 Hello everybody, thank you for coming today to attend this talk about Time-efficient aircraft isolation procedure with NLP techniques. I am very happy to be here I am the head of health management and predictive analytics in Arbus DS And the reason because I am very happy to be here is because sometimes I have the feeling that People consider that, okay, Arbus Airbus is a big player, but it's not a very innovative company Okay, you are a very big Manufacturing industry, but you are not an innovative company And I want to share with you today that really we are an innovative company And we are in fact one of the most or the biggest player in artificial intelligence in Spain Wait two minutes Okay So I was talking about health management Okay, and what is health management? Health management is manage an aircraft health status now Diagnosis and in the future prognosis It's like a doctor for a human It's only the same we take care about our aircraft in the same way that a doctor take care about their patients So This is so important because health of an aircraft you can consider is a very very critical thing But what is health management today? It's a little bit different because what we are doing today is self-manage the aircraft health status Mainly driven by an artificial intelligence and this is why I'm talking we are a very innovative company In fact in our team in Hetafe here We have more than 50 people working on that and This is one of the key strategy of our bus to use data and to use artificial intelligence to improve our health safety Health management and other incredible tasks that we are performing there But with a specific sample What is health management? Why so important? Obviously you can consider if you fight with your health management You have fun aircraft crash fact you can see With health management we monitor for example the engine's condition in order to avoid a Rotor bus that is a very critical situation on our graph as you can imagine But it's not only that because health management is one of the key contributor to the mission system Why because in a defense? Environment the mission is the key thing. We need to perform our mission in the right time and in the right way And health management is providing the information about health status of the different system in order the mission system can take decision about What is the right action to do? but on the other hand Maintenance is a very well loaded task for a person For example a fuel system that is located inside of the wing of very crap In order to perform maintenance that what you need to do you need to unfil all the fuel tank You need to put bends in order to remove all toxic vapors And after a somebody needs to be inside the technician inside of the wind and to perform the maintenance In a very uncomfortable situation because they need to be like that so For that is very important of management And the last point obviously something very important in every company in every industry is the cost Maintenance cost is the key driver for total cost of ownership of an aircraft And this is something crazy when you are thinking an aircraft development is costing More than 20 billion for heroes and maintenance cost is even more So for that in the maintenance strategy that We are using in our bus similar to another industry. We have started with reactive Fix something when it's broken No, it's seen not very good strategy for an aircraft because if you wait until that moment Maybe you have nothing to fix So for that we develop a very powerful a schedule maintenance plan It's called preventive maintenance when we define based on Theoretical models based on predicted operational use of aircraft so in a specific maintenance checks In order to perform inspection or review or repair or removal But it's the same case. It's not accurate enough because The operational of aircraft is not same as you predicted and sometimes you perform the action before it needed or later and You can have what you go back to reactive maintenance because you have the failure before it was predictive Theoretically or later. So you are Increasing the maintenance cost For that We go to the proactive maintenance This is more or less the state of the art in the most of the industry even all the people are talking about predictive maintenance What is proactive maintenance? Okay, we are taking data from the system and we are trying to eliminate or remove the facts Calculating the abnormal behavior or detecting anomalies It is much better approach, but it knows always the best because the same way something that is behaving abnormal It could not fail And you are wasting money and sometimes the degradation are so Quick that you cannot predict using an anomaly detector For that in the hype right now What we have is predictive maintenance Everybody talks about predictive maintenance And it's very very important in an aircraft Because obviously if you predict something before they happen, it is good for an aircraft in order to increase safety For predictive maintenance what we are understanding is to use advanced analytics and the data from the sensor of the aircraft in order to predict maintenance Events or to predict failures or to predict everything and this is a much more intelligent strategy But I have told you that we are innovative in Airbus so What we are looking for in the future is a prescriptive maintenance strategy and This is strategy is an artificial intelligent base health management system that is capable To diagnose or prognose something before they happen but not only It's a system that is capable to recommend you to your maintenance action is the Alexa of the maintenance as something and The system is capable to recommend exactly what you need to do and even more in the future the system will be completely automated and Will be on board on the aircraft and your aircraft will take will make its own decisions And will provide to the operational system the health status in order this operational system came Modifying its control law adapt to the new situation or even to vote for a safe mode For that is very very critical in the aerospace industry to go for that but the question is Do we need artificial intelligence to do that or? We are Lead by the hype about artificial intelligence. Okay, we are going to put artificial intelligence and any in everything and The response is yes. We need artificial intelligence in order to really have a good health management system if we Want to have this benefit increased fleet of availability? This is the number of hours regarding all the time that the aircraft is available to do something if we want to reduce maintenance cost or if we get our missions access and All under the same word safety first. We need to have a health management a driving system In order to detail that I once inspired but For some talk that I was attending Past year in mad lab expo. It's cold. Are you ready for predictive maintenance? And I have taken that an example about why it's needed artificial intelligence in health management and in maintenance area If we want to build a health model We can go for the traditional approach engineering driving. We call that We take old knowledge from the engineering guys the models the design data Manuas whatever they have and we combine with the data With the real data and we try to get some outputs But it doesn't work the most of the time because the system are very complex an aircraft is a very large Complex system and there is no engineering around the world that really can predict every interaction of an aircraft system It's impossible. For example, the fuel system the international sample Has interaction with the engines with the electrical system with the environmental system The operation of the aircraft is not always as predicted because the user Doesn't that not use the aircraft always in the same way And this is impossible to perform a theoretical model that is capable to detect or predict failures so Okay, so we therefore that are driving approach and we use much in learning and artificial intelligence or whatever Techniques that is database in order to do that So we get The real data we level they all put the behavior And we train an algorithm in order to get a model But he's the problem An aircraft is a very large complex machine so for example an aircraft could have as some of our aircraft could have 600,000 signals with rate up to kilo Earth We are generating terabytes of data per fly in a fleet with 100 aircraft or thousand aircraft It's something crazy the state of the AI technology machine learning technology It's not capable right now to get this data and to develop really accurate models For that in Airbus what we are doing right now is Combining the both of our both words the best of both words It's called hybrid We take the engineering knowledge engineering models the theoretical knowledge We take the radial data and we level the output and using machine learning We train models against the engineering models and we are capable to provide accurate algorithms Because in order to use an algorithm in an aircraft we need to perform trusted AI We need to be capable to explain to the authority how our artificial intelligence working if not, it's not possible to certify And we need to be capable really certify that if we want to use For example in predictive maintenance Authority are requesting at at least 99% of accuracy in order to certify a predictive mental assistant And this is something that we cannot get right now with the state of the art of AI technology only Probably you have Listen about digital twin because what I was talking about was digital twin Digital twin is when I have a Digital model of every part of the aircraft We use real data to fit that models and using machine learning artificial intelligence We can predict or detect something and this is the state of the art and for that is so important artificial intelligence Combined with physical digital models in order to Get health management in context of this Talk that we are presenting today, I am going to put an example with the field system field system is a very critical system because It's contributing to several catastrophic condition feel Extarvation feel exhaust or tongue explosion Any of this situation lead to a loss of an aircraft is one of the key maintenance Contributor in terms of cost and I talked about the cost of the maintenance of on an aircraft and It's one of the three principal Contributors to the aircraft and availability due to this main point. It's very Hard work to perform maintenance on the fuel system and it's a very critical system So you'll think okay. It's a problem But when the aircraft is landed for example, you want to perform the file isolation in order to get the root cost Analysis in order to replace some component that is faulty. Okay, the aircraft is landed You have all the data you analyze you look for the entry. Okay. I need to perform that maintenance action Okay, it takes time, but you can fix that Run is not so easy Because an aircraft is a very large complex system and when when the aircraft is landed There is no fault code We have hundred of thought of full code So the technician needs to go code for code performing the maintenance action until he gets their right problem And this is what we are solving here with another pectin nicks using much learning together with engineering knowledge Much much better than me For explaining that Welcome, Rocio to the stage and please Thank you, Miguel and thank you to all of you to be here I am very grateful for the opportunity and I'm really excited I'm going to talk to you about the root cost analysis problem The part of that analytics problem so The problem here is to identify the father of the all four codes that are that the the technician have when the When the they have the post flight report, I then I will explain you a little bit the terms that we are going to use With all of these a traditional approach was to use what is called expert knowledge That means that you have the knowledge and the experience of seasoned experts that have a deep understanding of the system Then with the expansion of the use of algorithms They they try to make an effort to do algorithms in order to automate this root cost analysis And also to make independent of individuals and the solution it was pretty poor because Finally it was like kind of algorithms hard-coded so as you might be thinking this is not a solution because if you you cannot hard-code all an aircraft because it's a very large complex system and Also, if you hard-code all the systems The output of the algorithm would only be valid for that specific aircraft in that specific configuration So in an industry as our space where companies have a lot of a great variety of products is not a solution Then with the digitalization and the machine The industry for point zero we get available more data Related with the physical properties of the systems and also of the status of the system so it was easier for us to use machine learning techniques and probabilistic models to do these kind of things So now before I explain you the the technical solutions I want to introduce you a little bit in the the the kind of data that we are working with The first I want you to know is what is a full code? Well, I want you to know that when an error or failure occurs in a flight it is registered So when a system has summer status that meds we in with certain Predefined conditions then an error message is triggered. This error message is what is a full code? It's only an alphanumeric code to identify a failure. So a full code is a failure That's all okay. Another thing that I want you to know is that each full code has its own a aircraft full station a Document that is a document where it is explained to the technicians How to identify which is the real equipment that is failing and how to solve this problem, okay? So in the in the AFI task what we would see is all the equipments that can affect to that failure, okay? and Another thing is that all the equipments are identified a with an nomenclature that is called fin for functional identification number, okay, and Fin only means equipment, okay? another term is a that when you make a flight You land on ground and then you have a report with all the full codes that have happened during this flight Okay, so in this is what it called was flight report in the post flight report You have all the failures triggered during the flight. So it's like a list of codes so once we have seen and The kind of data that we are working with we're going to see what is the amount of data that we have So for solving this problem, we had a 1504 possible full codes for a field system. This is all the full codes that can happen This is the theoretical ones. Okay, and then we have a The half of the of AFI proceed the documents Why is the half because an aircraft is symmetric you have a half of the fuel system in one wing and the other in the other wing So to solve a problem with a sensor in the right wing is the same procedure for the left wing Okay, so that's what that is why this number is the half of the other We have studied a data for 14 different aircraft during several years and we have a total of more or less 4,000 of post flight reports and in these post flight reports what we saw is that we only had 404 for the codes of the 1,500 possible full codes So the first approach that we took was to use a probabilistic method the bayesian network This kind of algorithms what they they aim to to determine our parent-side relationships between a set of data By learning a diet a cyclical graph, okay This is done by calculating the conditional probabilities of note to be the father of of its children So for that problem we use the post flight report if you remember the list of full codes for the flights So our input data were as you can see in the screen a metrics with ones and zeros One if the full code happened in that flight and zero if not, so that was our input metrics and Then a we built the the network once we had the network We also had to do a planning because there were a lot of relationships that Had a really low value of the conditional probability So we are not interested in this case So we we set a prune and we only take take took as bay lead Relationship the ones those ones that have more than an 80 percent of conditional probability So we had the network build but sadly we saw that This approach was not good enough. Why? Well, we saw that some relationships Were physically impossible because they had no Interface between the two equipment so it was impossible to have an impact between each other So That's a pain point in the probabilistic methods that you are not taking into account the knowledge of the engineering part It's only the data Another thing was that the data is biased because we use the data from the post flight report from six years So we have as I told you for 400 of full codes, but they can be a lot of more So if you have a new full code in an S flight and and you don't have it on your graph You won't know what is the impact of that for code? Okay? Now another important problem was that the real big failures that can cause you an accident Thanks, God. Do not happen too much time So the conditional probability would be really low and it would not pass the threshold of the planning So you won't see also the impact and and that's really dangerous That is why we thought why don't we use all the knowledge that we have at every company in order to make the root cause analysis Because probabilistic one is not enough Well That is why we introduced the knowledge the automated knowledge graph approach The first time that the knowledge graph was introduced was by Google when they tried to improve the indexing of the webs and they use two techniques The name entity recognition Mixed with the theory theory of graphs. So our approach is basically the same We are going to index all the equipments of our aircraft and there we're going to build the graph with the graph theory So how are we going to learn this relationship between the equipments? And that's the key point. We are going to use the AFI tasks To know how the how the equipments relates each other. Why? Well, the the AFI tasks are beaten by expert people of the system is the people that actually designed the system So they they perfectly know how the system behaves how the things relate each other So I think it's the right place to look for So we we will use natural language processing to extract the knowledge for the AFI task And with that knowledge we will build the graph in order to point out to the to the root cause So the first step is the natural language processing I want you to know that it's very complex to extract information because the AFI tasks are written in English free text So it's very difficult to extract the knowledge Another thing is that it has a lot of nomenclature related with our space Feel so there's no standard library that is strong enough for your for your algorithms. So we have to do it by ourselves The the good point is that the AFI tasks are very structured they have a the paragraphs it might and in each paragraph you have a The equipment that is the candidate to be the default and how to relate to each other's Also, it is ordered by importance. So as you can see in the slide the first Equipment that would appear for example would be a computer then a wiring and then a sensor because if you have a Sensor fail it does not make sense that the computer is also failed But if the computer is fail all the connections make sense that you have lost it and How do we build the graph? well, we have a We have a list of all the possible Equipments of the fuel system So we search for them for all the 700 tasks We were looking for the for the equipment and then we applied some regular expressions to to find how they relate each other For example, if the task tells you please first look at the first the equipment number one Then look at the equipment number two and then look at the equipment number three Then you will have three notes in a row Okay, but if they are fetish, please check Equipment number four and then the possible Equipments failed are the two the three or the four then you have one parent and three children of the same parent So that kind of this present Where the ones that we used to to develop the algorithm? once we had a The nlp done and we had like a dictionary with their parents and his children And then we only had to apply graph theory to to connect all of them So as I have told you we have used the AFI task. That's an Theoretical graph we have used all the possibilities. So how can we take into account the real operational data? well Once we have the graph done that is the the line that you can see right here Then we use the post flight report to see which are the real failures that you have in a flight With that failures you see which are the equipments that can cause that failure and then you extract from the theoretical graph Your notes and your edges that are affected Here you can see the result a The graph knowledge is right there a is all the systems that that can contribute to a field systems are Right here. So these are all the equipments of the field system It makes sense the results because as you can see in the in the center part You have a the system that is in charge of the communication of all the equipments so if you lose that system you lose everything then in the In the top you see the probes okay, the probes are sensors that Are monitoring the quantity of fuel in the system? So it makes sense that they are in the tip because they should not affect to another to the computer thing for example and In the middle is the wiring that is called harness that is which connects this probe to the communication system So as you can see it makes sense that in the center is the computer in the middle the the wiring and in the tip is the sensor That's the theoretical one. Okay, if we Wanted to to see for a specific flight We would have to take up a flight report and extract the note and that's what we have On the left. We have the PFR graph So when the technician is going to to make the maintenance it makes sense that he'll start on the middle because it's more related It didn't it do not make sense that he starts on the tip of the graph So what is the advantage is of you of using this algorithm? Well, the first and most important is that we are taking the knowledge of it was people of engineering people of experts people of of the company But making automated and without human intervention so another Benefit is that it makes sense as you have seen the structure of the graphs Has sense from the engineering point of view also is deterministic is not probabilistic So it's a great advantage and also it does not need to retrain Okay, you have this theoretical graph for each aircraft only by changing the AFI task that apply to each other So you would not need to retrain another thing is that we did like an study of What was the the maintenance workload that we were using with these Techniques and we discovered that it was like a 30 percent. So we were really contributing to to the maintenance tasks the only limitations that I saw to this algorithm is the The natural language processing because as I told you for there is no library for our space industry. So As it is written in free text There are some expressions that the algorithm is not able to detect So a few of the edges of the graph did not make sense because this kind of particularities But the rest I think that is a good approach to use artificial intelligence With the knowledge of the company in really if you want to close the talk Okay, so only I want to Say thank you to all of you and obviously if you are interested in to learn more about what we are doing In our bus related to artificial intelligence machine learning or whatever Please contact me by the application or later in the in the meeting and again Thank you so much to be attending