A New Weight-Based Dual Domain Adaptation Transfer Model for Bearing Fault Diagnosis Under Noisy and Cross-Domain Conditions

Abstract

The traditional fault diagnosis methods generally present poor diagnosis accuracy and robustness when faced with complex conditions that involve noisy interference and domain shift. Therefore, a new weight-based dual domain adaptation transfer model for bearing fault diagnosis is proposed. First, based on continuous wavelet transform, the temporal signals are transformed into time-frequency images (TFIs) for enhancing feature representation. Second, the TFIs are used as the input of the improved network which is based on dense and residual connections to complete feature extraction. Third, the proposed transfer model uses local maximum mean discrepancy (LMMD) to adjust data distribution between different working domains and batch nuclear-norm maximization (BNM) to improve the discriminability and diversity of the output matrix. Moreover, the weight controller is used to trade off the contributions of LMMD and BNM during training. Finally, the proposed frequency domain cut can be seen as a simple moving and cutting method to adjust each frequency spectrum of TFIs. In this process, the controlled weight factor is involved to further alleviate noise interference. Case studies show that the proposed model outperforms other methods and works well even in complex conditions mixed by noise and cross-domain.

Keywords

• Fault diagnosis
• dual domain adaptation
• transfer learning
• rotor bearing