Method and system for releasing insulator detection robot by unmanned aerial vehicle based on visual guidance

Abstract

The invention provides a method and system for releasing an insulator detection robot by an unmanned aerial vehicle based on visual guidance, and the method comprises the steps: enabling the unmanned aerial vehicle to recognize a whole insulator string through an insulator recognition model, and obtaining a target position; after the unmanned aerial vehicle moves to the upper ends of the insulators, tail end recognition is carried out, the insulators on the insulator string are recognized one by one, and the string falling position is determined; after string falling operation is completed, the robot is released at the target insulator through the optical auxiliary structure and the throwing main body structure, and insulation resistance detection is carried out. The design of the unmanned aerial vehicle automatic online and offline and robot putting strategy provides multiple beneficial effects, and the effects improve the interaction efficiency and safety between the unmanned aerial vehicle and the robot.

Description

Technical Field

[0001] The present invention belongs to the field of power transmission line maintenance, and in particular relates to a method and system for using a drone to drop an insulator inspection robot based on vision guidance. Background Art

[0002] In recent years, transmission line operation robots have developed rapidly and can replace humans to perform various transmission maintenance tasks, such as replacing insulators, tightening tension clamp bolts, and repairing broken conductors, which can greatly improve work efficiency. However, how to efficiently deploy robots on transmission lines is still an urgent problem to be solved. Most operation robots do not have the function of automatic online and offline functions, and mainly rely on external equipment to assist online and offline functions. They have extremely high requirements for human and online and offline auxiliary equipment, and when the geographical environment is complex, they have low adaptability and poor practicality, which greatly affects the efficiency of operation and maintenance. Therefore, while meeting the needs of live transmission line maintenance, ensuring that robots can go online and offline flexibly and quickly is also an urgent problem to be solved.

[0003] Thanks to the advantages of multi-rotor drones such as small size, light weight and low cost, drones are widely used in power transmission line operations. However, most drones in the current power transmission context can only complete environmental perception tasks such as monitoring, detection and testing, and cannot actively interact with target objects or target environments, making transmission line drones only have inspection functions but not maintenance functions. Unable to complete more in-depth maintenance work, if it needs to cooperate with other equipment, docking to achieve separation or transportation functions has become a necessary link for it to complete a mission cycle. This technology plays an important role in the fields of operation coordination, hardware distribution, emergency rescue, etc. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, the present invention provides a method and system for using a UAV to drop an insulator inspection robot based on vision guidance, which combines optical means and uses visual recognition technology to achieve the automated docking action of the UAV.

[0005] The present invention adopts the following technical solution.

[0006] The present invention provides a method for using a UAV to drop an insulator detection robot based on vision guidance, comprising:

[0007] The UAV uses the insulator recognition model to identify the entire insulator string and obtain the target position;

[0008] The drone moves to the upper end of the insulator and performs terminal identification, identifying the insulators on the insulator string one by one to determine the location of the string drop;

[0009] After completing the string drop operation, the robot is released at the target insulator using the optical auxiliary structure and the main structure to perform insulation resistance testing.

[0010] Preferably, the drone identifies the entire insulator string through the insulator identification model, including:

[0011] For the visible light image of insulators, the One-stage target detection algorithm and the Grabcut algorithm are combined to adaptively segment the insulator strings, highlight the features of the insulator strings, obtain the feature image, and calculate it using the depthwise separable convolution algorithm.

[0012] Preferably, the depthwise separable convolution algorithm comprises:

[0013]

[0014] SepConv(W n ,W d ,y) (i,j) =PointwiseConν (i,j) (W n ,DepthwiseConν (i,j) (W d ,y))

[0015] Where DepthwiseConv represents the depthwise convolution operation; W represents the weight of the convolution kernel (or filter); y represents the input feature map; (i, j) represents the coordinates of the output feature map; represents the sum of all weights of the convolution kernel, where K and L are the sizes of the convolution kernel in height and width respectively; W k,l represents the weight of the convolution kernel at position (k, l); y (i+1,j+1) Represents the value of the input feature map at position (i+1,j+1);

[0016] PointwiseConv represents the point-wise convolution operation; W m Represents the weight of point-by-point convolution; represents the sum of all input channels, where M is the number of channels of the input feature map; y (i,j,m) Represents the value of the input feature map at position (i, j) and channel m;

[0017] SepConv represents a separable convolution operation; W n Represents the weight of point-by-point convolution; W d Represents the weights of the depthwise convolution.

[0018] Preferably, the weights are initialized with He at the beginning of training, and the initial value of each weight is set to be a mean of 0 and a variance of The Gaussian distribution of C is sampled, where C in is the number of neurons in the input layer, and L2 regularization is used to limit the range of weights.

[0019] Preferably, the UAV identifying the entire insulator string through the insulator identification model further comprises:

[0020] Introducing holes in the convolution kernel, the calculation of the hole convolution receptive field is as follows:

[0021] N=n+(n-1)(r-1)

[0022] Among them, N is the receptive field range, n is the size of the original convolution kernel, and r is the void rate, which is 2;

[0023] In the hole convolution layer, the input feature map A im The dilated convolution operation is:

[0024] [i] = ∑ k A[i+rk]*I[k]

[0025] Where [i] represents the value of the output feature map at position i; ∑ k A[i+rk] represents the sum of the elements of all convolution kernels; the value of the input feature map at position [i+rk], where r is the void rate and k is the index of the convolution kernel element; I[k] represents the weight of the convolution kernel at position k.

[0026] Preferably, the drone moves to the upper end of the insulator to perform terminal identification, and the identification of the insulators on the insulator string one by one includes:

[0027] Using image processing technology, the distance between the center lines of the insulator string is measured, the relative positions between adjacent insulator pieces are calculated, the positions of the insulator pieces are identified and marked, and the string dropping operation is completed; the edge lines and position information of the insulator string are extracted in real time, and the center line of the insulator string is fitted, the distance between adjacent insulator pieces is calculated, and the position of each insulator is marked. The drone slowly moves along the fitted center line to the target insulator, and the robot is released to perform insulation resistance detection.

[0028] Preferably, the optical auxiliary structure comprises:

[0029] Visual markers are added to the edge of the robot to achieve automatic control of lifting and delivery, where the markers provide visual recognition features for the visual guidance system.

[0030] The present invention also provides a vision-guided drone-dropped insulator inspection robot system, the system being used in the vision-guided drone-dropped insulator inspection robot method, comprising:

[0031] An insulator string identification module is used to identify the entire insulator string and obtain the target position;

[0032] The delivery structure is used to deliver the robot when the target insulator is identified.

[0033] The present invention also provides a terminal, including a processor and a storage medium;

[0034] The storage medium is used to store instructions;

[0035] The processor is configured to operate according to the instructions to execute the steps according to the aforementioned method.

[0036] The present invention also provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the steps of the above method are implemented.

[0037] The beneficial effects of the present invention are that, compared with the prior art, the design of the automatic online and offline and robot deployment strategies of the drone of the present invention provides multiple beneficial effects, which improve the interaction efficiency and safety between the drone and the robot. The following are the main beneficial effects of the technical solution:

[0038] Improve docking accuracy: Using advanced vision algorithms and sophisticated sensor systems, the drone can accurately identify and align with passive side markers and emitters to achieve high-precision docking. This accuracy is especially important when operating in complex or restricted environments.

[0039] Automatic on-line and off-line: The robot on-line process is automated, reducing the difficulty of manual operation and improving efficiency. At the same time, the locking device design allows the drone to quickly grab or release the passive end, greatly reducing the time of docking and separation, making the task execution more efficient.

[0040] Strong adaptability: The use of image augmentation technology and one-stage target detection routes, such as YOLO and SSD algorithms, improves the detection speed, is suitable for real-time application scenarios, and ensures the accuracy and security of string detection.

[0041] Space saving and weight optimization: The lightweight base design makes the entire device lighter and easier to carry on drones, while also saving carrier space and optimizing energy efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Schematic diagram of the depthwise separable convolution process in the present invention;

[0043] Figure 2 It is a schematic diagram of the identification and delivery process in the present invention;

[0044] Figure 3It is a schematic diagram of the delivery structure in the present invention;

[0045] Figure 4 This is a schematic diagram of the terminal identification and delivery structure in the present invention;

[0046] Figure 5 This is a schematic diagram of alignment of marking points during delivery in the present invention;

[0047] Figure 6 This is a schematic diagram of the delivery and grabbing structure in the present invention. Figure 6-1 This is a schematic diagram of the grabbing / delivering device. Figure 6-2 To show the intention at the time of delivery, Figure 6-3 Schematic diagram of the hook during grabbing and lifting. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0049] Before the operation robot goes online, it is necessary to use a drone to assist in placing the rope bracket. However, the traditional manual operation method, that is, to observe and control the release of the insulator string through camera observation, is difficult and inefficient. To this end, through visual processing, a full convolutional network is used for image segmentation and semantic segmentation to achieve more refined binarization processing, thereby extracting the image features of the insulator string. Through adaptive threshold binarization, combined with deep learning (Deep-Learn), the accuracy and robustness of straight line fitting are improved, and the noise and outliers in the data are better handled, so as to calculate the offset to control the drone and realize vision-based automatic centering control. Finally, through the analysis of the electric field, the finite element analysis (FEA) combined with the neural network predicts the inner point, and the safe distance for the drone to release the rack is calculated. Finally, the robot is placed on the insulator string through the designed docking standard platform, and the drone carries the robot to the online and offline operation.

[0050] Through highly automated control systems and advanced visual algorithms, the drone can efficiently and quickly deploy the insulator string inspection robot.

[0051] Embodiment 1 of the present invention provides a method for deploying an insulator inspection robot using a drone based on vision guidance, comprising:

[0052] Step 1: The drone identifies the entire insulator string through the insulator identification model to obtain the target position;

[0053] For the visible light image of insulators, a two-stage insulator placement detection robot method is proposed by combining the One-stage target detection algorithm and the image segmentation Grabcut algorithm. First, the drone identifies the target insulator string through the insulator recognition model, determines the ROI with the recognition frame coordinates and predicts the foreground and background, and adaptively segments the insulator string based on the image segmentation algorithm, so that the insulator string appears as a clear part in the image, highlighting the characteristics of the insulator string, which is convenient for the subsequent straight line fitting algorithm to process and prepare for the subsequent terminal placement robot.

[0054] One-stage object detection is a method that directly predicts the category and location of an object in a single network pass. The category and location of the object are predicted simultaneously through a single network pass. The input image is feature extracted through a convolutional neural network (CNN); then, the category probability and bounding box coordinates of the object are predicted at each position on the feature map; then, the anchor box mechanism is used to set candidate boxes of multiple scales and aspect ratios in advance, and the offset and classification information of each candidate box are predicted; finally, non-maximum suppression (NMS) is used to remove duplicate boxes and retain the best detection results. This method simplifies the object detection process, has a high detection speed, and is suitable for real-time applications.

[0055] The numerous and various hardware near the insulator strings of transmission towers increase the workload of the recognition system and reduce the accuracy and speed of recognition. In order to increase the working speed, compared with conventional convolution, the use of depthwise separable convolution significantly reduces the number of parameters and improves computational efficiency.

[0056] like Figure 1 As shown in the figure, the depthwise separable convolution consists of two parts: channel-wise convolution and point-wise convolution. Channel-wise convolution performs convolution operations on each channel separately, that is, each channel is processed by a convolution kernel. Since the number of convolution kernels is the same as the number of channels of the previous level feature map, the number of feature map channels remains unchanged after the first convolution. At the same time, since each channel is operated independently, the same spatial information on different channels cannot be effectively utilized.

[0057] To this end, pointwise convolution is introduced to perform weighted combination of feature information of each channel to generate a new feature map to better extract multi-channel information. The main formula of depth-separable convolution is as follows:

[0058]

[0059] SepConv(W n ,W d ,y) (i,j) =PointwiseConν (i,j) (Wn ,DepthwiseConν (i,j) (W d ,y))(3)

[0060] DepthwiseConv represents the depth convolution operation; W represents the weight of the convolution kernel (or filter); y is the input feature map; (i, j) is the coordinate of the output feature map; represents the sum of all weights of the convolution kernel, where K and L are the sizes of the convolution kernel in height and width respectively; W k,l represents the weight of the convolution kernel at position (k, l); y (i+1,j+1) Represents the value of the input feature map at position (i+1,j+1).

[0061] PointwiseConv represents the point-wise convolution operation; W m Represents the weight of point-by-point convolution; represents the sum of all input channels, where M is the number of channels of the input feature map; y (i,j,m) Represents the value of the input feature map at position (i, j) and channel m.

[0062] SepConv represents a separable convolution operation; W n Represents the weight of point-by-point convolution; W d Represents the weights of the depthwise convolution.

[0063] The above weights are initialized with He at the beginning of training, assuming that the initial value of each weight is a mean of 0 and a variance of The Gaussian distribution of C is sampled, where C in is the number of neurons (or channels) input to this layer. The core idea of ​​this method is to maintain good gradient flow in the early stage of network training, thereby accelerating the convergence of the model and preventing gradient vanishing or gradient exploding problems. At the same time, in order to ensure the stability of the training process, L2 regularization is used to limit the range of weights, which will penalize excessive values ​​of weights and keep the weights in a smaller range.

[0064] The characteristic of adding a dilated convolution layer to the semantic segmentation network is that it introduces holes in the convolution kernel to expand the receptive field, thereby increasing the model's ability to capture contextual information. There will be a certain interval between the pixels of the feature map input by the dilated convolution, forming a loose arrangement. The size of these intervals is called the dilation rate or dilation coefficient.

[0065] The calculation of the receptive field of the dilated convolution is as follows:

[0066] N=n+(n-1)(r-1) (4)

[0067] Where N is the receptive field range, n is the size of the original convolution kernel. r is the dilation rate, which is a key parameter used to control the spacing between elements in the convolution kernel, thereby expanding the receptive field. Common values ​​are 1, 2, 3, and 4. 1 is equivalent to the standard convolution operation, 2 is suitable for scenes that require a moderate expansion of the receptive field, and 3 and 4 are used for more complex tasks. Here, 2 is selected. In the dilation convolution layer, the input feature map A is im The dilated convolution operation is

[0068] [i] = ∑ k A[i+rk]*I[k] (5)

[0069] [i] represents the value of the output feature map at position i; ∑ k A[i+rk] means summing the elements of all convolution kernels; the value of the input feature map at position [i+rk], where r is the dilation rate, k is the index of the convolution kernel element, which refers to the position or serial number of each element in the convolution kernel (or filter) in the convolution operation; I[k] represents the weight of the convolution kernel at position k.

[0070] Where i is the output feature map A out At each position, I is the filter. The size of the feature map after processing is a out for:

[0071]

[0072] where a in is the size of the feature map input, N is the size of the convolution kernel, p is the zero-padding area, and s is the stride. The zero-padding area refers to adding a zero-value area to the edge of the input feature map to ensure that the convolution operation does not lose edge information and control the size of the output feature map; and the stride is the size of the step that the convolution kernel slides on the input feature map. The step value determines the step size of the convolution operation and affects the size and amount of calculation of the output feature map.

[0073] Step 2: The drone moves to the upper end of the insulator to perform terminal identification, identify the insulators on the insulator string one by one, and determine the drop position of the string;

[0074] like Figure 2 As shown, after the mid-proximal process is completed, the distance information of the adjacent insulator pieces is calculated, and the extracted insulator positions are numbered.

[0075] Image processing technology is used to measure the distance between the center lines of the insulator string and calculate the relative position between adjacent insulator pieces. Combined with the image captured by the visual system, the position of the insulator piece is identified and marked. After the string drop operation is completed, the fitting algorithm extracts the edge line and position information of the insulator string in real time and fits the center line of the insulator string. Using this information, ensure alignment with the fitted center line of the insulator string to maintain accuracy and stability during the drop process. Next, the drone calculates the distance between adjacent insulator pieces and marks the position of each insulator. Subsequently, the drone slowly moves along the fitted center line to the target insulator and releases the robot for insulation resistance testing.

[0076] Step 3: After completing the string drop operation, use the optical auxiliary structure and the main structure to release the robot at the target insulator to perform insulation resistance detection.

[0077] like Figure 3-6 As shown in the figure, an optically assisted delivery device is designed. The entire delivery is divided into the delivery main structure and optical auxiliary means. The active side of the delivery is deployed on a multi-rotor drone, and its main structure is described as follows:

[0078] The optical auxiliary part is deployed on the machine to be transported. By adding visual markers to the edge of the robot, automatic control of lifting and delivery is achieved. The markers provide visual recognition features for the visual guidance system. Corner points are important visual features of markers. They are based on local feature points in the image. These feature points are usually located in prominent positions such as corners or edges in the image. The grayscale values ​​near these positions will change dramatically. Corner points are judged by detecting these changes. Corner points have the characteristics of high stability at different scales and rotation angles, so they play an important role in tasks such as image matching, target detection and positioning. Corner point recognition usually uses corner point detection algorithms, the most classic of which is the Harris corner point detector. This detector calculates the grayscale changes in the neighborhood around each pixel in the image and uses the definition of corner points, that is, points with large grayscale changes in multiple directions, to identify corner points.

[0079] The calculation of relative posture requires the identification of feature points on at least three planes. This mechanism adds an indicator light at each of the four corners of the docking mechanism quadrilateral. Each color hue can provide feature points for the visual recognition system, enabling it to obtain redundant plane calculation base points, thereby constructing a virtual docking platform plane and three-dimensional coordinate system.

[0080] In the process of visual guidance, the shape of the virtual quadrilateral formed by the four corners will change due to the different positions of the camera. The shape of the drone viewed from above is obtained through linear transformations such as translation, rotation, and scaling. The transformation matrix is ​​calculated with the help of the camera coordinate system o-xyz and the pixel coordinate system o-uv:

[0081]

[0082] u0 and v0 represent the pixel difference between the origin of the pixel coordinate system and the origin of the image coordinate system on the u axis and v axis, which are generally half of the image width and height pixel size. f is the focal length of the camera, a and b are constants, and the camera coordinate system o-xy is parallel and coplanar with the pixel coordinate system, so xy and uv can be converted to each other to calculate z, and then the relative attitude angle between the camera and the marker point can be obtained, and then how the camera should move to the target position is determined.

[0083] Since the visual features of each marker point are mainly displayed in shape and hue, after setting the thresholds according to the hue of the four colors, four windows are obtained and filtered and bitwise processed, and then integrated to finally calculate the virtual center normal vector for subsequent calculations. The direction of the passive end is preliminarily determined by the virtual center normal vector, and then more complete posture information is calculated by the visual recognition features provided by the four colors.

[0084] Embodiment 2 of the present invention provides a vision-guided drone-based insulator detection robot system, comprising:

[0085] An insulator string identification module is used to identify the entire insulator string and obtain the target position;

[0086] The delivery structure is used to deliver the robot when the target insulator is identified.

[0087] The beneficial effects of the present invention are as follows:

[0088] Improve docking accuracy: Using advanced vision algorithms and sophisticated sensor systems, the drone can accurately identify and align with passive side markers and emitters to achieve high-precision docking. This accuracy is especially important when operating in complex or restricted environments.

[0089] Automatic on-line and off-line: The robot on-line process is automated, reducing the difficulty of manual operation and improving efficiency. At the same time, the locking device design allows the drone to quickly grab or release the passive end, greatly reducing the time of docking and separation, making the task execution more efficient.

[0090] Enhanced operational safety: The design of the buffer pad can effectively absorb the impact force that may occur during the docking process, reduce damage to the drone and carrier, reduce the failure rate, and extend the life of the equipment.

[0091] Strong adaptability: The use of image augmentation technology and one-stage target detection routes, such as YOLO and SSD algorithms, improves the detection speed, is suitable for real-time application scenarios, and ensures the accuracy and security of string detection.

[0092] Space saving and weight optimization: The lightweight base design on the passive side makes the entire device lighter and easier to carry on drones, while also saving carrier space and optimizing energy efficiency.

[0093] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0094] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0095] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0096] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for deploying an insulator inspection robot using a UAV based on vision guidance, comprising: The UAV uses the insulator recognition model to identify the entire insulator string and obtain the target position; The drone moves to the upper end of the insulator and performs terminal identification, identifying the insulators on the insulator string one by one to determine the location of the string drop; After completing the string drop operation, the robot is released at the target insulator using the optical auxiliary structure and the main structure to perform insulation resistance testing.

2. The method for using a UAV to drop an insulator inspection robot based on vision guidance according to claim 1 is characterized in that: The UAV identifies the entire insulator string through the insulator identification model, including: For the visible light image of insulators, the One-stage target detection algorithm and the Grabcut algorithm are combined to adaptively segment the insulator strings, highlight the features of the insulator strings, obtain the feature image, and calculate it using the depthwise separable convolution algorithm.

3. The method for using a UAV to drop an insulator inspection robot based on vision guidance according to claim 2 is characterized in that: The depth-wise separable convolution algorithm includes: SepConv(In n ,IN d ,y) (i,j) =PointwiseConν (i,j) (IN n ,DepthwiseCon (i,j) (IN d ,y)) Where DepthwiseConv represents the depthwise convolution operation; W represents the weight of the convolution kernel (or filter); y represents the input feature map; (i, j) represents the coordinates of the output feature map; represents the sum of all weights of the convolution kernel, where K and L are the sizes of the convolution kernel in height and width respectively; W k,l represents the weight of the convolution kernel at position (k, l); y (i+1,j+1) Represents the value of the input feature map at position (i+1,j+1); PointwiseConv represents the point-wise convolution operation; W m Represents the weight of point-by-point convolution; represents the sum of all input channels, where M is the number of channels of the input feature map; y (i,j,m) Represents the value of the input feature map at position (i, j) and channel m; SepConv represents a separable convolution operation; W n Represents the weight of point-by-point convolution; W d Represents the weights of the depthwise convolution.

4. The method for using a UAV to drop an insulator inspection robot based on vision guidance according to claim 3 is characterized in that: The weights are initialized with He at the beginning of training, and the initial value of each weight is set to be a mean of 0 and a variance of The Gaussian distribution of C is sampled, where C in is the number of neurons in the input layer, and L2 regularization is used to limit the range of weights.

5. The method for using a UAV to drop an insulator inspection robot based on vision guidance according to claim 4 is characterized in that: The UAV identifying the entire insulator string through the insulator identification model also includes: Introducing holes in the convolution kernel, the calculation of the hole convolution receptive field is as follows: N=n+(n-1)(r-1) Among them, N is the receptive field range, n is the size of the original convolution kernel, and r is the void rate, which is 2; In the hole convolution layer, the input feature map A im The dilated convolution operation is: [i]=∑ k A[i+rk]*I[k] Where [i] represents the value of the output feature map at position i; ∑ k A[i+rk] represents the sum of the elements of all convolution kernels; the value of the input feature map at position [i+rk], where r is the void rate and k is the index of the convolution kernel element; I[k] represents the weight of the convolution kernel at position k.

6. The method for using a UAV to drop an insulator inspection robot based on vision guidance according to claim 5 is characterized in that: The drone moves to the upper end of the insulator to perform terminal identification, and identifies the insulators on the insulator string one by one, including: Using image processing technology, the distance between the center lines of the insulator string is measured, the relative positions between adjacent insulator pieces are calculated, the positions of the insulator pieces are identified and marked, and the string dropping operation is completed; the edge lines and position information of the insulator string are extracted in real time, and the center line of the insulator string is fitted, the distance between adjacent insulator pieces is calculated, and the position of each insulator is marked. The drone slowly moves along the fitted center line to the target insulator, and the robot is released to perform insulation resistance detection.

7. The method for using a UAV to drop an insulator inspection robot based on vision guidance according to claim 6 is characterized in that: The optical auxiliary structure comprises: Visual markers are added to the edge of the robot to achieve automatic control of lifting and delivery, where the markers provide visual recognition features for the visual guidance system.

8. A vision-guided drone-dropped insulator inspection robot system, the system being a system used in a vision-guided drone-dropped insulator inspection robot method according to any one of claims 1 to 7, comprising: An insulator string identification module is used to identify the entire insulator string and obtain the target position; The delivery structure is used to deliver the robot when the target insulator is identified.

9. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images