An Adaptive MVMD-Based Stacking Ensemble Framework for Bearing Reliability Assessment and Prediction

Abstract

Rolling element bearings are critical components in rotating machinery, yet assessing and predicting their reliability under heavy industrial noise remains challenging. Existing methods suffer from three major limitations: (1) single-channel signal processing and single-scale indicators lack robustness against non-stationary noise; (2) classical multi-channel decomposition methods, such as multivariate variational mode decomposition (MVMD), rely on empirical parameter tuning, which frequently leads to over- or under-decomposition; and (3) monolithic deep architectures and homogeneous ensemble models suffer from prediction drift and generalization bottlenecks during long-term temporal extrapolation. To address these issues, this paper introduces an automated framework that combines adaptive multi-channel signal purification with a heterogeneous stacking ensemble (HeteroStack-LR). Unlike conventional MVMD pipelines that fix [K,α] empirically, the Sequoia Optimization Algorithm (SOA) autonomously determines the globally optimal configuration, achieving a mean SNR of −2.08dB—a 1.88 to 2.50dB improvement over standard VMD/MVMD baselines—along with up to a 37.1% reduction in computation time. Rather than relying on conventional single-metric intrinsic mode function (IMF) selection, we construct a multi-domain hybrid index integrating the Fault Correlation Factor, energy ratio, and refined composite multiscale dispersion entropy (RCMDE) to robustly identify noise-resistant components, thereby enhancing denoising quality by 22.4% to 32.2% over single-scale criteria. Furthermore, contrasting with linear PCA-based reduction, Diffusive Topology Neighbor Embedding (D-TNE) effectively preserves the nonlinear manifold structure of degradation trajectories in a low-dimensional space. Finally, a heterogeneous stacked ensemble featuring an out-of-fold (OOF) leakage-prevention strategy and a logistic regression meta-learner is designed to suppress prediction drift while avoiding the over-parameterization typical of deep architectures. Experimental results across four bearing datasets demonstrate that HeteroStack-LR achieves a minimal MAE of 0.063 with a variance of ≤±0.002, outperforming state-of-the-art deep architectures (such as TCN, CNN-LSTM, BiLSTM-Attention, and Transformer) as well as classical baselines (Bi-LSTM, CNN, and LSSVM). Ablation studies confirm that removing SOA and MVMD degrades MAE by 12.7% and 19.0%, respectively, validating that the framework’s strength stems not from any isolated module, but from the end-to-end synergistic integration of signal purification and reliability prediction.