Rise of AI Conference 2019 · Berlin · Unicorn Stage

The First AI Product for Railway Infrastructure

Vlad Lata · Vice Chairman, KONUX

At Rise of AI 2019 in Berlin, Vlad Lata, Vice Chairman at KONUX, describes how his company built a predictive-maintenance product for railway switches. KONUX, then over 50 employees and having raised more than 50 million dollars, operates in Germany, Sweden and France and is expanding into Japan, China and the US. Lata frames the problem: in Germany alone 12 million passengers travel daily, one in five experiences a delay of more than six minutes, and half of passengers are dissatisfied, while maintenance rules date from the 1960s and 1970s and no longer match today's doubled train and cargo loads. He argues that building an AI product for such an industry requires three things together: domain expertise, IoT sensing, and AI, since existing railway systems were built for process control, not data analysis. KONUX attaches acceleration sensors to railway switches, converts the signal into vertical displacement to assess ballast condition, and applies predictive algorithms, following what Lata calls a strategic data acquisition cycle in which the product's immediate value to users drives more data and further product improvement.

Key takeaways

  • KONUX builds AI predictive-maintenance products for railway switches, a physical component responsible for about 20 percent of all delay minutes.
  • Vlad Lata argues that domain expertise, purpose-built IoT sensing and narrowly applied AI must combine, since legacy railway systems were built for process control, not data analysis.
  • KONUX follows a strategic data acquisition cycle, attributed to Andrew Ng, where immediate customer value drives more data and further product improvement.
  • A small neural network detects zero-crossings to correct sensor drift, converting acceleration traces into vertical displacement accurate to 0.1 millimeters versus laser measurement.
  • Customers require at least a 90-day prediction horizon because a temporary intervention train costs 80,000 to 100,000 euros, and KONUX's hit rate has risen from about 90 to above 95 percent.

Detailed notes

KONUX's mission and the railway problem

Lata introduces KONUX, then just over 50 employees, having raised more than 50 million dollars including from a US fund (NEA), a Munich VC fund, and more recently a strategic investor, and operating across Germany, Sweden and France while expanding into Japan, China and the US. KONUX's mission is to empower the railway system to operate with the highest availability, because railway beats all other transport modes on emissions, accidents, noise and climate impact. He cites that in Germany alone 12 million passengers travel daily, one in five experiences a delay of more than six minutes, and about half of passengers are dissatisfied with service. He attributes part of the problem to maintenance rules and protocols designed in the 1960s and 1970s that no longer fit today's doubled volume of passenger and cargo trains.

Three keys to building an AI product: domain expertise, IoT and AI

Lata says KONUX's first key learning was that building an AI product in this space depends on three things: domain expertise to understand what should be predicted and what data matters; IoT, because existing railway systems were designed roughly 20 years ago for real-time process control, not for the deeper, more precise data needed for data analysis; and AI itself, applied narrowly. KONUX focused on railway switches, the physical point where a train changes track, which are influenced by many unrecorded variables such as weather, installation method, train speed and weight, and metal processing. Because historic data on these variables mostly does not exist, KONUX works with what Lata calls "small data which is kind of big data in disguise," combining sparse sensor data with heavy domain expertise.

The strategic data acquisition cycle

The second learning, which Lata attributes to Andrew Ng (who led the Google and Baidu AI teams), is to design a strategic data acquisition cycle: gather field data, build a product that creates immediate value for the customer, use that value to gain more users, and use more users to gain more data, feeding back into a better product. For KONUX this means installing acceleration sensors on switches, giving customers immediate failure predictions, and in return receiving customers' maintenance history and other sensor data. Lata says this creates a defensible position because only KONUX holds the resulting learning, while keeping the company focused on delivering short-term customer value.

How the sensor and prediction algorithm work

KONUX's device, built by manufacturing partners to KONUX's design, is mounted next to the switch blade on the rail track. As a train passes, it produces an acceleration trace, which is converted (via double integration) into vertical displacement, indicating the condition of the ballast, the stones fixing the switch in place. Data shows that good ballast condition reduces failure rate by a factor of six. Because double integration of noisy acceleration data causes signal drift, KONUX trained a small neural network to detect zero-crossings in the acceleration trace as recalibration points; the resulting displacement curve matches direct laser measurements to within 0.1 millimeters. KONUX also classifies train types from acceleration traces, first using a simple vector of axle-distance characteristics and later a neural network trained on known schedules, and can reconstruct a full train schedule over a switch after one day of data.

Key metrics and business impact

Lata says customers need at least a 90-day prediction horizon to plan the closures and lead processes required to fix a switch, since a temporary intervention train costs 80,000 to 100,000 euros to run. KONUX's hit rate started around 90 percent and is now above 95 percent. Switches cause about 20 percent of all delay minutes, so improving them directly improves punctuality; KONUX also aims to extend asset lifetime beyond fixed replacement cycles and to make maintenance more efficient, citing a case where it validated five different maintenance measures in two to three weeks versus a prior six-month wait-and-see cycle. KONUX monetizes as "system as a service," charging a yearly per-asset fee benchmarked against the savings its solution generates.

Q&A highlights

Asked how a rollout is validated before scaling across a whole network, Lata describes starting from a customer's historical failure data to estimate how many assets and how much time are needed, then discovering failures occurring sooner than the historical data suggested, which built a feedback cycle with the customer. Asked whether success would eventually remove KONUX's own training data by preventing failures, Lata says removing the system would mean losing field visibility entirely, and describes KONUX's longer-term ambition as becoming an operating system for the railway that also informs schedule planning and new route construction. Asked about using the roughly 50 million dollars raised, he says KONUX has stayed conservative with prior rounds, given that industry change cycles and deals can take two to four years, to ensure sufficient runway.

Quotes

We don't want to just apply AI generic, we want to look at specific components of a product and where can it be beneficial for us.

Vlad Lata

Questions answered in this talk

How does KONUX predict failures in railway switches?

According to Vlad Lata at Rise of AI 2019, KONUX attaches acceleration sensors to switches, converts the signal into vertical displacement to assess ballast condition, and applies predictive algorithms including a small neural network, reaching over 95 percent hit rate with a 90-day prediction horizon.

Why do railway switches matter for train punctuality?

Vlad Lata explains at Rise of AI 2019 that switches cause about 20 percent of all delay minutes on the railway, so improving their maintenance directly improves overall punctuality.

What is a strategic data acquisition cycle in an AI product?

At Rise of AI 2019, Vlad Lata describes a cycle, attributed to Andrew Ng, in which a product creates immediate value for users, which drives more users and data, which in turn improves the product and enables new products, calling this approach central to KONUX's growth.

Mentioned

predictive maintenance, railway switch, ballast condition, small data, strategic data acquisition cycle, system as a service, zero-crossing detection, acceleration sensor

Keep exploring

More from Rise of AI Conference 2019

More on Enterprise AI & Adoption

Transcript

Auto-generated transcript, not edited. · 4,872 words

Show the full transcript

I will give you a short introduction of what we do but basically also want to give you kind of three takeaways when we talk about building companies or businesses around AI and what our key learnings are besides this if you want to know more about what we do feel free to approach me afterwards and we can everyone talk about that so a little bit about our company so we are something about over 50 employees now like it's all about the people that's why I always like to say this the first thing because this is the most important thing to us as Jesper said we've managed to get some significant investment over 50 million dollars the most which Nea out of the u.s.

we have some VC fund out of Munich we also had you Baba now more recently as a strategic on board and we are working across Europe so far so we're working in Germany Sweden France but also expanding Clos slowly into Japan China u.s. so other sides of the world bringing up technology they're right in our space it takes a bit more longer it's a bit more complex because we have regulations and so on so forth but I will tell you about that in a minute now what is our mission right and we spend a lot of time thinking about this at the beginning of a company and we said ok the goal is to empower the railway system to operate with the highest availability right so that means everything needs to be on time all maintenance is scheduled accordingly all the things work at they should because the railway is here to serve us the people right we need to go from A to B so the REA will spend power that and we want

to empower them to be able to do that now why are they away yeah pretty simple if you look at a number of different reasons for it right if you look at emissions if you look at accidents if you look at noise if you look at impact and climate change like railway beats all other modes of transport right doesn't matter if it's personal or if it's freight railways kind of the safest the most free the most efficient in terms of energy the most climate friendly so basically in our view as the generation who is kind of trying to save the planet right we were thinking that the railway is the one thing we want to support the most because it gives back the most to the pad now the problem with railway though today is that you have something like in Germany alone 12 million passengers traveling every day and there's one in every five passengers who is experiencing a delay which is higher than six minutes right I think everyone in this room can

say they were almost once delayed with a train so that's unfortunately still a big problem and at the same time we have half of the passengers traveling which are dissatisfied with the service right so I don't want to tell you what's the customer satisfaction no great that the deutsche bahn has because it can be gravely improved right but this is the proof that okay there is something needs to be done right things need to change and another thing about railway is that as with many industries which are very traditional right which has been here for a long long time there's a lot of rules and and protocols which have been designed in the 70s and the 60s that that describe how work should be done that describe house but in decisions should be taken and the problem is that those things don't rely anymore to what's happening today right you have doubled the number of trains running on the track you have doubled the load of cargo trains running on the track you cannot do

maintenance the same way right it doesn't work like that at one point you will reach a point where you're not being able to serve with the maintenance you're doing to ensure the availability that you need for a network so how to go about it right because it's a big problem it has a huge impact but you can assume it's not easy to solve right so the first thing that we learned when we started conics and when we were starting to solve this challenge was that it all comes down to three different things right of course number one you need to have the domain expertise so as we heard about AI today and and and talked about it I think it's clear that understanding the system you're trying to predict you're trying to moderate you're trying to analyze gives you a lot of edge right and understanding what the input and the output is for example helps you create a better algorithm so the domain expertise is key for us and then secondly you need to

pair that in our case with IOT right because we're talking about a network which is across Germany and there isn't any system out there they design to send us data right so all of the data that is being aggregated so far today's but systems that are being designed maybe let's say 20 years ago and those were designed for process control not for date afford analysis so there's two different things right I'm an electrical engineer of my background so I can tell you when we talk about process control it's much more about okay real-time information being able to take the decision when we talk about when we talk about data analysis we need much more deeper level of data depth but also much more precise information on specific points and third of course AI right because this is one of the problems we're not having information out there makes it difficult for you to think about okay how can I solve it right just taking a simple example we are focusing on railway switches right and

the railway switch is a point where the railway is changing the on the route the from one track to the other very simple physical systems if you look at it it has some make some some metal points and concrete components and then there's a train running over it and that's it right but the problem is that it can be influenced by so many variables like the weather how was it installed what speed the train is running what weight the train has how was the metal processed and so on so for there's a lot of different variables which of which data you don't have you don't know because this happened maybe 2025 years ago when the switch was installed so what would you need to do is you need to create some system that is able to react with an intelligent algorithm based on a small amount of data but then you add a lot of domain expertise to kind of make it more powerful right so we call it small data which is kind of

big data in disguise right because you don't have all the information you need but you need to make some decisions now second thing that we learned also in our process of working in the space is that ideally you manage to place yourself in a data a strategic data acquisition cycle and what I mean by that is something that I didn't invent so this is something we learned from the likes of Andrew and if you know him he's one of the most known people in AI he led the the Google AI team he led the by do AI teams and he said to think about it the following right at first you want to create some date you want to have some data gathered in the field right so in our case we want to have some accelerations and measurements or some trains pressing the switch based on the data you build a product you build a service that gets immediate value to your customer right so something that he can use to change something he

does immediately not something that he needs to wait first one year for the business case to realize but something he can use immediately by doing that you bring it you convince him to give it to more and more users and by the the user start to use the product even more and more you gain traction for more data right so it brings you back into the cycle for us for example we install the acceleration sensors we give them information about when the switch is going to fail they come back to us and say ok we also have this maintenance history and we also have this type of electrical sensor of data which we look at other components what can you say about that right and it puts us in this strategic data acquisition cycle where we constantly get more data build by the product bring more users to the product right and in this case you kind of create something which is very defensible because only you have the learning with the data but at

the same time you're very focused on delivering customer value in a short period of time this is very very important to us and I can recommend you there's even a talk online by just Google and ung and I think it's one of his main things on YouTube and he explained this also in his words and I think this is one of the the key things when we talk about AI to be aware that you need to create your own data acquisition strategy framework and basically based on that deliver valuable products for customers and drive more and more users to those products now in our case we have a device right which looks like this we don't manufacture it ourselves we have partners to do that for us we just own the design of it and the kind of the whole IP around it and it's something we attached to the rail track right so in this situation you see it next to the blade making people complicated the switch has more components one of it

is a blade these are the sleepers right and then we install the sensor there the train runs over it and then what you get is something like this right so what you see here is an IC e and I see passes the sensor creates an acceleration trace and then out of the acceleration trace we can convert it in so called vertical displacement so that how much the track is moving up and down when the train is passing over and that is information we take to look at the condition of the ballast the condition of the ballast is how fixed are the stones that are sitting underneath the switch right data has shown us that if you have a good balanced condition the number of failure that US theaters which reduces by a factor of six so basically the first revolution was okay let's ensure that we always have a good ballast quality in this case we ensure that we have a low probability of having faders right and there's much more complexity behind that on

our failures kind of relate to each other and how they can influence each other but this is kind of the basics in simple sentence and then we take this parameter or indicator and then apply some predictive algorithms to see how it's going to evolve in the future of course is not really a linear prediction but you can approximate it to the different ways and then you can see okay when in the future do you need to fix the asset a number of things I want to say here and number one is that this is it looks as a very simple problem to solve right you have acceleration sensor you make it vertical displacement so it's just like double integrating it right lets me know from from physics and then you do some prediction but come on that doesn't look that complicated the thing is that as we all a situation like this the devil is in the detail and if you think about just converting acceleration into vertical displacement into movement if you have worked

with acceleration sensors the challenge there is that you always have a residual noise in there and this causes your signal to drifts right and even if the noise is very small once you double integrator drifts a lot right so what we did here is we trained I don't know exactly which type of neural network viewers I have to be honest but a relatively small one right because we need to be careful in terms of how well we can control the the quality output of the algorithm to just recognize okay where is the zero crossing right because the zero-crossing gives us an indication that where we can recalibrate the acceleration signal we it enables us to bring the drift back and by recognizing that in the acceleration trace we were able to set some reference points in the in the vertical displacement and then create this curve which is very very we've had to compare it with laser measurements right which do exactly the direct measurement of displacement and the difference is zero point one millimeter

right so that the closest you can get to it but basically the point I was trying to make is that we take an AI component and we place it somewhere in our processing pattern where we need it right so we we don't have the data to afford to go into deep learning because no one has seen this type of information before right there's no historic data we cannot compare it to anything there's no patterns we can look for but we need to think of K which of the things that we need to solve in our algorithm can we solve with small AI components like right now what we're doing is we're taking the acceleration trace and we're trying with detecting based on it which type of train is it right so we can give you in theory actually also in practice we can give you a full train schedule over a switch after one day we can say okay this was I see one that was I see - that was 94 that was the

original train that was a cargo train because if you look at the acceleration traces there's certain patterns in there which give you indication what type of training days again for that we use at the beginning we started using something like let's say a very simple how to call it like imagine it as a vector right which describes the characteristics of a train which is kind of the distance between the axles the ratios between the X's on so forth and we just place that vector into a space and then every new train was a new factory right and the closet it was the specific type of vector it was a specific type of train like this was the beginning way of doing it then afterwards we moved into more often let's say again neural network type of of training because we had more information of the actual schedule we knew exactly what type of train was running over it and it was easier for us to train and say okay if this is happening this is

what type of thing you're going to see so coming back to my initial point is we are always the fans of saying okay we don't want to just apply AI generic we want to look at specific components of a product and where can it make it can it be beneficial for us and then of course it's important to be able to predict in the future there's a couple of a couple of key metrics right specific components need a specific time so that you are able to react but the customer is able to react right because it doesn't matter if I tell you okay this is going to fail in five days if not able to react anymore and when we talk about balanced condition a huge let's call it mini train has to come out has a temple the whole track the whole track needs to be closed does a lot of lead processes which have to happen before the failure can be solved so there we need at least 90 days prediction horizon so

that they can plan ahead if something is gets detected at the same time we need a relatively high hit rate for something which hasn't been done before because the data was not there again because our customers cannot afford to just run a train of that on the track it impacts availability for them so it causes delay minutes it cost a lot of money like a temping measure cost eighty to one hundred thousand euros just just running this 10 pin train out of it and then they have to be sure that when they do it it makes sense right because if not they're just creating a lot of additional cost due to false positives that our algorithm is is delivering and 90% is just at the beginning now we're somewhere at nine above 95% which we have to deliver towards them so that they can guarantee the availability of the track so basically at the end what are we trying to do we're trying to make trains be more punctual why is that relevant with the

switches because the switches cause 20% of all delays so like 1/5 of all delay minutes is caused by switches and this is something with if you if you if you manage to solve that you make the trains more punctual secondly we want to have assets that last longer right because most of them get replaced based on cycles and knowing when a component has reached let's say end-of-life or a place in its lifetime where it's not usable anymore should be also driven by data right not just purely by cycles and third we want to make maintenance more efficient so one of the things which I don't have our slides right now but what we've seen also was there was a specific curve of degradation right of things happening and then we recommended the customer to do a specific type of maintenance so he went out did the maintenance he came back realized okay nothing changed right then we recommend it again he went out again came back nothing change so we could validate kind of in

a matter of let's say two to three weeks five different maintenance measures to see which is which is the most effective one before us or customers did a maintenance waited six months to see what happened and after six months okay they realized okay to have to do it again but in that six months the degradation of the asset just increased increased increased because the maintenance that was done initially was not the right one and we're giving them the option to validate maintenance measure because we also have to learn right which are the most effective ones but we give them the option to do that in a very short period of time they can iterate they can improve and then they can decide what's the most effective so basically as we put it always into the voice of our customers predictive maintenance would help us will help us intervene at the right place of time at the right place at the right time and with right methods yeah thank you and you guys know that deutsche

bahn has just said that they want to improve the punctuality a lot yes so I guess that's only because of you guys we we hope to be a big part of that I think we're all looking forward to that any questions from the audience it's very hard your flesh the way I would have won yes I think we had a similar one before so obviously you have to get the data and then you get some prediction out of the data and then you have to test it and you just said it how does it work with such a project I mean I would assume you don't roll it out entirely across the whole tracks so how do you go step by step what do you prove and so we for example what we did at the beginning we we took the whole the the customers like blow job and they have a history of failures right it's just a history of the asset and what happened to this asset and what the poor maintenance they

done now I not argue it if it's accurate or not but it's there so we took that and look at okay what are the most common types of failures how frequently do they occur and and what type of switches they cover right so this gave us an indication when we look at the assets that we want to equip how much do we need for how long period of time to be able to disperse it in specific phenotypes so then we went back and said okay this is the number of assets that we want to have and this is a period of time that we recommend for a project to be able to see something in the data right what happened afterwards was that actually you figure out that these things fail more often than the previous data showed right so we were able to see even in shorter period of time specific failures occurring which we were expecting maybe to happen only six months later so then when we came back to the customers okay

we see this this in this and it was like okay we haven't even thought about that like why is this happening now and then by that we get into a cycle where every time something's happening they call us and say okay so we've been out there this is what we saw what you see in the data and then we go back and check because we might have a normal switch you don't know what they are every time we see something we go back to them and say okay this is what we've seen the data can you please validate for us right and I think once you've established that cycle with a customer learning becomes more easy right because before that it's all about said she said and then it makes complicated to decide and just following up on that I was wondering if you might replace yourself somehow because since you are preventing all those failures you miss the data in the future of those failures happening so you can't challenge your models anymore and

then what happens have you thought about that or I might just continue out of space here I think it's like the way there is like fairly long right because you would have to ensure that all the maintenance is done 100% in time and 100% the right way and so on so forth and even that then what happens is that by not having our system you basically take your eyes out right if you remove it and then even if you would like to do it you don't know when do you want to do it right because you're not going to be able to adapt the cycles in a generic way that you fix all your problems because the situation is so dependent out there on temperature how the Train is running like what all the maintenance is being done and all these parameters which you cannot control you will always need a set of eyes in the field to give you the information about that and then secondly for acids the maintenance is just the beginning

right when we think about long term we see ourselves as a type of operating system for the food for the railway system right giving them information not only about maintenance but also understanding ok how does the train schedule how is it evolving right what does make more sense to automate to gain more availability and what things should we look into when we talk about planning new routes right planning new types of constructions right now we have a certain it's all driven today it's all driven by specific things that we want to achieve like we want to have deutschland talked right so that's why we will plant roots a certain way but by having more data from the field will inform that decision even even better in the future and there's a lot of data services we want we want to do on top of what we do today but that's we need to get there right we need to get more data more systems out there and it's a step-by-step process right that's one question

so our current situation today is that we monetize it as we call it system as a service right so we have a per asset price and then the contract the contact itself is always a bit different because these are customers who like to pay on a yearly basis who like to pay maybe significant parts of money up front and so on so forth so I'm not going to that details but basically the structure behind it how we come down to the price is looking at okay how much are we really saving like what's the impact of the solution that's a proxy that we need we use to compute the price and then we have yearly pay a fee per asset right so for every switch that we monitor we get a yearly payment conics for quite some time and I'm really impressed by your development and if I see the numbers and you're now raised well you have raised a bit more before but something close to 50 million with 54 people I know salaries

in Munich are like hell but what is your plan to utilize that money to grow in the exponential way to to argue for for such an investment which is really incredible I think it's one thing that we've always tried for for the beginning is that we wanted our investors to buy into our long term vision right and a long term vision isn't just we want of course we want to be a profitable company we want to be completely sustainable and that's all like big parts of what we want to do but we also want to have a big impact on whatever industry and it's obvious thing and everyone knows by now at least or investors a lot that this is a thing with long cycles right it's one thing to get into a deal that's the takes like almost two years but also changing a maintenance process these are processes which have been there for 25 years and all of these have take a long time to change so what we do we say okay

we need to be sure that we can manage this type of pace right we need to be sure that we don't end up in a situation where we need to raise money and there's also like all the deals almost being there so what we do is say okay how can we be the most efficient what with what our resources that we have and how can we use it in the smartest way possible to get us there so for example we race now the CVP but before we races here is B right and even after we raised it until the end of last year we haven't spent more money generation or Series a we kept everything and they're very close monitoring making sure that we are delivering like all the milestones that we want to achieve with the buffer that things might take longer right and in this case if a tender like just goes from a ten out of two years to four years for us that's like double or triple the runway right so

we need to make sure that we have this capacity to really survive the change in the pace of the industry and that's one of the key drivers in terms of finances like achieving the long term vision we need a lot of we need some fuel to be able to get there but we need to be efficient about how we use that money

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.