A multi-input based full envelope acceleration schedule design method for gas turbine engine based on multilayer perceptron network

Abstract

Acceleration schedule design is a crucial task in gas turbine engine (GTE) control system design as it dramatically influences the acceleration performance of GTE. The corrected parameter based (CPB) method is the traditional solution for acceleration schedule design, which is simple to implement but could not fully exploit GTE acceleration performance in full envelope. It even fails to prevent the GTE from exceeding its surge margin boundary in some envelope points. In this paper, aiming to improve full envelope acceleration performance of GTE, a novel multi-input based (MIB) method for high-precision acceleration schedule design is proposed. Firstly, the full envelope acceleration schedule (FEAS) design problem is represented, and the CPB method is realized as a baseline. Then, the proposed MIB method is formulated, which integrates a combined input selection (CIS) strategy and a multilayer perceptron (MLP) network. With a weighted integration loss function to evaluate sensors of GTE, the CIS strategy determines appropriate inputs to design a high-precision and robust FEAS. The MLP network is employed to further enhance the FEAS precision. Finally, effectiveness of the proposed method is verified through a series of simulations and flight data verification. Compared with the CPB, the simulation results of acceleration processes under random envelope points indicate that the high-precision MIB method significantly improves acceleration performance of GTE and reduces its possibility of violating the surge margin boundary. Acceleration performances of the two methods are further compared under 1000 random envelope points, and similar results show the effectiveness of the MIB. Moreover, the CIS strategy performs better than other input selection strategies and contributes to enhancing the FEAS design precision and robustness. The MLP network is more qualified to improve FEAS precision than other classical machine learning models.