A self-adaptive deep learning framework for noise-robust fault diagnosis in high-speed rotating machinery with automated bayesian optimization

Abstract Data-driven fault diagnosis for rotating machinery faces challenges in deep feature extraction, noise interference, and hyperparameter sensitivity. To address these issues, we propose a novel deep learning framework integrating three key innovations: (1) A multi-channel parallel network combining CNN and TCN to extract spatiotemporal features via dense connections, preventing gradient vanishing; (2) An adaptive soft-threshold denoising module driven by self-attention mechanisms, dynamically suppressing noise while preserving critical fault signatures; (3) A Bayesian hyperparameter optimizer automating model configuration, significantly reducing manual tuning efforts. Additionally, a penalty-enhanced loss function is designed to improve classification of hard-to-distinguish faults. Extensive experiments on two bearing datasets (CWRU and WHUT) demonstrate superior performance: 98.84% accuracy on CWRU and 99.89% on WHUT, outperforming ResNet, DenseNet, and other benchmarks by 5%–30%. Especially, it outperforms the benchmark by 3%–50% in noise conditions. The framework shows strong potential for industrial predictive maintenance under noisy environments.