Career Profile
Interests span over Wireless Sensing (include WiFi, Bluetooth, and mmW Radar) for Indoor Localization and Human Activity Monitoring (HealthCare or Fitness), Resource Allocation using Deep Reinforcement Learning, Cognitive Radio and GPS/GNSS techniques.
Experiences
- WiFi CSI sensing for human activity monitor and indoor localization,
- 802.11mc Round Trip Time (RTT) & Extend Kalman Filter/ Particle Filter/ PDR for indoor localization,
- UWB Radar sensing on mobile phones,
- Andriod prototype application development.
Building WiFi sensing testbed for human activity recognition and indoor localization, result in 2 conference papers, 2 journal papers submission, several patents and selected into Mitsubishi Global OpenHouse 2020 and 30-anniversary highlights. Build WiFi sensing system with CSI (intel5300/ Atheros/ nexmon/ ESP32 tools), RSSI and 11ad beamSNR routers with time synchronization and offset calibration, using OpenPose tool as labels.
- Implement algorithms AoA-ToF-Doppler profile from commercial router CSI.
- Build an indoor localization system on a robot, using SLAM coordinates as labels.
- Bluetooth 5.2 AoA testbed with antenna switching.
- Cross-Modality human activity reasoning with CSI and mmW FMCW radar.
- 5G & 60G band WiFi sensing Fusion work as journal submission
- SPAWC sensor fusion indoor localization challenge, finished 2nd place with 20cm accuracy.
Wireless Sensing and Communication using Deep Learning techniques.
- Healthcare with Radio-Frequency monitoring: apply RSSI for sub-room level localization, fall and stroke detection using IMU and CSI AoA-ToF-Doppler profile, self-learning between IMU and CSI.
- DNN-assisted CFAR: ResNet assisted Constant False-Alarm Rate signal detection
- High-Loading DoA Estimation with Machine Learning: apply machine learning to estimate direction of arrival(DoA) at vector sensor to resolve the antenna offsets, nearby angles and high-loading issues in subspace methods. apply ResNet for Modular-Order-Estimation (MoE). apply hybrid DNN-MUSIC method to locate regions first using DNN and then MUSIC for super-resolution estimation. transfer learning with different unknown polarization angles.
- Narrowband Interference Mitigation Library: implement and evaluate interference mitigation library in presence of common interference, cover filter bank, FRFT, DSSS etc. design neural network to classify interference type based on spectrum.
- SC2 DARPA - Spectrum Collaboration Challenge: Apply Reinforcement Learning on dynamic channel access, implement Markov Decision Process(MDP), Deep Q Network(DQN) to share spectrum with hopping or intermittent nodes, apply Deep Recurrent Q Network(DRQN) to deal with partial observation, learn stochastic nodes and hidden nodes,explore fast and suboptimal converge cast algorithms.
Projects
2nd place CTW indoor location competition
- Mining CSI using RNN for centimeter-level localization, reaching under 10 cm accuracy.
Qualcomm Innovation Fellowship 2nd Round
- Dynamic spectrum sharing by Distributed Reinforcement Learning on Mobile Edges.
1st Prize of Openpower hackathon ARC, Virginia Tech
- Extend indoor positioning with neural network to IBM DLL multi-CPU platform and Nvidia Jetson Nano.
Obstacle Avoid Car with DQN in Unreal Engine & Airsim
- Cars start somewhere surrounded by obstacle, plan to get out without collide eventually by Reinforcement Learning.
Implementation of a Tiny MIMO Communication System with USRPs
- Implement 2X2 MIMO system on 4 Ettus N210 devices with GNU Radio Companion and some OOT programmings, able to successfully transmit text, image and audio.
Cuisine Prediction from Recipe, A classification problem on Kaggle
- Trial with cosine similarity, PCA reduction, xgboost(extreme gradient boosting toolkit), item split, LDA and grid search method in R or python codes, and finally reach top 5% with 82% accuracy.
Publications
IEEE ioT, 2022
IEEE SPAWC, 2021
IEEE SouthEast. (2021)
IEEE GLOBECOM WS (2020)
IEEE PLAN (2020)
IEEE VTC (2020)
IEEE DySPAN (2019)
IEEE RadarCom (2019)
IEEE MILCOM (2018)