Data-Driven Modeling of Aero-Engine Performance Degradation Models

Abstract

Periodic maintenance is a fundamental method of aero engine maintenance. However, frequent regular maintenance leads to resource wastage and inefficient maintenance. With the increasing technological maturity, integration, and complexity of aero engines, the costs of research and development are rising, creating an urgent demand for advanced Prediction and Health Management technology (PHM). PHM technology enables the transition from periodic maintenance to predictive-based maintenance by modeling aero-engine performance degradation, thereby allowing for life prediction and sensor parameter prediction. Data-driven modeling methods have gained popularity owing to their strong traceability, wide consideration range, and high freedom of adjustment; however, they currently suffer from low prediction accuracy and reliability. To address this issue, this study utilizes NASA’s open data set C-MAPSS for data preprocessing before inputting it into the LSTM training network to obtain function loss values in the forward direction. These values are then updated through the quantum particle swarm algorithm to achieve loss reduction optimization of key parameters to obtain the best QPSO-LSTM aero-engine performance degradation model. Additionally, the PCA method was used for dimensionality reduction modeling by calculating variance percentages of data eigenvalues which determined 11 dimensions as the lowest dimension for progressive reduction modeling. Ultimately, it was found that the QPSO-LSTM combined neural network remaining life prediction achieved a minimum RMSE value of 22.04, which is 43.76% higher than that of the basic model, presenting an excellent linear fitting relationship after crossing threshold values despite not conforming initially with true value trends at initial time nodes.

Keywords

• Data-driven
• aeroengine
• long short-term memory network
• particle quantum swarm algorithm