Remaining Useful Life Prediction of Aero-Engine Based on KSFA-GMM-BID-Improved Autoformer

Abstract

Addressing the limitation of traditional deep learning models in capturing the spatio-temporal characteristics of flight data and the constrained prediction accuracy due to sequence length in aero-engine life prediction, this study proposes an aero-engine remaining life prediction approach integrating a kernel slow feature analysis, a Gaussian mixture model, and an improved Autoformer model. Initially, the slow degradation features of gas path performance parameters over time are extracted through kernel slow feature analysis, followed by the establishment of a Gaussian mixture model to create a health state representation using Bayesian inferred distances for quantifying the aero-engine’s health status. Moreover, a spatial attention mechanism is introduced alongside the autocorrelation mechanism of the Autoformer model to augment the global feature extraction capacity. Additionally, a multilayer perceptron is employed to further elucidate the degradation trends, which enhances the model’s learning and predictive capabilities for extended sequences. Subsequently, experiments are conducted using authentic aero-engine operational data, comparing the proposed method with the standard Autoformer and Transformer models. The results demonstrate that the proposed method outperforms both models in swiftly and accurately predicting the remaining life of aero-engines with robustness and high prediction accuracy.