Emergent Charging Coordination in Electric Delivery Fleets
研究解决了电动配送车队中班次充电协调问题，通过使用本地控制的学习代理方法，无需中央调度即可实现高效协调。
0 citationsRead paper
Selected work
Representative Papers
Monitoring and Prediction of Mood in Elderly People during Daily Life Activities
This study addresses the challenge of real-time monitoring of emotional states in older adults by developing an emotion prediction system based on smart wearable devices. The system captures physiological signals via a wristband and integrates ecological momentary assessment (EMA) data collected through smartphones to train machine learning models capable of automatically and accurately identifying key emotional states—such as happiness and activeness—using only wearable-derived inputs. Experimental results demonstrate that the proposed approach achieves state-of-the-art accuracy on critical affective dimensions, substantially reducing reliance on subjective self-reports. This work thus offers a practical and unobtrusive technological pathway for continuous, passive emotional monitoring in aging populations.
0 citationsRead paper
Evaluation of Deep Learning Models for LBBB Classification in ECG Signals
This study addresses the clinical challenge of accurately differentiating left bundle branch block (LBBB) from strict LBBB (sLBBB) in patients undergoing cardiac resynchronization therapy (CRT) candidate selection. We propose a deep learning–based electrocardiogram (ECG) spatiotemporal feature modeling framework for three-class classification (healthy, LBBB, sLBBB). Systematic evaluation compares convolutional neural networks (CNNs), recurrent neural networks (RNNs), Transformers, and hybrid architectures; our novel design integrates multi-scale temporal modeling with channel-wise attention to enhance discriminative capability for pathological ECG morphologies. On public ECG datasets, the optimal model achieves 92.3% accuracy and a macro-F1 score of 0.94—substantially outperforming conventional metrics (e.g., QRS duration) and baseline models. The approach delivers clinically interpretable, automated sLBBB subtyping, offering a practical, objective tool to improve CRT patient selection efficiency and decision consistency.
0 citationsRead paper
Adaptive Gaussian Mixture Models-based Anomaly Detection for under-constrained Cable-Driven Parallel Robots
For under-constrained cable-driven parallel robots (CDPRs), external disturbances—such as gust winds or cable impacts—during point-to-point positioning may induce pose instability or cable tension loss, posing critical safety risks. To address this, we propose an unsupervised, real-time anomaly detection method relying solely on motor torque signals. Our core contribution is an adaptive Gaussian Mixture Model (GMM) that integrates Mahalanobis distance-based anomaly scoring with dynamically adjusted statistical thresholds, updated online via a sliding-window mechanism to enhance robustness against operational drift and environmental variability. Evaluated over 14 long-duration experimental trials, the method achieves a 100% true positive rate, an average true negative rate of 95.4%, and a mean detection latency of only 1 second—outperforming both conventional power-thresholding and non-adaptive GMM approaches.
0 citationsRead paper
Recent publications
Latest Papers
Emergent Charging Coordination in Electric Delivery Fleets
研究解决了电动配送车队中班次充电协调问题，通过使用本地控制的学习代理方法，无需中央调度即可实现高效协调。
0 citationsRead paper
Monitoring and Prediction of Mood in Elderly People during Daily Life Activities
This study addresses the challenge of real-time monitoring of emotional states in older adults by developing an emotion prediction system based on smart wearable devices. The system captures physiological signals via a wristband and integrates ecological momentary assessment (EMA) data collected through smartphones to train machine learning models capable of automatically and accurately identifying key emotional states—such as happiness and activeness—using only wearable-derived inputs. Experimental results demonstrate that the proposed approach achieves state-of-the-art accuracy on critical affective dimensions, substantially reducing reliance on subjective self-reports. This work thus offers a practical and unobtrusive technological pathway for continuous, passive emotional monitoring in aging populations.
0 citationsRead paper
Evaluation of Deep Learning Models for LBBB Classification in ECG Signals
This study addresses the clinical challenge of accurately differentiating left bundle branch block (LBBB) from strict LBBB (sLBBB) in patients undergoing cardiac resynchronization therapy (CRT) candidate selection. We propose a deep learning–based electrocardiogram (ECG) spatiotemporal feature modeling framework for three-class classification (healthy, LBBB, sLBBB). Systematic evaluation compares convolutional neural networks (CNNs), recurrent neural networks (RNNs), Transformers, and hybrid architectures; our novel design integrates multi-scale temporal modeling with channel-wise attention to enhance discriminative capability for pathological ECG morphologies. On public ECG datasets, the optimal model achieves 92.3% accuracy and a macro-F1 score of 0.94—substantially outperforming conventional metrics (e.g., QRS duration) and baseline models. The approach delivers clinically interpretable, automated sLBBB subtyping, offering a practical, objective tool to improve CRT patient selection efficiency and decision consistency.
0 citationsRead paper
Adaptive Gaussian Mixture Models-based Anomaly Detection for under-constrained Cable-Driven Parallel Robots
For under-constrained cable-driven parallel robots (CDPRs), external disturbances—such as gust winds or cable impacts—during point-to-point positioning may induce pose instability or cable tension loss, posing critical safety risks. To address this, we propose an unsupervised, real-time anomaly detection method relying solely on motor torque signals. Our core contribution is an adaptive Gaussian Mixture Model (GMM) that integrates Mahalanobis distance-based anomaly scoring with dynamically adjusted statistical thresholds, updated online via a sliding-window mechanism to enhance robustness against operational drift and environmental variability. Evaluated over 14 long-duration experimental trials, the method achieves a 100% true positive rate, an average true negative rate of 95.4%, and a mean detection latency of only 1 second—outperforming both conventional power-thresholding and non-adaptive GMM approaches.
0 citationsRead paper