Comparison of Static and Dynamic Neural Networks for Limit Cycle Oscillation Prediction

Abstract

A dynamic artificial neural network in the form of a multilayer perceptron with a delayed recurrent feedback connection is investigated to determine its ability to predict linear and nonlinear flutter response characteristics. Flight-test results show that limit cycle oscillation (LCO) response characteristics are strongly dependent on Mach number in the transonic flight regime. This effect is also evident in the classical transonic small-disturbance theory governing equations. A dynamic network is considered in order to examine the effects of sequential Mach-number dependence on the network's predictive capability. The architecture of a dynamic network allows for modeling data dependent on a sequentially or linearly increasing parameter (usually time, but in this case Mach number). The predictive capabilities are compared to those of a static artificial neural network. The network is developed and trained using linear flutter analysis and flight-test results from a fighter test. Eleven external store carriage configurations are used as training data, and three configurations are used as test cases. The dynamic network was successful in predicting the aeroelastic oscillation frequency and amplitude responses over a range of Mach numbers for two of the test cases. The dynamic network showed slightly better correlation to flight-test results for the typical LCO test case but slightly worse correlation for the flutter case. Predictions for the nontypical LCO test case were not good for either network.