Converter valve radiator cleaning path planning method based on regional feature recognition and neighborhood information

Abstract

The invention discloses a converter valve radiator cleaning path planning method based on regional feature recognition and neighborhood information. The invention relates to the technical field of converter valve radiator cleaning path planning, and aims to solve the problems of low cleaning efficiency and poor cleaning effect of traditional manual cleaning of a converter valve radiator. The method comprises the following steps: constructing a two-dimensional grid map containing obstacle and dirty area information, cleaning target points according to a dirty grade sequence, searching a feasible path by adopting an optimized RRT algorithm to obtain an initial path set, performing smooth optimization on the initial path, and performing collision detection and coverage verification on a final cleaning path. According to the method, by accurately identifying the pollution level and optimizing path planning, the path length is effectively shortened, the cleaning efficiency and safety are improved, and technical support is provided for automatic cleaning of the converter valve radiator.

Description

Technical Field

[0001] This invention relates to the field of converter valve radiator cleaning path planning technology, and in particular to a converter valve radiator cleaning path planning method based on regional feature recognition and neighborhood information. Background Technology

[0002] The converter valve cooling system is one of the most critical pieces of equipment in a high-voltage direct current (HVDC) transmission system, and its proper functioning directly impacts the safe operation of the entire system. The external converter valve cooler, as a vital auxiliary system within the cooling system, uses water spray or forced air convection to transfer heat from the converter valves carried in the internal circulation system to the external environment, ensuring continuous heat dissipation under complex operating conditions. However, external converter valve coolers are prone to accumulating dust and dirt, leading to decreased heat dissipation performance, increased cooling water temperature, and in severe cases, triggering alarms or even causing converter valve lock-up, threatening the safe operation of the HVDC transmission system.

[0003] Currently, cleaning of converter valve radiators mainly relies on manual water spraying, which suffers from limitations in cleaning angle, incomplete coverage, and low efficiency, making it difficult to meet the demand for efficient and stable cleaning. Existing research has proposed some improvements to the cleaning agents and industrial cleaning methods for the surface of converter valve external cooling radiators by changing the cleaning medium and cleaning methods, solving the problem that manual water spraying cannot completely remove surface dirt. In the actual cleaning process of converter valve external cooling radiators, although efficient cleaning methods are adopted, the louvers on the top of the radiator limit the space, preventing maintenance personnel from closely observing the cleaning effect on the radiator surface. Researchers have conducted in-depth research on image recognition of equipment surface dirt. For example, shadow recognition and segmentation technology is used for hyperspectral detection of dirt on insulator surfaces to reduce detection errors; a hybrid neural network insulator dirt level classification method is used to assess the severity of insulator dirt; and image processing technology based on dirt preprocessing reduces the interference of dirt on surface damage recognition. However, most existing research focuses on identifying surface contamination on porcelain insulators in transmission lines, with no reports on the identification and automated cleaning of surface contamination on converter valve external cooling radiators. Therefore, it is necessary to propose a cleaning path planning method for converter valve radiators based on regional feature recognition and neighborhood information to address the aforementioned issues. Summary of the Invention

[0004] The purpose of this invention is to provide a method for planning the cleaning path of converter valve radiators based on regional feature recognition and neighborhood information, so as to solve the problems of low cleaning efficiency and poor effect of traditional manual cleaning of converter valve radiators.

[0005] This invention provides a method for planning the cleaning path of a converter valve radiator based on regional feature recognition and neighborhood information, comprising the following steps:

[0006] Step 1: Construct a two-dimensional grid map of the converter valve radiator surface. The two-dimensional grid map contains obstacle information and contaminated area information. The obstacle information includes the location and size of the louver support column and the gap between the heat dissipation fins. The contaminated area information includes the location, area and contamination level of each contaminated area.

[0007] Step 2: Sort the cleaning target points of each dirty area according to the dirt level in the dirty area information;

[0008] Step 3: Using the optimized RRT algorithm, starting from the robot's initial position, and combining the obstacle information and dirty area information in the two-dimensional grid map, feasible paths for each of the cleaning target points are searched sequentially to obtain an initial path set;

[0009] Step 4: Perform smoothing optimization on the initial paths in the initial path set to obtain the final cleaning path;

[0010] Step 5: Perform collision detection and cleaning coverage verification on the final cleaning path. If there is a collision risk or coverage blind spot, return to step 2 to readjust the sorting of cleaning target points, and repeat steps 3 to 5 until the path meets the requirements.

[0011] Further, step one includes: obtaining CAD drawings of the converter valve radiator, extracting radiator structural information from the CAD drawings, the radiator structural information including the position of the louver support column, the distribution and size of the heat dissipation fins; obtaining information on the dirty areas on the surface of the converter valve radiator through visual recognition; dividing the radiator surface into grid units based on the radiator structural information and the dirty area information, constructing a two-dimensional grid map containing obstacles and dirty areas, the resolution of the grid units being set according to the radiator structural accuracy and cleaning requirements.

[0012] Further, information on contaminated areas on the surface of the converter valve radiator is obtained through visual recognition, including: acquiring an image of the converter valve radiator surface, the image covering the entire surface of the radiator; preprocessing the image, the preprocessing including image denoising, contrast enhancement, and distortion correction; constructing an HSV color space contamination feature model, using the HSV color space contamination feature model to locate contamination areas in the preprocessed image, the HSV color space contamination feature model determining the presence of contamination in a certain area based on the combined probability of hue, saturation, and brightness; using a convolutional neural network to identify the contamination level of the located contamination areas, the input of the convolutional neural network being the preprocessed image, and the output being the classification probability of the contamination level, the contamination levels being divided into 5 categories, and visual features for each contamination level being set based on converter station operation and maintenance experience and experimental data; and determining contamination area information based on the contamination area location results and contamination level identification results, the contamination area information including the location coordinates, area size, and corresponding contamination level of each contamination area.

[0013] Furthermore, an HSV color space contamination feature model is constructed. This model is used to locate contamination areas in the preprocessed image. The HSV color space contamination feature model determines the presence of contamination in a region based on the combined probability of hue, saturation, and brightness. This includes: converting the preprocessed image from the RGB color space to the HSV color space, separating the feature parameters of the hue, saturation, and brightness channels; establishing probability density models for hue, saturation, and brightness, respectively, based on the statistical regularity of color features in contamination areas; multiplying the probability densities of hue, saturation, and brightness to obtain the combined probability of the contamination color feature; and determining that the region is contamination when the combined probability exceeds a set threshold, recording its location coordinates and area.

[0014] Furthermore, the architecture of the convolutional neural network includes an input layer, a feature extraction layer, a feature aggregation layer, and an output layer. The feature extraction layer consists of multiple convolutional blocks, each containing a convolutional layer, a batch normalization layer, an activation function layer, and a pooling layer. The feature aggregation layer transforms the feature map into a fixed-length feature vector through a global average pooling layer, and then integrates the features through a fully connected layer. The output layer uses a softmax activation function to output the classification probability of the filth level.

[0015] Furthermore, in step two, the cleaning target points of each dirty area are sorted in descending order of dirt level. For dirty areas of the same dirt level, they are sorted in descending order of left to right and top to bottom according to their center coordinates. The cleaning target points are the center positions of each dirty area or key cleaning positions determined according to the dirt distribution density.

[0016] Furthermore, in step three, the optimized RRT algorithm includes a sampling strategy based on dirt density and a node expansion strategy based on obstacle neighborhood. The sampling strategy based on dirt density is used to increase the sampling probability of densely dirty areas, and the node expansion strategy based on obstacle neighborhood is used to avoid paths from approaching obstacles.

[0017] Furthermore, the sampling strategy based on contamination density includes: dividing the radiator surface into multiple grid cells according to the contamination area information in the two-dimensional grid map, calculating the contamination density of each grid cell, wherein the contamination density is the ratio of the area of ​​the contamination area within the grid cell to the total area of ​​the grid cells; and during the random sampling phase, allocating sampling probabilities based on the contamination density of each grid cell, with higher contamination density grid cells having a higher sampling probability, thereby concentrating sampling points more in densely contamination areas and reducing invalid sampling of clean areas.

[0018] Furthermore, the node expansion strategy based on obstacle neighborhood includes: defining an obstacle neighborhood, which is a region extending outward by a preset distance from the obstacle boundary, wherein the obstacle includes louver support columns and the roots of heat dissipation fins; during the node expansion process, when the expansion node corresponding to a randomly generated sampling point falls into the obstacle neighborhood, the expansion direction is adjusted, the direction of the line connecting the expansion node and the center of the obstacle is calculated, and the expansion node is offset away from the obstacle by a preset distance to obtain a new expansion node, ensuring that the new expansion node maintains a safe distance from the obstacle.

[0019] Furthermore, in step four, B-spline curves are used to smooth and optimize the initial path. By fitting the discrete nodes in the initial path to a continuously differentiable curve using B-spline curves, sharp angles in the path are eliminated, making the robot's motion trajectory smooth and reducing the impact of nozzle vibration on the cleaning effect. The order of the B-spline curve and the number of control points are set according to the complexity of the path and the smoothing requirements.

[0020] The beneficial effects of this invention are as follows: This invention accurately locates and classifies the contaminated areas on the surface of the converter valve radiator by identifying regional features, and optimizes the path planning algorithm by combining obstacle distribution and contamination density in the neighborhood information, thus achieving high efficiency and safety in cleaning paths. By prioritizing the cleaning of high-level contaminated areas, dynamically adjusting the target point sorting, and optimizing path smoothing, it effectively solves the problems of low efficiency and incomplete coverage in traditional manual cleaning. At the same time, collision detection and coverage verification ensure that the path has no collision risk and no cleaning blind spots, significantly improving the reliability and economy of automated cleaning of converter valve radiators, and providing technical support for the safe and stable operation of high-voltage direct current transmission systems. Attached Figure Description

[0021] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the converter valve radiator cleaning path planning method based on regional feature recognition and neighborhood information according to the present invention.

[0023] Figure 2 This is a diagram illustrating the architecture of a convolutional neural network-based model for identifying the surface contamination levels of a converter valve radiator.

[0024] Figure 3 The image shows the pollution level distribution, where (a) is the original image of the radiator and (b) is the pollution level distribution identification result.

[0025] Figure 4 The simulation diagrams show the path comparison, where (a) is the initial path of the traditional RRT algorithm and (b) is the path optimized by B-spline smoothing and neighborhood information. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0027] This invention addresses the problem of poor manual cleaning results for the surface of external cooling radiators of converter valves. It conducts research on path planning technology for converter valve radiator cleaning robots based on area feature recognition and neighborhood information. By intelligently identifying the distribution of dirt, the robot is precisely controlled to perform cleaning tasks, achieving efficient and comprehensive automated cleaning and solving the problems of low efficiency and poor cleaning results of traditional manual cleaning. Please refer to [link to relevant documentation]. Figure 1 The present invention provides a method for planning a cleaning path for a converter valve radiator based on regional feature recognition and neighborhood information, comprising the following steps:

[0028] Step 1: Construct a two-dimensional grid map of the converter valve radiator surface. The two-dimensional grid map contains obstacle information and contaminated area information. The obstacle information includes the location and size of the louver support column and the gap between the heat dissipation fins. The contaminated area information includes the location, area and contamination level of each contaminated area.

[0029] Step 2: Sort the cleaning target points of each dirty area according to the dirt level in the dirt area information.

[0030] Step 3: Using the optimized Rapidly Exploring Random Tree (RRT) algorithm, starting from the robot's initial position, and combining obstacle information and dirty area information in the two-dimensional grid map, feasible paths are sequentially searched for each of the cleaning target points to obtain an initial path set.

[0031] Step 4: Perform smoothing optimization on the initial paths in the initial path set to obtain the final cleaning path.

[0032] Step 5: Perform collision detection and cleaning coverage verification on the final cleaning path. If there is a collision risk or coverage blind spot, return to step 2 to readjust the sorting of cleaning target points, and repeat steps 3 to 5 until the path meets the requirements.

[0033] The following provides a detailed description of the converter valve radiator cleaning path planning method based on regional feature recognition and neighborhood information of the present invention.

[0034] The external cooling radiator of the converter valve operates in an outdoor environment all year round, and dust and dirt continuously accumulate on its surface.

[0035] Existing research has employed equivalent salt density methods, infrared or ultraviolet imaging techniques, and hyperspectral technology to detect surface contamination on transmission line insulators. Traditional contact-based detection methods, such as the equivalent salt density method, require power outages for sampling. Non-contact image recognition technologies, based on hyperspectral imaging, infrared thermal imaging, and deep learning, achieve accurate classification and grade assessment of contamination components by fusing material spectral characteristics, temperature field distribution, and spatial topological information. However, the aforementioned methods are costly, inefficient, and susceptible to environmental interference when detecting contamination on the external cooling radiators of converter valves. Therefore, this invention employs image recognition algorithms to identify surface contamination on the external cooling radiators of converter valves.

[0036] The surface color of the external cooling radiator of the converter valve changes as deposits accumulate, necessitating quantitative analysis of the color of the dirt on the radiator surface. In the field of color feature quantification analysis, color space models such as Red-Green-Blue (RGB), Hue-Saturation-Value (HSV), and Lab are widely used. Among them, the RGB color space follows the additive mixing principle of the three primary colors red, green, and blue, and is directly coupled with the physical photosensitivity mechanism of image acquisition devices; however, this model is extremely sensitive to changes in ambient light intensity and color temperature. The HSV color space represents color through three dimensions: hue, saturation, and lightness. Its perceptual dimension division conforms to the characteristics of human vision in CIE colorimetry and possesses strong light robustness. The Lab color space is constructed based on the CIE 1976 uniform color space standard, exhibiting device independence and demonstrating excellent accuracy in measuring color differences.

[0037] To address the complex operating conditions of the external cooling radiator of the converter valve, and considering the time-varying characteristics of ambient light conditions and the need for accurate differentiation between dirt and the radiator substrate color, a mathematical model of dirt color characteristics is constructed using the HSV color space model.

[0038] Let the characteristic parameters of the color of dirt on a certain area of ​​the surface of the converter valve's external cooling radiator in the HSV color space be (H, S, V). The hue of the dirt on the converter valve radiator surface depends on its composition; dust dirt is gray, and oily dirt is brown.

[0039] The hue H model follows a normal distribution, and the model expression is shown in Equation 1.

[0040] (1)

[0041] In Equation 1: Let H be the probability density of hue H. The mean of the filthy hue represents the filthy hue. The standard deviation of the hue indicates the degree of dispersion of this type of dirty hue; the smaller the value, the more concentrated the hue.

[0042] The saturation S model reflects the vividness of a color. Since dirt has a complex composition and is mostly a mixture of substances, its saturation is usually low. The model expression is shown in Equation 2.

[0043] (2)

[0044] In Equation 2: Let S be the probability density function for saturation S. The mean of saturation represents the average level of saturation for that type of saturation. The standard deviation of saturation reflects the fluctuation of saturation.

[0045] The brightness V model represents the lightness or darkness of a color. Dirt will reduce the brightness of the heat sink surface, making it look darker. The model expression is shown in Equation 3.

[0046] (3)

[0047] In Equation 2: Let V be the probability density function for brightness V. The mean value of the brightness due to dirt reflects the degree of brightness reduction caused by this type of dirt. The standard deviation of lightness reflects the range of lightness variation.

[0048] By multiplying the probability densities of hue, saturation, and brightness, a comprehensive probability model of the characteristics of dirty colors is obtained, which is used to determine whether the color of a certain area belongs to dirt, as shown in Equation 4.

[0049] (4)

[0050] when If the value exceeds the set threshold, the area is determined to be dirty.

[0051] The areas where dust and dirt accumulate on the external cooling radiator of the converter valve are relatively fixed, influenced by the position of the upper louvers and the season. By identifying and screening dirt images in certain areas, locations requiring focused cleaning can be quickly determined. The cleaning robot then automatically plans its path to reach these key cleaning areas and performs the cleaning. The surface dirt on the external cooling radiator of the converter valve exhibits regional and seasonal characteristics; image recognition is performed by capturing detailed information from localized areas and extracting local features.

[0052] Convolutional neural networks (CNNs) are a type of deep learning model, particularly suitable for extracting features from local regions. Therefore, this invention employs CNNs for the identification and classification of surface contamination levels in converter valve radiators. The architecture of the CNN-based model for identifying the surface contamination levels of converter valve radiators is as follows: Figure 2 As shown in the figure, the input layer receives a preprocessed image of the radiator surface. The feature extraction layer consists of multiple convolutional blocks, each containing a convolutional layer, a batch normalization layer, an activation function layer, and a pooling layer, progressively extracting local features from the image, including the edges, textures, and color distribution of contaminants. The feature aggregation layer transforms the feature map into a fixed-length feature vector through a global average pooling layer, and then integrates the features through a fully connected layer. The output layer uses the softmax activation function to output the classification probability of the contamination level. Based on converter station operation and maintenance experience and experimental data, the contamination level is divided into 5 categories, and the contamination classification and flushing time settings are shown in Table 1.

[0053] surface Sewage classification and flushing time settings

[0054] Serial Number Filth Level Visual features Rinse time Flushing water pressure 1. Level 0 The surface of the heat sink fins is a uniform light gray color with no obvious stains. 20 seconds 0.6 MPa 2. Level 1 Scattered gray spots appeared on the edges of the fins, accounting for less than 10%. 40 seconds 0.6 MPa 3. Level 2 A continuous layer of dirt, light brown in color, forms on the surface of the fins, accounting for 10%-30% of the total surface area. 60 seconds 0.8 MPa 4. Level 3 The entire surface of the fins is covered in filth, which is dark brown in color and has black hard lumps in some areas, accounting for 30%-50% of the surface. 120 seconds 1.0 MPa 5. Level 4 The dirt thickness exceeds half the fin gap, the surface shows cracks, the color is black, and it accounts for more than 50%. 180 seconds 1.2 MPa

[0055] Convolutional layers perform local correlation operations with the input feature map through convolution kernels to extract spatial features, as shown in Equation 5.

[0056] (5)

[0057] In Equation 5: This is the current convolutional layer number. For the first Layer Each feature map for Number of layer feature maps The kernel size is the convolution kernel size. For the first Layer The convolutional kernel corresponds to the first... Each input feature map Position weight, For the first Layer Each feature map in The pixel value of the location, This represents the offset corresponding to the sliding step size. For the first Layer Bias terms of each feature map, This is the activation function.

[0058] Pooling layers are used to reduce the dimension of feature maps, reduce computation, and enhance translation invariance. Max pooling is used, and the calculation is shown in Equation 6.

[0059] (6)

[0060] In Equation 6: For pooled window size, For the first Layer Each feature map in Position pooling results, For the first Layer Each feature map corresponds to a pixel value within a window.

[0061] The fully connected layer performs nonlinear mapping on the aggregated features for classification, as shown in Equation 7.

[0062] (7)

[0063] In Equation 7: Fully connected layer The value of each output node, The dimension of the input feature vector. For the first The output node and the first Connection weights for each input feature For the input feature vector, the first... One element, For bias terms, This is the activation function.

[0064] The output layer uses the softmax function to convert the output of the fully connected layer into classification probabilities, as shown in Equation 7.

[0065] (8)

[0066] In Equation 8: For the input image to belong to the first The probability of a certain level of filth. For the fully connected layer to the first The output value of the class The total number of categories, .

[0067] The unique structure of the louvered support columns and heat dissipation fins of the external cooling radiator of the converter valve, along with the regional distribution of contamination, necessitates that path planning be optimized by incorporating neighborhood information, including obstacle locations and the density of contamination areas. This invention utilizes the Fast Expanding Random Tree (RRT) algorithm, integrating neighborhood information to achieve path planning optimization.

[0068] RRT is a path planning algorithm based on probabilistic sampling. It constructs a search tree from the starting point to the destination by randomly sampling nodes in the environment space, ultimately finding a feasible path. Its basic process is as follows:

[0069] Initialization setting start point and the finish line Build a search tree Add the starting point as the root node of the tree. .

[0070] A sampling point is randomly generated in the environmental space. .

[0071] Find the node in the search tree T that is closest to the random sample point. .

[0072] from Towards Expand direction by one step new node .

[0073] Collision detection: judgment Did it collide with obstacles such as the louver support column or the radiator casing? If there was no collision, then... Add to search tree And record the parent node .

[0074] Endpoint determination: If and If the distance is less than a set threshold, the path search is successful, and the path from the parent node is retrieved by backtracking. arrive The path; otherwise, repeat steps 2-5 until the maximum number of iterations is reached.

[0075] Traditional Redirect Response (RRT) algorithms suffer from path redundancy and low search efficiency in complex obstacle environments, making them particularly unsuitable for scenarios with multiple contaminated areas and obstacles on the radiator surface. To address this, this invention optimizes the RRT algorithm by incorporating neighborhood information. A sampling strategy based on contaminated area density increases the sampling probability of densely contaminated areas, while a node expansion strategy based on obstacle neighborhood avoids collision risks caused by paths approaching obstacles too closely.

[0076] Based on the aforementioned contamination level distribution, the radiator surface is divided into multiple 50×50mm grid cells, and the contamination density of each grid cell is defined. This is the ratio of the area of ​​the contaminated region within the grid to the total area of ​​the grid, as shown in Equation 9.

[0077] (9)

[0078] In Equation 9: The area of ​​the contaminated zone within the grid. This represents the total area of ​​the grid.

[0079] During the random sampling phase, a weighted probability sampling strategy is adopted to determine the contamination density. Higher grid density, higher sampling probability The larger the value, the sampling probability is as shown in equation (10).

[0080] (10)

[0081] In Equation 10: Let be the sampling probability of the i-th grid. Let be the dirt density of the i-th grid. This represents the total number of grid points on the radiator surface. This strategy concentrates sampling points in areas with high levels of contamination, reduces invalid sampling in clean areas, and improves pathfinding efficiency.

[0082] The obstacle louver support columns and heat dissipation fin roots on the radiator surface have fixed positions and dimensions. The obstacle neighborhood is defined based on the obstacle boundary. To avoid collisions between the robot cleaning nozzle and the obstacle, a 10mm outward area is added. During the node expansion phase, if... If the object falls into the neighborhood of an obstacle, adjust the expansion direction: calculate... The direction of the line connecting the center of the obstacle will Offset 5mm away from the obstacle to obtain a new extended node. As shown in equation (11).

[0083] (11)

[0084] In formula 11: The coordinates of the obstacle's center; for and The Euclidean distance is calculated. This optimization ensures a safe distance of ≥5mm between the path and obstacles, reducing the collision rate.

[0085] Combining the structure of the external cooling radiator of the converter valve with the motion characteristics of the cleaning robot, the specific steps for path planning are as follows:

[0086] Based on the radiator CAD drawings and visual recognition results, a two-dimensional grid map containing obstacles, louver support columns, fins, and dirty areas is constructed, with the grid resolution set to 10×10mm.

[0087] Based on the level of contamination from high to low (Level 4 → Level 0), and within the same level, the cleaning target points of each contaminated area are sorted from left to right and from top to bottom according to the center coordinates of the contaminated area, to ensure that high-risk contaminated areas are cleaned first.

[0088] Using the optimized RRT method, starting from the lower left corner of the radiator at the robot's initial position, feasible paths to each target point are searched sequentially to obtain the initial path set. ,in Let be the path from the i-th target point to the i-th target point.

[0089] The initial path is smoothed and optimized using B-spline curves to eliminate sharp angles in the path, as shown in Equation (12), so that the robot's motion trajectory is continuous and differentiable, and the impact of nozzle vibration on the cleaning effect is reduced.

[0090] (12)

[0091] In Equation 12: Points on the B-spline curve; The nodes of the initial path; is the d-th order spline basis function; n is the number of control points.

[0092] The smoothed path undergoes collision detection and coverage verification. If collision risks or coverage blind spots are found, the target point sorting and sampling parameters are readjusted, and steps 3-4 are repeated until the path meets the requirements.

[0093] The present invention will be described below with reference to specific examples.

[0094] An industrial camera with a resolution of 2592×1944 and a frame rate of 30fps, equipped with an 8mm fixed-focus lens, was mounted on top of the cleaning robot, 1.5m above the radiator surface. The acquisition range covered the entire radiator surface. Images of the radiator surface were captured by the industrial camera, and after being processed by a CNN model, the distribution of dirt levels was obtained as follows: Figure 3 As shown, there are three main areas of contamination on the radiator surface: the lower left area is moderately contaminated (Level 2), covering approximately 0.8 m²; the central area is severely contaminated (Level 3), covering approximately 0.5 m²; and the upper right area is extremely severely contaminated (Level 4), covering approximately 0.3 m². The remaining areas are either lightly contaminated (Level 0-1) or clean. The contamination level identified by the CNN model is completely consistent with the actual judgment results, verifying the model's accuracy.

[0095] Based on the pollution level distribution, an optimized RRT algorithm is used for path planning, resulting in a comparison between the initial path and the smoothed path. Figure 4 As shown, the initial path planned by the traditional RRT algorithm has multiple redundant turns, with a path length of 18.6m. After B-spline smoothing and neighborhood information optimization, the path length is shortened to 15.2m, a reduction rate of 18.3%. Moreover, the path avoids obstacles such as louver support columns, and the minimum distance to obstacles is 8mm, which meets the safety requirements.

[0096] This invention addresses the problem of low efficiency and poor results of manual cleaning of external cooling radiators for converter valves. It proposes an automated cleaning technology based on regional feature recognition and neighborhood information, constructs an HSV color space dirt feature model, achieves accurate positioning of dirty areas through comprehensive probability judgment, and establishes a 5-level dirt classification system by combining a CNN model, providing a quantitative basis for differentiated cleaning.

[0097] We propose an optimization scheme for the RRT algorithm based on neighborhood information. By sampling the dirt density and expanding the obstacle neighborhood, the path length is shortened by 18.3% and the collision rate is reduced to below 0.1%, balancing path economy and safety.

[0098] The embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention.

Claims

1. A method for planning cleaning paths for converter valve radiators based on regional feature recognition and neighborhood information, characterized in that, Includes the following steps: Step 1: Construct a two-dimensional grid map of the converter valve radiator surface. The two-dimensional grid map contains obstacle information and contaminated area information. The obstacle information includes the location and size of the louver support column and the gap between the heat dissipation fins. The contaminated area information includes the location, area and contamination level of each contaminated area. Step 2: Sort the cleaning target points of each dirty area according to the dirt level in the dirty area information; Step 3: Using the optimized RRT algorithm, starting from the robot's initial position, and combining the obstacle information and dirty area information in the two-dimensional grid map, feasible paths for each of the cleaning target points are searched sequentially to obtain an initial path set; Step 4: Perform smoothing optimization on the initial paths in the initial path set to obtain the final cleaning path; Step 5: Perform collision detection and cleaning coverage verification on the final cleaning path. If there is a collision risk or coverage blind spot, return to step 2 to readjust the sorting of cleaning target points, and repeat steps 3 to 5 until the path meets the requirements.

2. The converter valve radiator cleaning path planning method based on regional feature recognition and neighborhood information according to claim 1, characterized in that, Step one includes: Obtain the CAD drawing of the converter valve radiator, and extract the radiator structure information from the CAD drawing. The radiator structure information includes the position of the louver support column, the distribution and size of the heat dissipation fins. Information on contaminated areas on the surface of the converter valve radiator is obtained through visual recognition; Based on the radiator structure information and the dirty area information, the radiator surface is divided into grid cells to construct a two-dimensional grid map containing obstacles and dirty areas. The resolution of the grid cells is set according to the radiator structure accuracy and cleaning requirements.

3. The converter valve radiator cleaning path planning method based on regional feature recognition and neighborhood information according to claim 2, characterized in that, Information on contaminated areas on the surface of the converter valve radiator is obtained through visual recognition, including: Acquire an image of the surface of the converter valve radiator, the image covering the entire area of ​​the radiator surface; The image is preprocessed, including image denoising, contrast enhancement, and distortion correction. An HSV color space dirt feature model is constructed, and dirt areas are located in the preprocessed image using the HSV color space dirt feature model. The HSV color space dirt feature model determines whether a certain area is dirty by combining the probabilities of hue, saturation and brightness. A convolutional neural network is used to identify the pollution level of the located contaminated area. The input of the convolutional neural network is the preprocessed image, and the output is the classification probability of the pollution level. The pollution level is divided into 5 categories. The visual features of each pollution level are set by combining the operation and maintenance experience of the converter station and experimental data. Based on the location results of the contaminated areas and the contamination level identification results, the contaminated area information is determined, which includes the location coordinates, area size, and corresponding contamination level of each contaminated area.

4. The converter valve radiator cleaning path planning method based on regional feature recognition and neighborhood information according to claim 3, characterized in that, An HSV color space contamination feature model is constructed. This model is used to locate contamination areas in the preprocessed image. The HSV color space contamination feature model determines the presence of contamination in a given area based on a combined probability of hue, saturation, and brightness. This includes: The preprocessed image is converted from the RGB color space to the HSV color space, and the feature parameters of the three channels of hue, saturation and lightness are separated. Probability density models for hue, saturation, and brightness are established respectively, and the probability density models are constructed based on the statistical regularity of color characteristics of the dirty area; Multiply the probability densities of hue, saturation, and brightness to obtain the combined probability of the dirty color feature; When the overall probability is greater than the set threshold, the area is determined to be a polluted area, and its location coordinates and area are recorded.

5. The converter valve radiator cleaning path planning method based on regional feature recognition and neighborhood information according to claim 3, characterized in that, The architecture of the convolutional neural network includes an input layer, a feature extraction layer, a feature aggregation layer, and an output layer. The feature extraction layer consists of multiple convolutional blocks, each containing a convolutional layer, a batch normalization layer, an activation function layer, and a pooling layer. The feature aggregation layer transforms the feature map into a fixed-length feature vector through a global average pooling layer, and then integrates the features through a fully connected layer. The output layer uses a softmax activation function to output the classification probability of the filth level.

6. The method for planning the cleaning path of a converter valve radiator based on regional feature recognition and neighborhood information according to claim 1, characterized in that, In step two, the cleaning target points of each dirty area are sorted in descending order of dirt level. For dirty areas of the same dirt level, they are sorted in descending order of left to right and top to bottom according to their center coordinates. The cleaning target points are the center positions of each dirty area or the key cleaning positions determined according to the dirt distribution density.

7. The method for planning the cleaning path of a converter valve radiator based on regional feature recognition and neighborhood information according to claim 1, characterized in that, In step three, the optimized RRT algorithm includes a sampling strategy based on dirt density and a node expansion strategy based on obstacle neighborhood. The sampling strategy based on dirt density is used to increase the sampling probability of densely dirty areas, and the node expansion strategy based on obstacle neighborhood is used to avoid the path from being too close to obstacles.

8. The converter valve radiator cleaning path planning method based on regional feature recognition and neighborhood information according to claim 7, characterized in that, The sampling strategy based on dirt density includes: Based on the contaminated area information in the two-dimensional grid map, the surface of the radiator is divided into multiple grid cells, and the contamination density of each grid cell is calculated. The contamination density is the ratio of the area of ​​the contaminated area in the grid cell to the total area of ​​the grid cells. During the random sampling phase, the sampling probability is assigned based on the contamination density of each grid cell. The grid cell with higher contamination density has a higher sampling probability, which makes the sampling points more concentrated in the contamination-dense areas and reduces invalid sampling of clean areas.

9. The converter valve radiator cleaning path planning method based on regional feature recognition and neighborhood information according to claim 7, characterized in that, The node expansion strategy based on obstacle neighborhood includes: Define an obstacle neighborhood, which is a region extending outward from the obstacle boundary by a preset distance. The obstacle includes a louver support column and the root of a heat dissipation fin. During node expansion, when the expanded node corresponding to the randomly generated sampling point falls into the neighborhood of the obstacle, the expansion direction is adjusted, the direction of the line connecting the expanded node and the center of the obstacle is calculated, and the expanded node is shifted away from the obstacle by a preset distance to obtain a new expanded node, ensuring that the new expanded node maintains a safe distance from the obstacle.

10. The method for planning the cleaning path of a converter valve radiator based on regional feature recognition and neighborhood information according to claim 1, characterized in that, In step four, B-spline curves are used to smooth and optimize the initial path. By fitting the discrete nodes in the initial path to a continuously differentiable curve using B-spline curves, sharp angles in the path are eliminated, making the robot's motion trajectory smooth and reducing the impact of nozzle vibration on the cleaning effect. The order of the B-spline curve and the number of control points are set according to the complexity of the path and the smoothing requirements.

Images