Lingli Cui

Rolling bearing fault diagnosis is confronted with significant technical challenges, including the difficult extraction of weak fault features and the insufficient fusion of one-dimensional (1D) temporal dynamics with two-dimensional time-frequency patterns. To address this issue, this article proposes a novel dual-branch fusion network that synergistically integrates a ResNet optimized with frequency dynamic convolutional attention with a gated recurrent unit for accurate bearing fault diagnosis, termed FDCA-GRU. Firstly, a parallel dual-branch architecture is constructed. The image branch encodes the 1D signal into a Markov transition field image, which is then fed into an optimized FDCA-ResNet backbone. Innovatively, frequency dynamic convolution (FDConv) is introduced in this branch to replace standard convolution, enabling it to adaptively learn filter characteristics in the frequency domain. This significantly enhances the perception of specific frequency fault patterns. Meanwhile, the convolutional block attention module (CBAM) is integrated into the residual blocks. Through its channel and spatial attention mechanisms, features are adaptively weighted to enhance the model’s response to discriminative feature regions. On the other hand, the temporal branch utilizes a GRU network to directly model the 1D stacked sequences, thereby efficiently capturing long-term dynamic temporal dependencies and evolutionary laws. The two branches are complementary in that the image branch emphasizes state-transition textures and frequency-related spatial patterns, whereas the temporal branch preserves amplitude evolution, transient impact order, and long-range sequential dependencies. Finally, the deep features extracted by the two branches are fused via cross-modal concatenation to achieve synergistic optimization of time-frequency spatial patterns and dynamic temporal information. Test results on the Beijing Jiaotong University Rail Autonomous Operation bogie datasets (BJTU-RAO) and HUST public datasets show that the proposed method achieves average diagnostic accuracies of 99.52 and 98.61%, respectively. Furthermore, ablation experiments systematically validate the synergistic effects and respective contributions of the dual-branch fusion architecture, FDConv, and CBAM, confirming the effectiveness of the proposed framework.

In aerospace bearing fault diagnosis under complex operating conditions, extracting non-stationary signal features and ensuring model generalization remain significant challenges. To address these issues, this study introduces WCSKDSC, a multi-scale feature fusion model that integrates physics-informed guidance with data-driven learning. Initially,the wavelet transform convolution (WTConv) module processes time-frequency images generated by Gramian Angular Difference Fields (GADF) through Daubechies wavelet-based multi-scale decomposition and residual fusion. Next, the Swin Kansformer module replaces the traditional MLP structure in feedforward networks with KAN, utilizing learnable spline basis functions and a dynamic grid adjustment strategy to transcend the constraints of static activation functions in mapping complex time-frequency patterns. Furthermore, the depth-wise separable convolution (DSC) module amplifies local time-frequency edge responses through group convolution and fosters cross-channel interaction via pointwise convolution. Experiments validate that the proposed method achieves an average accuracy of 99.85\% and 99.72\% in the four-classification and seven-classification tasks across two aerospace bearing datasets. When compared to existing approaches, the proposed method demonstrates stable recognition accuracy and robust generalization capabilities.

Traditional deep learning networks in fault diagnosis tasks struggle to effectively utilize unlabeled data, and their diagnostic performance is constrained by the limited availability of labeled samples. To address this issue, this paper proposes a semi-supervised method based on multi-modal time-frequency fusion and an improved Swin Transformer, termed the Contrastive Multi-modal Time-Frequency Swin Transformer for Semi-Supervised Fault Diagnosis (CMTFST-SFD). First, a multi-scale shift window attention fusion module is designed as the backbone network, performing parallel computations with a 7 × 7 fine-grained window and a 14 × 14 coarse-grained window. Subsequently, a three-channel time-frequency encoder and a multi-frequency fusion convolution module are constructed to extract diverse frequency characteristics. Finally, by integrating a joint contrastive-cross-entropy loss function, the network backbone is trained in the first stage using a combination of unlabeled and labeled data, followed by training the classification head solely with labeled data in the second stage, thereby fully leveraging all available data for comprehensive model training. The proposed method is evaluated on two bearing fault datasets under three different labeling rates, consistently achieving a recognition rate exceeding 97%. Compared to other advanced techniques, the proposed method demonstrates superior recognition rates and enhanced clustering performance.

The “black-box” nature of deep diagnostic models fundamentally undermines trust in their deployment for high-stakes scenarios. While post-hoc explanation methods are mainstream techniques for revealing decision-making logic, they primarily produce visual interpretations (e.g., heatmaps) without yielding tangible performance improvements in the models themselves. This paper proposes a post-hoc explanation augmented deep learning (PHExp-DL) framework that closes the loop between interpretability and performance enhancement in fault diagnosis. Capitalizing on the intrinsic cyclostationarity of rotating machinery vibrations, we introduce a cyclic orthogonal convolution (CyclicOrthoConv), a novel convolution operator that could embed fault-induced cyclostationarity priors. Crucially, a dynamic gradient feedback scheme propagates gradient-weighted class activation mapping (Grad-CAM)-derived feature contributions backward, steering convolutional neural network (CNN) feature learning toward discriminative fault signatures. This creates a self-correcting learning cycle where explanation-guided feedback continuously refines feature extraction. Experimental results confirm the remarkable robustness of the proposed method under unseen noise conditions. It outperforms the baseline CNN by margins of 21.66% and 18.86% in average accuracy for the two Cases, respectively, and also exceeds the performance of other state-of-the-art methods.

The poor generalization capability of deep diagnostic models remains the primary impediment to their deployment in safety-critical industrial systems. This article proposes a novel re-masked autoencoder knowledge distillation (Re-masked AE-KD) method for robust machinery fault diagnosis, i.e., the models are subjected to testing using data from unseen operating conditions. Specifically, we develop a self-supervised re-masked convolutional autoencoder (RCAE) that employs dual-stage masking, applied to both input features and latent representations, to enhance learning of compressed yet distortion-invariant embeddings. Furthermore, we construct a physically informed multiscale filter (PIMSF) that encodes prior knowledge of vibration signal amplitude modulation characteristics for precise fault extraction while integrating temporal-context encoding to capture time-varying dependencies. Finally, we propose a KD method, in which a teacher model processes physically meaningful multidomain features. Crucially, the teacher and the student models interact within a shared latent space and a bidirectional feature reconstruction process, enforcing consistency to preserve critical diagnostic information. Extensive comparative experiments demonstrate superior performance over several state-of-the-art methods under unseen operational conditions.