Available via license: CC BY 4.0
Content may be subject to copyright.
© The Author(s) 2023. Published by Oxford University Press on behalf of Central South University Press. This is an Open Access
article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/),
which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.
ORIGINAL UNEDITED MANUSCRIPT
BLE Beacons for Sample Position Estimation in A Life
Science Automation Laboratory
Haiping Wu 1, Steffen Junginger 2, Thomas Roddelkopf 2, Hui Liu 1 and Kerstin Thurow 3,*
1Institute of Artificial Intelligence & Robotics (IAIR), Key Laboratory of Traffic Safety on Track of
Ministry of Education, School of Traffic and Transportation Engineering, Central South University,
Changsha 410075, Hunan, China; 2Institute of Automation (IAT), University of Rostock, 18119
Rostock, Germany; and 3Center of Life Science Automation (celisca), University of Rostock, 18119
Rostock, Germany
*Correspondence: Kerstin.Thurow@celisca.de; Tel.: +49 381 7800
Abstract: Estimation of the sample position is essential for working process
monitoring and management in the life science automation laboratory. Bluetooth
low-energy (BLE) beacons have the advantage of low price, small size, and low-energy
consumption, which make them a promising solution for sample position estimation in
the automated laboratory. Several fingerprinting models have been proposed to achieve
indoor localization with the received signal strength (RSS) data. However, most of the
research depends on intensive beacon installation. Proximity estimation, which
depends entirely on one beacon, is more suitable for sample position estimation in large
automated laboratories. The complexity of the life science automation laboratory
environment brings challenges to the traditional path loss model (PLM), which is a
widely used radio wave propagation model-based proximity estimation method. In this
paper, BLE sensing devices for sample position estimation are proposed. The BLE
beacon-based proximity estimation is discussed in the framework of machine learning,
in which the support vector regression (SVR) is utilized to model the nonlinear
relationship between the RSS data and distance, and the Kalman filter is utilized to
decrease the RSS data deviation. The experimental results over different environments
indicate that the SVR outperforms the PLM significantly, and provides 1 m absolute
errors for more than 95% of the testing samples. The Kalman filter brings benefits to
stable distance predictions. Apart from proximity-based sample position estimation, the
proposed framework turned out to be effective in position estimation between parallel
workbenches and position estimation on an automated workstation.
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
Keywords: BLE beacon, position estimation, proximity estimation, received signal
strength (RSS), support vector regression (SVR), life science automation laboratory
1. Introduction
The past years have seen great developments in the degree of automation in laboratories
[1]. Apart from the workflow design, the state monitoring of the sample and automation
equipment becomes an inseparable part of the automated laboratories. For instance,
samples are transported between automation islands automatically by a transportation
robot. It is important to track the position of the sample during the automated working
process. The position information of the sample can bring benefits for working process
management, process visualization, and fast fault traceability.
As the sample moves between automated laboratories in a pre-designed path, the
sample position tracking task can be regarded as a special kind of proximity estimation.
Reference points can be added in the path of the sample, and the sample position
tracking can be achieved by estimating the distance of the sample and the reference
points. The requirements of the proximity estimation solution are analyzed as follows:
firstly, feasibility. The automation islands are distributed in laboratories of different
sizes, the proximity estimation solution should be feasible enough to be installed in
rooms of different sizes and environments. Secondly, low cost. The proximity
estimation solution should be cheap enough to cover the entire path. Thirdly,
power-efficient. For ease of installation and maintenance, the proximity estimation
solution should be characterized by low-power consumption. Finally, accuracy. The
proximity estimation should provide high accuracy to meet the requirement of sample
position estimation.
There are vast kinds of indoor distance estimation solutions such as ultrasonic-based
methods [2], computer vision-based methods [3], laser-based methods [4], inertial
navigation-based methods, magnetic-based methods, acoustic-based methods,
infrared-based methods, visible light-based methods, and son on [5]. With the
development of the Internet of Things, the wireless network has brought new
perspectives for indoor proximity estimation [5]. Wireless techniques such as Wi-Fi [6],
Radio Frequency Identification (RFID) [7], Bluetooth low energy (BLE) [8,9], and
ultra-wideband (UWB) [10] have been widely discussed in indoor localization
applications. A Survey of this topic can be found in [11]. Among the wireless
techniques, the BLE beacon has the advantages of flexible installation, low price, and
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
low-power consumption. The price of BLE beacons can be as low as a few dollars each.
A single button cell battery can power a BLE beacon for over one year. Distance
estimation with BLE beacons has attracted great attention in both academia and
industry [12].
In the application of BLE beacon-based proximity estimation, the BLE beacon
broadcasts the information about the beacon with Bluetooth signals, the receiver device
gets the information continuously. According to the radio wave propagation model, the
received signal strength (RSS) decreases with distance. The RSS values can be utilized
as indicators of the distance between the receiver device and the BLE beacons.
However, the RSS values are suffering from severe fluctuation as a result of the channel
noise, multipath effects, and shadow effects [13,14]. It is hard to get a high accuracy
proximity estimation with raw RSS values and the traditional propagation model.
In this paper, BLE beacon-based sensing solutions for sample position estimation in a
life science automation laboratory were proposed, and RSS data characters in those
environments were investigated. A machine learning framework for BLE beacon-based
proximity estimation was proposed and validated. The following paper is organized as
follows: Section 2 introduces the related work on BLE beacon-based indoor positioning
and proximity estimation. Section 3 provides the system concept, implementation of
the proposed solutions. Section 4 gives the performance of the proposed solutions over
five experiments. The discussion and conclusion are given in Section 5 and Section 6,
respectively.
2. Related Work
BLE is widely studied in different kinds of environments and tasks. Typical cases are
mobile ticketing in public transportation [15], wireless health monitoring in smart
hospitals [16], and energy management in smart homes [17]. An overview of signal
characteristics, hardware, software, and application scenes of the BLE is provided in
[18]. A detailed outlook for the future adoption of BLE is provided in [19]. In this
section, we focus on the recent advances in BLE beacon-based indoor positioning and
proximity estimation.
The fingerprinting models have attracted great attention in BLE beacon-based indoor
positioning recently [20-22]. The fingerprinting models ignore the propagation
mechanism of the wireless signal and pay attention to the RSS values directly. A
template matching-based BLE fingerprinting indoor positioning system was discussed
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
in [21]. The entire area was divided into several cells, RSS values of each cell were
regarded as fingerprints and recorded in the modeling phase. The Euclidean distances
between current RSS values and the recorded RSS values were calculated. The current
position is estimated according to the distances. A similar modeling process is known
as k-Nearest Neighbors (kNN), which has been widely tested in different environments
[23]. Some attempts have been made to improve the distance calculation in the
kNN-based indoor positioning, for instance, A kernel method is proposed to evaluate
the distance between the current RSS values and the recorded RSS values [24]. It
indicated that the kernel method outperforms the Euclidean distance and cosine
distance. In [25], a taxonomy of fingerprinting positioning was provided, fingerprinting
models including multilayer perceptron (MLP), random forest (RF), radial basis
function network (RBFN), support vector machines (SVM), k-NN, weighted k-Nearest
Neighbors (Wk-NN), kernel regression (KR), Gaussian process regression (GPR),
independent Gaussian distribution (IGD) and histogram matching method (HMM)
were evaluated over 11 different datasets, the best average errors of these algorithms
range between 2.417 m and 3.555 m. It turned out that the Wk-NN provides the best
average accuracy for large-scale environments, and the SVM is the best choice for
small environments.
Different from the BLE beacon-based indoor positioning, where RSS data from several
beacons are utilized at the same time, in the BLE beacon-based proximity estimation,
only one beacon is available. The path loss model is the most popular solution in
proximity estimation. The path loss model describes the relationship between the RSS
value and the distance based on the energy attenuation mechanism of the Bluetooth
signal. In [26], the BLE beacon-based proximity estimation was utilized in social
robotics. In the robotic application, the BLE beacon was carried by the user, the signal
receiver was connected to a social robot. The relationship between the RSS values and
the distance was described in [27]. A more widely adopted path loss model (PLM) is the
log-normal shadowing model proposed in [28]:
0
10 log
RSS n d C
(1)
where
d
is the distance between the BLE beacon and the receiver,
RSS
is the
received signal strength, n and C0 are parameters depending on the BLE beacon and
environment. The filters play important roles in the application of the log-normal
shadowing model. For instance, in [29], a special mean filter that removes the outlier
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
was executed to process the raw RSS data. In [30], the log-normal shadowing model
was utilized in the proximity estimation at a smart museum, and the Kalman filter was
implemented to improve the accuracy. The Kalman filter was also proved to be
effective in improving the indoor positioning accuracy in [31, 32]. In [33], particle
filtering was utilized in the BLE beacon-based smart parking system. In [34], several
Bayesian filtering techniques including Kalman filter, particle filter, and nonparametric
information filter were compared in large and small rooms. The test results indicated
that the filtering techniques can improve the log-normal shadowing model-based
proximity accuracy.
Machine learning methods showed promising results in Fingerprinting model-based
indoor positioning, however, most of the research was discussed in the framework of
classification, and relied on data from more than one beacon. The PLMs provided a
suitable framework for proximity estimation. However, the multipath propagation of
the Bluetooth signal and the complexity of the real application environment brings
challenges to the accuracy of proximity estimation. In this paper, we investigate the
BLE beacon-based position estimation in life science automation laboratory
environments. The innovation and contribution of this work are as follows: (1) A
machine learning framework that improves the BLE beacon-based proximity
estimation performance is proposed; (2) This study is an early attempt in the BLE
beacon-based sample position estimation in life science automation laboratory; (3)
Both the positioning of the sample between different workstations and the positioning
of the sample within a workstation are taken into consideration.
3. System Concept and Implementation
3.1 Sensing Device
The H2A BLE beacon (Moko Smart, Shenzhen, China) utilized in our system is shown
in Fig.1(a). The NRF52840 chipset is utilized as the Bluetooth module on the H2A BLE
beacon. The detailed information is given in Table 1. The H2A BLE beacon was
selected due to its low price, small size, and long battery life. The BLE beacon can be
installed on walls, workbenches, and other important positions as reference points for
proximity estimation. What is more, the BLE beacon can also be installed on the
sample container to achieve position estimation on the workbench. In our application,
the iBeacon protocol [35] and Gauss frequency shift keying are utilized, and the
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
advertising interval is set as 0.1s to make sure the beacon is discovered quickly. The
transmission power is set as +4 dBm to make sure the RSS is strong enough in the
complex life science automation laboratory environments.
In most BLE beacon-based indoor localization research, the smartphone is utilized as
the receiver. It is clear that the smartphone is not suitable for beacon data collection in
the sample position estimation application. An IoT node is utilized as the receiver node
instead [36]. As shown in Fig. 1(b), the modular IoT node consists of 4 layers, which
are the power layer, microcontroller – layer, sensor layer and Wi-Fi and Bluetooth-
communication layer. The power layer provides the power supply from a rechargeable
battery for the IoT node when operated without external power supply. The
microcontroller – layer is responsible for configuring and management of the other
layers. The sensor layer is designed for gas leakage detection in laboratory applications.
The communication layer is designed for Wi-Fi and Bluetooth communication, and in
this particular application, for beacon data collection. The ACN52840 Bluetooth 5
module (Aconno, Düsseldorf, Germany) is utilized to receive the advertising
information from the beacon, and the ESP-WROOM-02D Wi-Fi module (Espressif
Systems, Shanghai, China) is utilized to achieve communication with the IoT cloud. An
RSS register is utilized, and the RSS data is averaged over 10 samples to reduce data
fluctuation during data collection. More detailed information about the IoT node is
given in [36].
Figure 1 Sensing device: (a) BLE beacon; (b) IoT node
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
Table 1 Parameters of the H2A Beacons
Parameter Values / Description
Transmission Type Bluetooth 4.0 Low Energy
Transmission Power (dBm) -40, -20, -16, -12, -8, -4, 0, +3, +4
Advertising Interval (s) 0.1 to 10
Size Φ48×13.89 mm
Battery CR2477 (1100 mAh), replaceable
Default Battery Life Typical up to 30 months
Weight 24 g
Waterproof IP67
Compatibility iBeacon and Eddystone
Installation 3M double side sticker and Screw
3.2 RSS data estimation with Kalman filter
The Kalman filter is a widely used filtering technique in both industry and academia
[37]. The discrete Kalman Filter is utilized in this paper [38]. The state equation of the
target system is explained as
1 1 1
k k k k
x Ax Bu w
(2)
where
k
x
is the system state at the time step k, A is the state transition matrix, B is the
control input matrix,
1
k
u
is the control signal at the time step
1
k
, and
1
k
w
is the
process noise at the time step
1
k
. The measuring equation of the target system is
explained as
k k k
z Hx v
(3)
where
k
z
is the measured data at the time step k, H is the state observation matrix, and
k
v
is the measurement noise. The prediction step and the correction step are carried out
alternately during the filter process. The prediction step is explained as
1 1
ˆ ˆ
k k k
x Ax Bu
(4)
1
T
k k
P AP A Q
(5)
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
where
ˆ
k
x
is the prior state estimate at the time step k,
1
ˆ
k
x
is the posterior state
estimate at the time step
1
k
,
k
P
is the covariance of the prior estimate error, and Q
is the covariance of the process noise. The correction step is explained
T T
k k k
K P H HP H R
(6)
ˆˆ
ˆ
k k k k k
x x K z Hx
(7)
k k k
P I K H P
(8)
where
k
K
is the Kalman gain at the time step k, R is the covariance of the
measurement noise, and
k
P
is the covariance of the posterior estimate error. In the
RSS data filtering application, the system state is constant in a certain position, and
there is no control input, for equation (2), we have A=1, B=0, and 1
0
k
w
. Similarly,
for equation (5) we have Q=0 as the system state in a certain position is constant. The
measured RSS data is of the state directly, so we can get H=1 for equation (3). The
measured RSS data at the time step k is RSSk, therefore, we have
k k
z RSS
. The initial
value of the posterior state estimate in equation (4) is configured as the first measured
RSS data, that is
0 1
ˆ
x RSS
. The initial value of
1
k
P
in equation (5) is configured as 1
according to [38].
3.3 Proximity estimation with support vector regression (SVR)
A training set is defined as
, y , 1, 2, ,
i i
x i N
, where
i
x
is the independent
variable and
i
y
is the response variable. A linear function that describes the
relationship between the independent variable and the response variable is defined as
T
f x w x b
(9)
where w is weight vectors and b is the bias value. The modeling process of the support
vector regression is explained as the following constrained optimization problem
*
2
*
1
, , ,
*
*
1
2
. . ,
,
0, 0
i i
N
i i
i
w b
T
i i i
T
i i i
i i
w C
min
s t w x y
y w x
(10)
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
where
C
is the regularization parameter,
and
*
are slack variables, and
is the
tolerance for deviation [39]. The constrained optimization problem can be solved
according to the following dual problem
* *
1 1 1
*
1
*
1
*
1
max 2
. . 0
0 ,0
N N N
i i i j i i
i j i
N
i i i
i
N
i i
i
i i
x x
y
s t
C C
(11)
where
i
and
*
i
are dual variables. To achieve nonlinear mapping of the input data, the
kernel function is utilized in SVR. The original independent variable is mapped into
high-dimensional feature space [40]. With the kernel function, the trained model is
explained as
*
1
,
N
i i i
i
f x k x x b
(12)
The Gaussian kernel utilized in this study is explained as
,i
x x
i
k x x e
(13)
It is worth noting that the parameters
C
,
, and
should be configured in advance.
There are many solutions to configure the parameters including Bayesian optimization,
the grid search [41], and the heuristic optimization algorithm [42]. In this study, to make
a fair comparison with the PLM, the default parameters are utilized, in which
1
,
1.349
C IQR y, and
13.49
IQR y
[43].
IQR
represents the interquartile
range of the response variable [44].
4. Experimental Evaluation
The sample position estimation performance of the proposed solution in a life science
automation laboratory was evaluated through five comparison experiments. Firstly,
PLMs were developed to evaluate the characteristics of the BLE beacon at different
distances. The experiments were carried out in both the laboratory room and the
corridor, which are typical environments in the life science automation laboratory.
Secondly, the proximity estimation performance of the PLM and SVR were compared.
The best method was selected as the basic proximity estimation model for subsequent
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
experiments. Thirdly, the Kalman Filter was involved to decrease the RSS data
fluctuation. The effects of the Kalman Filter on proximity estimation error and variance
were examined. Fourthly, the performance of BLE beacon-based position estimation
between parallel workbenches was evaluated. Both the proximity estimation model and
prior knowledge of workbench position were utilized to achieve sample position
estimation. Finally, the BLE beacon-based sample position estimation was tested on an
automated workstation. The sample position was identified when it was moved
between different instruments on the automated workstation.
4.1 Modelling the RSS data with PLM
To examine the characteristics of the BLE beacon at different distances, RSS data
collection was carried out in the laboratory and corridor. The room is a typical life
science laboratory filled with devices, workbenches, laboratory cabinets and fume
hoods. The size of the empty area for data collection is 3×4 m. The width of the corridor
is 1.7 m. The experimental layout is designed according to [30]. The experimental setup
is given in Fig.2. As described in Section 3.1, the receiver node was utilized to collect
RSS data from the beacon. The receiver node and the beacon are fixed at the same
height of 0.9 m. There is no obstacle between the receiver node and the beacon. The
distance between the receiver node and the beacon was increased from 0.2 m to 3.0 m at
an interval of 0.2 m, 600 samples were collected at each distance. To model the true
working condition, laboratory staff is working as usual during the data collection. The
collected data in different environments are given in Fig. 3.
It can be found that the RSS data suffers from obvious fluctuation in both
environments. It tends to provide higher fluctuation when increasing the distance
between the receiver node and the beacon. The RSS data within 1 m is much more
stable than the RSS data farther than 1 m. The possible reason is that the BLE signal
propagation in a dense environment suffers from multipath fading, shadowing, and
fading [18]. The RSS data is more likely to be affected by walls, obstacles, and people
when the distance between the receiver node and the beacon increases.
The average RSS data in each distance is utilized to build the PLM in both
environments. In this study, the average is performed in linear scale. The parameters in
equation (1) are estimated by solving a least-squares nonlinear curve-fitting problem,
and the Quasi-Newton method was adopted. The parameters are calculated as follows:
for the corridor, n =1.737 and C= -45.473, for the laboratory, n = 1.675 and C= -46.811.
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
The fitted RSS is given in Fig. 3. It can be observed that the average RSS data deviates
from the fitted RSS data in both environments. The difference between average RSS
data and fitted RSS data increases as the distance between the receiver node and the
beacon increases. For instance, in the corridor, we can see significant differences
between average RSS data and the fitted RSS data at 1.4 m, 2 m, and 2.6 m. It is hard for
the PLM to describe the relationship between the distance and the RSS data in these
points. The characteristics of the life science laboratory may explain this phenomenon.
As above mentioned, both the corridor and laboratory are dense environments, signal
propagation can be easily affected by the walls, obstacles, and people, which brings big
challenges to the traditional PLM model.
Figure 2 Experimental Setup in different environments (a) corridor; (b) laboratory
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
Figure 3 Collected RSS data and fitted RSS in different environments: (a) corridor; (b)
laboratory
4.2 Comparison of PLM and SVR on Proximity Estimation
The experimental results indicated that the PLM model cannot describe the relationship
between RSS data and distance well. In this section, the SVR is developed to model the
RSS values. The RSS data described in Section 4.1 is utilized as the training data. That
is, the distance was increased from 0.2 m to 3.0 m at an interval of 0.2 m. 600 samples
were collected at each distance, and 9,000 training samples were collected in total.
During the training process, the raw RSS data was utilized as the input feature, and the
distances between the receiver node and the beacon were utilized as the output. The
Gaussian kernel was utilized to achieve nonlinear mapping between the RSS value and
distance. The parameters of SVR were configured as described in Section 3.3.
The proximity experiments were carried out to compare the performance of the PLM
and SVR in proximity estimation. The distance between the beacon and the receiver
node was set to 0.2 m, 0.5 m, 1.0 m, 1.5 m, 2.0 m, 2.5 m, and 3.0 m, respectively. 600
samples were collected at each distance, and 4,200 testing samples were collected in
total. The SVR and PLM were utilized to estimate the distances of all testing samples.
The estimated distances were compared with the true distances to identify the
proximity estimation error. The cumulative distribution function of absolute error is
demonstrated in Fig. 4. It can be found that the SVR performs better than the PLM
significantly in both environments. As shown in Fig. 4(a), for the corridor, the
maximum absolute error of the PLM is higher than 3.5 m, while that of the SVR is
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
smaller than 1.5 m. The SVR provides 1 m absolute errors for more than 95% of the
testing samples, while the PLM provides 1 m absolute errors for less than 75% of the
testing samples.
The histograms of the absolute error in different environments are shown in Fig. 5.
Compared to the PLM, the SVR can decrease big error samples and increase the small
error samples. The SVR can provide better accuracy in proximity estimation in both
environments. The possible reason is that the dense life science automation laboratory
environments increase the complexity of the BLE signal propagation. The SVR, which
has powerful nonlinear mapping abilities, can describe the nonlinear relationship
between the RSS data and distance better than the traditional mechanism model.
Figure 4 Cumulative distribution function of the absolute error in different
environments: (a) corridor; (b) laboratory
Figure 5 Histogram of absolute error in different environments: (a) corridor; (b)
laboratory
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
4.3 Improving Proximity Estimation Using Kalman Filter
As presented in Section 4.1, the raw RSS data suffers from obvious fluctuation, which
brings more challenges to proximity estimation. In this section, the Kalman Filter is
involved to smooth the testing samples. The parameters of the Kalman filter were set as
described in Section 3.2. The Raw RSS data and estimated data by Kalman filter at the
distance of 1 m and 2 m in different environments are shown in Fig. 6. It can be
observed that the raw data at 1 m and 2 m show obvious fluctuation in both
environments. The estimated data by the Kalman filter are close to the mean values and
robust to the extreme values. Compared to the Raw data, the data fluctuation of the
estimated data decreases significantly. For instance, in the corridor, the raw RSS data
fluctuation range at 1 m is -49.00 dBm to -58.00 dBm, while that of the estimated data
is -53.38 dBm to -55.00 dBm. The Kalman Filter is effective in smoothing the raw data
and decreasing the effect of the extreme values.
To evaluate the effect of the Kalman filter in improving the performance of the BLE
beacon-based proximity estimation, the Kalman filter is combined with the SVR. The
Raw RSS data is processed by the Kalman filter before putting into the SVR in the
experiment. The distance estimation performance with estimated RSS data and Raw
data in different environments are given in Fig. 7. The black dotted line demonstrates
the actual distance. The orange solid line depicts the range of predicted distance with
the raw data, and the blue solid line depicts the range of predicted distance with the
Kalman filtered data. The orange circles and blue points are the mean values of
predicted distance with the raw data and Kalman filtered data, respectively. It can be
found that the variation of the predicted distance is decreased significantly by the
Kalman filter in all distances and environments. However, when it comes to the mean
values of predicted distance, we cannot say that the Kalman filter brings benefits at all
distances.
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
Figure 6 The Raw RSS data and estimated data by Kalman filter at 1 m and 2 m in
different environments (a) corridor; (b) laboratory
To evaluate the effect of the Kalman filter more clearly, the proximity estimation
performance across different distances and environments is given in Table 2. To
investigate the performance comprehensively, evaluation metrics including Mean
Absolute Error (MAE), Root Mean Square Error (RMSE) [45], and Standard Deviation
of Error (SDE) are utilized. Among the three metrics, the MAE and RMSE reflect the
average error level across all samples, the SDE evaluates the stability of the predictions.
The best value for each evaluation metric is highlighted in bold. As shown in Table 2,
for the experiments in the corridor, compared to the raw data, the Kalman filtered data
provide smaller MAE and RMSE at 0.2 m, 0.5 m, 1.0 m, 2.0 m, and 3.0 m, while bigger
MAE and RMSE at 1.5 m and 2.5 m. The Kalman filter decreases the average error
level in 5 cases, but increases the average error level in the other 2 cases. Similar results
can also be observed in the laboratory. The possible reason is that the Kalman filter
achieves data estimation according to the raw data, the Kalman filter fails to converge
to the true value when most of the raw data is incorrect. A similar phenomenon has also
been reported in [33] in a smart parking application, in which particle filtering was
utilized to process the raw data. Apart from the mean error, compared to the raw data,
the Kalman filtered data provides smaller SDE in all distances and environments, the
Kalman filter brings benefit to stable distance predictions.
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
Figure 7 Distance estimation results in different environments: (a) corridor; (b)
laboratory
Table 2 Experimental Results (in meters) across various distances and environments
Environment
Distance
Raw data Kalman filter
MAE RMSE SDE MAE RMSE SDE
Corridor
0.2
m
0.0491
0.0709
0.0616
0.0238 0.0323 0.0262
0.5
m
0.2182
0.2184
0.0099
0.2154 0.2154 0.0002
1.0 m 0.2279 0.2840 0.2835 0.1582 0.1603 0.0258
1.5 m
0.5818 0.5932 0.1158 0.6452 0.6481 0.0614
2.0 m 0.0493 0.0828 0.0764 0.0158 0.0174 0.0078
2.5 m
0.4767 0.4902 0.1141 0.5442 0.5442 0.0025
3.0 m 0.5207 0.5248 0.0653 0.4790 0.4790 0.0053
Laboratory
0.2 m 0.1266
0.1341
0.0496 0.1416
0.1417
0.0028
0.5 m 0.0956 0.1052 0.0667 0.0776 0.0778 0.0054
1.0 m 0.1414 0.1605 0.1182 0.1271 0.1277 0.0116
1.5 m 0.1861 0.2348 0.1981 0.1044 0.1049 0.0094
2.0 m 0.4165
0.4498
0.3985 0.4063
0.4095
0.0619
2.5
m
0.3201
0.3764
0.1980
0.3588
0.3609
0.0385
3.0
m
0.6442
0.6547
0.1168
0.5997
0.5998
0.0115
4.4 Sample Position Estimation between Parallel Workbenches
In this section, we consider the BLE beacon-based position estimation between two
parallel workbenches, which are common scenarios in the life science laboratory.
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
Typical parallel workbenches are given in Fig. 8. Compared to conventional indoor
positioning applications, position estimation between two parallel workbenches has
several features. Firstly, the number of available beacons is limited. In the life science
automation laboratory scenarios, beacons are installed on important reference points
for proximity estimation. For instance, in our application, we install one beacon for
each workbench. So, there are not enough beacons for the traditional trilateration
positioning method. Secondly, the possible position is limited. Generally speaking, the
workbenches are only accessible on one side, which can exclude half of the possible
positions. Moreover, the possible position can be excluded by the wall, tables, and other
obstacles.
In this study, both the proximity estimation model and prior knowledge of workbench
position were utilized to achieve sample position estimation between parallel
workbenches. The sample position estimation experiment layout is given in Fig 9. As
shown in Fig. 9(a), the relationship between the position of the receiver node and the
beacons can be mathematically represented as:
2
2 2
1
2
2 2
2
x a y d
x b y d
(14)
where (a, 0) is the position of beacon 1, (b, 0) is the position of beacon 2, and (x, y) is
the position of the receiver node. d1 is the distance between the receiver node and
beacon 1, and d2 is the distance between the receiver node and beacon 2. The distance
between the beacons and the receiver node can be calculated by proximity estimation.
The parameters a and b can be configured according to the installation position of the
beacons. In this experiment, we set a=0.5, b=2.5 m. It can be found that we can get
more than one possible position by solving equation (14). To get the true position of the
sample, the impossible position is excluded according to
1 1
2 2
,
,
0
x x if x x
x x if x x
y
(15)
where
1
x
and
2
x
are configured according to the position of the wall and tables. In
this experiment, we set 1
0
x
, 2
4
x
. Generally speaking, we can get the final
position of the receiver node according to (15) and (16). However, it is worth noting
that (15) relies on the assumption that
1 2
b a d d
. Some extreme cases that have
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
high proximity estimation errors may not satisfy the assumption. In the experiment, the
last estimated position is utilized as the current estimation if the assumption is not
satisfied.
The experiment was carryout in the laboratory described in Section 4.1. Firstly, the
SVR models for proximity estimation were set up as described in Section 3.3 and
Section 4.2. Secondly, the distances between the two beacons and the receiver node
were set as demonstrated in Fig. 9(b), in which A, B, and C refer to the positions with
x-axis values of b,
2
a b, and a, respectively, while 0.5 m, 1.0 m, 1.5 m, and 2.0 m
refer to the y-axis values. 200 samples were collected in each position and 2,400 testing
samples were collected in total. Thirdly, the Kalman filter is utilized to smooth the raw
data as described in Section 3.2 and Section 4.3. The SVR models were utilized to
estimate the distance between the receiver node and beacons for all the testing samples.
Fourthly, the position of the receiver node was estimated according to (15) and (16).
The estimation errors at each position were evaluated by the Euclidean distance
between the estimated position and the target position. The MAE and SDE for sample
position estimation between parallel workbenches are given in Table 3. When
comparing the results with Table 2, it can be found that the MAE and SDE of sample
position estimation between parallel workbenches are higher than that of proximity
estimation. Compared to proximity estimation, positioning between two workbenches
is affected by the proximity estimation error of two beacons. What is more, as
explained in [30], the interface between two beacons may cause more challenges to
BLE-based indoor positioning. When comparing the performance at different
distances, it can be found that the system tends to provide smaller position estimation
errors when the distance between the receiver node and workbenches is smaller than 1
m, compared to the position estimation after 1 m. When using the raw data, the average
MAE over 12 positions is 0.808 m, while that of the Kalman filtered data is 0.752 m.
Specifically, the Kalman filtered data provides smaller MAE in 8 of the 12 testing
positions, and smaller SDE in all positions. Kalman filter brings benefits to stable
position estimation. It is also worth noting that Kalman filter increases the average error
level in some cases, the possible reason is that the Kalman filter achieves data
estimation according to the raw data, the Kalman filter fails to converge to the true
value when most of the raw data are wrong.
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
Figure 8 Parallel workbenches in life science laboratory: (a) two workbenches with the
same function and; (b) two workbenches with different functions.
Figure 9 Sample position estimation experiment layout (a) the layout of the laboratory
and; (b) the position of the receiver node
Table 3 Experimental Results (in meters) for sample position estimation between
parallel workbenches
Position Distance Raw data Kalman filter
MAE SDE MAE SDE
A 0.5 m 0.482 0.127 0.405 0.085
1.0 m 0.698 0.277 0.383 0.135
1.5 m 1.280 0.244 1.304 0.020
2.0 m 1.109 0.206 0.887 0.078
B 0.5 m 0.644 0.227 0.606 0.076
1.0 m 0.387 0.267 0.649 0.052
1.5 m 0.839 0.284 0.907 0.064
2.0 m 1.026 0.073 1.013 0.020
C 0.5 m 0.431 0.152 0.469 0.037
1.0 m 0.897 0.147 0.717 0.046
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
1.5 m 0.536 0.146 0.442 0.074
2.0 m 1.361 0.256 1.237 0.147
4.5 Sample Position Estimation on Workstation
In this section, the BLE beacon-based sample position estimation is tested on an
automated workstation. It is important to identify the position of the sample when it is
moved between different instruments on the automated workstation. The sample
position information is of great benefit to the sample traceability and automation
process monitoring. Compared to conventional indoor positioning application, sample
position estimation on the workbench has several features. Firstly, the sample is moved
between limit areas, and the position estimation can be regarded as a classification task
with prior knowledge of the workbench. In the classification task, the RSS data is
utilized as the input feature to identify the position of the sample. Secondly, the RSS
data is faced with high fluctuation, which may cause challenges to sample position
identification. We analyze the reason for the RSS data fluctuation as follows: On one
hand, the automated workstation is filled with various instruments and devices.
Obstacles on the Bluetooth signal propagation path cause more uncertainty to the RSS
data. On the other hand, the location variation inside a certain instrument or device can
also cause RSS data variation. For instance, there are several sample locations on the
workstation deck, and all locations of the sample frame belong to the same class. The
RSS data variance between different sample locations on the sample frame is an
important source of the RSS data variation. In this study, the Kalman filter is involved
to reduce the RSS data variation, and the SVM [8] is utilized to handle the classification
task.
The experiment is carried out on the automated workstation as shown in Fig. 10. The
BLE beacon was installed on the bottom of the sample. The sample is moved between
the labware positioner (area A) and the automated liquid dispensing device (area C) by
the robotic arm. The purpose of this experiment is to identify the position of the sample
according to the BLE beacon signal. The position of the receiver node is an important
factor that may affect the sample position estimation accuracy. Theoretically, when
choosing the installation position of the receiver node, we hope to maximize the RSS
data differences between area A and area C, while minimizing the RSS data differences
inside area A and area C. However, the installation position of the receiver node is
limited by multiple factors including installation space, power supply, and interference
with other devices. In this study, two options for installing the receiver node are
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
discussed. In the first option, the receiver node is installed on the bottom right of the
workstation. In the second option, the receiver node is installed on the mobile robot. As
there is a positioning mark beside the workbench, the mobile robot will reach a
relatively stable position when transporting the sample to the workstation. Compared to
the first option, the second option is more flexible, the workstation can share the
receiver node with other position estimation tasks in the life science automation
laboratory. Three tests were designed to evaluate the effectiveness of the proposed BLE
beacon-based position estimation method. In the first test, the receiver node is installed
on the bottom right of the workbench, and 2,700 training samples are collected from
area A and area C respectively. In the model testing phase, 2,700 testing samples are
collected from area A and area C, respectively. The red dot in Fig. 10(b) are possible
locations of the sample according to the design of the workstation. All possible
locations in area A and area C are included during the data collection. In the second test,
the receiver node is installed on the mobile robot, similar to test 1, 2,700 training
samples and 2,700 testing samples are collected, respectively. In the third test, all data
collected in test 1 is utilized as the training sample, and all data collected in test 2 is
utilized as testing samples. The Gaussian kernel was utilized in the experiments and the
parameters were determined by grid search. That is, the regularization parameter
C
and kernel scale
are searched in the range of
3 3
[10 , 10 ]
with a log scale, and 10
values were tested in each dimension. For test 1, we get
0.4642
C
, and
10
. For
test 2, we get
215.443
C
, and
0.464
. For test 3, we get
0.001
C
, and
215.443
.
The sample position estimation results are given in Table 4. It can be found from the
results that the Kalman filter is effective in improving the position estimation results in
all three tests. For instance, in test 1, there are 46 false identifications when using the
raw data, while only 2 false identifications when using the Kalman filtered data. The
Kalman filter can decrease the false identifications caused by extreme RSS values.
When comparing the results of test 3 with test 1 and test 2, it can be found that the
overall false identification rate of test 3 is higher than that of test 1 and test 2 when
using raw data. The possible reason is that the installation position of the receiver node
can affect the distribution of the RSS data. That is, there is a domain drift between the
training data and testing data. To achieve high sample position estimation accuracy
with raw data, it is a better choice to train a specific SVM model for each installation
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
option. Luckily, when using Kalman filtered data, there is no significant difference
between test 3 and the other two tests. The Kalman filter can decrease data variation
and therefore decrease the effect of domain drift. The Kalman filter provides more
flexibility for the installation position of the receiver node.
Figure 10 Layout of the automated workstation: (a) real view; (b) top view
Table 4 Experimental Results (in meters) for sample position estimation on
workstation
Test
number Area Estimation with raw data Estimation with Kalman filter
A C Accuracy A C Accuracy
Test 1 A 2,699 1 99.96% 2,700
0 100.00%
C 1 2,699 99.96% 0 2,700
100.00%
Test 2 A 2,668 32 98.81% 2,700
0 100.00%
C 14 2,686 99.48% 2 2698
99.93%
Test 3 A 5,367 33 99.39% 5,400
0 100.00%
C 184 5,216 96.59% 4 5,396
99.93%
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
5. Discussion
It is worth noting that the samples are assumed to be at rest during the proximity
estimation and position estimation in this study. A similar assumption was adopted by
[34], in which approx. 1200 samples were collected at each position. In this study 600
samples were collected at each position. To extend application scenarios, the
proximity estimation and sample position estimation should be discussed under moving
state in the future work.
In theory, the RSS data is recommended to be averaged over 36 sample points to reduce
the effect of fast fading during data collection [46]. However, practice
recommendations in [47] believe that increasing the window length beyond 10 or 15
has limited effect on reducing the error and introduces significant delay. Therefore, the
RSS data are averaged over 10 samples during data collection in this paper.
In this study, the SVR is compared with the path loss model, which is the most widely
used benchmark model. It can also be regarded as a comparison of the data driving
model and mechanism-based model. Actually, other data driving algorithms including
machine learning and deep learning methods can also be utilized to model the
nonlinear relationship between the RSS data and distance. It is possible to further
improve the sample position estimation accuracy by careful algorithm selection.
The construction materials and layout of the environment can affect the multipath
propagation and attenuation of Bluetooth signals. Data from different environment
have different probability distributions. In this study, the training data and testing data
are collected in the same environment. That is, comprehensive data collecting and
labeling are carried out in every environment. To decrease data collecting and
labeling work, the transfer learning framework can be used to model data from
different environment in the future work.
This study is an early attempt in the BLE beacon-based sample position estimation in
life science automation laboratory. Actually, there are multiple possible applications
according to the sample position estimation accuracy. Here we try to classify them
into 3 levels, which are presence detection, proximity detection, and localization. The
presence detection can be used to determine if the sample arrived at the target room.
The proximity detection can be used to provide the distance information between the
sample and some key position. For instance, the proximity detection can be used to
find the closest instrument for samples, which is beneficial to process optimization.
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
Localization can be utilized to provide high accuracy position information of the
samples, which is beneficial to personal security, collision avoidance and process
visualization. According to the result of this study, it appears that the BLE beacon is
capable of the first 2 levels while needs more research in the 3rd level tasks.
6. Conclusion
To achieve sample position estimation in the life science automation lab, flexible
sensing solutions based on BLE beacon and IoT node are proposed. The BLE
beacon-based sample position estimation is discussed in the framework of proximity
estimation. To improve the proximity estimation accuracy in the complex life science
automation laboratory, a machine learning proximity estimation framework is
proposed, in which the SVR is utilized to model the nonlinear relationship between the
RSS data and distance. A Kalman filter is utilized to decrease the RSS data deviation.
The proximity estimation results indicate that the SVR outperforms the PLM
significantly in both corridor and laboratory. The SVR provides 1 m absolute errors for
more than 95% of the testing samples. The Kalman Filter is effective in smoothing the
raw data and decreasing the effect of the extreme value. The Kalman filter brings
benefits to stable distance predictions. Apart from proximity-based sample position
estimation, the proposed framework turned out to be effective in position estimation
between parallel workbenches and position estimation on a workstation, which are
typical scenarios in the life science automation laboratory. For position estimation
between parallel workbenches, the proposed framework provides an average MAE of
0.752 m over 12 testing positions. For position estimation on a workstation, the
identification accuracies are higher than 99.93%.
Acknowledgments
This research was done within the Synergy Project ADAM (Autonomous Discovery of
Advanced Materials) funded by the European Research Council (grant number
856405). The authors would also like to thank Dr. Mohammed Faeik Ruzaij Al-Okby,
Dr. Sebastian Neubert, and Mr. Heiko Engelhardt for their help and technical support.
Conflicts of Interest: The authors declare no conflict of interest.
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
References
1. Thurow K, Zhang L, Liu H et al. Multi-floor laboratory transportation technologies based on intelligent mobile
robots. Transp. Saf. Environ. 2019;1:37–53.
2. Aparicio J, Álvarez FJ, Hernández Á et al. A review of techniques for ultrasonic indoor localization systems.
The Journal of the Acoustical Society of America 2019;145:1884–1884.
3. Morar A, Moldoveanu A, Mocanu I et al. A Comprehensive Survey of Indoor Localization Methods Based on
Computer Vision. Sensors 2020;20:2641.
4. Chan S-H, Wu P-T, Fu L-C. Robust 2D Indoor Localization Through Laser SLAM and Visual SLAM Fusion.
2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC). 2018, 1263–8.
5. Obeidat H, Shuaieb W, Obeidat O et al. A Review of Indoor Localization Techniques and Wireless
Technologies. Wireless Pers Commun 2021;119:289–327.
6. Luo RC, Hsiao T-J. Indoor Localization System Based on Hybrid Wi-Fi/BLE and Hierarchical Topological
Fingerprinting Approach. IEEE Transactions on Vehicular Technology 2019;68:10791–806.
7. Liu M, Wang H, Yang Y et al. RFID 3-D Indoor Localization for Tag and Tag-Free Target Based on
Interference. IEEE Trans Instrum Meas 2019;68:3718–32.
8. Lam CH, Jeon KE, Wong S et al. Distance Estimation Using BLE Beacon on Stationary and Mobile Objects.
IEEE Internet Things J 2022;9:4928–39.
9. Ye H, Yang B, Long Z et al. A Method of Indoor Positioning by Signal Fitting and PDDA Algorithm Using
BLE AOA Device. IEEE Sens J 2022;22:7877–87.
10. Morawska B, Lipinski P, Lichy K et al. Transfer learning-based UWB indoor localization using MHT-MDC
and clusterization-based sparse fingerprinting. Journal of Computational Science 2022;61:101654.
11. Zafari F, Gkelias A, Leung KK. A Survey of Indoor Localization Systems and Technologies. IEEE
Communications Surveys & Tutorials 2019;21:2568–99.
12. Sun X, Ai H, Tao J et al. BERT-ADLOC: A secure crowdsourced indoor localization system based on BLE
fingerprints. Appl Soft Comput 2021;104:107237.
13. Zhang K, Zhang Y, Wan S. Research of RSSI indoor ranging algorithm based on Gaussian - Kalman linear
filtering. 2016 IEEE Advanced Information Management, Communicates, Electronic and Automation Control
Conference (IMCEC). 2016, 1628–32.
14. Ayadi M, Zineb AB. Body Shadowing and Furniture Effects for Accuracy Improvement of Indoor Wave
Propagation Models. IEEE Trans Wireless Commun 2014;13:5999–6006.
15. Campos Ferreira M, Galvão Dias T, Cunha JF. Is Bluetooth Low Energy feasible for mobile ticketing in urban
passenger transport? Transp Res Interdiscip Perspect 2020;5:100120.
16. Omre AH, Keeping S. Bluetooth Low Energy: Wireless Connectivity for Medical Monitoring. J Diabetes Sci
Technol 2010;4:457–63.
17. Collotta M, Pau G. A Novel Energy Management Approach for Smart Homes Using Bluetooth Low Energy.
IEEE J Sel Areas Commun 2015;33:2988–96.
18. Jeon KE, She J, Soonsawad P et al. BLE Beacons for Internet of Things Applications: Survey, Challenges, and
Opportunities. IEEE Internet Things J 2018;5:811–28.
19. Mendoza-Silva GM, Torres-Sospedra J, Potortì F et al. Beyond Euclidean Distance for Error Measurement in
Pedestrian Indoor Location. IEEE Trans Instrum Meas 2021;70:1–11.
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
20. Parralejo F, Aranda FJ, Paredes JA et al. Comparative study of different BLE fingerprint reconstruction
techniques. 2021 International Conference on Indoor Positioning and Indoor Navigation (IPIN). IEEE, 2021,
1–8.
21. Dinh T-MT, Duong N-S, Sandrasegaran K. Smartphone-Based Indoor Positioning Using BLE iBeacon and
Reliable Lightweight Fingerprint Map. IEEE Sens J 2020;20:10283–94.
22. Shin B, Lee JH, Yu C et al. Received Signal Strength-Based Robust Positioning System in Corridor
Environment. IEEE Trans Instrum Meas 2022;71:1–15.
23. Pu Y-C, You P-C. Indoor positioning system based on BLE location fingerprinting with classification
approach. Appl Math Modell 2018;62:654–63.
24. Ng PC, Spachos P, She J et al. A Kernel Method to Nonlinear Location Estimation with RSS-based Fingerprint.
IEEE Trans Mob Comput 2022:1–1.
25. Aranda FJ, Parralejo F, Álvarez FJ et al. Performance analysis of fingerprinting indoor positioning methods
with BLE. Expert Syst Appl 2022;202:117095.
26. Alonso-Martín F, Castro-González A, Malfaz M et al. Identification and distance estimation of users and
objects by means of electronic beacons in social robotics. Expert Syst Appl 2017;86:247–57.
27. Seybold JS. Introduction to RF Propagation. John Wiley & Sons, 2005.
28. Kumar P, Reddy L, Varma S. Distance measurement and error estimation scheme for RSSI based localization
in Wireless Sensor Networks. 2009 Fifth International Conference on Wireless Communication and Sensor
Networks (WCSN). 2009, 1–4.
29. Chai S, An R, Du Z. An Indoor Positioning Algorithm using Bluetooth Low Energy RSSI. Atlantis Press, 2016,
274–6.
30. Spachos P, Plataniotis KN. BLE Beacons for Indoor Positioning at an Interactive IoT-Based Smart Museum.
IEEE Syst J 2020;14:3483–93.
31. Alsmadi L, Kong X, Sandrasegaran K et al. An Improved Indoor Positioning Accuracy Using Filtered RSSI
and Beacon Weight. IEEE Sens J 2021;21:18205–13.
32. Dinh T-MT, Duong N-S, Nguyen Q-T. Developing a Novel Real-Time Indoor Positioning System Based on
BLE Beacons and Smartphone Sensors. IEEE Syst J 2021;21:23055–68.
33. Mackey A, Spachos P, Plataniotis KN. Smart Parking System Based on Bluetooth Low Energy Beacons With
Particle Filtering. IEEE Syst J 2020;14:3371–82.
34. Mackey A, Spachos P, Song L et al. Improving BLE Beacon Proximity Estimation Accuracy Through
Bayesian Filtering. IEEE Internet Things J 2020;7:3160–9.
35. Zafari F, Papapanagiotou I, Devetsikiotis M et al. Enhancing the accuracy of iBeacons for indoor
proximity-based services. 2017 IEEE International Conference on Communications (ICC). 2017, 1–7.
36. Neubert S, Roddelkopf T, Al-Okby MFR et al. Flexible IoT Gas Sensor Node for Automated Life Science
Environments Using Stationary and Mobile Robots. Sensors 2021;21:7347.
37. Kalman RE. A new approach to linear filtering and prediction problems. Trans ASME D J Basic Eng
1960;82:35-54.
38. Welch G, Bishop G. An Introduction to the Kalman Filter. Siggraph Course 1997:1–16.
39. Smola AJ, Schölkopf B. A tutorial on support vector regression. Statistics and Computing 2004;14:199–222.
40. Zhou C, Yu W, Huang K et al. A New Model Transfer Strategy Among Spectrometers Based on SVR
Parameter Calibrating. IEEE Trans Instrum Meas 2021;70:1–13.
41. Dash RK, Nguyen TN, Cengiz K et al. Fine-tuned support vector regression model for stock predictions.
Neural Comput & Applic 2021, DOI: 10.1007/s00521-021-05842-w.
42. Yang Z, Wang Y, Kong C. Remaining Useful Life Prediction of Lithium-Ion Batteries Based on a Mixture of
Ensemble Empirical Mode Decomposition and GWO-SVR Model. IEEE Trans Instrum Meas 2021;70:1–11.
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
ORIGINAL UNEDITED MANUSCRIPT
43. MathWorks. Fit a support vector machine regression model - MATLAB fitrsvm - MathWorks China. 2022.
44. Shine P, Murphy MD, Upton J et al. Machine-learning algorithms for predicting on-farm direct water and
electricity consumption on pasture based dairy farms. Comput Electron Agric 2018;150:74–87.
45. Chai T, Draxler RR. Root mean square error (RMSE) or mean absolute error (MAE)? – Arguments against
avoiding RMSE in the literature. Comput Electron Agric 2014;7:1247–50.
46. Lee WCY. Estimate of local average power of a mobile radio signal. IEEE Transactions on Vehicular
Technology 1985;34:22–7.
47. Zanella A. Best Practice in RSS Measurements and Ranging. IEEE Communications Surveys & Tutorials
2016;18:2662–86.
Downloaded from https://academic.oup.com/tse/advance-article/doi/10.1093/tse/tdad033/7303314 by guest on 09 November 2023
