Fengyuan Zuo
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Fengyuan Zuo has a total of 41 co-authors. Here are the top co-authors by number of shared publications:
- Jinhai Liu
College of Information Science and Engineering, Northeastern University, Shenyang 110819, China
27 shared publications - Lei Wang
School of Mechatronic Engineering and Automation, Shanghai University, Shanghai, China
18 shared publications - Huaguang Zhang
Northeastern University
12 shared publications - Zhen Zhao
School of Information Science and Engineering, Northeastern University, Shenyang, China
9 shared publications - Mingrui Fu
Shenyang Paidelin Technology Company Ltd., Shenyang, China
6 shared publications - Xiangkai Shen
School of Automation Science and Electrical Engineering, Beihang University, Beijing, China
5 shared publications
Latest Publications
- 1 October 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 334
DecoupDiff: Spatially decoupled denoising diffusion model for industrial anomaly detection
- 1 October 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 334
DecoupDiff: Spatially decoupled denoising diffusion model for industrial anomaly detection
- 3 August 2026
- Published by Elsevier in Journal of Industrial Information Integration
- Vol. 53
Physics-guided blind super-resolution of MFL signals for pipeline inspection
- 3 August 2026
- Published by Elsevier in Journal of Industrial Information Integration
- Vol. 53
Physics-guided blind super-resolution of MFL signals for pipeline inspection
- 13 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Automation Science and Engineering
- Vol. 23 (15455955) , 9797-9808
A Domain-Incremental Learning Framework Based on Defect Replay and Gaussian Mixup for Industrial Defect Detection
Incremental learning is a critical yet challenging problem in automation engineering, especially across heterogeneous domains. Existing incremental learning methods utilize mixup and mosaic techniques to replay previous knowledge. However, the saliency discrepancy between replayed samples is often ignored, leading to suboptimal accuracy. To address this problem, we propose a novel replay-based incremental learning framework that replays previous defect samples based on mixup and mosaic. To dynamically assign samples to different image augmentation techniques, a saliency score calculation mechanism is proposed by using posterior probability and spatial location. To eliminate the step noise brought by mixup, we propose a Gaussian operator-improved mixup to smoothly merge previous foregrounds with current images. Additionally, a domain-attentive distillation loss is proposed to further alleviate forgetting. Experiments on three scenarios verify the improved accuracy of the proposed method, achieving improvements of 2.4% mAP on weld defect detection, 2.7% mAP on steel defect detection and 4.0% mAP on no-service rail defect detection, respectively.
- 13 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Automation Science and Engineering
- Vol. 23 (15455955) , 9797-9808
A Domain-Incremental Learning Framework Based on Defect Replay and Gaussian Mixup for Industrial Defect Detection
Incremental learning is a critical yet challenging problem in automation engineering, especially across heterogeneous domains. Existing incremental learning methods utilize mixup and mosaic techniques to replay previous knowledge. However, the saliency discrepancy between replayed samples is often ignored, leading to suboptimal accuracy. To address this problem, we propose a novel replay-based incremental learning framework that replays previous defect samples based on mixup and mosaic. To dynamically assign samples to different image augmentation techniques, a saliency score calculation mechanism is proposed by using posterior probability and spatial location. To eliminate the step noise brought by mixup, we propose a Gaussian operator-improved mixup to smoothly merge previous foregrounds with current images. Additionally, a domain-attentive distillation loss is proposed to further alleviate forgetting. Experiments on three scenarios verify the improved accuracy of the proposed method, achieving improvements of 2.4% mAP on weld defect detection, 2.7% mAP on steel defect detection and 4.0% mAP on no-service rail defect detection, respectively.
- 11 March 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 318
A hierarchical knowledge-embedded fine-grained network for pipeline weak defect detection
- 11 March 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 318
A hierarchical knowledge-embedded fine-grained network for pipeline weak defect detection
- Top Cited Papers
- 20 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (7) , 13590-13600
A Sim2Real Defect Inversion Method for Real-World Energy Transportation Systems Based on Intra- and Inter-Domain Interactive Learning
Defect inversion is a critical quantitative assessment technique to evaluate the integrity of energy transportation systems. In practice, nonstationary states and frequent situation switches in real-world industrial processes make it difficult to acquire ideal sensor signals with precise annotations, thereby compromising accurate defect inversion. The simulation system is a viable solution. However, the dramatic inter-domain disparity between the simulated and real systems poses challenges to existing studies. Moreover, intra-domain divergence within the real energy transportation systems is also unexplored. To address the above issues, an intra- and inter-domain interactive learning method is proposed for simulation-to-real (sim2real) defect inversion. Specifically, reality-aware feature stylization is proposed, which diversifies simulated signals with real style variants, so as to alleviate the domain gap initially. Then, bilevel inter-domain alignment is proposed to not only penalize hard-to-align samples in abstract feature space, but also innovatively exploits intrinsic geometric relations of sensor signals to guide inter-domain gaps. Next, physical-informed intra-domain alignment is proposed, where a similarity matrix is constructed to pursue the consistency between sensor signals and physical knowledge, and it also serves as pseudolabels to constrain inter-domain adaptation, so that inter- and intra-domain adaptation interact with united strength to enhance information exchange. Finally, the proposed method is systematically validated on an energy transportation experimental platform and competitive results are obtained, which show its promise in real-world industrial processes.
- 20 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (7) , 13590-13600
Top Cited PapersA Sim2Real Defect Inversion Method for Real-World Energy Transportation Systems Based on Intra- and Inter-Domain Interactive Learning
Defect inversion is a critical quantitative assessment technique to evaluate the integrity of energy transportation systems. In practice, nonstationary states and frequent situation switches in real-world industrial processes make it difficult to acquire ideal sensor signals with precise annotations, thereby compromising accurate defect inversion. The simulation system is a viable solution. However, the dramatic inter-domain disparity between the simulated and real systems poses challenges to existing studies. Moreover, intra-domain divergence within the real energy transportation systems is also unexplored. To address the above issues, an intra- and inter-domain interactive learning method is proposed for simulation-to-real (sim2real) defect inversion. Specifically, reality-aware feature stylization is proposed, which diversifies simulated signals with real style variants, so as to alleviate the domain gap initially. Then, bilevel inter-domain alignment is proposed to not only penalize hard-to-align samples in abstract feature space, but also innovatively exploits intrinsic geometric relations of sensor signals to guide inter-domain gaps. Next, physical-informed intra-domain alignment is proposed, where a similarity matrix is constructed to pursue the consistency between sensor signals and physical knowledge, and it also serves as pseudolabels to constrain inter-domain adaptation, so that inter- and intra-domain adaptation interact with united strength to enhance information exchange. Finally, the proposed method is systematically validated on an energy transportation experimental platform and competitive results are obtained, which show its promise in real-world industrial processes.
- 27 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Neural Networks and Learning Systems
- Vol. 37 (5) , 2321-2334
Leader-Based Multiexpert Neural Network for High-Level Visual Tasks
Remarkable progress has been achieved in the detection and segmentation of the baseline; however, for high-level visual tasks in complex scenes (e.g., dense, occlusion, scale diversity, high background noise, etc.), existing frameworks often fail to provide satisfactory performance. To further improve the object recognition ability, this article introduces a leader-based multiexpert mechanism into the detection and segmentation tasks. In this work, we first design a leader-based attention learning layer to fully integrate multilevel features from the backbone network, which can effectively obtain global semantics and assign instructions to detection experts. Then, we propose multiple feature pyramids with dual fusion paths to replace the traditional single pipeline using semantic and spatial allocators. With this strategy, we can further establish deep supervision for multiple experts during training and sufficiently utilize the multiexpert detection results from leaders’ assignments during reasoning, thereby comprehensively improving the performance of the model in complex scenarios. In the experiment, we established ablation studies and performance comparisons on COCO 2017 detection and segmentation tasks. Finally, we demonstrated the model’s performance in three complex application scenarios (remote sensing, autonomous driving, and industrial fields), and the results showed our advantages.
- 27 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Neural Networks and Learning Systems
- Vol. 37 (5) , 2321-2334
Leader-Based Multiexpert Neural Network for High-Level Visual Tasks
Remarkable progress has been achieved in the detection and segmentation of the baseline; however, for high-level visual tasks in complex scenes (e.g., dense, occlusion, scale diversity, high background noise, etc.), existing frameworks often fail to provide satisfactory performance. To further improve the object recognition ability, this article introduces a leader-based multiexpert mechanism into the detection and segmentation tasks. In this work, we first design a leader-based attention learning layer to fully integrate multilevel features from the backbone network, which can effectively obtain global semantics and assign instructions to detection experts. Then, we propose multiple feature pyramids with dual fusion paths to replace the traditional single pipeline using semantic and spatial allocators. With this strategy, we can further establish deep supervision for multiple experts during training and sufficiently utilize the multiexpert detection results from leaders’ assignments during reasoning, thereby comprehensively improving the performance of the model in complex scenarios. In the experiment, we established ablation studies and performance comparisons on COCO 2017 detection and segmentation tasks. Finally, we demonstrated the model’s performance in three complex application scenarios (remote sensing, autonomous driving, and industrial fields), and the results showed our advantages.
- 1 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (4) , 9427-9438
A X-Ray Based Dual-Expert Detection Method for Automatic Welding Defect Inspection
Automatic welding technology has been widely applied in the field of industrial production. In promoting high-quality intelligent manufacturing processes for the consumer market, precise detection of welding defects is a key link in enhancing product reliability. Considering the high demand for domain knowledge in welding defect inspection tasks, existing methods have shortcomings in fine-grained recognition. To overcome these weaknesses, this article proposes a dual-expert detection method for automatic welding defect inspection, inspired by the human multi-expert complementary evaluation process. The method process is mainly reflected in the following aspects. Firstly, a high-precision robot welding defect inspection system that integrates application platform based data collection, model training, and deployment has been effectively implemented. Then, diverse detection experts are constructed using complementary regression paradigms to obtain predictions from multiple perspectives. Finally, a dual-expert fusion scheme based on probabilistic ensembling is designed to fully integrate prediction information and obtain refined results. In the experiment, a database rich in diverse types of welding defects is applied for training and systematic testing, and multiple sets of refined indicators are evaluated. In addition, we also discussed edge deployment, failure case analysis, and the application of in consumer PCB defect detection. These practical application results indicate that the proposed method can achieve product level implementation, effectively reducing the threshold for industrial testing and enabling end-users to gain benefits.
- 1 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (4) , 9427-9438
A X-Ray Based Dual-Expert Detection Method for Automatic Welding Defect Inspection
Automatic welding technology has been widely applied in the field of industrial production. In promoting high-quality intelligent manufacturing processes for the consumer market, precise detection of welding defects is a key link in enhancing product reliability. Considering the high demand for domain knowledge in welding defect inspection tasks, existing methods have shortcomings in fine-grained recognition. To overcome these weaknesses, this article proposes a dual-expert detection method for automatic welding defect inspection, inspired by the human multi-expert complementary evaluation process. The method process is mainly reflected in the following aspects. Firstly, a high-precision robot welding defect inspection system that integrates application platform based data collection, model training, and deployment has been effectively implemented. Then, diverse detection experts are constructed using complementary regression paradigms to obtain predictions from multiple perspectives. Finally, a dual-expert fusion scheme based on probabilistic ensembling is designed to fully integrate prediction information and obtain refined results. In the experiment, a database rich in diverse types of welding defects is applied for training and systematic testing, and multiple sets of refined indicators are evaluated. In addition, we also discussed edge deployment, failure case analysis, and the application of in consumer PCB defect detection. These practical application results indicate that the proposed method can achieve product level implementation, effectively reducing the threshold for industrial testing and enabling end-users to gain benefits.
- 4 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-10
A Flexible Multitask Framework With Decoupled Training Paradigm for Radiographic Weld Images
Weld sign detection (WSD), weld edge segmentation (WES), and weld defect detection (WDD) are three major tasks in artificial intelligence-empowered radiographic weld images. Most of the existing methods focus on improving detection performance on a single task; the others focus on two tasks and currently still lack a more comprehensive detection method. We argue that simultaneously accomplishing three tasks can not only efficiently utilize the multitask measurement data for downstream industrial applications but also enhance the resource utilization of industrial deployment. To this end, we propose a flexible multitask framework for effectively addressing three tasks. First, a multitask framework is elegantly designed, where the shared features are used as a flexible interface for connecting the shared feature encoder (SFE) with multiple task-specific prediction heads. Particularly, a two-stage decoupled training method is proposed to improve the multitask method’s accuracy by separately pretraining the SFE and fine-tuning the task-specific prediction heads. We instantiate a multitask training-decoupled network (MT-TDN) on an open-source dataset GDXray and a real-world dataset private practical pipeline weld dataset (PPWD). The experimental results demonstrate that the proposed method achieves a higher balance of accuracy and efficiency on three tasks, including WDD (0.959 mAP), WES (0.941 mAP), and WSD (0.972 mAP).
- 4 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-10
A Flexible Multitask Framework With Decoupled Training Paradigm for Radiographic Weld Images
Weld sign detection (WSD), weld edge segmentation (WES), and weld defect detection (WDD) are three major tasks in artificial intelligence-empowered radiographic weld images. Most of the existing methods focus on improving detection performance on a single task; the others focus on two tasks and currently still lack a more comprehensive detection method. We argue that simultaneously accomplishing three tasks can not only efficiently utilize the multitask measurement data for downstream industrial applications but also enhance the resource utilization of industrial deployment. To this end, we propose a flexible multitask framework for effectively addressing three tasks. First, a multitask framework is elegantly designed, where the shared features are used as a flexible interface for connecting the shared feature encoder (SFE) with multiple task-specific prediction heads. Particularly, a two-stage decoupled training method is proposed to improve the multitask method’s accuracy by separately pretraining the SFE and fine-tuning the task-specific prediction heads. We instantiate a multitask training-decoupled network (MT-TDN) on an open-source dataset GDXray and a real-world dataset private practical pipeline weld dataset (PPWD). The experimental results demonstrate that the proposed method achieves a higher balance of accuracy and efficiency on three tasks, including WDD (0.959 mAP), WES (0.941 mAP), and WSD (0.972 mAP).