bootstrap navbar

ARNOLD YEUNG

machine learning.  mechatronics.

ABOUT ME. 

University of British Columbia, 2016
Bachelor of Applied Science with Distinction,
Mechanical Engineering (Mechatronics Option), Minor in Commerce

Based in Vancouver, Canada and Hong Kong, PRC
Interests include machine learning applications
and mechatronics systems
Experience in both industry and academia

Control Systems Engineer, ASM Pacific Technology Ltd.
Data Scientist and Administrative Manager, EEGlewave Inc.
Engineering Research Assistant, UBC Faculty of Medicine
Co-op Research Engineer, Cadex Electronics Inc.
Mechanical Engineer Intern, L.B. Foster Rail Technologies Corp.

Manufacturing and Prototyping

CSA Drafting Standards
Machine Shop
Hand and Power Tools
Electrical Equipment
Material Testing
Breadboard Prototyping
Rapid Prototyping

Engineering Software

MATLAB, Simulink
AutoCAD
SolidWorks
Siemens NX
Microsoft Visual Studio
Microsoft Office Suite (Word, PowerPoint, Excel, Project)

Programming

C, C++, C#
Python
Processing
Visual Basic (VBA)
VHDL (FPGA)
Microcontroller (Embedded C)
Machine Learning
Multi-Threaded Programming
Data Structures, Algorithms
Unified Modeling Language (UML)

Design and Project Management

Engineering Design Process
Risk Management
Project Scheduling
Technical Reports
Peer-Reviewed Publications
Patents
Business Plans
Market Research

PROJECTS.

algorithmic trading portfolio manager

Portfolio management of stocks typically requires observations of various statistical performances of the market in order to make buying/selling decisions. To maximize gain, statistical analysis and optimization may be used on existing data so that machine learning techniques may provide rule-based decision making.

Goal: Create a set of integrated applications using the Pandas, NumPy, and Scikit-Learn libraries for automated statistical analysis and prediction of stock portfolio performances.

The set of applications currently includes:
1. Stock portfolio statistical assessment and performance optimization
2. Market and portfolio growth simulation
3. Random Forest Regressor index return predictor
4. Q-Learning decision maker

This project is currently ongoing. Stay tuned for more updates.

Python, Machine Learning, Reinforcement Learning, Financial Analysis, Data Analysis, Risk Management, Statistics and Probability

EEG concussion assessment algorithm

EEG power spectral features and wavelet features have independently shown changes in brain activity due to concussions.

Goal: Identify optimal features for concussion classification based on a training set using different machine learning techniques such as SVM and Random Forest. MATLAB scripts are written to extract EEG features based on literature. These features are then trained and tested through a machine learning pipeline to determine its accuracy.


Our results show that the power spectral feature set had the higher sensitivity and the wavelet feature set had the higher specificity in classifying concussed subjects. When both feature sets were combined, the overall accuracy improved by 5% due to higher sensitivity, suggesting the relevance of both feature sets.

Presented at the Engineering in Medicine and Biology Society (EMBC), 2015 37th Annual International Conference of the IEEE.

Machine Learning, Signal Processing, MATLAB, Data Analysis, Technical Reports

3D environment laser scanner

Goal: Design a stationary tool capable of automated scanning of the surrounding environment, outputting a 3D image. The proposed design is a derivative of the 3D Environment Laser Scanner designed by Callan Mackay*.

Designed for dark environments, a line laser emits a red light at the surrounding environment. The webcam captures the light as the line laser and the webcam rotates. Each image frame is processed to extract the red line, creating a 3D point cloud when concatenated together.

The final design contrasts from Mackay’s design by providing an addition degree of freedom, allowing a greater vertical range of scan of the environment.

Visual C#, Processing, Actuators, SolidWorks, Rapid Prototyping, Machine Shop, Electrical Equipment, Breadboard Prototyping, Microcontroller (Embedded C), Microsoft Visual Studio, Technical Reports 

EZT teabag packager



Tea retailers express a need for an automated method of packaging specialty tea leaves in-store from bulk purchases. The EZT Teabag Packager aims to meet this demand.

Goal: Design a teabag packaging machine that may be used in a typical tea retailer environment. The EZT Teabag Packager is designed to meet the needs and requirements of both tea retailers and tea drinkers.

An alpha prototype proving the feasibility of the concept was designed and manufactured for presentation to tea retailers. Many tea retailers have shown interest in the purchase of the final product and 3 have produced Letters of Purchase Intent.

SolidWorks, Actuators, Machine Shop, Rapid Prototyping, CSA Drafting Standards, Engineering Design Process, Failure Modes and Effects Analysis (DFMEA), Gantt Chart, Prototype Testing, Technical Reports 

C++ elevator simulator

Goal: Design a real-time terminal-based program which simulates the movements of a 10-floor 2-elevator system. The Elevator Simulator operates and receives commands exactly like a real-world elevator system, maximizing efficiency and allowing passengers to reach their desired floor as fast as possible.

The program takes commands from either inside the elevator or outside the elevator. Additional commands such as generating and removing faults in specific elevators and restarting the entire simulation are also available.

Commands that are sent while the elevators are in operation are stored. An algorithm prioritizes the commands based on efficiency to ensure that the maximum number of passengers are served within the shortest time

C++, Object-Oriented Programming, Multi-Threaded Programming, Microsoft Visual Studio.

machine learning number recognition

Two machine learning algorithms were designed and developed from theory to classify images of handwritten digits. This project is a derivative of Exercises 3 and 4 from the Coursera online course, Machine Learning by Andrew Ng (Stanford University, 2013).

Goal: Develop MATLAB functions for supervised logistic regression and 3-layer neural network with regularization. A dataset of pixel information of handwritten digit images is then inputted into these functions to train the classifiers. The trained classifiers are then applied to a validation dataset to estimated accuracies of their classifications.

In total, 4000 training examples and 1000 test sets were used to obtain an accuracy of 90% for logistic regression and 92% for neural network. All developed scripts and functions are derived from the mathematical theory of the classifiers and do not rely on any existing machine learning libraries.

Machine Learning, MATLAB

trajectory generator and CNC controller design

Goal: Develop a set of instructions to generate a trajectory for a 2-axis CNC machine and design an appropriate controller based on the properties of the machine to maximize accuracy and minimize unstability in trajectory.

Different controllers are tested with the CNC machine and simulated using MATLAB, Simulink and Virtual CNC and compared to ensure that error is minimized. This project is composed of 3 mini-projects:

1. Modeling and Identification of Motion Control Mechanism
2. Digital Control of Motion Actuators
3. Simulation of Contouring Performance in Coordinated Two Axis Motion

Control Theory, MATLAB, Simulink, Technical Reports 

motion-controlled beat sequencer

Goal: Develop a software which reads real-time data from a 3-axis accelerometer and interprets the data into instructions. Instructions obtained from the accelerometer allow the user to hear instant playback of 1 of 6 audio files using state machines, as well as place the sounds at indicated time positions on a 2-bar beat for repetitive playback.

Innovation in music technology has led to new music production and performance tools. Many songs are now being produced through software and music performances are beginning to integrate interactions between music and the actions of performers.

The Motion-Controlled Beat Sequencer aims to combine the two. The beat sequencer translate physical movements of the user, obtained through a 3-axis accelerometer, into live music production. Furthermore, the software allows the user to organize the sounds indicated from the movements in a digital beat sequencer.

State Machines, Sensors, Visual C#, Microsoft Visual Studio

PUBLICATIONS.

Contact me for free access to articles.

Development of the Sports Organization Concussion Risk Assessment Tool (SOCRAT)

Arnold Yeung*, Vrinda Munjal*, and Naznin Virji-Babul
Brain Injury

The Sports Organization Concussion Risk Assessment Tool (SOCRAT) was developed to assist sport organizations in assessing the overall risk of concussion at a team level by identifying key risk factors. Risk factors were incorporated into the FMEA-inspired algorithm, resulting in an individual risk priority number (RPN) for each risk factor and an overall RPN that estimates the risk in the given circumstances. The SOCRAT can be used to analyse how different risk factors contribute to the overall risk of concussion. The tool may be tailored to organizations to provide: (1) an RPN for each risk factor and (2) an overall RPN that takes into account all the risk factors.

Failure Modes and Effects Analysis (FMEA), Risk Management,
Statistics and Probability, Peer-Reviewed Publications

Comparison of Foam-Based and Spring-Loaded Dry EEG Electrodes with Wet Electrodes in Resting and Moving Conditions 

Arnold Yeung, Harinath Garudadri, Carolyn Van Toen, Patrick Mercier,
Ozgur Balkan, Scott Makeig, and Naznin Virji-Babul
Engineering in Medicine and Biology Society (EMBC), 2015 37th Annual International Conference of the IEEE  

In this study, we compared the performances of foam-based and spring-loaded dry electrodes with wet electrodes under three different task conditions: resting state, walking, and cycling. Our analysis showed that signals obtained by the 2 types of dry electrodes and obtained by wet electrodes displayed high correlation for all conditions, while being prone to similar environmental and electrode-based artifacts. Overall, our results suggest that dry electrodes have a similar signal quality in comparison to wet electrodes during motion and may be more practical for use in mobile and real-time motion applications due to their convenience.

Sensors, Electrical Design, Electrical Equipment, Breadboard Prototyping,
Data Analysis, MATLAB, Peer-Reviewed Publications

CONTACT ME.

contact@arnoldyeung.com
linkedin.com/in/arnoldysyeung
github.com/arnoldysyeung