SDS 217: Aerospace Engineers and Data Science

Kirill Eremenko: Welcome back to the SuperDataScience Podcast, ladies and gentlemen, very excited to have you on the show today. You know how sometimes you hear the saying, “Well, this is not rocket science.”? When somebody is describing something that is supposedly easy or quite simple to get your head around, they might say, “Well, this is not rocket science. You could understand this. This is very easy to grasp.” Well, this time, it is rocket science because, on the show today, we have an aerospace engineer, Carlos Hervás García, who works for Airbus.

Kirill Eremenko: Very exciting episode, I had a great chat with Carlos, and what you will find out in this episode … well of course at the start we get quite carried away with the whole aerospace and orbital mechanics and the International Space Station, so quite an interesting chat, especially if you’re looking for a glimpse of the world of aerospace and what aerospace engineers actually do. But then in terms of data science, I found this episode very insightful to see how data science, machine learning, deep learning, and artificial intelligence can be used in the space of aerospace engineering. So what are the implications? What benefit do those technologies bring to this forefront of us exploring beyond our planet, to exploring interplanetary travel, and that is where Carlos actually applies his work, in interplanetary missions and interplanetary rockets and things like that. So it’ll be interesting what value those technologies bring there.

Kirill Eremenko: You will hear Carlos go specifically into two use cases. One for optimization with artificial intelligence, and the other one is for fault prevention detection and prevention, or error detection with artificial intelligence. I felt both use cases were very interesting. And of course we’ll dive into Carlos’ journey through his career. How he managed to combine his two passions, one for aerospace, one for artificial intelligence, into his one career and how he’s enjoying that, how he’s going about innovation at Airbus, how he’s helping the company succeed and thrive in the space of artificial intelligence. Very inspiring episode, can’t wait for you to hear all the amazing things that Carlos had to share. And without further ado, I bring to you Carlos Hervás García, aerospace engineer at Airbus.

Kirill Eremenko: Welcome back to the SuperDataScience Podcast, ladies and gentlemen. Today we’ve got a very special guest on the show, Carlos Hervás García calling in from the UK. Carlos, how are you going today?

Carlos Hervás: I’m good, thanks. I’m very good. How are you?

Kirill Eremenko: Very, very good as well, thank you. We were just chatting about the weather just before. I’m in Australia, you’re in the UK. It’s nice and sunny, beautiful start of summer here. How’s things in the UK in terms of weather right now?

Carlos Hervás: Not here, I’m afraid, but wet. It’s fine. We just have like nine months of weather now. So just fine.

Kirill Eremenko: But I heard you had some very nice couple of weeks during the summer with some sunny weather. Were you there for that?

Carlos Hervás: Yeah, yeah. This year we cannot complain, definitely. We’ve had three months of uninterrupted sun. That’s something I’ve never seen in the UK.

Kirill Eremenko: Three months.

Carlos Hervás: Yeah. Three months. It’s unbelievable.

Kirill Eremenko: Wow, amazing. Well that’s very, very exciting news. Okay. Well thank you for coming on the show. It’s a great honor to have you. You’re an aerospace engineer using deep learning and machine learning in Airbus. I’m very looking forward to digging into your career. How are you feeling about the podcast?

Carlos Hervás: I’m really excited. To me it’s an incredible opportunity to be talking to you, because, I mean, you would be surprised but I took your courses once upon a time. So this is really, really exciting for me, and then, yeah, I’m really happy to share with people what we’re working on and pushing forward.

Kirill Eremenko: Thank you. I want to say to the listeners that just before we started the podcast, we were chatting with Carlos, and he mentioned exactly that, that he had taken our SuperDataScience courses about a year ago. And I was so shocked because … I’m so impressed by Carlos’ career and you can, if you go to the show notes you can find his LinkedIn there, and have a look. Carlos, you’re like the leading machine learning and AI … you’re leading machine learning and AI applications, including RNS, LCM’s and [inaudible 00:05:39], so you’ve got a very impressive background. But I guess it all points us back to our humble beginnings that we all learn regardless of our … and that’s what you said, right? There’s always a desire and need to learn, regardless of your level. So I really appreciated that comment. I think that’s exciting for everybody out there.

Carlos Hervás: Well, it’s only some months ago that this move has started as well in the UK, at least in Airbus I mean. In other parts of Airbus, there are other people investigating and we are now a very well-established community. But yeah, in the UK, I saw this opportunity and I didn’t hesitate, I went for it. And it’s been a very intense year with a lot of work on my side, but very rewarding at the same time because I’ve been able to do this on your top and applied space to me, that’s extraordinary. I’m really happy about it.

Kirill Eremenko: Awesome, awesome. Just for the sake of our listeners, I want to clarify here. So Airbus is a big company, we all know Airbus. And working in aerospace, Carlos inevitably is working on projects where he won’t be able to disclose all of the details. So please forgive us for situations where Carlos might have to say sorry I can’t talk about that and we’re going to be extremely careful about that because we want to preserve the privacy of Airbus and all of their … the work that you’re doing. But at the same time we’d love to learn from your experiences in how you approach challenges in AI, how you learn yourself, and how your career went about. I think that’ll be fair for everyone.

Carlos Hervás: Yeah. Thanks very much, Kirill.

Kirill Eremenko: Thank you. Thank you. Alright. Let’s get started. Tell us a bit about … we know now that you’re an aerospace engineer. Tell us what does an aerospace engineer actually do? We’ve heard these … there’s this term before, I think everybody has their own conception of what’s an aerospace engineer does on a daily basis. What does your job look like and what do you do on a daily basis?

Carlos Hervás: Well, that’s a complicated question. So aerospace engineering, I’ll tell you, is very, very broad. It covers, or at least when you study in Spain, and even in France, because I studied in the two countries, it covers really a lot of disciplines. Ranging even from airports, aircraft engines, space, missiles, well all these fields belong to aerospace. So you obviously need to specialize to some extent, otherwise you’re a bit useless in the domain. Myself, I went for space and control engineering, and then when I joined Airbus, I did it on the space systems branch, which is maybe not so known in Airbus worldwide, because we’re mostly known for aircraft. But it’s also a very big part of Airbus and quite an important one. Also a leader in the world. And then I joined to space system division, where there’s also a big variety of things to do. So you can do from telecommunications satellites to Earth observation missions, also interplanetary missions. There used to be a part doing launches, as well. Now it’s a joint venture with another company, but there are many, many fields. And even within those fields, like I am working for the spacecraft, more precisely for interplanetary missions, I’ve been working –

Kirill Eremenko: Wow. That sounds like sci-fi. You’re building spaceships to go to other planets!

Carlos Hervás: Well, it’s actually … we call it interplanetary when the spacecraft travels around the solar system, and just reaches it’s ultimate target, which is not Earth, basically. And then, even within that, there is many different skills that are needed. So spacecraft has many different subsystems, so someone has to design the structure, the mechanical structure that will support the spacecraft. Someone has to design the propulsion models. Someone has to design the communications. Well when I say someone, I don’t mean a person, I mean a team. And then myself, I worked in the AOCS department, so that stands for Attitude and Orbit Control System, which for people to understand, is roughly the autopilot of the spacecraft. So you cannot drive or steer the wheel in the spacecraft. You need predefine what it has to do as a function of what it is and what it can see. This is the algorithms that we design and tune and validate ourselves. So that is what I have been doing for the past three years, I would say. Three and a half years.

Kirill Eremenko: Okay. That’s very exciting. You actually reminded me of … do you know Chris Hadfield, the Canadian astronaut?

Carlos Hervás: Yeah, it rings a bell.

Kirill Eremenko: I was taking his course on … where was this? On Master Class, masterclass.com, he has a course about space, space exploration. And one of the tutorials, he was describing how much resistance an aircraft experiences as it goes through the atmosphere. He mentioned the formula, I forgot it, it has the [inaudible 00:12:03] V squared S, or something. Do you know that formula?

Carlos Hervás: Yes. That’s the dynamic pressure of the air on you.

Kirill Eremenko: Yeah, is that correct?

Carlos Hervás: Yeah. [inaudible 00:12:14] coefficient, which is the drag coefficient, or when you’re looking at the thing that lift you up, then it’s the lift coefficient. I hope I said it right in English, because I studied this in French and in Spanish, but not in English.

Kirill Eremenko: Yeah. It sounded right, exactly. But you’re just talking about how, as a spacecraft goes through air, and the amount of drag it experiences is proportional to the density of the air, the area of the spacecraft, and the speed that it’s going, and I think it’s [inaudible 00:12:50] V squared S. But as you go higher up, there’s less and less density of air, and then in space, it’s a completely different story because there’s no drag. As an aerospace engineer, you have to take the whole thing into account, right? Before, as the rocket goes through … but in your case then if it’s like a satellite then it’s already deployed in space, and then you just need to calculate the situation when it’s in space, right?

Carlos Hervás: Yes, that’s correct. So basically even that could seem like very similar thing. We actually split it in two. So the long-term will be the bit that takes the spacecraft in outer space, so where there’s almost no atmosphere, so almost no drag. If you’re in low Earth orbit, then the drag is still a bit there. But it’s nothing like on a plane or on Earth. And you’re very, very far away from Earth, then there’s no drag, basically. But yeah, for these interplanetary missions, our mission starts once the launcher has put us in low Earth orbit. So we don’t necessarily care about drag that much. So we don’t need to design with that.

Carlos Hervás: But there is, if you wanted to reenter Earth, or even land on a planet with a dense atmosphere, then that’s a whole different story. And it’s quite challenging.

Kirill Eremenko: Just speaking of Earth orbits, I think it’s very interesting to know that the International Space Station, it feels like is up there in space, but it’s actually only about 400 kilometers from the surface of the Earth. And if you compare it to the radius of the Earth, the radius is about 6,400 kilometers, so it’s less than ten percent up. So if you take a globe, the international space station’s very close to the top of the Earth, but already there is so little atmosphere, there’s practically no atmosphere there. In fact Chris Hadfield says the international space station experiences drag equivalent to the weight of a piece of toast, like if you put a piece of toast in your hand, that’s how much drag the ISS experiences. What is a low Earth orbit? How high is that?

Carlos Hervás: Well, you can have spacecraft, I think maybe down to 200 kilometers up off of the surface. Obviously the lower you are, the more drag you have, and the sooner your satellite or spacecraft will decay. So eventually the drag will become so important that the velocity of your orbit … or rather your spacecraft will lose velocity in the orbit, so the orbit will shrink, and eventually you will reenter the Earth. Unless you need it, you don’t necessarily want to be very, very close to Earth, otherwise you need to compensate for the air drag by correcting your orbit with delta-V’s, and you’re using propellant.

Kirill Eremenko: You’re wasting fuel.

Carlos Hervás: The lower you are, the more propellant, yeah. The more propellant you need.

Kirill Eremenko: And do you know what I find really fascinating? It’s orbital mechanics. The whole notion that in space, if you’re going around the Earth, and in order to go to a lower orbit, you have to go faster. In order to go to … you have to be moving faster at a lower orbit. And if you go to a higher orbit, you have to slow down. So to move between orbits you actually either have to accelerate or the other way around. So to go to a higher orbit, you would have to shoot your thrusters in the opposite direction to decelerate and that would take you up to a higher orbit. Like that is crazy when you think about it, like how do you get your head around that?

Carlos Hervás: That’s actually slightly trickier, the math. It’s actually not that long ago that AI democratized to industrial applications. And for aerospace, this is even more true given it’s particularities. Given my passion about machine learning and my [inaudible 00:17:24] aerospace, just maintains for me to [inaudible 00:17:28] into it and be pioneering in this aspect. I was very likely to pick even the opportunity to do so. Well it took a lot of effort from my side and dedication, but definitely worth it. So say you want to go from a low Earth orbit to a geostationary orbit, so these are very particular orbits where the period of the orbit is similar to the rotation of the Earth, so 24 hours. So if you wanted to go from a low Earth orbit to that orbit, then you would need to put a positive delta-V on your vehicle to get there. So you would do, normally, now I’m not an orbital person, so I hope I don’t say anything [inaudible 00:18:23] more like attitude control, etc. So then what you would do is you accelerate at a given point, you raise the apogee of your orbit, and then … so you would go from a semi-circular or actually almost circular orbit to a elliptical orbit. And when you are on the far end, on the apogee, you need to again fire to circularize the orbit.

Carlos Hervás: So it’s a two-steps maneuver where you increase your speed at two different times. I think, if I remember work from school, that’s a Hohmann Transfer, which is the most basic maneuver you can do.

Kirill Eremenko: Wow, very interesting. I’ll have to read up about that, that is so cool. I find it very, very interesting. Very different, as well.

Carlos Hervás: By the way, just to point out, that this … in our department, it’s actually divided into guidance, navigation, and control, attitude and orbit control systems, and mission analysis. And it’s the people from mission analysis that do all this orbit data [inaudible 00:19:44], and by the way the capabilities in the UK are absolutely unique in that sense. So we’ve done the preliminary phase for most of the European Space Agency interplanetary missions, like BepiColombo, or in solar orbiter, things like that.

Kirill Eremenko: Wow, very impressive. Very impressive to hear. We got a bit carried away with all the orbital mechanics and space stuff. Let’s talk a little bit, or let’s direct our attitude, the conversation towards data science. So tell us a bit about how, to the extent of course that you can disclose, how, in all of this, do you apply, or do you leverage, machine learning, artificial intelligence. You mention on your LinkedIn that some of the things that you use are RNN’s, recurring neural networks, LSDM’s, long short-term memory, then [inaudible 00:20:41] encoders, genetic algorithms, you use Python, [inaudible 00:20:44], all these very advanced tools. How do they help you in your work?

Carlos Hervás: All these techniques are very useful to us in many different ways. So for instance, very obviously application cases, RNN’s, so recurring neural net, with their sense of state or memory, that’s very similar to the dynamic systems that we work with on a daily basis. So dynamic system have also states, and therefore RNN’s are a very good representation of dynamic systems or anything that you’re trying to control. I mean if you’re trying to control a dynamic system, well using RNN’s is something quite normal. Then also other axes of investigation are like meta models, so we have very complex models out there that we’ve used for simulation purposes, for design purpose, etc. What if you are able to capture this complexity and just run them on neural nets, on [inaudible 00:22:02] neural net, which are less computationally involved?

Carlos Hervás: Other applications, for instance, could be tuning, so you have a given algorithm that you’ve designed. You use some artificial intelligence techniques to tune those algorithms, to choose the values that make your pre-defined algorithm work better. Also another very obviously application would be … well, today, we basically design and tune our algorithms, but the machine learning in itself offers techniques to just declare some algorithms and then by using the data or the simulators or whatever, the algorithms are actually learned as opposed to specified. So as you can imagine, that’s also a very powerful approach that we’re investigating. For instance, one of the videos that got me really into wanting to investigate these techniques is a very simple and dummy video of the Atari game that probably most of the people here in this podcast have seen already. I think it was Deep Mind that was trying to beat a Atari game. And then you could see in the video, after 200 episodes, the policy was quite bad, then 400 episodes, it plays like me, but because I’m very bad, then 600, the neural net does something unexpected and [inaudible 00:23:53] on the left, and then gets the maximum score out of it.

Carlos Hervás: So we’re just hoping that perhaps, for some of the challenges that we have today, AI could find smarter solutions than what we do. Or not even smarter, but there are solutions that we have to [inaudible 00:24:16] because their complexity [inaudible 00:24:22], tuning, and validation. So perhaps in the future, this will not be a [inaudible 00:24:30], the AI or the neural net learns it by itself. So we’re working on that. That’s one axis. Another axis is a very big axis of big learning today, which is anomaly detection. So our spacecraft are incredibly complex machines that need to be operated, and you can imagine that you cannot interact with them like you do with a car. So they need to have a certain degree of autonomy. And they also need to be very safe. So they all have a subsystem that we call FDIR, so failure detection, isolation, and recovery subsystem. One of the things that we’re investigating is a new way of doing FDAR using these techniques. And that’s where most of the work has been focused so far. And this is where we’ve been trying different techniques like LSDM’s, [inaudible 00:25:39], even [inaudible 00:25:41]. Different techniques just to see if we can come up with more efficient and more robust engineering solutions for big [inaudible 00:25:54].

Kirill Eremenko: Okay, wow. Thanks. That’s very cool to hear. I really liked how you described, how you compared what you guys do in the space of optimizing different algorithms, different solutions, using artificial intelligence, how you compared it to using AI to win Atari games. Because ultimately indeed, it’s a fair comparison. It’s a very simplified approach, a very simplified problem when you’re trying to use an AI to beat an Atari game. But if you think about it, what you guys are doing in aerospace is you have certain problems, certain challenges, that you are finding solutions to, well how about we get an AI and train it to … just like we train an AI to play an Atari game, how about we train an AI, even in a simulator, doesn’t have to be a real rocket. You can’t afford to crash a real rocket, right? You gotta sometimes –

Carlos Hervás: Yeah, absolutely.

Kirill Eremenko: Use simulators. And then you train in a simulator to see what happens if it solves a problem this way, like it pulls these levers, or it opens the gas at this time, or during launch this valve is unavailable, or is used to 30% or 40%, what happens. Like it tweaks all the parameters, goes through multiple simulations, hundreds of thousands, as I imagine, and then it might come up with a better solution than humans come up with. Is that about right? Do you guys use a lot simulations for all these things?

Carlos Hervás: Yeah. That’s exactly the point. If you realize part of our work, once you’ve done the design, the conceptual design of the algorithms, and even the tuning, then the rest is, okay, simulate it, and let’s see how it goes. And that’s almost like playing a game. You have a simulator, in our case it’s not the flight simulator that most of people play with, but it’s a real simulator of the spacecraft where we validate all our algorithms. So why not using this simulator as a learning environment to our AI? And that’s really one of the axes, yeah.

Kirill Eremenko: That’s very cool. If you don’t mind me asking, how long do these simulations usually take before you can get some kind of … like with an Atari game it might take a day, a few days for it depending on the strength of your processors or you GPU’s and things like that, couple days before you get an answer. Do you end up running something for a year and waiting for results? On average, how long is the iteration process?

Carlos Hervás: Well, this is something, it’s a very difficult question to answer, and in fact I don’t think it has an answer now. Because it really depends on how ambitious you want to be with your AI. You can think, so the software, the onboard software that you put in spacecraft is massively complicated with many, many different functions, doing different [inaudible 00:29:06], and functions or algorithms or whatever you want to say. So you could, in theory, or in principle, you could replace all those algorithms by just one [inaudible 00:29:22] then train it from end to end, and that would take ages. We’re not at that point yet, because we’re trying to apply AI in a way that actually … in an industrial way. Not in a academic way. I don’t know if you see what I mean. So we really focusing on getting things to the level that we can actually use, rather than just impressing the world, let’s say.

Carlos Hervás: If you’re super-ambitious and try to replace every bit, I don’t know, that could take … it’s something we’ve not tried yet. Now if you are a bit more humble, at least for now and do it with certain bits, or certain functions of your spacecraft, the ones that maybe are trickier, or … the ones that are trickier, then we are talking about maybe five to six days of training, something like that. It does also depend on really what is the neural network [inaudible 00:30:41] behind what you’re trying to do.

Kirill Eremenko: Gotcha [inaudible 00:30:45]. Do you think, this is kind of a more futuristic question, do you think … you mentioned at this stage we have to be a bit more humble and build AI for specific parts of the spacecraft, specific tools inside that it can control rather than controlling the whole thing. But do you think, from what your experience in the industry, do you think that in the future we’re going towards something like something we see in Star Trek, where there’s one AI that is responsible for the whole spacecraft, and that it can talk to humans, and that it can open doors, and switch the air con on and off, and control thrusters and whatnot, as if the spacecraft has a mind of it’s own. Do you think that’ll ever happen?

Carlos Hervás: Well, I don’t know. To me, that’s the version on space of the almost perfect robot on Earth that will be able to talk to you and will be your companion for life. To me that’s as far as … or even more because space is a very conservative world thus far, so I could see how even on anything we have, this general artificial intelligence and still on space we are [inaudible 00:32:04] wait, it is not validated yet. So I think that’s very, very farfetched.

Kirill Eremenko: Alright. So first it has to happen on Earth, then it will happen in space?

Carlos Hervás: Well, these days, this seems to be a bit trendy, yeah.

Kirill Eremenko: Gotcha. Okay. Then moving on to the second application, or your second use-case you mentioned, which is anomaly detection. You mentioned a abbreviation, FDAR, was that fault detection and, what?

Carlos Hervás: Fault detection, isolation, and recovery.

Kirill Eremenko: Ah, okay. Fault detection, isolation, FDIR, and recovery. Okay, so that actually reminded me of, in the artificial intelligence course, when we … oh, no, in the deep-learning course, when we were creating it, there is a type of deep learning called Boltzmann Machines, and the use-case that is described for Boltzmann machines is actually similar, so it’s anomaly detection but at a nuclear power plant. And for our listeners who haven’t taken the course, I’ll just quickly describe, you could create a Boltzmann machine, it’s kind of like a deep-learning neural network, but rather than having an input and output, it’s just a deep learning … imagine like a circular deep learning neural network where it’s kind of like all interconnected, like you have on a circle, on an outer ring you have all these neurons, and then they’re all interconnected. And then that is used, rather than putting inputs into the neural network and then getting outputs, that is used to just generally model, every neuron models a parameter of the nuclear power plant, and then the neural network learns through observation how they interact together. And so it knows what are the normal states of this power plant and what are, therefore it also knows, what are the abnormal states. And as soon as something becomes abnormal, it can trigger an alarm very quickly.

Kirill Eremenko: I’m assuming the likelihood that that is exactly how you use neural network is very low, but at least I can see that there is already a way. Like I know in my head that there is a way that you could use neural networks for anomaly detection. Are you able to disclose any general value, any general value points that you get from using neural networks in anomaly detection? I know you can’t go into details on the exact algorithms of what you do, but in general, how do they help you with anomaly detection? Is it because they can process way more data and you don’t have to describe every single thing? Or is there something else?

Carlos Hervás: Okay, without going into much detail, so the general principle behind doing FDIR with deep learning is the fact that you can teach, or the neural net can learn what is nominal, so what is the nominal behavior of your equipment in space. And this can be learned by simulations, that anyways they are done for other purposes. So you can use all that data, which is massive, and just train neural nets on what is nominal. And then you use the neural net to flag when something is not behaving as seen in your envelope of simulations. So then there’s many ways, many different ways in which you can do that. The one you mentioned is a very valid one. There are, even recently NASA’s published papers where they do anomaly detection with LSDM’s, so future prediction and some sort of anomaly score, elaborated anomaly score. Then there are auto-encoders. There are several papers on how to do anomaly detection with auto-encoders. Also recently with [inaudible 00:36:17], etc. So there are many ways in which you can do that, but the main principle, or what we ultimately want to achieve, is something that is able to first detect that something is wrong, something is inconsistent inside your system, and then ultimately also isolate it, so be able to pinpoint where the source of the error is. And those two things are quite important.

Kirill Eremenko: Gotcha. Do you guys, moving a bit away from deep-learning. Do you guys ever use Byzantine fault tolerance?

Carlos Hervás: Sorry? To use what, sorry?

Kirill Eremenko: Byzantine fault tolerance.

Carlos Hervás: I’m not familiar with the name.

Kirill Eremenko: Okay, no, no worries. I just heard that … it’s kind of like a system that, when you have lots of different things in your circuit, lot’s of different, I guess, algorithms, and also different tools working at the same time, there’s a fault tolerance where if one of them fails, the whole spaceship still keeps working, or if one of them is saying the wrong thing, it still keeps working. I just read somewhere that the International Space Station has a requirement for at what level of fault tolerance does a spaceship have before it will allow it to dock with the International Space Station.

Carlos Hervás: We do work with that. So I just didn’t know the name. So the most common solution today for FDIR is sort of an extra system where you define, you do an analysis of what can go wrong, so you do a list, you make a list, so you go through your equipment, so sensor actuators, you define your mission availability requirements, your fault tolerance, degree of tolerance you have to have against [inaudible 00:38:29], etc. And then based on that, you start mapping on what are the type of failures that cannot cure, and what will be the impact on the system. So what do I need to monitor in order to pick up this failure and reconfigure so that I’m compliant with the safety and availability requirement. And that’s how you build your thing. So you can imagine this is a very complex thing to do. It’s complex and critical, as well. If you make a mistake there and then something unexpected happen, then you may not be covered by your FDIR design.

Carlos Hervás: One of the benefits of deep-learning here is that … or at least the methods we’re working on, is that they’re agnostic to this. So they just know about what is nominal, and if something isn’t nominal, then it will flag it. So this bit of change in the mindset, because you no longer need to worry about all the things that can go wrong, but you know that everything that is not nominal is wrong. There is even a safety change in mindset.

Kirill Eremenko: Gotcha. That’s very cool. And when you say nominal, you mean it’s the normal, like whatever is nominal, you mean it’s normal, and whatever is not nominal is not normal, therefore is wrong?

Carlos Hervás: Exactly. Nominal, so we use nominal in space to refer to the set of conditions that the spacecraft will operate normally. If you deviate from that, then something … well it’s not necessarily wrong, but is unsafe or unexpected, and therefore we would rather be careful about it, just in case we didn’t design properly for it.

Kirill Eremenko: Gotcha. That’s very cool. Well thanks so much for sharing those two main use-cases of deep learning in aerospace, so the optimization with artificial intelligence and anomaly detection FDIR. What I wanted to talk about next was your path, tell us, did you become an aerospace, like I’m just … such a interesting intersection of fields. Aerospace and artificial intelligence at the same time. So did you first become an aerospace engineer and then get into the field of AI, or the other way around? Were you into AI and deep learning first and then you went into aerospace engineering?

Carlos Hervás: No, actually I’m an aerospace engineer because [inaudible 00:41:25] I was really fascinated by … well I always wanted to be an astronaut first, and then as I grew, I started to understand that this may be a difficult to become an astronaut, but I still liked aircraft and space flight, and so on. So I went for it. Also in Spain is a very nice degree to have, so it opens the door to many possibilities. So that’s what I did. And then I went to Toulouse because it … the University there is also quite well-known for its links to industry and companies like Airbus, Thales, etc. I knew I wanted to be [inaudible 00:42:13] in … or to work for a big company that can work on these missions that have an impact on society. And then I made it, I made it there, I joined the UK straight after from uni, and then I spent some, good three years, three years and a half immersed in actual development work of interplanetary missions. So working for solar orbiter for three years. And then [inaudible 00:42:48], so this also a European Space Agency mission that is [inaudible 00:42:55] in France, but we do in the UK certain bits of the AOCS.

Carlos Hervás: And then after that, I was working a bit on the R&D strategy for the department, etc. And in my mind, at the same time, I was a really enthusiast about AI and machine learning, and I was really wanting to know more about it. So I kind of [inaudible 00:43:24] in my spare time. And then I was like it would be great if we could use these techniques as well for what we do, because I think they have a lot of potential. So then I started to come up with ideas on how we could use AI on this. And then Airbus is pushing now for a lot of innovation in the company just to make sure that we keep up with the pace of the world in terms of innovation. And there are very frequently open innovation calls.

Carlos Hervás: So in one of these innovation calls, I submitted one idea, and I got funding for it. So I started to work on it and this is … this was the first time that we were applying deep-learning techniques for engineering solutions in the UK. And that’s how it all began. Then after that, the interest and the hunger for AI has kept increasing. And then there was another innovation call in Airbus, they called it the AI campaign, so just a call for ideas on how we could use AI to make our products and services better, and our engineering better, and so on. So I submitted another idea, and I got the funding for it. So now basically I’m just steering these ideas and basically I made it to combine this LEOCS world that I truly love with my other love, which is AI and machine learning, and that’s really how I got to this situation.

Kirill Eremenko: Very inspiring. I love that journey, that path I got. Seriously, it’s so inspiring and I’m very excited for our listeners, because even if you’re not a hardcore data scientist and that’s all you do, machine learning, AI, if you are in a company, because there are lots of people listening to this who are just like curious about the space of data science and how to potentially build a career in it, but they don’t know really how to get into that, or how to get a job in that space, or how to bring that into a company. And this is a very good testament to that. Airbus, they didn’t come to you and say “Hey, Carlos, can you create some AI for us?” They just said, “Okay, here’s an open call for innovation, what can you come up with?” And that’s when you leveraged those things that you were exploring on the side on your own. That’s what I want to say to everybody listening to this, that if you have a passion for technology, for AI, deep learning, data science, machine learning, and you are hesitant about how you can use that in your work, most likely there are ways. There’s like a 90% chance there is something you can bring your company.

Kirill Eremenko: Companies like Airbus, I know when I worked at Deloitte, they had open calls for innovation. Large companies like that, they usually have these innovation labs, or innovation challenges, innovation brainstorming sessions, where you can bring your ideas. You just have to think, what is it that I have learned, how can I apply it in my work? How can I apply it to make this business better? Because if you work in a company, you probably love that company. You probably love what you’re doing, otherwise you shouldn’t be at that company. Like in your case, it’s obvious that you’re passionate about what you’re doing.

Kirill Eremenko: And then the other thing is if you work at a smaller company, like a mom and dad bakery store, or, I don’t know, like some smaller, even a pizza place or something like that, you can still … you don’t have to wait for an open call for innovation. If you have an idea, you bring it to your boss, you bring it to your manager. If it’s your business, you bring it into your business, to your partners or whoever, and you decide. You show that you can add value. It’s so easy these days to use these technologies such as artificial intelligence or even just simple machine learning to do a customer segmentation analysis, or analyze the workflow and see where the bottlenecks are and how you can improve them, or improve the product and monitor quality, improve the service, find out hours operation, where you’re analyze financial statements, and things like that. There’s lots of ways you can add value. And if you’re truly passionate about both, about your job and about what you’re learning on the side, it would be so cool to combine them. So I really admire this example, Carlos, that you gave us, where you combined two passions.

Carlos Hervás: I think there’s also another element that I perhaps not stressed enough. But I think we’re not at a point where there’s a lot of AI developed at companies or by people that are really knowledgeable about AI. But to me there’s a step between that, which is now used to solve impressive buy maybe not so meaningful problems, so that is that, but then there is the part where actually comes here to help solving real problems. And I would say for the type of things we do in Airbus, so not all of them, but many of them, they require as much AI knowledge as they require domain knowledge. What I means is if you try and build an anomaly detection deep learning solution without knowing what your equipment, your sensors actuators, what is behind all of those things, you would probably end up doing something maybe even not working or if it works, it would work not in a good way. So I would stress the importance today of having this dual knowledge of domain knowledge and AI skills in order to successfully apply AI in your work.

Kirill Eremenko: That’s really cool.

Carlos Hervás: I just thought it was important to stress that, because I just find that otherwise I would not have been able to do such a thing.

Kirill Eremenko: Yeah, totally agree. And I like what you mentioned that there’s a difference between artificial intelligence, that is the fancy AI that ends up in the news and that is developed mostly for academia which is really cool, but not really applicable to real-world challenges. And that type of AI versus artificial intelligence for business. AI for business is what is actually driving companies forward, changing the world, is the hands-on stuff. And it doesn’t have to be as fancy. It just has to … you’re right, if you have that domain knowledge, you combine it with the more simple and more accessible artificial intelligence, which you can do with Tensor [inaudible 00:50:58], and so on. And you can just like, even to the level of almost drag-and-drop. It’s kind of like … okay, it’s not drag-and-drop, but it’s add five lines of code and you have a recurrent neural network set up. That’s not really hard. You can learn that in a book.

Carlos Hervás: That’s how it worked for me.

Kirill Eremenko: Exactly, and that’s the fascinating part. We have an aerospace engineer on the podcast right now, and you’d think that this is the most complex that it gets. This is aerospace. This is where, rocket science. You’d think that you are … this is the time that you would need artificial, the most advanced artificial intelligence in the world. But no, even in aerospace, it is sufficient to have domain knowledge and then go out and learn some artificial intelligence, not the super-fancy, cutting-edge artificial intelligence in the world, but just artificial intelligence like RNN’s, LSTM’s, CNN’s maybe, and things like that, and apply it and see what happens. That’s very inspiring that if you can get such great results with this not super-fancy artificial intelligence, but like basic, very well-developed, well-tested artificial intelligence, then pretty much any other business in the world, or 98% of other businesses in the world should be able to replicate your success.

Carlos Hervás: Well, thanks very much, Kirill. But yeah, I think what you’re saying is true and basically with this combination of domain knowledge and artificial intelligence, I mean as you say, even if you were to apply the most or the fanciest AI on this domain, you would probably not have … were not be able to fly it at any point, because you need to build credibility, you need to build customer acceptance, let’s say. You need to go progressively on this, so you cannot … just step jump and say this is our formal solution where it’s all hand-crafted and everything and this is our new solution where you don’t understand the [inaudible 00:53:25], complete black box to you and you have no … so yeah. I think going step-by-step and this mixer between AI and domain knowledge is really, really important.

Kirill Eremenko: Gotcha. That’s very inspiring. Thank you so much for sharing that. We’re slowly coming to the end of the podcast, but let’s do a quick rapid fire questions so we get a better feel for what kind of tools you use and things like that, and algorithms. You ready for this?

Carlos Hervás: Yeah.

Kirill Eremenko: Okay, cool. So what tools do you use on a daily basis? Like software tools, programming languages, deep learning, AI, type of tools.

Carlos Hervás: Okay. So my aerospace type of tools are mostly MATLAB Simulink for everything that is analysis and modeling, and then for the AI is mostly Python with Tensor Flow, Scikit-Learn, [inaudible 00:54:30], and that’s it.

Kirill Eremenko: Okay, that’s cool. What techniques do you use most commonly? You mentioned LSTM’s, RNN’s, can you just give us the full list of your favorite techniques in this space?

Carlos Hervás: So when it comes to time series, obviously we tend to go for RNN’s. So either LSTM’s or GRU’s are often the prepared solutions. Then we also work with convolutional neural nets when trying some general truth adversarial networks [inaudible 00:55:15], as well, so having perhaps a discrete [inaudible 00:55:20] that is happening off CNN and a [inaudible 00:55:26] neural net at the end of it. Then also a simple concatenated dense layers for simple tasks have proven to work appropriately. That’s pretty much it. With time series, which is 95% of work, I would say.

Kirill Eremenko: Nice. Cool. So moving on to more experience and career type of questions. What is your one most favorite thing about being in the space of data science, in the space of artificial intelligence?

Carlos Hervás: Well to me, it’s really the power of what you can achieve, and also the feeling that it’s something completely … it’s changing mentalities. So I like to compare it to when people did missions to the moon just using paper and a pen, and then the computers were introduced into our daily jobs and that completely changed it. So I have the feeling that AI could become this other step change in our engineering world, and now very few people have AI that they use, so maybe in some years time the engineers will [inaudible 00:56:58].

Kirill Eremenko: Gotcha. And what is the thing that you’re looking forward to learning next? Is there like one big thing, one big challenge in your educational journey about artificial intelligence that you’re really looking forward to?

Carlos Hervás: Well, we’ve started investigating reinforcement learning not so long ago. My big challenge, my big personal challenge, is to dig into this field and take it further, and really make it a suitable paradigm for our job and our algorithms. So to me the main challenge is reinforcement learning.

Kirill Eremenko: That’s a really cool one. Finally, from what you’ve seen about data science and artificial intelligence, where do you think this field is going? What should our listeners look into to prepare for the future that’s coming ahead?

Carlos Hervás: I think the way I see it is AI and machine learning is being democratized in the last, I would say five years with the introduction of things like Tensor Flow, and more accessible computational power. Also all of the explosion on the internet of informational courses like the one you did, etc. Now the information is starting to be more available and that’s precisely how I got to jump on this. So to me the future is really about making not a few people but most of the people acquainted and familiar with the tools so they can incorporate them on their daily work and they can contribute and come up with solutions on AI. So the fact that you have it used by more people increases the possibilities that you can do with it. So this is how I think this is going to evolve, and of course it will be more and more like [inaudible 00:59:14] research, but to me the main change will be when instead of one person [inaudible 00:59:21], maybe 40% of engineering population will be able to do AI on their daily jobs.

Kirill Eremenko: Wow, and how far away do you think that is?

Carlos Hervás: Well, I don’t know, I hope in Airbus, so we are working for it and I’m going to be devoted, as well, that’s one of my challenges as well, to really be able to convince people at my level but also top management that we need empowered engineers with these tools so that we can all benefit from that. My hope is that, I don’t know, within two, three years at least, having people capable of doing this in every department. That will be an ambitious but very, very good timeframe for it.

Kirill Eremenko: Gotcha. Well Carlos, thank you so much, that sounds like a very bright future, and thank you for the rapid fire answer questions. We’ve slowly come to the end of the podcast, and really appreciate you sharing all the insights. But before I let you go, I would like to ask where can our listeners get in touch with you? What’s the best way to follow your career and find out what are some of the next amazing things that you’ll get up to?

Carlos Hervás: Well, I would suggest if you want to get in touch with me, do it by LinkedIn. I tend to answer everyone. I’m afraid I’m not such a Facebook user, so if someone contacts me by Facebook, it’s very likely that I won’t reply in two months. So LinkedIn is okay. Don’t be shy, if you have interesting things to say, if interesting feedback, or whatever you want to say, just get in touch with me.

Kirill Eremenko: Gotcha, fantastic, thank you. And I have one last question for you today. What is a book that you would like to recommend to our listeners for them to empower their careers?

Carlos Hervás: For me, if you really want to make it, to go through this learning curve, I think the book that helped me the most was Hands On-Machine Learning with Scikit-Learn and Tensor Flow. I think I bought it maybe seven or eight months ago and it’s got directly some pieces of code that I could implement directly and they would work. And he also explains very well, gives a good overview of different techniques. It comes with some application examples. It’s really well written, so I strongly recommend it.

Kirill Eremenko: Gotcha. Thank you. So that’s Hands-On Machine Learning with Scikit-Learn and Tensor Flow. Once again, Carlos, thank you so much for coming on the show today and for sharing all of your amazing insights. And such a great and unique story of aerospace and artificial intelligence. Thanks so much.

Carlos Hervás: Thanks. Thanks very much, Kirill, it’s really an honor for me to be talking to you today. Thanks genuinely, thanks for this.

Kirill Eremenko: Fantastic, mate. Thanks. Stay in touch. And everybody, make sure to follow Carlos and his career. 

Kirill Eremenko: So there you have it. That was Carlos Hervás García, aerospace engineer at Airbus. Quite a different episode to what we normally talk about, don’t you think? It would be interesting to know what your favorite part was, what was your main take-away from this episode? I probably enjoyed … I enjoyed hearing about the two different types of applications. I think that was very cool that Carlos distinguished between the two, that there’s this one where they’re optimizing and trying to find new ways, new solutions with artificial intelligence, and the comparison to playing a computer game with artificial intelligence was very useful, very insightful, because that’s what we do in our courses, we mostly find these enclosed environments and learn how artificial intelligence can learn that Atari computer game or something like that, and now we can see how those skills can be transferred to real life aerospace engineering, which is quite a complex thing. And then, on the other hand, the second application, which is fault prevention, fault loss, or anomaly detection and prevention and isolation and things like that. So that’s also very, very powerful when a human can’t keep their eyes on everything, on all the meters, all the gadgets and stuff like that, when an AI can take over and look for what’s normal and what’s abnormal and try to keep the system in a normal state.

Kirill Eremenko: So that was a very, very fun episode, I think. If you’d like to get in touch with Carlos, if you’re perhaps an aerospace engineer yourself and you would like to learn how you can apply better artificial intelligence in your job, definitely get in touch, otherwise you just want to follow along Carlos’ career, you can find his LinkedIn and all the things that we mentioned in this episode in the show notes at www.www.superdatascience.com/217, that’s 2-1-7, and there you can also find the transcript for this episode. On that note, thank you so much for being here today, I hope you enjoyed today’s journey into yet another career and a very different one, let alone somebody who still uses artificial intelligence and deep learning in their role. And on that note, I look forward to seeing you back here next time, until then, happy analyzing.