Welcome to the Pipelineers Podcast, episode 318, sponsored by InterSYS Corporation, providers and poems, the Pipeline Operations Excellence Management System, compliance and operations software for the Pipeline Control Center, to address control remandagement, scale and audit readiness. Find out more about poems at intersyscorp.com. We give away a customized Yeti Tumblr to one listener every episode. This week, our winner is Desi Orias with Vincaro Mid-String.
Congratulations, you're Yeti's on its way, to learn how you can win this signature prize and stick around to the end of the episode. This week, we're going to speak with Stuart Mitchell from PIPES Sense and talk about direct leak detection using pressure events. Stuart, welcome to the Pipelineers Podcast. Hi, Russell.
Good to be here. Looking forward to the discussion we're going to have. Yeah, absolutely. So before we dive in, if you would, tell us a little bit about who you are and your background and how you came to be working in leak detection.
Okay, yeah, sure. I'm Stuart Mitchell. My background, I guess, I've been working now 27 years in the oil and gas industry, mostly focused in the area of research and development. I've done a lot of work both onshore and offshore in this area.
So how I got into leak detection was that I've always been interested in developing new technologies and having spent most of my time offshore, a really interesting onshore pipeline tech gamer around leak detection. So having just next to one business and being a serial entrepreneur, I started up another business to go and do pipeline leak detection. And that's how we got here. Yeah, my grandpa called that serial entrepid northing being a businessman.
Oh, really? Okay. That's just a fancy name we made up feel better about it. Yeah, it makes us all feel better.
Anyways, look, I want to talk to you about your tech and what you're doing. So tell us a little bit about the company that you're with and just quick overview of your technology, what you're doing. Sure. Yeah, the company that I'm with now is called PIPES Sense and PIPES Sense, we really started the business here to help people enhance their pipeline performance through data analysis.
Part of that technology that we apply that data analysis to is pipeline leak detection. So our technology really has evolved through a couple of stages, key stages. So for those people out there that are not hugely familiar with pipeline leak detection, the first technique that we applied was called negative pressure weight. So just as a real quick explanation of that, think of sonar within a pipeline.
When you get a pipeline leak, you get a pulse in the pressure within the pipeline. And that pulse travels up and downstream and can be detected by pressure sensors. So that's the basic technology. Generally, there have been problems about using negative pressure weight in the industry.
So it's been around for a while. But the main problems with it are that if you want to pick up smaller leaks, you're very, very prone to having false positives. And false positives are bad because you're effectively telling the operator they've got a pipeline leak when they have it. So what we looked to do was try and first of all enhance that technique.
Couple of ways we do that. Do you mind if I get into that Russell now? Yeah, no, please. This is great.
Just keep going. I'll jump in when I have a question. Sure, no problem. So yeah, the first of those that we do is through the techniques is through applying a very high sample rate.
So we sample pressure on the pipeline at about a thousand samples a second. And what that enables us to do is get a very detailed, very granular view of any pressure events on the pipeline. So once we detect a pressure event on the pipeline, what we've got to establish is that really a leak event or is it just normal pipeline operation or something else that's going on. So the couple of ways that we do that is once we think we have pipeline leak, we then look at accuracy and timing between pipeline sensors.
So our technology, let's say we put sensors on the pipeline that are looking at pressure every 10 to 20 miles. In order to determine whether it's actually a leak event or it makes sense for a leak event, we do two things. We look for location of accuracy of the predicted leak. So by being able to accurately predict leak location, and we are accurate to be about 20 to 50 feet, we can actually check on the pipeline the pipeline map whether that location of a leak makes sense compared to what else is going on in the pipeline.
The second thing we do is we check timing. So a leak event typically will pick up on more than one field processing unit, one of our pieces of hardware on the pipeline. And what we'll do is we'll check the timing of the signal between multiple FBUs. And if that progression makes sense in terms of timing and timing is related to location, then we've got more confidence it's an actual leak event.
What this technique doesn't do is get away from pressure events on the pipeline that look very, very similar to leak events. Let's say you rapidly shut a valve or you rapidly open a valve and you get a very quick transient pressure change. That can still confuse a negative pressure wave system. So what we then looked at doing Russell's moving on from there and applying machine learning to the problem.
So what we're actually doing as I said is gathering an awful lot of data, so 1000 samples a second in real time on the pipeline in multiple places. So what we then do for a new technique to try and improve accuracy and do away with false positives is to run an ML technique, machine learning technique called convolutional neural networks. And what they're basically doing is that they're processing the data. And in this case we're looking for, to be able to identify the visualization of a leak event.
So we take a pressure reading, we turn that into an instantaneous frequency plot and then we look at that plot using convolutional neural networks. And that we do most of that processing for that part of it, what we call it on the edge. So edge computing, which is taking the computing away from the cloud and putting it actually location at the location of the pipeline. So once we've done that identification using the CNN locally at the pipeline, what we've determined is we've got basically an event, something that's different from normal pipeline operation.
We then take that data when we've identified the normal pipeline operation, we send that off to the cloud. So once we've gathered a number of different sets of data from individual field processing units on the pipeline, we then do some further analysis in the cloud. And what we're trying to do now is we're taking what are just pressure events on the pipeline, and that could be a pipeline leak, but it also could be a pressure surge. It could be a control valve that's faulty and chattering on the pipeline.
It could be any number of pipeline events that could have happened and we've detected. So then we use secondary processing in the cloud to try and classify those. And what we're doing is we're comparing different types of pressure events against a training set. So in our case, we've gathered something around 25 to 30,000 pressure sets, pressure training sets.
And those events, we then can compare to what we've just received from the pipeline when an event has been detected. So let me kind of jump in here, Stuart, because there's a lot of information for somebody listening to kind of the grab and process. And certainly we have some listeners that are, you know, leak detection guys that are familiar with negative pressure waves. So there's a couple of things that you guys are doing that's different than what I have heard from others.
One of them is, I mean, the first off, what's common in negative pressure wave is very high, simple rates of pressure and looking for specific kinds of signatures. That's common. One of the things I've seen some of the technologies do is they require two sensors and they compare the data between the two sensors to find a leak location. What I what I think you guys are doing that's unique.
And correct me if I don't have this right, one of the things you're doing is you don't require two signatures, you just require one. The other thing you're doing that's unique is you're combining edge processing to kind of identify the data sample of interest, if you will. And then the data sample of interest gets passed to the cloud and then you do further analysis in the cloud to eliminate the false positives. So one of the other things you're doing that's unique is deliberately eliminating false positives.
That's correct. And if we take, if we take the second part of it first, because I think that's that's the most important part of the technique to what we're doing here. As I said, we've moved from just purely negative pressure wave, which is just looking at you as you say, a high sample rate, a specific type of signature when you apply different filters to the data, that signature still stands out. So that's traditional negative pressure wave.
What we've moved on to do is to take those leak events, the pressure events on the pipeline, where the leak events are otherwise and turn those into instantaneous frequency plots. And by doing that and sending that off to the cloud, we can then use cloud processing, a lot of horsepower there, to really look through a large data set of different pressure events. So it enables us to, my favorite phrase is pick the wood out from the trees. And what we're trying to do here is really determine that we definitely have a leak event, because the signatures of a leak event when you use this type of machine learning technique are very, very different from anything else on the pipeline.
Yeah, you know, it's interesting because I've done, I've done podcasts on edge processing. And one of the things that, you know, it's interesting about edge processing is you can deal with very large data sets, which allows you to do a certain kind of analysis. The downside of that is if you push all that data to the cloud, then it very quickly becomes uneconomic. Absolutely.
Yeah. So the ability to kind of understand this is what I do at the edge, and this is what I do at the cloud, and this is how I move the data around and how I manage that data. That's pretty interesting and intriguing to me because that's something I've been talking about for a while. There's only two places where I've seen that done where it makes a lot of sense.
And this is the second one I've heard of. The other is in image processing, where you're taking very large data sets of images and then performing analysis in the cloud to find the data sets of interest. Yeah. And so we're, I mean, we're effectively doing the same.
So it's a great analogy to use their great comparison Russell because effectively we are doing analysis of images. We're turning the pipeline event effectively into an image. And we're doing that image processing or the turning of that event into an image at the edge, so we're using edge processing to do that. So then what that enables us to do is we're not kicking huge amounts of data to the cloud.
We're only passing data to the cloud, a small data set, when we get a pressure event. So we still don't know what that pressure event is, but we're still only passing data when we see a pressure event. So that's much more, much more uncommon. It takes off the load of sending 1000 samples of second, for instance, to the cloud.
The real weakness, Russell, if you don't do that and you don't do your pressure, you're processing at the edges, you come, you become very susceptible for data dropouts and loss of communication. So for a system where let's say you're doing all of your processing remotely or you're relying on interaction between units to do confirmation, then if one of those units drops out of comms or you cannot send that data for whatever reason to the cloud or common locations to comparison or processing, then you're effectively blind. Your system is down. But the duration that your comms are down, your leak detection is down.
Well, yeah, and the way that that's typically mitigated is you put on the remote device a very accurate clock and you're time syncing all of that data. But that is from a processing standpoint, that's burdensome. Yeah, it is. But actually, intriguingly Russell, we've managed to do that.
And maintain the ability to still process that data at the edge. So in fact, all of our data is time stamped as well. And the way that we do that is individual units uses a GPS time server. So we're connected through all of the units that commonly connected to the same GPS satellite.
So picking up the same time signature and everything is very, very, very accurately time stamped. Now, Well, for those of us that are knowledgeable of those details that you say that you make that sound easy and I know that's not easy. Yeah, it was a particularly complex process in developing the hardware. Yeah, the time syncing multiple devices to a millisecond and making sure that they don't drift by very much is not trivial.
Yeah, so I mean, we could get into the weeds here, but it involves a lot around managing data interrupts and processing interrupts in the correct ways that you do not get, you do not stop the ability to influx intake data and process it whilst you're searching for time stamps and time stamps at the same time. But it's something that was. Yeah, I just want to point that out. My peers that work in that they'll name know exactly what's involved in doing that.
Yeah, so as you can imagine in developing this is Russell, there was an awful lot of testing, a degree of frustration and a lot of going back and scratching our heads and working things out in the initial term, but we managed to work it out now. I mean, that's that system that hardware and software system without the machine learning. So the more traditional negative pressure wave are enhanced version of that that's been running now using that technique reliably for some three years. So that whole part of that whole process we're extremely comfortable with really moving on to the machine learning was A to get rid of false positives.
So really move towards zero false positives, a true zero on pipeline lead detection systems. But the second thing is, and the interesting thing is you've got so much data, you've got the ability to take that data and really break that down into different types of events on the pipeline. So really anything. Well, before we go there, Stuart, there's another question I want to kind of follow in the lead detection because the other thing that comes up for me when I hear this is, okay, so I'm doing the stuff I'm doing in the field.
I'm identifying a signature for analysis. I'm sending that signature for analysis to the cloud. The cloud is performing that analysis and then I've got to get my alarm back to the control center. Correct.
And all of that has to happen in a time frame. So how are you managing that kind of, if you will, the back end part of that process? So the back we built, a pretty flexible back end. So the first thing is that our system is initially cloud deployable.
So in its base form, it's a cloud deployed system that has a web interface. So typically, the sort of thing that you can pop up on your laptop or on your phone and have an interface there that will tell you basically that everything's functioning, all the units on the pipeline are functioning, okay? And if you've got a leak event, so we keep it very simple. Everything's functioning, okay?
You've got a problem or X marks the spot with a GPS location for a leak event. Now, how we ensure that that's communicated to the operator effectively in a timely fashion? First of all, our leak detection process end to end is, I would call it near real time. It's as near as you're going to get for leak detection typically within two to three minutes, or able to inform the operator for our system.
Now, people get heartburned about deploying systems on an outside cloud. Some of the pipeline operators rightly are wearing cyber security. So we developed the system so we can deploy that in our cloud, their cloud instance, and we can even deploy it locally on a server boarder within their own building. So that's kind of the mechanics of it.
We can take any of those methods and connect them to the control room using a straightforward OPC UA connection. And so it gives us a very flexible way to integrate that into their control room process. So now you've enabled a very flexible deployment, B, you can communicate directly with the control room. But the other part of this, this nice is we then set this up to communicate to individuals in the organization.
And we can have up to 20 individuals assigned per pipeline. And what we do is we can set up to send those guys emails, we can notify them by a text message, we can send them a GPS location in the field by a text. And so it's an incredibly flexible way, and we ensure that the information gets to the operator within two, three minutes across the organization. Well, what's really interesting to me about this is the fact that you, this is just interesting to me about negative pressure wave in general.
But the fact you can send a location is super useful. Yeah. Because having a leak alarm on a segment that's 20 miles long is very different than having a leak alarm on a segment that's 20 mile long saying you have a leak at mile marker X. That is, that's a very different kind of thing.
It impacts your whole response. Yeah, absolutely. I mean, as I said, leak location is, is important to us, very important to us for that reason, but also in being able to, to help to rule out false positives as well. So it's been something that we've been incredibly focused on.
And for me, the important steps that we go through in the technique that we use, we actually have two, two ways now going about leak location. So the more traditional way for us is we utilize those timestamps that we were talking about. And when we first put the system in, we go to the field with portable units and we, we do a calibration, what we call a pitch and catch. So we're putting a pressure sensor on the pipeline with a small ball quarter turn valve release valve.
And we're doing it at one location and then we're putting on a pressure sensor at a second location, a known distance away. And what we do is, I'm very old school. I don't particularly put much stay in simulation work. We simulate a leak.
We release a small amount of fluid, maybe five seconds, liquid or gas at the leak size, the leak threshold size, the client wants to operate out. What do they want to see that? So effectively, we're simulating the onset of a pipeline leak. What we do is we catch that between two sets of sensors and accurately timestamp when the first sensor sees it and the second sensor sees it.
So now we have a time between sensors. And we have a propagation rate because we know the distance. What we pair with that data is we obviously have pressure data. We pair that data, then we temperature readings on the pipeline at each location.
And we pair that if available with density readings. So what we're trying to establish here, we're trying to get a correlation between the generally the speed of sound of the fluid, the liquid or gas in the pipeline. And the propagation rate that we measured at the time. What we do is when we get a leak event that's been detected, all that data goes off the cloud.
But we also resample temperature and density of the got it at the time. And we use those to recalibrate that propagation rate on the fly. Why is that important? If you didn't do that, let's say you're working with NGLs, compressible, effectively compressible fluids at times.
They can be squishy. And the effectively, the density of those fluids and the speed of sound therefore is very, very susceptible to pressure and temperature change. So your propagation rate is very susceptible to those things. So if we're talking about trying to be very accurate with location 20 to 50 feet is what we achieve, you cannot do that without recalibrating your propagation rate on the fly.
You're simply relying too much on having a very steady state when you first calibrated to when you now operate in the system. Well, in practice for most pipelines running, particularly liquids, that density is changing. Even if it's the same liquid, it's still changing. Go around.
To the extent you're batching on a pipeline or anything else, then that just raises that complexity. But if you're able to calculate speed of transmission between two points, that's going to be able to get you back to density. Yeah, absolutely. And so we're not giving an absolute value.
But what we are doing is using the same parameters as when we put the system in on the pipeline and first got that propagation rate to recalibrate it at the time of the leak event. And that allows us to be incredibly accurate in location. Right. Yeah, no, it's fascinating.
The second way that we're doing that now is actually something that we've learned through the huge amount of data that we've gathered and the large number of leak events and simulated leak events we've gone through is that we're now also able to do a second check, which is basically trending on event intensity, because we're changing this now into a frequency plot, and we've got an intensity of that response. And as you can imagine, the closer you are to a pipeline event, a spontaneous event like a leak event, so it's initiated away from a sensor. The closer you are to that, the higher the response, so the magnitude of the response is going to be greater. So as you move away from that, you get a degradation in that magnitude.
And so through all our analysis now, we're able to actually trend the magnitudes at different sensor locations to be able to also pinpoint the location that way. So now we've got two ways of going out this. Yeah, it's really, it's really interesting what you guys are doing and how you're doing it. The other thing that, you know, is we're having our conversation kind of prepping for this call that came up for me is you use the term pressure event detection versus negative pressure wave to kind of make what you're doing distinct.
And you know, what comes up for me when I start thinking about that is just from an operations and maintenance standpoint, there's a lot of information you could get that you otherwise could not get because there's other kinds of pressure sent pressure signatures for things like valves that are leaking or flow control valves that are not properly tuned or not, you know, not that are chattering or just all kinds of other things are just mechanical things happening on the pipeline that this system could detect because once you get that pressure signature, you can look for other things other than just leaks. Yeah, absolutely. And that's what we found to be very important. And just focusing on leak events, you're actually missing out on on effective user efficiencies for all of that data that you've gathered.
Thousands samples a second can come back to that point. That's a lot of information about how that pipeline is operating. By moving away from just a focus on leak events to pressure events and the way that we can use machine learning to categorize this pressure events now, we can actually work with the operators and say, hey, look, outside of leaks, what else is important to you in operating the pipeline? What would help you with your pipeline operations, your pipeline efficiency?
And we can take samples of those type of events, teach the system to look for them and notify the operator when those are occurring as well. So in the end, what we're trending towards here Russell is being able to do an end-to-end lifetime pipeline performance analysis for the operator using the huge amount of data that we've gathered to feed that. And as I say, whether that's pipeline leak events, which are incredibly important to take those quickly, very, very important. But it's also everything else that's going on on the pipeline.
To be able to really put those into buckets for the operator and to help them to really understand what's happened throughout the lifetime and what effects pipeline life is where we're going with this. That's the ultimate goal of it. Yeah, I think that the opportunity for improving mechanical reliability for all the rotating equipment and control equipment is immense here. That's the kind of thing that there's not a lot of, well, I don't know, I'm getting a little over my skis here.
But when I think about the alternative means, vibration analysis and things like that, I don't think you necessarily eliminate that, but you actually can get confirming signals. And that can help you prioritize your maintenance activities. Absolutely. Yeah, it's a fascinating area to get into.
And the neat thing in the way that we've set the hardware up as well, despite what you said, the complexity of what we're doing and all the sort of pressure on the hardware to perform doing those complex activities, we've actually got spare capacity. So we build our hardware, we spare high speed and low speed sample input channels. So we can actually take other data from the operator as well. And if they wanted us to look at that and do that, we could do additional processing for them as well on that data.
But what you can then get is complete combined data set from multiple different sources and sensors along the pipeline. So to really build, as you said, let me know how I improve mechanical performance or mechanical lifetime of my pipeline. And we can do that through all these different data sets. Yeah, so Stuart, before we wrap up here, I just want to tell the listeners that you and I met at an event where I was sharing a panel talking about product commercialization, you were one of the members of that panel.
Where would you say that you guys are in the commercialization maturity? And what are you working on now in that process? So I guess the first part of this technique, we started developing back in 2018. So we've gone through probably all those years now, five years of pain and suffering and developing a new product and testing that.
So we do a lot of our own confirmation testing on our own test loop facility that we have. Actually, our facility is in Houston. But we've also been able to test this equipment now and the techniques in field for probably the last two to three years as well. So we've got an awful lot of experience of operating the system near commercial.
We've actually had an installation of the system. And you mentioned about running the system one ended as well. We've actually got an example of that offshore. And that's been offshore for some two, three years as well now.
So we're starting to work into our first four commercial projects onshore. So the system, I would say, is fully commercialized at this point. We've got enough data and trust over since 2018 to know that the system does what it says on the tin. And so where we are now is just into deployment and commercialization.
And really the focus for us, Russell, in doing that is the important part that's often missed is how do we integrate this with the operator? And so you can have a very mature technology. But if you don't work with the operator and how you deploy that and how they get benefit from it, it just falls on its sides. It doesn't get used.
So I guess that's the area that we're most focused on now and the operators that we're starting to work with is not only how do we digitally integrate the system with and how do we operational integrate system with them? Yeah, that's the thing that I added to Chris Alexander's maturity matrix is what I call the operations maturity, which is the part that the operator has to do to be able to actually deploy, operate, maintain the technology. Absolutely. And it's absolutely key.
I mean, we've had experience when we first put these systems in, in the operator or a guy in the field, you know, having picking up an issue and suddenly picking up a phone to one of the guys in the office here, because we were the guys in field that helped put them in and they've still got a phone number. Right. And it's like, this is the wrong way to go about this guys. You need a detail processor, you control room, which is a system that utilizes it to go check that you haven't got a false positive to really operate to build a whole capability whenever they deploy a technology.
Correct. They have to do that. Otherwise, they will not get the full benefit from the system. And I'm from your what you're telling me, that's where you are in your commercial ization journey is working with operators to get those kinds of things in place.
That's correct. Yeah. And I think that's the most critical set. So Stuart, if somebody wants to learn more about pipe sense and get in touch with you guys, how would they go about doing that?
So really simple. First of all, first of all, first point of contact, best source of information, our website, simplepipecents.com. Thank you to the website. Got a lot of information on there about different techniques that we're using this for.
Really quick example, Russell, just in very short sentences, we're using this to support real time leak detection during higher tests. We can track pigs in real time using these techniques. And we've even developed a technique to find pre-existing leak some pipelines when you put the system in, and you've already got a leak on the pipeline to be able to find that as well. So go to the website, find out all that information.
You can also contact us through the websites, but all our contact details on that. And as always, we'll put all that information in the show notes. Stuart, look, thank you for taking your time. And thank you for kind of walking us through the gory details, if you will, of what you guys are doing and how you're approaching this.
It certainly seems like you're doing some really interesting things. It'll be interesting to see where you are in another three years with us. Yep. Looking forward to the journey, Russell.
It's been a journey getting here, but we're very excited about what this technology can do. I hope you enjoyed this week's episode of the pipeline or podcast, my conversation with Stuart. Just a reminder before you go, you should register to win our customized pipeline or podcast study Tumblr. Simply visit pipeline podcast network.com slash win and enter yourself in the drawing.
If you would like to support the podcast, please leave us a review on Apple Podcast, Google Play, Spotify, wherever you happen to listen. You can find instructions at pipeline podcast network.com. If you have ideas, questions, or topics you'd be interested in, please let me know on the contact us page at pipeline podcast network.com. Reach out to me on LinkedIn.
Thanks for listening. I'll talk to you next week.