1. Introduction
Underwater Acoustic Sensor Networks (UASNs) have attracted much attention due to its application in many fields such as environmental monitoring, military defense, disaster prevention and ocean resource exploration [
1,
2,
3]. In UASNs, underwater nodes measure water phenomena’s values with equipped sensors, communicate with each other by acoustic signals and collaborate to send the sensed data to the sink node. In this process, nodes’ locations need to be aware to support topology control, routing decision, subsequent data processing, etc. [
4]. Therefore, localization is one of the most essential and fundamental services in UASNs.
Although there has been a lot of research for localization in terrestrial wireless sensor networks, they can’t be directly applied to UASNs due to the following reasons: (1) the acoustic communication has long propagation delay, and the acoustic signals often propagate with varying speeds in non-straight lines due to the inhomogeneity in terms of temperature, salinity and pressure; and (2) underwater nodes move with tides and currents continuously, resulting in frequent changes of network topology. These unique characteristics of underwater environment pose huge challenges for localization in UASNs.
Recently, many localization methods for UASNs have been widely explored [
5,
6]. Most of them assume that there are sufficient beacon nodes in the localization process. The unknown nodes perform localization using the spatial correlation between nodes (e.g., the distance estimations to neighbor beacon nodes). The localization performance is usually subject to the ranging accuracy and the algorithm efficiency. In many cases, UASNs are sparse since a limited number of nodes are deployed in a wide area. Moreover, the beacon nodes account for a small proportion in consideration of cost and efficiency. It is often the case that the unknown nodes can’t be localized due to the lack of sufficient beacon nodes, which results in low localization coverage.
In addition, ambient noise is an important factor that impacts localization performance, especially for range-based localization methods. Most existing works model the range noise as a Gaussian variable with fixed mean and variance, which can’t reflect the actual situations of harsh underwater surroundings. Generally, the range noises have two features: (1) the noises depend on the propagation path of the acoustic signal and change with the motion of underwater nodes [
3]; (2) the range error usually increases with the distance between two nodes, which implies that the noise is closely related to the distance.
The characteristic of continuous mobility of underwater nodes has its own disadvantage and advantage. On one hand, a localized node will become unlocalized when beacon nodes move out of its communication range; on the other hand, the continuous movement of each node implies that the locations in its trajectory are dependent temporally. For example, the current location of one node is closely related to its locations at previous and next time instants. However, this temporal dependency among locations is neglected in most existing works.
In fact, localization based on spatial correlation in UASNs, especially with insufficient beacon nodes in sparse deployments, can be improved significantly using temporal dependency information. As illustrated in
Figure 1, the location estimation of node C at time
is denoted as
. It moves with currents and becomes neighbors of beacon nodes A and B at time
t. Localization with two beacons will calculate two candidate locations
and
. It is obvious that
is the correct estimation under the assumption of low current velocity and short localization period.
To fuse spatial correlation and temporal dependency effectively, the approach commonly adopted is Bayesian filtering, which has been widely used in target tracking [
7,
8,
9,
10,
11]. It attempts to obtain the posterior probability distribution of the state based on the available measurements. Since the measurements such as distance estimations are nonlinear functions of the state, the analytic solutions of the state estimation are intractable. Hence, the approximation approaches are needed. The Monte Carlo sampling methods such as Particle Filtering (PF) approximate a posterior probability distribution by use of a set of weighted samples. This method is not suitable for resource-constrained UASNs due to huge computation and storage costs. Alternatively, variational inference aims to find some decomposable distribution over the state to approximate a posterior distribution. It can accommodate an arbitrary measurement model and the convergence can be guaranteed. Aiming at the above challenges in UASN localization, we propose a spatial–temporal variational filtering algorithm, which is named STVF. It provides real-time localization in a distributed manner. First, we use a general state transition model to characterize the mobility pattern of the nodes. In virtue of this model, each node can predict its coarse location based on its previous location. This coarse location is then updated according to Bayes rule. Specifically, the posterior probability distribution of the node location can be obtained by taking the predicted location as a priori information and incorporating the spatial measurements as the likelihood information. Considering that the spatial measurements are the nonlinear function of the node location, we adopt an optimization method based on variational inference, in which the state variables related to the node location are optimized in turn. After the convergence value of the node location is obtained, we calculate the confidence value of the nodes, label the nodes with high confidence as reference nodes and localize the rest nodes iteratively. Therefore, the localization accuracy and coverage are both guaranteed in the STVF algorithm.
The rest of this paper is organized as follows: in
Section 2, existing related works on UASN localization are briefly reviewed. Then, in
Section 3, the network model is given and the UASN localization problem is formulated.
Section 4 presents the design details of the STVF algorithm, in which the state evolution model and the measurement model are studied to support subsequent variational prediction and update phases. In
Section 5, performances of the STVF algorithm are evaluated by simulations and are compared with the similar SLMP algorithm. Finally, we conclude the paper in
Section 6.
2. Related Work
Existing localization schemes for UASN can be divided into two categories: range-free and range-based. While the range-free scheme is rarely adopted in realistic environments due to its poor localization accuracy, the range-based scheme has been paid more attention in recent literature. Ranging methods include Time of Arrival (TOA), Time Difference of Arrival (TDOA) , Angle of Arrival (AOA) and Received Signal Strength Indicator (RSSI) [
12]. RSSI relies on an accurate signal attenuation model, which is hard to obtain in the time-varying acoustic environment. AOA necessitates the nodes to be equipped with directional antennas, resulting in additional costs. TDOA needs complex computation and thus consumes more energy [
13]. In contrast, TOA is preferred in UASN localization since a long propagation delay provides high time resolution for distance estimation.
The TOA method gets the difference of message sending and receiving time, and then calculates distance by multiplying the time difference and acoustic speed. Its accuracy is vulnerable to multiple factors such as time synchronization, multipath effect and stratification effect. In SLSMP [
14] and STSL [
15], time synchronization is first performed to get clock skew and offset between beacon nodes and target nodes. Hence, the distance between two synchronized nodes can be estimated precisely. An alternative approach to time synchronization is using half of the round-trip time as the propagation delay [
16]. This method eliminates the impact of clock offset, but increases the communication overhead. Beniwal et al. [
17] proposed a time synchronization-free algorithm. They assumed that beacon nodes dive and rise in a vertical direction. Sensor nodes passively receive two localization messages from one beacon node and estimate the distance to the beacon node geometrically. In [
18], an EM algorithm is proposed to alleviate the degradation caused by multipath effect, in which line-of-sight (LOS) and non line-of-sight (NLOS) links are identified to support subsequent localization. Considering the acoustic speed varies with depths, a ranging method with stratification effect compensation is put forward in [
19], in which bias-free distance estimation can be obtained through integrating the function of depth. These methods are aimed at improving the ranging accuracy.
Once the distances to reference nodes are determined, a localization algorithm is executed to convert the distance estimations into the node location. In RLS [
20], a new message exchange mechanism is designed to enable fast response to events and reduce communication overhead. The nodes that detect an event send messages to a sink station, wherein the node locations are estimated by trilateration. 3DUL [
21] projects three neighbor beacon nodes of each unknown node onto a horizontal plane, and conduct trilateration when four triangles in the plane are all robust. This method has high energy consumption due to two-way TOA. In [
22], a novel hybrid network DR-OSN is proposed, which consists of the double-head nodes deployed on the sea surface and the moored underwater nodes linked to double-head nodes with mooring lines. The moored underwater nodes are first localized by leveraging the free drifting movement of their surface nodes with GPS module and then turn into beacon nodes to localize other underwater nodes and the floating nodes without GPS module. The whole localization process is based on trilateration and does not need presence of designated beacon nodes. Besides trilateration, Bian et al. [
23] proposed a hyperbola-based approach to eliminate the localization ambiguity existing in trilateration. MP-PSO [
24] searches the locations of the beacon nodes using particle swarm optimization technique and calculates the velocities of the unknown nodes based on the velocities of their neighboring beacons. Then, the current location of an unknown node can be estimated by adding its previous location with its velocity. In [
25], MDS-MAP (C, E) algorithm is proposed, in which the distance estimations within one-hop and two-hop nodes are obtained to form a distance matrix, and then the multi-dimensional scaling is performed to transform the matrix to the nodes’ locations.
Many localization algorithms reveal the weakness when they are applied in localization for large-scale UASNs. The situation of sparse beacon nodes brings new challenges. The unknown nodes can’t be localized due to a lack of sufficient beacon nodes. SLMP [
26] provides an iterative localization technique to solve this problem. The unknown nodes estimate the locations of the beacon nodes based on their mobility speed vectors and the locations received last time. The location of an unknown node is calculated by multilateration when the number of its reference nodes is more than 4. The localization confidence is then obtained based on the confidence of its reference nodes and its own localization accuracy. The localized node whose confidence value is larger than confidence threshold becomes a new reference node and helps other unknown nodes to localize themselves. MANCL [
27], TPS [
28] and TP-TSFL [
29] adopt a similar localization strategy. In TPS, three-dimensional Euclidean distance estimation is used to compute the distances to the reference nodes within two-hop range. This method is supplemented in MANCL using communication and vote mechanisms. Furthermore, MANCL utilizes DV-HOP distance estimation to search for new reference nodes. DRL [
30] presents a double rate scheme, in which low rate mode is used for estimating the distances to multi-hop reference nodes and a high rate mode helps to transmit a mass of data in the localization procedure. These algorithms improve localization coverage and have a low communication cost, but with low localization accuracy.
The algorithms mentioned above make the best of spatial correlation between nodes; however, none of them exploits temporal dependency. The characteristic of continuous movement with time can help to improve localization performance. In [
31], an offline localization algorithm that takes full advantage of all available distance estimations between nodes at different time instants is proposed. The factor graph is employed to express the temporal dependency of consecutive time instants and the distance constraints on node pairs. Although this method gains high localization accuracy, it does not offer real-time localization and is not suitable for long-term UASNs. JSL [
19] utilizes an IMM filter to alleviate the impact of the node mobility. It predicts the locations of the nodes based on that of previous time instant and conducts the correction according to the location measurements. Two filters running in parallel are used, which are Kalman filter for uniform moving and extended Kalman filter (EKF) for maneuvering, respectively. The final location estimation comes from the combination of the estimations from the two filters. The drawback of JSL is that the location measurements may introduce additional errors and EKF has a possibility of divergence. SLSMP [
14] alleviates localization error by applying Kalman filter and averaging filter. It performs localization only when the target node locates at fixed sending point, which has no universality.
4. Spatial–Temporal Variational Filtering Design
As described in
Section 3, in the Bayesian framework, the estimation of
requires a recursive update of posterior distribution of each unknown node. The estimation accuracy depends on the appropriate definition of the temporal evolution model of the state transition
and the spatial measurement model
. In particular, the definition of the two models needs take the irregularity of the continuous mobility and the spatial-temporal variation of the ambient noises into account.
4.1. State Evolution Model
Considering an accurate mobility model may cause the degradation of localization performance due to the deviation of the actual motion of nodes, in this paper, we employ a General State Evolution Model [
35,
36,
37] to fully characterize the complex mobility dynamics of the nodes. The state of the unknown node is represented by its location and velocity
, where
denote the locations of the unknown node in the
x- and
y-axis directions, and
denote the velocities of the unknown node in the
x- and
y-axis directions. The necessity of incorporating the velocity in the state of a node comes from the calculation of the confidence value of its measurements, as we will see later. It is important to note that there are no constraints on the velocities and directions of the unknown node, i.e., all the information that is available in the localization process are from the spatial and temporal observations.
In the state evolution model, the current state
is assumed to be Gaussian distributed with the expectation vector
and precision matrix
, that is,
. The expectation and precision are both assumed to be random variables, for capturing the uncertainty of the state and reflecting the dynamics of underwater environments as much as possible. The expectation
follows a Gaussian distribution
, which means that
transits from the expectation
of the previous state in a completely random manner and the transition uncertainty is controlled by the precision matrix
. For the convenience of recursively computing the posterior distribution in subsequent variational inference, the precision matrix
of the state is assumed to follow a conjugate Wishart distribution with the degrees of freedom
and the precision
, in which the uncertainty of the state
around its expectation
is captured. In summary, the state evolution model can be represented as follows:
where the state
is augmented with its expectation
and precision
, and the fixed parameters
,
and
collaboratively determine the ability to model the state transition between two successive time instants. Based on this model, the probability distribution of the state
can be obtained according to the Equation (
1), and serve as the a priori distribution of the following procedure.
4.2. Measurement Model
In the measurement model, the coarse estimation of the state is refined with the spatial observations, which mainly consist of the measurements from the neighbor reference nodes in . Generally, a one-way TOA method is adopted to produce the measurements due to two considerations: (1) it needs no additional hardware and thus saves costs greatly; (2) the unknown nodes passively listen to the beacon signals and only need a simple calculation, resulting in remarkable energy consumption reduction.
Supposing that the time synchronization has been achieved in the whole network, the traditional TOA method estimates the distances from the unknown node to its neighbor reference nodes by
, where
and
denote the sending and receiving time of the acoustic signal, respectively, and
v is a constant speed value predetermined. However, considering the varying speed and non-straight propagation, the traditional TOA method often suffers from large errors and causes the degradation of the localization performance. Alternatively, in STVF, we use the time difference as the measurement:
where
denotes the value of time difference from the unknown node to the j-th reference node, and
denotes the measurement noise that follows a Gaussian distribution with zero mean and precision
, namely,
. Hence, the measurement follows
. The actual propagation delay
depends on the distance between the unknown node and the j-th reference node
and current acoustic speed
, i.e.,
. The distance
can be further represented with
, where
stands for the state of the j-th reference node and A is a matrix for selecting the location information from
and
:
The speed
is assumed to be Gaussian distributed with the expectation
and the precision
, namely,
. This a priori information reflects the time-varying characteristic of the propagation speed and implies that the propagation speed within the communication range of the unknown node is assumed to be same. It also helps to compensate the measurement error caused by other uncertain factors such as multipath effect and stratification effect. To further capture the uncertainty of the measurements, the precision
of the Gaussian variable
is naturally assumed to be the conjugate Gamma distribution
, where
and
are the shape and scale parameters, respectively. In summary, the uncertainty of the measurements can be captured by the randomness of the acoustic speed and the precision scalar:
where the fixed parameters
,
,
and
provide the ability to model the complex dynamics of underwater environments.
4.3. Variational Inference
Given the state evolution model on Equation (
3) and the measurement model on Equation (
5), the original state
is extended to an augmented state
. Accordingly, the prediction and posterior distributions have the form of
and
, respectively. However, the analytic solutions of the posterior distribution are intractable due to the nonlinear function of the state in the measurement model. Alternatively, we seek to find a decomposable distribution
to approximate the posterior distribution
using a variational approach, in which the Kullback–Leibler (KL) divergence between the two distributions is minimized:
where
Under the condition of maximizing the log-likelihood of the measurement data
, the minimization of KL divergence is equivalent to the maximization of the lower bound of
[
38]:
We then substitute the factorized form of
into Equation (
7) and conduct optimization with respect to each distribution in turns:
where
denotes an expectation with respect to the
distribution. Suppose we keep the distributions
fixed,
can be regarded as a negative KL divergence between
and
. The minimization of this KL divergence produces the optimal solution of
:
This means the optimal solution is obtained by considering the log of the joint distribution over both the state and observation variables and then taking the expectation with respect to .
4.4. Variational Prediction
In the prediction phase, we seek to find a coarse estimation of the current state
based on previous state
and the state evolution model on Equation (
3). Assume we have obtained the approximate distribution
at time instant
, according to Equation (
1), the predictive distribution of the current state can be given as:
where
means the transition distribution from the previous state to the current state. We substitute the definition of each state variable into the transition distribution and rewrite it as:
Considering the constraint
, the predictive distribution takes the form as follows:
where
. Hence, the current state only depends on the variable
of the previous state through the transition distribution
. Assume
is Gaussian distributed,
, the marginal distribution
of the joint Gaussian distribution
is also a Gaussian distribution:
where the expectation vector
and the precision matrix
. Now the exact forms of all the components of
have been determined, and we can approximate it with the distribution
using the variational approach. According to Equation (
8), the distribution
is inferred as follows:
It is observed that the approximate distribution
only depends on the state variables
and
. By substituting Equations (
3) and (
11) into Equation (
12), we have:
where the subscript
is suppressed as
. Hence, the predictive variable
follows a Gaussian distribution
. Its expectation vector and precision matrix is derived as:
Similar with the deduction process in Equation (
12), the distribution
can be approximated as:
which means that
follows a Wishart distribution
with the parameters:
For the variables
and
, they do not obtain additional information from the transition distribution since they are independent of
and
. Therefore, the distributions of
and
remain the same as
and
in Equation (
5), with the following parameters:
The distribution of the variable
depends on
and
, which can be deduced as follows:
From Equations (
13)–(
16), we can see that each component of
depends on the expectations computed with respect to other components, which suggests that the parameters of each component distribution can be optimized iteratively. Nevertheless, we delay the iterative optimization to the following update phase because the spatial observations can be incorporated to refine the inferred locations of the unknown node.
4.5. Variational Update
In the measurement model, the measurement
is actually time of flight from the reference node
to the unknown node, which provides the spatial information to update the predicted variables
,
and
. The overall spatial information provided by all the reference nodes in
can be represented with the joint likelihood distribution:
where
denotes the number of the reference nodes in
. Next, we treat the predictive distribution
as a priori information, fuse it with the likelihood information and derive the posterior distribution as:
where the distribution form of
has been deduced in the prediction phase. Let
denotes the approximate posterior distribution. It is observed that the measurement
is independent of the variables
and
, thus the likelihood function does not contain the information for further updating them. The distributions of
and
remain the same as
and
, with the following parameters:
The involved expectations have the expressions:
The remaining expectations involve the joint optimization of the component distributions
and
, which can be derived with the same procedure as Equation (
12):
We can see that the logarithm likelihood function is nonlinear over the variables
and
, which means that the analytic solutions of
and
are intractable. Hence, we resort to an importance sampling method to compute the expectations approximately, wherein the Gaussian distribution is treated as the proposal distribution and the weights are calculated according to the likelihood function. Let
denote the samples of
and the corresponding weights, where
and
. Its expectation is given by:
For
, the samples are drawn by
and the weights are calculated by
. Its associated expectation and precision are approximated as:
Finally, the component distribution
can be derived as:
which implies that the variable
follows a Gamma distribution, and its shape and scale parameters have the expressions:
Equations (
18)–(
23) suggest that these parameters can take turns being updated until convergence. The pseudo-code of the whole localization process is summarized in Algorithm 1.
Algorithm 1: STVF localization algorithm. |
|
4.6. Iterative Localization
Considering the hardware costs and the deployment difficulties, the beacon nodes usually account for a small proportion of all the nodes in most UASNs. The network has low localization coverage, i.e., many unknown nodes can’t be localized due to lack of sufficient beacon nodes, especially for large-scale UASNs. To solve this problem, an iterative localization scheme is adopted, in which an unknown node that has been localized with high confidence can serve as a reference node and broadcast beacon signals to localize other unknown nodes. The confidence value of an unknown node is not only related to its localization error, but also related to the confidence values of its reference nodes. In STVF, the confidence value is calculated as:
where
D represents the sum of the estimated distances to all the reference nodes:
and
E represents the sum of errors between the estimated and measured distances for all reference nodes:
If the confidence value of the localized node is higher than the confidence threshold , the node is labeled as a new reference node. Then, it will broadcast beacon signals to localize other unknown nodes.
The variational update procedure reveals that the measurements from the reference nodes are vital for improving localization accuracy. The confidence value of a reference node reflects its own localization accuracy, whereas the quality of its measurement to the unknown node is also subject to other factors such as distance and velocity. Herein, we consider the confidence value of each measurement based on that of a reference node. Next, we present the method for updating the confidence value of the measurements.
Intuitively, the long distance between the unknown node and the reference node will incur large errors, i.e., the measurement error increases with the distance. Similarly, the reference nodes with high velocities usually move a long distance between two consecutive time instants, which will increase the localization error as well. Thus, the confidence value of a measurement from the reference node
can be defined as:
where
,
and
are the weight values that satisfy
.
denotes the confidence value of the
i-th reference node. The function
computes the normalization value of
as:
where
and
are the minimum and maximum values of all the measurements
. The function
computes the normalization value of
as:
where
and
are the minimum and maximum values of the velocities of all the reference nodes. The velocity of a reference node is calculated as
.
Based on Equations (
27)–(
29), the confidence values of all the measurements can be obtained. We sort the measurements according to their confidence values and select the former
measurements to be used in the update phase, where
denotes the threshold value of the reference nodes’ number and will be determined in the simulations.
6. Conclusions
Focusing on the problems in UASN localization, including continuous movements of nodes, varying acoustic speed and dynamic noises, we have presented STVF, a novel localization method that fuses spatial correlation and temporal dependency information. In STVF, the mobility patterns of the nodes are characterized by the general state evolution model, and the variation of the acoustic speed and the dynamics of the measurement noises are captured by the measurement model. The states of the unknown nodes are augmented by regarding the expectation, the precision, the acoustic speed and the measurement noise as variables. The posterior probability distribution of the state variables of each unknown node can be optimized jointly using a variational filtering technique. The localized nodes with high confidence are then labeled as new reference nodes. We evaluate the impacts of the threshold of the measurements’ number and the number of the samples. The simulation results give the suggestions of selecting proper values for the two parameters. We also evaluate the localization coverage and accuracy of STVF and SLMP under different parameters, including the node density, the confidence threshold and the standard deviation of the measurements. The simulation results show that STVF reduces the localization error dramatically while maintaining approximate localization coverage with SLMP. Moreover, STVF is robust to the change of the parameter settings and works well even in sparse UASNs.