0:15 Hi everyone. Today I want to make the case that you need to wait for AI to get smarter if you want to apply AI in engineering that is on manufacturing on design we simulate all you need is better data infrastructure. And to show you that I want to start with me just some weeks ago. That’s me working in a customer project. And we’ve been working for some weeks at that point building the infrastructure.

0:54 So we think at this point the model is ready for us to ask a lot of questions. So I did I said could you reverse reverse engineer our customer’s proprietary balancing algorithm then investigate the root causes behind typical correction errors at the factory. Notice this is quite a big ask. First cuz we didn’t have documentation about this algorithm. It had been long ago built. Not even our customers knew their insights.

Secondary because well this is not a task AI can do within a specific environment. For AI to succeed succeed at this task it had to understand the problem, the physics behind it. It had to plan it. It had to query the proper databases. It had to analyze that data, iterate, see the results, and keep going. So, we waited. And at some point, the result was out, and he had done it all.

2:11 Again, this wasn’t a summary or a chart that our AI was doing in our analytics tool. The AI was working at Google platform, reconstructing an algorithm. It did it almost perfectly, and then it went on and built new features. It built six ground beta. So, it created new beta from the raw sensor data. This is 100 kHz data, very difficult to work with. And it found there a signal that decades of expertise in our customer’s company hadn’t been able to yet.

2:50 This was such a warm moment for me during this project, a great great feeling in the project, but also for the first time, I feel I had a different relationship with my AI. I wasn’t asking it for junior engineering tasks. I felt now like a director of engineering. I was now asking for very high-level tasks that needed a team of specialists to be completed. Another way to put it, we probably use AI every day, all of us.

3:20 We ask it, “Debug this code with Cloud in our environment.” Maybe we are in our database, and we ask it to help out with a SQL query. Maybe we ask it to summarize a document or generate a chart. All of these are limited tools. The AI lives within our specific environment. We give it a snippet of context or an order, we get a response. The model here is not accessing all your data, is not able to plan with the data.

3:58 It’s the engineer, our self, doing the thinking. Data that I just showed you was a different animal. The kind of engine that needs to leave behind looks something like this. First, it will need to plan, and for that it needs a knowledge base. This is context about the physics, the rules, that the rule the the the the process that we want to model. We will have to know about the methods, about prior hypothesis, and this knowledge base will have to get updated with the new findings.

4:35 Once a task is planned, it will probably start by querying data. It has to have access to all our databases, from the summary ones to the raw data, 100 per hour sensor signal that we have available. Once it queries the data, it will write an and run code, it will generate analysis, it will read the results, validate or refute hypothesis, and start again. And it will keep looping until we find the the result of our task.

5:10 So, this was a different animal. It was such a powerful thing to experience, to be able to have an engine like this at your disposal and keep asking questions. But, when did we ask them to start with that? We’ll come back to this, but let me present myself. I’m Javier Blanco. I’ve been a data scientist for about more than a decade now. I work in companies like Orange or Jaguar and Rover, and for the past 5 years, I’ve been working at Quix.

We well, I’ve been lately the lead solutions engineer, and Quix is a British startup owned by F1 engineers. So, the one thing is the team behind the McLaren racing championship winning data platform. So, these guys were used to deal with thousands of sensors generating high frequency data, and they had to deal with it in real time. So, from that expertise came the Quix platform that enables engineers to to work on on on different engineering problems.

6:22 Lately, we are focusing on integrating AI with your data to basically provide the most capable AI for engineering. So, now that you know who we are, let’s go back to the problem. In case you didn’t know yet, but we were doing rotor balancing. So, our customer was one of the biggest European vacuum manufacturers of this type. They were doing over 50,000 rotors a year. And at the end of each rotor manufacturing, they would have to do a balancing process.

7:00 So, it looked like this. They spin up the rotor at a certain speed at which they take measurements. They take a couple of vectors, so we have amplitude and angle of the vibrations at a couple of points at those different measurements, and then with those measurements, they have their proprietary balancing algorithm to decide where to laser out little pieces of material to balance the rotor back with leaving.

7:34 If the process worked great, it should work on the perfect run. The perfect balancing pass should fix any vibration, and we should get a rotor. But reality is not like that. They take between two up to 10 passes to finally get the rotor up to the specification they consider okay. Each of these factors takes between five up to 20 minutes and average of over 10 minutes. So, this was becoming the bottleneck.

8:06 They were having to spend They were One of the spend quite a lot of money in in buying new balancing machines. So, we came in to try to improve the balancing balancing algorithm so that the average balancing passes would decrease drastically. That was the challenge. And first, we tried to solve it with their current data. We look at their historic data. We use ML. We use AI.

8:39 But the signal wasn’t there. We We weren’t able at first to find a signal that there wasn’t there. We We couldn’t improve it at first. We decided to invest in their infrastructure and that’s what I’ve been talking to you about. What do I mean with infrastructure? We simplified it into three pillars. That’s a nice way to to communicate it. The first pillar is data accessibility. The challenge in your data you do have your data, but sometimes it’s hard to deal with it.

9:18 Like raw sensor data in this case with 100 kHz. Dealing with that data requires an agent in the manufacturing machine in the balancing machine that is able to send that time series data. It requires a data platform with a proper ingestion pipelines. It requires a data lake ready to deal with that amount of data. So, you have to send that up. That’s what we are good at from our 1X expertise.

9:48 But, then you also need to invest on a knowledge base. This is the context that they are going to have and this is what our customer was providing. They told us about the physics, the engineering hypothesis they had. They told us about their history with the process. And they helped us create a virtual dream team. So, we create a set of virtual roles, a data engineer, data scientist, signal processing engineer, rotor line rotor line analysis.

10:21 This virtual dream team allow us to approach them the problem from different angles. We had the data accessibility we provided. We had the knowledge context our customer provided. And then we needed the autonomous execution. We needed an adaptive integration able to write code to run against the real data, to access the different databases, to plan, to integrate, and to update the knowledge base. And with that, we had our engine built.

11:02 We had an AI agent capable of reasoning with the real physics rules that were important for our process, and with the raw sensor data. That’s the kind of engine that I was telling you about. The powerful engine that can benefit from real knowledge on the problem, that has all the real raw data, and that it can iterate. With this, I have already told you I was able to feel like a director of engineering.

11:39 I I already mentioned that I asked Kitty to reverse engineer the proprietary balancing algorithm, but that’s just one of the many tasks I asked I asked Kitty to predict. With the vibration data on the third run, whether we will know it will fail within certain amount of runs. I asked Kitty to predict based on both curves raw files whether we have a better edge on predictive signal.

11:41 I asked Kitty to generate a correction on the balancing algorithm at scale with ML. I asked Kitty so many things, very high-level. You can see there prior classifier, stack optimization, anomaly detection, clustering, classifying of the balance big trajectory, quantify the different bearing manufacturers have failed. These were very high-level tasks. Tasks that normally a director of engineering would ask, and I was able to to ask him. It took us almost 3 months to build the infrastructure, and we did all the 50 or more investigations in less than a week.

13:03 So, that’s how it was for us. We reverse engineered the algorithm. We found the signal that no one had found decades before. We improved the correction methods, so we improved the balancing algorithm, the ultimate objective. We improved the throughput of the machines, so our customer wouldn’t have to buy new machines anymore. And it took us far more time to work on the infrastructure for this problem than to actually solve it.

13:35 That thing running costs of machines, so not taking into account the the capital extent that they are avoiding, they are about to save over 1 million a year. And we got all this in 1 week. If not necessarily about the problem, the key takeaways that we would like you to have fun and start with us about is that most engineering organizations already have the data and the model intelligence that they need to perform these kind of tasks.

14:12 What you need to work on, as we did here, is on the infrastructure. And there’s three ways to do it. One, your your AI should have access to all your data, including the raw one. You should be able to work with high-frequency sensor data because in many cases the signal is going to be there. Two, the AI has to have a structured domain knowledge about the problem.

14:42 You have to integrate with your ticketing system, with your meeting system. You have to bring the knowledge from your team to the AI. Three, you need to work on an autonomous execution so that this is an intelligent intelligent system that is able to work with the other two. With these three pillars, someone like me, a single executor, was able to perform those 15 investigation in over in that one 1 week.

15:13 And that takes us back to our point where no one in this room has a AI intelligence problem, it’s an infrastructure problem. The good thing is that they can be solved. Thank you. To learn more about the CDFAM Computational Design Symposium, access the archive of previous presentations, interviews with speakers, and information about future events around the world, visit CDFAM.com.