Baiyu Zhu

Baiyu Zhu

Auto-generated profile, based on publicly available information

Baiyu Zhu

State Key Laboratory of Industrial Control Technology, College of Control Science and Engineering, Zhejiang University

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
3 November 2025 in IEEE Access
Verified
Journal Article
Open access
BZBaiyu Zhu
Baiyu Zhu
YCYunqi Cao
Yunqi Cao
Dibo HouDibo Hou
Dibo Hou
PHPingjie Huang
Pingjie Huang
Accurate detection of internal defects in battery CT images is essential for ensuring the safety and reliability of energy storage systems. However, the limited number and imbalance of defect samples—especially rare cases such as electrode cracks and metallic inclusions—greatly affect the performance of deep learning models. To address this challenge, we propose HDA-SinGAN, a lightweight image generation framework capable of producing realistic and diverse defect samples from a single image while simultaneously reducing the parameters by 19.6%. HDA-SinGAN integrates High-Frequency Loss, a Dynamic Channel Scaling (DCS) mechanism, and an Adaptive Feature Aggregation (AFA) module, which together preserve fine details and enhance spatial feature extraction. Experimental results on industrial battery CT datasets show that HDA-SinGAN outperforms SinGAN in both image quality and efficiency. For cracks, it achieves improvements of 2.1% in SIFID, 2.5% in SSIM, and 27.5% in PSNR; for metallic inclusions, the gains are 1.8%, 1.9%, and 19.3%, respectively. Incorporating the generated samples into YOLOv8-n training further increases the mean Average Precision (mAP) by 10.6% for cracks and 9.4% for inclusions. These results indicate that HDA-SinGAN effectively balances training datasets and enhances the detection of rare battery defects, providing a lightweight and efficient solution for industrial applications.

17 March 2026 in Engineering Research Express
Verified
Journal Article
Closed access
BZBaiyu Zhu
Baiyu Zhu
Jinglun XuJinglun Xu
Jinglun Xu
YCYunqi Cao
Yunqi Cao
Dibo HouDibo Hou
Dibo Hou
PHPingjie Huang
Pingjie Huang
GZGuangxin Zhang
Guangxin Zhang
Computed Tomography (CT) imaging combined with deep learning models is increasingly employed for defect detection in lithium-ion batteries of electric vehicles. However, analyzing battery CT images presents major challenges, including complex electrode textures, the presence of tiny defects, and computational constraints. To address these issues, we propose GEG-YOLO, an optimized deep learning framework designed to balance detection accuracy and efficiency. The model integrates four key innovations: a Learnable Grouped Gabor Filter for enhanced texture feature extraction, an Enhanced Dual Attention Module (EDAM) to strengthen spatial-channel feature correlation, a P2 Detection Head specialized for tiny defect detection, and a Ghost-style Feature Aggregation Neck (GFA-Neck) for efficient multi-scale feature fusion with reduced parameters. Experimental results demonstrate that GEG-YOLO achieves a 96.1% mean average precision (mAP), representing a 4.8% improvement over YOLOv8, while reducing 11.3k parameters (41.4k with the P2 head) for a total reduction of 12.3%. With an inference speed of 9.5 ms per image, GEG-YOLO provides a lightweight yet effective solution for vision-based defect detection in CT imaging, advancing both methodological innovation and practical application in battery inspection.

26 June 2026 in Engineering Research Express
Verified
Journal Article
Closed access
QZQingle Zhang
Qingle Zhang
YZYe Zheng
Ye Zheng
ZLZhiqiang Li
Zhiqiang Li
JXJigen Xia
Jigen Xia
BZBaiyu Zhu
Baiyu Zhu
PHPingjie Huang
Pingjie Huang
Accurate quantitative evaluation of small inner-wall corrosion defects in natural gas pipelines is critical for ensuring structural integrity, yet remains a challenging task in ultrasonic nondestructive testing due to severe noise interference, weak defect echoes, and temporal overlap in electromagnetic acoustic transducer (EMAT) signals. Existing approaches struggle to simultaneously address unknown sparsity and impulse response uncertainty, limiting their applicability in practical scenarios. To overcome these limitations, this paper proposes a sparse blind deconvolution framework that explicitly models EMAT signals as the convolution of a sparse reflection sequence and an unknown impulse response. A key innovation is the introduction of a kurtosis-driven sparsity adaptive orthogonal matching pursuit algorithm, in which a statistical kurtosis-based termination criterion is embedded into the iterative process to automatically determine the number of effective echoes without prior sparsity information. In addition, an alternating iterative optimization strategy is developed to jointly estimate the reflection sequence and refine the impulse response, where the L-BFGS-B algorithm is employed to correct the transducer response under boundary constraints. This joint optimization mechanism effectively mitigates model mismatch and suppresses false echoes, significantly improving reconstruction accuracy and robustness. Simulation and experimental results demonstrate that the proposed method can reliably separate overlapping echoes, recover weak reflection components under low signal-to-noise ratio conditions, and achieve high-precision depth estimation for corrosion defects ranging from 1 mm to 7 mm. Compared with conventional methods, the proposed approach shows superior robustness to noise and initialization errors while maintaining strong physical interpretability. The proposed framework provides a data-efficient and practical solution for EMAT-based pipeline inspection, with significant potential for improving the reliability of in-service pipeline safety assessment.