Academic homepage

Reliability-aware multimodal intelligence for physical systems

I work on learning and perception problems where sensors are noisy, labels are limited, modalities may be missing, and long-term estimation can accumulate drift. My current work connects mechanical engineering, time-series AI, biomedical sensing, and robotic SLAM.

Robust multimodal learning Robotic perception LiDAR-visual-inertial SLAM Biomedical signal processing Reliability-aware AI

About me

I am an undergraduate student in Mechanical Design, Manufacturing and Automation at Central South University. My research interests are centered on physical-system AI: robust multimodal fusion, sensor reliability, robotic perception, and physiological time-series analysis.

I prefer projects that can be tested through reproducible code and clear evaluation protocols. Recent examples include bearing fault diagnosis under limited samples and strong noise, long-term LiDAR-visual-inertial SLAM with loop closure, and cuffless blood-pressure estimation from synchronized PPG and ECG signals.

  • Academic standing: weighted average 90.63/100; ranked 5/60 in major.
  • Accepted paper: SC-FAST-LIVO2, AIRC 2026 full paper and oral presentation.
  • Project role: PI of a National Undergraduate Innovation Training Program project.
  • Engineering basis: CAD, ANSYS, mechanics, ROS, sensing, and control practice.

Portfolio

Robust multimodal learning

MHFL-MCA

A multimodal heterogeneous feature-learning framework for bearing fault diagnosis under limited labels, load shift, and strong noise.

  • Modality-specific convolutional encoders
  • Bidirectional modality-cross-attention
  • Adaptive modality re-weighting
Robotic perception

SC-FAST-LIVO2

A loop-closure and pose-graph optimization extension for FAST-LIVO2, targeting long-term consistency in LiDAR-visual-inertial SLAM.

  • Scan Context loop detection
  • SE(3) pose graph optimization
  • Drift reduction in large-scale trajectories
Biomedical AI

SAQM-MedFuse

A quality-driven multimodal framework for cuffless blood-pressure estimation and risk stratification from synchronized physiological signals.

  • PPG + ECG multimodal fusion
  • Signal quality and missing-modality awareness
  • Uncertainty and conformal reliability analysis

Selected papers and manuscripts

SC-FAST-LIVO2: A Cross-Modal Loop Closure Framework for Long-Term SLAM with Reduced Drift

International Conference on Artificial Intelligence, Robotics, and Control (AIRC 2026)
Accepted · Full paper · Oral
Abstract excerpt. SC-FAST-LIVO2 augments FAST-LIVO2 with a Scan Context based loop-closure module and an SE(3) pose-graph optimization back end. The system extracts keyframes, searches loop candidates, adds geometric constraints, and feeds optimized corrections back to the map so that accumulated drift can be reduced during long-term mapping.
GitHub DeepWiki

MHFL-MCA: Multimodal Heterogeneous Feature Learning with Modality-Cross-Attention for Robust Fault Diagnosis with Limited Samples

Manuscript submitted to Advanced Engineering Informatics
Submitted manuscript
Abstract excerpt. MHFL-MCA studies rotating-machinery fault diagnosis when labeled samples are scarce and measurements are noisy. The method uses heterogeneous modality-specific encoders, bidirectional cross-attention, and adaptive re-weighting to learn complementary information from multimodal signals and improve low-shot, cross-load, and noise robustness.
GitHub DeepWiki

SAQM-MedFuse: Safety-aware Quality-driven Multimodal Fusion for Cuffless Blood Pressure Estimation and Risk Stratification

Manuscript submitted to Biomedical Signal Processing and Control
Submitted manuscript
Abstract excerpt. SAQM-MedFuse treats cuffless blood-pressure estimation as a reliability-sensitive multimodal problem. The framework combines PPG and ECG through signal-quality estimation, sparse quality-conditioned expert fusion, uncertainty-aware credibility aggregation, subject-adaptive personalization, and conformal reliability analysis for BP estimation and risk stratification.
GitHub DeepWiki

Open source and reproducibility

SAQM-MedFuse

Code and reproducibility resources for quality-aware biomedical multimodal fusion.

GitHub DeepWiki

MHFL-MCA

Code, datasets, and result records for robust multimodal fault diagnosis experiments.

GitHub DeepWiki

SC-FAST-LIVO2

Project documentation for Scan Context enhanced loop closure and pose graph optimization.

GitHub DeepWiki

Contact

Email 2617376@dundee.ac.uk
GitHub github.com/yazyeah