Mr. Wallner, artificial intelligence and machine learning are buzzwords that adorn the hype surrounding Industry 4.0. What do these terms have to offer apart from nice-sounding messages?

Philipp Wallner: Of course, I can only speak for our company MathWorks and our customers. But our customers from the mechanical and plant engineering sector are now very much looking at what can be put into practice in terms of Industry 4.0. Two areas of application that the industry is looking at extensively today are virtual commissioning based on model-based development and predictive maintenance based on machine data using artificial intelligence methods.

While conventional methods for virtual commissioning are aimed exclusively at using a model of the machine - i.e. a 'virtual machine' - to test the programs that will later be executed on an industrial control system in advance, model-based development goes one step further. Both the machine and the controller - i.e. the functionality that will later run on the controller - are implemented in the model and used throughout the entire development cycle for simulation, verification and automatic code generation. This means that the cost of creating the model is significantly lower than with traditional virtual commissioning.

And the models from the simulation then also serve as the starting point for the digital twin of the system?

Philipp Wallner: Correct. Algorithms that combine both domain expertise in the form of models and artificial intelligence technologies - such as machine learning or deep learning - have proven to be particularly effective for predictive maintenance, but also for other areas such as optimizing plant performance or energy consumption.

Keyword 'artificial intelligence' - where do you still see the biggest challenges?

Philipp Wallner: Firstly, there is the lack of sufficient error data. If you don't allow components, machines and systems to fail during operation, there is naturally no or very little fault data. However, this is absolutely essential in order to train the algorithm accordingly. Simulation models provide a remedy here by simulating different fault scenarios and thus generating synthetic fault data.

The second challenge that we often see is that machine learning algorithms are developed in special data science environments that do not offer any options for implementing the algorithms on corresponding target systems such as a PLC or an edge system. We will show how this is implemented with Matlab using a demo with industrial hardware at our stand.