1. Introduction
With the increasing demand for ocean observation and exploration, UWSNs have attracted much attention recently. UWSNs are widely used in marine environmental monitoring, ecological protection, sonar navigation, resource utilization, and military surveillance [
1].
In UWSNs, the proper deployment of nodes is very important to the communication performance of the whole network, including network coverage, connectivity, energy consumption, and lifespan. Furthermore, the node deployment can be divided into static, restricted movement, and completely free movement. Static deployment means that nodes do not move and are fixed in predefined positions. The movement range of those restricted mobile nodes is small. A fully mobile node means that all ordinary nodes move. Due to the water flow, underwater nodes will move. Compared with fixed deployment, the mobile deployment of ordinary nodes may be more practical. Specifically, the restricted mobile nodes such as Argo buoy [
2] can be adjusted freely at a specific depth to meet the needs of ocean exploration.
Some scholars try to predict the ocean current information and adjust the nodes’ depth to make the node drift with the ocean current by using the ocean currents’ velocity and direction at different depths to reach the target sea area [
3]. Although theoretically feasible, the ideal effect has yet to be achieved in practical application due to the ocean currents’ complex changes and the lack of accurate space–time prediction tools. For UWSNs, which are based on limited mobility nodes, it is a greater challenge to solve the problems of network structure destruction, communication performance degradation, and life-shortening caused by node drift.
Since the marine environment is complex and difficult to predict, there are few studies on ocean current prediction and fewer reports on layered ocean current prediction. In this paper, a new UWSN is proposed, which combines the layered model of the ocean currents’ uneven depth, the layered ocean current prediction model based on CNN–transformer, the node trajectory prediction model, and the periodic depth adjustment model based on the SOA to achieve superior network coverage and connectivity. In this paper, the ocean current environment at different depths is analyzed. Based on the current velocity and direction of ocean current, the error threshold of current velocity and direction in the layer was introduced to guide the effective stratification process and ensure the change of ocean current in the same layer within an acceptable range to realize the stable control of nodes.
In recent years, with the wide application of machine learning algorithms in various fields, the real-time in-situ prediction of ocean current [
4] based on long-term and short-term memory networks and transformer structures [
5] has been proven more accurate. Generally, the existing research on ocean current stratification prediction mostly adopts the principle of depth average stratification [
6], ignoring the slowness of velocity and direction change after the depth of ocean current increases. In [
7], the error threshold of the current velocity and direction in the layer was introduced to guide the division of the sea area’s actual depth layer and the scientific stratification of the shallow sea area was realized.
The CNN–transformer hybrid network model is used to predict regional ocean current changes and combines the advantages of the CNN model and transformer model: a CNN captures the spatial details of ocean current data and the transformer handles the long-distance dependence of time series to ensure the accuracy of time–space series prediction. Although, the existing algorithms have obvious advantages for some kinds of data. This study proposes a combination of a CNN and transformer for ocean current data containing time and space series. It fully uses the CNN’s local perception and the transformer’s parallel processing ability to improve the accuracy and efficiency of processing big data ocean current models.
The layered prediction model of ocean current was combined with the underwater drift model of sensor nodes and the motion trajectory of nodes at different depths was predicted by sensor dynamics analysis and the iterative method. Then, the SOA was used to determine the optimal depth of sensor nodes and make periodic adjustments to optimize the network coverage and connectivity.
The SOA [
8] is a swarm intelligence optimization algorithm proposed by Gaurav Dhiman and Vijay Kumar in 2018. It has strong global search ability and low computational complexity. It has been widely used in various constraint problems and engineering optimization tasks and has shown good results in engineering practice.
The wide application of UWSNs aims to monitor large sea areas. UWSNs face multiple challenges in network deployment, such as mobility and three-dimensional deployment of nodes, high underwater acoustic delay, and communication delay caused by node deployment in specific monitoring areas. The deployment methods of nodes mainly include deep adaptation, static, and mobile assistance. In this study, the deployment goal was to ensure network coverage and connectivity.
The main contributions of this paper are as follows:
- (1)
A CNN–transformer hybrid network model is proposed to predict ocean currents that can accurately predict time–space series data.
- (2)
The trajectory of nodes in the dynamic ocean environment was predicted, combined with the ocean current layered prediction model. Then, the SOA determined the optimal adjustment depth of sensor nodes in the period to optimize the coverage and connectivity of the UWSN.
- (3)
Aiming for a UWSN with limited mobility, a new solution is proposed to deal with node drift, alleviating the problems of network structure destruction, communication performance damage, and life-shortening.
The structure of this paper is as follows:
Section 2 reviews the previous research and
Section 3 summarizes the new UWSN.
Section 4 details the ocean current uneven depth stratification, CNN–transformer hybrid network ocean current prediction, sensor node trajectory prediction, and SOA periodic depth adjustment model. In
Section 5, the simulation experiment is described, and
Section 6 summarizes the proposed method.
2. Related Work
This section reviews the previous research related to the node deployment of UWSNs. With the increasing demand for marine exploration, the development of UWSNs has attracted many researchers to devote themselves to underwater detection and monitoring. However, UWSNs face challenges such as small bandwidth, long propagation delay, limited energy, and high deployment costs, so node deployment and depth adjustment have become key issues in the research. These problems directly affect network performance, energy consumption, coverage, environmental adaptability, data collection efficiency, security, and privacy protection. In brief, node deployment and depth adjustment involve many research fields, such as topology control, communication protocol, energy management, adaptive algorithms, and machine learning.
Through the following research, the performance and adaptability of UWSNs were better optimized. In [
9], Chaudhary et al. summarized the technologies of node deployment and data collection challenges in UWSNs. In [
10], Gola et al. aimed to comprehensively survey UWSNs, including their application, deployment technology, and routing algorithm. In [
11], Jiang et al. proposed a distributed node deployment algorithm based on virtual forces to improve the coverage of UWSNs and reduce energy consumption. In a complex underwater environment, the three-dimensional deployment of nodes is challenging. In [
12], Su et al. proposed a deployment scheme based on the Voronoi diagram, suitable for marine surveillance, marine monitoring, deep-sea archaeology, and oil monitoring. In [
13], Yan et al. proposed an annual wheel uneven node depth adjustment self-deployment optimization algorithm to enhance the coverage and reliability of UWSNs and solve the energy hole problem. The current node deployment and position technology did not consider the nodes’ mobility and network disconnection. In [
14], Choudhary et al. aimed to monitor node mobility to predict node position and ensure coverage and connectivity. Researchers have shown great interest in establishing UWSNs with mobile nodes. Some optimized locating technologies involve higher energy consumption and costs. In [
15], Nain et al. studied a range-based node position scheme with a hybrid optimization of a UWSN. In [
16], Latif et al. designed a framework to jointly optimize node propagation delay, throughput, network lifetime, and energy consumption to solve the problems of limited energy and the node mobility’s high energy consumption. To sum up, UWSNs, which have certain mobility and can adapt to the complex underwater environment, have attracted extensive attention from researchers.
The restricted mobile nodes have certain mobility, and the depth can be adjusted to adapt to the underwater environment. The movement of nodes is mainly subject to the current velocity and direction. Therefore, it is necessary to study the marine environment in depth to establish an accurate ocean current prediction model. In ocean current prediction, many methods have been put forward to predict the movement of ocean currents.
Information on surface ocean currents can be obtained by observation or numerical simulation. However, it is usually difficult for observation data to reflect the current ocean current situation accurately. Specifically, the prediction models of marine physical parameters based on the Regional Ocean Model System [
17,
18,
19], Naval Coastal Ocean Model [
20,
21], and Global Real-time Ocean Forecasting System are not only difficult to calculate but also take a long time to predict ocean current information in real-time. Traditionally, the airborne Acoustic Doppler Current Profiler (ADCP) [
22] has been used for the spatial prediction of ocean current velocity vectors, providing instantaneous ocean current profiles of one or more lines at a certain time. In addition, Franchi et al. [
23] proposed a method combining Kalman filtering with underwater vehicle navigation sensors or acoustic locating that can estimate the spatial distribution of ocean current velocity in real-time. However, this method cannot predict the time evolution of ocean current velocity. Therefore, Pawlowicz et al. [
24] put forward that the time prediction of ocean currents should adopt the harmonic method, which uses the ocean current data obtained by the ADCP to pre-calculate a group of harmonic components at a certain number of offshore positions. However, this prediction method relies on a specific ocean area with known ocean current data and cannot be updated in real-time. In recent years, machine learning techniques such as the Gaussian process [
25] and support vector machines [
26] have been used to predict ocean current information to improve prediction accuracy. In addition, prediction tools based on neural network technology have been proposed, such as artificial neural networks with a full connection layer or a recursive Long Short-Term Memory (LSTM) layer [
27]. However, these forecasting tools usually only train and test the data of the same area but not the data of other sea areas, so predicting other areas is not universal. Hence, this paper proposes an ocean current prediction scheme consisting of CNN and transformer models that can realize the real-time and accurate in-situ prediction of ocean currents without data training in specific areas.
Although the LSTM-transformer model shows good ability in the real-time and accurate in-situ prediction of ocean currents, ocean currents’ layered prediction at different depths has yet to be considered. Generally, most studies adopt the depth average stratification strategy, but this strategy can only partially meet the needs of shallow water and deep sea areas. Specifically, setting the depth layer too high will reduce the prediction accuracy of shallow water areas, while setting it too low will squander the calculation resources of deep water areas.
Therefore, this study introduces the error threshold of ocean current velocity and direction to guide the division of depth layers in actual sea areas. This method aims to divide more layers with smaller heights in shallow water areas while dividing fewer layers with larger heights in deep water areas.
This section mainly reviews the related research of UWSN node deployment and ocean current prediction. Traditional UWSN architecture includes static and mobile nodes. Static nodes are usually placed on the seabed and connected to the receiver unit. Nodes can move freely in the mobile auxiliary network, forming a dynamic topology. Mobile nodes usually need two transceivers to enhance their data collection capabilities. To sum up, the research of node deployment is devoted to improving the packet transmission rate to minimize the number of nodes and energy consumption and achieve maximum coverage and high network connectivity.
3. System Model
A plethora of research has been dedicated to static deployment and deployment strategies during the initial phase of network formation in UWSNs [
28]. We focus on the performance variations in a UWSN throughout its operational lifecycle. Within UWSNs, the drift velocities of ocean currents vary with depth, leading to disparate drift distances for nodes at different depths. This variation consequently impacts the network topology and overall performance. Therefore, a node depth adjustment scheme based on node trajectory prediction and a swarm intelligence optimization algorithm are introduced, which encompass a stratified ocean current model, a node trajectory prediction model, and a node depth adjustment model, all aimed at enhancing network coverage and connectivity. The UWSN architecture primarily consists of ordinary nodes and sink nodes. Ordinary nodes constituted by Argo floats are utilized to collect environmental data [
2]. In contrast, relay sink nodes are pivotal for information transmission within the UWSN [
29], as depicted in
Figure 1.
The layered ocean current model includes constructing the layered structure of ocean current and predicting ocean current parameters at different depths. Firstly, the ocean current properties at different depths are analyzed to determine the stratification standard. After that, the stratified ocean current data are used in the prediction model to form a stratified ocean current prediction model. The impact of ocean currents at different depths on node drift varies. Therefore, the trajectories of nodes at different layers can be derived by integrating ocean current stratification prediction data with kinematic equations. Finally, the optimal depth adjustment of nodes in the cycle is obtained by the SOA.
At the initial stage of the system, the original ocean current data are analyzed to determine the criteria for stratification. Then, based on these criteria, a deep stratification operation is carried out, and the layered ocean current model is established. The model covers the characteristics of ocean currents at different depth levels, including the rapid change of the surface layer and the relative stability of the deep layer.
Subsequently, the system adopts the ocean current prediction model constructed by the CNN–transformer hybrid neural network, which takes the data of the layered ocean current model as input to generate detailed and accurate ocean current prediction. The prediction model fully uses CNN’s advantages in extracting spatial features and combines the transformer’s global perception of long-time series data, which can effectively predict ocean currents’ spatial distribution and temporal changes.
Then, the system uses the node trajectory prediction model and ocean current prediction results to calculate the node trajectory in the future through the dynamic analysis of underwater nodes and the numerical incremental iteration method. This trajectory information is very important to guide the depth adjustment of nodes.
Finally, the system uses the SOA to intelligently adjust the node depth. This algorithm aims to improve network coverage and connectivity, simulates the seagull’s predation behavior strategy, and finds the best node depth adjustment scheme for ordinary nodes and sink nodes.
To sum up, the NAS-MP provides a complete solution, aiming to optimize the performance of the UWSN under dynamic ocean current change. This system integrates an accurate ocean current layered model, advanced ocean current prediction technology, node trajectory analysis, and intelligent optimization strategy to ensure the stable and efficient operation of a UWSN in complex marine environments.
The following key assumptions were put forward to ensure the feasibility and practicability of the model construction:
- (1)
All nodes in the network were isomorphic and had the same communication radius and perception radius.
- (2)
In the initial stage of the network, all nodes were randomly distributed in a square detection area with a side length of L.
- (3)
Since the current velocity in the vertical direction was small, the influence of the current velocity in the vertical direction on the network was ignored.
5. Simulation Results
5.1. Ocean Current Data Source
The ocean current data used in this study were obtained from the Copernicus Ocean Data Center [
35], and data from 1 January 2021 to 1 January 2024 were obtained. In order to meet the needs of the UWSN, ocean current data sets between 115° E and 119° E and 15.5° N and 19.5° N were downloaded. As shown in
Figure 9, these data were used as historical data to establish an ocean current prediction model and provide support for simulating the sea area of UWSN nodes.
Historical ocean current data were preprocessed and spatially interpolated and an ocean current data model covering part of the South China Sea was established. This part of the data was used to train the ocean current forecasting model and verify the feasibility of the node adjustment scheme.
5.2. CNN–Transformer Neural Network Structure and Model Training
The network structure of the ocean current prediction model included six layers: input layer, convolutional neural network layer, transformer layer, dense connection layer, discard layer, and output layer.
Table 1 summarizes the model’s frame structure, including the name of each level and its output shape.
The input layer data came from the ocean current data model, divided into an 80% training and a 20% verification set. The data were input in tensor form with the size of (None, 80, 256), where None represents the batch size, 80 represents the input window size of the time series, and 256 represents the number of features in the data. The data were first processed by the CNN layer, which used two convolution kernels with different sizes and the ReLU activation function to extract features from the input data automatically. Subsequently, the feature dimension was raised to the hidden layer’s size through the dense connection layer to increase the model’s feature extraction ability. Before entering the transformer layer, in order to avoid over-fitting, 20% of data points were discarded. The transformer layer consisted of two encoder layers and two decoder layers, and the multi-head self-attention mechanism was set to 4, aiming to extract more feature information from the time series of ocean current data. After the transformer layer processing, the characteristic length of the data became the size of the prediction window, that is, 40, which represented the size of the prediction window of the time series. Finally, the feature dimension of the data was mapped back to the original feature dimension size through the output layer. Hence, it could obtain the final prediction result.
In the training process, the Adam optimizer was adopted, and the loss function was defined as the mean square error (MSE) between the predicted label (ocean current velocity with dimension information of time and space) and the real label. This index could effectively evaluate the degree to which the model fit the real data. The Adam optimizer adopted an adaptive learning rate adjustment strategy, which had a large initial learning rate and helped accelerate the convergence process. With the training, the learning rate gradually decreased, which helped the model approach the local optimal solution stably.
Figure 10 shows the change in loss after 20 trainings. As the number of training iterations increased, both the training loss and the verification loss gradually decreased, indicating that the model’s performance continued to improve. At the 18th epoch, the model reached the minimum loss. However, too many training cycles could lead to over-fitting the model, increasing the loss.
5.3. Using SOA to Optimize and Adjust the UWSN
A sea area of 600 m × 600 m × 600 m was randomly selected from ocean current data as the activity and detection area of the UWSN. The initial positions of all ordinary nodes were evenly distributed in this area, while the bases of the four sink nodes were evenly placed at the bottom of the water. The depth layer where the sink nodes were located could be adjusted adaptively according to the needs of the network structure. Please refer to
Table 2 for the specific parameter settings of the model.
A 10-day simulation operation of the UWSN was conducted in the sea area. At the beginning of each cycle, the depth of the node was adjusted. Specifically, the predicted trajectory of nodes was analyzed to determine the best node adjustment strategy in each cycle, and then the network coverage and connectivity changes were calculated. In addition, the average value of these two metrics within the time frame was adopted as the optimization objective. A tent chaotic map was utilized to generate the initial solution space and compute the initial fitness. In the iterative process, the SOA algorithm generated a series of new non-dominated solutions. Then, the non-dominated quick sorting method based on the elite strategy was used to sort the old and new solutions. After completing the iteration, a boundary composed of dominant solutions could be obtained. According to the preset optimization weights, the optimal solution could be selected at the beginning of the cycle to determine the optimal adjustment strategy.
In this study, the network nodes were constructed using real data run in the sea area. According to the proposed optimal adjustment strategy, the adjustment scheme of nodes at the beginning of each cycle depended on the influence of predicted future ocean current changes on node movement. Therefore, the accuracy of the adjustment scheme and the actual effect of network optimization were affected by the accuracy of ocean current prediction. Real ocean current data were used instead of forecast data to verify the feasibility of the periodic node depth adjustment scheme to improve network performance.
Figure 11 shows the performance changes of the UWSN when the optimal weight ratio of coverage and connectivity was set to 5:5. With the operation of the network, the influence of ocean currents on sensor nodes led to the continuous drift of nodes, gradually reducing the network coverage. Implementing the optimized depth scheme reduced the extent of this decline, as shown in
Figure 11a. After introducing the optimization adjustment strategy of periodic depth adjustment, the node distribution was obviously improved, and the network’s average coverage was significantly improved compared with the unoptimized state. However, the performance fluctuation in individual periods was because the current periodic depth adjustment scheme only considered the network performance change before the next node adjustment and did not consider the influence of subsequent periods. Due to the spatial difference of the real-time ocean current environment, the network performance changes were inconsistent before and after optimization.
In addition, the overall movement trend of sensor nodes with only the depth adjustment function used for sea area detection tasks was consistent with the current ocean environment, which was also the main reason for the network performance fluctuation. As far as network connectivity is concerned, as shown in
Figure 11b, the network connectivity obviously improved after the introduction of the optimization adjustment strategy. In addition, the optimized network connectivity rate could always be maintained at a high level because the sink nodes could adjust their depth layer according to the distribution of nodes. Furthermore, the coverage and connectivity before and after network optimization were multiplied by 0.5 and then they were added together to comprehensively evaluate the change in network performance. As shown in
Figure 11c, the comprehensive performance of the optimized network was significantly improved, and the performance improvement effect was more significant with the increase in network running time.
During the periodic operation of the sensor network, ocean currents continuously affected the network’s coverage and connectivity. Since the node communication employed a multi-hop transmission method, it possessed good interference resistance against ocean currents. Therefore, by decreasing the connectivity optimization weight from 0.5 to 0.1 and increasing the coverage optimization weight from 0.5 to 0.9, we periodically operated the network to collect optimization data.
Figure 12 shows the influence of the optimization weight ratio of coverage and connectivity on the comprehensive performance. Under different weight ratios, the initial deployment of nodes was consistent. Because of the different settings of the weight ratio, the optimal solution selected after each iteration was different, which led to a difference in network performance. Although the network performance changes were different under different settings, the overall trend significantly improved compared with the unoptimized network. The proposed optimization scheme was highly versatile and could improve performance under different network requirements.
5.4. Comprehensive Performance Analysis of the Model
To verify the feasibility of the periodic node depth adjustment scheme for network performance optimization, real ocean current data were used instead of predicted ocean current data to eliminate the influence of deviation between predicted data and real data on network performance. Firstly, the predicted ocean current data were used to predict the periodic movement trajectories of the nodes. Then, based on these trajectories, the optimization algorithm generated a periodic depth adjustment scheme. Finally, the nodes were placed in the sea area and constructed with real ocean current data. The depth adjustment scheme generated from the predicted data was executed. This process allowed us to observe changes in network performance and simulate a more realistic application scenario.
Figure 13 shows the influence of the adjustment scheme generated by using the predicted data on the network performance when the optimization weight ratio of coverage and connectivity was set to 5:5. The results show that after optimization and adjustment, the network’s performance was significantly improved compared with that without adjustment. However, compared with
Figure 11, the optimized network’s performance was slightly reduced due to some errors between the predicted ocean current data and the actual ocean current.
In addition,
Figure 14 shows the influence of adjusting the optimized weight ratio on the comprehensive performance. Compared with
Figure 12, although the network performance after optimization was slightly reduced, compared with the situation without optimization and adjustment, the network performance after adding nodes was significantly improved. To sum up, our periodic node depth adjustment scheme could effectively weaken the influence of ocean currents on the UWSN, thus significantly improving the network performance in long-term operation.
5.5. Comparative Analysis of Node Distribution
We investigated the adjustment strategies for common nodes affected by ocean currents, focusing on dynamic node adjustment during the network’s operational process in response to ocean current movement.
Figure 15 presents a comparative analysis of the node distribution status within a UWSN subjected to dynamic ocean currents, contrasting the scenarios with and without applying the proposed adjustment scheme.
Figure 15a illustrates that the UWSN without depth adjustment exhibited the characteristics of a smaller coverage area and uneven distribution, manifesting as a smaller coverage area with regions of excessive node aggregation juxtaposed against areas of sparse deployment, which resulted in the formation of detection voids in the deployment space because the upper ocean currents were faster, while the deeper currents were more moderate in speed. On the other hand, the adjusted UWSN demonstrated a larger coverage area and a more uniform distribution of nodes compared to its unadjusted counterpart, as depicted in
Figure 15b. Attributed to our node depth adjustment scheme, based on ocean current prediction, it could enhance the network’s coverage by adjusting node depths while ensuring connectivity.