Publikationer från Uppsala Universitet
Not a member yet
127603 research outputs found
Sort by
Uncertainty Quantification in Simulation-Based Inference Using Deep Generative Models
No full textThis thesis explores uncertainty quantification in simulation-based inference (SBI) using deep generative models. By focusing on probabilistic frameworks, the study addresses the inverse problem inherent in SBI, where traditional likelihood functions are often intractable. Four generative models are examined: two Conditional Variational Autoencoders (cVAEs), one Masked Autoregressive Flow (MAF), and one Neural Spline Flow (NSF). These models are evaluated for their effectiveness in learning benchmark posterior distributions. While individual assessments are conducted, all models are compared using the Maximum Mean Discrepancy (MMD) metric, with the MAF model demonstrating the best overall performance. The work concludes by suggesting future research directions involving alternative generative models and evaluation metrics
A comparitive study of energy consumption in Zig and C in embedded systems and its effects on green IT.
No full textDenna studie jämför energieffektiviteten hos programmeringsspråken Zig och C för inbäddadesystem, med fokus på deras potential att bidra till grön IT. Genom att utföra ett kvantitativtexperiment på en Raspberry Pi Pico, där identiska algoritmer implementerades i båda språken,mättes energiförbrukningen under exekvering. Resultaten visade inga statistiskt signifikantaskillnader mellan maskinkod genererad av Zig och C, vilket tyder på att båda språken presterarlikvärdigt ur ett energiperspektiv. Slutsatsen är att valet mellan Zig och C för energieffektiva inbäddade system bör baseras påandra faktorer som inte är energiprestanda. Studien bidrar till forskningen om grön IT genom attbelysa hur programmeringsspråken som undersöks är likvärdiga alternativ för användning ienergieffektiva inbäddade system och ger även en grund för framtida forskning kringenergieffektivitet i inbäddade system. This study compares the energy efficiency of the programming languages Zig and C forembedded systems, focusing on their potential to contribute to Green IT. Through a quantitativeexperiment conducted on a Raspberry Pi Pico, where identical algorithms were implemented inboth languages, energy consumption during execution was measured. The results showed nostatistically significant differences between Zig and C, indicating that both languages performsimilarly from an energy perspective. The conclusion is that the choice between Zig and C for energy-efficient embedded systemsshould be based on factors other than energy performance. The study contributes to green ITresearch by highlighting how the examined programming languages are equivalent alternativesfor creating energy-efficient embedded systems and also furthermore provides a foundation forfuture research on energy efficiency in embedded systems
Straightness Evaluation of Steel Bars using Image Analysis Algorithms
No full textEnsuring the straightness of steel bars is a critical aspect in quality control in the steel industry, as even minor straightness deviations can lead to costly returns. Today, manual evaluation of straightness is done, which is subjective and error-prone. This thesis explores the development of an automated, camera-based solution for evaluating the straightness deviation of steel bars, with the aim of replacing manual inspection with a more objective and accurate system.The approach in this project combines image processing techniques with a dual-camera setup to capture bars up to six meters in length. A reliable image stitching method is required to combine images from the cameras into one, unified view for analysis. Stereo stitching, based on chessboard calibration, proved successful in creating geometrically accurate stitched images, though only within a specific region. Segmentation of steel bars from the background is the next step, where a single-frame background subtraction method was tested and evaluated across multiple configurations. While it achieved moderate accuracy (F1-scores up to 0.82), the accuracy was not enough for such a case, where pixel-level accuracy is needed for reliable straightness deviation measurements. Instead, a previously developed mask generation algorithm was used to segment the bars and use them for final measurements. The straightness deviation was calculated by measuring the maximum distance between the bottom edge of steel bars and a line drawn between the ends. Analysis of eight test steel bars showed that accurate measurements could be obtained within a specific region of the image, with an almost perfect correlation between the estimated and on-site measured values in that region. This work demonstrates the feasibility of automated straightness evaluation using image analysis algorithms. While the proposed solution is limited to a small region, it provides a strong foundation where some adjustments and installments can make it into a robust, accurate automated system for evaluation of steel bar straightness
Automated Inspection of Printed Circuit Board Assemblies : A project with Unibap Space Solutions
No full textThe purpose of this project was to develop a software program to automate and improve the incoming inspection of Printed Circuit Board Assemblies at the production facility of Unibap Space Solutions. This program first generates a binary segmentation mask of a Printed Circuit Board Assembly image using a trained DeepLabv3+ model. As an extension of the network DeepLabv3, the trained model combines Spatial Pyramid Pooling and an encoder-decoder. Once the segmentation is complete, the program performs a pixel-wise comparison of the binary mask against a reference template. The final output is a binary image in which differences and potential defects are marked in red. After the program’s inspection, it falls upon the user at Unibap to interpret the result and decide whether the Printed Circuit Board Assembly passes or fails the inspection. The main difficulty encountered in this project was achieving a sufficiently accurate segmentation. A dataset of images from various sources was used in training. However, evaluation metrics such as the F1 score and dice loss indicate that the model’s performance is not yet adequate for production use. For a model to be considered accurate a F1 score close to one is to be preferred and the F1 score obtained by the model was approximately 0.6. Retraining of the model with a more extensive dataset is suggested for increased accuracy. An automated inspection and defect detection program is highly desirable, as it has the potential to generate both financial and environmental benefits. However, such a program also carries the risk of unintended negative consequences. Therefore, this report includes a reflection on the project’s impact on a number of the Sustainable Development Goals
Timetable Optimization of ERTMS/ETCS Hybrid Train Detection
No full textThe introduction of Hybrid Train Detection, a modern signaling system for railways, has been widely discussed in the literature, highlighting its potential advantages and challenges. A key aspect of realizing the capacity benefits of this system lies in the optimization of train timetables. This project develops and implements a Mixed-Integer Linear Programming (MILP) model, augmented by simulation techniques, to analyze and optimize train schedules. By addressing this critical aspect, the research aims to provide insights into the operational improvements that Hybrid Train Detection can enable, contributing to more efficient railway management. Results show that the optimization framework can reduce delays by up to 50%, compared to today’s timetable, using moderate amounts of supplement time. However, improvements in congested scenarios are limited, primarily due to inaccurate modeling of knock-on delays and simplified assumptions about train headways. Additionally, while Hybrid Train Detection shows theoretical capacity benefits on open line segments, station bottlenecks and infrastructure constraints often prevent these gains from being realized in practice. These findings suggest that both optimization and Hybrid Train Detection offer promising potential for improving railway operations, but their effectiveness relies heavily on refined modeling and real-world calibration
Adult mental health outcomes of adolescent depression and co-occurring alcohol use disorder : a longitudinal cohort study
No full textDepression and alcohol use disorder (AUD) are frequently co-occurring in adolescence, which often goes undetected in routine care. While this may potentially compromise treatment effectiveness and lead to a less favourable long-term prognosis, few longitudinal studies have followed this group into adulthood. The aim of this study was to explore the risk for adult depression, anxiety disorders, suicidality, and AUD in adolescents with concurrent depression and AUD. The study was based on the Uppsala Longitudinal Adolescent Depression Study (ULADS), a Swedish prospective cohort study. Diagnostic interviews were conducted in adolescence (age 16-17) and adulthood (around age 30). Adolescents with concurrent depression and AUD (n = 38) were compared with peers having only depression (n = 189) or neither of the conditions (n = 144). Logistic regression was used to calculate adjusted odds ratios (aORs) with 95% confidence intervals (CIs). Adolescents with concurrent depression and AUD were more likely than their non-affected peers to experience adult depressive episodes (aOR, 5.33; 95% CI, 2.22-12.83), anxiety disorders (4.05; 1.77-9.27), suicidality (5.37; 2.28-12.66), and AUD (7.68; 2.59-22.81). Notably, 34% of adolescents with both depression and AUD subsequently experienced both these conditions as adults, compared to 7% of adolescents with only depression. Adolescents suffering only from depression were less likely than those with both conditions to experience suicidality (0.44; 0.21-0.95) and AUD in adulthood (0.18; 0.07-0.44). These findings underscore the clinical imperative to identify adolescents with this comorbidity. Recognition of the poor long-term prognosis can inform targeted interventions for this vulnerable group, ultimately improving health and well-being throughout the life course
Antimicrobial Resistance in Refugee Settings : Cultural and Operational Barriers to Stewardship and Implications for Humanitarian Health Systems
No full text Antimicrobial resistance (AMR) is an urgent and growing threat to global and humanitarian health systems, particularly in refugee settings where overcrowding, poverty, and fragile infrastructure accelerate the spread of resistant infections. This thesis explores the cultural and operational barriers to antimicrobial stewardship (AMS) in these contexts through a thematic synthesis of existing qualitative research and grey literature. Guided by the WHO health systems building-blocks framework and informed by a people-centred health perspective, the analysis identifies four key barriers that undermine stewardship efforts: gaps in community knowledge and health beliefs, systemic limitations in the health workforce, restricted access to healthcare services, and governance failures. The study reveals how social determinants of health, such as education, living situations and legal status, shape both access to antibiotics and patterns of use. The thesis argues that humanitarian actors have a pivotal role in adapting AMS strategies to the realities of displacement. This includes developing culturally appropriate health education tools, supporting task-sharing within overstretched health workforces and integrating AMS with broader health interventions. It calls for renewed urgency and cross-sector collaboration to prevent AMR from evolving into a widespread humanitarian crisis, especially in already marginalised populations
Mellan maskin och människa : Revisionsbranschens utmaningar i ett nytt tekniklandskap
No full textImplementeringen av generativ AI i organisationer har förändrat arbetsprocesser och de arbetskrav som ställs på arbetskraften. I uppsatsen undersöktes hur generativ AI implementerats i revisionsprocessen och vilka utmaningar kopplat till kompetens det kan ha givit upphov till. Det möjliggjordes genom att analysera hur arbetskraven upplevs har förändrats till följd av implementeringen av generativ AI och hur organisationens kompetensutvecklingsarbete har uppfattats hantera förändringen. Undersökningen tog sin form av en fallstudie som utgjordes av intervjuer med team leaders för att belysa implementeringen ur ett strategiskt perspektiv. Resultatet av intervjuundersökningen visade att arbetskraven upplevs har förändrats informellt och att organisationens kompetensutvecklingsarbete uppfattats bristfällig. Implementeringen av generativ AI i revisionsprocessen kan ha givit upphov till utmaningar i form av kompetensgap, vilket kan riskera att försämra revisionen ur ett kvalitet- och säkerhetsperspektiv
Resultatmanipulation vid Nyemissioner : En jämförande studie mellan Nasdaq Stockholm och Nasdaq First North
No full textDenna studie undersöker förekomsten av resultatmanipulation i samband med företrädesemissioner för svenska bolag noterade på Nasdaq Stockholm och Nasdaq First North samt skillnader i resultatmanipulation mellan dessa två marknadsplatser. Studien använder diskretionära periodiseringar som proxy för resultatmanipulation, där kvartalsdata för perioden 2018-2024 analyseras. Resultaten visar att företag på Nasdaq Stockholm tenderar att uppvisa positiva diskretionära periodiseringar i kvartalet före emissionen, medan företag på Nasdaq First North inte visar något tydligt samband gällande diskretionära periodiseringar kvartalet före emissionen. En multipel regressionsanalys visar att skillnaderna beträffande användandet av diskretionära periodiseringar mellan marknadsplatserna för kvartalet delvis kan förklaras av bolagsstorlek och emissionsstorlek. Även resultat som pekar på resultatmanipulation efter genomförd emission analyseras och diskuteras. Studien bidrar till förståelsen kring hur marknadsstruktur påverkar redovisningsbeteende i samband med kapitalanskaffning