Distribution network unmanned aerial vehicle inspection system based on front-end AI visual identification

Abstract

The invention relates to a distribution network unmanned aerial vehicle inspection system based on front-end AI visual identification, and belongs to the technical field of information processing. The inspection system comprises a server and one or more unmanned aerial vehicles in communication connection with the server. The server is equipped with an arithmetic unit and a control unit and is responsible for analyzing flight data of the unmanned aerial vehicle, generating a flight strategy model and issuing a control instruction. The operation of the inspection system is divided into three stages: in the preparation stage, an operator controls the unmanned aerial vehicle to execute a standard inspection route, and key data is recorded to generate a standard route file; in the modeling stage, a flight strategy model based on a deep Q network is trained by using a standard route file, a reward value is calculated through a reinforcement learning framework, the task quality of the unmanned aerial vehicle is evaluated, and an optimal strategy is generated; the unmanned aerial vehicle performs automatic inspection at the real flight stage, operates the flight strategy model in real time through the server, and dynamically adjusts the flight strategy or intervenes in control based on the current state of the unmanned aerial vehicle and the check point image quality.

Description

Technical Field

[0001] The present invention belongs to the field of information processing technology, and specifically relates to a distribution network drone inspection system based on front-end AI visual recognition. Background Art

[0002] As the scale and complexity of power grids increase, traditional manual inspection methods face many challenges in terms of efficiency, cost and safety. Distribution network inspection tasks require regular inspections of power grid equipment to ensure the safe operation of equipment and the timely detection of faults. However, traditional inspection methods often require manpower to go deep into remote or dangerous areas, which is not only time-consuming and labor-intensive, but also has potential safety hazards. In addition, manual inspections are difficult to guarantee in terms of inspection accuracy and data consistency, and the inspection results are easily affected by human factors. In recent years, drone technology has been widely used in the field of power inspections due to its flexible, maneuverable, efficient and low-consumption characteristics. However, current drone inspections mostly rely on preset programs or manual operations, and cannot flexibly adjust strategies according to on-site conditions, making it difficult to meet the needs of complex inspection scenarios. Therefore, combining artificial intelligence technology, especially front-end AI technology based on visual recognition, dynamically optimizing inspection strategies through real-time analysis of images collected by drones and flight status has become a research focus for improving inspection efficiency and intelligence.

[0003] According to the relevant public technologies, the technical solution with publication number CN116203973A proposes a rail inspection robot, which uses AI vision to generate a three-dimensional image of the inspection area and automatically judges abnormal situations that occur during the inspection process. The technical solution with announcement number AU2022263617B1 proposes a rescue drone, which is equipped with a perspective acquisition device and an AI recognition system, and can automatically identify dangerous scenes and dangerous objects through AI models to implement further rescue operations. The technical solution with publication number WO2015078284A1 proposes an inspection drone used in the community, which can perform complex inspection tasks on large areas and multi-story buildings by relaying information through a central processing terminal located on the ground.

[0004] The above technical solutions all propose technical solutions for using drones to inspect designated spatial areas, but there are currently few technical solutions that mention more complex inspection strategies to solve the uncertain conditions of drones during the inspection process.

[0005] The foregoing discussion of the background art is intended only to facilitate an understanding of the present invention. This discussion does not acknowledge or admit that any of the material referred to is part of the common general knowledge. Summary of the invention

[0006] The purpose of the present invention is to provide a distribution network drone inspection system based on front-end AI visual recognition. The inspection system includes a server and one or more drones that are communicatively connected to the server. The server is equipped with a computing unit and a control unit, which are responsible for analyzing drone flight data, generating a flight strategy model and issuing control instructions. The operation of the inspection system is divided into three stages: in the preparation stage, the operator controls the drone to execute a standard inspection route, records key data and generates a standard route file; in the modeling stage, the standard route file is used to train a flight strategy model based on a deep Q network, and the reward value is calculated through a reinforcement learning framework to evaluate the quality of the drone mission and generate the optimal strategy; in the actual flight stage, the drone automatically inspects, and the flight strategy model is run in real time through the server, and the flight strategy is dynamically adjusted or intervention control is performed based on the current state of the drone and the image quality of the checkpoint.

[0007] The present invention adopts the following technical solution: a distribution network drone inspection system based on front-end AI visual recognition, the inspection includes:

[0008] A server, and one or more drones communicatively connected to the server;

[0009] The server comprises:

[0010] The computing unit is configured to obtain and analyze the flight data of the UAV from the UAV; generate, save, and optimize a flight strategy model for evaluating the quality of the UAV's mission execution, and generate a control instruction for the UAV's flight control based on the evaluation result;

[0011] A control unit configured to communicate with one or more UAVs and send the control command to the target UAV;

[0012] The inspection system includes the following two working stages to realize the network distribution drone inspection process based on front-end AI visual recognition:

[0013] Preparation stage: The operator controls the drone to fly along the preset route, perform standard flight actions and flight routes, check the target checkpoints, record key data sets, and generate standard route files;

[0014] Modeling stage: training and establishing a flight strategy model based on the standard route file;

[0015] Actual flight stage: During the automated inspection process of the UAV, real-time data of the UAV is obtained and a real-time data set is generated; the flight strategy model is run to evaluate the quality of the UAV's execution of tasks during the automated inspection process, and subsequent flight strategies are adjusted.

[0016] Preferably, the key data set includes:

[0017] Flight route coordinates, i.e., the coordinates of the drone’s position at a preset interval;

[0018] Drone status, including drone attitude, speed, and pointing angle;

[0019] Checkpoint data: coordinates of key points and standard images of one or more checkpoints.

[0020] Preferably, the flight strategy model is constructed based on a deep Q network, comprising the following steps:

[0021] Input the standard route file as training data into the deep Q network;

[0022] The deep Q network then takes a large amount of UAV flight status data and a large number of checkpoint images as input, and uses a reinforcement learning framework to train and calculate the cumulative reward value Q(s,a) to train the flight strategy model to evaluate the quality of the current UAV's mission execution, and output the flight score and the optimal flight strategy;

[0023] The flight strategy model is trained by optimizing the loss function of the deep Q network and using the gradient descent method.

[0024] Preferably, the calculation of the reward value Q(s,a) is based on at least the following evaluation indicators:

[0025] The degree to which the flight route of the drone complies with the route specified in the standard route file;

[0026] The attitude quality of the drone during flight;

[0027] Image quality of the checkpoints captured by the drone.

[0028] Preferably, during the actual flight phase, the UAV performs automated inspections based on its stored flight program.

[0029] Preferably, during the actual flight phase, the UAV further includes receiving control instructions from the server, thereby executing intervention control of the control instructions.

[0030] Preferably, during the actual flight phase, the server performs the following steps to determine whether to intervene in the inspection work of controlling the drone:

[0031] S100: continuously acquires real-time data from drones and generates real-time data sets;

[0032] S200: running the flight strategy model to evaluate the quality of the UAV's execution of tasks during the automated inspection based on the real-time data set, thereby calculating a flight score;

[0033] S300: Determine the necessity of intervening in controlling the UAV based on the flight score;

[0034] S400: If it is determined to intervene in the control of the UAV, a corresponding control instruction is generated based on the flight score, and sent to the UAV to intervene in the control of the UAV to execute the control instruction.

[0035] Preferably, the flight strategy model is used to evaluate the quality of the UAV's mission execution, and the evaluation indicators included include:

[0036] The degree to which the flight route of the drone complies with the route specified in the standard route file;

[0037] The attitude quality of the drone during flight;

[0038] Image quality of the checkpoints captured by the drone.

[0039] Preferably, the drone is equipped with at least one image acquisition device, and the image acquisition device is composed of one or more of the following sensors or image devices: camera, video camera, thermal imaging camera, infrared camera, night vision sensor, depth camera, ranging sensor, laser ranging sensor, radio ranging sensor.

[0040] The beneficial effects achieved by the present invention are:

[0041] 1. This technical solution adopts a flight strategy model based on a deep Q network, combined with front-end AI visual recognition technology, which can analyze the flight status and image data collected by the drone in real time. By dynamically optimizing the inspection strategy, the drone can flexibly adjust the flight path and inspection actions in different scenarios, reduce manual intervention, and realize a fully automated inspection process. Compared with the traditional method that relies on preset routes or manual operations, this technical solution greatly improves the efficiency and intelligence level of inspection.

[0042] 2. This technical solution dynamically generates a task quality score by comparing standard images with real-time images and combining multi-dimensional flight status data. The design of the reward function comprehensively considers multiple factors such as the degree of flight route deviation, attitude stability, and image quality to ensure the high accuracy and comprehensiveness of the inspection task. This multi-dimensional evaluation mechanism can effectively improve the reliability of drone inspections and the accuracy of results, and reduce the possibility of omissions and misjudgments.

[0043] 3. The drone configured by this technical solution has the ability of automated inspection, and a real-time intervention control mechanism on the server side is also designed. Through continuous monitoring of flight scores, when it is found that the mission quality has declined or an abnormal situation has occurred, the server can immediately intervene in the control of the drone, adjust the flight strategy or terminate the mission to ensure the safety and integrity of the inspection mission. This mechanism improves the adaptability and fault tolerance of the inspection system to complex scenarios.

[0044] 4. The software and hardware parts of the inspection system of this technical solution adopt a modular design. The various working modules and components of the hardware part of the system, as well as the instructions, parameters, and algorithms of the software part can be conveniently replaced and / or upgraded later, thereby reducing the construction cost and maintenance cost of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The present invention can be further understood from the following description in conjunction with the accompanying drawings. The components in the figures are not necessarily drawn to scale, but the emphasis is placed on illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0046] Description of the accompanying drawings: 10-server; 20-drone; 40-network; 500-computer system; 502-bus; 504-processor; 506-main memory; 508-read-only memory; 510-storage device; 512-display; 514-input device; 516-cursor control device; 518-network device;

[0047] Figure 1 It is a schematic diagram of the configuration of the inspection system of the present invention;

[0048] Figure 2 This is a schematic diagram of the architecture and functions of the server described in an embodiment of the present invention;

[0049] Figure 3 Schematic diagram of the working stages of the inspection system in an embodiment of the present invention;

[0050] Figure 4 It is a schematic diagram of the working steps of the inspection system in the actual flight stage according to an embodiment of the present invention;

[0051] Figure 5 It is a schematic diagram of the framework of the computer system used by the computing unit in the embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with its embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. For those skilled in the art, after reviewing the following detailed description, other systems, methods and / or features of the present embodiment will become apparent. It is intended that all such additional systems, methods, features and advantages are included in this specification. Included within the scope of the present invention and protected by the appended claims. Additional features of the disclosed embodiments are described in the following detailed description, and these features will be apparent from the following detailed description.

[0053] The same or similar reference numerals in the drawings of the embodiments of the present invention correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right" and the like indicating the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, it is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or component referred to must have a specific orientation. The invention is constructed and operated in a specific orientation, so the terms describing the positional relationship in the drawings are only used for exemplary description and cannot be understood as a limitation of this patent. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0054] Embodiment 1: For example, a network distribution drone inspection system based on front-end AI visual recognition is proposed, and the inspection includes:

[0055] A server, and one or more drones communicatively connected to the server;

[0056] The server comprises:

[0057] The computing unit is configured to obtain and analyze the flight data of the UAV from the UAV; generate, save, and optimize a flight strategy model for evaluating the quality of the UAV's mission execution, and generate a control instruction for the UAV's flight control based on the evaluation result;

[0058] A control unit configured to communicate with one or more UAVs and send the control command to the target UAV;

[0059] The inspection system includes the following two working stages to realize the network distribution drone inspection process based on front-end AI visual recognition:

[0060] Preparation stage: The operator controls the drone to fly along the preset route, perform standard flight actions and flight routes, check the target checkpoints, record key data sets, and generate standard route files;

[0061] Modeling stage: training and establishing a flight strategy model based on the standard route file;

[0062] Actual flight stage: During the automated inspection process of the UAV, real-time data of the UAV is obtained and a real-time data set is generated; the flight strategy model is run to evaluate the quality of the UAV's execution of tasks during the automated inspection process, and subsequent flight strategies are adjusted.

[0063] Preferably, the key data set includes:

[0064] Flight route coordinates, i.e., the coordinates of the drone’s position at a preset interval;

[0065] Drone status, including drone attitude, speed, and pointing angle;

[0066] Checkpoint data: coordinates of key points and standard images of one or more checkpoints.

[0067] Preferably, the flight strategy model is constructed based on a deep Q network, comprising the following steps:

[0068] Input the standard route file as training data into the deep Q network;

[0069] The deep Q network then takes a large amount of UAV flight status data and a large number of checkpoint images as input, and uses a reinforcement learning framework to train and calculate the cumulative reward value Q(s,a) to train the flight strategy model to evaluate the quality of the current UAV's mission execution, and output the flight score and the optimal flight strategy;

[0070] The flight strategy model is trained by optimizing the loss function of the deep Q network and using the gradient descent method.

[0071] Preferably, the calculation of the reward value Q(s,a) is based on at least the following evaluation indicators:

[0072] The degree to which the flight route of the drone complies with the route specified in the standard route file;

[0073] The attitude quality of the drone during flight;

[0074] Image quality of the checkpoints captured by the drone.

[0075] Preferably, during the actual flight phase, the UAV performs automated inspections based on its stored flight program.

[0076] Preferably, during the actual flight phase, the UAV further includes receiving control instructions from the server, thereby executing intervention control of the control instructions.

[0077] Preferably, during the actual flight phase, the server performs the following steps to determine whether to intervene in the inspection work of controlling the drone:

[0078] S100: continuously acquires real-time data from drones and generates real-time data sets;

[0079] S200: running the flight strategy model to evaluate the quality of the UAV's execution of tasks during the automated inspection based on the real-time data set, thereby calculating a flight score;

[0080] S300: Determine the necessity of intervening in controlling the UAV based on the flight score;

[0081] S400: If it is determined to intervene in the control of the UAV, a corresponding control instruction is generated based on the flight score, and sent to the UAV to intervene in the control of the UAV to execute the control instruction.

[0082] Preferably, the flight strategy model is used to evaluate the quality of the UAV's mission execution, and the evaluation indicators included include:

[0083] The degree to which the flight route of the drone complies with the route specified in the standard route file;

[0084] The attitude quality of the drone during flight;

[0085] Image quality of the checkpoints captured by the drone.

[0086] Preferably, the drone is equipped with at least one image acquisition device, and the image acquisition device is composed of one or more of the following sensors or image devices: camera, video camera, thermal imaging camera, infrared camera, night vision sensor, depth camera, ranging sensor, laser ranging sensor, radio ranging sensor.

[0087] Specifically, as attached Figure 1 As shown, it is a schematic diagram of an exemplary embodiment of the inspection system; wherein the server 10 can be various types of computer devices, including but not limited to: traditional physical servers, virtual servers, cloud computing servers, edge computing devices or other electronic devices with data processing capabilities; the server 10 is connected to one or more drones 20 through a network 40 (only one drone is schematically shown in the figure), receives data sent by the drone 20, and sends control instructions to the drone 20; the type of the network 40 may be WiFi, radio, cellular network (such as 4G, 5G), satellite communication, Bluetooth, ZigBee or other wireless communication technologies suitable for data transmission; the communication bandwidth and communication distance of the network 40 can ensure real-time data exchange between the server and the drone, and support remote control, status monitoring and image data transmission of the drone;

[0088] For example, the operator may operate the drone 20 by using a remote controller or a control unit on the server 10 ;

[0089] Preferably, the operator can operate the drone 20 by using a remote controller or a control unit on the server 10; in the preparatory stage, the operator can control the drone 20 in multiple stages, execute flight along a preset inspection route, and hover and inspect at key points of the route. During the flight, the drone 20 records the flight status data in real time, including position coordinates, attitude information (such as pitch angle, yaw angle, roll angle), flight speed and other environmental parameters.

[0090] At key points on the inspection route, the drone 20 collects images of the target checkpoints and ensures that the captured images cover the main parts of the target equipment. To this end, the operator can adjust the drone's flight attitude, camera angle or hovering time to ensure that the collected images are clear and meet the requirements. The collected images will be stored as standard images of key points, and the coordinates and other necessary information of the checkpoints will be recorded to form a complete checkpoint data set.

[0091] After the drone 20 completes the inspection of the preset route, all collected data will be summarized as a key data set and uploaded to the server 10 through the communication module. The server 10 saves the key data set as a standard route file to provide high-quality training data for the subsequent modeling stage. Preferably, the key data set includes the flight route coordinates, the flight status information of the drone, and the standard image of each checkpoint, which is used to train and optimize the flight strategy model.

[0092] In a preferred embodiment, the flight route coordinates and additional information thereof may specifically include one or more of the following data:

[0093] Timestamps record the time information of each coordinate data to ensure that the flight data has clear timing characteristics. Timestamps can be used to accurately restore the flight path and speed changes of the drone.

[0094] Geographic coordinates use GPS or Beidou navigation system to record the longitude, latitude and altitude information of the drone at every moment during the flight to form a three-dimensional flight trajectory; illustratively, the coordinate format is (longitude, latitude, altitude), for example (120.385, 36.067, 50).

[0095] The heading angle records the current direction of the drone, usually expressed as a clockwise angle relative to the north (0° to 360°). The heading angle can help describe the specific direction of the drone during flight.

[0096] The velocity vector includes the velocity components of the drone in space, such as eastward velocity, northward velocity, and vertical velocity; velocity information can be used to assess the smoothness and efficiency of flight.

[0097] Coordinate accuracy describes the measurement accuracy of the flight path coordinates, usually expressed as an error range (e.g. ±0.5 meters).

[0098] Supplementary data may include additional information related to the flight environment, such as the location of surrounding obstacles, weather conditions (wind speed, temperature, humidity), etc., to further optimize flight strategy and safety.

[0099] Further, in a preferred embodiment, the flight status information includes flight parameters of the drone 20, such as current position, speed, altitude, attitude angle and other information; wherein the current position and altitude of the drone 20 can be calculated using the geographic coordinates and heading angle information as described above, and the speed and attitude angle information of the drone 20 are obtained through the drone 20's own sensors (such as GPS, inertial measurement unit), and are used for real-time flight control and navigation;

[0100] Preferably, the attitude angle refers to the tilt state of the UAV in space, including three key angles: roll angle, pitch angle and yaw angle. The value of the attitude angle and the change of the value can be used to determine the current flight state of the UAV 20, such as whether it is flying in a stable or bumpy state, or whether repeated corrections and directional instability occur during turning; and it can be further combined with the image data described below to comprehensively determine the cause of the image error.

[0101] Among them, the sampling frequency of key data defines the time interval for recording the above data. Preferably, it can be recorded 20 times per second, that is, 20 Hz, or a higher frequency. This is based on the performance of the processing system of the drone 20 and the communication performance between the drone 20 and the server 10. A higher sampling frequency can provide more detailed flight trajectory information, and relatively speaking, it will also generate a larger amount of data and a higher computing load on various parts of the system.

[0102] Furthermore, in a preferred embodiment, the standard image of the checkpoint refers to image data with benchmark reference significance that is captured and collected by the drone 20 at the target checkpoint according to the specified action and position in the preparation stage; the standard image is used to compare and score the real-time collected images in the actual flight stage to evaluate the quality of the drone's mission execution.

[0103] Preferably, the images collected according to the preset position, posture and / or image parameter requirements of the checkpoints are stored as reference images after subsequent quality screening, post-processing and annotation.

[0104] Preferably, the standard image includes a definition requirement, that is, the edges of the target object in the image are clear, the details are rich, and there is no blur or occlusion;

[0105] Preferably, the standard image includes composition specifications, including that key components of the target equipment (such as electric poles, transformers, etc.) are located in the center of the picture or in a specified area;

[0106] Preferably, the standard image includes shooting angle requirements, such as shooting angle and direction consistent with task requirements, to ensure the accuracy of subsequent real-time image comparison.

[0107] Preferably, the standard image includes environmental reference requirements, that is, contains sufficient environmental background information (such as surrounding topography, cable distribution, etc.) to facilitate identification and comparison.

[0108] Furthermore, the standard image includes the following multi-source data:

[0109] (1) Image data, including image formats such as RAW and PNG;

[0110] (2) Shooting parameters, including resolution (such as 1080P or higher), exposure value, white balance, focal length and other device shooting settings.

[0111] (3) Manually or automatically annotated information, including: target area information, annotating the key positions of target equipment components (such as transformers, insulators, wire connectors, etc.) in the image, defined as the target area ROI; and target status information, which can also show the health status of the equipment, such as whether there is obvious damage, pollution, rust, etc.

[0112] (4) Environmental data, including light intensity, which can be collected by light sensors; environmental variables such as wind speed, temperature, and humidity, which can be collected by environmental sensors;

[0113] Furthermore, after the image is acquired, including integrating and storing various data, all data related to the standard image are integrated and stored as a checkpoint record file, and are associated with the spatial position of the corresponding checkpoint;

[0114] At the same time, after obtaining sufficient and standard flight route coordinates, flight status information, and standard images of each checkpoint, the standard route file is integrated;

[0115] Furthermore, the construction and training of the flight strategy model is based on a neural network of a deep Q network, and the parameters and related functions involved in the deep Q network are defined as follows:

[0116] Define the initial state-action data (s, a, r, s′), where:

[0117] State s represents the current flight state of the drone, such as position, altitude, speed, etc.;

[0118] Action a represents the flight action that the drone can perform, such as adjusting direction, changing speed, etc.;

[0119] Reward r represents the feedback that the agent gets from the environment after performing action a in a certain state s;

[0120] Strategy P, describing the behavior rules of the drone to choose actions;

[0121] Among them, the strategy P can be composed of the standard route file combined with the control instruction set of the drone, and the technicians can first write an initial automatic flight program and store it in the drone's own control system. The strategy P can be an automatic flight program that the drone can autonomously execute when it is separated from the server;

[0122] State-action value function Q(s,a): represents the cumulative reward obtained after performing action a in state s. Through training, the reinforcement learning framework aims to optimize Q(s,a) to guide the agent to choose the best action.

[0123] Among them, the cumulative reward Q value is defined as:

[0124]

[0125] In the above formula, represents the expected cumulative reward when taking action a from state s based on strategy P and then following strategy P; preferably, multiple inspection segments in the inspection process can be set to calculate corresponding reward values; for example, it can be the cumulative reward in a preset flight route, or it can refer to the cumulative reward in the process of completing the inspection action after reaching a checkpoint.

[0126] Where δ is the discount factor, which ranges from [0 to 1] and is set by the technician; δ t represents the change with time step t, δ t Decreasing is used to indicate the change in the weight of future rewards, thereby reducing the impact of future returns; this method makes the algorithm pay more attention to recent returns rather than the distant future; E is the expected return value, E P That is, the expected return value when executing strategy P;

[0127] r t It is the sequence of reward values ​​expected to be obtained when executing each subsequent [state-action] pair according to strategy C; it is also determined by the technician when formulating the initial strategy;

[0128] Further, the training of the flight strategy model is implemented, specifically including the following steps:

[0129] (1) Initialize the deep Q network parameters: Randomly initialize the weight parameter θ of the deep Q network Q(s,a;θ) and the weight parameter θ′ of the target network Q′(s′,a′;θ′);

[0130] (2) Experience replay: Create an experience pool to store the quadruple (s, a, r, s′) of the state, action, reward, and next state of the drone during flight;

[0131] (3) Randomly sampling small batches of data: that is, randomly sampling small batches of data (s, a, r, s′) from the experience pool to break data correlation and improve training stability;

[0132] (4) Calculate the target Q value: Recursively define Q(s,a) through the Bellman equation, and we have:

[0133]

[0134] In the above formula, y is the target value, which is used to update the current Q value. It represents the estimated value of the return that can be obtained from the current state-action pair (s, a), including immediate rewards and future cumulative rewards.

[0135] r is the current reward, which means the reward obtained immediately after the drone performs action a in state s. It needs to be designed according to the requirements of the inspection task. The value of r can be continuous or discrete. Preferably, r includes positive rewards and negative rewards. For positive rewards, for example, when the image quality is high, a high reward is given, and as the image quality decreases, the reward decreases, or even becomes negative, that is, a negative reward. For another example, when the inspection flight route deviates slightly from the standard route, a high reward is given. The greater the degree of deviation, the lower the reward, or even a negative value.

[0136] Q′(s′,a′;θ′) is the target Q-value function, which represents the estimated reward when performing action a′ in the next state s′; here, Q′ is calculated by the target network; and max a′ Q′(s′,a′;θ′) means that in the next state s′, the Q value is calculated for all possible actions a′, and the largest Q value is selected. It represents the possible reward for selecting the optimal action under the estimation of the target network.

[0137] And θ′ is a parameter of the target Q network, and each θ′ value corresponds to a stable version of the target Q network; preferably, the fluctuation of the target Q network can be reduced by regular updates.

[0138] (5) Based on the reinforcement learning framework, a deep Q network is used to approximate Q(s, a), including the following sub-steps:

[0139] (5-1) Input current state s and action a;

[0140] (5-2) Output the current state-action pair Q(s,a);

[0141] (5-3) Loss function optimization: Set the target value y to:

[0142]

[0143] The loss function Lost(θ) of the cyclic optimization network is:

[0144] Lost(θ)=E (s,a,r,s′) [(yQ(s,a;θ)) 2 ];

[0145] Among them, θ tar is the target network parameter, which is periodically synchronized from the current network parameter θ;

[0146] (5-4) Use the gradient descent method to update the network parameters θ, that is,

[0147]

[0148] in, It represents the gradient of the loss function Loss(θ) with respect to the parameter θ, and indicates the direction and magnitude of the loss function as the network parameters change. The purpose of calculating the gradient is to find the parameter update direction that can reduce the loss function. λ is the learning rate, which is a hyperparameter that controls the size of each gradient descent update step. It can be specifically set by the technicians who train the model according to the convergence speed and stability of the training process.

[0149] Finally, a trained flight strategy model based on the deep Q network is generated, and the flight strategy model is used to calculate the optimal action a of the UAV in different states to form a complete flight strategy.

[0150] The trained flight strategy model can evaluate the current state of the drone and the corresponding action, and calculate the comprehensive value of the action Q(s,a). Through the output of the model, the deviation between the current action and the optimal strategy can be judged, thereby providing a score for the action. If the score of an action is high, it means that the action is in line with the optimal strategy and is reasonable; if the score is low, it means that the action may deviate from the optimal strategy and needs to be adjusted. This scoring mechanism enables the flight strategy model to not only help drones optimize their actions, but also monitor the quality of task execution in real time during operation.

[0151] Preferably, since the flight strategy model has a large capacity and requires high-intensity computing, it is generally deployed on the server side, or other computing sides, rather than the drone side. In actual flight, the drone relies on the preset program stored internally to complete autonomous flight and inspection tasks, which can ensure basic operations and task execution. In this case, the server only acts as an observer, using the flight strategy model to continuously evaluate the state and action of the drone, especially the images collected at the checkpoints, and calculate the score. If the behavior of the drone is consistent with the optimal strategy with a small deviation, and the quality of the acquired image is stable, the server does not need to intervene, but only needs to record the score and use it as a reference for subsequent task optimization.

[0152] When the drone's behavior score is low, such as multiple deviations from the standard route or substandard image quality at key points, the server will determine whether intervention control is required based on the score results. If intervention is required, the server will generate an optimized control strategy through the flight strategy model, generate control instructions and send them to the drone end, thereby adjusting the drone's flight behavior. For example, when the score shows that the image quality of a certain checkpoint is poor, the server may generate a hovering instruction or a strategy to return to the previous stable point for re-checking. This intervention mechanism ensures that while the drone is flying autonomously, it can be quickly adjusted when the quality of mission execution decreases to avoid mission failure.

[0153] Embodiment 2: This embodiment should be understood to include at least all the features of any one of the aforementioned embodiments and to be further improved thereon.

[0154] In a preferred embodiment, an artificial intelligence-based image analysis model is deployed in the computing unit to perform difference analysis between real-time images collected by the drone and preset standard images to evaluate the execution quality of the drone inspection task.

[0155] Preferably, the image analysis model is based on deep learning technology, and through image recognition and feature extraction methods, key features are extracted from real-time pictures and standard pictures, such as edge contours of equipment, texture details, color distribution, and relative positions of targets, etc. The model will match and compare these extracted features and calculate the similarity between the two pictures.

[0156] Preferably, the image analysis model uses a multi-layer convolutional neural network (CNN) as its core structure.

[0157] Preferably, the image analysis model includes an input layer, a feature extraction layer, a multi-layer convolutional layer, a pooling layer, a fully connected layer and an output layer.

[0158] The input layer is responsible for receiving the original pixel data of the standard image and the real-time image, and normalizing the data to ensure the stability and consistency of subsequent calculations.

[0159] The feature extraction layer consists of multiple layers of convolution operations, and each layer of convolution kernels automatically learns and extracts local features in the image, such as edges, corners, and texture patterns. After the convolution operation, nonlinearity is introduced through activation functions (such as ReLU) to enhance the model's ability to express complex features.

[0160] The pooling layer performs dimensionality reduction processing on the convolutional feature map, such as using maximum pooling or average pooling, to reduce computational complexity and enhance the model's robustness to position changes.

[0161] In the deep structure of the model, multi-layer convolutional networks gradually extract high-level semantic features, such as the overall shape of the device, structural details, and the spatial layout of the target. After multiple layers of stacked convolution and pooling operations, the extracted feature maps are flattened and input into the fully connected layer. The fully connected layer combines the extracted local features into global semantic features by learning specific weights for overall analysis of the image.

[0162] In order to accurately compare the real-time image with the standard image, the AI ​​model sets up a comparison module after the fully connected layer. This module quantifies the difference between the real-time image and the standard image by calculating the cosine similarity or Euclidean distance of the two sets of features. The comparison module outputs a similarity score that reflects the degree of match between the two images in terms of content.

[0163] In addition, the model also adopts a multi-task learning strategy, combining the target detection module and the anomaly detection module to enhance the analysis capability. The target detection module is used to identify and locate the key target parts in the image, ensuring that the model focuses on the target area during comparison; the anomaly detection module detects whether the device is damaged or blocked by obstacles by analyzing features such as texture changes and edge anomalies in the image.

[0164] Finally, the model outputs a comprehensive score based on the scoring mechanism, combining factors such as image similarity, target component integrity, and image clarity to generate an evaluation result. If the score is low, the model will mark possible problems and trigger the server to generate corresponding adjustment strategies. Through this refined model design, the AI ​​system can efficiently and reliably evaluate the quality of drone-collected images and support the optimization and adjustment of dynamic inspection tasks.

[0165] Furthermore, the image analysis model can be trained to capture special problems in the image, such as equipment damage, obstacle occlusion, etc., which will further affect the scoring results. When the score is lower than the preset threshold, the server can generate an adjustment strategy in combination with the deep Q network, such as re-hovering to shoot, adjusting the drone posture or additional shooting, thereby improving the image acquisition quality. This AI model-based image analysis method enables the drone inspection system to have the ability to automatically evaluate and dynamically adjust the execution effect of the task, greatly improving the reliability and intelligence level of the inspection task.

[0166] Since the image analysis model is not the focus of discussion in this technical solution, it will not be further described here.

[0167] Embodiment 3: This embodiment should be understood to include at least all the features of any one of the aforementioned embodiments and to be further improved thereon.

[0168] For example, as shown in the attached Figure 5As shown, an implementation of a computer system 500 used by the server is described; the computer system 500 can be applied to the data storage, calculation and result output process of each working module in the identification and judgment system.

[0169] Illustratively, computer system 500 includes a bus 502 or other communication mechanism for communicating information, and one or more processors 504 coupled with bus 502 for processing information; processor 504 may be, for example, one or more general-purpose microprocessors.

[0170] The computer system 500 also includes a main memory 506, such as a random access memory (RAM), a cache, and / or other dynamic storage device, coupled to the bus 502 for storing information and instructions to be executed by the processor 504; the main memory 506 may also be used to store temporary variables or other intermediate information during the execution of instructions to be executed by the processor 504; these instructions, when stored in a storage medium accessible to the processor 504, present the computer system 500 as a special-purpose machine customized to perform the operations specified in the instructions.

[0171] The computer system 500 may also include a read-only memory (ROM) 508 or other static storage device coupled to the bus 502 for storing static information and instructions for the processor 504; a storage device 510 such as a disk, an optical disk, or a USB drive (flash drive) will be coupled to the bus 502 for storing information and instructions.

[0172] Furthermore, coupled to the bus 502 may also include a display 512 for displaying various information, data, media, etc., and an input device 514 for allowing a user of the computer system 500 to control, manipulate and / or interact with the computer system 500 .

[0173] A preferred way to interact with the management system may be through a cursor control device 516, such as a computer mouse or similar control / navigation mechanism.

[0174] Furthermore, the computer system 500 may also include a network device 518 coupled to the bus 502; wherein the network device 518 may include, for example, a wired network card, a wireless network card, a switching chip, a router, a switch, and other components;

[0175] In general, the terms "engine", "component", "system", "database", etc. used in this document may refer to logic embodied in hardware or firmware, or to a collection of software instructions, possibly with entry and exit points, written in a programming language such as Java, C or C++; software components may be compiled and linked into executable programs, installed in a dynamic link library, or may be written in an interpreted programming language (such as BASIC, Perl or Python); it should be understood that software components may be called from other components or from themselves, and / or may be called in response to detected events or interrupts.

[0176] Software components configured to execute on a computing device may be provided on a computer-readable medium, such as a compact disc, digital video disc, flash drive, diskette, or any other tangible medium, or as a digital download (and may be initially stored in a compressed or installable format that requires installation, decompression, or decryption prior to execution); such software code may be stored in part or in whole on a memory device of the executing computing device for execution by the computing device; the software instructions may be embedded in firmware, such as an EPROM; it should also be understood that hardware components may be composed of connected logic units (such as gates and flip-flops), and / or may be composed of programmable units (such as a programmable gate array or processor).

[0177] Computer system 500 includes custom hardwired logic, one or more ASICs or FPGAs, firmware, and / or program logic that, in combination with the computer system, renders computer system 500 a special-purpose computing device, to implement the techniques described herein.

[0178] According to one or more embodiments, the techniques herein are performed by computer system 500 in response to processor 504 executing one or more sequences of one or more instructions contained in main memory 506; such instructions may be read into main memory 506 from another storage medium such as storage device 510; execution of the sequences of instructions contained in main memory 506 causes processor 504 to perform the processing steps described herein; in alternative embodiments, hard-wired circuitry may be used in place of or in combination with software instructions.

[0179] As used herein, the term "non-transitory media" and similar terms refer to any medium that stores data and / or instructions that cause a machine to operate in a specific manner; such non-transitory media may include non-volatile media and / or volatile media; non-volatile media include, for example, optical or magnetic disks, such as storage device 510; volatile media include dynamic memory, such as main memory 506.

[0180] Among them, common forms of non-transitory media include, for example, floppy disks, diskettes, hard disks, solid-state drives, magnetic tapes or any other magnetic data storage medium, CD-ROMs, any other optical data storage medium, any physical medium with a pattern of holes, RAM, PROM and EPROM, FLASH-EPROM, NVRAM, any other memory chip or cartridge, and network versions thereof.

[0181] Non-transient media are distinct from transmission media but may be used in conjunction with transmission media; transmission media participate in the transmission of information between non-transient media; for example, transmission media include coaxial cables, copper wires, and optical fibers, including the wires that make up bus 502; transmission media may also take the form of sound waves or light waves, such as radio waves and infrared data communications.

[0182] Although the present application has been described above with reference to various embodiments, it should be understood that many changes and modifications may be made without departing from the scope of the present application. That is, the methods, systems and devices discussed above are examples. Various configurations may appropriately omit, replace or add various processes or components. For example, in alternative configurations, the method may be performed in an order different from the order described, and / or various components may be added, omitted and / or combined. Moreover, the features described with respect to certain configurations may be combined in various other configurations, such as different aspects and elements of the configuration may be combined in a similar manner. In addition, as the technology develops, the elements therein may be updated, i.e., many elements are examples and do not limit the scope of the present disclosure or claims.

[0183] Specific details are given in the specification to provide a thorough understanding of the exemplary configurations including implementations. However, the configurations may be practiced without these specific details. For example, well-known circuits, processes, algorithms, structures, and techniques have been shown without unnecessary details to avoid obscuring the configurations. This description provides only example configurations and does not limit the scope, applicability, or configurations of the claims. On the contrary, the foregoing description of the configurations will provide those skilled in the art with an enabling description for implementing the described techniques. Various changes may be made to the functions and arrangements of the elements without departing from the spirit or scope of the present disclosure.

[0184] In summary, it is intended that the above detailed description is considered to be illustrative rather than restrictive, and it should be understood that the above embodiments should be understood to be only used to illustrate the present invention and not to limit the scope of protection of the present invention. After reading the contents of the present invention, the technician can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A distribution network drone inspection system based on front-end AI visual recognition, characterized in that: include: A server, and one or more drones communicatively connected to the server; The server comprises: The computing unit is configured to obtain and analyze the flight data of the UAV from the UAV; generate, save, and optimize a flight strategy model for evaluating the quality of the UAV's mission execution, and generate a control instruction for the UAV's flight control based on the evaluation result; A control unit configured to communicate with one or more UAVs and send the control command to the target UAV; The inspection system includes the following two working stages to realize the network distribution drone inspection process based on front-end AI visual recognition: Preparation stage: The operator controls the drone to fly along the preset route, perform standard flight actions and flight routes, check the target checkpoints, record key data sets, and generate standard route files; Modeling stage: training and establishing a flight strategy model based on the standard route file; Actual flight stage: During the automated inspection process of the UAV, real-time data of the UAV is obtained and a real-time data set is generated; the flight strategy model is run to evaluate the quality of the UAV's execution of tasks during the automated inspection process, and subsequent flight strategies are adjusted.

2. The inspection system according to claim 1, characterized in that: The key data sets include: Flight route coordinates, i.e., the coordinates of the drone’s position at a preset interval; Drone status, including drone attitude, speed, and pointing angle; Checkpoint data: coordinates of key points and standard images of one or more checkpoints.

3. The inspection system according to claim 2, characterized in that: The flight strategy model is built based on a deep Q network and includes the following steps: Input the standard route file as training data into the deep Q network; The deep Q network then takes a large amount of UAV flight status data and a large number of checkpoint images as input, and uses a reinforcement learning framework to train and calculate the cumulative reward value Q(s,a) to train the flight strategy model to evaluate the quality of the current UAV's mission execution, and output the flight score and the optimal flight strategy; The flight strategy model is trained by optimizing the loss function of the deep Q network and using the gradient descent method.

4. The inspection system according to claim 3, characterized in that: The calculation of the reward value Q(s,a) is based on at least the following evaluation indicators: The degree to which the flight route of the drone complies with the route specified in the standard route file; The attitude quality of the drone during flight; Image quality of the checkpoints captured by the drone.

5. The inspection system according to claim 4, characterized in that: During the actual flight phase, the UAV performs automated inspections based on its stored flight program.

6. The inspection system according to claim 5, characterized in that: During the actual flight phase, the UAV also includes receiving control instructions from the server, thereby executing intervention control of the control instructions.

7. The inspection system according to claim 6, characterized in that: During the actual flight phase, the server performs the following steps to decide whether to intervene in the inspection work of controlling the drone: S100: continuously acquires real-time data from drones and generates real-time data sets; S200: running the flight strategy model to evaluate the quality of the UAV's execution of tasks during the automated inspection based on the real-time data set, thereby calculating a flight score; S300: Determine the necessity of intervening in controlling the UAV based on the flight score; S400: If it is determined to intervene in the control of the UAV, a corresponding control instruction is generated based on the flight score, and sent to the UAV to intervene in the control of the UAV to execute the control instruction.

8. The inspection system according to claim 7, characterized in that: The flight strategy model evaluates the quality of the UAV's mission execution, and the evaluation indicators included include: The degree to which the flight route of the drone complies with the route specified in the standard route file; The attitude quality of the drone during flight; Image quality of the checkpoints captured by the drone.

9. The inspection system according to claim 8, characterized in that: The drone is equipped with at least one image acquisition device, and the image acquisition device is composed of one or more of the following sensors or image devices: camera, video camera, thermal imaging camera, infrared camera, night vision sensor, depth camera, ranging sensor, laser ranging sensor, radio ranging sensor.

10. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the functions of the inspection system as described in any one of claims 1 to 9 are performed.

Images