intelligent control systems
gaz turbine engine
V. I. Vasilyev, I. I. Idrisov, A. S. Makarov
A promising direction to solve problems of designing automatic control systems (ACS) of gas turbine engines (GTE) is their construction in a class of intelligent control systems ensuring the robustness, the adaptability and the fault-tolerance of GTE control under uncertainty conditions. To form the controlling actions to the GTE actuators usually selectors are used now in multivariable GTE ACS. This control method is used for control the fuel supply to the GTE combustion chamber.
The selectors are introduced in GTE ACS to eliminate the areas of the control channel joint work. Moreover, each of GTE control channels operates autonomously, and its parameters are usually chosen with little regard to the interaction with other channels. This allows us to provide the static accuracy and the given stability margins in the separate control channels.
For example, to control GTE by changing the fuel flow to the main combustion chamber at the maximum mode the controlled to adjust GTE parameters should not exceed their maximum values (the upper limit), and the selector should provide the minimum fuel flow consumption (selecting by the minimum). With selecting by the minimum, the implementation of regulation programs nmax = const, T*max = const, p*max = const and others is provided.
If, however, it is necessary to limit the minimum values (the lower limit) of the controlled variables, then the preference is to be given to the controller of the parameter providing the highest fuel consumption, i.e. to use selecting by the maximum. This principle is applied to match the discharge of GTE controllers, the restrictions on lower fuel consumption. The main advantage of the developed method to design GTE neural network (NN) controller, based on the use of the neural network ensemble in comparison with classical methods of ACS design with control channels selecting is an ability to adapt the characteristics of NN controller to real experimental data.
A problem of hardware NN implementation with use of FPGA is a complex technical problem. Its solution needs to solve the following tasks: the development of the blocks for elementary operations (multiplication, addition, activation functions calculation), executions the development of the block-based elementary operations units (single and multiinput) neurons, the development of the given NN structure based on the FPGA technology.
To develop the blocks of elementary operations it is necessary to fulfill the following requirements: the inputs, the weights and the biases of the neurons are the real numbers, the fractional part of which contains two digits after the decimal point, the activation functions are considered as the tangent and logistic (sigmoid) functions. The software and hardware implementation of GTE ACS was realized in CAD Quartus. The simulation of the NN controller with selecting GTE control channels and NN-model of GTE actuator in CAD Quartus was fulfilled, the standard error deviation by the main parameters in simulation experiments was amounted to 1.61% and 1.54%.