Fangyu Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Fangyu Liu has a total of 34 co-authors. Here are the top co-authors by number of shared publications:
- Fangmin Sun
Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China
7 shared publications - Hao Wang
Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China
7 shared publications - Weimin Li
School of Computer Engineering and Science, Shanghai University, Shanghai, P. R. China
7 shared publications - Ye Li
Chinese Academy of Sciences
6 shared publications - Xiang Li
School of Electronics and Information Engineering, South China Normal University, Foshan, China
5 shared publications - Qun Jin
Waseda University, Tokyo, Japan
3 shared publications
Latest Publications
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 72 (4) , 10678-10689
A Multi-Scale Network With Cross-View Gated Attention Enhancement for Human Activity Recognition
Human Activity Recognition (HAR) based on inertial sensors plays a pivotal role in consumer electronics applications such as health monitoring, fitness tracking, and human-computer interaction. However, achieving robust HAR performance remains challenging due to significant variations in activity duration and rhythms, restricted views in attention modeling, and information loss caused by excessive dimensionality compression. In this paper, we propose MSNCVGA, a multi-scale network enhanced by cross-view gated attention for inertial sensor-based HAR tasks. This method explicitly models sequence temporal scale diversity through multi-scale convolutions and a cross-scale feature fusion branch, while utilizing the cross-view gated attention mechanism to synergistically focus on complementary and dependent cross-view global key features among channel, sequence, and sample dimensions, effectively avoiding excessive information loss. Finally, a recurrent neural network is employed to integrate contextual features from different scale branches. Experiments on five public datasets demonstrate that MSNCVGA achieves better performance than the baseline methods, reaching F1-scores of 0.848, 0.989, 0.985, 0.642, and 0.979 on PAMAP2, UCI HAR, mHealth, RealWorld, and KU-HAR, respectively. Notably, the proposed model has only 0.54M parameters, providing a high-accuracy, low-latency solution suitable for deployment on resource-constrained consumer electronic devices.
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 72 (4) , 10678-10689
A Multi-Scale Network With Cross-View Gated Attention Enhancement for Human Activity Recognition
Human Activity Recognition (HAR) based on inertial sensors plays a pivotal role in consumer electronics applications such as health monitoring, fitness tracking, and human-computer interaction. However, achieving robust HAR performance remains challenging due to significant variations in activity duration and rhythms, restricted views in attention modeling, and information loss caused by excessive dimensionality compression. In this paper, we propose MSNCVGA, a multi-scale network enhanced by cross-view gated attention for inertial sensor-based HAR tasks. This method explicitly models sequence temporal scale diversity through multi-scale convolutions and a cross-scale feature fusion branch, while utilizing the cross-view gated attention mechanism to synergistically focus on complementary and dependent cross-view global key features among channel, sequence, and sample dimensions, effectively avoiding excessive information loss. Finally, a recurrent neural network is employed to integrate contextual features from different scale branches. Experiments on five public datasets demonstrate that MSNCVGA achieves better performance than the baseline methods, reaching F1-scores of 0.848, 0.989, 0.985, 0.642, and 0.979 on PAMAP2, UCI HAR, mHealth, RealWorld, and KU-HAR, respectively. Notably, the proposed model has only 0.54M parameters, providing a high-accuracy, low-latency solution suitable for deployment on resource-constrained consumer electronic devices.
- Top Cited Papers
- 22 March 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 320
Ontology-enhanced subgraph reasoning with prompt learning for inductive knowledge graph completion
- 22 March 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 320
Top Cited PapersOntology-enhanced subgraph reasoning with prompt learning for inductive knowledge graph completion
- Open Access
- 12 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 924-929
A Sensor-site Hybrid Algorithm Pipeline for Locomotion and Transportation Mode Recognition
- 12 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 924-929
Open AccessA Sensor-site Hybrid Algorithm Pipeline for Locomotion and Transportation Mode Recognition
- Open Access
- 12 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 1018-1023
MixMatch-based Semi-supervised Learning Approach for Cross-domain Locomotion and Transportation Mode Recognition
- 12 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 1018-1023
Open AccessMixMatch-based Semi-supervised Learning Approach for Cross-domain Locomotion and Transportation Mode Recognition
- 6 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (4) , 12527-12538
DBMNet: Dual-Branch Multi-Modal Network With Spatial-Temporal Fusion for Objective Physical Fatigue Assessment
Accurate objective physical fatigue (OPF) assessment is critical for enhancing safety in labor-intensive industries, optimizing athletic performance, and supporting personalized health management. With the advancement of wearable sensing technologies, wearable OPF assessment methods based on multi-source information fusion is emerging is gradually becoming an critical approach for sports and health monitoring. However, wearable OPF assessment methods face the following challenges: spatial modeling deficiency, ineffective multimodal fusion and limited long-range dependency capture. To address these limitations, we propose DBMNet, a dual-branch multi-modal network with spatial-temporal fusion for fatigue level classification. DBMNet captures temporal dynamics and spatial patterns separately from raw 12-lead electrocardiogram (ECG) and multi-site inertial measurement unit (IMU) signals. A novel Convolutional Additive Adaptive Cross-refinement Fusion (CAACF) module is introduced to enable adaptive, efficient, and interpretable fusion of heterogeneous modalities. We collected synchronized ECG and IMU data from 65 subjects during treadmill exercise following a modified Bruce protocol, annotated with both coarse- and fine-grained fatigue levels. Extensive experiments demonstrate that DBMNet achieves significant improvements over baseline and state-of-the-art methods, with up to 7% increase in classification accuracy and superior performance across multiple evaluation metrics. Moreover, DBMNet maintains a lightweight architecture suitable for deployment on mobile and wearable devices. This work provides an effective and scalable framework for real-time, objective fatigue monitoring using multi-source physiological signals.
- 6 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (4) , 12527-12538
DBMNet: Dual-Branch Multi-Modal Network With Spatial-Temporal Fusion for Objective Physical Fatigue Assessment
Accurate objective physical fatigue (OPF) assessment is critical for enhancing safety in labor-intensive industries, optimizing athletic performance, and supporting personalized health management. With the advancement of wearable sensing technologies, wearable OPF assessment methods based on multi-source information fusion is emerging is gradually becoming an critical approach for sports and health monitoring. However, wearable OPF assessment methods face the following challenges: spatial modeling deficiency, ineffective multimodal fusion and limited long-range dependency capture. To address these limitations, we propose DBMNet, a dual-branch multi-modal network with spatial-temporal fusion for fatigue level classification. DBMNet captures temporal dynamics and spatial patterns separately from raw 12-lead electrocardiogram (ECG) and multi-site inertial measurement unit (IMU) signals. A novel Convolutional Additive Adaptive Cross-refinement Fusion (CAACF) module is introduced to enable adaptive, efficient, and interpretable fusion of heterogeneous modalities. We collected synchronized ECG and IMU data from 65 subjects during treadmill exercise following a modified Bruce protocol, annotated with both coarse- and fine-grained fatigue levels. Extensive experiments demonstrate that DBMNet achieves significant improvements over baseline and state-of-the-art methods, with up to 7% increase in classification accuracy and superior performance across multiple evaluation metrics. Moreover, DBMNet maintains a lightweight architecture suitable for deployment on mobile and wearable devices. This work provides an effective and scalable framework for real-time, objective fatigue monitoring using multi-source physiological signals.
- Open Access
- 30 September 2025
- Published by MDPI AG in Electronics
- Vol. 14 (19) , 3905
Analysis of Sensor Location and Time–Frequency Feature Contributions in IMU-Based Gait Identity Recognition
Inertial measurement unit (IMU)-based gait biometrics have attracted increasing attention for unobtrusive identity recognition. While recent studies often fuse signals from multiple sensor positions and time–frequency features, the actual contribution of each sensor location and signal modality remains insufficiently explored. In this work, we present a comprehensive quantitative analysis of the role of different IMU placements and feature domains in gait-based identity recognition. IMU data were collected from three body positions (shank, waist, and wrist) and processed to extract both time-domain and frequency-domain features. An attention-gated fusion network was employed to weight each signal branch adaptively, enabling interpretable assessment of their discriminative power. Experimental results show that shank IMU dominates recognition accuracy, while waist and wrist sensors primarily provide auxiliary information. Similarly, the contribution of time-domain features to classification performance is the greatest, while frequency-domain features offer complementary robustness. These findings illustrate the importance of sensor and feature selection in designing efficient, scalable IMU-based identity recognition systems for wearable applications.
- 30 September 2025
- Published by MDPI AG in Electronics
- Vol. 14 (19) , 3905
Open AccessAnalysis of Sensor Location and Time–Frequency Feature Contributions in IMU-Based Gait Identity Recognition
Inertial measurement unit (IMU)-based gait biometrics have attracted increasing attention for unobtrusive identity recognition. While recent studies often fuse signals from multiple sensor positions and time–frequency features, the actual contribution of each sensor location and signal modality remains insufficiently explored. In this work, we present a comprehensive quantitative analysis of the role of different IMU placements and feature domains in gait-based identity recognition. IMU data were collected from three body positions (shank, waist, and wrist) and processed to extract both time-domain and frequency-domain features. An attention-gated fusion network was employed to weight each signal branch adaptively, enabling interpretable assessment of their discriminative power. Experimental results show that shank IMU dominates recognition accuracy, while waist and wrist sensors primarily provide auxiliary information. Similarly, the contribution of time-domain features to classification performance is the greatest, while frequency-domain features offer complementary robustness. These findings illustrate the importance of sensor and feature selection in designing efficient, scalable IMU-based identity recognition systems for wearable applications.
- 20 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (4) , 9340-9351
A Transfer Learning Method for Cross-Scenario Abnormal Running Posture Detection Based on Wearable Sensors
Running is a popular sport, but abnormal running postures can lead to injuries. With the growing popularity of wearable electronics, real-time detection of abnormal running postures has become increasingly feasible. However, the variability of real-world running scenarios poses significant challenges to the robustness of deep learning-based detection methods. Models trained on fixed scenarios often experience a substantial decline in performance when applied to different environments. To address these issues, this paper proposes a cross-scenario transfer learning (CSTL) framework to detect abnormal running postures by wearable inertial sensors, addressing the challenges of high time complexity and limited adaptability across various running scenarios. In the pre-training stage, we propose a Light-Norm Transformer (LNT) model to improve recognition accuracy using data collected from a single running scenario. In the fine-tuning stage, the pre-trained LNT model is transferred to adapt to five running scenarios with a contribution-guided freezing and dynamic adapter fine-tuning strategy, and the leave-one-out cross-validation method is adopted to evaluate the cross-scenario classification performance on each running scenario in the test module. The proposed CSTL method significantly reduces training time and enhances recognition accuracy, achieving an accuracy of 87.0% on the five running scenarios with only 10 epochs of fine-tuning, outperforming the non-transfer learning approach and achieving the optimal performance-efficiency balance. The proposed CSTL method enhances the adaptability and robustness of abnormal running posture detection systems in practical, real-world applications.
- 20 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (4) , 9340-9351
A Transfer Learning Method for Cross-Scenario Abnormal Running Posture Detection Based on Wearable Sensors
Running is a popular sport, but abnormal running postures can lead to injuries. With the growing popularity of wearable electronics, real-time detection of abnormal running postures has become increasingly feasible. However, the variability of real-world running scenarios poses significant challenges to the robustness of deep learning-based detection methods. Models trained on fixed scenarios often experience a substantial decline in performance when applied to different environments. To address these issues, this paper proposes a cross-scenario transfer learning (CSTL) framework to detect abnormal running postures by wearable inertial sensors, addressing the challenges of high time complexity and limited adaptability across various running scenarios. In the pre-training stage, we propose a Light-Norm Transformer (LNT) model to improve recognition accuracy using data collected from a single running scenario. In the fine-tuning stage, the pre-trained LNT model is transferred to adapt to five running scenarios with a contribution-guided freezing and dynamic adapter fine-tuning strategy, and the leave-one-out cross-validation method is adopted to evaluate the cross-scenario classification performance on each running scenario in the test module. The proposed CSTL method significantly reduces training time and enhances recognition accuracy, achieving an accuracy of 87.0% on the five running scenarios with only 10 epochs of fine-tuning, outperforming the non-transfer learning approach and achieving the optimal performance-efficiency balance. The proposed CSTL method enhances the adaptability and robustness of abnormal running posture detection systems in practical, real-world applications.
- Open Access
- 13 December 2024
- Published by Association for Computing Machinery (ACM)
- p. 38-46
Influencing Factors Mining and Modeling of Energy Expenditure in Running Based on Wearable Sensors
- 13 December 2024
- Published by Association for Computing Machinery (ACM)
- p. 38-46
Open AccessInfluencing Factors Mining and Modeling of Energy Expenditure in Running Based on Wearable Sensors
- 24 September 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 260
Exploring multi-granularity contextual semantics for fully inductive knowledge graph completion
- 24 September 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 260
Exploring multi-granularity contextual semantics for fully inductive knowledge graph completion