ZYZhaohui Yuan

Zhaohui Yuan

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Zhaohui Yuan has a total of 43 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    MTL-Frame: An End-to-End Multi-Task Learning Framework for Student Profiling and Composite-Score Estimation

    by Guifen Jiang, Muhua Tan, Zhaohui Yuan

    Composite-score estimation (or assessment auditing) and student profiling are two fundamental tasks in Educational Data Mining (EDM). However, existing studies often treat them as separate problems, typically adopting a sequential pipeline in which clustering is first performed and then used for downstream estimation. Such a fragmented paradigm limits the interaction between latent student-group structures and supervised outcome signals, particularly in low-data educational scenarios. To address this limitation, this study proposes “MTL-Frame”, an end-to-end multi-task learning framework that jointly optimizes student profiling and composite-score estimation. MTL-Frame integrates prototype-based contrastive clustering with context-aware regression to directly inject student profile priors into grade estimation. A dispersion regularization and dynamic loss weighting ensure training stability. Experiments conducted on a real-world blended English-course dataset involving 429 university students show that MTL-Frame outperforms representative single-task regression baselines, including XGBoost, Random Forest, SVR, and LSTM, achieving an RMSE of 1.7532, an MAE of 0.8168, and R2 = 0.9812. While this high R2 partly reflects the model’s ability to learn the deterministic scoring aggregation, the performance remains strong even when the final exam score is excluded from the inputs (R2 = 0.9681), confirming genuine estimation capability. Compared to the strongest single-task baseline (XGBoost), MTL-Frame reduces RMSE by 38.2%. The model also obtains a Silhouette Score of 0.3562, indicating its ability to generate meaningful student profiles while maintaining strong estimation accuracy. These results demonstrate that integrating unsupervised profiling priors into supervised estimation can improve model robustness and provide actionable insights for differentiated instructional intervention.

  • Open Access

    WAVE: Wall-Aligned Vector Embedding for Self-Supervised Learning of Electrocardiograms

    by Shurong Pan, Wenhan Liu, Qingyuan Wu, Cong Wang, Zhaohui Yuan

    Deep learning has achieved remarkable progress in electrocardiogram (ECG) analysis, but its heavy dependence on labeled data greatly increases annotation cost. This work proposes wall-aligned vector embedding (WAVE), a self-supervised learning framework that effectively extracts prior knowledge from unlabeled ECGs to reduce reliance on labels. WAVE fully leverages the diversity, synergy, and lead correlation of multi-lead ECGs by explicitly incorporating the correspondence between ECG leads and cardiac walls. Specifically, a multi-branch network captures lead-wise diversity; wall-wise synergy is modeled by concatenating leads from the same wall and projecting them via shared projection; and a dual alignment task is designed to learn correlations both within and across cardiac walls. Experimental results demonstrate that WAVE consistently surpasses all baselines under various evaluation settings, and maintains strong performance even when only a small fraction of labeled ECGs is available. Furthermore, components such as dual alignment, shared projection, wall-based concatenation, and mean target embedding are empirically verified to significantly enhance pretraining quality. In summary, WAVE learns highly informative ECG representations from unlabeled data, enabling low-cost and label-efficient ECG analysis for real-world cardiovascular diagnostics.

  • Open Access

    ACL-ECG: Anatomy-Aware Contrastive Learning for Multi-Lead Electrocardiograms

    by Wenhan Liu, Zhijing Wu, Zhaohui Yuan

    Deep learning has achieved impressive progress in automated electrocardiogram (ECG) analysis, yet its performance still relies heavily on large-scale labeled datasets. As ECG annotation requires cardiologists, this process is costly and time-consuming, limiting its scalability in clinical practice. Contrastive learning offers a promising alternative by enabling the extraction of generalizable representations from unlabeled ECG data. In this study, we propose Anatomy-Aware Contrastive Learning for ECG (ACL-ECG), a self-supervised method that incorporates cardiac anatomical relationships into contrastive learning. ACL-ECG employs a physiology-aware augmentation strategy to generate rhythm-preserving augmented views, including random scale cropping, cardiac-cycle masking, and temporal shifting. Furthermore, ECG leads are grouped into four anatomically meaningful regions—anterior, inferior, septal, and lateral—and region-level contrastive objectives are introduced to promote intra-region consistency while enhancing inter-region discriminability. Extensive evaluations of downstream tasks demonstrate that ACL-ECG consistently outperforms state-of-the-art contrastive baselines under linear probing, achieving improvements of up to 1.29% in the area under the receiver operating characteristic curve (AUROC) and 3.57% in the area under the precision–recall curve (AUPRC). Moreover, when fine-tuned using only 10% of labeled data, ACL-ECG attains a performance comparable to fully supervised training, effectively reducing annotation requirements by approximately 5∼8×. Ablation studies further confirm the contributions of both the physiology-aware augmentation strategy and the anatomy-aware contrastive objective. Overall, ACL-ECG enhances representation quality without increasing annotation burden, and provides a promising and anatomy-informed foundation for self-supervised ECG analysis in label-scarce settings.


  • SegEdit: image editing via semantic mask segmentation and shape injection within diffusion model

    by Ruikang Liu, Qingwen Xie, Nan Jiang, Zha Cheng, Zhaohui Yuan, Xiaohui Huang, Mengfei Duan


  • Open Access

    Addressing human speech characteristics in single-channel speaker extraction networks

    by Xinghuo Ye, Li Lin, Guifen Jiang, Zhaohui Yuan

    The objective of single-channel speaker extraction is to isolate the clean speech of a target speaker from a mixture of multiple speakers’ utterances. Conventionally, an auxiliary reference network is utilized to extract the speaker’s voiceprint features from the speech signal, which are then fed as input cues to the primary speech extraction network, thereby enhancing the extraction robustness. Nevertheless, existing studies have largely overlooked the sequential characteristics of speech signals, leading to a mismatch between the receptive field of the model and the inherent signal features. Moreover, prevalent speaker extraction architectures fail to account for the spectral distribution properties of speaker speech. To address these limitations, this study presents an extended version of our previous work ( DOI 10.36227/techrxiv.23849361.2 ), which proposes an innovative approach by extending the Convolutional Neural Network Next (ConvNeXt) model, originally designed for image processing, into the Time-Domain ConvNeXt model (TD-ConvNeXt). The TD-ConvNeXt model is integrated with Temporal Convolutional Networks (TCN) blocks to construct the core architecture of the speech extraction network. Additionally, the ConvNeXt model is adapted into a novel Spk block for the auxiliary network, which effectively learns the identity-related features from the reference speech and represents them as embedding vectors. This methodology enables high-fidelity extraction of the target speaker’s speech while maintaining excellent speech quality. The extraction network and the auxiliary reference network are jointly optimized via multi-task learning. Extensive experimental evaluations demonstrate that the proposed model substantially outperforms state-of-the-art methods in single-channel target speech extraction. Across various signal-to-noise ratios (SNRs), the proposed model achieves superior performance in terms of Scale-Invariant Signal-to-Distortion Ratio (SI-SDR), Perceptual Evaluation of Speech Quality (PESQ), and Short-Time Objective Intelligibility (STOI). Specifically, at an SNR of 5 dB, the proposed model attains SI-SDR, PESQ, and STOI scores of 14.83, 3.03, and 0.934, respectively.


  • Open Access

    Smartphone-Based Psychological Sensing: A Large-Scale Study on the Impact of Extreme Isolation

    by Zhaohui Yuan, Yuqing Cao, Zhong Chen

    The COVID-19 pandemic and associated isolation measures have greatly impacted mental health, especially among students. Previous attempts at using mobile sensors to analyze users' emotional states faced barriers including insufficient data and limited modalities. This study aims to address these limitations and derive insights on psychological changes under extreme isolation. We collected a large-scale multivariate dataset from 725 undergraduate students during the complete COVID-19 campus lockdown period. To our knowledge, this is the largest dataset on this population during an extended isolated period. Features were engineered from mobile sensor data to capture modalities including physical activity, sleep patterns, and social interaction. Additionally, self-reported assessments related to mental health conditions were compiled. This rich dataset was leveraged to develop a machine learning model based on autoencoders to detect emotional states. Comprehensive experiments indicate the model can accurately predict mental health changes using mobile sensor data. Our work has unique contributions in collecting large-scale isolated data, engineering informative modalities for modeling mental health, and providing a validated detection method. This can support rapid screening and intervention for mental health crises, especially those emerging from extreme events. The dataset and models open promising directions for big data analytics in mobile health and psychological research.


  • Open Access

    A Multi-Channel Monitoring and Maintenance Platform for Unmanned Surveillance Station Network

    by Yihao Zheng, Zhaohui Yuan

    In recent years, the demand for unmanned surveillance station networks across different fields has been steadily increasing. To address the challenges of operating and maintaining these stations, this study designs a multi-channel operation and maintenance platform for unmanned surveillance station network with private intranet. The platform comprises a hardware terminal system, a concurrent network relay system, and an operation and maintenance management system. The hardware terminal system utilizes an STM32 main control chip, integrating power supply, multi-sensor, and network modules to collect real-time data on station equipment status, power supply and distribution, network status, and environmental conditions. Data are transmitted via Ethernet or 4G network based on the TCP/IP protocol. The concurrent network relay system, based on the Linux operating system, employs epoll multiplexing IO technology and local and network sockets to achieve concurrent data processing and cross-network data transmission, paving the communication gaps between the private intranet and the internet. The operation and maintenance management system features a mobile app designed with a front-end and back-end separation architecture. The front end uses React Native and Ant Design mobile, while the back end is based on Node.js and Express, combined with a MySQL database. The system implements functions such as user management, data monitoring, fault analysis, and operation and maintenance management. A series of tests indicate that the platform operates stably, offers high real-time performance, and possesses strong processing capabilities. Through the platform, users can effectively conduct remote surveillance of stations, significantly improving operating and maintaining management efficiency.

  • Open Access

    Multiscale knowledge distillation with attention based fusion for robust human activity recognition

    by Zhaohui Yuan, Zhengzhe Yang, Hao Ning, Xiangyang Tang

    Knowledge distillation is an effective approach for training robust multi-modal machine learning models when synchronous multimodal data are unavailable. However, traditional knowledge distillation techniques have limitations in comprehensively transferring knowledge across modalities and models. This paper proposes a multiscale knowledge distillation framework to address these limitations. Specifically, we introduce a multiscale semantic graph mapping (SGM) loss function to enable more comprehensive knowledge transfer between teacher and student networks at multiple feature scales. We also design a fusion and tuning (FT) module to fully utilize correlations within and between different data types of the same modality when training teacher networks. Furthermore, we adopt transformer-based backbones to improve feature learning compared to traditional convolutional neural networks. We apply the proposed techniques to multimodal human activity recognition and compared with the baseline method, it improved by 2.31% and 0.29% on the MMAct and UTD-MHAD datasets. Ablation studies validate the necessity of each component.


  • Open Access

    GDCP-YOLO: Enhancing Steel Surface Defect Detection Using Lightweight Machine Learning Approach

    by Zhaohui Yuan, Hao Ning, Xiangyang Tang, Zhengzhe Yang

    Surface imperfections in steel materials potentially degrade quality and performance, thereby escalating the risk of accidents in engineering applications. Manual inspection, while traditional, is laborious and lacks consistency. However, recent advancements in machine learning and computer vision have paved the way for automated steel defect detection, yielding superior accuracy and efficiency. This paper introduces an innovative deep learning model, GDCP-YOLO, devised for multi-category steel defect detection. We enhance the reference YOLOv8n architecture by incorporating adaptive receptive fields via the DCNV2 module and channel attention in C2f. These integrations aim to concentrate on valuable features and minimize parameters. We incorporate the efficient Faster Block and employ Ghost convolutions to generate more feature maps with reduced computation. These modifications streamline feature extraction, curtail redundant information processing, and boost detection accuracy and speed. Comparative trials on the NEU-DET dataset underscore the state-of-the-art performance of GDCP-YOLO. Ablation studies and generalization experiments reveal consistent performance across a variety of defect types. The optimized lightweight architecture facilitates real-time automated inspection without sacrificing accuracy, offering invaluable insights to further deep learning techniques for surface defect identification across manufacturing sectors.