Tailorable topology generation method based on bionic neural collaborative optimization mechanism
By generating the topology of an unmanned swarm communication network through a biomimetic neural collaborative optimization mechanism, the problems of inflexible node role division and redundant connections in existing methods are solved, and an adaptive hierarchical topology structure is achieved, which improves network performance and resource utilization efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-23
- Publication Date
- 2026-04-21
AI Technical Summary
Existing methods for generating topology in unmanned swarm communication networks have fixed weights when assigning node roles, which cannot be dynamically adjusted. This results in redundant connections and poor dynamic adaptability, affecting network performance and resource utilization efficiency.
A biomimetic neural collaborative optimization mechanism is adopted to generate an initial link graph through spatial situational awareness, generate weight vectors based on community division and node characteristics, dynamically adjust node roles, prune redundant links, and form an adaptive hierarchical topology structure.
It realizes a dynamically adaptive, multi-layered, customizable hierarchical topology under decentralized and infrastructure-free conditions, which improves global connectivity, local efficiency and energy efficiency, and reduces redundant connections and energy consumption.
Smart Images
Figure CN121907696A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of cluster communication technology, and more specifically to a topology generation method based on a biomimetic neural collaborative optimization mechanism. Background Technology
[0002] In unmanned swarm communication networks, dynamic node deployment and topology optimization are crucial for improving network performance and resource utilization efficiency. Traditional topology generation methods are typically based on static network models, making it difficult to adapt to dynamically changing environments and node states. Furthermore, these methods often fail to fully utilize the multidimensional characteristics of nodes, such as energy, centrality, and resources, leading to insufficient network performance optimization. In recent years, research on the hierarchical structure of biological communities has provided a new approach to solving this problem. The hierarchical structure in biological communities demonstrates efficient information transfer and resource utilization between different levels. This hierarchical structure not only improves the stability and adaptability of the ecosystem but also reduces unnecessary energy consumption. Therefore, applying the concept of biological community hierarchies to unmanned swarm communication networks holds promise for achieving more efficient and stable network topologies. However, existing topology generation methods based on biological community stratification have the following shortcomings: insufficient weight optimization, as existing methods use fixed weights when assigning node roles, which cannot be dynamically adjusted according to network status, resulting in limited optimization effect of the fitness function; redundant connection problem, as the generated topology contains a large number of redundant connections, increasing communication overhead and energy consumption, and reducing network efficiency; poor dynamic adaptability, as existing methods lack a real-time feedback mechanism for network performance, making it difficult to dynamically adjust the topology according to network status, resulting in poor adaptability.
[0003] Therefore, how to provide a topology generation method that can dynamically adjust according to the network is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0004] In view of the above problems, the present invention is proposed to provide a biomimetic neural collaborative optimization mechanism-based pruning topology generation method that overcomes or at least partially solves the above problems.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for generating customizable topologies based on a biomimetic neural collaborative optimization mechanism includes the following steps: S1: Obtain the communication link status between various network nodes through spatial situational awareness and generate an initial link diagram; S2: Perform community partitioning based on the initial link graph; S3: Within each community, generate a weight vector of node features based on the node status of each network node. S4: Calculate the node fitness based on the weight vector and node features, and take the node with the best fitness as the center node; S5: Classify other nodes in the current community based on the central node, and prune the links between intermediate nodes; summarize the pruning results of each community to obtain the final topology network.
[0006] Preferably, in step S1, a trained first neural network model is used for spatial situational awareness. The training steps of the first neural network model include: S11: Obtain the first training set, the training samples include RSSI values corresponding to a specific communication distance and a specific obstacle thickness; S12: Construct the first neural network. The input layer is a feature vector composed of communication distance and obstacle thickness. Multiple hidden layers are used to learn features from the feature vector. The predicted RSSI value is output through the output layer and activation layer. S13: Optimize the model according to the preset loss function.
[0007] Preferably, S2 specifically includes: using graph theory to divide the nodes into disconnected connected components, with each connected component forming an independent community; The connected components can be one or more.
[0008] Preferably, in step S3, the step of generating the weight vector includes: S31: Obtain node status data; S32: Identify the ecological stage based on the node status data and obtain the initial weight; S33: Obtain the weight adjustment amount through the second neural network model based on the initial weights and the node state data; S34: Obtain the final weight vector based on the initial weights and weight adjustment amounts.
[0009] Preferably, S32 includes: S321: Calculate the ecological index based on the node status data:
[0010] in, α , β , γ , δ These are the weight coefficients corresponding to the states of each node. Network density represents the ratio of the number of nodes to the area of the region. This represents the average energy of the nodes; The variance of the node energy; This represents the average value of the node centrality. S322: Determine the current ecological stage based on the ecological index and a preset threshold, and call the initial weight corresponding to the current ecology.
[0011] Preferably, the training steps of the second neural network model include: S331: Obtain the second training set, which includes average energy level, energy distribution variance, network density, average connectivity, stage encoding value, and performance trend indicators. S332: Construct a second neural network, input the training samples into the second neural network to obtain the initial adjustment amount; obtain the final weight adjustment amount by performing amplitude limiting processing on the initial adjustment amount.
[0012] Preferably, S34 includes:
[0013] Where λ is the neural network adjustment coefficient and μ is the historical memory coefficient; To preserve historical weights, the average of the optimal weights from multiple historical rounds is used.
[0014] Preferably, the node fitness is:
[0015] in, e For energy, c For centrality, For resources, Centrality per unit distance; , , and These are the adaptive weight coefficients corresponding to the features of each node.
[0016] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a pruning topology generation method based on a biomimetic neural collaborative optimization mechanism, which can dynamically and adaptively generate multi-layered pruning and self-healing hierarchical topologies as nodes move, energy dissipates and task granularity changes under the conditions of no center and no infrastructure, while taking into account global connectivity, local efficiency and energy consumption optimization. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0018] Figure 1The present invention provides a flowchart of a method for generating customizable topologies based on the hierarchical structure of biological communities; Figure 2 The ecological-neural weight optimization flowchart provided by this invention; Figure 3 A schematic diagram illustrating the node role division and initial topology generation results provided by this invention; Figure 4 This is a schematic diagram of the topology trimming result provided by the present invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] This invention discloses a method for generating customizable topologies based on a biomimetic neural collaborative optimization mechanism, comprising the following steps: S1: Obtain the communication link status between various network nodes through spatial situational awareness and generate an initial link diagram; S2: Perform community partitioning based on the initial link graph; S3: Within each community, generate a weight vector of node features based on the node status of each network node. S4: Calculate the node fitness based on the weight vector and node features, and take the node with the best fitness as the center node; S5: Classify other nodes in the current community based on the central node, and prune the links between intermediate nodes; summarize the pruning results of each community to obtain the final topology network.
[0021] The steps of this invention will now be described in detail: S1 is the spatial situational awareness process. Through the movement and detection of unmanned vehicles, electromagnetic signal strength (RSSI) and geographic obstacle information are collected from multiple points in space. Based on a deep learning model, a nonlinear mapping model from location and obstacles to signal strength is constructed to generate a high-precision spatial electromagnetic situation map.
[0022] To further implement the above technical solution and train the first neural network model for spatial situational awareness, the training steps for the first neural network model include: Obtain the first training set, which includes RSSI values corresponding to specific communication distances and obstacle thicknesses; construct the first neural network, with the input layer being a feature vector composed of communication distance and obstacle thickness, multiple hidden layers for feature learning of the feature vector, and the output layer and activation layer outputting the predicted RSSI value; optimize the model according to the preset loss function.
[0023] The specific steps for building the first training set and the model, i.e., training, are as follows: S11: Deploy an unmanned vehicle node in an underground enclosed space, equipped with a wireless ad hoc network communication module as a receiving node; in addition, set a coordinate origin and deploy a terminal computer, which is connected to a wireless ad hoc network communication module as a transmitting node.
[0024] Understandably, the deployed autonomous vehicle nodes have onboard computers with ROS interfaces and positioning capabilities, enabling them to publish relative location information in real time. The deployed wireless ad hoc network communication module has signal transmission, reception, and RSSI value acquisition functions.
[0025] S12: Control the unmanned vehicle to move to different positions, stop at each position and continuously collect the electromagnetic signal strength RSSI and the relative position information [x, y] of the unmanned vehicle for 10 seconds, manually measure the obstacle thickness δ between the unmanned vehicle and the terminal, process the RSSI data and merge and save it with the position information and obstacle thickness.
[0026] In one embodiment, S12 specifically includes the following steps: S121: Sort the n RSSI data R = {r1, r2, ..., rn} collected at a certain location in ascending order to obtain the sorted data RS; S122: Remove the first k and last k data points from the sorted dataset RS, and calculate the average of the remaining data. , where k = n / 10; Specifically, if n=100 and k=10, then remove the first 10 and last 10 data points from the sorted data RS, and calculate the average of the remaining 80 data points.
[0027] S123: Combine and save the average value of the processed RSSI data with the location distance information. This yields the training samples at the current location, where .
[0028] S13: Construct a feedforward neural network with 6 fully connected layers. The input layer is a standardized 2D feature vector [d, δ], there are 5 hidden layers, and the output layer has 1 neuron with linear activation, outputting the predicted RSSI value. The loss function uses mean squared error as the optimization objective.
[0029] S2 begins biomimetic hierarchical role mapping. The first step is to simulate the niche division principle of biological communities based on the aforementioned electromagnetic situation map, and to divide the communication network nodes into communities.
[0030] To further implement the above technical solution, a graph theory method is used to divide the nodes into disconnected connected components, with each connected component forming an independent community. It can be understood that there are one or more connected components. The specific steps are as follows: S21, building upon S1, allows each network node (m agents) in space to generate its own RSSI map, referencing the mutual perception and communication between individuals in a biological community, based on a trained RSSI prediction model. The existence of a direct communication link between two nodes is determined by the RSSI map and the positions of other agents, and a topological connection graph of all agents is generated accordingly. It is understandable that when determining whether there is a direct communication link between two nodes, an agent generates its own RSSI map and calculates the RSSI value of the other node's location based on the RSSI prediction model constructed by S1. In this embodiment, it is assumed that if the RSSI value is less than -65dBm, the connection can be established.
[0031] S22: Use graph theory to classify the nodes, dividing them into M distinct connected components. Each connected component Ck is a set containing multiple nodes that can be interconnected through direct or indirect communication links. Ultimately, this results in M disconnected connected components C = {C1, C2, ..., CM}.
[0032] S3 is an intermediate stage in the biomimetic hierarchical mapping. This invention uses weights to determine fitness, thereby influencing node classification and pruning. To achieve dynamic adaptive weight optimization, an ecological-neural collaborative feedback control loop is introduced. The macroscopic state of the network is perceived by calculating a global ecological index, driving the adjustment of basic weights. Simultaneously, a lightweight neural network is used to perform real-time, non-linear fine-tuning of the weights, and this is fused with historical optimal weights, ultimately achieving online self-correction of the node fitness function. Specific steps include: S31: Obtain node state data; The node state data used in this embodiment includes network density, average energy, energy variance, and average centrality.
[0033] The specific method for calculating network density is as follows:
[0034] Where N represents the number of nodes in a certain community divided in S2, and A represents the spatial area covered by the communication of that community.
[0035] Average energy and energy variance can be calculated from the remaining charge of the nodes, as follows:
[0036] Among them, E i This represents the remaining power of the i-th node.
[0037] The method for calculating average centrality is as follows:
[0038] Among them, deg( i ) represents the number of edges connected to the i-th node.
[0039] S32: Identify the ecological stage based on node status data and obtain the initial weights for different ecological stages; calculate the global ecological index based on the node status data, using the following formula:
[0040] Where α, β, γ, and δ are the weight coefficients corresponding to the state of each node, which are set to 0.3, 0.3, 0.2, and 0.2 in this embodiment. Network density represents the ratio of the number of nodes to the area of the region. This represents the average energy of the nodes; The variance of the node energy; This represents the average value of the node centrality.
[0041] Based on the calculated ecological index E(t), the current ecological stage of the network is determined. It is divided into three stages: pioneer stage, competition stage, and climax stage. For example, if E(t) < θ1, it is in the pioneer stage, and the initial weights are adjusted to [0.15, 0.45, 0.25, 0.15]; if θ1 ≤ E(t) < θ2, it is in the competition stage, and the initial weights are adjusted to [0.35, 0.25, 0.30, 0.10]; if E(t) ≥ θ2, it is in the climax stage, and the initial weights are adjusted to [0.40, 0.20, 0.25, 0.15]. Here, θ1 and θ2 are thresholds, taken as 0.3 and 0.7 respectively. It can be understood that these weight coefficients are ecological weights in subsequent fusion, reflecting the characteristics of the network at the current stage, and are static weights based on the network state.
[0042] S33: Obtain the weight adjustment amount through the second neural network model based on the initial weights and node state data; the training steps of the second neural network model include: S331: Obtain the second training set, which includes average energy level, energy distribution variance, network density, average connectivity, stage encoding value, and performance trend indicators. S332: Construct a second neural network, input the training samples into the second neural network to obtain the initial adjustment amount; obtain the final weight adjustment amount by performing amplitude limiting processing on the initial adjustment amount.
[0043] The second neural network is a feedforward neural network with three fully connected layers. The input layer has 6 neurons, receiving a standardized state vector; hidden layer 1 has 24 neurons using the ReLU activation function; hidden layer 2 has 24 neurons using the ReLU activation function; and the output layer has 4 neurons using the tanh activation function. The optimizer is the Adam algorithm, with adaptive learning rate adjustment. The loss function is the mean squared error (MSE).
[0044] A six-dimensional state vector is constructed as the input to the neural network, including: average energy level, energy distribution variance, network density, average connectivity, stage encoding value (pioneer = 0.0, competition = 0.5, climax = 1.0), and performance trend index. The performance trend index characterizes the direction and magnitude of changes in the overall network performance across adjacent iterations, reflecting the long-term impact of current weight adjustments on network performance. The specific calculation method is as follows:
[0045] Here, Favg(t) represents the average fitness Fi of all nodes in a connected community at time t. In the initial stage of the system, Fi can be calculated using the initial weights Weco.
[0046] The encoded state vector is input into the neural network to obtain the original weight adjustment amount ΔWraw. Then, it is processed by amplitude limiting: ΔWneural = ΔWraw × 0.2, to ensure that the weight adjustment amount is limited to the range of [-0.2, 0.2].
[0047] S34: Obtain the final weight vector based on the initial weights and weight adjustment amounts.
[0048] Finally, the initial weights and the neural network adjustment values are combined to obtain the final optimal weights. The formula is as follows:
[0049] In this system, the neural network adjustment coefficient λ = 0.2, and the history memory coefficient μ = 0.1. The history weight memory (Whistory) is the arithmetic mean of the best weights from the most recent five rounds. In the initial stage of the system, when the number of historical rounds is less than the preset number, a degenerate initialization strategy is adopted for the historical weights, i.e., the current initial ecological weight is used as the initial value of the historical weights, gradually transitioning to a multi-round averaging method as the number of running rounds increases. The running rounds referred to here are the processing cycles during which the system performs a complete collection of network status, weight updates, and topology adjustments according to a preset period.
[0050] In this embodiment, based on the optimized final weights, the fitness values of all nodes are recalculated. The optimal weight vector for the current round is added to the weight history record. If the number of history records exceeds a preset capacity, the oldest record is removed, maintaining the time window characteristic of the history record. A feedback mechanism based on network performance is established, and the weight optimization effect is evaluated by monitoring the following indicators: the trend of average network fitness change, the stability of role partitioning, the balance of energy consumption, and the quality of topological connectivity. Based on the performance evaluation results, the neural network parameters are updated periodically to achieve continuous adaptive optimization.
[0051] S4 and S5 represent the classification and pruning process. Classification maps communication network nodes into four heterogeneous roles: center, relay, intermediate, and endpoint. Based on the optimized role classification, a graph pruning algorithm is used to strategically remove redundant links between intermediate nodes while maintaining network connectivity, resulting in a minimized backbone communication topology that combines high energy efficiency and high reliability.
[0052] The fitness calculation formula is as follows:
[0053] in, e For energy, c For centrality, For resources, Centrality per unit distance; , , and These are the adaptive weight coefficients corresponding to the features of each node.
[0054] Specifically, for each connected component, i.e., community, the node with the highest fitness is selected as the central node; For each connected component, select the node that is connected to only one other node as the terminal node; For each connected component, select those nodes that are not directly connected to the central node and are not terminal nodes as relay nodes.
[0055] Mark all nodes that are not central nodes, terminal nodes, or relay nodes as intermediate nodes.
[0056] It is understandable that each connected component has only one central node, multiple intermediate nodes, relay nodes, and terminal nodes that meet the conditions.
[0057] Furthermore, the specific steps for cutting include: S41: For each connected component Ck, identify and remove direct communication links between intermediate nodes. First, identify redundant connections. For each pair of intermediate nodes, check if there is a direct communication link between them. If there is a direct communication link between two intermediate nodes, remove it. This removal operation follows the principle of minimum redundancy to ensure the efficiency of network connectivity.
[0058] like Figure 3 The image shown is in Figure 2 Topology pruning was performed on the basis of comparison. Figure 2 The topology connections in the code have been optimized by removing redundant links. This avoids network congestion caused by complex topologies, and the final result is as follows: Figure 4 As shown.
[0059] S42: In each connected component, preserve the communication links between the supernode, relay node, and terminal node. Verify the pruned communication topology to ensure that each node can communicate with other nodes through at least one path.
[0060] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0061] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for generating customizable topologies based on a biomimetic neural collaborative optimization mechanism, characterized in that, Includes the following steps: S1: Obtain the communication link status between various network nodes through spatial situational awareness and generate an initial link diagram; S2: Perform community partitioning based on the initial link graph; S3: Within each community, generate a weight vector of node features based on the node status of each network node. S4: Calculate the node fitness based on the weight vector and node features, and take the node with the best fitness as the center node; S5: Classify other nodes in the current community based on the central node, and prune the links between intermediate nodes; summarize the pruning results of each community to obtain the final topology network.
2. The method for topology generation based on a biomimetic neural collaborative optimization mechanism as described in claim 1, characterized in that, In step S1, a trained first neural network model is used for spatial situational awareness. The training steps of the first neural network model include: Obtain the first training set, which includes RSSI values corresponding to a specific communication distance and a specific obstacle thickness; The first neural network is constructed. The input layer is a feature vector composed of communication distance and obstacle thickness. Multiple hidden layers are used to learn features from the feature vector. The predicted RSSI value is output through the output layer and activation layer. The model is optimized based on a preset loss function.
3. The method for topology generation based on a biomimetic neural collaborative optimization mechanism as described in claim 1, characterized in that, S2 specifically includes: using graph theory to divide the nodes into disconnected connected components, with each connected component forming an independent community; The connected components can be one or more.
4. The method for topology generation based on a biomimetic neural collaborative optimization mechanism as described in claim 1, characterized in that, In step S3, the step of generating the weight vector includes: S31: Obtain node status data; S32: Identify the ecological stage based on the node status data and obtain the initial weight; S33: Obtain the weight adjustment amount through the second neural network model based on the initial weights and the node state data; S34: Obtain the final weight vector based on the initial weights and weight adjustment amounts.
5. The method for topology generation based on a biomimetic neural collaborative optimization mechanism as described in claim 4, characterized in that, S32 includes: S321: Calculate the ecological index based on the node status data: in, α , β , γ , δ These are the weight coefficients corresponding to the states of each node. Network density represents the ratio of the number of nodes to the area of the region. This represents the average energy of the nodes; The variance of the node energy; This represents the average value of the node centrality. S322: Determine the current ecological stage based on the ecological index and a preset threshold, and call the initial weight corresponding to the current ecology.
6. The method for topology generation based on a biomimetic neural collaborative optimization mechanism according to claim 4, characterized in that, The training steps for the second neural network model include: S331: Obtain the second training set, which includes average energy level, energy distribution variance, network density, average connectivity, stage encoding value, and performance trend indicators. S332: Construct a second neural network, input the training samples into the second neural network to obtain the initial adjustment amount; obtain the final weight adjustment amount by performing amplitude limiting processing on the initial adjustment amount.
7. The method for topology generation based on a biomimetic neural collaborative optimization mechanism as described in claim 4, characterized in that, S34 includes: in, λ represents the optimal weight for the current round, λ is the neural network adjustment coefficient, and μ is the historical memory coefficient. To preserve historical weights, the average of the optimal weights from multiple historical rounds is used.
8. A method for topology generation based on a biomimetic neural collaborative optimization mechanism as described in claim 1 or 4, characterized in that, The node fitness is: in, e For energy, c For centrality, For resources, Centrality per unit distance; , , and These are the adaptive weight coefficients corresponding to the features of each node.