基于相似变换的涡扇发动机智能动态推力估计方法

An Intelligent Dynamic Thrust Estimation Method for Turbofan Engines Based on Similarity Transformation

关键词:  涡扇发动机  深层动态神经网络  相似变换  数据处理  推力估计

Key words:  Turbofan engine  Dynamic deep neural network  Similarity transformation  Data processing  Thrust estimation

摘要:

智能推力估计面临飞行包线大、工作状态多变带来的数据采集和处理问题，获得的训练数据难以覆盖整个飞行包线的各种过渡工作状态，为此本文提出一种基于相似变换的推力估计数据处理方法。通过机理分析选择推力估计器输入，以相似变换对推力估计的输入和输出数据进行处理，并设计了基于输入延迟的深层动态神经网络来实现动态推力估计。非训练数据区域的动态仿真结果表明，相似变换后，深层动态神经网络的最大推力估计误差降低了62.20%，平均误差降低了43.50%；未进行相似变换时，相比深层静态神经网络，深层动态神经网络的最大推力估计误差降低了43.42%，平均误差降低了2.35%，仿真结果表明了本文所提出的数据处理方法和动态推力估计结构有效性。

Abstract:

Data acquisition and processing problems aroused by wide flying envelope and multi-working-state variation are confronted with intelligent thrust estimation research. The available training data can hardly cover all the transition state in the full envelope. Therefore， a similarity principle based data processing method for thrust estimation is proposed. The thrust estimator inputs are selected by working principle analysis. The similar transform is applied to the inputs and output of the estimator， and a dynamic deep neural network （DDNN） with input delay is designed to realize the dynamic thrust estimation. Simulation results at the envelope different from the training data show that the maximum thrust estimation error and average thrust estimation error of the DDNN are decreased by 62.20% and 43.50% ， respectively， after similar transform data processing. Compared with the deep static neural network （DSNN）， the maximum thrust estimation error and average thrust estimation error of the DDNN are decreased by 43.42% and 2.35% ， respectively， without similar transform data processing. Simulation results show the effectiveness of the date processing method and the dynamic thrust estimator structure.