Method and system for detecting multi-scale target defects in power transmission line

Abstract

The invention relates to the technical field of power transmission line defect detection, in particular to a method and system for detecting multi-scale target defects in a power transmission line. According to the technical scheme, temperature information in a tunnel is obtained in real time, and the temperature difference in a preset time interval is calculated based on the temperature information; dividing the detection targets into I-class targets and II-class targets according to the sizes of the detection targets; training a defect detection model; the defect detection model comprises a classification module, a class I target defect detection model and a class II target defect detection model; the inspection unmanned aerial vehicle carries a camera to move along a pre-planned route, and aerial photography is carried out to obtain a to-be-inspected target image; a three-dimensional laser radar, an edge calculation module, a GNSS / IMU board card and an antenna connected with the GNSS / IMU board card are installed on the inspection unmanned aerial vehicle. By adopting the scheme provided by the invention, the training efficiency of the defect detection model, the defect detection efficiency and precision and the data acquisition efficiency and precision can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of transmission line defect detection, and in particular to a method and system for detecting multi-scale target defects in a transmission line. Background Art

[0002] In recent years, with the development of deep learning technology and the widespread application of drone equipment, the traditional manual inspection method of power transmission lines has gradually been replaced by the automatic inspection method of drones. The inspection method based on drone image recognition technology has the advantages of safety, efficiency and stability. In the target detection process for power transmission line components, there are problems such as the fusion of components with complex environmental backgrounds, and large differences in component scale and morphology, which makes it difficult for the trained defect detection model to achieve the expected accuracy, bringing great challenges to the target detection task. In addition, the existing drone inspection flight mainly relies on remote control by staff and is highly dependent on digital mobile networks. It is impossible to achieve high-precision autonomous positioning and obstacle avoidance flight in a weak network environment, which greatly affects the efficiency and accuracy of data collection. Summary of the invention

[0003] In order to overcome the above problems or partially solve the above problems, the purpose of the present application is to provide a method and system for detecting multi-scale target defects in transmission lines, so as to improve the detection accuracy and detection efficiency of multi-scale target defects in transmission lines.

[0004] This application is implemented through the following technical solutions:

[0005] In a first aspect, the present embodiment provides a method for detecting multi-scale target defects in a transmission line, the detection method comprising: dividing the detection target into Class I targets and Class II targets according to the size of the detection target; training a defect detection model; the defect detection model comprises a classification module, a Class I target defect detection model and a Class II target defect detection model; using a patrol drone equipped with a camera to move along a pre-planned route, and obtaining an image of the target to be inspected by aerial photography; the patrol drone is also equipped with a three-dimensional laser radar, an edge computing module, a GNSS / IMU board and an antenna connected to the GNSS / IMU board, the patrol drone has a built-in main control module, and the camera, the three-dimensional laser radar, the edge computing module and the GNSS / IMU board are respectively connected to the main control module; using a classification module to classify the image of the target to be inspected, and sending the image of the target to be inspected to the Class I target defect detection model and / or the Class II target defect detection model according to the classification result; the Class I target defect detection model determines whether the Class I target in the image of the target to be inspected has defects; the Class II target defect detection model determines whether the Class II target in the image of the target to be inspected has defects.

[0006] Based on the first aspect, in an embodiment of the present application, the Class I targets include: equalizing shielding rings and insulators, and the Class I target defect detection model is composed of a backbone network, a detail enhanced pyramid network, a spatial cross-domain feature enhancement network and a detection module.

[0007] Based on the first aspect, in an embodiment of the present application, the detail enhancement pyramid network adopts a feature fusion method of sub-pixel convolution, and the calculation formula of the sub-pixel convolution is: I H =PS(I L )I H =PS(I L )(1); In formula (1), I H Represents a high-resolution feature map, I L represents a low-resolution feature map, PS represents a periodic pixel reshoot algorithm, and its expression is: In formula (2), T represents the input feature, represents rounding down, mod(·) represents the remainder, x and y represent the coordinate points on the high-resolution image, r represents the upsampling multiple, and C represents the number of output channels.

[0008] Based on the first aspect, in an embodiment of the present application, the spatial cross-domain feature enhancement network includes a position information enhancement layer; the detection method includes: using the position information enhancement layer to detect the feature G at position i in the feature map G i After enhancement, the features are obtained The specific calculation process is as follows: Step 1: Set the input parameter q i , k j and v j ; Step 2: Calculate q i and k j The similarity of ij :s ij =F sim (q i ,k j ); Step 3: Normalize the similarity and get the weight value w ij :w ij =F norm (s ij ); Step 4: Enhanced features It can be expressed as: Among them, the feature map G is the output of the detail enhancement pyramid network; among them, q i =f q (G i )∈Q, represents the i-th query vector in the input query vector set Q; k j =f k (G j)∈K, represents the jth key vector in the input key vector set K; v j =f v (G j )∈V, represents the jth value vector in the input value vector set V; f q (·),f k (·) and f v (·) represent the conversion functions of query vector, key vector and value vector respectively; G j is the feature at position j in the feature map G, F sim is a function that uses the inner product to calculate the similarity, F norm is the normalization function, F mul is the weighted summation function.

[0009] Based on the first aspect, in an embodiment of the present application, the Class II target includes bolts, and the Class II target defect detection model is composed of a Swin-Transform network, a feature pyramid network, and a multi-representation fusion auxiliary detection module.

[0010] Based on the first aspect, in the embodiment of the present application, the feature fusion calculation formula of the multi-representation fusion auxiliary detection module is: In formula (3), and g i q They represent the input, output and geometric vector of query instance i respectively; and g j k denote the input and geometric vector of key instance j respectively; T v (·) represents the value obtained after linear transformation; S(·) represents the similarity function between point i and point j, and its function expression is: In formula (4), represents the appearance similarity between query instance i and key instance j, A geometric term representing the relative position of query instance i and key instance j.

[0011] Based on the first aspect, in an embodiment of the present application, the inspection drone equipped with a camera moves along a pre-planned route, including: controlling the inspection drone to move along the transmission line, and obtaining the straight-line distance between the inspection drone body and the transmission line in real time during the movement; when the straight-line distance is less than a preset safety distance, controlling the inspection drone to move away from the transmission line.

[0012] Based on the first aspect, in an embodiment of the present application, the inspection drone has a built-in classifier constructed based on a point cloud data classification algorithm; the detection method also includes: using a three-dimensional laser radar to scan the route environment near the transmission line in real time to obtain point cloud data; using a classifier to identify the point cloud data to determine whether there are detection targets or obstacles on and near the route; if there is a detection target, then taking aerial photos close to the detection target while maintaining a safe distance to obtain an image of the detection target; if there is an obstacle, then calculating the positional relationship and distance between the inspection drone and the obstacle and transmission line, and planning a safe obstacle avoidance route based on the positional relationship and distance.

[0013] In the second aspect, the present application provides a detection system for multi-scale target defects in transmission lines, and the detection system includes: a model training unit, used to train a defect detection model; a defect detection model, used to determine whether a target to be inspected has defects based on the target image to be inspected; an inspection drone, wherein the inspection drone is equipped with a camera and can move along a pre-planned route for aerial photography to obtain the image of the target to be inspected; the inspection drone is also equipped with a three-dimensional laser radar; the defect detection model includes a classification module, a Class I target defect detection model and a Class II target defect detection model; the classification module is used to classify the target image to be inspected, and send the target image to the Class I target defect detection model and / or the Class II target defect detection model according to the classification result; the Class I target defect detection model is used to determine whether a Class I target in the target image to be inspected has defects; the Class II target defect detection model is used to determine whether a Class II target in the target image to be inspected has defects; wherein the sizes of Class I targets and Class II targets are different.

[0014] Based on the second aspect, in an embodiment of the present application, the Class I targets include: equalizing shielding rings and insulators, and the Class I target defect detection model is composed of a backbone network, a detail enhanced pyramid network, a spatial cross-domain feature enhancement network and a detection module; the Class II targets include bolts, and the Class II target defect detection model is composed of a Swin-Transform network, a feature pyramid network and a multi-representation fusion auxiliary detection module.

[0015] Compared with the prior art, this application has at least the following advantages and beneficial effects:

[0016] The solution provided in this application divides the detection targets into Class I targets and Class II targets according to their sizes, distinguishes the detection targets with large size differences, and trains Class I target defect detection models and Class II target defect detection models respectively to identify the corresponding types of detection targets, so as to achieve the effect of "treating the symptoms with the right medicine", simplify the model structure, and improve the training efficiency, recognition efficiency and recognition accuracy of the model. In addition, the inspection drone used to collect images of the target to be inspected, based on its own 3D laser radar and GNSS / IMU modules, combined with the corresponding data processing and positioning algorithms, can enable the drone to achieve high-precision autonomous positioning and obstacle avoidance flight in a weak network environment, ensuring the safety of the drone itself and improving the efficiency and accuracy of data collection. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present application and should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0018] Figure 1 A schematic flowchart of a method for detecting multi-scale target defects in a transmission line is provided;

[0019] Figure 2 The schematic block diagram of the structure of the detection system for multi-scale target defects in transmission lines.

[0020] icon:

[0021] 1-Model training unit; 2-Defect detection model; 21-Classification module; 22-Class I target defect detection model; 23-Class II target defect detection model; 3-Inspection drone. DETAILED DESCRIPTION

[0022] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with examples and drawings. The illustrative implementation scheme of the present application and its description are only used to explain the present application and are not intended to limit the present application.

[0023] Example 1

[0024] This embodiment provides a method for detecting multi-scale target defects in a transmission line, such as Figure 1 As shown, the detection method comprises:

[0025] S1. According to the size of the detection target, the detection target is divided into Class I target and Class II target;

[0026] Specifically, the detection targets include bolts, voltage-equalizing rings, shock-absorbing hammers, yoke plates, parallel hanging plates, extension rings, adjustment plates, traction plates, right-angle hanging plates, voltage-equalizing shielding rings, U-shaped rings, insulators, etc. There is a huge size gap between various types of components. If a single recognition model is used, in order to ensure the recognition accuracy of the model, it is necessary to continuously optimize the structure of the model and improve the complexity of the model. In addition, the demand for training samples is huge and the initial workload is huge, which will seriously affect the training progress of the model.

[0027] For example, in this embodiment, voltage-equalizing shields, yoke plates, insulators, etc. that are similar in size and larger are classified into one category (Class I targets), and other small-sized hardware such as bolts, extension rings, U-shaped rings, right-angle hanging plates, parallel hanging plates, adjustment plates and traction plates that are similar in size and smaller are classified into Class II targets (this embodiment is for reference only, and the number of types is not limited to two).

[0028] S2, training defect detection model 2; the defect detection model includes a classification module 21, a class I target defect detection model 22 and a class II target defect detection model 23;

[0029] Specifically, the training of the defect detection model 2 mainly includes three parts: the training of the classification module (model) and the training of the class I target defect detection model 22 and the class II target defect detection model 23. The classification model 21 can be trained first, and then the classification model classifies and screens the original image samples, and the training samples are made according to the screened original image samples. The class I target defect detection model 22 and the class II target defect detection model 23 are trained according to different model training frameworks. In this embodiment, the model structure of the class I target defect detection model 22 and the class II target defect detection model 23 is described in detail. The classification module 21 can adopt existing convolutional neural network (CNN) or support vector machine (SVM) and other models, which will not be described in detail here.

[0030] Due to the limitation of the safety distance of drone inspection and the influence of the collection method, even for the same type of component targets, the scale difference shown in the image samples may be very large, which makes it very easy to miss the inspection during the inspection. Therefore, the Class I target defect detection model 22 constructed in this embodiment is improved based on the FCOS network, and is mainly composed of a backbone network, a detail enhancement pyramid network, a spatial cross-domain feature enhancement network and a detection module. The detail enhancement pyramid network and the spatial cross-domain feature enhancement network can achieve the effect of detail information enhancement and spatial cross-domain feature enhancement, so as to improve the detection and recognition accuracy of component targets of various morphological and scale characteristics.

[0031] Specifically, the feature fusion method of sub-pixel convolution is adopted in the detail enhancement pyramid network, and the calculation formula of the sub-pixel convolution is:

[0032] I H=PS(I L )I H =PS(I L ) (1)

[0033] In formula (1), I H Represents a high-resolution feature map, I L represents a low-resolution feature map, PS represents a periodic pixel reshoot algorithm, and its expression is:

[0034] In formula (2), T represents the input feature, represents rounding down, mod(·) represents the remainder, x and y represent the coordinate points on the high-resolution image, r represents the upsampling multiple, and C represents the number of output channels.

[0035] Among them, sub-pixel convolution is an upsampling technology used in convolutional neural networks. Its core is to rearrange pixels, increase the number of pixels by interpolating pixels at the sub-pixel level, and combine the feature maps of r×r channels with a size of w×h into a new feature map of size [w×r,h×r], thereby obtaining high-resolution detail information.

[0036] Specifically, the spatial cross-domain feature enhancement network includes a position information enhancement layer; the detection method includes: using the position information enhancement layer to detect the feature G at position i in the feature map G i After enhancement, the features are obtained The specific calculation process is as follows:

[0037] Set the input parameter q i , k j and v j ;

[0038] Calculate q i and k j The similarity of ij :s ij =F sim (q i ,k j );

[0039] Normalize the similarity and get the weight value w ij :w ij =F norm (s ij );

[0040] Enhanced features It can be expressed as:

[0041] Among them, the feature map G is the output of the detail enhanced pyramid network;

[0042] Among them, q i =f q (G i )∈Q, represents the i-th query vector in the input query vector set Q; k j =f k (G j )∈K, represents the jth key vector in the input key vector set K; v j =f v (G j )∈V, represents the jth value vector in the input value vector set V; f q (·),f k (·) and f v (·) represent the conversion functions of query vector, key vector and value vector respectively; G j is the feature at position j in the feature map G, F sim is a function that uses the inner product to calculate the similarity, F norm is the normalization function, F mul is the weighted summation function.

[0043] Since there is a lack of information interaction between objects of different scales in the feature fusion of the pyramid network, in this embodiment, the spatial position relationship between multi-scale objects can be mined based on the spatial cross-domain feature enhancement network, thereby realizing feature enhancement of multi-scale objects in the feature fusion stage, and further improving the detection accuracy of multi-scale objects.

[0044] Since the physical size of Class II targets is relatively small, the size of the target image collected by the drone while ensuring a safe shooting distance accounts for a very small proportion of the entire original image. In addition, due to the limitations of the installation position of the target (such as bolts) and the influence of the acquisition and shooting angle, its spatial form is diverse. In addition, affected by the lighting environment and environmental background, the pins are very easy to merge with other components and the environmental background, making them difficult to detect and identify. Therefore, the Class II target defect detection model 23 of this embodiment includes a Swin-Transform network, a feature pyramid network, and a multi-representation fusion auxiliary detection module. Among them, Swin-Transformer is used as the backbone network for feature extraction, and the feature pyramid network performs feature fusion based on the output of the Swin-Transformer network. In the detection stage, this embodiment adds a multi-representation fusion module for enhancing classification and regression features on the basis of the original three branches of classification, regression and centrality, so as to fuse the advantages of different representation methods. The feature fusion calculation formula is:

[0045]

[0046] In formula (3), and gi q They represent the input, output and geometric vector of query instance i respectively; and g j k denote the input and geometric vector of key instance j respectively; T v (·) represents the value obtained after linear transformation; S(·) represents the similarity function between point i and point j, and its function expression is:

[0047]

[0048] In formula (4), represents the appearance similarity between query instance i and key instance j, S G (g i q ,g j k ) is a geometric term representing the relative position of query instance i and key instance j.

[0049] In this embodiment, the advantages of different representation methods are integrated to enhance the features of the classification and regression detection parts of the original network. The features of the heterogeneous representation (such as the center point representation) are used as the key geometric vectors and the similarity is calculated with the query geometric vector (corresponding to the bounding box representation) to assist in enhancing the features of the main representation, thereby obtaining the classification and regression features assisted by the heterogeneous representation, which can fully utilize the advantages of the center point representation for small-size target detection.

[0050] S3, using the inspection drone 3 equipped with a camera to move along the pre-planned route, and obtaining the image of the target to be inspected by aerial photography; the inspection drone 3 is also equipped with a three-dimensional laser radar, an edge computing module, a GNSS / IMU board, and an antenna connected to the GNSS / IMU board. The inspection drone 3 has a built-in main control module, and the camera, three-dimensional laser radar, edge computing module and GNSS / IMU board are respectively connected to the main control module;

[0051] Specifically, in order to realize autonomous positioning of the UAV, this embodiment introduces a three-dimensional point cloud classifier (which can be obtained through model training) to classify discrete point cloud data into point cloud sets of different objects. Combined with three-dimensional laser radar and (visible light) camera, it can realize the recognition of targets such as transmission line towers, wires, buildings and vegetation, and the real-time construction of three-dimensional scenes. In addition, by adopting a positioning algorithm based on INS-GNSS, it can provide high-precision high-frequency position and attitude information for the inspection UAV 3, realize real-time autonomous positioning, and ensure the collection accuracy of image data.

[0052] The inspection drone 3 is equipped with a camera and moves along a pre-planned route, including: controlling the inspection drone 3 to move along the power line, and obtaining the straight-line distance between the inspection drone 3 and the power line in real time during the movement; when the straight-line distance is less than a preset safety distance, controlling the inspection drone 3 to stay away from the power line.

[0053] Furthermore, the edge computing module has a built-in classifier based on the point cloud data classification algorithm; the detection method also includes: using a three-dimensional laser radar to scan the route environment near the transmission line in real time to obtain point cloud data; using a classifier to identify point cloud data to determine whether there are detection targets or obstacles on and near the route; if there is a detection target, then approach the detection target to obtain an image of the detection target while maintaining a safe distance; if there is an obstacle, calculate the position relationship and distance between the inspection drone 3 and the obstacle and transmission line, and plan a safe obstacle avoidance route based on the position relationship and distance. Control the inspection drone 3 to move along the transmission line, and obtain the straight-line distance between the fuselage of the inspection drone 3 and the transmission line in real time during the movement; when the straight-line distance is less than the preset safety distance, control the inspection drone 3 to stay away from the transmission line.

[0054] Specifically, firstly, the route environment is scanned by a lidar to obtain point cloud data, and then the point cloud is classified and processed by a classifier to distinguish between detection targets and obstacles, and obtain the position information of each detection target (such as transmission lines, towers, etc.) and obstacles (such as tree obstacles, buildings, etc.), and calculate the three-dimensional spatial position relationship between the inspection UAV 3 and the obstacles and detection targets. Based on the three-dimensional spatial position relationship between obstacles and detection targets, combined with flight control algorithms and real-time feedback data from GNSS, cameras, etc., the inspection UAV 3 can fly autonomously along the (transmission) line to ensure the efficiency of image data collection.

[0055] S4, using the classification module to classify the target image to be inspected, and sending the target image to be inspected to the Class I target defect detection model and / or the Class II target defect detection model according to the classification result;

[0056] S5, judging whether there is a defect in the Class I target in the target image to be inspected by the Class I target defect detection model;

[0057] S6. The Class II target defect detection model is used to determine whether there are defects in the Class II target in the target image to be inspected.

[0058] Example 2

[0059] This embodiment provides a detection system for multi-scale target defects in a transmission line, such as Figure 2As shown, the detection system includes: a model training unit 1, which is used to train a defect detection model; a defect detection model 2, which is used to determine whether the target to be inspected has defects based on the target image to be inspected; an inspection drone 3, which is equipped with a camera and can move along a pre-planned route to obtain the target image to be inspected by aerial photography; the inspection drone is also equipped with a three-dimensional laser radar; the defect detection model includes a classification module 21, a class I target defect detection model 22 and a class II target defect detection model 23; the classification module 21 is used to classify the target image to be inspected, and send the target image to the class I target defect detection model 22 and / or the class II target defect detection model 23 according to the classification result; the class I target defect detection model 22 is used to determine whether the class I target in the target image to be inspected has defects; the class II target defect detection model 23 is used to determine whether the class II target in the target image to be inspected has defects; wherein the sizes of the class I target and the class II target are different.

[0060] Furthermore, the Class I targets include: equalizing shielding rings and insulators, and the Class I target defect detection model is composed of a backbone network, a detail enhanced pyramid network, a spatial cross-domain feature enhancement network and a detection module; the Class II targets include bolts, and the Class II target defect detection model is composed of a Swin-Transform network, a feature pyramid network and a multi-representation fusion auxiliary detection module.

[0061] The system provided in the embodiment of the present application can be used to execute the method in the above-mentioned embodiment 1. For details, please refer to embodiment 1 and will not be repeated here.

[0062] The specific implementation methods described above further illustrate the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only the specific implementation method of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for detecting multi-scale target defects in a transmission line, characterized in that: The detection method comprises: According to the size of the detection target, the detection targets are divided into Class I targets and Class II targets; Training a defect detection model; the defect detection model includes a classification module, a class I target defect detection model, and a class II target defect detection model; The inspection drone is equipped with a camera and moves along a pre-planned route to obtain an image of the target to be inspected by aerial photography; the inspection drone is also equipped with a three-dimensional laser radar, an edge computing module, a GNSS / IMU board, and an antenna connected to the GNSS / IMU board. The inspection drone has a built-in main control module, and the camera, three-dimensional laser radar, edge computing module, and GNSS / IMU board are respectively connected to the main control module; The target image to be inspected is classified by using a classification module, and the target image to be inspected is sent to a Class I target defect detection model and / or a Class II target defect detection model according to the classification result; The Class I target defect detection model is used to determine whether the Class I target in the target image to be inspected has defects; The Class II target defect detection model is used to determine whether there are defects in the Class II target in the target image to be inspected.

2. The method for detecting multi-scale target defects in a power transmission line according to claim 1, characterized in that: The Class I targets include: voltage-equalizing shielding rings and insulators. The Class I target defect detection model consists of a backbone network, a detail-enhanced pyramid network, a spatial cross-domain feature enhancement network and a detection module.

3. The method for detecting multi-scale target defects in a transmission line according to claim 2, characterized in that: The detail enhancement pyramid network adopts the feature fusion method of sub-pixel convolution, and the calculation formula of the sub-pixel convolution is: I H =PS(I L )I H =PS(I L ) (1) In formula (1), I H Represents a high-resolution feature map, I L represents a low-resolution feature map, PS represents a periodic pixel reshoot algorithm, and its expression is: In formula (2), T represents the input feature, represents rounding down, mod(·) represents the remainder, x and y represent the coordinate points on the high-resolution image, r represents the upsampling multiple, and C represents the number of output channels.

4. The method for detecting multi-scale target defects in a transmission line according to claim 2, characterized in that: The spatial cross-domain feature enhancement network includes a position information enhancement layer; the detection method includes: using the position information enhancement layer to detect the feature G at position i in the feature map G i After enhancement, the features are obtained The specific calculation process is as follows: Set the input parameter q i , k j and v j ; Calculate q i and k j The similarity of ij :s ij =F sim (q i ,k j ); Normalize the similarity and get the weight value w ij :w ij =F norm (s ij ); Enhanced features It can be expressed as: Among them, the feature map G is the output of the detail enhanced pyramid network; Among them, q i =f q (G i )∈Q, represents the i-th query vector in the input query vector set Q; k j =f k (G j )∈K, represents the jth key vector in the input key vector set K; v j =f v (G j )∈V, represents the jth value vector in the input value vector set V; f q (·),f k (·) and f v (·) represent the conversion functions of query vector, key vector and value vector respectively; G j is the feature at position j in the feature map G, F sim is a function that uses the inner product to calculate the similarity, F norm is the normalization function, F mul is the weighted summation function.

5. The method for detecting multi-scale target defects in a transmission line according to claim 1, characterized in that: The Class II targets include bolts, and the Class II target defect detection model is composed of a Swin-Transform network, a feature pyramid network and a multi-representation fusion auxiliary detection module.

6. The method for detecting multi-scale target defects in a power transmission line according to claim 5, characterized in that: The feature fusion calculation formula of the multi-representation fusion auxiliary detection module is: In formula (3), f i q 、f i ' q and g i q They represent the input, output and geometric vector of query instance i respectively; f j k and g j k denote the input and geometric vector of key instance j respectively; T v (·) represents the value obtained after linear transformation; S(·) represents the similarity function between point i and point j, and its function expression is: S(f i q ,f j k ,g i q ,g j k )=softmax j [S A (f i q ,f j k )+S G (g i q ,g j k )] (4) In formula (4), S A (f i q , f j k ) represents the appearance similarity between query instance i and key instance j, S G (g i q , g j k ) is a geometric term representing the relative position of query instance i and key instance j.

7. The method for detecting multi-scale target defects in a power transmission line according to claim 1, characterized in that: The inspection drone carries a camera and moves along a pre-planned route, including: Control the inspection drone to move along the transmission line, and obtain the straight-line distance between the inspection drone and the transmission line in real time during the movement; When the straight-line distance is less than the preset safety distance, the inspection drone is controlled to stay away from the power transmission line.

8. The method for detecting multi-scale target defects in a power transmission line according to claim 7, characterized in that: The edge computing module has a built-in classifier based on a point cloud data classification algorithm; The detection method further comprises: Use 3D laser radar to scan the route environment near the transmission line in real time to obtain point cloud data; Use the classifier to identify point cloud data and determine whether there are detection targets or obstacles on or near the route; If there is a detection target, take aerial photography close to the detection target while maintaining a safe distance to obtain the detection target image; If there are obstacles, the position relationship and distance between the inspection drone and the obstacles and power lines are calculated, and a safe obstacle avoidance route is planned based on the position relationship and distance.

9. A multi-scale target defect detection system in a transmission line, characterized in that: The detection system includes: A model training unit, used to train a defect detection model; A defect detection model is used to determine whether the object to be inspected has defects based on the image of the object to be inspected; Inspection drones are equipped with cameras and can move along pre-planned routes to obtain images of targets to be inspected through aerial photography. Three-dimensional laser radars are also installed on the inspection drones. The defect detection model includes a classification module, a Class I target defect detection model and a Class II target defect detection model; the classification module is used to classify the target image to be inspected, and send the target image to the Class I target defect detection model and / or the Class II target defect detection model according to the classification result; the Class I target defect detection model is used to determine whether there is a defect in the Class I target in the target image to be inspected; the Class II target defect detection model is used to determine whether there is a defect in the Class II target in the target image to be inspected; wherein, the sizes of the Class I target and the Class II target are different.

10. The multi-scale target defect detection system in the power transmission line according to claim 9, characterized in that: The Class I targets include: equalizing shielding rings and insulators. The Class I target defect detection model consists of a backbone network, a detail enhanced pyramid network, a spatial cross-domain feature enhancement network and a detection module. The Class II targets include bolts. The Class II target defect detection model consists of a Swin-Transform network, a feature pyramid network and a multi-representation fusion auxiliary detection module.

Images