A sea-land-air cross-domain heterogeneous cluster cooperative global maximization smoke cover scheduling method

Abstract

This invention discloses a method for coordinating and maximizing global smokescreen coverage across land, sea, and air heterogeneous clusters, belonging to the field of cross-domain heterogeneous cluster coordination technology. The method includes collecting multi-source data and preprocessing it to obtain standardized multi-source heterogeneous data. This standardized data is then input into an HSUnet model for feature extraction and fusion, outputting a global state feature vector. Based on this global state feature vector and combined with smokescreen diffusion patterns, a global optimization target model is established. This global optimization target model is used as the optimization task of a DPDQN-TER model, with the global state feature vector as input for training and inference, generating the optimal cross-domain collaborative scheduling strategy. This invention improves the collaborative smokescreen coverage capability of land, sea, and air heterogeneous clusters, achieving maximum global smokescreen coverage, and has strong practicality and promotional value.

Description

Technical Field

[0001] This invention relates to the field of cross-domain heterogeneous cluster collaboration technology, specifically to a global maximization smokescreen coverage scheduling method for cross-domain heterogeneous cluster collaboration across land, sea, and air. Background Technology

[0002] In modern public safety, emergency protection, and the cover of important targets, smoke screens serve as a low-cost, high-efficiency means of concealment and interference. They effectively block the monitoring paths of various optical monitoring, infrared detection, and high-precision detection equipment, providing crucial cover for the relocation of protected targets and adjustments to protective deployments. With the continuous upgrading of protection environments and needs, single-domain, single-type smoke screen release platforms are no longer sufficient to meet the global concealment requirements in complex scenarios. Cross-domain heterogeneous cluster collaborative smoke screen coverage, encompassing sea, land, and air, is gradually becoming the mainstream application. This cross-domain heterogeneous cluster consists of unmanned surface vessels (USVs), unmanned vehicles (UAVs), and unmanned aerial vehicles (UAVs). These platforms exhibit significant differences in smoke screen release capabilities, mobility, deployment range, and environmental adaptability, forming typical heterogeneous characteristics. While this heterogeneity enables multi-dimensional and wide-range smoke screen coverage, it also presents numerous technical challenges for cluster collaborative scheduling.

[0003] Currently, existing smokescreen coverage scheduling methods suffer from the following shortcomings: First, they lack effective cross-domain heterogeneous data fusion capabilities. Multi-source data perceived by various platforms exhibit modal differences, noise interference, and data redundancy, making it difficult for traditional data fusion methods to achieve accurate fusion of heterogeneous data, resulting in insufficient reliability of the data source for scheduling decisions. Second, their global optimization capabilities are weak. Most methods focus only on smokescreen coverage optimization in a single or local area, failing to fully consider the correlation and constraints of cross-domain collaboration. This easily leads to local coverage oversaturation and blind spots in global coverage, failing to maximize global smokescreen coverage. Third, dynamic adaptation... The response is poor. Environmental parameters in complex scenarios are constantly changing, and the smoke screen diffusion process is also affected by various factors in real time. Existing scheduling methods mostly adopt static scheduling strategies, which are difficult to respond quickly to environmental changes and cluster state changes, resulting in a continuous decline in smoke screen coverage. Fourth, the generation efficiency and optimization performance of collaborative scheduling strategies are insufficient. When facing the high-dimensionality, dynamics, and strong constraints of cross-domain heterogeneous clusters, traditional optimization algorithms are prone to problems such as slow convergence and local optima. Ordinary deep reinforcement learning methods are difficult to solve the problems of feature entanglement and low efficiency of experience replay, and cannot generate efficient cross-domain collaborative scheduling strategies.

[0004] To address the aforementioned issues, there is an urgent need for a cross-domain heterogeneous cluster collaborative global maximization smokescreen coverage scheduling method to solve the problems existing in traditional methods. Summary of the Invention

[0005] The purpose of this invention is to provide a method for scheduling cross-domain heterogeneous clusters across land, sea, and air to collaboratively maximize global smokescreen coverage. By fusing a Hybrid Self-attention U-Net (HSUnet) and a Disentangled Parameterized Deep Q-Network with Transformer-Enhanced Replay (DPDQN-TER), this method achieves accurate fusion of cross-domain heterogeneous data, optimized modeling of global smokescreen coverage, and efficient generation of dynamic collaborative scheduling strategies. This addresses the problems of poor data fusion performance, weak global optimization capabilities, insufficient dynamic adaptability, and suboptimal scheduling strategy performance in existing methods, thereby improving the collaborative smokescreen coverage capability of cross-domain heterogeneous clusters across land, sea, and air and maximizing global smokescreen coverage.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A cross-domain heterogeneous cluster collaborative global maximization smokescreen coverage scheduling method includes: Step 1: Utilize the sensor modules mounted on each platform in the cross-domain heterogeneous cluster of sea, land, and air to collect multi-source data in real time, and preprocess the data to obtain standardized multi-source heterogeneous data; Step 2: Input the standardized multi-source heterogeneous data into the HSUnet model, perform feature extraction and fusion through the HSUnet model, and output a global state feature vector including all key state information of the cross-domain heterogeneous cluster. The HSUnet model includes an encoder, a self-attention fusion module and a decoder, and is trained and optimized using a hybrid loss function. Step 3: Based on the global state feature vector and combined with the smoke screen diffusion law, establish a global optimization objective model that includes smoke screen coverage area, cross-domain cooperation cost, coverage unevenness and various constraints. Step 4: The global optimization target model is used as the optimization task of the DPDQN-TER model. The global state feature vector is used as the input of the model. Through model training and inference, the optimal cross-domain collaborative scheduling strategy is generated to maximize the global smoke screen coverage. The DPDQN-TER model consists of a de-entanglement parameterization module, a parameterized Q network, a Transformer enhanced replay buffer module, and a target network.

[0007] Furthermore, the cross-domain heterogeneous cluster of sea, land, and air includes unmanned surface vessels, unmanned vehicles, and unmanned aerial vehicles, and the multi-source data includes the self-state data of each platform, environmental state data, and target area data.

[0008] Furthermore, in step 1, the multi-source data is preprocessed, specifically as follows: The min-max standardization method is used to standardize the multi-source data to obtain standardized multi-source heterogeneous data.

[0009] Furthermore, in step 2, the encoder adopts the UNet encoder structure to perform hierarchical feature extraction on the standardized multi-source heterogeneous data; the self-attention fusion module is used to fuse the features extracted by the encoder; the decoder adopts the UNet decoder structure to upsample and reconstruct the fused feature vector and output a global state feature vector with uniform dimensions.

[0010] Further, in step 2, the hybrid loss function is: ; In the formula, Cross-entropy loss is used to optimize the feature classification accuracy of the model. Dice loss is used to address data imbalance and improve the accuracy of feature fusion. The weighting coefficients for the dice loss are determined through model training to find suitable values.

[0011] Furthermore, in step 3, the smoke screen coverage area is the union area of ​​the coverage areas formed after the smoke screens of each platform are released; the cross-domain collaboration cost includes communication cost, platform maneuver cost, and smoke screen consumption cost; the coverage unevenness is calculated based on the variance method.

[0012] Furthermore, in step 3, various constraints include: smoke release capability constraints, cross-domain communication latency constraints, environmental constraints, and platform maneuver constraints. Among these, the platform maneuver constraints include maneuver speed constraints, deployment location constraints, and upwind smoke generation constraints.

[0013] Furthermore, the deentanglement parameterization module adopts a variational autoencoder structure to decompose the global state feature vector into multiple independent feature factors, thereby eliminating the entanglement effects between features of different dimensions.

[0014] Furthermore, the parameterized Q-network adopts a deep neural network structure to calculate the Q-value of each scheduling action based on the unentangled feature factors.

[0015] Furthermore, the Transformer enhanced replay cache module adopts the encoder structure of Transformer to store empirical data during model training, and performs feature extraction and correlation analysis on the empirical data. During sampling, empirical data with high benefits and strong correlation are selected to accelerate the model convergence speed. The structure of the target network is consistent with the structure of the parameterized Q network and is used to periodically copy and update from the parameterized Q network.

[0016] In summary, the present invention has at least one of the following beneficial technical effects: 1. This invention applies the HSUnet model to cross-domain heterogeneous data fusion. By combining a hybrid loss function and a self-attention mechanism, it can effectively solve the problems of modal differences, noise interference, and redundancy in multi-source heterogeneous data, achieve accurate feature extraction and fusion of heterogeneous data, provide high-quality data source support for scheduling decisions, and improve the accuracy of scheduling decisions.

[0017] 2. This invention integrates the DPDQN-TER model, eliminates the influence of feature entanglement through the de-entanglement parameterization module, and improves the efficiency of experience replay by leveraging the Transformer to enhance the replay cache. It solves the problems of slow convergence, local optima, and weak generalization ability of traditional optimization algorithms and ordinary deep reinforcement learning methods in cross-domain heterogeneous cluster scheduling, and can quickly generate the optimal cross-domain collaborative scheduling strategy. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] like Figure 1 As shown, this invention provides a method for scheduling global maximization of smokescreen coverage through cross-domain heterogeneous cluster collaboration, including: Step 1: Utilize the sensor modules mounted on each platform in the cross-domain heterogeneous cluster of sea, land, and air to collect multi-source data in real time, and preprocess the data to obtain standardized multi-source heterogeneous data; Step 2: Input the standardized multi-source heterogeneous data into the HSUnet model, perform feature extraction and fusion through the HSUnet model, and output a global state feature vector including all key state information of the cross-domain heterogeneous cluster. The HSUnet model includes an encoder, a self-attention fusion module and a decoder, and is trained and optimized using a hybrid loss function. Step 3: Based on the global state feature vector and combined with the smoke screen diffusion law, establish a global optimization objective model that includes smoke screen coverage area, cross-domain cooperation cost, coverage unevenness and various constraints. Step 4: The global optimization target model is used as the optimization task of the DPDQN-TER model. The global state feature vector is used as the input of the model. Through model training and inference, the optimal cross-domain collaborative scheduling strategy is generated to maximize the global smoke screen coverage. The DPDQN-TER model consists of a de-entanglement parameterization module, a parameterized Q network, a Transformer enhanced replay buffer module, and a target network.

[0021] Next, the above steps will be explained in detail: In step 1, sensor modules on each platform within the cross-domain heterogeneous cluster (sea, land, and air) are used to collect multi-source data in real time and preprocess it to obtain standardized multi-source heterogeneous data. Specifically: First, it should be noted that the cross-domain heterogeneous cluster of sea, land, and air consists of unmanned surface vessels (USVs), unmanned vehicles (UAVs) on land, and unmanned aerial vehicles (UAVs) in the air. Each platform is equipped with corresponding sensor modules, enabling it to collect multi-source data in real time. Specifically, the multi-source data includes: 1. Self-status data: The three-dimensional position coordinates of each platform are collected via GPS positioning module. (i is the platform number, (Total number of platforms in the cluster), the remaining smoke level and maximum smoke release of each platform are collected through the smoke level detection module. The current maneuver speed of each platform is collected through maneuver status sensors. and maximum speed ; 2. Environmental Status Data: Real-time wind speeds in each area are collected via wind speed sensors. Real-time wind direction is collected through wind direction sensors. Ambient temperature is collected by a temperature sensor, and terrain parameters of each area are collected by a terrain sensor. 3. Target area data: The range, contour coordinates, and location information of key masking points of the target area are collected by the image acquisition module and radar sensor to determine the effective masking range of the target area.

[0022] Each platform transmits the collected multi-source data to the collaborative scheduling center via a cross-domain communication network. Upon receiving the heterogeneous data, the collaborative scheduling center first performs data standardization using the min-max standardization method to eliminate the influence of data units, facilitating subsequent model training and feature extraction. The standardization formula is: ; In the formula, The standardized data values, The original data values ​​to be standardized. The minimum value in this type of original data. This is the maximum value in the original data of this type.

[0023] In step 2, the standardized multi-source heterogeneous data is input into the HSUnet model. The HSUnet model performs feature extraction and fusion, outputting a global state feature vector that includes all key state information of the cross-domain heterogeneous cluster. Specifically: First, the specific structure and working process of the HSUnet model will be introduced: The HSUnet model consists of three parts: an encoder, a self-attention fusion module, and a decoder. It uses a hybrid loss function for training and optimization, which will be explained in detail below: 1. Encoder: The encoder structure adopts UNet, consisting of multiple convolutional and pooling layers. It performs hierarchical feature extraction on the preprocessed multi-source heterogeneous data. The convolutional layers use the ReLU activation function to enhance the model's non-linear fitting ability, and the pooling layers use max pooling to retain key features while reducing feature dimensionality. The feature extraction formula for the convolutional layers is as follows: ; In the formula, It is the feature map output by the convolutional layer. It is the number of convolution kernels. It is the first Each convolutional kernel weight, It is a two-dimensional convolution operation. It is the convolutional layer bias term. The first feature map input to the encoder k Each element value.

[0024] 2. Self-Attention Fusion Module: The feature maps of various data extracted by the encoder are input into the self-attention mechanism module. By calculating the attention weights between different features, the global correlation between heterogeneous data is captured. The module focuses on key features related to smoke screen coverage optimization and performs weighted fusion on various features to obtain a preliminary fused feature vector. The calculation formulas for attention weights and feature fusion are as follows: ; ; In the formula, It is the attention weight matrix. It is a query matrix. It is a key matrix. It is a key matrix transpose, These are the feature dimensions of the query matrix and the key matrix, and Softmax is the normalization function. It is the initial feature vector output after self-attention fusion. It is a value matrix.

[0025] 3. Decoder: The decoder adopts the UNet decoder structure, consisting of multiple deconvolutional and convolutional layers. It upsamples and reconstructs the feature vector output by the self-attention fusion module, outputting a global state feature vector with uniform dimensions. This feature vector contains all the key state information of the cross-domain heterogeneous cluster and is used for subsequent global optimization modeling and scheduling strategy generation. The deconvolution upsampling formula and feature reconstruction formula are as follows: ; ; In the formula, It is the upsampled feature map output from the deconvolution layer. These are the deconvolution kernel weights. It is a two-dimensional deconvolution operation. It is the bias term of the deconvolution layer. It is the globally consistent state feature vector output by the decoder. It's a convolution operation. Deconvolutional layers for initial feature vectors The output feature map after upsampling.

[0026] 4. Hybrid loss function optimization: ; In the formula, Cross-entropy loss is used to optimize the feature classification accuracy of the model. Dice loss is used to address data imbalance and improve the accuracy of feature fusion. This is the weighting coefficient for the dice loss, with a value of 0.5.

[0027] By processing the data using the HSUnet model, accurate fusion of multi-source heterogeneous data was achieved, eliminating the impact of noise interference and modal differences. The output global state feature vector provides reliable data source support for subsequent steps.

[0028] In step 3, based on the global state feature vector and combined with the smoke screen diffusion law, a global optimization objective model is established, including smoke screen coverage area, cross-domain collaborative cost, coverage unevenness, and various constraints. Specifically: The smoke screen coverage area, cross-domain collaboration costs, coverage unevenness, and various constraints are discussed in detail, as follows: 1. Smoke screen coverage area The combined area of ​​the coverage areas formed by the smoke screens released from each platform is used. Considering the influence of environmental factors (wind speed, wind direction, topography) on smoke screen diffusion, a modified smoke screen diffusion model is used to calculate the smoke screen coverage area of ​​a single platform. Then, the global smoke screen coverage area is obtained through union operation. for: ; ; In the formula, It is the first The coverage area formed by the smoke screen released by each platform It is the first The initial radius of smoke release from each platform. This is the environmental correction factor, with a value of 0.25. It is the real-time wind speed. It is the diffusion time after the smoke screen is released. yes The union of the smoke screen coverage areas of each platform is calculated, which is the global smoke screen coverage area.

[0029] 2. Costs of cross-domain collaboration It includes three parts: communication cost, platform maneuver cost, and smokescreen consumption cost. The calculation formula is as follows: ; In the formula, , , The weighting coefficients for communication cost, maneuver cost, and smokescreen consumption cost are 0.20, 0.35, and 0.45, respectively. Cross-domain communication costs It is the cost of platform mobility. It's the cost of the smokescreen.

[0030] 3. Calculate the global smoke screen coverage unevenness using the variance method. The target area is divided into multiple uniform grid cells, and the smoke cover intensity and non-uniformity of each grid cell are calculated. The variance of the coverage intensity of each grid cell is calculated using the following formula: ; In the formula, It is the total number of grid cells in the target area. It is the first Smoke coverage intensity of each grid cell It is the average smoke coverage intensity of all grid cells.

[0031] 4. Considering the performance of each platform and the impact of complex environments, the following constraints are determined: (1) Constraints on smoke release capacity: That is, the amount of smoke released by each platform. It cannot exceed its maximum release amount And it cannot be a negative number.

[0032] (2) Cross-domain communication delay constraints: Ensure that cross-domain communication latency between platforms does not exceed the maximum allowable latency. This is to avoid affecting the real-time performance of collaborative scheduling.

[0033] (3) Environmental constraints: When the wind speed exceeds the maximum permissible wind speed If the smoke spreads too quickly to provide effective coverage, stop releasing the smoke. When the wind direction exceeds the effective range, adjust the platform deployment position to ensure that the smoke screen can spread to the target area. For the minimum effective wind direction angle, This represents the maximum effective wind direction angle.

[0034] (4) Platform mobility constraints: Each platform's maneuvering speed must not exceed its maximum maneuvering speed. The deployment location of each platform cannot exceed its effective deployment area. At the same time, the first Azimuth of the release position of each platform with wind angle The included angle satisfy This enables the cross-domain heterogeneous cluster of sea, land, and air to achieve complete downwind smoke generation, ensuring that the smoke screen spreads stably along the wind direction and covers the target area.

[0035] Based on the above analysis, the global optimization objective function is: ; In the formula, These are weighting coefficients, with values ​​of 0.6, 0.25, and 0.15 respectively.

[0036] In step 4, the global optimization target model is used as the optimization task of the DPDQN-TER model, and the global state feature vector is used as the input of the model. Through model training and inference, the optimal cross-domain collaborative scheduling strategy is generated to maximize the global smokescreen coverage. Specifically: First, the DPDQN-TER model is introduced. It consists of four parts: a deentanglement parameterization module, a parameterized Q-network, a Transformer enhanced replay buffer module, and a target network. Specifically: 1. Deentanglement Parameterization Module: Employing a variational autoencoder structure, the global state feature vector is decomposed into multiple independent feature factors, eliminating the entanglement effects between features of different dimensions. This allows the model to more accurately capture the impact of each feature on the scheduling strategy. Through deentanglement parameterization, the global state feature vector is decomposed into multiple independent feature factors. Decomposed into ,in Let h be the number of mutually independent eigenfactors. The unentanglement decomposition formula is as follows: ; ; In the formula, It is the latent feature vector after unentanglement. It is the mean function of the latent eigenvectors. It is the variance function of the latent eigenvectors. Random noise that follows a standard normal distribution. It is the first Each independent feature factor It is the decoder function of the deentanglement module, used to convert the 1st digit of the latent feature vector into the 2nd digit. Each component Decoded into independent feature factors .

[0037] 2. Parameterized Q-Network: Employs a deep neural network structure, with inputs being the de-entangled feature factors and outputs being the scheduling actions. value, The value is used to measure the expected benefit of the action in the current state; parameterization. The network's output layer uses a linear activation function, while the hidden layers use the ReLU activation function. Parameterized modeling enhances the model's ability to fit high-dimensional state spaces. The Q-value calculation function is as follows: ; In the formula, It is the output value of the parameterized Q-network. Parameterize all trainable weight parameters of the Q-network. Deep neural network mapping functions, It is the scheduling action vector.

[0038] 3. Transformer Enhanced Replay Cache Module: Used to store empirical data during model training. ,in This is the current state. For the current action, It is the first The target three-dimensional position coordinates of each platform It is the first The target maneuver speed of each platform It is the first Smoke release from each platform It is the first The smoke screen release launch moment of each platform. For the immediate benefit of the current action, To determine the next state after an action, the Transformer enhanced replay cache module employs the Transformer encoder structure to extract features and perform correlation analysis on the stored empirical data. The empirical data is sorted according to its gain value and correlation, and sampling prioritizes empirical data with high gains and strong correlations. This improves the efficiency and quality of empirical replay and accelerates model convergence. The formulas for calculating empirical correlations and sampling weights are as follows: ; ; In the formula, Current experience status With target sampling state The correlation coefficient, The current state With target sampling state cosine similarity, This is the profit difference adjustment coefficient, with a value of 0.4. These represent the immediate benefits of current experience and target sampling experience, respectively. It is the first The sampling weights of the empirical data. It is the first The state feature vector of the empirical data, It is the real-time state feature vector input to the current model. It is the total amount of experience data stored in the replay cache module. Refers to the first The correlation coefficient between the state feature vector of the empirical data and the real-time state feature vector of the current model input. All in the replay cache module The sum of the correlation coefficients between each empirical data point and the real-time state feature vector.

[0039] 4. Target Network: Structure and Parameterization The network is completely consistent and used to compute the target. To avoid training fluctuations in the parameterized Q-network and improve model stability, the parameters of the target network are periodically copied and updated from the parameterized Q-network. ; In the formula, The target output by the target network Value, corresponding to the next state Next action Expected returns These are all the trainable weight parameters of the target network, periodically derived from parameterization. network Copy and update, The next state The feature factors after decomposition by the deentanglement parameterization module The next state The corresponding optimal scheduling action vector. The deep neural network mapping function refers to the target network.

[0040] The overall process will be explained below: The global state feature vector is input into the de-entanglement parameterization module to obtain the de-entangled feature factors, which are then parameterized. The network outputs each scheduling action based on the feature factors. Value, using Greedy strategy for selecting scheduling actions ,in To explore probabilities, used to balance the exploration and exploitation of the model. Execute actions. Get immediate benefits and the next state , to use empirical data Stored to the Transformer Enhanced Playback Cache module.

[0041] When the amount of empirical data in the replay buffer module reaches a preset quantity, a batch of empirical data is sampled through the Transformer-enhanced replay buffer module. The sampled empirical data is then input into the de-entanglement parameterization module to obtain the de-entangled feature factors, which are then parameterized. The network calculates the current state based on the feature factors of the current state. The target network calculates the target value based on the feature factors of the next state. The model parameters are optimized using the temporal difference error loss function to minimize the current value. Values ​​and Objectives The error between values. The formula for the time-series difference error loss function is as follows: ; in, This is the error loss value of the parameterized Q-network. It is the batch size of empirical data from a single sampling. It is a parameterized Q-network for the first 100 empirical data points Current Q-value prediction, It is the first The immediate benefits of a piece of experience data This is a discount factor with a value of 0.9. Is the target network in the 1st epoch? The next state of an experience The maximum target Q value corresponding to all possible actions. It is the first In the empirical data, the actions performed The next state feature vector.

[0042] Parameterization will be performed according to the update cycle T. The network parameters are copied to the target network, and the parameters of the target network are updated. This process is repeated until the model converges, at which point the model can stably output the optimal scheduling policy.

[0043] After the model training converges, the global state feature vector, which is collected in real time and fused with HSUnet, is input into the DPDQN-TER model. The de-entanglement parameterization module performs de-entanglement processing and parameterization on it. The network outputs each scheduling action. Value, selection The action with the highest value is used as the optimal scheduling strategy, which realizes efficient collaboration of cross-domain heterogeneous clusters across land, sea and air. It solves the problems of poor data fusion effect, weak global optimization capability and insufficient dynamic adaptability of existing technologies, realizes global maximization of smoke screen coverage, and improves the smoke screen shielding and protection capability in complex scenarios.

[0044] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0045] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0046] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0047] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0048] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A method for scheduling global maximization of smokescreen coverage through cross-domain heterogeneous cluster collaboration between sea, land, and air, characterized in that, include: Step 1: Utilize the sensor modules mounted on each platform in the cross-domain heterogeneous cluster of sea, land, and air to collect multi-source data in real time, and preprocess the data to obtain standardized multi-source heterogeneous data; Step 2: Input the standardized multi-source heterogeneous data into the HSUnet model, perform feature extraction and fusion through the HSUnet model, and output a global state feature vector including all key state information of the cross-domain heterogeneous cluster. The HSUnet model includes an encoder, a self-attention fusion module and a decoder, and is trained and optimized using a hybrid loss function. Step 3: Based on the global state feature vector and combined with the smoke screen diffusion law, establish a global optimization objective model that includes smoke screen coverage area, cross-domain cooperation cost, coverage unevenness and various constraints. Step 4: The global optimization target model is used as the optimization task of the DPDQN-TER model. The global state feature vector is used as the input of the model. Through model training and inference, the optimal cross-domain collaborative scheduling strategy is generated to maximize the global smoke screen coverage. The DPDQN-TER model consists of a de-entanglement parameterization module, a parameterized Q network, a Transformer enhanced replay buffer module, and a target network.

2. The method for scheduling global maximization of smokescreen coverage through cross-domain heterogeneous cluster collaboration between sea, land, and air as described in claim 1, is characterized in that... The cross-domain heterogeneous cluster of sea, land, and air includes unmanned surface vessels, unmanned vehicles, and unmanned aerial vehicles. The multi-source data includes the platform's own status data, environmental status data, and target area data.

3. The method for scheduling global maximization of smokescreen coverage through cross-domain heterogeneous cluster collaboration between sea, land, and air, as described in claim 1, is characterized in that... In step 1, the multi-source data is preprocessed, specifically as follows: The min-max standardization method is used to standardize the multi-source data to obtain standardized multi-source heterogeneous data.

4. The method for scheduling global maximization of smokescreen coverage through cross-domain heterogeneous cluster collaboration between sea, land, and air, as described in claim 1, is characterized in that... In step 2, the encoder adopts the encoder structure of UNet and is used to perform hierarchical feature extraction on the standardized multi-source heterogeneous data. The self-attention fusion module is used to fuse the features extracted by the encoder; the decoder adopts the UNet decoder structure to upsample and reconstruct the fused feature vector and output a global state feature vector with uniform dimensions.

5. The method for scheduling global maximization of smokescreen coverage through cross-domain heterogeneous cluster collaboration between sea, land, and air as described in claim 1, characterized in that... In step 2, the hybrid loss function is: ； In the formula, Cross-entropy loss is used to optimize the feature classification accuracy of the model. Dice loss is used to address data imbalance and improve the accuracy of feature fusion. This is the weighting coefficient for the dice loss.

6. The method for scheduling global maximization of smokescreen coverage through cross-domain heterogeneous cluster collaboration between sea, land, and air as described in claim 1, is characterized in that... In step 3, the smoke screen coverage area is the union area of ​​the coverage areas formed after the smoke screens of each platform are released; the cross-domain collaboration cost includes communication cost, platform maneuver cost, and smoke screen consumption cost; the coverage unevenness is calculated based on the variance method.

7. The method for scheduling global maximization of smokescreen coverage through cross-domain heterogeneous cluster collaboration between sea, land, and air, as described in claim 1, is characterized in that... In step 3, various constraints include: smoke release capability constraints, cross-domain communication latency constraints, environmental constraints, and platform maneuver constraints. Among them, the platform maneuver constraints include maneuver speed constraints, deployment location constraints, and upwind smoke generation constraints.

8. The method for scheduling global maximization of smokescreen coverage through cross-domain heterogeneous cluster collaboration between sea, land, and air as described in claim 1, characterized in that, The de-entanglement parameterization module adopts a variational autoencoder structure to decompose the global state feature vector into multiple independent feature factors, thereby eliminating the entanglement effects between features of different dimensions.

9. A method for scheduling global maximization of smokescreen coverage through cross-domain heterogeneous cluster collaboration between sea, land, and air, as described in claim 8, is characterized in that... The parameterized Q-network adopts a deep neural network structure and is used to calculate the Q-value of each scheduling action based on the unentangled feature factors.

10. A method for scheduling global maximization of smokescreen coverage through cross-domain heterogeneous cluster collaboration between sea, land, and air, as described in claim 9, is characterized in that... The Transformer enhanced replay cache module adopts the encoder structure of Transformer to store empirical data during model training and to perform feature extraction and correlation analysis on the empirical data; the structure of the target network is consistent with the structure of the parameterized Q network and is used to periodically copy and update from the parameterized Q network.

Images