Method and device for detecting action state of disconnecting switch

Abstract

The invention discloses a method and a device for detecting the action state of an isolating switch. The method comprises the following steps: acquiring a point cloud data set; constructing an initial detection model, and training the initial detection model by adopting the target loss function to obtain a target detection model; inputting the test data into the target detection model, obtaining first point cloud data corresponding to a conductive arm of the disconnecting switch corresponding to the test data and second point cloud data corresponding to an insulating pillar of the disconnecting switch corresponding to the test data; third point cloud data corresponding to a power transmission cable of the disconnecting switch corresponding to the test data and fourth point cloud data corresponding to a base of the disconnecting switch corresponding to the test data; and obtaining a vector corresponding to the first point cloud data, and determining an action state of the disconnecting switch at least according to the vector corresponding to the first point cloud data. The technical problems that manual inspection efficiency is low, potential safety hazards exist, and complex models and high computing resources are needed in the image processing technology in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the field of sensor technology, and in particular to a method for detecting an operating state of an isolating switch, a device for detecting an operating state of an isolating switch, and a computer-readable storage medium. Background Art

[0002] LiDAR is a long-distance measurement technology that obtains the position, shape and surface features of target objects by sending laser pulses and measuring their return time and other properties. With the development of LiDAR 3D scanning technology, the acquisition of 3D point cloud models has become more mature, and 3D point cloud models are widely used in unmanned driving, machine vision and medical research. Compared with traditional 2D images, the 3D point cloud data model obtained by LiDAR has the advantages of more object information, real-time performance and less influence by light.

[0003] The isolating switch is in the closed state when it is closed and energized, and in the open state when it is disconnected and not energized. The isolating switch is in the closing process when it changes from the disconnected and not energized state to the closed and energized state, and is in the disconnecting process when it changes from the closed and energized state to the disconnected and not energized state. The conductive arm of the double-column horizontal rotary isolating switch is fixed at the upper end of the insulating support. The operating mechanism drives the two conductive arms to rotate, thereby completing the opening and closing actions.

[0004] When the isolating switch is in the open state, the left and right conductive arms are parallel to each other. When the isolating switch is in the closing process, the operating mechanism drives the conductive arms on both sides to rotate. When the conductive arms on both sides rotate to the same horizontal line, the contacts at the front end of the conductive arms can contact and connect with each other to form a stable conductive path. At this time, the isolating switch is in the closed state. When the isolating switch is in the disconnecting process, the operating mechanism drives the conductive arms on both sides to rotate. When the conductive arms on both sides rotate to parallel positions, the isolating switch to the store path is disconnected. At this time, the isolating switch is in the open state.

[0005] As an important part of the power system, the substation is responsible for transforming voltage and current, and concentrating and distributing electric energy. The normal operation of the power facilities in the substation is related to the safety of the entire power grid. The isolating switch is one of the important power facilities in the substation. Ensuring the normal posture of the isolating switch is a very important part of the power work of the substation. The status detection of the traditional substation is mainly judged by manual inspection. This method has the problems of low efficiency, large workload, and certain safety hazards to the inspectors. With the intelligentization and automation of substation safety measures, the isolating switch status detection technology based on machine vision and image processing has developed rapidly. The judgment of the isolating switch posture by image processing technology ensures the safety of the inspectors and the reliability of the isolating switch status judgment.

[0006] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0007] The embodiments of the present invention provide a method for detecting the action status of an isolating switch, a device for detecting the action status of an isolating switch, and a computer-readable storage medium, so as to at least solve the technical problems in the related technology that manual inspection is inefficient and has potential safety hazards, and that image processing technology requires complex models and high computing resources.

[0008] According to one aspect of an embodiment of the present invention, a method for detecting the action state of an isolating switch is provided, comprising: obtaining a point cloud data set, the point cloud data set comprising a plurality of point cloud data corresponding to the isolating switch, each of the point cloud data comprising a plurality of frames of point cloud images obtained by scanning the isolating switch to be detected by a laser radar, each of the point cloud images comprising a plurality of point cloud data points, part of the point cloud data set being test data and the other part being training data, the isolating switch comprising at least a conductive arm, an insulating support, a transmission cable and a base; constructing an initial detection model, and training the initial detection model using a target loss function to obtain a target detection model, wherein the target detection model comprises a feature extraction module, a classification module and a segmentation module, the input end of the feature extraction module being the input end of the target detection model, the output end of the feature extraction module being respectively connected to the input end of the classification module and the input end of the segmentation module, the output end of the classification module and the output end of the segmentation module being the output end of the target detection model, and the feature extraction module comprising The method comprises a stacked local feature extraction submodule and an attention submodule, wherein the feature extraction module is used to extract point cloud features of the input data of the target detection model, the classification module is used to classify the output data of the feature extraction module, and the segmentation module is used to segment the output data of the feature extraction module; the test data is input into the target detection model to obtain first point cloud data corresponding to the conductive arm of the isolating switch corresponding to the test data, second point cloud data corresponding to the insulating support of the isolating switch corresponding to the test data, third point cloud data corresponding to the transmission cable of the isolating switch corresponding to the test data, and fourth point cloud data corresponding to the base of the isolating switch corresponding to the test data; a vector corresponding to the first point cloud data is obtained, and an action state of the isolating switch is determined at least according to the vector corresponding to the first point cloud data, the action state includes an open state, a closed state, a disconnection process, and a closing process, the open state is the state of the isolating switch when it is not powered on, and the closed state is the state of the isolating switch when it is powered on.

[0009] According to another aspect of an embodiment of the present invention, a detection device for the action state of an isolating switch is also provided, comprising: an acquisition module, used to acquire a point cloud data set, the point cloud data set comprising a plurality of point cloud data corresponding to the isolating switch, each of the point cloud data comprising a plurality of frames of point cloud images obtained by scanning the isolating switch to be detected by a laser radar, each of the point cloud images comprising a plurality of point cloud data points, part of the point cloud data set being test data and the other part being training data, the isolating switch comprising at least a conductive arm, an insulating support, a transmission cable and a base; a first processing module, used to construct an initial detection model, and to train the initial detection model using a target loss function to obtain a target detection model, wherein the target detection model comprises a feature extraction module, a classification module and a segmentation module, the input end of the feature extraction module being the input end of the target detection model, the output end of the feature extraction module being respectively connected to the input end of the classification module and the input end of the segmentation module, the output end of the classification module and the output end of the segmentation module being the output end of the target detection model, It includes a stacked local feature extraction submodule and an attention submodule, wherein the feature extraction module is used to extract point cloud features of the input data of the target detection model, the classification module is used to classify the output data of the feature extraction module, and the segmentation module is used to segment the output data of the feature extraction module; an input module is used to input the test data into the target detection model to obtain first point cloud data corresponding to the conductive arm of the isolating switch corresponding to the test data, second point cloud data corresponding to the insulating support of the isolating switch corresponding to the test data, third point cloud data corresponding to the transmission cable of the isolating switch corresponding to the test data, and fourth point cloud data corresponding to the base of the isolating switch corresponding to the test data; a second processing module is used to obtain a vector corresponding to the first point cloud data, and determine the action state of the isolating switch at least according to the vector corresponding to the first point cloud data, the action state includes an open state, a closed state, a disconnection process and a closing process, the open state is the state of the isolating switch when it is not powered on, and the closed state is the state of the isolating switch when it is powered on.

[0010] Optionally, the acquisition module includes: a scanning processing unit, used to scan and process the isolating switch through a laser radar to obtain a plurality of initial point cloud data; and a filtering and denoising processing unit, used to filter and denoise the plurality of initial point cloud data respectively to obtain the point cloud data set.

[0011] Optionally, the input module includes: a first input unit, used to input the test data into the first local feature extraction submodule to perform a first downsampling and a first feature extraction process on the test data to obtain a plurality of first point cloud data points; a second input unit, used to sequentially input the plurality of first point cloud data points into the first channel attention submodule to perform a first feature dimension weighted process on the plurality of first point cloud data points to obtain a plurality of second point cloud data points; a third input unit, used to input the plurality of second point cloud data points into the second local feature extraction submodule to perform a second downsampling and a second feature extraction process on the plurality of second point cloud data points to obtain a plurality of third point cloud data points; a fourth input unit, used to sequentially input the plurality of third point cloud data points into the second channel attention submodule to perform a first feature dimension weighted process on the plurality of first point cloud data points The plurality of third point cloud data points are weightedly processed by the second feature dimension to obtain a plurality of fourth point cloud data points; a fifth input unit is used to input the plurality of test data, the plurality of the second point cloud data points and the plurality of the fourth point cloud data points into the segmentation module to respectively perform segmentation processing on the plurality of the fourth point cloud data points to obtain a plurality of fifth point cloud data points, wherein the number of the fifth point cloud data points is the same as the number of point cloud data in the test data; a sixth input unit is used to input the plurality of the fourth point cloud data points into the classification module to respectively perform classification processing on the plurality of the fourth point cloud data points to obtain a plurality of sixth point cloud data points, wherein the plurality of the fifth point cloud data points and the plurality of the sixth point cloud data points form the first point cloud data, the second point cloud data, the third point cloud data and the fourth point cloud data.

[0012] Optionally, the first input unit includes: a first input subunit, used to input the test data into the downsampling layer VFS to perform voxel farthest point downsampling processing on the test data to obtain a first sampling point set, wherein the first sampling point set includes a plurality of first sampling points; a second input subunit, used to input the test data and the first sampling point set into the weighted mean grouping layer PAG to perform grouping processing with the first sampling point as the centroid point to obtain a second sampling point set, wherein the second sampling point set is a local feature point set corresponding to each of the first sampling points, and the second sampling point set includes a plurality of second sampling points; a third input subunit, used to input the second sampling point set into the local feature point set input point mesh layer to perform feature extraction processing on the plurality of second sampling points respectively, to obtain a plurality of first point cloud data points; and a fourth input subunit, used to input the plurality of second point cloud data points into the second local feature extraction submodule V The PP performs a second downsampling and a second feature extraction process on the plurality of second point cloud data points to obtain a plurality of third point cloud data points, including: a fifth input subunit, which is used to sequentially input the plurality of second point cloud data points into the downsampling layer to perform a voxel farthest point downsampling process on the second point cloud data points to obtain a third sampling point set, wherein the third sampling point set includes a plurality of third sampling points; a sixth input subunit, which is used to input the plurality of second point cloud data points and the third sampling point set into the weighted mean grouping layer to perform grouping process with the third sampling point as the centroid point to obtain a fourth sampling point set, wherein the fourth sampling point set is a local feature point set corresponding to each of the third sampling points, and the fourth sampling point set includes a plurality of fourth sampling points; and a seventh input subunit, which is used to input the fourth sampling point set into the local feature point set input point mesh layer to perform feature extraction process on the plurality of fourth sampling points respectively to obtain a plurality of third point cloud data points.

[0013] Optionally, the second input unit includes: an eighth input sub-unit, which is used to input the plurality of the first point cloud data points into the activation function layer in sequence to respectively calculate the weight of the feature dimension of each of the first point cloud data points, and obtain a plurality of first weight values; a ninth input sub-unit, which is used to input the plurality of the first weight values ​​and the corresponding first point cloud data points into the calculation layer to respectively perform point multiplication processing on the first weight values ​​and the corresponding first point cloud data points, and obtain a plurality of the second point cloud data points; a tenth input sub-unit, which is used to input the plurality of the third point cloud data points into the second channel attention sub-module in sequence to perform second feature dimension weighted processing on the plurality of the third point cloud data points, and obtain a plurality of fourth point cloud data points, including: inputting the plurality of the third point cloud data points into the activation function layer in sequence to respectively calculate the weight of the feature dimension of each of the third point cloud data points, and obtain a plurality of second weight values; an eleventh input sub-unit, which is used to input the plurality of the second weight values ​​and the corresponding third point cloud data points into the calculation layer to respectively perform point multiplication processing on the second weight values ​​and the corresponding third point cloud data points, and obtain a plurality of the fourth point cloud data points.

[0014] Optionally, the fifth input unit includes: a twelfth input subunit, which is used to input multiple second point cloud data points and multiple fourth point cloud data points into the first interpolation layer, using the second point cloud data points as the points to be interpolated and the fourth point cloud data points as the neighboring feature points to perform a first interpolation process on the second point cloud data points and the fourth point cloud data points to obtain a first data point set; a thirteenth input subunit, which is used to input the first data point set into the first convolution layer to perform feature extraction process on the first data point set to obtain a second data point set, wherein the second data point set includes multiple second data points; a fourteenth input subunit, which is used to input the second data point set and the test data into the second interpolation layer, using the test data as the points to be interpolated and the second data points as the neighboring feature points to perform a second interpolation process on the second data point set and the test data to obtain a third data point set, wherein the third data point set includes multiple third data points; a fifteenth input subunit, which is used to input the third data point set into the second convolution layer to perform feature extraction process on the multiple third data points to obtain multiple fifth point cloud data points.

[0015] Optionally, the sixth input unit includes: a sixteenth input subunit, used to input the multiple fourth point cloud data points into the third convolutional layer in sequence to perform feature extraction processing on the multiple fourth point cloud data points respectively, to obtain multiple seventh point cloud data points; a seventeenth input subunit, used to input the multiple seventh point cloud data points into the fully connected layer to perform classification processing on the seventh point cloud data points, to obtain multiple sixth point cloud data points.

[0016] Optionally, the second processing module includes: a calculation unit for calculating according to the formula Calculate the angle α between the conductive arm corresponding to the point cloud image of the i-th frame and the conductive arm in the closed state i , where u i =((x0,y0),(x i ,y i )), (x0, y0) is the coordinate corresponding to the end of the conductive arm fixed to the stationary insulating support, (x i ,y i ) is the coordinate corresponding to one end of the conductive arm rotating around the insulating support, v = ((x0, y0), (x v ,y v )),(x v ,y v ) is the coordinate corresponding to one end of the conductive arm rotating around the insulating support in the closed state, u i v is the vector u i The dot product of the vector v, |u i | is the vector u i The modulus of |v| is the modulus of vector v; the first determination unit is used to determine α i Is it greater than α i+n , in α i Greater than α i+n In the case of i+n The corresponding disconnector action state is the closing process, where α i+n is the angle between the conductive arm corresponding to the point cloud image of the i+nth frame and the conductive arm in the closed state, n is greater than 10; the second determining unit is used to determine the angle between the conductive arm corresponding to the point cloud image of the i+nth frame and the conductive arm in the closed state in the closed state in the i+nth frame ... i Less than α i+n In the case of i The corresponding isolating switch action state is the disconnection process; the third determining unit is used to determine the i Equal to α i+n In the case of i and α i+n Whether α is satisfied i+n =α i =0, at α i+n =α i = 0, determine α i The corresponding isolating switch has the action state of the closed state; the fourth determining unit is used to determine the i+n =α i ≠0, determine α i The corresponding operating state of the isolating switch is the open state.

[0017] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is further provided, wherein the computer-readable storage medium includes a stored program, wherein when the program is run, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned methods for detecting the operating status of the isolating switch.

[0018] According to another aspect of an embodiment of the present invention, a computer program product is provided, comprising computer instructions, wherein when the computer instructions are executed by a processor, the method for detecting the operating state of an isolating switch described in any one of the above isolating switches is executed.

[0019] According to another aspect of an embodiment of the present invention, a processor is further provided, the processor being used to run a program, wherein the program, when running, executes any one of the above-mentioned methods for detecting the operating state of an isolating switch.

[0020] In an embodiment of the present invention, a point cloud data set is obtained, the point cloud data set includes a plurality of point cloud data corresponding to the disconnector, each point cloud data respectively includes a plurality of frames of point cloud images obtained by scanning the disconnector to be detected by a laser radar, each point cloud image respectively includes a plurality of point cloud data points, part of the point cloud data set is test data, and the other part is training data, the disconnector includes at least a conductive arm, an insulating support, a transmission cable and a base; an initial detection model is constructed, and the initial detection model is trained using a target loss function to obtain a target detection model, wherein the target detection model includes a feature extraction module, a classification module and a segmentation module, the input end of the feature extraction module is the input end of the target detection model, the output end of the feature extraction module is respectively connected to the input end of the classification module and the input end of the segmentation module, the output end of the classification module and the output end of the segmentation module are the output end of the target detection model, and the feature extraction module includes stacked local feature extraction. submodule and attention submodule, the feature extraction module is used to extract point cloud features of the input data of the target detection model, the classification module is used to classify the output data of the feature extraction module, and the segmentation module is used to segment the output data of the feature extraction module; the test data is input into the target detection model to obtain the first point cloud data corresponding to the conductive arm of the isolating switch corresponding to the test data, the second point cloud data corresponding to the insulating support of the isolating switch corresponding to the test data, the third point cloud data corresponding to the transmission cable of the isolating switch corresponding to the test data, and the fourth point cloud data corresponding to the base of the isolating switch corresponding to the test data; the vector corresponding to the first point cloud data is obtained, and the action state of the isolating switch is determined at least according to the vector corresponding to the first point cloud data, the action state includes an open state, a closed state, a disconnection process and a closing process, the open state is the state of the isolating switch when it is not powered on, and the closed state is the state of the isolating switch when it is powered on. Through the technical solution provided by the present invention, the three-dimensional point cloud data of the isolating switch is obtained by using a laser radar, and combined with the voxel farthest point sampling and weighted mean grouping technology, the weight of the feature dimension useful for the current task is improved through the point cloud channel attention mechanism module, thereby achieving the technical effect of real-time automatic monitoring of the operating status of the isolating switch, improving the accuracy and efficiency of monitoring, and further solving the technical problems in related technologies such as low efficiency and safety hazards of manual inspections, and image processing technology requiring complex models and high computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0022] Figure 1 is a flow chart of a method for detecting an operating state of an isolating switch according to an embodiment of the present invention;

[0023] Figure 2 It is a technical roadmap of a method for detecting an operating state of an isolating switch according to an embodiment of the present invention;

[0024] Figure 3 is a structural diagram of a VPP module of a method for detecting an operating state of an isolating switch according to an embodiment of the present invention;

[0025] Figure 4 is a schematic diagram of a PSE module of a method for detecting an operating state of an isolating switch according to an embodiment of the present invention;

[0026] Figure 5 is a schematic diagram of point cloud feature information of a method for detecting an operating state of an isolating switch according to an embodiment of the present invention;

[0027] Figure 6 is a schematic diagram of coordinates of both ends of a conductive arm of an isolating switch according to an embodiment of the present invention;

[0028] Figure 7 is a schematic diagram of a conductive arm of an isolating switch according to an embodiment of the present invention;

[0029] Figure 8 is a schematic diagram of a disconnector action monitoring model structure according to an embodiment of the present invention;

[0030] Fig. 9 2 is a schematic diagram of a device for detecting the operating state of an isolating switch according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only 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 ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0032] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0033] As introduced in the background technology, manual inspection in the related technology is inefficient and has potential safety hazards, and image processing technology requires complex models and high computing resources. In view of the above defects, a method for detecting the action state of an isolating switch, a device for detecting the action state of an isolating switch, and a computer-readable storage medium are provided in an embodiment of the present invention.

[0034] According to an embodiment of the present invention, a method embodiment of a method for detecting the operating status of an isolating switch is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0035] Figure 1 is a flow chart of a method for detecting the operating state of an isolating switch according to an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0036] Step S202, obtaining a point cloud data set, the point cloud data set includes multiple point cloud data corresponding to the disconnector, each point cloud data includes multiple frames of point cloud images obtained by scanning the disconnector to be detected by a laser radar, each point cloud image includes multiple point cloud data points, part of the point cloud data set is test data, and the other part is training data. The disconnector includes at least a conductive arm, an insulating support, a transmission cable and a base.

[0037] In this embodiment, a point cloud data set including multiple point cloud data corresponding to the isolating switch can be obtained, which is convenient for subsequent analysis of the point cloud data, extraction of features of the isolating switch, and automatic identification and positioning of the isolating switch.

[0038] Step S204, construct an initial detection model, and use the target loss function to train the initial detection model to obtain a target detection model, wherein the target detection model includes a feature extraction module, a classification module and a segmentation module, the input end of the feature extraction module is the input end of the target detection model, the output end of the feature extraction module is respectively connected to the input end of the classification module and the input end of the segmentation module, the output end of the classification module and the output end of the segmentation module are the output end of the target detection model, the feature extraction module includes a stacked local feature extraction submodule and an attention submodule, the feature extraction module is used to extract point cloud features of the input data of the target detection model, the classification module is used to classify the output data of the feature extraction module, and the segmentation module is used to segment the output data of the feature extraction module.

[0039] Figure 2 is a technical roadmap of a method for detecting the operating state of an isolating switch according to an embodiment of the present invention, such as Figure 2 As shown, after constructing the initial disconnector action monitoring model, the model is trained to obtain the target monitoring model.

[0040] In this embodiment, an initial detection model is constructed, and a cross entropy loss function (i.e., a target loss function) is used to train the initial detection model to improve the accuracy and generalization ability of the model, thereby obtaining an isolating switch action monitoring model including a feature extraction module, a classification module, and a segmentation module. The key isolating switch action features can be extracted by the feature extraction module, and then the features are classified by the classification module. Finally, the monitoring area is segmented by the segmentation module to determine the specific location of the isolating switch action.

[0041] Step S206, input the test data into the target detection model to obtain the first point cloud data corresponding to the conductive arm of the disconnector corresponding to the test data, the second point cloud data corresponding to the insulating support of the disconnector corresponding to the test data, the third point cloud data corresponding to the transmission cable of the disconnector corresponding to the test data, and the fourth point cloud data corresponding to the base of the disconnector corresponding to the test data.

[0042] In this embodiment, all point cloud data points in a frame of disconnector point cloud data point set are input into the target detection model to obtain point cloud data divided into four components: disconnector conductive arm, insulating support, transmission cable and disconnector base. This can realize automatic recognition and segmentation of disconnectors and their components in the test data for further analysis and processing, thereby improving the accuracy and generalization ability of the target detection model.

[0043] Step S208, obtain the vector corresponding to the first point cloud data, and determine the action state of the disconnector at least based on the vector corresponding to the first point cloud data, the action state includes an open state, a closed state, a disconnection process and a closing process, the open state is the state of the disconnector when it is not powered on, and the closed state is the state of the disconnector when it is powered on.

[0044] like Figure 2 As shown, the trained disconnector action monitoring model is applied to determine the disconnector action status.

[0045] In this embodiment, the operating state of the isolating switch can be judged based on the vector corresponding to the first point cloud data, and abnormal conditions of the isolating switch can be discovered in time, such as abnormal opening or closing process, inconsistent opening and closing states, etc., so that timely measures can be taken for maintenance or processing to ensure the safe and stable operation of the power system.

[0046] As can be seen from the above, in an embodiment of the present invention, a point cloud data set is obtained, the point cloud data set includes multiple point cloud data corresponding to the disconnector, each point cloud data respectively includes multiple frames of point cloud images obtained by scanning the disconnector to be detected by a laser radar, each point cloud image respectively includes multiple point cloud data points, part of the point cloud data set is test data, and the other part is training data, the disconnector at least includes a conductive arm, an insulating support, a transmission cable and a base; an initial detection model is constructed, and the initial detection model is trained using a target loss function to obtain a target detection model, wherein the target detection model includes a feature extraction module, a classification module and a segmentation module, the input end of the feature extraction module is the input end of the target detection model, the output end of the feature extraction module is respectively connected to the input end of the classification module and the input end of the segmentation module, the output end of the classification module and the output end of the segmentation module are the output end of the target detection model, the feature extraction module includes a stacked local feature extraction submodule and an attention submodule, the feature extraction module is used to extract point cloud features of the input data of the target detection model, and the classification module is used to train the output data of the feature extraction module Classification processing is performed, and the segmentation module is used to segment the output data of the feature extraction module; the test data is input into the target detection model to obtain the first point cloud data corresponding to the conductive arm of the disconnector corresponding to the test data, the second point cloud data corresponding to the insulating pillar of the disconnector corresponding to the test data, the third point cloud data corresponding to the transmission cable of the disconnector corresponding to the test data, and the fourth point cloud data corresponding to the base of the disconnector corresponding to the test data; the vector corresponding to the first point cloud data is obtained, and the action state of the disconnector is determined at least according to the vector corresponding to the first point cloud data. The action state includes an open state, a closed state, a disconnection process, and a closing process. The open state is the state of the disconnector when it is not powered on, and the closed state is the state of the disconnector when it is powered on. The three-dimensional point cloud data of the disconnector is obtained by using the laser radar, and combined with the voxel farthest point sampling and weighted mean grouping technology. At the same time, the point cloud channel attention mechanism module is used to enhance the weight of the feature dimension useful for the current task, thereby achieving the technical effect of real-time automatic monitoring of the action state of the disconnector and improving the accuracy and efficiency of monitoring.

[0047] The technical solution provided by the present invention solves the technical problems in the related art that manual inspection is inefficient and has potential safety hazards, and that image processing technology requires complex models and high computing resources.

[0048] According to the above embodiment of the present invention, obtaining a data set includes: scanning the isolating switch through a laser radar to obtain a plurality of initial point cloud data; and filtering and denoising the plurality of initial point cloud data respectively to obtain a point cloud data set.

[0049] In this embodiment, a point cloud data set of an isolating switch can be constructed and data preprocessing can be performed. The three-dimensional spatial information of the isolating switch can be acquired by scanning the isolating switch with a laser radar, and a plurality of initial point cloud data can be obtained. Then, filtering and denoising can be performed to remove some noise points, making the point cloud data more accurate and clear. The obtained point cloud data set can be used for subsequent modeling, identification, positioning and other applications to help achieve precise control and monitoring of the isolating switch.

[0050] Specifically, a laser radar is used to scan the isolating switch to obtain the 3D point cloud data of the isolating switch and construct a point cloud data set of the isolating switch. Each point cloud image in the isolating switch point cloud data set is composed of continuous 3D point cloud images of the isolating switch, which contains at least 100 frames of point cloud images and at most 200 frames of point cloud images. The point cloud data point set P in each frame of the point cloud image is P={p1,p2,…,p n}, p represents a point cloud data point in P, the size of P is n×c, n is the number of point cloud data points contained in P, c is the number of feature dimensions of the point cloud data point, and c=3, that is, each point p contains (x, y, z) three-dimensional coordinate information.

[0051] In addition, the point cloud images in the disconnector point cloud dataset are filtered and denoised, and the point cloud positions of the disconnector conductive arms in the point cloud images are marked.

[0052] According to the above embodiment of the present invention, the feature extraction module includes a first local feature extraction submodule, a first channel attention submodule, a second local feature extraction submodule and a second channel attention submodule connected in sequence, and inputs the test data into the target detection model to obtain the first point cloud data corresponding to the conductive arm of the isolating switch corresponding to the test data, the second point cloud data corresponding to the insulating pillar of the isolating switch corresponding to the test data, the third point cloud data corresponding to the transmission cable of the isolating switch corresponding to the test data, and the fourth point cloud data corresponding to the base of the isolating switch corresponding to the test data, including: inputting the test data into the first local feature extraction submodule to perform a first downsampling and a first feature extraction processing on the test data to obtain a plurality of first point cloud data points; inputting the plurality of first point cloud data points into the first channel attention submodule in sequence to perform a first feature dimension weighted processing on the plurality of first point cloud data points to obtain a plurality of second point cloud data points; and The second point cloud data points are input into the second local feature extraction submodule to perform a second downsampling and a second feature extraction process on the multiple second point cloud data points to obtain multiple third point cloud data points; the multiple third point cloud data points are sequentially input into the second channel attention submodule to perform a second feature dimension weighted process on the multiple third point cloud data points to obtain multiple fourth point cloud data points; the multiple test data, the multiple second point cloud data points and the multiple fourth point cloud data points are input into the segmentation module to respectively perform segmentation process on the multiple fourth point cloud data points to obtain multiple fifth point cloud data points, and the number of the fifth point cloud data points is the same as the number of point cloud data in the test data; the multiple fourth point cloud data points are input into the classification module to respectively perform classification process on the multiple fourth point cloud data points to obtain multiple sixth point cloud data points, and the multiple fifth point cloud data points and the multiple sixth point cloud data points form the first point cloud data, the second point cloud data, the third point cloud data and the fourth point cloud data.

[0053] In this embodiment, a frame of disconnector point cloud data point set P = {p1, p2, ..., p nAll the point cloud data points in} are input into the disconnector action monitoring model, and the size of P is n×c, where n is the number of point cloud data points contained in the point cloud data point set P, and c is the number of feature dimensions of the point cloud data points. After P is processed by the first VPP module (i.e., the first local feature extraction submodule) for downsampling (i.e., the first downsampling) and feature extraction (i.e., the first feature extraction), the point cloud data point set P1 (i.e., multiple first point cloud data points) is obtained. After P1 is processed by the first VSE module (i.e., the first channel attention submodule) for feature dimension weighting of the point cloud data points (i.e., the first feature dimension weighting), the point cloud data set P2 (i.e., multiple second point cloud data points) is obtained. P2 is processed by the second VPP module (i.e., the second local feature extraction submodule) for downsampling (i.e., the second downsampling) and feature extraction (i.e., the second feature extraction) ) is processed by the second VSE module (i.e., the second channel attention submodule) to obtain a point cloud data point set P3 (i.e., a plurality of third point cloud data points). After P3 is processed by the second VSE module (i.e., the second channel attention submodule) for feature dimension weighting of the point cloud data points (i.e., the second feature dimension weighting), a point cloud data point set P4 (i.e., a plurality of fourth point cloud data points) is obtained. Finally, P4 is respectively input into the segmentation module ST and the classification module CF. After segmentation and classification, point cloud data of four components, namely, the first point cloud data, the second point cloud data, the third point cloud data and the fourth point cloud data, are obtained.

[0054] In the above embodiment of the present invention, the first local feature extraction submodule and the second local feature extraction submodule are the same, and the first local feature extraction submodule and the second local feature extraction submodule respectively include a downsampling layer, a weighted mean grouping layer, and a local feature point set input point mesh layer, and the test data is input into the first local feature extraction submodule to perform a first downsampling and a first feature extraction process on the test data to obtain a plurality of first point cloud data points, including: inputting the test data into the downsampling layer to perform a voxel farthest point downsampling process on the test data to obtain a first sampling point set, the first sampling point set including a plurality of first sampling points; inputting the test data and the first sampling point set into the weighted mean grouping layer to perform a grouping process with the first sampling point as the centroid point to obtain a second sampling point set, the second sampling point set being a local feature point set corresponding to each first sampling point, the second sampling point set including a plurality of second sampling points; inputting the second sampling point set into the local feature point set input point mesh layer The method comprises: performing feature extraction processing on multiple second sampling points respectively to obtain multiple first point cloud data points; inputting the multiple second point cloud data points into the second local feature extraction submodule to perform second downsampling and second feature extraction processing on the multiple second point cloud data points to obtain multiple third point cloud data points, including: inputting the multiple second point cloud data points into the downsampling layer in sequence to perform voxel farthest point downsampling processing on the second point cloud data points to obtain a third sampling point set, the third sampling point set including multiple third sampling points; inputting the multiple second point cloud data points and the third sampling point set into the weighted mean grouping layer to perform grouping processing with the third sampling point as the centroid point to obtain a fourth sampling point set, the fourth sampling point set being the local feature point set corresponding to each third sampling point, the fourth sampling point set including multiple fourth sampling points; inputting the fourth sampling point set into the local feature point set input point mesh layer to perform feature extraction processing on the multiple fourth sampling points respectively to obtain multiple third point cloud data points.

[0055] Figure 3 is a structural diagram of a VPP module of a method for detecting an operating state of an isolating switch according to an embodiment of the present invention, such as Figure 3 As shown in the figure, the isolating switch point cloud data point set P is input into the VPP module. After the input isolating switch point cloud information passes through the voxel farthest point downsampling layer VFS, n′ data points are sampled from P, and then the n′ data points after sampling are taken as the center points and input into the weighted mean grouping layer PAG for grouping to obtain n′ local feature point sets. Finally, the n′ sampling points and the corresponding local feature point sets are input into the point network layer Pointnet to extract features.

[0056] Obtain the first sampling point set, that is, input the above-mentioned isolating switch point cloud data point set P = {p1, p2, ..., p n} to the VFS layer, n is the number of point cloud data points in P, c is the number of feature dimensions of the point cloud data points in P, the point cloud space where the point cloud data points in P are located is divided into 4×4×4 voxel spaces, and the distance between each point is calculated in each voxel space using the following formula:

[0057] Among them, d ij For point p i and point p j The distance between i ,y i ,z i ) is point p i The position coordinates, (x j ,y j ,z j ) is point p j The position coordinates, p i and p j For any point in the voxel space, sample from each voxel space points, the specific steps are as follows: 1) Randomly select an unsampled voxel space, create a query point set T, which is used to temporarily store the sampling points, randomly select a point a from the selected voxel space and put it into T, and select a point b farthest from point a from the remaining points in the voxel space (that is, the points not in T) and put it into T. 2) Put the taken points a and b into T, and from the remaining points in the voxel space, take the point c farthest from T and put point c into T, where the method for solving c is as follows: calculate the distance d from all the remaining points in the voxel space to T, and the point with the largest distance d from T is point c, where the distance d is calculated by, for any point t among the remaining points in the voxel space, calculating the distance from t to all points in T, and then taking the minimum value of the distance between t and the points in T as the distance d from t to T. t 3) Repeat the above step 2 until a sample is found in the voxel space. 4) Repeat the above steps until all voxel spaces are sampled, merge all the sampling points sampled from the voxel spaces, and get a total of n′ sampling points, forming a sampling point data point set V′={v1′,v2′,…,v n ″} (i.e., the first set of sampling points) is the output of the VFS layer, and the size of V′ is n′×c.

[0058] It should be noted that the above-mentioned voxel space is cubes or cube grid units evenly distributed in three-dimensional space.

[0059] The second sampling point set is obtained, that is, the disconnector point cloud data point set P = {p1, p2, ..., p n} and the above sampling point data set V′={v1′,v2′,…,v n ″}, take the sampling point in V′ as the centroid point to perform grouping operation and obtain the local feature point set corresponding to the sampling point. The specific steps are as follows: 1) Randomly select an ungrouped sampling point v in V′ i ′ as the centroid point, set the spherical neighborhood radius r and the number of local feature points k for the grouping operation. 2) Find v in P i ′ in the ball neighborhood and v i ′The nearest k neighboring points are used as local feature points. These k neighboring points constitute v i ′’s local feature point set, find v i The formula of the local feature point set corresponding to ′ is as follows:

[0060] Where q(r,k) is v i ′’s local feature point set, ∥vv i ′∥ is the neighborhood point v and v i ′, s is the number of points in the set q(r,k), s≤k means that in v i ′, select k nodes with v i ′ is the closest neighboring point that is smaller than the radius r of the sphere neighborhood. If v i When the number of neighboring points in the ball neighborhood of ′ is greater than or equal to k, select the first k points that are related to v i ′The nearest neighboring points are used as local feature points. These k neighboring points constitute v i ′’s local feature point set, based on point v i ′ is completed; if v i If the number of neighboring points in the ball neighborhood of ′ is less than k, then v is obtained by calculation i The weighted mean point of all points in the ball neighborhood of ′ is as follows:

[0061] in, is the weighted mean point, v j v i Any point in the ball neighborhood of ′, s is v i The number of points in the ball neighborhood of ′, d j v i Any point v in the ball neighborhood of ′ j to v i ′, d s v i All points in the ball neighborhood of ′ to v i ′. The weighted mean point Copy it many times to get k points as v i ′ local feature points, and get v i′’s local feature point set, based on point v i ' is grouped. 3) Repeat the above steps until the local feature point sets corresponding to all sampling points in V' are obtained, the grouping operation is completed, and a point cloud data point set V" (i.e., the second sampling point set) of size n'×k×c containing sampling points and corresponding local feature point sets is obtained, where n' is the number of sampling points, k is the number of local feature points of the sampling points, and c is the number of feature dimensions of the sampling points; output the point cloud data point set V".

[0062] It should be noted that the above-mentioned centroid point is the center point of the spherical neighborhood used for grouping; the above-mentioned weighted mean point is a data point obtained by calculating the weight value according to the distance from the neighborhood point to the centroid, and then using the weight value to weightedly calculate the average value of the neighborhood point.

[0063] Specifically, the above-mentioned multiple first point cloud data points are obtained. The above-mentioned point cloud data point set V″ is input into the point network layer Pointnet. After the features are extracted by the fully connected layer, ReLu activation function and maximum pooling layer MaxPool of Pointnet, the feature dimension of the isolating switch point cloud data point is increased to c′, and the isolating switch point cloud data point set P1 with a size of n′×c′ after the feature dimension is expanded is output.

[0064] It should be noted that the above-mentioned Pointnet is a lightweight convolutional network suitable for point cloud data. It extracts features by fusing local information and consists of a fully connected layer, a ReLu activation function, and a maximum pooling MaxPool.

[0065] Taking the PAG layer parameter spherical neighborhood radius r=5 and the number of domain points k=10 as an example to illustrate this embodiment, a set of point cloud data points P is input, assuming that the size is 4096×3. After downsampling at the VFS layer, a sampling point data point set V′ is obtained, the size is 1024×3, and an ungrouped point cloud data point v is randomly selected from V′. i ′ is the centroid point for grouping operations, and v i ′ is the center of mass, the spherical neighborhood with radius r, and the distance v is found in P i ′The nearest k points in the spherical neighborhood are placed in v i ′, if the number of points in the spherical neighborhood is less than k, then calculate the weighted mean point of all points in the spherical neighborhood and copy the weighted mean point into v i The local feature points of ′ are concentrated to k points, and v i′ local feature point set, repeatedly select point cloud data points from the sampling point data set V′ for grouping operation. After all sampling points in the sampling point data set V′ have obtained the corresponding local feature point set, the grouping operation is completed, and a point cloud data point set V″ of size 1024×10×3 containing the sampling points and the corresponding local feature point set is obtained. Finally, after Pointnet is used to extract features, the point cloud data point set P1 is obtained as the output of the VPP module, with a size of 1024×128.

[0066] In the above-mentioned embodiment of the present invention, the first channel attention submodule and the second channel attention submodule are the same, and the first channel attention submodule and the second channel attention submodule respectively include an activation function layer and a calculation layer, and the plurality of first point cloud data points are sequentially input into the first channel attention submodule to perform first feature dimension weighted processing on the plurality of first point cloud data points to obtain a plurality of second point cloud data points, including: sequentially inputting the plurality of first point cloud data points into the activation function layer to respectively calculate the weight of the feature dimension of each first point cloud data point to obtain a plurality of first weight values; inputting the plurality of first weight values ​​and the corresponding first point cloud data points into the calculation layer to respectively perform first feature dimension weighted processing on the plurality of first point cloud data points The first weight value and the corresponding first point cloud data point are subjected to dot multiplication processing to obtain multiple second point cloud data points; the multiple third point cloud data points are sequentially input into the second channel attention submodule to perform second feature dimension weighted processing on the multiple third point cloud data points to obtain multiple fourth point cloud data points, including: the multiple third point cloud data points are sequentially input into the activation function layer to respectively calculate the weight of the feature dimension of each third point cloud data point to obtain multiple second weight values; the multiple second weight values ​​and the corresponding third point cloud data points are input into the calculation layer to respectively perform dot multiplication processing on the second weight value and the corresponding third point cloud data point to obtain multiple fourth point cloud data points.

[0067] In this embodiment, a feature dimension attention module PSE suitable for point cloud data is used to obtain the importance of the information dimension of each point cloud data point, that is, to obtain the weight value of the feature dimension of each point cloud data point, and then perform weighted calculation on the feature dimension of each point cloud data point according to this weight value, thereby increasing the weight of the feature dimension that is useful for the current task and reducing the weight of the feature dimension that is not very useful for the current task.

[0068] Figure 4 is a schematic diagram of a PSE module of a method for detecting an operating state of an isolating switch according to an embodiment of the present invention, such as Figure 4 As shown, the above-mentioned isolating switch point cloud data point set P1 is input, the feature dimension of each point cloud data point is calculated through the sigmoid activation function, and the weight value of the feature dimension of the point cloud data point is obtained. Finally, the weight value of the feature dimension of the point cloud data point is multiplied by its corresponding feature dimension to obtain the weighted isolating switch point cloud data point set P2.

[0069] The specific steps are as follows: 1) Input the disconnector point cloud data point set P1 of size n′×c′ to the PSE module, where n′ is the number of point cloud data points in P1, c′ is the number of feature dimensions of each point cloud data point, and use the sigmoid activation function to obtain the weight value of the feature dimension of each point cloud data point. The calculation formula is as follows: W i =σ(p i ), where W i For p i The weight value of the feature dimension, p i is the input point cloud data point, σ(·) represents the sigmoid function, and the calculation formula of σ(·) is as follows: 2) The weight value W of the feature dimension of the point cloud data point in P1 and the point cloud data point i Corresponding to the dot multiplication operation, the feature dimension of the point cloud data point is weighted to obtain the point cloud data point after the feature dimension is weighted. The calculation formula is as follows: Among them, p i (i.e., the first point cloud data point) is the point cloud data point in P1, W i (ie, the first weight value) is p i The corresponding feature dimension weight value, p i ′ is the point cloud data point after the feature dimension is weighted. After all the point cloud data points are calculated, the disconnector point cloud data point set P2 after the feature dimension is weighted is obtained.

[0070] Figure 5 is a schematic diagram of point cloud feature information of a method for detecting an operating state of an isolating switch according to an embodiment of the present invention, such as Figure 5 As shown in the figure, the dimension of the point cloud is generally 3D, but after the network extracts features, a lot of feature information will be added. By expanding the feature dimension of the point cloud, the number of feature dimensions of the point cloud will also increase. When extracting features from the point cloud data, the xyz coordinate information of the point cloud data does not participate in the calculation, and the extracted feature information is spliced ​​after the xyz coordinate information.

[0071] Taking the point cloud data set P1 with a size of 1024×128 as an example to illustrate this embodiment, the point cloud data set P1 is input into the PSE module, and the sigmoid function is used to calculate the weight value W of the feature dimension of each point cloud data point. i , and then the weight value W i Perform point multiplication with the feature dimension of the corresponding point cloud data point, weight the feature dimension of the point cloud data point, and obtain the point cloud data set P2 with weighted feature dimension.

[0072] In addition, multiple fourth point cloud data points are obtained by stacking the VPP module and the PSE module, and the above point cloud data point set P2 is input into the second VPP module for downsampling and feature extraction to obtain the point cloud data point set P3, and P3 is input into the second PSE module (i.e., the second channel attention submodule) to weight the feature dimensions of the point cloud data points to obtain the point cloud data point set P4.

[0073] It should be noted that the operation steps of the second VPP module are consistent with those of the first VPP module, and the operation steps of the second PSE module are consistent with those of the first PSE module, which will not be repeated here.

[0074] In the above embodiment of the present invention, the segmentation module includes a first interpolation layer, a first convolution layer, a second interpolation layer and a second convolution layer connected in sequence, and multiple test data, multiple second point cloud data points and multiple fourth point cloud data points are input into the segmentation module to respectively perform segmentation processing on the multiple fourth point cloud data points to obtain multiple fifth point cloud data points, including: inputting the multiple second point cloud data points and the multiple fourth point cloud data points into the first interpolation layer, taking the second point cloud data points as the points to be interpolated and the fourth point cloud data points as the neighboring feature points, performing the first interpolation processing on the second point cloud data points and the fourth point cloud data points to obtain to a first data point set; input the first data point set into a first convolutional layer to perform feature extraction processing on the first data point set to obtain a second data point set, the second data point set including a plurality of second data points; input the second data point set and the test data into a second interpolation layer to perform a second interpolation processing on the second data point set and the test data with the test data as the points to be interpolated and the second data points as the neighboring feature points to obtain a third data point set, the third data point set including a plurality of third data points; input the third data point set into a second convolutional layer to perform feature extraction processing on the plurality of third data points to obtain a plurality of fifth point cloud data points.

[0075] In this embodiment, the point cloud data set P4 is input into the segmentation module ST to perform a segmentation operation to classify different point cloud data points into different categories, so that the point cloud data can be analyzed and processed more deeply and accurately.

[0076] Specifically, first obtain the (x, y, z) coordinates of all point cloud data points from the above point cloud data point set P2. The (x, y, z) coordinates of all point cloud data points form a point cloud coordinate set N, where the size of N is n′. The points within N are the points to be interpolated, and the points in the point cloud data point set P4 are the neighboring feature points. The interpolation layer Interpolate is input for interpolation operation, and finally a Unit Pointnet network is used to extract features to obtain a point cloud data point set P5 (i.e., multiple second data point sets) of size n′×c′.

[0077] Afterwards, the (x, y, z) coordinates of all point cloud data points are obtained from the above data point set P. The (x, y, z) coordinates of all point cloud data points form a point cloud coordinate set N1, the size of N1 is n, and the points in N1 are used as points to be interpolated. The points in the point cloud data point set P5 are neighboring feature points and are input into the interpolation layer Interpolate for interpolation. Finally, features are extracted through a UnitPointnet network to obtain a point cloud data point set P6 (i.e., multiple fifth point cloud data points) of size n×(c+s), where s is the number of different categories of segmentation. At this time, the number of point cloud data points of P6 is restored to the same number of point cloud data points as the above point cloud data point set P, and the point cloud of P6 has been segmented into multiple different component categories.

[0078] It should be noted that the above-mentioned interpolation layer Interpolate is an interpolation method based on the weighted average of the inverse distance of k nearest neighbors, which is a common method in point cloud processing; the above-mentioned Pointnet is a lightweight convolutional network suitable for point cloud data, which extracts features by fusing local information, and consists of a fully connected layer, ReLu activation function and maximum pooling MaxPool; the above-mentioned Unit Pointnet is similar to 1×1 convolution. Unit Pointnet is a Pointnet without pooling operation. It uses the fully connected layer and ReLU to reorganize the feature vector of each point in the input point cloud data point set, and does not change the number of points.

[0079] In the above embodiment of the present invention, the classification module includes a third convolutional layer and a fully connected layer connected in sequence, and multiple fourth point cloud data points are input into the classification module to classify the multiple fourth point cloud data points respectively to obtain multiple sixth point cloud data points, including: inputting the multiple fourth point cloud data points into the third convolutional layer in sequence to perform feature extraction processing on the multiple fourth point cloud data points respectively to obtain multiple seventh point cloud data points; inputting the multiple seventh point cloud data points into the fully connected layer to classify the seventh point cloud data points to obtain multiple sixth point cloud data points.

[0080] In this embodiment, the point cloud data point set P4 is input into the classification module CF for classification operation, and multiple data points can be classified at the same time to obtain multiple classification results, thereby realizing rapid classification processing of multiple data points.

[0081] Specifically, the above point cloud data point set P4 is input into the classification module CF, and after the features are extracted by a point network layer Pointnet (i.e., the third convolutional layer), the classification is completed through a fully connected layer FC to obtain the four categories of information corresponding to the point cloud data, namely, the disconnector conductive arm, insulating pillar, transmission cable and disconnector base. This is combined with the above point cloud data point set P6 divided into different components to obtain the point cloud data of the four components that are divided and classified into the disconnector conductive arm, insulating pillar, transmission cable and disconnector base.

[0082] The embodiment is described by taking the point cloud data point set P with a size of 4096×3 as an example. A frame of disconnector point cloud data point set P is input. After downsampling and feature extraction by the first VPP module, a point cloud data set point P1 with a size of 1024×128 is obtained. After passing through the second PSE module, a point cloud data point set P2 with weighted feature dimensions of the point cloud data points is obtained. After downsampling and feature extraction by the second VPP module, a point cloud data point set P3 with a size of 256×512 is obtained. After passing through the second PSE module, a point cloud data point set P4 with weighted feature dimensions of the point cloud data points is obtained. The point cloud data point set P4 is input into the segmentation module. First, a point cloud coordinate set N with a size of 1024×3 is obtained from the point cloud data point set P2. An interpolation operation is performed with 1024 points in the point cloud coordinate set N as the points to be interpolated. The Unit After the Pointnet network extracts features, a point cloud dataset P5 with a size of 1024×128 is obtained. A point cloud coordinate set N1 with a size of 4096×3 is obtained from the disconnector point cloud dataset P. The 4096 points in the point cloud coordinate set N1 are used as interpolation points for interpolation. After the Unit Pointnet network extracts features, a point cloud data point set P6 with a size of 4096×(3+4) is obtained, where 3 is the coordinate information of the point cloud and 4 is the number of four categories into which the point cloud data is divided. The point cloud data set P4 is input into the classification module. After the features are extracted by a point network layer Pointnet, it passes through a fully connected layer FC with a size of (512, 128, 4), where 512 is the number of layers of the first hidden layer of FC, 128 is the number of layers of the second hidden layer of FC, and 4 is the number of layers of the output layer of FC (i.e. the number of classified categories). Finally, the four category information of the disconnector conductive arm, insulating pillar, transmission cable and disconnector base corresponding to the point cloud data is obtained. Combined with the point cloud data set P6 divided into different components, the point cloud data of the four components of the disconnector conductive arm, insulating pillar, transmission cable and disconnector base that are divided and classified are obtained.

[0083] In addition, the disconnector action monitoring model needs to be trained later. The specific steps are as follows: First, the disconnector point cloud data set is divided into a training set, a validation set, and a test set in a ratio of 6:2:2, and the disconnector action monitoring model is trained and verified. A batch of the disconnector 3D point cloud data training set is input into the disconnector action monitoring model for training. The training batch size batch_size=16, the training batch epoch=200, the initial learning rate is 0.001, and the loss value loss is calculated using the cross entropy loss function. After the training of all batches of the disconnector point cloud data training set is completed, the disconnector point cloud data validation set is input into the disconnector action monitoring model in batches to obtain the corresponding batch loss value batch_loss. During training and verification, the disconnector action monitoring model will automatically learn and adjust parameters according to each loss and batch_loss situation. When the loss and batch_loss values ​​tend to converge and the difference between them is very small, the disconnector action monitoring model training is completed.

[0084] It should be noted that the above cross entropy loss function is a loss function commonly used in deep learning model training. It is a loss function used to measure the difference between two probability distributions.

[0085] According to the above embodiment of the present invention, determining the action state of the disconnector at least according to the vector corresponding to the first point cloud data includes: according to the formula Calculate the angle α between the conductive arm corresponding to the i-th frame point cloud image and the conductive arm in the closed state i , where u i =((x0,y0),(x i ,y i )), (x0, y0) is the coordinate of the end of the conductive arm fixed to the stationary insulating support, (x i ,y i ) is the coordinate corresponding to the end of the conductive arm rotating around the insulating support, v = ((x0, y0), (x v ,y v )),(x v ,y v ) is the coordinate corresponding to one end of the conducting arm rotating around the insulating support in the closed state, u i v is the vector u i The dot product of the vector v, |u i | is the vector u i The modulus of |v| is the modulus of vector v; determine α i Is it greater than α i+n , in α i Greater than α i+n In the case of i+nThe corresponding disconnector action state is the closing process, where α i+n is the angle between the conductive arm corresponding to the i+nth frame point cloud image and the conductive arm in the closed state, n is greater than 10; i Less than α i+n In the case of i The corresponding disconnector action state is the disconnection process; i Equal to α i+n In the case of i and α i+n Whether α is satisfied i+n =α i =0, at α i+n =α i = 0, determine α i The corresponding disconnector is in the closed state; i+n =α i ≠0, determine α i The corresponding isolating switch is in the open state.

[0086] Figure 6 Schematic diagram of coordinates of both ends of the conductive arm of the isolating switch according to an embodiment of the present invention. Figure 6 As shown in the figure, when the double-column horizontal rotary disconnector is in the process of opening or closing, the conductive arm will only move in a plane. Input a point cloud video containing continuous point cloud images of the disconnector (the point cloud video of the disconnector point cloud image can be obtained in real time using a laser radar), and input each frame of the point cloud image in the point cloud video into the disconnector action monitoring model in chronological order to obtain the point cloud data set G = {g1, g2, …, g m}, g1 represents the point cloud data of the conductive arm of the disconnector in the first frame of the point cloud image, and m represents the number of frames contained in the point cloud video. The coordinates (x0, y0) and (x i ,y i ), where i represents the isolating switch conductive arm corresponding to the i-th frame point cloud image. The above coordinates can then be used to confirm the vector u of the isolating switch conductive arm corresponding to the i-th frame point cloud image on the x, y coordinate plane. i :u i =((x0,y0),(x i ,y i )), the vector v of the disconnector conducting arm in the closed state is set to: v = ((x0, y0), (x v ,y v )), calculate the vector u by the above formula iThe angle between the conductive arm of the disconnector corresponding to the i-th frame point cloud image and the conductive arm of the disconnector in the closed state can be obtained by i , compare the angle α corresponding to the i-th frame point cloud image in the point cloud image i The angle α corresponding to the i+10th frame point cloud image i+10 , if α i+10 Greater than α i , then the isolating switch is in the disconnection process; if α i+10 Less than α i , then the isolating switch corresponding to the i+10th frame point cloud image is in the closing process. i+10 =α i =0, the conductive arm of the disconnector corresponding to the i+10th frame point cloud image coincides with the conductive arm when it is in the closed state. At this time, the disconnector corresponding to the i+10th frame point cloud image is in the closed state, otherwise it is in the open state.

[0087] It should be noted that the frame rate of the point cloud image acquired by the above-mentioned laser radar is about 10 frames per second to 20 frames per second, so when comparing the point cloud images with a difference of 10 frames, the displacement of the conductive arm is more obvious; the above-mentioned isolating switch is in the closed state when it is closed and energized, and is in the open state when it is disconnected and not energized. The isolating switch is in the closing process when it changes from the disconnected and not energized state to the closed and energized state, and is in the disconnecting process when it changes from the closed and energized state to the disconnected and not energized state.

[0088] The embodiment is described by taking the input of a point cloud video containing 100 frames of continuous three-dimensional point cloud images of an isolating switch as an example. The three-dimensional point cloud images of the isolating switch are sequentially input into the isolating switch action monitoring model in chronological order. After the isolating switch action monitoring model processes the three-dimensional point cloud images of the isolating switch, it outputs the point cloud data of the isolating switch conductive arm and its corresponding vector u i , calculate the vector u i The angle α between the vector v of the disconnector conducting arm in the set closed state i .

[0089] Figure 7 Schematic diagram of a conductive arm of an isolating switch according to an embodiment of the present invention. Figure 7 As shown in the figure, the straight line is an abstract representation of the conductive arm of the disconnector from a top-down perspective, compared with the vector u corresponding to the point cloud image of the 10th frame 10 =((x0,y0),(x 10 ,y 10 ))The angle α between v 10 The vector u corresponding to the 20th frame point cloud image 20 =((x0,y0),(x 20 ,y 20))The angle α between v 20 , at this time α 20 Greater than α 10 The isolating switch is in the process of disconnection. Compare the vector u corresponding to the 50th frame point cloud image 50 =((x0,y0),(x 50 ,y 50 ))The angle α between v 50 The vector u corresponding to the 60th frame point cloud image 60 =((x0,y0),(x 60 ,y 60 ))The angle α between v 60 , at this time α 60 =α 50 , and α 60 If the value is not 0, the isolation switch is in the open state.

[0090] Figure 8 Schematic diagram of the disconnector action monitoring model structure according to an embodiment of the present invention. Figure 8 As shown in the figure, point cloud features are extracted by stacking the VPP module and the PSE module, and then the classification module ST and the segmentation module CF are used to classify and segment the disconnector point cloud data to realize the recognition of the point cloud area of ​​the disconnector conductive arm. Among them, the segmentation module ST realizes the point cloud segmentation function by extracting features through interpolation upsampling Interpolate and Unit Pointnet network, and the classification module CF realizes the point cloud classification function through a point network layer Pointnet and a fully connected layer FC.

[0091] It should be noted that the above-mentioned disconnector action monitoring model is used to identify and classify the information of the disconnector conductive arm and determine the action state of the disconnector.

[0092] Therefore, the technical solution provided by the above embodiment of the present invention solves the following problems: 1) The traditional detection of the action state of the isolating switch needs to rely on manual inspection, which has the disadvantages of low efficiency and certain safety hazards to the inspectors; 2) The action state of the isolating switch is judged by image processing technology, and the information level that can be obtained is relatively small, and a complex image processing model is often required, which requires higher computing resources. At the same time, it has the following beneficial effects: 1) A voxel farthest point sampling method is designed, which divides the point cloud space into multiple voxel spaces and then performs farthest point downsampling, and retains local feature information while downsampling; 2) A weighted mean grouping method is designed. During the grouping process, when it is necessary to copy the neighborhood feature points, the domain feature points are weighted according to the distance and then copied; 3) A point cloud channel attention mechanism module PSE is designed, which obtains the weight value of the feature dimension of each point cloud data point, and then weights the feature dimension of the point cloud data point according to this weight value, thereby increasing the weight of the feature dimension that is useful for the current task and reducing the weight of the feature dimension that is not very useful for the current task.

[0093] That is, the technical solution provided by the embodiment of the present invention comprehensively considers various factors and uses a laser radar to obtain the three-dimensional point cloud data of the isolating switch. The three-dimensional point cloud data contains the three-dimensional coordinate information of the isolating switch, has rich information levels, and has a simple data structure. It is more efficient and intuitive for real-time judgment of the isolating switch action information.

[0094] In addition, it should be noted that the present invention discloses a method for monitoring the action of an isolating switch based on laser point cloud, designs a local feature extraction module VPP, wherein a voxel farthest point sampling method and a weighted mean grouping method are designed and used to better retain the local feature information of the isolating switch conductive arm while extracting the point cloud features of the isolating switch; designs a feature dimension attention module PSE suitable for point cloud data, which is used to obtain the weight value of the feature dimension of each point cloud data point, and then performs weighted calculation on the feature dimension of each point cloud data point according to the weight value, thereby increasing the weight of the feature dimension that is useful for the current task and reducing the weight of the feature dimension that is not very useful for the current task; designs a method for judging the opening and closing state of the isolating switch, which is used to accurately judge the action state of the isolating switch; by calculating the vector corresponding to the point cloud information of the isolating switch conductive arm, and calculating the angle with the isolating switch conductive arm vector in the set closing state, and comparing the angle changes of the continuous frames, the action state of the isolating switch can be judged in real time and accurately, so as to solve the problems of insufficient efficiency and high labor cost in monitoring the action state of the isolating switch, and realize real-time automatic monitoring of the action state of the isolating switch.

[0095] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0096] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods of various embodiments of the present invention.

[0097] According to an embodiment of the present invention, there is also provided a device for detecting the operating state of an isolating switch for implementing the above-mentioned method for detecting the operating state of an isolating switch. Fig. 9 is a schematic diagram of a detection device for the operating state of an isolating switch according to an embodiment of the present invention, such as Fig. 9 As shown, the device includes: an acquisition module 1001, a first processing module 1003, an input module 1005 and a second processing module 1007. The detection device for the operating state of the isolating switch is described below.

[0098] The acquisition module 1001 is used to acquire a point cloud data set, which includes multiple point cloud data corresponding to the disconnector, each point cloud data includes multiple frames of point cloud images obtained by scanning the disconnector to be detected by a laser radar, and each point cloud image includes multiple point cloud data points. Part of the point cloud data set is test data, and the other part is training data. The disconnector includes at least a conductive arm, an insulating support, a transmission cable and a base.

[0099] The first processing module 1003 is used to construct an initial detection model and use the target loss function to train the initial detection model to obtain a target detection model, wherein the target detection model includes a feature extraction module, a classification module and a segmentation module, the input end of the feature extraction module is the input end of the target detection model, the output end of the feature extraction module is respectively connected to the input end of the classification module and the input end of the segmentation module, the output end of the classification module and the output end of the segmentation module are the output end of the target detection model, the feature extraction module includes a stacked local feature extraction submodule and an attention submodule, the feature extraction module is used to extract point cloud features of the input data of the target detection model, the classification module is used to classify the output data of the feature extraction module, and the segmentation module is used to segment the output data of the feature extraction module.

[0100] Input module 1005 is used to input the test data into the target detection model to obtain first point cloud data corresponding to the conductive arm of the disconnector corresponding to the test data, second point cloud data corresponding to the insulating support of the disconnector corresponding to the test data, third point cloud data corresponding to the transmission cable of the disconnector corresponding to the test data, and fourth point cloud data corresponding to the base of the disconnector corresponding to the test data.

[0101] The second processing module 1007 is used to obtain the vector corresponding to the first point cloud data, and determine the action state of the disconnector at least based on the vector corresponding to the first point cloud data. The action state includes an open state, a closed state, a disconnection process and a closing process. The open state is the state of the disconnector when it is not powered on, and the closed state is the state of the disconnector when it is powered on.

[0102] It should be noted here that the above-mentioned acquisition module 1001, the first processing module 1003, the input module 1005 and the second processing module 1007 correspond to steps S202 to S208 in the above-mentioned embodiments, and the four modules and the corresponding steps implement the same instances and application scenarios, but are not limited to the contents disclosed in the above-mentioned embodiments.

[0103] As can be seen from the above, in the scheme recorded in the above-mentioned embodiment of the present invention, the acquisition module can be used to first acquire a point cloud data set, the point cloud data set includes multiple point cloud data corresponding to the disconnector, each point cloud data respectively includes multiple frames of point cloud images obtained by scanning the disconnector to be detected by laser radar, each point cloud image respectively includes multiple point cloud data points, part of the point cloud data set is test data, and the other part is training data, the disconnector at least includes a conductive arm, an insulating support, a transmission cable and a base; then the first processing module can be used to construct an initial detection model, and the initial detection model is trained using a target loss function to obtain a target detection model, wherein the target detection model includes a feature extraction module, a classification module and a segmentation module, the input end of the feature extraction module is the input end of the target detection model, the output end of the feature extraction module is respectively connected to the input end of the classification module and the input end of the segmentation module, the output end of the classification module and the output end of the segmentation module are the output end of the target detection model, the feature extraction module includes a stacked local feature extraction submodule and an attention submodule, the feature extraction module is used to extract point cloud features of the input data of the target detection model, and the classification module is used to train the output of the feature extraction module. The output data is classified and processed, and the segmentation module is used to segment the output data of the feature extraction module; then the input module can be used to input the test data into the target detection model to obtain the first point cloud data corresponding to the conductive arm of the disconnector corresponding to the test data, the second point cloud data corresponding to the insulating pillar of the disconnector corresponding to the test data, the third point cloud data corresponding to the transmission cable of the disconnector corresponding to the test data, and the fourth point cloud data corresponding to the base of the disconnector corresponding to the test data; finally, the second processing module can be used to obtain the vector corresponding to the first point cloud data, and at least according to the vector corresponding to the first point cloud data, the action state of the disconnector is determined, and the action state includes an open state, a closed state, a disconnection process, and a closing process. The open state is the state of the disconnector when it is not powered on, and the closed state is the state of the disconnector when it is powered on, so as to obtain the three-dimensional point cloud data of the disconnector by using the laser radar, and combine the voxel farthest point sampling and weighted mean grouping technology. At the same time, the point cloud channel attention mechanism module is used to improve the weight of the feature dimension useful for the current task, thereby realizing the technical effect of real-time automatic monitoring of the action state of the disconnector and improving the accuracy and efficiency of monitoring.

[0104] The technical solution provided by the present invention solves the technical problems in the related art that manual inspection is inefficient and has potential safety hazards, and that image processing technology requires complex models and high computing resources.

[0105] In an optional embodiment, the acquisition module includes: a scanning processing unit, which is used to scan and process the isolating switch through a laser radar to obtain multiple initial point cloud data; and a filtering and denoising processing unit, which is used to filter and denoise the multiple initial point cloud data respectively to obtain a point cloud data set.

[0106] In an optional embodiment, the input module includes: a first input unit, which is used to input the test data into the first local feature extraction submodule to perform a first downsampling and a first feature extraction process on the test data to obtain a plurality of first point cloud data points; a second input unit, which is used to sequentially input the plurality of first point cloud data points into the first channel attention submodule to perform a first feature dimension weighted process on the plurality of first point cloud data points to obtain a plurality of second point cloud data points; a third input unit, which is used to input the plurality of second point cloud data points into the second local feature extraction submodule to perform a second downsampling and a second feature extraction process on the plurality of second point cloud data points to obtain a plurality of third point cloud data points; a fourth input unit, which is used to sequentially input the plurality of third point cloud data points into the second channel attention submodule The attention submodule performs weighted processing on the second feature dimension of the multiple third point cloud data points to obtain multiple fourth point cloud data points; the fifth input unit is used to input the multiple test data, the multiple second point cloud data points and the multiple fourth point cloud data points into the segmentation module to respectively perform segmentation processing on the multiple fourth point cloud data points to obtain multiple fifth point cloud data points, and the number of the fifth point cloud data points is the same as the number of point cloud data in the test data; the sixth input unit is used to input the multiple fourth point cloud data points into the classification module to respectively perform classification processing on the multiple fourth point cloud data points to obtain multiple sixth point cloud data points, and the multiple fifth point cloud data points and the multiple sixth point cloud data points form the first point cloud data, the second point cloud data, the third point cloud data and the fourth point cloud data.

[0107] In an optional embodiment, the first input unit includes: a first input subunit, which is used to input the test data into the downsampling layer VFS to perform voxel farthest point downsampling processing on the test data to obtain a first sampling point set, and the first sampling point set includes multiple first sampling points; a second input subunit, which is used to input the test data and the first sampling point set into the weighted mean grouping layer PAG to perform grouping processing with the first sampling point as the centroid point to obtain a second sampling point set, and the second sampling point set is a local feature point set corresponding to each first sampling point, and the second sampling point set includes multiple second sampling points; a third input subunit, which is used to input the second sampling point set into the local feature point set input point mesh layer to perform feature extraction processing on the multiple second sampling points respectively, and obtain multiple first point cloud data points; a fourth input subunit, which is used to input the multiple second point cloud data points into the second local feature extraction submodule The VPP performs a second downsampling and a second feature extraction process on the plurality of second point cloud data points to obtain a plurality of third point cloud data points, including: a fifth input subunit, which is used to sequentially input the plurality of second point cloud data points into the downsampling layer to perform a voxel farthest point downsampling process on the second point cloud data points to obtain a third sampling point set, wherein the third sampling point set includes a plurality of third sampling points; a sixth input subunit, which is used to input the plurality of second point cloud data points and the third sampling point set into the weighted mean grouping layer to perform grouping process with the third sampling point as the centroid point to obtain a fourth sampling point set, wherein the fourth sampling point set is a local feature point set corresponding to each third sampling point, and the fourth sampling point set includes a plurality of fourth sampling points; and a seventh input subunit, which is used to input the fourth sampling point set into the local feature point set input point mesh layer to perform feature extraction process on the plurality of fourth sampling points respectively to obtain a plurality of third point cloud data points.

[0108] In an optional embodiment, the second input unit includes: an eighth input sub-unit, which is used to input multiple first point cloud data points into the activation function layer in sequence to respectively calculate the weight of the feature dimension of each first point cloud data point, and obtain multiple first weight values; a ninth input sub-unit, which is used to input multiple first weight values ​​and corresponding first point cloud data points into the calculation layer to respectively perform point multiplication processing on the first weight values ​​and the corresponding first point cloud data points, and obtain multiple second point cloud data points; a tenth input sub-unit, which is used to input multiple third point cloud data points into the second channel attention sub-module in sequence to perform second feature dimension weighted processing on multiple third point cloud data points, and obtain multiple fourth point cloud data points, including: inputting multiple third point cloud data points into the activation function layer in sequence to respectively calculate the weight of the feature dimension of each third point cloud data point, and obtain multiple second weight values; an eleventh input sub-unit, which is used to input multiple second weight values ​​and corresponding third point cloud data points into the calculation layer to respectively perform point multiplication processing on the second weight values ​​and the corresponding third point cloud data points, and obtain multiple fourth point cloud data points.

[0109] In an optional embodiment, the fifth input unit includes: a twelfth input subunit, which is used to input multiple second point cloud data points and multiple fourth point cloud data points into the first interpolation layer, using the second point cloud data points as the points to be interpolated and the fourth point cloud data points as the neighboring feature points to perform a first interpolation process on the second point cloud data points and the fourth point cloud data points to obtain a first data point set; a thirteenth input subunit, which is used to input the first data point set into the first convolution layer to perform feature extraction process on the first data point set to obtain a second data point set, wherein the second data point set includes multiple second data points; a fourteenth input subunit, which is used to input the second data point set and test data into the second interpolation layer, using the test data as the points to be interpolated and the second data points as the neighboring feature points to perform a second interpolation process on the second data point set and the test data to obtain a third data point set, wherein the third data point set includes multiple third data points; a fifteenth input subunit, which is used to input the third data point set into the second convolution layer to perform feature extraction process on the multiple third data points to obtain multiple fifth point cloud data points.

[0110] In an optional embodiment, the sixth input unit includes: a sixteenth input subunit, used to input the multiple fourth point cloud data points into the third convolutional layer in sequence to perform feature extraction processing on the multiple fourth point cloud data points respectively, to obtain multiple seventh point cloud data points; a seventeenth input subunit, used to input the multiple seventh point cloud data points into the fully connected layer to perform classification processing on the seventh point cloud data points, to obtain multiple sixth point cloud data points.

[0111] In an optional embodiment, the second processing module includes: a calculation unit for calculating according to the formula Calculate the angle α between the conductive arm corresponding to the i-th frame point cloud image and the conductive arm in the closed state i , where u i =((x0,y0),(x i ,y i )), (x0, y0) is the coordinate of the end of the conductive arm fixed to the stationary insulating support, (x i ,y i ) is the coordinate corresponding to the end of the conductive arm rotating around the insulating support, v = ((x0, y0), (x v ,y v )),(x v ,y v ) is the coordinate corresponding to one end of the conducting arm rotating around the insulating support in the closed state, u i v is the vector u i The dot product of the vector v, |u i | is the vector u i The modulus of |v| is the modulus of vector v; the first determination unit is used to determine α i Is it greater than α i+n, in α i Greater than α i+n In the case of i+n The corresponding disconnector action state is the closing process, where α i+n is the angle between the conductive arm corresponding to the i+nth frame point cloud image and the conductive arm in the closed state, n is greater than 10; the second determining unit is used to determine the angle between the conductive arm and the conductive arm in the closed state in α i Less than α i+n In the case of i The corresponding isolating switch is in the disconnecting process; the third determining unit is used to determine the i Equal to α i+n In the case of i and α i+n Whether α is satisfied i+n =α i =0, at α i+n =α i = 0, determine α i The corresponding isolating switch is in the closed state; the fourth determining unit is used to determine the state of the isolating switch in α i+n =α i ≠0, determine α i The corresponding isolating switch is in the open state.

[0112] According to another aspect of an embodiment of the present invention, a computer program product is provided, comprising computer instructions, and when the computer instructions are executed by a processor, any one of the above methods for detecting the operating state of an isolating switch is executed. According to another aspect of an embodiment of the present invention, a processor is provided, and the processor is used to run a program, wherein when the program is run, any one of the above methods for detecting the operating state of an isolating switch is executed.

[0113] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned methods for detecting the action status of the isolating switch.

[0114] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for detecting the operating state of an isolating switch, characterized in that: include: Acquire a point cloud data set, the point cloud data set comprising a plurality of point cloud data corresponding to the disconnector, each of the point cloud data comprising a plurality of frames of point cloud images obtained by scanning the disconnector to be detected by a laser radar, each of the point cloud images comprising a plurality of point cloud data points, part of the point cloud data set being test data and the other part being training data, the disconnector comprising at least a conductive arm, an insulating support, a transmission cable and a base; Constructing an initial detection model, and using a target loss function to train the initial detection model to obtain a target detection model, wherein the target detection model includes a feature extraction module, a classification module and a segmentation module, the input end of the feature extraction module is the input end of the target detection model, the output end of the feature extraction module is respectively connected to the input end of the classification module and the input end of the segmentation module, the output end of the classification module and the output end of the segmentation module are the output end of the target detection model, the feature extraction module includes a stacked local feature extraction submodule and an attention submodule, the feature extraction module is used to extract point cloud features of the input data of the target detection model, the classification module is used to classify the output data of the feature extraction module, and the segmentation module is used to segment the output data of the feature extraction module; Input the test data into the target detection model to obtain first point cloud data corresponding to the conductive arm of the disconnector corresponding to the test data, second point cloud data corresponding to the insulating support of the disconnector corresponding to the test data, third point cloud data corresponding to the transmission cable of the disconnector corresponding to the test data, and fourth point cloud data corresponding to the base of the disconnector corresponding to the test data; Obtain a vector corresponding to the first point cloud data, and determine an action state of the isolating switch at least based on the vector corresponding to the first point cloud data, the action state comprising an open state, a closed state, a disconnection process, and a closing process, the open state being the state of the isolating switch when no power is supplied, and the closed state being the state of the isolating switch when power is supplied.

2. The method according to claim 1, characterized in that Get the dataset, including: Scanning the isolating switch by laser radar to obtain a plurality of initial point cloud data; The plurality of initial point cloud data are respectively subjected to filtering and denoising processing to obtain the point cloud data set.

3. The method according to claim 1, characterized in that The feature extraction module includes a first local feature extraction submodule, a first channel attention submodule, a second local feature extraction submodule and a second channel attention submodule connected in sequence, inputs the test data into the target detection model, obtains first point cloud data corresponding to the conductive arm of the disconnector corresponding to the test data, second point cloud data corresponding to the insulating support of the disconnector corresponding to the test data, third point cloud data corresponding to the transmission cable of the disconnector corresponding to the test data, and fourth point cloud data corresponding to the base of the disconnector corresponding to the test data, including: Inputting the test data into the first local feature extraction submodule to perform first downsampling and first feature extraction processing on the test data to obtain a plurality of first point cloud data points; Inputting the plurality of first point cloud data points into the first channel attention submodule in sequence to perform first feature dimension weighted processing on the plurality of first point cloud data points to obtain a plurality of second point cloud data points; Inputting the plurality of the second point cloud data points into the second local feature extraction submodule to perform second downsampling and second feature extraction processing on the plurality of the second point cloud data points to obtain a plurality of third point cloud data points; Inputting the plurality of the third point cloud data points into the second channel attention submodule in sequence to perform second feature dimension weighted processing on the plurality of the third point cloud data points to obtain a plurality of fourth point cloud data points; Inputting the plurality of test data, the plurality of the second point cloud data points and the plurality of the fourth point cloud data points into the segmentation module to respectively segment the plurality of the fourth point cloud data points to obtain a plurality of fifth point cloud data points, wherein the number of the fifth point cloud data points is the same as the number of the point cloud data in the test data; The multiple fourth point cloud data points are input into the classification module to classify the multiple fourth point cloud data points respectively to obtain multiple sixth point cloud data points. The multiple fifth point cloud data points and the multiple sixth point cloud data points form the first point cloud data, the second point cloud data, the third point cloud data and the fourth point cloud data.

4. The method according to claim 3, characterized in that The first local feature extraction submodule is the same as the second local feature extraction submodule, and the first local feature extraction submodule and the second local feature extraction submodule respectively include a downsampling layer, a weighted mean grouping layer, and a local feature point set input point network layer. The test data is input into the first local feature extraction submodule to perform a first downsampling and a first feature extraction process on the test data to obtain a plurality of first point cloud data points, including: Input the test data into the downsampling layer to perform voxel farthest point downsampling processing on the test data to obtain a first sampling point set, wherein the first sampling point set includes a plurality of first sampling points; Input the test data and the first sampling point set into the weighted mean grouping layer, and perform grouping processing with the first sampling point as the centroid point to obtain a second sampling point set, wherein the second sampling point set is a local feature point set corresponding to each of the first sampling points, and the second sampling point set includes a plurality of second sampling points; Inputting the second sampling point set into the local feature point set input point mesh layer to perform feature extraction processing on a plurality of the second sampling points respectively, to obtain a plurality of the first point cloud data points; Inputting the plurality of the second point cloud data points into the second local feature extraction submodule to perform second downsampling and second feature extraction processing on the plurality of the second point cloud data points to obtain a plurality of third point cloud data points, including: Inputting the plurality of the second point cloud data points into the downsampling layer in sequence to perform voxel farthest point downsampling processing on the second point cloud data points to obtain a third sampling point set, wherein the third sampling point set includes a plurality of third sampling points; Inputting a plurality of the second point cloud data points and the third sampling point set into the weighted mean grouping layer to perform grouping processing with the third sampling point as the centroid point, to obtain a fourth sampling point set, wherein the fourth sampling point set is a local feature point set corresponding to each of the third sampling points, and the fourth sampling point set includes a plurality of fourth sampling points; The fourth sampling point set is input into the local feature point set input point mesh layer to perform feature extraction processing on the plurality of fourth sampling points respectively, so as to obtain the plurality of third point cloud data points.

5. The method according to claim 3, characterized in that: The first channel attention submodule is the same as the second channel attention submodule, and the first channel attention submodule and the second channel attention submodule respectively include an activation function layer and a calculation layer, and the plurality of first point cloud data points are sequentially input into the first channel attention submodule to perform first feature dimension weighted processing on the plurality of first point cloud data points to obtain a plurality of second point cloud data points, including: Inputting the plurality of first point cloud data points into the activation function layer in sequence to respectively calculate the weight of the feature dimension of each of the first point cloud data points to obtain a plurality of first weight values; Inputting a plurality of the first weight values ​​and the corresponding first point cloud data points into the calculation layer to perform point multiplication processing on the first weight values ​​and the corresponding first point cloud data points respectively, to obtain a plurality of the second point cloud data points; Inputting the plurality of the third point cloud data points sequentially into the second channel attention submodule to perform second feature dimension weighted processing on the plurality of the third point cloud data points to obtain a plurality of fourth point cloud data points, including: inputting the plurality of the third point cloud data points sequentially into the activation function layer to respectively calculate the weight of the feature dimension of each of the third point cloud data points to obtain a plurality of second weight values; Input a plurality of the second weight values ​​and the corresponding third point cloud data points into the calculation layer to perform point multiplication processing on the second weight values ​​and the corresponding third point cloud data points respectively, so as to obtain a plurality of the fourth point cloud data points.

6. The method according to claim 3, characterized in that The segmentation module includes a first interpolation layer, a first convolution layer, a second interpolation layer and a second convolution layer connected in sequence, and the plurality of test data, the plurality of the second point cloud data points and the plurality of the fourth point cloud data points are input into the segmentation module to respectively segment the plurality of the fourth point cloud data points to obtain a plurality of fifth point cloud data points, including: Inputting a plurality of the second point cloud data points and a plurality of the fourth point cloud data points into the first interpolation layer, taking the second point cloud data points as the points to be interpolated and the fourth point cloud data points as the neighboring feature points, performing a first interpolation process on the second point cloud data points and the fourth point cloud data points, to obtain a first data point set; Inputting the first data point set into the first convolutional layer to perform feature extraction processing on the first data point set to obtain a second data point set, where the second data point set includes a plurality of second data points; Input the second data point set and the test data into the second interpolation layer, use the test data as the to-be-interpolated point and the second data point as the neighboring feature point, perform a second interpolation process on the second data point set and the test data, and obtain a third data point set, wherein the third data point set includes a plurality of third data points; The third data point set is input into the second convolutional layer to perform feature extraction processing on the plurality of third data points to obtain a plurality of fifth point cloud data points.

7. The method according to claim 3, characterized in that The classification module includes a third convolutional layer and a fully connected layer connected in sequence, and the plurality of fourth point cloud data points are input into the classification module to classify the plurality of fourth point cloud data points respectively to obtain a plurality of sixth point cloud data points, including: sequentially inputting the plurality of fourth point cloud data points into the third convolutional layer to perform feature extraction processing on the plurality of fourth point cloud data points respectively, to obtain a plurality of seventh point cloud data points; Input the plurality of the seventh point cloud data points into the fully connected layer to perform classification processing on the seventh point cloud data points to obtain the plurality of the sixth point cloud data points.

8. The method according to claim 1, characterized in that Determining the operating state of the disconnector at least according to the vector corresponding to the first point cloud data includes: According to the formula Calculate the angle α between the conductive arm corresponding to the point cloud image of the i-th frame and the conductive arm in the closed state i , where u i =((x0,y0),(x i ,y i )), (x0, y0) is the coordinate corresponding to the end of the conductive arm fixed to the stationary insulating support, (x i ,y i ) is the coordinate corresponding to one end of the conductive arm rotating around the insulating support, v = ((x0, y0), (x v ,y v )),(x v ,y v ) is the coordinate corresponding to one end of the conductive arm rotating around the insulating support in the closed state, u i v is the vector u i The dot product of the vector v, |u i | is the vector u i The modulus of |v| is the modulus of vector v; Determine α i Is it greater than α i+n , in α i Greater than α i+n In the case of i+n The corresponding disconnector action state is the closing process, where α i+n is the angle between the conductive arm corresponding to the point cloud image of the i+nth frame and the conductive arm in the closed state, and n is greater than 10; In α i Less than α i+n In the case of i The corresponding action state of the isolating switch is the disconnection process; In α i Equal to α i+n In the case of i and α i+n Whether α is satisfied i+n =α i =0, at α i+n =α i = 0, determine α i The corresponding operating state of the isolating switch is the closed state; In α i+n =α i ≠0, determine α i The corresponding operating state of the isolating switch is the open state.

9. A detection device for the operating state of an isolating switch, characterized in that: include: an acquisition module, for acquiring a point cloud data set, wherein the point cloud data set includes a plurality of point cloud data corresponding to the disconnector, each of the point cloud data includes a plurality of frames of point cloud images obtained by scanning the disconnector to be detected by a laser radar, each of the point cloud images includes a plurality of point cloud data points, a part of the point cloud data set is test data, and another part is training data, and the disconnector includes at least a conductive arm, an insulating support, a transmission cable and a base; A first processing module is used to construct an initial detection model and train the initial detection model using a target loss function to obtain a target detection model, wherein the target detection model includes a feature extraction module, a classification module and a segmentation module, the input end of the feature extraction module is the input end of the target detection model, the output end of the feature extraction module is connected to the input end of the classification module and the input end of the segmentation module respectively, the output end of the classification module and the output end of the segmentation module are the output end of the target detection model, the feature extraction module includes a stacked local feature extraction submodule and an attention submodule, the feature extraction module is used to extract point cloud features of the input data of the target detection model, the classification module is used to classify the output data of the feature extraction module, and the segmentation module is used to segment the output data of the feature extraction module; An input module, used to input the test data into the target detection model, to obtain first point cloud data corresponding to the conductive arm of the disconnector corresponding to the test data, second point cloud data corresponding to the insulating support of the disconnector corresponding to the test data, third point cloud data corresponding to the transmission cable of the disconnector corresponding to the test data, and fourth point cloud data corresponding to the base of the disconnector corresponding to the test data; The second processing module is used to obtain the vector corresponding to the first point cloud data, and determine the action state of the isolating switch at least based on the vector corresponding to the first point cloud data, the action state includes an open state, a closed state, a disconnection process and a closing process, the open state is the state of the isolating switch when it is not powered, and the closed state is the state of the isolating switch when it is powered.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein the program executes the method for detecting the operating state of the disconnector according to any one of claims 1 to 8.

Images