The topic of artificial intelligence (AI) has fascinated people for ages. What if one day machines are intelligent enough to take over the earth and we humans have to serve them as slaves? This is the stuff of which numerous novels and Hollywood classics are made. The reality, however, is quite different. Much less spectacular at first glance - but on closer inspection, at least as relevant from an industrial perspective.

For years, the requirements for more flexible production systems have been increasing due to ever shorter cycles on the consumer market. Whereas in the past, once a classic production system was put into operation, it always manufactured the same product for 20, 30 or more years, today modern production systems - i.e. machines, automation technology and sensor technology - are expected to be able to react flexibly to new market requirements - always with the ultimate goal of 'batch size 1' in mind.

This is difficult to achieve with traditional programming, parameterization and commissioning. Artificial intelligence methods could provide a remedy in the future, allowing the system and its components to adapt to new conditions through continuous learning - much like we humans do. But there is still a long way to go until then, and it will take a few more years. Nevertheless, artificial intelligence methods have already found their way into selected industrial applications.

Predictive maintenance applications are leading the way. While early applications only monitored predefined threshold values - such as for certain frequencies in the system - AI technologies such as machine learning are now increasingly being used to flexibly train the algorithm's evaluation system. Measured data from productive use - including those cases that led to errors or even system failure - are made available to the machine learning algorithm in order to automatically create a corresponding AI model. Development platforms such as Matlab provide the developer with numerous different machine learning methods as well as apps for selection and training.

The trailblazer: predictive maintenance

Figure 2: The Signal Analyzer app: It can be used to detect outliers, filter signals and extract features.

© Mathworks

The user selects from numerous widely used classification and regression algorithms, compares models using standard metrics - Support Vector Machines, Nearest Neighbor, Ensemble Methods or Decision Trees - and exports promising models for further analysis and integration. The user can easily and efficiently apply the complete workflow to the data with the Classification and Regression Learner Apps(Fig. 1). The appropriate pre-processing of the raw data is essential for success. Here too, Matlab provides tools with corresponding extensions to automatically detect outliers, filter signals, transfer signals from the time domain to the frequency domain(Figure 2), reduce the dimensionality of the data and extract features, for example.

The fully implemented and tested algorithm is then transferred to the production system, where it runs continuously 24 hours a day, 365 days a year - as is the case at Mondi in Gronau - and ensures that potential machine downtimes and production losses are detected before they occur.

Why are machine learning algorithms currently being used primarily for predictive maintenance applications, while other areas of application are still being ignored? A predictive maintenance system can be installed and operated in parallel with ongoing operations without actively intervening in them. The use of artificial intelligence for areas such as machine control or flexible reconfiguration during operation, which intervene directly in the production process, still requires a greater degree of experience and trust from most potential users. Furthermore, the business case for the introduction of predictive maintenance applications is comparatively simple. Every fault or even machine downtime that can be detected at an early stage and every unnecessary maintenance intervention that can be avoided saves money and contributes to the economic success of the predictive maintenance application.

Machine Learning, Deep Learning, Reinforcement Learning

Figure 3: Three forms of artificial intelligence could play a key role for mechanical and plant engineering in the future: Reinforcement Learning, Machine Learning and Deep Learning.

© Mathworks

Machine learning for predictive maintenance is just one form of artificial intelligence that could play a role in mechanical and plant engineering in the future. Overall, three groups of AI methods are currently emerging that will become increasingly important for industrial use in the coming years(Fig. 3):

• Machine learning: the basic idea is to find patterns in data (numerical values, images) in order to then make a prediction based on previously unknown data. Statistical methods enable machines to 'learn' tasks from data - without explicit programming.
• Deep learning: A form of machine learning in which neural networks with many layers learn the representations and tasks 'directly' from data.
• Reinforcement learning: Another form of machine learning in which the system receives feedback as to whether the right or wrong decision was made. With a sufficiently large number of repetitions, the system is finally able to predict the correct results.

Simulation a must

Simulation models play an important role in the training of AI algorithms. Measurement data from the field, which is not available to the extent required, is supplemented with values from the simulation. This applies in particular to error data, which is naturally not available in the quantity required to train the AI models. Especially in the field of reinforcement learning, simulation models are essential for providing feedback to the algorithm. They can be used to simulate control and decision-making algorithms for complex systems such as autonomous robots. Using deep neural networks, polynomials or look-up tables, the policies are implemented and trained by enabling interaction with environments represented by models in Matlab or Simulink.

Authors:
Dr. Rainer Mümmler is Senior Application Engineer at Mathworks;
Philipp Wallner is Industry Manager EMEA at Mathworks.