Converter valve assembly hidden danger automatic detection method and system based on three-dimensional vision

Abstract

The invention discloses a converter valve assembly hidden danger automatic detection method and system based on three-dimensional vision, and belongs to the technical field of direct current transmission equipment detection. The method comprises the following steps: acquiring three-dimensional point cloud and two-dimensional image data of a converter valve assembly through a non-contact three-dimensional visual sensor; preprocessing the point cloud, including region-of-interest cutting, filtering denoising and downsampling in the depth direction of a sensor measurement coordinate system; segmenting the point cloud based on a clustering algorithm to obtain a candidate component set; key parts are automatically identified by using specific three-dimensional shape features of the wires, the capacitors and the connecting strips; and calculating the distance according to the three-dimensional space coordinates of the identification parts, and comparing the distance with a safety threshold value to judge the potential discharge hazard. The system correspondingly comprises a data acquisition module, a data processing module and a hidden danger judgment module. According to the invention, non-contact, automatic and high-precision detection of hidden dangers of the converter valve assembly is realized, and the defects of strong subjectivity, low efficiency, poor safety and the like of manual visual inspection are effectively overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of converter station DC power transmission equipment detection, and particularly relates to a three-dimensional vision-based automatic detection method and system for hidden dangers of a converter valve assembly. BACKGROUND

[0002] As a core and key device in a high-voltage direct current power transmission system, the operation state of a converter valve directly relates to the safety and stability of the entire power grid. The converter valve is internally composed of thyristors, damping capacitors, voltage-sharing capacitors, resistors, reactors and other precise electrical components, and is electrically connected through a large number of wires and connecting strips. These components are subjected to the combined stress of electricity, heat and mechanical vibration during long-term operation, which may cause hidden dangers such as loose connection of connecting wires, displacement, insulation aging, and poor crimping of potential wires such as voltage-sharing electrodes. If these problems are not discovered in time, they can easily cause partial discharge or even short circuit faults, causing huge economic losses and social impact.

[0003] Currently, routine inspection and hidden danger screening of the converter valve assembly mainly rely on manual visual inspection by maintenance personnel entering the valve hall during annual maintenance. The traditional method has the following limitations:

[0004] Firstly, the detection accuracy is low, subjective and prone to missed detection. Manual visual inspection is difficult to accurately quantify the measurement of key parameters such as wire spacing and component displacement distance, especially for parts with complex spatial structure and strong concealment. The subjective experience, fatigue level and individual differences of the maintenance personnel directly affect the accuracy and consistency of the inspection results, and are prone to miss and misdiagnose subtle hidden dangers.

[0005] Secondly, the contradiction between low detection efficiency and tight maintenance time window is prominent. The annual maintenance time window of the converter station is usually very short, only 5 to 7 working days. During this period, maintenance personnel need to complete comprehensive screening of multiple valve halls and tens of thousands of components, with a huge workload. The traditional manual method is slow and cannot achieve rapid and comprehensive coverage, which forms a sharp contradiction with the limited maintenance time.

[0006] Finally, there is a safety risk and may introduce secondary hidden dangers. The frequent entry of maintenance personnel into the valve tower for inspection is itself a high-risk operation. At the same time, the personnel's entry and operation have the possibility of touching the precise components or connecting parts, which may in turn cause new looseness or damage, i.e. "secondary hidden dangers".

[0007] In the prior art, there are also some automatic detection methods based on three-dimensional point cloud. For example, the comparison file CN119989013A discloses a fuse rule component extraction and cluster growth-based conductor distance calculation method, which adopts a voxel-based Transformer network model for point cloud classification and extracts substation facility point clouds connected with conductors through a cluster growth algorithm. However, this method is mainly aimed at long-distance power transmission conductors in the substation scene, and has insufficient adaptability to scenes with complex internal structure of the converter valve, dense components, and more delicate spatial relationships. In particular, it relies on a complex deep learning model, has a large amount of calculation, and has limited recognition accuracy for components with specific shape characteristics such as conductors, capacitors, and connecting strips, which is difficult to meet the detection requirements of millimeter-level precision of the converter valve assembly.

[0008] Therefore, there is an urgent need for a method and system capable of realizing non-contact, automation, and high-precision detection of converter valve assemblies according to the characteristics of the converter valve assemblies to overcome the shortcomings of the prior art. SUMMARY

[0009] In view of the problems existing in the prior art, the present application provides a non-contact, automated, and high-precision converter valve assembly hidden danger automatic detection method and system to overcome the inherent defects of the existing detection methods based on manual visual inspection in terms of objectivity, efficiency, and safety.

[0010] The present application provides a three-dimensional vision-based converter valve assembly hidden danger automatic detection method, comprising the following steps:

[0011] S1, acquiring three-dimensional point cloud data and two-dimensional image data of a converter valve assembly to be detected by a non-contact three-dimensional vision sensor;

[0012] S2, preprocessing the three-dimensional point cloud data, the preprocessing at least including filtering and denoising and downsampling to obtain clear point cloud data;

[0013] S3, segmenting the preprocessed point cloud data based on a clustering algorithm to obtain a plurality of point cloud cluster sets corresponding to different candidate components;

[0014] S4, identifying key components in the converter valve assembly based on three-dimensional shape features of each point cloud cluster set, the key components at least including conductors, capacitors, and connecting strips; wherein the three-dimensional shape features include an elongated path feature of the conductor, a circular ring feature of the top end of the capacitor, and a planar strip feature of the connecting strip;

[0015] S5, calculating the spatial distance between the conductor and the surrounding capacitors and connecting strips according to the three-dimensional spatial coordinates of the point cloud cluster set corresponding to the identified key components;

[0016] S6. Compare the calculated spatial distance between the wire, capacitor, and connecting strip with a preset safe distance threshold. If the spatial distance is less than the safe distance threshold, it is determined that there is a potential discharge hazard.

[0017] Furthermore, in step S1, the non-contact 3D vision sensor is a 3D structured light camera, and the data collected includes: two-dimensional grayscale images, three-dimensional point cloud coordinates, point cloud reflection intensity information, and texture normal information.

[0018] Furthermore, the preprocessing in step S2 also includes performing the following before the filtering and denoising: setting a three-dimensional spatial threshold along the depth direction in the measurement coordinate system of the non-contact three-dimensional vision sensor to crop the region of interest of the original point cloud in order to filter out background interference and reduce the data range for subsequent processing.

[0019] Further, the downsampling in step S2 specifically involves: downsampling the filtered point cloud based on the octree structure, setting the octree depth dimension, calculating the centroid coordinates of the point cloud within each octree region as the voxel representative of that region, and filtering out regions with fewer than a preset number of points.

[0020] Step S3 specifically involves: using a radius search algorithm based on the octree to calculate the spacing between points in the point cloud, merging points with a spacing less than a preset distance threshold into the same connected component, and assigning different labels to different connected components, thereby completing the point cloud clustering and segmentation.

[0021] Further, step S4 is performed based on a pre-trained mathematical classification model; the mathematical classification model is obtained through the following training steps:

[0022] S41. After performing the preprocessing S2 and clustering segmentation S3 on the sample point cloud data collected from multiple angles, the obtained point cloud cluster sets are manually labeled to form positive and negative samples.

[0023] S42. Extract the three-dimensional shape features from the labeled sample point cloud data;

[0024] S43. Train a nonlinear mathematical classification model using labeled sample data;

[0025] Furthermore, in step S5, the three-dimensional spatial coordinates used to calculate the spatial distance are determined specifically in the following manner:

[0026] Based on the aforementioned three-dimensional shape features, the representative spatial geometric parameters of each key component are determined;

[0027] The representative spatial geometric parameters of the conductor are the coordinates of a set of points on its center line, the representative spatial geometric parameters of the capacitor are the coordinates of the center of its top ring, and the representative spatial geometric parameters of the connecting strip are the coordinates of a point on its surface center line.

[0028] Furthermore, the method also includes:

[0029] S7. The two-dimensional image data is segmented at the pixel level using a semantic segmentation model based on deep learning to identify the two-dimensional pixel regions of the key components; wherein, the semantic segmentation model is a DeepLabv3+ network model.

[0030] S8. The two-dimensional pixel region is fused with the three-dimensional point cloud data to assist and verify the recognition results of the key components in three-dimensional space.

[0031] Furthermore, after identifying the conductor in step S4, the method also includes completing the point cloud of the conductor. Specifically, using the fixed diameter of the conductor as prior knowledge, the method performs cross-sectional circle fitting on the point cloud of the conductor that is missing due to occlusion, so as to complete the missing point cloud data and thus accurately calculate the spatial position and center coordinates of the conductor.

[0032] Furthermore, the preprocessing in step S2 also includes wire point cloud extraction based on the two-dimensional image, specifically:

[0033] Read the two-dimensional grayscale image of the conductor and obtain the pixel coordinates of the conductor;

[0034] Calculate the 3D point cloud data index corresponding to the coordinates;

[0035] The initial point cloud of the traverse is extracted from the original 3D point cloud data according to the index, for subsequent processing in steps S2 to S6.

[0036] Furthermore, after extracting the traverse point cloud and filtering out stray point clouds, the process also includes traverse segmentation and segmentation-based calculation steps, specifically:

[0037] The identified guide point clouds are sorted by coordinate dimensions, and the beginning and end of the lines are removed.

[0038] Divide the point cloud data in the middle part into multiple segments;

[0039] Calculate the coordinates of the center point of each segment, and perform a radius search within the defined cylindrical area, using the center point of each segment as the center.

[0040] When the distance between the searched point and the center point is greater than the preset power line width threshold, the point is determined to be an obstacle point;

[0041] Density analysis is performed on the identified obstacle point cloud to locate the area with the highest density of obstacles;

[0042] Calculate the shortest distance between the densest point of obstacles and the power line, and compare it with a safety threshold. If the distance is less than the safety threshold, an alarm is triggered.

[0043] This invention also provides an automatic detection system for potential hazards in converter valve components based on three-dimensional vision, used to implement the above-described methods, comprising:

[0044] The data acquisition module includes a 3D structured light camera for performing step S1;

[0045] A data processing module, comprising at least one processor and a memory, wherein the memory stores a computer program, which, when executed by the processor, is used to implement steps S2 to S5;

[0046] The hazard assessment module consists of at least one processor and a memory. When the program is executed by the processor, it is used to implement step S6.

[0047] The beneficial effects of this invention are as follows:

[0048] (1) The detection process is made more objective and precise, effectively eliminating missed and incorrect detections. This invention completely replaces manual visual judgment that relies on the subjective experience of maintenance personnel by analyzing three-dimensional point cloud data and automatically identifying based on shape features. The system can perform precise and quantitative measurements of key parameters such as wire spacing and component offset distance at the millimeter level, eliminating the influence of individual differences and fatigue factors on the detection results, and fundamentally solving the problem of inconsistent standards and easy missed and incorrect detections in manual inspection.

[0049] (2) It greatly improves the detection efficiency and meets the stringent requirements of the maintenance window. Through automated three-dimensional data acquisition, point cloud processing and recognition algorithms, this invention achieves non-contact rapid detection. The detection time of a single component is controlled at the minute level, so that the comprehensive detection of the entire valve hall can be completed within half a working day. This highly efficient detection mode perfectly meets the actual needs of the converter station with short annual maintenance time and heavy tasks.

[0050] (3) Secondary safety hazards are avoided, and the safety of equipment and personnel is guaranteed. As a completely non-contact detection method is adopted, maintenance personnel do not need to frequently enter the valve tower. This not only eliminates secondary hazards such as component contact and loosening that may be caused by personnel entering, but also reduces the safety risks of maintenance personnel working in a high-voltage environment, and significantly improves the safety of the detection process.

[0051] (4) Improved the comprehensiveness and intelligence of detection. This invention comprehensively utilizes the fusion analysis of three-dimensional point clouds and two-dimensional images, combined with point cloud completion based on prior knowledge and wire segmentation calculation, effectively solving problems such as missing points and occlusion, and improving the detection accuracy in complex converter valve environments; compared with existing general methods based on deep learning, this invention optimizes the specific shape features of converter valve components (slender wires, circular capacitors, long strip connectors), resulting in higher recognition accuracy, better computational efficiency, and improved robustness and reliability of the entire system in complex field environments. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the cloud processing process; among which, Figure 1 (a) is a two-dimensional grayscale image of the measured area. Figure 1 (b) is a point cloud diagram. Figure 1 (c) is the texture normal map along the X-axis. Figure 1 (d) is the texture normal map along the Y-axis. Figure 1 (e) is the texture normal map along the Z-axis. Figure 1 (f) is the preprocessed point cloud map.

[0053] Figure 2 This is a schematic diagram of the region of interest (ROI) extraction.

[0054] Figure 3 This is a schematic diagram of point cloud data filtering and preprocessing.

[0055] Figure 4 This is a schematic diagram of feature classification.

[0056] Figure 5 This is a diagram of the point cloud effect extracted after filtering out stray point clouds.

[0057] Figure 6 This is a screenshot of the obstacle search effect.

[0058] Figure 7 This is a schematic diagram of the obstacle distance calculation results.

[0059] Figure 8 This is a schematic diagram of the system detection results. Detailed Implementation

[0060] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0061] This embodiment focuses on inspecting the converter valve assembly in the pole I valve hall of a converter station. This assembly includes key components such as wires, damping capacitors, and connecting strips. This embodiment provides an automatic detection system for potential problems in converter valve assemblies based on 3D vision. Its hardware mainly includes: a high-precision 3D structured light camera, precision servo motion control lead screw modules in both the X and Y axes, an industrial computer equipped with a high-performance GPU (as a processing and control unit), and a stable mounting bracket.

[0062] The implementation steps of the detection method are as follows:

[0063] S1. Data Acquisition:

[0064] The 3D structured light camera is mounted on the XY motion module using a mounting bracket, and its pose is adjusted to face the converter valve assembly to be inspected. After system startup, the control module moves the camera along a predetermined path to scan the assembly region by region. At each acquisition point, the camera simultaneously acquires a 2D grayscale image, 3D point cloud coordinates, point cloud reflection intensity information, and texture normal information of the measured area. After scanning is complete, all data is uploaded to an industrial computer. This step provides a multimodal data foundation for subsequent fusion analysis.

[0065] S2, Point Cloud Preprocessing:

[0066] The industrial computer preprocesses the acquired raw 3D point cloud data to remove noise, simplify the data, and extract key regions. The inputs, outputs, and intermediate processes of this step can be found in [link to relevant documentation]. Figure 1 The following is an example of a set of data.

[0067] First, the system acquires a two-dimensional grayscale image of the measured area, such as... Figure 1 As shown in (a) above. This image not only provides the basis for subsequent possible two-dimensional image analysis, but its texture information can also be used to assist in the analysis of three-dimensional point clouds.

[0068] The acquired raw 3D point cloud data contains reflection intensity information, such as... Figure 1 As shown in (b) in the figure, the shades of color represent the intensity of the point cloud reflection.

[0069] As a preferred embodiment, the method further includes a wire point cloud extraction step based on a two-dimensional image:

[0070] Read the 2D grayscale image of the conductor, obtain the pixel coordinates of the conductor, and calculate the 3D point cloud data index corresponding to the coordinates; the index is calculated by the formula X-1+Y*W, where X and Y are pixel coordinates and W is the image width; extract the initial point cloud of the conductor from the original 3D point cloud data according to the index, which can significantly reduce the data range for subsequent processing.

[0071] Data preprocessing operations: Subsequently, operations S21 (ROI cropping), S22 (filtering and denoising), and S23 (downsampling) are performed on the point cloud. The specific preprocessing operations in this embodiment are as follows:

[0072] S21. Detection Region Extraction (ROI Cropping): For example... Figure 2 As shown, in the measurement coordinate system of the 3D structured light camera, a three-dimensional spatial threshold is set along its depth direction (Z-axis). In this embodiment, thresholds δ1 (1, -3, -1.5) and δ2 (5, 3, 2) are set, and the point cloud coordinates P are... i (X i ,Y i Z i Points that do not meet this spatial range are filtered out. After processing, a point cloud of the region of interest with a spatial size of approximately 500×400×400 mm³ is obtained, effectively filtering out background interference;

[0073] S22, Filtering and noise reduction: such as Figure 3 As shown, a discrete filter is used to remove sparse outlier noise points. The specific operations include: analyzing the neighborhood of each point using statistical methods, mainly by setting a specified threshold, then removing points that do not meet the threshold; calculating the spatial neighborhood U distribution of each point and its neighbors, and then calculating the center point u and variance σ in the neighborhood U, where:

[0074]

[0075]

[0076] A standard range is defined using the variance of the distances to all points, thereby filtering out point clouds that do not meet this standard range. If the data represents the point cloud data before removal, then the data represents the point cloud data after removal. The judgment criteria are:

[0077]

[0078] Where T represents the set variance neighborhood threshold, the size of which depends on the number of points in the neighborhood. The KNN (K-Nearest Neighbor) search algorithm is used for point cloud neighborhood search; therefore, the main threshold parameters involved in the discrete filter are the number of KNN nearest neighbors K0 and the variance neighborhood coefficient T0. The radius filter describes a filtering method based on nearest neighbor radius search, mainly using point P... i (X i ,Y i Z iP is a circle with a specified search radius R0 as the center. A set number of elements within the circle, K1, is used as the criterion. If the number of elements within the search circle is less than K1, then P is removed. i By setting discrete filter parameters according to the actual scenario, discrete point filtering is completed to achieve a clearer point cloud;

[0079] S23. Downsampling: To reduce computational load, an octree structure is used to downsample the filtered point cloud. The octree depth dimension Θ = 0.015m is set, and the centroid coordinates of the point cloud within each octree region are calculated as the voxel representation of that region. Regions with fewer than 3 points are filtered out, thereby significantly reducing the number of point clouds while preserving contour features.

[0080] To demonstrate the geometric characteristics of point clouds, texture normal maps in the X, Y, and Z directions were generated based on the point cloud data, as shown below. Figure 1 As shown in (c), (d), and (e) of the diagram. These normal diagrams reflect the microscopic geometric orientation of an object's surface and are fundamental to calculating shape features. Specifically, from... Figure 1 The results of the texture normal map along the Z-axis shown in (e) verify that the camera's active structured light source is emitted along the Z-axis.

[0081] After the above series of preprocessing operations, a cleared preprocessed point cloud map that can be used for subsequent analysis is finally obtained, such as... Figure 1 As shown in (f) in the figure.

[0082] By comparison Figure 1 As can be seen from (b) and (f) in the figure, the preprocessing process effectively filtered out noise and obtained point cloud data with clear contours and obvious features, which laid a solid foundation for subsequent clustering, segmentation and recognition.

[0083] S3, Point Cloud Clustering and Segmentation:

[0084] The downsampled point cloud is segmented to separate different components. This embodiment uses an octree-based radius search clustering algorithm, setting the maximum interval distance D = 0.02m. The algorithm groups points in space with a distance less than this threshold into the same connected component and assigns a unique label to each connected component, thereby initially obtaining point cloud cluster sets corresponding to different components such as wires, capacitors, and connecting strips.

[0085] S4. Identification of key components:

[0086] like Figure 4 As shown, the specific component type is identified based on the three-dimensional shape features of each point cloud cluster set.

[0087] For a wire, its point cloud clusters are elongated. By searching its endpoints and extending the path within a set conical angle α, if the path length is greater than the threshold s, it is determined to be a wire.

[0088] At the top of the capacitor, its point cloud is circular. By fitting its outermost plane Π, the circular point cloud is projected onto this plane and fitted into a circle within the plane. The center of this circle is the coordinate of the capacitor.

[0089] For a connecting strip, its point cloud approximately converges to a plane, and the aspect ratio of the point cloud distribution in that plane is greater than a preset value, thus it is determined to be a connecting strip.

[0090] This step can be performed based on a pre-trained mathematical classification model M. The training method for this model is as follows:

[0091] S41. Collect a large number of sample point clouds. After processing through steps S2 and S3, manually label the clustering results to form a positive and negative sample dataset containing wires, capacitors, connecting strips and noise.

[0092] S42. Extract the aforementioned shape features from these data;

[0093] S43 is obtained by training using a nonlinear mathematical classification model (such as a support vector machine).

[0094] As a preferred embodiment, after identifying the wire in step S4, a wire segmentation step is also included:

[0095] like Figure 5 As shown, after extracting the guide point cloud, statistical filtering and the DBSCAN algorithm are used to filter out stray point clouds. Then, the identified guide point clouds are sorted by coordinate dimensions, and the beginning and end portions of the lines (such as the first 10% and the last 10%) are removed. The point cloud data in the middle is divided into three segments. The center point coordinates of each segment are calculated using the following formula:

[0096] ,

[0097] in Let i be the center point of the i-th segment. For the segment start point index, The length of the segment.

[0098] Simultaneously, after identifying the conductor, the process also includes step S4b of completing the point cloud of the conductor: using the fixed diameter of the conductor as prior knowledge, performing cross-sectional circle fitting on the point cloud of the conductor that is missing due to occlusion, in order to complete the missing point cloud data, thereby accurately calculating the spatial position and center coordinates of the conductor.

[0099] S5. Spatial distance calculation:

[0100] The spatial distance between components is calculated based on the three-dimensional spatial coordinates of the identified key components. In this embodiment, this is specifically achieved by determining the representative spatial geometric parameters of each key component: the representative spatial geometric parameters of the conductor are the coordinates of a set of points on its center line; the representative spatial geometric parameters of the capacitor are the coordinates of the center of its top ring; and the representative spatial geometric parameters of the connecting strip are the coordinates of points on its surface center line. By calculating the distances between these representative geometric parameters, the precise spatial distance between the components is obtained.

[0101] As another preferred embodiment, the method further includes a calculation step based on the segment center point:

[0102] A radius search is performed with the center point of each segment as the center, and the search range is limited to the cylindrical area of ​​the current segment to avoid affecting the beginning and end of the line; when the distance between the searched point and the center point is greater than the preset power line width threshold, the point is determined to be an obstacle point;

[0103] Radius search process:

[0104] (1) For each center point c, perform a search on all points p, where ||p - c|| <= R, and R is the search radius;

[0105] (2) For each point p that satisfies the above conditions, if ||p - c|| > d (where d is the distance threshold for judging obstacles), then p is considered an obstacle. The search results are as follows: Figure 6 As shown.

[0106] Density analysis is performed on the identified obstacle point cloud to locate the areas with the highest obstacle density; the shortest distance between the areas with the highest obstacle density and the power lines is calculated using the following formula:

[0107] ,

[0108] in( ) represents the coordinates of the densest point, ( ) represents the coordinates of the center point of the power line. The calculation results are as follows: Figure 7 As shown.

[0109] S6. Hazard Assessment:

[0110] The calculated spatial distance is compared with a preset safety distance threshold (e.g., 10mm as set according to design specifications). If the measured distance is less than the safety threshold, the system determines that there is a potential discharge hazard and issues an alarm in the software interface, prompting maintenance personnel to pay close attention. Figure 8 As shown.

[0111] As a preferred embodiment, the following steps are also included:

[0112] S7. Two-dimensional image-assisted analysis and verification:

[0113] Simultaneously, using the acquired 2D grayscale images, a semantic segmentation model based on DeepLabv3+ was employed for pixel-level segmentation. The Labelme tool was used to annotate the training images with polygons, generating JSON-formatted label files for model training. During inference, the pixel positions of the wires in the image could be calculated. In this embodiment, the average inference time was 5.42 seconds. Using different images for inference segmentation, the overall accuracy (aAcc) reached 98.56%, and the average accuracy (mAcc) reached 89.45%. These inference results fully validate the technical feasibility of using the DeepLabv3+ network for pixel-level wire segmentation. This 2D segmentation result was fused with 3D point cloud data to assist and verify the accuracy of the 3D recognition results in step S4.

[0114] S8, Conductor Point Cloud Completion:

[0115] To address the issue of missing point clouds due to occlusion of the conductor, after the conductor is identified in S4, the fixed known diameter of the conductor (e.g., 3mm) is used as prior knowledge to complete the missing point cloud on the conductor cross-section. That is, by fitting a complete cross-sectional circle, the coordinates of the center point of the conductor cross-section are corrected, so that it more accurately reflects the spatial position of the current cross-sectional circle of the conductor, thereby improving the accuracy of distance measurement in S5.

[0116] Through the above steps, this embodiment successfully achieves non-contact, automated, and high-precision testing of converter valve components. The system controls the testing time for a single component to within 5 minutes, and a comprehensive test of a single valve chamber can be completed within half a working day, effectively and reliably detecting potential discharge hazards such as excessively close spacing that are difficult to detect manually.

[0117] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept of the present invention, and all such modifications or additions should fall within the protection scope of the present invention.

Claims

1. An automatic detection method for potential hazards in converter valve components based on three-dimensional vision, characterized in that, Includes the following steps: S1. Acquire three-dimensional point cloud data and two-dimensional image data of the converter valve assembly to be tested using a non-contact three-dimensional vision sensor; S2. Preprocess the three-dimensional point cloud data, the preprocessing including at least filtering and denoising and downsampling, to obtain clearer point cloud data; S3. Based on the clustering algorithm, the preprocessed point cloud data is segmented to obtain multiple point cloud cluster sets corresponding to different candidate components; S4. Based on the three-dimensional shape features of each point cloud cluster set, identify the key components in the converter valve assembly. The key components include at least wires, capacitors, and connecting strips. The three-dimensional shape features include: the elongated path features of the wires, the circular ring features at the top of the capacitors, and the planar strip features of the connecting strips. S5. Based on the three-dimensional spatial coordinates of the point cloud cluster set corresponding to the identified key components, calculate the spatial distance between the wire and the surrounding capacitors and connecting strips, with the wire as the core. S6. Compare the calculated spatial distance between the wire, capacitor, and connecting strip with a preset safe distance threshold. If the spatial distance is less than the safe distance threshold, it is determined that there is a potential discharge hazard.

2. The method as described in claim 1, characterized in that, In step S1, the non-contact 3D vision sensor is a 3D structured light camera, and the data collected includes: two-dimensional grayscale images, three-dimensional point cloud coordinates, point cloud reflection intensity information, and texture normal information.

3. The method as described in claim 1, characterized in that, The preprocessing in step S2 also includes performing the following steps before the filtering and denoising: setting a three-dimensional spatial threshold along the depth direction in the measurement coordinate system of the non-contact three-dimensional vision sensor, and cropping the region of interest of the original point cloud to filter out background interference and reduce the data range for subsequent processing.

4. The method as described in claim 1 or 3, characterized in that, The downsampling in step S2 specifically involves: downsampling the filtered point cloud based on the octree structure, setting the octree depth size, calculating the centroid coordinates of the point cloud within each octree region as the voxel representation of that region, and filtering regions with fewer than a preset number of points. Step S3 specifically involves: using a radius search algorithm based on the octree to calculate the spacing between points in the point cloud, merging points with a spacing less than a preset distance threshold into the same connected component, and assigning different labels to different connected components, thereby completing the point cloud clustering and segmentation.

5. The method as described in claim 1, characterized in that, Step S4 is performed based on a pre-trained mathematical classification model; the mathematical classification model is obtained through the following training steps: S41. After performing the preprocessing S2 and clustering segmentation S3 on the sample point cloud data collected from multiple angles, the obtained point cloud cluster sets are manually labeled to form positive and negative samples. S42. Extract the three-dimensional shape features from the labeled sample point cloud data; S43. Train a nonlinear mathematical classification model using labeled sample data; Furthermore, in step S5, the three-dimensional spatial coordinates used to calculate the spatial distance are determined specifically in the following manner: Based on the aforementioned three-dimensional shape features, the representative spatial geometric parameters of each key component are determined; The representative spatial geometric parameters of the conductor are the coordinates of a set of points on its center line, the representative spatial geometric parameters of the capacitor are the coordinates of the center of its top ring, and the representative spatial geometric parameters of the connecting strip are the coordinates of a point on its surface center line.

6. The method as described in claim 1, characterized in that, The method further includes: S7. The two-dimensional image data is segmented at the pixel level using a semantic segmentation model based on deep learning to identify the two-dimensional pixel regions of the key components; wherein, the semantic segmentation model is a DeepLabv3+ network model. S8. The two-dimensional pixel region is fused with the three-dimensional point cloud data to assist and verify the recognition results of the key components in three-dimensional space.

7. The method as described in claim 1, characterized in that, After identifying the conductor in step S4, the method further includes completing the point cloud of the conductor. Specifically, using the fixed diameter of the conductor as prior knowledge, the point cloud of the conductor that is missing due to occlusion is fitted with a cross-sectional circle to complete the missing point cloud data, thereby accurately calculating the spatial position and center coordinates of the conductor.

8. The method as described in claim 1, characterized in that, The preprocessing in step S2 also includes wire point cloud extraction based on two-dimensional images, specifically: Read the two-dimensional grayscale image of the conductor and obtain the pixel coordinates of the conductor; Calculate the 3D point cloud data index corresponding to the coordinates; The initial point cloud of the traverse is extracted from the original 3D point cloud data according to the index, for subsequent processing in steps S2 to S6.

9. The method as described in claim 8, characterized in that, After extracting the traverse point cloud and filtering out stray point clouds, the process also includes traverse segmentation and segmentation-based calculation steps, specifically: The identified guide point clouds are sorted by coordinate dimensions, and the beginning and end of the lines are removed. Divide the point cloud data in the middle part into multiple segments; Calculate the coordinates of the center point of each segment, and perform a radius search within the defined cylindrical area, using the center point of each segment as the center. When the distance between the searched point and the center point is greater than the preset power line width threshold, the point is determined to be an obstacle point; Density analysis is performed on the identified obstacle point cloud to locate the area with the highest density of obstacles; Calculate the shortest distance between the densest point of obstacles and the power line, and compare it with a safety threshold. If the distance is less than the safety threshold, an alarm is triggered.

10. An automatic detection system for potential hazards in converter valve components based on three-dimensional vision, used to implement the method described in any one of claims 1-9, characterized in that, include: The data acquisition module includes a 3D structured light camera for performing step S1; A data processing module, comprising at least one processor and a memory, wherein the memory stores a computer program, which, when executed by the processor, is used to implement steps S2 to S5; The hazard assessment module consists of at least one processor and a memory. When the program is executed by the processor, it is used to implement step S6.

Images