Controle inteligente para um pâncreas artificial utilizando redes neurais artificiais

• Ano de defesa: 2021
• Autor(a) principal: Farias, João Lucas Correia Barbosa de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio Grande do Norte; Brasil; UFRN; PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA MECATRÔNICA
• Palavras-chave em Português: Pâncreas artificial; Diabetes Mellitus tipo 1; Controle inteligente não linear; Redes neurais artificiais; Linearização por realimentação
• Link de acesso: https://repositorio.ufrn.br/handle/123456789/33383

Type 1 Diabetes Mellitus is a disease that affects millions of people around the world. Recently, the incredible progress of embedded devices has given rise to proposals of devices that pump insulin subcutaneously, with the purpose of automatically regulating blood glucose level in diabetic patients. This way, the Artificial Pancreas could provide a better quality of life with more autonomy and comfort to the patients. The goal of this work is to design a nonlinear intelligent controller with a radial basis function (RBF) artificial neural network as an uncertainty estimator for an artificial pancreas. The IVP (Identifiable Virtual Patient) model for blood glucose regulation is used to simulate the dynamics of the virtual patient. The virtual patients and meals are randomly generated following normal distributions and the parameters of the patients vary in a sinusoidal way over the course of the simulation. The proposed control approach neither has knowledge of the system dynamics nor is alerted when a patient has a meal. The first controller analyzed was based on the feedback linearization (FBL) technique with a RBF estimator and a projection algorithm, and the second one was based on sliding modes control with a RBF estimator. On the first part of the tests, 200 virtual patients underwent a 7-day, 3-meal per day simulation. The controllers had equivalent performances with worst case scenario resulting in 115,97 mg/dL mean blood glucose and 97,14% of the time in normoglycemic regime. On the second part, 1 patient underwent a 63-day, 3-meal per day simulation with the goal of analyzing the long-term behavior of the controllers. In the worst case scenario, the simulations resulted in 119,20 mg/dL mean blood glucose and 93,67% of the time in normoglycemia. On this part, FBL technique showed better perfomance, suggesting that, in the long run, the projection algorithm provides greater stability in the update of the neural network weight vector. The results indicate that, due to its continuous learning and adaptation abilities, the proposed intelligent controller has proven to be fit for the problem of efficient blood glucose regulation in patients with type 1 diabetes mellitus.