Underwater Internet of Things destroy-resistant link optimization method combining self-guiding graph representation and AMF

Abstract

The invention discloses an underwater Internet of Things destroy-resistant link optimization method combining self-guided graph representation and AMF. The method comprises the following steps: 1) generating a scale-free topology and a candidate link set according to node positions, residual energy and underwater acoustic channel quality; 2) inputting the topology into a self-guided graph representation learning model, and applying contrast constraint through unsupervised training to obtain robust node embedding representation; 3) on the basis of candidate link two-end embedding, combining node degree, betweenness centrality, bridge edge coefficient and link success rate to construct feature representation, and generating an existence probability matrix through a probability mapping function; 4) inputting the matrix into adaptive multi-layer filtering (AMF), reserving edges according to Top-k to form a connected skeleton, setting a hierarchical threshold according to node hierarchy and load, and maintaining approximate power law degree distribution; and 5) applying a global connectivity constraint, and outputting an optimized link structure. According to the method, link optimization can be realized without supervision labels, network connectivity and survivability are improved in an attack environment, and the method is suitable for ocean monitoring and security scenes.

Description

Technical Field

[0001] This invention relates to the field of underwater Internet of Things (IoT) technology, and more particularly to an underwater IoT robust link optimization method using joint self-guided graph representation and AMF. Background Technology

[0002] Underwater IoT is a wireless network comprised of numerous intelligent sensor nodes and underwater mobile nodes, communicated via underwater acoustic communication technology. It is widely used in underwater pollution monitoring, ocean water sampling, tsunami and earthquake early warning, and intrusion detection. Because underwater IoT is typically deployed in unattended environments, potentially vulnerable to malicious attacks, nodes may fail due to energy depletion or network attacks. These failed nodes can disrupt network connections and lead to cascading failures throughout the network. Therefore, researching robust link optimization techniques for underwater IoT is crucial.

[0003] Commonly used methods for network topology control in underwater IoT include: (1) mobility-assisted link optimization; (2) optimizing node transmission power to control network structure; and (3) reinforcement learning-based link optimization. Mobility-assisted link optimization is a common approach in underwater IoT, where nodes reorganize the network topology by moving to new locations through depth adjustments. For example, in geolocation-based opportunistic routing, when data is transmitted to an empty node, a depth adjustment mode is switched, and the empty node is vertically moved to a new depth to restore the route. Anchoring node mobility can optimize the three-dimensional underwater topology, and some literature uses virtual force-based link optimization algorithms to improve network connectivity and coverage.

[0004] A search revealed Chinese patent publication number CN113708953B, which discloses a method for the resilience of underwater acoustic sensor networks based on node importance balancing. This method statistically analyzes the importance criteria of all nodes in the network, assesses the importance of each node, and identifies critical and non-critical nodes. The sink node determines a node importance balancing scheme based on the node importance identification results, broadcasts a node importance balancing control packet, and executes topology control and routing control to balance the importance of network nodes. A TOPSIS multi-criteria evaluation algorithm is proposed to evaluate the importance of each node and distinguish between critical and non-critical nodes. Then, the sink node formulates a balancing scheme: if a critical node has a non-critical node nearby, the non-critical node moves closer to share the load; otherwise, routing control reduces the forwarding probability of the critical node. Dong et al., in their paper "Learningautomaton-based energy-efficient and fault-tolerant topology evolution algorithm for underwater acoustic sensor network," proposed a highly energy-efficient and fault-tolerant underwater topology optimization algorithm called ETFLA. This method first uses a learning automata algorithm to optimize the transmission power, thereby obtaining the optimal power configuration for each node; on this basis, an improved priority growth mechanism is introduced to further enhance the fault tolerance performance of the underwater network.

[0005] Existing technologies, similar to the optimization methods described above, rely on manually set targets. Inappropriate selection of indicators or weights can easily lead to evaluation biases. They also lack adaptive learning and therefore have room for improvement. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a joint self-guided graph representation and AMF-based underwater IoT robust link optimization method. Its advantages include achieving link optimization without the need for supervised tags, improving network connectivity and robustness under attack environments, and making it suitable for marine monitoring and security scenarios.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] The joint self-guided graph representation and AMF underwater IoT survivability link optimization method include the following steps:

[0009] Step 1: The unmanned surface vessel (USV) acts as the data control center, collecting the position coordinates, remaining energy, and underwater acoustic channel quality information of all underwater nodes to construct an initial scale-free topology G = (V, E), A ∈ {0, 1}. |V|×|V| And generate a candidate link set.

[0010] Step 2: Input the initial scale-free network topology and node feature matrix X into the self-guided graph representation learning model, and apply contrast constraints between nodes through unsupervised training to obtain robust node embedding representations;

[0011] Step 3: Embed z of nodes at both ends of the candidate link i and z j By combining network topology characteristics and physical layer channel characteristics, a feature representation is constructed for each link in the candidate link set, and an existence probability matrix is ​​generated;

[0012] Step 4: Use adaptive multilayer filtering to filter the link existence probability matrix layer by layer;

[0013] Step 5: After obtaining the filtered link set, apply global connectivity constraints to obtain the final optimized link connection structure.

[0014] The present invention is further configured such that the candidate link set The expression is:

[0015]

[0016] Where V is the set of nodes, E is the set of edges, A is the adjacency matrix, and v i and v j These are two different nodes in the node set.

[0017] The present invention is further configured to input the initial scale-free network topology and node feature matrix X into a self-guided graph representation learning model, and apply contrastive constraints between nodes through unsupervised training to obtain robust node embedding representations, including:

[0018] Based on the collected network attributes of underwater nodes, namely location coordinates, depth, remaining energy, and underwater acoustic channel quality, the original input feature set X is obtained:

[0019]

[0020] In the formula, x1 is the input feature of node v1, x i It is node v i Input features, It is a d-dimensional real vector space;

[0021] A graph attention network is used as the encoder, and the original node features x i Using the adjacency matrix A as input, the node features x i Projection as a low-dimensional latent representation:

[0022]

[0023] In the formula, σ(·) is a nonlinear activation function, and N(i) represents the node v i The neighborhood group, For the l-th layer slave node v i to v j Attention coefficient, W (l) Let l be the weight matrix of the l-th layer. It is node v i Feature representation at layer l, It is node v i Feature representation at layer l+1;

[0024] Represent the node as h i Inputting the projection function g(·), mapping it to the representation space, yields the projected representation. Then Input the prediction function q(·) to obtain the prediction representation

[0025] Applying two types of random augmentations, adjacency perturbation and feature mask, to the graph structure G, we obtain the representations of the same node from two perspectives: online branch output prediction representation. Target branch output projection representation loss through self-guided comparison Force the same node to maintain consistent representation across different perspectives without introducing negative samples;

[0026] During training, only the parameters of the online branch are updated using gradient descent; the parameters of the target branch are updated iteratively with the online branch using an exponential moving average: φ←τφ+(1-τ)θ, where θ represents the online branch parameter, φ represents the target branch parameter, and τ is the momentum coefficient; after multiple iterations until convergence, the final node embedding representation is obtained:

[0027] The present invention is further configured such that the contrast loss The loss function is defined as:

[0028]

[0029] In the formula, |V| represents the total number of nodes in the graph, sg(·) represents stopping the gradient operation, and ‖·‖ represents the L2 norm of the vector.

[0030] The present invention is further configured to embed z based on the nodes at both ends of the candidate link. i and z j By combining network topology characteristics and physical layer channel characteristics, a feature representation is constructed for each link in the candidate link set, and an existence probability matrix is ​​generated, including:

[0031] The betweenness centrality property is used to characterize a node's control over shortest path transmission in a network; its expression is:

[0032]

[0033] In the formula, ψ et (i) represents all slave nodes v e to node v t In the shortest path, passing through node v i The number of paths; ψ et It is node v e and v t The number of all shortest paths between;

[0034] To characterize the criticality of a link in the network, a bridge-side coefficient is introduced;

[0035] The underwater acoustic channel transmission success rate is used to characterize the impact of the underwater environment on node characteristics. The main characteristics of the acoustic channel model are signal attenuation and environmental noise, which depend on the distance l the signal travels and the frequency f. The propagation loss A(l,f) is given by Urick's empirical formula:

[0036] 10logA(l,f) / A0=10ρlogl+10llogα(f);

[0037] A0 is a unit normalization constant, ρ is the propagation factor, and α(f) is the absorption coefficient; the absorption coefficient α(f) is expressed by Thorpe's formula:

[0038]

[0039] When only path loss gain and loss are considered, node v i and v j The success rate of link transmitting m-bit data between them ij for:

[0040]

[0041] In the formula, ξ(f) is the signal transmission power, ξ(f) is the ambient noise, and Δf is the receiver noise bandwidth.

[0042] Based on candidate links Embedding of two ends of the node (z i ,z j ), and combine node degree, betweenness centrality, and bridge edge coefficient. and link success rate ij Features such as these are used to construct the feature vector f of the candidate link. ij :

[0043]

[0044] In the formula, ·||· represents vector concatenation, and deg(i) and deg(j) are the nodes v and v, respectively. i and v j The degree;

[0045] After obtaining the feature vectors of candidate links, the link (v) is calculated using a logistic regression function. i ,v j Probability of existence:

[0046] p ij =σ(w T f ij +b);

[0047] In the formula, σ(·) is the Sigmoid function, w is the weight matrix, and b is the bias term; the probabilities of all candidate links are summed to form the existence probability matrix.

[0048] The present invention is further configured such that the expression for the bridge side coefficient is:

[0049]

[0050] in,

[0051]

[0052] In the formula, It is node v i and v j The bridge-side coefficient of the link, S is the node v i The neighbor set of node v, R is the set of nodes v. j The neighbor set of node v, where |S| and |R| are the values ​​of node v. i and v j The number of neighbors, μ sr It is an indicator of whether nodes s∈S and r∈R in the neighbor set are connected, ∑ s∈S ∑ r∈R μ sr ∑ is the number of existing connecting edges between neighbor sets S and R. s∈S μ js Represents node v j The total strength of connections between all nodes in set S, ∑ r∈R μ ri Represents node v i The total strength of connections between nodes in set R and all nodes in the set R, O = |S∩R|, represents the connection strength between node v. i and v j The number of common neighbors, ∈ is a moderating factor.

[0053] The present invention is further configured to employ adaptive multilayer filtering to perform layer-by-layer filtering on the link existence probability matrix, including:

[0054] First, based on connectivity constraints, the links of the nodes are sorted by probability. The top-k edges of each node are retained, that is, the k links with the highest probability for each node are retained to form the initial connected skeleton.

[0055] Based on the initial skeleton, hierarchical load constraints are further applied to maintain the scale-free nature of the topology. Based on the connected skeleton, different thresholds are adaptively set according to the load level of the nodes: high-load nodes typically have lower thresholds, while low-load nodes have higher thresholds. The selection set is defined as follows:

[0056]

[0057] In the formula, θ l(i) For node v i Adaptive threshold for the relevant level It is node v i The set of links after threshold filtering.

[0058] The present invention is further configured such that the initial connected skeleton expression is:

[0059]

[0060] In the formula, It is node v i The set of links that are connected to at least k neighbors.

[0061] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0062] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0063] The beneficial effects of this invention are as follows:

[0064] 1. This invention introduces a self-guided graph representation learning model. Without the need for manually setting multi-dimensional indicators, it combines scale-free topology and node features, and obtains robust embeddings in unsupervised contrastive learning through graph attention networks and random augmentation. Furthermore, it integrates features such as betweenness centrality, bridge edge coefficients, and link success rates to generate an existence probability matrix, thereby achieving end-to-end modeling of link importance, avoiding human bias, and improving the accuracy and generalization ability of link importance assessment.

[0065] 2. This invention combines the adaptive multilayer filtering (AMF) mechanism to transform the link robustness optimization problem into a hierarchical thresholding and global constraint screening problem. It can achieve adaptive selection and redundancy construction of critical links while maintaining an approximate power-law degree distribution, thereby enhancing the overall connectivity and topology stability of the network.

[0066] 3. This invention takes into account both random failure and malicious attack scenarios at the link layer. By using deep characterization and multi-layer screening for joint optimization, it can effectively distribute the communication load, reduce the risk of critical links being attacked, and demonstrate stronger resilience and better energy efficiency in complex and changeable underwater environments, thus having broad application prospects. Attached Figure Description

[0067] Figure 1 This is a schematic diagram illustrating the workflow of the joint self-guided graph representation and AMF underwater IoT survivability link optimization method proposed in this invention.

[0068] Figure 2 This is a schematic diagram of the underwater IoT system architecture proposed in this invention, which combines the joint self-guided graph representation and the underwater IoT robust link optimization method of AMF.

[0069] Figure 3 This is a schematic diagram of the generated topology of the joint self-guided graph representation and the underwater IoT survivability link optimization method of AMF proposed in this invention;

[0070] Figure 4 This is a schematic diagram of the degree distribution probability corresponding to the topology of the joint self-guided graph representation proposed in this invention and the underwater IoT survivability link optimization method of AMF;

[0071] Figure 5 This is a schematic diagram comparing the network connectivity of the joint self-guided graph representation proposed in this invention and the underwater IoT survivability link optimization method of AMF after being subjected to random attacks.

[0072] Figure 6 This is a schematic diagram comparing the cascading robustness of the joint self-guided graph representation proposed in this invention with the underwater IoT survivability link optimization method of AMF after being subjected to malicious attacks. Detailed Implementation

[0073] The technical solution of this patent will be further described in detail below with reference to specific embodiments.

[0074] The embodiments of this patent are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this patent, and should not be construed as limiting this patent.

[0075] Reference Figure 1 and Figure 2 The joint self-guided graph representation and AMF underwater IoT survivability link optimization method include the following steps:

[0076] Step 1: The unmanned surface vessel (USV) acts as the data control center, collecting the position coordinates, remaining energy, and underwater acoustic channel quality information of all underwater nodes. Based on the collected information, the USV constructs an initial scale-free topology G = (V, E), A ∈ {0, 1}. |V|×|V| Where V is the set of nodes, E is the set of edges, A is the adjacency matrix, and a candidate link set is generated.

[0077]

[0078] In equation (1), v i and v j These are two different nodes in the node set.

[0079] Step 2: Input the scale-free network topology and node feature matrix X into the self-guided graph representation learning model, and apply contrast constraints between nodes through unsupervised training to obtain robust node embedding representations.

[0080] Step 2.1: Based on the network attributes of the underwater nodes collected in Step 1, namely location coordinates, depth, remaining energy, and underwater acoustic channel quality, assign a value to each node v. i Construct the initial feature vector to obtain the original input feature set X:

[0081]

[0082] Where x1 is the input feature of node v1, x i It is node v i Input features, It is a d-dimensional real vector space.

[0083] Step 2.2: Use a graph attention network as the encoder, and the original node features x i Using the adjacency matrix A as input, the node features x i Projection as a low-dimensional latent representation:

[0084]

[0085] In equation (3), σ(·) is a nonlinear activation function, and N(i) represents the node v i The neighborhood group, For the l-th layer slave node v i to v j Attention coefficient, W (l) Let l be the weight matrix of the l-th layer. It is node v i Feature representation at layer l, It is node v i Feature representation at layer l+1.

[0086] Step 2.3: Represent the node h i Inputting the projection function g(·), mapping it to the representation space, yields the projected representation. The aim is to reduce interference from factors irrelevant to downstream tasks, allowing the learning objective to focus on a space more conducive to self-supervision. Then... Input the prediction function q(·) to obtain the prediction representation

[0087] Step 2.4: Apply two types of random augmentations, adjacency perturbation and feature mask, to the graph structure G to obtain the representation of the same node from two perspectives: online branch output prediction representation. Target branch output projection representation loss through self-guided comparison This forces the representation of the same node to remain consistent across different viewpoints without introducing negative samples. The loss function is defined as:

[0088]

[0089] In Equation (4), |V| represents the total number of nodes in the graph, ||·|| represents the L2 norm of the vector, and sg(·) indicates stopping the gradient operation, meaning that the parameters of the target branch do not participate in backpropagation and are only updated by the online branch through exponential moving average. This loss achieves stable and effective self-supervised training by minimizing the cosine distance between the predicted representation and the target representation of the same node from different perspectives, and avoids the dependence of traditional contrastive learning methods on a large number of negative samples.

[0090] Step 2.5: During training, gradient descent is used to update the parameters of only the online branches (encoder, projector, and predictor). The parameters of the target branch do not participate in backpropagation but are updated iteratively with the online branches via exponential moving average: φ←τφ+(1-τ)θ, where θ represents the online branch parameters, φ represents the target branch parameters, and τ is the momentum coefficient. After multiple iterations until convergence, the encoder and projector outputs of the online branches are taken as the final node embedding representation.

[0091] Step 3: Based on the node embedding representation z at both ends of the candidate link i and z j By combining network topology characteristics and physical layer channel characteristics, a feature vector is constructed for each link in the candidate link set, and an existence probability matrix is ​​generated.

[0092] Step 3.1: Construct network feature information for scale-free networks using betweenness centrality. Betweenness centrality reflects the degree to which a node acts as a mediator for the shortest path between other node pairs in a network. i Betweenness centrality is defined as:

[0093]

[0094] In equation (5), ψ et (i) represents all slave nodes v e to node v t In the shortest path, passing through node v i The number of paths; ψ et It is node v e and v t The number of all shortest paths between them.

[0095] Step 3.2: To characterize the criticality of a link in the network, a bridge-side coefficient is introduced, the expression of which is:

[0096]

[0097] in,

[0098]

[0099] In equations (6) and (7), It is node v i and v j The bridge-side coefficient of the link, S is the node v i The neighbor set of node v, R is the set of nodes v. j The neighbor set of node v, where |S| and |R| are the values ​​of node v. i and v j The number of neighbors, μ sr It is an indicator of whether nodes s∈S and r∈R in the neighbor set are connected, ∑ s∈S ∑ r∈R μ sr ∑ is the number of existing connecting edges between neighbor sets S and R. s∈S μ js Represents node v j The total strength of connections between all nodes in set S, ∑ r∈R μ ri Represents node v i The total strength of connections between nodes in set R and all nodes in the set R, O = |S∩R|, represents the connection strength between node v. i and v j The number of common neighbors, ∈ is a moderating factor.

[0100] Step 3.3: The underwater acoustic channel transmission success rate is used to characterize the impact of the underwater environment on node characteristics. The main characteristics of the acoustic channel model are signal attenuation and noise in the environment, which depend on the distance l the signal travels and the frequency f. The propagation loss A(l,f) is given by Urick's empirical formula:

[0101] 10logA(l,f) / A0=10ρlogl+10llogα(f) (8)

[0102] A0 is a unit normalization constant, ρ is the propagation factor, and α(f) is the absorption coefficient. The absorption coefficient α(f) is expressed using Thorp's formula:

[0103]

[0104] When only path loss gain and loss are considered, node v i and v j The success rate of a link transmitting m bits of data between them is:

[0105]

[0106] In equation (10), ξ(f) is the signal transmission power, ξ(f) is the ambient noise, and Δf is the receiver noise bandwidth. This metric reflects the impact of the physical layer on link reliability.

[0107] Step 3.4: Based on candidate links Embedding of two ends of the node (z i ,z j ), and combine node degree (deg) and betweenness centrality Bridge side coefficient and link success rate Features such as these are used to construct the feature vector f of the candidate link. ij :

[0108]

[0109] In equation (11), ·||· represents vector concatenation, and deg(i) and deg(j) are the nodes v i and v j The feature vector integrates node semantic representation, topological attributes, and physical layer characteristics.

[0110] Step 3.5: After obtaining the feature vectors of candidate links, calculate the link (v) using the logistic regression function. i ,v j Probability of existence:

[0111] p ij =σ(w T f ij+b) (12)

[0112] In equation (12), σ(·) is the Sigmoid function, w is the weight matrix, and b is the bias term. The probabilities of all candidate links are summed to form the existence probability matrix.

[0113] Step 4: To ensure that the candidate link probability matrix maintains near-scale-free characteristics as much as possible while satisfying global connectivity, an adaptive multilayer filter (AMF) is introduced to screen the links layer by layer.

[0114] Step 4.1: First, using connectivity as a constraint, sort the nodes by the probability of link existence, and retain the top-k edges for each node, that is, retain the k links with the highest probability for each node, forming the initial connected skeleton, the expression of which is:

[0115]

[0116] In equation (13), It is node v i The set of links that are connected to at least k neighbors.

[0117] Step 4.2: Based on the initial skeleton, further hierarchical load constraints are applied to maintain the scale-free nature of the topology. Based on the connected skeleton, different thresholds are adaptively set according to the node load level: high-load nodes typically have lower thresholds to retain more links and ensure robustness; low-load nodes have higher thresholds to limit redundant connections. The selection set is defined as follows:

[0118]

[0119] In equation (14), θ l(i) For node v i Adaptive threshold for the relevant level It is node v i The set of links after threshold filtering. This process ensures that the degree distribution of nodes at different levels in the network conforms to the scale-free property and can adaptively reflect the differences in node load.

[0120] Step 5: Obtain the candidate link set after AMF filtering. Then, global connectivity constraints are further applied to ensure that the final network topology maintains strong connectivity. Based on this, an optimized link connection structure that satisfies the balance between robustness and energy efficiency is obtained as the final topology control result.

[0121] Step 6: Conduct two tests on the optimized topology: random attack and malicious attack. Random node deletion is used to simulate a random attack, while deleting nodes in descending order of degree simulates a malicious attack. This invention utilizes cascading robustness. Two metrics, network connectivity and network resilience, are used to assess network resilience.

[0122]

[0123] In equation (15), C net (i) represents the number of nodes that can be connected to the USV after deleting the i-th node with the highest degree over a period of time.

[0124] The data required for this invention includes network topology and simulated node information. The network topology is generated based on the Barabási-Albert model combined with actual underwater deployment, maintaining scale-free characteristics. The specific steps are as follows: ① Surface USVs periodically send beacons; nodes receiving the beacons establish links with the USVs, forming an initial backbone network and marking backbone nodes. ② Each marked node broadcasts a query packet carrying its own coordinates to its neighborhood; when a neighboring node receives this packet for the first time, it selects a target and establishes a connection according to the priority connection rule and connection probability, forwarding the packet only once. ③ This process is recursively performed in the network; received query packets are directly discarded to avoid redundant propagation. Thus, the topology grows layer by layer from the surface to deep water, with nodes closer to the USVs having higher node degrees than nodes in other areas, maintaining overall scale-free characteristics.

[0125] Figure 3 This paper demonstrates the topology generated by the proposed "self-guided graph representation + AMF" robust link optimization method. Squares represent surface USVs, and dots represent underwater IoT nodes; the size of the dots distinguishes node degree, with larger dots indicating higher degree; solid lines represent effective connections between nodes. Meanwhile, Figure 4 Showing Figure 3 The degree distribution probability corresponding to the topology shown is as follows: it can be seen that most nodes have low degree and a few nodes have high degree, and the overall distribution is approximately power-law, indicating that the network still maintains its scale-free characteristics after adopting this method.

[0126] To verify the advantages of this invention over the prior art Figure 5 This paper presents a performance comparison of the method of this invention with three other underwater survivability topology optimization methods under random attacks. The horizontal axis represents the attack rate, i.e., the proportion of attacked nodes to the total number of nodes; the vertical axis represents the network connectivity after a random attack. Initial-BA is a method for generating scale-free topology based on the Barabási-Albert model. The results show that as the attack rate increases, leading to more node failures, the network connectivity of all methods decreases. However, the method of this invention maintains higher connectivity throughout, and its advantage is more significant under high attack rates.

[0127] To verify the advantages of this invention over the prior art Figure 6The results show a comparison between the method of this invention and three other underwater survivability topology optimization methods under malicious attacks. The horizontal axis represents the number of nodes, and the vertical axis represents the cascading robustness after the malicious attack. Initial-BA and ETFLA are both based on scale-free topology, and their cascading robustness is the lowest. TOPSIS improves network resilience to some extent by balancing node importance, thus outperforming the former two. The method of this invention introduces a self-guided graph representation learning model, which integrates node degree, betweenness centrality, bridge-side coefficient, and link success rate to construct link feature representations and predict links. Combined with the AMF mechanism, it determines the existence of links between nodes based on the link generation probability matrix, thereby achieving adaptive selection and redundancy construction of critical links. Therefore, it exhibits superior resilience among various comparison methods.

[0128] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0129] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above method.

[0130] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An underwater Internet of Things invulnerability link optimization method combining self-guided graph representation and AMF, characterized in that, The method comprises the following steps: Step 1: The water unmanned ship as a data control center collects the position coordinates, residual energy and underwater acoustic channel quality information of all underwater nodes, constructs an initial scale-free topology G=(V,E), A∈{0,1} |V|×|V| , and generates a candidate link set Step 2: inputting the initial scale-free network topology and the node feature matrix X into a self-guided graph representation learning model, applying contrast constraints between nodes through unsupervised training, and obtaining robust node embedding representation; Step 3: Constructing feature representation for each candidate link based on node embedding z at both ends of the candidate link i and z j , combining network topology characteristics and physical layer channel characteristics, constructing a feature representation for each link in the candidate link set, and generating a presence probability matrix; Step 4: performing layer-by-layer screening on the link existence probability matrix by using adaptive multi-layer filtering; Step 5: after obtaining the screened link set, a global connectivity constraint is applied to obtain the final optimized link connection structure.

2. The method of claim 1, wherein the method further comprises: the candidate link set The expression is: Wherein, V is a node set, E is an edge set, A is an adjacency matrix, v i and v j are two different nodes in the node set, respectively. 3.The method of claim 1, wherein the method further comprises: receiving, by the joint self-bootstrapping graph representation and AMF, a request for a new connection from a new node; and determining, by the joint self-bootstrapping graph representation and AMF, a new path for the new node based on the new connection request. The initial scale-free network topology and the node feature matrix X are input into a self-guided graph representation learning model, contrast constraints between nodes are applied through unsupervised training, and robust node embedding representation is obtained, including: According to the collected network properties of the underwater nodes, i.e., position coordinates, depth, residual energy and underwater acoustic channel quality, an original input feature set X is obtained: wherein x1is an input feature of node v1, x i is an input feature of node v i , is a d-dimensional real vector space; Using a graph attention network as the encoder, the original node features x i and the adjacency matrix A as input, the node features x i are projected into a low-dimensional latent representation: where σ(·) is a nonlinear activation function, N(i) denotes the neighbor set of node v i , is the attention coefficient from node v i to v j in the l-th layer, W (l) is the weight matrix in the l-th layer, is the feature representation of node v i in the l-th layer, is the feature representation of node v i in the (l+1)-th layer; Representing a node h i Input a projection function g(·) that maps to a representation space, resulting in a projected representation Subsequently, the projected representation is Input a prediction function q(·) that results in a predicted representation Applying two random augmentations, adjacency perturbation and feature mask, to the graph structure G, respectively, obtains the representation of the same node under two perspectives: online branch output prediction representation Target branch output projection representation Through self-guided contrastive loss Forcing the representation of the same node under different perspectives to be consistent without introducing negative samples; During training, only the online branch parameters are updated by gradient descent; the parameters of the target branch are updated by exponential moving average with the online branch iteration: φ <- τφ + (1 - τ) θ, where θ represents the online branch parameters, φ represents the target branch parameters, and τ is the momentum coefficient; after multiple iterations until convergence, the final node embedding representation is obtained: v i ∈V.

4. The method of claim 3, wherein the method further comprises: the contrastive loss The loss function is defined as: In the formula, |V| represents the total number of nodes in the graph, sg(·) represents the stop gradient operation, and ||·|| represents the two-norm of the vector.

5. The method of claim 1, wherein the method further comprises: node embeddings z at both ends of the candidate link i and z j , combined with network topology characteristics and physical layer channel characteristics, to construct a feature representation for each link in the candidate link set, and generate a presence probability matrix, comprising: The betweenness centrality is used to characterize the control ability of the node in the transmission of the shortest path in the network, and the expression is: wherein ψ et (i) is the number of paths from all nodes v e to node v t that pass through node v i ; ψ et is the number of all shortest paths between nodes v e and v t ; In order to characterize the key degree of the link in the network, a bridge edge coefficient is introduced; The underwater acoustic channel transmission success rate is used to represent the influence of the underwater environment on the node characteristics. The main characteristics of the acoustic channel model are signal attenuation and noise in the environment, which depend on the distance l and frequency f that the signal passes through. The propagation loss A(l, f) is given by the Urick empirical formula: 10logA(l,f) / A0=10ρlogl+10llogα(f); A0 is a unit normalization constant, p is a propagation factor, and a(f) is an absorption coefficient; the absorption coefficient a(f) is expressed by the Thorp formula: When only path loss gain and loss are considered, the link success probability of transmitting m bit data between nodes v i and v j is: ​ wherein is the signal transmission power, ξ(f) is the ambient noise, and Δf is the receiver noise bandwidth. Based on the candidate link Two end nodes embedded in the chain (z i ,z j ), combined with node degree, betweenness centrality, bridge coefficient And link success rate And other features, build a feature vector f ij : where •||• denotes vector concatenation, deg(i) and deg(j) are the degrees of nodes v i and v j respectively; After obtaining the feature vector of the candidate link, the link (v i ,v j ) existence probability is calculated by a logistic regression function: p ij = σ(w T f ij + b); where σ(·) is a sigmoid function, w is a weight matrix, and b is a bias term; and aggregating all candidate link probabilities to form a presence probability matrix 6. The method of claim 5, wherein the method further comprises: The bridge edge coefficient expression is: Wherein, In the formula, is the bridge coefficient of the link where node v i and v j are located, S is the neighbor set of node v i , R is the neighbor set of node v j , |S| and |R| are the number of neighbors of node v i and v j , μ sr is an indication of whether node s∈S is connected to node r∈R in the neighbor set, ∑ s∈S ∑ r∈R μ sr is the number of existing connection edges between the neighbor sets S and R, ∑ s∈S μ js represents the total connection strength between node v j and all nodes in the set S, ∑ r∈R μ ri represents the total connection strength between node v i and all nodes in the set R, O=|S∩R| represents the number of common neighbors of node v i and v j , ∈ is a regulating factor.

7. The method of claim 1, wherein the method further comprises: receiving, by the AMF, a request for a new PDU session from the UE; and sending, by the AMF, a request for a new PDU session to the joint self-bootstrapping graph representation and AMF underwater IIoT resilient link optimization server. The adaptive multi-layer filtering is used to perform layer-by-layer screening on the link existence probability matrix, including: First, sort the link existence probability of the node according to the connectivity constraint, and retain the Top-k edges of each node, i.e., retain the k links with the highest probability for each node, to form an initial connected skeleton; On the basis of the initial skeleton, further hierarchical load constraints are applied to maintain the scale-free property of the topology structure; on the basis of the connected skeleton, different threshold values are adaptively set according to the load level of the node: high-load nodes usually take lower threshold values, and low-load nodes take higher threshold values; the screening set is defined as: where θ l(i) is the node v i adaptive threshold of the hierarchy, is the node v i the set of links filtered by threshold. 8.The method of claim 7, wherein the method further comprises: receiving, by the AMF, a request for a new PDU session from the UE; and sending, by the AMF, a response to the request for the new PDU session to the UE. The initial connected skeleton expression is: wherein is a node v i a set of links to at least k neighbors.

9. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the method of any one of claims 1 to 8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 8.

Images