A Transfer Learning-Based Multimodal Feature Fusion Model for Bearing Fault Diagnosis

Abstract

Fault diagnosis based on single-modal features struggles to capture the coupling relationship between multiple fault factors, resulting in inferior diagnosis accuracy. To address this problem, a transfer learning-based multimodal feature fusion (TL-MMFF) model is proposed for fault diagnosis. First, a continuous wavelet transform (CWT)-based modal expression method is employed to transform raw vibration signals into time-frequency representations. Then, this high-resolution time-frequency modal can be utilized to capture transient vibration and energy changes in nonstationary signals. Second, a multimodal feature fusion strategy is proposed, which designs learnable parameters to dynamically weight the time-domain features of torque and the time-frequency features of vibration signals. This adaptive weighting strategy optimizes the fusion process based on the correlation of different modal feature sets, thereby enhancing the ability to describe fault characteristics. Third, a maximum mean discrepancy (MMD)-based transfer learning (TL) algorithm is designed to reduce the distribution differences between fused features under different operating conditions. Then, the model can identify fault characteristics across varying operating conditions. Finally, experiments on the Paderborn University dataset demonstrate that TL-MMFF achieves 99.1% accuracy and converges 30% faster than single-modal methods. These results validate the effectiveness of the model in integrating multimodal data and generalizing across domains.