Norfolk Southern installs "new" detection devices near the recent East Palestine derailment

[dehusman]

Euclid

Apparently, the system takes 1,000 photos of each passing railcar, and then “reads” those photos to find various problems. 

These type detectors on other railroads also have the "normal" detection sensors (hot journal, dragging equipment, wheel impact, etc) in addition to cameras or other visual sensors.  I would imagine that the NS's do also, they just weren't mentioned because it wasn't sexy new technology, so wasn't part of the press release.

[Euclid]

Apparently, the system takes 1,000 photos of each passing railcar, and then “reads” those photos to find various problems.  It is programed to know the normal conditions, and so it compares what is shown in the photos to what is normal. 

 

For instance to inspect for truck hunting, the system would analyze images of the trucks and look for locations of the truck details that would indicate that truck is wobbling in a rotational pattern about its axis on the truck center pin.  It would show a hunting truck to be abnormally changing its position as it passes the cameras. 

 

I doubt it would detect cracked wheels, but it could detect wheels with broken out material, including wheels with excessively flat spots on their treads. It would not hear the impact, but it could see the missing material on a condemnable flat spot.

 

For hot bearings, it could at least detect smoke or fire.  It also will easily detect dragging equipment.  That could be things like air hoses being low enough to touch grade crossing decking, or derailed-dragging wheelsets.  For critical defects like the latter, I would expect the system to issue an immediate alert to stop the train, and then the human subject-matter experts would address the issue. 

 

A large train could generate a quarter million photographs, and each photo could contain a very large number of target results.  I suppose the results would be presented in a range of charts and graphs from general overview to specific problems that call for varying degrees of action within varying future timeframes.

 

There are potential emergencies such as a derailed-dragging car, overheated bearing, or extreme wheel problem that the system would readily detect, but is unlikely to detect in time to avert a derailment or pileup; because too much time will pass before the train reaches the next detector tunnel. 

[dehusman]

Euclid

So this is only a visual inspection as the inspector in the link below is doing. 

Yes and no.

There are lots of things that a yard inspection won't catch because a yard inspection is done in a yard, with the car stationary and cold.

On the other hand there are lots of things that will only be caught with the car moving and there are defects that can happen after the car departs a yard (hot bearings, truck hunting, shifted loads, dragging equipment, etc.)

[Euclid]

Initially,  I had concluded that this new system was comprehensive sensing and detection with thermal, audio, and vibration detection for things like hot bearings, flat spots, broken wheels, cracked wheels, truck hunting, axle defects, overweight, etc.

 

But what this new inspection tunnel is actually doing, is only taking photos of the train.  So this is only a visual inspection as the inspector in the link below is doing.  The only difference is that the new tunnels will do it on the fly, and that will be much faster than the walking human inspector.   

 

However, I don’t see why the article talks about fewer false positives.  This type of inspection is not as urgent as detecting a suddenly developing running emergency such as a failing bearing, which can quickly cause a derailment.  So it is hard to understand how the human inspector can see what he believes to be a problem in need of attention, and have that decision turn out to be a “false positive.”  The walking human inspector is more likely to miss a problem rather than to see something he falsely believes to be a problem.  A walking human inspector missing a problems would be a “false negative.”

 

A false positive is more likely to result from an automatic hot bearing detector being calibrated to error on the side of catching every potential bearing issue versus on the side of preventing the unecessary stopping and delaying of trains.

 

Visual inspection by walking human inspector:

 

https://www.youtube.com/watch?v=96SZ0QleXjw

 

[Overmod]

CatFoodFlambe

All the technology in the world is useless if NS doesn't have enough employees to act on the information provided.

At least in this respect, NS is way ahead of the curve.  They have several levels of AI/ES running on the raw data and in any developing 'incidents' that might proceed to concern, and are actively developing more; there is nothing groundbreaking required to implement rapid and nuanced high-level reporting to dispatch and train crews -- likely far faster and with less confusion than routing 'big data' through some sort of incident desk so employees can pick up a phone.

Doesn't help the actual bad-bearing detection much, if at all -- but it does promise the necessary quick and positive response once they actually provide failing-bearing detection that works.

[CatFoodFlambe]

All the technology in the world is useless if NS doesn't have enough employees to act on the information provided.  

[dehusman]

Euclid

so I am wondering how it would have changed the outcome of the overheated bearing,

Probably not at all since visual band cameras won't detect roller bearing failures.  They will detect other stuff, but not failing bearings.

[Euclid]

dehusman

 

 

Euclid

Replacing those current detectors with more accurate and capable detectors, although spaced much further apart, seems to focus on a non-problem while ignoring the real problem of too much time and distance between detector inspections. 

 

 

Not all detectors detect the same things, they may detect different defects with different failure modes and different rates of development.  There are some defects that may be fine testing once a trip and there may be other defects that need to be tested every 10-20 miles.  For example wide load detectors are typically only installed ahead of restricted clearance points.

 

Yes, I understand that.  I am thinking mostly about the hot bearing detectors since NS is championing this new system and, and considering their experience with East Palestine putting their hot bearing detectors in the spotlight after failing to prevent the wreck at East Palestine.   

 

Now, they have introduced their new A.I. multi-function detector tunnels as being a powerful advancement in the wayside detection art, so I am wondering how it would have changed the outcome of the overheated bearing, and its presumed connection to the East Palestine wreck. 

 

It would be interesting if someone published a non-A.I. explanation of exactly what the new system detects, how it does so; and how it interacts with the “Subject Matter Experts” referred to in the news article. 

 

From those articles, I conclude that the new A.I. tunnel system merely hands the scan data to the Subject Matter Experts, and they alone will decide what action to take, if any.  So humans will decide according to the conflicting risks of acting on a false alarm versus causing a derailment by not acting.  I wonder if the A.I. will at least weigh those risks and rank them as a reference for the humans. 

[jeffhergert]

Currently defect detectors do not interact with PTC.  At most would be if a detector is linked to a hold signal that needs to be manually cleared if a defect is detected.  Other than that, PTC does not react to a detector alarm.

Not that it couldn't be integrated in the future. 

Jeff

[dehusman]

Euclid

Replacing those current detectors with more accurate and capable detectors, although spaced much further apart, seems to focus on a non-problem while ignoring the real problem of too much time and distance between detector inspections. 

Not all detectors detect the same things, they may detect different defects with different failure modes and different rates of development.  There are some defects that may be fine testing once a trip and there may be other defects that need to be tested every 10-20 miles.  For example wide load detectors are typically only installed ahead of restricted clearance points.

[dehusman]

blue streak 1

This question is not about flat spots that all posters have assumed. Ot can not be detected of a crack from axel outward to wheel tread.

Those are tested by ultrasonic detectors that "ping" the wheels and if they "ring" the wheel is good, if it doesn't the wheel may be cracked.  The UP N Platte yard is equipped with one of those that tests inbound coal trains.  It is affectionally known as the "crackhouse".

[Euclid]

What most people are talking about here is the status quo detector practice, with detectors placed at regular intervals of say 30 miles apart.  With the East Palestine wreck, the detector intervals have come to be questioned as to whether they should be shorter.

 

However this thread began with the topic of the new A.I. detector “tunnels” with greatly enhanced detection capabilities compared to the current state of the art.  But I doubt that these new detector tunnels are going to be placed at the suggested closer intervals than what is current practice of the historic detectors.

 

So, based on that observation, I wonder if the new A.I. tunnel detectors are missing the point of today’s detector shortcomings.  If the problem is detectors spaced too far apart, that means that they allow too much time for hot bearings and other impending failures to develop. 

 

Replacing those current detectors with more accurate and capable detectors, although spaced much further apart, seems to focus on a non-problem while ignoring the real problem of too much time and distance between detector inspections. 

[BaltACD]

blue streak 1

 

blue streak 1

One problem that am not aware is being addressed by all of these detection systems.  That is partial wheel failures.  Some wheels have a diameter up to 42 inches.  That means up to ~~ 131 inches of travel for one revolution. 36 inch wheel 113 inches of travel.  It does not appear that there is any visual system that can survey a wheel for any partial failure or loss of mass.  Am I missing something? 

 This question is not about flat spots that all posters have assumed. Ot can not be detected of a crack from axel outward to wheel tread. There can be fractures on the wheel itself that cannot be detected by a wild detector.  Although not as prevelant as flat spots there certainly have been instances.  

I believe the CSX super detector sites have high speed cameras looking at the axle wheel interface from their inside connection point - looking for cracks or other defects at that location of stress.  NS and the other carriers may have the same detection systems in place.

At present CSX has their 'super detector' sites on the main routes into Rice Yard at Waycross, GA - one of CSX's main car repair terminals as well as a high volume hump yard.  The super detector site also inspect the through trains that do not stop or work at Waycross.

[blue streak 1]

blue streak 1

One problem that am not aware is being addressed by all of these detection systems.  That is partial wheel failures.  Some wheels have a diameter up to 42 inches.  That means up to ~~ 131 inches of travel for one revolution. 36 inch wheel 113 inches of travel.  It does not appear that there is any visual system that can survey a wheel for any partial failure or loss of mass.  Am I missing something?

 This question is not about flat spots that all posters have assumed. Ot can not be detected of a crack from axel outward to wheel tread. There can be fractures on the wheel itself that cannot be detected by a wild detector.  Although not as prevelant as flat spots there certainly have been instances.  

[Overmod]

dehusman

Tree68

With PTC, I suppose the capability exists to enforce such a stop, assuming good communications all along the data path...

PTC is irrelevant.  Trains have been stopped by defect detectors automatically notifying the trains for the last 40 years.

Defect detectors haven't been 'notifying the trains'; they've been notifying the train crews.  What tree is saying is that PTC can physically enforce a detector indication with (presumably appropriate) train braking directly.

[dehusman]

tree68

With PTC, I suppose the capability exists to enforce such a stop, assuming good communications all along the data path

PTC is irrelevant.  Tains have been stopped by defect detectors automatically notifying the trains for the last 40 years.

[dehusman]

Euclid

QUESTION:  When the system finds a problem, does it make the decision to stop the train? 

Most of the newer systems I am familiar with the system tells the train to stop in most cases.   Older legacy systems may send results to a central desk and they tell the train to stop.

If so, does the system stop the train?

Yes.

If that is the case, do the humans decide whether the system command to stop the train is correct?

 

Yes to the extent that the stop and the results of the stop are reviewed to determine if the stop was valid or not (was a defect found).  If not then the the defect trigger conditions are validated to see if the they should be changed.

If the humans conclude that the stop command is incorrect, do they then intervene and overrule the system?

In the systems I am familiar with, yes, but not on a real time basis.

[jeffhergert]

These detector arrays may or may not give an immediate alert to stop.  I would guess they would if they detected a defect that required immediate attention.  However, I think these are designed to catch equipment that is in the early pase of failure, something that can make it to the next point that has car men.  That's assuming that they haven't cut off most of the car men and that someone doesn't try to "extend" the distance beyond the next repair point.

Broken wheels may set off a dragging equipment detector.  When opearting on concrete ties and a visual inspection for dragging equipment doesn't turn up a defect, an audible inspection is required.  The conductor places him/herself 20 axles ahead of the indicated axle, have the train pulled by 20 axles past the indicated axle, listening for any unusual sound from the wheels.

Jeff 

[tree68]

BaltACD

That is what Wheel Impact Load Detectors are for.

Anyone who's been trackside when a wheel with a flat spot is passing knows you can hear them from several cars away.

I've heard that an informal measure of the degree of a flat spot is if you can hear it from seven cars away, it's actionable.  You won't find that in any rule books.  The ultimate measure is a ruler.

One hundred thirty one inches is almost 11 feet.  An auto rack (a common place to find flat spots) is 80+ feet - which means that in one car length, that flat spot will hit 7 or 8 times.  

I'm sure the WILD detectors will sense such problems, even from the vantage point of a foot or so of actual detector.

As Balt points out, an out-of-balance wheel will also generate a certain amount of impact, especially at higher speeds.  It's up to the engineers (design types, not locomotive) to develop the impact parameters that require action - something they likely have already done.  

[BaltACD]

blue streak 1

One problem that am not aware is being addressed by all of these detection systems.  That is partial wheel failures.  Some wheels have a diameter up to 42 inches.  That means up to ~~ 131 inches of travel for one revolution. 36 inch wheel 113 inches of travel.  It does not appear that there is any visual system that can survey a wheel for any partial failure or loss of mass.  Am I missing something?

That is what Wheel Impact Load Detectors are for. A wheel that has lost mass will lose it from the rim/tread area and thus create a level of impact with every revolution of the wheel.  WILD's also detect slid flat wheels as their flat spots generate impact passing through the detector.

I am guessing that WILD's are a part of these 'super detector/inspection' sites.

[samfp1943]

blue streak 1

One problem that am not aware is being addressed by all of these detection systems.  That is partial wheel failures.  Some wheels have a diameter up to 42 inches.  That means up to ~~ 131 inches of travel for one revolution. 36 inch wheel 113 inches of travel.  It does not appear that there is any visual system that can survey a wheel for any partial failure or loss of mass.  Am I missing something?

 

  A calculation of process of 'action' for an 'ALERT of prioblem' might resemble in some manner the following:

Sensor delivers a' problem alert' > [ the signal delivers the alert to a 'point of notice']>  The ALERT is delivered to the Enginer of the affected train..              The Engineer makes an 'action' decision.  At that point,Physics of the envolved mechanisms take over, and train speed and time become key issues.  Train is stopped; with or without incident. (?)    

An oversimplied example(?) , but yoiu can get the idea.   

 

 

[blue streak 1]

One problem that am not aware is being addressed by all of these detection systems.  That is partial wheel failures.  Some wheels have a diameter up to 42 inches.  That means up to ~~ 131 inches of travel for one revolution. 36 inch wheel 113 inches of travel.  It does not appear that there is any visual system that can survey a wheel for any partial failure or loss of mass.  Am I missing something?

[Euclid]

The question of whether the new detector system stops the train if a problem is detected is not just a yes-or-no question.  The full “yes” answer might mean that the system finds a problem that requires an immediate stopping of the train.  So it notifies the team at the same time it stops the train.  Conceivably, it could stop the train just by triggering an automatic stop procedure. 

 

Or, a gray area answer to the question would be that the system could order the train to be stopped immediately by communicating that order to the team and the train crew.  Then they would have to react without questioning the order.

 

However, as I interpret the article in Progressive Rail and Railway Age, the new system will not stop the train if it is necessary, either by direct intervention forcing an automatic stop; or by a direct order to the team, which demands an immediate stop under the control of the train crew.  So this would be the full “No” answer to the yes-or-no question. 

 

Both of the two above-mentioned articles indicate that the new system will scan trains for a variety of problems and convey that scan data to humans who will evaluate it and decide the course of action.

 

According to both articles:

 

“The AI transmits the information to the NS Network Operations Center where the data is reviewed by subject-matter experts to identify and address issues. Critical defects found are flagged for immediate handling.”

 

So, the only question I see is the specificity of the required action that is implied to the subject-matter experts as the automatic system scans the train.  Depending on that specificity, it would seem that the subject-matter experts may not agree on what action should be taken if the AI data implies a problem that requires action.  They might also not agree with the data being provided by the automatic system.

 

Also, when the article says,“Critical defects found are flagged for immediate handling,” I assume that it is the subject matter experts that find and flag the critical defects, based on the raw data from the detection system. 

 

The articles imply that this new automatic system will reduce false positive detection conclusions to an all-time low rate.  This is critical because false positives delay trains.  I conclude that the historic potential for false positives is what drives the need for human review of the automatic warnings.  Such review will reduce actual false positives at times. However, it also motivates dismissing a true danger as being a false positive.   Doing that could lead to a derailment.

 

So, it seems to me that in the ideal world, the automatic system would be so perfect that there would be almost no possibility of generating a false positive warning of an equipment defect.  Then with so little chance of false positives, the system could be designed to make its stop decision and execute the stop without any human second guessing that could override that decision.  This would result in the fewest derailments and the fewest false positives. 

[jeffhergert]

If there is a question, we have to call the dispatcher who then contacts the bearing desk-the "back room" to check the detector.

But with going to talk on defect only, with some long trains at some detectors you might not hear a tone or exit message.  The detectors should give an indication that a detector, at least the full fledged defect detectors, has been tripped to the dispatcher.  They used to do so, but I'm not so sure they still do since they now have the "back room."

The old detectors on some lines that only talked on defect, no entrance or exit message, were tied into the signal system.  The system would hold a controlled signal at stop until manually cleared by the dispatcher if a defect or detector malfunction was detected.  There were many times when a detector failed to give a tone and announce a defect but the dispatcher would contact the train before clearing the hold signal.  That signal tie-in has been removed. 

Jeff  

[zugmann]

We can tone a repeat message on the detector.  

 

Or if that doesn't work, you tone up the detector helpdesk that will look it up for you. 

[BaltACD]

Paul of Covington

 

jeffhergert

And with 2 or 3 mile long trains there are a few detectors that are hard to hear any exit message. 

   Is there any way to send an inquiry to the detector such as requesting a repeat of last message?  Or can you contact dispatcher and get info?

Signal Maintainers can inquire through a Detector's history logs.  That I am aware of the History Log must be accessed On Site.

Train Dispatcher radio is configured that the dispatcher, if listening, is only listening through one of multiple radio transmitters on his territory.  These transmitters have an effective radius of 10 to 15 miles.  In addition to the Road Channel the Dispatcher radio consoles have access to the Dispatchers Channel and the MofW Channel.  A 100 mile long segment of railroad, like the RF&P Subdivison between Washington and Richmond will have between 8 and 10 radio transmitter locations.

Defect Detectors broadcast their reports over the Road Channel for the territory.  With longer trains, it may be necessary to increase the broadcast strength of the defect detectors radios.  I don't know what their broadcast strength is, however, I suspect is in the neighborhood of what the allowed strength for Citizen Band Radios is - which I think is on the order of 5 watts.

[Paul of Covington]

jeffhergert

And with 2 or 3 mile long trains there are a few detectors that are hard to hear any exit message.

   Is there any way to send an inquiry to the detector such as requesting a repeat of last message?  Or can you contact dispatcher and get info?

[tree68]

One example of a reason not to automatically stop a train - several trestles have burned because a car with a heat issue was stopped atop them.  Leaving the human element in the mix is important.

[jeffhergert]

We have combination defect detectors about every 20 +/- miles. These look for hot bearings, dragging equipment, and hot wheels. These detectors now only give an entrance message and then talk on defect only. 

We have dragging equipment detectors that talk on defect only at many intermediate block signals where there are concrete ties.

Hot wheels are the only alarm that may not require an immediate inspection.  It's allowed to run 30mph to the next detector.  A second alarm on the same axle requires inspection.  The car can't pass through a through truss bridge.  There are places because of this a train must stop and inspect before reaching the bridge.  Generally if it's the first detector after picking up cars, it's probably a missed handbrake.  If it's gone over a few detectors before an alarm, it might be sticking brakes.  Make a heavy reduction and release on the air brake and go 30 to the next detector.  Usually a hot wheel won't set off the defect tone, but alerts the bearing  desk, who alerts the dispatcher, who alerts us.

For the combination detectors, if a defect tone and message "defect detected" is given, we are to clear the detector, get the exit message indicating defect and location and stop and inspect. If there is the tone, but no other message, inspect the entire train.

For a dragging equipment detector that gives a defect tone, immediately stop and inspect.  If no specific axle is given, inspect the entire train.

For the detectors that only give an entrance message and talk on defect only, KEY trains must stop and inspect if no entrance message is received. 

I don't like the change to entrance/talk on defect only.  I feel it conditions crews to not hear any possible tone or message at the end of a train.  And with 2 or 3 mile long trains there are a few detectors that are hard to hear any exit message.

Jeff