About Me
Hey! I’m Xinyao (pronounced like Cynthia), a Ph.D. candidate mentored by Dr. Sara Behdad @UF. Alongside, I’m enriching my journey with a Master’s in ECE, specializing in signals and systems.
My research leverages machine learning to enhance safety and efficiency in human-robot collaboration scenarios. Research areas of interest to me include human intent and activity recognition, early prediction, generative learning.
Passionate about applying theory to tackle real problems, I’m on my way from researcher to engineer. The visual below maps my knowledge base, which I’ll keep fresh and updated. :)
Feel free to discuss any innovative ideas and collaborations. Let’s connect via email!
Projects
Published in the Journal of Computing and Information Science in Engineering
- Collected arm motion data in a human-robot collaboration experiment, and built Transformer and Bi-LSTM models in TensorFlow to classify human intents.
- Evaluated model performance by inputting data lengths from 20% to 100% in 20% intervals, visualizing the results with boxplots and heatmaps.
- Converted motion data into Euclidean distance features for a Hidden Markov Model (HMM) and an Autoregressive HMM to identify intent state transitions from uncertain to certain.
- Achieved the early prediction goal by predicting intent before state transitions, resulting in a 2% improvement for the Transformer model and a 6% increase for the Bi-LSTM model.
Published in the IEEE Transactions on Industrial Informatics
- Developed an unsupervised learning framework in TensorFlow that combines a sequential variational auto-encoder (Seq-VAE) for latent representation learning, a HMM for continuous action segmentation, and a nonlinear support vector machine kernel for recognition validation.
- Tested the framework on video-based disassembly tasks, achieving an average recognition accuracy of 91.52%, surpassing both autoencoder and principal component analysis benchmarks.
- Assessed the Seq-VAE’s spatio-temporal feature extraction capabilities through image reconstruction, latent traversal exploration, and t-SNE visualization.
- Analyzed the HMM’s discrete state segmentation efficacy using a hierarchical HMM tree, detailing the probability of each hidden state across different activity classes.
- Confirmed the framework’s robustness with unseen video data, effectively distinguishing between unexpected and predefined actions.
Published in the Journal of Manufacturing Science and Engineering
- Optimized the YOLOv4 algorithm by adjusting layer sizes and connections within the path aggregation network, improving screw detection accuracy in used electronics from 92% to 94%.
- Developed an EfficientNetv2 algorithm to classify three types of screws, facilitating tool recommendation for disassembly. This was achieved by expanding the class range and augmenting the dataset, which increased classification accuracy from 98% to 99%.
- Expanded the dataset using data augmentation techniques such as hue, exposure, saturation, and mosaic adjustments.
Awards
- Machine Learning (ECE) graduate certificate, University of Florida, 2023
- Certificates of Outstanding Merit, University of Florida, 2023
- Florida AWMA Scholarship, Florida Section of the Air and Waste Management Association, 2023
- Don Maurer Memorial Scholarship Award, University of Florida, 2023
Services
- Reviewer for the Journal of Mechanical Design (JMD), International Conference on Machine Learning (ICML) Workshop, and the International Design Engineering Technical Conferences & Computers and Information in Engineering Conference (IDETC)
- Member of the American Society of Mechanical Engineers (ASME) and the UF Society of Women Engineers (SWE)
More about me
-
When I'm not hitting work, I love to travel and take photos. Big on food and pets!