Cong Wang

Multivariate time series data are commonly collected by multi-channel sensors in measurement systems and are widely encountered in applications such as industrial monitoring, biomedical analysis, and human activity recognition, etc. A fundamental challenge is to effectively exploit these multi-channel observations to identify the operating states or safety conditions of complex systems. However, the high-dimensional nature of multivariate time series, the intricate relationships among variables, and the inherent complexity of temporal data continue to raise formidable challenges for method development. This paper introduces a novel framework, Dyntopo-Net, which integrates distributed dynamic modeling with topological information aggregation to address these challenges from a time series dynamics perspective. First, Dyntopo-Net employs distributed deterministic learning to model several decoupled multivariate time series subsystems. These subsystems are divided based on dependency relationships among multivariate time series variables, forming multiple topological nodes. Additionally, utilizing the dynamic pattern library obtained from modeling, real-time dynamic estimation at the node level generates category representations to achieve distributed dynamic pattern recognition. Finally, we reconsider the coupling relationships at each node to leverage the full dimensional information of multivariate time series. By utilizing topological relationships to aggregate synchronization errors at each node, we achieve effective fusion of the system’s dynamic characteristics. We conducted experimental validation on the UEA dataset. Results demonstrate that the proposed method achieves competitive performance, attaining the optimal average ranking among all comparison methods.

Objective. Rapid stratification of acute chest pain patients with non-ischemic electrocardiograms (ECGs) remains challenging. We developed and externally validated an interpretable cardiodynamicsgram (CDG)- major adverse cardiovascular events (MACE) Score to predict 30 d MACE. Approach. We proposed a three-step framework: (1) ECG dynamic analysis: using deterministic learning to model the ST-T repolarization process and derive CDG features that capture subtle repolarization abnormalities. We defined the temporal heterogeneity index and spatial heterogeneity index as quantitative CDG. (2) Ensemble model: training an XGBoost classifier on eight pre-specified variables with five-fold cross-validation. Patient-level splits were used; only the first emergency department ECG per patient entered the model. (3) Score derivation: transforming the ensemble into a sparse, globally interpretable score via SHAP-based variable contributions. To mitigate the demographic influence, age and gender were included as explicit covariates, and the study results were prespecified to be reported stratified by age and gender in both cohorts. Main results. Calibration and decision-analytic utility were assessed. Two independent cohorts ( n = 2836) were included. In Cohort-1 ( n = 2196; 23.27% MACE), the ensemble model achieved AUC 0.8441. Adding CDG dynamics to clinical variables improved discrimination compared with a clinical-only model (AUC 0.7963-0.8221). The derived CDG-MACE Score maintained discrimination (internal AUC 0.8221) and generalized well to Cohort-2 ( n = 640; 11.09% MACE; external AUC 0.8219). Using prespecified cutoffs from the training set (low ⩽ 9.52; high > 26.83), the internal low-risk group had negative predictive value (NPV) 99.22% and MACE 0.78%, while the external low-risk group achieved NPV 100%. Ablation analyses confirmed that CDG dynamics contributed independent signals beyond demographics. Significance. The CDG-MACE Score combines dynamic ECG modeling with a SHAP-linearized scoring system to achieve discrimination with global interpretability, enabling safe exclusion of low-risk patients without typical ischemic ECG changes. External validation suggests robustness and clinical utility; additional multicenter prospective studies and fairness monitoring are warranted.

In this article, a sampled-data fault identification (SDFI) scheme for nonlinear uncertain systems is proposed based on the deterministic learning approach, and the learning performance of the SDFI algorithm is analyzed. First, a learning-based estimator is designed, and a sampled-data (SD) linear time-varying (LTV) system is employed to describe the learning systems. Second, by constructing a special time-varying symmetric positive definite matrix, the exponential convergence property of the SD LTV system is derived. Third, explicit formulas for learning accuracy are established, which illuminate the relation between the learning performance, persistent excitation (PE) level of neural networks (NNs), and the parameters of the learning system. The main advantage of the theoretical results given in this article is that the parameters in learning accuracy formulas can be derived from measurable signals, allowing for effective evaluation of the learning performance of the SDFI scheme in practical applications. Simulation studies of a robot manipulator and a compressor system are included to show the effectiveness of the proposed method.