Anwendung von Verfahren der künstlichen Intelligenz im Bereich elektrischer Antriebe

This article deals with the possibilities of using methods and procedures that are becoming established in the rapidly growing field of artificial intelligence (AI). The focus is particularly on fault detection and condition assessment (predictive maintenance) of electrical machines. In principle, this is a very well-researched and published topic. A key point regarding the known and common procedures is that they are usually designed for a specific application or use case. Any major change or any other motor type usually requires a new definition of the relevant parameters of the procedure. Another problem is that in many cases, limit values or hysteresis thresholds must be defined. If these are set too strictly, warnings or faults will be detected reliably, but the probability of unjustified messages will increase. If the tolerance is too wide, the possibility of reliable fault detection is reduced. The use of self-learning and thus feature-recognizing methods opens up the possibility of circumventing the aforementioned difficulties and developing solutions that cover a significantly broader range of applications for variants of electric machines without application-dependent adjustment. Three different, practically tested examples are used to show the possibilities, but also the pitfalls, that arise in this context. A very transparent method based on an Arduino® board shows that simple and relatively universal fault detection is possible with very little effort (Figure 1). Another, significantly more powerful example demonstrates bearing condition detection. Two different approaches are compared. One variant consists of a hybrid, hierarchical AI approach (Hierarchical Extreme Learning Machine - HELM) consisting of an autoencoder (AE) and a single-class classifier. The alternative variant is based on a method called Generative Adversarial Network (GAN). Finally, an example based on fault detection using sampled commutator surfaces shows the danger that, if the target is chosen unfavorably or the approach is disadvantageous, the effort and benefit can become unbalanced. This prevents practical application. Overall, it is shown that the possibilities and methods from the field of artificial intelligence, far removed from the language models that are the focus of media attention, also offer interesting approaches for use in the field of electrical machines.