| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.82 MB | Adobe PDF |
Autores
Resumo(s)
Redes de sensores sem fios consistem num número variado de dispositivos eletrónicos (nós) que realizam monitorização de factores físicos ou ambientais. Dependendo do tipo de sensor acoplado ao nó, este pode executar inúmeras medições, tal como, temperatura, pressão, movimentos, entre outros. Estes dispositivos estarão sempre sujeitos a interferências devido a condições externas ambientais ou humanas, levando os sensores a relatar medições erradas tornando os relatórios não confiáveis. Quando estas medições inesperadas são produzidas, é necessário detectá-las e identificá-las em tempo real. Para tal serão combinadas diversas técnicas de redundância e fusão de dados que, em conjunto, permitirão melhorar a qualidade da monitorização. O trabalho desenvolvido para esta dissertação concentra-se sobre o domínio dos sistemas ciber-físicos e está incluído no projecto AQUAMON. Este projecto tem como objectivo principal o desenvolvimento de uma plataforma de recolha e tratamento de dados de sensores aquáticos. A monitorização contínua de ambientes aquáticos usando sensores de água é importante para diversas aplicações relacionadas com prevenção de acidentes, recursos hídricos e gestão da aquicultura e actividades recreativas. Portanto, é fundamental garantir a qualidade dos dados de monitorização para evitar falsos alarmes ou ignorar eventos relevantes. Este trabalho irá concentrar-se na implementação de uma plataforma de acordo com uma arquitectura previamente definida, denominada de ANNODE. Esta arquitectura utiliza técnicas de Aprendizagem Automática, mais especificamente Redes Neuronais, com o intuito de criar previsões, medir a qualidade destas, detectar falhas e corrigi-las, em ambiente off-line. No entanto o foco deste trabalho, não é apenas a implementação desta arquitectura, mas sim a modificação desta permitindo utilizá-la num ambiente on-line, possibilitando o processamento de medições em tempo-real, mantendo todos os aspectos centrais anteriores. Como será possível ver nos capítulos seguintes, este trabalho obteve resultados positivos e comparando-os com a arquitectura original observaram-se melhorias.
Wireless sensor networks consist of a set of devices equipped with sensors, with processing capabilities, and connected together in a wireless network, which perform the measurement of physical or environmental factors and transmit these measurements to some central node. Depending on the type of sensor attached to the node, it can perform various measurements such as temperature, pressure, velocity, etc. These devices will always be subject to interference due to external conditions, whether environmental or human, which can lead to poor communication between sensor nodes or can result in errors being introduced in measurements. When these unexpected measurements are produced, there is a need to detect and identify them, for which various methods can be implemented. This work will focus on the implementation of a monitoring platform in which the data processing methods for the detection and correction of these errors are based on neural networks. This work is done in the context of the AQUAMON project, which aims to create a dependable monitoring platform for application in aquatic environments. This work will focus on the implementation of a platform according to a previously defined architecture, called ANNODE. This architecture uses Machine Learning techniques, more specifically Neural Networks, in order to predict measurements and use this predictions to detect failures, measure the quality of measurements and correct erroneous measurements. The objective of the work was to implement a platform according to this architecture, therefore, rather than defining new methods for the detection of anomalies in sensor data, this work will exploit previous work done in AQUAMON on the definition of concrete methods for dependable monitoring, implementing them in a real-time platform, evaluating the implementation and proposing possible improvements.
Wireless sensor networks consist of a set of devices equipped with sensors, with processing capabilities, and connected together in a wireless network, which perform the measurement of physical or environmental factors and transmit these measurements to some central node. Depending on the type of sensor attached to the node, it can perform various measurements such as temperature, pressure, velocity, etc. These devices will always be subject to interference due to external conditions, whether environmental or human, which can lead to poor communication between sensor nodes or can result in errors being introduced in measurements. When these unexpected measurements are produced, there is a need to detect and identify them, for which various methods can be implemented. This work will focus on the implementation of a monitoring platform in which the data processing methods for the detection and correction of these errors are based on neural networks. This work is done in the context of the AQUAMON project, which aims to create a dependable monitoring platform for application in aquatic environments. This work will focus on the implementation of a platform according to a previously defined architecture, called ANNODE. This architecture uses Machine Learning techniques, more specifically Neural Networks, in order to predict measurements and use this predictions to detect failures, measure the quality of measurements and correct erroneous measurements. The objective of the work was to implement a platform according to this architecture, therefore, rather than defining new methods for the detection of anomalies in sensor data, this work will exploit previous work done in AQUAMON on the definition of concrete methods for dependable monitoring, implementing them in a real-time platform, evaluating the implementation and proposing possible improvements.
Descrição
Tese de mestrado em Engenharia Informática (Sistemas de Informação), Universidade de Lisboa, Faculdade de Ciências, 2020
Palavras-chave
Redes de sensores Confiabilidade Aprendizagem automática Previsão em tempo-real Redes Neuronais Teses de mestrado - 2020
