Neuro-fuzzy modeling for forecasting future dynamical behaviors of vibration testing in satellites qualification
- Paper number
IAC-08.C2.1.11
- Author
Dr. Ernesto Araujo, Instituto Nacional de Pesquisas Espaciais (INPE), Brazil
- Coauthor
Mr. Rogerio Marinke, Brazil
- Year
2008
- Abstract
A neuro-fuzzy modeling for forecasting the future dynamical behavior in vibration testing during satellite qualification is proposed in this paper. Vibration testing is employed for emulating vibrations present during the launching and it is one of the tasks carried out to verify the structure of the satellite and their sub-systems in order to appropriately support the launcher lift-off and to guarantee useful life when in orbit. There are different levels of excitation during vibration testing in order to verify and assure that the satellite and their sub-systems will support the efforts when in orbit or during the launching. Due to that estimating future dynamical behavior when using high amplitude testing signals is important to safe satellites or other space devices. Moreover, the analysis of the dynamical behavior can help not only to avoid breaks and other damages but also allows feasible adjustments in the structure model. The estimation of future dynamical behavior may be determined by using different techniques of system identification. Space systems, such as satellites, are inherently non-linear. While conventional identification techniques are adequate for models or systems linear in parameters, for systems that are usually nonlinear, identification methods used in linear systems are not appropriated and suitable nonlinear approaches should be used. The objective in this paper is to show the feasibility of employing a nonlinear identification technique denominated neuro-fuzzy modeling for forecasting the future behavior of vibration systems. Fuzzy sets is appropriate to deal with uncertainty, imprecise measures and incomplete information. Nevertheless, it does not allow learning by examples. In turn, artificial neural network are low-level computational algorithms presenting learning capacity. This approach is effective when processing numerical data and presents distributed computational characteristic allowing that each node in the network adjusts its connections to obtain the best possible input-output mapping after learning from data. When combining neural networks and fuzzy systems it is possible to obtain hybrid models with the capacities of learning, adaptation, optimization, robustness, dealing with large amounts of numerical data and, finally, knowledge representation through fuzzy rules, as well as the ability to deal with imperfect data. The neuro-fuzzy model is used to describe the dynamical behavior through data measured during the qualification of space systems in Integration and Testing Laboratory (LIT) at the National Institute of Space Research (INPE). The problem is composed of two parts. In the first one, the model uses part of signals of low amplitude for training the neuro-fuzzy system and then it is validated with the remaining set of data. Afterwards, this proposed neuro-fuzzy model is employed to estimate a distinct dynamical behavior when a new input signal of high amplitude is applied to the space system. The criterion for validation of the models adopted was Root Mean Square Error. Actual data with high and low amplitude signals were used for eliciting the neuro-fuzzy model. Results of the structural model used in the design of the satellite and of their sub-systems are confronted with the real behavior presented by the structure, allowing eventual adjustments. Results show the neuro-fuzzy modeling is a feasible solution for forecasting dynamic satellite behaviors under distinct exogenous input. It is also shown that the models have good capacity of generalization. These results were improved when used the variation of the signal of low amplitude as input.
- Abstract document
- Manuscript document
IAC-08.C2.1.11.pdf (🔒 authorized access only).
To get the manuscript, please contact IAF Secretariat.