Session: 05-05 Data-Driven Methods and AI for Diagnostics

Paper Number: 124076

124076 - IndRNN-Based Data-Driven Modeling Integrated With Physical Knowledge for Engine Performance Monitoring 

Monitoring the whole performance status of aircraft engines is of paramount importance for ensuring flight safety and prognostic health management. Currently, model-based monitoring methods fall into three categories: physical-based, data-driven, and hybrid. Physical-baesd models face challenges due to mismatches with individual engines as a result of manufacturing, assembly, installation and performance degradation. With the development of artificial intelligence and sensor technology, data-driven methods become a feasible option. Models constructed from sensor data of individual engines take into account the differences among engines and bridge the gap between them and engines with real-time data. Purely data-driven models, lacking in physical knowledge, face challenges in terms of poor interpretability. Hybrid methods, combining the strengths of both, are categorized into two types: one integrates physical knowledge into data-driven modeling, while the other involves using data-driven methods to compensate for physical models. Hybrid methods offer distinct advantages in engineering applications.

The objective of this work is to introduce a data-driven modeling approach that integrates the engine's physical knowledge. We validate its strength by utilizing running data from both civil aviation and military turbofan engines to predict thrust and exhaust gas temperature. Firstly, component networks are established for each engine component (e.g., fan, turbine, nozzle) using the independently recurrent neural network (IndRNN), attention mechanism, and residual network (ResNet). IndRNN is used to capture spatiotemporal features, the attention mechanism uncovers critical features within sequential data, and ResNet is employed to prevent the vanishing gradient problem caused by the excessive complexity of the whole engine model. Through feature matching, specific functions are assigned to each component network, analogous to real-world roles such as the combustion chamber's role in combustion rather than air compression. Subsequently, based on the physical spatial alignment of engine components, the data transfer between component networks is determined to establish the whole engine model. Data augmentation method is investigated to address the issue of non-uniform distribution of engine working states in the training data, thereby improving the accuracy throughout the entire working process. Normalization methods and model hyperparameter settings are also investigated. Finally, comparisons with recurrent neural network and long short-term memory-based purely data-driven models are conducted to demonstrate the advantages of this work.

The outcomes show that compared to purely data-driven models, this work achieves higher precision, stability, and interpretability throughout the entire engine working process, encompassing both steady states and transient states. The modeling method demonstrates remarkable flexibility and is applicable to engines of various structures, which is a potential area of future research. 

Presenting Author: Xiao Dasheng School of Power and Energy, Northwestern Polytechnical University

Presenting Author Biography: A Ph.D. Candidate at School of Power and Energy, Northwestern Polytechnical University.
Education background:
From September 2017 to March 2021, the author studied at the School of Power and Energy of Northwestern Polytechnical University and obtained a bachelor's degree in aircraft engine engineering;
Research interests: Machine learning, Prognostic health management; Simulation on the performance of aircraft engine.

Authors:

Dasheng Xiao School of Power and Energy, Northwestern Polytechnical University
Hong Xiao School of Power and Energy, Northwestern Polytechnical University
Zhanxue Wang School of Power and Energy, Northwestern Polytechnical University