13 August 2024 in Measurement Science and Technology
In the field of data-driven fault diagnosis (FD), deep learning methods have proven their excellent performance, especially when dealing with complex signals from rotating equipment such as bearings. However, fault features in vibration signals are often mixed with noise features and distributed at different frequency scales, posing challenges for effective feature extraction. In order to solve this problem, this paper proposes a high frequency-multiscale cascade network (HF-MSCN), which enhances the noise suppression and feature learning capability of the model by combining a high-frequency convolutional block (HFCB) with a multi-scale cascade block (MSCB). HFCB effectively suppresses high-frequency noise through wide convolutional layers and self-attention mechanisms while still retaining essential high-frequency fault signals. MSCB enhances the interaction between convolutional layers at different scales by cascading the layers at different scales and strengthens the model’s ability to capture subtle fault features, especially when processing periodic fault pulse signals. Finally, we investigate the internal functioning of the network using time–frequency analysis methods in signal processing to improve the interpretability of deep learning methods in FD applications and further verify the enhanced effect of HFCB and MSCB on feature extraction. We validate the effectiveness of HF-MSCN on the case western reserve university dataset as well as a self-constructed bearing composite fault dataset, and the experimental results demonstrate that the network exceeds the performance of six state-of-the-art fault diagnostic methods in high-noise environments.
1 January 2026 in IEEE Transactions on Automation Science and Engineering
11 August 2026 in Structural Health Monitoring
Extracting weak fault features obscured by environmental noise is a critical challenge. However, existing fault feature enhancement methods ignore the low-rank distribution of feature correlation structures, making it difficult to effectively extract critical fault information. To address this problem, a frequency-domain matrix rank transformation and time-domain spatial graph attention enhancement network (FMRT-TSGAEN) is proposed. First, an interpretable frequency-domain feature enhancement mechanism is designed, which establishes a transformation theory from low-rank to high-rank matrices by controlling the spectral condition number of the frequency-domain similarity matrix. Then, a dynamic graph structure is designed to capture the spatial dependencies of features across time scales. Meanwhile, a global context gating mechanism is constructed to calibrate global spatial information, thereby achieving global spatial fusion of cross-scale features. Finally, the time-frequency feature learning module is constructed through the residual structure, enabling the model to extract more comprehensive time-frequency information. The experimental results show that the accuracy of FMRT-TSGAEN on two simulated fault datasets reaches 95.63 and 93.83%, respectively. On the real industrial spinning machine bearing dataset, the model achieved excellent results under conditions of no noise, pink noise, Gaussian white noise, and impulse noise, with an average accuracy of 90.39%. Moreover, we prove and visualize changes in the rank of the frequency-domain correlation matrix, which not only reveal the physical principles behind frequency-domain feature enhancement but also effectively improve the model’s interpretability.
2 September 2025 in Journal of the Brazilian Society of Mechanical Sciences and Engineering