Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

2020 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE)

Wireless Positioning Network Location Prediction


Based on Machine Learning Techniques

Furat Al-Obaidy Najmeh Razfar Peyman Moeini


Department of Electrical,Computer, and Department of Electrical, Computer, and Department of Electrical, Computer, and
Biomedical Engineering Biomedical Engineering Biomedical Engineering
Ryerson University Ryerson University Ryerson University
Toronto,Canada Toronto, Canada Toronto, Canada
furat.alobaidy@ryerson.ca nrazfar@ryerson.ca pmoeini@peytec.com
Farah Mohammadi
Department of Electrical,Computer, and
Biomedical Engineering
Ryerson University
Toronto,Canada
fmohamma@ee.ryerson.ca

Abstract— Developing wireless communication technology as well strength can help to measure the sensor and the reader (i.e.
as the extensive usage of wireless networks imply that location between the transmitter and receiver) at WSN [3]. A frequent
applications have been made rendered necessary. Notably, this challenge that WSN systems are often confronted with is
position denotes the functionality and events to estimate a node of wireless reader’s behavioral variations in received signal strength
interest’s location. This study compares different machine
learning (ML) approaches to locate the accurate localization of a
indicator (RSSI) signals. Such a problem arises when readers
moving tag based on wireless sensor network (WSN) as the unable to get the tag queried in their range of reading. However,
wireless propagation medium to improve the prediction of it is possible to address this problem by augmenting the level of
positions. Our experimental recommends using neural networks power and seeking ways of ensuring that the distance between
method to design an appropriate intelligent WSN networks and readers and tags falls in an acceptable range and does not
seek an optimal position with high precisions. The results show the drastically tamper RSSI levels [4]. WSN localization, another
proposed method provides a prediction accuracy about 94%based common challenge, is attributed to factors causing environment
on neural network approach. interference as well as collision induced due to obstacles that
prevent the transmission of signals. Researchers have strived hard
Keywords— Neural network, Wireless position network, Machine
learning prediction, Received signal strength
to reduce these errors attributed to reasons similar to those that
find mention by other studies [5,6]. However, there is a need to
I. INTRODUCTION conduct new studies in order to derive results that are more
Object positioning location forms an important part of scalable. Therefore, the need for a new or enhanced algorithms
numerous applications like security, tracking, location-based that improve the WSN localization methods based on machine
services and navigation. A geometric description is a typical way learning (ML) techniques.
of describing locations spatially. Wireless positioning system for
indoor localization is a key enabling technology comprising of a
minimum of a couple of hardware components.
This includes a signal transmitter and a measuring unit. The
former denotes a simplistic version by transmitting beacon
signals [1-2]. The typical wireless sensors network (WSN)
structure for system localization is shown in fig.1, its consists
from hundreds of wireless tag nodes for locating the positions,
wireless tag readers for data collecting and processing then this
information are send into the end user for manipulation via
gateway unit and the internet. Moreover, routing MAC protocols
are used to handle and communicate between the wireless
network nodes and readers.
WSN based technology is predicted on existing networks for
seeking information pertaining to reader position. In turn, this
entails the utilization of networking infrastructure for determine
the tags’ angulation position. Therefore, a receiving signal’s Fig. 1. Shows typical wireless positioning system structure.

978-1-7281-5442-8/20/$31.00 ©2020 IEEE


Authorized licensed use limited to: University of Exeter. Downloaded on June 01,2021 at 22:30:29 UTC from IEEE Xplore. Restrictions apply.
2020 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE)

II. RELATED WORK appropriate in order to resolve problems with predictions


Several studies have proposed to identify a solution for WSN due to their computational flexibility and simplicity as
systems, and considered a gamut of factors such as accuracy, compared to other architectures. In this work, the
complexity and robustness with high scalability in indoor or supervised cascade forward MLP has been applied.
outdoor positioning. The study in [7] provides an interesting B. Adaptive Neuron-Fuzzy Inference System (ANFIS): It
assessment on indoor WSN technologies such as Wi-Fi, RFID, is a hybrid system that combines fuzzy logic with neural
UWB, Bluetooth, and ZigBee to resolve simple localization networks. ANFIS has been proven as a more accurate
problems. The authors have postulated the use of hybrid means of predicting as well as a fast turning speed for
mechanism as an effective approach to achieve reliable indoor data forecasting, which is why it has been adopted as the
localization system. However, for these methods in order to chosen methodology [16].
increase the localization capability tags advantages such as battery C. Support Vector Machine (SVM) technique was derived
life, scalability, latency, and reliability and other network will be from statistical learning theory and ML in order to
sacrificed to increase the localization accuracy. Furthermore, perform predictions, classification, as well as regressions
depending on the sample rate and method of localization the [17-21]. The SVM proposed in this study delivers the
localization accuracy gained would differ significantly. prediction of the tag position detection [22].
Ultra wideband (UWB) technologies such as Dec-wave is IV. METHODOLOGY
used for localization systems as [8]. In [9] MAC protocol has
been used to synchronize transmission in order to detect tags In this work, we present a methodology for making
sensors location in a WSN for an indoor environment. Similarly, predictions about the nodes’ position. The feasibility of applying
studies in [10] used a cloud infrastructure to link WSN based on prediction methodology based on ML within an WSN system has
adaptive clustering approach in order to support object tracking been ascertained. In order to accomplish that purpose, we load
and location. Different RFID indoor localization systems have the features of the model to ML, after normalizing the received
been proposed in extant research [11-14]; these include hybrid samples (i.e. RSSIs and measured nodes positions) as an input.
methods combining RFID with UWB, as observed by [11]. The The output is the estimation of sensor position at (x, y, z)
hybrid approach of combined devices (RFID and WSN) has coordinates. As part of this structure, supervised learning
been adopted to yield desirable findings [12]. In machine methods have been applied to the learning algorithms in order to
learning studies, the authors [13] used 3D RFID localization establish a relationship between the predictor model and the
method as well. In [14], neural network (NN) algorithm and target on the basis of these steps:
virtual reference tags were added to cluster-based movable tag • Data Collection And featuers processing:
localization (CMTL) method, in order to improve cluster-based
The data set is collected before the prediction operation.
RFID tag localization. Similarly, [15] proposed a deep learning
This collected data set is processed and divided into two
architecture combining long short-term memory network with
groups: training set and testing set. The ML inputs consist
convolutional NN for dynamically identifying and deriving
of two features: the measured position and RSSI signal
visceral attributes from data involving large-scale multipath
collected between each two readers. It is notable that
spectrums. Most localization methodologies mentioned above
target output denotes the desired position between each
often has poor localization accuracy and it is not reliable nor
these two nodes.
repeatable. Moreover, most of them are not sustainable for
critical industrial environments owing to severe radio channel • Data processing:
noise and harsh propagation conditions such as severe multipath, Input data gets normalized as part of this process. The
diffractions and reflections [Peyman’s paper]. To overcome process of normalization not only improved the efficiency
some issues related with previous researches such as poor of training, but also lowers the mistakes caused in terms of
localization accuracies, a proposal approach based on WSN- output activation. Under this step, the normalization is
ML techniques is presented as a main contribution in this work. performed process on input data sets. The normalize value
The paper makes mention of associated work in Section II, of feature (xni) can be obtained by applying Equation 1.
whereas the ML prediction techniques are discussed in section
III. Section IV discusses a methodology procedure, whereas =∑ (1)
section V includes the experimental evaluations for the proposed ( )
method. Finally, section VI concludes this paper. • Building the predction model:
III. ML PRIDICTION TECHNIQUES Fig.2 shows the proposed MLP architecture. This
design is able to yield a powerful performance. This
In order to make prediction for the physical location for
configuration can help ascertain the relationship between a data
sensors, three typical ML techniques are used. A succinct
set of training and testing set. It trains and creates a fitting
explanation of the SVM, MLP, and ANFIS methods is
network of samples set before yielding an estimated output.
presented below.
The intelligent prediction is used to learn a model called a
A. Multilayer Perceptron (MLP) algorithm: It is
predictor in this architecture. The involves a couple of phases:
categorized as a feedforward neural network approach
training phase (testing via validation samples of data set are
that is no feedback is allowed. The MLPs are deemed

Authorized licensed use limited to: University of Exeter. Downloaded on June 01,2021 at 22:30:29 UTC from IEEE Xplore. Restrictions apply.
2020 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE)

used to fit the model during the training stage), so that the conjugate gradient method scheme is shown in fig.5 with
model derives learnings from this data. Through this phase, the parameters being set in table II. In order to examine the
validation data set is also used in order to evaluate a learning accuracy of results in this prediction technique, we make a series
model in a proper manner so as to derive appropriate results. of attempts and compare their accuracy for different structures
For this reason, the model occasionally got the data (i.e. of ANFIS. This demonstrates that ANFIS scheme shows
validation data) tested. Thus, it implies that the validation does improved performance with 16 membership functions for each
impact a model, albeit indirectly. In the testing phase, test data input. As a consequence, the SVM predictor model is
set is only used once a model is completely trained (i.e. using implemented in accordance with the sequential minimal
the train and validation sets). As part of this step, we used our optimization approach that entails the use of kernel function. In
test sets as the testing mechanism. The verification of test this predictor, a kernel function, referred to as quadratic
samples is done on the basis of the desired position or target function, is used to conceal layers for the same effect.
output. In testing phase, test data set is only used once a model
TABLE I. PARAMETERS SETTING FOR WPN STRUCTURE.
is completely trained (i.e. using the train and validation sets).
For this step we used our test sets for testing mechanism. The Parameter Value
Network size 13×9 m2
test samples are verified depending on the target output (i.e.
No. of sensors (tags) 25
desired position). Frequency 6.49 GHz; Band = [6.24 - 6.74] GHz
Packet size 25 bytes
Initial energy 15876 joules= 1470mAh, 3Volt for a
Training Data Set Target Output battery type CR123A.
(Desired position) Transmission current 75 mA
Receiving current 130 mA
Distance between each node 4 m to 9.22 m
Input features Beacon Order (BQ) 4
Output results
(RSSI, Measured Predictor Beacon interval 2 sec
(Optimal position)
position) MAC Technique (protocol) IEEE 802.15.4
collecting the samples period 30-40 min

Testing data Set

Fig. 2. Block diagram for the ML prediction steps.

V. RESULTS AND PERFORMANCE EVALUATION


The experiment field geometry for this study has its location
area in the indoor facility, whereas the ground has dimensions
of 9 m by 13 m. The test is based on five different readers in
addition to 25 sensors or tags. All these readers are randomly Fig. 3. The proposed MLP Configuration.
distributed the area with different coordinates (x, y, z) of the
facility. They are assumed to be identical in size and type and no
obstacles or walls are expected to be found within the
experiment area. At the initial stages, readers are adjusted to
specific heights. Table I lists the WSN system’s characteristics.
The proposed ML model entails the use of specific tools such as
MATLAB [23] for estimating the desired position of nodes.
In the MLP model, the supervised cascade-forward is
applied. The configuration yielding the factors of minimum
performance include a hidden layer that in turn consists of 10
neurons. Furthermore, sigmoid function is regarded as a
function of activation. More specifically, in this study, we used
Levenberg–Marquardt training algorithm as part of the design.
It is notable that the output layer transfer function is selected to
be a liner, as shown in fig.3.The simulation model shows the fast
convergence of the selected MLP configuration where the
performance response is lowered to 1.0296e-5 following more
than 245 epochs. Fig.4 displays the performance results for
Fig.4. Results for comparing between desired distance and predicted
testing phase for our proposed design and makes comparison distance for 10 testing samples.
between 10 samples as example.
The model of the ANFIS prediction adopted by scaled

Authorized licensed use limited to: University of Exeter. Downloaded on June 01,2021 at 22:30:29 UTC from IEEE Xplore. Restrictions apply.
2020 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE)

TABLE IV. COMPARISON OF PREDICTION RESULTS SAME READERS HEIGHT.

Parameters MLP SVM ANFIS


No. of samples 7000
No. of training samples 4900
No. of testing samples 2100
Average Accuracy (%) 93.91 81.45 92.34

TABLE III. COMPARISON OF PREDICTION RESULTS ARBITRARY READERS


HEIGHT.
Parameters MLP SVM ANFIS
No. of samples 7000
No. of training
4900
samples
No. of testing samples 2100
Average Accuracy (%) 94.32 83.45 92.23
Fig.5. ANFIS design structure.
VI.CONCLUSION
TABLE II. ANFIS PARAMETERS SETTING.
This work aims to systematically investigate and forecast the
Parameters Description use of three supervised ML algorithms on the basis of WSN.
Number of Inputs 2 We achieved two experiments that showed the intelligent ML
Number of Outputs 1 techniques based WSN system ability to forecast the position of
Number of input membership 10
functions sensors position in a precise manner. Simulations are carried out
Number of Output membership 16 widely for ascertaining this approach’s optimal forecast
functions accuracy. Detectability of nodes positions is predicted in an
Number of Rules 16 accurate manner for the reader, irrespective of whether the
Fuzzy model type sugeno
height is the same or found to be higher. It is also observed that
MLP provides a better estimation with an increased accuracy of
The accuracy ratio is calculated in order to evaluate the around 94% on average. This implies that it is worthwhile to
efficiency of our proposed ML techniques for a localization consider the utility of WSN positioning for industrial
system. The accuracy percentage of predictions is defined as the applications. Moreover, this ML-based approach is an important
percentage of total samples that are correctly forecasted to the aspect of advanced WSNs, which is likely to pave the way for
total number of samples in the testing dataset as shown in several next-generation applications during the years to come.
equation (2). ACKNOWLEDGEMENT
.
%= × 100 (2) The authors would like to acknowledge our industrial partner
.

All the techniques of prediction were considered in a PeyTec through the support provided by the National Science
separate manner. Table III demonstrates this approach. and Engineering (NSERC-Engage) program.
According to the findings, the accuracy rate of MLP REFERENCES
algorithm is higher by almost 94.32% % in comparison
[1] Wen, Q., Liang, Y., Wu, C., Tavares, A. and Han, X., “Indoor localization
to algorithms used elsewhere. algorithm based on artificial neural network and radio-frequency
identification reference tags,” J. Advances in Mechanical Eng., vol.10,
• Sensitivity Analysis no.pp.1-12, 2018.
[2] Khan, Aftab et al. “Standing on the Shoulders of Giants: AI-driven
Calibration of Localization Technologies,” J. arXiv , 2019.
As part of this step, we gathered from the same
[3] Al-Obaidy, F., Momtahen, S. and Mohammadi, F., “Wireless Sensor
experimental conditions, although the readers are located at Networks Analysis based on MAC Protocols,”, In Proc. Conf. IEEE
same height (z- coordinate). Steps taken previously are Canadian Elect. Comput. Eng. (CCECE) pp. 1-4, 2019.
repeated under the new scenario as well. In this case, the MLP [4] A.M. Khalel, “Position Location Techniques in Wireless Communication
algorithm displays an increased height, as illustrated in Table Systems,” Master Thesis, Bleckinge Institute of Technol.,
IV. MPL provide us higher prediction accuracy 93.91 % on Karlskrona,Sweden, 2010.
average compares with others ML techniques. [5] F. Khelifi, A. Bradai, A. Benslimane, P. Rawat, and M. Atri, “A Survey
of Localization Systems in Internet of Things,” J . Mobile Netw. Appl.,
vol.24, no.3, pp. 761–785, 2019.
[6] D. Zhang, J. Zhou, M. Guo, J. Cao, and T. Li, "TASA: Tag-free activity
sensing using RFID tag arrays,” J. IEEE Trans. Parallel & Distributed
Syst., vol. 22, no.4, pp. 558-570, 2011.
[7] Oguntala, G., Abd-Alhameed, R., Jones, S., Noras, J., Patwary, M. and
Rodriguez, J., “Indoor location identification technologies for real-time

Authorized licensed use limited to: University of Exeter. Downloaded on June 01,2021 at 22:30:29 UTC from IEEE Xplore. Restrictions apply.
2020 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE)

IoT-based applications: An inclusive survey,” J. Comput. Science [16] Al-Obaidy, Furat, Yazdani, F., and Mohammadi, F., "Fault detection
Review, vol.30, pp.55-79, 2018. using thermal image based on soft computing methods: Comparative
[8] L. Yang, J. Han, Y. Qi, C. Wang, T. Gu, and Y. Liu, "Season: Shelving study," J. Microelectronics Rel., vol.71, pp. 56-64, 2017.
interference and joint identification in large-scale RFID systems," in Proc. [17] A. Dorostkar, A. Asad, M. Fathy, M. R. Jahed-Motlagh, & F.
IEEE INFOCOM, 2011, pp. 3092-3100. Mohammadi, “Low-power heterogeneous uncore architecture for future
[9] X. Li, “A Mac Layer Protocol For Smart Indoor Inventory Management 3D chip-multiprocessors,” ETRI Journal, vol.40, no.6, pp.759-773,2018.
System.” Master. Thesis, Ryerson University, Toronto, Canada, 2020. [18] A. Asad, M. Fazeli, M. R. Jahed-Motlagh, M. Fathy, & F. Mohammadi,
[10] Kamruzzaman, S.M., Jaseemuddin, M., Fernando, X. and Moeini, P., “ An Energy-Efficient Reliable Heterogeneous Uncore Architecture for
“Wireless Positioning Sensor Network Integrated with Cloud for Future 3D Chip-Multiprocessors. Journal of Circuits, Systems and
Industrial Automation.” In Proc. Conf. IEEE 42nd Local Comput. Netw. Computers,” pp.1950224,2019.
(LCN), pp. 543-546, 2017. [19] F. Al-Obaidy, A. Asad, and F. Mohammadi, “Reconfigurable hybrid
[11] R. D'Errico, M. Bottazzi, F. Natali, E. Savioli, S. Bartoletti, A. Conti, et cache hierarchy in 3D chip-multi processors based on a convex
al., "An UWB-UHF semi passive RFID System for localization and optimization method”, 32nd In Proc. IEEE Conf. Canadian Elect. &
tracking applications," in Proc. IEEE Int. Conf. RFID-Technol. Appl., Comput. Eng. (CCECE), pp. 1-6, 2019.
RFID-TA 2012, pp. 18-23. [20] A. Asad, O. Ozturk, M. Fathy, and M. R. Jahed-Motlagh, “Optimization-
[12] Z. Xiong, Z. Song, A. Scalera, E. Ferrera, F. Sottile, P. Brizzi, et al., based power and thermal management for dark silicon aware 3D chip
"Hybrid WSN and RFID indoor positioning and tracking system RFID multiprocessors using heterogeneous cache hierarchy,” J.
and near field communications in embedded systems," J. Eurasip Microprocessors & Microsystems, vol. 51, pp. 76-98, 2017.
Embedded Syst., vol. 2013, no.1, pp.6, 2013. [21] F. Al-Obaidy, A. Asad, and F. Mohammadi, “Power-Management based
[13] B. Jachimczyk, D. Dziak, and W. J. Kulesza, "RFID-Hybrid Scene on Reconfigurable Last-Cache level on Non-volatile Memories in Chip-
Analysis-Neural Network System or 3D Indoor Positioning optimal Multi processors”, 32nd In Proc. IEEE Conf. Canadain Elect. & Comput.
system arrangement approach," in Proc. IEEE Int. Conf. Inst. Meas. Eng. (CCECE), pp. 1-4, 2019.
Technol. (I2MTC) 2014, pp. 191-196. [22] M. Jo, H. Y. Youn, S. Cha and H. Choo, "Mobile RFID Tag Detection
[14] Choi, J.S. “Accurate and Cost-Efficient Object Localization Using Passive Influence Factors and Prediction of Tag Detectability," J. IEEE Sensors,
UHF RFID,” University of Texas at Arlington, Arlington, Ph.D. Thesis, vol. 9, no. 2, pp. 112-119, 2009.
USA, 2011 [23] Moore, Holly. MATLAB for Engineers. Pearson, 2017.
[15] Fan, X., Wang, F., Wang, F., Gong, W. and Liu, J., “When RFID Meets
Deep Learning: Exploring Cognitive Intelligence for Activity
Identification,” J. IEEE Wireless Communications, pp.2, 2019.

Authorized licensed use limited to: University of Exeter. Downloaded on June 01,2021 at 22:30:29 UTC from IEEE Xplore. Restrictions apply.

You might also like