JQJunyu Qi

Junyu Qi

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Junyu Qi has a total of 66 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Top Cited Papers

    Knowledge Library Network for Non-Exemplar Incremental Remaining Useful Life Prediction

    by Jianghong Zhou, Jun Luo, Junyu Qi, Yi Qin

    Aiming at the difference of various incremental remaining useful life (RUL) prediction tasks, a novel nonexemplar incremental learning (IL) method is proposed, named knowledge library network (KLN). First, a multilevel focal self-attention (MLFSA) is developed to model the multiscale temporal dependencies and extract the key features. On this basis, a dual domain fusion temporal convolutional network (DDFTCN) is designed to mine the degradation information from both time-domain and frequency-domain data. Next, a task similarity judgment strategy specifically tailored for IL scenarios is proposed to assess the discrepancies between different tasks. Based on the task differences, a network augmentation rule is proposed for implementing the feature extraction network augmentation or task branch network augmentation. The feature extraction network augmentation is employed for the tasks with significant differences, while the task branch network augmentation is utilized for the tasks with small differences. Especially for the task branch network augmentation, a retention and absorption distillation strategy is proposed to enhance the model's capability to retain previous knowledge while adapting to new task. The network augmentation network rule is integrated into DDFTCN to form the KLN. Finally, the experimental results on the life-cycle bearing and gear datasets demonstrate that DDFTCN achieve an average accuracy of 93.6% across both datasets, surpassing the existing prediction methods. The experimental results on the incremental RUL prediction tasks validate the superiority of KLN over the typical and state-of-the-art IL methods.


  • Helical Guided Wave Mode Decomposition With Local Peak Constraints for Aluminum Cable Sheath Health Monitoring

    by Bin Zhang, Zhuyun Chen, Jingyan Xia, Junyu Qi, Weiguo Chen

    Corrosion detection of high-voltage cable aluminum sheaths is of paramount importance as power systems continue to evolve. Ultrasonic guided waves technology, renowned for its high sensitivity, robust penetration capabilities, and comprehensive damage condition feedback, has been widely utilized in non-destructive pipeline testing. However, traditional axial and circumferential guided wave detection techniques fall short in addressing the complex structure of cable aluminum sheath. To tackle this challenge, this paper introduces a mode identification and separation method for helical ultrasonic guided wave. Firstly, the mechanism of the propagation of helical ultrasonic guided waves in cable aluminum sheaths is thoroughly investigated, and the feasibility of this approach is validated through simulation. Following this, a full-modal peak-frequency modulation dictionary for helical ultrasonic guided waves is constructed, grounded in the modal characteristics of these waves within the cable aluminum sheath. Lastly, a local peak constrained-based orthogonal matching pursuit method is developed to effectively extract and identify pertinent modes. Experimental results demonstrate that the proposed method can accurately extract the damage modes with an impressive modal identification accuracy of 99.48 ± 0.041%. This performance significantly outperforms state-of-the-art methods, highlighting its superiority in modal analysis.

  • Speed-Robust Feature Learning and Minor-Class Generation for Imbalanced Variable-Speed Bearing Fault Diagnosis

    by Lijuan Zhao, Haidong Shao, Junyu Qi, Yi Qin

    Variable-speed operation and class imbalance are two prevalent challenges in rolling bearing fault diagnosis; However, existing methods seldom tackle both simultaneously, which limits their diagnostic effectiveness. To address the above issues, this article proposes a variational information bottleneck-based speed-robust feature learning and minor-class feature generation (VIB-FGNet) for imbalanced fault diagnosis under variable-speed conditions. The proposed method extracts speed-invariant fault representations in the latent space through the designed multiple constraints, thereby effectively suppressing condition-related disturbances. On this basis, a class-level feature generation strategy is applied in the shared cross-speed latent space, dynamically augmenting minority-class fault features via reparameterized sampling to effectively alleviate sample imbalance. Experimental results demonstrate that across different variable-speed and sample-ratio settings, the proposed method exhibits significant advantages over state-of-the-art approaches, validating its effectiveness for variable-speed and imbalanced fault diagnosis tasks.


  • PhyMSNet: A Physics-Informed Multi-Scale Network for Robust and Interpretable Diagnosis of Bearings Under Variable Conditions

    by Guanze Li, Hongqi Lin, Zhuyun Chen, Junyu Qi, Weihua Li, Qiang Liu

    In engineering practice, rolling element bearings operate under varying conditions, generating nonstationary vibration signals with complex physical characteristics. Although deep learning methods using multi-head self-attention and multi-scale convolution improve feature representation, they often increase computational burden and offer limited physical interpretability, constraining diagnostic reliability in applications. To tackle this issue, PhyMSNet, a robust and interpretable physics-informed multi-scale network, is proposed. The core of the framework is the integration of physical priors with a multi-scale learning architecture. First, raw vibration signals are processed through cyclic spectrum analysis to extract modulation features intrinsically related to fault characteristics. These physically grounded features are then fed into a multi-scale convolutional network designed to capture fault signatures across different temporal resolutions. To further align model focus with underlying failure mechanisms, a Physics-informed Squeeze-and-Excitation Module (PhySE) is embedded within the network. This module leverages prior knowledge of system resonance and transient impulse patterns to adaptively recalibrate feature importance. Moreover, for enhanced transparency, the model decision logic is quantitatively explained using Shapley additive explanations (SHAP), identifying contributions of spectral and temporal features to the final diagnosis. Comprehensive experiments on a CNC fault dataset with varying speeds and loads demonstrate that PhyMSNet achieves superior diagnostic accuracy, stronger robustness to condition changes, and clearer interpretability compared to state-of-the-art deep learning methods.


  • Detail Feature Calibration Bilateral Segmentation Network and its Application in Gear Pitting Measurement

    by Yi Qin, Changyan Pei, Daiping Song, Junyu Qi

    Accurate and real-time detection of gear pitting is crucial for industrial equipment health monitoring.However, existing segmentation methods encounter a fundamental trade-off between computational efficiency and segmentation precision. Heavyweight models achieve high accuracy but are impractical for real-time use, whereas lightweight architectures often fail to capture fine defect details. To address these challenges, we propose a vision-based measurement framework, termed detail feature calibration bilateral segmentation network (DFCBNet), a lightweight segmentation network optimized for real-time gear inspection. The architecture incorporates three key components: (1) a detail feature calibration module, which enhances micro-defect sensitivity by calibrating multiscale semantic and spatial features; (2) a multiscale spatial detail attention module, which improves feature extraction for tiny pitting regions using parallel atrous convolutions and attention mechanisms; and (3) an adaptive feature fusion module, which adaptively compensates for detail loss during the decoding phase. Experimental results demonstrate that DFCBNet achieves 78.7% mean intersection over union (mIoU) on Cityscapes and 76.1% mIoU on Cambridge-driving Labeled Video Database, with real-time performance of 118 frames per second and only 2.38 M parameters. Furthermore, the proposed method demonstrates strong performance in practical gear-pitting measurement tasks, highlighting its potential for deployment in industrial online monitoring applications.

  • Semi-Supervised Transfer Graph Representation Learning with Few-Shot Adaptation for gearbox diagnostics under extraneous transient noise

    by Peng Chen, Jia Gao, Yuhao Wu, Changbo He, Ge Xin, Shuai Fan, Junyu Qi

    Gearboxes are critical mechanical components widely deployed in industrial applications, where their reliable operation directly impacts system safety and efficiency. However, conventional fault diagnostic approaches face significant challenges when operating under extraneous transient noise conditions, particularly with limited labeled fault samples. These challenges manifest as performance degradation with extremely sparse labeled datasets, vulnerability in pseudo-label generation mechanisms under intense transient noise, and inconsistent feature scale representations due to noise-induced interference. Furthermore, existing methods struggle to maintain diagnostic accuracy when confronted with both data scarcity and transient disturbances, often resulting in compromised model generalization and unreliable fault classification. To address these limitations, this research proposes the Semi-Supervised Transfer Graph Representation Learning with Few-Shot Adaptation (SSTGRL-FSA) framework, featuring three innovative components: a novel pseudo-label reliability enhancement mechanism leveraging systematic knowledge transfer from established source domains, an advanced label transmission and matching strategy exploiting homologous signal patterns across operational domains, and an integrated first-order Markov state probability transition matrix with amplitude-constrained scaling. SSTGRL-FSA significantly advances the field by effectively handling both labeled data scarcity and transient noise interference while enhancing model robustness through sophisticated temporal dependency modeling and stable feature scale maintenance, ultimately providing a more reliable and practical solution for industrial gearbox fault diagnosis under challenging operational conditions.