82837 research outputs found
Sort by
Dynamic Analysis and Stabilization of VSC-Based Multi-Terminal DC Grids with Power Flow Controllers
No full textThe high-voltage voltage-source converter (VSC)-based multi-terminal dc (MTDC) grid is a key technology for integrating offshore wind energy and enabling bulk power transmission
across nations and continents. The power flow controller (PFC) is essential for future MTDC grids to mitigate potential transmission bottlenecks and ensure economical operation. In
this research, the functionality of the PFC is extended to provide active damping of MTDC grid current oscillations caused by dc-side resonance in addition to power flow management. Eigenvalue and sensitivity analyses, based on comprehensive small-signal models, are conducted to evaluate the damping capability of the proposed PFC-integrated compensators and assess their dynamic coupling with the PFC control loops.
DC circuit breakers (DCCBs) are crucial for interrupting faults in MTDC grids. The current limiting inductors of DCCBs can significantly influence the dynamic performance and
stability of MTDC grids. However, the existing literature does not address the assessment and mitigation of the DCCB inductance impacts on the stability of an MTDC grid equipped with a PFC. In this research, eigenvalue and frequency response analyses are employed to assess the dynamic performance and stability of the MTDC grid, considering the effects of DCCBs inductances, PFCs, and converter station control parameters. Moreover, the capability of the PFC is expanded via the development of a PFC-based stabilizer to improve the dynamic performance of the MTDC grid and enhance its damping.
Back-to-back and point-to-point HVDC systems can enhance cross-regional transmission capacity and power system stability. However, their use in interconnecting two extremely
weak grids can compromise system stability. This research presents a comprehensive stability analysis of back-to-back and point-to-point HVDC systems to distinguish the root causes of instability mechanisms and identify the critical short-circuit ratio of the converter stations. Eigenvalue analysis, based on detailed small-signal modeling, showed that three distinct instability mechanisms, high-, low-, and medium-frequency instabilities, can compromise the operation of VSC stations under extremely weak grid conditions. Notably, medium-frequency instability observed in the point-to-point HVDC system during inverter operation is predominantly caused by the power feedforward compensation of the dc-bus voltage controller. Active compensators are designed based on participation factor analysis to mitigate the identified instabilities.
Connecting MTDC grid converter stations to very long ac transmission lines creates interfaces with weak or extremely weak ac systems. This research addresses the small-signal
stability of an MTDC grid equipped with a PFC under extremely weak ac system interconnections. Using a comprehensive linearized state-space model, eigenvalue analysis uncovers that beyond the well-documented low-frequency oscillations, an MTDC grid may experience
unstable medium-frequency oscillations under rectifier operation of droop voltage-controlled VSCs when interfaced with an extremely weak ac system. Moreover, it was shown that the most challenging scenario is when multiple droop voltage-controlled VSCs are interfaced to very weak ac systems, with unstable low- and medium-frequency oscillations observed at short-circuit ratios of 1.4 and 1.8, respectively. Simple yet efficient low- and medium-frequency oscillation compensators are proposed to stabilize the system, and their parameters design is detailed.
Multi-terminal medium-voltage dc (MVDC) grids are emerging as a promising solution for effectively integrating renewable energy resources and supplying the growing demand for
dc loads, such as EV charging stations (EVCSs). However, ensuring their stable operation under varying operating conditions remains a significant challenge. This research addresses the stability and dynamic interactions of a typical MVDC grid integrating EVCSs and PV generation via dual-active-bridge converters, an ac grid via a VSC, and a PFC. Using a comprehensive linearized state-space model, eigenvalue and frequency response analyses reveal that the MVDC system stability is jeopardized during the discharging operation of EVCSs, in the presence of the power feedforward compensation loop in the dc voltage controller of the VSC; such stability challenges become more pronounced in weak ac grid connections. A novel active damping controller integrated with the PFC control strategy is proposed to
stabilize the MVDC system, extending the PFC capabilities to offer system-level damping in addition to power flow management.
Extensive offline and real-time simulation studies verified the analytical results and the effectiveness of the proposed compensators in enhancing the stability of MTDC grids under
a wide range of practical operating conditions
Effective De-censoring of Survival Data via Budget Constrained Active Learning
No full textStandard supervised learning algorithms build a prediction model from a labeled dataset. Given a small set of labeled instances and a pool of unlabeled ones, a budgeted active learner will identify relevant unlabeled instances, then use its predefined budget to pay to acquire labels of those instances, which are then added to the training set, then used to produce a new performance model, with the goal of producing an effective model. We investigate different Budgeted Learning methods within the context of survival datasets that contain some (right) censored instances, which provide only a lower bound c on an instance’s time-to-event t. Fields such as medicine, finance, and engineering often involve survival datasets with many of these censored instances.
We explore various settings, where the learner can use its budget (1) to label a censored instance – eg, an oracle updates an instance censored at 2 years [censored 2], to be uncensored at 5 years [uncensored 5] – or (2) to partially label an censored instance, by specifying “k more years”. Here, if k = 1, the oracle would just update the label for this instance to be [censored, 3]. (If k = 4, it would state [uncensored 5]; presumably, k = 4 would cost more than k = 1; and case (1) above should be even more costly.) This approach reflects real-world data collection scenarios, where follow-up with censored instances does not always lead to uncensoring, and the amount of information the learner will receive is dependent on both the budget and the success of follow-up efforts to obtain data. Despite its relevance to real-world scenarios, to our knowledge, no other projects have analyzed this intersection of data selection with survival analysis and partial information, nor has anyone considered the budget in these settings.
We present both experimental and theoretical results on how state-of-the-art Budgeted Learn-\ning algorithms can be adapted to survival data, discussing the associated limitations. Our approach involves reducing these scenarios to variants of the maximum coverage problem while preserving asymptotic bounds and time complexity comparable to standard data selection methods. Further- more, empirical evaluations on multiple survival tasks demonstrate that our model outperforms other potential methods across various benchmarks
Experiences of Sexual Assault Among Transmen: A Feminist Narrative Inquiry
No full textThis study examines the sexual assaults of transmen through a feminist narrative inquiry that integrates embodiment, female masculinity, and structural analyses of gendered violence. Despite exceptionally high prevalence rates, transmen’s narratives remain largely absent from feminist rape theory, legal scholarship, and policy discourse. Drawing on in-depth, relational, trauma-informed interviews, the study analyzes how participants made meaning of their assaults and how their accounts illuminate the structures that produced vulnerability. The findings show that sexual violence against transmen followed well-established feminist accounts of rape as the enforcement of sexed hierarchy and the punishment of gender nonconformity. Participants’ experiences indicate that vulnerability did not arise from identity labels or transition status but from the social meaning of the female body, the policing of female masculinity, and the conditions that rendered their assaults misrecognized or narratively unintelligible. The analysis also identifies forms of testimonial and hermeneutical injustice that constrained participants’ ability to name their experiences. By situating transmen’s accounts within feminist theory and narrative epistemology, the study argues that their experiences are neither anomalous nor marginal but continuous with global patterns of sexual violence against gender-nonconforming girls and women. Centering these narratives strengthens feminist understandings of sex-based vulnerability and challenges discursive shifts that obscure the material and embodied dimensions of sexual violence
Learning from Video for Control
No full textThis thesis addresses the challenge of learning temporal and spatial abstractions from video data to enhance decision-making capabilities in agents. The ability to model agent-environment interactions can streamline decision-making by recognizing patterns and regularities. Reward-free, offline video-based interactions are the most abundantly and easily available form of data, containing sequences of high dimensional observations and action pairs. The goal of this thesis is to learn spatial and temporal abstractions that capture useful information for control from video data, in the absence of explicit reward or task information.
The first contribution of this thesis is the development of spatial abstractions that capture control-relevant information that discards unnecessary information. By focusing on high-dimensional observation data, we point out that capturing all information in visual inputs can hamper policy learning, especially when unrelated visual elements are present. The proposed approach, named Agent Controller Representations for Offline RL (ACRO), uses a multi-step action prediction objective to capture only the control-relevant information, which we demonstrate across new datasets with added visual noise.
The second contribution introduces a general mechanism for informational parsimony (meaning using as little information as possible from a given input), which we term InfoGating, to enhance discarding information irrelevant to a given objective. InfoGating optimizes how much information from observations is processed by selectively gating either input pixels or activations in intermediate layers, thereby discarding control-irrelevant information. Our experiments show that representations learnt through InfoGating improve generalization in noisy environments and significantly enhances robustness in policies when irrelevant visual information is present.
Finally, we introduce Video Occupancy Models (VOCs), a family of generative models that make temporally coarse, latent-space predictions. VOCs learn compact representations of high-dimensional observations and then temporally predict future states in a single-step, sidestepping the error-prone multistep rollouts often used in traditional temporal models. This temporal abstraction approach effectively estimates value functions, which we demonstrate by performing control within a model-based RL framework. Together, these contributions offer new insights into learning spatial and temporal abstractions of offline video data, with practical implications for scaling up methods that model high dimensional agent-environment interactions
Large Language Model Fine-tuning via Structured Data Exploitation and Instruction Generation
No full textLarge Language Models have transformed Natural Language Processing, achieving remarkable performance on complex tasks. However, fine-tuning these models remains challenging due to their reliance on vast datasets which may have complex structures. High-quality data and effective fine-tuning techniques are essential for ensuring accurate interpretation in downstream applications.
This thesis adopts a data-centric approach to enhance LLM fine-tuning, focusing on structured data exploitation and systematic instruction evolution. We introduce improved learning methods, including hierarchical label-aware modeling and structured data augmentation for Fine-Grained Named Entity Recognition, as well as supervised contrastive learning to enhance modality-specific and modality-invariant feature representation in Multimodal Sentiment Analysis. Additionally, we explore structured synthetic data generation by leveraging LLMs to create high-quality fine-tuning data. To enhance instruction diversity and complexity, we introduce instruction fusion, merging seed instructions to generate richer datasets and improve LLM adaptability. Furthermore, we propose a structured instruction decomposition framework that systematically restructures instructions to control complexity and refine LLM training. Finally, these techniques are integrated into a unified instruction evolution framework, ensuring progressive learning across various domains, including general instruction comprehension, mathematical reasoning, and code generation.
By optimizing structured data exploitation and instruction-driven synthetic data generation, this thesis advances LLM fine-tuning, improving adaptability, efficiency, and performance across real-world applications
Detecting Healthcare Insurance Fraud Using Markov Observation Model
No full textHealthcare insurance fraud poses a significant threat to both the financial sustainability
of insurance funds and the delivery of quality healthcare services. Rising costs, evolving
fraudulent schemes, and the limitations of traditional rule-based detection systems necessitate
more sophisticated analytical methods. This thesis introduces the Markov Observation
Model (MOM), an extension of the Hidden Markov Model framework, specifically
designed for sequential data where each observation depends on both its immediate
predecessor and a hidden state sequence. By leveraging counting processes and incorporating
three canonical fraud types—Switch Service Fraud, Overcharge Fraud, and Forged
Fraud—the MOM effectively captures nuanced behaviors commonly overlooked by conventional
techniques.
To evaluate the performance of the proposed model, a simulated healthcare insurance
dataset was created to mirror real-world complexities. The research details how MOM’s
transition probabilities and emission structures are estimated via an iterative Expectation-
Maximization algorithm and demonstrates how these estimates translate into practical
fraud detection. Comparative analysis against a Random Forest classifier illustratesMOM’s
improved accuracy in identifying illicit claim patterns, particularly in scenarios where
conventional supervised learning methods struggle to detect subtle anomalies.
The results highlight MOM’s potential to serve as a robust and scalable tool for healthcare
fraud detection, offering enhanced interpretability of providers’ “mindsets” and adaptability
to a broad range of fraudulent schemes. Concluding with a discussion on the
model’s limitations and avenues for future research, this thesis underscores the importance
of combining advanced statistical techniques with domain expertise to better safeguard
healthcare systems
Mapping the risks of sea ice change, shipping, and Peary and Dolphin-Union caribou movements to Inuvialuit and Inuit well-being in the Western Arctic
No full textThe Arctic is changing. Its physical and geopolitical landscapes are rapidly transforming due to increasing temperatures and loss of sea ice—both of which present new risks to the Western Arctic in particular, where the corridors of the Canadian Arctic Archipelago are shared by Inuvialuit and Inuit communities, the Peary and Dolphin-Union caribou herds. Due to warming temperatures, there has also been an increase in the number of shipping vessels in this space. Currently, concerns at the intersection of caribou movement, sea ice change, Arctic shipping, and Inuvialuit well-being are unaddressed in risk assessments, which inspect risks as independent from each other, and it is difficult to illustrate the nexus of the four issues on a map. To more accurately understand these concerns and explore possibilities for more comprehensive mapping of risk, I used the methodology of Community-Based Participatory Research to design this project and research questions. A participatory mapping workshop was planned in collaboration between the research team and Inuvialuit leaders and used narrative inquiry to link stories to ArcGIS data documentation. Comparing the literature review on caribou, sea ice, shipping, and Inuvialuit use with the workshop results, I concluded that the risk at the nexus of the four issues depends on overlapping seasons and locations which are not explored by assessments that compartmentalize individual risk factors. Participatory mapping has the potential to illustrate risks as interconnected; I reflect on the method’s capacity to express Indigenous observations and oral histories. While the participatory mapping workshop increased data access for the Inuit and knowledge exchange between actors through interactive touchscreen technology, it did not fully capture the impacts of overlapping seasons and locations. To support increased knowledge exchange, I recommend holding a similar workshop in Inuit and Inuvialuit communities and focusing on cartographic methods informed by Indigenous concepts of space and time. Knowledge exchange of risks in the Arctic—of critical importance for adapting to the ongoing ecological and geopolitical change in this region—can be improved if Inuvialuit have access to the data and the tools used in decision-making. Risks can be more accurately understood if participatory mapping in the communities is used to incorporate Indigenous Knowledge and Indigenous conceptual frameworks in the study of risk
Music in the Graphic Narrative Medium: A Research-Creation Exploration of Musical Representation in Comics through a Graphic Biography of Borys Liatoshynsky
No full textThis thesis explores the representation of music within the graphic narrative medium. The study addresses the question of how music, an auditory art form, can be effectively conveyed through the non-auditory visual and textual means in comics. While the intersection of music and literature has been extensively studied, the representation of music in comics remains a relatively new and niche field.
To achieve these goals, the thesis employs a research-creation methodology, specifically a research-for-creation approach. This involves both the scholarly investigation of existing comics and theories related to music and comics and the practical creation of a graphic biography, titled Borys Liatoshynsky: A Life in Reflections, in collaboration with a visual artist.
The findings of this thesis highlight multiple layers of musical representation in comics, drawing on both structural and expressive parallels between the two media. Through the analysis of existing works, it became evident that techniques such as visual rhythm, panel composition, and interaction between visual and verbal can effectively evoke musical qualities, such as the notion of counterpoint.
The graphic biography created as part of this thesis, along with the accompanying textual reflections, offers a different perspective on the creation of biographical narratives about musical figures and their works. Additionally, it provides insights into the creative process of an individual involved in the making of a graphic biography.
Ultimately, this research-creation endeavor demonstrates the potential of the comics to engage with music on structural, stylistic, and emotional levels, contributing to the interdisciplinary dialogue between music studies and comics studies
Multi-dimensional characterization of plant-based and dairy milks: Sensory attributes, consumer perceptions, and tribo-rheological properties
No full textPlant-based milk (PBM) has been viewed as a sustainable and healthier alternative to dairy milk (DM). However, differences in PBM composition lead to variations in physical properties, consequently affecting sensory properties, consumer acceptance, and perception. This research aimed to address current challenges by developing a fortified oat milk and comparing it with commercial PBMs and Dairy Milks (DMs) through sensory characterization, consumer perceptions, and instrumental analyses. In the first study, oat milks were fortified with β-glucan at a level that attains health benefits and protein at a level equivalent to cow’s milk. Sensory attributes and liking were evaluated by consumers (n=106) using the 9-point hedonic scale, Just-About-Right, and Check-All-That-Apply (CATA). Fortification with β-glucan significantly increased the overall liking of oat milks attributed to smoothness. However, fortification with β-glucan and oat protein significantly reduced the overall liking due to rancid, chalky, and sandy attributes. Conjoint Analysis (CA) confirmed that protein from plant-based sources, protein content greater than cow’s milk, and β-glucan at recommended health-benefiting levels were the most important attributes to consumers. Thus, improving the protein component and sensory attributes was essential for greater acceptance of fortified oat milks. The second study developed the fortified oat milk using β-glucan and pea protein, and determined the influence of extrinsic cues (labels, testing locations) on consumer perception and acceptance. Consumers (n=108) evaluated fortified oat milks for four consecutive days: Day 1 (label pre-exposure) in sensory laboratory; Day 2-3 (bottles labeled with either health or environmental benefits) through Home-Use-Test (HUT); Day 4 (post-exposure) in sensory laboratory. Attribute intensity and overall liking (9-point hedonic scale) were evaluated daily, and wellness perception (six dimensions assessed by CATA items) was evaluated on Days 1 and 4. Fortified oat milks with health and environmental labels were correlated with a greater number of wellness dimensions compared to unlabelled milks, while product liking did not significantly differ between labeled and unlabeled milks. Additionally, testing locations did not significantly influence product liking and perceived attribute intensities in fortified oat milk. This implied that although extrinsic cues enhanced positive experience, consumers remained consistent in their liking for fortified oat milk, meriting comparison of fortified oat milk with a broader range of commercial PBMs and DMs. In the third study, experienced sensory panelists (n=14) evaluated fortified oat milk, PBMs, and DMs (n=29) through Projective Mapping. Variations in textural attributes were perceived; moreover, six clusters revealed groups with similar texture profiles. Twelve milks were selected from these clusters for evaluation by consumers of both PBMs and DMs (n=109) using liking (9-point hedonic scale), sensory attributes (Rate-All-That-Apply), perceptions and uses (CATA), and dairy consumption motivation (4Ns Scale) based on factors ‘Natural,’ ‘Need,’ ‘Normal,’ and ‘Nice.’ Dairy and oat milks were most significantly liked, generating two consumer clusters: dairy-oat-highlikers (high liking for dairy and oat milk) and dairy-plant-modlikers (moderate liking for dairy and select plant-based milks). ‘Sweet,’ ‘smooth,’ and ‘creamy’ were the drivers of liking, while ‘simple,’ ‘traditional,’ ‘comforting,’ and varied product uses were associated with a mean liking increase. A key finding was that consumers viewed PBMs as distinct products with unique sensory attributes that conveyed specific emotions, perceptions, and uses, and achieved comparable liking to DM driven by the key sensory attributes. Moreover, continued consumption of dairy milk was due to a perceived ‘Need’ (p<0.0001).
Texture was identified as the most important attribute of PBM acceptance. Thus, in the fourth study, consumers (n=104) evaluated milks (n=5) through Temporal Dominance of Sensation (TDS), and the resulting texture perceptions were associated with particle size, viscosity, and tribological behavior by instrumental analysis. PBMs had a greater number of dominant sensory attributes, including thickness, smoothness, powdery, astringency, and creaminess, while DM only had smoothness and creaminess as TDS dominant perceptions. These sensory perceptions were correlated with the instrumental texture measures.
The novelty of this thesis was reflected in the development of fortified oat milk with improved sensory attributes and consumer liking, the consumer-based approach which provided relevant insights on consumer segments, perceptions, and consumption motivations, and in addressing the gap of limited use of dynamic methodologies and their integration to instrumental texture properties. This thesis is significant to academia and food industry by identifying drivers of liking, consumer perceptions, and insights on the PBM and DM product landscape
The Use of Breathwork as an Adjunctive Therapy for Individuals with Chronic Health Conditions
No full textIntroduction: People living with chronic health conditions often experience negative impacts on multiple aspects of their health and well-being, including disease-related physical and mental health symptoms, as well as reduced quality of life. Breathwork interventions, which involve intentional control or awareness of breathing, have gained widespread recognition as a promising therapy to complement traditional approaches to managing chronic diseases, due to their holistic health benefits, accessibility and potential for application across various chronic conditions.
Objectives: To advance the understanding of the impacts of breathwork interventions in individuals living with chronic health conditions by (i) providing a brief summary of the current literature on breathwork interventions for chronic conditions, (ii) conducting a scoping review to summarize the use and health impacts of breathwork interventions evaluated by randomized controlled trials (RCTs) in chronic disease populations, and (iii) using a qualitative descriptive approach to explore the perspectives of individuals with chronic health conditions who engaged in breathwork as part of an online mind-body wellness program.
Methods: A scoping review was conducted in Chapter Three, where five databases were searched from inception to November 2, 2023, including RCTs that assessed the impact of breathwork on health and wellness outcomes in adults with chronic conditions. Data was extracted on the clinical populations, intervention protocols, health and wellness outcomes evaluated, and safety, allowing for a comprehensive descriptive analysis of breathwork intervention delivery characteristics and impact across chronic diseases and health outcomes. Chapter Four employed a qualitative descriptive approach to conduct a secondary analysis of semi-structured interviews with participants who had various chronic health conditions and had completed a 12-week online mind-body wellness program, eMPower. Thematic analysis followed an in vivo and open coding scheme.
Results: Chapter Three yielded 69 RCTs, assessing 76 unique breathwork interventions across 31 chronic conditions and 118 health and wellness outcomes. Most interventions involved slow or deep breathing, often with an expiration-focus, and instructor guidance. Breathwork was most commonly studied in individuals with cardiovascular disease, cancer, gastroesophageal reflux disease and mental health conditions, and consistently demonstrated benefits for heart rate variability, respiratory rate, depression, anxiety, pain and fatigue. Adverse events were rare, when reported. Chapter Four included 20 participants with a range of chronic health conditions. Four primary themes emerged from our thematic analysis: (i) Losing and regaining control, (ii) Addressing the whole person, (iii) A breath away: accessible, on demand support, and (iv) Buying in: facilitators of acceptance. The first two themes capture participants’ motivations for joining eMPower, preconceptions of breathwork and its perceived holistic benefits. The latter two discuss the accessibility and suitability of breathwork for people living with chronic diseases, as well as factors contributing to its acceptance.
Conclusions: Breathwork interventions include a broad spectrum of practices, delivery methods, and potential clinical applications. It is supported by quantitative and qualitative evidence as an accessible and adaptable tool with promising holistic health benefits. Breathwork is well-suited for individuals with diverse chronic health conditions and offers a valuable complement to conventional chronic disease management when appropriately tailored and adapted to the individual needs of patients