Localização de vazamentos em tubulações enterradas usando uma rede Nneural de convolução
Carregando...
Data
2021-11-11
Autores
Boaventura, Otavio Duarte Zotelli
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A água um elemento essencial para todas as atividades humanas, porém, diversas redes urbanas de distribuição ao redor do mundo têm enfrentado dificuldades em provê-la à população. No Brasil, em média, 38,5% da água tratada é perdida antes de chegar ao consumidor. Os métodos atuais de detecção de vazamentos são trabalhosos ou requerem muita experiência ou qualificação técnica do operador. O objetivo desse trabalho é desenvolver um modelo de Aprendizado de Máquina, mais especificamente uma Rede Neural de Convolução (RNC), que simplifique o processo de localização de vazamentos em tubulações de água subterrâneas, calculando a distância entre o sensor e o vazamento a partir de medições realizadas na superfície. Para isso, construiu-se uma caixa com placas de madeira compensada de 1,6m x 1,6m x 0,8m preenchida com solo extraído da região de Ilha Solteira - São Paulo. Um tubo de PVC foi inserido a 0,3m de altura da base da caixa atravessando ao meio duas de suas laterais. Acoplou-se ao tubo um atuador eletromecânico para simular a vibração causada pelo vazamento. Foram coletados sinais de aceleração na superfície, variando a intensidade da vibração no tubo, varrendo uma malha de pontos espaçados 20cm entre si. Dos sinais, extraiu-se a densidade espectral e montou-se matrizes de entrada para a RNC com três sensores. Após uma busca aleatória de combinações em um espaço de hiper-parâmetros, são apresentadas as redes que conseguiram os melhores resultados. Algumas são analisadas em mais detalhes, mostrando as regressões com elas realizadas. O melhor modelo alcançou um erro médio absoluto de 0,103mm na distância predita do vazamento.
Water is an essential element for all human activities, but, around the world, several urban distribution networks have struggled to provide it to their population. In Brazil, around 38.5% of all treated water is lost before reach consumers. Current detection methods are laborious or require a very experienced or high qualified operator. The goal of this work is to develop a Machine Learning, more specifically a Convolutional Neural Network (CNN) model to simplify the process of locating leaks in underground water pipes, calculating the distance between sensor and leakage from measurements in the ground surface. For this, a 1.6m x 1.6m x 0.8m plywood box was built and filled with soil extracted from Ilha Solteria’s (SP) surroundings. A PVC pipe was placed inside the box at 0.3m height from its base, crossing the side of the box. An electromechanical actuator was coupled to the pipe to simulate the vibration caused by the leak. Acceleration signals were measured in the soil surface, varying the signal intensity and passing by a mesh of points 20cm apart from each other. From the signal, the Spectral Density was extracted and input matrices for the CNN were built using three sensors for each input. After an extensive hyper-parameter search, models that got the best results are presented. Few are analyzed in detail, showing regressions made by them. The best model reached a mean absolute error of 0, 103mm in the predicted leak distance.
Water is an essential element for all human activities, but, around the world, several urban distribution networks have struggled to provide it to their population. In Brazil, around 38.5% of all treated water is lost before reach consumers. Current detection methods are laborious or require a very experienced or high qualified operator. The goal of this work is to develop a Machine Learning, more specifically a Convolutional Neural Network (CNN) model to simplify the process of locating leaks in underground water pipes, calculating the distance between sensor and leakage from measurements in the ground surface. For this, a 1.6m x 1.6m x 0.8m plywood box was built and filled with soil extracted from Ilha Solteria’s (SP) surroundings. A PVC pipe was placed inside the box at 0.3m height from its base, crossing the side of the box. An electromechanical actuator was coupled to the pipe to simulate the vibration caused by the leak. Acceleration signals were measured in the soil surface, varying the signal intensity and passing by a mesh of points 20cm apart from each other. From the signal, the Spectral Density was extracted and input matrices for the CNN were built using three sensors for each input. After an extensive hyper-parameter search, models that got the best results are presented. Few are analyzed in detail, showing regressions made by them. The best model reached a mean absolute error of 0, 103mm in the predicted leak distance.
Descrição
Palavras-chave
Monitoramento da integridade estrutural, Inteligência artificial, Redes neurais (Computação), Águas vazamento, Leak location, Machine learning, Artificial neural networks, Convolutional neural networks