What our brains have long mastered is now set to revolutionize measurement data interpretation in ISAT: so-called “machine learning”.

While machines can perform complex mathematical calculations within a fraction of a second, they fail when it comes to intuitive questions such as recognizing a cat, which the human brain, on the other hand, solves playfully. Artificial intelligence (AI) can also be used to get machines to perform simple intuitive tasks. Machine learning” is an important tool in this context. Machine learning can be described as the ability of computer programs to learn from experience. The advantage over classical algorithms is that the user does not have to understand the physical processes comprehensively/completely, because this is done by the algorithm. Overall, there are three different types of machine learning: supervised, unsupervised and reinforcement learning.

In supervised learning, the correct results are given for the training data. The algorithm recognizes the correlations between data and result independently and adapts accordingly. Classical tasks are for example the classification of objects in groups or regression/predictions for the future. In unsupervised learning, the system attempts to learn without guidance in order to recognize the relationship between data and simplify complex data sets. In this process, for example, discrepancies can be detected and sorted out. In reinforced learning, the system must observe a given environment and analyze and perform possible actions itself. For correct actions it receives rewards based on classical conditioning, for incorrect actions it receives punishments (negative rewards). An example of this are robots that learn to walk by themselves. Depending on the algorithm and the task, several hundred data are needed for evaluation. For complex tasks, such as image recognition, you need significantly more data sets (tens of thousands to hundreds of thousands). The more data available for learning, the better the algorithm can generalize the underlying data and use it for evaluation.

Input for a “machine learning” algorithm developed by ISAT can be, for example, the primary signals obtained by means of acoustic sensors for different degrees of damage to a component. The user simply tells the algorithm which data set represents which degree of damage. The algorithm does everything else on its own, i.e. it independently determines suitable parameters that it needs for condition assessment and decision-making. Machine learning” is particularly suitable for complex measurement tasks where, for example, no satisfactory assignment to the component condition can be made from the measurement signal using conventional evaluation methods.

At ISAT, the application of machine learning algorithms opens up completely new possibilities for sensor development with regard to the quality of data interpretation and evaluation. But the method is not only suitable for sensor development. Thanks to the know-how of our employees in the field, we can also support your company with a very specific problem. For example, we can integrate their machine data into a “machine learning” algorithm, so that it can be predicted in time whether, for example, a defect in their machine is imminent. We have already been able to convince many of ISAT’s industrial partners of the potential of machine learning, as it delivers much more reliable measurement results than conventional methods.