Dissertations / Theses on the topic 'Neural Tensor Network'

With the increasing number of wireless devices and the phenomenal amount of data that is being generated by them, there is a growing interest in the wireless communications community to complement the traditional model-driven design approaches with data-driven machine learning (ML)-based solutions. However, managing the large-scale multi-dimensional data to maintain the efficiency and scalability of the ML algorithms has obviously been a challenge. Tensors provide a useful framework to represent multi-dimensional data in an integrated manner by preserving relationships in data across different dimensions. This thesis studies two new applications of tensors to ML for wireless communications where the tensor structure of the concerned data is exploited in novel ways. The first contribution of this thesis is a tensor learning-based low-complexity precoder codebook design technique for a full-dimension multiple-input multiple-output (FD-MIMO) system with a uniform planar antenna (UPA) array at the transmitter (Tx) whose channel distribution is available through a dataset. Represented as a tensor, the FD-MIMO channel is further decomposed using a tensor decomposition technique to obtain an optimal precoder which is a function of Kronecker-Product (KP) of two low-dimensional precoders, each corresponding to the horizontal and vertical dimensions of the FD-MIMO channel. From the design perspective, we have made contributions in deriving a criterion for optimal product precoder codebooks using the obtained low-dimensional precoders. We show that this product codebook design problem is an unsupervised clustering problem on a Cartesian Product Grassmann Manifold (CPM), where the optimal cluster centroids form the desired codebook. We further simplify this clustering problem to a $K$-means algorithm on the low-dimensional factor Grassmann manifolds (GMs) of the CPM which correspond to the horizontal and vertical dimensions of the UPA, thus significantly reducing the complexity of precoder codebook construction when compared to the existing codebook learning techniques. The second contribution of this thesis is a tensor-based bandwidth-efficient gradient communication technique for federated learning (FL) with convolutional neural networks (CNNs). Concisely, FL is a decentralized ML approach that allows to jointly train an ML model at the server using the data generated by the distributed users coordinated by a server, by sharing only the local gradients with the server and not the raw data. Here, we focus on efficient compression and reconstruction of convolutional gradients at the users and the server, respectively. To reduce the gradient communication overhead, we compress the sparse gradients at the users to obtain their low-dimensional estimates using compressive sensing (CS)-based technique and transmit to the server for joint training of the CNN. We exploit a natural tensor structure offered by the convolutional gradients to demonstrate the correlation of a gradient element with its neighbors. We propose a novel prior for the convolutional gradients that captures the described spatial consistency along with its sparse nature in an appropriate way. We further propose a novel Bayesian reconstruction algorithm based on the Generalized Approximate Message Passing (GAMP) framework that exploits this prior information about the gradients. Through the numerical simulations, we demonstrate that the developed gradient reconstruction method improves the convergence of the CNN model.
Master of Science
The increase in the number of wireless and mobile devices have led to the generation of massive amounts of multi-modal data at the users in various real-world applications including wireless communications. This has led to an increasing interest in machine learning (ML)-based data-driven techniques for communication system design. The native setting of ML is {em centralized} where all the data is available on a single device. However, the distributed nature of the users and their data has also motivated the development of distributed ML techniques. Since the success of ML techniques is grounded in their data-based nature, there is a need to maintain the efficiency and scalability of the algorithms to manage the large-scale data. Tensors are multi-dimensional arrays that provide an integrated way of representing multi-modal data. Tensor algebra and tensor decompositions have enabled the extension of several classical ML techniques to tensors-based ML techniques in various application domains such as computer vision, data-mining, image processing, and wireless communications. Tensors-based ML techniques have shown to improve the performance of the ML models because of their ability to leverage the underlying structural information in the data. In this thesis, we present two new applications of tensors to ML for wireless applications and show how the tensor structure of the concerned data can be exploited and incorporated in different ways. The first contribution is a tensor learning-based precoder codebook design technique for full-dimension multiple-input multiple-output (FD-MIMO) systems where we develop a scheme for designing low-complexity product precoder codebooks by identifying and leveraging a tensor representation of the FD-MIMO channel. The second contribution is a tensor-based gradient communication scheme for a decentralized ML technique known as federated learning (FL) with convolutional neural networks (CNNs), where we design a novel bandwidth-efficient gradient compression-reconstruction algorithm that leverages a tensor structure of the convolutional gradients. The numerical simulations in both applications demonstrate that exploiting the underlying tensor structure in the data provides significant gains in their respective performance criteria.

Diffusion tensor imaging (DTI) is a noninvasive MRI technique used to assess white matter (WM) integrity, fiber orientation, and structural connectivity (SC) using water diffusion properties. DTI techniques are rapidly evolving and are now having a dramatic effect on depression research. Major depressive disorder (MDD) is highly prevalent and a leading cause of worldwide disability. Despite decades of research, the neurobiology of MDD remains poorly understood. MDD is increasingly viewed as a disorder of neural circuitry in which a network of brain regions involved in mood regulation is dysfunctional. In an effort to better understand the neurobiology of MDD and develop more effective treatments, much research has focused on delineating the structure of this mood regulation network. Although many studies have focused on the structural connectivity of the mood regulation network, findings using DTI are highly variable, likely due to many technical and analytical limitations. Further, structural connectivity pattern analyses have not been adequately utilized in specific clinical contexts where they would likely have high relevance, e.g., the use of white matter deep brain stimulation (DBS) as an investigational treatment for depression. In this dissertation, we performed a comprehensive analysis of structural WM integrity in a large sample of depressed patients and demonstrated that disruption of WM does not play a major role in the neurobiology of MDD. Using graph theory analysis to assess organization of neural network, we elucidated the importance of the WM network in MDD. As an extension of this WM network analysis, we identified the necessary and sufficient WM tracts (circuit) that mediate the response of subcallosal cingulate cortex DBS treatment for depression; this work showed that such analyses may be useful in prospective target selection. Collectively, these findings contribute to better understanding of depression as a neural network disorder and possibly will improve efficacy of SCC DBS.

Cost modelling in construction is the art and science of developing a reliable and effective estimation of the tender price of a project. Cost estimation is an experiencebased task, which involves evaluations of unknown circumstances and complex relationships of cost-influencing factors. Researchers argue that cost model developments lack rigour and consistent conceptual framework within which the performance of different models may be compared and evaluated. This study analyses construction cost models by classifying them into three groups according to the techniques used. These include deterministic models (regression analysis); probabilistic models (Monte Carlo simulation); and artificial intelligence models (neural networks). This research investigates the development of two methodologies for tender price estimation of buildings utilising neural computing and regression techniques. The emphasis is to provide clients and practitioners with a reliable tool, which would offer trustworthy advice and prediction of tender prices at an early stage of a construction project. The analysis in this research is based upon a data set of 230 office projects, newly constructed in the UK between 1983 and 1997. The cost data of these buildings consists of tender prices and 13 other cost influencing factors. The data extracted using the Building Cost Information Service (BCIS) database of the Royal Institution of Chartered Surveyors (RICS). Questionnaire survey and interviews were adopted to identify, evaluate and rank cost significant factors according to their degree of influence on tender prices. The practitioners involved in this stage were UK based quantity surveyors. Some of these cost variables formulate the basis for developing the tender estimation models. Cluster analysis was conducted to categorise the data set into more homogeneous project groups based upon the cost variables. The hypothesis is that developing estimation models using project categories would yield better performance and more efficient models. Self-Organising Maps (SOM), a type of neural networks, is used for the cluster analysis. Seventeen neural networks and thirteen regression models are developed for tender price estimation using different parameters and cost factors. The performance and efficiency of these models are analysed and compared before and after the cluster analysis of the data set. On the other hand, sensitivity analysis is conducted by developing fifty-five models to evaluate the effectiveness of different combinationso f network parameterso n the accuracyo f tenderp rice estimation. The research findings indicate that, when the whole data set of 230 office projects is used, both methodologies produced low accuracy and failed to map the relationship between the tender price and the selected influencing cost factors. On the contrary, after clustering the data set into coherent groups using Kohonen neural networks, the performance of both RA and ANN models increased dramatically, with many estimation accuracies above 80% and 90%, which is highly satisfactory for tender price estimation at an early stage of a project. The outcomes imply that: (a) clustering the projects into homogeneous categories is significant and key for model performance and accuracy; (b) after cluster analysis there is no significant difference in the performance of RA and ANN models, although the RA outperformed the ANN in some models. The results also reveal that for both methodologies the accuracy of the estimation models that utilised two cost factors (project area and duration) outperformed the estimation models that used 13 cost factors, which is an indication that area and duration are the most dominant cost determinant variables.

Modern robotics and automation systems require high-level reasoning capability in representing, identifying, and interpreting the three-dimensional data of the real world. Understanding the world's geometric structure by visual data is known as 3D perception. The necessity of analyzing irregular and complex 3D data has led to the development of high-dimensional frameworks for data learning. Here, we design several sparse learning-based approaches for high-dimensional data that effectively tackle multiple perception problems, including data filtering, data recovery, and data retrieval. The frameworks offer generative solutions for analyzing complex and irregular data structures without prior knowledge of data. The first part of the dissertation proposes a novel method that simultaneously filters point cloud noise and outliers as well as completing missing data by utilizing a unified framework consisting of a novel tensor data representation, an adaptive feature encoder, and a generative Bayesian network. In the next section, a novel multi-level generative chaotic Recurrent Neural Network (RNN) has been proposed using a sparse tensor structure for image restoration. In the last part of the dissertation, we discuss the detection followed by localization, where we discuss extracting features from sparse tensors for data retrieval.
Doctor of Philosophy
The development of automation systems and robotics brought the modern world unrivaled affluence and convenience. However, the current automated tasks are mainly simple repetitive motions. Tasks that require more artificial capability with advanced visual cognition are still an unsolved problem for automation. Many of the high-level cognition-based tasks require the accurate visual perception of the environment and dynamic objects from the data received from the optical sensor. The capability to represent, identify and interpret complex visual data for understanding the geometric structure of the world is 3D perception. To better tackle the existing 3D perception challenges, this dissertation proposed a set of generative learning-based frameworks on sparse tensor data for various high-dimensional robotics perception applications: underwater point cloud filtering, image restoration, deformation detection, and localization. Underwater point cloud data is relevant for many applications such as environmental monitoring or geological exploration. The data collected with sonar sensors are however subjected to different types of noise, including holes, noise measurements, and outliers. In the first chapter, we propose a generative model for point cloud data recovery using Variational Bayesian (VB) based sparse tensor factorization methods to tackle these three defects simultaneously. In the second part of the dissertation, we propose an image restoration technique to tackle missing data, which is essential for many perception applications. An efficient generative chaotic RNN framework has been introduced for recovering the sparse tensor from a single corrupted image for various types of missing data. In the last chapter, a multi-level CNN for high-dimension tensor feature extraction for underwater vehicle localization has been proposed.

This thesis presents a fuzzy logic controller aimed at maintaining constant tension between two adjacent stands in tandem rolling mills. The fuzzy tension controller monitors tension variation by resorting to electric current comparison of different operation modes and sets the reference for speed controller of the upstream stand. Based on modeling the rolling stand as a single input single output linear discrete system, which works in the normal mode and is subject to internal and external noise, the element settings and parameter selections in the design of the fuzzy controller are discussed. To improve the performance of the fuzzy controller, a dynamic fuzzy controller is proposed. By switching the fuzzy controller elements in relation to the step response, both transient and stationary performances are enhanced. To endow the fuzzy controller with intelligence of generalization, flexibility and adaptivity, self-learning techniques are introduced to obtain fuzzy controller parameters. With the inclusion of supervision and concern for conventional control criteria, the parameters of the fuzzy inference system are tuned by a backward propagation algorithm or their optimal values are located by means of a genetic algorithm. In simulations, the neuro-fuzzy tension controller exhibits the real-time applicability, while the genetic fuzzy tension controller reveals an outstanding global optimization ability.

Made available in DSpace on 2017-06-01T18:20:30Z (GMT). No. of bitstreams: 0 Previous issue date: 2011-02-28
The success of artificial neural networks (ANN) applications as an alternative modeling technique to response surface methodology (RSM) has attracted interest from major industries such as pharmaceuticals, cosmetics, oil, food, petroleum and surfactants, among others. Development of production media is a strategic area for the industry of biosurfactants by to increase efficiency and reduce costs of the process. In this area, surface tension measurements and emulsification activity has been routinely used for indirect monitoring of biosurfactant production. In this paper, the capabilities of RNA-based modeling and MSR were compared in surface tension estimation of biosurfactant production media. The two techniques used experimental data from the central composite design with four axial points and three replicates at the central point. The concentrations of ammonium sulfate and potassium monobasic phosphate were used as independent variables. The surface tensions of cell-free broths, with 96 h, of biosurfactant production media by Candida lipolytica UCP 988 in sea water were used as response variable. The results demonstrated the superiority of the RNA-based methodology. The quadratic model obtained using MSR showed a coefficient of determination equal to 0.43 and highly significant lack of fit. The fit of the model RNA based on experimental data was excellent. Simulations with the model using the training, validation an test sets showed root mean squared error (rmse) of less than 0.05 and coefficients of determination higher than 0.99. In this context, the RNA-based estimation of surface tension from the constituents of biosurfactant production media showed to be an efficient, reliable and economical method to monitor the biosurfactant production. The work also showed the ability of the yeast Candida lipolytica UCP 0988 use corn oil and produce biosurfactants in extremely alkaline sea water (initial pH 14), supplemented with sources of nitrogen and phosphorus
O sucesso das aplicações de redes neurais artificiais (RNA) como técnica de modelagem alternativa à metodologia de superfície de resposta (MSR) tem atraído o interesse de grandes indústrias, como a farmacêutica, a de cosméticos, a de alimentos, a de petróleo e a de surfactantes, entre outras. Desenvolvimento de meios de produção é uma área estratégica para a indústria de biossurfactantes por aumentar a eficiência e reduzir custos do processo. Nesta área, determinações de tensão superficial e de atividade de emulsificação vem sendo usadas rotineiramente para monitoramento indireto da produção de biossurfactantes. No presente trabalho, as capacidades de modelagem de metodologia baseada em RNA e metodologia de superfície de resposta foram comparadas na estimação de tensão superficial de meios de produção de biossurfactante. As duas técnicas usaram dados experimentais obtidos de planejamento composto central, com 4 pontos axiais e 3 repetições no ponto central, tendo as concentrações de sulfato de amônio e fosfato monobásico de potássio como variáveis independentes e como variável resposta a tensão superficial de líquidos metabólicos, com 96 horas, livres de células, de meios de produção de biossurfactante por Candida lipolytica UCP 988. Os resultados demonstraram a superioridade da metodologia baseada em RNA. O modelo quadrático obtido usando MSR apresentou coeficiente de determinação igual a 0,43 e falta de ajuste altamente significativa. O ajuste do modelo baseado em RNA aos dados experimentais foi excelente. Simulações com o modelo usando os conjuntos de treinamento, validação e teste apresentaram raízes dos erros quadráticos médios (rmse) inferiores a 0,05 e coeficientes de determinação superiores a 0,99. Neste contexto, a estimação da tensão superficial baseada em RNA a partir dos constituintes de meios de produção de biossurfactantes mostrou ser um método eficaz, confiável e econômico para monitorar a produção de biossurfactantes. O trabalho mostrou também a capacidade da levedura Candida lipolytica UCP 0988 utilizar óleo de milho e produzir biossurfactantes em água do mar extremamente alcalina (pH inicial 14), suplementada com fontes de nitrogênio e fósforo

In this thesis, three methods are presented to model the decision process of whether or not to bid for a tender in offshore petroleum platform fabrication. A sample data and the assessment based on this data are gathered from an offshore petroleum platform fabrication company and this information is analyzed to understand the significant parameters in the industry. The alternative methods, &ldquo
Regression Analysis&rdquo
, &ldquo
Neural Network Method&rdquo
and &ldquo
Fuzzy Neural Network Method&rdquo
, are used for modeling of the bidding decision process. The regression analysis examines the data statistically where the neural network method and fuzzy neural network method are based on artificial intelligence. The models are developed using the bidding data compiled from the offshore petroleum platform fabrication projects. In order to compare the prediction performance of these methods &ldquo
Cross Validation Method&rdquo
is utilized. The models developed in this study are compared with the bidding decision method used by the company. The results of the analyses show that regression analysis and neural network method manage to have a prediction performance of 80% and fuzzy neural network has a prediction performance of 77,5% whereas the method used by the company has a prediction performance of 47,5%. The results reveal that the suggested models achieve significant improvement over the existing method for making the correct bidding decision.

Dans cette thèse, nous développons des algorithmes à haute performance pour certains calculs impliquant des tenseurs denses et des matrices éparses. Nous abordons les opérations du noyau qui sont utiles pour les tâches d'apprentissage de la machine, telles que l'inférence avec les réseaux neuronaux profonds. Nous développons des structures de données et des techniques pour réduire l'utilisation de la mémoire, pour améliorer la localisation des données et donc pour améliorer la réutilisation du cache des opérations du noyau. Nous concevons des algorithmes parallèles à mémoire séquentielle et à mémoire partagée.Dans la première partie de la thèse, nous nous concentrons sur les noyaux tenseurs denses. Les noyaux tenseurs comprennent la multiplication tenseur-vecteur (TVM), la multiplication tenseur-matrice (TMM) et la multiplication tenseur-tendeur (TTM). Parmi ceux-ci, la MVT est la plus liée à la largeur de bande et constitue un élément de base pour de nombreux algorithmes. Nous proposons une nouvelle structure de données qui stocke le tenseur sous forme de blocs, qui sont ordonnés en utilisant la courbe de remplissage de l'espace connue sous le nom de courbe de Morton (ou courbe en Z). L'idée clé consiste à diviser le tenseur en blocs suffisamment petits pour tenir dans le cache et à les stocker selon l'ordre de Morton, tout en conservant un ordre simple et multidimensionnel sur les éléments individuels qui les composent. Ainsi, des routines BLAS haute performance peuvent être utilisées comme micro-noyaux pour chaque bloc. Les résultats démontrent non seulement que l'approche proposée est plus performante que les variantes de pointe jusqu'à 18%, mais aussi que l'approche proposée induit 71% de moins d'écart-type d'échantillon pour le MVT dans les différents modes possibles. Enfin, nous étudions des algorithmes de mémoire partagée parallèles pour la MVT qui utilisent la structure de données proposée. Nos résultats sur un maximum de 8 systèmes de prises montrent une performance presque maximale pour l'algorithme proposé pour les tenseurs à 2, 3, 4 et 5 dimensions.Dans la deuxième partie de la thèse, nous explorons les calculs épars dans les réseaux de neurones en nous concentrant sur le problème d'inférence profonde épars à haute performance. L'inférence sparse DNN est la tâche d'utiliser les réseaux sparse DNN pour classifier un lot d'éléments de données formant, dans notre cas, une matrice de caractéristiques sparse. La performance de l'inférence clairsemée dépend de la parallélisation efficace de la matrice clairsemée - la multiplication matricielle clairsemée (SpGEMM) répétée pour chaque couche dans la fonction d'inférence. Nous introduisons ensuite l'inférence modèle-parallèle, qui utilise un partitionnement bidimensionnel des matrices de poids obtenues à l'aide du logiciel de partitionnement des hypergraphes. Enfin, nous introduisons les algorithmes de tuilage modèle-parallèle et de tuilage hybride, qui augmentent la réutilisation du cache entre les couches, et utilisent un module de synchronisation faible pour cacher le déséquilibre de charge et les coûts de synchronisation. Nous évaluons nos techniques sur les données du grand réseau du IEEE HPEC 2019 Graph Challenge sur les systèmes à mémoire partagée et nous rapportons jusqu'à 2x l'accélération par rapport à la ligne de base
In this thesis, we develop high performance algorithms for certain computations involving dense tensors and sparse matrices. We address kernel operations that are useful for machine learning tasks, such as inference with deep neural networks (DNNs). We develop data structures and techniques to reduce memory use, to improve data locality and hence to improve cache reuse of the kernel operations. We design both sequential and shared-memory parallel algorithms. In the first part of the thesis we focus on dense tensors kernels. Tensor kernels include the tensor--vector multiplication (TVM), tensor--matrix multiplication (TMM), and tensor--tensor multiplication (TTM). Among these, TVM is the most bandwidth-bound and constitutes a building block for many algorithms. We focus on this operation and develop a data structure and sequential and parallel algorithms for it. We propose a novel data structure which stores the tensor as blocks, which are ordered using the space-filling curve known as the Morton curve (or Z-curve). The key idea consists of dividing the tensor into blocks small enough to fit cache, and storing them according to the Morton order, while keeping a simple, multi-dimensional order on the individual elements within them. Thus, high performance BLAS routines can be used as microkernels for each block. We evaluate our techniques on a set of experiments. The results not only demonstrate superior performance of the proposed approach over the state-of-the-art variants by up to 18%, but also show that the proposed approach induces 71% less sample standard deviation for the TVM across the d possible modes. Finally, we show that our data structure naturally expands to other tensor kernels by demonstrating that it yields up to 38% higher performance for the higher-order power method. Finally, we investigate shared-memory parallel TVM algorithms which use the proposed data structure. Several alternative parallel algorithms were characterized theoretically and implemented using OpenMP to compare them experimentally. Our results on up to 8 socket systems show near peak performance for the proposed algorithm for 2, 3, 4, and 5-dimensional tensors. In the second part of the thesis, we explore the sparse computations in neural networks focusing on the high-performance sparse deep inference problem. The sparse DNN inference is the task of using sparse DNN networks to classify a batch of data elements forming, in our case, a sparse feature matrix. The performance of sparse inference hinges on efficient parallelization of the sparse matrix--sparse matrix multiplication (SpGEMM) repeated for each layer in the inference function. We first characterize efficient sequential SpGEMM algorithms for our use case. We then introduce the model-parallel inference, which uses a two-dimensional partitioning of the weight matrices obtained using the hypergraph partitioning software. The model-parallel variant uses barriers to synchronize at layers. Finally, we introduce tiling model-parallel and tiling hybrid algorithms, which increase cache reuse between the layers, and use a weak synchronization module to hide load imbalance and synchronization costs. We evaluate our techniques on the large network data from the IEEE HPEC 2019 Graph Challenge on shared-memory systems and report up to 2x times speed-up versus the baseline

Identifying new potential treatment options for medical conditions that cause human disease burden is a central task of biomedical research. Since all candidate drugs cannot be tested with animal and clinical trials, in vitro approaches are first attempted to identify promising candidates. Likewise, identifying other essential relations (e.g., causation, prevention) between biomedical entities is also critical to understand biomedical processes. Hence, it is crucial to develop automated relation prediction systems that can yield plausible biomedical relations to expedite the discovery process. In this dissertation, we demonstrate three approaches to predict treatment relations between biomedical entities for the drug repositioning task using existing biomedical knowledge bases. Our approaches can be broadly labeled as link prediction or knowledge base completion in computer science literature. Specifically, first we investigate the predictive power of graph paths connecting entities in the publicly available biomedical knowledge base, SemMedDB (the entities and relations constitute a large knowledge graph as a whole). To that end, we build logistic regression models utilizing semantic graph pattern features extracted from the SemMedDB to predict treatment and causative relations in Unified Medical Language System (UMLS) Metathesaurus. Second, we study matrix and tensor factorization algorithms for predicting drug repositioning pairs in repoDB, a general purpose gold standard database of approved and failed drug–disease indications. The idea here is to predict repoDB pairs by approximating the given input matrix/tensor structure where the value of a cell represents the existence of a relation coming from SemMedDB and UMLS knowledge bases. The essential goal is to predict the test pairs that have a blank cell in the input matrix/tensor based on the shared biomedical context among existing non-blank cells. Our final approach involves graph convolutional neural networks where entities and relation types are embedded in a vector space involving neighborhood information. Basically, we minimize an objective function to guide our model to concept/relation embeddings such that distance scores for positive relation pairs are lower than those for the negative ones. Overall, our results demonstrate that recent link prediction methods applied to automatically curated, and hence imprecise, knowledge bases can nevertheless result in high accuracy drug candidate prediction with appropriate configuration of both the methods and datasets used.

This paper is focus on flowers recognition in an image and class classification. Theoretical part is focus on problematics of deep convolutional neural networks. The practical part if focuse on created flowers database, with which it is further worked on. The database conteins it total 13000 plant pictures of 26 spicies as cornflower, violet, gerbera, cha- momile, cornflower, liverwort, hawkweed, clover, carnation, lily of the valley, marguerite daisy, pansy, poppy, marigold, daffodil, dandelion, teasel, forget-me-not, rose, anemone, daisy, sunflower, snowdrop, ragwort, tulip and celandine. Next is in the paper described used neural network model Inception v3 for class classification. The resulting accuracy has been achieved 92%.

Cette thèse s’inscrit dans le contexte de la surveillance des postes sous enveloppe métallique (PSEM) en courant continu (DC). La disponibilité de ces équipements étant primordiale pour leurs utilisateurs, il est nécessaire de disposer d’un outil de surveillance (monitoring) permettant de prévenir toute défaillance. Cet outil doit être capable de détecter et d’identifier les défauts présents, afin d’apporter une réponse adaptée. Depuis de nombreuses années, le monitoring des PSEM en AC est réalisé grâce à la mesure des décharges partielles (DP). Malheureusement, les connaissances des DP dans les PSEM en DC sont encore lacunaires, et les techniques d’identification des défauts sont intrinsèquement liées à l’environnement AC. De nouvelles techniques sont donc nécessaires en DC.Ce travail de thèse avait pour but de caractériser les décharges partielles dans les postes sous enveloppe métallique en tension continue, et de mettre en place une solution de reconnaissance automatique des défauts. Pour cela, un banc de mesure des décharges partielles a d’abord été mis en place. Afin de garantir la pertinence des résultats pour des systèmes industriels, les travaux ont été réalisés dans une section de PSEM sous tension continue. Le comportement des DP a été étudié pour deux types de défauts : des pointes sur le conducteur haute-tension et des particules libres métalliques. La caractérisation a porté sur l’influence de plusieurs paramètres : la nature et la pression du gaz, le niveau et la polarité de la tension. La mesure des DP a d’abord été réalisée en conformité avec la norme IEC 60270, permettant ainsi d’évaluer la pertinence de cette méthode pour les applications DC. La caractérisation a été complétée grâce à d’autres chaînes de mesure : une mesure de courant stationnaire, une mesure de courant haute-fréquence, une mesure de lumière, et une mesure des ondes ultra-haute fréquence (UHF). Le travail sur l’identification des défauts a d’abord consisté à construire une signature pertinente à partir des mesures de DP, puis à constituer une base de données, et enfin à implémenter un algorithme de reconnaissance automatique.Ces travaux ont montré que la méthode conventionnelle de mesure des DP présente certaines limites pour la détection des décharges partielles en DC, notamment pour les décharges couronne. Elle a tout de même permis de faire une bonne partie du travail de caractérisation. Les résultats obtenus avec les autres chaînes de mesure utilisées ont permis d’expliquer les lacunes de la méthode conventionnelle. Ils ont également permis un véritable apport pour la caractérisation des DP engendrées par des défauts de type pointe et particule. Enfin, une classification automatique efficace des défauts a été mise en place. Elle s’appuie sur le diagramme q(Δt) issu des données de la mesure conventionnelle des décharges partielles et sur un algorithme de réseau de neurones
The framework of this thesis is the monitoring of High-Voltage, Direct Current (HVDC) Gas-Insulated Substations (GIS). The availability of these equipment is crucial for electrical networks operators. That is why they need a preventive diagnosis tool. The solution must be able to detect and identify the insulation defects, so that an appropriate maintenance can be planned. The last 40 years have seen Partial Discharges (PD) measurement become a classic monitoring tool for AC GIS. Unfortunately, there is a lack of scientific information about PD in HVDC GIS, and the known defect identification techniques are very specific to the AC environment. New techniques are thus needed in DC.This thesis aimed to characterize partial discharges in DC gas-insulated substations, and to develop an automatic defect identification tool. The first step of this work was the development of a partial discharge measuring bench. The complete study has been performed in a GIS section, so that the results can be directly applied to industrial equipment. Two kinds of defect have been investigated: protrusions on the high-voltage conductor, and free metallic particles. The influence of parameters such as gas nature and pressure, voltage level and polarity has been evaluated. First, PD have been measured in conformity with the IEC 60270 standard, and the relevance of this method in a DC environment has been evaluated. Then, other measuring chains have been used to improve the characterization of partial discharges: a steady-state current measurement, a high-frequency current measurement, a light measurement and a measurement of Ultra-High Frequency (UHF) waves. Finally, a relevant signature for defect identification has been designed and extracted from DP recordings. A database has been constituted, and an automated recognition algorithm has been implemented.The results show that the conventional PD measurement technique is not fully adapted to partial discharges detection in DC, corona discharges being the most problematic situation. Nevertheless, this method has brought enough information to start the characterization of PD. The limitations of the conventional method have been explained thanks to the results of the other measurements. These other experimental results have led to an actual improvement of the characterization of protrusion and particle-generated partial discharges. An effective automated defect classification solution has been implemented. The signature is derived from the q(Δt) diagram that has been extracted from the data obtained with the partial discharge conventional measurement. The identification algorithm has a neural network structure

Les signatures spectrales des composants constitutifs présents dans les images hyperspectrales peuvent être significativement affectées par les variations des conditions atmosphériques, d'illumination ou d'environnement se produisant typiquement dans une image. Les algorithmes traditionnels de démélange spectral (Spectral Unmixing - SU) négligent la variabilité spectrale des composants constitutifs, ce qui propage des erreurs importantes tout au long du processus de démélange et compromet la qualité des abondances estimées. Par conséquent, des efforts importants ont été récemment consacrés à atténuer les effets de la variabilité spectrale dans les procédures de démélange. Cependant, de nombreux défis restent à relever pour savoir comment exploiter au mieux les informations a priori sur le problème afin d'améliorer à la fois la qualité et la robustesse des algorithmes de SU qui tiennent compte de la variabilité spectrale des composants. Dans cette thèse, de nouvelles stratégies sont développées pour aborder cette variabilité spectrale. Premièrement, une stratégie de régularisation multi-échelles basée sur la (sur)-segmentation des images est proposée pour explorer plus efficacement les informations spatiales sur les abondances. De nouveaux algorithmes sont ensuite proposés pour le démélange spéctral semi-supervisé et non-supervisé, ce qui se traduit par une amélioration des performances de reconstruction des abondances avec une complexité de calcul réduite. Ensuite, trois nouveaux modèles sont proposés pour représenter la variabilité spectrale des composants constitutifs, en utilisant des représentations paramétriques, tensorielles et basées sur des réseaux neuronaux pour les spectres de ces composants en chaque pixel de l'image. Le modèle paramétrique introduit des facteurs multiplicatifs dépendant des pixels dans une matrice des composants de référence pour modéliser une variabilité spectrale arbitraire, tandis que la représentation basée sur un tenseur permet d'exploiter la grande dimension des données en exploitant sa structure de rang faible sous-jacente. Les réseaux de neurones génératifs (tels que les variational autoencoders ou les generative adversarial networks) permettent enfin de modéliser la variété de faible dimension des signatures spectrales des matériaux, directement à partir des données observées. Les modèles proposés sont utilisés dans la conception de quatre nouveaux algorithmes de démélange non-supervisés et semi-supervisés. Enfin, nous donnons un bref aperçu des travaux qui étendent les stratégies proposées dans la thèse à de nouveaux problèmes en démélange et en dans l'analyse d'images hyperspectrales. Cela comprend l'utilisation de la régularisation d'abondance multi-échelles en démélange spectral non-linéaire, la modélisation de la variabilité spectrale, la prise en compte des changements soudains lors du démélange et la détection des changements dans les images hyperspectrales multitemporelles, ainsi que la prise en compte de la variabilité spectrale et des changements dans le problème de fusion d'images hyperspectrales et multispectrales
The spectral signatures of the materials contained in hyperspectral images, also called endmembers (EMs), can be significantly affected by variations in atmospheric, illumination or environmental conditions typically occurring within an image. Traditional spectral unmixing (SU) algorithms neglect the spectral variability of the endmembers, what propagates significant mismodeling errors throughout the whole unmixing process and compromises the quality of the estimated abundances. Therefore, significant effort have been recently dedicated to mitigate the effects of spectral variability in SU. However, many challenges still remain in how to best explore a priori information about the problem in order to improve the quality, the robustness and the efficiency of SU algorithms that account for spectral variability. In this thesis, new strategies are developed to address spectral variability in SU. First, an (over)-segmentation-based multiscale regularization strategy is proposed to explore spatial information about the abundance maps more effectively. New algorithms are then proposed for both semi-supervised and blind SU, leading to improved abundance reconstruction performance at a small computational complexity. Afterwards, three new models are proposed to represent spectral variability of the EMs in SU, using parametric, tensor, and neural network-based representations for EM spectra at each image pixel. The parametric model introduces pixel-dependent scaling factors over a reference EM matrix to model arbitrary spectral variability, while the tensor-based representation allows one to exploit the high-dimensional nature of the data by means of its underlying low-rank structure. Generative neural networks (such as variational autoencoders or generative adversarial networks) finally allow one to model the low-dimensional manifold of the spectral signatures of the materials more effectively. The proposed models are used to devise three new blind SU algorithms, and to perform data augmentation in library-based SU. Finally, we provide a brief overview of work which extends the proposed strategies to new problems in SU and in hyperspectral image analysis. This includes the use of the multiscale abundance regularization in nonlinear SU, modeling spectral variability and accounting for sudden changes when performing SU and change detection of multitemporal hyperspectral images, and also accounting for spectral variability and changes in the multimodal (i.e., hyperspectral and multispectral) image fusion problem

Coordenação de Aperfeiçoamento de Pessoal de Nível Superior
O objetivo deste trabalho é contribuir com o desenvolvimento de uma técnica baseada em sistemas inteligentes que possibilite a localização exata ou aproximada do ponto de origem de uma Variação de Tensão de Curta Duração (VTCD) (gerada por uma falta) em um sistema de distribuição de energia elétrica. Este trabalho utiliza um Phase-Locked Loop (PLL) com o intuito de detectar as faltas. Uma vez que a falta é detectada, os sinais de tensão obtidos durante a falta são decompostos em componentes simétricas instantâneas por meio do método proposto. Em seguida, as energias das componentes simétricas são calculadas e utilizadas para estimar a localização da falta. Nesta pesquisa, são avaliadas duas estruturas baseadas em Redes Neurais Artificiais (RNAs). A primeira é projetada para classificar a localização da falta em um dos pontos possíveis e a segunda é projetada para estimar a distância da falta ao alimentador. A técnica aqui proposta aplica-se a alimentadores trifásicos com cargas equilibradas. No desenvolvimento da mesma, considera-se que há disponibilidade de medições de tensões no nó inicial do alimentador e também em pontos esparsos ao longo da rede de distribuição. O banco de dados empregado foi obtido através de simulações de um modelo de alimentador radial usando o programa PSCAD/EMTDC. Testes de sensibilidade empregando validação-cruzada são realizados em ambas as arquiteturas de redes neurais com o intuito de verificar a confiabilidade dos resultados obtidos. Adicionalmente foram realizados testes com faltas não inicialmente contidas no banco de dados a fim de se verificar a capacidade de generalização das redes. Os desempenhos de ambas as arquiteturas de redes neurais foram satisfatórios e demonstram a viabilidade das técnicas propostas para realizar a localização de faltas em redes de distribuição.
The aim of this work is to contribute to the development of a technique based on intelligent systems that allows the accurate location of the Short Duration Voltage Variations (SDVV) origin in an electrical power distribution system. Once the fault is detected via a Phase-Locked Loop (PLL), voltage signals acquired during the fault are decomposed into instantaneous symmetrical components by the proposed method. Then, the energies of the symmetrical components are calculated and used to estimate the fault location. In this work, two systems based on Artificial Neural Networks (ANN) are evaluated. The first one is designed to classify the fault location into one of predefined possible points and the second is designed to estimate the fault distance from the feeder. The technique herein proposed is applies to three-phase feeders with balanced loads. In addition, it is considered that there is availability of voltage measurements in the initial node of the feeder and also in sparse points along the distribution power grid. The employed database was made using simulations of a model of radial feeder using the PSCAD / EMTDC program. Sensitivity tests employing cross-validation are performed for both approaches in order to verify the reliability of the results. Furthermore, in order to check the generalization capability, tests with faults not originally contained in the database were performed. The performances of both architectures of neural networks were satisfactory and they demonstrate the feasibility of the proposed techniques to perform fault location on distribution grids.

Fundação de Amparo à Pesquisa do Estado do Rio de Janeiro
O objetivo deste trabalho é conhecer e compreender melhor os imprevistos no fornecimento de energia elétrica, quando ocorrem as variações de tensão de curta duração (VTCD). O banco de dados necessário para os diagnósticos das faltas foi obtido através de simulações de um modelo de alimentador radial através do software PSCAD/EMTDC. Este trabalho utiliza um Phase-Locked Loop (PLL) com o intuito de detectar VTCDs e realizar a estimativa automática da frequência, do ângulo de fase e da amplitude das tensões e correntes da rede elétrica. Nesta pesquisa, desenvolveram-se duas redes neurais artificiais: uma para identificar e outra para localizar as VTCDs ocorridas no sistema de distribuição de energia elétrica. A técnica aqui proposta aplica-se a alimentadores trifásicos com cargas desequilibradas, que podem possuir ramais laterais trifásicos, bifásicos e monofásicos. No desenvolvimento da mesma, considera-se que há disponibilidade de medições de tensões e correntes no nó inicial do alimentador e também em alguns pontos esparsos ao longo do alimentador de distribuição. Os desempenhos das arquiteturas das redes neurais foram satisfatórios e demonstram a viabilidade das RNAs na obtenção das generalizações que habilitam o sistema para realizar a classificação de curtos-circuitos.
The objective of this work is to know and understand the unforeseen in the supply of electricity, when there are short duration voltage variations (SDVV). The required databases for the diagnosis of faults were obtained through simulations of a model of radial feeder through software PSCAD/EMTDC. This work uses a Phase-Locked Loop (PLL) in order to detect and perform the estimation SDVV automatic frequency, phase angle and amplitude of the voltage and current from the power grid. This research is developing two artificial neural networks: one to identify and another to locate the SDVV occurred in the distribution system of electricity. The technique proposed here applies to three-phase feeders with unbalanced loads, which can have side extensions triphasic, biphasic and monophasic. In developing the same, it is considered that there is availability of measurements of voltages and currents at the node of the initial feeder and also in some points scattered along the distribution feeder. The performances of the architectures of neural networks were satisfactory and demonstrate the feasibility of ANNs in obtaining the generalizations that enables the system for the classification of short circuits.

У дисертації розроблено спосіб безконтактного визначення струму в стінках одного підземного трубопроводу на основі однократного вимірювання напруженостей п’ятьма магнітними антенами, що знаходяться в двох блоках, без попередньої орієнтації бази вимірювальної системи в перпендикулярній до осі трубопроводу площині. Розроблено спосіб автоматичного профілювання горизонтальної складової напруженості магнітного поля при проходженні із вимірювачьною системою над трубопроводами в перпендикулярному відносно їх осей напрямку. Запропоновано використання трьохшарової нейронної мережі для безконтактного визначення струму в стінках одного і двох підземних трубопроводів на основі даних профілю напруженостей магнітного поля над цими трубопроводами. Розроблено спосіб, в якому передбачено використання умовних рівнянь і отримання на їх базі нормальних рівнянь для безконтактного визначення струмів в стінках підземних трубопроводів при перпендикулярному проходженні над ними. Запропоновано структурну схему і виготовлено систему для безконтактного визначення струму в стінках підземних трубопроводів, яка реалізує всі запропоновані способи визначення цих струмів. Виконано метрологічний аналіз розробленої системи безконтактного визначення струмів в підземних трубопроводах, розроблена установка, яка дає змогу проводити експериментальні дослідження метрологічних характеристик розробленої системи безконтактного визначення струмів в підземних трубопроводах, а також подібних їй приладів і систем. Визначено метрологічні показники розробленої системи при безконтактному визначенні струмів у стінках контрольованих трубопроводів. Проведені лабораторні, польові і промислові випробування розробленої системи, які підтвердили її працездатність і можливість використання для контролю ізоляційного покриття підземних трубопроводів на основі заникання струму вздовж траси.
У дисертації розроблено спосіб безконтактного визначення струму в стінках одного підземного трубопроводу на основі однократного вимірювання напруженостей п’ятьма магнітними антенами, що знаходяться в двох блоках, без попередньої орієнтації бази вимірювальної системи в перпендикулярній до осі трубопроводу площині. Розроблено спосіб автоматичного профілювання горизонтальної складової напруженості магнітного поля при проходженні із вимірювачьною системою над трубопроводами в перпендикулярному відносно їх осей напрямку. Запропоновано використання трьохшарової нейронної мережі для безконтактного визначення струму в стінках одного і двох підземних трубопроводів на основі даних профілю напруженостей магнітного поля над цими трубопроводами. Розроблено спосіб, в якому передбачено використання умовних рівнянь і отримання на їх базі нормальних рівнянь для безконтактного визначення струмів в стінках підземних трубопроводів при перпендикулярному проходженні над ними. Запропоновано структурну схему і виготовлено систему для безконтактного визначення струму в стінках підземних трубопроводів, яка реалізує всі запропоновані способи визначення цих струмів. Виконано метрологічний аналіз розробленої системи безконтактного визначення струмів в підземних трубопроводах, розроблена установка, яка дає змогу проводити експериментальні дослідження метрологічних характеристик розробленої системи безконтактного визначення струмів в підземних трубопроводах, а також подібних їй приладів і систем. Визначено метрологічні показники розробленої системи при безконтактному визначенні струмів у стінках контрольованих трубопроводів. Проведені лабораторні, польові і промислові випробування розробленої системи, які підтвердили її працездатність і можливість використання для контролю ізоляційного покриття підземних трубопроводів на основі заникання струму вздовж траси.

La fibrillation atriale (FA) est l’arythmie cardiaque soutenue la plus fréquemment rencontrée dans la pratique clinique. Pour la traiter, l’ablation par cathéter de zones cardiaques jugées responsables de soutenir l’arythmie est devenue la thérapie la plus utilisée. Un nouveau protocole d’ablation se base sur l’identification des zones atriales où les électrogrammes (EGM) enregistrés à l’aide d’un cathéter à électrodes multiples, appelé PentaRay, manifestent des décalages spatiotemporels significatifs sur plusieurs voies adjacentes. Ce phénomène est appelé dispersion spatio-temporelle (DST). L’intervention devient ainsi plus adaptée aux spécificités de chaque patient et elle atteint un taux de succès procédural de 95%. Cependant, à l’heure actuelle les zones de DST sont identifiées de manière visuelle par le spécialiste pratiquant l’ablation. Cette thèse vise à identifier automatiquement les sites potentiels d’ablation basée sur la DST à l’aide de techniques d’apprentissage statistique et notamment d’apprentissage profond adaptées. Dans la première partie, les enregistrements EGM sont classés par catégorie en DST vs. non-DST. Cependant, le rapport très déséquilibré entre les données issues des deux classes dégrade les résultats de classification. Nous abordons ce problème en utilisant des techniques d’augmentation de données adaptées à la problématique médicale et qui permettent d’obtenir de bons taux de classification. La performance globale s’élève ainsi atteignant des valeurs de précision et d’aire sous la courbe ROC autour de 90%. Deux approches sont ensuite comparées, l’ingénierie des caractéristiques et l’extraction automatique de ces caractéristiques par apprentissage statistique à partir d’une série temporelle, appelée valeur absolue de tension maximale aux branches du PentRay (VAVp). Les résultats montrent que la classification supervisée de VAVp est prometteuse avec des valeurs de précision, sensibilité et spécificité autour de 90%. Ensuite, la classification des enregistrements EGM bruts est effectuée à l’aide de plusieurs outils d’apprentissage statistique. Une première approche consiste à étudier les circuits arithmétiques à convolution pour leur intérêt théorique prometteur, mais les expériences sur des données synthétiques sont infructueuses. Enfin, nous investiguons des outils d’apprentissage supervisé plus conventionnels comme les réseaux de neurones convolutifs (RNC). Nous concevons une sélection de représentation des données adaptées à différents algorithmes de classification. Ces modèles sont ensuite évalués en termes de performance et coût de calcul. L’apprentissage profond par transfert est aussi étudié. La meilleure performance est obtenue avec un RNC peu profond pour la classification des matrices EGM brutes, atteignant 94% de précision et d’aire sous la courbe ROC en plus d’un score F1 de 60%. Dans la deuxième partie, les enregistrements EGM acquis pendant la cartographie sont étiquetés ablatés vs. non-ablatés en fonction de leur proximité par rapport aux sites d’ablation, puis classés dans les mêmes catégories. Les annotations de dispersion sont aussi prises en compte comme une probabilité à priori dans la classification. La meilleure performance représente un score F1 de 76%. L’agrégation de l’étiquette DST ne permet pas d’améliorer les performances du modèle. Globalement, ce travail fait partie des premières tentatives d’application de l’analyse statistique et d’outils d’apprentissage pour l’identification automatique et réussie des zones d’ablation en se basant sur la DST. En fournissant aux cardiologues interventionnels un outil intelligent, objectif et déployé en temps réel qui permet la caractérisation de la dispersion spatiotemporelle, notre solution permet d’améliorer potentiellement l’efficacité de la thérapie personnalisée d’ablation par cathéter de la FA persistante
Catheter ablation is increasingly used to treat atrial fibrillation (AF), the most common sustained cardiac arrhythmia encountered in clinical practice. A recent patient-tailored AF ablation therapy, giving 95% of procedural success rate, is based on the use of a multipolar mapping catheter called PentaRay. It targets areas of spatiotemporal dispersion (STD) in the atria as potential AF drivers. STD stands for a delay of the cardiac activation observed in intracardiac electrograms (EGMs) across contiguous leads.In practice, interventional cardiologists localize STD sites visually using the PentaRay multipolar mapping catheter. This thesis aims to automatically characterize and identify ablation sites in STD-based ablation of persistent AF using machine learning (ML) including deep learning (DL) techniques. In the first part, EGM recordings are classified into STD vs. non-STD groups. However, highly imbalanced dataset ratio hampers the classification performance. We tackle this issue by using adapted data augmentation techniques that help achieve good classification. The overall performance is high with values of accuracy and AUC around 90%. First, two approaches are benchmarked, feature engineering and automatic feature extraction from a time series, called maximal voltage absolute values at any of the bipoles (VAVp). Statistical features are extracted and fed to ML classifiers but no important dissimilarity is obtained between STD and non-STD categories. Results show that the supervised classification of raw VAVp time series itself into the same categories is promising with values of accuracy, AUC, sensi-tivity and specificity around 90%. Second, the classification of raw multichannel EGM recordings is performed. Shallow convolutional arithmetic circuits are investigated for their promising theoretical interest but experimental results on synthetic data are unsuccessful. Then, we move forward to more conventional supervised ML tools. We design a selection of data representations adapted to different ML and DL models, and benchmark their performance in terms of classification and computational cost. Transfer learning is also assessed. The best performance is achieved with a convolutional neural network (CNN) model for classifying raw EGM matrices. The average performance over cross-validation reaches 94% of accuracy and AUC added to an F1-score of 60%. In the second part, EGM recordings acquired during mapping are labeled ablated vs. non-ablated according to their proximity to the ablation sites then classified into the same categories. STD labels, previously defined by interventional cardiologists at the ablation procedure, are also aggregated as a prior probability in the classification task.Classification results on the test set show that a shallow CNN gives the best performance with an F1-score of 76%. Aggregating STD label does not help improve the model’s performance. Overall, this work is among the first attempts at the application of statistical analysis and ML tools to automatically identify successful ablation areas in STD-based ablation. By providing interventional cardiologists with a real-time objective measure of STD, the proposed solution offers the potential to improve the efficiency and effectiveness of this fully patient-tailored catheter ablation approach for treating persistent AF

Diffusion Tensor Imaging (DTI) is a magnetic resonance imaging (MRI)-based technique that has frequently been used for the identification of brain biomarkers of neurodevelopmental and neurodegenerative disorders because of its ability to assess the structural organization of brain tissue. In this work, I present (1) preclinical findings of a longitudinal DTI study that investigated asymptomatic high school football athletes who experienced repetitive head impact and (2) an automated pipeline for assessing the quality of DTI images that uses a convolutional neural network (CNN) and transfer learning. The first section addresses the effects of repetitive subconcussive head trauma on the white matter of adolescent brains. Significant concerns exist regarding sub-concussive injury in football since many studies have reported that repetitive blows to the head may change the microstructure of white matter. This is more problematic in youth-aged athletes whose white matter is still developing. Using DTI and head impact monitoring sensors, regions of significantly altered white matter were identified and within-season effects of impact exposure were characterized by identifying the volume of regions showing significant changes for each individual. The second section presents a novel pipeline for DTI quality control (QC). The complex nature and long acquisition time associated with DTI make it susceptible to artifacts that often result in inferior diagnostic image quality. We propose an automated QC algorithm based on a deep convolutional neural network (DCNN). Adaptation of transfer learning makes it possible to train a DCNN with a relatively small dataset in a short time. The QA algorithm detects not only motion- or gradient-related artifacts, but also various erroneous acquisitions, including images with regional signal loss or those that have been incorrectly imaged or reconstructed.

Industry 4.0 promotes the use of emergent technologies, such as Internet of Things (IoT), Big Data, Artificial Intelligence (AI) and cloud computing, sustained by cyber-physical systems to reach smart factories. The idea is to decentralize the production systems and allow to reach monitoring, adaptation and optimization to be made in real time, based on the large amount of data available at shop floor that feed the use of machine learning techniques. This technological revolution will bring significant productivity gains, resources savings and reduced maintenance costs, as machines will have information to operate more efficiently, adaptable and following demand fluctuations. This thesis discusses the application of supervised Machine Learning (ML) techniques allied with artificial vision, to implement an intelligent, collaborative and adaptive robotic inspection station, which carries out the Quality Control (QC) of Human Machine Interface (HMI) consoles, equipped with pressure buttons and Liquid Crystal Display (LCD) displays. Machine learning techniques were applied for the recognition of the operator’s face, to classify the type of HMI console to be inspected, to classify the state condition of the pressure buttons and detect anomalies in the LCD displays. The developed solution reaches promising results, with almost 100% accuracy in the correct classification of the consoles and anomalies in the pressure buttons, and also high values in the detection of defects in the LCD displays.
Indústria 4.0 promove o uso de tecnologias emergentes, como Internet of Things (IoT), Big Data, artificial intelligence (AI) e cloud computing, sustentadas por sistemas ciberfísicos, como o designio de alcançar o que chamam de fábricas inteligentes. A ideia é descentralizar os sistemas de produção e permitir que a monotorização, a adaptação e a otimização sejam feitos em tempo real, com base na grande quantidade de dados disponíveis no ambiente fabril que alimentam o uso de técnicas de machine learning (ML). Esta revolução tecnológica trará ganhos significativos de produtividade, economia de recursos e custos de manutenção mais reduzidos, pois as máquinas terão informações para operar com mais eficiência, adaptáveis e acompanhar as flutuações de procura. Esta tese discute a aplicação de técnicas supervisionadas de ML, aliadas à visão artificial, para a implementação de uma estação de inspeção robótica inteligente, colaborativa e adaptativa, que realiza o controlo de qualidade de consolas HMI, equipados com botões de pressão e displays LCD. Técnicas de ML foram aplicadas para o reconhecimento facial do operador, para classificação do tipo de console HMI a ser inspecionado, para classificar a condição do estado dos botões de pressão e deteção de anomalias nos displays LCD. A solução desenvolvida alcança resultados promissores, com quase 100 % de precisão na correta classificação das consolas e anomalias nos botões de pressão, e também valores elevados de acerto na deteção de defeitos nos displays LCD.

碩士
國立交通大學
電信工程研究所
103
The growing interests in multi-way or multi-channel data analysis have made the tensor factorization and classification a crucial issue in the areas of signal processing and machine learning. Conventionally, the neural network (NN) classifier is estimated from a set of input vectors or one-way observations. The multi-way observations are unfolded as the high-dimensional vectors for model training. As a result, the classification performance is constrained because the correlation or neighboring information in temporal or spatial domains among different ways is lost in the trained NN classifier. More parameters are required to learn the complicated data structure from multiple ways, trials or channels. This study presents a new tensor classification network (TCN) which combines tensor factorization and NN classification for multi-way feature extraction and classification. The proposed TCN can be viewed as a generalization of NN classifier for multi-way data classification where Tucker decomposition and nonlinear operation are performed in each hidden unit. Using this approach, the affine transformation in conventional NN is replaced by the tensor transformation. We generalize from vector-based NN classifier to tensor-based TCN where the multi-way information in temporal, spatial or other domains is preserved through projecting the input tensors into latent tensors. The projection over tensor spaces is efficiently characterized so that a very compact classifier could be achieved. The proposed TCN does not only construct a compact model but also reduce the computation time in comparison with the traditional NN classifier. The tensor error backpropagation algorithm is developed to efficiently establish a tensor neural network. Experimental results on image recognition over different datasets demonstrate that TCN could attain comparable or even better classification performance but with very few parameters and the reduced computation cost when compared with the traditional NN classifier.

碩士
國立高雄應用科技大學
電機工程系碩士班
93
The purpose of this thesis is to improve efficiency and quality for traditional steel bar factory. The factors of influencing effect and quality in manufacturing process are bottleneck of manufacturing process and scheduling process, bottleneck of materiel transmission, non optimum in control system, and fluctuating parameter as a result of alteration in manufacturing process and more. To renew entire apparatus costs a great deal. To renew entire apparatus costs a great deal. Particularly in Small and Medium Enterprises, it is a burden. Only gradual, partial rationalization and optimum can achieve at optimum efficiency and minimum spending. Based on that, firstly, we undergo rationalization, information over the internet and e-type in manufacturing process to find optimal information flow, scheduling flow and material transmission flow. And depending on establishing Ethernet control system in manufacturing process, we get e-type network for the keys bottleneck. Then the control engineers can input appropriate parameters in steel bar manufacturing process quickly and correctly through PLC/PC network. Furthermore, we use monitor system on remote controlling and monitoring for keys bottleneck in all manufacturing process. In order to reduce manpower expenses, enhance work efficiency, reduce the cost and enhance network integrative ability, we integrate expert system of production management and database aiming at management such as production, quality, scheduling process, workers etc,. We finally combine three major objectives--the foregoing keys system, monitoring system, and management system and apply to the cooperative factory. Through actual operation, revision and improvement, we get the practical result of e - type procreation process and remote monitor. The next, we are improving response and stability in manufacturing processes controller. In the part stressed optimum control in steel bar roller mill. Steel bar is a requirement in a building, and its quality affects dependability and stability of the building. How to ensure stable quality of procreation process, enhance efficiency and increase competitive edge is one of important topics nowadays. Steel pulling force and tensile force are influenced by tension control of a roller mill. So the focus is how to maintain constant in tension control. In a traditional roller mill tension controller always uses PID or linear quadratic. This method requires accurate math function, but it is just hard to get accurate math function in the steel bar manufacturing processes. So the controller depends on workers to adjust parameter in a traditional roller mill but the adjusted parameter maybe not the best. Our purpose is to add neural network into inherent PID controller loop in order to accelerate the system response and enhance product quality.

碩士
國立臺灣科技大學
高分子系
96
Abstract This thesis forces on diagnosing the fault tension in the melting spinning processing. We use the material of polypropylene (PP) in experiments and there are ten processing parameters, including three section extruder barrel temperatures, die temperature, metering pump temperature, spinning temperature, metering pump speed, the formation speed, cooling air speed and take-up velocity. First we find the best processing parameters to give the smaller tension variance by using the Taguchi method and the analysis of variance (ANOVA). Then, we measure the spinline tension of every factor in different processing parameter conditions, ten times for each condition. If the tension variance of spinline have three times over the normal range in ten experiments, we consider this tension is abnormal; Otherwise it is normal. In addition, we use statistical processing control (SPC) to choose three feature values and the back-propagation neural network (BPNN) to classify the fault processing parameters. The features include the average distance (RDIST), ALSLSC (Area between the pattern and least square line per least square line crossover expressed in terms of standard deviation) and skewness. The output layer has seven significance factors. The recognition rate can reach 100%. In order to complete the diagnosis system, we present procedures for classifying two factors in abnormal conditions we identify one of the two factors using the classifier for single factor, and locate the other with elimination of the identified factor from the neural network. The experiment results show that the proposed method can successfully classify the fault processing parameters in the melt spinning machines.

碩士
朝陽科技大學
營建工程系碩士班
96
This research is concentrated on the tensile strength of cold-formed steel members. By using the experimental data from the previous studies, the Artificial Neural Networks simulation was adopted for the prediction of thesile strength. The tested ultimate strengths of cold-formed steel angle sections were mainly used in the application of analysis, and the tested values of cold-formed steel channel sections were also utilized to verify the prediction of proposed process. The process of Back Propagation Network was selected for analysis. This network consists of three layers, input layer, hidden layer, and output layer. By applying the Matlab NN toolbox BPN structure, input layer has four neurons for input parameters, and output layer has only one output variable. It was found that Wu/Wc、 /L、An and Fu, these four input parameters provided better prediction as compared to the target values (tested ultimate strengths). The output variable is assigned to be Pult, as expected. Although the solution of Artificial Neural Networks does not have a physical model, neural network process can deal with the problem with complexity and nonlinearity characteristics. By applying the appropriate parameter inputting and network structure, the output can map the real world results very well. It was found that the predictions computed by using neural network can make a good agreement with the tested strengths of cold-formed steel tension members in this study.

碩士
國立臺灣科技大學
材料科學與工程系
101
The polymer film generally is transported by rollers in production and processing. Its nonuniform tension and oblique movement often downgrade the film itself and its products, such as deformation, sags, wrinkles, rail, poor pasting, uneven edge of slitting, wind up in irregularity, crimping, and slant on painting. Thus, it is prerequisite to maintain uniform tension and film alignment before the film enters a processing section or is in rewinding. This study aims to improve nonuniform tension and oblique movement of polymer films simultaneously. We plan to build a three-rollers and two-span setup in experiment to simulate a PET film rewinding system, mainly including a unwinder, rewinder, tension sensor, guiding roller and CCD(Charge Coupled Device). The neural network learning algorithm is applied to design the neural network controller. Since nonuniform tension and oblique movement may interact, their mathematical models are extremely difficult to derive. The neural network controller does not need a complicated mathematical model, and it’s based on the error of the actual output and the target output to adjust the weights. The tension data are acquired by the tension sensor and the lateral position of film is detected by CCD and image processing. The controller outputs a control signal to regulate the speed of the roller, driven by servo motors, to achieve uniform tension. The controller also controls a stepping motor to move the guiding roller for correcting the lateral deflection and aligning the moving film. The results reveal that the proposed controller is effective for control of the uniform tension and oblique guiding.

碩士
國立臺北科技大學
化學工程研究所
95
Thermodynamic data are required in the application of different chemical fields. Because the data is difficult to find so there is a few inconvenient for use. In the computer aided design field, thermodynamic data is collected and fitted into some equations. Those fitting equations are more convenient for use. In general, the thermodynamic data of different chemical substance are fitted into different correlation equations. If one generalized correlation equation can be used to describe all substances, it can increase the convenience for use. In our research, feedforward neural network is utilized to fit the surface tension, saturated vapor pressure and saturated densities data of twenty nonpolar fluids. Three individual generalized correlation equations are determined for these three kinds of thermodynamic qualities. The absolute average deviation of the generalized neural network of nonpolar fluid surface tension is 0.43085%. The absolute average deviation of the generalized neural network of nonpolar fluid saturated vapor pressure is 0.4361%. The absolute average deviation of the generalized neural network of nonpolar fluid saturated densities is 0.10093%. In result, the generalized artificial neural network of the research is practicable.

碩士
元智大學
工業工程與管理學系
96
Control chart pattern recognition is an important work in statistical process control. A control chart may present several unnatural patterns which including trends, sudden shifts, mixtures, and cyclic patterns. The occurrence of unnatural patterns implies that the process is affected by assignable causes, and corrective actions should be taken. Actually, the types of unnatural patterns which exist in real process are comprehensive. Identification correctly of unnatural patterns can greatly narrow the set of possible causes that must be investigated, and thus the diagnostic search time could be reduced in length. The purpose of this research was to develop two classifiers based on support vector machine (SVM) and artificial neural network (ANN) to classify unnatural patterns in yarn tension data. First, we apply some statistical features to extract distinguished features from raw data. The extracted features are used as the components of the input vectors. Secondly, we develop SVM-based and ANN-based classifiers for control chart pattern recognition. The performances of two recognizers using statistical features extracted from correlation coefficient as the components of the input vectors was investigated and compared. The results show that the SVM and ANN have similar recognition performances. Extensive comparisons indicate that the proposed recognizers perform better than that using raw data as inputs. Our research concluded that the extracted statistical features can reduce the input vectors while maintaining good levels of accuracy.

In this study, to increase inference speed of convolutional neural networks, the optimization technique low-rank tensor decomposition has been implemented and applied to AlexNet which had been trained to classify dog breeds. Due to a small training set, transfer learning was used in order to be able to classify dog breeds. The purpose of the study is to investigate how effective low-rank tensor decomposition is when the training set is limited. The results obtained from this study, compared to a previous study, indicate that there is a strong relationship between the effects of the tensor decomposition and how much available training data exists. A significant speed up can be obtained in the different convolutional layers using tensor decomposition. However, since there is a need to retrain the network after the decomposition and due to the limited dataset there is a slight decrease in accuracy.
För att öka inferenshastigheten hos faltningssnätverk, har i denna studie optimeringstekniken low-rank tensor decomposition implementerats och applicerats på AlexNet, som har tränats för att klassificera hundraser. På grund av en begränsad mängd träningsdata användes transfer learning för uppgiften. Syftet med studien är att undersöka hur effektiv low-rank tensor decomposition är när träningsdatan är begränsad. Jämfört med resultaten från en tidigare studie visar resultaten från denna studie att det finns ett starkt samband mellan effekterna av low-rank tensor decomposition och hur mycket tillgänglig träningsdata som finns. En signifikant hastighetsökning kan uppnås i de olika faltningslagren med hjälp av low-rank tensor decomposition. Eftersom det finns ett behov av att träna om nätverket efter dekompositionen och på grund av den begränsade mängden data så uppnås hastighetsökningen dock på bekostnad av en viss minskning i precisionen för modellen.

碩士
華梵大學
機電工程學系博碩專班
98
As growing demands on machining accuracy and miniaturization in the variety of manufacturing industries, micro wire-EDM has been one of key technologies for micromachining. However, wire bending and vibration are prone to occur due to unexpected disturbance of wire transport system, discharge forces and flushing pressure during machining and thereby leading to machining inaccuracy. Therefore, wire tension should be controlled tightly when the wire transport system is subject to undesired disturbance for the geometrical accuracy of micro wire-EDM. This paper presents a combined self-organizing fuzzy controller (SOFC) and neural-network-based disturbance observer (NNBDOB) for wire tension control. A neural-network-based disturbance observer is proposed to compensate for undesired and unknown disturbance. This intelligent control strategy has on-line learning capability and disturbance rejection robustness to suppress the variation of wire tension during the micro wire-EDM process. Experimental results not only demonstrate that the proposed controller can achieve smaller root mean square error (RMSE) than a PID controller but also indicate that the geometrical error of corner cutting and the geometrical accuracy of slit width can be significantly improved.

Orientação: Lénio Ribeiro ; co-orientação: António Martinho
A pressão arterial (PA) é um dos parâmetros vitais mais utilizados a nível mundial, tanto na monitorização anestésica, como em unidades de cuidados intensivos. Permite avaliar o estado hemodinâmico do paciente e tem uma grande importância no diagnóstico e prevenção de doenças cardiovasculares. O método de medição mais utilizado é o método oscilométrico, que tem determinadas desvantagens que influenciam a correta leitura da PA. Na tentativa de diminuir o erro associado a este, tentou-se desenvolver um método não invasivo e de leitura contínua, criando um sistema de redes neuronais artificiais (RNA), que utiliza como entradas do sistema as medições do traçado de eletrocardiograma (ECG) e irá gerar valores de PA ajustados, facilitando a monitorização dos pacientes. Os dados foram recolhidos em 54 cães submetidos a procedimentos cirúrgicos, que compreendeu o registo individual de 13 parâmetros de monitorização: peso, frequência cardíaca (FC), pressão arterial sistólica, média e diastólica (PAS, PAM e PAD, respetivamente), amplitude e duração das ondas e segmentos no ECG. Os dados foram submetidos a análise estatística para averiguar o nível de correlação (cc) entre variáveis. O sistema de RNA criado, consiste no conjunto de entradas (peso e os parâmetros do ECG) com maior correlação com as saídas (PAS, PAM e PAD). As variáveis que apresentaram maior correlação com a PA foram o peso (ccPAS=0,43; ccPAM= 0,38; ccPAD= 0,33), a Dur. P (ccPAS=0,18; ccPAM= 0,25; ccPAD= 0,30), o intervalo P-R (ccPAS=0,19; ccPAM= 0,07; ccPAD= 0,08), a Dur. QRS (ccPAS=0,11; ccPAM= 0,09; ccPAD= 0,14), e a Amp. R (ccPAS=0,37; ccPAM= 0,35; ccPAD= 0,44). Foi criada uma RNA de 5 entradas e três saídas, uma arquitetura de 3 camadas, com 12 neurónios em cada uma das camadas ocultas. Esta rede obteve um índice de regressão linear de 0.93 quando comparado com os valores de PA obtidos pelo método oscilométrico. A construção de um método de medição de PA, baseado em RNA, foi possível. Este é um método inovador, nunca tendo sido realizado um estudo com a mesmas características, e permite uma medição da PA de forma não invasiva e contínua, com uma elevada taxa de sucesso.
Blood pressure (BP) is one of the most widely used vital parameters worldwide, both in anesthesia monitoring and in intensive care units. It allows assessing the hemodynamic status of the patient and is of great importance in the diagnosis and prevention of cardiovascular diseases. The most widely used measurement method is the oscillometric method, which has certain disadvantages that influence the correct reading of the BP. In an attempt to reduce the error, we developed a non-invasive and continuous reading method, creating an artificial neural network (ANN) system, which uses the electrocardiogram (ECG) to generate adjusted BP values, facilitating patient monitoring. We collected data in 54 dogs submitted to surgical procedures, wich included the individual recording of 13 monitoring parameters: weight, heart rate (HR), systolic, mean and diastolic blood pressure (SBP, MBP and DBP, respectively), amplitude and duration of waves and segments on the ECG. The data was submitted to statistical analysis to ascertain the level of correlation (cc) between variables. The ANN system created consists of the set of inputs (weight and ECG parameters) with the highest correlation with the outputs (SBP, MBP and DBP). The variables with the highest correlation with BP were weight (ccSBP = 0.43, ccMBP = 0.38, ccDBP = 0.33), Dur. P (ccSBP = 0.18, ccMBP = 0.25, ccDBP = 0.30), the P-R range (ccSBP = 0.19; ccMBP = 0.07; ccDBP = 0.08), Dur. QRS (ccSBP = 0.11, ccMBP = 0.09, ccDBP = 0.14), and Amp. R (ccSBP = 0.37, ccMBP = 0.35, ccDBP = 0.44). A 5-input, 3-output ANN was created, a 3-layer architecture with 12 neurons in each of the hidden layers. We managed to built an innovative method with a linear regression index of 0.93, when compared to the BP values obtained by the oscillometric method. The construction of a method for BP measurement, based on ANN was possible. There isn’t any study performed, with the same characteristics, so this is an innovative method that allows a measurement of BP in a non-invasive and continuous way, with a high success rate.

The aim of this thesis is to review the current state of machine learning in music composition and to train a computer on Beatles' songs using research project Magenta from the Google Brain Team to produce its own music. In order to explore the qualities of the generated music more thoroughly, we restrict our- selves to monophonic melodies only. We train three deep learning models with three different configurations (Basic, Lookback, and Attention) and compare generated results. Even though the generated music is not as interesting as the original Beatles, it is quite likable. According to our analysis based on musically informed metrics, generated melodies differ from the original ones especially in lengths of notes and in pitch differences between consecutive notes. Generated melodies tend to use shorter notes and higher pitch differences. In theoretical background, we cover the most commonly used machine learning algorithms, introduce neural networks and review related work of music generation. 1