Few-shot bearing fault diagnosis: a transfer learning framework via time–frequency information fusion

Abstract

To address the challenges of sparse monitoring data and insufficient training samples in single-sensor rolling-bearing fault diagnosis, this paper presents an end-to-end lightweight diagnostic framework leveraging multi-domain feature fusion and transfer learning. Recognizing that data scarcity-induced overfitting severely undermines the generalization of single-sensor systems, an overlapping sampling technique is first employed to augment the training set. To overcome the limitations of individual transforms in capturing multiscale and non-stationary signal characteristics, a parallel multi-branch convolutional neural network architecture is developed. By integrating transfer learning, this architecture maps raw 1D vibration signals into a suite of complementary 2D time–frequency representations-including short-time Fourier transform, continuous wavelet transform, Hilbert–Huang transform, and S-transform-to extract deep discriminative features independently. Furthermore, an enhanced Dempster–Shafer (D–S) evidence theory is integrated to perform decision-level fusion. To overcome the closed-world assumption inherent in conventional D–S theory, this strategy leverages information entropy to dynamically compute an adaptive confidence coefficient (Formula presented) (Formula presented), explicitly quantifying epistemic uncertainty. This mechanism aims to better exploit the complementary information from different time–frequency representations under single-sensor constraints, while preventing overconfident misclassifications of unknown faults. Experimental results on the Case Western Reserve University benchmark dataset indicate that the proposed method achieves competitive and stable performance in few-shot scenarios, consistently outperforming five state-of-the-art diagnostic models in both accuracy and generalization capability.