Rolling bearing life prediction based on a dual-branch network with parallel and cross-modal feature fusion

To tackle the poor degradation feature extraction and high life prediction errors in rolling bearings, the method based on dual-branch network parallel and cross-modal feature fusion is proposed. First, the raw signal is denoised using an extended Kalman filter (EKF). Time-domain (TD) and frequency-domain (FD) features are extracted from the denoised signal, and sensitive features are selected based on monotonicity. Simultaneously, continuous wavelet transform (CWT) is applied to convert the denoised signal into time-frequency (TF) maps. To capture both long-term dependencies and spatial dimensional information, a dual-branch parallel network integrating a temporal convolutional network (TCN) and EfficientNetV2 is designed. The TCN branch processes temporal dimensions and the EfficientNetV2 branch handles spatial features. Furthermore, a cross-attention (CA) mechanism is introduced to effectively integrate spatiotemporal features and reduce information loss or conflict. Finally, bi-directional gated recurrent unit (BiGRU) is used to predict the bearing life. The validation is conducted using the PHM 2012 and Xi’an Jiaotong University (XJTU) bearing datasets. In comparison with different methods, the mean squared error (MSE) and mean absolute error (MAE) of this method are reduced by 16.9% and 8.3%, respectively. The results confirm the superior predictive accuracy and stability of the proposed method across multiple datasets.