Selected work

Representative Papers

Adaptive Double-Booking Strategy for Outpatient Scheduling Using Multi-Objective Reinforcement Learning

Mar 07, 2026

This study addresses the significant inefficiencies and inequities in outpatient care caused by patient no-shows, which conventional fixed double-booking strategies fail to mitigate due to their inability to adapt to dynamic environments and individual heterogeneity. To overcome these limitations, we propose an adaptive double-booking framework grounded in multi-objective reinforcement learning. The approach integrates a multi-head attention-based soft random forest to predict individual no-show risk, which is then embedded into the state representation of a Markov decision process. We further design a multi-policy proximal policy optimization algorithm augmented with a KL divergence–based τ-rule to enable selective knowledge transfer across policies, enhancing both convergence and solution diversity. Additionally, SHAP values are employed to improve the interpretability of scheduling decisions. Experimental results demonstrate that our method substantially outperforms traditional heuristic strategies in alleviating clinic congestion and mitigating the adverse effects of no-shows.

0 citationsRead paper

Federated Anomaly Detection and Mitigation for EV Charging Forecasting Under Cyberattacks

Nov 22, 2025

To address the dual challenges of degraded load forecasting accuracy and data privacy leakage in electric vehicle (EV) charging infrastructure under cyberattacks, this paper proposes a synergistic framework integrating anomaly detection, attack mitigation, and federated learning. The method innovatively combines a distributed LSTM autoencoder for anomaly detection, interpolation-driven anomaly correction, and a privacy-preserving federated LSTM network—enabling decentralized collaborative modeling without sharing raw data. Evaluated on real-world DDoS attack data, the framework achieves a 15.2% improvement in R² score, restores 47.9% of post-attack forecasting performance, attains 91.3% anomaly detection accuracy, and maintains a low false positive rate of 1.21%. These results demonstrate substantial gains in forecasting robustness and system resilience against adversarial cyber threats.

0 citationsRead paper

Uncertainty-Aware Federated Learning for Cyber-Resilient Microgrid Energy Management

Nov 22, 2025

To address the joint challenge of ensuring economic efficiency and operational reliability in microgrids under false data injection attacks, photovoltaic (PV) forecasting uncertainty, and anomalous measurements, this paper proposes a robust, uncertainty-aware energy management framework integrating federated learning. The method features a two-stage cascaded attack detection mechanism leveraging autoencoder reconstruction error and quantified prediction uncertainty; privacy-preserving distributed PV power forecasting via federated LSTM; and a two-stage robust optimal scheduling scheme incorporating multi-signal fusion analysis. Experimental results demonstrate that under severe attacks, the framework achieves a 93.7% recovery rate in forecasting accuracy, reduces false alarm rate by 58%, lowers operational cost by 5%, and mitigates economic losses by 34.7%. These outcomes significantly enhance system resilience and the synergistic balance between security and economic performance.

0 citationsRead paper

Satellite Image Utilization for Dehazing with Swin Transformer-Hybrid U-Net and Watershed loss

Aug 31, 2025

Atmospheric haze severely degrades the clarity and information fidelity of satellite imagery, hindering accurate remote sensing analysis. To address this, we propose SUFERNOBWA, a lightweight haze-removal framework integrating Swin Transformer and U-Net. Methodologically, it incorporates a SwinRRDB module to jointly model global contextual dependencies and local fine-grained structures, and employs a composite loss function comprising L2 loss, guided loss, and a novel watershed-aware loss—enhancing edge sharpness and structural preservation. Key innovations include: (1) efficient heterogeneous fusion of Swin Transformer and U-Net; (2) a watershed-perceptive loss specifically designed for remote sensing image characteristics; and (3) a multi-scale feature adaptive fusion strategy. Evaluated on RICE and SateHaze1K benchmarks, SUFERNOBWA achieves 33.24 dB PSNR and 0.967 SSIM, outperforming state-of-the-art methods across all metrics.

0 citationsRead paper

Surface Defect Detection with Gabor Filter Using Reconstruction-Based Blurring U-Net-ViT

Aug 31, 2025

To address the challenges of strong background interference and weak defect features in texture surface defect detection, this paper proposes a Gabor-enhanced hybrid U-Net–ViT model. It leverages Gabor filtering to pre-extract orientation- and frequency-sensitive texture priors, synergistically integrating U-Net’s local detail reconstruction capability with ViT’s global contextual modeling. A Gaussian-weighted Dice loss is introduced to improve segmentation boundary accuracy, and a salt-and-pepper masking strategy is designed to enhance edge perception of defects. Adaptive optimization of Gabor parameters and multi-scale feature fusion further boost robustness in localizing and segmenting minute defects. Evaluated on MVTec-AD, Surface Crack Detection, and Marble Surface Anomaly Dataset, the model achieves an average AUC of 0.939. Ablation studies confirm statistically significant contributions of each component to overall performance.

0 citationsRead paper

Recent publications

Latest Papers

Adaptive Double-Booking Strategy for Outpatient Scheduling Using Multi-Objective Reinforcement Learning

Mar 07, 2026

This study addresses the significant inefficiencies and inequities in outpatient care caused by patient no-shows, which conventional fixed double-booking strategies fail to mitigate due to their inability to adapt to dynamic environments and individual heterogeneity. To overcome these limitations, we propose an adaptive double-booking framework grounded in multi-objective reinforcement learning. The approach integrates a multi-head attention-based soft random forest to predict individual no-show risk, which is then embedded into the state representation of a Markov decision process. We further design a multi-policy proximal policy optimization algorithm augmented with a KL divergence–based τ-rule to enable selective knowledge transfer across policies, enhancing both convergence and solution diversity. Additionally, SHAP values are employed to improve the interpretability of scheduling decisions. Experimental results demonstrate that our method substantially outperforms traditional heuristic strategies in alleviating clinic congestion and mitigating the adverse effects of no-shows.

0 citationsRead paper

Federated Anomaly Detection and Mitigation for EV Charging Forecasting Under Cyberattacks

Nov 22, 2025

To address the dual challenges of degraded load forecasting accuracy and data privacy leakage in electric vehicle (EV) charging infrastructure under cyberattacks, this paper proposes a synergistic framework integrating anomaly detection, attack mitigation, and federated learning. The method innovatively combines a distributed LSTM autoencoder for anomaly detection, interpolation-driven anomaly correction, and a privacy-preserving federated LSTM network—enabling decentralized collaborative modeling without sharing raw data. Evaluated on real-world DDoS attack data, the framework achieves a 15.2% improvement in R² score, restores 47.9% of post-attack forecasting performance, attains 91.3% anomaly detection accuracy, and maintains a low false positive rate of 1.21%. These results demonstrate substantial gains in forecasting robustness and system resilience against adversarial cyber threats.

0 citationsRead paper

Uncertainty-Aware Federated Learning for Cyber-Resilient Microgrid Energy Management

Nov 22, 2025

To address the joint challenge of ensuring economic efficiency and operational reliability in microgrids under false data injection attacks, photovoltaic (PV) forecasting uncertainty, and anomalous measurements, this paper proposes a robust, uncertainty-aware energy management framework integrating federated learning. The method features a two-stage cascaded attack detection mechanism leveraging autoencoder reconstruction error and quantified prediction uncertainty; privacy-preserving distributed PV power forecasting via federated LSTM; and a two-stage robust optimal scheduling scheme incorporating multi-signal fusion analysis. Experimental results demonstrate that under severe attacks, the framework achieves a 93.7% recovery rate in forecasting accuracy, reduces false alarm rate by 58%, lowers operational cost by 5%, and mitigates economic losses by 34.7%. These outcomes significantly enhance system resilience and the synergistic balance between security and economic performance.

0 citationsRead paper

Satellite Image Utilization for Dehazing with Swin Transformer-Hybrid U-Net and Watershed loss

Aug 31, 2025

Atmospheric haze severely degrades the clarity and information fidelity of satellite imagery, hindering accurate remote sensing analysis. To address this, we propose SUFERNOBWA, a lightweight haze-removal framework integrating Swin Transformer and U-Net. Methodologically, it incorporates a SwinRRDB module to jointly model global contextual dependencies and local fine-grained structures, and employs a composite loss function comprising L2 loss, guided loss, and a novel watershed-aware loss—enhancing edge sharpness and structural preservation. Key innovations include: (1) efficient heterogeneous fusion of Swin Transformer and U-Net; (2) a watershed-perceptive loss specifically designed for remote sensing image characteristics; and (3) a multi-scale feature adaptive fusion strategy. Evaluated on RICE and SateHaze1K benchmarks, SUFERNOBWA achieves 33.24 dB PSNR and 0.967 SSIM, outperforming state-of-the-art methods across all metrics.

0 citationsRead paper

Surface Defect Detection with Gabor Filter Using Reconstruction-Based Blurring U-Net-ViT

Aug 31, 2025

To address the challenges of strong background interference and weak defect features in texture surface defect detection, this paper proposes a Gabor-enhanced hybrid U-Net–ViT model. It leverages Gabor filtering to pre-extract orientation- and frequency-sensitive texture priors, synergistically integrating U-Net’s local detail reconstruction capability with ViT’s global contextual modeling. A Gaussian-weighted Dice loss is introduced to improve segmentation boundary accuracy, and a salt-and-pepper masking strategy is designed to enhance edge perception of defects. Adaptive optimization of Gabor parameters and multi-scale feature fusion further boost robustness in localizing and segmenting minute defects. Evaluated on MVTec-AD, Surface Crack Detection, and Marble Surface Anomaly Dataset, the model achieves an average AUC of 0.939. Ablation studies confirm statistically significant contributions of each component to overall performance.

0 citationsRead paper