黑料福利网

Alexander Teeuwen has been working as a software engineer at Gooiland Elektro for over seven years, focusing on industrial automation for HVAC (heating, ventilation, and air conditioning) systems on luxury yachts. He joined the expert track of the Mastering Data & AI program to deepen his technical knowledge, particularly in machine learning. Daniel Kapitan (DK) spoke with him about the impact of the program and the Mastering project he completed to build a digital twin.

DK: To start with 鈥� what made you choose the Mastering Data Science & AI program at EAISI Academy?

鈥淲ith a background in industrial automation, I started working at Gooiland Elektro as a software engineer. After seven years, I sought a way to broaden my knowledge. In the OT (operational technology) world, AI is still underutilized, despite the many lessons we can learn from IT (information technology). I was looking for a solid, technical program to bridge that gap. EAISI Academy appealed to me because the expert track provides a strong technical foundation and teaches you how to apply AI practically.鈥�

DK: How did you experience the structure of the program?

Before starting this program, I had no background in statistics, so I had to make an effort to get up to speed on that quickly. The content on training and validating algorithms was also entirely new to me. So there was plenty to learn right from the first module 鈥� and I enjoyed it.

The second module, Applications, was my favorite. Working in a team on a project challenged me to deal with the complexity of collaborating on a single problem. That wasn鈥檛 due to the team 鈥� they were great 鈥� but because I was still trying to grasp the material and my project fully. That made it difficult to communicate your thoughts to others 鈥� you were still figuring things out yourself. I would recommend this module to anyone interested in the topic. I鈥檇 love to do it again (with the same team, of course) to build a better version of the project using everything I鈥檝e learned.

The third and final module, Mastering, was both the most valuable and the most challenging. In this phase, you work independently on a project of your own choice, applying all the knowledge you've gained. Luckily, you鈥檙e not entirely on your own 鈥� the support from a mentor helps to keep you on track. What stood out to me was that the material was starting to feel more familiar. I could run experiments where the outcomes often matched my expectations 鈥� a clear sign that I was beginning to understand it all.

DK: What was your Mastering project about?

鈥淢y goal was to develop a digital twin of an air handling unit, based on data from a real-life installation, not a mathematical model. At Gooiland Elektro, I develop control software for yacht climate systems. These systems are only assembled and tuned once they鈥檙e on board. With an accurate digital twin, that tuning could be done elsewhere, which would make configuration more efficient and comfortable. My project was based on a scientific paper on data-driven digital twins, and I successfully reproduced the results.鈥�

DK: As your mentor, I found it impressive to see how, once you got going, you ran a lot of experiments using HVAC system data. What exactly did you do, and how did you experience it?

鈥淚 also explored a technique to identify which variables are responsible for which outcomes 鈥� for example, activating a humidifier increases humidity. That worked to an extent, but the method couldn鈥檛 deal with feedback loops, and climate systems are full of them. Take the cooling valve: if it opens further, the temperature drops, which then affects other parts of the system. That made it clear this method wasn鈥檛 yet suitable.

To build the digital twin, I compared two prediction techniques: a simple regression model and a neural network. Initially, both gave similar results. However, after further analysis, it became clear that the neural network was more effective at handling the internal dependencies of the system. For example, if the cooling water temperature increases, the whole system鈥檚 dynamics change 鈥� something the neural network could deal with much better.

The biggest challenge came when the neural network had to make predictions outside the known data range. In one test where the cooling valve was adjusted, the model predicted a temperature increase instead of a decrease. This turned out to be because the training data included the existing controller, which regulates everything toward specific setpoints. For a useful digital twin, you need to exclude that control behavior. So the current challenge is: how do you effectively switch off the existing controller?鈥�

DK: Now that you鈥檝e completed the program, what would you say to future participants?

鈥淒on鈥檛 cheer too soon. I had moments where I was thrilled because the prediction almost perfectly matched the actual data, only to find out later that there was still a leak in the data. A model will always give you some output, and it may seem reasonable, but always ask yourself: Is this the outcome I expected? And can I easily break the model, or does it hold up?

Things won鈥檛 always go your way, and at some point, your program or computer will crash. But when things do work, those moments are pure gold.鈥�