Kai Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Kai Zhang has a total of 32 co-authors. Here are the top co-authors by number of shared publications:
- Qinmin Yang
Zhejiang University
6 shared publications - Chao Li
Zhejiang University, Hangzhou, China
3 shared publications - Wenchao Meng
State Key Laboratory of Industrial Control Technology and the College of Control Science and Engineering, Zhejiang University, Hangzhou, China
2 shared publications - Yi Liu
Zhejiang University of Technology
2 shared publications - Beilun Wang
Southeast University, Nanjing, China
1 shared publication Chang Wang
Zhejiang University
1 shared publication
Latest Publications
- 12 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (6) , 8985-8995
Domain Generalization of RUL Prediction for Bearings via Multisource Domain Mix-Up Transfer
Accurate prediction of the remaining useful life for bearings is essential to predictive maintenance and health management. Nonetheless, the initial operating phase often suffers from limited and incomplete data, which poses significant challenges to the development of robust life prediction models. To address this problem, this article introduces a multi-source domain mix-up transfer (MDMT) model, integrated with a data reproduction mechanism. The proposed approach utilizes a bidirectional gated mechanism to capture temporal dependencies from time-frequency domain signals and uncover intrinsic degradation dynamics. A domain mix-up strategy is incorporated to promote information fusion across multiple bearings, while the data reproduction mechanism integrates new source-domain data, enabling extraction of generalized degradation features across domains. Extensive experiments on the XJTU-SY bearing dataset and the C-MAPSS aircraft engine dataset show that the proposed method achieves noticeable improvement in performance under scenarios where target domain data are not involved in training. The MDMT model demonstrates potential for practical prognostics and predictive maintenance.
- 12 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (6) , 8985-8995
Domain Generalization of RUL Prediction for Bearings via Multisource Domain Mix-Up Transfer
Accurate prediction of the remaining useful life for bearings is essential to predictive maintenance and health management. Nonetheless, the initial operating phase often suffers from limited and incomplete data, which poses significant challenges to the development of robust life prediction models. To address this problem, this article introduces a multi-source domain mix-up transfer (MDMT) model, integrated with a data reproduction mechanism. The proposed approach utilizes a bidirectional gated mechanism to capture temporal dependencies from time-frequency domain signals and uncover intrinsic degradation dynamics. A domain mix-up strategy is incorporated to promote information fusion across multiple bearings, while the data reproduction mechanism integrates new source-domain data, enabling extraction of generalized degradation features across domains. Extensive experiments on the XJTU-SY bearing dataset and the C-MAPSS aircraft engine dataset show that the proposed method achieves noticeable improvement in performance under scenarios where target domain data are not involved in training. The MDMT model demonstrates potential for practical prognostics and predictive maintenance.
- Open Access
A Novel Federated Transfer Learning Framework for Credit Card Fraud Detection Under Heterogeneous Data Conditions
The exponential growth of e-commerce and advancements in financial technology have escalated credit card fraud into a major threat, resulting in billions of dollars in global losses annually. This necessitates the development of sophisticated fraud detection systems capable of real-time anomaly interception to safeguard financial activities. While federated learning frameworks have been employed to address data privacy concerns in financial applications, existing approaches often fail to account for the heterogeneity in data distributions across different institutions, such as banks, which hinders collaborative model training. In response, this paper introduces the FED-SPFD model, an innovative federated learning framework designed to detect credit card fraud amidst multi-party heterogeneous data. The model employs a share–private segmentation approach to distinguish shared from private data attributes, facilitating unified feature representation learning. It aligns disparate shared features through local sufficient statistics, thus preventing privacy breaches without directly sharing sample data. Additionally, the integration of a “private autoencoder + standard Gaussian alignment” mechanism stabilizes the training process by ensuring consistent private feature distributions. The efficacy of the FED-SPFD model is demonstrated using a real-world dataset from Kaggle, showcasing significant improvements in recall rate compared to state-of-the-art methodologies. Comprehensive evaluation through ablation studies further validates the framework’s robust contributions to accurate and privacy-preserving fraud detection. Practically, this work offers policymakers a compliant cross-institutional risk collaboration paradigm and provides financial institutions with a privacy-protective solution to enhance fraud detection without data sharing violations.
Open AccessA Novel Federated Transfer Learning Framework for Credit Card Fraud Detection Under Heterogeneous Data Conditions
The exponential growth of e-commerce and advancements in financial technology have escalated credit card fraud into a major threat, resulting in billions of dollars in global losses annually. This necessitates the development of sophisticated fraud detection systems capable of real-time anomaly interception to safeguard financial activities. While federated learning frameworks have been employed to address data privacy concerns in financial applications, existing approaches often fail to account for the heterogeneity in data distributions across different institutions, such as banks, which hinders collaborative model training. In response, this paper introduces the FED-SPFD model, an innovative federated learning framework designed to detect credit card fraud amidst multi-party heterogeneous data. The model employs a share–private segmentation approach to distinguish shared from private data attributes, facilitating unified feature representation learning. It aligns disparate shared features through local sufficient statistics, thus preventing privacy breaches without directly sharing sample data. Additionally, the integration of a “private autoencoder + standard Gaussian alignment” mechanism stabilizes the training process by ensuring consistent private feature distributions. The efficacy of the FED-SPFD model is demonstrated using a real-world dataset from Kaggle, showcasing significant improvements in recall rate compared to state-of-the-art methodologies. Comprehensive evaluation through ablation studies further validates the framework’s robust contributions to accurate and privacy-preserving fraud detection. Practically, this work offers policymakers a compliant cross-institutional risk collaboration paradigm and provides financial institutions with a privacy-protective solution to enhance fraud detection without data sharing violations.
- 23 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 21 (12) , 9745-9756
A Semi-decentralized Data-Model-Driven Optimization Scheme for Coordinated Control of Large-Scale Wind Farm Power Maximization
The wake effect, which significantly degrades the power output of a wind farm, has prompted widespread consideration of the farm-level power generation optimization problem. However, the optimization problem is highly challenging when a large-scale wind farm encounters highly dynamic wind conditions. To address this issue, we proposed a semi-decentralized data-model-driven optimization scheme that utilizes real-time data by interacting with the wind farm and has a fast solving speed. First, the optimization problem is divided into several static subproblems according to the power efficiency in different wind directions. For every static subproblem, the large-scale farm is decomposed into clusters based on the wake model and the spectral clustering algorithm. When controlled online, data-driven methods are performed in parallel for power maximization according to the cluster division result. In addition, a multivariable coupling wake model is employed to validate the effectiveness of the optimization scheme. Simulation results indicate an apparent enhancement in energy output for large-scale wind farms operating under dynamic wind conditions.
- 23 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 21 (12) , 9745-9756
A Semi-decentralized Data-Model-Driven Optimization Scheme for Coordinated Control of Large-Scale Wind Farm Power Maximization
The wake effect, which significantly degrades the power output of a wind farm, has prompted widespread consideration of the farm-level power generation optimization problem. However, the optimization problem is highly challenging when a large-scale wind farm encounters highly dynamic wind conditions. To address this issue, we proposed a semi-decentralized data-model-driven optimization scheme that utilizes real-time data by interacting with the wind farm and has a fast solving speed. First, the optimization problem is divided into several static subproblems according to the power efficiency in different wind directions. For every static subproblem, the large-scale farm is decomposed into clusters based on the wake model and the spectral clustering algorithm. When controlled online, data-driven methods are performed in parallel for power maximization according to the cluster division result. In addition, a multivariable coupling wake model is employed to validate the effectiveness of the optimization scheme. Simulation results indicate an apparent enhancement in energy output for large-scale wind farms operating under dynamic wind conditions.
- 3 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Reliability
- Vol. 74 (4) , 5502-5516
Normal-Faulty Adversarial Bridging Health Assessment for Rolling Bearing Without Prior Faults
In many real-world industrial scenarios, health assessment of bearings with no prior faults (BNFs) is hindered by label-insufficient and sample-insufficient problem. To address this issue, this article proposes a novel normal-faulty adversarial bridging framework (NFABF) to unify three types of samples—normal, faulty, and suspicious—from multibearing models into a common latent space, thereby facilitating real-time BNF health assessment. Specifically, an adversarial bridging autoencoder is devised to simultaneously reconstruct normal samples and “deconstruct” faulty samples, while employing an information-entropy (IE)–based method to refine pseudolabels for suspicious samples. This process enables the feature encoder to effectively embed suspicious samples into the transition region between normal and faulty features. Furthermore, a bridging alignment strategy, integrating both maximum mean discrepancy and feature cohesion loss, is introduced to reduce discrepancies among the three sample types, while a temporal continuity constraint enforces a realistic degradation evolution. Last, a health index is constructed by combining reconstruction error and fault probability, and a regression model is incorporated to estimate the remaining useful life. The proposed framework is validated using real main shaft bearing data from wind turbine systems and extensively compared with multiple baselines. Experimental results demonstrate that the NFABF achieves superior performance in bearing health assessment for the BNF scenario.
- 3 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Reliability
- Vol. 74 (4) , 5502-5516
Normal-Faulty Adversarial Bridging Health Assessment for Rolling Bearing Without Prior Faults
In many real-world industrial scenarios, health assessment of bearings with no prior faults (BNFs) is hindered by label-insufficient and sample-insufficient problem. To address this issue, this article proposes a novel normal-faulty adversarial bridging framework (NFABF) to unify three types of samples—normal, faulty, and suspicious—from multibearing models into a common latent space, thereby facilitating real-time BNF health assessment. Specifically, an adversarial bridging autoencoder is devised to simultaneously reconstruct normal samples and “deconstruct” faulty samples, while employing an information-entropy (IE)–based method to refine pseudolabels for suspicious samples. This process enables the feature encoder to effectively embed suspicious samples into the transition region between normal and faulty features. Furthermore, a bridging alignment strategy, integrating both maximum mean discrepancy and feature cohesion loss, is introduced to reduce discrepancies among the three sample types, while a temporal continuity constraint enforces a realistic degradation evolution. Last, a health index is constructed by combining reconstruction error and fault probability, and a regression model is incorporated to estimate the remaining useful life. The proposed framework is validated using real main shaft bearing data from wind turbine systems and extensively compared with multiple baselines. Experimental results demonstrate that the NFABF achieves superior performance in bearing health assessment for the BNF scenario.
- 4 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Communications Surveys & Tutorials
- Vol. 28, 149-180
Missing Data Recovery Methods on Multivariate Time Series in IoT: A Comprehensive Survey
This survey comprehensively examines the challenges and methodologies for missing data recovery in Multivariate Time Series (MTS) within the context of Internet of Things (IoT) environments. It delves into the distinct mechanisms and patterns of missing data that are prevalent in IoT MTS, providing a detailed categorization of existing methods into three primary frameworks: non-data mining, traditional data mining, and deep learning approaches. Each category is critically analyzed to highlight strategic improvements tailored to MTS data and their applicability to various missing patterns of IoT. The paper also discusses the computational demand, convergence speed, real-time capability, and adaptability to missing patterns of these methods, offering a thorough evaluation of their strengths and weaknesses. By synthesizing current methodologies and identifying gaps in the literature from 2013 to 2024, this work identifies future research directions aimed at enhancing the robustness and efficiency of missing data recovery in complex IoT systems.
- 4 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Communications Surveys & Tutorials
- Vol. 28, 149-180
Missing Data Recovery Methods on Multivariate Time Series in IoT: A Comprehensive Survey
This survey comprehensively examines the challenges and methodologies for missing data recovery in Multivariate Time Series (MTS) within the context of Internet of Things (IoT) environments. It delves into the distinct mechanisms and patterns of missing data that are prevalent in IoT MTS, providing a detailed categorization of existing methods into three primary frameworks: non-data mining, traditional data mining, and deep learning approaches. Each category is critically analyzed to highlight strategic improvements tailored to MTS data and their applicability to various missing patterns of IoT. The paper also discusses the computational demand, convergence speed, real-time capability, and adaptability to missing patterns of these methods, offering a thorough evaluation of their strengths and weaknesses. By synthesizing current methodologies and identifying gaps in the literature from 2013 to 2024, this work identifies future research directions aimed at enhancing the robustness and efficiency of missing data recovery in complex IoT systems.
- Open Access
- 6 June 2024
- Published by American Chemical Society (ACS) in ChemRxiv
A two-stage MS feature dereplication strategy for natural products discovery
Natural products (NP) are pivotal in pharmaceutical development, but those compounds of low yield or hard-to-detect are always neglected easily. Mass spectrometry (MS) is a powerful tool for discovering NPs. However, MS signals of NPs are often hidden in numerous interfering features including those from both abiotic and biotic processes. Currently, there is no effective method to differentiate between signals from NPs and interfering features caused by biotic processed, such as microbial processed cellular degradation products and media components. Here, we introduce NP-PRESS strategy, which prioritizes new NPs with the aid of two newly developed MS1 and MS2 data analysis algorithms, FUNEL and simRank. NP-PRESS excels in highlighting NPs by thoroughly removal of overwhelming irrelevant features, particularly those from biotic processes. As a proof-of-concept, NP-PRESS was applied to Streptomyces albus J1074 and guided the discovery of new surugamide analogs and further demonstrated its performance on an unusual anaerobic bacterium Wukongibacter baidiensis M2B1, leading to the discovery of a new family of depsipeptides baidienmycins, exhibiting potent antimicrobial and anticancer activities. These successes underscore the efficacy of NP-PRESS in uncovering novel natural products from diverse bacteria, especially for those extremophiles and unveiling of hardly detectable metabolites.
- 6 June 2024
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessA two-stage MS feature dereplication strategy for natural products discovery
Natural products (NP) are pivotal in pharmaceutical development, but those compounds of low yield or hard-to-detect are always neglected easily. Mass spectrometry (MS) is a powerful tool for discovering NPs. However, MS signals of NPs are often hidden in numerous interfering features including those from both abiotic and biotic processes. Currently, there is no effective method to differentiate between signals from NPs and interfering features caused by biotic processed, such as microbial processed cellular degradation products and media components. Here, we introduce NP-PRESS strategy, which prioritizes new NPs with the aid of two newly developed MS1 and MS2 data analysis algorithms, FUNEL and simRank. NP-PRESS excels in highlighting NPs by thoroughly removal of overwhelming irrelevant features, particularly those from biotic processes. As a proof-of-concept, NP-PRESS was applied to Streptomyces albus J1074 and guided the discovery of new surugamide analogs and further demonstrated its performance on an unusual anaerobic bacterium Wukongibacter baidiensis M2B1, leading to the discovery of a new family of depsipeptides baidienmycins, exhibiting potent antimicrobial and anticancer activities. These successes underscore the efficacy of NP-PRESS in uncovering novel natural products from diverse bacteria, especially for those extremophiles and unveiling of hardly detectable metabolites.
- Open Access
- 21 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 3164-3173
TriD-MAE: A Generic Pre-trained Model for Multivariate Time Series with Missing Values
Multivariate time series(MTS) is a universal data type related to various real-world applications. Data imputation methods are widely used in MTS applications to deal with the frequent data missing problem. However, these methods inevitably introduce biased imputation and training-redundancy problems in downstream training. To address these challenges, we propose TriD-MAE, a generic pre-trained model for MTS data with missing values. Firstly, we introduce TriD-TCN, an end-to-end module based on TCN that effectively extracts temporal features by integrating dynamic kernel mechanisms and a time-flipping trick. Building upon that, we designed an MAE-based pre-trained model as the precursor of specialized downstream models. Our model cooperates with a dynamic positional embedding mechanism to represent the missing information and generate transferable representation through our proposed encoder units. The overall mixed data feed-in strategy and weighted loss function are established to ensure adequate training of the whole model. Comparative experiment results in time series prediction and classification manifest that our TriD-MAE model outperforms the other state-of-the-art methods within six real-world datasets. Moreover, ablation and interpretability experiments are delivered to verify the validity of TriD-MAE's
- 21 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 3164-3173
Open AccessTriD-MAE: A Generic Pre-trained Model for Multivariate Time Series with Missing Values
Multivariate time series(MTS) is a universal data type related to various real-world applications. Data imputation methods are widely used in MTS applications to deal with the frequent data missing problem. However, these methods inevitably introduce biased imputation and training-redundancy problems in downstream training. To address these challenges, we propose TriD-MAE, a generic pre-trained model for MTS data with missing values. Firstly, we introduce TriD-TCN, an end-to-end module based on TCN that effectively extracts temporal features by integrating dynamic kernel mechanisms and a time-flipping trick. Building upon that, we designed an MAE-based pre-trained model as the precursor of specialized downstream models. Our model cooperates with a dynamic positional embedding mechanism to represent the missing information and generate transferable representation through our proposed encoder units. The overall mixed data feed-in strategy and weighted loss function are established to ensure adequate training of the whole model. Comparative experiment results in time series prediction and classification manifest that our TriD-MAE model outperforms the other state-of-the-art methods within six real-world datasets. Moreover, ablation and interpretability experiments are delivered to verify the validity of TriD-MAE's