Instrumentation & Measurement Magazine 26-4 - 38

Gaussian data; however,
the data collected from the
sensors embedded in the
chair are non-Gaussian,
which leads to misclassification
results. In [7], a
supervised machine learning
approach is introduced
to identify a person's posture
based on a sensors
network embedded in the
wheelchair. This approach
uses the Condensed Nearest-Neighbors
approach
for data filtering, the
Kennard-Stone algorithm
to balance the data, PCA
for dimensionality reduction,
and KNN algorithm for posture classification. Results
indicated that this combined approach achieves an average
accuracy of over 75%. This could be due to the small-sized
and unbalanced data used in this study. Most of the developed
detection schemes for sitting posture recognition are
generally designed using shallow supervised techniques
that need labeled data in training. However, getting labeled
data is not obvious and is time-consuming. Thus, this study
aims to design a semi-supervised data-driven detector for
sitting posture monitoring that does not require labeled
data.
Fig. 1. Schematic of the proposed strategy.
Data-driven Wheelchair User's Posture
Detection and Isolation Strategy
The sitting posture monitoring approach is performed in four
steps: empirical model construction using anomaly-free data,
residual generation using the constructed model, posture sitting
detection, and isolation (Fig. 1).
Methodology
This paper presents a data-driven approach for detecting
and identifying wheelchair users' posture using unlabeled
data. Also, this research studies the detection capacity of
the independent component analysis (ICA)-based monitoring
approach on datasets of limited size. Specifically,
the proposed approach combines the advantages of the
ICA model and KD-based monitoring chart to obtain good
detection. Unlike PCA employing orthogonal principal
components, ICA utilizes a linear non-orthogonal coordinate
system, where the directions depend on higher-order
statistics to handle non-Gaussianity in the data. The KD
monitoring chart is applied to the residuals generated from
the ICA for abnormal event detection. The employment of
the KD-based chart is expected to improve the detection of
the ICA-based approach. Once the ICA-KD approach detects
the abnormal posture, a contribution plot is conducted
to identify the type of sitting posture. Experiments based on
a public dataset provided in [7] demonstrate that the proposed
ICA-KD approach can effectively reduce the false
alarm rate and reach a 99% detection accuracy.
The following section briefly describes the preliminary
materials, including the ICA and the KD anomaly detector.
Then, the coupled ICA-KD technique that identifies the
sitting position in a wheelchair is presented. After that, we
assess the performance of the proposed approach using a
publicly available dataset. Finally, we offer conclusions to
this study.
38
Independent Component Analysis-based Monitoring Approach:
ICA is a data-driven technique that considers
statistical parameters of the higher order for extracting important
non-Gaussian features from the process data. Once the
data X=[xl
,x2,...xn]T
with XR is available, the ICA model
mn
*

may be represented as [9]:
X = AS + F
(1)
where A, S and F represent the mixing matrix, the matrix with
independent components (ICs) and the residual matrix, respectively.
Since there are two unknown entities in (1), the aim
of ICA is to find a separating matrix W which can be computed
from ˆS as follows:
S ˆ  WX
(2)
The initial step in ICA involves a normalization process
where the measurements in data X are scaled to have
zero mean. Next, the whitening step is carried out to remove
cross-correlation between the variables, and the whitening
transformation is represented using the following expression:
Z = QX
where Z denotes the whitening matrix, Q Λ U Λ refers to
 1

,
the diagonal matrix and U is the eigenvector matrix computed
from covariance of X. From (1), it is known that X=AS. Using
this, (3) can be expressed as:
IEEE Instrumentation & Measurement Magazine
June 2023
T
(3)
Instrumentation & Measurement Magazine 26-4

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 26-4
