Publication: Modelado ubicuo del comportamiento para la asistencia de personas mayores en el hogar
Loading...
Identifiers
Publication date
2013-09
Defense date
2013-09-20
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
El envejecimiento de la población está teniendo un impacto muy relevante en
los sistemas de salud en todo el mundo. Debido a esta situación, hoy día existe
un creciente interés en el desarrollo de sistemas automáticos de supervisión y
alarma que faciliten vivir de forma independiente a las personas mayores durante
el mayor tiempo posible. Los sistemas de inteligencia ambiental proporcionan una
manera sencilla de mejorar la sensación de seguridad de las personas que viven de
forma independiente, y permiten a las personas mayores ser más autosuficientes,
fomentando su autonomía.
Unos indicadores muy comunes para evaluar el bienestar físico y cognitivo de
las personas mayores son los cambios o desviaciones en sus Actividades de Vida
Diaria (AVDs) y en sus patrones de comportamiento. El conjunto de actividades
que se conocen como AVDs son aquellas que se llevan a cabo con una periodicidad
diaria y se consideran definitorias del comportamiento general de un individuo.
Dentro de esta definición se incluyen actividades tales como asearse, alimentarse o
bañarse. Recientes desarrollos relativos a las tecnologías de monitorización hacen
posible instrumentar cualquier entorno de un modo sencillo y barato, ofreciendo
un mecanismo para poder controlar en todo momento la situación de una vivienda
real.
En esta tesis doctoral se ha hecho uso de redes de sensores inalámbricos para
registrar el comportamiento de las personas en su propia vivienda. Empleando
sensores digitales se han controlando tanto el movimiento de los usuarios, como
el uso que hacían de electrodomésticos, muebles o puertas. Modelar el comportamiento
humano no es un problema trivial, debido a que los datos generados por los
sensores son generalmente ruidosos y tienen un comportamiento no determinista,
y además cada usuario realiza cada actividad específica de un modo personal y
diferente. Modelos probabilísticos clásicos basados en el algoritmo de estimación
de máxima verosimilitud han demostrado ofrecer un buen comportamiento en este
dominio, sin embargo, dichos modelos son propensos a sobreadaptarse a los datos
y generalmente requieren de la aceptación de suposiciones muy estrictas acerca de la distribución de los atributos.
En esta tesis doctoral, se aborda el problema de modelar el comportamiento humano a través de dos enfoques de aprendizaje diferentes. Se propone un sistema
para el reconocimiento de las AVDs mediante el uso de aprendizaje supervisado,
y un método de detección de anomalías que identifica estadísticamente patrones
de comportamiento atípicos de un modo no supervisado.
Para abordar el problema del reconocimiento de actividades, se presentan dos
propuestas basadas en los modelos ocultos de Markov (HMM), a saber: esquemas
HMM híbridos y un método de inferencia bayesiano para entrenar un HMM estándar.
Un HMM es un modelo probabilístico temporal que puede ser empleado
para reconocer actividades humanas de un modo eficaz, pero que requiere asumir
condiciones de independencia muy estrictas entre los atributos, que muy probablemente
no representan la distribución real de los datos. En este trabajo se muestra
cómo la combinación de las capacidades discriminativas de clasificadores bien
conocidos, como son las redes de neuronas y las máquinas de vectores de soporte,
junto a la capacidad de los HMM de capturar las relaciones temporales de
los datos, ofrece una mejor tasa de reconocimiento y un marco más flexible para
modelar el espacio de características de los sensores.
También se demuestra que las técnicas de simulación estocástica, como el método
Montecarlo basado en cadenas de Markov (MCMC), pueden ser empleadas
para estimar eficazmente los parámetros de un HMM haciendo uso de estadística
bayesiana. Se evitan además los problemas que típicamente acompañan a los
métodos clásicos de inferencia basados en máxima verosimilitud, a saber: son
numéricamente sensibles, dependen fuertemente de las estimaciones iniciales de
los parámetros y su tipo de búsqueda suele conducir a soluciones subóptimas del
problema.
En lo relativo al método propuesto para la detección no supervisada de anomalías
en los patrones del comportamiento, se presenta un algoritmo basado en
estadística bayesiana que se puede emplear para identificar comportamientos atíatípicos
en una persona de forma eficaz. Los patrones de comportamiento de las personas
se modelan estadísticamente mediante tres probabilidades que presentamos
en este trabajo. Dichas probabilidades son definitorias de: (1) cúando se activan
los sensores, (2) en qué orden y (3) durante cuánto tiempo. El método se basa en
una estimación aproximada de las métricas que definen los patrones diarios de las
personas, y en unas probabilidades a priori que reflejan un conocimiento previo
acerca de cómo deben ser dichas métricas. Estas estimaciones se emplean para
detectar signos anómalos en el comportamiento que pudieran indicar un cambio
en la salud del usuario.
Finalmente, se detalla una propuesta para aplicar los métodos de modelado del
comportamiento presentados en múltiples entornos monitorizados, sin necesidad
de obtener datos de entrenamiento de cada entorno particular. Tanto los métodos
supervisados como los no supervisados, cuando se optimizan para una vivienda,
no pueden ser automáticamente empleados para otra vivienda distinta, debido a las diferencias tanto entre el diseño de casas como en el comportamiento de sus
habitantes. En esta tesis se presenta un método de transferencia de información
para los problemas de reconocimiento de actividades y detección de anomalías
que permite emplear el conocimiento adquirido en diferentes viviendas bajo un
enfoque bayesiano, y sirve como base para la instalación de los sistemas en un
entorno monitorizado completamente nuevo, ofreciendo una solución para aplicar
los métodos a mayor escala.
En este trabajo de tesis doctoral, se evalúan las diferentes propuestas usando
tres conjuntos de datos generados por el sistema de monitorización presentado,
y otros tres conjuntos de datos públicos, ampliamente usados en la literatura.
Los resultados experimentales han demostrado cómo los modelos HMM híbridos
pueden reconocer eficazmente las actividades del usuario, mejorando significativamente
tanto a su versión puramente discriminativa como a los modelos generativos.
También se ha demostrado cómo el rendimiento de un HMM puede mejorar
significativamente si se emplea inferencia bayesiana bajo un algoritmo MCMC.
Este método ofrece una búsqueda mucho más completa en el espacio de soluciones,
y permite integrar la estimación de la cadena oculta dentro del proceso
de inferencia. Finalmente, los resultados de los experimentos en transferencia de
información confirman que existe la posibilidad de diseñar métodos que permitan
abstraer las capacidades de detección de un entorno automáticamente, siendo
posible tener algoritmos de modelado que puedan funcionar en meta espacios de
características, y de este modo permitan trasladar el aprendizaje desde un entorno
instrumentado a otro.
Para concluir, esta tesis doctoral confirma que es posible modelar el comportamiento
humano de una forma eficaz mediante un conjunto de sensores inalámbricos
sencillos y baratos, empleando tanto enfoques supervisados como no supervisados. -------------------------------------------------------------------------------------------------------------------------------
Population aging is currently having a significant impact on health care systems. As demographics age and the burden of healthcare on society increases, the need for automated, ambient monitoring and alarming systems becomes more predominant. Automated and ambient systems provide an easy way for increasing the safety and the sense of security of people living on their own, and allow the elderly to be self-reliant longer, fostering their autonomy. Common indicators for assessing the cognitive and physical wellbeing of an elderly person are the changes in the Activities of Daily Living (ADLs) and in the behavior patterns. Activities referred to as ADLs are everyday tasks essential to self-care and independent living, such as cooking, toileting and showering. Recent developments in sensing technology make it possible to inexpensively equip existing homes with sensors, therefore allowing a continuous monitoring system. In this work, we use wireless sensor networks to capture the behavior of the inhabitants in their homes. Digital state-change sensors are used to measure the motion of the inhabitants and the use of appliances, home devices, pieces of furniture and doors. Modeling human behavior from such sensor data is a challenging problem because the data is ambiguous, unsegmented, noisy and also because there are changes in how each person performs a specific action. Probabilistic models using classical, maximum-likelihood estimation methods are known to work well in this domain, but they are prone to overfitting and typically require strong model assumptions, not offering a flexible solution. In this thesis, we address the human behavior modeling problem through the use of two different learning approaches.We propose an ADLs recognition system based on supervised learning and an anomaly detection approach to statistically identify unusual behavior patterns in an unsupervised way. In order to address the activity recognition problem, we present two approaches based on hidden Markov models (HMM), namely: hybrid HMM schemes and a standard HMM trained by full bayesian inference. HMM is a temporal probabilistic model that can be effectively used for recognizing human activities, but requires to apply strong model assumptions which most likely does not represent the true distribution of the data, as the complete independence of every sensor. We showed that the combination of the discriminative capabilities of a machine learning scheme and the superior dynamic time warping abilities of the HMM can offer better recognition performance and provide a more flexible framework to model the sensors feature space. We also show how Markov Chain Monte Carlo (MCMC) techniques can be used to properly estimate the parameters of an HMM in a bayesian configuration, overcoming usual drawbacks posed by classical expectation-maximization algorithms: the use of a recurrent forward-backward algorithm is numerically sensitive, it can get stuck as it approaches a local optima and it can easily overfit the training data. Regarding our unsupervised anomaly detection proposal, we present a method based on bayesian statistics that can be effectively applied for identifying abnormal human behavior. Behavioral patterns of the residents are statistically estimated based on three probabilistic features that we introduce, namely: sensor activation likelihood, sensor sequence likelihood and sensor duration likelihood. The method relies on an approximate estimation of the living patterns of the user, and on prior knowledge that reflects our belief of how such patterns should be, to detect abnormal behavior signs which could reflect changes in health status of the user. Finally, we introduce an approach for applying our behavior modeling methods in multiple homes without the need of having training data from each home. Both supervised or unsupervised methods, when trained for one home, can not automatically be used in another home, due to differences in the layout of the places and the behavior of the inhabitants.We present a transfer learning method that allows us to use data from other homes to recognize ADLs and estimate the living patterns of the user in a new home. This makes it possible to apply the models discussed on a large scale, therefore providing a broadly applicable solution to efficient care giving of elderly. In this thesis, we evaluate our approaches using three datasets recorded by our monitoring systems in real world settings and three datasets publicly available. Experimental results of the hybrid HMM models demonstrate that hybrid schemes can outperform other classical sequential pattern recognition approaches, showing how the combination of discriminative and generative models is more accurate than either of the models on their own. Also, fully bayesian inference using MCMCalgorithms has been proved to be a more flexible and accurate technique to estimate the HMM parameters for our domain. In the anomaly detection domain, results show that abnormal behavior signs can be statistically identified based on several probabilistic features and Bayesian statistics provides a very consistent way to reason under the uncertainty of human behavior. Finally, experiments on transfer learning show how there exists the possibility to devise methods that allow to abstract the sensing capabilities of an environment quite automatically, being possible to have behavior modeling systems operating in a meta space that allows the methods to be translated from one sensing configuration to another. Moreover, the work presented here further demonstrates that accurate human behavior modeling can be achieved by a set of simple and cheap state-change sensors installed in a wireless network, using both supervised and unsupervised approaches.
Population aging is currently having a significant impact on health care systems. As demographics age and the burden of healthcare on society increases, the need for automated, ambient monitoring and alarming systems becomes more predominant. Automated and ambient systems provide an easy way for increasing the safety and the sense of security of people living on their own, and allow the elderly to be self-reliant longer, fostering their autonomy. Common indicators for assessing the cognitive and physical wellbeing of an elderly person are the changes in the Activities of Daily Living (ADLs) and in the behavior patterns. Activities referred to as ADLs are everyday tasks essential to self-care and independent living, such as cooking, toileting and showering. Recent developments in sensing technology make it possible to inexpensively equip existing homes with sensors, therefore allowing a continuous monitoring system. In this work, we use wireless sensor networks to capture the behavior of the inhabitants in their homes. Digital state-change sensors are used to measure the motion of the inhabitants and the use of appliances, home devices, pieces of furniture and doors. Modeling human behavior from such sensor data is a challenging problem because the data is ambiguous, unsegmented, noisy and also because there are changes in how each person performs a specific action. Probabilistic models using classical, maximum-likelihood estimation methods are known to work well in this domain, but they are prone to overfitting and typically require strong model assumptions, not offering a flexible solution. In this thesis, we address the human behavior modeling problem through the use of two different learning approaches.We propose an ADLs recognition system based on supervised learning and an anomaly detection approach to statistically identify unusual behavior patterns in an unsupervised way. In order to address the activity recognition problem, we present two approaches based on hidden Markov models (HMM), namely: hybrid HMM schemes and a standard HMM trained by full bayesian inference. HMM is a temporal probabilistic model that can be effectively used for recognizing human activities, but requires to apply strong model assumptions which most likely does not represent the true distribution of the data, as the complete independence of every sensor. We showed that the combination of the discriminative capabilities of a machine learning scheme and the superior dynamic time warping abilities of the HMM can offer better recognition performance and provide a more flexible framework to model the sensors feature space. We also show how Markov Chain Monte Carlo (MCMC) techniques can be used to properly estimate the parameters of an HMM in a bayesian configuration, overcoming usual drawbacks posed by classical expectation-maximization algorithms: the use of a recurrent forward-backward algorithm is numerically sensitive, it can get stuck as it approaches a local optima and it can easily overfit the training data. Regarding our unsupervised anomaly detection proposal, we present a method based on bayesian statistics that can be effectively applied for identifying abnormal human behavior. Behavioral patterns of the residents are statistically estimated based on three probabilistic features that we introduce, namely: sensor activation likelihood, sensor sequence likelihood and sensor duration likelihood. The method relies on an approximate estimation of the living patterns of the user, and on prior knowledge that reflects our belief of how such patterns should be, to detect abnormal behavior signs which could reflect changes in health status of the user. Finally, we introduce an approach for applying our behavior modeling methods in multiple homes without the need of having training data from each home. Both supervised or unsupervised methods, when trained for one home, can not automatically be used in another home, due to differences in the layout of the places and the behavior of the inhabitants.We present a transfer learning method that allows us to use data from other homes to recognize ADLs and estimate the living patterns of the user in a new home. This makes it possible to apply the models discussed on a large scale, therefore providing a broadly applicable solution to efficient care giving of elderly. In this thesis, we evaluate our approaches using three datasets recorded by our monitoring systems in real world settings and three datasets publicly available. Experimental results of the hybrid HMM models demonstrate that hybrid schemes can outperform other classical sequential pattern recognition approaches, showing how the combination of discriminative and generative models is more accurate than either of the models on their own. Also, fully bayesian inference using MCMCalgorithms has been proved to be a more flexible and accurate technique to estimate the HMM parameters for our domain. In the anomaly detection domain, results show that abnormal behavior signs can be statistically identified based on several probabilistic features and Bayesian statistics provides a very consistent way to reason under the uncertainty of human behavior. Finally, experiments on transfer learning show how there exists the possibility to devise methods that allow to abstract the sensing capabilities of an environment quite automatically, being possible to have behavior modeling systems operating in a meta space that allows the methods to be translated from one sensing configuration to another. Moreover, the work presented here further demonstrates that accurate human behavior modeling can be achieved by a set of simple and cheap state-change sensors installed in a wireless network, using both supervised and unsupervised approaches.
Description
Keywords
Inteligencia ambiental, Ancianos, Asistencia en el hogar, Computación ubicua