Heterogeneous unmanned equipment resource recommendation method based on large model and attention network

Abstract

The invention discloses a heterogeneous unmanned equipment resource recommendation method based on a large model and an attention network, and the method comprises the steps: carrying out the narrative scene generation based on a large-scale language model, i.e., deducing the causal relationship between event concepts through the large-scale language model according to event cue word narration, and generating a logic scene, outputting task target set data of logic scene structuring; task target set data is converted into a standardized numeric feature matrix through feature engineering, and a high-quality training sample pair is generated in combination with a corresponding heterogeneous unmanned equipment resource configuration scheme; training a deep attention recommendation model by using the high-quality training sample pair, wherein the deep attention recommendation model comprises a target individual feature encoder, a multi-head attention fusion module and a resource configuration recommendation layer; and obtaining task target set data of the target scene, and obtaining a configuration scheme of the heterogeneous unmanned equipment resource recommendation quantity of the target scene. According to the invention, high intelligence of large-scale heterogeneous unmanned equipment resource scheduling is realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent scheduling and optimal allocation technology of unmanned equipment resources, specifically involving a heterogeneous unmanned equipment resource recommendation method based on a large model and attention network. Background Technology

[0002] With the increasing strategic importance of the ocean, the complexity and scale of maritime operations are constantly increasing. Missions such as maritime search and rescue, resource exploration, and environmental monitoring often require the coordinated operation of various types of unmanned equipment (such as drones, unmanned surface vessels, and unmanned underwater vehicles). The key to improving operational efficiency and reducing costs lies in how to quickly, efficiently, and intelligently configure these heterogeneous unmanned equipment resources according to dynamically changing mission requirements.

[0003] Existing technologies mainly suffer from the following limitations: Methods based on rules or traditional optimization algorithms. For example, resource allocation based on expert systems, genetic algorithms, or particle swarm optimization. These methods can achieve certain results when the task scenarios are relatively simple and fixed. However, when facing large-scale, highly dynamic, and multi-objective complex tasks, they often suffer from low computational efficiency, difficulty in real-time response, and poor adaptability to unexpected events. Furthermore, algorithm models usually require extensive manual design and adjustment, resulting in limited generalization ability. Traditional machine learning methods. Some studies have attempted to use traditional machine learning models for decision support, but these models struggle to capture the deep, non-linear, and complex relationships between task objectives and resource capabilities. For example, they may fail to effectively distinguish the priorities of different task objectives, leading to insufficient response to key objectives. General deep learning methods. While deep learning excels in pattern recognition, standard neural networks often employ simple information aggregation methods when dealing with complex situations consisting of multiple independent task objectives, failing to dynamically measure the importance of each task objective, thus affecting the accuracy of the final resource allocation scheme. Training data acquisition and quality bottlenecks. The success of deep learning methods highly depends on large-scale, high-quality labeled data. In the field of unmanned equipment resource scheduling, obtaining real and diverse mission scenario data is extremely difficult and costly.

[0004] Therefore, there is an urgent need for an intelligent recommendation method that can deeply understand complex mission situations, dynamically assess the importance of each objective, and be trained based on high-fidelity simulation data to quickly generate high-quality configuration schemes, in order to solve the bottlenecks in efficiency, accuracy and generalization ability in the current resource allocation of heterogeneous unmanned equipment. Summary of the Invention

[0005] The purpose of this invention is to provide a resource recommendation method for heterogeneous unmanned equipment based on a large model and attention network. This method aims to solve the problems of low resource allocation efficiency, low intelligence, untimely response to key targets, and poor generalization ability of existing technologies when dealing with complex dynamic tasks, due to insufficient model capabilities and low training data quality.

[0006] According to a first aspect of the present invention, a heterogeneous unmanned equipment resource recommendation method based on a large model and an attention network is provided, the method comprising: Narrative scene generation based on large-scale language models involves using large-scale language models to infer causal relationships between event concepts based on event prompts, generating logical scenes, and outputting structured task target set data of logical scenes. Feature engineering is used to convert the task target set data into a standardized numerical feature matrix, and combined with the corresponding heterogeneous unmanned equipment resource allocation scheme, high-quality training sample pairs are generated. A deep attention recommendation model is trained using high-quality training samples. The deep attention recommendation model includes: The target individual feature encoder is used to map the feature vector of each task target in the feature matrix to a high-dimensional feature embedding vector, and combine the high-dimensional feature embedding vectors of all task targets to form an embedding matrix. The multi-head attention fusion module is used to perform self-attention calculation on the embedding matrix using the multi-head attention mechanism, dynamically evaluate the importance of each task objective in the global situation, and output an attention output matrix containing contextual information. The resource allocation recommendation layer is used to aggregate the attention output matrix and map it to obtain a vector of recommended resource quantities for heterogeneous unmanned equipment. Obtain the task target set data of the target scene, use the trained deep attention recommendation model to generate a heterogeneous unmanned equipment resource recommendation quantity vector, and finally obtain the configuration scheme of the heterogeneous unmanned equipment resource recommendation quantity of the target scene.

[0007] Furthermore, narrative scene generation based on large-scale language models specifically includes: Input: Natural language instructions are input into a pre-trained large language model to leverage the natural language understanding and generation capabilities of the large language model to transform event prompts into structured task target set data. The natural language instructions describe the event type and requirements of the scene through event prompts and specify the scene parameters that need to be set by the large language model. The scene requirements include a scene description and a description of the task targets in the scene. The scene parameters include the location, priority, and environmental parameters of the task targets. Processing: Large language models, based on their knowledge base, understand the causal relationships between event concepts in natural language instructions and deduce a logical scene layout; Output: A large language model generates and outputs a structured set of task objectives that describes the entire event in detail and contains the event's inherent logic.

[0008] Furthermore, feature engineering includes feature extraction and feature matrix construction and filling; The feature extraction is as follows: For a given task scenario, determine the set of task objectives; For each task objective, extract one D f The feature vector contains spatial location features, dynamic features, and attribute features; among which, dynamic features include velocity and heading codes, and attribute features include target type and priority. Set a maximum number of task objectives N max The feature vectors of all task objectives in the task scenario are integrated into a shape of ( N max , D f ) characteristic matrix X If the total number of task objectives in the task scenario is N < N max If the value is zero, then fill it with the zero vector.

[0009] Furthermore, feature engineering also includes feature standardization, which involves standardizing the feature matrix.

[0010] Furthermore, for each set of mission objectives, an expert rule system or optimization algorithm is used to calculate a corresponding heterogeneous unmanned equipment resource allocation scheme.

[0011] Furthermore, the target individual feature encoder includes: Received feature matrix X' , characteristic matrix X' Each row represents a task objective; the individual target feature encoder consists of several fully connected layers and a ReLU activation function, independently processing the feature matrix. X' Each row represents the feature vector of each task objective. f' i Mapped to d model High-dimensional feature embedding vector h i : h i = Encoder(f' i ) = ReLU( W 2· ReLU( W 1· f' i + b 1) + b 2); In the formula, W 1 and W 2 is the learnable weight matrix. b 1 and b 2 represents a learnable bias; The high-dimensional feature embedding vectors of all task objectives constitute the embedding matrix. H Its shape is ( N max , d model ); The multi-head attention fusion module includes: For embedding matrix H Self-attention computation is performed to dynamically evaluate the importance of each task objective in the global situation. First, the computation of a single attention head is defined, which involves transforming the embedding matrix using a linear transformation. H Mapping to query, key, and value spaces: Q = H · W Q ; K = H · W K ; V = H · W V ; In the formula, Q It's a query. K It is a key. V It is worth it. W Q , W K and W V It is a learnable weight matrix; The output of the attention head is calculated using a scaled dot product attention mechanism: Attention Q , K , V = softmax( ( Q · K ) / sqrt( dk ) ) · V ; Among them, Attention ( Q , K , V ) represents the output of a single attention head; d k It is the dimension of the key vector. sqrt This represents the square root operation; softmax is the activation function. T Indicates matrix transpose; Multi-head attention mechanisms include k One point of attention, k Multiple attention heads execute in parallel, and each attention head learns different subspace features; The outputs of all attention heads are concatenated and subjected to a linear transformation to obtain the final attention layer output Z: Z = MultiHead( H = Concat(head1, ..., head) k ) · W O ; In the formula, MultiHead represents the multi-head attention mechanism, and head i Indicates the first i The output of each attention head; Concat indicates concatenation; W O The weight matrix is ​​a learnable matrix; Output Z The shape is still ( N max , d model It contains contextual information about the relationships between all task objectives; The resource allocation recommendation layer includes: The resource allocation recommendation layer is a feedforward neural network that outputs attention. Z By aggregating the vectors, we obtain a vector representing the global situation. c : c = Mean( Z ); In the formula, Mean represents average pooling; global situation vector c Input resource configuration recommendation layer OutputLayer, get a M dimensional output vector Y pred : Y pred = OutputLayer(c ) = ReLU( W out · c + b out ); In the formula, W out It is the weight matrix of the input resource allocation recommendation layer. b out It is the bias of the input resource configuration recommendation layer; Heterogeneous Unmanned Equipment Resource Recommendation Quantity Vector Y pred Each value in Y predj Represents the corresponding number j The recommended quantity of resources for each type of unmanned equipment is determined, and the ReLU activation function is used to ensure that the output value is non-negative. M This represents the total number of heterogeneous unmanned equipment resource types.

[0012] Furthermore, the mean squared error loss function is adopted, and the parameters of the deep attention recommendation model are iteratively updated through backpropagation.

[0013] According to a second aspect of the present invention, a heterogeneous unmanned equipment resource recommendation system based on a large model and attention network is provided, applying the heterogeneous unmanned equipment resource recommendation method based on a large model and attention network as described in any one of the first aspects. The system includes: The large-model-based scene generation module is used to generate narrative scenes based on a large language model. That is, the large language model infers the causal relationship between event concepts based on the event prompt words to generate logical scenes and outputs structured task target set data of logical scenes. The multimodal feature engineering module is used to convert the task target set data into a standardized numerical feature matrix through feature engineering, and generate high-quality training sample pairs by combining the corresponding heterogeneous unmanned equipment resource configuration scheme. A deep attention recommendation model module is used to train a deep attention recommendation model using high-quality training samples. The deep attention recommendation model includes: The target individual feature encoder is used to map the feature vector of each task target in the feature matrix to a high-dimensional feature embedding vector, and combine the high-dimensional feature embedding vectors of all task targets to form an embedding matrix. The multi-head attention fusion module is used to perform self-attention calculation on the embedding matrix using the multi-head attention mechanism, dynamically evaluate the importance of each task objective in the global situation, and output an attention output matrix containing contextual information. The resource allocation recommendation layer is used to aggregate the attention output matrix and map it to obtain a vector of recommended resource quantities for heterogeneous unmanned equipment. The inference application module is used to acquire the task target set data of the target scene, generate a heterogeneous unmanned equipment resource recommendation quantity vector using a trained deep attention recommendation model, and finally obtain the configuration scheme of the heterogeneous unmanned equipment resource recommendation quantity of the target scene.

[0014] According to a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the heterogeneous unmanned equipment resource recommendation method based on large models and attention networks as described in any one of the first aspects.

[0015] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the heterogeneous unmanned equipment resource recommendation method based on a large model and attention network as described in any one of the first aspects.

[0016] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: Strong adaptability and generalization ability: By introducing a large model as a data engine, the core bottlenecks of unavailable real-world data and low-quality traditional simulation data are solved. The model can be trained in massive, diverse and highly realistic scenarios, effectively bridging the gap between simulation and reality, giving it strong generalization ability, and enabling it to effectively cope with various complex and sudden task scenarios that are not seen during training but conform to real-world logic.

[0017] High decision-making quality: By introducing a multi-head attention mechanism, the model can accurately identify key objectives in complex tasks, make more targeted resource allocations, and significantly improve the completion rate of key tasks in complex and dynamic environments.

[0018] Fast response speed: The end-to-end model design and optional parallel inference framework enable the system to respond to new task situations in near real-time, meeting the needs of high timeliness scenarios.

[0019] High level of automation and intelligence: It realizes full-process automation from the original situation to the recommended solution, reduces the reliance on human expert experience, and achieves a high degree of intelligence in the large-scale unmanned equipment resource scheduling without the need for real data. Attached Figure Description

[0020] Figure 1This is a schematic diagram of the overall framework of a heterogeneous unmanned equipment resource recommendation method based on a large model and attention network according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the process of generating a dynamic scene based on a large model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a deep attention recommendation model according to an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the performance of a method according to an embodiment of the present invention; Figure 5 This is a schematic block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0021] The following detailed description is merely exemplary in nature and is not intended to limit the disclosed technology or its application and use. Furthermore, it is not intended to be bound by any express or implied theory presented in the foregoing technical fields, background art, or the following detailed description.

[0022] In the following detailed description of the embodiments, numerous specific details are set forth in order to provide a more thorough understanding of the disclosed technology. However, it will be apparent to those skilled in the art that the disclosed technology can be practiced without these specific details. In other instances, well-known features have not been described in detail to avoid unnecessarily complicating the description.

[0023] Terms such as "comprising" and "including" indicate that, in addition to the units and steps that are directly and explicitly stated in the specification, the technical solution of the present invention does not exclude the presence of other units and steps that are not directly or explicitly stated. Terms such as "first" and "second" do not indicate the order of the units in terms of time, space, size, etc., but are merely used to distinguish the units.

[0024] This invention provides an intelligent recommendation method for heterogeneous unmanned equipment resources based on a large model and a deep attention network. It combines a large model with deep learning and utilizes an attention mechanism to collaboratively optimize and intelligently recommend resources for heterogeneous unmanned equipment such as aerial, surface, and underwater systems. This method constructs an end-to-end deep learning model that can automatically learn from raw mission situation data and directly output recommended unmanned equipment resource configuration schemes. Its core points are: Introducing large-scale models to construct high-fidelity training data: Addressing the fundamental challenges of obtaining real-world data and the low quality of traditional simulation data, this invention utilizes large-scale language models to generate diverse and realistic task scenarios with complex causal logic through high-level instructions, thereby creating high-quality training data. This solves the data dependency problem in deep learning model training, effectively bridging the gap between simulation and reality, and enhancing the model's generalization ability and robustness in the real world.

[0025] Introducing a multi-head attention mechanism: The model can simulate the decision-making process of human experts. When analyzing the overall situation, it automatically and dynamically allocates different "attention levels" to different task objectives (such as rescued people in different locations or exploration points of different values), giving priority to high-priority or high-threat objectives, thereby generating more targeted and efficient resource allocation schemes.

[0026] This invention achieves end-to-end processing of "situational understanding-intelligent recommendation": Unlike traditional multi-stage and multi-module separation methods, this invention integrates feature extraction, situational understanding and solution recommendation into a unified neural network framework, reducing information loss during the transmission of information between different modules and improving the integrity of decision-making and response speed.

[0027] Therefore, this invention can be widely applied to various civilian scenarios, such as maritime search and rescue, marine resource exploration, and marine environmental monitoring.

[0028] like Figure 1 As shown, the heterogeneous unmanned equipment resource recommendation method based on large models and attention networks of the present invention specifically includes the following steps: Narrative scene generation based on large-scale language models involves using large-scale language models to infer causal relationships between event concepts based on event prompts, generating logical scenes, and outputting structured task target set data of logical scenes. Feature engineering is used to convert the task target set data into a standardized numerical feature matrix, and combined with the corresponding heterogeneous unmanned equipment resource allocation scheme, high-quality training sample pairs are generated. A deep attention recommendation model is trained using high-quality training samples. The deep attention recommendation model includes: The target individual feature encoder is used to map the feature vector of each task target in the feature matrix to a high-dimensional feature embedding vector, and combine the high-dimensional feature embedding vectors of all task targets to form an embedding matrix. The multi-head attention fusion module is used to perform self-attention calculation on the embedding matrix using the multi-head attention mechanism, dynamically evaluate the importance of each task objective in the global situation, and output an attention output matrix containing contextual information. The resource allocation recommendation layer is used to aggregate the attention output matrix and map it to obtain a vector of recommended resource quantities for heterogeneous unmanned equipment. Obtain the task target set data of the target scene, use the trained deep attention recommendation model to generate a heterogeneous unmanned equipment resource recommendation quantity vector, and finally obtain the configuration scheme of the heterogeneous unmanned equipment resource recommendation quantity of the target scene.

[0029] This invention uses a deep learning model to intelligently recommend the required configuration of various heterogeneous unmanned equipment resources and their quantities based on a series of input task objective information.

[0030] The framework of the heterogeneous unmanned equipment resource recommendation system based on large models and attention networks of this invention is as follows: Figure 1 As shown, the system mainly comprises four core modules: a scene generation module based on a large model, a multimodal feature engineering module, a deep attention recommendation model module, and an inference application module. The system receives externally input task target data, processes it, and outputs a configuration scheme containing recommended quantities of various unmanned equipment resources.

[0031] (1) Scene generation module based on large model To address the scarcity of real-labeled data and the low quality of traditional simulated data, such as Figure 2 As shown, this module introduces large models, aiming to create large-scale, diverse, high-fidelity training scenarios with inherent logic.

[0032] ① Narrative scene generation based on large-scale language models: This process leverages the powerful natural language understanding and generation capabilities of large models to transform high-level event descriptions into structured task data.

[0033] Input: Input a high-level natural language instruction (Prompt) into a pre-trained large-scale language model (LLM). For example: Generate a scenario of a chain of maritime accidents caused by Typhoon Haishen. Scenario requirements: A cargo ship breaks apart and sinks, causing containers to float and some crew members to fall into the water; simultaneously, a nearby small fishing boat loses contact due to severe sea conditions. Please set reasonable initial positions, priorities, and environmental parameters.

[0034] Processing: Based on its knowledge base, LLM understands the causal relationships between concepts such as "typhoon" and "chain reaction" and automatically deduce a logical scene layout.

[0035] Output: LLM generates structured data (such as JSON format) that describes the entire event in detail. This data not only contains random parameters but also the event's inherent logic.

[0036] ② Generation and labeling of training sample pairs: The complex, high-fidelity scenes created by large models are transformed into training sample pairs (X, Y_true) that can be used by the models.

[0037] Input features (X): The set of task objectives described in the JSON is transformed into a normalized feature matrix X through subsequent feature engineering.

[0038] Output label (Y_true): For this large model, which generates a more realistic complex scenario, use an expert rule system or a small-scale optimization algorithm to calculate an "ideal" resource allocation scheme for unmanned equipment, Y_true.

[0039] ③ Data persistence: The generated scene data is stored in a structured format to form a large-scale, high-quality training dataset containing both simple and complex scenes.

[0040] (2) Multimodal feature engineering module This module utilizes feature engineering for heterogeneous data to transform raw, diverse task target data into standardized numerical tensors that can be processed by neural networks.

[0041] ①Feature extraction: For a given task scenario, including N The set of task objectives is represented as: T = { t 1, t 2, ..., t n}

[0042] From each task objective t i Extract one D f 3D eigenvectors f i : f i = [ p x , p y , p z , v sin(θ), cos(θ), prio , type , ...]; in, p x , p y , p zSpatial location features; v sin(θ) and cos(θ) are dynamic features (velocity and heading encoding); prio , type These are attribute features (such as priority and target type), which are mapped to numerical values.

[0043] ② Construction and filling of the feature matrix: Due to the number of task objectives in each scenario N Not fixed, set a maximum target number N max The target features of each scene are integrated into a shape of ( ). N max , D f The input matrix of ) X If the actual target number N < N max Then, use a zero vector to fill in the blanks to ensure uniform size.

[0044] ③ Feature standardization: To accelerate model convergence, the feature matrix is... X Standardize the data so that its mean is 0 and its variance is 1: X ' = ( X – μ ) / σ ; in, μ and σ These are the mean and standard deviation of the training data along the feature dimension, respectively.

[0045] (3) Core recommendation model based on deep attention network The structure of a deep neural network model is as follows: Figure 3 As shown.

[0046] ①Target Individual Feature Encoder: The model receives the processed feature matrix. X' Each row of the matrix represents a task objective. The encoder (typically composed of several fully connected layers and ReLU activation functions) processes each row independently, converting the original feature vector of each objective into a single row. f' i Mapped to d model High-dimensional "feature embedding" vector h i : h i = Encoder( f'i ) = ReLU( W 2· ReLU( W 1· f' i + b 1) + b 2); In the formula, W 1 and W 2 is the learnable weight matrix. b 1 and b 2 represents a learnable bias.

[0047] The embedding vectors of all targets constitute the embedding matrix. H Its shape is ( N max , d model ).

[0048] ② Multi-head attention fusion module: This module is a key innovation of this invention, used to dynamically assess the importance of each target in the current global situation. It affects the embedding matrix. H Perform self-attention computation. First, define the computation of a single attention head. This is done through a linear transformation... H Mapping to the query, key, and value space: Q = H · W Q , K = H · W K , V = H · W V ; in, W Q , W K , W V It is a learnable weight matrix.

[0049] The output of the attention head is calculated using a scaled dot product attention mechanism: Attention Q , K , V = softmax( ( Q · K ) / sqrt( d k ) ) ·V ; in, d k It is the dimension of the key vector, used for scaling to prevent the gradient from being too small; sqrt This represents the square root operation; softmax is the activation function. T Represents matrix transpose; Attention ( Q , K , V ) represents the output of a single attention head.

[0050] Multi-head attention mechanisms execute the above processes in parallel. k Second-rate( k Each head learns different subspace features: head i = Attention( H · W Q i , H · W K i , H · W V i ); In the formula, head i Indicates the first i The output of each attention head.

[0051] Then, the outputs of all heads are concatenated and linearly transformed to obtain the final attention layer output Z: MultiHead( H = Concat(head1, ..., head) k ) · W O ; Among them, W Q i 、 W K i 、 W V i 、 W O All are learnable weight matrices; MultiHead represents a multi-head attention mechanism; Concat represents concatenation.

[0052] Output Z The shape is still ( N max , d modelHowever, it contains contextual information about the relationships between all targets.

[0053] ③ Resource allocation recommendation layer (Output Layer): This layer is a feedforward neural network that receives the global situation vector output by the attention module and maps it to the final output. First, the attention output... Z Aggregation is performed, for example through average pooling, to obtain a vector representing the global situation. c : c = Mean( Z ); In the formula, Mean represents average pooling.

[0054] Then, the global situation vector c Input the recommendation layer and get a M dimensional output vector Y pred , M The total number of unmanned equipment resource types possessed by our side: Y pred = OutputLayer( c ) = ReLU( W out · c + b out ); In the formula, W out It is the weight matrix of the input resource allocation recommendation layer. b out It is the bias of the input resource configuration recommendation layer.

[0055] vector Y pred Each value in Y predj Represents the corresponding number j Recommended quantities of each type of unmanned equipment resource. Use the ReLU activation function to ensure the output value is non-negative.

[0056] (4) Model training and optimization ① Data loading and partitioning: The generated simulated dataset is divided into a training set and a validation set.

[0057] ②Loss function: Mean squared error loss is used to measure the amount of resources recommended by the model. Y pred With the "ideal" configuration of the dataset Ytrue The difference between them: L ( Y pred , Y true ) = (1 / M ) · Σ( Y predj – Y truej )²; In the formula, L It indicates a loss.

[0058] ③ Optimizer: The loss function is calculated using modern optimization algorithms such as Adam through backpropagation. L For all learnable parameters of the model θ (including all) W and b The gradient of ) is calculated, and the parameters are iteratively updated to minimize the loss: θ (t+1) = θ t – η · θ · L ( Y pred , Y true ); in, η It is the learning rate.

[0059] ④ Training cycle: Repeat the above process over multiple epochs until the model's performance on the validation set stabilizes and the loss no longer decreases significantly. After training is complete, save the model parameters.

[0060] (5) Reasoning Application Module ① Model Deployment: Deploy the trained model (pth file) and the feature normalization parameters (pkl file) into the real-world application environment.

[0061] ② Parallelization for acceleration: For complex scenarios requiring extremely high real-time performance, different types of tasks (such as aerial targets and surface targets) or task clusters from different geographical regions can be distributed across different computing processes. Each process loads the same model, independently performs inference on its assigned subset of tasks, and finally aggregates the recommendation results from all processes. This parallel processing can significantly reduce the decision latency for large-scale tasks.

[0062] ③ Application Interface: Provide an interface to receive new, real-time task target data (JSON format), invoke the model for inference, and output a clear list of resource recommendations.

[0063] To verify the effectiveness of the method of the present invention, an embodiment of the present invention constructed a simulation experiment of joint maritime search and rescue.

[0064] Experimental scenario: A ship accident occurs in a certain sea area, resulting in multiple search and rescue targets. The details of the mission targets are shown in Table 1.

[0065] Table 1 Task Details Table

[0066] Our resources are shown in Table 2.

[0067] Table 2 Our Resources Table

[0068] Execution process: 1. Input the real-time information (location, type, priority, etc.) of the above 5 search and rescue targets into the system of this invention.

[0069] 2. The feature engineering module converts the data into a standardized feature matrix.

[0070] 3. Deep attention recommendation model processing: In the calculation, the attention module assigns the highest attention weight to the three people who fell into the water with the "emergency" priority, followed by the sunken hull with the "high" priority, while paying the least attention to the wreckage area with the "low" priority.

[0071] 4. Based on this situational understanding of the high level of attention to key objectives, the recommendation layer outputs the following recommended configuration: {Type B vertical take-off and landing search UAV: ​​2 units, Type C rescue material delivery UAV: ​​1 unit, Type D multibeam sonar unmanned underwater vehicle: 1 unit, Type A long endurance reconnaissance UAV: ​​1 unit}.

[0072] Results Analysis. The system prioritized deploying the Type B UAV, best suited for rapid search, and the Type C unmanned surface vessel, capable of rescue operations, to handle the most urgent personnel targets; it also assigned the specialized Type D underwater vehicle to locate the sunken ship; and used a Type A UAV to continuously monitor the entire area (including the low-priority wreckage area).

[0073] Performance comparison. Figure 4The heatmap clearly reveals the adaptability of each method under different task scenarios. Although all methods perform well in simple, static scenarios, the performance of traditional methods (changing from dark green to light green) significantly degrades as task complexity, dynamism, and resource constraints increase. In contrast, this invention, with its attention mechanism, maintains a high coverage of key targets (always dark green) in all demanding scenarios, demonstrating good robustness and wide applicability.

[0074] The present invention proposes an intelligent resource recommendation method for heterogeneous unmanned equipment based on deep attention networks and large models, which has the following technical effects: Strong adaptability and generalization ability: By introducing a large model as a data engine, the core bottlenecks of unavailable real-world data and low-quality traditional simulation data are solved. The model can be trained in massive, diverse and highly realistic scenarios, effectively bridging the gap between simulation and reality, giving it strong generalization ability, and enabling it to effectively cope with various complex and sudden task scenarios that are not seen during training but conform to real-world logic.

[0075] High decision-making quality: By introducing an attention mechanism, the model can accurately identify key objectives in complex tasks, make more targeted resource allocations, and significantly improve the completion rate of key tasks in complex and dynamic environments.

[0076] Fast response speed: The end-to-end model design and optional parallel inference framework enable the system to respond to new task situations in near real-time, meeting the needs of high timeliness scenarios.

[0077] High level of automation and intelligence: It realizes full-process automation from the original situation to the recommended solution, reduces the reliance on human expert experience, and achieves a high degree of intelligence in the large-scale unmanned equipment resource scheduling without the need for real data.

[0078] Figure 5 This is a schematic block diagram of an electronic device according to one or more embodiments of the present invention. The electronic device 200 includes a memory 210, a processor 220, and a computer program 230 stored in the memory 210 and executable on the processor 220. The execution of the computer program 230 causes the vehicle-to-grid frequency control method for calculating time delay compensation in a new energy power system, as described in the above method embodiments, to be performed. Exemplarily, the electronic device 200 may be a controller or a computer.

[0079] Alternatively, the present invention can also be implemented as a computer-readable storage medium storing a program for causing a computer to execute the frequency control method for vehicle-to-grid interaction in a new energy power system with time delay compensation as described in the above-described method embodiments. Here, various types of computer-readable storage media can be used, such as disks (e.g., magnetic disks, optical disks, etc.), cards (e.g., memory cards, optical cards, etc.), semiconductor memory (e.g., ROM, non-volatile memory, etc.), and tapes (e.g., magnetic tape, cassette tape, etc.).

[0080] Where applicable, the various embodiments provided by the present invention may be implemented using hardware, software, or a combination of hardware and software. Furthermore, where applicable, without departing from the scope of the invention, the various hardware and / or software components described herein may be combined into composite components comprising software, hardware, and / or both. Where applicable, without departing from the scope of the invention, the various hardware and / or software components described herein may be divided into sub-components comprising software, hardware, or both. Additionally, where applicable, it is contemplated that software components may be implemented as hardware components, and vice versa.

[0081] Software (such as program code and / or data) according to the invention can be stored on one or more computer-readable storage media. It is also contemplated that the software identified herein can be implemented using one or more networked and / or otherwise general-purpose or special-purpose computers and / or computer systems. Where applicable, the order of the various steps described herein can be changed, combined into compound steps, and / or divided into sub-steps to provide the features described herein.

[0082] The embodiments and examples presented herein are provided to best illustrate embodiments of the invention and its particular applications, thereby enabling those skilled in the art to practice and use the invention. However, those skilled in the art will understand that the above description and examples are provided merely for ease of illustration and example. The descriptions presented are not intended to cover all aspects of the invention or to limit the invention to the precise forms disclosed.

Claims

1. A heterogeneous unmanned equipment resource recommendation method based on a large model and an attention network, characterized in that, The method comprises: narrative scene generation based on a large language model, i.e. the large language model narrates according to event cues, infers the causal relationship between event concepts to generate a logical scene, and outputs a structured task target set data; standardized numerical feature matrix is converted from the task target set data through feature engineering, and a high-quality training sample pair is generated in combination with a corresponding heterogeneous unmanned equipment resource allocation scheme; a deep attention recommendation model is trained using the high-quality training sample pair, the deep attention recommendation model comprising: a target individual feature encoder for mapping the feature vector of each task target in the feature matrix into a high-dimensional feature embedding vector, and combining the high-dimensional feature embedding vectors of all task targets to form an embedding matrix; a multi-head attention fusion module for performing self-attention calculation on the embedding matrix using a multi-head attention mechanism, dynamically evaluating the importance of each task target in the global situation, and outputting an attention output matrix containing context information; a resource allocation recommendation layer for aggregating the attention output matrix and mapping to obtain a heterogeneous unmanned equipment resource recommendation quantity vector; obtaining the task target set data of the target scene, generating the heterogeneous unmanned equipment resource recommendation quantity vector using the trained deep attention recommendation model, and finally obtaining the configuration scheme of the heterogeneous unmanned equipment resource recommendation quantity of the target scene.

2. The method of claim 1, wherein, Narrative scene generation based on a large language model, specifically comprising: input: input natural language instructions into a pre-trained large language model to convert event cue narration into structured task target set data using the natural language understanding and generation capabilities of the large language model; wherein the natural language instructions describe the scene event type and scene requirements through event cue narration, and specify the scene parameters to be set by the large language model, the scene requirements include scene description and task target description in the scene, and the scene parameters include the location, priority and environment parameters of the task target; processing: the large language model understands the causal relationship between event concepts in the natural language instructions according to its knowledge base, and deduces a logical scene layout; output: the large language model generates and outputs structured task target set data that describes the entire event in detail and contains the inherent logic of the event.

3. The method of claim 1, wherein, Feature engineering includes feature extraction and feature matrix construction and filling; wherein the feature extraction is as follows: for a given task scene, determine the task target set; For each task target, extract one D f The feature vector comprises a spatial position feature, a dynamic feature and an attribute feature; the dynamic feature comprises a speed and a heading encoding, and the attribute feature comprises a target type and a priority. Set a maximum number of task goals N max , the feature vectors of all task goals in the task scene are integrated into a feature matrix with a shape of N max , D f ) if the total number of task goals in the task scene is X N < N max , the zero vector is used for padding.​ 4. The method of claim 3, wherein, Feature engineering also includes feature standardization, i.e. standardizing the feature matrix.

5. The method of claim 1, wherein, For each task target set data, a corresponding heterogeneous unmanned equipment resource allocation scheme is calculated using an expert rule system or an optimization algorithm.

6. The method of claim 1, wherein, The target individual feature encoder comprises: receiving the feature matrix X' , the feature matrix X' each row of the feature matrix X' represents a task objective; the objective individual feature encoder Encoder includes several fully connected layers and ReLU activation functions, independently processes each row of the feature matrix X' , and maps the feature vector f' i of each task objective to a d model dimensional high-dimensional feature embedding vector h i : h i = Encoder( f' i ) = ReLU( W 2 · ReLU( W 1 · f' i + b 1) + b 2) wherein W 1 and W 2 are learnable weight matrices, b 1 and b 2 are learnable biases; The high-dimensional feature embedding vectors of all task objectives constitute an embedding matrix H , which is shaped as N max , d model ); The multi-head attention fusion module comprises: On embedding matrices H Self-attention is computed, dynamically assessing the importance of each task objective in the global situation; first, define the computation of a single attention head, i.e. map the embedding matrices H to the query, key and value spaces through linear transformations: Q = H · W Q ； K = H · W K ； V = H · W V ； wherein Q is a query, K is a key, V is a value, W Q , W K and W V is a learnable weight matrix; The output of the attention head is calculated through the scaled dot-product attention mechanism: Attention( Q , K , V ) = softmax( ( Q · K ) / sqrt( d k ) ) · V ; wherein Attention Q , K , V ) denotes the output of a single attention head; d k is the dimension of the key vector, sqrt denotes the square root operation; softmax is an activation function; T denotes the matrix transpose; The multi-head attention mechanism includes k one attention head, k Each attention head learns different subspace features in parallel. The outputs of all attention heads are spliced and linearly transformed to obtain the final attention layer output Z: Z = MultiHead( H ) = Concat(head1,...,head k ) · W O ; In the formula, MultiHead represents a multi-head attention mechanism, head i represents the output of the i first attention head; Concat represents concatenation; W O is a learnable weight matrix. Output Z The shape of the output N max , d model is still a tree (Fig. 2) that contains all the context information after the interdependencies between the task goals are resolved. The resource allocation recommendation layer comprises: The resource configuration recommendation layer is a feedforward neural network that takes the attention outputs Z aggregates them to get a vector representing the global situation c : c = Mean( Z ); wherein Mean represents average pooling; The global state vector is updated as follows c An input resource configuration recommendation layer OutputLayer is inputted to obtain an output vector of dimension M Y pred :​ Y pred = OutputLayer( c ) = ReLU( W out · c + b out ); In the formula, W out is a weight matrix of the input resource configuration recommendation layer, b out is a bias of the input resource configuration recommendation layer. Isomorphic unmanned equipment resource recommendation quantity vector Y pred Each value in Y predj Represents the recommended number of corresponding unmanned equipment resources of the type, and the ReLU activation function is used to ensure that the output value is non-negative; j M The total number of isomorphic unmanned equipment resource types.​ 7. The method of claim 6, wherein, The mean square error loss function is adopted, and the parameters of the deep attention recommendation model are iteratively updated through backpropagation.

8. A heterogeneous unmanned equipment resource recommendation system based on a large model and an attention network, characterized in that, The application discloses a heterogeneous unmanned equipment resource recommendation method based on a large model and an attention network. A large model-based scene generation module is configured to generate a narrative scene based on a large language model, that is, the large language model describes an event prompt word, infers a causal relationship between event concepts to generate a logical scene, and outputs a task target set data in a logical scene structure. A multi-modal feature engineering module is configured to convert the task target set data into a standardized numerical feature matrix through feature engineering, and generate a high-quality training sample pair in combination with a corresponding heterogeneous unmanned equipment resource configuration scheme. A deep attention recommendation model module is configured to train a deep attention recommendation model using the high-quality training sample pair, wherein the deep attention recommendation model comprises: A target individual feature encoder is configured to map a feature vector of each task target in the feature matrix into a high-dimensional feature embedding vector, and combine high-dimensional feature embedding vectors of all task targets to form an embedding matrix. A multi-head attention fusion module is configured to perform self-attention calculation on the embedding matrix using a multi-head attention mechanism, dynamically evaluate the importance of each task target in the global situation, and output an attention output matrix containing context information. A resource configuration recommendation layer is configured to aggregate the attention output matrix and map to obtain a heterogeneous unmanned equipment resource recommendation quantity vector. An inference application module is configured to obtain task target set data of a target scene, generate a heterogeneous unmanned equipment resource recommendation quantity vector using the trained deep attention recommendation model, and finally obtain a configuration scheme of the target scene.

9. An electronic device, comprising: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the steps of the large model and attention network-based heterogeneous unmanned equipment resource recommendation method in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the steps of the large model and attention network-based heterogeneous unmanned equipment resource recommendation method in any one of claims 1-7.

Images