Introduction

The Computer Engineering Capstone Project requires the creation of an AR game system, where the software visualisation must synchronise with real-time sensor events to play out a shooter game.

Out of the 5 possible components, I contributed towards the Hardware component heavily. I also assisted the other members in their components to ensure smooth development in the project.

Contributions

1. Designed the prototype, circuit layout andfinal design of the laser guns, motion sensing handbands and laser targets.
   • The battery solution for each device is designed based on the nature of use; lithium ion batteries are used together with buck converters to provide a reliable power supply based on the energy consumption of the devices.
   • The laser gun incorporates a set of Lithium Ion battery pack, laser diode and IR emitting diode to ensure long-range detection and ease of aim.
   • The motion-sensing handbands incorporates the accelerometer on the MPU6050 to generate accelerometer readings. The readings are then relayed to the neural network model to predict the arm movement.
   • The laser target incorporates an IR receiving diode to receive IR emissions. The firmware of the device incorporates a state machine to handle different events based on the encoded message of the gun.
2. Assisted my teammates in other tasks, including:
   • Created a method to record accelerometer readings based on a button hold.
   • Assisted in setting up a SSH tunnel over the NUS sunfire server through the private network to bypass the firewall.
   • Assisted in setting up the MQTT broker to work with the Unity application.

Reflections

1. Be open to ideas by other developers; you would never know until you try!
2. Do not settle for substandard products; consider all the needs and requirements of the product and what solution can reliably tackle them all!
3. Upon a mind-expanding experience working with my fellow colleagues, I was inspired to pick up Neovim for software development work :^)

Introduction

This project involves the study of smart gas meter data collected by various IOT devices and accummulated into a gateway. We trained and validated various models including ARIMA, Multilayer Perceptron and LSTM to compare the training performances between the tree models. We concluded that for the purpose of prediciting future gas meter prices, Multilayer Perceptron provided the best performance. The feed-forward network and back-propagation algorithm associated with the model enabled it to successfully learn non-linear training data to predict forward in time.

Contributions

1. Conducted research on the ARIMA model and published findings into a notebook.
2. Wrote code to clean the training data using data filtering techniques, such as list comprehension or regex matching.
3. Conducted research on the various anomalies found in the dataset, by including temperature and rain data found online to draw correlation between weather conditions and days of suspected malfunctionsin the smart meter.

Reflections

1. Sometimes, data interpolation is not the way to go. It may conceal valuable information that are worthy of consideration; by ignoring gaps in data you risk the opportunity to investigate why they might be the way they are. In this case, data interpolation would have concealed the fact that the smart meter may be faulty.

Introduction

The project involves the creation of a full-stack IOT solution to an existing and prevalent societal issue; sitting postures of working adults in front of the computer leading to unhealthy curvatures forced upon the spine.

Contributions

1. Conducted research on Quaternions and the Madgwick Orientation Filter for a more efficient representation of the position of the spine in 3D space.
2. Implemented an efficient version of the Madgwick Orientation Filter on the sensortag firmware, by reducing precision to reduce packet size under delicate balance of sufficient accuracy over speed.
3. Created a new BLE characteristic and service from scratch to serve quaternion data generated on the device’s motion processing unit.
4. Programmed the BLE data gateway to process BLE data using AsyncIO library.

Reflections

1. Math is related to science!? Just kidding about the surprise. The true learning is that an acute understanding of the right math can simplify a problem to a solvable one.

Introduction

Source Code
Created a Discord Bot to share my workload as the severely overworked secretary and chairperson of the Art division of the club.

Contributions

• Developed a Discord Bot to automate administrative tasks that fall under my purview as the Productions Director in NUS Comic and Animation Society.
• The bot features tracking of birth dates, automated extraction of artworks from the Discord Channel, as well as running a month-long drawing activity (Digitised Inktober) with an elaborate point system.
• Technical Skills: Discord API, Google Drive API, Google Sheet API, Numpy and Pandas.

Reflection

• There is a need to practice proper security protocols in sensitive projects, including ones that handle the use of API credentials, personal information and access to Google accounts. This project taught me to practice vigilance and be sensitive to potential caveats during the development process.

Introduction

This project involves the investigation of the explainability of various tree-based supervised learning models in the prediction of housing prices across an array of features. The models subjected to investigation include XGBoost, Random Forest and Multiple Linear Regression with TreeSHAP augmentation.

We concluded that generally, the approach with lower model complexity tend to sacrifice accuracy for explanability. However, the use of advanced explaination techniques such as SHAPley values can reveal the impact of each feature in the prediction results using game theoretic approaches, without full dependence on the model. This means that a model with greater complexity can be used, and the explanability of the results can be derived in complement by the SHAPley explanability model. The findings of our research can help Singaporeans gain a better understanding of the relation between various common factors such as house pricing, floor area and location and the pricing of HDB blocks.

Contributions

1. Suggested the project topic and narrowed down the investigation to tree-based model, to compare explanability between decision rules and SHAPley explanability model.
2. Conducted research on SHAP, and used it on the XGBoost model to derive SHAPley values to explain the findings.
3. Contributed heavily to report writing and review.

Reflections

1. Explainability is a very important factor in machine learning; there need be an element of trust built between the model and the user so that its results can be relied upon. Model explainability is also valuable in drawing insights from the learning process of the model, which can then be applied widely in similar areas of research and investigation.

Report

Introduction

Programmed firmware for a remote-control robot using Keil v5 IDE and the CMSIS-RTOS2 library. Implemented a multi-threaded program with synchronisation using message queues and thread flags.

Contributions

1. Programmed a majority of the firmware, including thread synchronisation, interfacing with ultrasonic sensors, LED arrays, motor control and self-driving mode.
2. Experimented with memory allocation and low-level device firmware so as to allow clean implementation of library functions from bare-metal code.
3. Influenced my team to use Lithium Ion batteries, which provided a good supply of current to allow the motors to run at higher PWM frequency.

Reflection

• Battery is everything.
• It is better to solder your circuitory so that electrical contacts do not come loose.

Introduction

Slides This project aims to track students within the Rainbow Centre on their location in real time. This involves the development of a mobile app, website and a sensor network to enable real-time geolocation via RSSI triangulation, then updating the results to the frontend interface.

Contributions

• Proposed the overall architecture of the project.
• Developed various CRUD endpoints to interface with Firebase database.
• Implemented SocketIO and MQTT server alongside the REST server to communicate with both edge devices and client frontend in real time.
• Optimised the sensor data processing with multiprocessing and queues; hence effectively reducing sensor update interval from 4s to 0.5s.
• Developed various form pages, as well as a map viewer to insert and view sensors at their various positions in real life, at real time.
• Deployed the frontend and backend on AWS Amplify and AWS EC2 container instances with NGINX Reverse Proxy for load balancing and SSL certificate management.
• Technical Skills: Starlette, FastAPI, Gunicorn, SocketIO, AWS IoT SDK, Firebase Authentication, Firebase Real time Database and AsyncIO

Reflections

1. Perhaps the most rewarding project I have undertaken. In the context where the safety of the children are time-sensitive, latency was everything. Looking back at the late nights spent designing the multiprocessing pipeline to reduce latency and research on NGINX to deploy the product to production stage, I am proud to say that I have learnt alot from my experience with EGSC.

Introduction

Source Code
Designed and developed a Seat Occupancy System in hawker centres as part of the NUS Orbital Program, complete with a working sensor network, Telegram Bot and a Desktop Map Editor. Inspired by lunch under social distancing measures during the COVID Pandemic.

Contributions

• Implemented a cheap, low power and reliable BLE sensor network, comprising of ESP32 sensor motes and implementing star topology for transmission of seat occupancy data.
• Developed a desktop program to facilitate the conversion of seat occupancy data to a legible JPEG file for viewing and monitoring. Allows restaurant owners to upload their floor plan to launch to the backend for live occupancy updates.
• Technical Skills: Telegram bot Python Library, C++, Python, Redis, Tkinter and Inno Setup Scripting Language. Knowledge with ESP32, battery design and Wireless Sensor Networks were also applied.

Reflections

This project sparked my interests in the Internet of Things. It also first proved to me that the world is full of problems that are waiting to be solved; we just need to open our eyes and get ready to take action!

Introduction

Source Code
Modernised the painful process of compiling artbook pages on Photoshop to a powerful Python script. The program extracts image data from responses towards a Google Form, checks against registration details, and programmatically compiles all entries into artbooks for online distribution.

Contributions

• Wrote the entire program from scratch
• Technical Skills: Win32ComAPI with Photoshop, Python Image Library, Pandas, GSpread and PyDrive.

Reflections

I was shocked that Photoshop can be scripted from a Python codebase. It turned out to be extremely convenient for this use case.

Introduction

Designed a robot consisting of a Raspberry Pi 4 Model B and an Arduino Uno, as well as a LiDAR module. The robot is capable of autonomous mapping of the environment using a spinning laser LiDAR module, and actuates motor controls based on commands sent from a remote computer over a SSL link.

Contributions

Programmed the Raspberry Pi in C to enable the following features of the robot:

1. Interfacing of LiDAR driver code in C++ to support selective disabling of motor spin to reduce power consumption, for instance when the vehicle stops moving. (I used File IO, not proud of that)
2. Implemented PID control based on hall effect sensor inputs for incremental self-correction of vehicle driving trajectory.
3. Incorporated button key controls to control the robot remotely. Modified the skeleton code to create a new worker thread to handle keypresses using ncurses.
4. Designed the robot design so that it is compact with no loose wiring.

Reflection

• Nothing is impossible. Try any crazy idea!
• Ubuntu makes life easier

Introduction

The project requirement is to program firmware for TI SensorTags such that each sensor is able to detect another sensor that is proximate to itself within a certain period of time. If the sensor fulfills said condition, both sensors are said to be in close contact.

Contributions

1. Contributed to the implementation of the Quorum-Based Discovery Algorithm, based on the following paper.
2. Worked with the Contiki OS to program the TI Sensortags, using protothreads and RTimers to implement a power-efficient state machine.
3. Used Cooja to simulate network traffic so that the optimal scan period and algorithm can be derived in view of battery life, duty cycle and reliability.

Reflection

• Reliability is the cornerstone of network devices!
• Naming of variables and commenting helps to communicate your algorithms smoothly, especially with other developers!