ANN-Based Control of a Multiboat Group for the Deployment of an Underwater Sensor Network

Abstract

Underwater sensor networks (USNs) can be used for several types of commercial and noncommercial applications. However, some constraints resulting from the nature of aquatic environments severely limit their use. Due to constraints such as large propagation latency, low-bandwidth capacity, and short-distance communications, a large number of USN nodes are deployed to provide reliability in most applications. In this study, an unattended deployment approach based on the use of an autonomous boat group is proposed. A map of the deployment zone and optimal locations of USN nodes are fed into the onboard computers of the boat group. After processing these data and determining paths to be followed, the boat group deploys sensor nodes at predetermined locations. During the deployment, the boat group is controlled by an artificial neural network- (ANN-) based control system for reducing path errors. A set of performance evaluations is given to prove efficiency of the proposed control system. Performance results show that the boat group can successfully follow a predefined path set and deploy USN nodes. The tradeoffs between energy consumptions, end-to-end delay, and number of hops between underwater relay nodes of energy-efficient USN are also examined. The results indicate that increasing the number of hops reduces the total energy consumption and the end-to-end delay.

1. Introduction

Most of the Earth's surface is covered by water. Sensing and monitoring of aquatic environments are used in several potential applications including oceanographic data collection for fish and mussels growth observation, pollution detection, oil monitoring, seismic and volcanic prediction, coastal surveillance, and different objectives such as exploration, protection, and commercial exploitation [1]. A low-cost approach for all these applications is the use of underwater sensor networks (USNs). In USNs, nodes can be equipped with sensors for the measurement of different parameters such as depth, conductivity, oxygen, pH, turbidity, temperature, and salinity. In USNs, the transmission of information can be realized in the form of acoustic, electromagnetic, or optical waves. All the wireless communication techniques, acoustic, optical, and electromagnetic communications, have pros and cons [2].

Due to the nature of aquatic environments, a number of key constraints like the low-bandwidth capacity, large propagation latency, high bit-error rate, and short-distance radio communications must be taken into account [3]. These key constraints call for the deployment of a large number of sensor nodes for a successful USN. Its key advantage, the low attenuation of acoustic waves in water, makes acoustic communication the most widely used technique in USNs. Although the deployment of USNs can be realized using a number of different approaches, it remains a significant challenge due to the inherent difficulties posed by the underwater communication channel. In addition, budget-related constraints, arising from the high cost of each node, may severely limit the number of nodes that can be used. Therefore, the employment of efficient deployment approaches is necessary for the achievement of successful USNs in real applications.

In this study, an autonomous multi-boat-group-assisted deployment approach is proposed. It is based on the use of artificial neural networks (ANNs) for the control of the multiboat group. Before a deployment mission, a map of the deployment zone and optimal locations of USN nodes, provided by the USN deployment module, are fed into the onboard computers of the boats. The boats process the data and determine the paths that they are going to follow in order to deploy the USN nodes at the predetermined optimal locations. As proven by the simulation studies presented in this study, the use of ANNs reduces the path errors considerably compared to the traditional control approaches. We also investigate the tradeoffs between energy consumption, end-to-end delay, and number of hops for the case of USN. Although similar problems were investigated in previous works [4–6], the obtained results are not sufficient for the particular scenario examined in this paper. Numerical results in this study demonstrate the theoretical derivations.

The remainder of the paper is organized as follows. Section 2 presents a detailed literature survey on underwater sensor network deployment approaches. Section 3 presents a novel strategy for the deployment of USNs and gives the results obtained from a set of simulation studies, which were carried out for the evaluation of this approach. Section 4 provides a detailed description for the implementation of an artificial neural network designed to control a multiboat group that is responsible for the deployment of underwater sensor networks, while it also presents an overview of the performance evaluation results. Finally, Section 5 concludes the paper.

2. Related Work

In recent years, USNs (e.g., cf. Figure 1) have considerably drawn attention of the research communities. Although they offer numerous advantages in several different application scenarios, they come with many issues that need to be addressed. Therefore, factors associated with the success or failure regarding the implementation of USNs should be identified and subsequently handled in an appropriate way. As it has been shown by both simulation studies and field tests, one of the most important phases of all the sensor network implementations is the deployment phase which can be realized in an attended or unattended manner [7, 8]. Since USN implementations are much harder than other sensor network implementations and they are associated with several performance related factors, the attended deployment approaches offer considerable advantages over the unattended ones, including cost-effectiveness.

Figure 1

A typical underwater sensor network.

The most important objectives of an efficient USN deployment are (i) to provide full-connectivity coverage and (ii) to optimize the USN lifetime. However, the performance of USNs is greatly limited by the low-bandwidth and high propagation delay of underwater acoustic communications [9, 10]. In this regard, several network-related factors such as the packet reception rate and the delay and throughput, as well as node related factors, such as the transceiver power, the battery, and the CPU power, should be taken into consideration [11]. In [9], Felamban et al. define an optimal node placement strategy for USNs in order to obtain maximum coverage and connectivity with minimum transmission loss. By taking the characteristics of underwater acoustic channels, the authors formulated this problem as a nonlinear programming model. The authors proved that the number of USN nodes required to cover a definite volume depends on the operating frequency. When it is increased, more nodes are required with short internode distances in order to maintain a transmission loss threshold. Due to the abovementioned limitations of USNs, most studies have been focused on specific techniques to enhance USN performance metrics. Ibrahim et al. propose the use of multiple surface-level gateways to improve end-to-end delay and reduce energy consumption by distributing traffic loads [10, 12]. The authors show the significant advantages of their approach.

USN deployments can be realized using different manned or unmanned vehicles including boats and aerial vehicles. Before deployment, the deployment locations of sensor nodes and a map of the deployment site are loaded to the onboard computer of the vehicle or are provided to the human operator. This can be realized online during deployment, too. To realize USN deployments, unmanned autonomous vehicles need an accurate navigation system. In [13], Erckens et al. present a special navigation system that plans an optimal navigation course and efficiently controls an unmanned boat vessel called “Avalon boat” on this course. The navigation system is based on a novel path planner that generates the fastest path to a given destination and is able to avoid both static and dynamic obstacles [13]. While a boat moves on the sea or ocean, it is exposed to several uncontrollable factors that may affect its direction. For unmanned autonomous boats, it is difficult to obtain an accurate model and it requires a lot of sensors to feed a proper controller [14, 15]. A simple boat model which relates the most important variables and concentrates on data that can be provided by low-cost sensors is proposed in [14]. As proven in many field tests, autonomous boat navigation systems cannot easily replace human operators due to uncontrollable factors such as wind and waves [16, 17]. On the other hand, knowledge of sailing experts can be transformed into a fuzzy interference system in order to steer sails and rudder according to weather conditions and target [16].

However, all the aforementioned deployment algorithms are either driven by the deployment cost or performance related factors like the network lifetime. Furthermore, they do not focus on the navigation accuracy of the vehicles that deploy the sensor nodes. Our approach differs from the related studies since it combines a performance-driven deployment scheme with an ANN-based path-following system used for the realization of an USN deployment.

3. Efficient Node Placement for Balanced Energy Consumption and End-to-End Delay of a Multiple-Hop USN

In the following, we present a simulation study to reveal the communication related tradeoffs arising for an acoustic wave USN deployed in a model dam. The USN consists of two entities: the relay nodes and the surface gateway. A sufficient number of nodes are deployed to cover a certain underwater volume, in this case the volume of an orthogonal shallow dam model. The surface of this dam is 3.391 km², while the depth is 26.3 m. Each node can be deployed in any position, while the location of the gateway node is known. Since the sensor nodes communicate with each other using acoustic links, the path loss is caused by two phenomena: the energy (geometric) spreading and the wave absorption. The energy spreading mainly depends on the transmission range of the acoustic waves [19], while the wave absorption is frequency-dependent; that is, high frequency signals are more vulnerable to absorption loss due to the conversion of acoustic energy to heat. The nodes in the USN are powered by batteries and hence they should be properly placed to reduce path loss and energy consumption. Then, these nodes will maintain their location using various means of location and depth adjustments.

Two-dimensional (2D) shallow-water acoustic communications were considered, since the depth is set to 26.3 m. In this case, the acoustic signals propagate within a cylinder bounded by the surface and the sea floor; that is, cylindrical spreading appears and the signal spreads horizontally due to low depth [20]. We assume that all nodes are equipped with homogenous transceivers and have a cylindrical based communication with a radius of d, where d is assumed to be constant for all nodes. Two nodes are connected if the node interdistance is less than or equal to $d_{\max}$ . The number of nodes is inversely proportional to the volume covered by the nodes. We further assume that all nodes in the network transmit with a uniform transmission power. Let the node N be located at a distance of D from gateway. In general, this node sends its data packets to the gateway over a multiple-hop path, denoted by $h_{1} h_{2} h_{3} \dots h_{n}$ [6].

Underwater sound propagation is described using the Sonar equations [19]. The source signal level (which is related to the transmission power intensity and hence to the transmission power $P_{t}$ of the transceiver) of an emitted underwater acoustic signal can be expressed by the following passive sonar equation [20]:

\begin{matrix} SL = TL + NL + SNR - DI, \end{matrix}

(1)

where

TL

is the transmission loss,

NL

is the ambient noise level,

SNR

is the signal-to-noise ratio at the receiver, and

DI

is the directivity index. In general,

NL

decreases with increasing frequency and also decreases at great depths, since most noise sources are at the surface. Thus, it is greater in shallow water than in deep water, since noise is trapped between water floor and surface. Factors contributing to the ambient noise level in shallow water networks include waves, shipping traffic, wind level, biological noise, seaquakes, and volcanic activity. The impact of each of these factors on the ambient noise level depends on the particular setting. In this paper, we consider an average value for the ambient noise level to be 50 dB as a representative shallow water case, based on several studies of shallow water noise measurements [19, 21] and our particular scenario. By assuming omnidirectional hydrophones, the directivity index is set to zero. Through the above assumptions, we can express the source level

SL

intensity as a function of

TL

and

SNR

only:

\begin{matrix} SL = TL + SNR + 50 . \end{matrix}

(2)

Since cylindrical spreading applies for shallow waters, the transmission loss in dB is approximated as [19]

\begin{matrix} TL = 10 \log d + a d \times 1 0^{- 3} + A, \end{matrix}

(3)

where d accounts for the distance between source and receiver in meters and a is the frequency-dependent medium absorption coefficient. Note that A denotes the transmission anomaly, which results from degradation of the acoustic intensity caused by multiple path propagation, refraction, diffraction, and scattering of sound. This value is usually between 5 and 10 dB and is higher for shallow-water horizontal links, where multipath effects are higher. Furthermore,

SL

is related to the transmitted signal intensity at 1 m from the source according to the following expression:

\begin{matrix} SL = 10 \log \frac{I_{t}}{1 μ Pa}, \end{matrix}

(4)

where the intensity

I_{t}

of an underwater signal is in

μ Pa

. Solving for

I_{t}

yields

\begin{matrix} I_{t} = 1 0^{SL / 10} \times 0.67 \times 1 0^{- 18}, \end{matrix}

(5)

in Watts/m², where the constant converts

μ Pa

into Watts/m². The transmission power

P_{t}

needed to achieve an intensity

I_{t}

at a distance of 1 m from the source in the direction of the receiver is expressed as [19]

\begin{matrix} P_{t} = 2 π \times 1 m \times H \times I_{t}, \end{matrix}

(6)

in Watts, where H is the water depth in m.

We assume 16-quadrature amplitude modulation (QAM) and orthogonal frequency division multiplexing (OFDM) encoding method. Then, the bit error rate (BER) is derived as [22]

\begin{matrix} P_{b}^{16 QAM} = \frac{3}{2 k} erfc (\sqrt{\frac{k}{10} \frac{E_{b}}{N_{0}}}), \end{matrix}

(7)

where k is

lo g_{2} 16

and

E_{b} / N_{0} = SNR (B_{N} / R)

is the energy per-bit-to-noise power spectral density ratio,

B_{N}

is the noise bandwidth in Hz, and R is the data rate in bps.

The propagation time in seconds for a distance d in meters is

\begin{matrix} t_{prop} = \frac{d}{c_{acw}}, \end{matrix}

(8)

where

c_{acw}

is the propagation velocity of the acoustic waves in water and depends on environmental parameters, such as water temperature, salinity, and depth. Contention-based medium access control (MAC) protocols, for example, aloha and carrier sense multiple access (CSMA), unsatisfactorily perform in USNs [23]. Hence, the time division multiple access (TDMA) MAC protocol is considered [5], in order to evaluate the end-to-end delay in the string USN. This protocol allocates to each node an exclusive time-slot for communication and guarantees collision-free media access. This behavior allows for the reduction of the transceiver state switches and preamble transmissions to save more energy. The end-to-end delay of multiple-hop string network is given by [5]

\begin{matrix} Delay = n (t_{prop (i, i + 1)} + t_{pkt}), \end{matrix}

(9)

where

t_{prop (i, i + 1)}

is the propagation time between nodes i and

i + 1

, n is the number of hops, and

\begin{matrix} t_{pkt} = \frac{Packet length}{Data rate} \end{matrix}

(10)

is the time required to transmit data packet.

3.1. USN Simulation Results

In this subsection, we investigate the influence of placing middle nodes to convey information from a source to a destination on the energy consumption and the end-to-end delay. Based on the measurements of medium absorption in shallow water conducted at temperatures between 4°C and 20°C [24] and considering that $f =$ 40 KHz, the value of the average medium absorption a is set to 8.96 dB/km. The mean depth of the orthogonal dam is $H =$ 26.3 m and the transmission anomaly A was set to 5 dB. Moreover, the propagation velocity $c_{acw}$ of acoustic waves in water can be empirically estimated using the formula in [25]. In particular, the mean water temperature is set to 15°C and the propagation velocity is equal to 1466 m/s. In addition, the packet length, the data rate for each node, and the noise bandwidth are set to 512 bit, 1 Kbps (this bit rate is well within the bit rates of current hydrophones [26]), and 1 kHz, respectively.

Figure 2 demonstrates the total energy consumption (normalized transmission power $P_{t}$ ) of the USN, when sending a single packet. In order to obtain the required transmission power $P_{t}$ for signal transmissions at a given distance d and frequency f, one can initially compute the transmission loss $TL$ in (3). Then, using (2) and (5), the source level $SL$ and the intensity $I_{t}$ of an underwater signal can be computed, respectively. Using (6), the corresponding transmission power $P_{t}$ needed to achieve intensity $I_{t}$ can be computed. The normalized transmission power is obtained by dividing all the values of $P_{t}$ by the maximum value of $P_{t}$ . Since the surface of the orthogonal dam is 3.391 km², it is assumed that the sides of the orthogonal surface are $a =$ 1 km and $b =$ 3.391 km. Hence, reasonable distances between nodes of 50 m, 100 m, 200 m, and 400 m are considered at depth of 26.3 m, while the BER is set to 10⁻⁹. The number of nodes within the dam required for full coverage is related to the distance between the nodes. For instance, considering that this distance is constant and equal to 400 m, 2 nodes can be placed alongside a, while 8 nodes can be placed alongside b. Thus, overall 16 nodes are required. The results indicate that the total energy consumption is reduced, as the number of hops increases. In particular, the total energy consumption reduces by approximately 10%, 15%, 30%, and 50% for links with distances of 50 m, 100 m, 200 m, and 400 m, respectively. These results also show that the total energy consumption insignificantly changes, when the number of hops is greater than 10.

Figure 2

The normalized energy consumption of the USN as a function of the number of the hops for different distances between nodes.

Using (8)–(10), Figure 3 depicts the normalized end-to-end delay in string USN as a function of the number of hops for distances between nodes of 50 m, 100 m, 200 m, and 400 m at depth of 26.3 m. The normalization is obtained by dividing all the values of the delay by its maximum value. One observes that the end-to-end delay almost linearly increases with the augmentation of hops. However, the distance between nodes slightly affects the end-to-end delay. Thus, the number of hops should be carefully determined to compromise energy consumption and end-to-end delay.

Figure 3

The normalized end-to-end delay of the USN as a function of the number of the hops for different distances between nodes.

4. Artificial Neural Network-Based Control of a Multiboat Group

An ANN is essentially a parallel distributed processor made up of simple processing units that has an inherent trend for storing empirical knowledge and making it available for use [27]. Multilayer feed-forward neural networks (NNs) are widely used, while their training is based on the minimization of an error function. Specifically, backpropagation (BP) is a well-known training method that is used in the case of multilayer feed-forward NNs. BP learning uses the gradient descent procedure to determine the connection weights [28].

The multilayer feed-forward NNs consist of one input layer, one or more hidden layers, and one output layer. Each layer is composed of neurons which are connected to the neurons of the neighboring layers. Each connection link is characterized by a weight. The evolution of the online supervised training of multilayer perceptrons has been further raised by the development of the BP algorithm. The following equations summarize the main steps of the BP algorithm. Consider

\begin{matrix} o_{j} = f ({net}_{j}) = f (x) then {net}_{j} = \sum_{j}^{i} w_{j i} o_{i} + θ_{j} \end{matrix}

(11)

\begin{matrix} E_{p} = \frac{1}{2} \sum_{j \in out .} {(t_{p j} - o_{p j})}^{2} \end{matrix}

(12)

\begin{matrix} δ_{p j} = (t_{p j} - o_{p j}) \\ Δ_{p} w_{j i} = - ε (\frac{\partial E_{p}}{\partial w_{j i}}) \\ Δ_{p} θ_{j} = - ε (\frac{\partial E_{p}}{\partial θ_{j}}), \end{matrix}

(13)

where j is the layer number and i is the neuron number;

o_{j}

is the output,

ne t_{j}

is the weighted sum, and

θ_{j}

is the bias, respectively, while

w_{j i}

is the weight characterizing the connection; ε is the learning rate,

δ_{p j}

represents the error value in layer j,

t_{p j}

is the target output, and

o_{p j}

is the actual output. Equation (12) is used to estimate the total error in the output layer for the pth sample pattern, the error that is eventually minimized by changing the weights and biases using gradient descent (13).

The optimization method known as Levenberg-Marquardt algorithm [29] was used for the development of the proposed NN. This technique, due to Levenberg [30] and Marquardt [31], is a combination of the following two methods [32, 33]: (i)

the Gauss-Newton's method approaches rapidly near to a global or local minimum but sometimes may deviate;

(ii)

the Gradient descent method certainly approaches through a proper selection of step size parameter but does slowly.

Figures 4 and 5 illustrate the block diagrams and the architecture of the developed ANN.

Figure 4

The block diagram of the proposed ANN. The proposed NN model is used to drive an autonomous multiboat group for following a given path.

Figure 5

The structure of the proposed multilayer NN model was prepared using Neural Network Toolbox of MATLAB [18].

4.1. Performance Evaluation of the Proposed ANN Model

As already mentioned, this study proposes the use of an ANN model for the determination of the path trajectories of multiboat groups used for the deployment of USNs. Lake Buyukcekmece, Istanbul, Turkey, was selected as the application area. It is the major water resource of Istanbul, Turkey, and is located in the Architect Sinan District of Istanbul, on the coast of the Marmara Sea. The lake is 7 kilometers long, 2 kilometers wide, with an average depth of 8.6 meters. It covers an area of 28.47 km². The Buyukcekmece Dam was built on the sea side of the lake. Besides that, the lake is available for fishing, energy generation, tourism, and irrigation.

In this simulation study, three autonomous boats belonging to a multiboat group in the Lake Buyukcekmece were initially considered at three different static points, while a static target point was also considered. The main objective of this study was that these boats should track a predefined path through which each boat arrives at the target point and then the second and the third boats track the same path to the opposite direction in order to go back to their original positions, while the first boat tracks a different path to go to a new static target point.

The simulated results were obtained using MATLAB Neural Network Toolbox [18] and they were compared to their corresponding actual datasets. Eloquent figures showing both actual and estimated trajectories, marking the coordinates during the progressive movement of the boats, are given for the evaluation of the effectiveness of the proposed method. The data used to develop the NN model were based on a virtual x-y plane coordinate system, where “X” and “Y” were the input variables, while “ $X_{t}$ ” (standing for X of target) and “ $Y_{t}$ ” (standing for Y of target) were the output variables. The paths of the boats on the virtual x-y plane are shown in Figure 6. The length unit of the virtual plain grid corresponds to 120 meters in either direction.

Figure 6

Boats paths on the employed virtual x-y plane.

Boat 1 was initially located at the virtual x-y plane position (10, 76), Boat 2 at (24, 50), and Boat 3 at (44, 21). There was a target point which was located at the static position (27, 34). All three boats followed a specific, predefined path to reach the target point. Then, the second and the third ones went back to their original positions, (10, 76) and (44, 21), respectively, while the first one tracked a different path to go to a new target point located at (34, 34). The full dataset used in this study is given in Table 1 and comprises 164 sets of input-output variables values.

Table 1

Dataset used for the ANN model. The column label “Step” refers to the sequential progressive movement of each boat, while “ $B_{j}$ ,” $j = 1,2, 3$ , corresponds to the movement from the original location of each boat (indicated by j) to the first target, at (27, 34), while “ $B_{j}$ _back,” $j = 1,2, 3$ , corresponds to the next mission, that is, return to original locations for Boat 2 and Boat 3, while movement to a new target, at (34, 34), for Boat 1.

Step	Boat	X	Y	$X_{t}$	$Y_{t}$

1	$B_{1}$	10	76	10	75
2		10	75	10	74
3		10	74	10	73
4		10	73	10	72
5		10	72	10	71
6		10	71	10	70
7		10	70	10	69
8		10	69	10	68
9		10	68	10	67
10		10	67	10	66
11		10	66	10	65
12		10	65	10	64
13		10	64	10	63
14		10	63	10	62
15		10	62	10	61
16		10	61	10	60
17		10	60	10	59
18		10	59	10	58
19		10	58	10	57
20		10	57	10	56
21		10	56	10	55
22		10	55	10	54
23		10	54	10	53
24		10	53	10	52
25		10	52	10	51
26		10	51	10	50
27		10	50	10	49
28		10	49	10	48
29		10	48	10	47
30		10	47	10	46
31		10	46	11	46
32		11	46	12	46
33		12	46	13	46
34		13	46	14	46
35		14	46	15	46
36		15	46	16	46
37		16	46	17	46
38		17	46	18	46
39		18	46	19	46
40		19	46	19	45
41		19	45	19	44
42		19	44	19	43
43		19	43	19	42
44		19	42	19	41
45		19	41	19	40
46		19	40	19	39
47		19	39	19	38
48		19	38	19	37
49		19	37	19	36
50		19	36	19	35
51		19	35	19	34
52		19	34	20	34
53		20	34	21	34
54		21	34	22	34
55		22	34	23	34
56		23	34	24	34
57		24	34	25	34
58		25	34	26	34
59		26	34	27	34

60	$B_{2}$	24	50	24	49
61		24	49	24	48
62		24	48	24	47
63		24	47	24	46
64		24	46	24	45
65		24	45	25	45
66		25	45	26	45
67		26	45	27	45
68		27	45	27	44
69		27	44	27	43
70		27	43	27	42
71		27	42	27	41
72		27	41	27	40
73		27	40	27	39
74		27	39	27	38
75		27	38	27	37
76		27	37	27	36
77		27	36	27	35
78		27	35	27	34

79	$B_{3}$	44	21	43	21
80		43	21	42	21
81		42	21	41	21
82		41	21	40	21
83		40	21	39	21
84		39	21	38	21
85		38	21	37	21
86		37	21	36	21
87		36	21	35	21
88		35	21	34	21
89		34	21	33	21
90		33	21	32	21
91		32	21	31	21
92		31	21	30	21
93		30	21	29	21
94		29	21	28	21
95		28	21	27	21
96		27	21	27	22
97		27	22	27	23
98		27	23	27	24
99		27	24	27	25
100		27	25	27	26
101		27	26	27	27
102		27	27	27	28
103		27	28	27	29
104		27	29	27	30
105		27	30	27	31
106		27	31	27	32
107		27	32	27	33
108		27	33	27	34

109	$B_{1}$ _back	27	34	28	34
110		28	34	29	34
111		29	34	30	34
112		30	34	31	34
113		31	34	32	34
114		32	34	33	34
115		33	34	34	34

116	$B_{2}$ _back	27	34	27	35
117		27	35	27	36
118		27	36	27	37
119		27	37	27	38
120		27	38	27	39
121		27	39	27	40
122		27	40	27	41
123		27	41	27	42
124		27	42	27	43
125		27	43	27	44
126		27	44	27	45
127		27	45	26	45
128		26	45	25	45
129		25	45	24	45
130		24	45	24	46
131		24	46	24	47
132		24	47	24	48
133		24	48	24	49
134		24	49	24	50

135	$B_{3}$ _back	27	34	27	33
136		27	33	27	32
137		27	32	27	31
138		27	31	27	30
139		27	30	27	29
140		27	29	27	28
141		27	28	27	27
142		27	27	27	26
143		27	26	27	25
144		27	25	27	24
145		27	24	27	23
146		27	23	27	22
147		27	22	27	21
148		27	21	28	21
149		28	21	29	21
150		29	21	30	21
151		30	21	31	21
152		31	21	32	21
153		32	21	33	21
154		33	21	34	21
155		34	21	35	21
156		35	21	36	21
157		36	21	37	21
158		37	21	38	21
159		38	21	39	21
160		39	21	40	21
161		40	21	41	21
162		41	21	42	21
163		42	21	43	21
164		43	21	44	21

The developed ANN model is a 1-input layer, 2 hidden layers (comprising 24 and 22 nodes, resp.), and 1-output layer. As already mentioned, the Levenberg-Marquardt optimization algorithm was used since it presents certain advantages against the simple BP [34]. Table 2 shows the test dataset, while Table 3 shows the corresponding ANN results for the test data. The comparison of the training dataset with the corresponding ANN results for Boat 1, Boat 2, and Boat 3 is given in Figures 7, 8, and 9, respectively. Figure 10 shows a combined view of the training datasets and ANN obtained positions for all three boats.

Table 2

Dataset for test.

$X_{t}$	$Y_{t}$

10	47
11	46
18	46
19	45
19	35
20	34
24	47
27	45
36	21
28	21
27	22
27	29
30	34
27	43
24	45
27	29
27	22
28	21
36	21

Table 3

ANNs results.

$X_{t}$	$Y_{t}$

10.265003	46.587496
11.908684	46.203325
18.681534	45.940821
19.034220	44.207650
19.310665	34.966706
20.778849	34.432972
23.709556	47.529866
25.610589	45.114751
36.037393	21.191301
28.188850	21.252901
27.281613	22.342490
27.129286	29.131046
30.808445	33.931133
26.750776	42.699318
24.745857	45.002582
27.129286	29.131046
27.281613	22.342490
28.188850	21.252901
36.037393	21.191301

Figure 7

The comparison of the training set and ANN results for Boat 1. Since the ANN has driven the boat very successfully, the ANN results and the dataset coincide. Therefore, it is difficult to locate some symbols in this figure.

Figure 8

The comparison of the training set and ANN results for Boat 2. Basically, overlapping or close symbols indicate that the ANN has successfully driven the boat to the target points in the path.

Figure 9

The comparison of the training set and ANN results for Boat 3. As it can be seen, the proposed ANN has driven the boat successfully except for a few points.

Figure 10

The comparison of the training set and ANN results for all boats. Overlapping points indicate that the ANN has successfully driven the boats to the target points in the path.

Before entering the ANN model, the input dataset was normalized for more reliable results and then rescaled to the original dataset. The activation functions used for the neurons of the proposed ANN model were tangent-sigmoid functions for the 2 hidden layers and linear transfer functions for the output layer. In addition, specific algorithms such as the one proposed in [35] can be implemented in order to speed up the convergence of ANN training phase.

The performance of the NN model and regression between target and NN output are shown in Figures 11 and 12, respectively.

Figure 11

ANN training performance. Epoch is basically one step iteration for all the input dataset in order to train the neural network. In BP algorithm, for each epoch, a different network model with a different set of weights is created. Typically, a large enough value of epoch is chosen to ensure mean squared error converges and overfitting is not observed.

Figure 12

Regression between target and NN output. The regression values measure the correlation between outputs and targets. Overlapping or close points mean a close relationship.

5. Conclusion

In this paper, an unattended deployment approach for underwater sensor networks has been proposed. The proposed approach combines an efficient node placement strategy with an artificial neural network- (ANN-) based control system in order to drive an autonomous multiboat for deploying underwater sensor network (USN) nodes. The node placement strategy provides balanced energy consumption and reduces end-to-end delay of the USN nodes.

Performance results in this paper have shown that the proposed ANN-based control system can drive an autonomous multiboat group to enable them to follow a predefined path set successfully in order to deploy USN nodes. In addition, a set of performance evaluations has been performed to show the efficiency of the proposed node placement strategy. The tradeoffs between energy consumption, end-to-end delay, and number of hops in string network have been also investigated. The results revealed that increasing the hops in the acoustic link almost linearly increases end-to-end delay but reduces the total network energy consumption. These results can be used to determine the number of hops that satisfy the energy consumption and end-to-end delay requirements.

Footnotes

Conflict of Interests

The authors declare that there is no conflict of interests regarding the publication of this paper.

Acknowledgment

This research has been supported by Yildiz Technical University Scientific Research Projects Coordination Department, Project no. 2013-04-04-KAP02.

References

Heidemann

Stojanovic

Zorzi

Underwater sensor networks: applications, advances and challenges

Philosophical Transactions of the Royal Society A 2012 370 1958 158 175

10.1098/rsta.2011.0214

2-s2.0-84855482959

Vasilescu

Kotay

Rus

Dunbabin

Corke

Data collection, storage, and retrieval with an underwater sensor network

Proceedings of the 3rd International Conference on Embedded Networked Sensor Systems (SenSys '05)

2005

San Diego, Calif, USA

ACM

154 165

10.1145/1098918.1098936

Liu

Zhou

Cui

J. H.

Prospects and problems of wireless communication for underwater sensor networks

Wireless Communications and Mobile Computing 2008 8 8 977 994

10.1002/wcm.654

2-s2.0-54949148390

Gao

Foh

C. H.

Cai

On the selection of transmission range in underwater acoustic sensor networks

Sensors 2012 12 4 4715 4729

10.3390/s120404715

2-s2.0-84860123703

Cui

Madan

Goldsmith

Lall

Energy-delay tradeoffs for data collection in TDMA-based sensor networks

Proceedings of the IEEE International Conference on Communications (ICC '05)

May 2005

Seoul, Republic of Korea

3278 3284

2-s2.0-24144479998

Felemban

Energy-delay tradeoffs for underwater acoustic sensor networks

Proceedings of the 1st International Black Sea Conference on Communications and Networking (BlackSeaCom '13)

July 2013

Batumi, Georgia

45 49

Younis

Akkaya

Strategies and techniques for node placement in wireless sensor networks: a survey

Ad Hoc Networks 2008 6 4 621 655

10.1016/j.adhoc.2007.05.003

2-s2.0-39149099807

Akyildiz

I. F.

Sankarasubramaniam

Cayirci

Wireless sensor networks: a survey

Computer Networks 2002 38 4 393 422

10.1016/S1389-1286(01)00302-4

2-s2.0-0037086890

Felamban

Shihada

Jamshaid

Optimal node placement in underwater Wireless Sensor Networks

Proceedings of the 27th IEEE International Conference on Advanced Information Networking and Applications (AINA '13)

March 2013

Barcelona, Spain

492 499

10.1109/AINA.2013.40

2-s2.0-84881042228

10.

Ibrahim

Al-Bzoor

Liu

Ammar

Rajasekaran

Cui

General optimization framework for surface gateway deployment problem in underwater sensor networks

Eurasip Journal on Wireless Communications and Networking 2013 2013 1, article 128

10.1186/1687-1499-2013-128

2-s2.0-84878023085

11.

Akyildiz

I. F.

Pompili

Melodia

Underwater acoustic sensor networks: research challenges

Ad Hoc Networks 2005 3 3 257 279

10.1016/j.adhoc.2005.01.004

2-s2.0-13944252629

12.

Ibrahim

Cui

J.-H.

Ammar

Surface-level gateway deployment for underwater sensor networks

Proceedings of the Military Communications Conference (MILCOM '07)

October 2007

Orlando, Fla, USA

1 7

10.1109/MILCOM.2007.4454996

2-s2.0-47949113744

13.

Erckens

Büsser

Pradalier

Siegwart

R. Y.

Avalon: navigation strategy and trajectory following controller for an autonomous sailing vessel

IEEE Robotics and Automation Magazine 2010 17 1 45 54

10.1109/MRA.2010.935792

2-s2.0-77949625171

14.

Cruz

Alves

Auto-heading controller for an autonomous sailboat

Proceedings of the IEEE Oceans '10

May 2010

Sydney, Australia

1 6

15.

Tuna

Arkoc

Koulouras

Potirakis

S. M.

Navigation system of an unmanned boat for autonomous analyses of water quality

Elektronika Ir Elektrotechnika 2013 19 8 3 7

10.5755/j01.eee.19.8.5387

16.

Stelzer

Pröll

John

R. I.

Fuzzy logic control system for autonomous sailboats

Proceedings of the IEEE International Conference on Fuzzy Systems

July 2007

London, UK

1 6

10.1109/FUZZY.2007.4295347

2-s2.0-50249139990

17.

Tuna

Arkoc

Gulez

Continuous monitoring of water quality using portable and low-cost approaches

International Journal of Distributed Sensor Networks 2013 2013 11

249598

10.1155/2013/249598

2-s2.0-84879387426

18.

Neural Network Toolbox—MATLAB

http://www.mathworks.com/products/neural-network/

19.

Urick

R. J.

Principles of Underwater Sound 1983

Mcgraw-Hill

20.

Stojanovic

Preisig

Underwater acoustic communication channels: propagation models and statistical characterization

IEEE Communications Magazine 2009 47 1 84 89

21.

Glegg

S. A. L.

Pirie

LaVigne

A study of ambient noise in shallow water

http://www.oe.fau.edu/acoustics/

22.

Proakis

J. G.

Digital Communications 1995 3rd

New York, NY, USA

McGraw Hill

23.

Farrell

A. C.

Peng

Performance of IEEE 802.11 mac in underwater wireless channels

Procedia Computer Science 2012 10 62 69

10.1016/j.procs.2012.06.012

24.

Fisher

F. H.

Simmons

V. P.

Sound absorption in sea water

Journal of Acoustical Society of America 1977 62 3 558 564

10.1121/1.381574

25.

Marczak

Water as a standard in the measurements of speed of sound in liquids

Journal of the Acoustical Society of America 1997 102 5 2776 2779

10.1121/1.420332

2-s2.0-0030680104

26.

Underwater Acoustic Modem http://www.link-quest.com

27.

Haykin

S. O.

Neural Networks and Learning Machines 2008 3rd

New York, NY, USA

Prentice Hall

28.

Fazayeli

Wang

Liu

Back-propagation with chaos

Proceedings of the IEEE International Conference Neural Networks and Signal Processing (ICNNSP '08)

June 2008

Zhenjiang, China

5 8

10.1109/ICNNSP.2008.4590298

2-s2.0-51849141786

29.

Roweis

Levenberg-Marquardt optimization

http://www.cs.nyu.edu/~roweis/notes/lm.pdf

30.

Levenberg

A method for the solution of certain non-linear problems in least squares

Quarterly of Applied Mathematics 1944 2 164 168

MR0010666

31.

Marquardt

An algorithm for least-squares estimation of nonlinear parameters

SIAM Journal on Applied Mathematics 1963 11 2 431 441

MR0153071

32.

Ranganathan

The Levenberg-Marquardt Algorithm 2004

Mountain View, Calif, USA

Honda Research Institute

33.

Kılıçaslan

Tuna

Gezer

Gulez

Arkoc

Potirakis

S. M.

ANN-based estimation of groundwater quality using a wireless water quality network

International Journal of Distributed Sensor Networks 2014 2014 8

458329

10.1155/2014/458329

34.

Zhao

Xin

Ren

Guo

Application and comparison of BP neural network algorithm in MATLAB

Proceedings of the International Conference on Measuring Technology and Mechatronics Automation

March 2010

590 594

10.1109/ICMTMA.2010.492

2-s2.0-77953141215

35.

Quan

D. M.

Ogliari

Grimaccia

Leva

Mussetta

Hybrid model for hourly forecast of photovoltaic and wind power

Proceedings of the of IEEE International Conference on Fuzzy Systems (FUZZ '13)

July 2013

Hyderabad, India

1 6

10.1109/FUZZ-IEEE.2013.6622453