A Wireless Sensor Network with Soft Computing Localization Techniques for Track Cycling Applications.

sensors Article

A Wireless Sensor Network with Soft Computing Localization Techniques for Track Cycling Applications Sadik Kamel Gharghan 1,2, *, Rosdiadee Nordin 1 and Mahamod Ismail 1 1

2

*

Department of Electrical, Electronic and Systems Engineering, Faculty of Engineering and Built Environment, Universiti Kebangsaan Malaysia, UKM Bangi, Selangor 43600, Malaysia; [email protected] (R.N.); [email protected] (M.I.) College of Electrical and Electronic Engineering Techniques, Middle Technical University, Al Doura 10022, Baghdad, Iraq Correspondence: [email protected] or [email protected]; Tel.: +60-17-232-0776

Academic Editor: Lyudmila Mihaylova Received: 15 April 2016; Accepted: 27 June 2016; Published: 6 August 2016

Abstract: In this paper, we propose two soft computing localization techniques for wireless sensor networks (WSNs). The two techniques, Neural Fuzzy Inference System (ANFIS) and Artificial Neural Network (ANN), focus on a range-based localization method which relies on the measurement of the received signal strength indicator (RSSI) from the three ZigBee anchor nodes distributed throughout the track cycling field. The soft computing techniques aim to estimate the distance between bicycles moving on the cycle track for outdoor and indoor velodromes. In the first approach the ANFIS was considered, whereas in the second approach the ANN was hybridized individually with three optimization algorithms, namely Particle Swarm Optimization (PSO), Gravitational Search Algorithm (GSA), and Backtracking Search Algorithm (BSA). The results revealed that the hybrid GSA-ANN outperforms the other methods adopted in this paper in terms of accuracy localization and distance estimation accuracy. The hybrid GSA-ANN achieves a mean absolute distance estimation error of 0.02 m and 0.2 m for outdoor and indoor velodromes, respectively. Keywords: cycling; distance estimation; optimization technique; soft computing; WSN

1. Introduction One of the key challenges in wireless sensor networks (WSNs) is localization [1]. Localization of sensor nodes in WSNs is essential since it reflects spatial context with the data gathered by sensor nodes and used in applications [2]. There are several applications related to location knowledge in WSNs, such as target tracking [3], person tracking, monitoring [4], unmanned aerial vehicles [5], patient fall detection [6], wild forest areas [7], agriculture [8], disasters [9], and environment management [10]. Given the various constraints in WSN networks, accuracy is the main challenge in WSN localization techniques [11]. These techniques differ based on physical location (e.g., x and y coordinates in a two-dimensional Cartesian map or distance between nodes); types of environment (e.g., outdoor or indoor); network topology (e.g., centralized, decentralized, or remote sensing); and wireless protocol and cost (e.g., Bluetooth, Wi-Fi, ZigBee, or RF). Sensor nodes (SNs) in most WSN applications are powered by a battery and the amount of energy consumed by the nodes determines the network’s lifespan. For future Internet of Things (IoT) applications, reducing energy consumption and prolonging the battery life of SNs have become compulsory. The power consumption of sensor nodes can be reduced by controlling the transmitted power of the wireless protocols. This control can be achieved through an accurate distance estimation between the nodes in the WSN. One method of reducing sensor node power consumption is by exploiting the location or distance between the nodes. Sensors 2016, 16, 1043; doi:10.3390/s16081043

www.mdpi.com/journal/sensors

received signal strength (RSS) [16], acoustic energy [17], and global positing system (GPS) [18]. In contrast, the range-free method has a low level of accuracy for the estimation of the sensor node location while being cost-effective [19]. It depends on the connectivity between fixed nodes, known as anchor nodes, and a further mobile or stationary sensor node. The range-free method determines the position the sensor node without achieving distance estimation [20]. Examples of this approach Sensors 2016, 16,of 1043 2 of 23 are the centroid localization technique [21], weighted centroid localization technique [22], hop-count-based localization [23], and the pattern matching method [24]. Artificial intelligence There are several techniques localization. techniques can Neural be classified into localization techniques have beenused usedininWSN previous research,These including Artificial Network two categories: classical localization techniques and artificial intelligence techniques that use soft (ANN) [25], Neural Fuzzy Inference System (ANFIS) [26], Fuzzy logic [27], and optimization computing optimization algorithms, as shown Figure 1.Swarm Classical localization techniques been algorithms, such as Genetic Algorithms [28],inParticle Optimization (PSO) [29],have Bacterial introduced in previous research works and include range-based and range-free. The range-based Foraging Algorithm (BFA) [30], and Gravitational Search Algorithm (GSA) [31]. method can be used to estimate distance or angle between sensor nodes. This method has aishigh However, in most previousthe localization techniques, the localization or distance accuracy still level of accuracy but requires extra hardware. Examples of this approach is time of arrival (ToA) [12], not satisfactory. The distance estimation accuracy of track cycling is essential for paving the way for phasestudies of arrival (POA) angle of arrival (AoA) [14], the timetransmitted difference of arrival [15], received more related to[13], energy saving in WSNs, where power of(TDoA) the sensor nodes can signal strength (RSS) [16], acoustic energy [17], and global positing system (GPS) [18]. In contrast, be controlled according to the accurate distance measurement between the coach and the bicycle on the track. range-free a low level accuracy for the estimation of can the sensor node location the With method accuratehas localization, theofpower consumption of the WSN be improved and the while being cost-effective [19]. It depends on the connectivity between fixed nodes, known as anchor battery life of the sensor nodes can be extended. This paper presents range-based localization using nodes, and a further mobile or stationary sensor node. The range-free method determines the position RSSI with two soft computing techniques (i.e., ANFIS and ANN) to improve the distance estimation of the sensor node the without achieving distance estimation [20]. Examples of this approach the accuracy between mobile node (i.e., bicycle) while moving on the cycle track and theare coach centroid localization technique [21], weighted centroid localization technique [22],hop-count-based (located at the centre of the track cycling field) in both outdoor and indoor velodromes. Three localization [23], and the pattern matching [24]. Artificial intelligence(BSA), localization optimization algorithms PSO, GSA, and method Backtracking Search Algorithm were techniques combined have been used in previous research, including Artificial Neural Network (ANN) [25], Neural individually with ANN to improve the distance estimation accuracy. The PSO, GSA, and BSAFuzzy were Inference to System (ANFIS) [26], Fuzzy logicof[27], and inoptimization as Genetic designed determine the optimum number neurons each hidden algorithms, layer and thesuch learning rate of Algorithms [28], Particle Swarm Optimization (PSO) Bacterial Algorithm [30], ANN. A high level of distance estimation accuracy was[29], achieved thatForaging was better than in the(BFA) previous and Gravitational Search Algorithm (GSA) [31]. works. Wireless sensor node localization Artificial intelligence methods

Classical methods

Range-based localization

Range-free localization

Hop-count based localization

Fuzzy logic

ANFIS

PSO

ANN

Pattern matching localization

RSSI

TOA

TDoA

AOA

GPS

Figure 1. Classification of localization techniques. Figure 1. Classification of localization techniques.

2. Related Cycling Localization Techniques However, in most previous localization techniques, the localization or distance accuracy is still Several studies different approaches wireless bicyclethe track have not satisfactory. The adopting distance estimation accuracy offor track cyclingpositioning is essentialon forapaving way for been conducted. The location and speed of an athlete were monitored in [32] using GPS as part of a more studies related to energy saving in WSNs, where the transmitted power of the sensor nodes can be controlled according to the accurate distance measurement between the coach and the bicycle on the track. With accurate localization, the power consumption of the WSN can be improved and the battery life of the sensor nodes can be extended. This paper presents range-based localization using RSSI with two soft computing techniques (i.e., ANFIS and ANN) to improve the distance estimation accuracy between the mobile node (i.e., bicycle) while moving on the cycle track and the coach (located at the centre of the track cycling field) in both outdoor and indoor velodromes. Three optimization algorithms PSO, GSA, and Backtracking Search Algorithm (BSA), were combined individually with ANN to improve the distance estimation accuracy. The PSO, GSA, and BSA were designed to determine the optimum number of neurons in each hidden layer and the learning rate of ANN. A high level of distance estimation accuracy was achieved that was better than in the previous works.

Sensors 2016, 16, 1043

3 of 23

2. Related Cycling Localization Techniques Several studies adopting different approaches for wireless positioning on a bicycle track have been conducted. The location and speed of an athlete were monitored in [32] using GPS as part of a server-based mobile coaching system (MCS). An MCS was used in the Advanced and Adaptive Network Technology (ANT) wireless protocol to monitor the biomechanical and physiological parameters of the bicycle and the cyclist, respectively. In research work [33], a remote mobile monitoring system (RMMS) was implemented using a ZigBee WSN. This RMMS was able to monitor physiological and biomechanical parameters using several sensors that transmitted their data to the coordinator node, which communicated with a small laptop that was fixed to the bicycle handle bar. Bicycle location was estimated using GPS, which was installed on the bicycle laptop. Location information was sent to a remote server to be monitored by a remote user. In [34], the authors combined a wheel sensor and a compass sensor to determine bicycle location instead of using GPS in order to reduce the cost and the power consumption of the bicycle area network (BAN). In particular, a bicycle may be used in places other than common cycling environments (for example in dense forests and between tall buildings in a city), where GPS is not useful [34]. However, this method is not sufficiently accurate to use in place of GPS. Zhan et al. [35] used GPS and a cellphone to track bicycle trips in outdoor environments. The proposed system saved a significant amount of energy (57%) by adjusting the duty cycling of the GPS. A similar approach (i.e., GPS) was also used in several studies on cycling positioning [36–38]. However, the GPS method is inefficient for indoor positioning, although it is useful in outdoor environments when LOS between a satellite and a receiver is available without any barriers [39]. In addition, GPS methods are affected by several factors that result in high localization errors of 1–30 m [40]. Moreover, GPS has other limitations, such as a poor battery life, and it is very expensive [41] when deployed in large quantities. Wireless bicycle sensor nodes with a Wi-Fi infrastructure have been proposed by Pias et al. [42] to estimate bicycle location in Cambridge city centre (2.5 km ˆ 2.5 km). The proposed system consisted of three subsystems: mobile nodes, the central server, and the communication infrastructure. The mobile nodes communicated with the trackside servers, which in turn transferred the data to the central server to determine bicycle location on the track. However, localization with Wi-Fi is an expensive and impractical method that requires a huge infrastructure, such as a large number of Wi-Fi clients, routers and gateways in a city. Shin et al. [43] used a ZigBee WSN to monitor the position of a cyclist on a cycle path. A path length of approximately 13 km was considered in their study. A total of 30 wireless router nodes were distributed along the cycle path. The router nodes along the path constantly recorded the bicycle position and the biomedical parameters of the cyclist. Bicycle location was determined based on the communication between the bicycle node and the router nodes. The router nodes transmitted the unique Medium Access Control (MAC) address of the bicycle and the router ID to the server through a gateway, whereby the latitude and longitude of the routers and the gateways were known in advance and recorded in the database of the server. Thus, the router node should be aware of the position of the bicycle, which would be recorded to determine bicycle location on the path. The authors stipulated that approximately 18 s would be required to transfer the data from the bicycle to the server because of the considerable number of routers along the path. During this time delay, the bicycle could communicate with the next router, and the location of the bicycle would be recorded in the database along with that of the previous router. Consequently, a positioning error would occur. Because of the aforementioned challenges and limitations for localization accuracy, the proposed soft computing techniques based on optimization algorithms (i.e., PSO, GSA, and BSA) were introduced in this research to improve the distance estimation accuracy of the bicycle moving along the cycle track. 3. Mobile Node System Model The proposed bicycle wireless sensor network (BWSN) topology is shown in Figure 2. The topology consisted of three ZigBee (XBee Series 2) anchor nodes, one ZigBee router node (RN), and

Sensors 2016, 16, 1043

4 of 23

two ZigBee sensor nodes (SNs). The RN and SNs, which moved along with the bicycle, are hereafter called the “mobile node”. The anchor nodes, namely AN1, AN2, and AN3, were mounted on a surface 1.5 m above the ground to avoid ground signal reflection or the Fresnel zone [44], while the router Sensors 16, 1043 4 of 22 node was2016, fixed under the bicycle seat, which was 0.85 m above the ground. AN1, which was located at the centre of the track cycling field, handled the reception of the bicycle parameters (i.e., speed, at the centre of the track cycling field, handled the reception of the bicycle parameters (i.e., speed, cadence, and torque) and sent a beacon signal to the mobile node for distance measurement. The data cadence, and torque) and sent a beacon signal to the mobile node for distance measurement. The data received by AN1 were displayed on the coach’s laptop or tablet to allow the coach to monitor the received by AN1 were displayed on the coach’s laptop or tablet to allow the coach to monitor the cyclist’s performance. AN1 could be connected to the coach’s laptop or any smart device, thus did cyclist’s performance. AN1 could be connected to the coach’s laptop or any smart device, thus did not encounter any issue related to power consumption. The AN2 and AN3 were powered from the not encounter any issue related to power consumption. The AN2 and AN3 were powered from the main AC source. Both AN2 and AN3 were not connected to AN1 but they operated independently. main AC source. Both AN2 and AN3 were not connected to AN1 but they operated independently. TheThe anchor nodes AN3were were2727mm away from AN1, as shown in Figure 3. Thenodes anchor anchor nodesAN2 AN2 and and AN3 away from AN1, as shown in Figure 3. The anchor nodes AN2 and AN3 were fixed at the north and south sides of the track to establish a consistent AN2 and AN3 were fixed at the north and south sides of the track to establish a consistent communication link with thethe bicycle node. AN2 andand AN3 could notnot be located at the easteast or west sides communication link with bicycle node. AN2 AN3 could be located at the or west because the mobile the track have been theat farthest distance (130 (130 m) from them, sides because the node mobileonnode on thewould track would have at been the farthest distance m) from which would have resulted in a loss of communication. AN1, AN2 and AN3 sent a beacon signal them, which would have resulted in a loss of communication. AN1, AN2 and AN3 sent a beacon to thesignal mobile for distance estimation were conducted both outdoor and tonode the mobile node for distancepurposes. estimationExperiments purposes. Experiments were in conducted in both indoor environments. outdoor and indoor environments. Anchor node

Mobile node (located on the bicycle) Router node

Speed/cadence Sensor node

Torque sensor node

AN2

Wireless link

Beacon

Master AN1 node Wireless link

Data Transfer

AN3 Anchor node

Figure2.2. BWSN BWSN topology. topology. Figure

Sensors 2016, 16, 1043

Coach tablet computer

5 of 22

3.1. Outdoor Experiment The outdoor experiment was performed on a track cycling field (the Cheras velodrome, located in the middle of Kuala Lumpur, Malaysia). AN1 was located at the centre of the cycling field. The anchor nodes AN2 and AN3 were 27 m away from AN1 and the mobile node received the beacons North East from the anchor nodes to collect the RSSI data, as shown in Figure 3. The track cycling field area was 130 m × 65 m and AN1(coach) 333 m long. The minimum and maximum distances between AN1 and the mobile Mobile node AN2 Half65 A m (length of the velodrome), node were 32 m (width of the velodrome) and (bicycle)respectively, as measured from the centre of the field (Figure 3). Half B AN3 into two symmetrical halves (A and B), and the The bicycle track area was divided measurements were only performed for the first half (A). Half B was excluded due to the resemblance West between the two halves. Half A was selected because it represented the starting line (pursuit line) for South cycle training and competitions. The mobile node used the RSSI values to determine the distance between itself and AN1 according to the soft computing techniques. The RSSI was measured for 18 pre-defined positions in half A. A total of 20 samples were recorded at each position, with one sample per second. Each sample contained one data packet, and each data packet contained 34 bytes. Therefore, 900 samples (300 for each anchor node) of RSSI values of the three anchor nodes were collected and used for training, testing, and validating ANN and ANFIS to determine the distance to AN1. The collected RSSI values were used for training (70%) [45,46], testing (15%), and validating Figure 3. The outdoor experimental setting at the track cycling field (Cheras velodrome). (15%) ANN and ANFIS to establish a relationship input RSSI values and the predicted Figure 3. The outdoor experimental setting at between the track the cycling field (Cheras velodrome). physical distance. The measured RSSI values at the mobile node for three anchor nodes can be plotted -50 with respect to the number of samples in the outdoor velodrome, as shown in Figure 4. -55

(dBm)

-60 -65

Sensors 2016, 16, 1043

5 of 23

3.1. Outdoor Experiment

Sensors 2016, 16, 1043

5 of 22

The outdoor experiment was performed on a track cycling field (the Cheras velodrome, located in the middle of Kuala Lumpur, Malaysia). AN1 was located at the centre of the cycling field. The anchor nodes AN2 and AN3 were 27 m away from AN1 and the mobile node received the beacons from the anchor nodes to collect the RSSI data, as shown in Figure 3. The track cycling field area was 130 North m ˆ 65 m and 333 m long. The minimum and maximum distances between AN1 and the mobile East node were 32 m (width of the velodrome) and 65 m (length of the velodrome), respectively, as measured from the centre ofAN1(coach) the field (Figure 3). Mobile node AN2 Half A The bicycle track area was divided into two symmetrical halves (A and B), and the measurements (bicycle) were only performed for the first half (A). Half B was excluded due to the resemblance between the Half B two halves. Half A was selected becauseAN3 it represented the starting line (pursuit line) for cycle training and competitions. The mobile node used the RSSI values to determine the distance between itself and West AN1 according to the soft computing techniques. The RSSI was measured for 18 pre-defined positions in half A. A total of 20 samples were recorded at each position, Southwith one sample per second. Each sample contained one data packet, and each data packet contained 34 bytes. Therefore, 900 samples (300 for each anchor node) of RSSI values of the three anchor nodes were collected and used for training, testing, and validating ANN and ANFIS to determine the distance to AN1. The collected RSSI values were used for training (70%) [45,46], testing (15%), and validating (15%) ANN and ANFIS to establish a relationship between the input RSSI values and the predicted physical distance. The measured RSSI values at the mobile node for three anchor nodes can be plotted with respect to the number of samples in the outdoor velodrome, as shown in Figure 4. Figure 3. The outdoor experimental setting at the track cycling field (Cheras velodrome). -50 -55

RSSI (dBm)

-60 -65 -70 -75 RSSI 1 RSSI 2 RSSI 3

-80 -85

0

50

100

150

200

250

300

Samples Figure values at at the themobile mobilenode nodefor forthree threeanchor anchornodes nodes a number of samples Figure 4. 4. Measured Measured RSSI values onon a number of samples for for each anchor node (300 samples) in the outdoor velodrome. each anchor node (300 samples) in the outdoor velodrome.

3.2. 3.2. Indoor IndoorExperiment Experiment The The indoor indoor experiment experiment was was performed performedin in the the sports sports hall hall (Dewan (Dewan Gemilang) Gemilang) of of the the University University Kebangsaan Malaysia, Bangi (Figure 5) to represent a quarter of the cycle track area as Kebangsaan Malaysia, Bangi (Figure 5) to represent a quarter of the cycle track area as no no indoor indoor velodrome is available in the country. The velodrome geometry is symmetrical; therefore the velodrome is available in the country. The velodrome geometry is symmetrical; therefore the considered considered area was resemblance the closest resemblance quarter an indoor velodrome. to area indoor areaindoor was the closest to a quarterto ofaan indoorofvelodrome. Due to areaDue restrictions, restrictions, the maximum number of points was reduced to 13 points. In addition, furniture and the maximum number of points was reduced to 13 points. In addition, furniture and objects were objects were excluded from the measurement area, which was almost similar to that of an actual excluded from the measurement area, which was almost similar to that of an actual velodrome. The velodrome. of 36 m × 34 Given the length of not thisthe building was not building hadThe an building area of 36had m ˆan 34area m. Given that them. length of that this building was same as that of the same as that of a velodrome, a diagonal distance was considered to obtain the maximum distance a velodrome, a diagonal distance was considered to obtain the maximum distance between the mobile between the mobile node and AN1 (coach), as shown in Figure 5. AN1 was located at one corner of the hall and anchor nodes AN2 and AN3 were 15 m and 40 m away from AN1, as shown in Figure 5. The RSSI values of AN1, AN2, and AN3 were collected by a mobile node from 13 pre-defined locations. A total of 20 samples were recorded at each position with one sample per second. Each

Sensors 2016, 16, 1043

6 of 23

node and AN1 (coach), as shown in Figure 5. AN1 was located at one corner of the hall and anchor nodes AN2 and AN3 were 15 m and 40 m away from AN1, as shown in Figure 5. The RSSI values of AN1, AN2, and AN3 were collected by a mobile node from 13 pre-defined locations. A total of 20 Sensors 2016, 16, 1043 6 of 22 Sensors 2016, 16, 1043 of 22 samples were recorded at each position with one sample per second. Each sample contained one6data packet, and each data packet contained 34 bytes. Therefore, 780 samples (260 for each anchor node) of sample contained one data packet, and each data packet contained 34 bytes. Therefore, 780 samples sample contained one data packet, and each data packet contained 34 bytes. Therefore, 780 samples RSSI values the three anchor nodes were of collected andanchor used for training, validating (260 for eachofanchor node) of RSSI values the three nodes were testing, collectedand and used for (260 for each anchor node) of the RSSIdistance values to of AN1. the three anchor nodes were used collected and used for ANN and ANFIS to determine The collected RSSI were for training (70%), training, testing, and validating ANN and ANFIS to determine the distance to AN1. The collected training, testing, and validating ANN andand ANFIS to determine the distance to AN1. The collected testing (15%), and (15%) ANN ANFIS to establish a relationship input RSSI were used forvalidating training (70%), testing (15%), and validating (15%) ANN and between ANFIS tothe establish RSSI were used for training (70%), testing (15%), and validating (15%) ANN and ANFIS to establish RSSI values and the predicted physical distance. The measured RSSI values at the mobile node for a relationship between the input RSSI values and the predicted physical distance. The measured RSSI athree relationship between the input RSSI values and the predicted physical distance. The measured RSSI anchor nodes can be plotted with respect to the number of samples in the indoor velodrome as values at the mobile node for three anchor nodes can be plotted with respect to the number of samples values at the mobile node for three anchor nodes can be plotted with respect to the number of samples shown in Figure 6. in the indoor velodrome as shown in Figure 6. in the indoor velodrome as shown in Figure 6.

RSSI RSSI(dBm) (dBm)

Figure 5. 5. Sports Sports hall hall (Dewan (Dewan Gemilang) Gemilang) for for the the indoor indoor experiment. experiment. Figure Figure 5. Sports hall (Dewan Gemilang) for the indoor experiment. -35 -35 -40 -40 -45 -45 -50 -50 -55 -55 -60 -60 -65 -65 -70 -70 -75 -75 -80 -80 0 0

RSSI 1 RSSI 1 RSSI 2 RSSI RSSI 23 RSSI 3

30 30

60 60

90 90

120 120

150 150

Samples Samples

180 180

210 210

240 260 240 260

Figure 6. Measured RSSI values at the mobile node for the three anchor nodes with respect to a Figure 6. at at thethe mobile node for the anchor nodes nodes with respect to a number Figure 6. Measured MeasuredRSSI RSSIvalues values mobile node forthree the three anchor with respect to a number of samples for each anchor node (260 samples) in the indoor velodrome. of samples for each anchor node (260 samples) in the indoor velodrome. number of samples for each anchor node (260 samples) in the indoor velodrome.

4. Soft Computing-Based Localization Techniques 4. Soft Computing-Based Localization Techniques Soft computing techniques are employed to solve complex numerical optimization problems as Soft computing techniques are employed to solve complex numerical optimization problems as well as non-linear and non-differentiable systems [47]. There are several categories of soft computing, well as non-linear and non-differentiable systems [47]. There are several categories of soft computing, for example ANFIS, ANN, Support Vector Machine (SVM), Fuzzy Logic (FL), and Optimization for example ANFIS, ANN, Support Vector Machine (SVM), Fuzzy Logic (FL), and Optimization Algorithms (OA) [47]. Each category of soft computing also has a fine-grained set; for example, PSO, Algorithms (OA) [47]. Each category of soft computing also has a fine-grained set; for example, PSO, GRA, BSA, GA, BFA, Artificial Bee Colony (ABC), Ant Colony Optimization (ACO), and Differential

Sensors 2016, 16, 1043

7 of 23

Sensors 2016, 16, 1043

7 of 22

4. Soft Computing-Based Localization Techniques GSA, and BSA with ANN backpropagation (BP) to accurately estimate the distance between bicycles Soft computing techniques employed solve complextechniques numericalwill optimization problems as on the cycle track and the coach.are The selected to soft computing be explained in detail well non-linear and non-differentiable systems [47]. There are several categories of soft computing, in theasfollowing subsections. for example ANFIS, ANN, Support Vector Machine (SVM), Fuzzy Logic (FL), and Optimization 4.1. ANFIS Techniques Algorithms (OA) [47]. Each category of soft computing also has a fine-grained set; for example, PSO, GRA,ANFIS BSA, GA, Artificial Colony (ABC), Ant Colony Optimization Differential is a BFA, technique for Bee the evolution of self-organizing neuro-fuzzy (ACO), systemsand [48]. ANFIS is Evolution (DE) are in non-linear the OA category. Thisalgorithms, paper focuses on ANFIS and is thethe hybridization of data. PSO, employed to perform estimation which in this case collected RSSI GSA, and BSA with ANN backpropagation (BP) to accurately estimate the distance between bicycles ANFIS has been used in research [26,49–53] to estimate the location of the nodes or distance between on the in cycle trackInand coach. selected soft computing techniques be explained in detail in nodes WSNs. thisthe work, theThe ANFIS structure shown in Figure 7 waswill adopted. The ANFIS input thethe following subsections. is three RSSI values from three anchor nodes AN1, AN2, and AN3 for outdoor and indoor as

shown in Figures 4 and 6. 4.1. ANFIS Techniques For each input, three, five, and seven inputs membership function (mfs) were trained and tested. is membership a technique for the evolution self-organizing neuro-fuzzy systems [48]. ANFIS is Two ANFIS types of function, namelyofthe triangle membership (trimf) and generalized bell employed to perform non-linear estimation algorithms, which in this case is the collected RSSI data. membership (gbellmf), were also used in ANFIS. Different numbers and types of membership ANFIS been used inminimum research [26,49–53] to estimate oftesting, the nodes distance between functionhas give different distance error values.the Forlocation training, andorvalidating ANFIS, WSNs.of InRSSI this work, theare ANFIS structure shown in(outdoor) Figure 7 was adopted. The(indoor) ANFIS input is anodes largein number samples required. 900 samples and 780 samples of RSSI the three RSSI values from three anchor nodes AN1, AN2, and AN3 for outdoor and indoor as shown values were used for training, testing, and validating ANFIS to accurately determine the distance in Figuresthe 4 and 6. on the track and the coach. between bicycle

Figure gbellmfs. Figure 7. 7. The The adopted adopted ANFIS ANFIS structure structure for for seven seven gbellmfs.

4.2. ANN Techniques For each input, three, five, and seven inputs membership function (mfs) were trained and tested. is membership an information processing system whichmembership has been developed a generalized Two ANN types of function, namely the triangle (trimf ) andasgeneralized bell mathematical model),ofwere human nervesDifferent [54]. ANN-based arefunction able to membership (gbellmf alsobiological used in ANFIS. numbers localization and types oftechniques membership model the complex mathematical between the input variables in current work) and give different minimum distancerelationship error values. For training, testing, and (RSSI validating ANFIS, a large target variable In this work,900 a BPsamples neural network Levenberg–Marquardt (LM) number of RSSI(distance). samples are required. (outdoor)type andand 780 the samples (indoor) of RSSI values training algorithm were selected for training, testing, and validating. The LM training algorithm were used for training, testing, and validating ANFIS to accurately determine the distance betweenwas the selected because it gives minimum bicycle on the track and the coach. localization error, as proven in [55], in addition to its speed and efficiency. However, the LM algorithm requires a considerable amount of working memory [56]. For 4.2. ANNtesting, Techniques training, and validating ANN, the same amounts of RSSI values which were used in ANFIS couldANN be used to accurately determine the distance between bicycle onas theatrack and the is for an ANN information processing system which has beenthe developed generalized coach for outdoor and indoor velodromes. In the ANN parameters the number of inputs, number mathematical model of human biological nerves [54]. ANN-based localization techniques are able of to hidden layers, number of neurons in each hidden layer, rate, and theinnumber outputs model the complex mathematical relationship between the learning input variables (RSSI current of work) and must determined before testing, and validating paper, a four layer ANN targetbe variable (distance). Intraining, this work, a BP neural network the typedata. and In thethis Levenberg–Marquardt (LM) architecture was built to determine the distance from the mobile node to the coach, based on the RSSI measurements of the anchor nodes. These layers included an: (i) input layer; (ii) first hidden layer; (iii) second hidden layer; and (iv) output layer, as shown in Figure 8. The input layer consisted of

Sensors 2016, 16, 1043

8 of 23

training algorithm were selected for training, testing, and validating. The LM training algorithm was selected because it gives minimum localization error, as proven in [55], in addition to its speed and efficiency. However, the LM algorithm requires a considerable amount of working memory [56]. For training, testing, and validating ANN, the same amounts of RSSI values which were used in ANFIS could be used for ANN to accurately determine the distance between the bicycle on the track and the coach for outdoor and indoor velodromes. In the ANN parameters the number of inputs, number of hidden layers, number of neurons in each hidden layer, learning rate, and the number of outputs must be determined before training, testing, and validating the data. In this paper, a four layer ANN architecture was built to determine the distance from the mobile node to the coach, based on the RSSI Sensors 2016, 16, 1043 8 of 22 measurements of the anchor nodes. These layers included an: (i) input layer; (ii) first hidden layer; (iii) second layer;the and (iv) anchor output nodes, layer, as shown in Figure The input layer consisted of three RSSI hidden values from three and the neurons in 8. this layer only acted as a buffer three RSSI values from the three anchor nodes, and the neurons in this layer only acted as a buffer for for distributing the input signals RSSIi (i = 1, 2, 3, …, n) to the neurons in the first hidden layer. The distributing input signalsagainst RSSIi (ithe = 1,strengths 2, 3, . . . , n) the neurons in the first layer. Theby input input RSSIithe were weighted of to particular connections wijhidden and summated each RSSI against the strengths particular connections wij and layer summated each neuron neuron ofweighted the first hidden layer to passofthe output of the first hidden to theby neurons of the i were ofsecond the first hidden layerThe to pass theofoutput of thelayer first hidden layer to the neurons of the second hidden hidden layer. inputs the second were weighted against the strengths of particular layer. The inputs of the second layer wereneuron weighted against thelayer strengths of particular connections connections wiz and summated by each of the second to calculate the output yk in the wfourth and summated by each neuron of the second layer to calculate the output y in the fourth layer. layer. The first and second hidden layers used the tansigmoidal activation functions to cover all iz k The first and second hidden layers used the tansigmoidal activation functions to cover all ranges of the ranges of the negative RSSI values, whereas the output layer employed the linear activation functions negative values, whereas output The layerfirst employed the linear activation functions of to number cover theof to coverRSSI the positive values ofthe distance. and second hidden layers consisted positive values of distance. The first and second hidden layers consisted of number of neurons in each neurons in each hidden layer. The number of neurons in each hidden layer and learning rate were hidden layer. The of neurons in each hidden layer and BSA) learning rate were selected based on selected based onnumber the optimization algorithms (PSO, GSA, because this parameter selection the optimization (PSO, and BSA) because this parameter selection was notprovide secure and was not securealgorithms and subject to GSA, the trial-and-error method, which does not always the subject to the trial-and-error method, which does not always provide the optimum solution. The PSO, optimum solution. The PSO, GSA, and BSA algorithms address such a problem by determining the GSA, BSA of algorithms a problem by determining thelearning best number of ANN. neuronsThus, in each best and number neurons address in each such hidden layer and the optimum rate of the hidden layer and the optimum learning rate of ANN. performance can be improved. performance of ANN can be improved. In this case, Thus, these the algorithms couldofbeANN hybridized with ANN Intothis case, these algorithms could bewhich hybridized ANN threePSO-ANN different algorithms, form three different algorithms, were with known as to theform hybrid algorithm, which hybrid were known algorithm, as the hybrid PSO-ANN algorithm,algorithm hybrid GSA-ANN algorithm, hybrid BSA-ANN GSA-ANN and hybrid BSA-ANN through which ANNand was able to achieve a algorithm minimumthrough distancewhich error.ANN was able to achieve a minimum distance error. Input layer RSSIi RSSI 1 (AN1)

RSSI 2 (AN2)

RSSI 3(AN3)

Hidden layer1 wij

Σ

Hidden layer 2 wiz

Output layer

Σ

Σ

Σ

Σ

Σ

Σ

Σ

Σ

Σ

Σ

Σ

Σ

Σ

wik

Σ

yk

Distance between AN1 and moving bicycle on the track

Figure 8. The architecture of the ANN algorithm. Figure 8. The architecture of the ANN algorithm.

4.3. Heuristic Algorithms 4.3. Heuristic Algorithms Heuristic algorithms are a technique which tries to find a good solution (near optimal) at a Heuristic algorithms are a technique which tries to find a good solution (near optimal) at a sensible computational cost without the ability to undertake either optimality or feasibility, or even sensible computational cost without the ability to undertake either optimality or feasibility, or even in in several cases to clarify how close to the optimum solution [57]. Because the classical WSN several cases to clarify how close to the optimum solution [57]. Because the classical WSN localization localization methods provided a high localization error, ANN was adopted in this paper to improve the estimated localization error. Due to the learning capabilities and flexible modelling of ANN, it is possible to perform lesser errors in determining the distance between the coach and the bicycle on the cycle track without detailed knowledge of the surroundings. With sufficient ANN parameters and a large amount of RSSI values, ANN is capable of representing the relationship between inputs (RSSI) and output (distance between coach and the bicycle). The heuristic algorithms, such as PSO,

Sensors 2016, 16, 1043

9 of 23

methods provided a high localization error, ANN was adopted in this paper to improve the estimated localization error. Due to the learning capabilities and flexible modelling of ANN, it is possible to perform lesser errors in determining the distance between the coach and the bicycle on the cycle track without detailed knowledge of the surroundings. With sufficient ANN parameters and a large amount of RSSI values, ANN is capable of representing the relationship between inputs (RSSI) and output (distance between coach and the bicycle). The heuristic algorithms, such as PSO, GSA, and BSA, were hybridized with ANN to determine the optimum ANN parameters (i.e., number of neurons in each hidden layer and the learning rate). Selecting these parameters is not secure and subject to trial and error, which in return gives a high distance estimation error. Using the parameter settings of heuristic algorithms as shown in Table 1, we executed the PSO, GSA, and BSA algorithms and obtained the fitness function for 10, 20, 30, 40, and 50 population sizes. Several population sizes were implemented to allow each algorithm to select the population that could achieve the minimum fitness function and elapsed time. Nevertheless, there is no exact algorithm to provide an accurate result for all optimization problems. Some optimization algorithms provide a better solution for some specific problems compared to others. From the execution of the algorithms, it was revealed that the hybrid GSA-ANN achieved minimum errors compared to the other algorithms. Consequently, a hybrid GSA-ANN was considered in the current work. The training processes of ANN were repeated several times using a large number of epochs (i.e., 1000 iterations) until the error between the actual and predicted distances reached the minimum. The GSA algorithm was proposed in 2009 by Rashedi et al. [57]. This algorithm relies on Newtonian gravity: “Every particle in the universe attracts every other particle with a force that is directly proportional to the product of their masses and inversely proportional to the square of the distance between them” [57]. The GSA mathematical principle is based on the law of Newtonian gravity and the laws of motion as in the following: F“G

M1 M2 R2

(1)

where F is the gravitational force, R is the distance between the first particles mass (M1 ) and second particles mass (M2 ), and G is the constant value of the gravitational. Newton’s Second Law states that acceleration a is inversely proportional to mass M and directly proportional to force F as follows: a“

F M

(2)

Because of the effect of declining gravity, the real value of the “gravitational constant (G)” depends on the real age of the universe. Equation (3) provides the decrease of the gravitational constant with age [58]: ˆ ˙β t0 βă1 (3) G ptq “ G pt0 q ˆ t where G(t) is the gravitational constant at time t and G(t0 ) is the gravitational constant at the first cosmic quantum-interval of time t0 [58]. The positions of the agents are initialized (i.e., the masses are randomly selected within the given search interval). The position of the ith agent can be defined by: ´ ¯ Xi “ Xi1 , . . . . . . , Xid , . . . . . . , Xik , f or i “ 1, 2, . . . , N

(4)

where N is the number of agents, Xid is the position of ith agent in the dth dimension and k is the space dimension. To compute the fitness function of GSA, a root mean square error (RMSE) can be used to

Sensors 2016, 16, 1043

10 of 23

determine the best and the worst fit for each iteration. The computations aim to minimize the problems and determine the masses of each agent as follows [59]: g f n f1 ÿ RMSE “ e e2 n

(5)

i “1

best ptq “ worst ptq “ mi ptq “

min

jPt1,...,Nu

f it j ptq

(6)

f it j ptq

(7)

f iti ptq ´ Worst ptq best ptq ´ Worst ptq

(8)

max

jPt1,...,Nu

m ptq Mi ptq “ ř N i j“1 mi ptq

(9)

where e is the estimated distance error and n is the number of samples. The actual distance was obtained based on measurement, whereas the estimated distance was obtained using the hybrid GSA-ANN. The gravitational constant G at iteration t was computed as follows: G ptq “ G0 ep´αt{Tq

(10)

The total force computation in different directions in the ith agent, the velocity and acceleration calculation, and the position of the agents in the next iteration are as follows: Fijd ptq “ G ptq

¯ M pi ˆ Maj ´ d X j ptq ´ Xid ptq Rij ` ε

Fid ptq “

ÿ

rand j Fijd ptq

(11) (12)

jPKbest,j‰i

aid ptq “

Fid ptq Mi ptq

(13)

vid pt ` 1q “ randi ˆ vid ptq ` aid ptq

(14)

xid pt ` 1q “ xid ptq ` vid pt ` 1q

(15)

The details of the operation of the hybrid GSA-ANN based on the previous equations are shown in the flow chart in Figure 9.

+1 = +1 =

+

× +

(14)

+1

(15)

details SensorsThe 2016, 16, 1043of the operation of the hybrid GSA-ANN based on the previous equations are shown 11 of 23 in the flow chart in Figure 9. START Initialize values for number of hidden layers (N1and N2) and learning rate (LR) using Eq.4, G0 and α Agent = 20, Iteration =1000 i=1

i ≤ agent

Update G, best and worst using Eqs. (6, 7, and 10) Calculate the masses (M) for each agent using Eqs. (8 and 9)

No

Yes Choice N1, N2, and LR i=i+1

Evaluate the fitness for each agentusing RMSE Eq (5)

Run ANN Calculated the fitness function (RMSE)

Calculate the force and total force Using Eqs.(11 and 12) Calculate the acceleration for each agent using Eq. (13) Evaluate the velocity by using Eq. (14) Update the position by using Eq. (15) i=1

n ≤ Iteration

Yes No

i ≤ agent

No

n=n+1

Yes

Select best values of N1, N2, and LR

Choice N1, N2, and LR i=i+1

Run ANN END Calculated the fitness function (RMSE)

Figure 9. Flow chart of hybrid GSA-ANN.

Table 1. Parameters of the heuristic algorithms. Parameter

PSO

GSA

BSA

Population size 3 Iteration c1 and c2 w F G0 and α

10, 20, 30, 40, and 50 100 1.494 0.7 -

10, 20, 30, 40, and 50 100 1, 0.2

10, 20, 30, 40, and 50 100 3 -

5. Results and Discussion In this section the results of the soft computing techniques will be discussed in terms of distance estimation accuracy. The ANFIS results will be presented first, followed by those of the hybrid PSO-ANN, hybrid GSA-ANN, and hybrid BSA-ANN. The outcomes of the hybrid GSA-ANN will be considered for comparison with the previous works because it achieved the minimum distance error relative to the other soft computing techniques.

Sensors 2016, 16, 1043

12 of 23

5.1. ANFIS Techniques As previously mentioned, the collected RSSI values were used for training and testing ANFIS (70% for training, 15% for testing, and 15% for validating) to establish a relationship between the input RSSI values and the output physical distances. As a result, the distance error was obtained for different numbers and types of the membership function for outdoor and indoor, as shown in Table 2. The table shows MAE and RMSE, whereby the minimum error occurred when seven membership functions were selected for each input. The distance estimation accuracy is poor for the indoor scenario case, which is expected due to the presence of multipath scatters and reflectors that dominate an indoor environment. In addition, the gbellmf type is better than trimf for both outdoor and indoor. Therefore, seven gbellmfs were considered in this study for outdoor and indoor velodromes. For each ANFIS input, the membership function is shown in Figures 10 and 11 for outdoor and indoor, respectively. Figure 12 shows the comparison transient characteristics for seven gbellmfs after testing and validating data for outdoor and indoor velodromes. The figure shows that there is no significant difference between the estimated and the actual distance for outdoor and indoor velodromes. The red points represent the estimated distance (FIS out), whereas the blue points and plus signs are the actual distance (testing and validating data). For outdoor environments, most of the estimated distances matched the actual distances, as can be seen in Figure 12a (testing data) and Figure 12b (validation data). However, for indoor environments the difference between the estimated distance and the actual distance was very small, as shown in Figure 12c (testing data) and Figure 12d (validation data). This is due to the same reason mentioned previously, i.e., the presence of multipath scatters and reflectors. Sensors 2016, 16, 1043 12 of 22

0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 -72 -70 -68 -66 -64 -62 -74

mf2 mf3 mf4 mf5 mf6 mf7 1 mf1 mf2 mf3 mf4 mf5 mf6 mf7 mf1 0.81 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0

-80

-75

-70

-65

-60

Degree of membership Degree of membership

12 of 22

mf1 mf2 mf3 mf4 mf5 mf6 mf7 1 mf1 mf2 mf3 mf4 mf5 mf6 mf7 0.81



Sensors 2016, 16, 1043

mf1 mf2 mf3 mf4 mf5 mf6 mf7 1 mf1 mf2 mf3 mf4 mf5 mf6 mf7 0.81 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0

-70

-65

-60

-55

mf1 mf2

mf3

mf4 mf5 mf6 mf7

mf1 mf2

mf3

mf4 mf5 mf6 mf7

RSSI (dBm) -75 2-70 -65 -60 -70 RSSI -65 3 (dBm) -60 -55 RSSI RSSI RSSI (b) 2 (dBm) (c) 3 (dBm) (a) 1 (dBm) (b) (c) (a) Figure 10. ANFIS seven gbellmfs input of outdoor velodrome (a) RSSI 1; (b) RSSI 2; and (c) RSSI 3. Figure 10. ANFIS seven gbellmfs input of outdoor velodrome (a) RSSI 1; (b) RSSI 2; and (c) RSSI 3. Figure 10. ANFIS seven gbellmfs input of outdoor velodrome (a) RSSI 1; (b) RSSI 2; and (c) RSSI 3.

0.81

mf1 mf2 mf3 mf4

mf5 mf6 mf7

mf1 mf2 mf3 mf4

mf5 mf6 mf7

0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0

-70

-60 -50 -70 RSSI -601 (dBm) -50

-40 -40

1 0.81

mf1 mf2 mf3 mf4 mf5 mf6 mf7 mf1 mf2 mf3 mf4 mf5 mf6 mf7

0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0

-70 -70

-65

-60

-55

RSSI (dBm)-55 -65 2-60 RSSI (b) 2 (dBm)

-50 -50


1

-80



RSSI-68 1 (dBm) -74 -72 -70 -66 -64 -62

1 0.81 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0

0 -70 -70

-65

-60 -55 -65 RSSI -603 (dBm) -55

(a) 1 (dBm) (c) 3 (dBm) RSSI RSSI (a) (b) (c) Figure 11. ANFIS seven gbellmfs input of indoor velodrome (a) RSSI 1; (b) RSSI 2; and (c) RSSI 3. Figure 11. ANFIS seven gbellmfs input of indoor velodrome (a) RSSI 1; (b) RSSI 2; and (c) RSSI 3. Figure 11. ANFIS seven gbellmfs input of indoor velodrome (a) RSSI 1; (b) RSSI 2; and (c) RSSI 3.

-50 -50

0 -70

-60

-50

RSSI 1 (dBm) (a)

0.2

Degr

Degr

Degr

0.2

0 -70

-40

-65

-60

-55

RSSI 2 (dBm)

0.2

-50

0 -70

(b)

Sensors 2016, 16, 1043

-65

-60

-55

RSSI 3 (dBm) (c)

-50 13 of 23

Figure 11. ANFIS seven gbellmfs input of indoor velodrome (a) RSSI 1; (b) RSSI 2; and (c) RSSI 3.

(a)

(b)

(c)

(d)

Figure of transient transientcharacteristics characteristicsfor forANFIS ANFIS(a)(a) testing and validation data Figure 12. 12. Comparison Comparison of testing and (b)(b) validation data for for the the outdoor velodrome and (c) testing and (d) validation for the indoor velodrome. outdoor velodrome and (c) testing and (d) validation for the indoor velodrome.

5.2. Hybrid HeuristicTable Algorithms-ANN 2. ComparisonTechniques of distance estimation accuracy based on ANFIS. Several population sizes were simulated in Matlab for the hybrid GSA-ANN, hybrid PSO-ANN, Outdooralgorithm to select the optimum Indoor and hybrid BSA-ANN to allow each heuristic number of neurons in ANFIS Method 3 trimf 5 trimf 7 trimf 3 gbellmf 5 gbellmf 7 gbellmf

MAE (m)

RMSE (m)

MAE (m)

RMSE (m)

2.486 1.588 0.3634 2.335 0.4695 0.022

3.938 2.647 0.907 3.713 0.796 0.062

3.581 1.91 1.004 2.927 1.4269 0.284

4.725 3.05 2.107 3.978 2.476 1.108

5.2. Hybrid Heuristic Algorithms-ANN Techniques Several population sizes were simulated in Matlab for the hybrid GSA-ANN, hybrid PSO-ANN, and hybrid BSA-ANN to allow each heuristic algorithm to select the optimum number of neurons in each hidden layer and learning rate of ANN. Thereby, the minimum fitness function could be achieved. Table 3 shows the neurons of the hidden layers (N1 and N2) and LR, which were obtained from the implementation of each heuristic algorithm in Matlab based on different population sizes for outdoor and indoor velodromes. The fitness functions of the hybrid GSA-ANN, hybrid PSO-ANN, and hybrid BSA-ANN for different population sizes are shown in Figures 13–15, respectively, for outdoor and indoor environments.

Sensors 2016, 16, 1043

13 of 22

each hidden layer and learning rate of ANN. Thereby, the minimum fitness function could be each hidden layer and learning rate of ANN. Thereby, the minimum fitness function could be achieved. Table 3 shows the neurons hidden layers (N1 N2) andfitness LR, which werecould obtained each hidden layer and learning rateof ofthe ANN. Thereby, the and minimum function be achieved. Table 3 shows the neurons of the hidden layers (N1 and N2) and LR, which were obtained from the implementation of each heuristic algorithm in Matlab based on different population sizes achieved. Table 3 shows the neurons of the hidden layers (N1 and N2) and LR, which were obtained from the implementation of each heuristic algorithm in Matlab based on different population sizes for outdoor and indoor velodromes. The fitness functions of the hybrid GSA-ANN, hybrid sizes PSOfrom the implementation of each heuristic in Matlab based on different population for outdoor and indoor velodromes. The algorithm fitness functions of the hybrid GSA-ANN, hybrid PSOANN, and hybrid BSA-ANN for different population sizes are shown in Figures 13–15, respectively, Sensors 2016, 16, 1043 14 of 23 for outdoor and indoor velodromes. The fitness functions of the hybrid GSA-ANN, PSOANN, and hybrid BSA-ANN for different population sizes are shown in Figures 13–15,hybrid respectively, for outdoor and indoor environments. ANN, and hybrid BSA-ANN for different population sizes are shown in Figures 13–15, respectively, for outdoor and indoor environments. for outdoor and indoor environments. 10_agent 3.5 10_agent 3.5 20_agent 20_agent 30_agent 10_agent 3.53 30_agent 40_agent 20_agent 3 40_agent 50_agent 30_agent 3 2.5 50_agent 40_agent 2.5 50_agent 2.52 2 2 1.5 1.5 1.51 4; 1.2342 1 4; 1.2342 1 0.5 4; 1.2342 0.5 0.50 0 0 10 20 30 40 50 60 70 80 90 100 0 0 10 20 30 40 50 60 70 80 90 100 Iteration 0 10 20 30 40 50 60 70 80 90 100 Iteration

Fitness function Fitness function Fitness function


2.5 10_agent 2.5 10_agent 20_agent 2.5 20_agent 30_agent 10_agent 2 30_agent 40_agent 20_agent 2 40_agent 50_agent 30_agent 2 50_agent 40_agent 1.5 50_agent 1.5 1.5 1 1 1 3; 0.1742 0.5 3; 0.1742 0.5 3; 0.1742 0.5 0 0 0 10 20 30 40 50 60 70 80 90 100 0 0 10 20 30 40 50 60 70 80 90 100 Iteration 0 10 20 30 40 Iteration 50 60 70 80 90 100

(a) (b) Iteration Iteration (a) (b) (a) (b) Figure 13. GSA fitness function versus iteration for (a) outdoor and (b) indoor velodromes.

Figure Figure13. 13. GSA GSA fitness fitnessfunction functionversus versusiteration iterationfor for(a) (a)outdoor outdoorand and(b) (b)indoor indoorvelodromes. velodromes. Figure 13. GSA fitness function versus iteration for (a) outdoor and (b) indoor velodromes. 3 3 3 2.5 2.5 2.5 2 2 2 1.5 1.5 1.5 1 1 15; 1.2453 15; 1.2453 1 0.5 15; 1.2453 0.5 0.5 0 0 0 10 20 30 40 50 60 70 50 60 70 0 0 10 20 30 40 Iteration 0 10 20 30 40 Iteration 50 60 70



1 10 swarms 1 0.9 10swarms swarms 20 1 0.9 20swarms swarms 30 swarms 10 0.8 30swarms swarms 0.9 40 swarms 20 0.8 40swarms swarms 0.7 50 swarms 30 0.8 0.7 swarms 4050swarms 0.6 0.7 50 swarms 0.6 0.5 0.6 0.5 0.4 0.5 0.4 0.3 0.4 0.3 0.2 0.3 0.2 0.1 12, 0.2 0.1 12, 0 0.1 0 0 10 2012, 30 40 50 60 70 80 90 100 0 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40Iteration 50 60 70 80 90 100 Iteration

10 swarms 10swarms swarms 20 20swarms swarms 10 30 swarms 30swarms swarms 20 40 swarms 40swarms swarms 30 50 swarms swarms 4050swarms 50 swarms

80 90 100 80 90 100 80 90 100

Iteration Iteration (a) (b) (a) (b) (b) Figure 14. PSO(a) fitness function versus iteration for (a) outdoor and (b) indoor velodromes.

Figure 14. PSO fitness function versus iteration for (a) outdoor and (b) indoor velodromes. Figure 14. PSO fitness function versus iteration for (a) outdoor and (b) indoor velodromes. Figure 14. PSO fitness function versus iteration for (a) outdoor and (b) indoor velodromes. 3 10_popsize 3 10_popsize 20_popsize 3 2.5 20_popsize 10_popsize 30_popsize 2.5 30_popsize 20_popsize 40_popsize 2.52 40_popsize 30_popsize 50_popsize 2 50_popsize 40_popsize 2 50_popsize 1.5 1.5 1.51 12; 1.2431 1 12; 1.2431 1 0.5 12; 1.2431 0.5 0.50 0 0 10 20 30 40 50 60 70 80 90 100 50 60 70 80 90 100 0 0 10 20 30 40 Iteration 0 10 20 30 40 Iteration 50 60 70 80 90 100



3 10_popsize 3 10_popsize 20_popsize 3 2.5 20_popsize 30_popsize 10_popsize 2.5 30_popsize 40_popsize 20_popsize 2.52 40_popsize 50_popsize 30_popsize 50_popsize 2 40_popsize 50_popsize 2 1.5 1.5 1.51 1 1 76; 0.1742 0.5 76; 0.1742 0.5 76; 0.1742 0.50 0 0 10 20 30 40 50 60 70 80 90 100 50 60 70 80 90 100 0 0 10 20 30 40 Iteration 0 10 20 30 40Iteration 50 60 70 80 90 100

(a) (b) Iteration Iteration (a) (b) Figure 15. BSA (a) fitness function versus iteration for (a) outdoor and (b) (b) indoor velodromes.

Figure 15. BSA fitness function versus iteration for (a) outdoor and (b) indoor velodromes. Figure 15. BSA fitness function versus iteration for (a) outdoor and (b) indoor velodromes. Figure 15. BSA fitness function versus iteration for (a) outdoor and (b) indoor velodromes.

Sensors 2016, 16, 1043

15 of 23

Table 3. Neurons in each hidden layer and learning rate of ANN based on heuristic algorithms for different population sizes. Population Size

GSA-ANN

Parameters

PSO-ANN

BSA-ANN

Outdoor

Indoor

Outdoor

Indoor

Outdoor

Indoor

10

N1 N2 LR

19 19 0.2541

15 17 0.3429

12 18 0.1877

9 16 0.4477

14 19 0.2505

11 18 0.6652

20

N1 N2 LR

17 17 0.8004

10 16 0.7394

18 16 0.0709

7 18 0.5864

10 19 0.3492

15 18 0.223

30

N1 N2 LR

6 12 0.5864

14 11 0.4764

17 18 0.4533

14 14 0.5027

15 19 0.8871

15 19 0.6608

40

N1 N2 LR

18 16 0.5947

10 14 0.5152

17 16 0.6127

15 16 0.4206

17 17 0.7139

8 17 0.9743

50

N1 N2 LR

14 13 0.5487

13 10 0.535

11 18 0.9407

16 19 0.7194

16 18 0.4737

17 19 0.4655

N1: number of neurons in hidden layer 1, N2: number of neurons in hidden layer 1, and LR: learning rate of ANN. Sensors 2016, 16, 1043 14 of 22

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

BSA GSA PSO

1.5

10

20

1.4 1.35 1.3 1.25

3; 0.1742 0

BSA GSA PSO

1.45

Fitness function

Fitness function

Once the the adopted adopted population population size size was was tested tested for for each each heuristic heuristic algorithm, algorithm, itit was was found found that that the the Once 20 population size gave the minimum fitness function (based on RMSE) for all heuristic algorithms 20 population size gave the minimum fitness function (based on RMSE) for all heuristic algorithms for the the outdoor outdoor velodrome, velodrome, whereas whereas the the population population size size was was 20 20 for for the the hybrid hybrid PSO-ANN PSO-ANN and and hybrid hybrid for BSA-ANN and and 50 50 for forthe thehybrid hybridGSA-ANN, GSA-ANN, which whichgave gavethe theminimum minimumfitness fitnessfunction function for forthe theindoor indoor BSA-ANN velodrome. On the other hand, GSA-ANN achieved the minimum fitness function compared to the the velodrome. On the other hand, GSA-ANN achieved the minimum fitness function compared to other heuristic outdoor andand indoor. GSA-ANN was able reach bestthe fitness other heuristic algorithms algorithmsfor for outdoor indoor. GSA-ANN was to able to the reach bestfunction fitness at 20 (outdoor) and 50 (indoor) population sizes, as shown in Figure 16a,b, respectively. Based on the function at 20 (outdoor) and 50 (indoor) population sizes, as shown in Figure 16a,b, respectively. outcomes of the hybrid GSA-ANN, ANN was trained, tested, and validated using the parameters Based on the outcomes of the hybrid GSA-ANN, ANN was trained, tested, and validated using the that achieved theachieved minimum function (i.e., N1 = N2 = 17, LR= =17, 0.8004 parameters that thefitness minimum fitness function (i.e., N1and = N2 and for LR outdoor = 0.8004 and for N1 = 13, N2 = 10, and LR = 0.535 for indoor) as shown in Table 3. These parameters improved the outdoor and N1 = 13, N2 = 10, and LR = 0.535 for indoor) as shown in Table 3. These parameters ANN operation during the training, testing, and validation phases, which resulted in a high distance improved the ANN operation during the training, testing, and validation phases, which resulted in accuracy. aestimation high distance estimation accuracy.

30

40

50

60

70

80

90 100

Iteration

4;

1.2 0

10

20

30

40

50

60

70

80

90 100

Iteration (b)

(a)

Figure function comparison of GSA, BSAand for BSA (a) outdoor (b) indoor Figure 16. 16.Fitness Fitness function comparison of PSO, GSA,and PSO, for (a)and outdoor and velodromes. (b) indoor velodromes. Table 3. Neurons in each hidden layer and learning rate of ANN based on heuristic algorithms for different population sizes. Population Size

Parameters N1

GSA-ANN Outdoor Indoor 19 15

PSO-ANN Outdoor Indoor 12 9

BSA-ANN Outdoor Indoor 14 11

Sensors 2016, 16, 1043

16 of 23

Figure 17 shows the testing and validation data of ANN. The actual distance (i.e., the target on the x-axis) for the outdoor and indoor velodrome is plotted against the estimated distance Sensors 2016, 16, 1043 15 of 22 (i.e., output on the y-axis). The regression coefficient (R) of determination between the actual and distance a good indicator for the investigation of the prediction performance of the hybrid GSAestimatedisdistance is a good indicator for the investigation of the prediction performance of the hybrid ANN algorithm. From From FigureFigure 17a,b 17a,b for the outdoor velodrome, the Rthe values were were 0.9999 (testing) and GSA-ANN algorithm. for the outdoor velodrome, R values 0.9999 (testing) 0.99991 (validation). For For Figure 17c,d for for thethe indoor velodrome, the R Rvalues and 0.99991 (validation). Figure 17c,d indoor velodrome, the valueswere were0.99606 0.99606(testing) (testing) and and 0.99587 0.99587 (validation). (validation). The The regression regression coefficient coefficient results results suggested suggested aa close close agreement agreement between between the the actual and estimated distances. actual and Training: R=0.99991

Training: R=0.9999 Data Fit Y= T

60

Output ~= 1*Target + 0.0085

Output ~= 1*Target + 0.0095

60

Data Fit Y= T

55 50 45 40 35

55 50 45 40 35

35

40

45

50

55

60

35

40

45

Target

(a)

60

Training: R=0.99587

Data Fit Y= T

45

Output ~= 0.99*Target + 0.22

Output ~= 0.99*Target + 0.21

40

55

(b)

Training: R=0.99606 45

50

Target

35 30 25 20 15 10 5

40

Data Fit Y= T

35 30 25 20 15 10 5

10

20

Target

(c)

30

40

10

20

30

40

Target

(d)

Figure 17. The performance of the hybrid GSA-ANN algorithm for (a) testing and (b) validation data Figure 17. The performance of the hybrid GSA-ANN algorithm for (a) testing and (b) validation data for the outdoor velodrome and (c) testing and (d) validation for the indoor velodrome. for the outdoor velodrome and (c) testing and (d) validation for the indoor velodrome.

The hybrid GSA-ANN was used to optimize the ANN operation by selecting the optimum The was used optimize ANN operation optimum values of hybrid neuronsGSA-ANN in each hidden layer to and learningthe rate. Although GSAby is selecting a famous the optimization values of neurons in each hidden layer and learning rate. Although GSA is a famous optimization technique [60] it has not been hybridized with ANN in the previous literature to address the technique [60] it has notinbeen hybridized with ANN in the the previous literature to address the localization problems WSNs. In the current work, GSA technique outperforms PSOlocalization and BSA, problemsitinachieves WSNs. aIn theconvergence current work, the GSA technique PSO and4 BSA, whereby whereby fast with a lesser number ofoutperforms iterations, i.e., 3 and iterations and a minimum RMSE of 0.1742 and 1.2342 for outdoor and indoor, as shown in Figure 16a,b, respectively. Consequently, a lesser computational time was accomplished for convergence through this algorithm relative to other adopted algorithms. ANN was trained, tested, and validated offline using the measured RSSI values as inputs and the actual distance between the coach and the bicycle

Sensors 2016, 16, 1043

16 of 22

on the cycle track as output. Since the training, testing, and validating of ANN was done offline and not implemented in real time, there were no issues related to delay characteristics. One advantage of the proposed hybrid GSA-ANN is that it does not need training for any new distance estimation Sensors 2016, 16, 1043 17 of 23 when applied in real time. Once the model is trained offline, any new inputs of RSSI values from the anchor nodes can be given in real time to the neural network and the corresponding distance it achieves a can fast be convergence with a lesser numbertime of iterations, i.e.,phase 3 andwas 4 iterations and a estimation obtained. However, the training in the offline 23 s for BSA-ANN minimum RMSE of 0.1742 and 1.2342 for outdoor and indoor, as shown in Figure 16a,b, respectively. and 33 s for GSA-ANN and PSO-ANN. The training time of the adopted hybrid heuristic algorithms Consequently, a lesser time was accomplished forofconvergence through this[61], algorithm have performed thecomputational Bayesian Regularization (BR) algorithm ANN in research work whereby relative to other adopted ANN was trained, tested, and validated offline using the the training time was 751algorithms. s. measured RSSI values as inputs and the actual distance between the coach and the bicycle on the 6. Results cycle track asComparison output. Since the training, testing, and validating of ANN was done offline and not implemented in real time, there were no issues related to delay characteristics. One advantage of the Figure 18 shows the comparison of the distance estimation error between the soft computing proposed hybrid GSA-ANN is that it does not need training for any new distance estimation when techniques considered in the current work in terms of MAE and RMSE for outdoor and indoor applied in real time. Once the model is trained offline, any new inputs of RSSI values from the anchor velodromes. Both RMSE and MAE can be used when GSA is hybridized with ANN. RMSE and MAE nodes can be given in real time to the neural network and the corresponding distance estimation can be employed together to detect the variation in the errors in a set of predictions. MAE will always can be obtained. However, the training time in the offline phase was 23 s for BSA-ANN and 33 s be smaller than or equal to RMSE; the greater the difference between them, the greater the variance for GSA-ANN and PSO-ANN. The training time of the adopted hybrid heuristic algorithms have in the discrete errors in the sample. If MAE is equal to RMSE, then all the errors have the same performed the Bayesian Regularization (BR) algorithm of ANN in research work [61], whereby the magnitude. training time was 751 s. It can be observed that there was a small difference between MAE and RMSE for the case of as shown in Figure 18, which was expected due to the lesser signal attenuation. In contrast, 6. outdoor, Results Comparison the gap between RMSE and MAE was very huge, suggesting that the variance error in terms of Figure 18 shows comparison ofThe the figure distance estimation error between soft computing accuracy is large forthe indoor samples. shows that MAE and RMSE the for outdoor are better techniques considered in the current work in terms of MAE and RMSE for outdoor and indoor than indoor for all adopted techniques, whereas the hybrid GSA-ANN was best for outdoor and velodromes. Both RMSE and MAE can be used when GSA is hybridized with ANN. RMSE and indoor velodromes. MAE of hybrid GSA-ANN was improved by 27% and 16% relative toMAE ANFIS canseven be employed to detect the variation in the errors a set ofvelodrome, predictions.whereas MAE will gbellmfs together and hybrid BSA-ANN, respectively, for theinindoor thealways MAE of be hybrid smallerGSA-ANN than or equal to RMSE; the greater the difference between them, the greater the variance in was improved by 21% relative to the hybrid BSA-ANN for the outdoor velodrome. the discrete errors in the sample. If MAE is equal to RMSE, then all the errors have the same magnitude. MAE

1.243

RMSE

1.25 1

0.5

0.75

0.175 0.208

0.25

0.1742

0.087

0.062

0.2066 0.0279

0.022

0.022

0.0218

GSA_ANN (outdoor)

0.2475

PSO_ANN (outdoor)

0.284

ANFIS 7 gebellmf (outdoor)

0.5

BSA_ANN(outdoor)

MAE and RMSE in meters

1.274 1.108

1.5

GSA_ANN (indoor)

PSO_ANN (indoor)

BSA_ANN(indoor)

ANFIS 7 gebellmf (indoor)

0

Soft computing techniques Figure 18. 18. Comparison of the distance estimation error for for adopted softsoft computing techniques for for Figure Comparison of the distance estimation error adopted computing techniques outdoor and indoor velodromes. outdoor and indoor velodromes.

Thebe hybrid GSA–ANN algorithm be difference compared between with previous [9,30,39,46,51,53,61–79] It can observed that there was a can small MAEworks and RMSE for the case of in terms of localization or distance error to validate our proposed system. Similar studies based on outdoor, as shown in Figure 18, which was expected due to the lesser signal attenuation. In contrast, soft computing considered for the purposes of comparison. The of RSSI thedifferent gap between RMSE andtechniques MAE was were very huge, suggesting that the variance error in terms performance was samples. employedThe as figure the input, whereas the and location theoutdoor target are node or the accuracy is largemetric for indoor shows that MAE RMSEoffor better distance the nodes in the network was the usedhybrid as the GSA-ANN output. Most of these studies usedand ANN than indoorbetween for all adopted techniques, whereas was best for outdoor indoor velodromes. MAE of hybrid GSA-ANN was improved by 27% and 16% relative to ANFIS seven gbellmfs and hybrid BSA-ANN, respectively, for the indoor velodrome, whereas the MAE of hybrid GSA-ANN was improved by 21% relative to the hybrid BSA-ANN for the outdoor velodrome.

Sensors 2016, 16, 1043

18 of 23

The hybrid GSA–ANN algorithm can be compared with previous works [9,30,39,46,51,53,61–79] in terms of localization or distance error to validate our proposed system. Similar studies based on different soft computing techniques were considered for the purposes of comparison. The RSSI Sensors 17 Sensors2016, 2016,16, 16,1043 1043 17ofof22 22 performance metric was employed as the input, whereas the location of the target node or the distance between theLogic nodes in the network was used as the output. Mostand of these studies used ANN and Fuzzy and techniques to localization accuracy, some of used optimization andFuzzy Fuzzy Logic techniques toimprove improve localization accuracy, and some ofthem them used optimization Logic techniques to localization accuracy, and some of them used optimization algorithms such PSO Figures 19 and 20 show algorithms such as as improve PSO or or hybrid hybrid techniques. techniques. Figures 19 (outdoor) (outdoor) and 20 (indoor) (indoor)algorithms show the the such as PSO or hybrid techniques. Figure 19 (outdoor) and Figure 20 (indoor) show the comparison comparison error comparisonof ofthe thehybrid hybridGSA-ANN GSA-ANNwith withthose thoseworks. works.The Themean meanlocalization localizationor ordistance distance errorisis of the hybrid GSA-ANN with those The meanbetween localization or distance error is studies. considered a considered aaperformance metric for the our and considered performance metric forworks. thecomparison comparison between ourwork work andprevious previous studies.Our Our performance metric for the comparison between our work and previous studies. Our proposed hybrid proposed proposedhybrid hybridGSA–ANN GSA–ANNalgorithm algorithmoutperforms outperformsthe thealgorithms algorithmsor ortechniques techniquesof ofthe theother otherstudies studies GSA–ANN algorithm outperforms algorithms techniques the other studies with its average with distance errors 0.0218 m m and indoor velodromes, with its its average average distance errors of ofthe 0.0218 m and andor0.2066 0.2066 m for forofoutdoor outdoor and indoor velodromes, distance errors of 0.0218 m and 0.2066 m for outdoor and indoor velodromes, respectively. respectively. respectively.

Mean error error (m), (m), log log Mean

10 10

66 2.2787 2.2787

22 0.9886 0.9886

11

0.8025 0.8025

0.6938 0.6938

0.49 0.49

0.34574 0.34574

0.1 0.1

0.05412 0.05412

0.03976 0.03976 0.022 0.022

0.0218 0.0218

0.01 0.01

Figure Figure19. 19.Comparison Comparisonof ofGSA-ANN GSA-ANNtechniques techniqueswith withprevious previousworks worksfor foroutdoor outdoorenvironments. environments. Figure 19. Comparison of GSA-ANN techniques with previous works for outdoor environments.

Mean error error (m) (m) Mean

4.5 4.5 44 3.5 3.5 33 2.5 2.5 22 1.5 1.5 11 0.5 0.5 00

44

3.04 3.04

33

2.8 2.8 22

22

22 1.65 1.65 1.298 1.298 11

0.99 0.99

0.9 0.9

0.87 0.87 0.43 0.43

0.3 0.3

0.3 0.3

0.208 0.208 0.2066 0.2066

Figure Figure20. 20.Comparison Comparisonof ofGSA-ANN GSA-ANNtechniques techniqueswith withprevious previousworks worksfor forindoor indoorenvironments. environments. Figure 20. Comparison of GSA-ANN techniques with previous works for indoor environments.

7. 7.Conclusions Conclusions 7. Conclusions Two Twosoft softcomputing computingtechniques techniquesfor forWSN WSNdistance distanceestimation estimationwere werepresented presentedin inthis thispaper paperfor for Two soft computing techniques for WSN distance estimation were presented in this paper for outdoor and indoor velodromes. These techniques aimed to determine the distance between the outdoor and indoor velodromes. These techniques aimed to determine the distance between the outdoor and indoor velodromes. These techniques aimed to determine the distance between the bicycle bicycleposition positionwhile whilemoving movingon onthe thecycle cycletrack trackand andthe thecoach. coach.The The first method was based on ANFIS, bicycle position while moving on the cycle track and the coach. Thefirst firstmethod methodwas wasbased basedon onANFIS, ANFIS, whereas whereasthe thesecond secondapproach approachadopted adoptedheuristic heuristicalgorithms algorithmssuch suchas asGSA, GSA,PSO, PSO,and andBSA BSAhybridized hybridized whereas the second approach adopted heuristic algorithms such as GSA, PSO, and BSA hybridized with withANN. ANN.The TheANN ANNalgorithm algorithmwas wasfurther furtherimproved improvedby bycombining combiningwith withGSA, GSA,PSO, PSO,and andBSA BSAto to with ANN. The ANN algorithm was further improved by combining with GSA, PSO, and BSA to select the optimum number of neurons in the adopted two hidden layers and to select the optimum select the optimum number of neurons in the adopted two hidden layers and to select the optimum learning learningrate rateof ofANN. ANN.This Thisresulted resultedin inan animprovement improvementof ofthe thedistance distanceerror errorbetween betweenthe thebicycle bicycleand and the thecoach. coach.The Thesoft softcomputing computingtechniques techniqueswere werecompared comparedwith witheach eachother otherto toselect selectthe thebest bestalgorithm algorithm that thatgave gavethe theminimum minimumdistance distanceerror. error.The Theresults resultsindicated indicatedthat thatthe thehybrid hybridGSA-ANN GSA-ANNwas wasmore more convenient convenientthan thanthe theother otheralgorithms algorithmsadopted adoptedin inthis thiswork workin interms termsof ofdistance distanceestimation estimationaccuracy. accuracy.

Sensors 2016, 16, 1043

19 of 23

select the optimum number of neurons in the adopted two hidden layers and to select the optimum learning rate of ANN. This resulted in an improvement of the distance error between the bicycle and the coach. The soft computing techniques were compared with each other to select the best algorithm that gave the minimum distance error. The results indicated that the hybrid GSA-ANN was more convenient than the other algorithms adopted in this work in terms of distance estimation accuracy. The comparison results showed that the MAE of the hybrid GSA-ANN outperformed the previous works for both outdoor and indoor environments. Therefore, GSA-ANN is suitable in both indoor and outdoor environments and can be applied to any static or mobile WSN node. The limitation of this study lies in the possibility of implementing the hybrid GSA-ANN in real-time as the implementation of ANN requires a considerable amount of memory. The huge computation memory comes at the expense of limited memory size and processor speed of the microcontrollers. In such a case, the mobile node (bicycle) requires a microcontroller with a high speed and large memory size such as an Arduino Due. However, using an Arduino Due leads to high power consumption, large size and extra weight, all of which are considered critical issues in bicycle sensor nodes. The high level of power consumption will eventually reduce the battery life of the sensor nodes. In addition, the extra size and weight on the bike increases aerodynamic resistance, which consequently reduces the bike’s speed and induces fatigue in the athlete during cycling. These considerations are critical in competitive events. It is expected that once quantum computing is in place and Moore’s Law has achieved its saturation in future, the use of high computing will be possible for small sensor applications. Acknowledgments: This work was supported by Universiti Kebangsaan Malaysia with grant ref. INOVASI-2014-015.

No.:

Author Contributions: Sadik Kamel and Rosdiadee Nordin conceived and designed the experiments. Sadik Kamel performed the experiments. Rosdiadee Nordin and Sadik Kamel conducted the data analysis. All authors wrote the paper. Conflicts of Interest: The authors declare no conflict of interest.

References 1. 2. 3. 4.

5.

6. 7. 8. 9.

So-In, C.; Permpol, S.; Rujirakul, K. Soft computing-based localizations in wireless sensor networks. Pervasive Mob. Comput. 2016, 29, 17–37. [CrossRef] Zhang, T.; He, J.; Zhang, Y. Secure sensor localization in wireless sensor networks based on neural network. Int. J. Comput. Intell. Syst. 2012, 5, 914–923. [CrossRef] Blumrosen, G.; Hod, B.; Anker, T.; Dolev, D.; Rubinsky, B. Enhancing RSSI-based tracking accuracy in wireless sensor networks. ACM Trans. Sens. Netw. (TOSN) 2013, 9, 29. [CrossRef] Bhuvaneswari, P.; Vaidehi, V.; Saranya, M.A. Distance based transmission power control scheme for indoor wireless sensor network. In Transactions on Computational Science XI; Springer: Berlin, Germany, 2010; Volume 6480, pp. 207–222. Rossi, M.; Brunelli, D.; Adami, A.; Lorenzelli, L.; Menna, F.; Remondino, F. Gas-drone: Portable gas sensing system on UAVs for gas leakage localization. In Proceedings of the IEEE Sensors 2014, Valencia, Spain, 2–5 November 2014; pp. 1431–1434. Lo, G.; González-Valenzuela, S.; Leung, V.C.M. Wireless body area network node localization using small-scale spatial information. IEEE J. Biomed. Health Inform. 2013, 17, 715–726. [CrossRef] [PubMed] Zhao, J.; Xi, W.; He, Y.; Liu, Y.; Li, X.-Y.; Mo, L.; Yang, Z. Localization of wireless sensor networks in the wild: Pursuit of ranging quality. IEEE/ACM Trans. Netw. (TON) 2013, 21, 311–323. [CrossRef] Karim, L.; Nasser, N.; Mahmoud, Q.H.; Anpalagan, A.; Salti, T.E. Range-free localization approach for M2M communication system using mobile anchor nodes. J. Netw. Comput. Appl. 2015, 47, 137–146. [CrossRef] Gu, S.; Yue, Y.; Maple, C.; Wu, C. Fuzzy logic based localisation in wireless sensor networks for disaster environments. In Proceedings of the 18th International Conference on Automation and Computing (ICAC), Loughborough, UK, 7–8 September 2012; pp. 1–5.

Sensors 2016, 16, 1043

10.

11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28.

29.

30. 31.

32. 33.

20 of 23

Payal, A.; Rai, C.; Reddy, B. Analysis of some feedforward artificial neural network training algorithms for developing localization framework in wireless sensor networks. Wirel. Pers. Commun. 2015, 82, 2519–2536. [CrossRef] Gu, Y.; Lo, A.; Niemegeers, I. A survey of indoor positioning systems for wireless personal networks. IEEE Commun. Surv. Tutor. 2009, 11, 13–32. [CrossRef] Tarrío, P.; Bernardos, A.M.; Casar, J.R. An energy-efficient strategy for accurate distance estimation in wireless sensor networks. Sensors 2012, 12, 15438–15466. [CrossRef] [PubMed] Robles, J.J. Indoor localization based on wireless sensor networks. AEU Int. J. Electron. Commun. 2014, 68, 578–580. [CrossRef] Wen, F.; Liang, C. Fine-grained indoor localization using single access point with multiple antennas. IEEE Sens. J. 2015, 15, 1538–1544. Xiong, H.; Chen, Z.; Yang, B.; Ni, R. TDoA localization algorithm with compensation of clock offset for wireless sensor networks. China Commun. 2015, 12, 193–201. [CrossRef] Maddumabandara, A.; Leung, H.; Liu, M. Experimental evaluation of indoor localization using wireless sensor networks. IEEE Sens.J. 2015, 15, 5228–5237. [CrossRef] Liu, Y.; Hu, Y.H.; Pan, Q. Distributed, robust acoustic source localization in a wireless sensor network. IEEE Trans. Signal Process. 2012, 60, 4350–4359. [CrossRef] Chizari, H.; Poston, T.; Abd Razak, S.; Abdullah, A.H.; Salleh, S. Local coverage measurement algorithm in GPS-free wireless sensor networks. Ad Hoc Netw. 2014, 23, 1–17. [CrossRef] Pratap, K.S.; Eric, H.-K.W.; Jagruti, S. DuRT: Dual rssi trend based localization for wireless sensor networks. IEEE Sens. J. 2013, 13, 3115–3123. Pivato, P.; Palopoli, L.; Petri, D. Accuracy of RSS-based centroid localization algorithms in an indoor environment. IEEE Trans. Instrum. Meas. 2011, 60, 3451–3460. [CrossRef] Sugeno, M. An introductory survey of fuzzy control. Inf. Sci. 1985, 36, 59–83. [CrossRef] Akyildiz, I.F.; Su, W.; Sankarasubramaniam, Y.; Cayirci, E. Wireless sensor networks: A survey. Comput. Netw. 2002, 38, 393–422. [CrossRef] Ma, D.; Er, M.J.; Wang, B. Analysis of hop-count-based source-to-destination distance estimation in wireless sensor networks with applications in localization. IEEE Trans. Veh. Technol. 2010, 59, 2998–3011. Cheng, L.; Wu, C.; Zhang, Y.; Wu, H.; Li, M.; Maple, C. A survey of localization in wireless sensor network. Int. J. Distrib. Sens. Netw. 2012, 2012, 12. [CrossRef] Zhao, L.-z.; Wen, X.-b.; Li, D. Amorphous localization algorithm based on BP artificial neural network. Int. J. Distrib. Sens. Netw. 2015, 2015, 6. [CrossRef] Lin, C.-M.; Mon, Y.-J.; Lee, C.-H.; Juang, J.-G.; Rudas, I.J. ANFIS-based indoor location awareness system for the position monitoring of patients. Acta Polytech. Hung. 2014, 11, 37–48. Velimirovic, A.S.; Djordjevic, G.L.; Velimirovic, M.M.; Jovanovic, M.D. Fuzzy ring-overlapping range-free (FRORF) localization method for wireless sensor networks. Comput. Commun. 2012, 35, 1590–1600. [CrossRef] Chagas, S.H.; Martins, J.B.; De Oliveira, L.L. An approach to localization scheme of wireless sensor networks based on artificial neural networks and genetic algorithms. In Proceedings of the IEEE 10th International New Circuits and systems Conference (NEWCAS), Montreal, QC, Canada, 17–20 June 2012; pp. 137–140. Liu, Z.-K.; Zhong, L. Node self-localization algorithm for wireless sensor networks based on modified particle swarm optimization. In Proceedings of the 27th IEEE Chinese Control and Decision Conference (CCDC), Qingdao, China, 23–25 May 2015; pp. 5968–5971. Kulkarni, R.V.; Venayagamoorthy, G.K. Bio-inspired algorithms for autonomous deployment and localization of sensor nodes. IEEE Trans. Syst. Man Cybern. Part C Appl. Rev. 2010, 40, 663–675. [CrossRef] Krishnaprabha, R.; Gopakumar, A. Performance of gravitational search algorithm in wireless sensor network localization. In Proceedings of the National Conference on Communication, Signal Processing and Networking (NCCSN), Palakkad, India, 10–12 October 2014; pp. 1–6. Baca, A.; Kornfeind, P.; Preuschl, E.; Bichler, S.; Tampier, M.; Novatchkov, H. A server-based mobile coaching system. Sensors 2010, 10, 10640–10662. [CrossRef] [PubMed] Walker, W.; Praveen Aroul, A.L.; Bhatia, D. Mobile health monitoring systems. In Proceedings of the 31st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Minneapolis, MN, USA, 2–6 September 2009; pp. 5199–5202.

Sensors 2016, 16, 1043

34. 35.

36. 37.

38.

39. 40.

41.

42.

43.

44.

45.

46. 47. 48.

49.

50. 51. 52. 53.

21 of 23

Eisenman, S.B.; Miluzzo, E.; Lane, N.D.; Peterson, R.A.; Ahn, G.-S.; Campbell, A.T. Bikenet: A mobile sensing system for cyclist experience mapping. ACM Trans. Sens. Netw. (TOSN) 2009, 6, 6. [CrossRef] Zhan, A.; Chang, M.; Chen, Y.; Terzis, A. Accurate caloric expenditure of bicyclists using cellphones. In Proceedings of the 10th ACM Conference on Embedded Network Sensor Systems, Toronto, Canada, 6–9 November 2012; pp. 71–84. Mi, Q.; Stankovic, J.A.; Stoleru, R. Practical and secure localization and key distribution for wireless sensor networks. Ad Hoc Netw. 2012, 10, 946–961. [CrossRef] Liebner, M.; Klanner, F.; Stiller, C. Active safety for vulnerable road users based on smartphone position data. In Proceedings of the IEEE Intelligent Vehicles Symposium (IV), Gold Coast, QLD, Australia, 23–26 June 2013; pp. 256–261. Liu, X.; Li, B.; Jiang, A.; Qi, S.; Xiang, C.; Xu, N. A bicycle-borne sensor for monitoring air pollution near roadways. In Proceedings of the IEEE International Conference on Consumer Electronics-Taiwan (ICCE-TW), Taipei, Tawan, 6–8 June 2015; pp. 166–167. Luoh, L. ZigBee-based intelligent indoor positioning system soft computing. Soft Comput. 2014, 18, 443–456. [CrossRef] Marita, T.; Negru, M.; Danescu, R.; Nedevschi, S. Stop-line detection and localization method for intersection scenarios. In Proceedings of the IEEE International Conference on Intelligent Computer Communication and Processing (ICCP), Cluj-Napoca, Romania, 25–27 August 2011; pp. 293–298. Nabil, M.; Ramadan, R.; ElShenawy, H.; ElHennawy, A. A comparative study of localization algorithms in WSNs. In Advanced Machine Learning Technologies and Applications; Hassanien, A., Salem, A.-B., Ramadan, R., Kim, T.-H., Eds.; Springer Berlin Heidelberg: Cario, Egypt, 2012; Volume 322, pp. 339–350. Pias, M.; Xu, K.; Hui, P.; Crowcroft, J.; Yang, G.-H.; Li, V.O.; Taherian, S. Sentient bikes for collecting mobility traces in opportunistic networks. In Proceedings of the 1st ACM International Workshop on Hot Topics of Planet-Scale Mobility Measurements, Kraków, Poland, 22–25 June 2009; p. 5. Shin, H.Y.; Un, F.L.; Huang, K.W. A sensor-based tracking system for cyclist group. In Proceedings of the Seventh International Conference on Complex, Intelligent, and Software Intensive Systems (CISIS), Taichung, Taiwan, 3–5 July 2013; pp. 617–622. Ndzi, D.L.; Harun, A.; Ramli, F.M.; Kamarudin, M.L.; Zakaria, A.; Shakaff, A.Y.M.; Jaafar, M.N.; Zhou, S.; Farook, R.S. Wireless sensor network coverage measurement and planning in mixed crop farming. Comput. Electron. Agric. 2014, 105, 83–94. [CrossRef] Aboelela, E.H.; Khan, A.H. Wireless sensors and neural networks for intruders detection and classification. In Proceedings of the International Conference on Information Networking (ICOIN), Bali, Indonesia, 1–3 February 2012; pp. 138–143. Kumar, S.; Jeon, S.M.; Lee, S.R. Localization estimation using artificial intelligence technique in wireless sensor networks. J. Korea Inf. Commun. Soc. 2014, 39, 820–827. [CrossRef] Jang, H.; Topal, E. A review of soft computing technology applications in several mining problems. Appl. Soft Comput. 2014, 22, 638–651. [CrossRef] Volosencu, C.; Curiac, D.-I. Efficiency improvement in multi-sensor wireless network based estimation algorithms for distributed parameter systems with application at the heat transfer. EURASIP J. Adv. Signal Process. 2013, 2013, 1–17. [CrossRef] Nekooei, S.M.; Manzuri-Shalmani, M.T. Location finding in wireless sensor network based on soft computing methods. In International Conference on Control, Automation and Systems Engineering (CASE), Singapore, 30–31 July 2011; 2011; pp. 1–5. Larios, D.; Barbancho, J.; Molina, F.J.; León, C. Lis: Localization based on an intelligent distributed fuzzy system applied to a WSN. Ad Hoc Net. 2012, 10, 604–622. [CrossRef] Oussalah, M.; Alakhras, M.; Hussein, M. Multivariable fuzzy inference system for fingerprinting indoor localization. Fuzzy Sets Syst. 2015, 269, 65–89. [CrossRef] Mona, Y.-J. WSN-based indoor location identification system (LIS) applied to vision robot designed by fuzzy neural network. J. Digit. Inf. Manag. 2015, 13, 37. Palipana, S.; Kapukotuwe, C.; Malasinghe, U.; Wijenayaka, P.; Munasinghe, S. Localization of a mobile robot using ZigBee based optimization techniques. In Proceedings of the IEEE 6th International Conference on Information and Automation for Sustainability (ICIAfS), Beijing, China, 27–29 September 2012; pp. 215–220.

Sensors 2016, 16, 1043

54. 55.

56. 57. 58. 59. 60.

61.

62. 63.

64.

65. 66. 67. 68. 69. 70. 71.

72.

73.

74. 75.

22 of 23

Gholami, M.; Cai, N.; Brennan, R.W. An artificial neural network approach to the problem of wireless sensors network localization. Robot. Comput. Integr. Manuf. 2013, 29, 96–109. [CrossRef] Payal, A.; Rai, C.; Reddy, B. Artificial neural networks for developing localization framework in wireless sensor networks. In Proceedings of the International Conference on Data Mining and Intelligent Computing (ICDMIC), New Delhi, India, 5–6 September 2014; pp. 1–6. Kukolj, D.; Levi, E. Identification of complex systems based on neural and takagi-sugeno fuzzy model. IEEE Trans. Syst. Man Cybern. Part B Cybern. 2004, 34, 272–282. [CrossRef] Rashedi, E.; Nezamabadi-Pour, H.; Saryazdi, S. GSA: A gravitational search algorithm. Inf. Sci. 2009, 179, 2232–2248. [CrossRef] Mansouri, R.; Nasseri, F.; Khorrami, M. Effective time variation of G in a model universe with variable space dimension. Phys. Lett. A 1999, 259, 194–200. [CrossRef] Shuaib, Y.M.; Kalavathi, M.S.; Rajan, C.C.A. Optimal capacitor placement in radial distribution system using gravitational search algorithm. Int. J. Electr. Power Energy Syst. 2015, 64, 384–397. [CrossRef] Chatterjee, A.; Roy, K.; Chatterjee, D. A gravitational search algorithm (GSA) based photo-voltaic (PV) excitation control strategy for single phase operation of three phase wind-turbine coupled induction generator. Energy 2014, 74, 707–718. [CrossRef] Kumar, S.; Lee, S.R. Localization with RSSI values for wireless sensor networks: An artificial neural network approach. In Proceedings of the 1st International Electronic Conference on Sensors and Applications 1, Seoul, Korea, 1–16 June 2014; pp. 1–6. Chuang, P.-J.; Jiang, Y.-J. Effective neural network-based node localisation scheme for wireless sensor networks. IET Wirel. Sens. Syst. 2014, 4, 97–103. [CrossRef] Nanda, M.; Kumar, A.; Kumar, S. Localization of 3D WSN using mamdani sugano fuzzy weighted centriod approaches. In Proceedings of the IEEE Students’ Conference on Electrical, Electronics and Computer Science (SCEECS), Bhopal, India, 1–2 March 2012; pp. 1–5. Feng, X.; Gao, Z.; Yang, M.; Xiong, S. Fuzzy distance measuring based on RSSI in wireless sensor network. In Proceedings of the 3rd International Conference on Intelligent System and Knowledge Engineering, Xiamen, China, 17–19 November 2008; pp. 395–400. Yun, S.; Lee, J.; Chung, W.; Kim, E.; Kim, S. A soft computing approach to localization in wireless sensor networks. Expert Syst. Appl. 2009, 36, 7552–7561. [CrossRef] Yu, C.; Zhang, Y.; Zhang, J.; Liu, Y. Research of self-calibration location algorithm for ZigBee based on pso-rssi. In Electronics and Signal Processing; Springer: Berlin, Germany, 2011; pp. 91–99. Gharghan, S.K.; Nordin, R.; Ismail, M.; Ali, J.A. Accurate wireless sensor localization technique based on hybrid PSO-ANN algorithm for indoor and outdoor track cycling. IEEE Sens. J. 2016, 16, 529–541. [CrossRef] Tewolde, G.S.; Kwon, J. Efficient WiFi-based indoor localization using particle swarm optimization. In Advances in Swarm Intelligence; Springer: Berlin, Germany, 2011; pp. 203–211. Brunato, M.; Battiti, R. Statistical learning theory for location fingerprinting in wireless LANS. Comput. Netw. 2005, 47, 825–845. [CrossRef] Mestre, P.; Coutinho, L.; Reigoto, L.; Matias, J.; Correia, A.; Couto, P.; Serodio, C. Indoor location using fingerprinting and fuzzy logic. In Eurofuse 2011; Springer: Berlin, Germany, 2011; pp. 363–374. Nerguizian, C.; Nerguizian, V. Indoor fingerprinting geolocation using wavelet-based features extracted from the channel impulse response in conjunction with an artificial neural network. In Proceedings of the IEEE International Symposium on Industrial Electronics (ISIE), Vigo, Spain, 4–7 June 2007; pp. 2028–2032. Azenha, A.; Peneda, L.; Carvalho, A. A neural network approach for radio frequency based indoors localization. In Proceedings of the IECON 38th Annual Conference on Industrial Electronics Society, Montreal, QC, Canada, 25–28 October 2012; pp. 5990–5995. Wu, B.-F.; Jen, C.-L.; Chang, K.-C. Neural fuzzy based indoor localization by kalman filtering with propagation channel modeling. In Proceedings of the IEEE International Conference on Systems, Man and Cybernetics (ISIC), Montreal, QC, Canada, 7–10 October 2007; pp. 812–817. Irfan, N.; Bolic, M.; Yagoub, M.C.; Narasimhan, V. Neural-based approach for localization of sensors in indoor environment. Telecommun. Syst. 2010, 44, 149–158. [CrossRef] Rahman, M.S.; Park, Y.; Kim, K.-D. RSS-based indoor localization algorithm for wireless sensor network using generalized regression neural network. Arab. J. Sci. Eng. 2012, 37, 1043–1053. [CrossRef]

Sensors 2016, 16, 1043

76.

77. 78.

79.

23 of 23

Moreno-Cano, M.; Zamora-Izquierdo, M.A.; Santa, J.; Skarmeta, A.F. An indoor localization system based on artificial neural networks and particle filters applied to intelligent buildings. Neurocomputing 2013, 122, 116–125. [CrossRef] Brás, L.; Pinho, P.; Borges Carvaloh, N. Evaluation of a sectorised antenna in an indoor localisation system. IET Microw. Antennas Propag. 2013, 7, 679–685. [CrossRef] Thongpul, K.; Jindapetch, N.; Teerapakajorndet, W. A neural network based optimization for wireless sensor node position estimation in industrial environments. In Proceedings of the International Conference on Electrical Engineering/Electronics Computer Telecommunications and Information Technology (ECTI-CON), Chaing Mai, Thailand, 19–21 May 2010; pp. 249–253. Gogolak, L.; Pletl, S.; Kukolj, D. Indoor fingerprint localization in WSN environment based on neural network. In Proceedings of the IEEE 9th International Symposium on Intelligent Systems and Informatics (SISY), Subotica, Serbia, 8–10 September 2011; pp. 293–296. © 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/).

Actuator Network Design for Mobile Control Applications.

Energy-efficient ZigBee-based wireless sensor network for track bicycle performance monitoring.

A Wearable Wireless Sensor Network for Indoor Smart Environment Monitoring in Safety Applications.

Design of a Humidity Sensor Tag for Passive Wireless Applications.

A CMOS pressure sensor tag chip for passive wireless applications.

Design of a wireless sensor network platform for tele-homecare.

An Effective Cuckoo Search Algorithm for Node Localization in Wireless Sensor Network.

An efficient and adaptive mutual authentication framework for heterogeneous wireless sensor network-based applications.

NLOS Scenario.

Reputation-based secure sensor localization in wireless sensor networks.

Parametric Loop Division for 3D Localization in Wireless Sensor Networks.

SOUNET: Self-Organized Underwater Wireless Sensor Network.

A Wireless Pressure Sensor Integrated with a Biodegradable Polymer Stent for Biomedical Applications.

A Low Cost Bluetooth Low Energy Transceiver for Wireless Sensor Network Applications with a Front-end Receiver-Matching Network-Reusing Power Amplifier Load Inductor.

Performance Evaluation of Localization Accuracy for a Log-Normal Shadow Fading Wireless Sensor Network under Physical Barrier Attacks.

Disruption Tolerant Network-Based Message Forwarding for a River Pollution Monitoring Wireless Sensor Network Application.

Wireless Sensor Network-Based Service Provisioning by a Brokering Platform.

WiSPH: a wireless sensor network-based home care monitoring system.

A Wireless Sensor Network-Based Approach with Decision Support for Monitoring Lake Water Quality.

Analyzing comprehensive QoS with security constraints for services composition applications in wireless sensor networks.

Implementation of a wireless sensor network for heart rate monitoring in a senior center.

A Comparative Study of Anomaly Detection Techniques for Smart City Wireless Sensor Networks.

An Embedded Multi-Agent Systems Based Industrial Wireless Sensor Network.

A data acquisition protocol for a reactive wireless sensor network monitoring application.