Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

Value Driven Data Science

Value Driven Data Science artwork

Episode 130: [Value Boost] The Digital Twin Blueprint

Nov 4, 20269 minAudio
ShareXLinkedInFacebookWhatsAppEmail

Building a digital twin without a clear objective is like showing up at a construction site without a blueprint. You might produce something, but it's unlikely to be structurally sound, and it will certainly cost more than it should. In this Value Boost episode, Dr Greg Biegel joins Dr Genevieve Hayes to share a practical framework for mapping a physical system to its digital representation - one of the first and most critical steps in building a digital twin. You'll discover: The three things to write down before you build anything [01:45] How to decide which parts of a physical system to model and which to ignore [03:27] The four types of models that make up a digital twin [04:00] How to work backwards from a decision to the data and models you actually need [05:25] Guest Bio Dr Greg Biegel is the author of Digital Twins in Action and has been working as a researcher, software developer, architect and product manager across academia, startups and large enterprises for the past 26 years. He currently works as a Solution Architect with Amazon Web Services. Links Connect with Greg on LinkedIn Connect with Genevieve on LinkedIn Be among the first to hear about the release of each new podcast episode by signing up HERE

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show.

[00:00:00] Dr Genevieve Hayes: Hello, and welcome back to Value-Driven Data Science, where data professionals become strategic experts. I'm Dr. Genevieve Hayes, and I'm here again with Dr. Greg Biegel, author of Digital Twins in Action. Last week, Greg and I explored what digital twins actually are and how they can be used to enable better executive decision-making. [00:00:24] Today, in this Value Boost episode, you'll learn how to map a physical system to its digital representation, one of the first steps involved in building a digital twin. Welcome back, Greg. [00:00:37] Dr Greg Biegel: Hi, Genevieve. Thanks for having me [00:00:39] Dr Genevieve Hayes: If you wanna build a physical structure, be it a house, a bridge, or a factory, the first thing you invariably need to do is to create a blueprint. [00:00:49] You can't just show up at the construction site and start building, or you risk wasting a lot of time and money on a project that returns nothing more than a structurally unsound pile of raw materials. The same is true when building a digital representation of that structure, also known as a digital twin. [00:01:07] However, the creators of digital twins have one major advantage over the designers of their physical counterparts. Rather than designing something from scratch, data scientists need only create a map from physical reality to the digital space. Of course, that's easier said than done, and the question of how much of the physical world to translate into the digital is where things start to get tricky. [00:01:32] Greg, obviously it's rarely going to be feasible to create a digital twin that's a perfect replica of a physical system. But how do you determine which decisions to support with your digital twin? [00:01:45] Dr Greg Biegel: I think the first thing that you need to do is write down the objective that you're trying to achieve and put a dollar amount a-against it. So not in terms of technology, so not something like, "I want be-better visibility," but rather, "I lose 12 hours a month to bearing failures at $15,000 an hour." [00:02:06] Once you've done that, set indicators and a baseline and define those leading and lagging indicators that will help you make that better decision. So a leading indicator is your early warning, something like your vibration is fluctuating beyond two meters per second squared, or bearing temperature is moving more than three degrees in a week. [00:02:27] And a lagging indicator is a business outcome, the reduction of downtime from 12 hours a month to three. Thirdly, write that decision out twice. So the current version, the way in which you make your decision currently, and then the way that you would improve your decision by building your digital representation. [00:02:45] So your current might be, if six months have passed, replace the bearing, and your improved decision would be when your trend indicates 80% probability of a failure within four weeks, order a bearing and schedule the replacement. And I think if you can't write that second sentence, then you're not really ready to start building a digital twin [00:03:06] Dr Genevieve Hayes: so once you've decided on the decisions you want your digital twin to support, how do you go about representing them using models and creating that mapping from the physical world to the digital space? [00:03:18] Dr Greg Biegel: Yeah. So I think it's really important to look at each decision and look at the sources of data that you have to support that decision, and look at it in two axes. [00:03:27] Priority against your objectives and then the effort to integrate that into your digital model. And then start with working on the high priority, low effort pieces of data first, and then moving later into high priority, high effort second. And then be honest that you might not ever get to the low priority high effort pieces of data to integrate into your model [00:03:51] Dr Genevieve Hayes: How do you choose the model? Are they usually machine learning models or Is there some other alternative special type of model that tends to be used? [00:04:00] Dr Greg Biegel: Yeah. So I think the model that you build of the real world exists in a number of different formats. So there's an ontological model. So in the book, I talk about using a knowledge graph to model that ontology. So that is the entities of interest within your physical system, the attributes that those entities have, and the relationships between them. [00:04:21] That's very important for contextualizing data. There's also mathematical models. So modeling a physical system can often be done by a set of differential equations. And I use the example of a water tank being modeled with a differential equation in my book. There's also physics-based modeling. [00:04:41] Depending on the type of system, if you're modeling electricity generation or a mechanical system you can use the laws of physics to build that model of input and output of your system. And then finally, there's the visual model. So a visual model might be a two-dimensional visualization, such as a satellite image, or it might be a three-dimensional representation of that physical system. [00:05:08] So I think the modeling of that physical system exists in many different dimensions [00:05:13] Dr Genevieve Hayes: So once you've got your decision, would you then say to yourself, "Okay what are all the steps that go into these decisions, and how can I represent each of those using some sort of mathematical model?" [00:05:25] Dr Greg Biegel: Yeah, once you know the decision that you want to make, you then work backwards from what data you would need to support that decision. Once you know the data that you need to support your decision, you would then look at the data sources that you have available. And those might exist in different silos within the organization that you can bring together into one system and make accessible to your digital twin. [00:05:51] Or you might need to add additional instrumentation to get the data at a high enough frequency to support your decision-making. So that might involve deploying new sensor infrastructure. It might involve accessing different APIs. Quite often within an industrial context, we have data that's locked away behind OEM systems, so you might need to look at a way to liberate that data. [00:06:15] So it's really about working backwards from the data that you need to support your dec- and being laser-focused, only modeling and using data that you actually need and the objective that you are focusing on [00:06:28] Dr Genevieve Hayes: Can you walk us through a quick example of what that whole working backwards from the decision to the models would look like? [00:06:35] Dr Greg Biegel: Yeah, sure. So in the book, I use the example of my home because I think it's something that most readers can probably identify with. I live in a very dry city, so I really would like to reduce the water usage within that home. Unfortunately, I only receive a water bill from the water corp every quarter. [00:06:55] So my information that I have in order to make a decision about the consumption of water in the home is very infrequent. Now, I can go out and look at the water meter every day. However, I find that I tend not to do that. So I wanted a way to get the data from that meter much more frequently. [00:07:16] And because it's a mechanical system and it doesn't have any electronics, I looked at a edge AI-powered camera that would read the meter and then transmit the reading on a much higher frequency. So I could look at the consumption of water and make a decision every day as to how I could minimize that consumption. [00:07:38] The model for that as that reading comes in it's just a number. Being able to contextualize that within a knowledge graph, so to say that this reading relates to this water meter, and I then extended the meters onto other individual taps to get a more granular consumption view. [00:07:58] But the relationship of measurements to parts of the home, so specific taps, then provides that ontological model that I can then use to make better decisions. So I know that, this tap is watering this part of the garden, and it's using this much water, and then I combine that with other sensors soil moisture. [00:08:21] So I can then make a better decision about how I can optimize the usage of water [00:08:26] Dr Genevieve Hayes: If a data scientist was about to start mapping their first physical system to a digital twin what's the one thing they absolutely must get right in this process? [00:08:36] Dr Greg Biegel: I think they absolutely need to start with a decision and not a technology. So be very clear on What decision you want to improve, and then work backwards from that and use that to influence the model that you build [00:08:49] Dr Genevieve Hayes: And that's it for today's conversation with Greg. If you haven't already, listen to our previous episode where we explored how digital twins can be used to enable better executive decision-making. Thanks for joining me again, Greg. [00:09:04] Dr Greg Biegel: Thanks very much, Genevieve [00:09:06] Dr Genevieve Hayes: And for those in the audience, thanks for listening. I'm Dr. Genevieve Hayes, and this has been Value-Driven Data Science.

Open the transcript file

Hosts and Guests

Hosts

  • Genevieve Hayes Host Feed-stated

Guests

  • Greg Biegel Guest Feed-stated

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. Episode 132: [Value Boost] Making a Causal Argument Without Experiments

    Nov 18, 202615 minAudio

    Almost every analysis a data scientist presents makes a claim about cause and effect, whether they say so or not. But most organisations don't have the data, time or budget to validate these claims by running a formal...

    Credit the people on this episode

  2. Episode 131: Building a Causal Mindset in a Predictive World

    Nov 11, 202627 minAudio

    A machine learning model can tell you which of your customers are likely to cancel. It can't tell you what would happen if you intervened to change that. Prediction and causation are two completely different problems,...

    Credit the people on this episode

  3. Episode 129: Building Digital Twins Executives Can Trust

    Oct 28, 202627 minAudio

    When the cost of failure is low, experimentation is often the best way to make decisions. But when a wrong call could cost millions or endanger lives, you need an entirely different approach. This is where digital twins...

    Credit the people on this episode

  4. Episode 128: [Value Boost] The Three Types of Memory Every AI Agent Needs

    Oct 21, 202613 minAudio

    One of the most frustrating things about working with AI is when it seems to suddenly forget everything it knew about you and your work. This is annoying when using AI chatbots. When building AI systems, it can be...

    Credit the people on this episode

  5. Episode 127: From Algorithms to Agents and What It Means for Data Science

    Oct 14, 202623 minAudio

    The agentic AI era hasn't made algorithms irrelevant. It has just changed where data scientists need to focus their attention. And the data scientists who understand that distinction, and can explain it to stakeholders...

    Credit the people on this episode