Bed posture classification by Neural network and Bayesian network using noninvasive sensors

Name: Waranrach Viriyavit

keyword: Bed classification

LCSH: Older people

; Neural network

; Bayesian network

; Elderly care

; Elderly behavior

Abstract: The elderly population of the world continues to increasing rate. In Thailand, the proportion of elderly people is also growing up to 14.9% in 2014. The national statistical office of Thailand reports that percentage of living alone elderly is rising from 3.6 in 2002 to 10.4 in 2014. Hence, it needs more geriatric care. Elderly have a high risk of falling down when they attempt to get out of bed in order to go to a bathroom. 7.8% of them are hospitalized. This accident has a high risk of serious injury such the bone fracture. To prevent an accident around a bed, one of the effective approaches is the ability to detect gestures on a bed, then the system provides enough time for assist his/her movement. Such a monitoring system will help to reduce the burden of nurses and caregivers. Moreover, a hospitalized elder is usually restricted on a bed with cable or tubes. Then they have a high risk of losing the skill of activity in daily life because of less mobility. Corresponding to the concern of loss functional ability in elderly, a non-invasive sensor is appropriate to be used for monitoring the elderly behavior. In previous works, some studies use commercial pressure mat system to classify postures on a bed for a privacy reason, unlike the system with a camera. However, those of studies require a large number of sensing array which is not practical and costly. Therefore, our approach uses a sensor panel, which consists of only four sensors i.e. two piezoelectric sensors and two pressure sensors. The sensor panel is applied under the mattress in thoraces area. Our approach collects data from elderly patients in hospital with five different postures i.e., out of bed, sitting, lying down, lying left, and lying right. Neural Network approach is used to classify 5 postures and evaluate feature input, i.e. 4 inputs, 120 inputs, 4 inputs with normalized signal, 120 inputs with normalized signal. The 4 inputs are transaction signal from 4 sensors i.e., right piezoelectric signal, left piezoelectric signal, right pressure signal, right pressure signal. The 120 inputs are accumulated signal data in one second time slots. To eliminate the effect of weight and bias between different types of sensors, the unity based normalization (or feature scaling) method is used to normalize sensor data into the range of 0 to 1. The results of 120 inputs with normalized signal reach up to 100% of accuracy. In the full dataset, the accuracy decreases from 100% to 94.10% because of noise and unclean dataset. To eliminate to unexpected result of the output posture from the Neural Network model, the Bayesian Network is adopted to estimate the likelihood of the consecutive postures. We then combine the results from both Neural Network probability and Bayesian probability by the weight arithmetic mean. The experimental results yield the maximum accuracy up to 94.65% when the coefficient of Bayesian probability and a Neural Network are set to 0.7 and 0.3 respectively. Our approach uses only 4 sensors without losing much in performance when comparing to the previous approaches

Thammasat University. Thammasat University Library

Address: BANGKOK

Email: preserv@tu.ac.th

Name: Virach Sornlertlamvanich

Role: advisor

Name: Waree Kongprawechnon

Role: co-advisor

Created: 2016

Modified: 2564-05-19

Issued: 2021-05-18

วิทยานิพนธ์/Thesis

application/pdf

DOI: 10.14457/TU.the.2016.675

eng

DegreeName: Master of Engineering

Level: Master's Degree

Descipline: Information and Communication Technology for Embedded Systems

Grantor: Thammasat University

©copyrights Thammasat University

ลำดับที่.	ชื่อแฟ้มข้อมูล	ขนาดแฟ้มข้อมูล	จำนวนเข้าถึง	วัน-เวลาเข้าถึงล่าสุด
1	2404waranrach.pdf	3.88 MB	1	2021-07-15 17:49:34

ใช้เวลา

-0.980098 วินาที