Xihui Zhang

Dr. Xihui Zhang


Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
22 November 2025 in Sensors
Verified
Journal Article
Open access
PDF file
XZXihui Zhang
Xihui Zhang
LZLinrun Zhang
Linrun Zhang
MZMeng Zhang
Meng Zhang
ZZZhenxi Zhang
Zhenxi Zhang
PLPeiru Li
Peiru Li
FZFeng Zhou
Feng Zhou
Automatic Modulation Recognition (AMR) is a critical technology for intelligent wireless communication systems, but the deployment of high-performance deep learning models is often hindered by their substantial computational and memory requirements. To address this challenge, this paper proposes a multi-level knowledge distillation network, namely MLD-Net, for creating a lightweight and powerful AMR model. Our approach employs a large Transformer-based network as a teacher to guide the training of a compact and efficient Reformer-based student model. The knowledge contained in the large model is transferred across three distinct granularities: at the output level, to convey high-level predictive distributions; at the feature level, to align intermediate representations; and at the attention level, to propagate relational information about signal characteristics. This comprehensive distillation strategy empowers the student model to effectively emulate the teacher’s complex reasoning processes. Experimental results on the RML2016.10A benchmark dataset demonstrate that MLD-Net achieves state-of-the-art performance, outperforming other baseline models across a wide range of signal-to-noise ratios while requiring only a fraction of the parameters. Extensive ablation study further confirms the collaborative contribution of each distillation level, validating that the proposed MLD-Net is an effective solution for developing lightweight and efficient AMR networks for edge deployment.

8 July 2026 in Sensors
Verified
Journal Article
Open access
PDF file
XZXihui Zhang
Xihui Zhang
MZMeng Zhang
Meng Zhang
SWSun Wen
Sun Wen
YJYinuo Ji
Yinuo Ji
RCRuihan Chen
Ruihan Chen
Tao LiuTao Liu
Tao Liu
Existing target sorting methods suffer severe performance degradation or even failure under inherent severe spectrum overlap, homogeneous protocol parameters, and scarce single-source points in Synchronous Non-Orthogonal Frequency Hopping (SNOFH) scenarios. To address this challenge, this paper proposes a passive sorting framework for SNOFH UAV swarm signals based on multi-station relative hopping time difference. The proposed framework constructs a spatial-location-driven sorting feature system, designs a kernel joint distribution adaptation module to eliminate inter-station measurement discrepancies, and develops a multi-scale wavelet-based method to achieve sub-sampling level hopping time extraction, reducing the dependence on prior FH parameters and hardware radio frequency fingerprints. Experimental comparisons between the proposed and reference sorting methods are conducted on a simulated SNOFH dataset to validate the performance of the proposed sorting framework. The experimental results show that the proposed method achieves the highest sorting accuracy of 98%, outperforming adopted baselines in most SNOFH cases. The proposed method exhibits favorable robustness with noise interference, clock-synchronization error, carrier-frequency offset and multipath influence. It is a suitable choice for UAV swarm sorting under regular and slow-varying UAV formations.

29 September 2026 in Sensors
Verified
Journal Article
Open access
PDF file
XZXihui Zhang
Xihui Zhang
XZXinyuan Zhang
Xinyuan Zhang
WLWenke Li
Wenke Li
JLJianqing Li
Jianqing Li
JLJiaxu Liu
Jiaxu Liu
RF signal classification provides modulation information that can support wireless spectrum awareness after signal acquisition and segmentation. This paper presents a self-weighted convolutional neural network for classifying communication-signal modulation formats from in-phase/quadrature (IQ) sequences. In the proposed multi-branch block, the normalized output channels of one convolutional branch are multiplied by learnable scaling coefficients before aggregation. Stacked blocks are followed by squeeze-and-excitation recalibration and a fully connected classification head. The network is evaluated on a MATLAB-generated 12-class IQ dataset under variations in network depth, symbol rate, carrier-frequency offset, and transfer settings, and is compared with Inception, ResNet50, CaiT, and ViT for Small-Size Datasets. An institute-provided IQ dataset with the same modulation categories provides an additional within-dataset evaluation. On the synthetic dataset, the proposed network achieves an average test accuracy of 91.91%, 0.52 percentage points above Inception. Fine-tuning on a small target-domain subset achieves 91.81% average accuracy while reducing the observed convergence requirement from 21 to 5 epochs compared with training from scratch on the target dataset. On the institute-provided dataset, the proposed, transfer, and Inception networks achieve 93.61%, 92.75%, and 92.32%, respectively. These results support the proposed method as a modulation-aware RF signal-classification component for wireless spectrum-awareness workflows under the evaluated synthetic and institute-provided data settings.