Unmanned aerial vehicle group cooperative inspection control method for power station inspection

Abstract

The invention discloses an unmanned aerial vehicle group cooperative inspection control method for power station inspection, and the method comprises the steps: determining an unmanned aerial vehicle group according to a power station inspection task; performing task allocation of the heterogeneous unmanned aerial vehicle cluster by using a multi-objective optimization algorithm; performing unmanned aerial vehicle path planning by adopting an improved dynamic multi-population particle swarm algorithm; based on a graph theory and master-slave control, unmanned aerial vehicle group attitude consistency control is realized; integrating a target detection technology based on deep learning, and constructing an inspection anomaly detection model; performing remote fault diagnosis analysis on the real-time inspection data through an inspection anomaly detection model to obtain an inspection anomaly result; optimizing the inspection anomaly detection model according to the target condition of the real-time inspection data; and dynamically controlling task allocation and path planning of the unmanned aerial vehicle group according to an abnormal inspection result. Through the scheme, potential problems can be found and processed in time, the intelligent control level of power station inspection is improved, and the efficiency and accuracy of unmanned aerial vehicle group cooperative inspection are improved.

Description

Technical Field

[0001] The present invention generally relates to the field of artificial intelligence technology. More specifically, the present invention relates to a collaborative inspection control method for a group of unmanned aerial vehicles for power station inspection. Background Art

[0002] At present, drone inspections are widely used in the fields of electricity, oil, natural gas, transportation, environmental protection, etc. By using drones for inspections, vehicle emissions and road traffic congestion can be reduced, and the environment and health can be protected. Power stations are an important part of the power system, and the normal operation of power stations is the basis for ensuring the safe, reliable and economical operation of the power system. With the increasing number of power stations and their larger scale, the types of equipment that need to be inspected are diverse and large in number, and the workload and difficulty of operation and maintenance are increasing. In order to improve the efficiency of inspections, drone swarm inspections have been applied to power station inspections, but due to the complex environment, large number of equipment, and high safety requirements of power stations, drone swarm inspections still face a series of challenges.

[0003] Power station inspections face difficulties such as narrow flight airspace, diverse inspection contents, high operational risks, and strong interference from various magnetic fields. There are also coordination issues in task allocation, path planning, and obstacle avoidance capabilities.

[0004] In view of this, there is an urgent need to provide a collaborative inspection control method for drone swarms for power station inspections, so as to improve the efficiency and accuracy of collaborative inspections by drone swarms. Summary of the invention

[0005] In order to at least solve one or more of the technical problems mentioned above, the present invention proposes a collaborative inspection control method for a group of drones used for power station inspection in multiple aspects.

[0006] In the first aspect, the present invention provides a method for collaborative inspection control of a drone swarm for power station inspection, comprising: determining a drone swarm according to the task of the power station inspection; using a multi-objective optimization algorithm to perform task allocation for a heterogeneous drone swarm; using an improved dynamic multi-swarm particle swarm algorithm to perform drone path planning; based on graph theory and master-slave control, achieving attitude consistency control of the drone swarm; constructing an inspection anomaly detection model based on a target detection algorithm based on deep learning; wherein the inspection anomaly detection model includes a self-obstacle avoidance module; using a drone swarm that simulates the passive inertial emergency obstacle avoidance of a pigeon swarm and a binocular vision system to detect the The autonomous obstacle avoidance module performs flight verification, three-dimensional reconstruction and path planning to perform obstacle avoidance training and optimization on the autonomous obstacle avoidance module; obtains real-time inspection data collected during the inspection of the drone group; performs remote fault diagnosis and analysis on the real-time inspection data through the inspection anomaly detection model to obtain inspection anomaly results; optimizes the inspection anomaly detection model according to the target conditions of the real-time inspection data, the target conditions including small target detection, complex background and / or multi-scale changes; and dynamically controls the task allocation and path planning of the drone group through a reinforcement learning algorithm according to the inspection anomaly results.

[0007] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for controlling the collaborative inspection of a swarm of drones for power station inspection as described in any one of the first aspects is implemented.

[0008] Through the above-provided method for cooperative inspection control of drone swarms for power station inspection, the embodiment of the present invention uses a multi-objective optimization algorithm to assign tasks to drone swarms, and adopts an improved dynamic multi-swarm particle swarm algorithm for drone path planning, which can efficiently assign tasks to drones of different types and capabilities, and plan the optimal path to improve inspection efficiency and coverage. The application of reinforcement learning algorithms and obstacle avoidance strategies improves the autonomy and adaptability of drones in complex environments and reduces the risk of accidents. It helps to continuously optimize the operating efficiency and safety of drone swarms in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0010] Figure 1 An exemplary flow chart of a method for controlling a swarm of drones for power station inspection in accordance with some embodiments of the present invention is shown. DETAILED DESCRIPTION

[0011] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0012] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0013] Figure 1 An exemplary flow chart of a method 100 for controlling a group of drones for power station inspection in accordance with some embodiments of the present invention is shown.

[0014] like Figure 1 As shown, in method 100, step 101 determines a drone group according to the task of the power station inspection. In one embodiment scenario, it may include:

[0015] First, determine the task requirements of the power station inspection. Then, configure the type and quantity of drones, as well as the sensors and communication equipment for the drone swarm according to the power station environment and the task requirements. This ensures that the drone swarm can meet the specific inspection task requirements.

[0016] In preparation for the inspection mission, the communication network is crucial to ensure the coordinated work of the drone swarm and the real-time analysis of data. Therefore, a 5G communication network is also needed for the drone swarm to perform ultra-long-distance low-latency control and real-time data return.

[0017] Further, the process proceeds to step 102: using a multi-objective optimization algorithm to perform task allocation for a heterogeneous UAV cluster.

[0018] Use multi-objective optimization algorithms, such as genetic algorithms or particle swarm algorithms, to perform task allocation for heterogeneous UAV clusters. These algorithms can simultaneously consider multiple objectives, such as track distance cost, track threat cost, track energy cost, and multi-UAV collaborative performance.

[0019] In one embodiment, a multi-objective optimization algorithm is used to perform task allocation for a heterogeneous drone cluster, including the following steps:

[0020] First, the positions and velocities of the drone swarm are initialized, where the position of each drone represents a potential task allocation solution.

[0021] Then, a fitness evaluation is performed on each UAV, wherein the fitness evaluation method includes: evaluating the total flight distance using a track distance cost function, evaluating the total exposure time of the threat area in the flight path using a track threat cost function and / or evaluating the total energy consumption using a track energy consumption cost function.

[0022] Track distance cost (f d ) is: Among them, d ij represents the track length of the i-th UAV on the j-th track segment;

[0023] Track threat cost (f t ) is:

[0024]

[0025] Among them, t ij It represents the distance that the i-th UAV passes through the threat area on the j-th track segment, and N represents the penalty coefficient.

[0026] Track energy cost (f e ) is:

[0027]

[0028] e ij =32.44+20log 10 (D ij )+20log 10 (F),

[0029] Where F represents the radio frequency, D ij It represents the distance between the i-th UAV and the base station at the j-th track point.

[0030] Spatial coordination index (d safe ) is:

[0031]

[0032] Among them, min dis j (i,i nei ) indicates the i-th UAV and its nearest neighboring UAV i that has not been traversed on the j-th track segment. nei The Euclidean distance between .

[0033] After evaluating the fitness of each drone, the individual optimal solution and the global optimal solution are updated for each drone. Based on the combination of particle swarm optimization (PSO) and multi-objective optimization, the position and speed of the drone are updated by tracking the individual and global optimal solutions. The formula includes:

[0034] Location Updates:

[0035] Speed ​​Update:

[0036] in, is the position of drone i at time t, is the speed, w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, p i is the individual optimal position of drone i, p g is the global optimal position.

[0037] In another embodiment, an improved multi-objective optimization genetic algorithm NSGA-III algorithm can also be used, and the improved algorithm improves the performance of the algorithm by using a crossover strategy based on individual evaluation.

[0038] The formula for the individualized assessment index (CAD) is:

[0039]

[0040] The individual diversity indicator formula is:

[0041] Among them, SDE(p i ) represents individual p i transfer density.

[0042] Individual convergence index formula:

[0043]

[0044] Among them, Disc(p i ,P) represents individual p i and the Euclidean distance between the ideal point z′.

[0045] The detailed framework of UAV cluster task allocation based on the multi-objective optimization algorithm provided by the present invention covers trajectory planning, cost model and individual evaluation strategy, thereby realizing efficient collaborative task allocation of UAV clusters.

[0046] In another embodiment scenario, suppose there are 5 different types of drones that need to perform inspection tasks in a certain area of ​​a power station. Task allocation first minimizes the following three objectives: track distance cost (f d ): Total flight distance. Track threat cost (f t ): Total exposure time of the threat area in the flight path. Track energy cost (f e ): total energy consumption. The specific implementation steps are:

[0047] Initialization: Randomly initialize the position and velocity of the drone swarm. The position of each drone represents a potential task allocation solution.

[0048] Fitness evaluation: The fitness of each UAV (solution) is evaluated using the three objective functions mentioned above.

[0049] Update individual and global optimal solutions: For each drone, update its individual optimal solution and global optimal solution.

[0050] Update position and speed: Update the position and speed of the drone based on the individual optimal solution and the global optimal solution.

[0051] Repeat the above steps until the stopping condition is met (such as reaching the maximum number of iterations or the quality of the solution meets the preset requirements).

[0052] By using the multi-objective optimization algorithm provided by the present invention, such as genetic algorithm or particle swarm algorithm, task allocation for heterogeneous drone clusters can be performed. This can solve the problem of how to efficiently allocate tasks to drones of different types and capabilities, and how to plan the optimal path to improve inspection efficiency and coverage.

[0053] Further, in step 103: an improved dynamic multi-swarm particle swarm algorithm is used to perform UAV path planning.

[0054] In one embodiment, an improved dynamic multi-population particle swarm algorithm is used to perform drone path planning, and the steps include:

[0055] Create multiple subgroups, each subgroup contains a certain number of particles, assign a random initial position and velocity to each particle, and define the optimization objective function of the problem. Calculate the fitness value of each particle's objective function, where the objective function includes: track distance cost function, track threat cost function, and track energy consumption cost function.

[0056] For each particle, update the speed and position according to its current position and best historical position. In each subgroup, find the particle with the highest fitness and update the subgroup extreme value l best. Dynamically adjust the size of the subgroup according to its performance.

[0057] Information is shared between different subgroups, allowing particles to migrate to other subgroups to promote population diversity and global search capabilities. Among all subgroups, the particle with the highest fitness is found and the global extreme value gBest is updated.

[0058] Stop the iteration according to the preset termination condition; and

[0059] Return the global extremum as the optimal solution for the drone path planning.

[0060] The formula for updating the speed and position includes:

[0061] Update speed:

[0062] Update location:

[0063] in, represents the velocity of particle i in the dth dimension in the kth iteration, represents the position of particle i in the dth dimension in the kth iteration, p id represents the individual extreme value of the particle swarm so far, p gd represents the global extreme value of the particle swarm so far, c1 and c2 represent learning factors; ω represents the weight coefficient.

[0064] It should be noted that according to the fitness values ​​of the above-mentioned embodiment, the median value in ascending order is used to demarcate the sub-populations, and dynamic adjustments are implemented in each generation. The smaller fitness value is the "top particle" and the larger fitness value is the "bottom particle"; then the same number of particles are extracted from the two to form a local PSO model, and multiple (for example, 3) populations are balanced in size, namely the "dominant group" closer to the optimal solution, the "mixed group" of the local PSO model, and the "inferior group".

[0065] The improved dynamic multi-swarm particle swarm algorithm provided by the present invention can effectively provide a solution for drone path planning, taking into account multiple goals at the same time to achieve optimal task allocation. This method can ensure that when a drone swarm performs a task, it not only considers a single goal, but also comprehensively considers multiple goals to achieve the best overall performance.

[0066] Further, in step 104: based on graph theory and master-slave control, attitude consistency control of the drone swarm is achieved.

[0067] In one embodiment, based on graph theory and master-slave control, realizing attitude consistency control of a drone swarm may include:

[0068] Establish the attitude dynamics model of the drone group, including the position, velocity, acceleration and attitude angle of each drone. The attitude angle includes: roll angle, pitch angle and yaw angle. The specific formula includes:

[0069] x i =A i x i +B i u i , where x i represents the state vector of drone i, u i Represents the control input, A i and B i It is the system matrix of the drone.

[0070] Graph theory is used to define the communication network topology between drones, and an undirected graph G = {v, ε} is constructed, where v is the graph node set and ∈ is the graph edge set; nodes represent drones and edges represent communication connections between drones.

[0071] Calculate the Laplace matrix of the graph L = DA, where D is the degree matrix and A is the adjacency matrix. The specific Laplace matrix (also called the Laplace matrix of the graph) is derived from the adjacency matrix and degree matrix of the graph. For an undirected graph G = (V, E), where V is the vertex set and E is the edge set, the Laplace matrix is ​​defined as follows:

[0072] The degree matrix D is a diagonal matrix with the elements d on the diagonal ii Represents vertex v i The degree of vertex v i The number of connected edges. The mathematical expression is:

[0073] Among them, a ij is an element in the adjacency matrix A, representing the vertex v i With vertex v j Are there edges connecting them?

[0074] The degree matrix A is a square matrix. If the vertex v i With vertex v j There is an edge between them, then a ij =1, otherwise a ij =0.

[0075] The matrix L is symmetric and semi-positive definite, its eigenvalues ​​are all non-negative, and at least one eigenvalue is zero. The corresponding eigenvector is the all-one vector e=(1,1,...,1) T This eigenvector states that if the graph is connected, then there exists a globally consistent metric that makes the metrics of all nodes equal.

[0076] Furthermore, a master-slave controller is created, wherein the master-slave controller includes controlling the attitude angle and angular velocity of the drone group to track the attitude angle and angular velocity of the leader drone through consistency control. According to the master-slave control theory, a consistency control law is designed to ensure that the attitude angle and angular velocity of the drone group can track the attitude angle and angular velocity of the leader drone.

[0077] Adjust the control parameters to optimize the attitude consistency of the drone group, including control gain and damping ratio. The control gain k is a positive constant used to adjust the strength of the control force. A larger k value can speed up the convergence of the system, but it may also increase the oscillation of the system.

[0078] According to the above embodiments of the present invention, those skilled in the art can realize attitude consistency control of a drone swarm, ensuring that the drone swarm can maintain formation and work in coordination when performing tasks. This method can further ensure that when a drone swarm performs a task, it not only considers a single goal, but also comprehensively considers multiple goals to achieve the best overall performance.

[0079] Further, in step 105: based on the target detection algorithm of deep learning, a patrol anomaly detection model is constructed. The patrol anomaly detection model includes an autonomous obstacle avoidance module; the autonomous obstacle avoidance module is subjected to flight verification, three-dimensional reconstruction and path planning by using a pigeon-like passive inertial emergency obstacle avoidance drone cluster and a binocular vision system, so as to perform obstacle avoidance training and optimization on the autonomous obstacle avoidance module.

[0080] In one embodiment, a patrol anomaly detection model is constructed based on a deep learning target detection algorithm. The specific implementation steps may include:

[0081] Collect infrared image data of power station equipment. Then preprocess the infrared image data, including denoising, normalization, and enhancement. In the specific implementation, it can include correcting the RGB values ​​of the infrared image of the power station equipment, and mapping the corrected RGB values ​​to the Lab space to obtain the data set required for clustering. This step can standardize the image data to make it more suitable for subsequent analysis and processing.

[0082] Using a deep convolutional neural network, features are extracted from the infrared image data to identify abnormal conditions, such as abnormally hot areas.

[0083] The model is trained using annotated infrared image datasets, and the weights of the deep convolutional neural network are optimized using the back propagation algorithm and gradient descent method. The deep convolutional neural network includes multiple convolutional layers, pooling layers, and fully connected layers. Specifically:

[0084] The convolution layer formula includes f(x)=max(0,w*x+b), where w is the convolution kernel weight, * represents the convolution operation, b is the bias term, x is the input feature map, and f(x) is the output feature map.

[0085] The pooling layer formula includes f(x)=max i,j x i,j , where x i,j is the element of the local position (i, j) of the input feature map, and f(x) is the output feature map; i and j represent the row and column indices in the two-dimensional space.

[0086] The fully connected layer formula includes y=Wa+b, where W is the weight matrix, a is the output of the previous layer, b is the bias vector, and y is the output vector.

[0087] The loss functions include: Among them, y o,c is the one-hot encoding of the true label, p o,c is the probability that the model predicts class c.

[0088] In one embodiment, the above process optimizes the weights of the deep convolutional neural network by using a back propagation algorithm and a gradient descent method, and the implementation may include:

[0089] Starting from the output layer, we go backwards through each layer of the network and calculate the gradient of the loss function with respect to the weights.

[0090] Use the gradient descent algorithm to update the network weights to minimize the loss function. The formula includes:

[0091] Among them, w is the weight, is the partial derivative of the loss function with respect to the predicted value, is the partial derivative of the predicted value with respect to the weight.

[0092] Update weights in, is the learning rate, which controls the step size of weight update.

[0093] The weights and biases in the deep convolutional neural network are updated according to the calculated gradient and learning rate.

[0094] The loop is iteratively updated until the performance of the deep convolutional neural network on the validation set no longer improves or reaches a preset number of iterations.

[0095] In addition to the steps implemented in the above embodiments, in some embodiments, the performance of the model can also be evaluated using a validation set, including indicators such as accuracy and recall, and the model structure and parameters can be adjusted according to the evaluation results to improve the accuracy of detection. The model can also implement real-time monitoring and abnormal alarm functions. Based on the inspiration of the above embodiments of the present invention, those skilled in the art can perform verification and evaluation according to actual needs, as well as implement abnormal alarm methods, and the present invention does not impose too many restrictions on this.

[0096] By building an efficient deep learning-based anomaly detection model, we can use it for power station inspections. This model can automatically identify abnormal heating areas and other fault signs in equipment, improving the efficiency and safety of power station operations.

[0097] In one embodiment, by using a pigeon-like passive inertial emergency obstacle avoidance drone cluster and a binocular vision system to perform flight verification, three-dimensional reconstruction and path planning on the autonomous obstacle avoidance module, the method may include:

[0098] The obstacle avoidance behavior of the drone swarm is verified by setting up two scenarios: emergency obstacle avoidance of stationary obstacles on the ground and emergency obstacle avoidance of moving obstacles in the air.

[0099] In 3D reconstruction, image processing algorithms such as color segmentation and / or object detection algorithms are first used to detect obstacles in the image in real time and determine their positions at the pixel level. Stereo matching algorithms are then used to obtain the depth information of the obstacles to determine the 3D coordinates of the obstacles.

[0100] The drone adjusts its heading according to the calculated expected obstacle avoidance angle, passes through the obstacle from the largest gap between obstacles, and completes the obstacle avoidance. obsi The calculation formula may include:

[0101] Among them, ψ fly is the heading angle of the drone, θ obsi is the angle between the drone and the obstacle, R vis is the field of view radius of the drone, R obsi is the distance between the drone and the obstacle.

[0102] The UAV heading is controlled using the proportional-integral-derivative (PID) control law. The formulas include:

[0103]

[0104] Among them, K p ,K I ,K D ,K S are proportional control gain, integral control gain, differential control gain and inertial stability control gain respectively; τ is the time delay caused by the visual sensor detection of the control system, θ err =ψ fly -θ exp is the error between the current flight direction of the UAV and the expected flight direction, θ exp is the desired heading angle.

[0105] Update the heading angle according to the control rate of the heading controller, and update the spatial position of the drone. The formula includes:

[0106] Among them, V represents the current speed of the drone.

[0107] By integrating the self-obstacle avoidance module into the inspection anomaly detection model, the obstacle avoidance behavior of the pigeon flock is used for training, and the obstacle avoidance strategy of the drone is adjusted to improve its adaptability in complex environments. This enables the module to automatically adjust the flight path of the drone to avoid collision when an obstacle is detected, ensuring efficient and safe execution of inspection tasks.

[0108] In some embodiments, a binocular vision system can be used to reconstruct a three-dimensional scene to obtain three-dimensional information of the environment. Based on the three-dimensional reconstruction results, the flight path of the drone can be optimized to ensure that the path avoids obstacles and achieves the predetermined inspection target.

[0109] Through the above pigeon-like passive inertial emergency obstacle avoidance strategy, combined with the binocular vision system for 3D reconstruction and path planning, the obstacle avoidance capability of the drone is improved. This helps the drone to fly safely in complex environments and reduce the risk of accidents caused by obstacles.

[0110] Further, in step 106, real-time inspection data collected during the drone group inspection is obtained. Through the inspection anomaly detection model constructed in step 105, step 107 is executed: through the inspection anomaly detection model, remote fault diagnosis and analysis are performed on the real-time inspection data to obtain inspection anomaly results. Including preprocessing of inspection images in real-time inspection data, including denoising, contrast enhancement, etc., to improve the detectability of small targets. Use deep learning models, such as convolutional neural networks (CNN), to automatically extract features from raw data. Use 5G networks to realize real-time video monitoring and remote fault diagnosis and analysis of drone inspections, improve real-time understanding and response capabilities of power station status, and adapt to environmental changes and task requirements.

[0111] Further, in step 108, the inspection anomaly detection model is optimized according to target conditions of the real-time inspection data, wherein the target conditions include small target detection, complex background and / or multi-scale changes.

[0112] In one embodiment, a multi-scale convolutional neural network (CNN) structure, such as Feature Pyramid Networks (FPN), can be used to fuse feature maps of different scales and enhance the detection ability of small targets and multi-scale changes. Attention mechanisms, such as self-attention or channel attention, are added to CNN to focus on key areas and targets in the image. Contextual information can also be used to enhance the model's ability to recognize small targets in complex backgrounds. This can be achieved by introducing a global context network (such as GCN). In order to analyze the image in more detail, super-resolution technology can be applied to improve image resolution and enhance the detail information of small targets, thereby improving detection accuracy. For small target detection, the anchor frame mechanism is optimized, and a smaller anchor frame is designed to adapt to the characteristics of small targets.

[0113] Through the above technical solutions, the inspection anomaly detection model of the drone swarm can be effectively optimized, and the detection capabilities of small targets, complex backgrounds and multi-scale changing targets can be improved, thereby improving the efficiency and accuracy of inspections.

[0114] Further, in step 109: according to the abnormal inspection results, the task allocation and path planning of the drone group are dynamically controlled through a reinforcement learning algorithm.

[0115] In one embodiment, according to the inspection abnormality results, the task allocation and path planning of the drone group are dynamically controlled by a reinforcement learning algorithm, and the implementation steps include:

[0116] Define the state space of the drone swarm, including the position, speed, mission status, and environmental information of the drones. The state space is the basis for decision-making in the reinforcement learning algorithm. Then define the action space of the drone swarm, including movement direction, speed change, mission selection, etc.

[0117] When the drone swarm performs inspection tasks, the behavior of the drone swarm is evaluated through a reward function, wherein the reward function is associated with indicators such as task completion, energy consumption, and obstacle avoidance effect. The policy network is trained using a deep reinforcement learning method to learn the optimal strategy through interaction with the environment; wherein the deep reinforcement learning method can be selected from algorithms such as Q-learning and Deep Q-Network (DQN).

[0118] Based on the learned optimal strategy, tasks are dynamically assigned to the drone swarm and paths are planned to complete the tasks; actions output by the execution strategy network are executed and strategies are adjusted based on environmental feedback.

[0119] Through the above steps and formulas, the task allocation and path planning of the drone swarm can be dynamically controlled to improve the efficiency and adaptability of the drone swarm. This reinforcement learning-based method enables the drone swarm to autonomously learn and optimize task execution strategies in complex environments. Through intelligent task allocation and path planning, the drone swarm can be controlled to cover the inspection area more efficiently, reducing duplication of work and time waste.

[0120] In summary, this technical solution can improve the autonomy and adaptability of drones in complex environments and reduce the risk of accidents by integrating reinforcement learning algorithms and obstacle avoidance strategies. The application of real-time monitoring and deep learning technology makes fault detection more accurate and improves the intelligent control level of power station inspections, which helps to discover and deal with potential problems in a timely manner, thereby significantly improving the efficiency and accuracy of drone group collaborative inspections and reducing operation and maintenance costs.

[0121] Based on the above, the present invention also discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by one or more processors, the collaborative inspection control method of a group of drones for power station inspection described in the multiple embodiments or implementation methods described in the previous text is implemented.

[0122] Although multiple embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art may conceive of many changes, modifications, and alternatives without departing from the thought and spirit of the present invention. It should be understood that in the process of practicing the present invention, various alternatives to the embodiments of the present invention described herein may be adopted. The appended claims are intended to define the scope of protection of the present invention, and therefore cover equivalents or alternatives within the scope of these claims.

Claims

1. A method for controlling a group of drones for power station inspection, characterized in that: include: Determine a drone group according to the tasks of the power station inspection; Use multi-objective optimization algorithms to allocate tasks for heterogeneous drone clusters; The improved dynamic multi-swarm particle swarm algorithm is used for UAV path planning; Based on graph theory and master-slave control, the attitude consistency control of the drone swarm is realized; Based on the deep learning target detection algorithm, a patrol anomaly detection model is constructed; wherein the patrol anomaly detection model includes an autonomous obstacle avoidance module; flight verification, three-dimensional reconstruction and path planning of the autonomous obstacle avoidance module are performed by using a pigeon-like passive inertial emergency obstacle avoidance drone cluster and a binocular vision system; obstacle avoidance training and optimization are performed on the autonomous obstacle avoidance module; Obtain real-time inspection data collected by drone swarms during inspections; Through the inspection anomaly detection model, remote fault diagnosis and analysis is performed on the real-time inspection data to obtain inspection anomaly results; Optimizing the inspection anomaly detection model according to the target conditions of the real-time inspection data, wherein the target conditions include small target detection, complex background, and multi-scale changes; According to the abnormal inspection results, the task allocation and path planning of the drone swarm are dynamically controlled through a reinforcement learning algorithm.

2. The method according to claim 1, characterized in that Determining a drone group according to the task of the power station inspection includes: Determine the task requirements of the power station inspection; According to the power station environment and the mission requirements, configure the type and quantity of drones, as well as the sensors and communication equipment carried by the drone group; Build a 5G communication network for drone swarms.

3. The method according to claim 1, characterized in that The multi-objective optimization algorithm is used to perform task allocation for heterogeneous UAV clusters, including: Initialize the position and velocity of the drone swarm, where the position of each drone represents a potential task allocation solution; Performing fitness evaluation on each UAV, wherein the fitness evaluation method includes: evaluating the total flight distance using a track distance cost function, evaluating the total exposure time of the threat area in the flight path using a track threat cost function, and evaluating the total energy consumption using a track energy consumption cost function; For each drone, update its individual optimal solution and global optimal solution; Based on the combination of particle swarm optimization (PSO) and multi-objective optimization, the position and speed of the UAV are updated by tracking the individual optimal solution and the global optimal solution.

4. The method according to claim 1, characterized in that: The improved dynamic multi-swarm particle swarm algorithm is used to perform UAV path planning, and the steps include: Create multiple subgroups, each containing a certain number of particles, randomly assign an initial position and velocity to each particle, and define the optimization objective function of the problem; Calculating the fitness value of each particle's objective function, wherein the objective function includes: a track distance cost function, a track threat cost function, and a track energy consumption cost function; For each particle, update the velocity and position based on its current position and best historical position; In each subgroup, find the particle with the highest fitness and update the subgroup extreme value lbest; Dynamically adjust the size of subgroups based on their performance; Information sharing between different sub-swarms allows particles to migrate to other sub-swarms to promote population diversity and global search capabilities; Among all subgroups, find the particle with the highest fitness and update the global extreme value gBest; Stop the iteration according to the preset termination conditions; Return the global extremum as the optimal solution for the drone path planning.

5. The method according to claim 1, characterized in that The method of realizing attitude consistency control of a swarm of drones based on graph theory and master-slave control includes: Establishing the attitude dynamics model of the drone group, including the position, velocity, acceleration and attitude angle of each drone, wherein the attitude angles include: roll angle, pitch angle and yaw angle; The communication network topology between drones is defined using graph theory, and an undirected graph G = {v, ε} is constructed, where v is the graph node set and ∈ is the graph edge set; nodes represent drones, and edges represent communication connections between drones; Compute the Laplace matrix of the graph L = DA, where D is the degree matrix and A is the adjacency matrix; Creating a master-slave controller, wherein the master-slave controller includes controlling the attitude angle and angular velocity of the drone group through consistency control and being able to track the attitude angle and angular velocity of the leader drone; Control parameters are adjusted to optimize the attitude consistency of the UAV swarm, including control gains and damping ratios.

6. The method according to claim 1, characterized in that The target detection algorithm based on deep learning constructs an inspection anomaly detection model, and the steps include: Collect infrared image data of power station equipment; Preprocessing the infrared image data, including denoising, normalization, and enhancement; Using a deep convolutional neural network, extracting features from the infrared image data to identify abnormal heating areas; Use the labeled infrared image dataset to train the model; The deep convolutional neural network weights are optimized by back propagation algorithm and gradient descent method.

7. The method according to claim 6, characterized in that The method of optimizing the weights of the deep convolutional neural network by back propagation algorithm and gradient descent method includes: Starting from the output layer, go backwards through each layer of the network, calculate the gradient of the loss function with respect to the weights, and use the gradient descent algorithm to update the network weights to minimize the loss function; Update the weights and biases in the deep convolutional neural network according to the calculated gradient and learning rate; The loop is iteratively updated until the performance of the deep convolutional neural network on the validation set no longer improves or reaches a preset number of iterations.

8. The method according to claim 1, characterized in that The method of performing flight verification, three-dimensional reconstruction and path planning on the autonomous obstacle avoidance module by utilizing a pigeon-like passive inertial emergency obstacle avoidance drone cluster and a binocular vision system comprises: Image processing algorithms using color segmentation and / or object detection algorithms to detect obstacles in the image in real time and determine their location at the pixel level; Use stereo matching algorithm to obtain the depth information of obstacles to determine the three-dimensional coordinates of obstacles; The drone adjusts its heading according to the calculated desired obstacle avoidance angle, passes through the obstacle through the largest gap between obstacles, and completes obstacle avoidance. The UAV heading is controlled using the proportional-integral-derivative (PID) control law; The heading angle is updated according to the control rate of the heading controller, and the spatial position of the drone is updated.

9. The method according to claim 1, characterized in that: According to the inspection abnormality results, the task allocation and path planning of the drone group are dynamically controlled by a reinforcement learning algorithm, and the steps include: Define the state space of the drone swarm, including the position, speed, mission status and environmental information of the drones; Define the action space of the drone swarm, including movement direction, speed change, and mission selection; Evaluate the behavior of the drone swarm through a reward function, where the reward function is associated with indicators such as task completion, energy consumption, and obstacle avoidance effect; Using a deep reinforcement learning method to train a policy network to learn the optimal policy through interaction with the environment; wherein the deep reinforcement learning method includes a Q-learning algorithm; Dynamically assign tasks to drone swarms based on the learned optimal strategy and plan paths to complete the tasks; and Execute the actions output by the policy network and adjust the policy based on the environment feedback.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the collaborative inspection control method of a swarm of drones for power station inspection as described in any one of claims 1 to 9 is implemented.

Images