Prediction of the static force characteristic of annular gaps using neural networks

Authors

Maximilian M. G. Kuhr  - Technische Universität Darmstadt, Germany
Peter F. Pelz - Technische Universität Darmstadt, Germany

Abstract

Turbomachinery, such as centrifugal pumps, are developed and operated in the area of conflict between effort, availability and acceptability. The Availability is essentially determined by the force characteristic of annular gaps, such as seals and bearings. In general, these properties depend on the respective operating point of the annulus, i.e. shaft speed, mean axial fluid velocity, circumferential fluid velocity at the gap inlet, the lubricant, the gap geometry as well as the rotor position in the lubrication gap. However, the identification of these properties is often associated with great experimental or numerical effort. In the development phase of modern turbomachinery, however, a time-effective calculation of the dynamic properties is crucial. Simplified calculation methods such as the Reynolds differential equation and the bulk-flow model do exist, but it has been shown that even these simplified one or two-dimensional simulations may require a large amount of computing time. To address this, a neural network for the prediction of the static force characteristic of annular gaps is presented. Using MATLAB's Deep Learning Toolbox, a suitable neural network structure is selected in preliminary tests. Subsequently, the network is trained and validated on the basis of experimental and numerical data. Explicit care is taken to ensure that the validation is only carried out for selected operating points that have not previously been included in the training of the network. Apart from the effort of training, the neural network is able to predict the static force characteristics of annular gaps instantaneously, without significant uncertainties compared to the previously mentioned calculation methods.