Aero-engine multivariable reinforcement learning control method based on input and output information

Abstract

The invention discloses an aero-engine multivariable reinforcement learning control method based on input and output information. The method comprises the steps of selecting an engine multivariable state based on input / output information, constructing an output variable of a multivariable execution mechanism action, designing a deep neural network based on a depth deterministic strategy gradient algorithm, setting a control performance-oriented reward function, setting a deep neural network convergence condition, updating the deep neural network according to experience in the experience playback set, and finally achieving intelligent control over the aero-engine through the designed deep neural network. According to the method, the multivariable input / output information of the aero-engine is considered, a DDPG deep reinforcement learning multivariable controller with active interaction and autonomous exploration capabilities is constructed, the fuel flow and the nozzle throat area are adjusted, the control strategy is corrected and perfected in real time in the learning process, and high-level and strong-robustness control of two key variables of the rotating speed and the pressure ratio of the low-pressure rotor of the aero-engine is realized.

Description

Technical field [0001] The present invention belongs to the field of aviation engine control, and in particular to an aeronautical multivariate strengthening learning control method based on input output information. Background technique [0002] Avio engine has high complexity and strong coupling, plus complex working environment, is very demanding in various aspects of control accuracy, adjustment time, stability and fuel consumption. In terms of aerospace engines for multivariate characteristics, the single input single output control system is limited, and only a single control target can meet the demand indicators, and the complex coupling association can not process multiple targets simultaneously. Therefore, multi-input multi-output (multivariable) control is inevitable, and by increasing control variables, the overall operational performance of the aeroengine can be effectively improved. However, traditional multivariate control systems typically combine multiple single v...

AI Technical Summary

Problems solved by technology

However, there is little research on aeroengine control based on deep reinforcement learning methods, and no multivariable control design using input / output information

Images