Mr. Wallner, which technologies are interesting for AI implementation?

Philipp Wallner: In the broad field of application of artificial intelligence, we see that two techniques in particular are becoming increasingly present in industry - machine learning and deep learning. Both are based on the approach of using statistical models applied to large amounts of data instead of a parametric model - i.e. formulated equations.

And what prevents users from actually implementing these methods?

Philipp Wallner: These methods require significant amounts of existing measurement data in order to train the statistical models. However, this data is very often simply not available. In particular, error data, i.e. measurement data from previous error cases, is usually lacking in practice to the extent that it would be required.

And what about the industrial sector?

Philipp Wallner: Another reason why AI has not yet been widely implemented is the fact that it is still a relatively new technology as far as industrial production is concerned. Here, machine and plant manufacturers first need to gain the necessary experience in suitable projects - because not every problem is automatically suitable for the use of machine learning or deep learning.

The integration of AI algorithms into production systems is also a challenge. MathWorks offers a comprehensive range of tools both for the development of machine learning and deep learning algorithms and for their implementation in production.

What are the most exciting fields of application in manufacturing and production?

Philipp Wallner : The most common use case for AI that we currently see is for predictive maintenance - in other words, the predictive maintenance of machines and systems. Machine learning or deep learning models are used here to predict future developments from historical data and optimize maintenance intervals. The algorithms can be integrated into the production system, where they run online as a digital twin. Overall, we are seeing a shift towards more data-based services on the market.

Another use case is visual quality control. Here, too, deep learning networks are increasingly being used in combination with conventional image processing and machine vision algorithms.

Overall, however, increasingly flexible production with the goal of 'batch size 1' also plays a key role here. For example, the production flow is often distributed dynamically to different production modules with the help of optimization algorithms; machine learning or deep learning helps to predict the corresponding load limits of the modules.

What range of tools does Mathworks offer for implementing AI?

Philipp Wallner : With the 'Matlab R2018b' release, we offer the deep learning toolbox, which provides engineers and scientists with a framework to design complex network architectures more easily and increase the performance of deep learning models. We have also joined the ONNX community to enable collaboration between users of Matlab and other deep learning frameworks. Users are also supported by the Deep Network Designer app, which can be used to create complex network architectures or modify pre-trained networks for transfer learning. Direct support for Nvidia, Intel and ARM libraries also enables higher performance when training and executing networks - and higher training performance is also possible thanks to support from cloud providers. Users can tackle the problem of missing error data with Simulink by simply simulating it.