Rotating machinery is crucial in industrial production, and its stable operation directly affects both efficiency and safety. However, diverse operating conditions, non-stationary signals, and complex environments pose significant challenges for fault diagnosis. To address these issues, we propose a multi-granularity information fusion network to improve fault classification in complex scenarios. FusionNet first applies Continuous Wavelet Transform to decompose time-series signals, enabling in-depth extraction of key time-frequency features. Next, the multi-granularity feature extraction module focuses on multi-scale and sparse features, dynamically capturing local variations in fault information to enhance and supplement the overall sequence while effectively suppressing noise interference. After that, the Swin-Transformer with window-based interaction adaptively fuses these multi-granularity features, progressively integrating both local details and global dependencies. Finally, the Multi-Granularity Feature Aggregation Module combines features from different scales layer-by-layer and incorporates channel attention to highlight critical fault patterns. Experiments conducted with two datasets, the benchmark CWRU, and a self-developed dataset, show that FusionNet achieves state-of-the-art performance for fault diagnosis. In CWRU, it achieves an average accuracy of 99.67% in four operating conditions. In the self-developed Paper Delivery Structure Coupling dataset, it reaches 99.44%, outperforming conventional models like LSTM by 30.90%.

The coupling torque signal contains essential information about the operating condition of the motor‐follower mechanical system. Artificial Intelligence methods have been effective in diagnosing coupling faults. However, due to internal and external excitations caused by the operating environment and neighboring components, existing coupling fault diagnosis models often struggle with poor feature extraction, low recognition accuracy, and weak generalization performance. To overcome these limitations, a multihead self‐attention mechanism‐enhanced empirical mode decomposition (EEMD)–convolutional neural network (CNN)–bidirectional long short‐term memory (BiLSTM) model is proposed. Empirical mode decomposition (EMD) is first applied to extract spatial features from the data. Next, the multihead self‐attention mechanism captures essential internal characteristics. CNN is then used for further spatial feature extraction, and BiLSTM is employed to extract temporal features, enabling effective spatiotemporal feature fusion. Torque and vibration signals of three typical coupling faults—loosening, rough contact, and misalignment—are collected using a motor‐coupling‐paper handling mechanism testbed and compared with traditional models, including long short‐term memory (LSTM), CNN, and Random Forest. The proposed method shows an average F 1‐Score improvement of 31.43%, 40.07%, and 31.71%, respectively, indicating superior noise robustness and better generalization. Furthermore, the dimensionality reduction results of each network layer are visualized using the T‐stochastic neighbor embedding (T‐SNE) method, which reveals clear feature patterns and confirms the model’s reliability and effectiveness.