"uuid","repository link","title","author","contributor","publication year","abstract","subject topic","language","publication type","publisher","isbn","issn","patent","patent status","bibliographic note","access restriction","embargo date","faculty","department","research group","programme","project","coordinates"
"uuid:5cdae823-66c5-40e2-b624-29da6a3d8456","http://resolver.tudelft.nl/uuid:5cdae823-66c5-40e2-b624-29da6a3d8456","Automated Carnegie Staging of the Human Embryo in 3D Ultrasound using Deep Learning","Niemantsverdriet, Ruben (TU Delft Mechanical, Maritime and Materials Engineering)","Vos, F.M. (mentor); Bastiaansen, Wietske (mentor); Klein, Stefan (mentor); Rousian, Melek (mentor); Delft University of Technology (degree granting institution)","2024","The periconceptional period, encompassing the embryonic phase, is a critical window where a majority of reproductive failures, pregnancy complications, and adverse pregnancy outcomes arise. The Carnegie staging system comprises 23 stages which are based on embryonic morphological development. This allows for the assessment of normal and abnormal embryonic development during this critical period. In-utero Carnegie staging using three-dimensional (3D) ultrasound scans visualized with virtual reality offers valuable insights but is currently a time-consuming manual process. To address this, we propose a deep learning approach for Carnegie staging in 3D ultrasound scans.
We used a dataset comprising 1413 3D ultrasound scans from the Rotterdam Periconceptional Cohort, annotated with Carnegie stages spanning from stages 13 to 23, including fetal subjects. Various training strategies were explored. We compared a metric regression approach, which considers the ordered nature of the Carnegie stages by treating the Carnegie stages as a continuous variable, with a multi-class classification approach, treating stages as independent categories. Additionally, we evaluated the influence of using a loss function accommodating the categorical nature of the Carnegie stages in the metric regression approach and examined the impact of incorporating embryonic size in the model input. Ultimately, a regression approach using the Mean Squared Error (MSE) loss function emerged as the optimal choice.
This model achieved a classification accuracy of 0.59 and a Root Mean Squared Error (RMSE) of 0.62 on the test set. This performance is comparable to an intermediate human rater, which achieved an accuracy of 0.63 and a RMSE of 0.65. Our findings represent a significant step towards the development of an automated Carnegie staging method, offering the potential for a more comprehensive evaluation of the critical embryonic phase in the clinic.","Machine Learning; Deep Learning; embryonic development; 3D Ultrasound","en","master thesis","","","","","","","","","","","","Biomedical Engineering | Medical Physics","",""
"uuid:47c2baa6-b7a2-439b-b3c5-d0748c05621f","http://resolver.tudelft.nl/uuid:47c2baa6-b7a2-439b-b3c5-d0748c05621f","Meta-learning for few-shot on-chip sequence classification","den Blanken, Douwe (TU Delft Electrical Engineering, Mathematics and Computer Science)","Frenkel, C.P. (mentor); Makinwa, K.A.A. (graduation committee); Verhelst, M. (graduation committee); Delft University of Technology (degree granting institution)","2023","The growing interest in edge computing is driving the demand for more efficient deep learning models that fit into resource-constrained edge devices like Internet-of-Things (IoT) sensors. The challenging limitations of these devices in terms of size and power has given rise to the field of tinyML, focusing on enabling low-cost machine learning on edge devices. Up until recently, the work in this space was primarily focused on static inference scenarios. However, a prominent issue with this is that models cannot adapt post-deployment, leading to robustness issues with shifting data distributions or the introduction of new features in the data. However, at the edge, full on-device retraining, or communicating all new data to a central server, is infeasible: this necessitates the development of data-efficient learning algorithms to adapt locally and autonomously from streaming data. This challenge at the intersection of edge computing and data-efficient learning is currently an open challenge.
In this thesis, we propose to solve this challenge with meta-learning. To clarify in which way the application of meta-learning is the most suitable for edge hardware, for the first time, a principled approach for meta-learning at the edge is outlined and investigated in three parts.
The first part of this thesis details the selection of a suitable neural network architecture for few-shot learning over sequential data. By not being fixated on one architecture from the start, it is possible to explore different approaches to learning over sequences of temporal data, leading to the identification of the most effective architecture for generalizing from limited temporal examples. The quantitatively evaluated architectures are a recurrent neural network (RNN), a gated recurrent unit (GRU), a long-short-term memory (LSTM) and a temporal convolutional network (TCN). We show that TCNs outperform all architectures, while GRUs and LSTMs have a lower activation memory requirement. However, the latter require a linearly increasing number of multiplications with input sequence length, while it scales logarithmically for TCNs. Our results show that TCNs therefore provide the most favorable trade-off for low-cost temporal feature extraction at the edge.
The second part of the thesis focuses on the algorithmic developments of the few-shot learning setup. Building on recent results from machine learning research, we highlight how meta-learning techniques primarily rely on learning high-quality features that generalize well. Taking into account hardware-driven considerations such as memory and compute overheads and through detailed quantitative analyses, we demonstrate that the best performance-cost trade-off is reached with a simple supervised pre-training scheme, where on-chip learning is performed by comparing the outputs of a TCN-based feature extractor with Manhattan distance. We also analyze the impact of quantization on this trade-off and, accordingly, we select a scheme with 4-bit logarithmic weights and 4-bit unsigned activations...","meta-learning; Temporal Convolutional Neural networks; Edge Computing; digital design; Machine Learning; ASIC; silicon; Deep Learning; Edge AI","en","master thesis","","","","","","","","2025-09-28","","","","Electrical Engineering","","52.0022, 4.3736"
"uuid:13733886-4d6f-4d5b-bfe3-9cde77614f38","http://resolver.tudelft.nl/uuid:13733886-4d6f-4d5b-bfe3-9cde77614f38","Preoperative assessment of the histopathological growth patterns of colorectal liver metastasis on CT using artificial intelligence","van Gurp, Samuel (TU Delft Mechanical, Maritime and Materials Engineering)","Starmans, Martijn (mentor); Voigt, Kelly (graduation committee); Klein, Stefan (graduation committee); Vos, F.M. (mentor); Delft University of Technology (degree granting institution)","2023","Background: Histopathological growth patterns (HGP) are a biomarker for predicting survival and systemic treatment effectiveness in colorectal liver metastasis (CRLM). Currently, HGP assessment in CRLM requires the resection specimen. Predicting the HGP from preoperative medical imaging could allow more personalised care and better outcomes. Methods: 252 patients underwent CRLM resection between 2004 and 2018 without receiving any systemic treatment. Patients were characterised as having either pure desmoplastic growth (dHGP) or any other type of growth pattern combination (non-dHGP) (21% dHGP; 79% non-dHGP). These categories were chosen because pure desmoplastic growth is predictive of better overall survival. regions of interest were automatically extracted using a UNet based segmentation model. These ROIs were passed to a radiomics model and a deep learning model to classify between dHGP/non-dHGP and predict the fraction of dHGP. Results: The best-performing classification method was the radiomic approach achieving an AUC of 0.67 (95% CI: 0.58-0.78), whereas the best-performance deep learning model achieved an average AUC value of 0.59 (95% CI: 0.53-0.65). Additionally, regression predicting the fraction of dHGP failed, with the predicted values showing no significant correlation with the actual value. Conclusions: Radiomics can be used to assess HGP, however further improvements in predictive performance are needed before these methods can be applied.","Colorectal Liver Metastases; Histopathological Growth Pattern; CT scan; Machine Learning; Deep Learning; Radiomics","en","master thesis","","","","","","","","","","","","Biomedical Engineering | Medical Physics","",""
"uuid:0822c597-d3fc-4047-93f3-3bb850d14c13","http://resolver.tudelft.nl/uuid:0822c597-d3fc-4047-93f3-3bb850d14c13","Leveraging Related Datasets to Improve Model Performance on an Underrepresented Target Population","Ries, Maxmillan (TU Delft Electrical Engineering, Mathematics and Computer Science)","Tax, D.M.J. (mentor); Reinders, M.J.T. (graduation committee); Scharenborg, O.E. (graduation committee); Delft University of Technology (degree granting institution)","2023","Training deep learning models for time-series prediction of a target population often requires a substantial amount of training data, which may not be readily available. This work addresses the challenge of leveraging multiple related sources of time series data in the same feature space to improve the prediction performance of a deep learning model for a target population. Specifically, we focus on a scenario where the target dataset, representing the desired target population, is underrepresented, while the source datasets consist of mismatched populations that are sufficiently representative for training a deep learning model. In this study, we explore state-of-the-art techniques, including transfer learning, ensemble learning, and domain adaptation to leverage source datasets towards a target population using real-world medical data. Additionally, we investigate the use of model performance-derived baselines as a heuristic to quantify the magnitude of the distribution mismatch between a source(s) and a target. Our results demonstrate that a set of well-defined baselines can effectively quantify the distribution mismatch and provide insights into the choice of leveraging technique for a given mismatch scenario. Furthermore, our results show that all state-of-the-art techniques can be employed to leverage related source datasets towards the target, though the performance of these techniques varies depending on the characteristics of the distribution mismatch. Eventually, we discuss the applicability of this research to new scenarios, along with avenues for future research.","Machine Learning; Deep Learning; Timeseries data; Medical data; Data Science","en","master thesis","","","","","","","","","","","","Computer Science","",""
"uuid:55a159a7-461a-490e-bc73-5194c0ed3b4e","http://resolver.tudelft.nl/uuid:55a159a7-461a-490e-bc73-5194c0ed3b4e","Does text matter?: Extending CLIP with OCR and NLP for image classification and retrieval","Sassoon, Jordan (TU Delft Electrical Engineering, Mathematics and Computer Science)","Zhao, Zilong (mentor); Chen, Lydia Y. (mentor); Lukina, A. (graduation committee); Delft University of Technology (degree granting institution)","2023","Contrastive Language-Image Pretraining (CLIP) has gained vast interest due to its impressive performance on a variety of computer vision tasks: image classification, image retrieval, action recognition, feature extraction, and more. The model learns to associate images with their descriptions, a powerful method which allows it to perform well on unseen domains. Often, the descriptions fail to capture text which is contained within the image, a source of information which could prove useful for a handful of computer vision tasks. This limitation requires finetuning in domains where contained text is important. In fact, CLIP has mixed performance on Optical Character Recognition (OCR). This paper proposes a novel architecture: OSBC (OCR Sentence BERT CLIP), which combines CLIP and a custom text extraction pipeline, composed of an OCR model, and a Natural Language Processing (NLP) model. OSBC uses the text contained within images as an additional feature when performing image classification and retrieval. We tested the model on multiple datasets for each task, occasionally outperforming CLIP when images contained text, while maintaining finetunability, and improving the model's robustness. In addition, OSBC was designed to be generalizable, meaning it is expected to perform well on unseen domains without finetuning, though this was not achieved in practice.","zero-shot learning; Deep Learning; Machine Learning; Computer Vision; CLIP; Transformers","en","bachelor thesis","","","","","","","","","","","","Computer Science and Engineering","CSE3000 Research Project",""
"uuid:1be54eb0-34f0-4292-9e38-7cd5f7e27b32","http://resolver.tudelft.nl/uuid:1be54eb0-34f0-4292-9e38-7cd5f7e27b32","Feature extraction and classification on heart rate time series for cardiovascular diseases","Beekhuizen, Michael (TU Delft Electrical Engineering, Mathematics and Computer Science; TU Delft Pattern Recognition and Bioinformatics)","Reinders, M.J.T. (mentor); Naseri Jahfari, A. (mentor); Martinez, Jorge (graduation committee); Tax, D.M.J. (graduation committee); Ghorbani, R. (graduation committee); Delft University of Technology (degree granting institution)","2023","Cardiovascular diseases are one of the primary causes of mortality worldwide. Paroxysmal atrial fibrillation is a specific type that is difficult to detect and diagnose in a short time frame. To overcome this, we investigated if long-term wearable data can be used for the detection of heart diseases. The BigIdeasLab_STEP dataset and long-term Fitbit data from the ME-TIME study were used to examine this.
Our analysis showed a correlation between the window/stride size and accuracy when performing activity classification with the BigIdeasLAB_STEP dataset. Moreover, variability was found between subjects due to differences in the physical structure of their hearts. Normalization proved to be a crucial step to minimize the subject variability and significantly improved performance. Grouping subjects and performing classification inside a group helped to improve performance and decrease inter-subject variability. Integrating handcrafted features in deep learning networks also improved classification performance.
Analysis of the long-term Fitbit data showed that there is a difference between individuals based on their health condition. Classification of individual peaks was possible and worked best when utilizing a time series-specific support vector machine and grouping peaks together. Grouping peaks per week from a person and calculating a percentage of heart disease-predicted peaks also worked relatively well to distinguish between heart disease and reference subjects. Like with the previous dataset, normalization proved to be a crucial step to minimize subject variability.
The findings indicate that heart rate time series can be utilized for classification tasks like predicting activity or the detection of heart diseases. However, normalization and grouping techniques need to be chosen carefully to minimize the issue of subject variability.","Deep Learning; Machine Learning; Heart rate; clustering; Cardiovascular disease","en","master thesis","","","","","","","","","","","","Computer Science | Bioinformatics","",""
"uuid:f970b944-7912-4843-a1e5-1d55008d4a90","http://resolver.tudelft.nl/uuid:f970b944-7912-4843-a1e5-1d55008d4a90","Classification of primary liver tumors with radiomics and deep learning based on multiphasic MRI","Goedhart, Aisha (TU Delft Mechanical, Maritime and Materials Engineering)","Vos, F.M. (graduation committee); Starmans, M. P.A. (mentor); Klein, Stefan (graduation committee); Delft University of Technology (degree granting institution)","2023","Primary liver cancer is a commonly diagnosed cancer and accurate diagnosis is crucial for treatment planning. To differentiate between malignant and benign liver tumors, contrast-enhanced MRI is typically used as it provides information over multiple contrast phases. However, diagnosis based on MRI is challenging. In this study, automatic classification is used to distinguish common primary liver tumors.
Imaging data from 102 patients with malignant (hepatocellular carcinoma) and benign (focal nodular hyperplasia and hepatocellular adenoma) primary liver tumors was used for binary classification through radiomics and deep learning approaches. The radiomics method was applied with the use of the open-source toolbox WORC. The deep learning model was based on the ResNet-10 architecture. The data input consisted of individual and combined phases of contrast-enhanced T1-weighted and T2-weighted MRI.
The highest performance values were found for the radiomics approach that combined the precontrast, arterial, portal venous, and delayed contrast phases together with T2-weighted MRI, with an AUC of 0.92. The deep learning model scored an AUC of 0.83 with this data input, however substantial overfitting occurred due to the limited sample size.
In conclusion, the radiomics classifiers based on combined contrast-enhanced T1-weighted and T2-weighted MRI can differentiate malignant from benign primary liver tumors with limited data samples. The classification task is too complex with the given data when using a ResNet-10 model and should be applied to an extended dataset.","Radiomics; Deep Learning; ResNet; MRI; Post-contrast T1; Liver cancer; Machine Learning","en","master thesis","","","","","","","","","","","","Biomedical Engineering | Medical Physics","",""
"uuid:be08d8c2-4fd6-405b-8861-804985cbecd5","http://resolver.tudelft.nl/uuid:be08d8c2-4fd6-405b-8861-804985cbecd5","Vulnerability prealerting by monitoring the online repositories of open source projects","Westfalewicz, Andrzej (TU Delft Electrical Engineering, Mathematics and Computer Science; TU Delft Software Technology)","Proksch, S. (mentor); Bruntink, Magiel (mentor); Delft University of Technology (degree granting institution)","2023","Software security plays a crucial role in the modern world governed by software. And while closed source projects can enjoy a sense of confidentiality when addressing security issues, open source projects undertake them publicly even though just as many projects rely on them. In 50% of documented cases, the vulnerabilities could have been spotted almost 20 days before their disclosure leaving plenty of time for a potential attacker to exploit the weakness.
Based on the results of a basic text search, we conclude that the majority of security-related activity is in reaction to known vulnerabilities and that maintainers are not always mentioning security terms when fixing exploits. We also confirm that many security-labeled issues are not pushed to vulnerability systems, even though the maintainers realize their security aspect. Then, while commit classification models can spot security-related commits automatically, the models struggle in realistic scenarios, and no particular feature or sampling method is vastly better than the others. Nonetheless, we evaluated the state-of-the-art models which spot security-related commits with an F1 score of 0.36.
Given the findings, we conclude that security-related activity is hard to automatically distinguish from everyday development activity and that manual review is required to spot these traces. Proposed methods can make this review easier. We suggest that more attention should be given to open source security to avoid early public traces of vulnerabilities.","Open source software; Machine Learning; Deep Learning; Commit representation; Source code embedding; Software security; Software vulnerability analysis; Vulnerabilities; Security advisories; Vulnerability Management","en","master thesis","","","","","","","","","","","","Computer Science","",""
"uuid:06a525ef-a9a8-4899-bbdb-cf2925808dae","http://resolver.tudelft.nl/uuid:06a525ef-a9a8-4899-bbdb-cf2925808dae","An Empirical Assessment on the Limits of Binary Code Summarization with Transformer-based Models","Al-Kaswan, Ali (TU Delft Electrical Engineering, Mathematics and Computer Science)","van Deursen, A. (mentor); Devanbu, Prem (mentor); Izadi, M. (mentor); Sawant, Anand Ashok (graduation committee); Verwer, S.E. (graduation committee); Delft University of Technology (degree granting institution)","2022","Reverse engineering binaries is required to understand and analyse programs for which the source code is unavailable. Decompilers can transform the largely unreadable binaries into a more readable source code-like representation. However, many aspects of source code, such as variable names and comments, are lost during the compilation and decompilation processes. Furthermore, by stripping the binaries, more informative symbols/tokens, including the function names, are also removed from the binary.
Reverse engineering is time-consuming, much of which is taken up by labelling the functions with semantic information. Therefore, we propose a novel code summarisation method for decompiled and stripped decompiled code. First, we leverage the existing BinSwarm dataset and extend it with aligned source code summaries. Next, we create an artificial demi-stripped dataset by removing the identifiers from unstripped decompiled code. To train our model for summarising code using this dataset, we fine-tune a pre-trained CodeT5 model for the code summarisation task on the given dataset. Furthermore, we investigate the performance of the input types, the impact of data duplication and the importance of each aspect present in the source code on the model performance. Moreover, we design and present some intermediate-training objectives to increase the model performance.
We present the following findings:
Firstly, we find that the model generates good summaries for decompiled code, with similar performance to source C code. Compared to summarising decompiled code, the quality of the demi-stripped model is significantly lower but still usable. Stripped performed worse and produced mostly incorrect and unusable summaries.
Secondly, we find that deduplication greatly reduces the performance of the model, putting the performance of decompiled code roughly in line with other decompiled datasets. Thirdly, we found that the loss of identifiers causes a drop in the BLEU-4 score of 35\%, with another 25\% decrease attributable to the increase of decompilation faults caused by stripping. Lastly, we show that our proposed deobfuscation intermediate-training objective improves the model's performance by 0.54 and 1.54 BLEU-4 on stripped and demi-stripped code, respectively.","Machine Learning; Cyber Security; Reverse engineering; Natural Language Processing; Transformers; Deep Learning","en","master thesis","","","","","","","","","","","","Computer Science","",""
"uuid:b8ca8774-47f3-40c1-bc7a-97bce1e176a1","http://resolver.tudelft.nl/uuid:b8ca8774-47f3-40c1-bc7a-97bce1e176a1","Model-specific Explainable Artificial Intelligence techniques: State-of-the-art, Advantages and Limitations","Khan, Arghem (TU Delft Electrical Engineering, Mathematics and Computer Science)","Lal, C. (mentor); Conti, M. (mentor); Delft University of Technology (degree granting institution)","2022","Artificial Intelligence (AI) and Machine learning (ML) applications are being widely used to solve different problems in different sectors. These applications have enabled the human-effort and involvement to be very low. The AI/ML systems
make their own predictions and do not require a great deal of human help. However, over the last few years several incidents of the developed systems
have led to questions regarding the transparency of those AI/ML systems. Without expertise, it is not always as straightforward to understand
certain predictions. This pressing issue has led to the emerging topic of Explainable Artificial Intelligence (XAI). In this research, we will present the current work on a specific type of XAI, namely model-specific XAI. Model-specific XAI techniques are particular to certain types of ML techniques. We will look into several recent model-specific XAI techniques and provide the advantages
and disadvantages. Within similarities we find that there is a set of general requirements that the techniques should adhere to (expertise, bias, time, privacy
and performance). We characterize the techniques in feature-based, concept-based and logic-based. With regard to future work, there is room for
improvement on several areas. For example, this includes work from exploring hybrid techniques to investigating how current techniques can improve
the privacy.","Artificial Intelligence; Machine Learning; Explainable Artificial Intelligence; Model-Specific; Deep Learning","en","bachelor thesis","","","","","","","","","","","","Computer Science and Engineering","CSE3000 Research Project",""
"uuid:d278202d-fe0b-4679-8b74-63d3c2f57495","http://resolver.tudelft.nl/uuid:d278202d-fe0b-4679-8b74-63d3c2f57495","Explainable Artificial Intelligence Techniques for the Analysis of Reinforcement Learning in Non-Linear Flight Regimes","de Haro Pizarroso, Gabriel (TU Delft Aerospace Engineering)","de Visser, C.C. (graduation committee); van Kampen, E. (mentor); Mooij, E. (graduation committee); Delft University of Technology (degree granting institution); Carlos III University of Madrid (degree granting institution)","2022","Reinforcement Learning is being increasingly applied to flight control tasks, with the objective of developing truly autonomous flying vehicles able to traverse highly variable environments and adapt to unknown situations or possible failures. However, the development of these increasingly complex models and algorithms further reduces our understanding of their inner workings. This can affect the safety and reliability of the algorithms, as it is difficult or even impossible to determine which are their failure characteristics and how they will react in situations never tested before. It is possible to remedy this lack of understanding through the development of eXplainable Artificial Intelligence and eXplainable Reinforcement Learning methods like SHapley Additive Explanations. In this thesis, this tool is used to analyze the strategy learnt by an Actor-Critic Incremental Dual Heuristic Programming controller architecture when presented with a series of pitch rate and roll rate tracking tasks in a variety of non-linear flying conditions, which include flying close to the stall regime, experiencing non-linear reference signals, reaching the control surface deflection limits, and flying with large sideslip angles. This same controller architecture has been previously explored with the same analysis tool but limited to the nominal linear flight regime, and it was observed that the controller learnt linear control laws, even though its Artificial Neural Networks should be able to approximate any function. This thesis shows that, even in the non-linear flight regime, it is still more optimal for this controller architecture to learn quasi-linear control laws, although it seems to continuously modify the linear slopes as if it was an extreme case of the gain scheduling technique. Additionally, a more complex Reinforcement Learning architecture based on the Soft Actor-Critic algorithm is also explored with the same analysis tool, to demonstrate its usability in the presence of non-linear control laws and to improve our understanding of this offline state-of-the-art algorithm.","Machine Learning; Reinforcement Learning; Deep Learning; Adaptice Critic Designs; Explainable Artificial Intelligence; SHAP; Flight Control","en","master thesis","","","","","","","","","","","","Aerospace Engineering","",""
"uuid:4ac5c33f-efa4-47d3-b06e-d49110016890","http://resolver.tudelft.nl/uuid:4ac5c33f-efa4-47d3-b06e-d49110016890","Digital Soil Mapping based on PDFs of Cone Penetration Tests and Vibro Cores using Image Processing and Machine Learning","Ordeman, Sam (TU Delft Mechanical, Maritime and Materials Engineering)","van Rhee, C. (mentor); Talmon, A.M. (mentor); Soleymani Shishvan, M. (mentor); Nuttall, Jonathan (mentor); Pisano, F. (graduation committee); Delft University of Technology (degree granting institution)","2022","Digital Soil Mapping (DSM) of soil types in geotechnical project areas is a top priority. These maps are often used in decision making and can have significant consequences related to costs and risks. Usually, these maps are generated by digital soil models that interpolate soil types at known locations. In practice, conventional spatial interpolation techniques are still often used for DSM of soil types, such as inverse distance weighting and kriging. However, conventional models are not well suited for predicting or interpolating soil types because of their inability to deal with categorical data properly. Besides, the design of the conventional models does not allow for incorporating the abundance of meaningful covariate information that is available nowadays. The flexibility of machine learning algorithms vanquish both problems and has become increasingly popular for DSM of soil properties in recent years. The results of machine learning techniques for DSM of soil properties are promising and generally outperform conventional models. However, few studies have used machine learning for DSM of soil types and is therefore still a relatively unknown field. Moreover, at the time of writing, there are no studies that use sequence models for DSM of soil properties or types. Hence, the author proposes to introduce a new method for DSM of soil types, namely a Long Short-Term Memory (LSTM) network. The intuition behind this introduction is that the spatial correlation can be captured in sequences and can improve soil type prediction.
Real project data from a cable burial project is used to evaluate and compare the performance of the conventional interpolation methods triangulation and kriging, the machine learning models random forest and XGBoost, and the newly proposed deep learning model LSTM. The project data consist of 757 vibro cores (VC), 718 cone penetration test (CPT), bathymetry data and sub-bottom profilers. The geotechnical data, i.e. VCs and CPTs, is received on separate PDF pages that require to be digitized first. This thesis describes a simple yet precise manner to extract this data from the PDFs. The VCs and CPTs are provided with a soil type interpretation and can be used directly for developing the models. The data is split into a training set to develop/train the models and a test set for evaluation. Ultimately, the best performing model is used to build a 3D stratigraphic soil model for the project area with associated prediction accuracies.
All state-of-the-art techniques outperform the conventional models and especially in predicting minority classes. The best performing model is random forest with an overall accuracy of 85.44\% and is comparable to the performance of XGBoost of 85.11\%. LSTM network achieved a slightly lower accuracy of 84.27\%. The results show that LSTM is suitable for DSM of soil types and has considerable potential for improvement as only a few possibilities of the model have been examined.","Digital Soil Mapping; Machine Learning; Deep Learning; Spatial Interpolation; Image Processing; LSTM; Random Forest; XGBoost; Kriging; Trenching","en","master thesis","","","","","","","","2024-02-28","","","","Offshore and Dredging Engineering","",""
"uuid:f46e50cc-5c3c-47be-a060-ffe075872547","http://resolver.tudelft.nl/uuid:f46e50cc-5c3c-47be-a060-ffe075872547","Machine and deep learning models for vehicle routing problems: a literature review","Bijvoet, Bas (TU Delft Mechanical Engineering)","Atasoy, B. (mentor); Karademir, C. (mentor); Delft University of Technology (degree granting institution)","2021","Vehicle routing problems (VRPs), a generalization of the traveling salesman problem, are extensively studied combinatorial optimization problems for their practical application. Many solution methods, e.g., exact and heuristic algorithms, have been proposed in the last few decades but require relatively much computation time due to the NP-hard nature of the VRP. Additionally, to build such algorithms, much expert knowledge is required. The recent developments in a subfield of machine learning, deep learning, make it possible to solve routing problems in a purely data-driven manner or assist heuristic methods. This requires less problem-specific knowledge and can outperform the traditional solution methods in terms of objective value and computation time. In this literature review, introductions for the vehicle routing problem and machine learning are given first. Then, an existing categorization for algorithmic machine learning structures is described. The main body of this report reviews recent machine and deep learning methods to solve static and dynamic routing problems. The reviewed literature is summarized in tables that indicate the main characteristics of each work. Moreover, the results of several machine and deep learning based solution methods for the static VRP are compared to each other. Lastly, the challenges and opportunities for future research of deep learning based solution methods for the vehicle routing problem are discussed. The main challenges of these methods are scalability, generalization, and adaptability. Therefore, future research could be focused on addressing these challenges to improve deep learning methods for practical routing applications.","Vehicle Routing Problem; Machine Learning; Deep Learning; Mobility-on-demand; Dynamic Routing","en","student report","","","","","","Literature Assignment Multi-Machine Engineering Report number: 2021.MME.8591","","","","","","Mechanical Engineering | Multi-Machine Engineering","",""
"uuid:6a780003-cb20-4a8c-ac90-2867274a65d6","http://resolver.tudelft.nl/uuid:6a780003-cb20-4a8c-ac90-2867274a65d6","Pedestrian Detection and Tracking for Mobile Robots in Human Environments","Marcelis, N.H.H. (TU Delft Mechanical, Maritime and Materials Engineering)","Ferreira de Brito, B.F. (mentor); Alonso Mora, J. (graduation committee); Delft University of Technology (degree granting institution)","2021","With the performance of current motion planning methods being highly dependent on the quality of the perception system, robust 3D multi-object detection and tracking are vital for autonomous driving applications. Despite all the advancements in 2D and 3D object detectors, robust tracking of pedestrians in dense scenarios is still a challenging subject for small Automated Guided Vehicles (AGVs). Most research in the field of object detection and tracking focuses on autonomous cars, neglecting the design challenges that come with small AGVs.
This thesis presents a real-time multi-modal multi-pedestrian detection and tracking pipeline for small mobile robots. The framework integrates five RGB-D cameras and a LiDAR sensor to achieve real-time pedestrian detection and tracking. The system relies on state-of-the-art 2D and 3D object detectors, a sensor fusion and filtering scheme, and a 3D object tracker. Moreover, to improve detection and tracking performance, we have collected a pedestrian dataset tailored for small AGVs. We use this dataset to train the 3D pedestrian detector and evaluate the performance of the pedestrian detectors and tracker. Evaluation of the proposed framework demonstrated the ability to robustly detect and track multiple pedestrians up to a distance of 10 meters. We open-sourced our framework at: https://github.com/bbrito/amr_navigation.","Machine Learning; Deep Learning; Object Detection; Object Tracking; Computer Vision; Autonomous Vehicles; Intelligent Vehicles","en","master thesis","","","","","","","","","","","","Mechanical Engineering | Biomechanical Design - BioRobotics","",""
"uuid:438a9f4b-0f06-486d-8f68-dd203cae2af4","http://resolver.tudelft.nl/uuid:438a9f4b-0f06-486d-8f68-dd203cae2af4","Reinforcement Learning approach for decision-making in driver control shifting for semi-autonomous driving","Latoškinas, Evaldas (TU Delft Electrical Engineering, Mathematics and Computer Science; TU Delft Software Technology)","Li, Y. (mentor); Spaan, M.T.J. (graduation committee); van Deursen, A. (graduation committee); Delft University of Technology (degree granting institution)","2021","Semi-autonomous driving innovations aim to bridge the gap to fully autonomous driving by co-operating with human drivers to lead to optimal choices on who should drive in different scenarios by offering different automation levels. However, in the present day, known semi-autonomous driving solutions do not generalise to every complex case of driver and AI interaction. This limitation prompted research in attempting to solve the problem using artificial intelligence and machine learning techniques. This paper focuses on providing a reinforcement learning approach to solve one specific decision-making scenario of the driver initiating a shift of control to a different automation level. The decision problem was formulated as a Markov Decision Process, and the problem was solved both by a baseline handcrafted decision tree and a learned reinforcement learning policy using the DQN algorithm. The two policies were compared based on safety, comfort and efficiency metrics in a simulated driving environment. The results were indicative that a reinforcement learning policy generally ensured safety \& comfort and has shown increased efficiency over the baseline policy, however, it faced efficiency & comfort issues in outlier cases.","Reinforcement Learning; semi-autonomous vehicles; Markov Decision Process; Deep Learning; Decision Making; Machine Learning","en","bachelor thesis","","","","","","","","","","","","Computer Science and Engineering","CSE3000 Research Project",""
"uuid:67e2f68d-cd37-44fe-b4a1-0855516c102f","http://resolver.tudelft.nl/uuid:67e2f68d-cd37-44fe-b4a1-0855516c102f","The future of interventional cardiology: Machine learning algorithms for solving diagnostic and therapeutic challenges","van der Loo, Wouter (TU Delft Mechanical, Maritime and Materials Engineering)","Scherptong, Roderick (mentor); Dijkstra, Jouke (mentor); Harlaar, J. (graduation committee); Delft University of Technology (degree granting institution); Universiteit Leiden (degree granting institution); Erasmus Universiteit Rotterdam (degree granting institution)","2021","Coronary artery disease (CAD) is one of the leading causes of death and disability worldwide. In CAD, the coronary arteries, that supply the myocardium with oxygen, are narrowed or even blocked by a process called atherosclerosis. Invasive coronary angiography (ICA) is the gold standard for the diagnosis of CAD, as well as for intraprocedural guidance of percutaneous coronary interventions. At this moment, stenosis severity is determined via visual inspection by a cardiologist. This method has several important drawbacks: a significant inter- and intra-rater variability and a high positive prediction bias. Currently available additional assessment techniques improve results, however, they come with prolonged procedural time, complication risks and increased costs. Additionally, a relevant number of syndromes exists that cannot be diagnosed sufficiently with these techniques. Automated software that uses all information within the ICA images and relates it to the context of complaints and outcomes, could be a valuable tool for improvement. Since machine learning (ML) can find relations between patient groups based on images, this could be a possible solution. However, despite promising results, clinical implementation of ML is still limited. Limited clinical applicability of developed algorithms and a lack of large high-quality datasets are the cause of this. In this research it was investigated if ML can help in solving the diagnostic and therapeutic challenges encountered in interventional cardiology (IC) and what is needed to apply ML to ICA images and other sources of coupled medical data. This was done in several steps. First the expectations and perceived barriers by interventional cardiologists on ML-based algorithms in clinical practice were assessed. Next, the feasibility of health insurance code-based querying of electronic health records (EHRs) for the creation of ML datasets was assessed. Third, a proof-of-concept study on creating a dataset and a deep learning network for predicting lesion significance on ICA images was carried out. Last, a roadmap for the curation of data for the development of ML models in IC was created.
Generally, interventional cardiologists have positive expectations for applying ML in their clinical practice. Furthermore, the willingness to collaborate in the development and clinical validation of ML algorithms is high. This is essential for translating ML models to clinical practice. Health insurance code-based querying of EHRs is not a feasible approach for creating datasets for the complex syndromes which currently pose diagnostic challenges in IC. However, the EHRs hold valuable data for the development of ML datasets. The same goes for the data collection strategy in the proof-of-concept study and it was shown that it is feasible to train deep learning networks on this data.
ML shows promising results for solving the diagnostic and therapeutic challenges encountered in interventional cardiology and directions for further research were identified. For the creation of algorithms that can be applied in clinical practice, close collaboration between ML professionals and clinicians is needed. Besides this, further research is needed to develop scalable strategies for the creation of large datasets, containing adequately labelled patients that represent the real-life population.
To achieve these goals, four state-of-the-art Neural Network architectures (BLSTM, OPGRU, TDNN-BLSTM, and TDNN-OPGRU) are built and trained on a Dutch dataset containing 15 different speaking styles. Then, the trained architectures are used for three experiments. The first experiment tests the architectures on a 15-speaking style containing the same 15 speaking styles as the training set. The results of that training set are used to calculate the phoneme error rate (PER), analyse the errors, and identify error-prone phonemes. The second and third experiments follow the same methodology on the same-trained architectures, however, are tested on a clear speaking style test set and a spontaneous speaking style test set. The results are used to analyse and compare the performances of the models on different speaking styles. This research shows that the BLSTM and OPGRU are a poor choice for Dutch PR, as the results show relatively high PERs for these two architectures. The hybrid acoustic models TDNN-BLSTM and TDNN-OPGRU show to be good models for Dutch PR, as they gain low PERs. The results in this research set a new benchmark for Dutch PR, as the TDNN-BLSTM and TDNN-OPGRU show the lowest recognition accuracy found in existing Dutch PR literacy. Additionally, this research shows that recognising Dutch clear speech is less complex than recognising Dutch spontaneous speech, which is in line with existing word- and phoneme recognition research on other languages. Besides that, this research makes recommendations for which of the hybrid models is more effective to use when using the different speaking styles as input. Finally, this research finds the phonemes /I/, /a/, /k/, /s/, and the schwa as error-prone phonemes in Dutch PR, as they have high, above average individual PERs and contribute above-average to the overall PER on both the TDNN-BLSTM and TDNN-OPGRU. This research also makes recommendations on where future Dutch PR could focus on. Setting a benchmark for Dutch PR, analysing different speaking styles, and identifying error-prone phonemes hopefully triggers future researchers on further improving Dutch PR.
This thesis focusses on the use of autoencoders and generative adversarial networks (GANs) for one-class classification problems involving image data. Autoencoders can learn encoding and decoding functions for samples from the target dataset. These encoding and decoding functions are, however, expected to not perform well for non-target samples, as they have never been seen during the training phase. This makes it possible to separate target and non-target data. For GANs, the discriminator is used to distinguish between target and non-target data.
Autoencoders and GANs are evaluated extensively in this report. Their behavior, desired parameters and strengths and weaknesses are evaluated by performing experiments. The main findings are that GANs do not perform well for one-class classification tasks, because of mode collapse and insufficient sampling of the non-target data. Even for extremely simple datasets these issues were observed. Autoencoders are shown to perform much better and behave according to the theoretical expectations.","one-class; classification; high-dimensional; Autoencoder; GAN; Wasserstein Autoencoder; Pattern Recognition; Machine Learning; Deep Learning","en","master thesis","","","","","","","","","","","","Computer Science | Data Science and Technology","",""
"uuid:5028e05e-661f-4791-9d1f-e50d877aa24f","http://resolver.tudelft.nl/uuid:5028e05e-661f-4791-9d1f-e50d877aa24f","Human Activity Recognition using a Deep Learning Algorithm for Patient Monitoring","Friđriksdóttir, Esther (TU Delft Mechanical, Maritime and Materials Engineering)","French, Paddy (mentor); Bonomi, Alberto (mentor); Hunyadi, Bori (graduation committee); Delft University of Technology (degree granting institution)","2019","Physical activity and mobility are important indicators of the recovery process of patients in the general ward of the hospital. Currently, monitoring mobility of hospitalized patients relies largely on direct observation from the caregivers. Accelerometers have the potential to quantify physical activity of patients objectively and without obstructing their daily routines. Human Activity Recognition (HAR) is a technique used to assess the type of activity an individual subject is carrying out based on sensor readings and has been extensively studied. However, the literature shows that HAR methodologies have been largely developed to recognize activities typical for healthy subjects. This means that activities performed at a slow and irregular pace, such as by a symptomatic patient or an elder, are scarcely considered to design HAR methods. Using HAR for patient monitoring would allow clinically meaningful metrics, such as time spent ambulating or in sedentary behaviour each day, to be obtained automatically. This may offer a convenient solution to enable caregivers automatically monitoring the recovery process of patients.
The aim of this work was to develop an accurate HAR model to recognize activities typical for hospitalized patients. A data collection study was conducted with 20 healthy volunteers in a simulated hospital environment. A single triaxial accelerometer placed on the trunk was used to measure body movement and recognize seven activity types: Lying in bed, upright posture, walking, walking with walking aid, wheelchair transport, stair ascent and stair descent. Features from both time and frequency domain were extracted and used to train three machine learning (ML) classifiers (Naïve Bayes, Random Forest, Support Vector Machine). Additionally, a deep neural network (DNN) consisting of a three convolutional layers and a Long Short-Term Memory layer was developed.
The performance of the DNN model was evaluated on holdout data and compared to the performance of the feature-based ML classifiers. The DNN model reached a higher classification accuracy than the latter approaches (F1-score= 0.902 vs. 0.821). All the models showed a large number of misclassification between the walking with or without walking aid class. By combining these two classes the DNN model reached an F1-score 0.946, compared to F1-score 0.856 of the best feature-based ML approach represented by a support vector machine classifier.
This work shows for the first time the value of applying deep-learning techniques to improve the accuracy of feature-based ML classifiers for addressing the problem of HAR using a single triaxial accelerometer in simulated hospital conditions.","Human Activity Recognition; Machine Learning; Deep Learning; Accelerometer; Classification; Patient Monitoring","en","master thesis","","","","","","","","2024-08-27","","","","","",""
"uuid:12693ba0-61a5-4430-99fa-3f72bed91a08","http://resolver.tudelft.nl/uuid:12693ba0-61a5-4430-99fa-3f72bed91a08","Exploring Gravitational Waves Recordings with Machine Learning Techniques","Diab Montero, Hamed (TU Delft Civil Engineering & Geosciences)","Schmelzbach, Cedric (mentor); Ferraioli, Luigi (mentor); Meier, Men-Andrin (mentor); Broggini, Filippo (mentor); Giardini, Domenico (mentor); Delft University of Technology (degree granting institution); ETH Zürich (degree granting institution); Rheinisch-Westfälische Technische Hochschule (degree granting institution)","2019","The study of Gravitational Waves (GWs) opened a new window of possibilities to improve our understanding of the Universe. GWs provide suitable astronomical messengers for studying events that were not possible before through electromagnetic radiation, or in other cases complementing their observations. Ground-based interferometers like LIGO have been recording multiple GW events since the first detections in 2015. Despite the success of Earth-based observatories, the space limitations and noise sources on Earth point toward the need of building a spaceborne interferometer. The Laser Interferometer Space Antenna (LISA) is a planned project that will provide us with such a detector and will allow gaining access to lower frequency bands and more types of GW sources. To make the most out of LISA’s strengths, it is important to identify and develop alternative data analysis tools which are more appropriate for low latency searches of GWs than the current ones in use. Machine Learning techniques are a promising candidate since they can provide high accuracies, higher speeds, and a lower computational cost. Therefore, they can be used for the development of Low Latency Detectors (LLD) of GWs, which will be used to analyze the LISA recordings. I propose to build a prototype LLD by using a Sliding Window Algorithm, which makes use of Convolutional Neural Networks (CNNs) as its classification mechanism. To implement the LLD, I first create datasets composed of synthetic GW recordings of two different GW source types: Galactic Binaries (GBs) and Merging Blackhole Binaries (MBHBs). Then, I transform these recordings originally represented only in the time domain, into the frequency domain, and the time-frequency domain and train two different ML architectures (CNNs and Fully-Connected Neural Networks) using both the original and the transformed data. A performance evaluation is done to select the best combination of ML architecture and domain representation for solving the detection task. The chosen combination is then used as the classifier mechanism of the LLD acting in windows of five days duration. The LLD is tested on one-year-long recordings with different levels of noise. The analysis suggests that the time-frequency domain representations offer the most promising results for detecting both types of sources (GBs and MBHBs) reaching high accuracies in recordings with low to moderate signal-to-noise ratio (SNR).","Gravitational Waves; Machine Learning; Neural Networks; Deep Learning; Spectrogram; Laser Interferometer Space Antenna (LISA); Classification; Low Latency Detection","en","master thesis","","","","","","","","","","","","Applied Geophysics | IDEA League","",""
"uuid:3f98717d-08d7-47df-b2dc-6167874cf9ed","http://resolver.tudelft.nl/uuid:3f98717d-08d7-47df-b2dc-6167874cf9ed","Deep vs Shallow Reinforcement Learning for low dimensional continuous control tasks","Arnaoutis, Vasos (TU Delft Mechanical, Maritime and Materials Engineering; TU Delft Delft Center for Systems and Control)","Babuska, R. (mentor); de Bruin, T.D. (mentor); Delft University of Technology (degree granting institution)","2019","Deep Learning performance dependents on the application and methodology. Neural Networks with convolutional layers have been a great success in multiple tasks trained under Supervised Learning algorithms. For higher dimensional problems, the selection of a deep network architecture can significantly improve the accuracy of the network, however for low dimensional problems this might not be true. Shallow Neural Networks have successfully matched the performance of Deep Neural Networks in multiple tasks in the past and have been shown to be expressive enough to represent low dimensional continuous control problems. Through the thesis, the performance and expressiveness of Shallow and Deep networks is compared for low-dimensional continuous control tasks. The thesis begins by comparing the two network architectures in a Supervised Learning algorithm and progresses towards state-of-
the-art Reinforcement Learning algorithms. The thesis provides an empirical approach
towards comparison of neural networks and makes conclusions that can support the selection of a network architecture for continuous control applications using Deep Reinforcement Learning algorithms.","Deep Learning; Reinforcement Learning; Continuous control; Low dimensional system; Machine Learning; Shallow; Deep; Neural Networks","en","master thesis","","","","","","","","","","","","Mechanical Engineering | Systems and Control","",""
"uuid:6389d77c-007d-455f-8e84-10a4f9b57a9d","http://resolver.tudelft.nl/uuid:6389d77c-007d-455f-8e84-10a4f9b57a9d","Deep End-to-end Network for 3D Object Detection in the Context of Autonomous Driving","Jargot, Dominik (TU Delft Mechanical, Maritime and Materials Engineering; TU Delft Cognitive Robotics)","Gavrila, Dariu (mentor); Roth, Markus (mentor); Kober, Jens (graduation committee); Kooij, Julian (graduation committee); Kok, Manon (graduation committee); Delft University of Technology (degree granting institution)","2019","Nowadays, autonomous driving is a trending topic in the automotive field. One of the most crucial challenges of autonomous driving research is environment perception. Currently, many techniques achieve satisfactory performance in 2D object detection using camera images. Nevertheless, such 2D object detection might be not sufficient for autonomous driving applications as the vehicle is operating in a 3D world where all the dimensions have to be considered. In this thesis a new method for 3D object detection, using deep learning approach is presented. The proposed architecture is able to detect cars using data from images and point clouds. The proposed network does not use any hand-crafted features and is trained in an end-to-end manner. The network is trained and evaluated with the widely used KITTI dataset. The proposed method achieves an average precision of 81.38%, 67.02%, and 65.30% on the easy, moderate, and hard subsets of the KITTI validation dataset, respectively. The average inference time per scene is 0.2 seconds.","3D object detection; Thesis; Intelligent Vehicles; Deep Learning; Machine Learning; Camera; Lidar","en","master thesis","","","","","","","","","","","","Mechanical Engineering | Vehicle Engineering","",""
"uuid:1624a31f-7976-425a-a2b6-d6937cc39895","http://resolver.tudelft.nl/uuid:1624a31f-7976-425a-a2b6-d6937cc39895","Rotation invariant filters in CNNs: applied to segmentation of aerial images for land-use classification","Dhar, Aniket (TU Delft Electrical Engineering, Mathematics and Computer Science)","van Gemert, J.C. (mentor); Reinders, M.J.T. (graduation committee); Scharenborg, O.E. (graduation committee); van der Maas, Daan (mentor); Delft University of Technology (degree granting institution)","2018","Convolutional neural networks are showing incredible performance in image classification, segmentation, object detection and other computer vision applications in recent years. But they lack understanding of affine transformations to input data. In this work, we introduce rotational invariant
convolutional neural networks that learn rotational invariance by design, and not from data. We build rotation invariant filters through parametric learning of linear combination of a basis set of filters, rather than modelling the filters ourselves. Our approach combines the learning capability of CNNs with custom filter selection. We show stability in performance under rotations in input images. We first validate our findings for classification on MNIST and then for
multi-class semantic segmentation on the DeepGlobe 2018 Satellite Image Understanding Challenge.","Computer Vision; Deep Learning; Machine Learning; Convolutional Neural Networks","en","master thesis","","","","","","","","","","","","Electrical Engineering | Embedded Systems","",""
"uuid:066c00e8-b94c-4f5e-992a-d2ffbc3543ed","http://resolver.tudelft.nl/uuid:066c00e8-b94c-4f5e-992a-d2ffbc3543ed","DeepSleep: A sensor-agnostic approach towards modelling a sleep classification system","Rao, Shashank (TU Delft Electrical Engineering, Mathematics and Computer Science; TU Delft Multimedia Computing)","Cesar, Pablo (mentor); Ali, Abdallah (mentor); Hanjalic, A. (graduation committee); Hung, H.S. (graduation committee); Delft University of Technology (degree granting institution)","2018","Sleep is a natural state of our mind and body during which our muscles heal and our memories are consolidated. It is such a habitual phenomenon that we have been viewing it as another ordinary task in our day-to-day life. However, owing to the current fast-paced, technology-driven generation, we are letting ourselves be sleep-deprived, giving way to serious health concerns such as depression, insomnia, restlessness, apnea and Alzheimer’s. Polysomnography (PSG) studies are used for diagnosing and treating sleep-related disorders. Although the PSG studies are considered as the gold standard, they are obtrusive and do not allow for long-term monitoring. Various wearables have been manufactured to help people monitor their sleep-health. However, these devices have been shown to be inaccurate.
The ubiquitous sensor technology employed by the wearables provides large volumes of data, recorded in the most natural setting of the user. There is an opportunity to make use of the highly available sensor data to model a sleep scoring system that could help individuals monitor their sleep-health from the comfort of their home. In this thesis, we aim to alleviate this problem by attempting to bridge the gap between the highly accurate but obtrusive medical diagnosis (PSG) and the non-intrusive yet inaccurate wearables.
In this work, we propose DeepSleep, a deep neural net-based sleep classification model using an unobtrusive BCG-based heart sensor signal. Our proposed model’s architecture uses the combination of CNN and LSTM layers to perform self-feature extraction and sequential learning respectively. We show that our model can classify sleep stages with a mean f1-score of 74% using the BCG signal. We employ a 2-phase training strategy to build a pre-trained model to tackle the limited dataset size and test the transferability of the model on other types of heart-signal. With an average classification accuracy of 82% and 63% using ECG and PPG based heart signal respectively, we show that our pre-trained model can be used in the transfer learning setting as well. Lastly, with the help of a user study of 16 subjects, we show that the objective sleep quality metrics correlate with the perceived sleep quality reported by the subjects with a correlation score of 푟 = 0.43.
Although our proposed model’s performance is not yet comparable to the medical standards, we show that it is possible to monitor our sleep-health using the wearable signals with the least domain knowledge and preprocessing techniques. The prediction and performance of our DeepSleep model show that it is able to learn the biological rules of sleep wherein it always follows a Deep or REM stage with a transitional Light stage. Our model treats the classification problem sequentially, thus, identifying important sleep parameters like the onset of sleep cycles and time spent in different sleep stages which are time-dependent factors. Furthermore, our user study, conducted using the SATED questionnaire, provides an insight into the difference in the user’s perceived sleep quality and model’s estimation. It shows that an automated classification system needs to incorporate various external factors such as environmental and ambient conditions to be able to strongly correlate with the perceived or subjective quality. We further discuss the future research gaps and opportunities that could improve the model’s performance and also extend it to other domains like irregular heart-beat and apnea detection. We consider this work to be a starting point for research into sleep and heart health using non-intrusive wearable sensors and deep neural network-based architectures.","Machine Learning; Deep Learning; Sleep-state classification; Sleep System; Ubiquitous; Wearable Technology; long short-term memory networks; Convolutional Neural Networks; Transfer learning; User preferences; Human health; Sleep deprivation; Sensor data; Recurrent Neural Network; Sleep quality measurement; User perception study; subjective assessment; Artificial Intelligence; Explainable Machine Learning","en","master thesis","","","","","","","","","","","","Computer Science","",""
"uuid:dbf3e77c-e624-47ef-b951-3f1948b1609a","http://resolver.tudelft.nl/uuid:dbf3e77c-e624-47ef-b951-3f1948b1609a","Scalability Analysis of Predictive Maintenance Using Machine Learning in Oil Refineries","Helmiriawan, Helmi (TU Delft Electrical Engineering, Mathematics and Computer Science; TU Delft Quantum & Computer Engineering)","Al-Ars, Zaid (mentor); Delft University of Technology (degree granting institution)","2018","Modern refineries typically use a high number of sensors that generate an enormous amount of data about the condition of the plants. This generated data can be used to perform predictive maintenance, an approach to predict impending failures and mitigate downtime in refineries. This research analyzes the scalability of machine learning methods for predictive maintenance solution in an oil refinery. It can be done by modeling the normal behavior of the plant and use the prediction error to identify anomalies which might potentially become failures. Several methods and learning algorithms are explored in this research to model the normal behavior of multiple components in the plant. The experiments are performed by using historical process data from a crude distiller unit at Shell Pernis Refinery. The results show that the proposed approach using multiple targets model is able to predict multiple components in the plant. It is not only able to detect anomalies but also identify the faulty component. Furthermore, it reduces the required time to model the normal behavior of the plant which improves the scalability of the predictive maintenance approach in the refinery.","Machine Learning; Predictive Maintenance; Anomaly Detection; Deep Learning","en","master thesis","","","","","","","","2018-12-31","","","","Computer Science","",""
"uuid:13d6eb0e-4b34-4a41-86a8-9cbab66ecfa0","http://resolver.tudelft.nl/uuid:13d6eb0e-4b34-4a41-86a8-9cbab66ecfa0","Time series forecast in non-stationary environment with occurrence of economic bubbles: Bitcoin Price prediction","Garbacz, Mateusz (TU Delft Electrical Engineering, Mathematics and Computer Science)","Loog, Marco (mentor); Reinders, Marcel (graduation committee); Bozzon, Alessandro (graduation committee); Delft University of Technology (degree granting institution)","2018","Being capable to foresee the future of a given financial asset as an investor, may lead to significant economic profits. Therefore, stock market prediction is a field that has been extensively developed by numerous researchers and companies. Recently, however, a new branch of financial assets has emerged, namely cryptocurrencies. As a representative of these tokens, we chose the largest and most popular cryptocurrency, called Bitcoin. Its value is characterized by with non-stationary behaviour and occurrence of speculative bubbles, which cause a rapid explosion of the price, followed by a major crisis and market panic.
Currently, most of the research community does not take these issues into account, while predicting its price, which may lead to wrong conclusions or unstable results. Therefore, in this thesis, we take a step back and reconsider how does the environment influence model's performance and how to use this knowledge to implement more accurate forecast in the future. Moreover, by designing an appropriate methodology and employing semantic features from online text sources, such as Twitter, Reddit and online news portals, we attempt to build a robust prediction system that offers stable performance regardless of the market fluctuations.
Executed experiments prove that non-stationarity negatively influences the results, causing the deterioration of model's performance over time. Furthermore, it appears that there may be certain properties of economic bubbles that facilitate more efficient prediction, as well as some predictors have an ability to successfully forecast the beginning of a market crisis. However, these findings are based on individual observations, which need to be confirmed by further research. In addition, by designing an appropriate methodology, we prevented performance deterioration, caused by price signal non-stationarity. Although, the semantic features based on online sources did not boost the robustness of the system significantly, combined with the suitable system's design, they lead to improvement in the overall performance of the predictor.","Bitcoin; Bubble; Prediction; Machine Learning; Deep Learning; Non-stationarity","en","master thesis","","","","","","","","","","","","","",""
"uuid:27037d74-d49b-4dfe-bcad-ccf0a0bfd957","http://resolver.tudelft.nl/uuid:27037d74-d49b-4dfe-bcad-ccf0a0bfd957","Machine Learning for Predictive Maintenance: A Boeing 747 Bleed Air Valves case study","IJzermans, Erik (TU Delft Aerospace Engineering)","Verhagen, W.J.C. (mentor); Curran, R. (mentor); Delft University of Technology (degree granting institution)","2018","The newest generation of aircraft has seen a strong increase in sensor data generated on-board. The available data has the potential to indicate the health state of individual components based on which their maintenance requirements can be determined, a maintenance strategy called Condition Based Maintenance. Predictive Maintenance is a specific condition based maintenance strategy that aims to determine these requirements in advance by predicting failures from the sensor data. It has the potential to reduce costly unanticipated maintenance or unnecessarily conservative maintenance. In the short term it could add significant value for components that are currently subject to a reactive maintenance policy. In the long term it could potentially disrupt the traditional maintenance practice of periodic inspection. One of the main challenges in applying Predictive Maintenance in the aviation industry is translating the large amounts of sensor data into a reliable failure prediction, a process called prognostics.
In this study, state-of-the-art machine learning, and specifically deep learning models, have been investigated for their potential for prognostics. A case study has been performed at KLM Royal Dutch Airlines on the Boeing 747 Bleed Air Valves, traditionally some of the most challenging components from a maintenance perspective. It has been shown that fully self-learning algorithms can be used for prognostics, enabling the implementation of one of the first real-life predictive maintenance implementations.","Machine Learning; Deep Learning; Predictive Maintenance; Prognostics","en","master thesis","","","","","","","","2023-07-16","","","","Aerospace Engineering | Transport and Operations","",""
"uuid:2e203eee-4c38-4c86-a92a-db94d0ffc34c","http://resolver.tudelft.nl/uuid:2e203eee-4c38-4c86-a92a-db94d0ffc34c","Side-Channel Attacks using Convolutional Neural Networks: A Study on the performance of Convolutional Neural Networks on side-channel data","Samiotis, Ioannis Petros (TU Delft Electrical Engineering, Mathematics and Computer Science)","Picek, Stjepan (mentor); van der Lubbe, Jan (graduation committee); Hanjalic, Alan (graduation committee); Delft University of Technology (degree granting institution)","2018","Side-Channel Attacks, are a prominent type of attacks, used to break cryptographic implementations on a computing system. They are based on information ""leaked"" by the hardware of a computing system, rather than the encryption algorithm itself. Recent studies showed that Side-Channel Attacks can be performed using Deep Learning models. In this study, we examine the performance of Convolutional Neural Networks, on four different datasets of side- channel data and we compare our models with conventional Machine Learning algorithms and a CNN model from literature. We found that CNNs have the potential to achieve high accuracy performance (99.8%), although their capacity is heavily influenced by the use case. We also found that certain Machine Learning algorithms can outperform CNNs in certain cases, leaving an open debate on the performance gains of the latter.","Side-Channel Attacks; Deep Learning; Convolutional Neural Networks; Machine Learning; optimization algorithms; Classification; cybersecurity","en","master thesis","","","","","","","","2018-05-31","","","","","",""
"uuid:9bce7f8a-8d69-4b48-98c4-fc4e6600b63d","http://resolver.tudelft.nl/uuid:9bce7f8a-8d69-4b48-98c4-fc4e6600b63d","Segmenting and Detecting Carotid Plaque Components in MRI","van Hilten, Arno (TU Delft Mechanical, Maritime and Materials Engineering)","de Bruijne, Marleen (mentor); Sedghi Gamechi, Zahra (mentor); Niessen, Wiro (graduation committee); Kooij, Julian (graduation committee); Delft University of Technology (degree granting institution)","2018","Cardiovascular diseases and stroke are currently the leading causes of death worldwide. Atherosclerotic plaque is a mostly asymptotic vascular disease, but rupture of an atherosclerotic plaque in the carotid artery could lead to stroke. Automated segmentation of plaque components could help improve risk assessment by producing fast and reliable results while saving costs.
In this thesis two extensive comparisons have been made. First supervised classifiers are compared in the pixel-wise segmentation task of plaque components. In this comparison five conventional machine learning techniques and one deep learning architecture have been evaluated: linear and quadratic Bayes normal classifiers, linear logistic classifier, random forest and a U-net architecture. In the second comparison classifiers are evaluated in a detection task for their ability to learn with weakly labelled data. This is done within the multiple instance learning (MIL) framework. In addition to conventional multiple instance learning algorithms, a new MIL adaptation of the deep learning architecture, MIL U-net, is proposed and evaluated.
In the pixel-wise segmentation tasks the U-net architecture was the best overall classifier after the addition of 93 extra training patients to the original 20 training patients. A good inter-rater agreement was found for the haemorrhage class (ICC = 0.684) and the calcification class (ICC = 0.627). In the detection task the supervised methods, trained with one-sided noise, outperformed multiple instance classifiers such as MIL-Boost and the proposed MIL U-net. In this task both random forest and the linear logistic classifier obtained a fair Cohen's kappa (0.419 and 0.445 respectively) for detection of calcification per slice. The same classifiers obtained good correlation (Cohen's kappa 0.717 and 0.666 respectively) for haemorrhage detection per slice.","Machine Learning; Deep Learning; Multiple Instance Learning; Segmentation; Detection; Plaque Components; Carotid Artery; MRI","en","master thesis","","","","","","","","","","","","Mechanical Engineering","",""
"uuid:0be4a32d-f9ef-4ff9-b894-d3f9126b8990","http://resolver.tudelft.nl/uuid:0be4a32d-f9ef-4ff9-b894-d3f9126b8990","Een Semantisch Connectionistisch Redeneersysteem (SCORE)","Sadon, A.P.J.","Koppelaar, H. (mentor); Frietman, E. (mentor); Kerckhoffs, E. (mentor); Klaassen, A. (mentor)","1990","Het onderzoek heeft zich gericht op de ontwikkeling van een Semantisch Connectionistisch Redeneersysteem (SCORE). SCORE kan de basis vormen voor een volwaardig expertsysteem. De, voor een expertsysteem van elementair belang zijnde, fundamentele inferentie principes, kennisacquisitie principes en kennisrepresentatievormen hebben in SCORE de meeste aandacht gehad. Binnen SCORE vindt de probleembeschrijving op een dieper niveau plaats dan de traditionele expertsystemen. Dit wordt het subsymbolisch niveau genoemd, en is geïnspireerd op de theorie van Neurale Netwerken. Op het subsymbolische niveau wordt in termen van een groot aantal, 'zwakke' regels het probleem gedefinieerd. Deze regels zijn, in tegenstelling tot de regels binnen het traditionele AI-Paradigma, ‘overrulebaar’: de regels worden geïnterpreteerd als adviezen en niet als autoritaire uitspraken. Het doel van Connectionistisch inferentiemechanisme is, na een waarneming, dié oplossing te vinden waarbij zoveel mogelijk adviezen worden gehonoreerd. In het geval van inconsistenties, onvolledigheden en onzekerheden wordt dus, zonder extra aanpassingen, óok automatisch gezocht naar een zo goed mogelijke oplossing. Na een waarneming wordt het inferentieproces automatisch aangevangen doordat de subsymbolen hierdoor uit evenwicht worden gebracht. Alle subsymbolen gaan nu actief op zoek naar hun aandeel in de oplossing. Een stabiele situatie representeert voor ieder subsymbool (dus het expertsymbool), een waarde die zo goed mogelijk bij de waarneming aansluit. Deze set van waarden wordt beschouwd als de (of een) oplossing van het probleem. SCORE is zó opgezet dat stabiele situaties zo goed mogelijk overeenkomen met legale patronen (in dit rapport legale afleidingen genoemd), d.w.z. symboolcombinaties zoals de expert die aan het netwerk heeft geleerd. Logische, door de expert gedefinieerde, bij elkaar behorende expertsymbolen (bijvoorbeeld de symbolen kleur[groen]; kleur[rood]; kleur[geel]) worden units genoemd en staan onder de hoede van sequentiële unitbesturingen. De unit is in SCORE het kleinst ondeelbare element en kan in het geval van een parallelle implementatie nog worden toegekend aan een processor. Voordat inferentie kan plaatsvinden moet de expert, zoveel mogelijk rekening houdend met duidelijk gedefinieerde semantische eisen, een kennisstructuur, d.w.z. een structuur tussen de verschillende units, definiëren. De structuur heeft invloed op de performance van het systeem, maar is niet van cruciaal belang. Het is bedoeld om het aantal relaties tussen de subsymbolen in te perken. Vervolgens moet het netwerk worden getraind met voorbeelden. Deze voorbeelden worden de legale afleidingen. Aan de hand van deze voorbeelden bepaalt SCORE of er semantische eisen worden geschonden. Indien dat het geval is, worden er automatisch (voor de gebruiker onzichtbaar) dan wel extra units, dan wel extra subsymbolen geïntroduceerd. SCORE zorgt er automatisch voor dat de kennis voortdurend consistent blijft. Om de betekenis van de relaties tussen de subsymbolen niet te verliezen is de kanstheorie zover als mogelijk in SCORE verwerkt. Het inferentiemechanisme is gericht op het, uitgaande van de waarneming, vinden van een bijbehorend trainingsvoorbeeld. Als er meerdere voldoen, of als er geen enkele voldoet, dan moet een zo goed mogelijke worden gevonden. Om met dat zoekproces niet in lokale optima terecht te komen, is 'Simulated Annealing', een afgeleide van de tweede hoofdwet van de thermodynamica, toegepast. Aan het geheel is een kennisacquisitiemodule gekoppeld waarmee inferentieresultaten in de verbindingen van het netwerk kunnen worden verwerkt. Hiermee kan het op de praktijk worden afgestemd.","Artificial Intelligence; Machine Learning; Neural Netwerken; Neural Networks; Connectionist Systems; Deep Learning; Expert systems","nl","master thesis","","","","","","","","","Electrical Engineering, Mathematics and Computer Science","Informatica - Kennisgestuurde systemen","","","",""