Unmanned system intelligent networking method in complex structure environment

Abstract

The invention relates to an intelligent networking method for an unmanned system in a complex structure environment, which comprises the following steps of: constructing a channel attenuation prediction model, processing an environment map through rasterization, defining grid attributes, and establishing an attenuation model library in combination with a random forest algorithm and multiple parameters; and matching the key parameters by using the decision tree, outputting an RSSI predicted value, and generating a task initialization map. Adaptive networking is realized, and a layered reinforcement learning architecture is adopted: an execution layer processes an initial task map through a full connection layer, and outputs an optimal node deployment position; and the decision layer combines the RSSI value, the task emergency degree and the node state weight, outputs a layering result, optimizes the weight through deep reinforcement learning, and finally constructs an environment-task adaptive ad hoc network. The problem that existing cluster networking is insufficient in intelligence and poor in flexibility when facing a complex weak communication scene is solved, the networking stability and rationality are improved, the forwarding strategy can be adaptively adjusted, and the problem that when facing different tasks, the networking adjustment flexibility is low is solved.

Description

Technical Field

[0001] This invention relates to an unmanned swarm cooperative communication technology, and more particularly to an intelligent networking method for unmanned systems in complex structural environments. Background Technology

[0002] Unmanned swarms offer significant advantages in complex, enclosed environments, but this also places higher demands on their collaborative communication capabilities. On one hand, the swarm's self-organizing network must possess high flexibility, automatically planning network routes based on the swarm's spatial distribution and environmental conditions to ensure network stability. On the other hand, conventional self-organizing networks only consider the communication status of individual nodes and cannot adjust accordingly to task requirements, resulting in insufficient network flexibility. In complex, weak-connectivity scenarios, they are prone to task interruptions due to varying data transmission volumes. Therefore, a certain degree of task adaptability is required for the network.

[0003] The patent "A Dynamic Clustering Method for Airborne Unmanned Swarm Self-Organizing Networks" (CN118612814A) proposes a dynamic clustering method for airborne unmanned swarm self-organizing networks, belonging to the field of flight ad hoc network communication technology. It can set the cluster leader UAV based on a weighted sum of four indicators: the current remaining energy of a node, the ideal node degree difference, the average link holding time between neighboring nodes, and the average speed difference, thereby adjusting the hierarchical structure of the unmanned swarm nodes, improving the clustering efficiency of the ad hoc network, and enhancing robust security. The patent "A Multi-Agent Routing Algorithm for Unmanned Aerial Vehicle Self-Organizing Networks" (CN116113008A) proposes a multi-agent routing algorithm for unmanned aerial vehicle self-organizing networks. Each UAV node is considered an agent. When an agent generates a data packet destined for node 'd', a reinforcement learning model is activated to obtain routing actions and distribute the data. The agent receiving the data packet replies to the sending agent with an ACK packet according to the requirements of the reinforcement learning model, carrying reward information to assist the sending agent. The agent completes the reinforcement learning iteration and checks whether it is the destination node of the data packet. If it is not the destination node, the reinforcement learning algorithm is started to continue forwarding the data packet. The data packet eventually reaches the destination node after being forwarded hop by hop by several agents, completing the transmission of the data packet. Each agent completes the perception of the network environment by continuously interacting with other agents, so that when it is necessary to forward data packets, it can make the optimal routing decision through the trained reinforcement learning model. The patent "A Method for Constructing an Aircraft Cluster Self-Organizing Network Based on Optoelectronic Hybrid Link" (CN110061772A) realizes the networking of several aircraft clusters. The radio frequency link is only used to transmit relevant control information, which has the characteristics of small transmission bandwidth requirement, long transmission distance and high network reliability. The free-space optical link utilizes information from other nodes obtained through the radio frequency self-organizing network to guide it through the initial coarse aiming process of the ATP system. Then, it uses the ATP system itself to achieve precise aiming, forming a free-space optical transmission network. The patent "Intelligent Multi-hop Routing Method for UAV Swarm Networks Based on Multi-Agent Cooperation" (CN114499648A), compared to traditional Q-learning, employs a cooperative multi-agent value decomposition network method to calculate the joint value function to update network evaluation parameters, effectively reducing communication latency for UAVs in complex and changing environments.

[0004] Clearly, while the aforementioned technologies offer solutions to the shortcomings of multi-unit unmanned cluster networking, the task environments described are mostly open environments where obstacles have negligible impact on communication. When the task scenario shifts to a complex, weak-connectivity scenario, relying solely on assessing the state of the unmanned cluster nodes is insufficient, leading to robustness and connectivity issues in the constructed self-organizing network structure. Furthermore, these technologies, focusing on resolving the network's own state, struggle to respond to different task types. In complex, weak-connectivity environments, this problem is amplified, impacting task success rates and making them unsuitable for establishing collaborative communication networks in such scenarios. Summary of the Invention

[0005] To address the shortcomings of existing cluster networking in terms of insufficient intelligence and poor flexibility when facing complex weak connectivity scenarios, a novel intelligent networking method for unmanned systems in complex structural environments is proposed. This method implements a dynamic loading mechanism for path loss models that adapts to the environment, and solves the technical challenges by introducing a task-adaptive self-organizing network structure generation technique based on multi-agent reinforcement learning.

[0006] The technical solution of this invention is as follows:

[0007] A method for intelligent networking of unmanned systems in complex structural environments includes the following steps:

[0008] Phase 1: Channel Attenuation Prediction Model Design. First, the environmental map is rasterized. Based on this, and according to the obstacle type information from the sensing system, attributes are defined for each raster unit, and a corresponding data structure is designed to store these attribute parameters. Then, based on a database of communication parameter models for different scene materials, a random forest algorithm is used to construct a channel attenuation prediction method. Following this, based on signal attenuation model theory, a relationship model is established between communication signal attenuation and key parameters such as the distance between nodes, obstacle thickness, number of obstacles, environmental fading coefficient, and dielectric constant. An attenuation model library is then established to achieve communication semantic storage for different scenes and media materials.

[0009] Based on the established random forest RSSI prediction model, each branch of the decision tree corresponds to a key parameter of the model. The model adaptively matches the key parameters in the model relationship with the communication parameters of the obstacles in the environment to determine the most practical attenuation model and key coefficients in the current scenario. The model outputs the RSSI prediction value at the receiving node, thereby constructing the task initialization map.

[0010] Phase Two, Adaptive Network Design: This phase includes two layers: environment adaptation and task adaptation. A hierarchical reinforcement learning architecture, comprising a decision layer and an execution layer, is constructed using multiple fully connected layers to enable online learning and adaptation of the network to changes in tasks and the environment. The algorithm initializes the task map as state, embedding it into the execution layer network. It then constructs a node hierarchical allocation method based on the actual communication environment, outputting the optimal deployment location of nodes in the current environment, thus completing node deployment while maintaining the original topology. The decision layer determines whether to update the overall topology based on the received signal strength values ​​within the site fed back by the communication components of each sub-cluster and the urgency of the current communication task. In the decision layer, the energy, centrality, speed similarity, current communication state, and current execution task level of each node are used as weights to describe the node state. The node state in the current environment is used as auxiliary input to construct the input quantities for the decision layer. After fitting by the decision layer neural network, the node hierarchical results are output. Finally, deep reinforcement learning methods are used to learn the optimal node weight allocation scheme, thereby constructing an environment-task adaptive self-organizing network.

[0011] Furthermore, the process of constructing the channel attenuation prediction model using the random forest algorithm is as follows:

[0012] Step 1: Based on the selected signal attenuation model theoretical formula, collect geometric propagation data, including distance, thickness of the penetrating obstacle, and material data of the penetrating obstacle;

[0013] Step 2: Construct a path-RSSI value dataset, where the path includes the location of the transmitting / receiving nodes, the transmitted signal strength, the transmitted signal frequency, the interval distance, and the obstacle penetration attribute; the RSSI value is used as a label and read from the receiving node.

[0014] Step 3: Use the random forest regression model as the predictor, divide the dataset and validation set according to the proportion, train the model according to the one-hot encoding method, and obtain the random forest RSSI prediction model.

[0015] Furthermore, each grid cell in Phase 1 has its attributes defined, including location information, physical attributes, and electromagnetic propagation attributes.

[0016] Furthermore, signal attenuation model theories include free space attenuation model, Lee indoor attenuation model, and small-scale fading model.

[0017] Preferably, in the random forest regression model of step 3, the number of decision trees is 300, the maximum depth is 10, and the minimum number of splits is 2.

[0018] The beneficial effects of this invention are as follows:

[0019] Faced with complex, enclosed spatial structures, it can proactively adapt to environmental changes and adjust its network topology to ensure reliable point-to-point communication between nodes in different environments, thus improving network stability and rationality. Furthermore, the prediction model is largely consistent with the actual communication signal distribution model in the environment. Simultaneously, it can adaptively adjust forwarding strategies to meet different levels of task requirements, addressing the issue of low flexibility in network adjustments when dealing with various tasks. Attached Figure Description

[0020] Figure 1 This is a diagram of the intelligent networking method in a complex weak connectivity environment according to the present invention;

[0021] Figure 2 This is a diagram of the environment-task adaptive networking method of the present invention;

[0022] Figure 3 This is a diagram showing the results of an implementation example of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0024] Complex weak-connectivity intelligent networking technology establishes a complex dynamic model of multi-person cooperative node distance and communication radius, as well as the task environment, to realize the task-environment adaptive multi-person cooperative communication topology evolution mechanism, such as... Figure 1 As shown.

[0025] The steps of complex weak communication intelligent networking technology can be broken down into two parts: channel attenuation prediction model design and adaptive networking.

[0026] Phase 1: Channel Attenuation Prediction Model Design. First, the environmental map is rasterized. Based on this, obstacle type information (e.g., concrete, metal walls) from a perception system including LiDAR and vision modules is used to define attributes for each raster cell, including location information, physical properties, and electromagnetic propagation properties. Corresponding data structures are designed to store these attribute parameters. Next, based on a database of communication parameter models for different scene materials, a random forest algorithm is used to construct a channel attenuation prediction method. Then, based on indoor and outdoor attenuation theoretical models, mainly including free space attenuation models and Lee indoor attenuation models, a relationship model is established between communication signal attenuation and key parameters such as distance between nodes, obstacle thickness, number of obstacles, environmental fading coefficient, and dielectric constant. An attenuation model library is built to achieve communication semantic storage for different scenes and media materials.

[0027] The process of constructing the channel attenuation prediction model using the random forest algorithm is as follows:

[0028] Step 1: Based on the selected signal attenuation model theoretical formula (free space attenuation model, Lee indoor attenuation model, small-scale fading model), collect geometric propagation data, including distance, thickness of penetrating obstacles, and material data of penetrating obstacles;

[0029] Step 2: Construct a path-RSSI (Received Signal Strength Indicator) dataset, where the path includes the location of the transmitting / receiving node, the transmitted signal strength, the transmitted signal frequency, the spacing distance, and the obstacle penetration attribute; the RSSI value is used as a label and read from the receiving node.

[0030] Step 3: Use the random forest regression model as the predictor, divide the dataset and validation set in an 8:2 ratio, with 300 decision trees, a maximum depth of 10, and a minimum number of splits of 2. Train the model using one-hot encoding to obtain the random forest RSSI prediction model.

[0031] Based on the established random forest RSSI prediction model, each branch of the decision tree corresponds to a key parameter of the model. The model adaptively matches the key parameters in the model relationship with the communication parameters of the obstacles in the environment to determine the most practical attenuation model and key coefficients in the current scenario. The model outputs the RSSI prediction value at the receiving node, thereby constructing the task initialization map.

[0032] Phase Two, Adaptive Network Design, comprises two layers: environment adaptation and task adaptation. A hierarchical reinforcement learning architecture, consisting of a decision layer and an execution layer, is constructed using multiple fully connected layers to enable the network to learn and adapt online to changes in tasks and the environment. Figure 2 As shown, the algorithm initializes the task map as the state, embeds it into the execution layer network, constructs a node hierarchical allocation method based on the actual communication environment, and outputs the optimal deployment position of nodes in the current environment, completing node deployment without changing the original topology. The decision layer decides whether to update the overall topology based on the received signal strength values ​​in the field fed back by the communication components of each sub-cluster and the urgency of the current communication task. In the decision layer, the energy, centrality, speed similarity, current communication state, and current execution task level of each node are used as weights to describe the node state. The node state in the current environment is used as an auxiliary input to construct the input of the decision layer. After fitting by the decision layer neural network, the node hierarchical result is output. Then, deep reinforcement learning is used to learn the optimal node weight allocation scheme, thereby constructing an environment-task adaptive self-organizing network.

[0033] In a real-world environment, node deployment strategies and communication status verification are performed. One node is connected to a computer as a base station node to obtain communication status feedback. The remaining nodes are placed on a mobile, unmanned platform and move along a predetermined route according to the generated deployment strategy. Figure 3 As shown, the results indicate that the overall trend is known, and the observed fluctuations are mainly attributed to the impact of unpredictable measurement content and personnel movement on the site environment. Simultaneously, the task completion performance was observed, showing good performance in terms of latency, transmission rate, and packet loss rate, with results of 30.137 ms, 5.36 Mbits / s, and 20.3% respectively in multi-hop transmission tasks.

[0034] The above-described embodiments are merely one implementation of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for intelligent networking of unmanned systems in complex structural environments, characterized in that, Includes the following steps: Phase 1, Channel Attenuation Prediction Model Design: First, the environmental map is rasterized. Based on this, according to the obstacle type information fed back by the perception system, the attributes of each raster unit are defined, and the corresponding data structure is designed to store these attribute parameters. Then, based on the communication parameter model database of different scenario materials, a channel attenuation prediction method is constructed using the random forest algorithm. Then, based on the signal attenuation model theory, a relationship model is established between communication signal attenuation and key parameters such as the distance between nodes, obstacle thickness, number of obstacles, fading coefficient of the environment, and dielectric constant. An attenuation model library is established to realize the storage of communication semantics for different scenarios and different media materials. Based on the established random forest RSSI prediction model, each branch of the decision tree corresponds to a key parameter of the model. The model adaptively matches the key parameters in the model relationship with the communication parameters of the obstacles in the environment to determine the most practical attenuation model and key coefficients in the current scenario. The model outputs the RSSI prediction value at the receiving node, thereby constructing the task initialization map. Phase Two, Adaptive Network Design: This phase includes two layers: environment adaptation and task adaptation. A multi-layered fully connected architecture is used to construct a hierarchical reinforcement learning architecture comprising a decision layer and an execution layer. This architecture enables the network to learn and adapt online to changes in tasks and the environment. The algorithm initializes the task map as state and embeds it into the execution layer network. It then constructs a node hierarchical allocation method based on the actual communication environment, outputting the optimal deployment location of nodes in the current environment, thus completing node deployment while maintaining the original topology. The decision layer determines whether to update the overall topology based on the received signal strength values ​​within the site fed back by the communication components of each sub-cluster and the urgency of the current communication task. In the decision layer, the energy, centrality, speed similarity, current communication state, and current execution task level of each node are used as weights to describe the node state. The node state in the current environment is used as an auxiliary input to construct the input of the decision layer. After fitting by the decision layer neural network, the node hierarchical result is output. Finally, deep reinforcement learning methods are used to learn the optimal node weight allocation scheme, thereby constructing an environment-task adaptive self-organizing network.

2. The intelligent networking method for unmanned systems in complex structural environments according to claim 1, characterized in that, The process of constructing a channel attenuation prediction model using the random forest algorithm is as follows: Step 1: Based on the selected signal attenuation model theoretical formula, collect geometric propagation data, including distance, thickness of the penetrating obstacle, and material data of the penetrating obstacle; Step 2: Construct a path-RSSI value dataset, where the path includes the location of the transmitting / receiving nodes, the transmitted signal strength, the transmitted signal frequency, the interval distance, and the obstacle penetration attributes; The RSSI value is used as a tag and read from the receiving node; Step 3: Use the random forest regression model as the predictor, divide the dataset and validation set according to the proportion, train the model according to the one-hot encoding method, and obtain the random forest RSSI prediction model.

3. The intelligent networking method for unmanned systems in complex structural environments according to claim 1, characterized in that, In Phase 1, each grid cell has its attributes defined, including location information, physical attributes, and electromagnetic propagation attributes.

4. The intelligent networking method for unmanned systems in complex structural environments according to claim 1, characterized in that, Signal attenuation model theory includes free space attenuation model, Lee indoor attenuation model, and small-scale fading model.

5. The intelligent networking method for unmanned systems in complex structural environments according to claim 2, characterized in that, In the random forest regression model of step 3, the number of decision trees is 300, the maximum depth is 10, and the minimum number of splits is 2.

Images