Chalmers Open Digital Repository
Not a member yet
26247 research outputs found
Sort by
Skalmodell av elvägssystem: Konstruktion och design av skalmodell
No full textOmställningen till eldrivna fordon medför ökade krav på eldistribution och laddningsinfrastruktur. Ett sätt att möjliggöra längre körsträckor är konduktiv laddning via elektrifierade vägar, vilket undersöks i detta projekt genom utveckling av en skalmodell. Skalmodellen som baserats utifrån en inköpt bilbana modifierades och användes för att samla in mätdata, samtidigt som en simuleringsmodell anpassades för att efterlikna systemets beteende. Resultaten visar att beroende på hur energiförsörjningen och bilarna utformas kan den förväntade körsträckan öka, vilket belyser potentialen hos denna teknik i framtida transportsystem
Remaining useful life classification of ECUs in trucks using a transformer encoder model
No full textOriginally developed for natural language processing, transformer models have achieved state-of-the-art results in tasks such as machine translation and text classification. This has led to increasing interest in applying the transformer architecture to sequential data across multiple other domains. This thesis takes a binary classification approach to investigate whether a transformer encoder model can be used to classify the remaining useful life of electronic control units (ECUs) in Volvo trucks. The model is trained on operational data and faults related to the ECU, to predict whether an ECU is likely to fail within the following three years. The performance of the transformer model is evaluated against traditional machine learning classifiers, including logistic regression, LGBM, Extra Trees, and Random Forest. In addition to standard metrics, a custom cost metric is introduced to reflect the real-world impact of false positives and false negatives. Results show that the transformer encoder outperforms traditional models across all evaluation metrics, particularly when used with ensemble methods. However, the transformer encoder still underperformed compared to a naive classifier on the custom cost metric. This work serves as a starting point for improving the decision-making process in ECU refurbishment
Rare Events in Reaction-Diffusion Systems Field-theoretical Approximations and Monte Carlo Simulations
No full textRare events are events that have near-zero probability of occurring. Despite their
apparent irrelevance, when they do occur, they can have substantial, and even catastrophic,
repercussions. In this thesis, we study rare events in reaction-diffusion systems,
a class of mathematical models that finds various applications in physics and
life sciences. We employ both theoretical and computational methods to determine
the tails of the probability distribution describing the state of system.
Firstly, we follow existing literature to derive a quantum-mechanical description for
entirely classical reaction-diffusion systems, called the Doi-Peliti formalism. We express
the time evolution of the systems as a Feynman path integral, which we then
evaluate at the saddle point to obtain a semiclassical approximation for the probability
distribution, and a closed-form leading-order expression for the tails.
Secondly, we tailor a lesser-known Monte Carlo algorithm for rare probability estimation,
called adaptive multilevel splitting, to compute the probability distribution
of reaction-diffusion processes. We derive some theoretical results regarding its efficiency,
discuss practical implementation choices, and benchmark its performance
against well-understood examples.
Lastly, we compare the semiclassical approximation to the computational results,
determining under which conditions the former succeeds or fails
Testing of Field Grading Materials For HVDC Cable Joints
No full textAbstract
HVDC cable systems are designed to transmit large amounts of electricity over long distances with minimal losses. Most failures in these systems happen because of issues with cable accessories like joints and terminations. They are caused by strong electric fields appearing in such components due to geometrical features of the designs and inhomogeneities in the structure of the insulation system incorporating various materials. To enhance the performance of cable accessories, electric field control in the insulation is crucial. To address this, so-called field grading materials (FGMs) can be used to even out the electric field distributions. Thanks to their non-linear properties, these materials secure normal operations of the cable system and also are capable of handling events like lightning overvoltage and switching impulses. This thesis focuses on exploring electrical and mechanical properties of FGMs by testing samples of materials with different types and concentration of fillers at various temperatures and electric field strength. The focus is on a comparative analysis, where the properties of EPDM (Ethylene Propylene Diene Monomer) are taken as a reference. New materials with good electrical and mechanical properties are selected based on specific criteria, followed by measurements of their non-linear conductivity. Experimental results show that three selected materials met the criteria for
nonlinear conductivity. To further evaluate their performance as a base for building a large scale component, electrothermal simulations of a 525 kV DC cable joint made of those materials have been conducted for different voltage levels including nominal, type test and lightning impulse voltages. The simulation results show minimal differences in the field grading effects caused by the selected FGMs applied in different locations of the cable joint, likely because of similar non-linear behavior of their properties in the studied ranges of temperatures and electric stresses
Artificiell Intelligens användning inom drönarskanning och livedata: En kartläggning över befintliga modeller och branschperspektiv
No full textDenna studie undersöker integrationen av artificiell intelligens (AI) och drönarteknik i
Pontarius arbetsflöden, med fokus på att förbättra inspektioner av infrastruktur, såsom
asfalt och byggnader. Genom att kombinera kvalitativa och kvantitativa
forskningsmetoder utforskar projektet både tekniska och organisatoriska utmaningar
kopplade till implementeringen av Artificiell Intelligens. En fallstudiedesign används
för att ge en djupgående förståelse av de specifika förutsättningarna för en effektiv
AI-användning.
Forskningen bygger på semistrukturerade intervjuer med nyckelintressenter, tekniska
dataanalys av AI-modeller som DeepCrack, samt granskning av interna arbetsflöden
och riktlinjer. Denna metodkombination möjliggör en omfattande förståelse för
potentiella hinder och möjligheter vid AI-integrering. Genom att säkerställa validitet
och tillförlitlighet via metodtriangulering levererar studien handlingsbara insikter som
kan överföras till andra organisationer med liknande utmaningar.
Resultaten syftar till att bidra till utvecklingen av anpassningsbara och skalbara
metoder för teknikimplementering, tillämpliga inte bara för Pontarius utan även för
andra sektorer. Studien förväntas ge värdefulla perspektiv på både den tekniska
effektiviteten och den organisatoriska beredskapen som krävs för en framgångsrik
AI-integrering, och därigenom främja innovation och förbättrad
infrastrukturförvaltnin
Generating Representative Driving Cycles
No full textThe transportation industry is rapidly evolving. Customer expectations and environmental sustainability demands are quickly shifting, making performance optimisation of Heavy-Duty Vehicles (HDVs) increasingly critical. An essential aspect of the optimisation process is having Driving Cycles (DCs) that are representative of real-world driving patterns for accurate vehicle verification and validation.
In this thesis, different advanced data analytics methods were implemented and evaluated on their ability to generate DCs representative of real-driving patterns of HDVs. The implemented methods belonged to three main categories explaining the general technique of how DCs are constructed: Speed Acceleration State (SAS) methods, Micro-Trip (MT) methods, and Kinematic Segment (KS) methods.
To quantitatively assess method effectiveness, two main metrics were used: Characteristic Parameters (CPs), and Speed Acceleration Probability Distribution (SAPD). The CPs describe different statistical characteristics of driving behaviour which were compared between generated cycles and the operational data. Additionally, a comparison between generated cycles and the Vehicle Energy Consumption Calculation Tool (VECTO) was also made to provide a performance baseline. CPs were evaluated using Relative Difference (RD), while SAPD was evaluated with RD and Earth Mover’s Distance (EMD). EMD measures distribution dissimilarities, and its addition to the evaluation offers a more reliable SAPD assessment than what has been done in previous research.
The results showed that all implemented methods outperform the VECTO baseline both in terms of CP and SAPD representativeness, highlighting the need for fine-tuned cycles. The SAS and MT methods demonstrated superior performance compared to the KS methods. Especially in terms of SAPD representativeness and computational efficiency. While the KS methods showed significant limitations, the SAS and MT methods achieved highly promising CP and SAPD representations of the operational data. The SAS and MT methods were concluded as viable methods for DC generation and are the methods suggested to continue researching.
The thesis aims to serve as a DC generation framework that future researchers can follow, including detailed descriptions of all necessary methodological steps. The framework details a systematic approach for creating representative DCs adaptable for any driving profile or vehicle type
A Biometric Recognition-Based Authentication System: Development of a Secure Prototype Focusing on Facial Recognition
No full textTraditional password-based authentication is increasingly inadequate to protect online services, as it remains vulnerable to brute-force attacks, credential theft, and social engineering tactics. These threats are further exacerbated by common user behaviors, namely password reuse across different platforms and weak password selection. In response, facial recognition has emerged as a promising alternative, offering an intuitive and user-friendly form of authentication. However, existing facial recognition systems present their own challenges, including vulnerability to spoofing, privacy concerns, and inconsistent performance in different environments. This thesis presents the design and implementation of an authentication system, with a primary focus on a facial recognition system developed to address these limitations. A comparative analysis of available models identified YuNet as the most effective model for face detection, and FaceNet for face recognition. The system architecture incorporates a Flask-based backend with a responsive and user-friendly frontend interface, enabling secure user registration and login. To enhance accessibility, multiple alternative login methods have been incorporated, allowing users without camera access to authenticate securely. Additionally, to strengthen data security and privacy, all sensitive user information is securely encrypted before being stored. The system’s performance was evaluated using standard classification metrics and carefully optimized to achieve an effective balance between security and usability. The result is an authentication framework that addresses both user convenience, modern security demands, and ethical concerns
Strategier för att minimera översvämningar i områden med begränsad mark; en utvärdering av underjordiska dagvattenmagasins effektivitet och ytbehov
No full textPå grund av de rådande klimatförändringarna blir stadsplanering och fokus på hållbara städer
alltmer viktigt för att hantera konsekvenserna av mer intensiva regnfall i städerna. Ett stadsideal
till hållbara städer är förtätning. Detta examensarbete fokuserar på dagvattenhantering eftersom
översvämningar är en av de konsekvenser av klimat i förändring som påverkar täta stadsmiljöer.
Syftet med examensarbetet är att belysa strategier för att minimera översvämningar i områden
med begränsad mark. Vidare kommer tre typer av underjordiska dagvattenmagasin
dimensioneras i ett referensområde i Smålandsstenar och utvärderas med avseende på
effektivitet och lämplighet. Studien bygger på faktasamling, beräkningar, simuleringar i ScalGo
Live samt jämförelser mellan de underjordiska dagvattenmagasinen. För att behandla
rapportens syfte kommer vi att tillämpa de tre mest förekommande underjordiska
dagvattenmagasinen i branschen på referensprojektet. Fastigheten som används i
referensprojektet har haft stora problem med vattensamlingar som har lett till översvämningar.
Fastigheten har också utmaningen att det finns begränsad mark för att implementera en lösning
för dagvattenhantering. Rapporten avgränsar sig till att utvärdera de dagvattenmagasin som är
mest förekommande i branschen. För att utvärdera dessa har de dimensionerats för att hantera
den mängd vatten som beräknats vara förväntad maxvolym på fastigheten. Dagvattenmagasinen
har sedan jämförts med hjälp av jämförelsefaktorerna: anläggningsdjup, anläggningsyta,
marktäckning och livslängd. Jämförelsefaktorerna samt magasinens för- och nackdelar har
analyserats för att förstå dess effektivitet och hur väl de lämpar sig på områden med begränsad
mark. Andra lösningar utöver underjordiska dagvattenmagasin har även diskuterats för att
jämföra med underjordiska dagvattenmagasin.
Resultatet visar på att underjordiska dagvattenmagasin fungerar väl i urbana områden med
begränsad tillgång till mark. Detta eftersom de är effektiva i förhållande till dess
jämförelsefaktorer; anläggningsdjup, anläggningsyta, marktäckning och livslängd. Dess
effektivitet baseras även på att de kan hantera erforderlig maxvolym. Dock är inte samtliga
underjordiska dagvattenmagasin som utvärderats lämpliga för referensområdet på grund av att
regnbäddar inte uppfyller kraven på framkomlighet. Ytterligare en slutsats som har dragits är
att dagvattenmagasin är en del av lösningen för problematiken med översvämningar i urbana
områden men dessa behöver kombineras med andra åtgärder och strategier. Detta eftersom
olika lösningars lämplighet skiljer sig mellan olika områden. Rapportens resultat skapar bättre
beslutsunderlag inför val av dagvattenmagasin. Vidare kan det konstateras att det krävs att
dagvattenhantering är en central del av städernas klimatanpassningsplaner för att uppnå en
hållbar stad i framtiden, samt att åtgärder för dagvattenhantering behöver implementeras i redan
befintlig bebyggelse
Design Optimization of Balancing Holes in a Centrifugal Pump
No full textCentrifugal pumps are widely used in fluid transport systems due to their ability to deliver high flow rates in a compact, low-maintenance design. Despite their welldefined performance characteristics, a major operational challenge is the generation of axial force, a force imbalance caused by the pressure difference across the front (suction) and back (pressure) sides of the impeller. If the axial force is not properly investigated during the design process, it can cause bearing failure, increase maintenance requirements, and significantly reduce both hydraulic efficiency and pump lifespan. Balancing holes are commonly introduced to mitigate axial force by reducing the pressure imbalance. While effective in reducing axial force, balancing holes introduce fluid recirculation from the pressure side to the suction side of the impeller, leading to a reduction in hydraulic performance. Therefore, identifying an optimal balancing hole configuration is essential to minimize axial force without compromising overall efficiency.
This study investigates the development of a multi-objective target function using both scalarization and Pareto-based methods to identify an optimal balancing hole design configuration that reduces axial force while maintaining the hydraulic performance. CFD simulations were carried out using a steady-state approach with implicit local time stepping, in conjunction with a non-linear k−ε turbulence model and the frozen rotor approach, to capture the internal flow behaviour and to extract performance metrics—head, power, and axial force—for all generated configurations during the optimization process. To explore the design space effectively, Sobol sequence sampling was used to generate a diverse set of initial configurations. The target function was formulated by normalizing head and power at the Best Efficiency Point (BEP) and axial force across part-load, BEP, and over-load conditions, then iteratively optimized using Bayesian optimization and target function scoring to identify an optimal balancing hole design. A Pareto front analysis was finally conducted to analyze the trade-offs between hydraulic performance and axial force reduction for the generated configurations throughout this study.
The findings revealed that the formulated target function using scalarized method effectively guided the optimization process toward configurations that reduced axial force by up to 40.59 % while maintaining head and power within 99.98 % and 99.87 % of the reference, respectively. Iterative Bayesian optimization and Pareto front analysis consistently led to a key design feature, which was also supported by the flow field visualization
Emerging Architectures for Chemical Language Modeling
No full textIn recent years, language modeling architectures have become increasingly prominent
in the field of generative chemistry, offering new approaches for the de novo
design and optimization of small molecules. This thesis presents a comparative
study of two emerging architectures: the decoder-only Transformer and the Mamba
architecture, and a conventional Recurrent Neural Network with LSTM cells. The
investigation explores how choices in training data, including a targeted medicinal
dataset (ChEMBL) and a chemically broad dataset (PubChem), as well as data
augmentation via randomized SMILES representations, influence generative capacity
and chemical space coverage. In addition to this, task-specific optimization of
models through reinforcement learning is studied, and the models are compared with
respect to their ability to generate diverse molecules with desired properties.
Through pretraining experiments, it is shown that while the Mamba and RNN architectures
reach their optimum performance significantly faster, the decoder-only
Transformer achieves the highest validity and uniqueness in molecular generation.
Training on PubChem, as opposed to ChEMBL, generally enhances validity and
uniqueness but tends to reduce novelty, indicating a trade-off between chemical
space saturation and innovation. As for data augmentation through randomization
of SMILES, this helped all models refrain from memorizing the dataset, resulting in
higher novelty across architectures and datasets.
Reinforcement learning experiments further reveal that all three architectures are
capable of optimizing toward specific molecular properties, with the decoder-only
Transformer and Mamba each exhibiting distinct strengths depending on the optimization
task. Regarding the pretraining condition’s effect on reinforcement learning,
ChEMBL-trained models outperformed those trained with PubChem on multiple
tasks, and all architectures, but especially Mamba, benefitted from being pretrained
with randomized SMILES. Notably, even reduced-parameter models, such
as a downsized decoder-only Transformer variant, perform competitively relative to
larger architectures