Maritime Opportunistic Network Routing Strategies for Assessing Link Connectivity Based on Deep Learning

Abstract

In opportunistic networks, where link performance is often highly variable or extreme due to the intermittent nature of communication links between nodes, there may never be a continuous and complete path between the receiver and the sender, and packets of information can only be stored and carried by the movement of nodes, which then look for forwarding opportunities when they meet. Existing routing protocols for opportunistic networks suffer from problems such as excessive network memory consumption or insufficient link prediction that is focused on link connectivity determination. In this paper, we propose an efficient opportunistic network routing protocol that evaluates the historical values of encounter probability, movement posture, and acquired resource availability of all nodes within the communicable range based on link prediction. The intelligent prediction of link connectivity state provides a reliable aid for routing decisions, which can provide longer-period communication in the ocean; the consideration of nodes’ comprehensive attributes establishes the priority of message forwarding, avoids duplicate transmissions and route invalidation phenomena, and effectively improves the success rate of message delivery. It also reduces the transmission latency and routing overhead compared to the existing schemes.

1. Introduction

An opportunistic network mainly consists of nodes with limited energy; such nodes have various kinds of network cards such as WiFi, Bluetooth, 4G, 5G, satellite, etc., which come with functions such as data collection, computation, and transmission [1]. Any facility that carries a communication network can be a node, and, in maritime opportunistic networks, communication relies mainly on ships or fixed base stations at sea, such as buoys and platforms, which become nodes, as previously described. In opportunistic networks, the transmission of messages from the source node to the destination node also requires the assistance of relay nodes to store and carry the messages [2]; the nodes utilize the multi-hop mode to transmit all the data to the relay node, which fuses all the data to the network base station and, ultimately, sends the data required by the user to the user [3]. Due to the intermittent communication between nodes, the resources around the relay nodes are dynamic, changing, and uncertain [4]. Also, the maritime environment is different from the land, the meteorological conditions are complex and changeable, and the network users are highly mobile and sparsely distributed. The design of the maritime opportunistic network routing should contain a reasonable selection of communication resources to carry out an effective evaluation of the relay nodes [5] and to determine the forwarding priority of all the nodes in the candidate relay set, to avoid or inhibit the duplication of transmission [6].

In the current research that has bee nproposed for solving routing schemes for opportunistic networks, most works are based on replicating data copies for wide dissemination in order to improve the packet transmission performance [7], but this can lead to the wastage of network resources or even network congestion. There are also schemes that utilize link prediction to transform link availability between pairs of nodes into a probabilistic representation as a metric for routing decision making [8]. But the link prediction in which the sequence learner employed is concerned with learning the correlation in time series, while, in opportunistic networks with intermittent communication, more attention should be paid to link connectivity to cope with the highly variable link performance and improve the communication period. In [9], an opportunistic network routing scheme based on network connectivity assessment is proposed, where a breadth-first algorithm constructs a connectivity graph between a source node and a target node to determine the two nodes’ connectivity. But this algorithm has high space complexity, consumes a large amount of memory, and is not suitable for dealing with large-scale data. In a weighted graph, there is no guarantee that the path with the smallest weight can be found.

This research focuses on addressing the above limitations. The proposed deep learning-based link connectivity assessment for opportunistic network routing strategies firstly performs link prediction for link connectivity between pairs of nodes, takes into account the further forwarding state of the nodes in the future moments, and, on this basis, determines the forwarding priority of the nodes based on their comprehensive attributes in the set of candidate relays, then designs the routing scheme. The 6G networks will incorporate AI-driven dynamic spectrum access and spectrum agility mechanisms, where links adapt dynamically based on real-time spectrum conditions, interference levels, and coexistence requirements [10]. The link connectivity prediction model in this paper could be extended to predict spectrum availability, helping 6G networks to optimize routing based on real-time spectrum usage. Hence, deep learning-based link prediction will be crucial in 6G’s ultra-dynamic, software-defined network environments. The following are the major contributions of this work:

• Intelligent prediction of link state can effectively determine the risk of link breakage, thus assisting routing decisions and improving communication efficiency [11];
• In link prediction, more attention is paid to the link connectivity status in the ship characterization dataset (a ship characterization dataset consists of the communication data from multiple ships at a given moment in time) rather than the time series, in order to cope with the highly variable nature of the links in the maritime opportunistic network;
• Determining the forwarding priority of a node based on its attributes can avoid the repeated transmission of messages and reduce the waste of network resources. This can also avoid the situation in which the current node has no available neighboring nodes to forward packets to the destination within the shortest path, and can only be broadcasted all the time by certain nodes with high-priority messages, resulting in invalid routes, which provides a longer period of communication in the ocean;
• The special feature of opportunistic networks is the random movement of nodes, and, in this paper, we take into account the movement posture of nodes when evaluating them. This leads to a higher probability of forwarding a message to a node that is more convergent in its motion posture or to other nodes that are closer to the target node.

This paper is divided into six major sections. Section 2 introduces the main attribute factors of the evaluated nodes, including the node’s historical value encounter probability, the node’s posterior resource availability, and the node’s movement posture; Section 3 proposes the link prediction-based routing algorithm process in this paper; Section 4 specifically describes the link prediction algorithm based on Convolutional Neural Networks (CNNs) and an Efficient Additive Attention (EAA) mechanism, in which the ship communication data are first extracted from the attribute features by the CNN, and then processed using the EAA to process the feature representations and obtain the prediction results of link connectivity; Section 5 compares the algorithm in this paper with the replication-based routing SprayAndWait and node attribute-based routing RSNA in terms of network performance with different buffer sizes. This paper’s algorithm effectively improves the message delivery rate and reduces resource consumption and network overhead.

2. Comprehensive Attribute Evaluation of Nodes

2.1. Encounter Probability of Historical Values of Nodes

The biggest challenge facing marine communication networks is that nodes are in a mobile state and the link quality between nodes is constantly changing. Good link quality can improve the success rate of information reception: the higher the reception success rate, the more reliable the communication. Neighboring nodes form a link between them, and the link quality is inextricably linked to the interaction between pairs of nodes. From the perspective of data delivery, the more active a node is in the network, the higher the quality of its information carrier because it has a higher probability of encountering other nodes in the network, including the destination node [12]. In this paper, we use the node’s history value encounter probability to describe the interaction between pairs of nodes, as shown in Equation (1).

$$p_{l,r}={\displaystyle \frac{e_{l,r}}{{\sum }_{c=1,c\ne l}^n{e}_{l,c}}},$$

(1)

where $e_{l,r}$ denotes the number of direct encounters between nodes l and r, and $e_{l,c}$ (1 ≤ c ≤ n) denotes the total number of encounters between all pairs of nodes in the period T.

The ability of nodes to forward packets can be effectively measured using $p_{l,r}$ to improve link quality.

2.2. Acquired Resource Availability of Nodes

In opportunistic networks, a node’s resources are fixed and non-renewable. In routing, if the node currently carrying a message cannot reach the destination node within the communication range and there is no available neighboring node to relay it, then certain nodes with high priority will frequently perform data transmission, resulting in the early exhaustion of resources, which reduces the lifetime of the network and leads to the phenomenon of invalid routing [12]. Alternatively, when a node has insufficient available resources, it will not be able to continue to relay the message, and the source message that it processed will also be lost. Therefore, resource metrics should be involved as a key factor in the forwarding decision: the richer the node’s acquired resource availability, the higher its participation in delivering messages [5]. In this paper, the acquired resource availability of a node is defined as the device capacity (DC), which is expressed as a percentage of the node’s available buffer, as shown in Equation (2).

$$\mathrm{DC}={\displaystyle \frac{B_r\left(t\right)}{B_{init}}},$$

(2)

where $B_r$ is the remaining buffer size of the node and $B_{init}$ is the initial buffer size of the node.

Utilizing DC can effectively evaluate the node’s ability to forward messages, use the node’s limited resources more rationally, and reduce the loss rate of message transmission.

2.3. Kinematic Posture of Nodes

In Section 2.2, the forwarding ability of the node as the receiving party of the message is evaluated in the process of transmitting the data, whether it has enough device capacity for forwarding the message or not. In Section 2.3, the degree of willingness of the node to forward when it is the sending party of the message will be evaluated, and the node is more willing to use its resources to forward the message to the node that is related to it [13]. The special feature of the opportunistic network is the random movement of nodes, due to which the message cannot be directly transmitted to the destination node. The relay node is utilized to store and forward the message, and the message can be transmitted to the destination node hop by hop through the relay node. During the node’s movement, it is more likely to forward the message to the nodes that are more convergent with it in terms of its movement posture or other nodes that are closer to the destination node in terms of their movement posture, and the degree of willingness to forward the message to these nodes is also stronger. In this paper, we use the dot product to calculate the direction similarity (DS) on the node movement posture, as shown in Equation (3).

$$\mathrm{DS}=\left|\right|d_i\left|\right|\left|\right|d_j\left|\right|cos\left(\theta \right),$$

(3)

where $d_i$ and $d_j$ denote the direction vector of the node carrying the message and the direction vector of the node in the candidate relay set, respectively, and $\theta$ denotes the angle between the motion directions of the two nodes, $\theta$$ϵ$(0,$\pi$).

When $\theta$ is smaller, DS is larger, and nodes that are judged to be more similar in terms of movement posture have a higher degree of willingness for nodes to forward messages to them.

3. Routing Algorithms

In this paper, we propose an opportunistic network routing strategy based on link connectivity assessment in the following steps:

Step 1. Network nodes are initialized and the packet is sent to all available nodes in the neighbor list of the source node $N_l$. If the neighbor list contains the destination node $N_d$, then the message is directly transmitted to $N_d$. Otherwise, it will enter the relay node selection process;

Step 2. Firstly, determine the connectivity status of the relay node $N_r$ and its neighbor $N_{rf}$ according to the link prediction method, which is introduced in Section 4. If the predicted connectivity is 1, then there exists an encounter between the two. Then, from the historical encounter probability of the node $p_{l,r}$, the link availability of the node pair can be defined as shown in Equation (4).

$$p_{la}=\left\{\begin{array}{c}{p}_{lr},\mathrm{Link}\mathrm{availability}\mathrm{of}\mathrm{node}{N}_r\mathrm{and}\mathrm{node}{N}_l\hfill \\ 0,\mathrm{else}\hfill \end{array}\right.,$$

(4)

If the link availability $p_{la}$ > 0 for relay node $N_r$ and its neighbor $N_{rf}$ means that $N_r$ has next hop, this also identifies the nodes in the candidate relay set that do not have more relay nodes to the destination node and narrows the candidate relay set;

Step 3. To further narrow down the candidate relay set, calculate DC($N_r$, $N_{rf}$) and filter out the dead nodes with insufficient acquired resources;

Step 4. Use the weighting function U($N_r$, $N_{rf}$) to determine the priority in the remaining candidate relay set, as shown in Equation (5):

$$\mathrm{U}(N_r,N_{rf})=\alpha p_{la}+\beta DC+\gamma DS,$$

(5)

where $\alpha$ + $\beta$ + $\gamma$ = 1 [14]. In order for all three node attributes mentioned in Section 3 to be involved in routing decisions, it is better to take $\alpha \approx \beta \approx \gamma$. Based on DC($N_r$, $N_{rf}$) combined with the node’s motion posture, DS($N_r$, $N_{rf}$) can further identify the nodes that seem to be qualified, but there are no more subsequently available candidates to relay the data grouping to avoid data loss;

Step 5. Forward the message to the relay node Nr with higher priority based on the weighting function. The message is delivered by $N_r$ and goes back to the judgment in Step 1 if the neighbor list of $N_r$ does not contain the destination node $N_d$; it enters a new round of relay node selection. Finally, the message is passed hop by hop until it reaches the destination node.

To deal with the possible selfish behavior of the nodes, a reputation score is set for each node in the routing decision: the node will be rewarded with credit for forwarding the packet, and the credit score will be reduced if the node appears to refuse to forward the packet or prioritize its own transmission behavior. Only nodes with higher reputations are prioritized for data forwarding.

4. Intelligent Prediction of Link States

To improve the reliability of routing decisions, this paper proposes a link intelligent prediction scheme (CNN–EAA) based on the combination of a Convolutional Neural Network and attention mechanism for judging the link connectivity status to assist subsequent routing decisions. The algorithm is useful for opportunistic networks where nodes have high mobility to extract valid information about the nodes and focus on learning the connectivity status of the links. Applying this link prediction algorithm before routing decisions can effectively improve the communication efficiency and communication period of opportunistic networks.

4.1. Effective Additive Attention (EAA)

The EAA consists of four hierarchical stages of different scales $\{{\displaystyle \frac{1}{4}},{\displaystyle \frac{1}{8}},{\displaystyle \frac{1}{16}},{\displaystyle \frac{1}{32}}\}$, each of which is consistent and composed of a Convolutional Encoder and a SwiftFormer encoder [15], with a downsampling layer introduced between two consecutive phases to halve the spatial size and increase the feature latitude. The structure of EAA is shown in Figure 1. The Convolutional Encoder consists of a 3 × 3 depth convolution [16] and two point-state convolutions [17]. The initial module of the SwiftFormer encoder consists of a 3 × 3 depth convolution followed by a point-state convolution [16], followed by an efficient additive attention module. Finally, the output features are fed into a linear module that consists of two 1 × 1 point-state convolution layers, batch normalization, and GeLU activation [15].

In SwiftFormer’ s attention module, the input embedding matrix x is transformed into a query (Q) and a key (K) via two matrices $W_q$, $W_k$, where Q, $K\in R^{n\times d}$, $W_q$, $W_k\in R^{d\times d}$, n is the length of tokens, and d is the dimension of embedding vectors. Next, the query matrix Q is multiplied with the learnable parameter vector $w_a\in R^d$ to learn the attention weight of the query and generate the global attention query vector $\alpha \in R^n$, as in Equation (6):

$$\alpha =Q·w_{\alpha }/\sqrt{d},$$

(6)

The query matrix is then pooled based on the learned attention weights, resulting in a single global query vector $q\in R^d$, as follows in Equation (7):

$$q=\sum _{i=1}^n{\alpha }_i·Q_i,$$

(7)

Next, the interaction between the global query vector $q\in R^d$ and the key matrix $K\in R^{n\times d}$ is encoded to form a global context via element-level multiplication ($R^{n\times d}$). A linear transformation layer for query–key interactions is employed to learn the hidden representation of the tokens. The output of the attention module $\widehat{x}$ can be described using Equation (8):

$$\widehat{x}=\widehat{Q}+T(K·q),$$

(8)

where $\widehat{Q}$ denotes the normalized query matrix and T denotes the linear transformation [15].

4.2. Convolutional Neural Networks (CNN)s

The CNN consists of two Conv1D (one-dimensional convolutional layers) combined with a fully connected layer and an attention mechanism. It merges a MaxPooling layer and a Rectified Linear Unit (ReLU) layer between two consecutive convolutional modules to fully capture spatial dependencies [8]. A Conv1D operation with input channel 1, output channel 16, convolution kernel length 3, and step size 1 is applied to the sample set in the first convolutional layer to filter the noise and extract the valid information. Figure 2 shows an example of a simple one-dimensional convolution operation. Denote $y_t^{(i,0)}$ as the input matrix of the 0^th convolutional layer in the $i^{th}$ region matrix at time t. Then, the $k^{th}$ network layer is represented as shown in Equation (9):

$$y_t^{(i,k)}=\sum _{iϵx_t}f\left(y_t^{(i,k-1)}\right)\otimes W_{i,t}^k+b_t^k,$$

(9)

where $x_t$ denotes a set of input data, $W_{i,t}^k$ is the convolution kernel, f is the nonlinear (rectified linear unit) activation function, ⊗ is the elemental product, and $b_t^k$ and $y_t^{(i,k)}$ denote the deviation of the convolution module k and the output value, respectively [18].

The convolutional layer is followed by a nonlinear activation function ReLU to improve the model’s ability to learn complex structures, as shown in Equation (10).

$$f\left(x\right)=max(0,x),$$

(10)

A MaxPooling layer is added to the CNN model, which is a downsampling method that reduces the spatial dimensionality value of the feature attribute map by half and retains the most important features [19]. The pooling operation is shown in Equation (11), and the input data feature set is partitioned into connected 1 × 2 modules by means of the window free-sliding. An activation function accelerates the convergence of the model.

$$y_t^{(l,k)}=subsmapling\left(y_{t-1}^{(i,k)}\right),$$

(11)

The second layer of convolution is similar to the first layer, by applying a Conv1D convolution operation with 16 input channels, 32 output channels, convolution kernel length of 3, and step size of 1. This layer increases the complexity of the features and further reduces the data dimensions.

4.3. CNN–EAA

Figure 3 depicts the basic process of the CNN–EAA algorithm. The maritime vessel feature dataset with multi-attribute factors is used as the input to the CNN convolutional layer. Since there is no public dataset for maritime communication, this paper obtains it by constructing a maritime ship communication model, which has been introduced in Ref. [20]. Due to the limitation of the length of the article,

Table 1. Selected ship datasets.

ShipID	${\mathit{lon}}_{\mathit{t}}$	${\mathit{lat}}_{\mathit{t}}$	${\mathit{v}}_{\mathit{t}}$ (km/h)	${\mathit{\theta }}_{\mathit{t}}$	SNR (dB)	Link_Connectivity
1	113.1691	−62.0405	236.0666	14.12	13.16	0
2	146.0954	84.7322	12.8562	0.63	11.44	0
3	−134.2308	81.7312	305.6865	5.53	22.96	1
4	148.8293	−2.6385	336.2375	0.92	23.85	1
5	47.6353	54.0007	244.3446	1.94	5.60	1
6	−144.8721	−64.0963	272.7864	16.47	14.69	1
7	−79.7506	−14.3215	267.5276	13.89	13.36	1
8	16.7949	75.0855	141.2017	6.34	19.39	0
9	164.5252	52.1651	235.9720	19.01	21.28	1
10	167.3653	82.7882	61.6272	0.68	22.64	0

lists the communication data of the first group of ships generated in the first second; 80% of the collected sample data are used for the training set and 20% for the test set.

Before entering the convolutional layer, the input tensor dimensions are extended by the unsqueeze layer to adapt to the input format of the one-dimensional convolutional layer, and sample features can be extracted by using the CNN convolutional layer. The pooling module reduces the amount of data in the hidden intermediate layer without losing the scale-invariance of the features, and, at the same time, at each layer, the convolution module and pooling module utilize activation functions to accelerate model convergence. After the convolution is complete, the output of the convolutional layer is spread to a 2D tensor for input to the first fully connected layer.

If the best performance may not be achieved by simply applying convolutional operations to the information, the EAA (Efficient Additive Attention) mechanism [15] is used to weigh the input features so that the model can pay more attention to the important features and finally complete the classification task. Therefore, an unsqueeze layer is added between the CNN and EAA to map the features extracted by the CNN to a higher-dimensional space to fit the input format of the attention layer. The features adjusted after weighting by the EAA of the attention mechanism are then passed through the squeeze layer to recover the dimensionality in order to input to the second fully connected layer for classification.

To prevent the overfitting phenomenon, a dropout layer is also added between the CNN and EAA, which can effectively reduce the overfitting problem in deep learning training models [21].

The training of the CNN–EAA model on ship feature attributes can lead to the prediction of link connectivity, which assists in providing longer-period communication in the ocean and improving the efficiency of information transmission.

5. Simulation Results and Performance Analysis

To verify the rationality and efficiency of the proposed method, this section compares the network performance of routing in this paper with replication-based routing SprayAndWait [22] and node attribute-based routing RSNA [12] with different buffer sizes, analyzing them specifically.

5.1. Simulation Environment Parameters

In this paper, simulation experiments are carried out in the MATLAB 2018a software environment. Using a random walk model in an opportunistic network, the nodes’ corresponding geographic locations at different times are combined to represent the node’s movement trajectory, and each node in the network can be used as a source node, relay node, and destination node [23]. A distance threshold is used to define whether data exchange occurs when two nodes are in contact, and, to ensure the comparability of the experiments, the same packet caching and deletion strategies are used and the packet generation rate is adjusted based on the network load. In addition, it is assumed that all nodes are not selfish or malicious and that they can freely collaborate with each other to exchange information. The main parameter settings of the simulation scenario are shown in

Table 2. Simulation parameters.

Parameters	Values	Description
Node Deployment Area	100 m × 100 m	Square monitoring area
Transmission Speed	200 kbps	The speed at which a node transmits a message
Message Interval	10 s	Packet delivery cycle
Maximum Message Survival Time	1000 s	Packet survival time limit to prevent packets from looping indefinitely in the network
Nodal Velocity	0∼20 km/h	Random speed range

.

In the simulation experiments, packet delivery rate, average transmission delay, and routing overhead are used to evaluate the network performance of different routing protocols, which are defined in detail as follows:

• Packet Delivery Ratio: the ratio between the total number of packets delivered to the destination and the total number of packets sent from the source node, reflecting the impact of routing on network performance;
• Average Transmission Delay: the average value of the time taken by a packet to reach the destination node from the source node, which is affected by the frequency of encounters between nodes, packet storage waiting time, etc., and reflects the timeliness of the routing policy in delivering messages [24];
• Routing Overhead: Routing overhead is usually related to the number of packet forwards and network bandwidth consumption, and is used to measure the resource usage efficiency and scalability of the routing policy.

The packet delivery rate, routing overhead, and average delay calculations are mainly dependent on the contact between nodes, the exchange of packets, and the store and forward process.

5.2. Comparative Analysis of Routing Performance with Different Buffers

The routing overhead of the three algorithms with different buffers is shown in Figure 4. As the buffer increases, each node may store more packets and participate in more transmission and forwarding processes, so the routing overhead of all three algorithms gradually increases. With different buffers, SprayAndWait has a large number of message copies in the network, which generates a large routing overhead and consumes a lot of network resources. RSNA decreases the routing overhead compared to SprayAndWait, but there is still some transmission of useless information. The routing strategy proposed in this paper reduces the transmission of useless messages and greatly reduces the routing cost, which is less than SprayAndWait and RSNA under different buffers.

The average transmission delay under different buffers of the three algorithms is shown in Figure 5. In SprayAndWait, the node carrying the message needs to wait for the relay forwarding from the auxiliary node, which increases the message transmission delay. The routing strategy in this paper determines the forwarding priority of all the candidate neighbor nodes more reasonably than RSNA, which effectively reduces the transmission delay, and the average transmission delay under different buffers is less than both SprayAndWait and RSNA.

The packet delivery rates of the three algorithms under different buffers are shown in Figure 6. SprayAndWait does not take into account the prediction and motion posture of relay nodes, and the packet delivery rate is relatively low in general. The further prediction of relay nodes by RSNA improves the packet delivery rate to a certain extent compared to SprayAndWait. The routing strategy proposed in this paper improves the further prediction conditions for relay nodes based on RSNA, effectively reduces the number of message replicas, and makes the packet delivery rate higher than both SprayAndWait and RSNA under different buffers.

From the above experimental results, it can be concluded that the routing strategy based on deep learning to evaluate link connectivity proposed in this paper improves the network performance under different buffer sizes, effectively increases the message delivery rate, and reduces the resource consumption and network overhead compared to SprayAndWait and RSNA. Thus, it is an ideal choice for transmitting data in the dynamically changing maritime opportunistic network.

6. Conclusions

In this paper, we propose a routing strategy for assessing link connectivity based on deep learning. Firstly, the CNN–EAA model is used to judge the link connectivity state: CNN can extract the effective features of the ship nodes, and EAA makes the model pay more attention to the important features and focuses on learning the connectivity prediction result of the link. Finally, the intelligent prediction of the link state is combined with the comprehensive attributes of the nodes, taking into account the node’s historical value of the probability of encounter, the availability of resources in the latter day and the movement situation, and the design of the routing scheme, which effectively improves the message delivery rate and reduces the resource consumption and network overhead. In this paper, we design a routing scheme based on link prediction, which can better adapt to the maritime opportunistic network environment where the link state is highly changeable. Furthermore, the prediction of broken or potentially unavailable links also improves the reliability of communication; it improves the traditional link prediction method based on time-series learning [25] and the link connectivity judgment method, which is computationally complex and occupies a large amount of memory.