A fault diagnosis network that integrates physical feature enhancement and attention mechanisms-PFA-Net is proposed, providing a structurally clear and high-performance solution for intelligent fault diagnosis.
Abstract
As critical components in rotating machinery, the operational status of rolling bearings directly affects the safety and reliability of equipment. However, in real-world industrial environments, factors such as strong background noise, multi-source coupled vibrations, and complex transmission paths often mask the early failure characteristics of bearings, making it difficult for traditional methods to achieve accurate diagnosis. While existing signal-processing-based diagnostic methods possess clear physical significance, they exhibit poor adaptability under non-stationary, high-noise conditions; conversely, purely data-driven deep learning methods, despite their powerful feature extraction capabilities, face issues such as poor interpretability, heavy reliance on labeled data, and limited generalization ability. To address this, this paper proposes a fault diagnosis network that integrates physical feature enhancement and attention mechanisms-PFA-Net. This network consists of a cascaded Feature Decoupling Enhancement Module (FDEM) and a Feature Aggregation Module (FAM). The FDEM comprises a Quadratic Convolution Feature Enhancement Submodule (QCFEM) and a Physical Feature Decoupling Submodule (PFDM). the former enhances higher-order nonlinear features in time-frequency images through learnable quadratic convolution operations, while the latter introduces physical operators such as the Fourier transform and the Hilbert transform to decouple fault pulses and envelope information from the enhanced features into interpretable physical metrics, thereby effectively highlighting fault modes in noisy environments. FAM, on the other hand, integrates an improved residual block with a Swin Transformer. It enhances local key features through a convolutional attention module (CBAM) and captures global dependencies using a shift-window self-attention mechanism, achieving efficient synergy between local and global features. Through optimization via a dynamic joint loss function, PFA-Net significantly improves fault diagnosis accuracy in high-noise environments while ensuring physical interpretability. Experimental results demonstrate that this method exhibits excellent robustness and generalization capabilities under complex operating conditions, providing a structurally clear and high-performance solution for intelligent fault diagnosis.
A novel intelligent fault diagnosis framework, termed PGDS-CLNet, is proposed in this study, which is an end-to-end trainable diagnostic network after standard signal normalization and segmentation.
Fanlong Zhu, Junyu Lai, Peiwen Lu et al.· Advances in Mechanical Engin...· 0 citations
Traditional bearing fault diagnosis deep learning models rely on sufficient annotated data and powerful computing hardware, exhibiting poor anti-noise performance and weak generalization under small sample and mixed noise industrial conditions. Moreover, purely data-driven models fail to incorporate bearing contact mechanical laws, leading to low physical interpretability. This work develops a CAMHP framework coupling Hertz contact-based physics-informed neural networks (HPINNs) and physical-guided Comba attention mechanism (CAM). The VMD-HHT-RFE-MI preprocessing pipeline is utilized to purify noisy vibration signals. The main innovations of CAMHP are:(1) Embedding Hertz contact equations into PINN training and makes key mechanical parameters learnable to guarantee physically consistent diagnosis results.(2) Designing a lightweight CAM with dual feedback and block parallelism, realizing O(n) low-complexity temporal feature extraction guided by real-time contact state parameters. Validated on CWRU and JNU bearing datasets, CAMHP reaches 99.89% and 99.73% average diagnosis accuracy respectively. In mixed noise and small sample test scenarios, the proposed method consistently surpasses comparative algorithms with fast inference speed, demonstrating its robustness and interpretability for real-world rotating machinery fault monitoring.
A robust and noise-resilient bearing fault diagnosis framework that integrates advanced signal processing with hybrid deep learning techniques is presented, demonstrating strong robustness and generalization capability.
Sujit Kumar, Manish Kumar, Bam Bahadur Sinha· International Journal of Dyn...· 0 citations
Fault diagnosis of rolling bearings is crucial for operational safety. However, the scarcity of labeled data and significant domain shifts are two key challenges. Existing studies neglect robustness to physical disturbances and the interpretability of diagnostic decisions. To address these issues, this paper proposes a physics-informed cross-domain fault diagnosis framework. First, based on bearing fault modulation mechanisms, multi-domain features are extracted. A physics-regularized feature selection strategy combining Sparse Group Lasso with stability selection and Random Forest is used to retain fault-relevant features. Recognizing that velocity variations are a major source of domain shifts, order analysis is introduced as a physics-prior alignment method. Deep CORAL and pseudo-label self-training are combined for unsupervised knowledge transfer. Finally, multi-level interpretability analysis is embedded to quantify domain shift, track adaptation geometry and explain final decisions. Two case studies demonstrate its robustness, trustworthiness and engineering applicability.
Yi Xie, Ruyang Zheng, Xuepeng Guo et al.· Eksploatacja I Niezawodnosc-...· 0 citations
Rolling bearings are critical components in rotating machinery, and their failures may cause unexpected downtime and safety risks. However, conventional deep diagnostic models are often difficult to deploy on resource-constrained edge devices because of their high computational cost and memory consumption. This paper proposes a lightweight time–frequency attention network (LTFANet) for multi-fault diagnosis of rolling bearings on an edge platform. The proposed model directly processes one-dimensional vibration signals and employs multi-scale depthwise separable convolutions to capture impact and periodic fault features with low computational complexity. A lightweight frequency branch is introduced to enhance fault-frequency representation, while an efficient channel attention module adaptively emphasizes fault-sensitive features. Moreover, a severity-aware multi-task extension is introduced to jointly identify the fault location and degradation level. To further improve edge inference efficiency, knowledge distillation, structured pruning, and TensorRT-based acceleration are integrated into the deployment pipeline. Experiments on CWRU-10 and Paderborn achieve 97.20% and 90.25% accuracy, respectively, while LTFANet contains only 0.020 M parameters and requires 0.610 M FLOPs. Knowledge distillation increases the CWRU-10 accuracy to 98.50%, and the severity-aware extension achieves 95.18% severity accuracy. On the NVIDIA Jetson Nano, the pruned TensorRT FP16 implementation achieves an average inference latency of 0.520 ms and a throughput of 1923.08 samples/s. The framework provides an effective solution for real-time and low-cost bearing condition monitoring at the edge.
In this study, a hybrid deep learning architecture is proposed for robust vibration-based fault diagnosis in industrial machinery by jointly modeling time-domain, frequency-domain, and temporal dynamics. In the time domain, convolutional layers combined with an Enhanced Gated Attention (EGA) mechanism emphasize informative signal components while suppressing noise. Temporal evolution is modeled using Neural Ordinary Differential Equations (Neural ODEs), enabling smooth and stable continuous-time feature representations. In parallel, a Fourier Neural Operator (FNO) extracts frequency-domain characteristics, augmented with gated attention to focus on fault-related spectral patterns. Long Short-Term Memory (LSTM) layers capture long-range dependencies, while Squeeze-and-Excitation (SE) blocks adaptively recalibrate channel-wise feature responses. A multi-scale attention-based fusion module integrates domain-specific representations and auxiliary features to enhance discrimination under varying operating conditions. The proposed model is evaluated on the SUBFv1.0 dataset through extensive ablation studies and experiments under multiple noise levels, achieving 98.41% accuracy in noise-free conditions and maintaining performance above 91% even at 5 dB SNR. Unlike existing multi-path approaches that combine heterogeneous features in a loosely coupled or discrete manner, the proposed architecture uniquely integrates multi-domain feature learning with bidirectional attention mechanisms and continuous-time temporal dynamics, enabling coherent cross-domain interaction and robust fault characterization.
Canan Taştimur· Information Technology and C...· 0 citations