Haiyi Wu

Mr. Haiyi Wu

orcid

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
25 November 2025 in IEEE Internet of Things Journal
Verified
Journal Article
Closed access
HWHaiyi Wu
Haiyi Wu
KZKai Zhao
Kai Zhao
YXYong Xiong
Yong Xiong
WYWei Yao
Wei Yao
XCXu Chi
Xu Chi
Previous studies on human activity recognition (HAR) using mmWave radar in IoT scenarios have relied on complex network architectures and heavy preprocessing, making them computationally expensive and difficult to deploy on resource-constrained edge devices (e.g., smart homes and healthcare). To address this, we propose an LWDSC-Net, a lightweight wavelet dual-branch coupling network that embodies a novel, physically interpretable design paradigm. Inspired by classic signal processing, the LWDSC-Net embeds a 1-D Haar wavelet transform (WT) as a front-end layer, enabling an end-to-end, structured decoupling of raw signals to construct a dual-branch structure with low and high frequencies (LF and HF) without external manual preprocessing. This signal–architecture co-design dedicates each branch to its respective frequency components, with the multilevel WT concept further extended to the HF branch. For the LF component branch, we propose an improved lightweight temporal convolutional network (LightTCN), which builds the feed-forward network using depthwise convolution. The LightTCN reduces the depth of module stacking and efficiently models long-term global dependencies across the entire action sequence. For the HF components branch, we propose the wavelet-domain depthwise separable convolution (WD-DSC) module. It enhances micromotion features by combining secondary wavelet reconstruction with wavelet-domain convolution and further refines local detail information through deep convolutional layers. Experimental results show that the LWDSC-Net achieves an accuracy of 98.76% on the public dataset while significantly reducing the number of model parameters and computational complexity. More importantly, deployment on an NVIDIA Jetson Nano confirms its excellent real-world performance, featuring ultralow latency and power consumption, which highlights its outstanding potential for edge deployment.

10 November 2025 in IEEE Wireless Communications Letters
Verified
Journal Article
Closed access
KZKai Zhao
Kai Zhao
HWHaiyi Wu
Haiyi Wu
YXYong Xiong
Yong Xiong
LZLeiji Zhu
Leiji Zhu
MXMinhao Xu
Minhao Xu
In frequency division duplex (FDD) mode massive multiple-input multiple-output (MIMO) systems, the base station (BS) must obtain accurate downlink channel state information (CSI) fed back from the user equipment (UE). However, the CSI feedback overhead increases rapidly as the number of antennas grows. Although recent deep learning (DL)-based CSI feedback methods improve feedback performance, their architectures are usually complex and inefficient, hindering practical deployment. To address these issues, we propose StarCANet, a compact and efficient neural network featuring a novel star coordinate attention module (StarCAM) that combines the star operation (element-wise multiplication) with the coordinate attention (CA) mechanism. This design enables the network to effectively capture the non-linear correlations and long-range dependencies of CSI. Experimental results demonstrate that StarCANet achieves highly competitive performance, outperforming the state-of-the-art (SOTA) networks in the indoor scenarios, while its efficient and scalable architecture provides flexibility for practical deployment.

17 March 2026 in Electronics
Verified
Journal Article
Open access
PDF file
KZKai Zhao
Kai Zhao
HWHaiyi Wu
Haiyi Wu
WYWei Yao
Wei Yao
YXYong Xiong
Yong Xiong
In frequency division duplex (FDD) massive multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) systems, the sub-array multi-user (MU) hybrid beamforming architecture is highly attractive because of its low hardware cost and high energy efficiency. However, downlink channel state information (CSI) acquisition and hybrid beamformer optimization remain challenging due to the large feedback overhead and the non-convexity of the beamforming design. To address these issues, we propose an end-to-end deep learning (DL) framework that jointly optimizes pilot training, CSI feedback, and hybrid beamforming, overcoming the limitations of conventional independently designed modules. At the core of the network, we introduce the star efficient location attention (StarELA) module, which combines the implicit high-dimensional representation capability of star operations (element-wise multiplication) with the fine-grained feature localization of efficient location attention (ELA). In addition, for wideband digital beamformer generation, we exploit inter-subcarrier correlation and design a frequency–domain seed generation and interpolation upsampling strategy, which significantly reduces network parameters. Experimental results show that the proposed method approaches the upper-bound performance of conventional hybrid beamforming with ideal CSI, while consistently outperforming existing benchmark methods.

7 April 2026 in Electronics
Verified
Journal Article
Open access
PDF file
HWHaiyi Wu
Haiyi Wu
KZKai Zhao
Kai Zhao
WYWei Yao
Wei Yao
YXYong Xiong
Yong Xiong
Millimeter-wave (mmWave) radar point clouds offer a privacy-preserving solution for Human Activity Recognition (HAR), but their inherent sparsity and noise limit single-modal performance. While multimodal fusion mitigates this issue, existing methods often suffer from severe negative transfer during visual degradation and incur high computational costs, unsuitable for edge devices. To address these challenges, we propose Tac-Mamba, a lightweight cross-modal state space model. First, we introduce a topology-guided distillation scheme that uses a Spatial Mamba teacher to extract structural priors from visual skeletons. These priors are then explicitly distilled into a Point Transformer v3 (PTv3) radar student with a modality dropout strategy. We also developed a Trust-Aware Cross-Modal Attention (TACMA) module to prevent negative transfer. It evaluates the reliability of visual features through a SiLU-activated cross-modal bilinear interaction, smoothly degrading to a pure radar-driven fallback projection when visual inputs are corrupted. Finally, a Lightweight Temporal Mamba Block (LTMB) with a Zero-Parameter Cross-Gating (ZPCG) mechanism captures long-range kinematic dependencies with linear complexity. Experiments on the public MM-Fi dataset under strict cross-environment protocols demonstrate that Tac-Mamba achieves competitive accuracies of 95.37% (multimodal) and 87.54% (radar-only) with only 0.86M parameters and 1.89 ms inference latency. These results highlight the model’s exceptional robustness to modality missingness and its feasibility for edge deployment.