Unmanned aerial vehicle management method based on sensing integration
By deploying deep learning models on drones for control parameters optimization and resource allocation, the problem of drones adaptively tracking targets and providing high-quality communication services in synesthesia integrated operations is solved, and the perceived rate and communication rate are maximized.
Patent Information
- Application Number
- CN202510380279.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-28
- Publication Date
- 2025-06-10
AI Technical Summary
During the integrated synesthesia operation process, drones have difficulties in adaptively tracking targets and providing high-quality communication services to ground users, especially when perceiving the difference in target motion trajectory and the spatiotemporal heterogeneity of ground users' communication requirements.
By using the drone as an integrated aerial synesthesia platform, it generates and executes control parameter sample data, obtains synesthesia performance indicators, and builds a deep learning model for training, deploys it on the drone, and schedules the deep learning model for drone trajectory optimization and synesthesia resource allocation.
It realizes the optimization of control parameters, trajectory optimization and synesthesia resource allocation of drones in the integrated synesthesia operation process, and quickly formulates the optimal joint strategy to maximize the perception rate and communication rate, and ensure perception accuracy.
Smart Images
Figure CN120128951A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of integrated communication and sensing, and particularly relates to a method for managing unmanned aerial vehicles (UAVs) based on integrated communication and sensing. Background Art
[0002] With the 5th Generation Mobile Communication Network (5G) entering the large-scale commercial deployment stage, the global mobile communication technology has entered a new era. However, people have further demands and expectations for future mobile communication technology. For this reason, academia and industry at home and abroad have started to explore the requirements and key technologies of the 6th Generation Mobile Communication Network (6G), aiming to achieve more comprehensive, efficient, and secure mobile communication applications in the future. In the 6G technology vision, full coverage, full spectrum, full application, and strong security are the four most core features. Among the potential key technologies, integrated sensing and communication (ISAC) technology, air-ground integrated networking, etc. are considered very important technologies. Integrated sensing and communication technology, as the name implies, refers to the organic integration of communication technology and sensing technology. By sensing and analyzing elements such as network status, environmental information, and user requirements, intelligent network management and resource optimization can be achieved. With the development of wireless networks, the available spectrum resources are becoming increasingly scarce, and the frequency bands used by communication and radar are gradually converging. At the same time, with the rapid development of digital signal processing technology and hardware, there are many commonalities in the signal processing of communication and radar.
[0003] In an actual network, due to the differences in the movement trajectories of sensing targets and the spatio-temporal heterogeneity of the communication requirements of ground users, it is a difficult problem for UAVs to adaptively track different targets for sensing and at the same time provide high-quality communication services for ground users. At the same time, in order to ensure the communication requirements of the served users and the sensing ability of the target movement state, it is also quite important for UAVs to reasonably allocate power and bandwidth resources to achieve a balance between sensing and communication capabilities. Summary of the Invention
[0004] The present invention provides a method for managing UAVs based on integrated communication and sensing to solve the technical problems existing in the prior art.
[0005] A method for managing UAVs based on integrated communication and sensing includes:
[0006] Use the unmanned aerial vehicle (UAV) as an integrated aerial communication and sensing platform to simultaneously achieve the state perception of the target object and provide communication services for ground users; among them, the height between the UAV and the ground is fixed during the integrated communication and sensing operation.
[0007] Generate sample data of control parameters during the process of the UAV performing the integrated communication and sensing operation task, and execute the sample data of control parameters to obtain the communication and sensing performance indicators of the sample data of control parameters.
[0008] Construct a deep learning model, and use the sample data of control parameters and the communication and sensing performance indicators of the sample data of control parameters to train the deep learning model to obtain the trained deep learning model.
[0009] Deploy the trained deep learning model on the UAV, and during the process of performing the integrated communication and sensing operation task, schedule the deployed deep learning model to optimize the UAV trajectory and allocate communication and sensing resources to complete UAV management.
[0010] Furthermore, generating sample data of control parameters during the process of the UAV performing the integrated communication and sensing operation task, and executing the sample data of control parameters to obtain the communication and sensing performance indicators of the sample data of control parameters, including:
[0011] Generate sample data of control parameters during the process of the UAV performing the integrated communication and sensing operation task; among them, the sample data of control parameters includes angle, horizontal flight distance, bandwidth allocation decision, and power allocation decision; at any moment, fly according to the angle and horizontal flight distance in the sample of control parameters, and the flight trajectory of the UAV can be obtained.
[0012] Arrange all the sample data of control parameters in a random order to obtain the sorted sample data of control parameters; among them, the starting position corresponding to the first sample data of control parameters is the starting position of the UAV, and the starting position corresponding to each other sample data of control parameters is the position of the UAV after executing the previous sample data of control parameters.
[0013] Based on the sorted sample data of control parameters, sequentially obtain the communication and sensing performance indicators corresponding to each sample data of control parameters; among them, the communication and sensing performance indicators include the weighted sum of the total perception rate corresponding to multiple target objects and the total communication rate of multiple ground users.
[0014] Furthermore, constructing a deep learning model includes: constructing a convolutional neural network model to obtain a deep learning model.
[0015] Furthermore, using the sample data of control parameters and the communication and sensing performance indicators of the sample data of control parameters to train the deep learning model to obtain the trained deep learning model, including:
[0016] For all control parameter sample data, based on the sorted control parameter sample data, replace the angle and horizontal flight distance in each control parameter sample data with the UAV position after executing the control parameter sample data to obtain the processed control parameter sample data;
[0017] Initialize the hyperparameters of the deep learning model, encode the hyperparameters of the deep learning model into vectors to obtain model parameter vectors, and repeatedly obtain multiple different model parameter vectors;
[0018] For any model parameter vector, based on the processed control parameter sample data and the corresponding communication-sensing performance metrics, obtain the loss function value corresponding to each model parameter vector;
[0019] According to the loss function values corresponding to each model parameter vector, obtain the optimal model parameter vector;
[0020] For the model parameter vector, based on the optimal model parameter vector, perform local search using the optimal cooperation search strategy to obtain the model parameter vector after local search;
[0021] For the model parameter vector after local search, perform guided search using the multi-information fusion strategy to obtain the model parameter vector after guided search;
[0022] For the model parameter vector after guided search, perform global search using the adaptive mutation strategy to obtain the model parameter vector after global search;
[0023] Determine whether the training end condition is satisfied. If so, based on the model parameter vector after global search, re-obtain the optimal model parameter vector, and use the hyperparameters in the optimal model parameter vector as the final hyperparameters of the deep learning model to obtain the trained deep learning model. Otherwise, return to the step of obtaining the loss function value.
[0024] Furthermore, for any model parameter vector, based on the processed control parameter sample data and the corresponding communication-sensing performance metrics, obtaining the loss function value corresponding to each model parameter vector includes:
[0025] For any model parameter vector, apply the hyperparameters included in the model parameter vector to the deep learning model, use the processed control parameter sample data as the input of the deep learning model, and obtain the actual output of the deep learning model;
[0026] Use the communication-sensing performance metrics corresponding to the control parameter sample data as the expected output of the deep learning model. According to the actual output and the expected output of the deep learning model, use the cross-entropy loss function to obtain the loss function value corresponding to the model parameter vector.
[0027] Furthermore, for the model parameter vector, based on the optimal model parameter vector and adopting the optimal collaborative search strategy for local search, the model parameter vector after local search is obtained, including:
[0028] Determine the adaptive inertia weight based on the current training times;
[0029] Perform local search on the model parameter vector according to the adaptive inertia weight and the optimal model parameter vector to obtain the model parameter vector after local search.
[0030] Furthermore, for the model parameter vector after local search, adopt the multi-information fusion strategy for guided search to obtain the model parameter vector after guided search, including:
[0031] For the model parameter vector after local search, obtain the step size control factor according to the loss function value corresponding to the model parameter vector;
[0032] For the model parameter vector after local search, fuse the model parameter vector with other model parameter vectors to obtain the first fusion information;
[0033] For the model parameter vector after local search, separately fuse the model parameter vector with the optimal model parameter vector to obtain the second fusion information;
[0034] Perform guided search on the model parameter vector according to the step size control factor, the first fusion information, and the second fusion information to obtain the model parameter vector after guided search.
[0035] Furthermore, for the model parameter vector after guided search, adopt the adaptive mutation strategy for global search to obtain the model parameter vector after global search, including:
[0036] Determine the mutation decision probability based on the current training times;
[0037] For any model parameter vector after guided search, randomly generate a mutation decision factor for the model parameter vector between (0, 1);
[0038] Judge whether the mutation decision factor is less than the mutation decision probability. If so, perform mutation search on the model parameter vector to obtain the model parameter vector after global search. Otherwise, directly use the original model parameter vector as the model parameter vector after global search.
[0039] Furthermore, judge whether the training end condition is satisfied, including:
[0040] Determine whether the current number of training times is greater than or equal to the preset maximum number of training times. If so, it is determined that the training end condition is satisfied; otherwise, it is determined that the training end condition is not satisfied.
[0041] Furthermore, deploy the trained deep learning model on the unmanned aerial vehicle (UAV), and during the process of performing the integrated communication and sensing operation task, schedule the deployed deep learning model to optimize the UAV trajectory and allocate communication and sensing resources, including:
[0042] Deploy the trained deep learning model on the UAV;
[0043] At any moment, randomly generate parameter values within the upper and lower limits of the control parameters, form the parameter values into a vector to obtain a control parameter vector, and obtain multiple control parameter vectors;
[0044] For any one of the control parameter vectors, schedule the deployed deep learning model to analyze the control parameter vector and determine the communication and sensing performance indicators corresponding to each control parameter vector;
[0045] With the goal of maximizing the communication and sensing performance indicators, use an intelligent optimization algorithm to optimize multiple control parameter vectors until the optimization number condition is satisfied, and output the optimal control parameter vector;
[0046] Based on the parameter values included in the optimal control parameter vector, perform UAV trajectory optimization and communication and sensing resource allocation.
[0047] A UAV management method based on integrated communication and sensing provided by the present invention can effectively optimize the control parameters of the UAV during the integrated communication and sensing operation process by combining the prediction ability of deep learning technology with control parameter sample data, perform UAV trajectory optimization and communication and sensing resource allocation, and quickly formulate an optimal joint strategy, so as to simultaneously achieve the optimal communication capacity and sensing performance, and finally maximize the sensing rate and communication rate on the premise of ensuring the sensing accuracy. Description of the Drawings
[0048] The drawings herein are incorporated into the specification and form a part of this specification, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention.
[0049] Figure 1 It is a flowchart of a UAV management method based on integrated communication and sensing provided by an embodiment of the present invention.
[0050] Through the above-mentioned drawings, specific embodiments of the present invention have been shown, and there will be more detailed descriptions hereinafter. These drawings and text descriptions are not intended to limit the scope of the inventive concept in any way, but to illustrate the concept of the present invention to those skilled in the art by referring to specific embodiments. Detailed implementation manners
[0051] Here, exemplary embodiments will be described in detail, and examples thereof are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present invention. On the contrary, they are merely examples of apparatuses and methods consistent with some aspects of the present invention as detailed in the appended claims.
[0052] The embodiments of the present invention will be described in detail below with reference to the drawings.
[0053] As Figure 1 shown, a method for managing an unmanned aerial vehicle (UAV) based on integrated communication and sensing provided by an embodiment of the present invention includes:
[0054] S1. Using the UAV as an aerial integrated communication and sensing platform to simultaneously implement state perception of a target object and provide communication services for ground users; wherein, the height between the UAV and the ground is fixed during the integrated communication and sensing operation.
[0055] S2. Generating sample data of control parameters during the process of the UAV executing the integrated communication and sensing operation task, and executing the sample data of the control parameters to obtain the communication and sensing performance indexes of the sample data of the control parameters.
[0056] S3. Constructing a deep learning model, and training the deep learning model by using the sample data of the control parameters and the communication and sensing performance indexes of the sample data of the control parameters to obtain the trained deep learning model.
[0057] S4. Deploying the trained deep learning model on the UAV, and during the process of executing the integrated communication and sensing operation task, scheduling the deployed deep learning model to optimize the UAV trajectory and allocate communication and sensing resources to complete the management of the UAV.
[0058] A method for managing an unmanned aerial vehicle based on integrated communication and sensing provided by the present invention can effectively optimize the control parameters of the UAV during the integrated communication and sensing operation process by combining the prediction ability of deep learning technology with the sample data of the control parameters, optimize the UAV trajectory and allocate communication and sensing resources, quickly formulate an optimal joint strategy, so as to simultaneously achieve the optimal communication capacity and sensing performance, and finally maximize the sensing rate and communication rate on the premise of ensuring the sensing accuracy.
[0059] In the embodiment of the present invention, generating sample data of control parameters during the process of the UAV executing the integrated communication and sensing operation task, and executing the sample data of the control parameters to obtain the communication and sensing performance indexes of the sample data of the control parameters includes:
[0060] Generate control parameter sample data during the process of the drone performing integrated communication and sensing operation tasks; among them, the control parameter sample data includes angle, horizontal flight distance, bandwidth allocation decision, and power allocation decision; at any moment, fly with the angle and horizontal flight distance in the control parameter sample, and the flight trajectory of the drone can be obtained.
[0061] Randomly arrange all the control parameter sample data to obtain the sorted control parameter sample data; among them, the starting position corresponding to the first control parameter sample data is the starting position of the drone, and the starting position corresponding to each other control parameter sample data is the position of the drone after executing the previous control parameter sample data.
[0062] Based on the sorted control parameter sample data, sequentially obtain the communication and sensing performance indicators corresponding to each control parameter sample data; among them, the communication and sensing performance indicators include the weighted sum of the total sensing rates corresponding to multiple target objects and the total communication rates of multiple ground users.
[0063] It should be noted that due to the characteristics of the deep learning model, the input data needs to be relatively fixed. Therefore, the number of target objects and the number of ground users should remain unchanged. Further, when the number of target objects and the number of ground users decrease, the information of the corresponding target objects or ground users can be set to be empty, and optimization can also be achieved, but the optimization effect will decrease. Therefore, data in this case can also be collected to train the deep learning model.
[0064] In the embodiment of the present invention, constructing a deep learning model includes: constructing a convolutional neural network model to obtain a deep learning model. It should be noted that in addition to using a convolutional neural network model to construct a deep learning model, other neural networks can also be used to construct a deep learning model.
[0065] In the embodiment of the present invention, training the deep learning model with the control parameter sample data and the communication and sensing performance indicators of the control parameter sample data to obtain the trained deep learning model, including:
[0066] For all the control parameter sample data, based on the sorted control parameter sample data, replace the angle and horizontal flight distance in each control parameter sample data with the position of the drone after executing the control parameter sample data to obtain the processed control parameter sample data.
[0067] Initialize the hyperparameters of the deep learning model (such as weights, which can be obtained by the method of randomly initializing within the upper and lower limits), and encode the hyperparameters of the deep learning model into a vector to obtain a model parameter vector, and repeat to obtain multiple different model parameter vectors.
[0068] For any model parameter vector, based on the processed control parameter sample data and the corresponding communication-sensing performance metrics, obtain the loss function value corresponding to each model parameter vector;
[0069] According to the loss function value corresponding to each model parameter vector, obtain the optimal model parameter vector;
[0070] For the model parameter vector, based on the optimal model parameter vector, perform local search using the optimal cooperation search strategy to obtain the model parameter vector after local search;
[0071] For the model parameter vector after local search, perform guided search using the multi-information fusion strategy to obtain the model parameter vector after guided search;
[0072] For the model parameter vector after guided search, perform global search using the adaptive mutation strategy to obtain the model parameter vector after global search;
[0073] Determine whether the training end condition is satisfied. If so, based on the model parameter vector after global search, re-obtain the optimal model parameter vector, and use the hyperparameters in the optimal model parameter vector as the final hyperparameters of the deep learning model to obtain the trained deep learning model. Otherwise, return to the step of obtaining the loss function value.
[0074] In the prior art, the gradient descent intelligent optimization algorithm is often used to optimize the deep learning model, which easily leads to being trapped in local optima, further resulting in poor optimization effect of the deep learning model and inability to effectively implement the UAV trajectory and communication-sensing resource allocation. Therefore, the embodiment of the present invention provides a new intelligent optimization algorithm with strong algorithm convergence ability and global search ability, which can effectively improve the UAV trajectory optimization ability and communication-sensing resource allocation ability.
[0075] In the embodiment of the present invention, for any model parameter vector, based on the processed control parameter sample data and the corresponding communication-sensing performance metrics, obtaining the loss function value corresponding to each model parameter vector includes:
[0076] For any model parameter vector, apply the hyperparameters included in the model parameter vector to the deep learning model, and use the processed control parameter sample data as the input of the deep learning model to obtain the actual output of the deep learning model;
[0077] It should be noted that when using the processed control parameter sample data as the input of the deep learning model, the position information of the target object and the ground user should also be incorporated into the input of the deep learning model. That is, the input of the deep learning model includes (the future position information of the drone (i.e., the position information predicted based on the current drone position information, the horizontal flight angle, and the horizontal flight distance in the control parameter sample data), the position information of the target object, the position information of the ground user, bandwidth allocation, and power allocation).
[0078] Use the communication sensing performance index corresponding to the control parameter sample data as the expected output of the deep learning model. According to the actual output and the expected output of the deep learning model, use the cross-entropy loss function to obtain the loss function value corresponding to the model parameter vector.
[0079] In the embodiment of the present invention, for the model parameter vector, based on the optimal model parameter vector and using the optimal collaborative search strategy for local search, the model parameter vector after local search is obtained, including:
[0080] Based on the current training times, determine the adaptive inertia weight as: where w t represents the adaptive inertia weight in the t-th training process, r 1 represents the first random number uniformly distributed in [0, 1], T represents the preset maximum number of training times, w max represents the preset maximum value of the adaptive inertia weight, w min represents the preset minimum value of the adaptive inertia weight;
[0081] According to the adaptive inertia weight and the optimal model parameter vector, perform local search on the model parameter vector, and the model parameter vector after local search is: where represents the n-th model parameter vector in the t-th training process, n = 1, 2,..., N, N represents the total number of model parameter vectors, represents the model parameter vector after local search, represents the optimal model parameter vector, r 2 represents the second random number generated based on the normal distribution N(0, 1), α 1 represents the collaborative control factor, and α 1 = step 1 * exp(-30t / T) 10 step 1 represents the first search step size, and exp represents the exponential function with the natural constant e as the base.
[0082] The optimal collaborative search strategy provided by the present invention can effectively enable the model parameter vector to learn the position information of the optimal model parameter vector in the solution space, and can greatly improve the convergence speed of the algorithm.
[0083] In the embodiment of the present invention, for the model parameter vector after local search, a multi-information fusion strategy is adopted to perform guided search to obtain the model parameter vector after guided search, including:
[0084] For the model parameter vector after local search, the step size control factor is obtained according to the loss function value corresponding to the model parameter vector as: where τ i,d represents the step size control factor corresponding to the d-th dimension parameter of the i-th model parameter vector after local search, d = 1, 2,..., D, D represents the total dimension of the parameters, δ represents the attraction constant between vectors, represents the d-th dimension parameter of the i-th model parameter vector after local search in the t-th training process, represents the d-th dimension parameter of the optimal model parameter vector, dist ibest represents the Euclidean distance between the i-th model parameter vector after local search and the optimal model parameter vector, f best represents the quality degree of the optimal model parameter vector, and the quality degree = 1 / (loss function value + 0.001);
[0085] For the model parameter vector after local search, the model parameter vector is fused with other model parameter vectors to obtain the first fusion information as: where, represents the d-th dimension parameter of the m-th model parameter vector after local search in the t-th training process, represents the d-th dimension parameter of the first fusion information;
[0086] For the model parameter vector after local search, the model parameter vector is separately fused with the optimal model parameter vector to obtain the second fusion information as: where β represents the optimal direction search weight, and π represents pi, r 4 represents a random number between (0, 1), represents the d-th dimension parameter of the second fusion information;
[0087] According to the step size control factor, the first fusion information, and the second fusion information, guided search is performed on the model parameter vector to obtain the model parameter vector after guided search as: where r 3 represents a random number between (0, 1), Denote the d-th parameter of the model parameter vector after the i-th guided search.
[0088] The multi-information fusion strategy provided by the present invention can fuse the information of other model parameter vectors and learn the information at the optimal position, so as to more effectively explore the unknown area within the optimal area, thereby improving the search ability for local optimal solutions and effectively improving the search accuracy of the algorithm even in the later stage of the algorithm.
[0089] In the embodiment of the present invention, for the model parameter vector after the guided search, an adaptive mutation strategy is adopted for global search to obtain the model parameter vector after global search, including:
[0090] Based on the current training times, determine the mutation decision probability as: where v represents the probability control factor, e represents the natural constant, and Γ() represents the complete gamma function;
[0091] For any model parameter vector after the guided search, randomly generate a mutation decision factor for the model parameter vector between (0, 1);
[0092] Judge whether the mutation decision factor is less than the mutation decision probability. If so, perform mutation search on the model parameter vector to obtain the model parameter vector after global search. Otherwise, directly use the original model parameter vector as the model parameter vector after global search.
[0093] The mutation search is:
[0094]
[0095] where represents the model parameter vector after the k-th guided search, represents the model parameter vector the corresponding mutation search value, σ represents the mutation coefficient, UB represents the upper bound vector composed of all dimension upper threshold values, LB represents the lower bound vector composed of all dimension lower threshold values, and λ represents the scaling factor, represents a random number between [-2.5λ, 2.5λ], e represents the natural constant, θ represents the upper limit value of the scaling factor, and ξ represents the adjustment parameter of the scaling factor.
[0096] Optionally, when the loss function value of the mutation search value decreases, the mutation search value can be used as the model parameter vector after global search. Otherwise, the original model parameter vector is directly used as the model parameter vector after global search.
[0097] The adaptive mutation strategy provided by the present invention can provide a relatively large mutation probability in the early stage of the algorithm, enhancing the possibility of the algorithm searching for the global optimal position. In the later stage of the algorithm, there is also a certain probability of mutation, enabling the ability to always jump out of the local optimum.
[0098] Optionally, when the model parameter vector changes each time, boundary crossing processing can be performed on the model parameter vector to ensure that the parameters are always valid.
[0099] In the embodiments of the present invention, determining whether the training end condition is satisfied includes:
[0100] Determining whether the current number of training times is greater than or equal to the preset maximum number of training times. If so, it is determined that the training end condition is satisfied; otherwise, it is determined that the training end condition is not satisfied.
[0101] In the embodiments of the present invention, deploying the trained deep learning model on a drone, and during the process of performing the integrated communication and sensing operation task, scheduling the deployed deep learning model for drone trajectory optimization and communication and sensing resource allocation, including:
[0102] Deploying the trained deep learning model on a drone;
[0103] At any moment, randomly generate parameter values within the upper and lower limits of the control parameters, form a vector with the parameter values to obtain a control parameter vector, and obtain multiple control parameter vectors;
[0104] For any one control parameter vector, schedule the deployed deep learning model to analyze the control parameter vector and determine the communication and sensing performance index corresponding to each control parameter vector;
[0105] With the goal of maximizing the communication and sensing performance index, and using an intelligent optimization algorithm to optimize multiple control parameter vectors until the optimization number condition is satisfied, and output the optimal control parameter vector;
[0106] Based on the parameter values included in the optimal control parameter vector, perform drone trajectory optimization and communication and sensing resource allocation.
[0107] Optionally, the intelligent optimization algorithm can select existing algorithms (such as particle swarm algorithm, genetic algorithm, etc.), or can also select the algorithm provided by the embodiments of the present invention to achieve fast optimization. It should be noted that since the intelligent optimization algorithm generally aims to minimize the loss function value, the reciprocal of the communication and sensing performance index can be taken as the loss function value, so as to achieve intelligent optimization, and then perform drone trajectory optimization and communication and sensing resource allocation according to the optimal control parameters.
[0108] Other embodiments of the present invention will be readily apparent to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. The present invention is intended to cover any variations, uses, or adaptations of the invention following the general principles of the invention and including known common general knowledge or conventional technical means in the technical field not disclosed by the present invention. It should be understood that the present invention is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present invention is only limited by the appended claims.
Claims
1. A drone management method based on synaesthesia integration, characterized in that: include: The drone is used as an aerial synaesthesia integration platform to simultaneously realize the state perception of the target object and provide communication services to ground users; the altitude between the drone and the ground is fixed during the synaesthesia integration operation; Generate control parameter sample data when the UAV performs the synaesthesia integrated operation task, execute the control parameter sample data, and obtain the synaesthesia performance index of the control parameter sample data; Constructing a deep learning model, and training the deep learning model using the control parameter sample data and the synaesthesia performance index of the control parameter sample data to obtain a trained deep learning model; The trained deep learning model is deployed on the drone. In the process of executing the synaesthesia integrated operation task, the deployed deep learning model is scheduled to optimize the drone trajectory and allocate synaesthesia resources to complete the drone management.
2. The drone management method based on synaesthesia integration according to claim 1 is characterized in that: Generate control parameter sample data when the UAV performs the synaesthesia integrated operation task, execute the control parameter sample data, and obtain the synaesthesia performance indicators of the control parameter sample data, including: Generate control parameter sample data when the UAV performs the synaesthesia integration task; the control parameter sample data includes angle, horizontal flight distance, bandwidth allocation decision and power allocation decision; at any time, the UAV is flown at the angle and horizontal flight distance in the control parameter sample to obtain the flight trajectory of the UAV; Arrange all control parameter sample data in random order to obtain sorted control parameter sample data; wherein the starting position corresponding to the first control parameter sample data is the starting position of the UAV, and the starting position corresponding to each other control parameter sample data is the position of the UAV after executing the previous control parameter sample data; Based on the sorted control parameter sample data, the synaesthesia performance index corresponding to each control parameter sample data is obtained in turn; wherein the synaesthesia performance index includes a weighted sum of a total perception rate corresponding to multiple target objects and a total communication rate of multiple ground users.
3. The drone management method based on synaesthesia integration according to claim 1 is characterized in that: Constructing a deep learning model, including: constructing a convolutional neural network model to obtain a deep learning model.
4. The drone management method based on synaesthesia integration according to claim 1 is characterized in that: The deep learning model is trained using the control parameter sample data and the synaesthesia performance index of the control parameter sample data to obtain a trained deep learning model, including: For all control parameter sample data, based on the sorted control parameter sample data, the angle and horizontal flight distance in each control parameter sample data are replaced with the position of the UAV after executing the control parameter sample data, so as to obtain the processed control parameter sample data; Initialize hyperparameters of a deep learning model, encode the hyperparameters of the deep learning model into a vector, obtain a model parameter vector, and repeatedly obtain multiple different model parameter vectors; For any model parameter vector, based on the processed control parameter sample data and the corresponding synaesthesia performance index, the loss function value corresponding to each model parameter vector is obtained; According to the loss function value corresponding to each model parameter vector, the optimal model parameter vector is obtained; For the model parameter vector, according to the optimal model parameter vector, the optimal collaborative search strategy is used to perform a local search to obtain the model parameter vector after the local search; For the model parameter vector after local search, a multi-information fusion strategy is used to conduct guided search to obtain the model parameter vector after guided search; For the model parameter vector after the guided search, an adaptive mutation strategy is used to perform a global search to obtain the model parameter vector after the global search; Determine whether the training end conditions are met. If so, re-obtain the optimal model parameter vector based on the model parameter vector after the global search, and use the hyperparameters in the optimal model parameter vector as the final hyperparameters of the deep learning model to obtain the deep learning model after training. Otherwise, return to the step of obtaining the loss function value.
5. The method for managing unmanned aerial vehicles based on synaesthesia integration according to claim 4 is characterized in that: For any model parameter vector, based on the processed control parameter sample data and the corresponding synaesthesia performance index, the loss function value corresponding to each model parameter vector is obtained, including: For any model parameter vector, apply the hyperparameters contained in the model parameter vector to the deep learning model, and use the processed control parameter sample data as the input of the deep learning model to obtain the actual output of the deep learning model; The synaesthesia performance index corresponding to the control parameter sample data is taken as the expected output of the deep learning model. According to the actual output and expected output of the deep learning model, the cross entropy loss function is used to obtain the loss function value corresponding to the model parameter vector.
6. The method for managing unmanned aerial vehicles based on synaesthesia integration according to claim 4 is characterized in that: For the model parameter vector, a local search is performed based on the optimal model parameter vector and the optimal collaborative search strategy to obtain the model parameter vector after the local search, including: Determine the adaptive inertia weight based on the current number of training times; According to the adaptive inertia weight and the optimal model parameter vector, a local search is performed on the model parameter vector to obtain a model parameter vector after the local search.
7. The method for managing unmanned aerial vehicles based on synaesthesia integration according to claim 6, characterized in that: For the model parameter vector after local search, a multi-information fusion strategy is used to conduct guided search to obtain the model parameter vector after guided search, including: For the model parameter vector after the local search, the step size control factor is obtained according to the loss function value corresponding to the model parameter vector; For the model parameter vector after the local search, the model parameter vector is fused with other model parameter vectors to obtain first fused information; For the model parameter vector after the local search, the model parameter vector is separately fused with the optimal model parameter vector to obtain second fused information; A guided search is performed on the model parameter vector according to the step size control factor, the first fusion information and the second fusion information to obtain the model parameter vector after the guided search.
8. The method for managing unmanned aerial vehicles based on synaesthesia integration according to claim 7 is characterized in that: For the model parameter vector after the guided search, an adaptive mutation strategy is used to perform a global search to obtain the model parameter vector after the global search, including: Determine the probability of mutation decision based on the current number of training times; For any model parameter vector after guided search, a mutation decision factor is randomly generated between (0,1) for the model parameter vector; Determine whether the mutation decision factor is less than the mutation decision probability. If so, perform mutation search on the model parameter vector to obtain the model parameter vector after global search. Otherwise, directly use the original model parameter vector as the model parameter vector after global search.
9. The method for managing unmanned aerial vehicles based on synaesthesia integration according to claim 8, characterized in that: Determine whether the training end conditions are met, including: It is determined whether the current training number is greater than or equal to the preset maximum training number. If so, it is determined that the training end condition is met, otherwise it is determined that the training end condition is not met.
10. The drone management method based on synaesthesia integration according to claim 2 is characterized in that: The trained deep learning model is deployed on the drone. During the execution of the synaesthesia integrated operation task, the deployed deep learning model is scheduled to optimize the drone trajectory and allocate synaesthesia resources, including: Deploy the trained deep learning model on the drone; At any time, randomly generate parameter values within the upper and lower limits of the control parameters, and form the parameter values into vectors to obtain a control parameter vector, and obtain multiple control parameter vectors; For any control parameter vector, schedule the deployed deep learning model to analyze the control parameter vector and determine the synaesthesia performance index corresponding to each control parameter vector; The goal is to maximize the synaesthesia performance index, and use an intelligent optimization algorithm to optimize multiple control parameter vectors until the optimization times condition is met, and then output the optimal control parameter vector; Based on the parameter values contained in the optimal control parameter vector, the UAV trajectory optimization and synaesthesia resource allocation are carried out.