Research
I am a PhD Student in Data Science at Tsinghua Shenzhen International Graduate School (SIGS), Tsinghua University, working in Data Science and Information Technology under the supervision of Prof. Ercan Engin Kuruoglu. My research lies at the intersection of graph signal processing, robust state estimation, and statistical signal processing.
Graph Signal Processing and Time-Varying Graphs
My current PhD research focuses on time-varying graph models and graph-based adaptive learning for drone networks. I am interested in how changing graph structure can represent interactions among distributed agents and how graph-domain information can support filtering, learning, anomaly detection, and decision making in dynamic networks.
More broadly, I am interested in combining model-based signal processing with data-driven learning for networked and multi-agent systems.
Robust State Estimation
A major part of my previous and continuing work concerns robust recursive estimation when measurements contain persistent Gaussian background noise together with impulsive or heavy-tailed disturbances.
Additive Gaussian-Student’s t filtering
I developed a robust Kalman-filtering approach that explicitly models measurement noise as the sum of Gaussian and Student’s t components. This work appeared at ICASSP 2026.
Additive Gaussian-GSM filtering
I am extending this idea to a broader additive Gaussian-Gaussian scale mixture (GSM) framework. The work introduces a unified model for hybrid measurement noise and develops inference strategies for the resulting latent-scale estimation problem.
Statistical Signal Processing
My broader interests include statistical modeling, Bayesian inference, and learning for signal processing. I am especially interested in methods that remain reliable under model mismatch, non-Gaussian uncertainty, and changing network structure.
Acoustic signal processing is one application area in which I have used these ideas. During an internship at Shokz, I studied microphone-array beamforming and implemented real-time speech-enhancement pipelines in waveform and time-frequency domains. A related patent application studies lightweight real-time speech enhancement for virtual microphone arrays using sparse Bayesian direction estimation, fixed-parameter beamforming, and spectral-domain noise reduction.
Structured Analytic-Residual Correction (SARC)
SARC is a lightweight multichannel speech-enhancement method that combines a unit-response analytic estimate with spatial-residual-guided neural correction. It retains target-canceling channel differences to help a compact correction network suppress interference remaining in the analytic estimate.
Status: Submitted to ICASSP 2027.
Research Themes
- Graph signal processing and time-varying graphs
- Robust Bayesian filtering and Kalman-type estimation
- Statistical signal processing and Bayesian inference
- Heavy-tailed and hybrid noise modeling
- Distributed and multi-agent learning
- Microphone arrays, beamforming, and speech enhancement