基于内嵌物理约束神经网络模型的航空发动机数字工程模型

An Aeroengine Digital Engineering Model Based on Physics-Embedded Neural Networks

关键词:  航空发动机  数字工程模型  内嵌物理约束神经网络  性能参数预估  特征处理

Key words:  Aeroengine  Digital engineering model  Physics-embedded neural network  Performance parameter prediction  Feature processing

摘要:

基于数字化模型的航空发动机故障诊断与健康管理系统是航空发动机数字化智能化的重要应用，用于航空发动机状态监测和性能预估的数字化模型是健康管理系统的核心之一。本文给出了一种融合航空发动机领域知识与神经网络模型的策略，构建了内嵌物理约束的神经网络架构，基于该架构建立了用于航空发动机推力预估的数字模型。此外，给出了一种特征筛选方式，并利用不同数据集对模型进行了验证。计算结果表明：数字模型推力预估的平均相对误差和峰值相对误差均小于常规神经网络模型。在一定的模型规模下，基于架构的数字模型的峰值相对误差仅为常规神经网络模型的1/4。通过物理约束，克服了数据驱动模型对大数据的依赖，指导了神经网络层的超参数设置。

Abstract:

A digital model-based prognostics and health management （PHM） system is crucial for digitalization， intelligence in aeroengine. Among all digital models， an aeroengine performance digital model is one of the basic modules for PHM system， which is used for condition monitoring and performance prediction on aeroengine. In this work， a strategy for creating a performance digital model to predict aeroengine thrust is given. The strategy is to combine aeroengine domain knowledge and artificial neural networks， which is to create an architecture for tailoring the neural network model with physical information. More， the given model is designed to address feature selection. The application of the given model to aeroengine thrust prediction demonstrates its effectiveness in accuracy with the different testing datasets. Compared with the conventional neural network， the average relative error of the architecture-based model is small， and the max relative error of the architecture-based model is only 1/4 of it under the same model size. With physical constraint， the model is less reliant on training data， and the number of layers and the hyperparameters in the neural networks model are intervened.