1. Introduction
Energy harvesting is a promising solution to provide self-sustainability and extend the lifetime for energy-limit wireless sensor networks (WSNs) [
1,
2]. Thus, it has attracted much attention from researchers in recent years [
3]. However, the energy harvesting process from the natural environment and the radio frequency signals [
4] is instable, due to the time change of the day, the season or other factors [
5]. Wireless energy transfer (WET) [
6,
7], as a friendly means of compensating energy, can transfer energy from some energy-rich sensor nodes to others with energy-hungry sensor nodes so as to enhance the overall network performance [
8]. Meanwhile, due to the broadcast nature of wireless communications, the data signals of simultaneous transmissions cannot avoid interfering with each other in the same frequency band [
9]. As a result, it decreases the network performance.
Because of these considerations, we investigate the energy harvesting WSNs and concentrate on the delay minimization problem of the WSNs with interference channels. The delay of every data link is determined by the information rate on the link, which is monotonically decreased as the rate of the link for the fixed data flow over it [
10]. The information rate is monotonically increasing in SINR. We focus on the capacity assignment problem, which is similar to Bertsekas et al. [
10]. In particular, compared with the special case, in which information and energy transfer channels are orthogonal to each other [
11], we consider the general case of the communication model. In other words, the data transmission channels interfere with each other. This is a more realistic and meaningful model for the capacity assignment problem.
Therefore, by considering the energy consumption and power allocation for the fixed data flow, we formulate the capacity assignment problem in the energy harvesting WSNs with interference channels as a non-convex optimization problem. It is constrained by data flow conservation conditions, information rate requirements, energy and power consumption. Employing the relatively high SINR, the non-convex optimization problem can be transformed into a convex optimization problem by convex approximation in “log-sum-exp” form [
12]. The solution properties of the transformed capacity allocation problem are derived by Lagrange duality. Then, it is available to search the optimal Lagrange multiplier and obtain the optimal solution to minimize total delay for the energy harvesting WSNs with interference channels in a time slot. Finally, we solve the approximate convex problem by the CVX solver [
13].
Our study is related to and based on the previous classical works on a capacity allocation problem in communication networks [
10]. In [
14], the simultaneous routing and resource allocation (SRRA) are investigated. A capacitated multi-commodity flow model is used to describe the data flows in the wireless networks. The optimization problem is solved by the dual-decomposition method. A general flow-based analytical framework is presented in [
15]. In order to balance aggregate user utility, total network cost, power control, rate allocation, routing and congestion control are jointly optimized in wireless networks. Channel-aware decision fusion by MIMO channels is investigated in an existing large antenna-array at the decision fusion center [
16]. A decentralized multi-sensor estimation problem is studied in [
17]. In a WSN with a coherent MAC, the detection and estimation of a zero-mean Gaussian signal is investigated in [
18]. In [
19], sensors simultaneously report sensed data to a fusion center with multiple antennas in a WSN and a Gaussian mixture channel model is adopted to attain a general fading description of the channels collective between the sensors and the fusion center. A machine learning based method is proposed for joint scheduling and power control in [
20]. However, the previous works have not considered the energy harvesting and energy cooperation. Fouladgar et al. [
21] investigates the optimization problem of simultaneous information and energy flows in graph-based communication networks with energy transfer. In [
22], a model of multi-hop information transmission and energy transfer in TDMA-based multi-hop WSNs is proposed. Among previous studies, the most related to ours is that in [
11], which investigates the delay minimization problem in the energy harvesting wireless communication networks with energy transfer. However, though Gurakan et al. [
11], Fouladgar et al. [
21] and Xu et al. [
22] study the optimization problem of the joint information transmission and energy transfer, they neglect the interference among the data links. These motivate us to consider a general capacity assignment problem which is to minimize total delay in the energy harvesting WSNs with interference channels.
Lagrange method is a powerful studied tool which has been widely applied for the resource allocation problem in wireless networks [
11,
23]. It is worth noting that, although we utilize a similar mathematical approach to that in [
11] for modeling and solving the capacity assignment problem, our study is significantly different from the previous studies: the previous studies only consider a special case where the data transmission channels are orthogonal to each other, rather than consider the impact of data transmission interference. However, the more realistic case is that data transmission channels interfere with each other, which is one of the critical issues to be tackled in this study. Therefore, we need to remodel the capacity assignment problem for the energy harvesting WSNs with interference channels in a time slot.
In this paper, our main contributions are as follows:
We investigate a general and meaningful model of capacity assignment problem where the data links interfere with each other in the energy harvesting WSNs.
Considering the relatively high SINR, we transform the non-convex optimization problem into a convex one by convex approximation, and also derive the optimal solution properties by Lagrange duality.
Numerical results show that the interference signals significantly affect the network performance; the energy transfer can help to decrease the total network delay.
The rest of this paper is structured as follows.
Section 2 introduces the network model and problem formulation.
Section 3 investigates the capacity assignment problem with interference channels in a time slot.
Section 4 demonstrates the performance results. Finally,
Section 5 concludes the paper.
Notations. Throughout this paper, matrices and vectors are denoted by boldface uppercase and lowercase letters, respectively. stands for the natural logarithms. All numbers, vectors and matrices take real values in this paper. For a vector , is the ith element; similarly, denotes the th entry of matrix .
2. System Model and Problem Formulation
In this study, each sensor node not only has the capability of harvesting energy and sensing data from the ambient environment, but it also can transmit or receive energy and data. As the data transmission channels interfere with each other, the interference signals among the data flow may be unavoidable. Hence, we consider an energy harvesting WSNs model with interference channels as shown in
Figure 1.
Let be a directed graph modeling N sensor nodes which are placed randomly and seamlessly in a certain area. The vertices set V = {} is composed of one sink node and N sensor nodes. The edges set E is composed of the communication links between the sensor nodes, i.e., , if and only if a node can send a message to a node with the power constraint .
A data collection tree
[
24] is constructed for the energy harvesting WSNs with sink
at level 0. It is an acyclic spanning subgraph of
where
and
. In the data collection tree
T, each sensor node
can collect the sensing data from the area of interest and then store it for future transmission in a data buffer. Each sensor node
has to send the sensing data to sink
periodically in multi-hop fashion and half-duplex mode under the interference channel. Sensor nodes
and
are siblings if they have the same parent. Note that a sensor node can be either a transmitter, a relay or a receiver, which is determined by its location in WSNs. For brevity, the ordered pair
is replaced by
in the following sections (Throughout the paper, we denote sensor node indices by the first subscripts
i,
j and
n. The subscript
i and
j denote the start node and the end node at each link (i.e., data link and energy link), respectively.).
We consider the following interference model to characterize the relationships among data links in tree-based energy harvesting WSNs. For any data link
l, let
be the depleted power which transmits data flow from the sensor node
to the sensor node
in a time slot. We employ
as transmission power vector. Then, the received SINR of data link
l is
where
is the channel gain from the transmitter of data link
to the receiver of data link
l, which is dependent on various factors such as path loss, shadowing and fading effects. Particularly,
is the gain of primary link
l and
denotes the channel noise power [
25]. We adopt a simple distance based on the path-loss model to calculate the data link gains as
, where
is constant between 3 and 4, which depends on the ambient conditions [
26]. We assume that the channel gain remains constant and does not change over the time slot.
To illustrate,
Figure 1 shows a tree-based energy harvesting WSNs with interference channels. In the figure, there are only five active links at the first time slot since we employ half-duplex sensor nodes [
26]. Meanwhile, the network has five energy cooperation links, which can transfer energy to sensor nodes’ required energy. It guarantees that the sensing data can be successfully sent to the receivers at the time slot. In
Figure 1, we assume that the active link
is the primary link, the receiver
not only receives the data flow signal from the transmitter
, but also receives the interference signals from other transmitters
,
,
and
. The interference signals are represented by red dashed lines with arrows. Meanwhile, the sensor node
can transfer energy to the sensor node
through the energy link
. At the same time, other receivers also receive interference signals from active links’ transmitters except themselves. For brevity, we do not label them in
Figure 1.
2.1. Network Data Flow Model
Let us denote the data link as (The data link can be denoted or l, they can be interchangeable in this paper). The topology of data flows can be described by an matrix . The entries of matrix can be defined by , which is incident with sensor node n and data link l. More precisely, each entry is defined as
Let us define as the set of incoming data links to sensor node and as the set of outgoing data links from sensor node , respectively. Assume that the data flow on each data link follows the uniform distribution . The set of data flows is referred to as the L-dimensional flow vector. The divergence vector associated with the data flow vector is an N-dimensional vector which indicates the nonnegative amount of outside data flow injected into the sensor node . Suppose that the data flow is lossless over links. For every sensor node , the flow conservation conditions can be expressed as
The data flow conservation through the total WSNs can be rewritten as
Moreover, the data flow over each data link l cannot exceed the information carrying capacity , i.e.,
2.2. Network Energy Flow Model
In this section, we present the energy model for the case where each sensor node has a single energy harvest in a time slot. Notice that we only consider harvested energy from the ambient environment and transferred energy from the neighbor sensor node, and ignore the energy contributed by interference in this paper.
2.2.1. Energy Harvesting Model
Each sensor node powered can harvest energy from the ambient environment. Since the transmission consumption is the most significant amount of energy, we only account for energy consumption of transmitting data in this study. It is assumed that the energy harvesting sensor node has a capacity battery
which is enough for transmitting the data. The capacity of storage is considered to be constant, i.e., energy outage and circuitry cost are negligible. Since energy harvesting sources have a random nature, the energy arrivals are considered as an independent and identically distributed (
) Poisson distribution
with parameter
[
27,
28]. We assume that the energy arrivals occur only once in a time slot. Let
be the harvested energy of a sensor node
in a time slot,
. The harvested energy in a time slot can be exploited only in a later time slot.
2.2.2. Energy Cooperation Model
Energy cooperation depends upon the statistics of the energy harvesting and the energy consumption of the sensor nodes. In general, for a sensor node
, the more data flow is transmitted, the more energy is required. In order to replenish the energy of energy-hungry sensor nodes, the technique of wireless energy cooperation [
29] is adopted in our study. It is assumed that the energy is unidirectionally transferred from the sensor node
to the sensor node
in a time slot, the transfer efficiency is
,
, due to energy loss in transmission and conversion.
2.2.3. Energy Flow Model
In the previous analysis, we utilize
N-dimensional vector
to present the harvested energy vector for the WSNs. In the energy transfer process, the wireless energy links are similar to data links. The wireless energy link
q is also denoted as an ordered pair
in energy routing. The energy can be sent from the sensor node
to the sensor node
over energy link
q,
, if the energy of the sensor node
is not enough energy to operate. The energy transfer efficiency is
on each energy link
q where
. It implies that
amount of energy is transferred on wireless energy link
q from the sensor node
to the sensor node
, and the sensor node
receives
amount of energy. The request for energy transfer is known in advance, whereas the amount of transferred energy is unknown. The topology of energy flow can be denoted by an
matrix
. The entries of the matrix
can be defined by
, which is incident with sensor node
n and wireless energy link
q. More specifically, each entry
can be described as [
11]
We define and as the set of outgoing and incoming wireless energy links at the sensor node , respectively. The variable is the amount of energy transferred. Let vector be the L-dimensional energy flow vector.
2.3. Communication Model
For the energy harvesting WSNs with interference channels, we focus on minimizing the total delay and enhancing the network performance in order to ensure that sensing data on each data link can reach the sink as quickly as possible. It is similar to [
10,
11]; we assume that each time slot is large enough and the delay on the data link
l follows the
queueing model in this work. It can be defined as
where
is the amount of data flow and
is the information carrying capacity of communication link
l in which
.
According to the Shannon formula [
11,
30], the information carrying capacity (or information rate)
of data link
l can be expressed as
where all logarithms in our study are taken to the base
e.
At every sensor node
, the total power depleted (In contrast to transmission power consumption, the energy consumption of sensing data is ignored in our study.) on transmission data link
l and energy link
q are constrained by the usable energy as:
Let
, where
, which only distinguish the outgoing links at each sensor node
n. Hence, the energy availability constraints in Equation (
9) can be rewritten as
Notice that the power and energy can be interchangeable in a unit of time slot in this paper.
3. Capacity Assignment Problem in Energy Harvesting WSNs with Interference Channels
We consider the capacity assignment problem in WSNs with interference channels for a single energy harvesting sensor node in a time slot. Assume that the data flow assignments on all data links are fixed and available for harvested energy and transferred energy. The total delay D in a WSNs is
Hence, the goal of minimizing total delay in the energy harvesting WSNs with interference channels can be written as
As shown in
Figure 1, because the data transmission signals of active links interfere with each other, each data flow signal cannot perform interference cancelation and is treated as an additive noise compared with the primary link signal. By utilizing the information rate
in Equation (
8), the minimizing total delay in the energy harvesting WSNs with interference channels is
By analyzing Equation (13), we find that the minimizing of the total delay depends on the maximizing of the information carrying capacity . Meanwhile, because the information carrying capacity is a monotonically increasing function of , the maximizing of information carrying capacity depends on the maximizing of the .
Note that the optimization problem (13) is non-convex since both the objective function (
13a) and the constraint condition (
13c) are non-convex in terms of transmission power vector
, and it is not straightforward to attain the optimal solution. Therefore, we need to study the fundamental properties of the optimization problem (13) and transform it into the convex optimization problem.
3.1. Convex Approximation
We can get a convex approximation for capacity assignment problem with interference channels when the SINRs are relatively high (e.g., SINRs ≥ 5 or 10). The information carrying capacity (or information rate)
by using the Equation (
1) can be rewritten as
Let
, i.e.,
for
, we define
where the functions
are concave in the vector
.
With the approximation information carrying capacity formula, the optimization problem (13) can be reformulated as
where the objective function (
16a) is a convex function in the new variable
[
12]. The information carrying capacity constraint (
16c) is convex function in
and
. This means that the optimization problem (16) is a convex optimization problem and the global optimal solution can be found.
Remark 1. Here, we use the approximationwhich is reasonable for the optimization problem (13), since. This implies that the approximation is an underestimate and a tighter constraint for the information carrying capacity. Therefore, the solution of convex problem (16) is always feasible for the original optimization problem (13).
3.2. Properties of Capacity Assignment Problem with Interference Channels
For convex optimization problem (16), we form the dual problem by introducing Lagrange multiplier , and . The Lagrangian function is given by
The Lagrangian function (
17) corresponds to Lagrange dual function
as
The dual optimization problem is
The KKT optimality conditions hold for the convex optimization problem (16), thus we have
where
The complementary slackness conditions are
We extend Lemmas 1 and 2 in [
11] and derive some properties about the optimal power allocation with interference channels as follows.
Lemma 1. The feasibility of the convex optimization problem (16) requires.
Proof. The proof is a similar procedure in [
11]. If the convex optimization problem (16) is feasible, the objective function (
16a) must be bounded. The constraint condition (
16c) for any data link
l means that the objective function (
16a) is unbounded. Thus, the constraint condition (
16c) must strictly satisfy the inequalities for all data link
l. From Equation (
24), we can conclude that
. □
Lemma 2. At each sensor node, the optimal power allocation with interference channels among data links satisfies Proof. The proof is a similar procedure in [
11]. Combining Equation (
20) and Lemma 1, we attain
Since the outgoing links
l and
i reside in the same sensor node
n, we have
Thus, we can conclude that Equation (
26) holds. □
In the next subsections, we separately solve the convex optimization problem (16) under two cases, i.e., no energy transfer and energy transfer.
3.3. Case without Energy Transfer
As energy transfer does not occur in this case, we have . Thus, the convex optimization problem (16) becomes only in respect of as follows:
Since we employ half-duplex WSNs, the optimization problem can be considered active data links in the energy harvesting WSNs with interference channels as
If the optimization problem (30) is feasible, then it requires
which we assume that it holds. Similar to Equations (
17) and (30) corresponding to Lagrangian function
with
is
Meanwhile, the KKT optimality condition is
and the complementary slackness condition is
From Equation (
33), we have
For the total energy constraint condition Equation (
30b), the optimal power allocation can be found by searching the optimal
.
Remark 2. The constraint condition (30c) is not included in the Lagrangian function (32), since the constraint condition (30c) will always hold when the convex optimization problem (30) is feasible. 3.4. Case with Energy Transfer
Next, we solve the case with energy transfer, which implies for some energy links q. The convex optimization problem (16) becomes
According to the half-duplex mode, the optimization problem (37) which has active data links in the energy harvesting WSNs with interference channels can be written as
As in
Section 2.2.2, it is assumed that some energy
is transferred from the sensor node
to the sensor node
over energy link
q. Since sensor node
only transfers energy and does not transmit data, the energy causality constraint condition on sensor node
is denoted as
Therefore, by combining Equations (
36) and (
39), we can attain optimal power allocations if we find the optimal
.
The Lagrangian method can provide some ideas and in-depth insight into the above-defined optimization problem. However, it is difficult to find a closed-form optimal solution. Therefore, we use the CVX solver [
13] to tackle the optimization problems (30) and (38) in this paper.