Study and application of neural network techniques for detecting and adapting predictive control loops subject to abrupt system faults

• Ano de defesa: 2021
• Autor(a) principal: Cavalcanti, Dayse Maranhão
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Engenharia elétrica; Controle preditivo; Redes neurais (Computação); Tolerância a falha (Engenharia)
• Link de acesso: https://repositorio.udesc.br/handle/UDESC/15304

The use of Model Predictive Control (MPC) on aircraft flight control consists on finding the optimal control sequence to follow a reference trajectory by predicting the aircraft movement in real time. To such extent, the main focus of this work is to influence the behavior of a didactic plant. herefore, it is suggested the study of the system, that is, its structure, performance, restrictions, and requirements, obtaining the nominal and fault models on state-space representation, and design and implementation of a multi-model fault-tolerant controller. From the system model,the MPC predicts the process output over a time horizon or time window. That said, in the initial or zero time, it is performed an optimization process to find the best control signal that takes the output signal to the desired value in a small time period, saving the first value of the control law and observing the system response. Then, the beginning of the horizon moves to the corresponding time of that value, restarting the procedure. The fault tolerance, on the other hand, occurs through the diagnosis and consequent switching to the most appropriate predefined prediction model by an Artificial Neural Network (ANN). Two types of ANNs are studied: Multilayer Perceptron and Fully Connected Cascade. Which can recognize hidden patterns and correlations in raw data through algorithms. Also, a comparison with the classical one-model structure is discussed, and the obtained results show that the approach with multiple models is promising when it is possible to determine the dynamics models and perform a precise fault diagnosis.