.png){kind=icon}
.png){kind=icon}
Return
AWT-DDFN: an end-to-end adaptive wavelet transform-dual domain fusion network for bearing fault diagnosis
.png){kind=icon}
PRE
AI
Z
C
M
DOI：10.1016/j.ymssp.2026.114827.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep learning has been widely used in intelligent fault diagnosis of rolling bearings. Existing fault diagnosis methods usually use multi-scale wavelet transforms to convert raw 1D vibration signals into time–frequency representations as model inputs. However, such preprocessing brings high computational cost, inevitable information loss and tedious manual parameter tuning, which hinders real end-to-end diagnosis. To overcome these drawbacks, we propose an adaptive wavelet transform–dual domain fusion network (AWT-DDFN), a complete end-to-end framework that directly processes raw vibration signals without manual preprocessing. AWT-DDFN adopts a learnable wavelet convolution layer to replace fixed wavelet transforms, realizing joint optimization of feature extraction and fault diagnosis. Meanwhile, the network uses a dual-branch structure to capture complementary discriminative features: a frequency-domain branch based on pointwise convolution and a time-domain branch combining grouped convolution and dilated convolution. This structure effectively reduces model complexity and parameters, making the model lightweight and robust to noise and trivial signal variations. The features of the two branches are fused by element-wise addition before being sent to the classifier for fault identification, forming a seamless end-to-end diagnosis pipeline. Experiments on PU and CWRU datasets show that AWT-DDFN achieves superior diagnosis performance and stronger noise robustness than state-of-the-art methods, with fewer parameters and lower computation cost. By discarding handcrafted preprocessing, the proposed framework provides a lightweight, end-to-end and deployable intelligent diagnosis solution with strong practical value for industrial scenarios under complex working conditions and strong noise.
Keywords:
Bearing fault diagnosis
Dual domain fusion
End-to-end
Lightweight
Learnable wavelet convolution
Journal
IF:
8.9
Papers:
1.2W
Citations:
6.6W
Organization
No organization information available
Cited Papers