An insulation sub image detection task-oriented precision and energy consumption collaborative optimization method

Abstract

The present application relates to a kind of precision and energy consumption collaborative optimization methods for insulator image detection task, comprising: A) lightweight deep neural network model and high-precision deep neural network model are respectively deployed on unmanned aerial vehicle side and edge server side, and the end side collaborative detection framework with unmanned aerial vehicle-edge server as main body is constructed;B) insulator image detection task generated in real time on unmanned aerial vehicle is modeled according to time slot, and the energy consumption, delay, precision factor handled by unmanned aerial vehicle side and edge server side are modeled;C) the end-to-end delay constraint of task is modeled, and the precision and energy consumption collaborative optimization problem facing task is formalized into the mathematical model of optimization problem;D) construct image detection task unloading algorithm based on particle swarm optimization, obtain optimal task unloading strategy and the resolution configuration of unloading task image processing.This method can meet the delay constraint while making the precision and energy consumption of insulator image detection task reach joint optimization.

Description

Technical Field

[0001] This invention relates to the field of power line inspection technology, specifically to a method for the coordinated optimization of accuracy and energy consumption for insulator image detection tasks. Background Technology

[0002] The normal operation of the power system is a crucial guarantee for national economic development, with the stable operation of transmission lines being particularly critical. As my country's power system continues to develop and electricity consumption increases dramatically, the scale of transmission line construction is also constantly expanding. Transmission line equipment, due to prolonged exposure to the outdoor environment, experiences accelerated wear and tear, making it prone to power outages. For example, insulators exposed to prolonged damp conditions may spontaneously explode, causing power outages and economic losses. Therefore, the inspection and maintenance of transmission line equipment is extremely important. However, traditional manual inspections are inefficient and unsuitable for outdoor environments, failing to meet the inspection needs of large-scale transmission line deployments. To improve inspection efficiency, drones equipped with high-definition cameras can be remotely controlled to acquire aerial images of insulators. Furthermore, the application of deep learning technology has significantly improved the accuracy of automatic insulator identification and defect detection in aerial images.

[0003] Image processing using deep learning typically requires significant computational resources and energy. However, the computational resources and battery capacity of drones used in power line inspection are limited, making it impossible to deploy high-precision deep neural network detection models. Existing technologies usually deploy high-precision models on cloud servers, uploading aerial images of insulators captured by the drone to the cloud server for remote detection. While this method effectively alleviates the limitations of drone computational resources and energy, it is constrained by network bandwidth, and the task offloading process also incurs transmission latency and energy consumption, making it unsuitable for real-time detection. Other technologies have explored task offloading methods to reduce energy consumption and latency, but these models are relatively abstract and do not fully explore the relationship between detection accuracy and image resolution, making them unsuitable for the accuracy and energy consumption optimization problem of the insulator aerial image detection task studied. Summary of the Invention

[0004] The purpose of this invention is to provide a method for jointly optimizing accuracy and energy consumption for insulator image detection tasks. This method can achieve the joint optimality of accuracy and energy consumption for insulator image detection tasks while satisfying time delay constraints.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a method for jointly optimizing accuracy and energy consumption for insulator image detection tasks, comprising the following steps:

[0006] A) Deploy the trained lightweight deep neural network model and high-precision deep neural network model for insulator aerial image detection tasks on the UAV and edge server respectively to build an edge-end collaborative detection framework with UAV-edge server as the main body.

[0007] B) Divide the time into time slots, and model the insulator image detection task generated in real time on the UAV according to the time slots. Model the energy consumption, latency and accuracy factors of the UAV-side processing and edge server-side processing of the insulator image detection task.

[0008] C) Model the end-to-end time delay constraint of the insulator image detection task, and formalize the problem of co-optimization of accuracy and energy consumption for the insulator image detection task into a mathematical model of optimization problem;

[0009] D) Construct an image detection task unloading algorithm based on particle swarm optimization to obtain the optimal task unloading strategy and the image processing resolution configuration for the unloading task.

[0010] Furthermore, step A specifically includes the following steps:

[0011] A1) Data preprocessing: collect aerial images of insulators, label them, and construct a training dataset;

[0012] A2) The lightweight deep neural network MobileNet and the high-precision deep neural network YOLO-V5 are trained using the training dataset constructed in step A1, respectively, to obtain the trained lightweight deep neural network model and the high-precision deep neural network model.

[0013] A3) Deploy the lightweight deep neural network model trained in step A2 on the drone, and deploy the high-precision deep neural network model trained in step A2 on the edge server. The high-precision deep neural network model deployed on the edge server has N input image resolution configuration modes, and data transmission between the drone and the edge server is realized simultaneously.

[0014] Furthermore, step B specifically includes the following steps:

[0015] B1) Divide the time into time slots. Within each time slot t, the UAV generates F frames of images, and each frame is denoted as P. f For each frame image P, f∈1,F], f Modeling is performed to obtain image P f Data volume D f Unloading decision

[0016] B2) Detection accuracy of offloading image detection tasks to edge servers for processing Energy consumption Modeling is performed to assess the detection accuracy of image detection tasks processed on UAVs. Energy consumption Modeling is performed.

[0017] Furthermore, choose r i For image resolution mode calculation where i = 0, 1, ..., N, the data size D of the f-th frame image is... f for:

[0018] D f =σ·r i 2 (1)

[0019] Where σ is a constant, representing the number of data bits carried by one pixel;

[0020] The uninstallation decision is: This indicates that the detection task of the f-th frame image selects the i-th processing mode, i = 0, 1, ..., N, where i = 0 indicates processing at the UAV end, and i = 1, 2, ..., N indicates selecting the i-th processing mode on the edge server;

[0021] For tasks performed on drones, there is only one input image resolution setting and detection accuracy. and energy consumption Let it be a constant, that is

[0022] For tasks offloaded to edge servers for execution, detection accuracy Modeled as a function of the input image resolution, i.e.

[0023]

[0024] Where α and β are constants; when the task is offloaded to the edge server for execution, the energy consumption includes the energy consumption for data transmission from the drone and the energy consumption for execution on the edge server. Ignoring the energy consumption generated on the edge server, the total energy consumption for offloading is... for:

[0025]

[0026] Where ψ is the energy consumed to transmit 1 bit of data.

[0027] Furthermore, step C specifically includes the following steps:

[0028] C1) The delay caused by performing the detection task of the insulator image of the f-th frame on the UAV. The delay caused by offloading to the edge server Model them separately;

[0029] C2) Model the end-to-end average delay, average accuracy, and average energy consumption of the insulator image detection task within time slot t for F frames;

[0030] C3) formally represents the problem of co-optimizing accuracy and energy consumption for insulator image detection tasks as a mathematical model of optimization problems.

[0031] Furthermore, for tasks performed on a drone, the input image resolution is fixed, therefore the execution delay is set to a constant, i.e.

[0032] For tasks offloaded to edge servers for execution, latency consists of two parts: transmission latency and execution latency on the edge server; transmission latency is determined by the amount of data transmitted and network bandwidth. Where B is the network bandwidth within the current time slot t; the edge server execution latency is determined by the input image resolution, specifically:

[0033]

[0034] Where η and θ are constants, the total delay for task unloading calculation is:

[0035] The end-to-end latency L of the image detection task in frame f is... f for:

[0036]

[0037] The average delay for detecting F-frame insulator images is:

[0038]

[0039] The end-to-end energy consumption E of the f-th frame image detection task f for:

[0040]

[0041] The average energy consumption for detecting F-frame insulator images is:

[0042]

[0043] The accuracy A of the image detection task in the f-th frame f for:

[0044]

[0045] The average accuracy of F-frame insulator image detection is:

[0046]

[0047] The problem is defined as co-optimizing average energy consumption and accuracy under the constraint of end-to-end average time delay in F-frame insulator image detection, i.e.

[0048] min A+δE (11)

[0049] Where δ is a constant, representing the trade-off between detection accuracy and energy consumption;

[0050] The constraints are:

[0051]

[0052]

[0053]

[0054] Furthermore, step D specifically includes the following steps:

[0055] D1) Input to the image detection task offloading algorithm based on particle swarm optimization: the amount of data D of each frame f in the F frames generated within time slot t. f The network bandwidth parameter B is used as the algorithm input, and the maximum number of iterations T is set. max ;

[0056] D2) Algorithm initialization: Set the unloading strategy of the F-frame image to the position of a particle, set the image detection accuracy and energy consumption as the objective function, set the boundary of the optimal solution and the velocity and position of the initial sample, and use multiple sets of random data as the initial values ​​of the particle swarm.

[0057] D3) Calculate the fitness of particles in the particle swarm in parallel based on the input, that is, the accuracy value and energy consumption value corresponding to each unloading decision, to obtain the local optimal solution and the global optimal solution;

[0058] D4) Iterates and updates the particle's position based on the local optimum, the global optimum, and the particle velocity;

[0059] D5) Determine if the number of iterations has been reached. If yes, stop the iteration, output the position of the optimal particle, i.e. the optimal unloading decision scheme, and the joint value of image detection accuracy and energy consumption corresponding to the optimal solution; otherwise, continue the iteration.

[0060] Furthermore, in step D2, J groups of particles are set to form a particle swarm, the boundary of the optimal solution is set, and the velocity V of the example is initialized. j and position W j , j∈[1,J];

[0061] Set the unloading strategy for F-frame images Let W be the position of a particle. jWe set the joint A+δE objective function of image detection accuracy and energy consumption as the fitness of the particles;

[0062] In step D3, the fitness of particles in the particle swarm is calculated in parallel based on the input, i.e., the accuracy and energy consumption value corresponding to each unloading decision. Then, the local optimum Q of particle j is selected. j The global optimal solution Q for all particles g ;

[0063] In step D4, the particle's position and velocity are updated based on the local optimum, the global optimum, and the particle's velocity. The calculation formula is as follows:

[0064]

[0065]

[0066] Where ω1 is a constant, and ω2 and ω3 are two random numbers between [0,2] used to increase the randomness of the position.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] 1) This invention addresses the detection of insulators by aerial imagery and constructs a collaborative detection framework between UAVs and edge servers. Different models are deployed on the UAVs and edge servers, which solves the problems of low accuracy of UAV detection alone and high latency of edge server detection alone.

[0069] 2) This invention models the power consumption, task latency, and detection accuracy of UAVs in aerial image detection of insulators, and constructs an unloading algorithm for aerial image detection of insulators based on particle swarm optimization. Under the constraint of task latency, it coordinates the optimization of UAV power consumption and detection accuracy, and can balance accuracy and power consumption according to actual needs to achieve lower power consumption and higher accuracy.

[0070] 3) This invention proposes an end-edge collaborative video processing paradigm that can be deployed on resource-constrained terminal devices such as drones and cameras, and achieve real-time dynamic video task processing with the collaboration of edge servers. Attached Figure Description

[0071] Figure 1 This is a flowchart illustrating the method implementation of an embodiment of the present invention;

[0072] Figure 2 This is an architecture diagram of the edge-end collaborative detection framework built in this embodiment of the invention, which is based on a drone-edge server. Detailed Implementation

[0073] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0074] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0075] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0076] like Figure 1 As shown, this embodiment provides a method for jointly optimizing accuracy and energy consumption for insulator image detection tasks, including the following steps:

[0077] A) Deploy the trained lightweight deep neural network model and high-precision deep neural network model for insulator aerial image detection tasks on the UAV and edge server respectively, and construct an edge-end collaborative detection framework with UAV-edge server as the main body, such as... Figure 2 As shown.

[0078] Deep neural networks, as an emerging technology, have demonstrated significant advantages in image detection accuracy. With technological advancements, many image detection network models with different strengths have emerged. Some lightweight detection models, such as MobileNet, prioritize detection speed, featuring simpler network structures and fewer parameters, sacrificing some model accuracy for faster processing; they have lower accuracy but are faster. Other high-precision detection models, such as YOLO and Faster-RCNN, prioritize model accuracy, possessing a larger number of parameters, resulting in higher accuracy but longer processing times. Using the same aerial image dataset of insulators, both detection models were trained. Deploying the lightweight detection model on a drone meets the requirements for task latency, while deploying the high-precision detection model on an edge server meets the requirements for task detection accuracy.

[0079] Step A specifically includes the following steps:

[0080] A1) Data preprocessing: collect aerial images of insulators, label them, and construct a training dataset.

[0081] Specifically, the collected aerial images of insulators were labeled, and the coordinate and category information of the insulator detection boxes were saved in a text file. The dataset was divided into training and testing sets for later use.

[0082] A2) Place the training dataset constructed in step A1 into a high-performance platform and train the lightweight deep neural network MobileNet and the high-precision deep neural network YOLO-V5 respectively, thereby obtaining the trained lightweight deep neural network model and the high-precision deep neural network model.

[0083] A3) Deploy the lightweight deep neural network model trained in step A2 on the drone, and deploy the high-precision deep neural network model trained in step A2 on the edge server. The high-precision deep neural network model deployed on the edge server has N input image resolution configuration modes, and data transmission between the drone and the edge server is realized simultaneously.

[0084] In this embodiment, the edge server supports N=5 input modes, namely, input image resolutions of 640x640, 800x800, 960x960, 1020x1020 and 1280x1280, and simultaneously realizes data transmission between the drone and the edge server.

[0085] B) Divide the time into time slots, and model the insulator image detection task generated in real time on the UAV according to the time slots. Model the energy consumption, latency, accuracy and other factors of the UAV-side processing and edge server-side processing of the insulator image detection task.

[0086] Step B specifically includes the following steps:

[0087] B1) Divide the time into time slots. Within each time slot t, the UAV generates F frames of images, and each frame is denoted as P. f , f∈[1,F], for each frame image P f Modeling is performed to obtain image P f Data volume D f Unloading decision

[0088] If you choose r i The image resolution mode (i = 0, 1, ..., N) is calculated, then the data size D of the f-th frame image is... f for:

[0089] D f =σ·r i 2 (1)

[0090] Where σ is a constant, representing the number of data bits carried by one pixel.

[0091] The uninstallation decision is: The detection task of the f-th frame image selects the i-th processing mode (i = 0, 1, ..., N), where i = 0 indicates processing at the UAV end, and i = 1, 2, ..., N indicates processing at the edge server.

[0092] B2) Detection accuracy of offloading image detection tasks to edge servers for processing Energy consumption Modeling is performed to assess the detection accuracy of image detection tasks processed on UAVs. Energy consumption Modeling is performed.

[0093] For tasks performed on drones, there is only one input image resolution setting and detection accuracy. and energy consumption It can be set as a constant, that is

[0094] For tasks offloaded to edge servers for execution, detection accuracy Modeled as a function of the input image resolution, i.e.

[0095]

[0096] Where α and β are constants determined by the specific detection model. When the task is offloaded to the edge server for execution, the resulting energy consumption includes the energy consumption for data transmission from the drone and the energy consumption for execution on the edge server. Since this invention focuses on optimizing energy consumption on the drone side, the energy consumption generated on the edge server is ignored, thus the total energy consumption for offloading is... for:

[0097]

[0098] Where ψ is the energy consumed to transmit 1 bit of data.

[0099] C) Model the end-to-end time delay constraint for the insulator image detection task, and formalize the problem of co-optimization of accuracy and energy consumption for the insulator image detection task into a mathematical model of optimization problem.

[0100] Step C specifically includes the following steps:

[0101] C1) The delay caused by performing the detection task of the insulator image of the f-th frame on the UAV. The delay caused by offloading to the edge server Model them separately.

[0102] For tasks performed on a drone, the input image resolution is fixed, therefore the execution delay can be set to a constant, i.e.

[0103] For tasks offloaded to edge servers, latency consists of two parts: transmission latency and execution latency on the edge server. Transmission latency is determined by the amount of data transmitted and network bandwidth. Where B is the network bandwidth within the current time slot t. The edge server execution latency is determined by the input image resolution, specifically:

[0104]

[0105] Where η and θ are constants, the total delay for task unloading calculation is:

[0106] C2) Model the end-to-end average delay, average accuracy, and average energy consumption of the insulator image detection task within time slot t for F frames.

[0107] The end-to-end latency L of the image detection task in frame f is... f for:

[0108]

[0109] The average delay for detecting F-frame insulator images is:

[0110]

[0111] The end-to-end energy consumption E of the f-th frame image detection task f for:

[0112]

[0113] The average energy consumption for detecting F-frame insulator images is:

[0114]

[0115] The accuracy A of the image detection task in the f-th frame f for:

[0116]

[0117] The average accuracy of F-frame insulator image detection is:

[0118]

[0119] C3) formally represents the problem of co-optimizing accuracy and energy consumption for insulator image detection tasks as a mathematical model of optimization problems.

[0120] The problem is defined as co-optimizing average energy consumption and accuracy under the constraint of end-to-end average time delay in F-frame insulator image detection, i.e.

[0121] min A+δE (11)

[0122] Here, δ is a constant representing the trade-off between detection accuracy and energy consumption, and the value of δ can be adjusted according to actual needs.

[0123] The constraints are:

[0124]

[0125]

[0126]

[0127] D) Construct an image detection task offloading algorithm based on particle swarm optimization to obtain the optimal task offloading strategy and image processing resolution configuration for the offloading task, thereby achieving the joint optimal balance between accuracy and energy consumption of the detection task while satisfying the time delay constraint.

[0128] Based on the idea of ​​particle swarm optimization algorithm, this invention creatively designs a task offloading algorithm that coordinates the optimization of insulator image detection accuracy and energy consumption under time delay constraints.

[0129] Step D specifically includes the following steps:

[0130] D1) Input to the image detection task offloading algorithm based on particle swarm optimization: the amount of data D of each frame f in the F frames generated within time slot t. f Parameters such as network bandwidth B are used as inputs to the algorithm, and the maximum number of iterations T is set. max .

[0131] D2) Algorithm initialization: Set up a particle swarm of I groups of particles, set the boundary of the optimal solution, and initialize the velocity V of the sample. j and position W j , j∈[1,J]. Multiple random sets of data are used as the initial values ​​for the particle swarm. An unloading strategy for the F-frame images is defined. Let W be the position of a particle. j The joint A+δE objective function of image detection accuracy and energy consumption is set as the fitness of the particle and calculated according to equation (11).

[0132] D3) Calculate the fitness of particles in the particle swarm in parallel based on the input, i.e., the accuracy and energy consumption value corresponding to each unloading decision, and then select the local optimum Q of particle j. j The global optimal solution Q for all particles g This allows us to obtain both the local optimal solution and the global optimal solution.

[0133] D4) Iterates, updating the particle's position and velocity based on the local optimum, global optimum, and particle velocity, using the following formula:

[0134]

[0135]

[0136] Where ω1 is a constant, and ω2 and ω3 are two random numbers between [0,2] used to increase the randomness of the position.

[0137] D5) Determine if the number of iterations has been reached. If yes, stop the iteration, output the position of the optimal particle, i.e. the optimal unloading decision scheme, and the joint value of image detection accuracy and energy consumption corresponding to the optimal solution; if no, repeat steps D3-D4 to continue the iteration.

[0138] This invention proposes a method for co-optimizing accuracy and energy consumption in insulator image detection tasks. This method addresses the problems of low efficiency and difficulty in ensuring real-time performance in insulator image detection tasks in scenarios such as UAV power line inspection, achieving high-precision, low-energy detection under time-delay constraints. To meet the real-time requirements of safety early warning, and considering the generally weak computing power of edge devices, a lightweight algorithm model for insulator aerial image detection is deployed on the UAV. Simultaneously, to meet the high-precision detection requirements, a high-precision insulator aerial image detection model is deployed on an edge server. Within each time slot, the proposed algorithm dynamically generates task offloading decisions on the UAV and the image processing resolution configuration for offloading tasks, determining which aerial images need to be offloaded to the edge server for processing and which images need to be processed on the UAV. This collaboratively optimizes the accuracy and energy consumption of the detection task while ensuring time-delay constraints, enabling real-time UAV detection of insulators.

[0139] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0140] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0141] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0142] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0143] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for jointly optimizing accuracy and energy consumption for insulator image detection tasks, characterized in that, Includes the following steps: A) Deploy the trained lightweight deep neural network model and high-precision deep neural network model for insulator aerial image detection tasks on the UAV and edge server respectively to build an edge-end collaborative detection framework with UAV-edge server as the main body. B) Divide the time into time slots, and model the real-time insulator image detection task generated on the UAV according to the time slots. Model the energy consumption, latency, and accuracy factors of the UAV-side processing and edge server-side processing of the insulator image detection task. C) Model the end-to-end time delay constraint of the insulator image detection task, and formalize the problem of co-optimization of accuracy and energy consumption for the insulator image detection task into a mathematical model of optimization problem; D) Construct an image detection task unloading algorithm based on particle swarm optimization to obtain the optimal task unloading strategy and image processing resolution configuration for the unloading task; Step C specifically includes the following steps: C1) for the first The latency caused by performing frame insulator image detection tasks on a drone. The latency caused by unloading to the edge server for execution Model them separately; C2) Regarding time slots Inside The end-to-end average latency, average accuracy, and average energy consumption of the frame insulator image detection task are modeled. C3) The problem of co-optimizing accuracy and energy consumption for insulator image detection is formalized into a mathematical model of optimization problem; Step D specifically includes the following steps: D1) Input to the image detection task offloading algorithm based on particle swarm optimization: (Time slot) Internally generated Each frame in the image Data volume Network bandwidth The parameters are used as input to the algorithm, setting the maximum number of iterations. ; D2) Algorithm initialization: Setting The unloading strategy for the frame image is the position of a particle. The image detection accuracy and energy consumption are set as the objective function. The boundary of the optimal solution and the velocity and position of the initial sample are set. Multiple sets of random data are used as the initial values ​​of the particle swarm. D3) Calculate the fitness of particles in the particle swarm in parallel based on the input, that is, the accuracy value and energy consumption value corresponding to each unloading decision, to obtain the local optimal solution and the global optimal solution; D4) Iterates, updating the particle's position based on the local optimum, global optimum, and particle velocity; D5) Determine if the number of iterations has been reached. If yes, stop the iteration, output the position of the optimal particle, i.e. the optimal unloading decision scheme, and the joint value of image detection accuracy and energy consumption corresponding to the optimal solution; otherwise, continue the iteration.

2. The method for jointly optimizing accuracy and energy consumption for insulator image detection tasks according to claim 1, characterized in that, Step A specifically includes the following steps: A1) Data preprocessing: collect aerial images of insulators, label them, and construct a training dataset; A2) The lightweight deep neural network MobileNet and the high-precision deep neural network YOLO-V5 are trained using the training dataset constructed in step A1, respectively, to obtain the trained lightweight deep neural network model and the high-precision deep neural network model. A3) Deploy the lightweight deep neural network model trained in step A2 on the drone, and deploy the high-precision deep neural network model trained in step A2 on the edge server. The high-precision deep neural network model deployed on the edge server has... It offers multiple input image resolution configuration modes, enabling data transmission between the drone and the edge server.

3. The method for jointly optimizing accuracy and energy consumption for insulator image detection tasks according to claim 1, characterized in that, Step B specifically includes the following steps: B1) Divide time into time slots, each time slot Generated on the internal drone Frame images, each frame image is denoted as , For each frame of image Modeling is performed to obtain images. Data volume Unloading decision ; B2) Detection accuracy of offloading image detection tasks to edge servers for processing Energy consumption Modeling is performed to assess the detection accuracy of image detection tasks processed on UAVs. Energy consumption Modeling is performed.

4. The method for jointly optimizing accuracy and energy consumption for insulator image detection tasks according to claim 3, characterized in that, choose The image resolution mode is calculated, then the first... Data volume of a frame image for: （1） in, It is a constant representing the number of data bits carried by one pixel; The uninstallation decision is: ; Indicates the first The frame image detection task is selected as the first. Types of processing modes , This indicates that the processing is done on the drone end. This indicates that the first option is selected on the edge server. This mode of processing; For tasks performed on drones, there is only one input image resolution setting and detection accuracy. and energy consumption Let it be a constant, that is , ; For tasks offloaded to edge servers for execution, detection accuracy Modeled as a function of the input image resolution, i.e. （2） in, and The energy consumption is constant; when the task is offloaded to the edge server for execution, the energy consumption includes the energy consumption for data transmission from the drone and the energy consumption for execution on the edge server. Ignoring the energy consumption generated on the edge server, the total energy consumption for offloading is... for: （3） in, The energy consumed to transmit 1 bit of data.

5. The method for jointly optimizing accuracy and energy consumption for insulator image detection tasks according to claim 1, characterized in that, In step D2, set The particles form a particle swarm, the boundary of the optimal solution is set, and the velocity of the sample is initialized. and location , ; set up Frame image offloading strategy The position of a particle Setting a combined approach to image detection accuracy and energy consumption The objective function serves as the fitness of the particle; In step D3, the fitness of particles in the particle swarm is calculated in parallel based on the input, i.e., the accuracy and energy consumption value corresponding to each unloading decision, and then particles are selected. Its own local optimal solution and the global optimal solution for all particles ; In step D4, the particle's position and velocity are updated based on the local optimum, the global optimum, and the particle's velocity. The calculation formula is as follows: in, It is a constant. and These are two random numbers between [0,2], used to increase the randomness of the position.

Images