Neural network based high fault tolerance and anti-mobility underwater acoustic routing decision optimization method

Abstract

This invention discloses a highly fault-tolerant and mobility-resistant underwater acoustic routing decision optimization method based on neural networks, comprising: initializing the underwater acoustic sensor network; periodically updating the neighbor node list and generating multi-dimensional feature vectors for each node; inputting the multi-dimensional feature vectors into a neural network model to obtain three decision indicators, weighting and fusing them to obtain a comprehensive routing score for each neighbor node and ranking them to generate a candidate node sequence; integrating loop avoidance and opportunistic routing mechanisms to select the next-hop node; repeatedly executing packet forwarding and fault detection and recovery mechanisms until the packet is successfully delivered to the sink node; and updating the neural network weight parameters based on historical routing decision data for the next round. This invention achieves intelligent and highly fault-tolerant routing in dynamic underwater acoustic environments, extending the overall network lifetime.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic communication and network intelligent decision-making, and in particular to a highly fault-tolerant and mobility-resistant underwater acoustic routing decision optimization method based on neural networks, for achieving efficient and reliable data transmission in complex and dynamic underwater environments. Background Technology

[0002] With the increasing demands from fields such as marine resource development, environmental monitoring, and underwater security, underwater acoustic sensor networks are becoming increasingly important as a key infrastructure for realizing wide-area information sensing and transmission underwater. However, the inherent physical characteristics of the underwater environment, such as slow sound wave propagation speed, narrow channel bandwidth, large latency, high bit error rate, and significant multipath effects, coupled with the limited energy of nodes and the difficulty in replenishing them after deployment, pose serious challenges to traditional underwater routing protocols designed for terrestrial wireless networks or based on geometric models.

[0003] Traditional protocols designed for static or quasi-static networks, such as vector-based forwarding (VBF) or depth-based routing (DBR), typically rely on fixed or slowly changing topology information. When high-speed mobile nodes such as autonomous underwater vehicles (AUVs) are present in the network, these protocols struggle to adapt to rapid topology changes, often leading to route discovery failures and a sharp drop in packet delivery rates. Furthermore, underwater acoustic links are susceptible to intermittent interruptions due to environmental interference, and existing protocols often lack effective dynamic fault tolerance and rapid recovery mechanisms, resulting in insufficient network robustness. In addition, traditional routing decisions often rely on single or a few metrics (such as remaining energy or distance to the sink node), failing to fully utilize the multi-dimensional state information of nodes for joint optimization, making it difficult to achieve a good balance between data delivery rate, end-to-end latency, and network energy efficiency.

[0004] In recent years, the application of artificial intelligence technologies, especially machine learning methods, to network protocol design has become a research hotspot, aiming to endow networks with intelligent decision-making and self-optimization capabilities. However, existing intelligent routing schemes either rely on centralized training and decision-making, making them difficult to adapt to the distributed and low-communication-overhead characteristics of underwater acoustic networks; or they employ overly simplistic models that fail to capture the nonlinear characteristics of complex underwater dynamic environments; or they lack effective integration with classic fault-tolerant mechanisms such as opportunistic routing and predictive forwarding, resulting in limited overall performance improvement.

[0005] Therefore, there is an urgent need for an underwater acoustic network routing protocol that can deeply integrate intelligent decision-making and network dynamic characteristics. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a highly fault-tolerant and mobility-resistant underwater acoustic routing decision optimization method based on neural networks.

[0007] The specific technical solution is as follows:

[0008] A highly fault-tolerant and mobility-resistant underwater acoustic routing decision optimization method based on neural networks includes the following steps:

[0009] S1: Deploy underwater acoustic sensor network nodes, initialize network communication parameters, and neural network models within each non-convergence node;

[0010] S2: Each node periodically updates its list of neighboring nodes and collects network state information to generate multi-dimensional feature vectors;

[0011] S3. Adaptive Routing Decision: Input the multidimensional feature vector into the neural network model to obtain three decision indicators, and perform weighted fusion to obtain a comprehensive routing score for each neighbor node; sort the neighbor nodes according to the comprehensive routing score to generate a candidate node sequence; integrate loop avoidance mechanism and opportunistic routing mechanism to intelligently select the next hop node;

[0012] S4. Packet forwarding and execution of fault detection and recovery mechanisms: Integrating packet lifetime management, fast link failure switching and adaptive retransmission strategies to ensure network communication reliability in the event of node failure or link interruption.

[0013] S5: Repeat S4 until the data packet is successfully delivered to the aggregation node; during this process, historical routing decision data is stored in the experience replay buffer of each node on the transmission path.

[0014] S6: Based on the historical routing decision data, the neural network weight parameters are updated through reinforcement learning for the next round of adaptive routing decision-making.

[0015] Furthermore, in S1, the network communication parameters include: the node's initial energy, transmit power, receive power, idle power, and communication range.

[0016] Furthermore, the multidimensional feature vector includes: energy state, link reliability, node mobility, improved metric of the current node relative to the sink node, distance between nodes, packet priority, depth difference, and network congestion estimation.

[0017] Furthermore, in S3, the three decision indicators include: forwarding priority, reliability score, and energy efficiency score; the dynamic weight coefficients used for the weighted linear fusion of the three decision indicators are adaptively adjusted according to the energy level of the current node and the network congestion status.

[0018] Furthermore, in S3, the integrated loop avoidance mechanism and opportunistic routing mechanism, and the intelligent selection of the next-hop node, specifically involve:

[0019] If the highest value of the comprehensive routing score has reached the activation threshold, the neighbor node with the highest comprehensive routing score is selected as the next hop node, and the loop avoidance mechanism is applied to score and penalize the visited node; otherwise, if there is no suitable next hop node, the opportunistic routing mechanism is activated, and a mobile node is randomly selected from the neighbor nodes as an opportunistic relay.

[0020] Furthermore, in S4, when forwarding data packets, if the energy value of the next-hop node is lower than the preset energy value, or if the link between the current node and the next-hop node is interrupted, the next-hop node selected in S3 is determined to be invalid, and the next-hop node is selected again.

[0021] Furthermore, the fault detection and recovery mechanism is implemented through the following sub-steps:

[0022] S4.1: The current node sends a data packet to the selected next-hop node and starts an ACK timer to wait for an acknowledgment message;

[0023] S4.2: Determine whether the current node has received an ACK from the next-hop node within the preset timeout period. If it has, the data packet forwarding is considered successful, and proceed to S5. If it has not received an ACK, the currently selected next-hop node is considered to be invalid or the link is interrupted, and proceed to S4.3.

[0024] S4.3: The current node reads the candidate node sequence from the local cache, removes invalid nodes, and obtains the updated candidate node sequence;

[0025] S4.4: Determine if there are other available nodes in the updated candidate node sequence. If so, select a new neighbor node as the next hop node in descending order of comprehensive routing score, update the local routing table, and return to execute S4.1.

[0026] If not, the current node is determined to be in a routing hole or isolation state. At this time, the current node stops forwarding and sends a link failure notification message to the next hop node, triggering the next hop node to execute the retransmission mechanism or re-make routing decisions.

[0027] Furthermore, following dynamic topology adaptation, for nodes marked as moving, a random waypoint movement model is used to simulate their motion: the moving node randomly selects the next target location and moves towards the target point at a random speed, and then reselects the target after arriving, or periodically reselects the target point;

[0028] All nodes need to periodically update their neighbor node list, evaluate link quality based on underwater acoustic propagation characteristics, and update link status information; the neural network model combines historical information on node movement trends to make short-term predictions of topology change trends, thereby pre-adjusting routing strategies.

[0029] Furthermore, in S6, based on the historical routing decision data stored in the experience replay buffer, a batch of historical state-action-reward tuples are periodically randomly sampled from the experience replay buffer using the experience replay technique in the reinforcement learning mechanism, and the neural network weight parameters are updated using the backpropagation algorithm based on gradient descent.

[0030] A high-fault-tolerant and mobility-resistant underwater acoustic route decision optimization system based on neural networks is used to implement the aforementioned high-fault-tolerant and mobility-resistant underwater acoustic route decision optimization method based on neural networks. The system includes: a data input and processing layer, a feature extraction and decision layer, an execution and optimization layer, and a learning and feedback layer.

[0031] The data input and processing layer is responsible for the perception and collection of basic data, including: an environment perception module, a neighbor status collection module, a link quality detection module, and a self-status detection module, providing multi-dimensional raw network status information for subsequent processing;

[0032] The feature extraction and decision layer includes: a feature extractor, a neural network decision model, and a decision fusion unit; the feature extractor processes the original network state information, generates an eight-dimensional feature vector, and inputs it into the core neural network decision model; the neural network decision model includes an input layer, a hidden layer, and an output layer, which output three decision indicators; the decision fusion unit performs a weighted calculation on the three decision indicators to obtain a comprehensive routing score.

[0033] The execution and optimization layer includes: an executor, a packet forwarding module, a fault tolerance mechanism execution module, and a performance monitoring module. Based on the comprehensive routing score, the executor selects the next hop and drives packet forwarding. At the same time, it triggers the fault tolerance mechanism execution module when necessary and evaluates the network transmission performance in real time through the performance monitoring module.

[0034] The learning and feedback layer includes: an experience replay buffer, an online learning engine, and a model parameter update module. The experience replay buffer is used to store historical routing decision data. The online learning engine is used to calculate the optimization gradient. The model parameter update module performs neural network model parameter updates and feeds the updated parameters back to the neural network model, thereby forming a closed-loop adaptive optimization system.

[0035] The beneficial effects of this invention are:

[0036] This invention constructs an adaptive routing decision-making mechanism based on a distributed neural network, comprehensively utilizing multi-dimensional state information such as node energy, link quality, mobility, and network congestion to achieve intelligent and highly fault-tolerant routing in dynamic underwater acoustic environments. This method not only significantly improves packet delivery rate and reliability under node movement and link instability conditions, but also effectively reduces end-to-end transmission latency and balances network energy consumption by integrating opportunistic routing and online learning optimization, thereby greatly extending the overall network lifetime. Compared to traditional protocols, this invention has stronger environmental adaptability and decision-making flexibility, enabling continuous self-optimization of routing performance with lower computational and communication overhead while ensuring communication quality. It provides an efficient and robust solution for reliable underwater acoustic communication in complex marine scenarios. Attached Figure Description

[0037] Figure 1 This is a flowchart of a high-fault-tolerant and mobility-resistant underwater acoustic routing decision optimization method based on neural networks, as described in this embodiment of the invention.

[0038] Figure 2 This is a flowchart of the fault detection and recovery mechanism executed in S4 in an embodiment of the present invention.

[0039] Figure 3 This is a schematic diagram of the structure of the high fault tolerance and mobility-resistant underwater acoustic routing decision optimization system based on neural networks in an embodiment of the present invention. Detailed Implementation

[0040] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The objectives and effects of the present invention will become clearer as a result. The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0041] like Figure 1 As shown, a highly fault-tolerant and mobility-resistant underwater acoustic routing decision optimization method based on neural networks includes the following steps:

[0042] S1: Initialization of the underwater acoustic sensor network (hereinafter referred to as the network). Deploy underwater acoustic sensor network nodes according to the network size, set the location of the sink node, and configure the network communication parameters of the nodes; initialize a distributed local neural network model (hereinafter referred to as the model) for each non-sink node in the network.

[0043] Based on their mobility, underwater acoustic sensor network nodes can be classified into fixed nodes and mobile nodes. Based on their functional roles, locations and deployments, capabilities and resources, underwater sensor network nodes can be classified into convergence nodes (or "central control nodes," typically deployed on the surface) and distributed intelligent nodes (typically deployed underwater, including underwater sensor nodes and underwater communication nodes). Network communication parameters include: node initial energy, transmit power, receive power, idle power, and communication range.

[0044] Each node's neural network model includes an input layer, hidden layers, and an output layer, used to output three decision indicators. Each layer is initialized, and the connection weights and biases of the neural network model are randomly initialized. In this embodiment, the input layer, serving as the interface for the model to interact with external data, has 8 neurons and receives input signals. The hidden layer is the core processing unit of the neural network model, equipped with 16 neurons and using the ReLU activation function. The output layer, as the final decision-making stage of the neural network model, has 3 neurons and uses the Sigmoid activation function. These three neurons correspond to the three output decision indicators, each with an output value between 0 and 1, representing a certain probability or degree of that indicator.

[0045] S2: Multidimensional State Feature Extraction: The underwater acoustic sensor network nodes periodically perform a neighbor discovery process, generating a locally stored list of neighbor nodes. Each node periodically collects network state information (its own and detected neighbor node state information) to generate a multidimensional feature vector. In this embodiment, an eight-dimensional feature vector is generated: energy state, link reliability, node mobility, improvement metric of the current node relative to the sink node, inter-node distance, packet priority, depth difference, and network congestion estimation. The eight-dimensional feature vector is detailed below:

[0046] ① Energy state: This includes the percentage of remaining energy of a node, which is obtained by calculating the ratio of the current node's remaining energy to its initial energy.

[0047] ②Link reliability: The signal-to-noise ratio (SNR) from the current node to each neighboring node is calculated based on the underwater acoustic propagation attenuation model. At the same time, the communication probability is calculated by combining the underwater acoustic propagation model and the distance between nodes. The reliability of the link from the current node to each neighboring node is evaluated by combining the SNR and the communication probability.

[0048] ③ Node mobility: The current node's own mobility attributes are identified and marked by node type and motion state. In this embodiment, fixed nodes are 0 and mobile nodes are 1.

[0049] ④ Improvement metric relative to the sink node: The improvement metric measures the dispersion of nodes in the underwater acoustic sensor network. A larger improvement metric value indicates a greater difference in distance between the current node and its neighboring nodes to the sink node, indicating a more dispersed node distribution; conversely, a smaller improvement metric value indicates a more concentrated node distribution. Specifically, the difference between the Euclidean distance from the current node to the sink node and the Euclidean distance from its neighboring nodes to the sink node is calculated. A positive difference indicates that the neighboring node is closer to the sink node, and a larger difference indicates a higher improvement metric, resulting in higher transmission efficiency when selecting that neighboring node for forwarding. A negative difference indicates that the neighboring node is farther from the sink node, resulting in lower transmission efficiency when selecting that neighboring node for forwarding.

[0050] ⑤ Distance between nodes: Normalize the distance between nodes to the communication range; in this embodiment, the distance between nodes is calculated using three-dimensional Euclidean distance.

[0051] ⑥ Packet priority: Read the application layer priority identifier of the data packets to be forwarded within the current node; the priority is set according to the urgency of the data packets and application requirements.

[0052] ⑦ Depth difference: Calculate the absolute difference in depth (z coordinate) between the current node and the candidate neighbor nodes.

[0053] ⑧ Network congestion estimation: Estimate the degree of local network congestion based on the length of the data buffer queue (the queue of data packets to be sent, stored in the node buffer) in this node.

[0054] S3: Adaptive Routing Decision: The extracted eight-dimensional feature vector is input into the neural network model of this node. After processing by the hidden layer activation function (ReLU activation function performs non-linear fusion and abstraction of features) and the output layer (using Sigmoid activation function to output scalar values ​​between 0 and 1), three decision indicators are obtained: forwarding priority, reliability score and energy efficiency score. The three decision indicators are weighted and fused to calculate the comprehensive routing score of each neighbor node. The neighbor nodes are sorted according to the comprehensive routing score to generate a candidate node sequence.

[0055] The routing decision integrates both loop avoidance and opportunistic routing mechanisms, intelligently selecting the next-hop node: it determines whether the highest value of the comprehensive routing score has reached the activation threshold. If so, it selects the neighbor node with the highest comprehensive routing score as the next-hop node, and applies the loop avoidance mechanism to penalize the visited node (specifically, during the selection process, it checks whether the node has already appeared in the packet forwarding path history; if so, it applies a penalty factor to the comprehensive routing score to avoid routing loops); if not, there is no suitable next hop, and the opportunistic routing mechanism is activated, randomly selecting a mobile node from the neighbor nodes as an opportunistic relay (i.e., a temporary relay node), utilizing its mobility to wait for or create new routing opportunities.

[0056] The comprehensive route score is calculated using a weighted linear fusion algorithm, expressed as follows:

[0057]

[0058] Among them, S j P represents the comprehensive routing score of the j-th neighbor node. fwd P represents the forwarding priority score for the next-hop node output by the neural network. rel P represents the link reliability score output by the neural network. eff The energy efficiency score represents the output of the neural network; α, β, and γ are dynamic weight coefficients, and satisfy α+β+γ=1.

[0059] The dynamic weighting coefficients of the weighted linear fusion algorithm are adaptively adjusted based on the current node's energy level and network congestion. For example, when a node's remaining energy is below 20%, the system automatically increases the value of γ (e.g., setting γ=0.6) to prioritize low-energy paths, thereby extending network lifetime; while in emergency data transmission mode, the value of α is increased to prioritize forwarding efficiency.

[0060] S4: Packet forwarding and execution of fault detection and recovery mechanisms: Integrating packet lifetime management, fast link failure switching and adaptive retransmission strategies to ensure network communication reliability in the event of node failure or link interruption.

[0061] When forwarding data packets, if the energy value of the next-hop node is lower than the preset energy value, or if the link between the current node and the next-hop node is interrupted, the next-hop node selected by S3 is determined to be invalid, and the next-hop node selection is performed again.

[0062] like Figure 2 As shown, the specific execution flow of the fault detection and recovery mechanism is as follows:

[0063] S4.1: The current node sends a data packet to the selected next-hop node and starts an ACK timer to wait for an acknowledgment message.

[0064] S4.2: The node continuously monitors the acknowledgment status of the sent data packets and the quality of the link signal: it determines whether the current node has received an ACK from the next-hop node within the preset timeout period. If it receives an ACK, it determines that the data packet forwarding was successful and proceeds to S5; if it does not receive an ACK, it determines that the currently selected next-hop node is invalid (i.e., the currently selected next-hop node is recorded as the invalid node) or the link is interrupted and proceeds to S4.3.

[0065] Specifically, the preset timeout refers to the time to live (TTL) of a data packet. If the timeout occurs, the data packet is discarded to prevent network congestion.

[0066] S4.3: The current node reads the candidate node sequence from the local cache, removes invalid nodes, and obtains the updated candidate node sequence.

[0067] S4.4: Determine if there are other available nodes in the updated candidate node sequence. If so, select a new neighbor node as the next hop in descending order of comprehensive routing score, update the local routing table, and return to execute S4.1.

[0068] If not, it means that the current node has failed to change the next hop multiple times in a row (i.e., the candidate node sequence has been traversed but it is still unsuccessful). It is determined that the current node is in a routing hole or isolation state. At this time, the current node stops forwarding and sends a link failure notification message (NACK) to the next hop node, triggering the next hop node to execute the retransmission mechanism or re-make routing decisions.

[0069] Furthermore, the current node dynamically adjusts the packet retransmission strategy based on the link reliability with the selected next-hop node: when the link reliability score output by the neural network is low (e.g., below 0.5), the maximum number of retransmissions is automatically increased (e.g., from 3 to 5) and the retransmission waiting time is extended; conversely, when the link quality is good, the number of retransmissions is reduced to reduce the risk of network congestion.

[0070] S5: Repeat S4 until the data packet is successfully delivered to the aggregation node; each node maintains an experience replay buffer. During this process, the experience replay buffer of each node on the transmission path stores historical routing decision data; each routing decision data includes an eight-dimensional feature vector at the time of the decision, the final action selected for the next hop, and the result feedback of the action. The result feedback includes key performance indicators such as data packet delivery success rate, end-to-end latency, and energy consumption.

[0071] S6: Based on the historical routing decision data stored in the experience replay buffer, a batch of historical state-action-reward tuples (Mini-batch) is randomly sampled from the experience replay buffer periodically using the experience replay technique in the reinforcement learning mechanism. The neural network weight parameters are updated using the stochastic gradient descent optimizer (SGD) for the next round of adaptive routing decision-making, so as to eliminate the correlation between data samples, ensure that the stored historical experience can effectively guide the long-term optimization of the model, and improve the adaptability and accuracy of routing decisions.

[0072] Preferably, a backpropagation algorithm based on gradient descent is used to update the neural network weight parameters. Specifically, the target score is calculated based on actual feedback, and a loss function L (e.g., mean squared error function) is constructed to measure the deviation between the "predicted comprehensive route score" and the "actual comprehensive route score." The value of L is the mean squared error (MSE) between the comprehensive route score predicted by the neural network and the target reward value calculated based on the actual transmission results (e.g., success or delay). By calculating the gradient vector of the loss function with respect to the weight parameters of each layer of the neural network, the error is backpropagated from the output layer to the input layer using the chain rule. The connection weights and biases of neurons in each layer are fine-tuned along the gradient descent direction according to a preset learning rate to minimize the prediction error, enabling the model to more accurately evaluate neighboring nodes and adapt to long-term changes in the network environment. The expression for iteratively updating the model weight parameters along the gradient descent direction according to the preset learning rate is as follows:

[0073]

[0074] In the formula, This represents the updated model weight parameters. This represents the model weight parameters before the update. This represents the learning rate, which controls the step size for parameter updates. This represents the gradient of the loss function with respect to the current weight parameters, indicating the direction of the fastest error reduction.

[0075] Furthermore, the method of this invention follows dynamic topology adaptation. Specifically, for mobile nodes marked as moving (such as AUVs), a random waypoint movement model is used to simulate their movement: the mobile node randomly selects the next target location and moves towards the target point at a random speed. After arriving, it reselects a target, or periodically reselects a target point. All nodes need to periodically perform a neighbor discovery process, evaluate link quality based on underwater acoustic propagation characteristics, and update the list of neighbor nodes and the corresponding link status information. The neural network model can combine historical information such as node speed and direction, which reflect the node's movement trend, to make short-term predictions of topology change trends, thereby pre-adjusting the routing strategy.

[0076] Furthermore, this invention implements performance monitoring and feedback: the aggregation node collects and monitors network-wide or regional performance indicators in real time, including packet delivery success rate, end-to-end latency, energy consumption, and node survival rate; a performance evaluation system is established to comprehensively evaluate the effectiveness of the routing protocol; performance indicators are converted into evaluation signals and input into the neural networks of each node through global or local feedback mechanisms, influencing the online learning process of each node and forming a closed-loop optimization system of "decision-execution-evaluation-optimization"; routing policy parameters are dynamically adjusted based on feedback results to improve the system's adaptability and continuously enhance the overall adaptability and robustness of the routing protocol.

[0077] To achieve the aforementioned highly fault-tolerant and mobility-resistant underwater acoustic routing decision optimization method based on neural networks, this invention also proposes a highly fault-tolerant and mobility-resistant underwater acoustic routing decision optimization system based on neural networks, such as... Figure 3 As shown, the system architecture is logically divided into four mutually coordinating functional layers: data input and processing layer, feature extraction and decision-making layer, execution and optimization layer, and learning and feedback layer.

[0078] The data input and processing layer is responsible for the perception and collection of basic data, specifically including: environment perception module, neighbor status acquisition module, link quality detection module, and self-status detection module, providing multi-dimensional raw network status information for subsequent processing.

[0079] The feature extraction and decision layer comprises a feature extractor, a neural network decision model, and a decision fusion unit. The feature extractor processes the raw network state information, generating an eight-dimensional feature vector, which is then input into the core neural network decision model. The neural network decision model consists of an input layer (8 neurons), a hidden layer (16 neurons, using the ReLU activation function), and an output layer (3 neurons, using the Sigmoid activation function), outputting three decision metrics. The decision fusion unit performs a weighted calculation on the three decision metrics to obtain a comprehensive routing score.

[0080] The execution and optimization layer includes: an executor, a packet forwarding module, a fault tolerance mechanism execution module, and a performance monitoring module. Based on the comprehensive routing score, the executor selects the next hop and drives packet forwarding. Simultaneously, it triggers the fault tolerance mechanism execution module when necessary, and the performance monitoring module evaluates network transmission performance in real time.

[0081] The learning and feedback layer includes: an experience replay buffer, an online learning engine, and a model parameter update module. The experience replay buffer stores historical routing decision data, the online learning engine calculates the optimization gradient, and the model parameter update module performs model parameter updates and feeds the updated parameters back to the neural network model, thus forming a closed-loop adaptive optimization system.

[0082] This invention constructs a lightweight distributed neural network model that comprehensively utilizes multi-dimensional information such as node energy, link quality, mobility, and congestion status for online intelligent decision-making. It also embeds opportunistic routing, link prediction, and adaptive recovery mechanisms to efficiently cope with dynamic changes in underwater network topology and link uncertainties. This significantly reduces communication latency and extends network lifetime while ensuring high data delivery rates, providing an effective solution for reliable underwater acoustic communication in complex marine environments.

[0083] The above-described high-fault-tolerant and mobility-adaptive underwater acoustic sensor network routing system based on neural networks is merely one embodiment of the present invention. The present invention is applicable to any adaptive routing method employing distributed neural network decision-making, multi-dimensional state feature fusion, and online feedback optimization mechanisms. Any modifications, equivalent substitutions, and improvements made to node feature dimensions, neural network structure, routing decision weights, opportunistic routing triggering conditions, and online learning strategies within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0084] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A neural network-based high fault-tolerant and anti-mobility underwater routing decision optimization method, characterized in that, Includes the following steps: S1: Deploy underwater acoustic sensor network nodes, initialize network communication parameters, and neural network models within each non-convergence node; S2: Each node periodically updates its list of neighboring nodes and collects network state information to generate multi-dimensional feature vectors; The multidimensional feature vector includes: energy state, link reliability, node mobility, improved metric of the current node relative to the sink node, distance between nodes, packet priority, depth difference, and network congestion estimation; the improved metric is used to measure the distribution divergence of nodes in the underwater acoustic sensor network. S3: Adaptive routing decision: Input the multidimensional feature vector into the neural network model to obtain three decision indicators, including: forwarding priority score, reliability score and energy efficiency score; perform weighted fusion to obtain the comprehensive routing score of each neighbor node; the dynamic weight coefficients used for the weighted linear fusion of the three decision indicators are adaptively adjusted according to the energy level of the current node and the network congestion status. The neighboring nodes are sorted according to the comprehensive routing score to generate a candidate node sequence; a loop avoidance mechanism and an opportunistic routing mechanism are integrated to intelligently select the next-hop node; specifically: it is determined whether the highest value of the comprehensive routing score has reached the activation threshold. If so, the neighboring node with the highest comprehensive routing score is selected as the next-hop node, and the loop avoidance mechanism is applied to penalize the visited node; if not, there is no suitable next-hop node, the opportunistic routing mechanism is activated, and a mobile node is randomly selected from the neighboring nodes as an opportunistic relay. S4: Packet forwarding and execution of fault detection and recovery mechanism: Integrates packet lifetime management, fast link failure switching and adaptive retransmission strategy to ensure network communication reliability in the event of node failure or link interruption; When forwarding packets, if the energy value of the next hop node is lower than the preset energy value, or the link between the current node and the next hop node is interrupted, the next hop node selected by S3 is determined to be invalid, and the next hop node is reselected. The fault detection and recovery mechanism is implemented through the following sub-steps: S4.1: The current node sends a data packet to the selected next-hop node and starts an ACK timer to wait for an acknowledgment message; S4.2: Determine whether the current node receives an ACK from the next-hop node within the preset timeout period. If it does, the data packet is forwarded successfully, and proceed to S5. If it does not receive an ACK, the currently selected next-hop node is deemed to be invalid or the link is interrupted, and proceed to S4.

3. The preset timeout period refers to the lifespan of the data packet. If the timeout period expires, the data packet is discarded to prevent network congestion. S4.3: The current node reads the candidate node sequence from the local cache, removes invalid nodes, and obtains the updated candidate node sequence; S4.4: Determine if there are other available nodes in the updated candidate node sequence. If so, select a new neighbor node as the next hop node in descending order of comprehensive routing score, update the local routing table, and return to execute S4.

1. If not, the current node is determined to be in a routing hole or isolation state. At this time, the current node stops forwarding and sends a link failure notification message to the next hop node, triggering the next hop node to execute the retransmission mechanism or re-make routing decisions. The current node dynamically adjusts the retransmission strategy of data packets based on the link reliability between itself and the selected next-hop node: when the link reliability score output by the neural network is lower than the preset reliability threshold, the maximum number of retransmissions is automatically increased and the retransmission waiting time is extended; otherwise, the number of retransmissions is reduced to reduce the risk of network congestion. S5: Repeat S4 until the data packet is successfully delivered to the aggregation node; during this process, historical routing decision data is stored in the experience replay buffer of each node on the transmission path. S6: Based on the historical routing decision data, the neural network weight parameters are updated through reinforcement learning for the next round of adaptive routing decision-making.

2. The method of claim 1, wherein, In S1, the network communication parameters include: the node's initial energy, transmit power, receive power, idle power, and communication range. 3.The neural network based high fault-tolerant and anti-mobility underwater routing decision optimization method according to claim 1, wherein, Following dynamic topology adaptation, for nodes marked as moving, a random waypoint movement model is used to simulate their movement: the moving node randomly selects the next target location and moves towards the target point at a random speed. After arriving, it reselects the target, or periodically reselects the target point. All nodes need to periodically update their neighbor node list, evaluate link quality based on underwater acoustic propagation characteristics, and update link status information. The neural network model combines historical information on node movement trends to make short-term predictions of topology change trends, thereby pre-adjusting routing strategies. 4.The neural network based high fault-tolerant and anti-mobility underwater routing decision optimization method according to claim 1, wherein, In step S6, based on the historical routing decision data stored in the experience replay buffer, a batch of historical state-action-reward tuples are periodically randomly sampled from the experience replay buffer using the experience replay technique in the reinforcement learning mechanism, and the neural network weight parameters are updated using the backpropagation algorithm based on gradient descent.

5. A neural network based high fault-tolerant and anti-mobility underwater routing decision optimization system for implementing the neural network based high fault-tolerant and anti-mobility underwater routing decision optimization method of any one of claims 1-4, characterized in that, include: Data input and processing layer, feature extraction and decision-making layer, execution and optimization layer, learning and feedback layer; The data input and processing layer is responsible for the perception and collection of basic data, including: an environment perception module, a neighbor status collection module, a link quality detection module, and a self-status detection module, providing multi-dimensional raw network status information for subsequent processing; The feature extraction and decision layer includes: a feature extractor, a neural network decision model, and a decision fusion unit; the feature extractor processes the original network state information, generates an eight-dimensional feature vector, and inputs it into the core neural network decision model; the neural network decision model includes an input layer, a hidden layer, and an output layer, which output three decision indicators; the decision fusion unit performs a weighted calculation on the three decision indicators to obtain a comprehensive routing score. The execution and optimization layer includes: an executor, a packet forwarding module, a fault tolerance mechanism execution module, and a performance monitoring module. Based on the comprehensive routing score, the executor selects the next hop and drives packet forwarding, while triggering the fault tolerance mechanism execution module and evaluating network transmission performance in real time through the performance monitoring module. The learning and feedback layer comprises an experience replay buffer, an online learning engine, and a model parameter updating module; the experience replay buffer is used to store historical routing decision data, the online learning engine is used to calculate an optimization gradient, the model parameter updating module is used to perform neural network model parameter updating and feed back the updated parameters to the neural network model, thereby forming a closed-loop adaptive optimization system.

Images