 Good afternoon, ladies and gentlemen. Thanks for having me. My name is Arshad Kazmi, and we at VMT develop software solution tunnels and underground navigation system, mostly for mechanized tunneling and also for drill and blast tunnels. So today's topic is relevant to our latest development for our navigation system, which we did named as a moving station, where we use some machine learning techniques to determine the position of the TBM during the advance. I will go into more details in the following slides. So basically, we at VMT, we provide the navigation system since 25 years now. And the system was working before while having a total station on this tunnel wall, like you see here, and then a target unit on the front side of the machine and then determining the position. This was quite a rigorous job for a surveyor and the shift engineers to basically deal with the total station time to time when the machine is going forward, then shift the total station from the back to the front. What we did is basically, we enhanced the system called as moving station, keeping the total station with the TBM, which travels together with the TBM. This has many benefits for the customer. It determines automatic TBM position. TBM is basically a tunnel boring machine. I may be some of you know. I don't know. It provides a geometrical position as well, and there is automatic tri-bug, basically automatic leveling system there, which basically automatically levels the total station to determine the position. And we need basically backside prism to make the orientation of the total station minimum, at least two backside prisms, and this is the only thing which normally job site or the surveyor or the TBM operator has to do. Other than that, this automated product, which is called as moving station, provides all the solution autonomously. Basically, we enhance the system from our previous existing standard system, and so it means that customer can anytime switch between two modes if they want. How it works? Basically, tunnel boring machine is working in two cycles. The first cycle is basically when machine is starting or when machine is standing still. Total station is in a more relaxed form or the state. We can determine the position easily. But during the excavation, when machine starts here, for example, you see in this visualization, this is the cycle number one when we get the position. But as soon as machine starts advancing, then there is a problem of determining the precise position. And there are issues getting the free stationing plus the vibration issue. We are working on that. But in the meanwhile, we use some sensor fusion technique and machine learning technique to determine the position precisely to help the operator to steer the machine more optimally. The topic of the presentation, like I said in the beginning, using some machine learning techniques. So here, the cycle number two is basically will be addressed in this presentation. Here, you see a typical scenario from the tunnel-boring machine. Predator stands in the control cabin. He tried to steer the machine with those push ramps, potential meters, exerting some pressure. Here, you see a navigation GUI coming from the tunnel-boring machine itself, the data from the machine. And that screenshot basically shows our universal navigation system, which provides the position to steer the machine optimally. And the challenge is, like I said in the beginning, how to determine the position when the machine is in the operation mode, when the operator is steering it. So we use multi-sensors technique plus machine learning to basically determine the position. With the multi-sensors, actually, we had the issue that how operator tried to steer the machine. For example, here, you see this is a typical scenario. The machine is standing still. We got a position from total station. And on the below graph, you see there are green dots, which are basically representing the machine position when machine is standing still. And blue is a simple interpolation between these green dots to get the real TBM position. And red is basically showing how operator tried to steer. So if you look on this graph, for example, here, operator is really steering hard to get back to the zero deviation, so to get back to the zero offset. And machine is reacting a little bit later, because it's like you are drilling. And this quasi-static behavior of the machine was a big challenge how to model this. And we found some correlation in the history data using this history data that how can we determine the position. Because the future position of the machine is mostly dependent on the history data. And this gave us impression that we can probably deploy some machine learning algorithm and can use it for determining the position during the advance. So here is, again, the graph that how operator tried to steer. But the real blue dots, the real position of the machine is basically going straight. But the operator is trying to steer left and right. That basically brings us to the point that using this history data, since we have a problem which is basically numerical data. And it belongs to a regression problem using a classical machine learning technique, simple time series models. We basically model position of the machine as precisely as it could be possible to get a nice position, which really helped operator to steer machine better. So this is how the algorithm looks like. We have some unknown target function, probably output given by some inputs. We trained the model based on different job sites data. And then when in a real job site, we deploy this model, we get the information from our tuner software, and then we predict the position and we display it. Moving forward, basically, there are different kinds of machines I will not go into detail. For example, single shield with articulation slender to steer it more better than there are hard rock machine. So we basically use all the machine, the data, from all these machines to post-process the data to get the impression if we get those results, which we are am to have. So here you see the green, basically line is basically representing the predicted point. And the orange is basically the desired output. And you see the prediction is providing more close data or the output compared to the real TBM position. And the next question comes, most of the time we see that machine learning is easy to do or handle with the offline data. But when it goes to real deployment, then there's a big issue how to handle this data in the real time. There basically what we did, we simply made an API interface using some libraries from Python Flask. And this post method, basically talking with tuners, give me certain history data. I provide you the predicted point in future where the machine will end up. And as you see on the tuner's navigation GUI, there is this orange marked prediction. This is active status, for example, of the machine. That prediction is active. And this is live position being displayed from the predicted from Python server, which is basically interface between tuners and providing the position. There are definitely a lot of challenges to handle the data, which we handled it quite well. And with that, basically, I will quickly go to the results from the real job site from online data. This is one kind of machine, which is basically EPB Earth Pressure Balance Machine with articulation cylinders. Here you see the blue line is coming from the total station, blue dots, simply interpolated data between to get the impression that this was the real position of the machine. And green is basically the predicted position. And this provided us an accuracy of under 10 millimeter or so. And this was quite promising to basically use this moving station with the prediction as a product. We have results from other job sites as well. As you see here, the green is always more closer to the blue from real total station position. And then we used, basically, when we predict, we can only predict the history response from the machine. But in order to know what is the exact position of the machine right now, I mean, how the machine is reacting to the steering behavior during the advance, like I said, the quasi-static behavior in the beginning, this gives us only the front point of the machine in the ground where this will end up. But the rear point of the machine is a little bit more dependent on the steering parameters which operator applies. So using some sensor-fewing technique, data from these push ramps, which operator is basically pushing from the thrust cylinders. And in clearometer data and certain other sensors, we basically use this sensor-fewing technique basically to model the rear point of the machine. And here you see that these olive lines basically representing the rear point of the machine, which basically quite directly impact of the steering on the machine. This basically gave us really a good outcome to basically give the results, which are basically required for the job site. And then this is also another example. So we basically use this product for one job site, which we already completed the drive for around five kilometer or so. And then at the moment, we are using this product, Tunis moving station, for three job sites with single-shield machine and then single-shield machine with cutting wheel adjustment cylinder, three other job sites. So all together is around eight job sites, which we are using at the moment with this Tunis moving station. And this provided really great benefits to the customer, especially saving the time for the shift engineers, for the surveyor, to not necessary to be on the machine. And there is really no damage to the original or the important product, which is basically segments lining the tunnel itself. And it also reduces the wear and tear. And this also brings us to the point where we are looking forward to have autonomous navigation solution plus this autonomous steering in future, which is coming. And with this innovative product, basically using together machine learning techniques plus the sensor fusion gave us a direction where we can think about how can we autonomously provide a navigation system.