Data Science in the Energy Industry

Key Takeaways

e B hello everyone and thank you for joining today's webinar my name is Ree and I'm going to be a moderator for today's session we're going to kick off today's session in a couple of minutes we're just waiting so everyone has a chance to join uh in the meanwhile though we'd love to hear hear from you so let us know where you're joining from using the chat or the comments depending on what platform you're watching on and yeah tell us something that you'd like to learn in today's webinar uh we are going to be taking questions throughout but we will also be having a Q&A section towards the end so if you have any questions uh throughout please let us know using the chat and the comments and I will save them for the Q&A at the end um also to note this this session is being recorded and the recording will be emailed to everyone that's registered for the event I'll be posting a link so you can register for the event and get the uh recording and resources sent to you afterwards as well I'll be posting that on LinkedIn as well as in the chat shortly uh but yeah until we hit our start time I will leave you with background music so enjoy hello everyone and thank you for joining today's webinar my name is reys and I'm going to be your moderator for today's session we're going to kick off in about 30 seconds or so we've just been waiting so everyone has had a chance to join uh let us know where you're joining from using the chat or the comments depending on where you're watching from and yeah tell us something that you'd like to learn from today's webinar a few bits of housekeeping before we kick off the session is being recorded and the recording will be emailed to everyone that's registered for the event in the next few days um I've put a link so that you can register and get the recording and resources emailed to you in the chat but I'll also be reposting it shortly so that you can sign up for that we are going to have have a Q&A towards the end of the session so if you have any questions uh please type them in as soon as they spring to mind and I will save them for the Q&A brilliant I think that's everything from me so I'm to hand you over to your host for today's session Richie Richie please take it away hello and welcome everyone this is Richie I got to say just looking at uh where everyone's joining from around the world I'm always very impressed with people who uh joining from a very early time zone or very late time zone uh well done for making it here all right so if you try and list the most important skills for a data scientist most people will agree you need some sort of statistical skills and the ability to use data tools whether it's Python and SQL or a business intelligence platform and probably you'd throw in some soft skills like communication and self- motivation but something that often gets overlooked is that you really need domain knowledge of the business you're in you need to understand what your data means and what the goal of any analysis so this is something I've definitely fallen over in my career I've transitioned from one industry to a completely different one said oh I know how to do data science it'll be easy and then spent months being unproductive because I didn't understand the data so today you're going to acquire some domain knowledge you can learn all about what happens with data science in the energy industry and I hope that for any of you considering working in this industry it's going to save you having to ask a few silly questions once you start your job and and we have two guests from Pinnacle Technologies so Ryan Meers is a product manager and his work focuses on digital transformation the application of machine learning technology product development and strategic Consulting and Drew Waters is director of data science so he's an electrical engineer turned data scientist and his technical work focuses on mechanical reliability natural language processing and basian statistical methods so I'm really excited to hear about what uh hear about what goes on at Pinnacle and in the broader energy industry so I think uh over to you Ryan thank you Richie uh thank you for everybody um joining around the world very exciting to see that um I'm gonna go ahead and we'll start the presentation um brief agenda um we'll briefly cover who is Pinnacle um and really where do we fit in in the industry um as a whole throughout energy um and specifically ility and why that matters um talk a little bit about the enery industry uh in general um and then we'll start getting into more specific use cases right so we'll be a blend of domain expertise as well as some technical um and then at the end we'll talk a little bit about various data roles and plenty of Q&A at the end for whatever y'all are interested in talking about uh so who's Pinnacle uh really we're pretty much um the largest data analytics company that is solely dedicated to reliability and maintenance um we serve um customers around the world um and their large industrial operations for the most part that's everything from refining to chemical plants to mining facilities water Wastewater uh even you know wood Mills and Sawmills and and some other Industries as well U and really the the unique perspective that we take is taking data and using that to help our customers make better decisions those could be micro decisions on when do I do maintenance on an asset or which assets should I put sensors on to then get more data to do more data science to large scale decisions on when they should be shutting their plant down uh what should they be doing when they shut down where should they be investing Capital Etc and we bring a lot of data into it a lot of data science various techniques um some interesting challenges in this industry which we we'll cover in just a little bit but it's a little bit about who we are um why reliability matters um it's something quite frankly that you know I myself take for granted every day as well as most people um in the US at the very least um and throughout the world um reliability is basically you get something when you want it at as low a cost as possible right that's everything from electricity to fuel to food to stuff that you uh order on Amazon um having things um you know when you need them um and at reasonable cost is just it's important to society and allows us to you know keep um improving um and when any particular subset of of the of the energy industry um or any particular facility doesn't operate reliably right Society um even micro areas can break down you know example we have listed here is um couple years ago we had a historic freeze got extremely cold um the right measures weren't in place we had a variety of different you know facilities and assets go down led to a lot of Texas not having access to electricity which then had cascading effects a lot of different things happened um and then more in the industries that we serve um some other examples um there BP Texas City back in 2005 there was an explosion actually a series of explosions that occurred due to poor reliab ility um and and really you know injured a lot of people um some people died um and unfortunately benefit of that is it led to more rigorous programs being put in place and right as as the years have gone on getting more and more data driven and using that to help make plants safer um another example is a flint Water Crisis um in Flint Michigan um there was a where they they changed the where they were getting their Source water um there were some different contaminants and constituents in the water led to things corroding ended up leak and then lead got into the water and other contaminants really bad for um you know everybody who who was drinking that water and so reliability is again something we take for granted incredibly important um and we'll be getting into some of those use cases as we go through um before we dive deep into that and you some of the work we do at Pinnacle just taking a step back high level energy right it is makes the world go round um it Powers everything from our vehicles to our homes um a lot of by products that come from the various Industries also or you know things we use every day things we don't even really think about but like plastics for example um almost everything uh that we have around the world is some way somehow impacted by this industry there's a lot of different sub sectors in it and we're not going to necessarily cover everything there but some of the main ones to call out are um in oil and gas um you know Upstream where we're getting oil out of the ground um Midstream where we're transporting that and moving that um around and then Downstream which is things like refineries or or or chemical plants where that gets processed and you can actually do something useful and valuable with that um there's other aspects like with power generation so generating electricity some of which may come from Renewables some of that may come from oil and gas products um and even things like mining where a little bit of a stretch but also a lot of raw materials that are produced right that that Cascades throughout the industry so there a lot of different verticals that's why they're as you see here in the columns the things you'll kind of see throughout all of these is there's there's different kind of functional groups there's right the process itself or the technology that's being used to to produce or distribute um energy or energy sources there's the supply chain so the logistics between all these different facilities um every raw material coming in what happens outside of the facility and even within a given facility there's a lot of logistics and things that happen as well operations so you know people who are actually running the facility um and making decisions and just doing the day-to-day that needs to happen if we need to turn on move something and isolate Etc and then lastly reliability and maintenance which is that's the space that we play in is how do we help a facility produce as much as it can be as safe as it can and do that at the lowest cost possible um so that's that's the space that we play in and so we touched these verticals and other verticals as well um but wanted to give you all that context it's a good way to kind of think about the industry go yeah I can t about this so uh first of all I really appreciate Richie's uh introduction ction of of the topic of of really this is about talking about domain knowledge and the intersection of domain knowledge with data science uh my background is in data science my my PhD is in electrical engineering I've been working as a data scientist for H several years before joining Pinnacle and I had no domain knowledge coming into my my role at Pinnacle and Ryan was incredibly helpful to me in understanding our data better and understanding how those things work uh but yeah having that domain knowledge and being able to marry that with good technical data science skills is absolutely crucial in any role that you're going to have uh certainly crucial for the stuff that we're going to talk about today uh so what we're going to do is we're going to go through a couple of examples of big data science problems that we've solved within this industry um and some of those things are going to talk about uh asset degradation so if I've got a piece of equipment at a Refinery or some other kind of manufacturing plant uh trying to figure out when those things are likely to fail and cause problems when they're going to break down and and not work well uh being able to re make recommendations about when I should go out and do maintenance operations on certain pieces of equipment on one hand I don't want to over maintain and overspend but at the same time I don't necessarily want my equipment to fail because of you know loss of production or potential accidents that can happen uh we've worked in areas like natural language processing uh we're not going to talk a lot about NLP today but this is a very near and dear topic to my heart I work a lot in this uh but being able to mine textual data in our industry is a very very big deal in being able to take words and and derive insights from those uh classification and clustering looking for patterns within data where do we find similarities in safety incidences that have occurred in the past or different types of fail failures that occur and what we can learn and leverage from that uh forecasting is a big deal for us looking at uh what a plant's long-term production uh capabilities are going to look like and and being able to actionized those uh down to Logistics so uh things like trying to figure out when we should order spare parts when we should uh do certain things so that when when problems arise we can respond quickly to them uh so we'll be talking about a variety these things yeah and um some of the common data challenges so depends on which customer we're working with and kind of maybe even which subgroups and different use cases um for a lot of use cases um there's not always a lot of data when we say a lot of data um having a few sensors worth of data right that we're not talking about like Google level data or Amazon level data generally speaking unless you're looking across the Enterprise um for one of our large customers might have 30 plus facilities and that being said um throughout the industry we're seeing a lot of our customers they investing time in standing up data laks and various data pipelines some customers don't have that established fully or it's not something that anyone can really make use of um they're so they're going through a lot of IT projects to do those types of things or maybe they've got it partially implemented some data is available but maybe not the data that we'd be looking for to solve a very specific reliability problem again not the case always this just something that we're seeing a lot of investment go into is sustaining this type of stuff people know data is valuable um the other thing we've also seen is some of our customers our customers know data is valuable what the some of the challenges come into is okay what can I do with my data right that's another thing is and this is where the do this is where the domain expertise really comes into play with knowing what is possible with data science but then knowing hey here are the problems that we need to go solve here's the value of solving those problems and here's all the little nuances that'll bite you as you you know deploy neural networks or other fun things yeah but very good you want to cover the other two uh yeah I can do so uh with our data too so a lot of times we're we're dealing in situations where there's not a lot of volume of data like Ryan mentioned uh but even when we have a lot of data that data is noisy and sometimes it's not reliable we're going to give a lot of concrete examples about that uh going forward but uh those are two things that we do have to worry about uh with our industry too we have to work around concerns that our customers have with the data that they they already have access to a lot of times there's a lot of distrust regarding the data uh when we talk about inspection data that's very much a real thing where there might be certain periods of time where they say hey we don't trust anything that's in here should we throw it away should we not throw it away um or where we just have gaps we have like data that's missing we don't know what's going on in this one particular portion of the facility with high accuracy or um the data is not in structured or unstructured data sources there's a lot of data that still lives in people's heads yeah or paper documents paper documents or whatever the case is yeah all that kind of stuff uh finally there's a lot of tension between uh subject matter expertise in our industry and and data science principles uh we we've staked out a very good space uh at at Pinnacle on trying to marry those two things together and we'll give some very concrete examples about that uh but you will have a certain amount of tribal knowledge that exists among the veteran Engineers who have been working in this industry for 40 50 60 years um and then you know up andc coming data science techniques come in and try to revolutionize what they've been doing for a long time and doing very successfully uh but trying to marry those two things together is not always straightforward that takes a lot of uh you know both the industry knowledge and then also the statistical knowledge to be able to pull that off yeah I'll say that what we founded most successful is blending the two together yeah ummes help point and point out hey you're missing this really critical piece of information worst case we get it from them but best case we go poke around and find that and then vice versa right data science is pointing out interesting things they maybe didn't think about and then they can they can then reason here's what is actually happening that's causing that and then we can go help the customer make a better decision on how to prevent that issue and so bringing about those worlds together is where the magic happens this is great uh so one of the first examples we're going to talk about is uh asset degradation uh so when you're operating say an oil and gas Refinery you're you're taking crude oil and you're turning this into uh uh into products like gasoline or jet fuel or or whatever else uh the chemicals that you're you're operating on the crude oil the the things that you're doing chemically to to transform that crude oil into useful things uh it's a fairly corrosive process it's some pretty nasty stuff that you're working with there um and so what I'm showing over on the right hand side of this particular slide are places where degradation and corrosion are happening on on assets so I've got a pipe in the upper right uh over time after that pipe carries a lot of this chemical uh content uh that pipe is going to thin out over time it's going to corrode it's going to rot out uh eventually this gets to the point where we're now springing Leakes and best case scenario that leak results in hey we have to shut down our facility for a little bit we have to do some repairs we maybe didn't get to produce as as much product as we wanted during that day worst case scenario that can start a fire it can cause an explosion it can be something very very dangerous um down on the right um I've got this picture in here just to highlight that there are different types of degradation that we deal with in our industry uh so in the upper right this pipe is just carrying corrosive chemicals that that rots out uh the pipe over time uh down below is a special case called corrosion under insulation so this is where we've wrapped a pipe in in insulating material because we want to maintain temperature for this uh if a little bit of moisture gets underneath that insulation it can cause very very rapid corrosion uh not so much because of the like corrosive chemical properties of that just more more mechanical type corrosion that happens there uh so we deal with different types of of corrosion but the end result is the same after time when the acid gets to the point where it's no longer thick enough to hold the product uh we have problems and we'd like to avoid that there's a lot of things that'll determine what kind of damage we're dealing with how quickly that damage is going to occur uh some of those I've listed on here so the uh the chemical content of the stream that the asset is is working with uh the Metallurgy of our asset so uh if we're dealing with carbon steel things will manifest a certain way if our pipe is stainless steel it's a different thing uh we have other alloy materials that are more high-end for dealing with certain things and they they will all behave differently based on on on that interaction between the chemical content of the stream and and the Metallurgy at the asset uh operating conditions matter so knowing what the temperature knowing what the pressure is uh knowing what the ambient temperature is what's the humidity in the area that we're dealing with all of that stuff is is very important to us another thing that is a problem for us is that corrosion can be very localized and by a localized I mean very non-uniform so in some cases for certain types of damage you might expect that damage to be fairly uniform over the content of say a pipe or a pressure vessel whatever asset you're dealing with other times it could be very pin level corrosion that you're dealing with so even if you're looking at one portion of the asset you may not know anything about the real problem area of the asset and that's a problem that's something that we have to deal with so uh let's say I want to get down into modeling now so I'm a data scientist I want to do data science modeling on this and uh what will happen is that over time facilities will go out and they'll conduct inspection work on on a particular asset say I'm dealing with a piece of pipe uh there's a piece of machinery called an ultrasonic thickness monitor that an inspector can bring bring out into the field uh they'll they'll connect this up to the pipe they'll take a reading on it and they'll get some measurement of thickness for that particular pipe at that location at that particular point in time and this is a small area that they're looking at this is maybe at most like a foot of uh uh width that you're looking at on a pipe that could be very very very long um and what might happen over time is I get something that looks like that graph over on the right uh every Red Dot that I've got on there is an inspection data point that was taken uh my gray cone that I'm drawing here is something of our our expectations going into this problem this is how we expected corrosion to happen on this uh the Red Data ideally would be living inside of that cone but uh as as we're showing here it's not doing that because it almost never does uh what I want to point out on this too is that um the data that we're dealing with those red dots they're quantitative this is a machine that is taking a measurement and returning a value there is confidence with that value but there's also some noise that happens with this data right here so depending on the skill of the inspector that goes out and does that inspection that data point may be more or less reliable if you have a very seasoned inspector who's very very good at their job and very precise with things that might be a very very small error bar if this is a newer inspector or somebody who's not very familiar with the equipment or the facility that may be a very wide error bar that we're dealing with here so there's there's noise in this data um other things that can happen the equipment can be malfunctioning so it can be returning faulty measurements uh you will have times too where even if the inspector is doing their job correctly they may have taken the measurement of the wrong location so they may be a couple of feet down so it's not necessarily the same data that you're getting over time there's just a lot of things that go into this problem that are are difficult for us but the real problem that we're interested in solving here is hey I've got some data that tells me about thickness over time and because I'm I'm expecting this asset to thin out over time I'm kind of expecting a general downward Trend I should be getting thinner over time uh as things go on and what I'm really ultimately after is I want to know when this thing is going to get so thin that I can't use it reliably anymore uh that is not zero thickness that's going to be some kind of a safety margin that's built into that uh and so what we're after is can I like have a data model that tells me hey based on your data based on what you expect to have happen when is this thing going to fail now one of the things that uh I actually throw this out as a job interview question uh quite often so I know we've got some data science students on on the call so uh this this is one for you guys uh I will throw this problem out to prospective data scientists who apply for jobs at Pinnacle uh and as you can see I've got five data points on this plot right here and I'll ask them hey like go solve this problem for me go tell me like how you would go about figuring out when the thickness of this asset is going to get to the point that it's not usable anymore and I'll give him some value of like hey when when is it gonna get down to 300 now when I give that problem I'm not looking for something as complicated as what we do at Pinnacle like we've invested a lot of R&D in the modeling that we're about to talk about with you guys but uh I'm basically looking for an answer like hey I'm going to use use a regression model or maybe if they're really fancy I'm going to use a gan process or or something like that if the candidate says I'm going to use a neural network on this five data point example right here you're probably not getting a second interview and so uh just just something for the students out there I know neural networks are definitely like the really cool thing and if you're working at a Google or a Facebook they're absolutely state-of-the-art best thing that you can do uh but it's good to know other things that you can uh throw at data problems like this too so just just throwing that out there but the two things that I'm going to balance out when I solve this problem here are one I've got my data it's noisy data it's infrequent data I don't necessarily have a lot of it uh but I also need to balance in the role of the subject matter expert so again tremendous domain knowledge in this industry we have people who have been working on these corrosion problems for 70 80 years I mean we've just got this wealth of stuff in the the industry that we have to to draw on uh and so what the subject matter expert brings to the table is they know based on the parameters of this asset what it's doing what kind of stuff it's operating on how much temperature it has they know what that theoretical expectation for corrosion is uh plus or minus they maybe have worked on things that are similar uh but one of the downsides of that subject matter expert is that that General theoretical laboratory knowledge may not translate to the exact asset that I'm working with so I want to balance these two things out uh when I solve this problem so uh what am I going to do here well uh as when we were introduced my my background is Invasion statistics so of course we're going to solve this with Invasion approach um so vasion statistics real quick review I have some data that I observe uh so my data is y in the diagram that I'm showing here and what I really want to know is hey if I've got thickness data Y what is the true thickness of my asset or what is the true rate of corrosion that I'm seeing on this asset given that observed data that I've collected uh now my subject matter expert comes here in the role of being a uh aasan prior uh for this so I I have some kind of a prior belief on what thickness should look like over time for my asset what corrosion should look like on this asset I've got quantitative inspection data that is my my observed data what bay Rule and what basian statistics allow me to do is go backwards and say hey I've got data that tells me this and a a prior belief that says this what do I think now now that I've observed my data so how do I get back to my ex how do I get back to that underlying ground truth reality uh based on potentially noisy data and and some uh potentially noisy prior beliefs about what should happen here uh so I'm going to marry my prior and my data together I'm going to end up with a posterior distribution that tells me something about uh what's what's really going on based on on all things that I have access to and what I can do with that is I can come up with a cool model like this you want to talk to this one yeah this is this is one of the models that we use um which we we call our it's called a lifetime variability curve um but you can think about it as just it's a it's a very fancy regression model that's got some various aspects to it that work in this domain very well um but basically a lot of times we like early in an asset life we may not have a lot of data we're going to lean heavily on prior prior knowledge which could come from anme it could come from another machine learning model a variety of different ways we get that and as we get information we're using that to update what our beliefs are and what you see in the middle chart um is we collected some data which is a little bit less severe than what the original expectation was from our subject matter expert and as a result of that we don't have a lot of data points so we're not going to overfit to the data itself um knowing the there's error in the data and also the processes are very variable these things are not operating exactly the same way all the time they're highly Dynamic um so we see a wider band of uncertainty increase in our projection meaning we don't have a lot of confidence exactly of what's going on um but that's going to help us understand when is it valuable to go collect more data Big Challenge we Face data is not free so it might be like I could totally predict failure if you go scanned every single asset and spent hundreds of millions of dollars but that is just not simply practical to do on a regular basis um and so that help that's another big benefit of this knowing when do I go collect more data um and then lastly as we see here we've collected some more data at an appropriate point in time we've seen that there's a shift right there's something that changed in the process or maybe we got we're using a little bit better techniques we're finding more of the damage um now um which is which is a a challenge um and in this case we see that there's we have a little bit tighter Trend it matches a little bit more what we expect but we got a little bit of a life extension to an extent um just based on the the lower degradation operation we saw earlier in life and so this is always going to be updated um there's also variations of this model we can take where we bring in other information like process data um and other time series so there's there's varying approaches we can take here um that covers kind of One Core approach we take there and then I'll turn over to Drew another use case for this like I mentioned collecting this data costs money and it's not like it cost a dollar um it costs a significant check of change but once we've got these types of analytics run across a variety of assets and across all the different um locations where you might collect this type of data we can start optimizing where we go collect it yeah so so as R saying data collection is is costly this isn't something that comes for free this isn't a sensor that lives out in the field this is a a human who has to go out there spend a significant amount of time to do this this costs money uh storing it costs money like maintaining it cost money so there's there's a cost associated with all of this now what we found uh going over I mean facility after facility of client data is that there's a lot of inspection that's being done that is useful and there's a lot of inspection being done that is not useful um so as a good example if I go back to my my plot over here and let's say I talk about that uh that third plot over on the right if my data points are in a perfect line and every time that I am like collecting a data point I'm seeing exactly what I expect to see collecting more data isn't super useful for me that posterior distribution that shaded confidence area is going to get down to laser beam uh pretty quick and what that's essentially telling me is that I don't need to go collect more data there it's like I know what's going on I've got a good handle of things now in a case where my uncertainty is quite wide say in the second plot that we've got here uh there I do want to collect more data it's like I don't know what's going on here data can help resolve some of those issues uh but what can we do uh above and beyond things so once we run our model the model gives us an idea of when things are likely to fail well there's a couple of things that we can do with that there are interesting one is we can start telling them hey for future inspections these locations the subset of locations are like something with high uncertainty we need to know more about it there's risk that's associated with these locations you should go collect more data here uh but in other cases we can say you don't need to collect anything here right now like maybe in 10 years 20 years you might actually need to worry about this again but right now you're good Focus your resources where they matter and go look at the things that that are important uh so with the uh the pipe that I've drawn over on the right red pipe every blue location is one location that I'm inspecting uh what we might find is that for every 10 locations that they could inspect only one of them really matters we can actually get quite a bit of reduction uh in the amount of inspection that they do and that's a 10x cost savings now if they can uh do that and do it well uh now on the other hand what will happen is because the inspection locations only represent a fraction of the total surface area of the asset say I'm talking about a pipe uh I may have only sampled maybe one to two percent of the length of that large pipe out in the field uh which means there's a whole lot of pipe left that could be experiencing other problems uh we can ALS o do some need optimization work and some Orban optimization work uh where we can say hey based on what you've seen already and based on concerns that your subject matter expert had how bad could things get in those red regions of the pipe the stuff that we haven't inspect haven't inspected yet uh a lot of times what will happen is well based on the data that we've got and based on the concerns of theem not really that much can happen there you're okay you don't need to go do more inspection work you're fine uh you're fine here on this particular asset for other assets though what we'll find is that there is quite a bit of risk that that hasn't been uh dealt with so if I take some measurement data and I'm starting to see some really like wide variant on corrosion rates I'm starting to see some really scary stuff when I go and project over that unobserved region we're gonna say hey alarm Bells like there's a good chance that something really bad is happening out here given how little work you've done so far and given what you've seen already um you should go do more and you should bring maybe more uh uh better inspection techniques rather than just look at Point locations maybe do a bigger scan of this area if there's a real threat of uh you know particular types of degradation and corrosion that you're worried about there now when we've done this on real studies this is the the fun one here uh we typically find that about 70% of the inspections that that customers are doing are not value adding uh and that's been as high as 90% with some customers there's quite a bit of like room for growth here and uh for cost savings that you can have here uh but in about 10% of the assets that we've looked at uh not enough work was being done so on one end they're spending money where they shouldn't be spending it they're failing to spend money where they should be spending it you might be experiencing failures as a result which are costly yeah and so ultimately we can come in there and try to like optimize things for them uh tell them where they should do things and where they shouldn't yeah on the whole they end up spending less money generally and then they're also right mitigating risk as well so it's big big value ad um of of data analytics with corrosion good stuff actually there's more corrosion yeah more corrosion we're going to do more corrosion now so uh one thing that we mentioned in the previous uh portion of of the talk was uh we're combining data and we're combining a subject matter expert's opinion of what should be happening with a particular type of asset and so the may come around and say hey for this asset that's carrying this material that's made of this metalurgy operating at this temperature operating this pressure we expect to see 0.005 inches of degradation per year that may be something that they say but they wouldn't be surprised if it was maybe .03 or you know up to 0.07 or something like that uh so they'll give some like error bars on that that's the uh cone that we drew previously over here that's that's kind of the uncertainty that the might have uh and the smme estimate what the subject matter expert comes up with is based on a variety of things it's based on tribal knowledge it's based on uh uh laboratory condition type of work there's a variety of Industry studies that have been done you know over the years and some models quote unquote that exist so here's a fun anecdotal story right here so uh one particular type of corrosion that happens out in the field is called self fation this is when you're carrying sulfur content uh in in your material and the way that we understand sulfidation today is that based on the temperature of the stream that you're dealing with and based on the amount of sulfur that you have in your stream uh that tells you something about what your corrosion would be and this depends again on the Metallurgy of the asset that you're dealing with so in the plot that we're showing here I've got temperature by corrosion rate uh and we have a line for every different Metallurgy that we've got so carbon steel this is the cheapest material that they can throw out here and that's going to have the the highest corrosion rate it's not super resistant to this uh if they start using more expensive metallurgies they can bring that corrosion rate down uh but the major Trends are like as I increase my temperature I expect to see worse corrosion just because things are going to eat things through things faster when when temperature is active um and also as sulfur content increases you're going to expect to see an increase in corrosion here we're just showing one particular value of sulfur content so6 uh here but like as you increase that sulfur content things are going to get much worse now the fun story on this one so this uh curves like this were first introduced in the industry back in 1961 uh so these very very old amount of tribal knowledge that we're we're dealing with here was based on lab data it was based on experience it was based on like best practice knowledge at the time now by 1986 what they found out is that these curves were 2.5 times too excessive so uh if it was predicting say uh 10 millimeters of corrosion per year you could be pretty sure it was really about 2 .5 millimeters of corrosion they were uh uh too extreme in what they were doing now conservativism is a very good thing in our industry we want to air on the side of caution we want to be scared about the things that could be happening in the field but the downside of that is that if I'm overly concerned about corrosion it means I'm G to go do more inspection work I'm gonna spend more resources I'm gonna spend more money and then two at some point you realize that hey stuff's way over conservative and now I kind of start if I'm a business person I start not trusting the technical as much when it is too inflated yeah right um and then then you might end up missing something that really matters yeah so uh and so what we were driving at with the the next part of the talk is can we do something better than just rely on these standard industry curves where we leverage big data and in some meaningful way uh now the the analysis that we're going to talk about is for a very specific type of asset in the industry it's called a reformer unit uh we have a fairly large comprehensive data set of of reformers uh throughout the the world right now uh about 20 facilities 36 reformer units uh consisting of a very very large number of assets back of the envelope the calculation that we did said this was about uh 10% of the US reformer Market about 2% of worldwide reformers but again our overall goals here are can we leverage data and can we come up with a better predictor for corrosion based on all the things that the industry normally cares about and then can we do a comparison with real uh industry standard methods and see what we come up with uh data that we're going to use in this model right here uh so we've got our inspection data that we've talked about before so this is thickness and and when the thickness measurement was recorded uh we know which facility we're dealing with and we know where they're located that's going to be a big deal we'll talk about that more in a minute uh we've got conditions of the assets so what temperature are they running at what's their pressure what Metallurgy are they constructed from are they insulated do they have uh pwht is a post weld heat treatment it's it's something that gets applied selectively to certain assets uh we have some knowledge of chemical information so we know what our sulfide content is we know what the ammonia cont content is we know what the water content is uh things like that are typically available for a lot of this at least the the theoretical value is what we think we were going to run into now if I threw this content at you and I said go solve this problem a lot of you would probably be coming up with exactly what we did and that is treated like a supervised machine learning model so what we're going to do in supervised machine learning is we're going to Simply take an algorithm and we're going to start spoon feeding it example data uh so we're going to say hey here's an asset it experienced this corrosion rate had this temperature it had this pressure it had this Metallurgy it had this sulfide content it had this water mole percentage uh and like as we feed more and more examples to the machine the machine is going to start to figure out well okay what things matter for predicting corrosion which things don't matter and how do they all relate together uh that's that's how we tackle this problem so uh very very simple regression supervised machine learning problem this is very data science bread and butter um and we we simply just keep feeding all of these examples to the machine as it like starts to figure these things out really important thing to know with our model here is we do not balance any type ofme uh information in this we validated our model uh with the subject matter experts but we don't actually like tell anything to the machine about that the machine has no idea what corrosion is the machine doesn't understand what Metallurgy is it doesn't understand what temperature is it's just a data table that it's going to process on uh anything you want to say there or no just if if we wanted to we could also bring in otherme knowledge or industry tool data as another attribute in the model that that could help predict and we've talked about doing this this is this on our TBD list uh what we're going to Output from this ultimately we want our model to understand okay how do I map all this input data to a corrosion rate and then uh we're going to get some some fun stuff with that so uh corrosion rate estimates uh think of something like a random Forest regression model or like a gradient boosting regression model uh when everything is done I'm going to get an estimate of what corrosion should be for any new thing that I throw into this model but I'm also going to get uncertainty bounds on this and this is important because uh even with the data that we've got we may have assets that have the exact same conditions but experience slightly different corrosion uh so we want to know okay well what are the the bounds on what corrosion could look like for this or yeah or the conditions of the plant are changing every day yes right so corrosion is never this perfect line it might be flat and then it's severe or like everything in between at different points so there's musing yeah UND quantifying the uncertainty of what's going on um is hugely valuable for um bottom line impact ultimately yeah and so when we query the model uh we're going to get an estimated corrosion rate that's the Blue Line say that we've got on this graph over here but we also get a confidence interval around that uh so we have an idea of okay well what are the what are the bounds on this like how how best case scenario worst case scenario what are we looking at uh model actually fits really well so like most of our predictions kind of fall within the 95% confidence interval of uh what our model produces that's a good sanity check for us um variable importance comes out of our model so people who are recy hit learn users on here you know exactly what I'm talking about uh we can query the model and say hey when you were doing your learning What mattered and what didn't and what matters depends a little bit on the type of asset that we're looking at and the particular type of damage that we're looking at but for one particular example my feature importance might look like the graph that I've got over here where for this particular example temperature mattered quite a bit pressure mattered quite a bit location matters a lot now what do we mean by this so we mentioned before this is 20 different uh oil and gas facilities that we're looking at um some of those are operated by big major oil companies and the names would be things that you would recognize there uh now it turns out the location is hugely important like who is operating the facility matters a lot a uh a facility that is operated by one major refiner may look very different than a facility operated by another major refiner even though they're producing the same product with the same process and the same equipment and and really we see this as like it's a proxy variable that blends a lot of things together like you know facilities that have a very consist consistent feed like what they're bringing in and processing versus some facilities may get opportunity stuff that's hey man this stuff's real cheap let's run it but then it starts tearing your stuff apart and they see all sorts of different things um so there's a lot of different variables that go go into that um but that was a really interesting finding that that came out yeah good stuff and I know we're getting close on time so I'll get through uh model validation we do all the normal data science things we do fivefold cross validation to verify that this model is doing what we expected to do and we're not grotesquely overfitting to our data um results when we do this actually look pretty good so on average we're off by about 2.2 thousand of an inch per year in terms of the corrosion rates that we predict that's actually really good for our industry right there uh and then the other fun experiment that we did is we compared this against a real human corrosion so uh we gave them a variety of data regarding the asset it was the same stuff that our algorithm had but then they also could dig deeper into the data they could go uh pull up diagrams of how this facility operates and they could leverage a lot of that tribal knowledge that they had that are methodology didn't have access to uh we had a subset of about 20 assets that we gave them and generally our methodology outperformed the human uh by a reasonable margin so on average the the was off by about five uh thousands of an inch per year we were off by 3.1 it was a nice validation of of our methodology that this works well and again the intent of this is not to try to replace that human corrosion they bring a tremendous amount of value to the table uh but we see this methodology as really a tool to kind of augment them and make their life easier or also right you can use this also as kind of a screening approach right you can Mass run this model on an entire Enterprise worth of data and then start following up with the highest risk areas or or cross validate withes and the idea would be you continually improve this um and and as you keep getting more and more data cool cool uh okay and then our last use case um so then we're going to talk about we talk a lot about like corrosion and like fixed equipment we'll talk a little B about stuff that moves um so we're going to talk about um an example project we did for a uh low press compressor um and so this was this is an asset that is compressing fluid and it in this case we have a lot more data than like the low data case with corrosion uh there are about 250 sensors or or time series available to us um and there was a specific failure event that happened and so the challenge was hey can you tell me before this thing was going to fail and like how far in advance can you tell me that this thing you know had an issue that happened to it um so there's a few different things we did starting with anomaly detection and then there's a little bit of other work we did there as well um as I mentioned there was a failure that happened at the end of January um a year or two ago um and like I mentioned we had 250 different time series extremely high volume data to the point where like my computer couldn't handle it so we started getting into the realm of big data is um at least in this industry uh for a particular assets we had a lot of different information again a lot of cleansing and things that need to happen here right sensors may go bad or be or the asset might be off just because hey the facility was down that's not necessarily a