Financial trading data presents a unique challenge for anomaly detection due to its high dimensionality and often lack of labelled anomalous examples. Nevertheless, it is of great interest for financial institutions to gain insight into potential trading activities that might lead to financial losses and reputational damage. Given the complexity and unlabelled nature of this financial data, deep learning models such as the Transformer model are particularly suited for this task.   This work investigates the application of a Transformer-based autoencoder for anomaly detection in unlabelled financial transaction data with sequential characteristics. To assess the model's ability to detect anomalies and analyse the effects of class imbalance, synthetic anomalies are injected into the dataset. This creates a controlled environment where the model's performance can be evaluated but also the affects of imbalance can be investigated.    Two approaches are particularly explored for anomaly detection purposes: an unsupervised approach and a semi-supervised approach that explicitly leverages the presence of anomalies in the training data. Experiments suggest that while the unsupervised approach can detect anomalies with distinctive features, its performance suffers when anomalies are included in the training data since the model tends to reconstruct them. Conversely, the semi-supervised approach effectively addresses this limitation, demonstrating a clear advantage in the presence of class imbalance. While synthetic anomalies enable controlled evaluation and class imbalance analysis, generalizability to real-world financial data requires true anomalies.

This thesis delves into the use of Artificial Intelligence (AI) for quality control in manufacturing systems, with a particular focus on anomaly detection through the analysis of torque measurements in rotating mechanical systems. The research specifically examines the effectiveness of torque measurements in quality control of locks, challenging the traditional method that relies on human tactile sense for detecting mechanical anomalies. This conventional approach, while widely used, has been found to yield inconsistent results and poses physical strain on operators. A key aspect of this study involves conducting experiments on locks using torque measurements to identify mechanical anomalies. This method represents a shift from the subjective and physically demanding practice of manually testing each lock. The research aims to demonstrate that an automated, AI-driven approach can offer more consistent and reliable results, thereby improving overall product quality. The development of a machine learning model for this purpose starts with the collection of training data, a process that can be costly and disruptive to normal workflow. Therefore, this thesis also investigates strategies for predicting and minimizing the sample size used for training. Additionally, it addresses the critical need of trustworthiness in AI systems used for final quality control. The research explores how to utilize machine learning models that are not only effective in detecting anomalies but also offers a level of interpretability, avoiding the pitfalls of black box AI models. Overall, this thesis contributes to advancing automated quality control by exploring the state-of-the-art machine learning algorithms for mechanical fault detection, focusing on sample size prediction and minimization and also model interpretability. To the best of the author’s knowledge, it is the first study that evaluates an AI-driven solution for quality control of mechanical locks, marking an innovation in the field. / Denna avhandling fördjupar sig i användningen av Artificiell Intelligens (AI) för kvalitetskontroll i tillverkningssystem, med särskilt fokus på anomalidetektion genom analys av momentmätningar i roterande mekaniska system. Forskningen undersöker specifikt effektiviteten av momentmätningar för kvalitetskontroll av lås, vilket utmanar den traditionella metoden som förlitar sig på människans taktila sinne för att upptäcka mekaniska anomalier. Denna konventionella metod, som är brett använd, har visat sig ge inkonsekventa resultat och medför fysisk belastning för operatörerna. En nyckelaspekt av denna studie innebär att genomföra experiment på lås med hjälp av momentmätningar för att identifiera mekaniska anomalier. Denna metod representerar en övergång från den subjektiva och fysiskt krävande praxisen att manuellt testa varje lås. Forskningen syftar till att demonstrera att en automatiserad, AI-driven metod kan erbjuda mer konsekventa och tillförlitliga resultat, och därmed förbättra den övergripande produktkvaliteten. Utvecklingen av en maskininlärningsmodell för detta ändamål börjar med insamling av träningsdata, en process som kan vara kostsam och störande för det normala arbetsflödet. Därför undersöker denna avhandling också strategier för att förutsäga och minimera mängden av data som används för träning. Dessutom adresseras det kritiska behovet av tillförlitlighet i AI-system som används för slutlig kvalitetskontroll. Forskningen utforskar hur man kan använda maskininlärningsmodeller som inte bara är effektiva för att upptäcka anomalier, utan också erbjuder en nivå av tolkningsbarhet, för att undvika fallgroparna med svart låda AI-modeller. Sammantaget bidrar denna avhandling till att främja automatiserad kvalitetskontroll genom att utforska de senaste maskininlärningsalgoritmerna för detektion av mekaniska fel, med fokus på prediktion och minimering av mängden träningsdata samt tolkbarheten av modellens beslut. Denna avhandling utgör det första försöket att utvärdera en AI-driven strategi för kvalitetskontroll av mekaniska lås, vilket utgör en nyskapande innovation inom området.

<p dir="ltr">Distracted driving is at the forefront of the leading causes of road accidents. Therefore, research advancements in Driver Monitoring Systems (DMS) are vital in facilitating prevention techniques. These systems must be able to detect anomalous driving behavior by evaluating deviations from some predefined normal driving behavior. This thesis proposes an improved contrastive learning approach that introduces a hybrid loss function combining triplet loss and supervised contrastive loss, as well as improvements to the projection head of the framework. It progresses the architecture by performing a multi-threshold severity calculation and data processing using an exponential moving average technique. Due to the unbounded possibilities of anomalous driving behaviors, the proposed framework was tested on the Driver Anomaly Detection (DAD) dataset that incorporates multi-modal and multi-view inputs in an open set recognition setting. The test set of the DAD dataset has anomalous actions that are unseen by the trained model; therefore, high precision on such a dataset demonstrates success on any other closed-set recognition task. The proposed framework achieved an impressive accuracy, reaching 94.14\%, AUC-ROC at 0.9787, and AUC-PR at 0.9781 on the test set. These findings contribute to in-vehicle monitoring by providing a scalable and adaptable framework suitable for real-world conditions.</p>

Utförda hacker-attacker på moderna fordon belyser ett behov av snabb detektering av hot inom denna miljö, särskilt när det förekommer en trend inom denna industri där moderna fordon idag kan klassas som IoT-enheter. Det förekommer kända fall av attacker där en angripare förmår stoppa fordon i drift, eller ta bromsar ur funktion, och detta har påvisats ske fjärrstyrt. Denna studie undersöker detektion av utförda attacker, på en riktig bil, genom studie av CAN bus meddelanden. De två modellerna CUSUM, från området Change Point Detection, och Random Forests, från området maskininlärning, tillämpas på riktig datamängd, för att sedan jämföras på simulerad data sinsemellan. En ny hypotesdefinition introduceras vilket möjliggör att evalueringsmetoden Conditional expected delay kan nyttjas för fallet Random Forests, där resultat förmås jämföras med evalueringsresultat från CUSUM. Conditional expected delay har inte tidigare studerats för metod av maskininlärning. De båda metoderna evalueras också genom ROC-kurva. Sammantaget förmås de båda metoderna jämföras sinsemellan, med varandras etablerade evalueringsmetoder. Denna studie påvisar metod och hypotes för att brygga de två områdena change point detection och maskininlärning, för att evaluera de två enligt gemensamt motiverade parametervärden. / There are known hacker attacks which have been conducted on modern vehicles. These attacks illustrates a need for early threat detection in this environment. Development of security systems in this environment is of special interest due to the increasing interconnection of vehicles and their newfound classification as IoT devices. Known attacks, that have even been carried out remotely on modern vehicles, include attacks which allow a perpetrator to stop vehicles, or to disable brake mechanisms. This study examines the detection of attacks carried out on a real vehicle, by studying CAN bus messages. The two methods CUSUM, from the field of Change Point Detection, and Random Forests, from the field of Machine Learning, are both applied to real data, and then later comparably evaluated on simulated data. A new hypothesis defintion is introduced which allows for the evaluation method Conditional expected delay to be used in the case of Random Forests, where results may be compared to evaluation results from CUSUM. Conditional expected delay has not been studied in the machinelarning case before. Both methods are also evaluated by method of ROC curve. The combined hypothesis definition for the two separate fields, allow for a comparison between the two models, in regard to each other's established evaluation methods. This study present a method and hypothesis to bridge the two separate fields of study, change point detection, and machinelearning, to achieve a comparable evaluation between the two.

Nearly all present-day commercial intrusion detection systems are based on a hierarchical architecture. In such an architecture, the root node is responsible for detecting intrusions and for issuing responses. However, an intrusion detection system (IDS) based on a hierarchical architecture has many single points of failure. For example, by disabling the root node, the intrusion-detection function of the IDS will also be disabled. To solve this problem, an IDS inspired by the human immune system is proposed. The proposed IDS has no single component that is responsible for detecting intrusions. Instead, the intrusion-detection function is divided and placed within mobile agents. Mobile agents act similarly to white blood cells of the human immune system and travel from host to host in the network to detect intrusions. The IDS is fault-tolerant because it can continue to detect intrusions even when most of its components have been disabled. / Computer Science (School of Computing) / M. Sc. (Computer Science)

The thesis deals with the attack detection possibilities and the nonstandard behaviour. It focuses on problems with the IDS detection systems, the subsequent classification and methods which are being used for the attack detection. One part of the thesis presents the existing IDS systems and their properties which are necessary for the successful attack detection. Other parts describe methods to obtain information from the operating systems Microsoft Windows and it also analyses the theoretical methods of data abnormalities. The practical part focuses on the design and implementation of the HIDS application. The final application and its detection abilities are tested at the end of the practical part with the help of some model situations. In the conclusion, the thesis sums up the gained information and shows a possible way of the future development.

Cette étude vise à développer un système d’aide au diagnostic (CAD) pour la détection de lésions épileptogènes, reposant sur l’analyse de données de neuroimagerie, notamment, l’IRM T1 et FLAIR. L’approche adoptée, introduite précédemment par Azami et al., 2016, consiste à placer la tâche de détection dans le cadre de la détection de changement à l'échelle du voxel, basée sur l’apprentissage d’un modèle one-class SVM pour chaque voxel dans le cerveau. L'objectif principal de ce travail est de développer des mécanismes d’apprentissage de représentations, qui capturent les informations les plus discriminantes à partir de l’imagerie multimodale. Les caractéristiques manuelles ne sont pas forcément les plus pertinentes pour la tâche visée. Notre première contribution porte sur l'intégration de différents réseaux profonds non-supervisés, pour extraire des caractéristiques dans le cadre du problème de détection de changement. Nous introduisons une nouvelle configuration des réseaux siamois, mieux adaptée à ce contexte. Le système CAD proposé a été évalué sur l’ensemble d’images IRM T1 des patients atteints d'épilepsie. Afin d'améliorer la performance obtenue, nous avons proposé d'étendre le système pour intégrer des données multimodales qui possèdent des informations complémentaires sur la pathologie. Notre deuxième contribution consiste donc à proposer des stratégies de combinaison des différentes modalités d’imagerie dans un système pour la détection de changement. Ce système multimodal a montré une amélioration importante sur la tâche de détection de lésions épileptogènes sur les IRM T1 et FLAIR. Notre dernière contribution se focalise sur l'intégration des données TEP dans le système proposé. Etant donné le nombre limité des images TEP, nous envisageons de synthétiser les données manquantes à partir des images IRM disponibles. Nous démontrons que le système entraîné sur les données réelles et synthétiques présente une amélioration importante par rapport au système entraîné sur les images réelles uniquement. / This work represents one attempt to develop a computer aided diagnosis system for epilepsy lesion detection based on neuroimaging data, in particular T1-weighted and FLAIR MR sequences. Given the complexity of the task and the lack of a representative voxel-level labeled data set, the adopted approach, first introduced in Azami et al., 2016, consists in casting the lesion detection task as a per-voxel outlier detection problem. The system is based on training a one-class SVM model for each voxel in the brain on a set of healthy controls, so as to model the normality of the voxel. The main focus of this work is to design representation learning mechanisms, capturing the most discriminant information from multimodality imaging. Manual features, designed to mimic the characteristics of certain epilepsy lesions, such as focal cortical dysplasia (FCD), on neuroimaging data, are tailored to individual pathologies and cannot discriminate a large range of epilepsy lesions. Such features reflect the known characteristics of lesion appearance; however, they might not be the most optimal ones for the task at hand. Our first contribution consists in proposing various unsupervised neural architectures as potential feature extracting mechanisms and, eventually, introducing a novel configuration of siamese networks, to be plugged into the outlier detection context. The proposed system, evaluated on a set of T1-weighted MRIs of epilepsy patients, showed a promising performance but a room for improvement as well. To this end, we considered extending the CAD system so as to accommodate multimodality data which offers complementary information on the problem at hand. Our second contribution, therefore, consists in proposing strategies to combine representations of different imaging modalities into a single framework for anomaly detection. The extended system showed a significant improvement on the task of epilepsy lesion detection on T1-weighted and FLAIR MR images. Our last contribution focuses on the integration of PET data into the system. Given the small number of available PET images, we make an attempt to synthesize PET data from the corresponding MRI acquisitions. Eventually we show an improved performance of the system when trained on the mixture of synthesized and real images.

L’imagerie hyperspectrale, qui consiste à acquérir l'image d'une scène dans un grand nombre de bandes spectrales, permet de détecter des cibles là où l'imagerie couleur classique ne permettrait pas de conclure. Les imageurs hyperspectraux à acquisition séquentielle sont inadaptés aux applications de détection en temps réel. Dans cette thèse, nous proposons d’utiliser un imageur multispectral snapshot, capable d’acquérir simultanément un nombre réduit de bandes spectrales sur un unique détecteur matriciel. Le capteur offrant un nombre de pixels limité, il est nécessaire de réaliser un compromis en choisissant soigneusement le nombre et les profils spectraux des filtres de l'imageur afin d’optimiser la performance de détection. Dans cet objectif, nous avons développé une méthode de sélection de bandes qui peut être utilisée dans la conception d’imageurs multispectraux basés sur une matrice de filtres fixes ou accordables. Nous montrons, à partir d'images hyperspectrales issues de différentes campagnes de mesure, que la sélection des bandes spectrales à acquérir peut conduire à des imageurs multispectraux capables de détecter des cibles ou des anomalies avec une efficacité de détection proche de celle obtenue avec une résolution hyperspectrale. Nous développons conjointement un démonstrateur constitué d'une matrice de 4 filtres de Fabry-Perot accordables électroniquement en vue de son implantation sur un imageur multispectral snapshot agile. Ces filtres sont développés en technologie MOEMS (microsystèmes opto-électro-mécaniques) en partenariat avec l'Institut d'Electronique Fondamentale. Nous présentons le dimensionnement optique du dispositif ainsi qu'une étude de tolérancement qui a permis de valider sa faisabilité. / Hyperspectral imaging, which consists in acquiring the image of a scene in a large number of spectral bands, can be used to detect targets that are not visible using conventional color imaging. Hyperspectral imagers based on sequential acquisition are unsuitable for real-time detection applications. In this thesis, we propose to use a snapshot multispectral imager able to acquire simultaneously a small number of spectral bands on a single image sensor. As the sensor offers a limited number of pixels, it is necessary to achieve a trade-off by carefully choosing the number and the spectral profiles of the imager’s filters in order to optimize the detection performance. For this purpose, we developed a band selection method that can be used to design multispectral imagers based on arrays of fixed or tunable filters. We use real hyperspectral images to show that the selection of spectral bands can lead to multispectral imagers able to compete against hyperspectral imagers for target detection and anomaly detection applications while allowing snapshot acquisition and real-time detection. We jointly develop an adaptive snapshot multispectral imager based on an array of 4 electronically tunable Fabry-Perot filters. The filters are developed in MOEMS technology (Micro-Opto-Electro-Mechanical Systems) in partnership with the Institut d'Electronique Fondamentale. We present the optical design of the device and a study of tolerancing which has validated its feasibility.

Orientador: Carlos Roberto Minussi / Banca: Francisco Villarreal Alvarado / Banca: Maria do Carmo Gomes da Silveira / Banca: Luciana Cambraia Leite / Banca: Ricardo Menezes Salgado / Resumo: Esta pesquisa visa o desenvolvimento de um método para análise da estabilidade transitória de sistemas de energia eletrica multimaquinas, por meio de uma rede neural ART-ARTMAP Fuzzy Euclidiana Modificada com Treinamento Continuado. Esta arquitetura apresenta tres diferenciais em e relação a outras já utilizadas para abordar tal problema: (1) a rede iniciada com apenas um neuronio ativado e vai se expandindo durante todo o o treinamento/análise, (2) possui um módulo de treinamento continuado e (3) a o possui um módulo de deteção de intruso. No primeiro diferencial, a redeé iniciada com um neuronio e vai se expandindo de acordo com a aquisição de conhecimento, isto faz com que esta se torne muito mais rápida e que o gasto computacional se torne mínimo. Com o módulo de treinamento continuado, a rede neural consegue armazenar novos dados sem a necessidade de realizar o retreinamento. Já o módulo de detecção de intruso faz com que, ao ser apresentada a rede uma configuração "estranha", a rede execute um treinamento específico para que esta configuração, com um número mínimo de entradas, seja incorporada definitivamente à rede neural. A aplicação para a rede proposta nesta pesquisa, foi a análise de estabilidade transitória, considerando-se o modelo clássico (estabilidade de primeira oscilação), para um sistema composto por 10 máquinas síncronas, 45 barras e 73 linhas de transmissão / Abstract: This doctoral research aims to develop a method to analyze the transient stability of multimachine eletric power systems, through a neural network Modified Euclidean Fuzzy ART-ARTMAP with Continuous Training. The architecture presented has three differences in relation to others used to deal with this problem: (1) the network starts with only one neuron activated and expands throughout the training/analysis, (2) has a continuous training module and (3) has an intrusion detection module. The first difference, is the fact that it starts with a neuron and expands according to knowledge acquisition of the network, and causes it to become much faster and the computational expenses becomes minimum. With continuous training mod- ule, the neural network can store the new data without the need for the retraining. The intrusion detection module causes, when presented to the network a strange configuration, the network to carry out a specific training for this configuration with a minimum total of inputs so that the configu- ration is definitely incorporated to the neural network. The application for this network, in this research, was to analyze the transient stability consid- ering the classical model (stability of first oscillation) to a system composed of 10 synchronous machines, 45 buses and 73 transmission lines / Doutor

Networks of distributed, remote sensors are providing ecological scientists with a view of our environment that is unprecedented in detail. However, these networks are subject to harsh conditions, which lead to malfunctions in individual sensors and failures in network communications. This behavior manifests as corrupt or missing measurements in the data. Consequently, before the data can be used in ecological models, future experiments, or even policy decisions, it must be quality controlled (QC'd) to flag affected measurements and impute corrected values. This dissertation describes a probabilistic modeling approach for real-time automated QC that exploits the spatial and temporal correlations in the data to distinguish sensor failures from valid observations. The model adapts to a site by learning a Bayesian network structure that captures spatial relationships among sensors, and then extends this structure to a dynamic Bayesian network to incorporate temporal correlations. The final QC model contains both discrete and continuous variables, which makes inference intractable for large sensor networks. Consequently, we examine the performance of three approximate methods for inference in this probabilistic framework. Two of these algorithms represent contemporary approaches to inference in hybrid models, while the third is a greedy search-based method of our own design. We demonstrate the results of these algorithms on synthetic datasets and real environmental sensor data gathered from an ecological sensor network located in western Oregon. Our results suggest that we can improve performance over networks with less sensors that use exhaustive asynchronic inference by including additional sensors and applying approximate algorithms. / Graduation date: 2013