A Multi-Head Attention-Based Deep Learning Method for Condition Monitoring in Rotating Machinery Diagnostics Using RVMD-FNDNLP Feature Preprocessing

Abstract Rotating machinery condition monitoring is essential for reliable fault detection and maintenance decision-making in industrial systems. However, nonstationary bearing vibration measure-ments often produce statistical feature representations with non-linear coupling, redundancy, and class overlap, thereby limiting diagnostic reliability.In addition, conventional feature extraction and dimensionality reduction methods often suffer from feature redundancy and limited class separability. To address these is-sues, this paper proposes a bearing condition-monitoring method based on regularized variational mode decomposition (RVMD) and farthest-nearest distance-based neighborhood and locality projections (FNDNLP). By incorporating a regularization strate-gy into variational mode decomposition, RVMD improves de-composition efficiency and enhances fault-relevant modal infor-mation. Multiple statistical features are then extracted from the selected informative modes to construct a high-dimensional mon-itoring feature set. Subsequently, FNDNLP learns a supervised low-dimensional projection from the training data to reduce redundant and coupled information while improving intraclass compactness and interclass separability. Finally, a temporal con-volutional network with multi-head attention (TCN-MHA) mod-els dependencies in the projected representation. The framework is evaluated on the CWRU and HUST bearing platforms using complete-pipeline comparisons, controlled classifier evaluations, reduced-dimension sensitivity analysis, and component-wise ab-lation. It achieves mean test accuracies of 98.89 ± 0.03% and 98.10 ± 0.42% on CWRU and HUST, respectively. The results support the effectiveness of the integrated framework under the evaluated datasets and conditions.