measurement of well-being in the built
environment: sensor network and AI
In recent years the need for technology systems composed by sensor network and
artificial intelligence (AI) to measure people well-being has risen dramatically in
the healthcare sector with a rapid growth in the aging population. In order to
gather knowledge about human health and well-being, the capacity to quantify and
measure the activities of individuals in their home environment with sensors has
become an important factor. Hence combination of using information obtained from
smart homes and analyzing them with AI based algorithms would be great solution
to keep safe independent elderly people in home environment. The objective of
this thesis is to define a measurement setup to compute and distinguish abnormal
behavior from normal characterized by training machine learning algorithms (in
specific K-means, agglomerative and spectral algorithms) and the unlabelled data
obtained from Passive Infrared (PIR) motion sensors and door sensors which are
installed in voluntary participants’ houses in Italy, The Netherlands, Switzerland
and Norway. Our blind approach proposes an unsupervised learning-based solution
that can be seen as representing real-world use case where it is not possible to
provide accurate labelling and annotation of the sensor data. To do so, first, a
review of the available information is undertaken, so as to obtain information and
improve awareness. Later feature engineering techniques were developed to assess
other details from the document to better explain the situation. Then different
clustering algorithms have been applied on datasets to detect abnormalities in the
activity of daily living of the elderly, using real and simulated data, promising
accuracy test and metrics results have been reported, which indicates two out of
three algorithms which applied are valid to create accurate similarity group for
participants. After the training machine learning, providing 1 and 0.82 F1-score
(range of F1-score between 0 and 1) furthermore 100 % and 71 % sensitivity for
combined datasets with simulated data for K-means and Agglomerative clustering
algorithms, respectively, while spectral clustering algorithms shown only 0.75 F1-
score and 61 % sensitivity for combined datasets with simulated data, also average
Silhouette coefficient (range of Silhouette Coefficient between -1 and 1) obtained
from spectral algorithms for combined datasets with simulated datasets is 0.04 which
is much lower than Agglomerative and K-means clustering algorithms with 0.54 and
0.61, respectively. Plus, the result obtained from just real datasets shows even lower
than zero which is -0.07 that make the algorithm unreliable. However , Agglomerative
and K-means clustering algorithms shown acceptable Silhouette Coefficient for real
datasets with 0.5 and 0.35 respectively.
In recent years the need for technology systems composed by sensor network and
artificial intelligence (AI) to measure people well-being has risen dramatically in
the healthcare sector with a rapid growth in the aging population. In order to
gather knowledge about human health and well-being, the capacity to quantify and
measure the activities of individuals in their home environment with sensors has
become an important factor. Hence combination of using information obtained from
smart homes and analyzing them with AI based algorithms would be great solution
to keep safe independent elderly people in home environment. The objective of
this thesis is to define a measurement setup to compute and distinguish abnormal
behavior from normal characterized by training machine learning algorithms (in
specific K-means, agglomerative and spectral algorithms) and the unlabelled data
obtained from Passive Infrared (PIR) motion sensors and door sensors which are
installed in voluntary participants’ houses in Italy, The Netherlands, Switzerland
and Norway. Our blind approach proposes an unsupervised learning-based solution
that can be seen as representing real-world use case where it is not possible to
provide accurate labelling and annotation of the sensor data. To do so, first, a
review of the available information is undertaken, so as to obtain information and
improve awareness. Later feature engineering techniques were developed to assess
other details from the document to better explain the situation. Then different
clustering algorithms have been applied on datasets to detect abnormalities in the
activity of daily living of the elderly, using real and simulated data, promising
accuracy test and metrics results have been reported, which indicates two out of
three algorithms which applied are valid to create accurate similarity group for
participants. After the training machine learning, providing 1 and 0.82 F1-score
(range of F1-score between 0 and 1) furthermore 100 % and 71 % sensitivity for
combined datasets with simulated data for K-means and Agglomerative clustering
algorithms, respectively, while spectral clustering algorithms shown only 0.75 F1-
score and 61 % sensitivity for combined datasets with simulated data, also average
Silhouette coefficient (range of Silhouette Coefficient between -1 and 1) obtained
from spectral algorithms for combined datasets with simulated datasets is 0.04 which
is much lower than Agglomerative and K-means clustering algorithms with 0.54 and
0.61, respectively. Plus, the result obtained from just real datasets shows even lower
than zero which is -0.07 that make the algorithm unreliable. However , Agglomerative
and K-means clustering algorithms shown acceptable Silhouette Coefficient for real
datasets with 0.5 and 0.35 respectively.