An intelligent serial fault diagnosis method integrating optimized VMD and joint time-frequency features for POD thrust bearings

Thrust bearings are critical load-bearing components in podded propulsion (POD) systems, and their operational health directly governs vessel safety. However, existing vibration-based fault diagnosis methods suffer from parameter sensitivity, inconsistent feature extraction, and limited adaptability to complex operating conditions. This paper proposes a novel three-stage serial intelligent fault diagnosis framework termed OVMD-TDF-XGBoost (OTX) to address these challenges. In the first stage, an Optimal Variational Mode Decomposition (OVMD) method is developed, which introduces a Pearson correlation coefficient-based criterion for adaptive mode number determination and employs the Peacock Optimization Algorithm (POA) to jointly optimize the penalty factor and fidelity coefficient, thereby eliminating the reliance on manual parameter tuning inherent in conventional VMD. In the second stage, a standardized joint Time-Frequency Domain Feature (TDF) extraction strategy is established, constructing a 20-dimensional feature matrix comprising 10 time-domain statistics and 10 frequency-domain parameters from the optimally decomposed modal components, which provides a unified and robust feature representation across varying fault patterns. In the third stage, XGBoost is employed as the classifier, leveraging its second-order Taylor expansion-based optimization and built-in regularization to achieve efficient and accurate multi-class fault recognition. Comprehensive experiments are conducted on a purpose-built POD thrust bearing test rig under 14 fault conditions spanning two load levels, and the results demonstrate that the OTX method achieves an average diagnostic accuracy of 96.57%, outperforming EEMD-PNN, DWT-RF, and STFT-SVM by approximately 10–14 percentage points. Cross-validation on the public Case Western Reserve University (CWRU) bearing dataset further confirms the generalization capability of the proposed method, achieving an overall accuracy of 99.69%, with per-category accuracy exceeding 98% across all fault categories and reaching 100% in seven out of ten categories. The proposed OTX framework offers a systematic and adaptive solution for bearing fault diagnosis in marine propulsion systems.