Chalmers Open Digital Repository
Välkommen till Chalmers öppna digitala arkiv!
Här hittar du:
- Studentarbeten utgivna på lärosätet, såväl kandidatarbeten som examensarbeten på grund- och masternivå
- Digitala specialsamlingar, som t ex Chalmers modellkammare
- Utvalda projektrapporter
Forskningspublikationer, rapporter och avhandlingar hittar du i research.chalmers.se
Enheter i Chalmers ODR
Välj en enhet för att se alla samlingar.
Senast inlagda
Annotation-free Learning for Sensor Fusion in ADAS
(2025) Björkman, Maria; Tvingby, Ludvig
Vehicle automation has the potential to significantly improve road safety. Achieving
comprehensive vehicle perception requires systems that optimally combine information
from multiple sensor modalities. Such systems leverage the strengths of each
modality while compensating for their weaknesses. By continuously encoding and
fusing information from cameras, LiDARs, RADARs and the motion of the egovehicle,
a dynamic representation of the surrounding environment can be created
and maintained.
A major challenge for these systems is the large amount of annotated data required
for training, as manual labelling creates a significant bottleneck for scalability. In
this study, a pre-training task for a multi-modal machine learning model was implemented
and evaluated. To circumvent labour-intensive labelling, self-supervision
was employed, with both the model input and the supervision signal involving
annotation-free data. The pre-training aimed to learn general features related to
sensor pose changes by predicting ego-vehicle pose changes using odometry data.
To assess pre-training performance, the features were then used as initial weights
for fine-tuning a perception model. The performance of the perception model using
baseline weights trained on annotated data was similar to that using weights trained
on annotation-free data, indicating that the proposed method is viable. However,
further testing is required to establish statistical significance. Future work could
explore implementing attention-based methods for feature matching between scene
representations to improve model performance.
Simulations of Dielectric Frequency Response of Bushings for a Non-Destructive On-site Defect Identification
(2025) Kulandaivel, Nigalyadevi
Bushings are integrated components of high voltage transformers providing connections of windings to external circuits outside the transformer shell while providing mechanical and insulation support. During normal operation, bushings are consistently influenced by operating voltage, load current, and voltage stress due to transient over-voltages during natural or switching phenomena in the power system network. These stresses gradually degrade the bushing insulation and eventually cause failure, The insulation state should be monitored and maintained periodically. Dielectric Frequency Response (DFR) measurements yielding capacitance and loss factor values in frequency domain is one of the most popular diagnostics methods, which provides fruitful information about insulation conditions and possible defects that may be present within the bushing insulation. The aim of the thesis project is to develop a transformer bushing model using COMSOL Multiphysics to perform simulations of DFR for a real scale transform bushing geometry. The model was implemented based on electrostatics physics and current continuity through the insulation structure. The loss factor and the capacitance values were computed in the frequency window typical for practical measurements. The results of the simulations conducted using the developed model are validated by comparing them with measured DFR data. Furthermore, possible defects, which
may appear in practice (conductive layers on the insulation, gas bubbles in the insulation bulk) were introduced in the model and frequency dependent loss factor and capacitance values were computed for each type of defects. The results are compared with the reference (defect free) case to identify and interpret the dielectric response behavior. The sensitivity study conducted by varying the properties of the defects indicate that the developed model provides a tool for capturing presence of defects
in the insulation by analyzing changes in the DF response.
Validating a Sea Ice Drift Retrieval Algorithm Through Remote Sensing and In-Situ Buoy Comparisons
(2025) Desjonqueres , Thibault
This Master’s thesis validates a sea ice drift retrieval algorithm developed at Chalmers
(referred to as the Chalmers algorithm) using in situ buoy data collected during the
ARTofMELT 2023 cruise in the Fram Strait from May to June 2023.
The ARTofMELT 2023 expedition aimed to document melt onset in the Arctic
Ocean, collecting extensive data on sea ice dynamics near the Marginal Ice Zone
(MIZ) with seven OpenMETbuoys and three SIMBA buoys. These buoys were
deployed on first-year ice floes, with detailed measurements of horizontal motion,
rotation, deformation, wave action, and temperature. This deployment enabled
analysis of ice floe movement, convergence, divergence, and the interaction between
these dynamics and SAR imagery.
Low-noise Radarsat Constellation Mission (RCM) and Sentinel-1A SAR images,
captured in dual polarization (HH+VV or HH+HV), were acquired to match the
study area both spatially and temporally. Integrating in situ weather data, wave
measurements, sea ice concentration, and weather reanalysis with satellite imagery
observations of ice conditions around the buoys enabled a thorough interpretation
of drift patterns, floe decay, and wave signatures associated with the buoys.
The Chalmers drift algorithm was successfully applied to two pairs of RCM images,
each with a 10 m resolution, out of eight total image pairs. These pairs were cho sen based on criteria including a time gap of 1-24 hours between images, identical
resolution, and a maximum 5-minute time gap between the buoy and the images.
The remaining six pairs contained Sentinel-1A images with a resolution of 100-150 m.
Although the drift estimates from the Chalmers algorithm were not statistically
significant, they demonstrated the algorithm’s capability to produce coherent and
realistic drift estimates, evidenced by the alignment of drift vector azimuths with ac tual buoy drift and realistic sea ice drift distances. To better assess the algorithm’s
performance, future work should test it on a larger dataset, including the six re maining Sentinel-1A image pairs. Additionally, exploring the effects of increased
time gaps between buoy and satellite images, and between the two satellite images,
would provide further insights into the algorithm’s capabilities
Machine Learning for Wind Power Prediction. A Comparative Analysis of Traditional Machine Learning Models and Graph Neural Network for Wind Power Prediction and Forecasting in Wind Farms
(2025) Antonsson, Emma; Vind, Lisa
This thesis presents a comparative study of machine learning (ML) models and the deep learning (DL) model graph neural network (GNN) for wind power prediction and short- to medium-term forecasting in wind farms. Using high-resolution SCADA data from a 16-turbine onshore wind farm in Sweden, along with re-analysis and forecast weather datasets, various models including Random Forest (RF), eXtreme Gradient Boosting (XGBoost), k-nearest neighbours (kNN), Multi-Layer Perceptron (MLP), and GNN were trained and evaluated. Two baseline approaches, the farm’s
theoretical power curve and FLORIS wake model, were used as references. Results show that ML models outperform baseline models in predicting wind power output, with GNN achieving the best overall performance, although all ML models perform similarly. The ability of the models to generalize from wind power prediction to forecasting is however limited. The findings indicate that re-analysis data with low spatial resolution fails to adequately capture local weather conditions necessary for accurate power prediction. The study also investigates the effects of input feature selection, temporal resolution, and multi-task learning on model performance.
Furthermore, it identifies challenges related to input data quality, particularly in the estimation of global wind conditions from SCADA-based measurements. These results underscore the potential of ML methods for wind power applications and highlight the critical importance of accurately representing global weather data, as well as accounting for discrepancies between training data and forecast data.
Steering behavior-based fatigue detection: Evaluation and implementation for drowsy driver warning system
(2025) Hassan Ananda Kumar, Sanjana
This research was conducted to analyze steering behavior as an alternative approach
to detecting driver drowsiness. The thesis examined standard metrics such as steering wheel angle, steering wheel angle rate, yaw rate, and lane position—variables
directly and indirectly related to steering behavior—to assess how much information
they bear about drowsiness. A derived metric, the steering reversal rate, was also
analyzed to further explain the effects of drowsiness. These metrics demonstrated
a strong correlation with driver drowsiness, which was subjectively measured using
the Karolinska Sleepiness Scale (KSS).
Based on this analysis, two parameters derived from the steering reversal rate—micro
and macro corrections—were used to develop two methods for detecting drowsiness.
These parameters were significant because, as a driver becomes drowsier, the frequency of micro-corrections tends to decrease, while macro-corrections increase.
Two methods have been developed to detect a drowsy driver based on the above
analysis. The first method employed a logistic regression model, using the absolute
values of micro and macro corrections to directly correlate with the KSS ratings.
This approach did not account for temporal patterns and treated the data independently of its time-series nature. In contrast, the second method incorporated a
time-series perspective by evaluating changes in micro and macro correction rates
over time rather than relying on their absolute values. During development, it
was observed that vehicle speed significantly influenced steering behavior. At lower
speeds, even non-drowsy drivers exhibited more macro corrections and fewer microcorrections. However, at speeds above 65 km/h, non-drowsy drivers typically made
more micro-corrections and fewer macro-corrections. This insight enhanced the robustness of the second method, wherein vehicle speed was considered one of the
contributing factors in analyzing driver steering behavior. Additionally, the second
method involved a learning phase for each individual driver, allowing for personalized threshold values. This driver-specific calibration improved adaptability.
Overall, the second method of real-time analysis of changes in micro and macro
correction rates proved to be a more effective and reliable approach, yielding better
results than the current system(based on lane distance) for detecting driver drowsiness.