A noise-aware three-view time-frequency representation and deep–shallow feature fusion framework for bearing fault diagnosis

Abstract

Rolling bearing fault diagnosis under strong noise remains a challenging problem because defect-induced vibration responses are typically non-stationary, weakly impulsive, and easily submerged by background interference. In addition, fault categories with different damage severities often exhibit similar dominant periodic structures, which further increases the difficulty of fine-grained discrimination. To address these issues, this paper proposes a noise-aware bearing fault diagnosis framework based on three-view time–frequency representation and deep–shallow feature fusion. First, long vibration sequences are segmented by overlapping sliding windows to generate fixed-length samples. To construct a controlled noisy benchmark, random Gaussian noise and sparse impulsive interference are injected at the sample level according to the specified stochastic protocol. Second, three complementary continuous wavelet transform views are constructed from the raw noisy signal, the denoised signal, and the envelope signal, respectively. In this way, raw observation texture, denoised impulsive structures, and modulation-related information are jointly encoded into a unified three-channel time–frequency tensor. Third, an enhanced multi-branch convolutional neural network is designed to extract robust deep representations from the three-view input. The network integrates multi-scale receptive fields, adaptive soft-threshold shrinkage, and channel-energy attention to suppress redundant responses and emphasize discriminative fault-related structures. Fourth, to preserve physical interpretability and complement the deep feature space, 16 handcrafted statistical descriptors are extracted from the noisy samples, ranked by ReliefF, and the most discriminative descriptors are fused with the deep representation. Finally, a radial-basis-function support vector machine is employed to construct a stable decision boundary in the fused feature space. The evaluation is restricted to the selected ten-class Case Western Reserve University configuration and the specified synthetic-noise protocol. The framework demonstrates robust benchmark performance under these controlled conditions, whereas cross-machine, cross-dataset, and measured-industrial-noise generalization remain to be part of future work.