<p>This study presents a rigorously evaluated diagnostic–prognostic framework for Remaining Useful Life (RUL) prediction in rotating machinery, extending a previously validated probabilistic diagnostic backbone with a multi-layer bidirectional Long Short-Term Memory (BiLSTM) architecture. Probabilistic degradation states are modeled using Gaussian Mixture Models (GMMs), and bidirectional temporal learning is employed to capture long-horizon degradation dynamics in both forward and backward directions. To ensure a fair and controlled comparison, the proposed model is systematically benchmarked against representative CNN-, GRU-, and Transformer-based architectures under identical feature representations, training configurations, and leakage-aware five-fold cross-validation protocols. Results demonstrate that the BiLSTM consistently achieves superior predictive performance across MAE, RMSE, R², and ± 10% tolerance accuracy metrics. Statistical significance is confirmed through paired hypothesis testing with Holm correction, and large effect sizes are observed based on fold-wise paired differences. To enhance transparency, post-hoc interpretability analyses are conducted, including occlusion-based feature attribution and state-conditioned trajectory evaluation, providing insight into key prognostic indicators and the influence of latent degradation states on RUL estimation behavior. Computational analysis further reveals a favorable accuracy–complexity trade-off, supporting practical offline training and near-real-time inference. Overall, the proposed framework provides an interpretable, statistically validated, and computationally feasible approach for predictive maintenance and life-cycle management of rotating machinery.</p>