Actualmente, los riesgos de diferente comportamiento (variabilidad y vinculados a eventos) pueden afectar a las edificaciones multifamiliares, generando importantes sobrecostos y desviaciones presupuestales. Esta tesis se enfoca en la elaboración de una metodología pauteada que sirva como herramienta aplicativa para cualquier proyecto de edificio multifamiliar, independientemente de su ubicación. Se ha diferenciado el análisis de los riesgos bajo las dos tipologías mencionadas por el PMBOK en su sexta edición (variabilidad y vinculados a eventos), como una tendencia novedosa. En la investigación primeramente se desarrolla una encuesta de juicio experto para identificar los riesgos más frecuentes en la muestra de estudio (distrito de Miraflores, ciudad de Lima). Seguidamente, se realiza una entrevista a los profesionales responsables de los tres proyectos que se toman como caso de estudio. En las mismas se recaban todos los datos necesarios para realizar a continuación el análisis cualitativo y el análisis cuantitativo de los diferentes riesgos investigados. Para el análisis cuantitativo se utiliza el método de Monte Carlo, a través del software @RISK. En cada proyecto se obtienen los valores de los riesgos en forma de contingencia, relacionados con el Costo Directo del presupuesto. Finalmente, a través de las convoluciones, se determinan los valores de la contingencia de los riesgos generales, o combinación simultánea de los riesgos de variabilidad y vinculados a eventos. Se puede concluir que esta metodología es muy beneficiosa al implementarse en etapas tempranas o de planificación, pues permite cuantificar los riesgos, en aras de mitigarlos y controlarlos durante la ejecución del proyecto. / Currently, the risks of different behavior (variability and linked to events) can affect multifamily buildings, generating significant cost overruns and budget deviations. This thesis focuses on the elaboration of a methodology that serves as an application tool for any multi-family building project, regardless of its location. The risk analysis has been differentiated under the two typologies mentioned by the PMBOK in its sixth edition (variability and linked to events), as a novel trend. The investigation first develops an expert judgment survey to identify the most frequent risks in the study sample (Miraflores district, Lima city). Next, an interview is conducted with the professionals responsible for the three projects that are taken as a case study. They collect all the necessary data to carry out the qualitative analysis and the quantitative analysis of the different risks investigated. The Monte Carlo method is used for quantitative analysis, through the @RISK software. In each project the values of the risks are obtained in the form of contingency, related to the Direct Cost of the budget. Finally, through the convolutions, the values of the contingency of the general risks, or simultaneous combination of the risks of variability and linked to events are determined. It can be concluded that this methodology is very beneficial when implemented in early stages or planning, since it allows quantifying the risks, in order to mitigate and control them during the execution of the project. / Tesis

Studying the preload in a screw joint has been the focus of today’s industry. The manufacturer reflects that demand by investigating different opportunities and techniques to develop this area. There are four different ways of controlling the tightening of bolts and joints to achieve the required clamp force that can hold for a specific preload. Torque control, angle control, gradient control, and ultrasonic clamp-force or elongation control. Many studies do exist about the first three mentioned techniques. However, there are a small number of studies for the ultrasonic clamp-force technique, and there is no study focusing on the usage of machine learning in that technique. This study investigates the use of machine learning to find the optimal frequency used to transmit the ultrasonic signals into the bolt for calculating the bolt elongation. Two machine learning models have been constructed, presenting two approaches: one for one-dimensional data (1D-CNN) and one for two-dimensional data (2D-CNN). The models classify the received signals (echos) with different frequencies into either accepted or non-accepted signals to get the optimal frequencies to be used later on, in the bolt elongation process. Both the 1D-CNN and 2D-CNN show an accepted performance of around 85% accuracy. The results indicate that there does exist a pattern in these ultrasonic signals that are useful for classifying them into accepted and non-accepted frequencies, so the usage of machine learning for the problem is feasible. / Att studera förspänningen i en skruvförband har varit i fokus för dagens industri. Tillverkaren speglar den efterfrågan genom att undersöka olika möjligheter och tekniker för att utveckla detta område. Det finns fyra olika sätt att kontroller åtdragningen av bultar för att uppnå den erforderliga klämkraften som kan hålla för en specifik förspänning. Vridmomentkontroll, vinkelkontroll, gradientkontroll och ultraljudskontroll av klämkraft. Det finns många studier om de tre förstnämnda teknologier. Det finns dock ett litet antal studier för ultraljudsklämkraftstekniken, och det finns ingen studie som fokuserar på användningen av maskininlärning i den tekniken. Denna studie undersöker användningen av maskininlärning för att hitta den optimala frekvensen som används för att beräkna bultens förlängning. Två maskininlärningsmodeller har konstruerats, som presenterar två metoder: en för endimensionell data (1D-CNN) och en för två-dimensionella data (2D-CNN). Modellerna klassificerar de mottagna signalerna (ekon) med olika frekvenser i antingen accepterade eller icke-accepterade signaler för att få de optimala frekvenserna att användas senare, i bultförlängningsprocessen. Både 1D-CNN och 2D-CNN visar en accepterad prestanda på cirka 85% noggrannhet. Resultaten indikerar att det finns ett mönster i dessa ultraljudssignaler som är användbara för att klassificera dem i accepterade och icke-accepterade frekvenser, så användningen av maskininlärning för problemet är genomförbar.

Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) / A detecção de anomalias em vídeos de vigilância é um tema de pesquisa recorrente em visão computacional. Os métodos de aprendizagem profunda têm alcançado o estado da arte para o reconhecimento de padrões em imagens e o Autocodificador Convolucional (ACC) é uma das abordagens mais utilizadas por sua capacidade em capturar as estruturas 2D dos objetos. Neste trabalho, a detecção de anomalias se refere ao problema de encontrar padrões em vídeos que não pertencem a um conceito normal esperado. Com o objetivo de classificar anomalias adequadamente, foram verificadas formas de aprender representações relevantes para essa tarefa. Por esse motivo, estudos tanto da capacidade do modelo em aprender características automaticamente quanto do efeito da fusão de características extraídas manualmente foram realizados. Para problemas de detecção de anomalias do mundo real, a representação da classe normal é uma questão importante, sendo que um ou mais agrupamentos podem descrever diferentes aspectos de normalidade. Para fins de classificação, esses agrupamentos devem ser tão compactos (densos) quanto possível. Esta tese propõe o uso do ACC como uma abordagem orientada a dados aplicada ao contexto de detecção de anomalias em vídeos. Foram propostos métodos para o aprendizado de características espaço-temporais, bem como foi introduzida uma abordagem híbrida chamada Autocodificador Convolucional com Incorporação Compacta (ACC-IC), cujo objetivo é melhorar a compactação dos agrupamentos normais. Além disso, foi proposto um novo critério de parada baseado na sensibilidade e sua adequação para problemas de detecção de anomalias foi verificada. Todos os métodos propostos foram avaliados em conjuntos de dados disponíveis publicamente e comparados com abordagens estado da arte. Além do mais, foram introduzidos dois novos conjuntos de dados projetados para detecção de anomalias em vídeos de vigilância em rodovias. O ACC se mostrou promissor na detecção de anomalias em vídeos. Resultados sugerem que o ACC pode aprender características espaço-temporais automaticamente e a agregação de características extraídas manualmente parece ser valiosa para alguns conjuntos de dados. A compactação introduzida pelo ACC-IC melhorou o desempenho de classificação para a maioria dos casos e o critério de parada baseado na sensibilidade é uma nova abordagem que parece ser uma alternativa interessante. Os vídeos foram analisados qualitativamente de maneira visual, indicando que as características aprendidas com os dois métodos (ACC e ACC-IC) estão intimamente correlacionadas com os eventos anormais que ocorrem em seus quadros. De fato, ainda há muito a ser feito para uma definição mais geral e formal de normalidade, de modo que se possa ajudar pesquisadores a desenvolver métodos computacionais eficientes para a interpretação dos vídeos. / The anomaly detection in automated video surveillance is a recurrent topic in recent computer vision research. Deep Learning (DL) methods have achieved the state-of-the-art performance for pattern recognition in images and the Convolutional Autoencoder (CAE) is one of the most frequently used approach, which is capable of capturing the 2D structure of objects. In this work, anomaly detection refers to the problem of finding patterns in images and videos that do not belong to the expected normal concept. Aiming at classifying anomalies adequately, methods for learning relevant representations were verified. For this reason, both the capability of the model for learning automatically features and the effect of fusing hand-crafted features together with raw data were studied. Indeed, for real-world problems, the representation of the normal class is an important issue for detecting anomalies, in which one or more clusters can describe different aspects of normality. For classification purposes, these clusters must be as compact (dense) as possible. This thesis proposes the use of CAE as a data-driven approach in the context of anomaly detection problems. Methods for feature learning using as input both hand-crafted features and raw data were proposed, and how they affect the classification performance was investigated. This work also introduces a hybrid approach using DL and one-class support vector machine methods, named Convolutional Autoencoder with Compact Embedding (CAE-CE), for enhancing the compactness of normal clusters. Besides, a novel sensitivity-based stop criterion was proposed, and its suitability for anomaly detection problems was assessed. The proposed methods were evaluated using publicly available datasets and compared with the state-of-the-art approaches. Two novel benchmarks, designed for video anomaly detection in highways were introduced. CAE was shown to be promising as a data-driven approach for detecting anomalies in videos. Results suggest that the CAE can learn spatio-temporal features automatically, and the aggregation of hand-crafted features seems to be valuable for some datasets. Also, overall results suggest that the enhanced compactness introduced by the CAE-CE improved the classification performance for most cases, and the stop criterion based on the sensitivity is a novel approach that seems to be an interesting alternative. Videos were qualitatively analyzed at the visual level, indicating that features learned using both methods (CAE and CAE-CE) are closely correlated to the anomalous events occurring in the frames. In fact, there is much yet to be done towards a more general and formal definition of normality/abnormality, so as to support researchers to devise efficient computational methods to mimetize the semantic interpretation of visual scenes by humans.

CAPES / Este trabalho tem como objetivo propor um método reconhecimento de pessoas em imagens aéreas obtidas a partir de Veículos Aéreos Não Tripulados de pequeno porte. Esta é uma aplicação de grande interesse, pois pode ser inserida em diversas situações tanto civis quanto militares como, por exemplo, missões de busca e salvamento. O uso de Veículos Aéreos Não Tripulados autônomos tende a aumentar com o barateamento desta tecnologia. Assim, esta tecnologia pode sobressair sobre outras utilizadas atualmente, como satélites e voos com grandes aeronaves. Para o reconhecimento de pessoas em imagens aéreas de forma autônoma, este trabalho propõe métodos na forma de Sistemas de Reconhecimento de Padrões (SRP) aplicados ao reconhecimento de imagens. Para este métodos, foram testadas quatro técnicas de aprendizado de máquina: Redes Neurais Convolucionais, HOG+SVM, Cascata Haar e Cascata LBP. Além disso, a fim de possibilitar o reconhecimento de pessoas em imagens aéreas em tempo real, foram testadas e avaliadas técnicas de detecção e segmentação de objetos: Mapa de Saliências e o Processamento de Imagens Térmicas de baixa resolução (PIT). Neste trabalho foram avaliadas as taxas de reconhecimento dos SRPs, além do seu tempo de processamento em um sistema embarcado de baixo custo e em uma Base de Controle Móvel (BCM). Os resultados de reconhecimento mostraram a efetividade das Redes Neurais Convolucionais, com uma acurácia de 0,9971, seguido do HOG+SVM com 0,9236, Cascata Haar com 0,7348 e por fim, Cascata LBP com 0,6615. Em situações onde foi simulado a oclusão parcial, as Redes Neurais Convolucionais atingiram Sensibilidade média 0,72, HOG+SVM de 0,50 e as Cascatas 0,20. Nos experimentos com os SRPs (algoritmos de segmentação e detecção juntamente com as técnicas de reconhecimento), o Mapa de Saliências pouco afetou as taxas de reconhecimento, quais ficaram muito próximas das obtidas no experimentos de reconhecimento. Já o Processamento de Imagens Térmicas de baixa resolução apresentou dificuldades em executar uma segmentação precisa, obtendo imagens com variação na translação, prejudicando a precisão do sistema. Por fim, este trabalho propõe uma nova abordagem para implementação de um SRP para o reconhecimento de pessoas em imagens áreas, utilizando Processamento de Imagens Térmicas juntamente com as Redes Neurais Convolucionais. Este SRP une altas taxas de reconhecimento com desempenho computacional de ao menos 1 fps na plataforma BCM. / This work aims to propose a method for people recognition in Small Unmanned Aerial Vehicles aerial imagery. This is an application of high interest, it can be used in several situations, both civilian and military, as search and rescue missions. The use of Unmanned Aerial Vehicles autonomously tends to increase with the cheapening of this technology, supporting search and rescue missions. Thus, this technology can excel over others currently used, as satellites and flights with large aircraft. For autonomous people recognition, this work proposes new methods as Pattern Reconigition System (PRS) applied to image recognition, applying it in aerial images. Four Pattern Reconigition techniques were tested: Convolutional Neural Networks, HOG+SVM, Haar Cascade and LBP Cascade. Furthermore, in order to achieve recognition of people in aerial images in Real-Time target and detection techniques were tested and evaluated: Saliency Maps and Low-resolution Thermal Image Processing (TIP). In this work were considered recognition rates of the methods and their computational time in a low-cost embedded system and a Mobile Ground Control Station (MGCS). The recognition results shown the Convolutional Neural Network potential, where an accuracy of 0.9971 was achieved, followed by HOG + SVM with 0.9236, Haar Cascade with 0.7348 and LBP Cascade with 0.6615. In situations simulated partial occlusion, where was the CNNs achieved average Sensitivity of 0.72, HOG+SVM with 0.50 and both Cascades 0.20. In experiments with PRS (targeting and detection algorithms with the recognition techniques), the Saliency Map had little influence in recongition rates, it was close to the rates achieved in recognition experiments. While the Low-resolution Thermal Image Processing had difficulties in segmentation process, where translation variantions occured, it harmed the system precision. Lastly, this work proposes a new approach for PRS implementation for people recognition in aerial imagery, using TIP with CNN. This PRS combines high rates of recognition with an computational performace of, at least, 1 fps in MGCS plataform.

CAPES / Este trabalho tem como objetivo propor um método reconhecimento de pessoas em imagens aéreas obtidas a partir de Veículos Aéreos Não Tripulados de pequeno porte. Esta é uma aplicação de grande interesse, pois pode ser inserida em diversas situações tanto civis quanto militares como, por exemplo, missões de busca e salvamento. O uso de Veículos Aéreos Não Tripulados autônomos tende a aumentar com o barateamento desta tecnologia. Assim, esta tecnologia pode sobressair sobre outras utilizadas atualmente, como satélites e voos com grandes aeronaves. Para o reconhecimento de pessoas em imagens aéreas de forma autônoma, este trabalho propõe métodos na forma de Sistemas de Reconhecimento de Padrões (SRP) aplicados ao reconhecimento de imagens. Para este métodos, foram testadas quatro técnicas de aprendizado de máquina: Redes Neurais Convolucionais, HOG+SVM, Cascata Haar e Cascata LBP. Além disso, a fim de possibilitar o reconhecimento de pessoas em imagens aéreas em tempo real, foram testadas e avaliadas técnicas de detecção e segmentação de objetos: Mapa de Saliências e o Processamento de Imagens Térmicas de baixa resolução (PIT). Neste trabalho foram avaliadas as taxas de reconhecimento dos SRPs, além do seu tempo de processamento em um sistema embarcado de baixo custo e em uma Base de Controle Móvel (BCM). Os resultados de reconhecimento mostraram a efetividade das Redes Neurais Convolucionais, com uma acurácia de 0,9971, seguido do HOG+SVM com 0,9236, Cascata Haar com 0,7348 e por fim, Cascata LBP com 0,6615. Em situações onde foi simulado a oclusão parcial, as Redes Neurais Convolucionais atingiram Sensibilidade média 0,72, HOG+SVM de 0,50 e as Cascatas 0,20. Nos experimentos com os SRPs (algoritmos de segmentação e detecção juntamente com as técnicas de reconhecimento), o Mapa de Saliências pouco afetou as taxas de reconhecimento, quais ficaram muito próximas das obtidas no experimentos de reconhecimento. Já o Processamento de Imagens Térmicas de baixa resolução apresentou dificuldades em executar uma segmentação precisa, obtendo imagens com variação na translação, prejudicando a precisão do sistema. Por fim, este trabalho propõe uma nova abordagem para implementação de um SRP para o reconhecimento de pessoas em imagens áreas, utilizando Processamento de Imagens Térmicas juntamente com as Redes Neurais Convolucionais. Este SRP une altas taxas de reconhecimento com desempenho computacional de ao menos 1 fps na plataforma BCM. / This work aims to propose a method for people recognition in Small Unmanned Aerial Vehicles aerial imagery. This is an application of high interest, it can be used in several situations, both civilian and military, as search and rescue missions. The use of Unmanned Aerial Vehicles autonomously tends to increase with the cheapening of this technology, supporting search and rescue missions. Thus, this technology can excel over others currently used, as satellites and flights with large aircraft. For autonomous people recognition, this work proposes new methods as Pattern Reconigition System (PRS) applied to image recognition, applying it in aerial images. Four Pattern Reconigition techniques were tested: Convolutional Neural Networks, HOG+SVM, Haar Cascade and LBP Cascade. Furthermore, in order to achieve recognition of people in aerial images in Real-Time target and detection techniques were tested and evaluated: Saliency Maps and Low-resolution Thermal Image Processing (TIP). In this work were considered recognition rates of the methods and their computational time in a low-cost embedded system and a Mobile Ground Control Station (MGCS). The recognition results shown the Convolutional Neural Network potential, where an accuracy of 0.9971 was achieved, followed by HOG + SVM with 0.9236, Haar Cascade with 0.7348 and LBP Cascade with 0.6615. In situations simulated partial occlusion, where was the CNNs achieved average Sensitivity of 0.72, HOG+SVM with 0.50 and both Cascades 0.20. In experiments with PRS (targeting and detection algorithms with the recognition techniques), the Saliency Map had little influence in recongition rates, it was close to the rates achieved in recognition experiments. While the Low-resolution Thermal Image Processing had difficulties in segmentation process, where translation variantions occured, it harmed the system precision. Lastly, this work proposes a new approach for PRS implementation for people recognition in aerial imagery, using TIP with CNN. This PRS combines high rates of recognition with an computational performace of, at least, 1 fps in MGCS plataform.

In this masters thesis we focused on using machine learning methods for detecting money laundering in financial transaction networks, in order to demonstrate that it can be used as a complement or instead of the more commonly used rule based systems. The graph learning method graph convolutional networks (GCN) has been a hot topic in the field since they were shown to scale well with data size back in 2018. However the typical GCN models cannot use edge features, which is why this thesis combines the GCN model with a node and edge neural network (NENN) in order to solve this problem. This new method will be compared towards an already established machine learning method for financial transactions, namely the tree boosting method (XGBoost). Because of confidentiality concerns for financial transactions data, the machine learning algorithms will be tested on two carefully constructed synthetically generated data sets, which from agent based simulations resembles real financial data. The results showed the viability and superiority of the new implementation of the GCN model with it being a preferable method for connectivly structured data, meaning that a transaction or account is analyzed in the context of its financial environment. On the other hand the XGBoost method showed better results when examining transactions independently. Hence it was more accurately able to find fraudulent and non fraudulent patterns from the transactional features themselves. / I detta examensarbete fokuserar vi på användandet av maskininlärningsmetoder för att detektera penningtvätt i finansiella transaktionsnätverk, med målet att demonstrera att dess kan användas som ett komplement till eller i stället för de mer vanligt använda regelbaserade systemen. Grafinlärningsmetoden \textit{graph convolutional networks} (GCN) som har varit ett hett ämne inom området sedan metoden under 2018 visades fungera bra för stora datamängder. Däremot kan inte en vanlig GCN-modell använda kantinformation, vilket är varför denna avhandling kombinerar GCN-modellen med \textit{node and edge neural networks} (NENN) för att mer effektivt detektera penningtvätt. Denna nya metod kommer att jämföras med en redan etablerad maskininlärningsmetod för finansiella transaktioner, nämligen \textit{tree boosting} (XGBoost). På grund av sekretessanledningar för finansiella transaktionsdata var maskininlärningsalgoritmerna testade på två noggrant konstruerade syntetiskt genererade datamängder som från agentbaserade simuleringar liknar riktiga finansiella data. Resultaten visade på applikationsmöjligheter och överlägsenhet för den nya implementationen av GCN-modellen vilken är att föredra för relationsstrukturerade data, det vill säga när transaktioner och konton analyseras i kontexten av deras finansiella omgivning. Å andra sidan visar XGBoost bättre resultat på att examinera transaktioner individuellt eftersom denna metod mer precist kan identifiera bedrägliga och icke-bedrägliga mönster från de transnationella funktionerna.

L’approximation de phénomènes physiques, tels qu’une simulation de liquides en informatique graphique, requiert l’utilisation de méthodes complexes nécessitant des temps de calcul et une quantité de mémoire importants. Malgré les avancées récentes dans ce domaine, l’écart en réalisme entre un liquide simulé et la réalité demeure encore aujourd’hui considérable. Cet écart nous séparant du réalisme souhaité nécessite des modèles numériques de simulation dont la complexité ne cesse de croître. L’objectif ultime est de permettre à l’utilisateur de manipuler ces modèles de simulation de liquides sans la nécessité d’avoir une connaissance accrue de la physique requise pour atteindre un niveau de réalisme acceptable et ce, en temps réel. Plusieurs approches ont été revisitées dans les dernières années afin de simplifier ces modèles ou dans le but de les rendre plus facilement paramétrables. Cette thèse par articles encadre bien les trois projets constituant nos contributions dans le but d’améliorer et de faciliter la génération de simulations de liquides en informatique graphique. Tout d’abord, nous introduisons une approche hybride permettant de traiter séparément le volume de liquide non-apparent (i.e., en profondeur) et une couche de particules en surface par la méthode de calcul Smoothed Particle Hydrodynamics (SPH). Nous revisitons l’approche par bandes de particules, mais cette fois nouvellement appliquée à la méthode SPH qui offre un niveau de réalisme supérieur. Comme deuxième projet, nous proposons une approche permettant d’améliorer le niveau de détail des éclaboussures de liquides. En suréchantillonnant une simulation de liquides existante, notre approche est capable de générer des détails réalistes d’éclaboussures grâce à la dynamique de balistique. En complément, nous proposons une méthode de simulation par vagues permettant de reproduire les interactions entre les éclaboussures générées et les portions quasi-statiques de la simulation existante. Finalement, le troisième projet introduit une approche permettant de rehausser la résolution apparente d’un liquide par l’apprentissage automatique. Nous proposons une architecture d’apprentissage inspirée des flux optiques dont l’objectif est de générer une correspondance entre le déplacement des particules de simulations de liquides à différentes résolutions (i.e., basses et hautes résolutions). Notre modèle d’apprentissage permet d’encoder des caractéristiques de hautes résolutions à l’aide de déformations pré-calculées entre deux liquides à différentes résolutions et d’opérations de convolution basées sur le voisinage des particules. / The approximation of natural phenomena such as liquid simulations in computer graphics requires complex methods that are computationally expensive. Despite recent advances in this field, the gap in realism between a simulated liquid and reality remains considerable. This disparity that separates us from the desired realism requires numerical models whose complexity continues to grow. The ultimate goal is to provide users the capacity and tools to manipulate these liquid simulation models to obtain acceptable realism. In the last decade, several approaches have been revisited to simplify and to allow more flexible models. In this dissertation by articles, we present three projects whose contributions support the improvement and flexibility of generating liquid simulations for computer graphics. First, we introduce a hybrid approach allowing us to separately process the volume of non-apparent liquid (i.e., in-depth) and a band of surface particles using the Smoothed Particle Hydrodynamics (SPH) method. We revisit the particle band approach, but this time newly applied to the SPH method, which offers a higher level of realism. Then, as a second project, we propose an approach to improve the level of detail of splashing liquids. By upsampling an existing liquid simulation, our approach is capable of generating realistic splash details through ballistic dynamics. In addition, we propose a wave simulation method to reproduce the interactions between the generated splashes and the quasi-static portions of the existing liquid simulation. Finally, the third project introduces an approach to enhance the apparent resolution of liquids through machine learning. We propose a learning architecture inspired by optical flows by which we generate a correspondence between the displacement of the particles of liquid simulations at different resolutions (i.e., low and high resolutions). Our training model allows high-resolution features to be encoded using pre-computed deformations between two liquids at different resolutions and convolution operations based on the neighborhood of the particles.

No description available.