Digitalization and new technologies, such as digital twins and artificial intelligence, allow industrial companies to develop speed, efficiency, quality and flexibility in an unprecedented way. In this blog post, we cover first what is a digital twin and then dive into artificial intelligence and how these two together can benefit your overall operations.
A digital twin is a digital representation that simulates virtually a real-life object, process or system. Digital twins often consist of:
- Laws of physics
- Material properties
- Virtualized sensors
Ideally, the digital design tools integrate into the real-world control of the production facility or the product or prototype in question. This permits testing different production scenarios and validating the changes before taking the new features into production.
Machine learning and digital twin improve one another
The first benefit of a digital twin is the ability to produce simulated data. A virtual environment can go through an infinite number of repetitions and scenarios. The simulated data produced can then be used to train the AI model. This way the AI system can be taught potential real-world conditions, that might otherwise be very rare or still in the testing phase.
The second benefit is the ability to plan and test new features. The digital twin should represent reality, but it can produce a view into the future. Are you thinking about investing in a new production line? Are you looking into augmenting your data operations with machine learning? You can virtually create this world of tomorrow for you and test scenarios. The tests can be tweaked and done as many times as finding the most optimal solution will take.
Finally, adding machine learning to any industrial process will make the process more intelligent by getting more accurate data and predictions, and understanding also visual and unstructured data. By adding machine learning into your workflow you don’t only open up possibilities to discover previously unseen patterns in your data but also create a single learning-system that can manage complex data.
Digital twins unlock more advanced forms AI
Most of the machine learning today is supervised learning, where the model learns from labeled examples. There are other forms of learning too, that could permit finding unforeseen patterns in the data. One of these is called reinforcement learning, where the model learns in an unsupervised way from rewards when taking actions in a given (simulated) environment.
However, in most of the widely known cases of reinforcement learning, the conditions wouldn’t be possible in the real world. Even cutting-edge reinforcement learning models require a lot of experience to get good. To give an example, The OpenAI Five neural network took 180 years of effective play time to train and still lost to professional players of the game.
So far much of the cutting-edge reinforcement learning works only in games, as the amount of repetition would not be possible in real life. In the digital twin environment you can repeat a scenario or do a test without breaking the system so many times, that reinforcement learning agents can find novel ways to get the reward. In practice this could mean for example, finding out new ways to optimize a mobile network.
How does AI learn
|Supervised (deep) learning||Learn from examples|
|Active learning||Ask for human feedback|
|Transfer learning||Ability to transfer knowledge from one task to another|
|Reinforcement learning||Automated feedback (rewards)|
|Unsupervised learning||Need no feedback|
Key learnings from past projects
We’ve outlined a few learnings from our client work that helps implementing AI into digital twins. To succeed in building an intelligent digital twin, you will need:
- Rich, diverse data from as many different scenarios as possible. This means that possible situations are tested and planned systematically. Different methods can be used in order to remove the possible bias in the data.
- A clear desired outcome. Having a well-thought target in mind will help arriving there. However, starting today will help you understand the requirements faster, as often they are hard to perceive beforehand.
- A rule set of what is acceptable and what is punishable. In a human-in-the-loop AI system people are the ones that give feedback to the model on whether or not its suggestion is suitable. In reinforcement learning the model learns from the feedback it gets: from the rewards.
- Build trust into the AI system by working with people: We believe any AI system is eventually made to assist humans. In our view people should always be responsible, especially in cases where the data can contain a bias or cases that can’t be learned from the data (e.g. they are too rare etc.) In these cases human-assisted data interpretation might be the best way to build trust into the AI system and improve it.
How to get started
Starting to leverage machine learning to improve your digital twin will be an important step in staying technically top-notch. Before you get started, remember to ask yourself these questions:
- What hardware should we invest in today to unlock automation tomorrow?
- What is the first point of autonomy in our machines that would bring a significant business advantage?
- Which autonomous processes are realistic today? Is this feasible? Is legislation / ecosystem ready?
- How much are we looking to invest money / time / labor?
Roadmap to pole position?
- Who should we partner with?