DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
Interlocking 3D-printed structures
No full textIn 3D printing, which is often based on additive manufacturing, it can be very
challenging to fabricate large objects when they exceed the build volume of most
printers. This project delves into the creation of modular, interlocking components,
known as burr puzzles which are printed in smaller sizes and assembled into larger
structures seamlessly. A 3D burr puzzle is a complex interlocking structure that is
composed of a series of notched pieces and a movable key component. After
assembling, the puzzle attains stability and interlocks perfectly without the assistance
of adhesives or mechanical attachment. The high combinatorial complexity of the
puzzle is challenging to solve due to the strict sequence of assembling. Through
iterative design and testing, interlocking mechanisms are developed to enhance the
overall structural stability while reducing the number of parts. This method also results
in stronger and more stable connections between parts. One of the key advantages of
using interlocking parts is that it helps overcome the challenge of printing large
structures, which can be difficult due to printer size limitations. By dividing a large
structure into smaller interlocking pieces, it improves the efficiency of printing and
assembling. Additionally, this approach paves the way for the customization of
intricate designs that would be hard to achieve using traditional techniques. A
comparison of mechanical properties of the interlocking and solid homogeneous are
demonstrated in Abaqus simulations, combined with experimental validation. The
results demonstrate that interlocking mechanisms provide a viable alternative for
assembling large structures without the use of adhesives or fasteners. The simulation
in Abaqus reveals that the interlocking structures retains about 50% of the mechanical
properties as compared to a solid homogeneous structure while maintaining flexibility
in design and scalability. Furthermore, the study highlights the beauty of modularity
of interlocking design, allowing damaged sections to be replaced without printing the
entire structure. Ultimately, this project demonstrates how 3D printing and
interlocking mechanisms can transform the way parts are connected and assembled,
offering a more innovative and efficient solution for a wide range of applications.Bachelor's degre
Efficient surface passivation of Si for optoelectronic devices
No full textThe efficiency of optoelectronic devices is strongly dependent on the charge carrier lifetime at the semiconductor surface. Solar cell is a typical example of optoelectronic device that converts light into electrical energy. Improving the efficiency of solar cells remains a key challenge in the promotion and implementation of sustainable and renewable energy. This report investigates the enhancement of charge carrier lifetime in silicon (Si) optoelectronic devices with the application of various types of surface passivation techniques. In addition to surface passivation, wafer fabrication processes, such as annealing, are also applied to examine the impact on charge carrier lifetime improvement. The charge carrier lifetime is measured using quasi-steady-state photoconductance (QSSPC) method. The resulting measurements are then tabulated and analyzed to identify the most effective approach for charge carrier lifetime improvement.Bachelor's degre
A transformer-based method for forecasting stock market
No full textStock price prediction remains a challenging task in finance due to the complexity and volatility of markets. This report introduces a novel enhancement to the
Transformer-based model PatchTST for stock price prediction by adding a Cross Channel Convolution layer. While PatchTST efficiently processes long sequences
by dividing data into patches, it struggles to capture complex interdependencies
between multiple features. The cross-channel convolution layer addresses this by
learning relationships across features, improving the model’s ability to capture intricate feature interactions. The proposed hybrid model combines the strengths
of both a vanilla PatchTST model and convolution layer, offering a solution for
financial time series forecasting. Validation on a stock price dataset shows that
the enhanced model outperforms traditional methods and other machine learning
models, including vanilla PatchTST, significantly improving forecasting accuracy.
The key contribution lies in the model’s ability to handle multi-dimensional time
series data, capturing both long-term dependencies and feature interactions essential for accurate predictions. This approach provides a robust, interpretable forecasting tool for investors, traders, and portfolio managers, while also paving the
way for future research on hybrid deep learning models for time series forecasting
applications.Bachelor's degre
Generative auto-bidding with value-guided explorations
No full textAuto-bidding, with its strong capability to optimize bidding decisions within dynamic and competitive online environments, has become a pivotal strategy for advertising platforms. Existing approaches typically employ rule-based strategies or Reinforcement Learning (RL) techniques. However, rule-based strategies lack the flexibility to adapt to time-varying market conditions, and RL-based methods struggle to capture essential historical dependencies and observations within Markov Decision Process (MDP) frameworks. Furthermore, these approaches often face challenges in ensuring strategy adaptability across diverse advertising objectives. Additionally, as offline training methods are increasingly adopted to facilitate the deployment and maintenance of stable online strategies, the issues of documented behavioral patterns and behavioral collapse resulting from training on fixed offline datasets become increasingly significant. To address these limitations, this paper introduces a novel offline Generative Auto-bidding framework with Value-Guided Explorations (GAVE). GAVE accommodates various advertising objectives through a score-based Return-To-Go (RTG) module. Moreover, GAVE integrates an action exploration mechanism with an RTG-based evaluation method to explore novel actions while ensuring stability-preserving updates. A learnable value function is also designed to guide the direction of action exploration and mitigate Out-of-Distribution (OOD) problems. Experimental results on two offline datasets and real-world deployments demonstrate that GAVE outperforms state-of-the-art baselines in both offline evaluations and online A/B tests. By applying the core methods of this framework, we proudly secured first place in the NeurIPS 2024 competition, 'AIGB Track: Learning Auto-Bidding Agents with Generative Models'.Published versionThis research was partially supported by Kuaishou, Research Impact Fund (No.R1015-23), and Collaborative Research Fund (No.C1043- 24GF)
Action recognition for real-world applications
No full textHuman action recognition (HAR) plays an important role in various real-world applications such as intelligent surveillance, human-computer interaction, and healthcare monitoring. Recent advances in skeleton-based representations and graph convolutional neural networks (GCNs) have led to significant improvements in action recognition performance. However, the domain gap across different datasets, which is caused by variations in motion patterns, temporal alignment, camera angles, and action definitions, remains a major challenge, especially under unsupervised domain adaptation (UDA) settings.
In this study, we propose a novel method for skeleton-based HAR under UDA settings, building on a skeleton-cutmix framework. This approach generates cross-domain mixed skeleton samples by exchanging some part of the bones between source and target skeleton sequences, and assigns pseudo labels based on a dynamic label mixing strategy guided by a Beta distribution. We also reproduce the recover-and-resample method and make a modification to it. Additionally, we try to incorporate action priors into the mixing process to constrain cross-class combinations, enhancing the plausibility of synthesized actions.
We conduct extensive experiments on multiple cross-dataset sub-settings to evaluate our approach. The proposed unsupervised skeleton-cutmix demonstrates competitive performance and achieves the highest average accuracy across settings. Furthermore, the action prior skeleton-cutmix explores the integration of prior knowledge to improve semantic consistency during augmentation. These results highlight the potential of data-driven augmentation strategies in addressing domain shift in skeleton-based HAR.Bachelor's degre
Physics-informed machine learning for enhancing safety in tunnelling projects
No full textThe rapid pace of urbanization and population growth has accelerated the demand for underground tunnelling. With the advent of digitalization and big data, the tunnelling industry is undergoing substantial transformation, increasingly relying on collaborative technological approaches to meet the need for interdependent, tightly regulated, and "intelligent" design solutions. In confined underground environments, tunnelling operations must be optimized to ensure safety, cost-effectiveness, and minimal disruption to existing subsurface and surface infrastructure. The extensive data generated by contemporary tunnel boring machines (TBMs) presents a significant opportunity for the application of machine learning (ML) to enhance decision-making through timely and relevant insights. However, the use of ML in tunnel construction is hindered by several challenges: (1) Limited historical data on tunnel-induced damages, particularly settlement-related data, (2) Complex and uncertain geological conditions, and (3) The inherent "black-box" nature of ML models, which often lack interpretability in engineering contexts. In response to these limitations, this research aims to develop a series of physics-informed machine learning (PIML) methods to enhance safety control in tunnelling projects.
This study explores the application of PIML in tunnelling safety enhancement through four main stages: (1) Evaluation of system reliability with stochastic simulation, (2) Optimization of underground safety with fuzzy-robust programming, (3) Control of operational parameters with deep learning networks, and (4) Generation of operational strategy with reinforcement learning. The key findings are summarized as follows: (1) The stochastic simulation approach facilitates reliable system evaluation in large-diameter tunnelling projects, effectively addressing geological uncertainties and modeling errors; (2) Fuzzy-robust optimization significantly enhances structural safety, where robustness addresses unavoidable uncertainties in geotechnical conditions by adopting conservative solutions, while fuzzy logic seeks optimal solutions by balancing objectives and constraints; (3) PIML methods significantly improve the interpretability of relationships between variables, enhancing model reliability and predictive accuracy, particularly in multi-step forecasting with limited datasets; (4) Integrating physics-based insights into reinforcement learning enhances TBM operation efficiency, offering transformative potential for automated, reliable, and efficient tunnel construction.
Additionally, techniques such as explainable artificial intelligence (XAI), generative design, and parametric design are integrated to improve model performance. XAI techniques validate the rationality of PIML models, while generative design enables engineers to understand and control the optimization process, thereby fostering greater adoption of ML methods in tunnelling. In summary, the proposed PIML-based safety enhancement methodologies substantially improve the reliability and efficiency of tunnelling projects, advancing the automation, control, and intelligent decision-making in tunnel construction to new levels.Doctor of Philosoph
AI-powered study assistant for self-revision and interactive learning guide
No full textArtificial Intelligence (AI) chatbots have undergone significant evolution,
progressing from early rule-based systems to more sophisticated AI-driven models
leveraging natural language processing (NLP) and deep learning. Initially, chatbots
relied on simple predefined scripts and pattern matching, limiting their ability to engage
in meaningful dialogue. Over time, advancements in machine learning, particularly the
development of transformer-based models, enabled chatbots to generate more natural
and context-aware responses. Despite these advancements, challenges such as lack of
document-based context, and an inability to generate structured study materials still
exists and this project aims to tackle these limitations.
This web-based application uses artificial intelligence to transform how students
learn and revise their work. The application allows users to upload and work with
multiple PDFs, including lecture notes and slides, and perform a variety of tasks such as
creating dynamic study guides, precisely answering questions based on sections of the
documents, and simulating exam questions. The platform facilitates the setting up of
multiple chat histories for efficient content management. Moreover, the use of
speech-to-text technology provides a hands-free user experience. This application aims
to improve learning effectiveness and give students a comprehensive, interactive study
guide.Bachelor's degre
Real-time location tracking and fall detection system
No full textDue to the growing demand for real time elderly monitoring, various advanced
tracking and fall detection systems have been developed. The aim of this study is to
propose an elderly tracking system based on Ultra-Wideband (UWB) positioning
complemented with Bluetooth Low Energy (BLE) wireless communication that
integrates accelerometer and gyroscope motion sensing and the data transmission
using Message Queuing Telemetry Transport (MQTT), for high accuracy in elderly
care facilities. The elderly wear the DWM1001 tag that constantly send out the
motion and positioning data and thus provide precise localization and real time fall
detection. To overcome the signal fluctuation ridden by conventional Received
Signal Strength Indicator (RSSI) based BLE tracking, the system uses UWB
positioning to achieve an accuracy of ± 10 cm on average.
Th accelerometer and gyroscope fusion-based threshold fall detection algorithm has
demonstrated high accuracy in distinguishing normal movements from critical falls,
with a very low rate of false positives. On the other hand, battery life of the system
varies from 7 to 8 hours on a 650 mAh battery based on update interval and power
optimization techniques. We validate that this BLE-UWB system is able to offer a
scalable, reliable and real-time elderly care environment tracker. Future
developments could be with machine learning based fall classification, migration of
health analytics to the cloud and further energy efficiency to make it possible for a
longer term and wider deployment in assisted living facilities and to hospitals.Bachelor's degre
System of faith
No full textThis report documents the development of System of Faith, a 3D animated short film produced as part of the Final Year Project at the School of Art, Design & Media, Nanyang Technological University. Exploring themes of belief, doubt, and identity, the film is set within a gothic cathedral environment and features an alien protagonist confronting priests and masked believers. My contributions included character asset creation, VFX simulations (FLIP fluids, soft-body dynamics, fire/smoke), scene assembly, and compositing, as well as project coordination. The technical pipeline utilised Maya, ZBrush, Houdini, Redshift, Nuke, and Premiere Pro. Key challenges addressed include intensive simulation workflows, motion-capture integration, and adaptation to a cloud-based Redshift rendering system due to licensing constraints. The report also covers pre-production ideation, environment and character design, lighting, and audio collaboration. This body of work reflects significant growth in both artistic vision and technical proficiency, demonstrating effective interdisciplinary collaboration and project management.Bachelor's degre
Effects of regional white matter hyperintensities and β-amyloid on domain-specific cognition and progression to dementia
No full textWhite matter hyperintensities (WMHs) and cerebral β-amyloid (Aβ) have been characterized as clinically significant biomarkers associated with greater cognitive decline and incidence of Alzheimer’s Disease (AD) dementia. However, it remains unclear how their regional manifestations co-contribute to domain-specific cognition and dementia onset. We investigated 200 cognitively normal (CN) and 523 individuals with mild cognitive impairment (MCI). We first quantified regional WMHs and Aβ accumulation in the four cerebral lobes. Next, we evaluated the effects of both WMHs and Aβ in each lobe on memory, executive function (EF), language, and visuospatial function. We used Cox proportional hazard models to determine the contributions of both regional WMHs and Aβ to dementia progression. In CN individuals, greater WMHs in parietal and temporal regions were associated with poorer EF beyond Aβ. In MCI individuals, greater Aβ burden in all lobes were associated with poorer memory, EF, and language abilities beyond WMHs. Lastly, both greater occipital WMHs and Aβ predicted progression to dementia. Temporo-parietal WMHs may drive early decline in EF beyond regional Aβ, while occipital WMHs play a critical role in disease progression to AD dementia beyond regional Aβ, highlighting the complex interplay of regional WMHs and Aβ on domain-specific cognitive and clinical function.Published versionData collection and sharing for this project was funded by the Alzheimer's Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012)