Structure-Aware Unsupervised Anomaly Detection for Spacecraft Telemetry with Adaptive EVT Thresholding
本文提出了一种无监督的航天器遥测异常检测框架，通过增量式月度重训练、统计模型选择和自适应极值理论阈值控制，解决了实际中缺乏标注数据的问题。
0 citationsRead paper
Selected work
Representative Papers
AI-Enabled Capabilities to Facilitate Next-Generation Rover Surface Operations
Current planetary rovers operate at speeds of only ~10 cm/s, severely limiting deep-space exploration efficiency. To address this, we propose an intelligent high-speed autonomous navigation architecture that innovatively integrates FASTNAV for long-range obstacle detection, the CISRU multi-robot collaborative framework, and ViBEKO/AIAXR deep learning–based terrain classification. Leveraging computer vision, deep learning, and multi-agent cooperative control, the system achieves centimeter-scale obstacle identification and sub-meter semantic terrain classification in Mars-analog environments. Field validation attains Technology Readiness Level (TRL) 4, demonstrating a rover speed increase to 1.0 m/s while significantly enhancing operational safety and mission execution efficiency. This work overcomes key bottlenecks in traditional visual perception and distributed control, providing critical technological foundations for future high-speed autonomous exploration on Mars.
0 citationsRead paper
Reinforcement Learning for Automated Cybersecurity Penetration Testing
To address the low automation level, high maintenance cost, and inefficient path exploration in web application penetration testing, this paper proposes an automated security testing framework integrating reinforcement learning (RL) with geometric deep learning. Methodologically, we design a graph neural network (GNN)-based agent that perceives web topology and leverages geometric priors to compress the action space, enabling joint optimization of tool selection and penetration path planning. The agent is co-trained and validated in both simulated environments and real vulnerable web applications (e.g., WebGoat). To the best of our knowledge, this is the first work embedding GNNs into an RL framework for penetration path planning. Experimental results demonstrate a 32.7% increase in vulnerability detection rate and a 41.5% reduction in average detection steps, significantly improving both testing efficiency and maintainability.
0 citationsRead paper
Automated Detection of Clinical Entities in Lung and Breast Cancer Reports Using NLP Techniques
This study addresses the time-consuming, error-prone manual extraction of key clinical information from lung and breast cancer reports, which hinders the full utilization of healthcare data. To this end, we propose an end-to-end clinical natural language processing (NLP) system. Methodologically, we introduce the first joint application of the uQuery context-aware parsing engine and a fine-tuned RoBERTa model—bsc-bio-ehr-en3—on Spanish electronic health records (EHRs), enabling negation detection, temporal modeling, and patient-level semantic association. The system integrates named entity recognition (NER), standardized mapping to SNOMED CT and OMOP ontologies, and annotation via the Doccano platform to generate structured outputs. Evaluated on 600 real-world clinical reports, our approach achieves F1-scores exceeding 92% for both MET (metastasis) and PAT (pathology) critical entities, demonstrating strong cross-cancer generalizability and robustness in automated clinical information structuring.
0 citationsRead paper
Recent publications
Latest Papers
Structure-Aware Unsupervised Anomaly Detection for Spacecraft Telemetry with Adaptive EVT Thresholding
本文提出了一种无监督的航天器遥测异常检测框架，通过增量式月度重训练、统计模型选择和自适应极值理论阈值控制，解决了实际中缺乏标注数据的问题。
0 citationsRead paper
AI-Enabled Capabilities to Facilitate Next-Generation Rover Surface Operations
Current planetary rovers operate at speeds of only ~10 cm/s, severely limiting deep-space exploration efficiency. To address this, we propose an intelligent high-speed autonomous navigation architecture that innovatively integrates FASTNAV for long-range obstacle detection, the CISRU multi-robot collaborative framework, and ViBEKO/AIAXR deep learning–based terrain classification. Leveraging computer vision, deep learning, and multi-agent cooperative control, the system achieves centimeter-scale obstacle identification and sub-meter semantic terrain classification in Mars-analog environments. Field validation attains Technology Readiness Level (TRL) 4, demonstrating a rover speed increase to 1.0 m/s while significantly enhancing operational safety and mission execution efficiency. This work overcomes key bottlenecks in traditional visual perception and distributed control, providing critical technological foundations for future high-speed autonomous exploration on Mars.
0 citationsRead paper
Reinforcement Learning for Automated Cybersecurity Penetration Testing
To address the low automation level, high maintenance cost, and inefficient path exploration in web application penetration testing, this paper proposes an automated security testing framework integrating reinforcement learning (RL) with geometric deep learning. Methodologically, we design a graph neural network (GNN)-based agent that perceives web topology and leverages geometric priors to compress the action space, enabling joint optimization of tool selection and penetration path planning. The agent is co-trained and validated in both simulated environments and real vulnerable web applications (e.g., WebGoat). To the best of our knowledge, this is the first work embedding GNNs into an RL framework for penetration path planning. Experimental results demonstrate a 32.7% increase in vulnerability detection rate and a 41.5% reduction in average detection steps, significantly improving both testing efficiency and maintainability.
0 citationsRead paper
Automated Detection of Clinical Entities in Lung and Breast Cancer Reports Using NLP Techniques
This study addresses the time-consuming, error-prone manual extraction of key clinical information from lung and breast cancer reports, which hinders the full utilization of healthcare data. To this end, we propose an end-to-end clinical natural language processing (NLP) system. Methodologically, we introduce the first joint application of the uQuery context-aware parsing engine and a fine-tuned RoBERTa model—bsc-bio-ehr-en3—on Spanish electronic health records (EHRs), enabling negation detection, temporal modeling, and patient-level semantic association. The system integrates named entity recognition (NER), standardized mapping to SNOMED CT and OMOP ontologies, and annotation via the Doccano platform to generate structured outputs. Evaluated on 600 real-world clinical reports, our approach achieves F1-scores exceeding 92% for both MET (metastasis) and PAT (pathology) critical entities, demonstrating strong cross-cancer generalizability and robustness in automated clinical information structuring.
0 citationsRead paper