Unmanned aerial vehicle cooperative detection task execution method and device based on cross-level attention, equipment and medium

Abstract

The application discloses a method and device for cooperative detection task execution of a UAV based on cross-level attention, equipment and a medium, relates to the technical field of cooperative detection of a UAV, and comprises the following steps: acquiring multi-source heterogeneous sensing data of each UAV; performing intra-modal self-attention processing, cross-modal cross-attention processing and fusion on the multi-source heterogeneous sensing data, performing multi-hop transmission and aggregated updating on the fused features; generating an adaptive behavior tree by using a strategy network; controlling each UAV to perform a detection action by using the adaptive behavior tree, and obtaining a reward signal; evaluating the value of the detection action, obtaining an action value, and constructing a generalized advantage function; updating the parameters of the strategy network by using the generalized advantage function until a termination condition is met, and obtaining a target control instruction; and controlling each UAV to perform a cooperative detection task by using the target control instruction, improving the fusion capability of multi-source heterogeneous information, and improving the cooperative detection task execution efficiency in a high-dimensional action space.

Description

Technical Field

[0001] This invention relates to the field of UAV collaborative detection technology, and in particular to a method, apparatus, equipment and medium for performing UAV collaborative detection tasks based on cross-level attention. Background Technology

[0002] Multi-UAV collaborative sensing is an important method for utilizing multiple UAVs to jointly acquire intelligence and monitor the situation. With the development of sensing technology, a single UAV can carry different types of sensing devices, such as radar, infrared sensors, electronic support measures reconnaissance units, and laser rangefinders. The output data from different sensing devices differ significantly in format, dimension, and semantic attributes. When processing this data using methods such as direct feature concatenation or weighted averaging, it is difficult to adaptively select highly correlated sensor information based on the detection target and mission scenario. This can easily lead to the fused features containing a lot of redundant information and affect their discriminative ability.

[0003] In multi-drone communication, communication bandwidth is usually limited, and the connectivity of the communication network changes with the drone's location and mission status. When using a fully connected communication method, communication resource consumption increases rapidly with the number of drones; when using a fixed communication topology, it is difficult to adjust the communication targets according to the changing value of the information carried by different drones, which easily leads to low-value information occupying communication bandwidth and reducing the efficiency of information exchange between multiple drones.

[0004] In multi-UAV action decision-making, cooperative detection missions need to handle path planning and sensor scheduling simultaneously. There are many types of action parameters, and the action space expands with the increase in the number of UAVs, sensors, and action parameters. When directly outputting discrete action sequences or low-dimensional continuous action parameters, it is difficult to form hierarchical joint decisions for navigation control and sensor scheduling. When using a pre-fixed behavior tree, the node types and topology of the behavior tree cannot be adaptively adjusted according to the mission scenario.

[0005] Existing attention mechanisms are typically applied to only one level of sensor feature selection, multi-agent communication object selection, or action strategy optimization. When multi-sensor fusion, communication scheduling between UAVs, and detection action decision-making are processed independently, there is a lack of unified information filtering and transmission relationships between different levels. This can easily lead to semantic loss during cross-level information transmission and limit the overall effectiveness of multi-UAV collaborative detection.

[0006] As can be seen from the above, how to achieve cross-level collaborative processing of multi-source heterogeneous sensor data, communication information between UAVs, and UAV detection actions, improve the fusion capability of multi-source heterogeneous information, and improve the execution efficiency of collaborative detection tasks in high-dimensional action space are problems to be solved in this field. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for executing UAV collaborative detection tasks based on cross-level attention. This method enables cross-level collaborative processing of multi-source heterogeneous sensor data, communication information between UAVs, and UAV detection actions, improving the fusion capability of multi-source heterogeneous information and increasing the efficiency of collaborative detection task execution in high-dimensional action spaces. The specific solution is as follows: Firstly, this application discloses a method for performing UAV cooperative detection tasks based on cross-level attention, including: Acquire multi-source heterogeneous sensor data from various drones; The multi-source heterogeneous sensing data is subjected to feature encoding, intra-modal self-attention processing, cross-modal cross-attention processing and fusion to obtain the fused features of each UAV. The fused features are then subjected to multi-hop propagation and aggregation update to obtain the global feature vector of each UAV. Based on the fused features and the global feature vector, a local observation time series is constructed, and an adaptive behavior tree is generated using the policy network and the local observation time series. Using the adaptive behavior tree, control commands for each UAV are determined, and based on the control commands, each UAV is controlled to perform detection actions to obtain reward signals; The value of the detection actions performed by the UAV is evaluated to obtain the action value. Based on the reward signal and the action value, a generalized advantage function is constructed. Determine whether the preset termination condition is met; If the preset termination condition is not met, the generalized dominance function is used to update the parameters of the policy network, and the process jumps to the step of obtaining multi-source heterogeneous sensor data of each UAV until the preset termination condition is met and the target control command of the UAV is obtained. Using the target control commands, each UAV is controlled to perform a collaborative detection mission.

[0008] Optionally, acquiring the multi-source heterogeneous sensor data of each UAV includes: Obtain multi-source heterogeneous sensing data from each UAV from the multi-source heterogeneous sensing layer; The multi-source heterogeneous sensing layer includes different types of heterogeneous sensing devices; The different types of heterogeneous sensing devices include radar for acquiring in-phase or orthogonal complex data, infrared sensing devices for acquiring thermal radiation imaging, electronic support measures reconnaissance units for acquiring pulse descriptor sequences, and laser ranging devices for acquiring range sequences.

[0009] Optionally, the multi-source heterogeneous sensing data is subjected to feature encoding, intra-modal self-attention processing, cross-modal cross-attention processing, and fusion to obtain the fused features of each UAV, including: By using type-aware encoders corresponding to different types of sensors, feature encoding is performed on multi-source heterogeneous sensor data to obtain the encoded features of various sensors. By using an intramodal self-attention module, the encoded features are processed by intramodal self-attention to obtain intramodal features of various sensors; By utilizing a cross-modal cross-attention module, the intra-modal features are processed and fused through cross-modal cross-attention to obtain the fused features of each UAV.

[0010] Optionally, the fused features are subjected to multi-hop propagation and aggregation update to obtain the global feature vectors of each UAV, including: Based on the communication relationships between the drones, construct communication connection edges between the drones; The attention weights are determined using a graph attention network and based on the fused features and the communication connection edges. The attention weights are filtered based on a preset threshold to obtain the filtered attention weights; A dynamic communication topology for the UAV is constructed using the communication connection edges corresponding to the selected attention weights. Based on the dynamic communication topology, the fused features are subjected to multi-hop transmission and aggregation update to obtain the global feature vector of each UAV.

[0011] Optionally, an adaptive behavior tree is generated using the policy network and the local observation time series, including: The local observation time series is input into the policy network to obtain the behavior tree structure parameters; The behavior tree structure parameters include the root node type probability distribution, the number of child nodes, the type probability distribution of each child node, and the parameter values ​​of each atomic action node. The behavior tree structure parameters are parsed to generate adaptive behavior trees for each UAV.

[0012] Optionally, the adaptive behavior tree is used to determine the control commands for each UAV, and based on the control commands, each UAV is controlled to perform detection actions to obtain reward signals, including: The adaptive behavior tree is input to the tree executor so that the tree executor recursively executes each node in the adaptive behavior tree according to the topology of the adaptive behavior tree and outputs control commands for each UAV. The control commands include path planning commands and sensor scheduling commands; Based on the path planning instructions and the sensor scheduling instructions, each UAV is controlled to perform detection actions and obtain reward signals.

[0013] Optionally, the value assessment of the detection actions performed by the UAV to obtain the action value includes: Obtain the global status information of the drone; The global status information includes the position information, flight attitude information, and sensor operating status information of each UAV; The global state information is input into the evaluation network so that the evaluation network can be used to evaluate the value of the detection actions performed by each UAV and obtain the action value.

[0014] Secondly, this application discloses a UAV cooperative detection mission execution device based on cross-level attention, comprising: The data acquisition module is used to acquire multi-source heterogeneous sensor data from each UAV. The data processing module is used to perform feature encoding, intra-modal self-attention processing, cross-modal cross-attention processing and fusion on the multi-source heterogeneous sensor data to obtain the fused features of each UAV, and to perform multi-hop transfer and aggregation update on the fused features to obtain the global feature vector of each UAV. An adaptive behavior tree generation module is used to construct a local observation time series sequence based on the fused features and the global feature vector, and generate an adaptive behavior tree using a policy network and the local observation time series sequence. The reward signal determination module is used to determine the control commands for each UAV using the adaptive behavior tree, and control each UAV to perform detection actions based on the control commands to obtain a reward signal. The function construction module is used to evaluate the value of the detection actions performed by the UAV, obtain the action value, and construct a generalized advantage function based on the reward signal and the action value. The judgment module is used to determine whether the preset termination condition is met. The loop module is used to update the parameters of the policy network using the generalized dominance function if the preset termination condition is not met, and then jump to the step of acquiring the multi-source heterogeneous sensor data of each UAV until the preset termination condition is met and the target control command of the UAV is obtained. The task execution module is used to control each UAV to perform a collaborative detection task using the target control commands.

[0015] Thirdly, this application discloses an electronic device, including: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned method for performing UAV collaborative detection tasks based on cross-level attention.

[0016] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed method for executing a UAV cooperative detection task based on cross-level attention.

[0017] Beneficial Effects: This application provides a method for UAV collaborative detection task execution based on cross-level attention, including acquiring multi-source heterogeneous sensor data from each UAV, thus improving the completeness of the representation of heterogeneous sensing information collected by each UAV; performing feature encoding, intra-modal self-attention processing, cross-modal cross-attention processing, and fusion on the multi-source heterogeneous sensor data to obtain fused features of each UAV; performing multi-hop propagation and aggregation updates on the fused features to obtain global feature vectors of each UAV, thus improving the fusion degree of multi-source heterogeneous sensor information and enhancing the aggregation capability of each UAV for multi-hop propagated information; constructing a local observation time series based on the fused features and the global feature vectors, and generating an adaptive behavior tree using a policy network and the local observation time series; thus improving the temporal utilization of the fused features and global feature vectors and enhancing the adaptability of the behavior tree generation; using the adaptive behavior tree, determining the control commands for each UAV, and controlling each UAV to perform detection actions based on the control commands, thereby obtaining... The reward signal improves the matching between control commands and the adaptive behavior tree, and enhances the feedback capability of the detection action execution process. The value of the detection actions performed by the UAV is evaluated to obtain the action value. Based on the reward signal and the action value, a generalized dominance function is constructed. This improves the completeness of the evaluation of the detection action effect and enhances the ability of the generalized dominance function to represent the quality of actions. It determines whether the current condition meets the preset termination condition, avoiding invalid iterations after the preset termination condition has been met. If the preset termination condition is not met, the generalized dominance function is used to update the parameters of the policy network, and the process jumps to the step of acquiring multi-source heterogeneous sensor data from each UAV until the preset termination condition is met, obtaining the target control command for the UAV. This improves the matching degree between the policy network parameters and the iterative feedback, ensuring that the loop process stops when the preset termination condition is met. The target control command is used to control each UAV to perform cooperative detection tasks, thus improving the consistency of the control process when each UAV performs cooperative detection tasks. Attached Figure Description

[0018] 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.

[0019] Figure 1 This application discloses a flowchart of a UAV collaborative detection task execution method based on cross-level attention. Figure 2 This application discloses an architecture diagram of a UAV collaborative detection mission execution system based on cross-level attention. Figure 3 This is a schematic diagram of an HSCA module structure disclosed in this application; Figure 4 This is a schematic diagram of a DGAC module structure disclosed in this application; Figure 5 This is a flowchart illustrating the execution of a UAV collaborative detection task based on cross-level attention, as disclosed in this application. Figure 6 This is a schematic diagram of the structure of a UAV collaborative detection mission execution device based on cross-level attention disclosed in this application; Figure 7 This is a schematic diagram of the structure of an electronic device disclosed in this application. Detailed Implementation

[0020] 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.

[0021] Multi-UAV collaborative sensing is an important method for using multiple UAVs to jointly acquire intelligence and monitor the situation. With the development of sensing technology, a single UAV can carry different types of sensing devices, such as radar, infrared sensors, electronic support measures (ESR) reconnaissance units, and laser rangefinders. The formats, dimensions, and semantic attributes of the output data from different sensing devices vary significantly. When processing this data using methods such as direct feature concatenation or weighted averaging, it is difficult to adaptively select highly correlated sensor information based on the detection target and mission scenario. This can easily lead to the fused features containing a lot of redundant information, affecting their discriminative ability. Regarding communication between multiple UAVs, communication bandwidth is usually limited, and the connectivity of the communication network changes with the UAV's position and mission status. When using a fully connected communication method, communication resource consumption increases rapidly with the number of UAVs; when using a fixed communication topology, it is difficult to adjust the communication targets according to the changing value of the information carried by different UAVs, easily allowing low-value information to occupy communication bandwidth and reducing the efficiency of information exchange between multiple UAVs. In terms of multi-UAV action decision-making, collaborative detection tasks require simultaneous processing of path planning and sensor scheduling. There are many types of action parameters, and the action space expands with the number of UAVs, sensors, and action parameters. When directly outputting discrete action sequences or low-dimensional continuous action parameters, it is difficult to form hierarchical joint decisions for navigation control and sensor scheduling. When using pre-fixed behavior trees, the node types and topology of the behavior trees cannot be adaptively adjusted according to the task scenario. Existing attention mechanisms are usually only applied to one level of sensor feature selection, multi-agent communication object selection, or action strategy optimization. When multi-sensor fusion, communication scheduling between UAVs, and detection action decision-making are processed independently, there is a lack of unified information filtering and transmission relationships between levels, which easily leads to semantic loss during cross-level information transmission and limits the overall effect of multi-UAV collaborative detection. As can be seen from the above, how to achieve cross-level collaborative processing of multi-source heterogeneous sensor data, communication information between UAVs, and UAV detection actions, improve the fusion capability of multi-source heterogeneous information, and improve the execution efficiency of collaborative detection tasks in high-dimensional action spaces are problems to be solved in this field.

[0022] See Figure 1 As shown, this embodiment of the invention discloses a method for executing UAV cooperative detection tasks based on cross-level attention, which may specifically include: Step S11: Acquire multi-source heterogeneous sensor data from each UAV.

[0023] In this embodiment, multi-source heterogeneous sensing data of each UAV is acquired from the multi-source heterogeneous sensing layer; the multi-source heterogeneous sensing layer includes different types of heterogeneous sensing devices; the different types of heterogeneous sensing devices include radar for acquiring in-phase or orthogonal complex data, infrared sensing device for acquiring thermal radiation imaging, electronic support measures reconnaissance unit for acquiring pulse descriptor sequence, and laser ranging device for acquiring range sequence.

[0024] This application proposes a UAV cooperative detection mission execution system based on cross-level attention, and the system architecture diagram is shown below. Figure 2 As shown, this application relies on the CTDE (Centralized Training with Decentralized Execution) framework, which introduces attention mechanisms at three levels: sensor level, agent level, and action level. At the same time, it builds the internal correlation logic between the three levels of attention mechanisms to complete the end-to-end adaptive information filtering and transmission from multi-source heterogeneous sensor raw data to multi-UAV collaborative control commands.

[0025] from Figure 2 As can be seen from the architecture, the system contains three core functional layers: a multi-source heterogeneous sensor layer, a cross-level attention collaborative decision engine as the core module, and a multi-UAV execution layer. The whole system relies on a centralized training and decentralized execution paradigm to complete the training and operation process.

[0026] For the multi-source heterogeneous sensor layer: a radar carried by a single UAV can output I / Q (In-phase / Quadrature) complex data with a dimension of 128×128×3; an infrared sensor can output a thermal radiation image with a dimension of 256×256×1; an ESM (Electronic Support Measures) reconnaissance unit can output a pulse descriptor sequence with a dimension of 512×1; and a laser rangefinder can output a range sequence with a dimension of 64×1. Each heterogeneous sensing device operates autonomously and collects the raw detection information at the corresponding moment.

[0027] Step S12: Perform feature encoding, intra-modal self-attention processing, cross-modal cross-attention processing, and fusion on the multi-source heterogeneous sensing data to obtain the fused features of each UAV. Perform multi-hop transfer and aggregation update on the fused features to obtain the global feature vector of each UAV.

[0028] In this embodiment, the multi-source heterogeneous sensor data is subjected to feature encoding, intra-modal self-attention processing, cross-modal cross-attention processing, and fusion to obtain fused features for each UAV. This includes: using type-aware encoders corresponding to different types of sensors to encode the multi-source heterogeneous sensor data, obtaining coded features for each type of sensor; using an intra-modal self-attention module to process the coded features, obtaining intra-modal features for each type of sensor; using a cross-modal cross-attention module to process and fuse the intra-modal features, obtaining fused features for each UAV; constructing communication connection edges between each UAV based on their communication relationships; using a graph attention network and based on the fused features and the communication connection edges to determine attention weights; filtering the attention weights based on a preset threshold to obtain filtered attention weights; constructing a dynamic communication topology for the UAVs using the communication connection edges corresponding to the filtered attention weights; and performing multi-hop propagation and aggregation updates on the fused features based on the dynamic communication topology to obtain global feature vectors for each UAV.

[0029] The cross-level attention collaborative decision engine includes the HSCA (Heterogeneous Sensor Cross-Attention) module, the DGAC (Dynamic Graph Attention Communication) module, and the ABT (Adaptive Behavior Tree) module. It also incorporates an Actor-Critic dual-network architecture and a PPO (Proximal Policy Optimization) training component. These three attention units at different levels achieve deep collaboration: the output of HSCA serves as the node input features for DGAC; the features optimized through iterative iterations in DGAC serve as the decision basis for ABT; and the results of ABT use policy gradients to back-correct the attention weight parameters of HSCA and DGAC.

[0030] The schematic diagram of the HSCA module structure is as follows: Figure 3 As shown, the HSCA module addresses the question of "what to focus on," which involves autonomously selecting the most relevant features from multi-source, multi-type sensor data for deep fusion. This unit employs a three-tiered, progressive architecture: "category-aware encoding, intra-modal self-attention, and cross-modal interactive attention."

[0031] The HSCA module includes: a first layer and Type-Aware Encoders. Considering the differences in data format and physical properties among the four types of sensors, a dedicated neural network architecture is matched to each to complete the feature encoding process: Radar data uses an encoding module combining CNN (Convolutional Neural Network) and LSTM (Long Short-Term Memory). Spatial features are extracted through two-dimensional convolutional layers with channel numbers varying in sequence (3, 64, 128), and temporal features are extracted using a LSTM network with an input dimension of 128 and an output dimension of 256, ultimately outputting a feature vector with a dimension of 256. Infrared data uses ResNet-18 (18-layer Residual... The network (18-layer residual network) serves as the encoding backbone, using residual modules to acquire multi-scale spatial features, and then performing global pooling operations through GlobalAvgPool, outputting a feature vector with a dimension of 256. For ESM data, a bidirectional LSTM is selected as the encoding structure, stacking two layers of Bi-LSTM that map from 512 dimensions to 256 dimensions to capture forward and backward temporal correlation features, outputting a feature vector with a dimension of 256. For laser data, an MLP (Multilayer Perceptron) is used to complete the encoding mapping, and FC (fully connected layers) with dimensions varying sequentially from 64 to 128 to 256 are used to complete the feature transformation, outputting a feature vector with a dimension of 128.

[0032] This type of perceptual encoder combines the physical properties and inherent data distribution patterns of different sensors, selects the network architecture with the best adaptability to complete feature extraction, and maps multi-source heterogeneous data to a semantic feature space with unified dimensions, providing a reliable input basis for the subsequent attention operation module.

[0033] The second layer is Intra-Sensor Self-Attention. Self-attention operations are performed independently on the encoded features of each sensor class. Through Query-Key-Value mapping and Softmax normalization, deep optimization extraction of features within the same sensor is achieved. Different attention head parameters are matched to different sensors: radar, infrared, and ESM each have 4 attention heads, while the laser module has 2 attention heads. The number of attention heads is flexibly adjusted according to the feature complexity and information weight of the corresponding sensor.

[0034] The third layer: Cross-Sensor Cross-Attention (Cross-modal cross-attention module), which can set up to 8 attention heads. This structure is the innovation of the HSCA component. When calculating cross-modal interaction weights, it uses the following mechanism: data from a single sensor class is used as the query sequence input, and data collected by all other sensors are used as key and value sequences respectively to participate in the calculation, in order to calculate the cross-attention weights between different sensors. The cross-attention weights can be calculated according to the following formula: ; in, For the first The sensor is for the first Cross-attention weights of each sensor, For normalized exponential functions, For the first The first drone carried The query projection vector of each sensor, For the first The first drone carried The key projection vector of each sensor, This indicates the transpose operation. The dimension of a single attention head.

[0035] Based on cross-attention weights, the first one can be determined according to the following formula. Cross-modal characteristics corresponding to each sensor: ; in, For the first The first drone carried Cross-modal characteristics corresponding to each sensor The total number of sensors, For the first In the drone The projection vector of the values ​​from each sensor.

[0036] The dimension of each attention head is set as follows: =64. After multi-head attention concatenation and layer normalization, the final fused feature is obtained from the multi-source features. The calculation formula is: ; in, For the final fusion features, For layer normalization processing, For splicing operations, For the first The first feature output by the attention head in the drone. For the first The first of the drones Features of the attention head output.

[0037] Fusion features output by the HSCA module It not only encompasses complementary feature information from sensors from different sources, but also explicitly encodes the semantic relationships between different sensors and their contribution to the current task by using attention weights.

[0038] The DGAC module primarily addresses the core issue of "communication target selection," specifically, how to adaptively select the interaction targets and transmitted content with the highest information value in a multi-UAV swarm scenario. This component employs a three-layer architecture design: "graph attention weight calculation, attention threshold selection, and multi-hop message passing." A schematic diagram of the DGAC module structure is shown below. Figure 4 As shown, the specific steps are as follows: Step 1: Graph Attention Weight Calculation. The communication relationships of the multi-UAV system are modeled as a dynamic graph structure. ,in, For a moment Animated GIFs For a moment A collection of drone nodes, For a moment The set of communication connection edges. Node features are the fused features output by the HSCA module. The attention weights between agents are calculated using GAT (Graph Attention Network): ; in, For the first The drone and the first The unnormalized attention coefficients among the drones For a linear rectified activation function with leakage, This is a trainable attention parameter vector. Let be a trainable linear transformation matrix. For the first The fusion characteristics of individual drones For the first The fusion characteristics of individual drones This indicates the input range for vector concatenation. This represents the vector concatenation operation.

[0039] Normalizing the unnormalized attention coefficients can be achieved using the following formula: The drone targets the first Attention weights for each drone: ; in, For the first The drone targets the first Attention weights for each drone It is a natural exponential function. For the first The drone and the first Unnormalized attention coefficients among drones.

[0040] The second step is attention threshold filtering. This application designs a topology dynamic filtering strategy based on attention thresholds, retaining only those with attention weights exceeding the threshold. Communication connection edge: ; in, For a moment The filtered set of communication connection edges This is the communication threshold. The communication threshold can be set to 0.15 and can be adjusted within the range of 0.1 to 0.3 depending on the communication bandwidth limit.

[0041] Attention threshold filtering can truncate communication edges with low attention weight values, preventing low-value information from crowding out communication bandwidth and achieving a dynamic adaptive sparsity effect in the communication topology.

[0042] Step 3, Multi-hop Message Passing: Building upon the sparsified communication topology, multi-hop message passing is then performed using multiple layers... The multi-hop message passing mechanism with a value of 2 allows each drone to aggregate neighbor node information and iteratively update its own representation. The corresponding calculation formula is: ; in, For the first The global feature vector updated by aggregating data from all drones. Indicates the first A collection of neighbors of a drone.

[0043] Through two layers of messaging, a single drone can obtain information about other drones that have been relayed through multiple hops by utilizing a limited number of direct neighbors.

[0044] The communication complexity of the DGAC module is ,in, The number of edges after filtering. For feature dimensions. Due to attention threshold filtering... ,in, For the number of agents, the communication complexity of DGAC is much lower than that of fully connected communication. .

[0045] Since attention threshold filtering results in a significantly smaller number of communication edges after filtering compared to fully connected communication, communication overhead can be reduced.

[0046] Experiments show that DGAC reduces communication overhead from 72.4 KB / s in MAPPO (Multi-Agent Proximal Policy Optimization) full-connection communication to 38.2 KB / s, a reduction of 47.2%.

[0047] Step S13: Based on the fused features and the global feature vector, construct a local observation time series sequence, and use the policy network and the local observation time series sequence to generate an adaptive behavior tree.

[0048] In this embodiment, a local observation time series is constructed based on the fused features and the global feature vector. The local observation time series is input into the policy network to obtain behavior tree structure parameters. The behavior tree structure parameters include the root node type probability distribution, the number of child nodes, the type probability distribution of each child node, and the parameter values ​​of each atomic action node. The behavior tree structure parameters are parsed to generate an adaptive behavior tree for each UAV.

[0049] The ABT (Adaptive Behavior Tree) module primarily handles the "decision content generation" function, which involves achieving hierarchical and efficient reasoning and judgment within a high-dimensional continuous action domain. This unit employs an overall architecture of "policy network generating behavior tree parameters + hierarchical tree structure parsing + action space dimensionality compression".

[0050] Behavior trees can be divided into four different types of nodes: Selector (selection node, runs child nodes in order until a child node returns a success status and then terminates), Sequence (sequence node, runs child nodes in order until a child node returns a failure status and then terminates), Parallel (parallel node, starts all child nodes synchronously, waits for all nodes to complete before returning the result), and Action (atomic action node, responsible for executing the specific control commands of the drone).

[0051] The Actor network relies on the local observation time series stored by each drone. The sequence encompasses the fused features obtained through HSCA operation and the global context representation updated through DGAC iteration. The final output is the structural parameter vector corresponding to the behavior tree, which specifically includes three types of content: the probability distribution of the type to which the root node belongs (which can be selected from three types: selector, sequencer, and parallelizer), the total number of child nodes and the type probability corresponding to each child node, and the specific parameter values ​​of each atomic action node (such as the angle range of the scanning sector, the dwell time of the sensor, the threshold parameters for target tracking, etc.).

[0052] The behavior tree generation module parses the structured parameters output by the Actor network and constructs a hierarchical behavior tree in a top-down manner. The maximum depth of this tree structure is set to [value missing]. The number of branches at each node level can be adaptively adjusted according to the complexity of the task. The behavior tree execution module performs recursive operation in a top-down manner according to the tree topology: when accessing the selector node, it will try the execution of each child node in sequence until a success signal is returned; when accessing the sequence node, it will run each child node in sequence until a failure signal is returned; when accessing the parallel node, it will trigger the execution flow of all child nodes simultaneously; when accessing the action node, it will call the corresponding atomic action execution module to complete the operation.

[0053] The ABT module performs hierarchical compression of the motion space. The dimensions of the original motion space can be expressed as follows: ; in, The dimension of the original action space. For the number of drones, The number of categories for action parameters. The number of sensors carried by drones increases exponentially with the increase of parameters.

[0054] ABT can compress the motion space to: ; in, This represents the action space dimension corresponding to the adaptive behavior tree. The number of node types. The number of categories for node parameters.

[0055] Based on the above parameters, the adaptive behavior tree can compress the action space to 200 dimensions, with a compression ratio of no less than 85%. Relying on the hierarchical architecture of the behavior tree, the strategy first determines the upper-level decision intent, such as "full-domain area search", "dynamic target tracking", and "target type recognition", and then gradually breaks it down into specific atomic actions at the lower level, such as "setting the scanning sector angle to 120 degrees" and "adjusting the dwell time to 2 seconds", thereby completing a highly adaptable transformation from decision intent to execution action.

[0056] Step S14: Using the adaptive behavior tree, determine the control commands for each UAV, and based on the control commands, control each UAV to perform detection actions to obtain a reward signal.

[0057] In this embodiment, the adaptive behavior tree is input to the tree executor, so that the tree executor recursively executes each node in the adaptive behavior tree according to the topology of the adaptive behavior tree and outputs control commands for each UAV. The control commands include path planning commands and sensor scheduling commands. Based on the path planning commands and the sensor scheduling commands, each UAV is controlled to perform detection actions to obtain reward signals.

[0058] In this embodiment, a deep association pattern is established between three levels of attention modules to achieve cross-level information filtering and transmission linkage: Forward information flow (perception → communication → decision): fused features obtained through HSCA module processing. It can be used as the node input feature of the DGAC module, providing higher-precision single-sensor dimensional features for the graph attention operation process; the representation updated iteratively by the DGAC module (It covers global scene context information and can be used as the decision layer input of the ABT module to provide a global view of situational awareness for the behavior tree construction process; the behavior tree generated by the ABT module outputs specific path planning and sensor scheduling instructions through the tree executor to drive the UAV to carry out detection operations.)

[0059] Step S15: Evaluate the value of the detection actions performed by the UAV to obtain the action value, and construct a generalized advantage function based on the reward signal and the action value.

[0060] In this embodiment, global state information of the UAV is acquired; the global state information includes the position information, flight attitude information and sensor operation status information of each UAV; the global state information is input into the evaluation network so that the evaluation network can be used to evaluate the value of the detection actions performed by each UAV and obtain the action value.

[0061] In this embodiment, a CTDE (Computer-to-Demand) approach can be used to train and execute the UAV cooperative detection strategy. During the training phase, the evaluation network can acquire the position information, flight attitude information, and sensor operating status information of each UAV, and can also acquire target-related parameters to form global state information. The evaluation network evaluates the value of the multi-UAV joint detection actions based on the global state information to obtain the time-based assessment. The corresponding action value.

[0062] Backward gradient flow (decision-making → communication → perception): The output of the ABT unit generates a reward signal through environmental influences. The Critic network evaluates the value of joint actions and calculates the GAE dominance function. The policy gradient is propagated backward along the Actor network to the DGAC unit, updating the parameters of the graph attention. and The selection scheme for the communication topology is optimized; the gradient is further backpropagated to the HSCA unit to update the parameters of the type-aware encoder and cross-attention, optimizing the sensor data filtering and fusion process. The attention parameters at the three levels are optimized end-to-end within a unified policy gradient framework.

[0063] The role of cross-level attention collaboration is manifested in the following ways: cross-attention at the sensor level can endow fused features with perceptual characteristics for target categories, providing high-quality node feature representations for communication filtering at the agent level of UAV swarms; graph attention at the agent level enables each UAV to grasp the global situation dynamics, providing contextual support with collaborative attributes for planning and decision-making at the action level; tree attention at the action level uses policy gradients to back-update the attention parameters of the first two levels, promoting mutual promotion and joint iterative optimization of the attention modules at the three levels.

[0064] Step S16: Determine whether the preset termination condition is met.

[0065] Step S17: If the preset termination condition is not met, the generalized dominance function is used to update the parameters of the policy network, and the process jumps to the step of obtaining multi-source heterogeneous sensor data of each UAV until the preset termination condition is met and the target control command of the UAV is obtained.

[0066] Step S18: Use the target control command to control each UAV to perform a collaborative detection mission.

[0067] In this embodiment, the specific process of executing the UAV collaborative detection task based on cross-level attention is as follows: Figure 5 As shown, the specific steps are as follows: Step 1: Initialize the multi-UAV collaborative detection scenario. Configure a UAV swarm of 4 to 16 drones, and preset the detection performance parameters for various payloads: radar effective detection range 15km, field of view 120°, infrared detection equipment range 8km, electronic support measurement device detection range 30km, and laser detection unit range 5km. Import a raster map of the mission area containing four typical terrain types: urban, mountainous, marine, and desert, and complete the initial communication topology construction of the swarm. Simultaneously, set the hyperparameters related to the distributed partially observable Markov decision process: discount factor. The value is 0.99, which is the communication trigger threshold. Set to 0.15, the maximum allowed nesting depth for behavior trees. It is 5.

[0068] Step 2: Each UAV completes the acquisition of multi-source heterogeneous sensor data. Each UAV simultaneously activates four types of sensors to collect initial observations at the corresponding time. ;in, For the first The initial observations of the drone at the corresponding time. For in-phase or orthogonal complex data acquired by radar, Thermal radiation images acquired by an infrared sensor. Pulse descriptor sequences acquired by the electronic support measures reconnaissance unit. Distance sequences collected by a laser rangefinder.

[0069] Step 3: The HSCA unit completes the deep fusion of features from multiple sensor sources, sequentially performing category-aware encoding, intramodal self-attention computation, and cross-modal interactive attention computation with 8 attention heads, ultimately outputting a 256-dimensional context-aware fusion feature. .

[0070] Step 4: The DGAC module completes the construction of the dynamic communication topology. First, the corresponding attention weights are calculated using a graph attention network, and then the communication threshold is set. The weighting is set to 0.15, followed by a K-order message passing mechanism with a hop count of 2, and finally the updated situational awareness results for each UAV are output. .

[0071] Step 5: The Actor network generates an adaptive behavior tree, based on local observation history. Output behavior tree structure parameters: node type, number of child nodes, and atomic action parameters.

[0072] Step 6: The behavior tree completes the parsing and execution process of the detection action, covering navigation parameter adjustment (running speed). Heading angle Flight altitude The behavior tree outputs the action space dimension, which consists of two types of control variables: sensor state scheduling (on / off / beam pointing / operating mode). It is 200.

[0073] Step 7: Use the Critic network to estimate the value of joint actions and complete the GAE advantage function. The solution calculation.

[0074] Step 8: Determine if the termination condition is met (reaching the maximum training 5000 rounds or achieving a detection success rate >95% for 100 consecutive rounds). If met, output the optimal policy and terminate; otherwise, update the policy parameters using the PPO algorithm. ,in, The parameter is The update target of the strategy network, For near-end clipping target items, These are the weighting coefficients for the value objective items. For value objectives, The weighting coefficients for the entropy objective term are... The target term is entropy. Then return to step 2 to continue.

[0075] In this embodiment, the preset termination condition may include reaching the maximum number of training rounds of 5000, or having a detection success rate greater than 95% for 100 consecutive training rounds. If the maximum number of training rounds is reached, or the detection success rate is greater than 95% for 100 consecutive training rounds, then the preset termination condition is determined to be met; if the maximum number of training rounds is not reached, and the detection success rate is not greater than 95% for 100 consecutive training rounds, then the preset termination condition is determined to be unmet.

[0076] Key hyperparameter configuration data are shown in Table 1: Table 1 Key Hyperparameter Configuration Data

[0077] This application deploys attention mechanisms across three different dimensions: HSCA, DGAC, and ABT, establishing an integrated information filtering architecture encompassing object perception, interaction selection, and action decision-making. The three attention mechanisms are deeply integrated through forward information flow and backward gradient updates: the cross-attention module at the perception unit level outputs more accurate node feature representations for the graph attention module at the multi-agent interaction level; the graph attention module at the multi-agent interaction level provides contextual reference information about the global scene for the tree attention module at the decision output level; and the policy gradient at the decision output level iteratively optimizes the parameters of the first two attention modules through backpropagation.

[0078] This shows that the three different levels of information filtering processes are incorporated into a unified attention architecture for end-to-end collaborative optimization, effectively avoiding the information transmission loss and performance shortcomings caused by the separate optimization of each level in traditional solutions.

[0079] Test results show that the system achieves a global detection accuracy of 93.6%, which is 10.1 percentage points higher than the MAPPO-ATTN algorithm that only introduces an attention module at a single level, and 15.0 percentage points higher than the baseline MAPPO method without an attention mechanism. Ablation tests demonstrate that removing all three types of attention modules simultaneously reduces the overall system performance by 23.2 percentage points, fully confirming the crucial role of the cross-level attention collaboration mechanism.

[0080] The HSCA module proposed in this application matches an appropriate neural network architecture (CNN combined with LSTM, ResNet-18, Bi-LSTM, MLP) based on the physical properties and data format differences of various sensors, projecting multi-source heterogeneous data into a semantic feature space with unified dimensions. Subsequently, it completes the deep association and fusion of information from different sensors and adaptive weight allocation by using an 8-head intramodal self-attention mechanism and a cross-modal interactive attention module. Among them, the cross-attention weight parameters can explicitly characterize the degree of semantic association between different sensors and their actual contribution to the current target task.

[0081] Therefore, the fused feature vector with a dimension of 256 not only covers the complementary information of multiple source sensing devices, but also dynamically adapts the contribution ratio of different sensing devices according to the target category and scene (for example, the radar weight is 0.42 when detecting targets with high radar cross section, and the electronic support measures weight is 0.35 when identifying targets with low probability of interception), and finally constructs a situational representation with context-aware characteristics.

[0082] Experimental results show that the proposed method achieves a target detection success rate of 93.6%, which is 15.0 percentage points higher than the conventional feature direct stitching scheme; the false alarm rate has decreased from 11.2% to 6.8%, a reduction of 39.3%. Attention weight visualization results confirm that the HSCA component can autonomously adjust the weight allocation of different sensors according to the category of the target to be detected, effectively improving the comprehensive utilization efficiency of multi-source heterogeneous sensor information.

[0083] The DGAC module proposed in this application utilizes a graph attention mechanism to calculate the attention weights between different agents, retaining only communication links with weights exceeding a threshold of 0.15 and cutting off the transmission paths of low-value information. Through the filtering effect of the attention threshold, the final number of communication edges is far lower than the total number of edges n(n-1) / 2 in a fully connected scenario. Combined with a K=2 layer multi-hop message passing mechanism, a single UAV only needs to communicate directly with a small number of adjacent nodes to implicitly obtain global situational information.

[0084] Therefore, the communication bandwidth consumption decreased from 72.4 KB / s in the MAPPO fully connected communication mode to 38.2 KB / s, a decrease of 47.2%. In the scenario where the drone swarm size expands from 4 to 16 drones, the communication overhead only increases by 120.8%, far lower than the 360% increase corresponding to the fully connected architecture, and its complexity approximately conforms to the variation law of O(n log n).

[0085] Real-world performance tests show that, for combat scenarios with limited communication bandwidth, this research significantly reduces communication overhead while maintaining cooperative detection effectiveness (detection accuracy of 93.6%), giving the platform better scalability and resilience. Even if a local communication link is lost, DGAC can dynamically adjust and reconstruct the communication network structure to ensure that collaborative working capabilities remain intact.

[0086] The ABT module proposed in this application can map the output of the Actor network into a hierarchical behavior tree architecture. By combining and arranging four types of nodes—selectors, sequencers, parallelizers, and executors—it transforms the originally exponentially growing action space. Compress to =4×10×5=200, with a space compression ratio of no less than 85%. The hierarchical architecture of the behavior tree allows the strategy to prioritize the determination of high-level decision-making objectives, such as "area search", "target tracking" and "feature recognition", and then gradually break them down into low-level executable atomic operations, such as "setting the scanning sector range" and "adjusting the beam dwell time", thereby completing a highly reliable conversion from decision intention to specific action.

[0087] Therefore, the output dimension of the policy network is reduced from thousands to 200 dimensions, effectively reducing the complexity of the policy search space. The number of training epochs required for the policy to reach convergence is reduced by 42.6%, and a detection success rate of 93.6% can be achieved with only 5000 training epochs.

[0088] Experimental results show that the average detection time was reduced from 18.2 seconds for MAPPO to 12.5 seconds, a decrease of 31.3%. The modular architecture adopted by the behavior tree gives the policy better interpretability and reusability. Subtree modules can be shared between different task scenarios, effectively reducing the additional cost of policy migration.

[0089] The three-layer attention mechanism proposed in this invention completes the end-to-end joint optimization process within a unified training framework of centralized training and distributed execution. The cross-attention weights of the heterogeneous sensor cross-attention module, the graph attention parameters of the dynamic graph attention communication module, and the behavior tree generation parameters of the adaptive behavior tree decision module share the same set of policy gradient signals. Collaborative parameter updates are achieved through the backpropagation mechanism of the proximal policy optimization algorithm. The three layers of attention have mutually beneficial effects: sensor-level attention optimization effectively improves the node representation quality at the communication level; communication-level attention optimization significantly improves the accuracy of global context information at the decision level; and the policy feedback output from the decision level provides clear guidance for the optimization direction of the first two attention layers.

[0090] Therefore, it can be seen that the three-level attention mechanism can achieve collaborative optimization and autonomously adapt to diverse battlefield scenarios (cities, mountains, oceans, deserts) and target categories (fighter jets, helicopters, ships, ground vehicles) without the need to design separate response plans for different scenarios.

[0091] Experimental results show that in a test set covering 12 typical application scenarios, HA-MAPPO-BT maintained a detection success rate of no less than 90% in all test scenarios, demonstrating excellent environmental adaptability. Scalability test results confirm that when the number of drones in the cluster increased from 4 to 16, the corresponding detection success rate increased from 94.2% to 97.8%, fully demonstrating the practical value of this method in large-scale cluster operation scenarios.

[0092] In this embodiment, the type-aware encoder can be replaced according to the actual sensors carried by the UAV. The radar encoder can use ViT (Vision Transformer) or Swin Transformer (Shifted Window Transformer) to replace the combined structure of the convolutional neural network and long short-term memory network; the infrared encoder can use a ConvNeXt convolutional network to replace the 18-layer residual network; the electronic support measures encoder can use Transformer-XL (Transformer with Extra Long Context) to replace the bidirectional long short-term memory network; and the laser encoder can use a one-dimensional convolutional neural network to replace the multilayer perceptron. The output dimension of the replaced type-aware encoder only needs to match the input dimension of the cross-modal cross-attention module, for example, it can be unified to 256 dimensions.

[0093] In this embodiment, when a new sensor type is introduced, a corresponding type-aware encoder can be set according to the modal characteristics of the new sensor data. A query projection matrix, key projection matrix, and value projection matrix corresponding to the new sensor are added to the cross-modal attention module, and the input dimension of the feature stitching layer is adjusted according to the number of sensor types. The encoded features of the new sensor can interact with the encoded features of existing sensors through the cross-modal attention module without needing to pre-set fixed feature fusion weights.

[0094] In this embodiment, the graph attention network in the dynamic graph attention communication module can be replaced with GraphSAGE (Graph Sample and Aggregate), GIN (Graph Isomorphism Network), or TransformerConv graph convolutional structure. The communication threshold can be adjusted within the range of 0.1 to 0.3 according to the communication bandwidth limit, or the information bottleneck communication compression method can be used to further compress the transmitted information. The dynamic communication topology can be updated in every execution step, or it can be updated at preset intervals of a number of execution steps.

[0095] In this embodiment, the node types of the adaptive behavior tree can also include Decorator nodes and Subtree reference nodes. Decorator nodes can be used to perform conditional judgments or loop control, while subtree reference nodes can be used to reuse and combine behavior trees. When adding node types, the corresponding node type probability dimension can be added to the output layer of the policy network, while keeping the parsing process of the behavior tree structure parameters unchanged.

[0096] In this embodiment, the near-end policy optimization algorithm can be replaced by policy gradient algorithms such as SAC (Soft Actor-Critic), TD3 (Twin Delayed Deep Deterministic Policy Gradient), or TRPO (Trust Region Policy Optimization). For the selection of discrete nodes in the behavior tree, DQN (Deep Q-Network) or Gumbel-Softmax can be combined for hybrid training; the evaluation network can also adopt a dual-evaluation network structure to reduce value assessment bias.

[0097] In this embodiment, the centralized training and distributed execution methods can be adjusted to either a fully distributed or a hierarchical centralized approach depending on communication conditions. In a fully distributed approach, each UAV independently completes training and decision-making, without needing to evaluate the network to obtain global state information. In a hierarchical centralized approach, regional command nodes can be set up to perform multi-UAV collaborative scheduling within a local area. When the communication link is stable, a centralized execution method can also be used; when communication resources are limited or there is interference, a fully distributed execution method can be used.

[0098] In this embodiment, the Softmax attention in the heterogeneous sensor cross-attention module can be replaced with LinearAttention or KernelAttention, the node attention in the dynamic graph attention communication module can be replaced with EdgeAttention or HypergraphAttention, and the node selection in the adaptive behavior tree can use a greedy selection method based on value functions instead of the probability sampling method.

[0099] In this embodiment, the cross-level attention collaboration structure can also be used for multi-robot collaborative search and rescue, autonomous vehicle fleet collaborative perception, or multi-camera collaborative tracking. Multi-robot collaborative search and rescue scenarios can utilize visual sensing devices, LiDAR, infrared detectors, and gas detection modules; autonomous vehicle fleet collaborative perception scenarios can utilize camera devices, millimeter-wave radar, LiDAR, and ultrasonic sensing units; and multi-camera collaborative tracking scenarios can utilize visible light, infrared, and thermal imaging cameras. For different application scenarios, the type-aware encoder can be replaced to adapt to sensor types while maintaining the processing relationships between heterogeneous sensor cross-attention, dynamic graph attention communication, and adaptive behavior trees.

[0100] In this embodiment, during the training phase, a server equipped with an NVIDIA A100 graphics processor can be used, and training scenarios including urban, mountainous, marine, and desert terrains, as well as daytime, dusk, and nighttime lighting conditions, can be constructed using Unity3D or Gazebo simulation platforms. The training cycle can be 24 to 48 hours. During the runtime phase, the policy network can be deployed on an NVIDIA Jetson Xavier NX or Jetson AGX Orin embedded computing device mounted on a drone, and the time for a single inference can be less than 10 milliseconds.

[0101] In this embodiment, the dynamic graph attention communication module can transmit data through the DDS (Data Distribution Service) middleware, and the required communication bandwidth can not exceed 50 KB / s. Various sensors can access data through the ROS (Robot Operating System) node, and the heterogeneous sensor cross-attention module can act as a robot operating system node to subscribe to the data collected by each sensor and output fused features.

[0102] This application synchronously applies the attention mechanism to three levels: the perception module layer, the agent layer, and the behavior determination layer, building a fully automated information selection path of "feature filtering - interaction subject adaptation - execution sequence generation" to achieve cross-level information identification and transmission linkage; it realizes deep cognitive fusion of heterogeneous sensing devices: using a category-aware coding component to map multi-source heterogeneous data to a unified semantic representation space, relying on intra-modal self-attention mechanism and cross-modal interactive attention to achieve deep information interaction and dynamic weight adjustment between sensing units, so that the fused features have scene perception attributes oriented towards target attributes and combat scenarios; it realizes low-energy collaborative interaction of multiple agents: dynamically constructing the interaction topology based on the graph attention framework, using attention threshold filtering to complete the sparsity processing of the interaction link, and after multiple rounds of information... The system achieves implicit consistency of the global state through information interaction, significantly reducing communication overhead while ensuring no loss of collaborative performance. It enables hierarchical and efficient decision-making in high-dimensional action spaces: relying on the adaptive behavior tree generation unit, the output of the policy network is transformed into a hierarchical behavior tree structure. By combining selection nodes, sequence nodes, parallel nodes, and execution nodes, hierarchical dimensionality reduction of the action space is completed (the dimensionality reduction ratio is not less than 85%), improving the policy convergence rate and decision execution efficiency. It achieves efficient connection and collaborative iteration of the three-layer attention mechanism: using a training framework of centralized training and distributed execution, the attention assignment of the perception node level, intelligent decision body level, and execution action level is jointly optimized end-to-end under a unified policy gradient framework. The attention modules of the three levels empower each other and update synchronously.

[0103] As can be seen, this application provides a method for UAV cooperative detection task execution based on cross-level attention, including acquiring multi-source heterogeneous sensor data from each UAV, thus improving the completeness of the representation of heterogeneous sensing information collected by each UAV; performing feature encoding, intra-modal self-attention processing, cross-modal cross-attention processing, and fusion on the multi-source heterogeneous sensor data to obtain fused features of each UAV; performing multi-hop transfer and aggregation update on the fused features to obtain global feature vectors of each UAV, thus improving the fusion degree of multi-source heterogeneous sensor information and enhancing the aggregation ability of each UAV for multi-hop transferred information; constructing a local observation time series based on the fused features and the global feature vectors, and generating an adaptive behavior tree using a policy network and the local observation time series; thus improving the temporal utilization of the fused features and global feature vectors and enhancing the adaptability of the behavior tree generation; using the adaptive behavior tree, determining the control commands for each UAV, and controlling each UAV to perform detection actions based on the control commands to obtain rewards. The signal improves the matching between control commands and adaptive behavior trees, and enhances the feedback capability of the detection action execution process. The detection actions performed by the UAV are evaluated to obtain action value. Based on the reward signal and the action value, a generalized dominance function is constructed. This improves the completeness of the evaluation of the detection action effect and enhances the ability of the generalized dominance function to represent the quality of actions. It determines whether the current condition meets the preset termination condition, avoiding invalid iterations after the preset termination condition has been met. If the preset termination condition is not met, the generalized dominance function is used to update the parameters of the policy network, and the process jumps to the step of acquiring multi-source heterogeneous sensor data from each UAV until the preset termination condition is met, obtaining the target control command for the UAV. This improves the matching degree between the policy network parameters and iterative feedback, ensuring that the loop process stops when the preset termination condition is met. The target control command is used to control each UAV to perform collaborative detection tasks, thus improving the consistency of the control process when each UAV performs collaborative detection tasks.

[0104] See Figure 6 As shown, this embodiment of the invention discloses a UAV collaborative detection task execution device based on cross-level attention, which may specifically include: Data acquisition module 11 is used to acquire multi-source heterogeneous sensor data from each UAV; Data processing module 12 is used to perform feature encoding, intramodal self-attention processing, cross-modal cross-attention processing and fusion on the multi-source heterogeneous sensing data to obtain the fused features of each UAV, and to perform multi-hop transfer and aggregation update on the fused features to obtain the global feature vector of each UAV. The adaptive behavior tree generation module 13 is used to construct a local observation time series based on the fused features and the global feature vector, and generate an adaptive behavior tree using the policy network and the local observation time series. The reward signal determination module 14 is used to determine the control commands for each UAV using the adaptive behavior tree, and control each UAV to perform detection actions based on the control commands to obtain a reward signal; The function construction module 15 is used to evaluate the value of the detection actions performed by the UAV, obtain the action value, and construct a generalized advantage function based on the reward signal and the action value. Judgment module 16 is used to determine whether the preset termination condition is met. The loop module 17 is used to update the parameters of the policy network using the generalized dominance function if the preset termination condition is not met, and then jump to the step of obtaining the multi-source heterogeneous sensor data of each UAV until the preset termination condition is met and the target control command of the UAV is obtained. The task execution module 18 is used to control each UAV to perform a collaborative detection task using the target control command.

[0105] In some specific embodiments, the data acquisition module 11 may specifically include: The perception data acquisition submodule is used to acquire multi-source heterogeneous sensing data from each UAV from the multi-source heterogeneous perception layer. The multi-source heterogeneous perception layer includes different types of heterogeneous sensing devices. The different types of heterogeneous sensing devices include radar for acquiring in-phase or orthogonal complex data, infrared sensing devices for acquiring thermal radiation imaging, electronic support measures reconnaissance units for acquiring pulse descriptor sequences, and laser ranging devices for acquiring range sequences.

[0106] In some specific embodiments, the data processing module 12 may specifically include: The feature encoding submodule is used to encode the features of multi-source heterogeneous sensor data using type-aware encoders corresponding to different types of sensors, so as to obtain the encoded features of various types of sensors. The intramodal processing submodule is used to perform intramodal self-attention processing on the encoded features using the intramodal self-attention module to obtain intramodal features of various sensors; The cross-modal processing submodule is used to perform cross-modal cross-attention processing and fusion on the intra-modal features using the cross-modal cross-attention module to obtain the fused features of each UAV.

[0107] In some specific embodiments, the data processing module 12 may further include: The communication connection edge construction submodule is used to construct communication connection edges between drones based on the communication relationships between them. The attention weight determination submodule is used to determine attention weights based on the fused features and the communication connection edges using a graph attention network. The attention weight filtering submodule is used to filter the attention weights based on a preset threshold to obtain the filtered attention weights. The dynamic communication topology construction submodule is used to construct the dynamic communication topology of the UAV using the communication connection edges corresponding to the filtered attention weights. The aggregation update submodule is used to perform multi-hop transmission and aggregation update of the fused features based on the dynamic communication topology to obtain the global feature vector of each UAV.

[0108] In some specific embodiments, the adaptive behavior tree generation module 13 may specifically include: The structural parameter determination submodule is used to input the local observation time series into the policy network to obtain the behavior tree structural parameters; the behavior tree structural parameters include the root node type probability distribution, the number of child nodes, the type probability distribution of each child node, and the parameter values ​​of each atomic action node; The behavior tree parsing submodule is used to parse the behavior tree structure parameters and generate adaptive behavior trees for each UAV.

[0109] In some specific embodiments, the reward signal determination module 14 may specifically include: The tree execution submodule is used to input the adaptive behavior tree to the tree executor, so that the tree executor recursively executes each node in the adaptive behavior tree according to the topology of the adaptive behavior tree and outputs control commands for each UAV; the control commands include path planning commands and sensor scheduling commands; The detection action execution submodule is used to control each UAV to perform detection actions based on the path planning instructions and the sensor scheduling instructions, and to obtain reward signals.

[0110] In some specific embodiments, the function construction module 15 may specifically include: The global status information acquisition submodule is used to acquire the global status information of the UAV; the global status information includes the position information, flight attitude information and sensor operation status information of each UAV. The value assessment submodule is used to input the global state information into the evaluation network so that the evaluation network can be used to assess the value of the detection actions performed by each UAV and obtain the action value.

[0111] Figure 7This is a schematic diagram of an electronic device provided in an embodiment of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the method for executing a UAV cooperative detection task based on cross-level attention, as disclosed in any of the foregoing embodiments.

[0112] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0113] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored on it include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

[0114] The operating system 221 manages and controls the various hardware devices on the electronic device 20 and the computer program 222 to enable the processor 21 to perform operations and processing on the data 223 in the memory 22. The operating system 221 can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the UAV cooperative detection task execution method based on cross-level attention disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the UAV cooperative detection task execution device based on cross-level attention and transmitted from external devices, as well as data collected by its own input / output interface 25.

[0115] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0116] Furthermore, this application also discloses a computer-readable storage medium storing a computer program. When the computer program is loaded and executed by a processor, it implements the steps of the UAV cooperative detection task execution method based on cross-level attention disclosed in any of the foregoing embodiments.

[0117] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0118] The foregoing has provided a detailed description of the method, apparatus, device, and medium for performing UAV collaborative detection tasks based on cross-level attention provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for executing UAV collaborative detection tasks based on cross-level attention, characterized in that, include: Acquire multi-source heterogeneous sensor data from various drones; The multi-source heterogeneous sensing data is subjected to feature encoding, intra-modal self-attention processing, cross-modal cross-attention processing and fusion to obtain the fused features of each UAV. The fused features are then subjected to multi-hop propagation and aggregation update to obtain the global feature vector of each UAV. Based on the fused features and the global feature vector, a local observation time series is constructed, and an adaptive behavior tree is generated using the policy network and the local observation time series. Using the adaptive behavior tree, control commands for each UAV are determined, and based on the control commands, each UAV is controlled to perform detection actions to obtain reward signals; The value of the detection actions performed by the UAV is evaluated to obtain the action value. Based on the reward signal and the action value, a generalized advantage function is constructed. Determine whether the preset termination condition is met; If the preset termination condition is not met, the generalized dominance function is used to update the parameters of the policy network, and the process jumps to the step of obtaining multi-source heterogeneous sensor data of each UAV until the preset termination condition is met and the target control command of the UAV is obtained. Using the target control commands, each UAV is controlled to perform a collaborative detection mission.

2. The UAV cooperative detection task execution method based on cross-level attention according to claim 1, characterized in that, The acquisition of multi-source heterogeneous sensor data from each UAV includes: Multi-source heterogeneous sensing data of each UAV is acquired from the multi-source heterogeneous sensing layer; the multi-source heterogeneous sensing layer includes different types of heterogeneous sensing devices; the different types of heterogeneous sensing devices include radar for acquiring in-phase or orthogonal complex data, infrared sensing devices for acquiring thermal radiation imaging, electronic support measures reconnaissance units for acquiring pulse descriptor sequences, and laser ranging devices for acquiring range sequences.

3. The UAV cooperative detection task execution method based on cross-level attention according to claim 1, characterized in that, The multi-source heterogeneous sensor data is subjected to feature encoding, intra-modal self-attention processing, cross-modal cross-attention processing, and fusion to obtain the fused features of each UAV, including: By using type-aware encoders corresponding to different types of sensors, feature encoding is performed on multi-source heterogeneous sensor data to obtain the encoded features of various sensors. By using an intramodal self-attention module, the encoded features are processed by intramodal self-attention to obtain intramodal features of various sensors; By utilizing a cross-modal cross-attention module, the intra-modal features are processed and fused through cross-modal cross-attention to obtain the fused features of each UAV.

4. The UAV cooperative detection task execution method based on cross-level attention according to claim 1, characterized in that, The fused features are subjected to multi-hop propagation and aggregation update to obtain the global feature vectors of each UAV, including: Based on the communication relationships between the drones, construct communication connection edges between the drones; The attention weights are determined using a graph attention network and based on the fused features and the communication connection edges. The attention weights are filtered based on a preset threshold to obtain the filtered attention weights; A dynamic communication topology for the UAV is constructed using the communication connection edges corresponding to the selected attention weights. Based on the dynamic communication topology, the fused features are subjected to multi-hop transmission and aggregation update to obtain the global feature vector of each UAV.

5. The UAV cooperative detection task execution method based on cross-level attention according to claim 1, characterized in that, An adaptive behavior tree is generated using the policy network and the local observation time series, including: The local observation time series is input into the policy network to obtain the behavior tree structure parameters; the behavior tree structure parameters include the root node type probability distribution, the number of child nodes, the type probability distribution of each child node, and the parameter values ​​of each atomic action node; The behavior tree structure parameters are parsed to generate adaptive behavior trees for each UAV.

6. The UAV cooperative detection task execution method based on cross-level attention according to claim 1, characterized in that, Using the adaptive behavior tree, control commands for each UAV are determined. Based on these control commands, each UAV is controlled to perform detection actions to obtain reward signals, including: The adaptive behavior tree is input to the tree executor, so that the tree executor recursively executes each node in the adaptive behavior tree according to the topology of the adaptive behavior tree and outputs control commands for each UAV; the control commands include path planning commands and sensor scheduling commands. Based on the path planning instructions and the sensor scheduling instructions, each UAV is controlled to perform detection actions and obtain reward signals.

7. The method for executing UAV cooperative detection missions based on cross-level attention according to any one of claims 1 to 6, characterized in that, The value assessment of the detection actions performed by the UAV, to obtain the action value, includes: Acquire global status information of the UAVs; the global status information includes the position information, flight attitude information, and sensor operating status information of each UAV. The global state information is input into the evaluation network so that the evaluation network can be used to evaluate the value of the detection actions performed by each UAV and obtain the action value.

8. A UAV collaborative detection mission execution device based on cross-level attention, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous sensor data from each UAV. The data processing module is used to perform feature encoding, intramodal self-attention processing, cross-modal cross-attention processing and fusion on the multi-source heterogeneous sensor data to obtain the fused features of each UAV, and to perform multi-hop transfer and aggregation update on the fused features to obtain the global feature vector of each UAV. An adaptive behavior tree generation module is used to construct a local observation time series sequence based on the fused features and the global feature vector, and generate an adaptive behavior tree using a policy network and the local observation time series sequence. The reward signal determination module is used to determine the control commands for each UAV using the adaptive behavior tree, and control each UAV to perform detection actions based on the control commands to obtain a reward signal. The function construction module is used to evaluate the value of the detection actions performed by the UAV, obtain the action value, and construct a generalized advantage function based on the reward signal and the action value. The judgment module is used to determine whether the preset termination condition is met. The loop module is used to update the parameters of the policy network using the generalized dominance function if the preset termination condition is not met, and then jump to the step of acquiring the multi-source heterogeneous sensor data of each UAV until the preset termination condition is met and the target control command of the UAV is obtained. The task execution module is used to control each UAV to perform a collaborative detection task using the target control commands.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the UAV cooperative detection task execution method based on cross-level attention as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer programs; wherein, when the computer programs are executed by a processor, they implement the UAV cooperative detection task execution method based on cross-level attention as described in any one of claims 1 to 7.

Images