Дисертації з теми "Data-driven techniques"

One of the key goals of current research in data-driven computer animation is the synthesis of new motion sequences from existing motion data. This thesis presents three novel techniques for synthesising the motion of a virtual character from existing motion data and develops a framework of solutions to key character animation problems. The first motion synthesis technique presented is based on the character's locomotion composition process. This technique examines the ability of synthesising a variety of character's locomotion behaviours while easily specified constraints (footprints) are placed in the three-dimensional space. This is achieved by analysing existing motion data, and by assigning the locomotion behaviour transition process to transition graphs that are responsible for providing information about this process. However, virtual characters should also be able to animate according to different style variations. Therefore, a second technique to synthesise real-time style variations of character's motion. A novel technique is developed that uses correlation between two different motion styles, and by assigning the motion synthesis process to a parameterised maximum a posteriori (MAP) framework retrieves the desire style content of the input motion in real-time, enhancing the realism of the new synthesised motion sequence. The third technique presents the ability to synthesise the motion of the character's fingers either o↵-line or in real-time during the performance capture process. The advantage of both techniques is their ability to assign the motion searching process to motion features. The presented technique is able to estimate and synthesise a valid motion of the character's fingers, enhancing the realism of the input motion. To conclude, this thesis demonstrates that these three novel techniques combine in to a framework that enables the realistic synthesis of virtual character movements, eliminating the post processing, as well as enabling fast synthesis of the required motion.

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2017.
This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.
Cataloged from student-submitted PDF version of thesis.
Includes bibliographical references (pages 153-162).
Interactive visualizations are a popular medium used by scientists to explore, analyze and generally make sense of their data. However, with the overwhelming amounts of data that scientists collect from various instruments (e.g., telescopes, satellites, gene sequencers and field sensors), they need ways of efficiently transforming their data into interactive visualizations. Though a variety of visualization tools exist to help people make sense of their data, these tools often rely on database management systems (or DBMSs) for data processing and storage; and unfortunately, DBMSs fail to process the data fast enough to support a fluid, interactive visualization experience. This thesis blends optimization techniques from databases and methodology from HCI and visualization in order to support interactive and iterative exploration of large datasets. Our main goal is to reduce latency in visualization systems, i.e., the time these systems spend responding to a user's actions. We demonstrate through a comprehensive user study that latency has a clear (negative) effect on users' high-level analysis strategies, which becomes more pronounced as the latency is increased. Furthermore, we find that users are more susceptible to the effects of system latency when they have existing domain knowledge, a common scenario for data scientists. We then developed a visual exploration system called Sculpin that utilizes a suite of optimizations to reduce system latency. Sculpin learns user exploration patterns automatically, and exploits these patterns to pre-fetch data ahead of users as they explore. We then combine data-prefetching with incremental data processing (i.e., incremental materialization) and visualization-focused caching optimizations to further boost performance. With all three of these techniques (pre-fetching, caching, and pre-computation), Sculpin is able to: create visualizations 380% faster and respond to user interactions 88% faster than existing visualization systems, while also using less than one third of the space required by other systems to store materialized query results.
by Leilani Battle.
Ph. D.

Our society is becoming more digitalized for each day. Now, we are able to gather data from individual users with higher resolution than ever. With the increased amount of data on an individual user level, we can analyze their behavior. This is of interest in many different domains, for example service providers wanting to improve their service for their customers. If they know how their service is used, they have more insight in how they can improve. But, it also imposes additional difficulties. When we reach the individual user, the irregularities in the regular behavior makes it harder to model the normal behavior. In this thesis, we explore data-driven techniques to model and analyze user behaviors. We aim to evaluate existing as well as develop novel technologies to identify approaches that are suitable for use on an individual user level. We use both supervised and unsupervised learning methods to model the user behavior and evaluate the approaches on real world electricity consumption data. Firstly, we analyze household electricity consumption data and investigate the use of regression to model the household's behavior. We identify consumption trends, how data granularity affects modeling, and we show that regression is a viable approach to model user behavior. Secondly, we use clustering analysis to profile individual households in terms of their electricity consumption. We compare two dissimilarity measures, how they affect the clustering analysis, and we investigate how the produced clustering solutions differ. Thirdly, we propose a sequential clustering algorithm to model evolving user behavior. We evaluate the proposed algorithm on electricity consumption data and show how the produced model can be used to identify and trace changes in the user's behavior. The algorithm is robust to evolving behaviors and handles both dynamic and incremental aspects of streaming data.

Data acquisition systems used in remote environmental monitoring equipment and biological sensor nodes rely on limited energy supply soured from either energy harvesters or battery to perform their functions. Among the building blocks of these systems are power hungry Analogue to Digital Converters and Digital Signal Processors which acquire and process samples at predetermined rates regardless of the monitored signal's behavior. In this work we investigate power efficient event driven data acquisition and processing techniques by implementing an asynchronous ADC and an event driven power gated Finite Impulse Response (FIR) filter. We present an event driven single slope ADC capable of generating asynchronous digital samples based on the input signal's rate of change. It utilizes a rate of change detection circuit known as the slope detector to determine at what point the input signal is to be sampled. After a sample has been obtained it's absolute voltage value is time encoded and passed on to a Time to Digital Converter (TDC) as part of a pulse stream. The resulting digital samples generated by the TDC are produced at a rate that exhibits the same rate of change profile as that of the input signal. The ADC is realized in 0.35mm CMOS process, covers a silicon area of 340mm by 218mm and consumes power based on the input signal's frequency. The samples from the ADC are asynchronous in nature and exhibit random time periods between adjacent samples. In order to process such asynchronous samples we present a FIR filter that is able to successfully operate on the samples and produce the desired result. The filter also poses the ability to turn itself off in-between samples that have longer sample periods in effect saving power in the process.

L'objectif de ce travail de thèse est de proposer une approche permettant l'automatisation complète du processus de transformation de modèles pour le développement d'entrepôts de données. L'idée principale est de réduire au mieux l'intervention des experts humains en utilisant les traces de transformations réalisées sur des projets similaires. L'objectif est d'utiliser des techniques d'apprentissage supervisées pour traiter les définitions de concepts avec le même niveau d'expression que les données manipulées. La nature des données manipulées nous a conduits à choisir les langages relationnels pour la description des exemples et des hypothèses. Ces langages ont l'avantage d'être expressifs en donnant la possibilité d'exprimer les relations entres les objets manipulés mais présente l'inconvénient majeur de ne pas disposer d'algorithmes permettant le passage à l'échelle pour des applications industrielles. Pour résoudre ce problème, nous avons proposé une architecture permettant d'exploiter au mieux les connaissances issues des invariants de transformations entre modèles et métamodèles. Cette manière de procéder a mis en lumière des dépendances entre les concepts à apprendre et nous a conduits à proposer un paradigme d'apprentissage dit de concepts-dépendants. Enfin, cette thèse présente plusieurs aspects qui peuvent influencer la prochaine génération de plates-formes décisionnelles. Elle propose, en particulier, une architecture de déploiement pour la business intelligence en tant que service basée sur les normes industrielles et les technologies les plus récentes et les plus prometteuses
This thesis aims at proposing an end-to-end approach which allows the automation of the process of model transformations for the development of data warehousing components. The main idea is to reduce as much as possible the intervention of human experts by using once again the traces of transformations produced on similar projects. The goal is to use supervised learning techniques to handle concept definitions with the same expressive level as manipulated data. The nature of the manipulated data leads us to choose relational languages for the description of examples and hypothesises. These languages have the advantage of being expressive by giving the possibility to express relationships between the manipulated objects, but they have the major disadvantage of not having algorithms allowing the application on large scales of industrial applications. To solve this problem, we have proposed an architecture that allows the perfect exploitation of the knowledge obtained from transformations' invariants between models and metamodels. This way of proceeding has highlighted the dependencies between the concepts to learn and has led us to propose a learning paradigm, called dependent-concept learning. Finally, this thesis presents various aspects that may inuence the next generation of data warehousing platforms. The latter suggests, in particular, an architecture for business intelligence-as-a-service based on the most recent and promising industrial standards and technologies

This thesis explores state of the art robust multivariate statistical methods and data driven techniques to holistically perform prognostics and health management (PHM). This provides a means to enable the early detection, diagnosis and prognosis of future asset failures. In this thesis, the developed PHM methodology is applied to wind turbine drive train components, specifically focussed on planetary gearbox bearings and gears. A novel methodology for the identification of relevant time-domain statistical features based upon robust statistical process control charts is presented for high frequency bearing accelerometer data. In total, 28 time-domain statistical features were evaluated for their capabilities as leading indicators of degradation. The results of this analysis describe the extensible multivariate “Moments’ model” for the encapsulation of bearing operational behaviour. This is presented, enabling the early degradation of detection, predictive diagnostics and estimation of remaining useful life (RUL). Following this, an extended physics of failure model based upon low frequency SCADA data for the quantification of wind turbine gearbox condition is described. This extends the state of the art, whilst defining robust performance charts for quantifying component condition. Normalisation against loading of the turbine and transient states based upon empirical data is performed in the bivariate domain, with extensibility into the multivariate domain if necessary. Prognosis of asset condition is found to be possible with the assistance of artificial neural networks in order to provide business intelligence to the planning and scheduling of effective maintenance actions. These multivariate condition models are explored with multivariate distance and similarity metrics for to exploit traditional data mining techniques for tacit knowledge extraction, ensemble diagnosis and prognosis. Estimation of bearing remaining useful life is found to be possible, with the derived technique correlating strongly to bearing life (r = .96).

Thesis: Ph. D., Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center, 2019
Cataloged from PDF version of thesis.
Includes bibliographical references (pages 141-145).
In the design and operation of urban mobility systems, it is often desirable to understand patterns in traveler demand. However, demand is typically unobserved and must be estimated from available data. To address this disconnect, we begin by proposing a method for recovering an unknown probability distribution given a censored or truncated sample from that distribution. The proposed method is a novel and conceptually simple detruncation technique based on sampling the observed data according to weights learned by solving a simulation-based optimization problem; this method is especially appropriate in cases where little analytic information about the unknown distribution is available but the truncation process can be simulated.
The proposed method is compared to the ubiquitous maximum likelihood (MLE) method in a variety of synthetic validation experiments where it is found that the proposed method performs slightly worse than perfectly specified MLE and competitively with slight misspecified MLE. We then describe a novel car-sharing simulator which captures many of the important interactions between supply, demand, and system utilization while remaining simple and computationally efficient. In collaboration with Zipcar, a leading car-sharing operator in the United States, we demonstrate the usefulness of our detruncation method combined with our simulator via a pair of case studies. These tools allow us to estimate demand for round trip car-sharing services in the Boston and New York metropolitan areas, and the inferred demand distributions contain actionable insights.
Finally, we extend the detruncation method to cover cases where data is noisy, missing, or must be combined from different sources such as web or mobile applications. In synthetic validation experiments, the extended method is benchmarked against kernel density estimation (KDE) with Gaussian kernels. We find that the proposed method typically outperforms KDE, especially when the distribution to be estimated is not unimodal. With this extended method we consider the added utility of search data when estimating demand for car-sharing.
by Evan Fields.
Ph. D.
Ph.D. Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center

The core of many typical large-scale industrial infrastructure consists of hundreds or thousands of systems that are similar in their basic design and purpose. For instance, District Heating (DH) utilities rely on a large network of substations to deliver heat to their customers. Similarly, a factory may require a large fleet of specialized robots for manufacturing a certain product. Monitoring these systems is important for maintaining the overall efficiency of industrial operations by detecting various problems due to faults and misconfiguration. However, this can be challenging since a well-understood prior model for each system is rarely available. In most cases, each system in a fleet or network is fitted with a set of sensors to measure its state at different time intervals. Typically, a data-driven model for each system can be used for their monitoring. However, not all factors that can possibly influence the operations of each system in a fleet or network has an associated sensor. Moreover, sufficient instances of normal, atypical and faulty behavior are rarely available to train such a model. These issues can impede the effectiveness of a system level data-driven model. Alternatively, it can be assumed that since all the systems in a fleet or network are working on a similar task, they should all behave in a homogeneous manner. Any system that behaves differently from the majority is then considered as an outlier. This is referred to as the global model at the fleet or network level. While the approach is simple, it is less effective in the presence of non-stationary working conditions. Hence, both system level and global modeling approaches have their limitations.  This thesis investigates system level and fleet or network level (global) models for large-scale monitoring, and proposes an alternative way which is referred to as a reference-group based approach. Herein, the operational monitoring of each system, referred to as a target, is delegated to a reference-group, which consists of systems experiencing a comparable operating regime along with the target. Thus, the definition of a normal, atypical or faulty operational behavior in a target system is described relative to its reference-group. In this sense, if the target system is not behaving operationally in consort with the systems in its reference-group, then it can be inferred that this is either due to a fault or because of some atypical operation arising at the target system due to its local peculiarities. The application area for these investigations is the large-scale operational monitoring of thermal energy systems: networks of district heating (DH) substations and fleets of heatpumps. The current findings indicate three advantages of a reference-group based approach. The first is that the reference operational behavior of any system in the fleet or network does not need to be predefined. The second is that it provides a basis for what a system’s operational behavior should have been and what it is. In this respect, each system in the reference-group provides an evidence about a particular behavior during a particular time period. This can be very useful when the description of a normal, atypical and faulty operational behavior is not available. The third is that it can detect potential atypical and faulty operational behavior quicker compared to global models of outlier detection at the fleet or network level.

This doctoral thesis is the result of the supervision and collaboration of the University of Basilicata, the Polytechnic of Bari, and the enterprise Master Italy s.r.l. The main research lines explored and discussed in the thesis are: sustainability in general and, more specifically, manufacturing sustainability, the Industry 4.0 paradigm linked to smart (green) manufacturing, model-based assessment techniques of manufacturing processes, and data-driven analysis methodologies. These seemingly unrelated topics are handled throughout the thesis in such a way that it reveal how strongly interwoven and characterised by transversality they are. The goal of the PhD programme was to design and validate innovative assessment models in order to investigate the nature of manufacturing processes and rationalize the relationships and correlations between the different stages of the process. This composite model may be utilized as a tool in political decision-making about the long-term development of industrial processes and the continuous improvement of manufacturing processes. The overarching goal of this research is to provide strategies for real-time monitoring of manufacturing performance and sustainability based on hybrid thermodynamic models of the first and second order, as well as those based on data and machine learning. The proposed model is tested on a real industrial case study using a systemic approach: the phases of identifying the requirements, data inventory (materials, energetic, geometric, physical, economic, social, qualitative, quantitative), modelling, analysis, ad hoc algorithm adjustment (tuning), implementation, and validation are developed for the aluminium alloy die-casting processes of Master Italy s.r.l., a southern Italian SME which designs and produces the accessories and metal components for windows since 1986. The thesis digs in the topic of the sustainability of smart industrial processes from each and every perspective, including both the quantity and quality of resources used throughout the manufacturing process's life cycle. Traditional sustainability analysis models (such as life cycle analysis, LCA) are combined with approaches based on the second law of thermodynamics (exergetic analysis); they are then complemented by models based on information technology (big-data analysis). A full analysis of the potential of each strategy, whether executed alone or in combination, is provided. Following a summary of the metrics relevant for determining the degree of sustainability of industrial processes, the case study is demonstrated using modelling and extensive analysis of the processes, namely aluminium alloy die casting. After assessing the sustainability of production processes using a model-based approach, we move on to the real-time application of machine learning analyses with the goal of identifying downtime and failures during the production cycle and predicting their occurrence well in advance using real-time process thermodynamic parameter values and automatic learning. Finally, the thesis suggests the use of integrated models on various case studies, such as laser deposition processes and the renovation of existing buildings, to demonstrate the multidisciplinarity and transversality of these issues. The thesis reveals fascinating findings derived from the use of a hybrid method to assessing the sustainability of manufacturing processes, combining exergetic analysis with life cycle assessment. The proposed theme is completely current and relevant to the most recent developments in the field of industrial sustainability, combining traditional model-based approaches with innovative approaches based on the collection of big data and its analysis using the most appropriate machine learning methodologies. Furthermore, the thesis demonstrates a highly promising application of machine learning approaches to real-time data collected in order to identify any fault source in the manufacturing line beginning with sustainability measures generated from exergetic analysis and life cycle analysis. As such, it unquestionably represents an advancement above earlier information depicted in the initial state of the art. In actuality, manufacturing companies that implement business strategies based on smart models and key enabling technologies today have a higher market value in terms of quality, customisation, flexibility, and sustainability.

We present a novel approach to training data-driven dependency parsers on incomplete annotations. Our parsers are simple modifications of two well-known dependency parsers, the transition-based Malt parser and the graph-based MST parser. While previous work on parsing with incomplete data has typically couched the task in frameworks of unsupervised or semi-supervised machine learning, we essentially treat it as a supervised problem. In particular, we propose what we call agnostic parsers which hide all fragmentation in the training data from their supervised components. We present experimental results with training data that was obtained by means of annotation projection. Annotation projection is a resource-lean technique which allows us to transfer annotations from one language to another within a parallel corpus. However, the output tends to be noisy and incomplete due to cross-lingual non-parallelism and error-prone word alignments. This makes the projected annotations a suitable test bed for our fragment parsers. Our results show that (i) dependency parsers trained on large amounts of projected annotations achieve higher accuracy than the direct projections, and that (ii) our agnostic fragment parsers perform roughly on a par with the original parsers which are trained only on strictly filtered, complete trees. Finally, (iii) when our fragment parsers are trained on artificially fragmented but otherwise gold standard dependencies, the performance loss is moderate even with up to 50% of all edges removed.
Wir präsentieren eine neuartige Herangehensweise an das Trainieren von daten-gesteuerten Dependenzparsern auf unvollständigen Annotationen. Unsere Parser sind einfache Varianten von zwei bekannten Dependenzparsern, nämlich des transitions-basierten Malt-Parsers sowie des graph-basierten MST-Parsers. Während frühere Arbeiten zum Parsing mit unvollständigen Daten die Aufgabe meist in Frameworks für unüberwachtes oder schwach überwachtes maschinelles Lernen gebettet haben, behandeln wir sie im Wesentlichen mit überwachten Lernverfahren. Insbesondere schlagen wir "agnostische" Parser vor, die jegliche Fragmentierung der Trainingsdaten vor ihren daten-gesteuerten Lernkomponenten verbergen. Wir stellen Versuchsergebnisse mit Trainingsdaten vor, die mithilfe von Annotationsprojektion gewonnen wurden. Annotationsprojektion ist ein Verfahren, das es uns erlaubt, innerhalb eines Parallelkorpus Annotationen von einer Sprache auf eine andere zu übertragen. Bedingt durch begrenzten crosslingualen Parallelismus und fehleranfällige Wortalinierung ist die Ausgabe des Projektionsschrittes jedoch üblicherweise verrauscht und unvollständig. Gerade dies macht projizierte Annotationen zu einer angemessenen Testumgebung für unsere fragment-fähigen Parser. Unsere Ergebnisse belegen, dass (i) Dependenzparser, die auf großen Mengen von projizierten Annotationen trainiert wurden, größere Genauigkeit erzielen als die zugrundeliegenden direkten Projektionen, und dass (ii) die Genauigkeit unserer agnostischen, fragment-fähigen Parser der Genauigkeit der Originalparser (trainiert auf streng gefilterten, komplett projizierten Bäumen) annähernd gleichgestellt ist. Schließlich zeigen wir mit künstlich fragmentierten Gold-Standard-Daten, dass (iii) der Verlust an Genauigkeit selbst dann bescheiden bleibt, wenn bis zu 50% aller Kanten in den Trainingsdaten fehlen.

Numerical simulation, the use of computers to run a program which implements a mathematical model for a physical system, is an important part of today technological world. It is required in many scientific and engineering fields to study the behaviour of systems whose mathematical models are too complex to provide analytical solutions and it makes virtual evaluation of systems responses possible (virtual twins). This drastically reduces the number of experimental tests for accurate designs of the real system that the numerical model represents. However these virtual twins, based on classical methods which make use of a rich representations of the system (ex. finite element method), rarely allows real-time feedback, even when considering high performance computing, operating on powerful platforms. In these circumstances, the real-time performance required in some applications are compromised. Indeed the virtual twins are static, that is, they are used in the design of complex systems and their components, but they are not expected to accommodate or assimilate data so as to define dynamic data-driven application systems. Moreover significant deviations between the observed response and the one predicted by the model are usually noticed due to inaccuracy in the employed models, in the determination of the model parameters or in their time evolution. In this thesis we propose different methods to solve these handicaps in order to perform real-time monitoring and control. In the first part Model Order Reduction (MOR) techniques are used to accommodate real-time constraints; they compute a good approximation of the solution by simplifying the solution procedure instead of the model. The accuracy of the predicted solution is not compromised and efficient simulations can be performed (digital twins). In the second part Data-Driven modelling are employed to fill the gap between the parametric solution computed by using non-intrusive MOR techniques and the measured fields, in order to make dynamic data-driven application systems, DDDAS, possible (Hybrid Twins).
La simulación numérica, el uso de ordenadores para ejecutar un programa que implementa un modelo matemático de un sistema físico, es una parte importante del mundo tecnológico actual. En muchos campos de la ciencia y la ingeniería es necesario estudiar el comportamiento de sistemas cuyos modelos matemáticos son demasiado complejos para proporcionar soluciones analíticas, haciendo posible la evaluación virtual de las respuestas de los sistemas (gemelos virtuales). Esto reduce drásticamente el número de pruebas experimentales para los diseños precisos del sistema real que el modelo numérico representa. Sin embargo, estos gemelos virtuales, basados en métodos clásicos que hacen uso de una rica representación del sistema (por ejemplo, el método de elementos finitos), rara vez permiten la retroalimentación en tiempo real, incluso cuando se considera la computación en plataformas de alto rendimiento. En estas circunstancias, el rendimiento en tiempo real requerido en algunas aplicaciones se ve comprometido. En efecto, los gemelos virtuales son estáticos, es decir, se utilizan en el diseño de sistemas complejos y sus componentes, pero no se espera que acomoden o asimilen los datos para definir sistemas de aplicación dinámicos basados en datos. Además, se suelen apreciar desviaciones significativas entre la respuesta observada y la predicha por el modelo, debido a inexactitudes en los modelos empleados, en la determinación de los parámetros del modelo o en su evolución temporal. En esta tesis se proponen diferentes métodos para resolver estas limitaciones con el fin de realizar un seguimiento y un control en tiempo real. En la primera parte se utilizan técnicas de Reducción de Modelos para satisfacer las restricciones en tiempo real; estas técnicas calculan una buena aproximación de la solución simplificando el procedimiento de resolución en lugar del modelo. La precisión de la solución no se ve comprometida y se pueden realizar simulaciones efficientes (gemelos digitales). En la segunda parte se emplea la modelización basada en datos para llenar el vacío entre la solución paramétrica, calculada utilizando técnicas de reducción de modelos no intrusivas, y los campos medidos, con el fin de hacer posibles los sistemas de aplicación dinámicos basados en datos (gemelos híbridos).
La simulation numérique, c'est-à-dire l'utilisation des ordinateurs pour exécuter un programme qui met en oeuvre un modèle mathématique d'un système physique, est une partie importante du monde technologique actuel. Elle est nécessaire dans de nombreux domaines scientifiques et techniques pour étudier le comportement de systèmes dont les modèles mathématiques sont trop complexes pour fournir des solutions analytiques et elle rend possible l'évaluation virtuelle des réponses des systèmes (jumeaux virtuels). Cela réduit considérablement le nombre de tests expérimentaux nécessaires à la conception précise du système réel que le modèle numérique représente. Cependant, ces jumeaux virtuels, basés sur des méthodes classiques qui utilisent une représentation fine du système (ex. méthode des éléments finis), permettent rarement une rétroaction en temps réel, même dans un contexte de calcul haute performance, fonctionnant sur des plates-formes puissantes. Dans ces circonstances, les performances en temps réel requises dans certaines applications sont compromises. En effet, les jumeaux virtuels sont statiques, c'est-à-dire qu'ils sont utilisés dans la conception de systèmes complexes et de leurs composants, mais on ne s'attend pas à ce qu'ils prennent en compte ou assimilent des données afin de définir des systèmes d'application dynamiques pilotés par les données. De plus, des écarts significatifs entre la réponse observée et celle prévue par le modèle sont généralement constatés en raison de l'imprécision des modèles employés, de la détermination des paramètres du modèle ou de leur évolution dans le temps. Dans cette thèse, nous proposons di érentes méthodes pour résoudre ces handicaps afin d'effectuer une surveillance et un contrôle en temps réel. Dans la première partie, les techniques de Réduction de Modèles sont utilisées pour tenir compte des contraintes en temps réel ; elles calculent une bonne approximation de la solution en simplifiant la procédure de résolution plutôt que le modèle. La précision de la solution n'est pas compromise et des simulations e caces peuvent être réalisées (jumeaux numériquex). Dans la deuxième partie, la modélisation pilotée par les données est utilisée pour combler l'écart entre la solution paramétrique calculée, en utilisant des techniques de réduction de modèles non intrusives, et les champs mesurés, afin de rendre possibles des systèmes d'application dynamiques basés sur les données (jumeaux hybrides).

Cette thèse traite de la problématique de la détection de rupture dans le cadre séquentiel où le signal est supposé être observé en temps réel et le phénomène passe de son état de départ "normal" à un état post-changement "anormal". Le défi de la détection séquentielle est de minimiser le délai de détection, soumis à une limite tolérable de fausse alarme. L'idée est de tester séquentiellement l'existence d'une rupture par l'écriture récursive de la statistique de détection en fonction du score, qui remplace le Log-Likelihood Ratio lorsque la distribution des données est inconnue. La procédure de détection repose ainsi sur une statistique récursive, un seuil de détection et une règle d'arrêt. Dans un premier travail, nous considérons la statistique score-CUSUM et proposons d'évaluer la performance de détection de certains seuils de détection. Deux seuils sont issus de la littérature, et trois nouveaux seuils sont construits par une méthode basée sur la simulation: le premier est constant, le second instantané et le troisième est une version dynamique "data-driven" du précédent. Nous définissons rigoureusement chacun des seuils en mettant en évidence les différentes notions du risque de fausse alarme contrôlé suivant le seuil. Par ailleurs, nous proposons une nouvelle règle d'arrêt corrigée pour réduire le taux de fausse alarme. Nous effectuons ensuite une étude de simulation pour comparer les différents seuils et évaluer la règle d'arrêt corrigée. Nous constatons que le seuil empirique conditionnel est le meilleur pour minimiser le délai de détection tout en maintenant le risque toléré de fausse alarme. Cependant, sur des données réelles, nous recommandons d'utiliser le seuil data-driven car c'est le plus simple à construire et à utiliser pour une implémentation pratique. Dans la seconde partie, nous appliquons notre méthodologie de détection data-driven sur des signaux physiologiques, à savoir des signaux temporels enregistrés au niveau du faisceau supérieur du trapèze de 30 sujets effectuant différentes activités bureautiques. La méthodologie est sujet-activité dépendante; elle inclut l'estimation on-line des paramètres du signal et la construction du seuil data-driven sur le début du signal de chaque activité de chaque sujet. L'objectif était d'identifier des changements de régimes au cours d'une activité afin d'évaluer le niveau de sollicitation du muscle et la variabilité du signal EMG, qui sont liés à la fatigue musculaire. Les résultats obtenus ont confirmé l'aisance de notre méthodologie et la performance et praticité du seuil data-driven proposé. Par la suite, les résultats ont permis la caractérisation de chaque type d'activité en utilisant des modèles mixtes
This thesis deals the problem of change-point detection in the sequential framework where the signal is assumed to be observed in real time and the phenomenon changes from its "normal" starting state to an "abnormal" post-change state. The challenge of sequential detection is to minimize the detection delay, subject to a tolerable false alarm limit. The idea is to sequentially test for the existence of a change-point by recursively writing the detection statistic as a function of the score, which replaces the Log-Likelihood Ratio when the data distribution is unknown. The detection procedure is thus based on a recursive statistic, a detection threshold and a stopping rule. In a first work, we consider the score-CUSUM statistic and propose to evaluate the detection performance of some detection thresholds. Two thresholds come from the literature, and three new thresholds are constructed by a method based on simulation: the first is constant, the second instantaneous and the third is a dynamic "data-driven" version of the previous one. We rigorously define each of the thresholds by highlighting the different notions of the controlled false alarm risk according to the threshold. Moreover, we propose a new corrected stopping rule to reduce the false alarm rate. We then perform a simulation study to compare the different thresholds and evaluate the corrected stopping rule. We find that the conditional empirical threshold is the best to minimize the detection delay while maintaining the tolerated risk of false alarms. However, on real data, we recommend using the data-driven threshold as it is the easiest to build and use for practical implementation. In the second part, we apply our data-driven detection methodology to physiological signals, namely temporal signals recorded at the level of the upper trapezium beam of 30 subjects performing different office activities. The methodology is subject-activity dependent; it includes the on-line estimation of the signal parameters and the construction of the data-driven threshold on the start of the signal of each activity of each subject. The objective was to identify regime changes during an activity in order to assess the level of muscle solicitation and EMG signal variability, which are associated with muscle fatigue. The results obtained confirmed the ease of our methodology and the performance and practicality of the proposed data-driven threshold. Subsequently, the results allowed the characterization of each type of activity using mixed models

There has been extensive research done in the area of Semiparametric Regression. These techniques deliver substantial improvements over previously developed methods, such as Ordinary Least Squares and Kernel Regression. Two of these hybrid techniques: Model Robust Regression 1 (MRR1) and Model Robust Regression 2 (MRR2) require the choice of an appropriate bandwidth for smoothing and a mixing parameter that allows a portion of a nonparametric fit to be used in fitting a model that may be misspecifed by other regression methods. The current method of choosing the bandwidth and mixing parameter does not guarantee the optimal choices in either case. The immediate objective of the current work is to address this process of choosing the optimal bandwidth and mixing parameter and to examine the behavior of these estimates using 3D plots. The 3D plots allow us to examine how the semiparametric techniques: MRR1 and MRR2, behave for the optimal (AVEMSE) selection process when compared to data-driven selectors, such as PRESS* and PRESS**. It was found that the structure of MRR2 behaved consistently under all conditions. MRR2 displayed a wider range of "acceptable" values for the choice of bandwidth as opposed to a much more limited choice when using MRR1. These results provide general support for earlier fndings by Mays et al. (2000).

The University of ManchesterCarlos Alberto Duran VillalobosDoctor of Philosophy in the Faculty of Engineering and Physical SciencesDecember 2015This thesis presents an innovative batch-to-batch optimisation technique that was able to improve the productivity of two benchmark fed-batch fermentation simulators: Saccharomyces cerevisiae and Penicillin production. In developing the proposed technique, several important challenges needed to be addressed:For example, the technique relied on the use of a linear Multiway Partial Least Squares (MPLS) model to adapt from one operating region to another as productivity increased to estimate the end-point quality of each batch accurately. The proposed optimisation technique utilises a Quadratic Programming (QP) formulation to calculate the Manipulated Variable Trajectory (MVT) from one batch to the next. The main advantage of the proposed optimisation technique compared with other approaches that have been published was the increase of yield and the reduction of convergence speed to obtain an optimal MVT. Validity Constraints were also included into the batch-to-batch optimisation to restrict the QP calculations to the space only described by useful predictions of the MPLS model. The results from experiments over the two simulators showed that the validity constraints slowed the rate of convergence of the optimisation technique and in some cases resulted in a slight reduction in final yield. However, the introduction of the validity constraints did improve the consistency of the batch optimisation. Another important contribution of this thesis were a series of experiments that were implemented utilising a variety of smoothing techniques used in MPLS modelling combined with the proposed batch-to-batch optimisation technique. From the results of these experiments, it was clear that the MPLS model prediction accuracy did not significantly improve using these smoothing techniques. However, the batch-to-batch optimisation technique did show improvements when filtering was implemented.

A major factor that negatively impacts flight operations globally is adverse weather. To reduce the impact of adverse weather, avoidance procedures such as finding an alternative flight path can usually be carried out. However, such procedures usually introduce extra costs such as flight delay. Hence, there exists a need for alternative flight paths that efficiently avoid adverse weather regions while minimising costs. Existing weather avoidance methods used techniques, such as Dijkstra’s and artificial potential field algorithms that do not scale adequately and have poor real time performance. They do not adequately consider the impact of weather and its avoidance on passengers. The contributions of this work include a new development of an improved integrated model for weather avoidance, that addressed the impact of weather on passengers by defining a corresponding cost metric. The model simultaneously considered other costs such as flight delay and fuel burn costs. A genetic algorithm (GA)-based rerouting technique that generates optimised alternative flight paths was proposed. The technique used a modified mutation strategy to improve global search. A discrete firefly algorithm-based rerouting method was also developed to improve rerouting efficiency. A data framework and simulation platform that integrated aeronautical, weather and flight data into the avoidance process was developed. Results show that the developed algorithms and model produced flight paths that had lower total costs compared with existing techniques. The proposed algorithms had adequate rerouting performance in complex airspace scenarios. The developed system also adequately avoided the paths of multiple aircraft in the considered airspace.

The data generated in a chemical industry is a reflection of the process. With the modern computer control systems and data logging facilities, there is an increasing ability to collect large amounts of data. As there are many underlying aspects of the process in that data, with its proper utilization, it is possible to obtain useful information for process monitoring and fault diagnosis in addition to many other decision making activities. The purpose of this research is to utilize the data driven multivariate techniques of Principal Component Analysis (PCA) and Independent Component Analysis (ICA) for the estimation of process parameters. This research also includes analysis and comparison of these techniques for fault detection and diagnosis along with introduction, explanation and results from a new methodology developed in this research work namely Hybrid Independent Component Analysis (HICA).The first part of this research is the utilization of models of PCA and ICA for estimation of process parameters. The individual techniques of PCA and ICA are applied separately to the original data set of a waste water treatment plant (WWTP) and the process parameters for the unknown conditions of the process are calculated. For each of the techniques (PCA and ICA), the validation of the calculated parameters is carried out by construction of Decision Trees on WWTP dataset using inductive data mining and See 5.0. Both individual techniques were able to estimate all parameters successfully. The minor limitation in the validation of all results may be due to the strict application of these techniques to Gaussian and non-Gaussian data sets respectively. Using statistical analysis it was shown that the data set used in this work exhibits Gaussian and non-Gaussian behaviour.In the second part of this work multivariate techniques of Principal Component Analysis (PCA) and Independent Component Analysis (ICA) have been used for fault detection and diagnosis of a process along with introduction of the new technique, Hybrid Independent Component Analysis (HICA). The techniques are applied to two case studies, the waste water treatment plant (WWTP) and an Air pollution data set. As reported in literature, PCA and ICA proved to be useful tools for process monitoring on both data set, but a comparison of PCA and ICA along with the newly developed technique (HICA) illustrated the superiority of HICA over PCA and ICA. It is evident from the fact that PCA detected 74% and 67% of the faults in the WWTP data and Air pollution data set respectively. ICA successfully detected 61.3% and 62% of the faults from these datasets. Finally HICA showed improved results by the detection of 90% and 81% of the faults in both case studies. This showed that the new developed algorithm is more effective than the individual techniques of PCA and ICA. For fault diagnosis using PCA, ICA and HICA, contribution plots are constructed leading to the identification of responsible variable/s for a particular fault. This part also includes the work done for the estimation of process parameters using HICA technique as was done with PCA and ICA in the first part of the research. As expected HICA technique was more successful in estimation of parameters than PCA and ICA in line with its working for process monitoring.

Ce travail présente une méthodologie de test hybride complète pour évaluer l'écoulement dans turbomachines. Axée sur la minimisation des temps de test et des exigences en instrumentation, la méthodologie intègre de manière stratégique des mesures expérimentales standard avec des simulations numériques, en utilisant des processus gaussiens.La méthodologie réduit systématiquement à la fois les efforts d'instrumentation et les temps de test, fournissant des métriques d'incertitude comparables aux méthodologies traditionnelles. Appliquée initialement à un compresseur axial haute pression de référence (H25) puis à un ventilateur à ultra-haut taux de dilution (ECL5 UHBR) dans des conditions de test aveugles, la méthodologie démontre sa robustesse, son adaptabilité et des réductions significatives des points de mesure et des temps de test conduisant à un impact direct sur les coûts des campagnes expérimentales.Pour le compresseur axial H25, le cadre proposé se révèle capable de prédire les champs d'écoulement, mettant en évidence le compromis entre les mesures et l'exactitude de prédiction du flux. Les résultats du test aveugle du ventilateur ECL5 UHBR valident l'efficacité de la méthodologie dans les évaluations aérodynamiques et démontrent des économies de temps d'au moins une heure par condition de fonctionnement.La conception d'expériences a priori permet une réduction d'au moins 50% des mesures, surpassant l'échantillonnage aléatoire, et assiste efficacement dans la planification de campagnes expérimentales. L'échantillonnage adaptatif In situ surpasse l'échantillonnage aléatoire jusqu'à 44%, démontrant une détection précise des phénomènes d'écoulement et des applications prometteuses dans la réalisation d'exigences expérimentales. La nature modulaire et adaptable de la méthodologie la positionne pour une application étendue tant dans les environnements académiques qu'industriels, tandis que son exploitation ouvre des voies pour inférer des quantités d'écoulement non mesurées ou améliorer l'évaluation des performances.Ce travail introduit un changement de paradigme dans la planification de campagnes expérimentales, optimisant les budgets de mesure de manière stratégique à l'avance ou améliorant la précision dynamiquement au cours d'une campagne, mettant en évidence le potentiel des tendances entraînées par l'apprentissage automatique pour façonner de nouvelles voies de recherche
This work presents a complete hybrid testing methodology for assessing the flow in turbomachinery components. Focused on minimizing testing times and instrumentation requirements, the methodology strategically integrates standard experimental measurements with numerical simulations, specifically employing Multi-Fidelity Gaussian Processes, Sparse Variational Gaussian Processes, and adaptive Bayesian optimization.The methodology systematically reduces both instrumentation efforts and testing times, providing uncertainty metrics comparable to traditional methodologies. Applied initially to a benchmarked axial high-pressure compressor (H25) and afterwards to an ultra-high bypass ratio fan (ECL5 UHBR) in blind test conditions, the methodology demonstrates robustness, adaptability, and significant reductions in measurement points and testing times leading to a direct impact in experimental campaign costs.For the H25 axial compressor, the proposed framework proves capable of predicting flow fields, emphasizing the trade-off between high-fidelity measurements and mean flow prediction accuracy. The ECL5 UHBR fan blind test results validate the methodology's efficiency in aerodynamic assessments and demonstrates time savings of at least one hour per operating condition.The a priori Design of Experiments achieves at least a 50% reduction in measurements, outperforming random sampling, and effectively assists in experimental campaign planning. The In situ adaptive sampling outperforms random sampling by up to 44%, showcasing accurate detection of flow phenomena and promising applications in achieving high accuracy experimental demands. The modular and adaptable nature of the methodology positions it for broad application in both academic and industrial settings, while its exploitation opens paths to infer unmeasured flow quantities or improve performance evaluation measurements.This work introduces a paradigm shift in experimental campaign planning, optimizing measurement budgets strategically beforehand or enhancing accuracy dynamically during a campaign, emphasizing the potential of machine learning-driven trends in shaping new research paths

Face aux défis actuels liés à la conception de capteurs d'images à forte résolution comme la limitation de la consommation électrique, l'augmentation du flux de données ainsi que le traitement de données associé, on propose, à travers cette thèse, un capteur d'image novateur asynchrone à échantillonnage non uniforme.Ce capteur d’images asynchrone est basé sur une matrice de pixels événementiels qui intègrent un échantillonnage non uniforme par traversée de niveaux. Contrairement aux imageurs conventionnels, où les pixels sont lus systématiquement lors de chaque trame, les pixels événementiels proposés sont consultés que lorsqu’ils contiennent une information pertinente. Cela induit un flux de données réduit et dépendant de l’image.Pour compléter la chaîne de traitement des pixels, on présente également une architecture numérique de lecture dédiée conçue en utilisant de la logique asynchrone et destinée à contrôler et à gérer le flux de données des pixels événementiels. Ce circuit de lecture numérique permet de surmonter les difficultés classiques rencontrées lors de la gestion des demandes simultanées des pixels événementiels sans dégrader la résolution et le facteur de remplissage du capteur d’images. En outre, le circuit de lecture proposé permet de réduire considérablement les redondances spatiales dans une image ce qui diminue encore le flux de données.Enfin, en combinant l'aspect échantillonnage par traversée de niveau et la technique de lecture proposée, on a pu remplacer la conversion analogique numérique classique de la chaîne de traitement des pixels par une conversion temps-numérique (Time-to-Digital Conversion). En d'autres termes, l'information du pixel est codée par le temps. Il en résulte une diminution accrue de la consommation électrique du système de vision, le convertisseur analogique-numérique étant un des composants les plus consommant du système de lecture des capteurs d'images conventionnels
In order to overcome the challenges associated with the design of high resolution image sensors, we propose, through this thesis, an innovative asynchronous event-driven image sensor based on non-uniform sampling. The proposed image sensor aims the reduction of the data flow and its associated data processing by limiting the activity of our image sensor to the new captured information.The proposed asynchronous image sensor is based on an event-driven pixels that incorporate a non-uniform sampling crossing levels. Unlike conventional imagers, where the pixels are read systematically at each frame, the proposed event-driven pixels are only read when they hold new and relevant information. This induces a reduced and scene dependent data flow.In this thesis, we introduce a complete pixel reading sequence. Beside the event-driven pixel, the proposed reading system is designed using asynchronous logic and adapted to control and manage the flow of data from event pixels. This digital reading system overcomes the traditional difficulties encountered in the management of simultaneous requests for event pixels without degrading the resolution and fill factor of the image sensor. In addition, the proposed reading circuit significantly reduces the spatial redundancy in an image which further reduces the data flow.Finally, by combining the aspect of level crossing sampling and the proposed reading technique, we replaced the conventional analog to digital conversion of the pixel processing chain by a time-to-digital Conversion (TDC). In other words, the pixel information is coded by time. This results in an increased reduction in power consumption of the vision system, the analog-digital converter being one of the most consuming reading system of conventional image sensors components

Recommendation systems have been widely utilized in e-commerce settings to aid users through their shopping experiences. The principal advantage of these systems is their ability to narrow down the purchase options in addition to marketing items to customers. However, a number of challenges remain, notably those related to obtaining a clearer understanding of users, their profiles, and their preferences in terms of purchased items. Specifically, recommender systems based on collaborative filtering recommend items that have been rated by other users with preferences similar to those of the targeted users. Intuitively, the more information and ratings collected about the user, the more accurate are the recommendations such systems suggest. In a typical recommender systems database, the data are sparse. Sparsity occurs when the number of ratings obtained by the users is much lower than the number required to build a prediction model. This usually occurs because of the users’ reluctance to share their reviews, either due to privacy issues or an unwillingness to make the extra effort. Grey-sheep users pose another challenge. These are users who shared their reviews and ratings yet disagree with the majority in the systems. The current state-of-the-art typically treats these users as outliers and removes them from the system. Our goal is to determine whether keeping these users in the system may benefit learning. Thirdly, cold-start problems refer to the scenario whereby a new item or user enters the system and is another area of active research. In this case, the system will have no information about the new user or item, making it problematic to find a correlation with others in the system. This thesis addresses the three above-mentioned research challenges through the development of machine learning methods for use within the recommendation system setting. First, we focus on the label and data sparsity though the development of the Hybrid Cluster analysis and Classification learning (HCC-Learn) framework, combining supervised and unsupervised learning methods. We show that combining classification algorithms such as k-nearest neighbors and ensembles based on feature subspaces with cluster analysis algorithms such as expectation maximization, hierarchical clustering, canopy, k-means, and cascade k-means methods, generally produces high-quality results when applied to benchmark datasets. That is, cluster analysis clearly benefits the learning process, leading to high predictive accuracies for existing users. Second, to address the cold-start problem, we present the Popular Users Personalized Predictions (PUPP-DA) framework. This framework combines cluster analysis and active learning, or so-called user-in-the-loop, to assign new customers to the most appropriate groups in our framework. Based on our findings from the HCC-Learn framework, we employ the expectation maximization soft clustering technique to create our user segmentations in the PUPP-DA framework, and we further incorporate Convolutional Neural Networks into our design. Our results show the benefits of user segmentation based on soft clustering and the use of active learning to improve predictions for new users. Furthermore, our findings show that focusing on frequent or popular users clearly improves classification accuracy. In addition, we demonstrate that deep learning outperforms machine learning techniques, notably resulting in more accurate predictions for individual users. Thirdly, we address the grey-sheep problem in our Grey-sheep One-class Recommendations (GSOR) framework. The existence of grey-sheep users in the system results in a class imbalance whereby the majority of users will belong to one class and a small portion (grey-sheep users) will fall into the minority class. In this framework, we use one-class classification to provide a class structure for the training examples. As a pre-assessment stage, we assess the characteristics of grey-sheep users and study their impact on model accuracy. Next, as mentioned above, we utilize one-class learning, whereby we focus on the majority class to first learn the decision boundary in order to generate prediction lists for the grey-sheep (minority class). Our results indicate that including grey-sheep users in the training step, as opposed to treating them as outliers and removing them prior to learning, has a positive impact on the general predictive accuracy.

Traffic congestion is a one hundred billion dollar problem in the US. The cost of congestion has been trending upward over the last few decades, but has experienced slight decreases in recent years partly due to the impact of congestion reduction strategies. The impact of these strategies is however largely experienced on freeways and not arterials. This discrepancy in impact is partially linked to the lack of real-time, arterial traffic information. Toward this end, this research effort seeks to address the lack of arterial traffic information. To address this dearth of information, this effort developed a methodology to provide accurate estimates of arterial performance measures to transportation facility managers and travelers in real-time. This methodology employs transmitted point sensor data to drive an online, microscopic traffic simulation model. The feasibility of this methodology was examined through a series of experiments that were built upon the successes of the previous, while addressing the necessary limitations. The results from each experiment were encouraging. They successfully demonstrated the method's likely feasibility, and the accuracy with which field estimates of performance measures may be obtained. In addition, the method's results support the viability of a "real-world" implementation of the method. An advanced calibration process was also developed as a means of improving the method's accuracy. This process will in turn serve to inform future calibration efforts as the need for more robust and accurate traffic simulation models are needed. The success of this method provides a template for real-time traffic simulation modeling which is capable of adequately addressing the lack of available arterial traffic information. In providing such information, it is hoped that transportation facility managers and travelers will make more informed decisions regarding more efficient management and usage of the nation's transportation network.

Les approches de Pronostic et gestion de la santé des systèmes (Prognostics and Health Management : PHM) guidées par les données sont fortement dépendantes de la disponibilité et de la qualité d’historiques de défaillances, une exigence souvent difficile à satisfaire dans les systèmes de surveillance en conditions industrielles. Cette divergence crée un écart significatif entre les méthodologies de PHM et leur application pratique sur des systèmes réels. L’accent prédominant mis sur les ensembles de données unimodales dans les travaux de recherche en PHM basée sur les données met en lumière le potentiel des données multimodales pour combler cet écart.Cette thèse explore l’intégration des données multimodales afin d’améliorer les méthodes et les algorithmes de PHM appliqués aux machines industrielles. Elle aborde de manière systématique des défis cruciaux tels que l’absence de données, les données bruitées, les données clairsemées et irrégulières, le déséquilibre des classes et la rareté des données de fonctionnement jusqu’à la défaillance. Elle propose des méthodologies innovantes qui intègrent plusieurs modalités de données et exploitent l’expertise spécifique au domaine pour créer des modèles prédictifs robustes.Les contributions principales de la thèse se déclinent comme suit :1. Apprentissage basé sur l’attention intermodale : une nouvelle méthode d’apprentissage multimodal conçue pour atténuer les limites posées par les données manquantes et bruitées. Elle permet d’intégrer des informations provenant de multiples modalités, améliorant ainsi la précision et la robustesse des modèles prédictifs.2. Méthodologie de diagnostic multimodal assisté par les connaissances d’experts : cette méthodologie combine l’expertise du domaine avec l’apprentissage multimodal pour permettre des diagnostics complets, améliorant ainsi la détection et la classification des défauts dans les machines industrielles.3. Approche de pronostic basée sur des graphes : cette approche innovante construit des trajectoires de fonctionnement jusqu’à la défaillance à partir de données incomplètes en utilisant des techniques basées sur les graphes, offrant une avancée significative dans le domaine du pronostic de défaillances.Ces méthodologies ont été rigoureusement validées sur des données de simulation ainsi que sur des données industrielles provenant d’hydro-générateurs, démontrant des améliorations significatives des algorithmes de PHM et de maintenance prédictive. Les résultats soulignent le potentiel des données multimodales pour améliorer considérablement la fiabilité et l’efficacité des modèles de PHM.Cette thèse apporte un cadre complet pour exploiter diverses sources de données et l’expertise du domaine, promettant de transformer les stratégies de maintenance et de réduire les coûts opérationnels dans diverses industries. Les résultats ouvrent la voie à des recherches futures et à des applications pratiques, positionnant l’intégration des données multimodales comme une avancée essentielle dans le domaine du PHM
Prognostics and Health Management (PHM) with data-driven techniques is heavily dependent upon the availability of extensive and high-quality datasets, a requirement often challenging to fulfill in industrial condition monitoring environments. This discrepancy creates a significant gap between state-of-the-art PHM methodologies and their practical application in real-world scenarios. The prevailing focus in data-driven PHM research on unimodal datasets highlights the potential of multimodal data to bridge this gap.This thesis explores the integration of multimodal data to advance PHM models for industrial machines. It systematically addresses pivotal challenges such as data missingness and noise, sparse and irregular datasets, class imbalance, and the scarcity of run-to-failure data. The research develops innovative methodologies that incorporate multiple data modalities and harness domain-specific expertise to create robust predictive models.The primary contributions of this research include:1. Cross-modal attention-based learning: A new multimodal learning method is designed to mitigate the limitations posed by missing and noisy data. It allows integrating information across multiple modalities, thereby enhancing the accuracy and robustness of predictive models.2. Expert-knowledge-assisted multimodal diagnostics methodology: This methodology combines domain expertise with multimodal learning to enable comprehensive diagnostics, thereby improving fault detection and classification in industrial machinery.3. Graph-based prognostics approach: This innovative approach constructs run-to-failure trajectories from incomplete data using graph-based techniques, offering a significant advancement in failure prognostics.These methodologies were rigorously validated using both simulation and industrial dataset of hydrogenerators, demonstrating significant improvements in PHM and predictive maintenance capabilities. The results underscore the potential of multimodal data to significantly enhance the reliability and efficiency of PHM methods and algorithms.This thesis proposes a comprehensive framework for leveraging diverse data sources and domain expertise, promising to transform maintenance strategies and reducing operational costs across various industries. The findings pave the way for future research and practical implementations, positioning multimodal data integration as a pivotal advancement in the field of PHM

The Bayesian network (BN) formalism is a powerful representation for encoding domains characterised by uncertainty. However, before it can be used it must first be constructed, which is a major challenge for any real-life problem. There are two broad approaches, namely the hand-crafted approach, which relies on a human expert, and the data-driven approach, which relies on data. The former approach is useful, however issues such as human bias can introduce errors into the model. We have conducted a literature review of the expert-driven approach, and we have cherry-picked a number of common methods, and engineered a framework to assist non-BN experts with expert-driven construction of BNs. The latter construction approach uses algorithms to construct the model from a data set. However, construction from data is provably NP-hard. To solve this problem, approximate, heuristic algorithms have been proposed; in particular, algorithms that assume an order between the nodes, therefore reducing the search space. However, traditionally, this approach relies on an expert providing the order among the variables --- an expert may not always be available, or may be unable to provide the order. Nevertheless, if a good order is available, these order-based algorithms have demonstrated good performance. More recent approaches attempt to ''learn'' a good order then use the order-based algorithm to discover the structure. To eliminate the need for order information during construction, we propose a search in the entire space of Bayesian network structures --- we present a novel approach for carrying out this task, and we demonstrate its performance against existing algorithms that search in the entire space and the space of orders. Finally, we employ the hand-crafting framework to construct models for the task of diagnosis in a ''real-life'' medical domain, dementia diagnosis. We collect real dementia data from clinical practice, and we apply the data-driven algorithms developed to assess the concordance between the reference models developed by hand and the models derived from real clinical data.

[ES] El principal desafío en los motores turbina de gas empleados en aviación reside en aumentar la eficiencia del ciclo termodinámico manteniendo las emisiones contaminantes por debajo de las rigurosas restricciones. Ésto ha conllevado la necesidad de diseñar nuevas estrategias de inyección/combustión que operan en puntos de operación peligrosos por su cercanía al límite inferior de apagado de llama. En este contexto, el concepto Lean Direct Injection (LDI) ha emergido como una tecnología prometedora a la hora de reducir los óxidos de nitrógeno (NOx) emitidos por las plantas propulsoras de los aviones de nueva generación. En este contexto, la presente tesis tiene como objetivos contribuir al conocimiento de los mecanismos físicos que rigen el comportamiento de un quemador LDI y proporcionar herramientas de análisis para una profunda caracterización de las complejas estructuras de flujo de turbulento generadas en el interior de la cámara de combustión. Para ello, se ha desarrollado una metodología numérica basada en CFD capaz de modelar el flujo bifásico no reactivo en el interior de un quemador LDI académico mediante enfoques de turbulencia U-RANS y LES en un marco Euleriano-Lagrangiano. La resolución numérica de este problema multi-escala se aborda mediante la descripción completa del flujo a lo largo de todos los elementos que constituyen la maqueta experimental, incluyendo su paso por el swirler y entrada a la cámara de combustión. Ésto se lleva a cabo través de dos códigos CFD que involucran dos estrategias de mallado diferentes: una basada en algoritmos de generación y refinamiento automático de la malla (AMR) a través de CONVERGE y otra técnica de mallado estático más tradicional mediante OpenFOAM. Por un lado, se ha definido una metodología para obtener una estrategia de mallado óptima mediante el uso del AMR y se han explotado sus beneficios frente a los enfoques tradicionales de malla estática. De esta forma, se ha demostrado que la aplicabilidad de las herramientas de control de malla disponibles en CONVERGE como el refinamiento fijo (fixed embedding) y el AMR son una opción muy interesante para afrontar este tipo de problemas multi-escala. Los resultados destacan una optimización del uso de los recursos computacionales y una mayor precisión en las simulaciones realizadas con la metodología presentada. Por otro lado, el uso de herramientas CFD se ha combinado con la aplicación de técnicas de descomposición modal avanzadas (Proper Orthogonal Decomposition and Dynamic Mode Decomposition). La identificación numérica de los principales modos acústicos en la cámara de combustión ha demostrado el potencial de estas herramientas al permitir caracterizar las estructuras de flujo coherentes generadas como consecuencia de la rotura de los vórtices (VBB) y de los chorros fuertemente torbellinados presentes en el quemador LDI. Además, la implementación de estos procedimientos matemáticos ha permitido tanto recuperar información sobre las características de la dinámica de flujo como proporcionar un enfoque sistemático para identificar los principales mecanismos que sustentan las inestabilidades en la cámara de combustión. Finalmente, la metodología validada ha sido explotada a través de un Diseño de Experimentos (DoE) para cuantificar la influencia de los factores críticos de diseño en el flujo no reactivo. De esta manera, se ha evaluado la contribución individual de algunos parámetros funcionales (el número de palas del swirler, el ángulo de dichas palas, el ancho de la cámara de combustión y la posición axial del orificio del inyector) en los patrones del campo fluido, la distribución del tamaño de gotas del combustible líquido y la aparición de inestabilidades en la cámara de combustión a través de una matriz ortogonal L9 de Taguchi. Este estudio estadístico supone un punto de partida para posteriores estudios de inyección, atomización y combus
[CA] El principal desafiament als motors turbina de gas utilitzats a la aviació resideix en augmentar l'eficiència del cicle termodinàmic mantenint les emissions contaminants per davall de les rigoroses restriccions. Aquest fet comporta la necessitat de dissenyar noves estratègies d'injecció/combustió que radiquen en punts d'operació perillosos per la seva aproximació al límit inferior d'apagat de flama. En aquest context, el concepte Lean Direct Injection (LDI) sorgeix com a eina innovadora a l'hora de reduir els òxids de nitrogen (NOx) emesos per les plantes propulsores dels avions de nova generació. Sota aquest context, aquesta tesis té com a objectius contribuir al coneixement dels mecanismes físics que regeixen el comportament d'un cremador LDI i proporcionar ferramentes d'anàlisi per a una profunda caracterització de les complexes estructures de flux turbulent generades a l'interior de la càmera de combustió. Per tal de dur-ho a terme s'ha desenvolupat una metodología numèrica basada en CFD capaç de modelar el flux bifàsic no reactiu a l'interior d'un cremador LDI acadèmic mitjançant els enfocaments de turbulència U-RANS i LES en un marc Eulerià-Lagrangià. La resolució numèrica d'aquest problema multiescala s'aborda mitjançant la resolució completa del flux al llarg de tots els elements que constitueixen la maqueta experimental, incloent el seu pas pel swirler i l'entrada a la càmera de combustió. Açò es duu a terme a través de dos codis CFD que involucren estratègies de mallat diferents: una basada en la generación automàtica de la malla i en l'algoritme de refinament adaptatiu (AMR) amb CONVERGE i l'altra que es basa en una tècnica de mallat estàtic més tradicional amb OpenFOAM. D'una banda, s'ha definit una metodologia per tal d'obtindre una estrategia de mallat òptima mitjançant l'ús de l'AMR i s'han explotat els seus beneficis front als enfocaments tradicionals de malla estàtica. D'aquesta forma, s'ha demostrat que l'aplicabilitat de les ferramente de control de malla disponibles en CONVERGE com el refinament fixe (fixed embedding) i l'AMR són una opció molt interessant per tal d'afrontar aquest tipus de problemes multiescala. Els resultats destaquen una optimització de l'ús dels recursos computacionals i una major precisió en les simulacions realitzades amb la metodologia presentada. D'altra banda, l'ús d'eines CFD s'ha combinat amb l'aplicació de tècniques de descomposició modal avançades (Proper Orthogonal Decomposition and Dynamic Mode Decomposition). La identificació numèrica dels principals modes acústics a la càmera de combustió ha demostrat el potencial d'aquestes ferramentes al permetre caracteritzar les estructures de flux coherents generades com a conseqüència del trencament dels vòrtex (VBB) i dels raigs fortament arremolinats presents al cremador LDI. A més, la implantació d'estos procediments matemàtics ha permès recuperar informació sobre les característiques de la dinàmica del flux i proporcionar un enfocament sistemàtic per tal d'identificar els principals mecanismes que sustenten les inestabilitats a la càmera de combustió. Finalment, la metodologia validada ha sigut explotada a traves d'un Diseny d'Experiments (DoE) per tal de quantificar la influència dels factors crítics de disseny en el flux no reactiu. D'aquesta manera, s'ha avaluat la contribución individual d'alguns paràmetres funcionals (el nombre de pales del swirler, l'angle de les pales, l'amplada de la càmera de combustió i la posició axial de l'orifici de l'injector) en els patrons del camp fluid, la distribució de la mida de gotes del combustible líquid i l'aparició d'inestabilitats en la càmera de combustió mitjançant una matriu ortogonal L9 de Taguchi. Aquest estudi estadístic és un bon punt de partida per a futurs estudis de injecció, atomització i combustió en cremadors LDI.
[EN] Aeronautical gas turbine engines present the main challenge of increasing the efficiency of the cycle while keeping the pollutant emissions below stringent restrictions. This has led to the design of new injection-combustion strategies working on more risky and problematic operating points such as those close to the lean extinction limit. In this context, the Lean Direct Injection (LDI) concept has emerged as a promising technology to reduce oxides of nitrogen (NOx) for next-generation aircraft power plants In this context, this thesis aims at contributing to the knowledge of the governing physical mechanisms within an LDI burner and to provide analysis tools for a deep characterisation of such complex flows. In order to do so, a numerical CFD methodology capable of reliably modelling the 2-phase nonreacting flow in an academic LDI burner has been developed in an Eulerian-Lagrangian framework, using the U-RANS and LES turbulence approaches. The LDI combustor taken as a reference to carry out the investigation is the laboratory-scale swirled-stabilised CORIA Spray Burner. The multi-scale problem is addressed by solving the complete inlet flow path through the swirl vanes and the combustor through two different CFD codes involving two different meshing strategies: an automatic mesh generation with adaptive mesh refinement (AMR) algorithm through CONVERGE and a more traditional static meshing technique in OpenFOAM. On the one hand, a methodology to obtain an optimal mesh strategy using AMR has been defined, and its benefits against traditional fixed mesh approaches have been exploited. In this way, the applicability of grid control tools available in CONVERGE such as fixed embedding and AMR has been demonstrated to be an interesting option to face this type of multi-scale problem. The results highlight an optimisation of the use of the computational resources and better accuracy in the simulations carried out with the presented methodology. On the other hand, the use of CFD tools has been combined with the application of systematic advanced modal decomposition techniques (i.e., Proper Orthogonal Decomposition and Dynamic Mode Decomposition). The numerical identification of the main acoustic modes in the chamber have proved their potential when studying the characteristics of the most powerful coherent flow structures of strongly swirled jets in a LDI burner undergoing vortex breakdown (VBB). Besides, the implementation of these mathematical procedures has allowed both retrieving information about the flow dynamics features and providing a systematic approach to identify the main mechanisms that sustain instabilities in the combustor. Last, this analysis has also allowed identifying some key features of swirl spray systems such as the complex pulsating, intermittent and cyclical spatial patterns related to the Precessing Vortex Core (PVC). Finally, the validated methodology is exploited through a Design of Experiments (DoE) to quantify the influence of critical design factors on the non-reacting flow. In this way, the individual contribution of some functional parameters (namely the number of swirler vanes, the swirler vane angle, the combustion chamber width and the axial position of the nozzle tip) into both the flow field pattern, the spray size distribution and the occurrence of instabilities in the combustion chamber are evaluated throughout a Taguchi's orthogonal array L9. Such a statistical study has supposed a good starting point for subsequent studies of injection, atomisation and combustion on LDI burners.
Belmar Gil, M. (2020). Computational study on the non-reacting flow in Lean Direct Injection gas turbine combustors through Eulerian-Lagrangian Large-Eddy Simulations [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/159882
TESIS

S tím, jak se neustále vyvíjejí nové technologie pro energeticky náročná průmyslová odvětví, stávající zařízení postupně zaostávají v efektivitě a produktivitě. Tvrdá konkurence na trhu a legislativa v oblasti životního prostředí nutí tato tradiční zařízení k ukončení provozu a k odstavení. Zlepšování procesu a projekty modernizace jsou zásadní v udržování provozních výkonů těchto zařízení. Současné přístupy pro zlepšování procesů jsou hlavně: integrace procesů, optimalizace procesů a intenzifikace procesů. Obecně se v těchto oblastech využívá matematické optimalizace, zkušeností řešitele a provozní heuristiky. Tyto přístupy slouží jako základ pro zlepšování procesů. Avšak, jejich výkon lze dále zlepšit pomocí moderní výpočtové inteligence. Účelem této práce je tudíž aplikace pokročilých technik umělé inteligence a strojového učení za účelem zlepšování procesů v energeticky náročných průmyslových procesech. V této práci je využit přístup, který řeší tento problém simulací průmyslových systémů a přispívá následujícím: (i)Aplikace techniky strojového učení, která zahrnuje jednorázové učení a neuro-evoluci pro modelování a optimalizaci jednotlivých jednotek na základě dat. (ii) Aplikace redukce dimenze (např. Analýza hlavních komponent, autoendkodér) pro vícekriteriální optimalizaci procesu s více jednotkami. (iii) Návrh nového nástroje pro analýzu problematických částí systému za účelem jejich odstranění (bottleneck tree analysis – BOTA). Bylo také navrženo rozšíření nástroje, které umožňuje řešit vícerozměrné problémy pomocí přístupu založeného na datech. (iv) Prokázání účinnosti simulací Monte-Carlo, neuronové sítě a rozhodovacích stromů pro rozhodování při integraci nové technologie procesu do stávajících procesů. (v) Porovnání techniky HTM (Hierarchical Temporal Memory) a duální optimalizace s několika prediktivními nástroji pro podporu managementu provozu v reálném čase. (vi) Implementace umělé neuronové sítě v rámci rozhraní pro konvenční procesní graf (P-graf). (vii) Zdůraznění budoucnosti umělé inteligence a procesního inženýrství v biosystémech prostřednictvím komerčně založeného paradigmatu multi-omics.

Air pollution has become a significant issue, especially in high-density urban areas. Since the first industrial revolution in the 19th century, human activities have badly affected the earth and the environment. The world has become inhabitable day by day due to inappropriate human activities, construction patterns, and unsustainable cities development. Moreover, due to these activities, environmental air quality has become worst day by day. World Health Organization (WHO) provided evidence that urban air pollution is becoming the main threat to human health, especially in high-density countries like China and India. Changes in air pollution are basically affected by low-frequency and high-frequency pollution in other places and their weather conditions. In that situation, changes in meteorological parameters and spatial attributes should be considered and identify the correlation between them. People are paying more attention to changes in air quality and its control. Air pollution forecasting is one of the essential preventive steps to control air pollution. Based on the pollutant forecasting results, the relevant department and policymakers get the early information on pollutant concentration in a particular location. This information can help them adjust some critical measures according to pollutant emission sources to control pollutant emissions and their adverse impact on public health. However, pollution forecasting has become challenging due to its complexities with time-space nonlinearities, weather conditions, and spatial-temporal impact of nearby locations. In order to address these issues, this research work proposes various approaches, which are summarized as follows. In the first contribution, this research work presents a neural network based Convolutional Long Short term Memory-Sparse Denoising Autoencoder (CLS) model to forecast the PM2.5 level under meteorological conditions. The CLS model identifies the vast dataset’s hidden features, performs pollutants’ temporal modeling, and reconstructs the predicted output in the dynamic fine-tuning layer to get robust prediction results in a real time environment. The proposed model has experimented with different datasets, and its results show the model’s efficiency in air quality modeling. In the second contribution, this research work developed Temporal Convolutional Denoising Autoencoder (TCDA) network, a hybrid PM2.5 prediction framework that can perform rapid extraction of complex dataset’s features, handle missing values and improve PM2.5 prediction results. The model can reconstruct the corrupted, missing values and handle the different patterns of missing values to enhance the short-term PM2.5 forecasting results. In the third contribution, this research developed a Multi-Directional Temporal Convolutional Artificial Neural Network (MTCAN) model to impute and forecast PM2.5 pollutant concentration in a single training process. The main idea of the multi-directional properties of MTCAN is to maintain the temporal correlation within the features’ measurement and meteorological and pollutant variables to impute PM2.5 missing values. The MTCAN model performs feature learning and sequential modeling simultaneously with a wide range of past observations for long-term forecasting, minimizing memory size requirement and training cost. In the fourth contribution, this research work presents a newly developed multi-step ahead pollution forecasting model with a multi-input, multi-output learning process. The proposed model can work effectively under meteorological conditions and spatial impact. The proposed model has better long-term forecasting accuracy as compared to the traditional statistical and machine learning models. The proposed Multi-Output Long Short-Term Memory (LSTM) Autoencoder (M-LSTMA) accumulates each step prediction value to perform multi-step ahead forecasting for multiple pollutants in a single training process. The results show the model’s effectiveness, where we need to know the overall air pollution level for a particular area. In the fifth contribution, this research work proposed a novel PM2.5 forecasting model named as Multi-Output Temporal Convolutional Network Autoencoder (MO TCNA), which serves both the PM2.5 and PM10 pollutants forecasting for various locations instead of performing single output and site-specific pollutant forecasting for an overall idea of pollution level for a particular region. The proposed work developed a Recursive-Multi-Input, Multi Output (R-MIMO) strategy to improve multiple pollutant forecasting accuracy for the long-term period for different sites in a region. Experimental results indicate that the proposed models are superior to baseline single-output and multi-output forecasting models, which proves their effectiveness in regional air quality modeling. The efficiency of all the proposed models has experimented with two datasets for evaluation, and the comparative results illustrate the efficiency of all the models for an effective environmental decision support system.

In recent years, the segmentation process has undergone numerous changes, once with the advances in data mining. Knowledge discovery can automatize and provide better insights into customer trends and dynamics. The objective of the paper is to improve the quality of the marketing segmentation for company T. More specifically, the research question it plans to answer is whether data mining techniques deliver a better segmentation model than intuitive approaches. The segmentation steps comprise the identification of the necessary variables, the selection of the relevant ones to conduct the segmentation and the usage of artificial neural networks to predict future outcomes. To this end, the work makes use of web scraping (based on Google searches), K-means clustering and artificial neural networks.
Mestrado em Marketing

Considering the competitive environment in which companies operate nowadays and the importance of customer relationship management (CRM), it is crucial to analyse customer-related data to gain knowledge and insights about them in order to increase their retention and company’s performance. The presented investigation resulted from a curricular internship carried out at Inova+, a management consulting firm specialised in supporting the growth of organizations. In this sense, the aim of this investigation is to support the CRM system and the customer’s management strategies of Inova+, contributing to the improvement and strengthening of relations between the company and its customers. For this purpose, a quantitative methodology using analytical tools, namely data mining tools, was adopted to study various dimensions of CRM. In this context, this investigation focused on four main aspects under analysis, which allowed to obtain a more detailed knowledge about the company's customers. Initially, the observation of KPIs regarding the CRM and the company's performance through the construction of dashboards. Secondly, a time-series forecasting model for prospective revenues was applied. Additionally, an identification of customer segments according to their purchasing behaviour through the application of a RFM model and a clustering analysis was carried out. Finally, significant factors that influence the probability of adjudication of a commercial proposal were identified, such as the country, type of organisation and economic sector of the client company, as well as the service associated.
Considerando o ambiente competitivo em que as empresas operam atualmente e a importância do customer relationship management (CRM), é crucial analisar os dados relacionados com clientes para adquirir mais conhecimento e obter importantes insights sobre os mesmos, a fim de aumentar a sua retenção e o desempenho da empresa. A investigação apresentada resultou de um estágio curricular realizado na empresa Inova+, uma consultora especializada no apoio ao crescimento de organizações. Neste sentido, o objetivo desta investigação visa apoiar o sistema CRM e as estratégias de gestão de clientes da Inova+, contribuindo para a melhoria e fortalecimento das relações entre a empresa e os seus clientes. Para esse efeito, uma metodologia quantitativa utilizando ferramentas analíticas, nomeadamente ferramentas de data mining, foi adotada para estudar várias dimensões do CRM. Neste contexto, esta investigação focou-se em quatro aspetos principais em análise, que permitiram obter um conhecimento mais detalhado sobre os clientes da empresa. Inicialmente, a observação de KPIs relativos ao CRM e ao desempenho da empresa através da construção de dashboards. Em segundo lugar, foi aplicado um modelo de previsão de séries temporais relativo ao volume de negócios potencial. Adicionalmente, foram identificados segmentos de clientes de acordo com o seu comportamento de compra através da aplicação de um modelo RFM e foi desenvolvida uma análise de clustering. Por fim, foram identificados fatores significativos que influenciam a probabilidade de adjudicação de uma proposta comercial, tais como o país, tipo de organização e setor económico da empresa cliente, bem como o serviço associado.

Wireless mobile communications were initially a way for people to communicate through low data rate voice call connections. As data enabled devices allow users the ability to do much more with their mobile devices, so to will the demand for more reliable and pervasive wireless data. This is being addressed by so-called 4th generation wireless systems based on orthogonal frequency division multiplexing (OFDM) and multiple-input multiple-output (MIMO) antenna systems. Mobile wireless customers are becoming more demanding and expecting to have a great user experience over high speed broadband access at any time and anywhere, both indoor and outdoor. However, these promising improvements cannot be realized without an e±cient design of the receiver. Recently, receivers utilizing iterative detection and decoding have changed the fundamental receiver design paradigm from traditional separated parameter estimation and data detection blocks to an integrated iterative parameter estimator and data detection unit. Motivated by this iterative data driven approach, we develop low complexity iterative receivers with improved sensitivity compared to the conventional receivers, this brings potential benefits for the wireless communication system, such as improving the overall system throughput, increasing the macro cell coverage, and reducing the cost of the equipments in both the base station and mobile terminal. It is a challenge to design receivers that have good performance in a highly dynamic mobile wireless environment. One of the challenges is to minimize overhead reference signal energy (preamble, pilot symbols) without compromising the performance. We investigate this problem, and develop an iterative receiver with enhanced data-driven channel estimation. We discuss practical realizations of the iterative receiver for SISO-OFDM system. We utilize the channel estimation from soft decoded data (the a priori information) through frequency-domain combining and time-domain combining strategies in parallel with limited pilot signals. We analyze the performance and complexity of the iterative receiver, and show that the receiver's sensitivity can be improved even with this low complexity solution. Hence, seamless communications can be achieved with better macro cell coverage and mobility without compromising the overall system performance. Another challenge is that a massive amount of interference caused by MIMO transmission (spatial multiplexing MIMO) reduces the performance of the channel estimation, and further degrades data detection performance. We extend the iterative channel estimation from SISO systems to MIMO systems, and work with linear detection methods to perform joint interference mitigation and channel estimation. We further show the robustness of the iterative receivers in both indoor and outdoor environment compared to the conventional receiver approach. Finally, we develop low complexity iterative spatial multiplexed MIMO receivers for nonlinear methods based on two known techniques, that is, the Sphere Decoder (SD) method and the Markov Chain Monte Carlo (MCMC) method. These methods have superior performance, however, they typically demand a substantial increase in computational complexity, which is not favorable in practical realizations. We investigate and show for the first time how to utilize the a priori information in these methods to achieve performance enhancement while simultaneously substantially reducing the computational complexity. In our modified sphere decoder method, we introduce a new accumulated a priori metric in the tree node enumeration process. We show how we can improve the performance by obtaining the reliable tree node candidate from the joint Maximum Likelihood (ML) metric and an approximated a priori metric. We also show how we can improve the convergence speed of the sphere decoder (i.e., reduce the com- plexity) by selecting the node with the highest a priori probability as the starting node in the enumeration process. In our modified MCMC method, the a priori information is utilized for the firrst time to qualify the reliably decoded bits from the entire signal space. Two new robust MCMC methods are developed to deal with the unreliable bits by using the reliably decoded bit information to cancel the interference that they generate. We show through complexity analysis and performance comparison that these new techniques have improved performance compared to the conventional approaches, and further complexity reduction can be obtained with the assistance of the a priori information. Therefore, the complexity and performance tradeoff of these nonlinear methods can be optimized for practical realizations.

PV modules are engineered to produce electricity for 30+ years and are being deployed worldwide in ever more and ever bigger PV plants. Continuous quality assurance and performance analysis are the cornerstone for long-term reliability to maximize financial and energy returns. In today’s highly competitive Operation and Maintenance (O&M) market, employing and maintaining extensive networks of on-site sensors for remote monitoring purposes, proves challenging. Within this framework, data-driven solutions play a leading role to turn raw data from the field into reliable actionable insights. PV plant’s data from SCADA and monitoring systems is constantly subject to quality issues and the uncertainty related to it is directly reflected on the quality and reliability of the performance metrics used. In this work, the impact of the quality of the most relevant input parameters (i.e., output energy and irradiation) for the calculation key performance indicators (KPIs) is evaluated and different data cleaning and imputation techniques are benchmarked. The main objective of this work is to improve the quality of PV performance analysis by minimizing the negative effects of using incomplete and/or corrupted time-series as input for the calculation of PV plant KPIs (such as Performance Ratio and Availability). This objective is achieved through the assessment of different data sources with different intrinsic quality. In chapter 2, the methodology and data used are explained. Then, in chapter 3, as a pre-liminary data analysis, raw data from on-site sensors was compared with satellite-derived data to define and validate its uncertainty values. Special emphasis is given to irradiance sensors (pyranometers and reference cells), being the plane of array (POA) irradiance one of the variables with the greatest impact on performance evaluation. Later, in chapter 4, a consistent data quality analysis is proposed to assess the sensors’ health status to proceed with the corresponding cleaning procedure. At this stage, the concept of ‘virtual sensor’ is introduced, that solves the problem of having incomplete raw data by generating time-series with no missing data that efficiently combine on-site measurements with satellite data. Finally, in chapter 5, the advantage of performing data imputation using Machine Learning (ML) techniques is demonstrated by applying three good-performing algorithms (Random Forest, Bagging and Gradient Boosting Regressor) to replace missing data with highly accurate predicted values.

Access to abstract permanently restricted to Ball State community only.
Background : a survey for rate adaptation -- ABEP metric and channell assessment -- ABEP-based adaptive rate control -- Performance evaluation.
Access to thesis permanently restricted to Ball State community only.
Department of Computer Science