Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Kai Zhong has a total of 88 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Tearing mode instability prediction in EAST tokamak using multi-task contrastive learning and multimodal diagnostics

    by Wenchong Yao, Kai Zhong, Ting Lan, Weihua Wang, Zhihao Liu, Haiqing Liu, Yuntao Song, Hailin Zhao

    Accurate prediction of tearing-mode (TM) characteristics from heterogeneous diagnostics is important for tokamak monitoring and control-oriented analysis. In this work, we present MTCF-Net, a multimodal framework for joint prediction of TM existence, frequency, and strength using Mirnov probe signals and multi-channel electron cyclotron emission (ECE) measurements from EAST. The model uses three parallel feature extractors to process magnetic time-domain signals, magnetic time-frequency representations, and ECE spatiotemporal signals. We introduce a physics-motivated contrastive regularization term to encourage consistency between complementary magnetic representations while preserving the difference between magnetic perturbation features and thermodynamic responses. We also use a classification-guided multi-task objective to reduce unstable regression outputs during TM-absent intervals. The dataset contains 68 expert-annotated EAST discharges with TM activity and is split by shot number into training, validation, and independent test sets. On the independent 11-shot test set, MTCF-Net achieves an F 1-score of 0.991 for TM detection and an R 2 value of 0.966 for both frequency and strength prediction on held-out test shots. Ablation and noise-injection studies suggest that the contrastive regularization mainly improves the stability of strength prediction, while multimodal fusion improves frequency prediction. These results indicate that MTCF-Net is a promising framework for TM prediction on the curated EAST dataset, while maintaining low inference latency.

  • Graph Continual Learning Based Lithium-Ion Battery State of Health Prediction With Longer-Range Patterns Discovering

    by Kai Zhong, Yinglai Deng, Darong Huang, Jiaqiang Tian, Xinghua Liu, Peng Wang

    Accurate state-of-health (SOH) prediction of lithium-ion batteries is critical for ensuring operational safety, reliability, and extending service lifetime in energy storage systems. Existing methods for predicting the SOH of lithiumion batteries have predominantly focused on per-cycle estimation, often overlooking two critical longer-range patterns: latent temporal dependencies across cycles within single battery and continual sharing knowledge across battery packs, both of which are critical for capturing degradation patterns. To address this, we propose a graph continual learning framework with longer-range patterns discovering. First, to explicitly characterize cross-cycle temporal dependencies, a temporal graph representation is formulated through an enhanced multi-order dynamic time warping, wherein directed edges deterministically encode inter-cycle relational patterns. Then, an edge-oriented network is designed to depict temporal representations and aggregation along directed edges, further augmented with a global cell state to guide the capture of persistent memory across batteries. After that, a replay continual learning strategy is adopted to mitigate catastrophic forgetting, combined with lightweight graph distillation to significantly reduce memory overhead while ensuring practical deployment. Finally, the proposed method achieves an average root mean square error (RMSE) of 0.693% in a real-world lithium-ion battery dataset, yielding a 23.09% average improvement over baseline methods, even with only 1.25% of original data volume, thereby demonstrating its effectiveness and strong generalization capability.

  • Self-Adapting Federated Continual Learning With Gradient Causal Constraint for Fault Propagation Path Identification

    by Kai Zhong, Xiang Peng, Zhenhua Fan, Darong Huang, Shixiang Lu

    In modern industrial scenarios, federated learning (FL) has been widely adopted due to its advantages in privacy preservation and distributed modeling. However, existing FL approaches rely on a single update paradigm, which significantly hampers communication efficiency. Moreover, the server preserves historical aggregation states to model causal propagation across training rounds, instead of performing memoryless per-round aggregation. For this end, we propose a self-adapting federated continual learning with gradient causal constraint (GCC-SFCL). First, the selection of synchronous or asynchronous communication across FL rounds is reformulated as continual learning task, where an experience accumulation mechanism records historical aggregation states to model long-term temporal dependencies among client updates. Based on the aggregated context, we further design a self-adapting scheduling strategy driven by multidimensional client state information, enabling the server to dynamically select optimal aggregation. Furthermore, a causality-steering gradient constraint is introduced to explicitly regulate gradient updates, improving the interpretability of the causal model and allowing the capture of fault propagation paths. Experimental results of a simulation case and the real-world case demonstrate that GCC-SFCL significantly improves not only communication efficiency but also the accuracy of fault propagation path identification.

  • Continuous Causal Learning of Multimode Industrial Processes for Comprehensive Root Cause Analysis

    by Kai Zhong, Yingcheng Xu, Xiaoming Zhang, Hongtian Chen, Enrico Zio

    Accurately inferring the causality among variables is essential for diagnosing the root causes of industrial process faults. The multimode characteristics of industrial processes augment the complexity of fault evolution processes, which present challenges to traditional methods. To this end, the continuous causal learning framework is proposed for comprehensive root cause diagnosis (RCD) of multimode processes. Firstly, the temporal feature interference aided causal discovery network (TFI-CDN) is designed to achieve collaborative quantification of multivariate causality, which can facilitate continuous learning while avoiding pairwise causal modeling. Then, to address the “catastrophic forgetting” issue in continuous learning, an elastic weight consolidation (EWC) strategy is introduced to balance the newly emerged modes and previously learned modes through adaptive parameter regularization during sequential model updating. Subsequently, the RCD index impact score (IS) is designed to locate the real root cause according to the causality variations from normal to faults with the Cramer-von Mises (CVM) test. In addition, the propagation paths are identified through the strength and time lag (SL) metric by involving causality SL simultaneously. Finally, the proposed method significantly outperforms the existing methods in terms of RCD on both simulated and real-world multimode processes.


  • Hierarchical Causal Graph Neural Networks With Cascading Failure Analysis for Complex Systems Root Cause Identification

    by Kai Zhong, Yingcheng Xu, Yinglai Deng, Wenkai Hu, Fan Yang

    Modeling the evolution process of industrial cascading faults by capturing causal relationships has achieved widespread attention in the attack detection and security control field. However, the highly coupled nature of numerous sensors and control loops in complex systems makes it difficult for existing root cause identification (RCI) methods to accurately and clearly characterize the fault-induced causal relationships. To address this challenge, this article proposes the hierarchical causal graph neural networks (HCGNNs) to reveal the root cause in an end-to-end manner with the aid of cascading failure analysis. First, the cause–effect weighted adjacency matrix is designed to expand the connections between nodes from correlations to causalities in a graph neural network (GNN). On this basis, the proposed hierarchical network structure can collaboratively learn the intraunit and interunit causalities, which significantly reduces the spurious and redundant causal relations while improving the modeling efficiency. Additionally, a new cascading failure analysis method is formulated to quantify the root cause and clarify propagation paths from a more objective perspective. Finally, the effectiveness and superiority of the proposed method are verified by the Tennessee Eastman platform and the real-world coal mill group.


  • Trustworthy Open Set Domain Generalization Network for Unknown Fault Diagnosis Under Unseen Conditions

    by Zhenhua Fan, Bing Yan, Che Xu, Shixiang Lu, Kai Zhong

    Advanced intelligent fault diagnosis methods based on domain generalization leverage multi-source sensor data to overcome the limitations the limitations of domain adaptation-based models regarding target data demand in the training stage, which resolves domain shift problems under unseen conditions in industrial instrumentation. However, previous studies primarily focused on the closed-set diagnosis with same label space shared by the training and testing data, which struggle to address the problem of new fault identification under intricate industrial dynamics. To overcome this obstacle, a trustworthy Evidential Open-Set Domain Generalization Network (EOSDGN) is proposed for open-set fault diagnosis under unseen conditions. In the EOSDGN method, an evidential deep classifier is constructed to quantify the uncertainty of predicted results, an evidential domain discriminator is employed to integrate the data from diverse source domains, and an evidential uncertainty calibration is established to reconcile the misleading evidence and class-based evidence assignments. The EOSDGN model effectively addresses both domain and label shift challenges by integrating domain-invariant features extraction and uncertainty quantification of the predicted probabilities, which enables efficient classification of known faults while simultaneously facilitating the identification of unknown faults. The effectiveness of the EOSDGN model has been substantiated using public and practical datasets. Experimental findings demonstrate that the EOSDGN model surpasses the performance of the state-of-the-art models.


  • Multistage Consistency-Constrained Dynamic Causal Learning for Evolution-Revealed Root Cause Analysis

    by Kai Zhong, Juhao Yang, Xiaoming Zhang, Darong Huang

    Exploring the causality among variables is crucial for uncovering the fault root cause. However, the existing methods are difficult to achieve a fine characterization of the causality evolutionary process. To this end, this paper proposes a multi-stage consistency constrained dynamic causal learning for root cause identification. Firstly, the CPD-DenStream is proposed to objectively describe the evolution process of faults, which lay the foundation for the following causal learning and root cause analysis. Then, the designed causal Transformer introduces a variable-level attention mechanism, which can accurately capture the causality among variables through the continuous dynamic learning. Subsequently, the F-Diff causal loss and interaction redundancy based dynamic causal learning are combined to guide the learning and adjustment of causality, making them more compatible throughout the entire evolution process of faults. Finally, the proposed method is evaluated on Tennessee Eastman process and the real-world coal mill dataset, achieving minimum improvement 8.7% in F1-score, 9.8% in precision, 8.4% in recall compared with baseline methods. These results demonstrate the superiority of the proposed method in capturing multi-stage causality, thereby advancing the practical implementation of root cause analysis systems.