Selected Projects项目经历

MSc Dissertation · Imperial College London硕士毕业论文 · 帝国理工学院

Designed a Transformer-based diffusion framework for personalised trajectory prediction in complex car-following scenarios.面向复杂跟车场景，设计并实现 Transformer + Diffusion 轨迹预测框架，用于刻画车辆交互和未来驾驶行为不确定性。

• Built modelling datasets from highD / nuPlan, including trajectory cleaning, vehicle-pair matching and dynamic feature extraction.基于 highD / nuPlan 数据构建车辆交互建模数据集，完成轨迹清洗、前后车匹配、时间切片和动态特征提取。
• Modelled long-horizon leader-follower interactions and future-driving uncertainty through Transformer + Diffusion architecture.利用 Transformer 捕捉前后车长期交互依赖，并通过扩散模型生成多模态未来轨迹。
• Ablations showed map-context features reduced validation displacement-error loss by over 20%.引入车道线、道路边界、交通规则等地图上下文后，验证集位移误差损失降低 20% 以上。

Project Leader · In collaboration with World Bank Group项目组长 · 与世界银行 (World Bank Group) 合作

Led a team converting more than 40 million truck GPS records into interpretable event-state representations for logistics analysis.带队处理 4000 万条以上印度卡车 GPS 轨迹，将原始记录转化为可解释的事件-状态数据，用于物流效率分析。

• Coordinated weekly reviews with the World Bank Group and delivered a final A-rated project.对接 World Bank Group 项目团队，负责需求沟通、进度汇报、任务分工与成果整合，最终项目评价为 A。
• Combined speed, dwell time, location change, ignition state and logistics-domain rules to classify truck behaviour.结合速度、停留时长、位置变化、点火状态和物流场景规则，识别行驶、停止、怠速、装卸货和异常停留等状态。
• Built GIS visualisations for trajectory monitoring, stop hotspots and abnormal-operation analysis.搭建 GIS 可视化工具，展示车辆轨迹、状态分布、停留热点和异常运行情况，支持监控与调度分析。

Project Leader · Imperial College London项目组长 · 帝国理工学院

Built interpretable predictive models for vehicle CO2 emissions and translated model evidence into policy recommendations.建立可解释的 CO2 排放预测模型，并基于 SHAP 分析把模型证据转化为可执行的政策建议。

• SHAP identified vehicle type, mileage, engine efficiency as key drivers.SHAP 分析识别出车型、行驶里程、发动机效率为关键驱动因素。
• Connected explainability results to emission taxation, EV subsidy and high-emission-vehicle regulation strategies.为碳税、电车补贴和高排放车辆监管提供循证参考，体现可信 AI 在政策分析中的作用。

Undergraduate Final Year Project · First-Class本科毕业设计 · 一等

Applied persistent homology to spectrograms to build robust audio fingerprints under pitch shifts and rhythm perturbations.用持续同调从频谱图中提取拓扑特征，构建对音高变化和节奏扰动更鲁棒的音频指纹。

• +30% accuracy vs. Shazam on pitch-shifted audio.在变调音频下相较 Shazam 精度高 30% 以上。
• Converted geometric/topological structure into algebraic invariants.将频谱图的几何 / 拓扑结构转化为可量化的代数不变量。
• Connected abstract mathematical structure with practical media-recognition algorithms.把抽象数学理论转化为实际媒体识别算法组件。