REXpanded-LSTM: A Recurrent Expansion Framework for Milling Machine Degradation Prognostics

Abstract

Accurate prediction of complex, nonlinear degradation processes is critical for prognostics and health monitoring applications. In this work, we propose Recurrent Expansion eXpanded Long Short-Term Memory (REXpanded-LSTM), a novel recurrent neural network framework that recursively expands hidden states via Principal Component Analysis (PCA) and incorporates previous predictions into augmented input sequences. This iterative procedure allows the network to progressively refine temporal representations and capture higherorder dependencies in highly non-stationary and noisy signals. The method is evaluated on a challenging synthetic dataset designed to mimic the degradation of milling machines, capturing phenomena such as exponential amplitude growth, nonlinear frequency drift, impulsive tool impacts, and strong stochastic noise. Experimental results demonstrate that the REXpandedLSTM outperforms conventional LSTM models, achieving a root mean squared error of 0.201 compared to 0.253, a mean absolute error of 0.185 compared to 0.236, and an $R^{2}$ of −4.43 compared to −7.61. The predictions closely follow the true Health Index (HI) trajectories and exhibit stable convergence across training rounds. The framework achieves these improvements without requiring extensive preprocessing or manual feature engineering, offering a robust and generalizable approach for sequential time-series prediction in milling machine prognostics, condition monitoring, and other data-driven modeling applications.