DSMDTN: A Data-Selective Multiscale Dual Transfer Network for Fault Diagnosis of Key Components in Rotating Machinery

Rotating machinery often operates under varying working conditions, which poses significant challenges to achieving reliable bearing fault diagnosis using traditional deep learning-based models. To enhance the diagnostic performance for rolling bearings across diverse operational conditions and noisy environments, a data-selective multiscale dual transfer network (DSMDTN) with a data selector (DS), multiscale concatenation U-Net (MCU-Net), and dual classifier (DC) is proposed. The DS module employs a comprehensive scoring mechanism that integrates math, entropy, and anomaly scores to selectively identify high-quality source samples for model training. Meanwhile, the MCU-Net module incorporates gated convolutional (gated-conv) blocks and convolutional blocks to extract multiscale domain-invariant features and dynamically adjust feature importance. In addition, the DC module comprises separate source and target classifiers that jointly minimize distribution discrepancy and classification loss. The effectiveness of the proposed DSMDTN is validated through experiments on the public Case Western Reserve University (CWRU) dataset and the proprietary PT dataset collected from a PT500mini test bed. The experimental results demonstrate that DSMDTN achieves higher accuracy and exhibits stronger transfer capability compared to several state-of-the-art intelligent models across various transfer tasks and under different noise levels.