Local-Global Correction of Simplified EHSV Model using ANN

Modern flight control systems design demands accurate computer models, able to reproduce the behavior of on-board equipment. These high-fidelity models have a large computational cost in terms of memory and time, since highly detailed. On the contrary, for development of diagnostic and prognostic strategies simplified models are needed, since they require smaller computational costs with an adequate level of accuracy. However these low-fidelity models with concentrated parameters, developed from the physic of the problem with different linearizations, are not sufficiently representative in several aspects. The current trend is to make adjust to this model, in order to simulate non-linearity such saturations, water hammer’s rise of pressure, oil leaks. This approach provides increasingly accurate results, but also breaks the concept of a simplified model. The aim of this work is to build simplified models representing the fluid-dynamic behavior of the electro-hydraulic servo-valve (EHSV), which is the most non-linear and complicated to model and the most delicate and prone to wear and failure. Such models need to be simplified, i.e. requiring little computational effort, to be able to run in real time as actuator monitoring. Solutions could be to replace or supplement this type of modelling with artificial neural networks (ANN). These can be interpreted as a data fitting of arbitrary and unknown functions and in the case of n variables. It is basically a modelling strategy that does not rely on the physic of the problem per se, but it finds the relationship that subsists between the input variables and the output of the high-fidelity model. The idea is to start with a suitably two-gain model, add a global correction using ANN techniques, and transform it into a high-fidelity output. For it to be able to work within different operating ranges, it is necessary to make a second correction with n local networks, trained for different temperature and pressure, around an operating condition. A mathematical model is introduced to enable the HF fluid-dynamic model to be sensible to changes in dynamic viscosity and density due to oil’s pressure and temperature. Finally, a Simulink model is generated, capable of performing responses comparable with the detailed model. The results are then compared with the outputs, at the same conditions, of the high-fidelity fluid-dynamic model of EHSV to evaluate the error.