Detecção e isolamento de falhas em um sistema de tanques acoplados utilizando redes neurais artificias

• Ano de defesa: 2020
• Autor(a) principal: Duarte, Mateus Ânderson Barreto
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal Rural do Semi-Árido; Brasil; Centro de Engenharias - CE; UFERSA; Programa de Pós-Graduação em Engenharia Elétrica
• Palavras-chave em Português: Detecção e Isolamento de Falhas; Identificação de sistemas; MLP; NNARX; Redes Neurais Artificiais; Fault Detection and Isolation; Systems Identification; Artificial Neural Networks; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: https://repositorio.ufersa.edu.br/handle/prefix/6726

The growing demand for economic and environmental efficiency in the industrial sector, associated with the search for performance, quality, reliability and safety of processes has made fault detection and diagnosis systems increasingly important. Besides, when faults are detected and diagnosed in advance, it is possible to prevent the progression of abnormal events and, consequently, reduce the loss of productivity. In this context, the importance of having a system capable of efficiently detecting and diagnosing such faults is highlighted. Therefore, the present study aims to develop a Fault Detection and Isolation system (FDI), composed of structures that use Artificial Neural Networks (ANNs), which are trained offline by mathematical software. A case study is developed from the mathematical model of a Quanser® coupled tank system. The methodology is divided into two main stages, the first corresponds to the identification of the system, which is made by Neural Network AutoRegressive with eXogenous input (NNARX). While the second stage concerns the development of a classifier of faults based on residual values, for this step, a Neural Network Multilayer Perceptron (MLP) was used. For both stages the neural networks were trained with a backpropagation algorithm. Thus, the DIF system was developed to act, through simulations, in the detection and isolation of thirteen faults (structural, in sensors and actuator). Furthermore, the simulation results proved to be satisfactory for tested faults, since the classifier was able to classify all simulated faults. The three ANNs (ANN 1, ANN 2 and ANN 3) implemented in this dissertation, presented a satisfactory Mean Square Error (MSE) performance index