[pt] Esta tese de doutorado relaciona a estimação da diferenciação fracionária, como medida de memória longa, com o intervalo de tempo entre observações contíguas de uma série temporal. Em teoria, o grau de diferenciação é constante em relação à diminuição da freqüência de observação, não importando se para diminuir a freqüência de observação ignore-se as observações intermediárias ou agregue-se as observações temporalmente. Entretanto, para o caso de se obter séries amostradas a uma freqüência mais baixa através de se ignorar observações intermediárias, observamos nesta tese, através de simulações Monte Carlo, um corportamento diverso. Quando se amostra toda n-ésima observação de uma série, n>1, nota-se um considerável vício de estimação do grau de diferenciação (ou parâmetro de memória longa). O viés é em direção de zero, sendo positivo para valores negativos do parâmetro de memória longa e negativo para valores positivos do parâmetro de memória longa, d. Para valores positivos de d, o viés tem natureza aproximadamente quadrática, diminuindo para valores de d próximos de zero ou 0,5 e sendo mais intenso para valores em torno de 0,25. Para valores negativos de d, o viés é tal que a estimativa fica sempre bem próxima de zero, ou seja, é da magnitude de d. Ao considerarmos o efeito de aliasing (em que componentes de período menor que o intervalo de observação são observados como se tivessem freqüências mais baixas) conseguimos fórmulas heurísticas que explicam satisfatoriamente esse vício, produzindo resultados bastante semelhantes ao verificado nas simulações Monte Carlo. Por outro lado, se a diminuição na freqüência de observação é induzida por agregação temporal, não há vício considerável na estimação, como também mostramos atrvés de simulações Monte Carlo. Propõe-se nesta tese ainda uma maneira de melhorar a estimação da memória longa através da combinação de estimativas da série amostrada a diferentes freqüências. Em alguns casos, consegue-se reduções de até 30% no desvio-padrão da estimativa combinada em relação à original, sem causar viés significativo. / [en] This thesis investigates the relationship between the estimation of the fractional integration, as a measure of long memory, and the time interval between observations of a time series. In theory, the fractional integration is invariant to the frequency of observation. However, skip- sampling induces a considerable bias in the estimation, as shown by Monte Carlo simulations. The aliasing effect explains the bias and suggests formulas for it, which yield results very close to the simulated ones. On the other hand, temporal aggregation does not induce relevant bias to the long memory estimation. In addition, a combination of estimates from the same data sampled at different rates is proposed, achieving in some cases reduction of 30% in the root mean squared estimation error. / [es] Esta tesis de doctorado relaciona la estimación de la diferenciación fraccionaria, como medida de memoria prolongada, con el intervalo de tiempo entre observaciones contíguas de una serie de tiempo. En teoría, el grado de diferenciación es constante en relación a la disminución de la frecuencia de observación, sin importar que para disminuir la frecuencia de observación se ignoren las observaciones intermedias o se agreguen observaciones temporalmente. Sin embargo, en esta tesis se observa, a través de simulaciones Monte Carlo, un comportamiento diverso en el caso de obtener series muestreadas a una frecuencia más baja ignorando observaciones intermedias. Cuando se muestrea la n-ésima observación de una serie, n>1, se nota un considerable sesgo de estimación del grado de diferenciación (o parámetro de memoria longa). El sesgo está en dirección de cero, siendo positivo para valores negativos del parámetro de memoria prolongada y negativo para valores positivos del parámetro de memoria prolongada, d. Para valores positivos de d, el sesgo tiene una naturaleza aproximadamente cuadrática, disminuyendo para valores de d próximos de cero o 0,5 y siendo más intenso para valores en torno de 0,25. Para valores negativos de d, el sesgo es tal que la estimativa está siempre bien próxima de cero, o sea, es de la magnitude de d. Al considerar el efecto de aliasing (en que componentes de período menor que el intervalo de observación son observados como se tuvieran frecuencias más bajas) conseguimos fórmulas heurísticas que explican satisfactoriamente ese sesgo, produciendo resultados bastante semejantes a los obtenidos en las simulaciones Monte Carlo. Por otro lado, si la disminución en la frecuencia de observación se induce por agregación temporal, no hay sesgo considerable en la estimación, como también mostramos a través de simulaciones Monte Carlo. Se propone en esta tesis una forma de mejorar la estimación de la memoria prolongada a través de la combinación de estimativas de la serie amostrada a diferentes frecuencias. En algunos casos, se consiguen reducciones de hasta 30% en la desviación estándar de la estimativa combinada en relación a la original, sin causar sesgo significativo.

Dans cette thèse, nous proposons des techniques de gestion de données pour économiser l’énergie dans les réseaux de capteurs périodiques basés sur l’architecture de clustering. Premièrement, nous proposons d’adapter le taux d’échantillonnage du capteur à la dynamique de la condition surveillée en utilisant le modèle de one-way ANOVA et des tests statistiques (Fisher, Tukey et Bartlett), tout en prenant en compte l’énergie résiduelle du capteur. Le deuxième objectif est d’éliminer les données redondantes générées dans chaque cluster. Au niveau du capteur, chaque capteur cherche la similarité entre les données collectées à chaque période et entre des périodes successives, en utilisant des fonctions de similarité. Au niveau du CH, nous utilisons des fonctions de distance pour permettre CH d’éliminer les ensembles de données redondantes générées par les nœuds voisins. Enfin, nous proposons deux stratégies actif/inactif pour ordonnancer les capteurs dans chaque cluster, après avoir cherché la corrélation spatio-temporelle entre les capteurs. La première stratégie est basée sur le problème de couverture des ensembles tandis que la seconde prend avantages du degré de corrélation et les énergies résiduelles de capteurs pour ordonnancer les nœuds dans chaque cluster. Pour évaluer la performance des techniques proposées, des simulations sur des données de capteurs réelles ont été menées. La performance a été analysée selon la consommation d’énergie, la latence et l’exactitude des données, et la couverture, tout en montrant comment nos techniques peuvent améliorer considérablement les performances des réseaux de capteurs. / In this thesis, we propose energy-efficient data management techniques dedicated to periodic sensor networks based on clustering architecture. First, we propose to adapt sensor sampling rate to the changing dynamics of the monitored condition using one-way ANOVA model and statistical tests (Fisher, Tukey and Bartlett), while taking into account the residual energy of sensor. The second objective is to eliminate redundant data generated in each cluster. At the sensor level, each sensor searches the similarity between readings collected at each period and among successive periods, based on the sets similarity functions. At the CH level, we use distance functions to allow CH to eliminate redundant data sets generated by neighboring nodes. Finally, we propose two sleep/active strategies for scheduling sensors in each cluster, after searching the spatio-temporal correlation between sensor nodes. The first strategy uses the set covering problem while the second one takes advantages from the correlation degree and the sensors residual energies for scheduling nodes in the cluster. To evaluate the performance of the proposed techniques, simulations on real sensor data have been conducted. We have analyzed their performances according to energy consumption, data latency and accuracy, and area coverage, and we show how our techniques can significantly improve the performance of sensor networks.

In the thesis there are described the implementations of various types of filters to remove disturbing signals, which often degrade the ECG signal. In particular, it is a zero isoline fluctuations and power network interference. It is used a principle of the Lynn’s linear filters. The individual filters are designed in a recursive and non-recursive implementation. Then there is described a time-varying linear Lynn's filter for removing drift of zero isoline signal. The thesis also includes filters with minimized calculating time of response, by sampling rate conversion method for both interference types. In conclusion there is an experimental study of the filter implementation for ECG signal with false and real interferences.

No description available.

Im Fokus dieser kumulativ angefertigten Dissertation stehen vier methodenorientierte biomechanische Studien, in welchen die potentiellen Fehlerquellen analysiert werden, die beim Einsatz von Inertialsensoren in der biomechanischen Gang- und Laufanalyse auftreten können. In den einzelnen Beiträgen werden die Einflüsse der Inertialsensoraufnahmefrequenz (Studie I) und des Messbereichs der Beschleunigungssensoren (Studie II) auf die kinematischen, kinetischen und räumlich-zeitlichen Parameter systematisch untersucht. Des Weiteren wird sich kritisch mit der Genauigkeit verschiedener Detektionsmethoden des initialen Bodenkontaktes (Studie III) sowie mit der Aussagekraft der maximalen Eversionsgeschwindigkeit (Studie IV) auseinandergesetzt. Um ein umfassendes Bild der Einflussgrößen zu erhalten, wurde in den Studien II, III und IV untersucht, ob die Materialcharakteristik der Laufschuhsohle die Genauigkeit der biomechanischen Parameter beeinflusst. Zudem wurde in Studie III geprüft, welchen zusätzlichen Effekt der Laufstil (Vor- und Rückfußlaufen) auf die Genauigkeit der initialen Bodenkontaktbestimmung hat sowie welchen Einfluss die Bewegungsgeschwindigkeit (Gehen und Laufen) auf die maximale Eversionsgeschwindigkeit nehmen kann (Studie IV). Die Ergebnisse der vier Untersuchungen werden am Ende dieser Arbeit in einem gemeinsamen Kontext diskutiert. Auf Grundlage der Erkenntnisse konnte eine Übersicht erstellt werden, welche sowohl die Mindestanforderungen an Inertialsensoren als auch die Einflussgrößen auf die Genauigkeit der biomechanischen Parameter enthält. Mit diesem Überblick erhalten Nutzer von Inertialsensoren (z.B. Sportler, Trainer, Mediziner und Wissenschaftler) bei der Planung einer Bewegungsanalyse die Unterstützung, die Sensoren mit der passenden Sensorspezifikation in Kombination mit den präzisesten Auswertealgorithmen auszuwählen. Zudem können die Informationen aus dieser Dissertation dazu genutzt werden, Erkenntnisse bereits publizierter Studien kritisch zu hinterfragen. / In previous studies, inertial sensors were used to investigate kinematic, kinetic, and spatio-temporal parameters during walking and running. The present cumulative doctoral thesis consists of four methodological studies. Two of the studies examine the influence of inertial sensor sampling rate (study I) and accelerometer operating range (study II) on the accuracy of biomechanical parameters. Another study investigated whether different published foot strike detection methods can accurately detect the time of initial ground contact (study III). The final study examined whether a single gyroscope can be used to accurately determine peak eversion velocity (study IV). In order to obtain a comprehensive view of the influencing factors, studies II, III and IV also investigated whether the material characteristics of the running shoe sole also influence the accuracy of the biomechanical parameters. Additionally, the effect of running style (forefoot or rearfoot) on the accuracy of foot strike detection methods was investigated in study III, and the effect of locomotion speed (walking, running slow up to running fast) on the accuracy of peak eversion velocity was examined in study IV. The results of the four investigations will be summarized and discussed in a common context. Based on the findings, an overview was prepared which contains both the minimum requirements for inertial sensors and also the influencing variables on the accuracy of the biomechanical parameters. This overview may assist users of inertial sensors (e.g. athletes, trainers, physicians, or scientists) in planning gait and running analyses to select inertial sensors with the appropriate specification in combination with the most accurate algorithms. In addition, the information from this dissertation can be used to critically consider the findings of published studies.

Recently, Deep Learning has revolutionized many fields, where one such area is Voice Activity Detection (VAD). This is of great interest to sectors of society concerned with detecting speech in sound signals. One such sector is the police, where criminal investigations regularly involve analysis of audio material. Convolutional Neural Networks (CNN) have recently become the state-of-the-art method of detecting speech in audio. But so far, understanding the impact of noise and sampling rates on such methods remains incomplete. Additionally, there are evaluation metrics from neighboring fields that remain unintegrated into VAD. We trained on four different sampling rates and found that changing the sampling rate could have dramatic effects on the results. As such, we recommend explicitly evaluating CNN-based VAD systems on pertinent sampling rates. Further, with increasing amounts of white Gaussian noise, we observed better performance by increasing the capacity of our Gated Recurrent Unit (GRU). Finally, we discuss how careful consideration is necessary when choosing a main evaluation metric, leading us to recommend Polyphonic Sound Detection Score (PSDS).