48440 research outputs found
Sort by
Creating an Agent-Based Model to Evacuate Fuller Labs
No full textBuilding evacuations are critical for safety during emergencies, yet we cannot efficiently test evacuations in the real world. To help with this problem we created an agent-based evacuation simulator that we used to recreate Fuller Laboratories at WPI. Agents, modeled as circles with speed and size parameters, navigate via a pre-generated vector map to the nearest door, avoiding obstacles along the way. This allows us to rapidly test different room layouts, exit configurations, and obstacle placement to see how they affect Fuller evacuation times. Our results show that having students already in the lobby as well as the table by Upper Perrault Lecture Hall do not significantly impact evacuation times. However, the loss of a door increases the time it takes to evacuate a building. We recommend that to ensure student safety, WPI prioritize ensuring that all exits remain functional. In the future, our simulation can be improved by adding in a social force, allowing the agents to make decisions on which door to go to instead of only travelling to the closest door
Role of Hyaluronic Acid in Endothelial and Cancer Cell Interactions in-vitro
No full textMetastasis, the leading cause of cancer-related mortality, involves the adhesion and migration of cancer cells across the vascular endothelium. This study examined the role of hyaluronic acid (HA) within the endothelial glycocalyx (GCX) in modulating cancer cell interactions under physiologically relevant shear stress conditions. Human lung microvascular endothelial cells (HLMVECs) were cultured and exposed to static, 5, 15, or 30 dynes/cm² shear stress using a parallel plate flow system, followed by the introduction of GFP-labeled MDA-MB-231 breast cancer cells. HA expression, GCX thickness, cancer cell adhesion, clustering, and transendothelial migration (TEM) were quantified through fluorescence imaging and image analysis. The results demonstrated that increasing shear stress significantly enhanced HA expression and GCX thickness, while cancer cell attachment, TEM, and clustering were markedly reduced at higher shear levels. These findings suggest that shear stress–induced upregulation of HA reinforces the endothelial barrier and diminishes metastatic cancer cell behavior. Further studies involving targeted HA degradation are warranted to confirm its protective role in metastasis
Antimicrobial-Loaded Bacterial Cellulose for Chronic Wound Applications
No full textDue to the lack of existing chronic wound coverings, this study investigates the antimicrobial properties of vancomycin-loaded bacterial cellulose (BC) with various alginate concentrations and its impact on innate BC characteristics. Several techniques were utilized to investigate the effectiveness of each modified BC model. Fourier Transform Infrared Spectroscopy and X-Ray Diffraction were performed to ensure antimicrobial agent presence and preserve the BC structural integrity, respectively. Zone of inhibition was conducted to assess the ability of the loaded BC model to inhibit Staphylococcus aureus growth. Finally, the water holding capacity and water retention of BC were evaluated to ensure that these properties were maintained following modification
CS MQP - Performance Architecture Team - PowerSense NVIDIA
No full textThis Major Qualifying Project focuses on enhancing NVIDIA's telemetry tools used by the Performance Architecture Team to test and optimize performance on their Tegra Systems. We worked on five key projects to improve the functionality and efficiency of these tools. First, we added support for ATF and BPMP to an existing telemetry tool allowing for the collection of additional performance metrics. Second, we implemented system configuration capabilities allowing for users to modify chip settings through the command line. Third, we created a repository to collect benchmarks to be used as baselines during future testing. Fourth, we create a simulation framework for additional Low Power Idle states to predict the impact on performance before being implemented in hardware. Finally, we added support for visualizing some of the data one of the team's tools collect to a different visualization tool, centralizing data storage and improving processing efficiency. These improvements allow for more comprehensive data collection, quicker analysis and better testing workflows for chip performance optimization. Through these improvements, the Performance Architecture Team and others can now collect more telemetry data faster than before, supporting NVIDIA's ongoing efforts to optimize chip performance for both speed and power efficiency
Predicting Students' Mental Health Using Facial Expression Data
No full textThe COVID-19 pandemic has significantly impacted college students' mental health, increasing rates of stress, depression, and anxiety. With NSF support, the well-being of over 100 students was monitored for three months continuously across multiple academic terms from 2020 to 2021. Participants completed weekly surveys assessing their emotional state (PANAS), perceived stress (PSS), depressive symptoms (CESD), and engagement (OSE), as well as monthly interviews scheduled and held using Zoom. Simultaneously, students' facial expressions were recorded during online learning sessions and analyzed using OpenFace software to extract Facial Action Units (AUs), as our features. We developed several Machine Learning models, including Linear Regression, Support Vector Regression/Classification, Decision Tree, Gradient Boosting, Decision Tree, Random Forest, and Logistic Regression to predict mental health outcomes from the facial action units as our features. Our results showed that tree-based models significantly outperformed linear models in the regression task, with Random Forest achieving the highest R2 scores for predicting depression (0.531), stress (0.458), and negative affect (0.368), as well as lower RMSE values. For classification, Random Forest and Gradient Boosting achieved high accuracy (0.759-0.762) and F1 scores (0.771-0.772) for identifying students at risk of depression. We also transcribed and reviewed several interviews, later performing thematic analysis on these subject interviews. As for the qualitative data using the subject interviews, we found three major themes among subjects being issues with the barriers to virtual learning, COVID-19-related stressors, and work-life balance. These findings suggest that facial expression data can enhance mental health prediction and monitoring in remote educational settings, with interview data aiding in understanding what drives certain mental health concerns. This study demonstrates the potential of the integration of video analytics and self-report data to support mental health detection. Future work will explore real-time applications and the use of multi-modal data with physiological and behavioral indicators, as well as using the combination of facial expression, survey, and interview data to cross-analyze and better understand the meaning of these data in how it can provide support in mental health predictions
A Language Anchor-Guided Method for Robust Noisy Domain Generalization
No full textRecent advances in machine learning have underscored the critical challenge posed by domain shift, where models trained on one distribution often struggle when applied to previously unseen environments. This issue is particularly acute as variations in image conditions—such as lighting, pose, and background—or other domain-specific factors lead to significant shifts in feature representations even within the same class. Traditional approaches that assume consistency between training and test data distributions are increasingly inadequate when large networks overfit to spurious correlations, which can severely degrade model performance. Moreover, the presence of noisy labels that commonly occur in natural datasets exacerbates the problem, rendering the application of models in real-world scenarios ineffective. Researchers are thus turning to innovative strategies that focus on robust feature extraction and the alignment of intrinsic representations across diverse domains, aiming to enhance the reliability and interpretability of learned models. In this project, we aim to develop a novel method that leverages external domain expertise to guide the learning process and effectively address the challenges posed by domain shift and label noise
SMARTER LEARNING ENVIRONMENTS
No full textEnvironmental conditions in classrooms play a critical role in shaping students' health, cognitive function, and academic performance. A novel environmental data collection and analysis system was developed and deployed to take detailed samples and provide informative visuals of temperature, humidity, air quality, noise, and lighting. Perceptions of environmental factors were evaluated through outreach surveys and one-on-one interviews. Recommendations were proposed to the Universidad de Cádiz to help address deficiencies found in ambient classroom conditions and promote student well-being and learning
Stock Market Simulation 2509
No full textThis project was a six-week simulation of the stock market. The goal of the project was to gain knowledge of the stock market and its trends, and to gain experience by trading with paper money. The project used current and accurate numbers from the stock market to ensure the simulation was realistic. Two trading methods were tested: technical trading and hedged shorting. Stock simulations were performed with numerous investment options and portfolios with a starting $100,000 principal. Each strategy started with an identical portfolio. For the technical trading strategy, the return was 8.47%, while the hedged shorting strategy had a return of 7.03%. The total change of the S&P 500 index over the same six-week period was 5.46%. The results from the simulations indicated that technical trading was more profitable than the hedged shorting strategy, and both strongly outperformed the overall market. This project provided participants with trading experience that will benefit them in future investments
Feasibility Study on Establishing Smart Park Technologies in Acadia National Park
No full textTaking inspiration from smart homes and cities, national parks across the United States are investigating smart technology to improve park management. Acadia National Park currently uses non-live traffic sensors throughout the park, requiring park staff to physically retrieve the data. We explored the feasibility of implementing live updating smart sensors in the park by testing availability of cellular networks at each sensor using a custom-built smart device. Out of the locations tested, 24 out of 30 had network connectivity and due to this, we recommend that Acadia further pursues smart technologies. Live monitoring in the park will greatly improve park management and enhance visitor experience through real time updating traffic data and tracking long term trends
Survivor Tracking: A Wireless-Based Approach For Rescuers
No full textTurkiye’s location on multiple major tectonic fault lines and Istanbul’s dense urbanization puts the country and particularly Istanbul at major risk of a devastating earthquake. As recently as 2023, the Kahramnmaraș earthquake affected 11 cities, claiming the lives of at least 50,000 people, prompting the deployment of over 11,000 search and rescue (SAR) personnel. SAR operators work in harsh climates and often have limited access to communication and technical tools for rescue. Our project seeks to propose a system for cellular tracking in an infrastructureless environment, and identify a niche in SAR operations technology where this technology could be deployed. We accomplished these goals by researching existing cellular tracking to gather ideas for the new technology, interviewing SAR operators to identify the niche and their demands for future technologies, and creating proposal documents for technology recommendations. We found the best system to use was a portable base station with directional antennas for the SAR triage process, as our interviews identified triage to be the area of SAR with the largest room for improvement. Essential characteristics for the device, as specified by SAR operators, include producing no false positives, resilience in all weather conditions, and a long battery life. Our project's findings aim to speed up the triage process especially for local SAR teams to increase the number of lives saved