Identification and Prediction of Abnormal Behaviour Activities of Daily Living in Intelligent Environments

bumpedappleΚινητά – Ασύρματες Τεχνολογίες

21 Νοε 2013 (πριν από 3 χρόνια και 8 μήνες)

53 εμφανίσεις

Identification and Prediction of

Abnormal Behaviour Activities

of Daily Living in Intelligent

Environments


By

SAWSAN MOUSA MAHMOUD


A thesis submitted in partial fulfilment of the requirements of

Nottingham Trent University for the degree of

Doctor of
Philosophy

May 2012


Abstract


The aim of this research is to investigate e
ffi
cient mining of useful

information from a
sensor network forming an Ambient Intelligence

(AmI) environment. In this thesis, we
investigate methods for supporting

independent livi
ng of the elderly (and speci
fi
cally
patients

who are su
ff
ering from dementia) by means of equipping their home

with a simple sensor network to monitor their behaviour and identify

their Activities of
Daily Living (ADL). Dementia is considered to be

one of
the most important causes of
disability in the elderly. Most

patients would prefer to use non
-
intrusive technology to
help them to

maintain their independence. Such monitoring and prediction would

allow
the caregiver to see any trend in the behaviour of th
e elderly

person and to be informed of
any abnormal behaviour.


Employing a sensor network system allows us to extract daily behavioural

patterns of the
occupant in an Intelligent Inhabited Environment

(IIE). This information is then used to
build a
behavioural

model of the occupant which ultimately is applied to predict the
future

values representing the expected occupancy in the monitored environment.

Challenges of employing wired and wireless sensor network

have been widely
researched. However, pat
tern analysis and prediction

of sensory data is becoming an
increasing scienti
fi
c challenge and

this research investigates appropriate means of pattern
mining and

prediction within the IIE.


Door entry and occupancy sensors are used to extract the movement

patterns of the
occupant. These sensors produce long sequences of

data as binary time series, indicating
presence or absence of the occupant

in diff
erent areas. It is essential to convert these
binary series

into a more
f
exible and e
ffi
cient format before

they are processed

for any further analysis and prediction. Di
ff
erent ways of representing

and visualizing the
large sensor data sets in a format suitable for

predicting and identifying the behaviour
patterns are investigated.




A two
-
stage integration o
f Principal Component Analysis (PCA) and

Fuzzy Rule
-
Based
System (FRBS) is proposed to identify important

information regarding outliers or
abnormal behaviours in ADLs. In

the
fi
rst stage, binary dissimilarities or distance
measures are used to

measure the

distances between the activities. PCA is then applied to

find

two indices of Hotelling's
T
2
and Squared Prediction Error (SPE).

In the second stage of the process, the calculated indices are provided

as inputs to FRBSs
to model them heuristically. They ar
e used to

identify outliers and classify them. The
proposed system identi
fi
es

user activities and helps in distinguishing between the normal
and

abnormal behavioural patterns of the ADLs.


Data provided for this investigation was from real environments and

from a previously
developed simulator. The simulator was modi
fi
ed to

include trending behaviour in the
activities of daily living. Therefore,

in the occupancy signal generated by the simulator,
both seasonality

and trend are included in occupant's
movements. Prediction models

are built through Recurrent Neural Networks (RNN) after converting

the occupancy
binary time series. RNN have shown a great ability

in
fi
nding the temporal relationships
of input patterns. In this thesis,

RNN are compared to ev
aluate their abilities to accurately
predict

the behaviour patterns. The experimental results show that Echo

State Network (ESN) and Non
-
linear Autoregressive netwoRk with

eXogenous (NARX)
inputs correctly extract the long term prediction

patterns of the o
ccupant and
outperformed the classical Elman network.