Aufrufe
vor 13 Monaten

PuK - Process Technology & Components 2023

  • Text
  • Harnischcom
  • Maintenance
  • Compressed
  • Monitoring
  • Hydrogen
  • Vacuum
  • Components
  • Compressors
  • Website
  • Valves
  • Pumps
A technical trade magazine with a history of 60 years.

Leading article assume

Leading article assume that the neural network can represent the underlying principles. It will therefore obtain corresponding good results for new, unknown input data as well. Results of such a neural network with corresponding training are shown in Figure 3. The Convolutional Neural Network (CNN) used in this case is composed of a series of mathematical convolutions. Here the individual entries in the respective convolution matrix are the variables to be optimised. With this construct, the same processing steps are applied at every position of the frequency spectrum. Hereby the principle that can be determined at a point in the spectrum from its surrounding values must be able to define whether this is a regular peak. Such peaks do not have to be “learned” separately for each frequency value, but are executed in the same way for every frequency. As a result, the network then supplies a value between 0 and 1 for each frequency, reflecting how “certain” the network is that there is a real peak at the respective position. As a design parameter, one now still has to choose a suitable threshold at which one assumes that there is in fact a peak. Figure 3 presents an examination of the influence of this threshold. A signal of three vibrations (green markings) and additional white noise was examined for this purpose. Here the ratio of the vibration amplitude to the average noise intensity and the selected threshold were varied. The average noise intensity increases from no noise on the left to intense noise on the right. The required threshold for the positive identification of a peak (red markings) increases from top to bottom. One can see that increasing noise causes additional false positive identifications (see the right-hand column in particular). The number of false results can be reduced by choosing a higher threshold. However, note that setting the threshold too high can also cause false negative results (existing peaks are missed) (see the bottom row in particular). When a suitable threshold is selected, all real peaks can be identified and there are few false positive results, even with intense noise. Overall it was shown in this case that the detection of peaks using neural networks produced better results than comparable, conventional peak detection methods. Especially in cases with high noise levels, neural networks were able to produce signifi cantly better results. This example shows how AIbased algorithms can complement/ improve classic approaches or support manual evaluations with additional information in the evaluation of vibration spectrums. Such methods also have potential for subsequent evaluation steps. A neural network could provide comparable support in obtaining a more exact, undisturbed value for the respective amplitudes of the individual frequencies. In contrast to manual evaluation, algorithms can, for example, take into account the information from numerous different window functions at the same time. Additional application possibilities for AI-based methods can also be identified for the subsequent evaluation of the peaks that are found. For example, grouping multiple peaks into a basic vibration and their corresponding harmonics is a task one should be able to automate using machine learning. Overall, AI-based methods have great potential for the automation of evaluation steps that currently have to be performed manually as a rule. More elaborate evaluations, which are currently carried out manually in isolated cases only, will therefore be automated in the future and available for live monitoring. Consequently, increasingly small changes in the spectrum can be evaluated. Changes (and possible impending damage) can be detected earlier, and discerning between different changes (for example, different types of damage) can be improved. Prof. Dr.-Ing. Eberhard Schlücker Prof. (ret.), advisor on hydrogen and energy issues and Dominik Haspel 12 PROCESS TECHNOLOGY & COMPONENTS 2023

14th – 15th June 2023 / Globana Trade Center Leipzig/ Schkeuditz Save the Date The largest national meeting for industrial valves & sealing technologies / 08th – 09th November 2023 / Jahrhunderthalle Bochum DIAM-DDM.DE

Wählen Sie die gewünschte Fachzeitschrift

fng MAGAZIN - Food · Nonfood · Getränke · Tobacco