over a three-year period for a 200 m section of track, showing how the data can be localised by using features from the data sets themselves.
- Fig 4. Vertical wear rates (mm per year) corresponding to the wear measurements shown in Fig 3.
- Fig 5. Software that automatically aligns several measurements of rail profiles in a specified manner, finds the difference between the measurements and a reference, and calculates the average difference, can make a job possible that otherwise would have been at best frustrating.
We would like to thank the many railway and metro operators with whom we have worked and from whose problems we have learnt. None are mentioned by name lest some regard this as criticism. They should in fact be congratulated for identifying the value of what they had and seeking a means of realising this value.