Dataset for System Identification of an Advanced Geared Turbofan Model Using Neural Networks

System Identification of an Advanced Geared Turbofan Model Using Neural Networks This repository contains the data, code (also available in GitHub) and models for the paper "System Identification of an Advanced Geared Turbofan Model Using Neural Networks," presented at the AIAA Scitech 2025 Forum. The project focuses on modeling the dynamics of an advanced geared turbofan engine using both classical system identification techniques and various deep learning architectures. The data are the .mat files. The res.zip contains the prediction results and models.zip contains the weights of the networks. Overview The primary goal of this research is to develop a high-fidelity simulation model of a turbofan engine using only input-output data. We explore and compare two main approaches: Classical System Identification: A "gray-box" Hammerstein model structure, where a nonlinear static model (quasi-steady) is identified using multivariate orthogonal functions, followed by the estimation of dynamic parameters for the unsteady behavior. Deep Learning: A "black-box" approach using several neural network architectures to model the engine's dynamic response. The performance of these models is evaluated on their ability to predict five key engine outputs: Fuel flow (W_f) Net thrust (F_net) Low-pressure shaft speed (N_1) High-pressure shaft speed (N_2) Turbine temperature (T_45) The results demonstrate that a novel neural differential architecture significantly outperforms both the classical system identification method and other neural network approaches.基于神经网络的先进齿轮传动涡扇发动机模型系统辨识 本仓库包含发表于AIAA Scitech 2025论坛的论文《基于神经网络的先进齿轮传动涡扇发动机模型系统辨识》（System Identification of an Advanced Geared Turbofan Model Using Neural Networks）的配套数据、代码（代码亦托管于GitHub平台）与模型文件。本项目旨在结合经典系统辨识技术与多种深度学习架构，对先进齿轮传动涡扇发动机的动力学特性开展建模研究。 本研究所用数据存储于.mat格式文件中。res.zip压缩包包含模型预测结果，models.zip压缩包则存放各神经网络的权重参数。 ## 研究概述 本研究的核心目标是仅依靠输入输出数据，构建高精度的涡扇发动机仿真模型。本次研究探索并对比了两大类建模方案： 1. **经典系统辨识**：采用“灰箱”哈默斯坦（Hammerstein）模型架构，首先通过多元正交函数辨识得到非线性静态准稳态模型，随后针对非稳态特性估计动态参数。 2. **深度学习**：采用“黑箱”建模思路，借助多种神经网络架构对发动机的动态响应进行建模。 上述两类模型的性能将通过对五项核心发动机输出参数的预测能力进行评估，具体参数如下： - 燃油流量（Fuel flow, W_f） - 净推力（Net thrust, F_net） - 低压轴转速（Low-pressure shaft speed, N_1） - 高压轴转速（High-pressure shaft speed, N_2） - 涡轮温度（Turbine temperature, T_45） 研究结果表明，一种新型神经微分架构的建模性能显著优于经典系统辨识方法以及其他各类神经网络方案。