Elastic Ranking-Driven Time Series Feature Optimization for Aeroengine Remaining Useful Life Prediction

This paper proposes a hierarchical recurrent neural network framework that leverages a context-gated mechanism and the elastic sorting of time series indicators, addressing the strong dynamic coupling characteristics of aeroengine operational time series data. This framework addresses the coupling of multi-physical field signals and the interleaved cross-scale degradation characteristics of engine time series data. The framework achieves high-precision predictions of residual service life under complex operating conditions by deeply integrating multi-modal feature engineering with hierarchical time series modeling. An elastic ranking algorithm for time series indexes is designed to adaptively fuse multiple sensors' time domains, frequency domains, and nonlinear characteristics, thereby constructing a dynamic feature space sensitive to load conditions. This approach addresses the limitations of traditional fixed feature combinations in characterizing variable load conditions. Additionally, a multi-scale stack dilation convolution module is introduced to exponentially enlarge the receptive field, allowing for the capture of long-range dependencies in weakly degraded signals. A bidirectional time-frequency converter decomposes the signal into low-frequency trend components and high-frequency anomaly components, enabling coupling characterization through cross-scale feature interactive gating. Furthermore, context-gated cyclic units based on hierarchical cyclic networks are constructed to hierarchically integrate multi-modal features while dynamically adjusting time-dependent weights, thereby enhancing the model's causal reasoning ability for long sequence degradation patterns. In four typical experiments conducted on the C-MAPSS dataset, the ERCG-HRNN demonstrates significant improvements in RMSE and Score compared to existing methods.