Rolling Bearing Fault Diagnosis via Meta-BOHB Optimized CNN-Transformer Model and Time-Frequency Domain Analysis

Bearing fault diagnosis encounters limitations including insufficient accuracy, elevated model complexity, and demanding hyperparameter optimization. This research introduces a diagnostic framework combining variational mode decomposition (VMD) and fast Fourier transform (FFT) for extracting comprehensive temporal-spectral characteristics from vibration data. The methodology employs a hybrid deep learning architecture integrating convolutional neural networks (CNNs) with Transformers, where CNNs identify local features while Transformers capture extended dependencies. Meta-learning-enhanced Bayesian optimization and HyperBand (Meta-BOHB) is utilized for efficient hyperparameter selection. Evaluation on the Case Western Reserve University (CWRU) dataset using 5-fold cross-validation demonstrates a mean classification accuracy of 99.91% with exceptional stability (±0.08%). Comparative analysis reveals superior performance regarding precision, convergence rate, and loss metrics compared to existing approaches. Cross-dataset validation using Mechanical Fault Prevention Technology (MFPT) and Paderborn University (PU) datasets confirms robust generalization capabilities, achieving 100% and 98.75% accuracy within 5 and 7 iterations, respectively. Ablation studies validate the contribution of each component. Results demonstrate consistent performance across diverse experimental conditions, indicating significant potential for enhancing reliability and reducing operational costs in industrial fault diagnosis applications. The proposed method effectively addresses key challenges in bearing fault detection through advanced signal processing and optimized deep learning techniques.