Location and positioning in the railway (part 3)

Other sources of error

The errors in GNSS/INS measurement are not the only source of positional error - a set of Lat/Long coordinates in a certain datum is probably no use to anyone, these need either attribution against a network model (to get some kind of engineering geography reference and linear measure) or a separate device to interpret them (such as a handheld GNSS device to direct maintenance workers to their target). If we start with the latter, we send a maintenance worker out with a pair of coordinates alone to fix a fault - will they be able to find it? This ultimately depends on what other information they have, the accuracy of the point we give them and the type of fault:

1. A missing fastener on a piece of single track (i.e., one line only) in a nice open sky environment - the worker will be able to get near the fault using a handheld device and find it by walking up and down the track - the fault will be visible.
   Yes

2. A top defect on a 4-track railway in a nice open sky environment - this is where it becomes tricky, armed with a long/lat (with an error), the worker can use their handheld device to position themselves nearby - with a further error from that device. The worker realistically has no idea which line the fault could be on and so may have to walk up and down each track to spot the defect (assuming they can get access). If the defect is severe enough, this may be visible while standing on the other tracks.
   Maybe

3. A minor gauge fault in a tunnel with two lines - there is no source of GNSS inside the tunnel and low levels of lighting, the worker will find it harder to navigate to the defect, which may be hard to see. Here the worker really needs more information, some kind of linear measure (metres from the start of the tunnel for example). Unlikely

4. An ultrasonic defect in an urban canyon area of railway with multiple tracks - because of interference such as multipath and/or possible lack of a GNSS signal at all, the error in the handheld device position will be large. Coupled with the error in the measurement position, realistically the worker needs more information (or a long time on site) to find the defect as it is most likely not visible with the human eye and requires test equipment. The worker would have to test a large area of track to find the defect meaning it would be time prohibitive.
   No

So, if the worker had information about which track the defect was on, then they would have a better chance in the last 3 of these scenarios and in all scenarios some kind of linear measure would be advantageous so that in the event of a poor position from the handheld device, they can orient themselves locally (e.g. from the start of tunnel or a set of points etc.). There is of course still a chance that they will fail to find the defect within the time allotted. So how do we know which track the defect was recorded on?

Map matching

Map Matching is the process of calculating the path traversed by a train through a network model, this could be done by finding the nearest line to the long/lat - but we already know that we have a potential positional error in the measurement point, the nearest track might not be the right one. We could also have not just a positional error in the network model (i.e., track geography incorrect or an offset), but we could also have topological errors (i.e., tracks ordered incorrectly, missing tracks, additional tracks), as such, we must be careful as to how that attribution is done. When doing this with IM data, sophisticated algorithms are employed, these don’t just return the closest line, but they return the best fit path through a section in time - which is a much better measure. But of course, this could still be wrong, how do we tell? The answer is - we don’t. The maintenance worker (potentially) goes out and can’t find the fault, it isn’t fixed - worse still a speed restriction could be applied to the wrong line (and NOT applied to the right one).

The point is that whatever happens with IM data currently - some of the results are just plain wrong. This leads to an additional question in choosing the positioning system - what percentage of time is it acceptable for the position to be incorrect? This is the fundamental part that is missing in standards and spec sheets today. Yes, system X can give you 1cm + 1PPM in clear open sky, yes, the drift of the data in the absence of GNSS can be 10cm over 1km (given certain criteria and caveats - not least the speed of the train) - but what percentage of the data points produced by the overall system have the incorrect attribution? Currently, the latter is a number that is relatively large - we are probably talking somewhere between 1 and 10 % for the better systems. This may seem like a small number, but for an IM shift covering 200 miles, this means between 2 and 20 miles of the track has an ambiguous (or worse - incorrect) status.

The answer to the question - what percentage of time is it acceptable for the position to be incorrect, is quite easy to write. The answer is - given sufficient measurement runs within the maximum allotted time frame for recording, provided that at least one of these measurement runs is correct - then every piece of track must have a valid measurement (and therefore position). Basically, the percentage of time answer is zero - or we must introduce counter measures (such as speed restrictions etc.).

So, what is the mitigation for this? Record it again, and again, and again and hope that it is right at some point? So how do you detect whenthe position is wrong? This is done by user validation, so not only are multiple redundant shifts being run, but users trawl through positional (and measurement) data to validate it. This is of course starting to sound like we are no longer in the age of AI!

As we have said, the errors in GNSS depend upon a number of factors - not least satellite visibility, they also depend on environmental factors (ionospheric events for example) as well as the type of equipment used (and the maintenance of that equipment) and can even depend on the time of day. All of this leads to a realisation that the worst part of the error with a GNSS based system is that the errors are not predictable. Seen above is a typical histogram of the error spread of a typical GNSS receiver, as we can see the error varies and is not well distributed.