Topic: Open-Set Classification for Human Activity Recognition Using Wrist-Worn Accelerometers• Adapted open-set classification methods, originally developed for computer vision, to the accelerometer-based human activity recognition field. This novel approach enabled the recognition of unknown human activities outside the training dataset, enhancing the system's applicability in real-world scenarios.• Designed and implemented a scalable and efficient data processing pipeline using multiprocessing and the Dask library to handle and analyse large datasets, ensuring the system's ability to scale with increasing data volumes.• Investigated and applied synthetic data generation techniques to generate diverse negative activity samples, improving the model's ability to detect and handle out-of-distribution activities.

Topic: Activity Recognition Based on the Wrist-Worn Axivity AX3• Conducted extensive research on human activity recognition using wrist-worn accelerometer data, focusing on developing robust and accurate classification models for real-world applications.• Designed and implemented a highly scalable and efficient data processing pipeline for feature extraction from large datasets. • Applied innovative time-series data transformation techniques to simulate various accelerometer positions and conditions. This approach enhanced the model's generalisation ability.• Developed advanced classification models to accurately classify distinct human activities. These models demonstrated significant improvements over baseline performance, with accuracy gains ranging from 10% to 25% on unseen data.• Conducted a comprehensive comparative analysis of training datasets, simulation techniques, and models to identify potential areas for future research and opportunities for further improvement. This analysis provided valuable insights into the strengths and limitations of various approaches, guiding the direction of ongoing research efforts.

• Developed an intelligent anomaly detection system for shotcrete in tunneling, leveraging unlabeled time-series data from over 7 sensors. Achieved a true positive rate exceeding 90% and a false positive rate below 5%, effectively identifying previously undetected issues and reducing material waste.• Implemented a shotcrete operational efficiency evaluation model for productivity estimation through cross-functional team collaboration.• Presented technical concepts, solution capabilities, and analytical findings to internal and external stakeholders, utilizing advanced data visualization tools.• Conducted research and built a pipeline for real-time problematic spray forecasting.

• Optimized the sentiment-driven factor for risk measurement in small-scale trading.• Developed an automated trading system, including a financial news crawler, a multi-factor trading model, and a MySQL backend.• Earned approximately 60% cumulative returns from mid-2017 to mid-2019.

• Engineered a data collection pipeline to scrape over 30,000 financial news, reports, and social media entries.• Developed a sentiment-driven factor using machine learning for quant trading. • Collaborated with the internal strategy team to implement and backtest trading strategies.

• Developed a collaborative control system based on OpenCV for high-speed motion tracking using 32 pan-tilt gimbal IP cameras.

• Mentored laboratory sessions for the Operating Systems course.• Graded students’ assignments in the Embedded System and Microcomputer Principle course.