Identificação do modelo dinâmico de uma incubadora neonatal mediante a aplicação de uma rede neural convolucional unidimensional

Abstract

This work puts forward the identification of the temperature dynamics of a neonatal incubator with the aid of a one-dimensional convolutional neural network (1D-CNN). The goal is to use the convolutional network in order to carry out a high-dimensional multi-output nonlinear regression which estimates the parameters of a single-input single-output (SISO) dynamic model capable of depicting the temperature’s activity over time. The neural network learns a mapping between the vectors that represent the input-output dynamics of first order plus dead-time (FOPDT) models and their respective parameters. The methodology entails initially obtaining the response signals from thousands of models created via MATLAB® software simulation, applying a step signal into the incubator in order to perform an identification with a traditional optimization method in MATLAB® thus obtaining an aditional dynamic model whose time constant is used to design a pseudo-random binary sequence (PRBS) signal, obtaining a response signal from the incubator via measurement experiment using the PRBS signal as the system input, training the 1D-CNN with the simulation signals, executing the prediction of the physical system’s parameters and lastly testing the identified model’s performance. That test consisted of plotting a graph containing the response signals of the physical system and the identified models (by using the neural network and the optimization method) and of evaluating the mean squared error (MSE) and the error standard deviation. Furthermore, the models created through simulation were also estimated, for the purpose of validating the neural network’s training and assessing the disparity between the results obtained with these models and the ones obtained with the neonatal incubator dynamic model. The MSE value resulting from the proposed dynamic model (CNN-model) was aproximately 0.25, whilst the identified-by-optimization model’s MSE (PRBS-model) was about 2.10.