Publication: Implementación hardware de una red neuronal
Loading...
Official URL
Full text at PDC
Publication Date
2020
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
En la actualidad, para poder tener una calidad de vida aceptable los enfermos de diabetes mellitus deben controlar su glucemia, es decir la concentración de glucosa presente en la sangre. El tratamiento para controlar la glucemia se basa en una alimentación equilibrada, actividad física, hábitos generales de higiene, patrones de sueño saludables, medicación, controles periódicos por el endocrinólogo, seguimiento y control diario de los niveles de concentración de glucosa en sangre y suministro diario de insulina. La insulina tarda en hacer efecto, por lo tanto, lo ideal es conocer cómo va a ser la evolución glucémica del paciente para poder suministrar la insulina con tiempo suficiente como para evitar las hiperglucemias, niveles elevados de glucosa en sangre, o las hipoglucemias, bajos niveles de glucosa en sangre. Conocer con antelación cuál va a ser el índice glucémico le permite al paciente controlarlo mejor modificando las cantidades de insulina que se inyecta.
El objetivo principal de este Trabajo de Fin de Grado es el estudio e implementación en hardware de redes neuronales profundas para realizar la predicción de la glucemia a 30 minutos vista. En concreto, se van emplear las redes neuronales LSTM que proporcionan predicciones de mayor precisión frente a las obtenidas con otras técnicas clásicas, sobre todo en lo que a series temporales de datos se refiere. En el estudio de estos modelos se toma como entrada diversos parámetros como son: la glucemia, hidratos de carbono e insulina en sangre. En nuestro dataset, por cada registro de las características, se tiene un histórico desde dos horas atrás hasta la actual. En este trabajo primero explicamos el estudio que se ha realizado en Python para decidir la arquitectura de la red neuronal que se implementará posteriormente en hardware. Esta arquitectura incluye el número capas, el tipo de estas y el número de neuronas que las componen.
A continuación presentamos el diseño e implementación hardware. Primero se explican los aspectos generales del diseño, como los protocolos de comunicación o la representación de los datos utilizados, y posteriormente se detalla el diseño siguiendo una aproximación jerárquica top-down, empezando por la interfaz del sistema y finalizando con los módulos más internos.
En el último capítulo se expone tanto los resultados de sintetizar el diseño hardware sobre una FPGA Xilinx Virtex-6 ML-605 Evaluation Board, como los resultados experimentales de las pruebas realizadas a la red neuronal hardware.
Nowadays patients diagnosed with Diabetes Mellitus must control their glycaemic levels or quantity of glucose in their blood in order to have an acceptable life quality. Glycaemic control treatment is based on: a balanced diet, personal hygiene measures, physical activity, pharmacological-medication, periodic meetings with and endocrinologist, strict control of glycaemic levels (daily, every 8 hours, or if necessary, every hour), insulin administration. Taking in account that insulin takes some time to produce its effects, it would be helpful to know or predict an individual glycaemic level curve in order to prevent hyper (excess) or hypo (deficient) glycaemia. A patient who knows his own glycaemic behaviour in advance can take a better control of his disease by administering himself a higher o lower quantity of insulin, just as much as he needs. The main goal of this study is the analysis and implementation of deep neural network in order to predict glycaemic levels 30 minutes in advance. More specifically, LSTM neural networks will be used in this study as they lead to more precise predictions in comparison with others classic techniques, specially to time series data. In this study, it has taken as an entry diverse parameter such: the blood sugar, carbon hydrates and insulin in blood. In our dataset, for each register of the characteristics, it has a historical record from two hours before up to the present. The research carried out in Python is explained to decide the neural network architecture that it is going to be implemented subsequently in the hardware. This architecture includes the number and the type of layers, the number of neurons that compose them. Within the hardware design and implementation, we first explain the general aspects of the design, such as the communication protocols or the representation of the data used; and then details of the design in a top-down hierarchical approach are given, starting with the interface with the outer module and ending with the more internal modules. The last chapter presents both the results of synthesizing the hardware design on a Xilinx Virtex-6 ML-605 Evaluation Board FPGA, and the experimental results of the hardware neural network tests.
Nowadays patients diagnosed with Diabetes Mellitus must control their glycaemic levels or quantity of glucose in their blood in order to have an acceptable life quality. Glycaemic control treatment is based on: a balanced diet, personal hygiene measures, physical activity, pharmacological-medication, periodic meetings with and endocrinologist, strict control of glycaemic levels (daily, every 8 hours, or if necessary, every hour), insulin administration. Taking in account that insulin takes some time to produce its effects, it would be helpful to know or predict an individual glycaemic level curve in order to prevent hyper (excess) or hypo (deficient) glycaemia. A patient who knows his own glycaemic behaviour in advance can take a better control of his disease by administering himself a higher o lower quantity of insulin, just as much as he needs. The main goal of this study is the analysis and implementation of deep neural network in order to predict glycaemic levels 30 minutes in advance. More specifically, LSTM neural networks will be used in this study as they lead to more precise predictions in comparison with others classic techniques, specially to time series data. In this study, it has taken as an entry diverse parameter such: the blood sugar, carbon hydrates and insulin in blood. In our dataset, for each register of the characteristics, it has a historical record from two hours before up to the present. The research carried out in Python is explained to decide the neural network architecture that it is going to be implemented subsequently in the hardware. This architecture includes the number and the type of layers, the number of neurons that compose them. Within the hardware design and implementation, we first explain the general aspects of the design, such as the communication protocols or the representation of the data used; and then details of the design in a top-down hierarchical approach are given, starting with the interface with the outer module and ending with the more internal modules. The last chapter presents both the results of synthesizing the hardware design on a Xilinx Virtex-6 ML-605 Evaluation Board FPGA, and the experimental results of the hardware neural network tests.
Description
Trabajo de Fin de Grado en Ingeniería de Computadores, Facultad de Informática UCM, Departamento de Arquitectura de Computadores y Automática. Curso 2019/2020