Intelligent control method and system for mold cleaning

Abstract

The application provides an intelligent control method and system for mold cleaning, and belongs to the technical field of electronics. It solves the problem of unsatisfactory cleaning effect of existing manual cleaning. The intelligent control method for mold cleaning comprises the following steps: obtaining a three-dimensional point cloud model of a mold to be cleaned; after cleaning the mold to be cleaned from top to bottom and from left to right, identifying complex regions, deep cavity regions and hole site regions, and identifying the pose information of the mold to be cleaned through visual recognition technology, determining the 3D positions of the complex regions, the deep cavity regions and the hole site regions based on the identified pose information; enhancing water pressure, and spraying and washing the 3D positions of the complex regions, the deep cavity regions and the hole site regions at high pressure and different angles; judging whether there is a dirty region in the mold to be cleaned, and if there is a dirty region, spraying and washing the 3D position of the dirty region at high pressure until there is no dirty region. The application can improve the cleaning effect of the mold.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of electronics, and relates to an intelligent control method and system for mold cleaning. BACKGROUND

[0002] Mold parts need to be cooled by cutting fluid during machining on a machine tool, and several centimeters to several millimeters of iron filings will appear during milling. These iron filings will adhere to the surface of the workpiece or be left in the mold gaps or holes and grooves, affecting the normal work of other processes. Therefore, these iron filings and other dirt need to be cleaned before being transferred to other processes. Currently, manual cleaning is still mainly used in mold cleaning processes, which requires 30-60 minutes of time. This not only occupies the utilization rate of the machine tool and affects production efficiency, but also consumes a lot of manpower and material resources, and makes the workers full of dirt during the cleaning process. Moreover, manual cleaning may miss some dirty areas, and the cleaning effect is not ideal. SUMMARY

[0003] The present application aims to solve the above problems in the prior art, and provides an intelligent control method and system for mold cleaning, which solves the technical problem of how to improve the cleaning effect of the mold.

[0004] The object of the present application can be achieved by the following technical scheme: an intelligent control method for mold cleaning, comprising the following steps:

[0005] S1. Reconstructing a three-dimensional point cloud model of the mold to be cleaned by visual recognition technology;

[0006] S2. Cleaning the mold to be cleaned from top to bottom and from left to right;

[0007] S3. After cleaning, identifying complex areas, deep cavity areas and hole areas according to visual complex area recognition algorithm, visual deep cavity recognition algorithm and visual hole recognition algorithm based on the three-dimensional point cloud model obtained in step S1, and identifying the pose information of the mold to be cleaned by visual recognition technology, and determining the 3D positions of the complex areas, deep cavity areas and hole areas based on the identified pose information;

[0008] S4. Enhancing the water pressure, and performing high-pressure spray cleaning at different angles for the 3D positions of the complex areas, deep cavity areas and hole areas determined in step S3;

[0009] S5. Again performing visual recognition on the mold to be cleaned by visual recognition technology, and determining whether there are dirty areas in the mold to be cleaned, and if there are dirty areas, performing high-pressure spray cleaning for the 3D positions of the dirty areas until there are no dirty areas.

[0010] The mold part is milled on a machine tool, and then is cleaned by the intelligent control method for mold cleaning. First, a three-dimensional point cloud model of the mold to be cleaned is established, and then the mold to be cleaned is fully sprayed and washed, so that most of the iron filings and dirt can be cleaned. The order of the above two steps can be changed. After the full-spray washing is completed, in order to ensure the cleaning effect of the mold to be cleaned, the complex region, the deep cavity region and the hole region are identified, and the spray and washing are carried out in each special fixed position for these regions, so as to ensure that these regions can be fully washed. After the special washing, the mold to be cleaned is identified again, and it is confirmed whether the mold to be cleaned is clean. If there is still a dirty area, the position of the dirty area is washed again until there is no dirty area in the mold to be cleaned, and the cleaning is completed. The method realizes the automatic cleaning of the mold by the visual recognition technology, and can identify the complex region, the deep cavity region, the hole region and the dirty region and carry out special fixed-point cleaning, so as to improve the cleaning effect of the mold.

[0011] In the intelligent control method for mold cleaning, in step S1, the three-dimensional reconstruction of the mold to be cleaned by the visual recognition technology includes:

[0012] The mold to be cleaned is placed on a rotating table, and the full-range image information of the mold to be cleaned is collected by controlling the rotating table to rotate at multiple angles of 360 degrees, so as to obtain a multi-view image sequence of the mold to be cleaned;

[0013] The SIFT algorithm is used to detect the feature points and corresponding feature descriptors of each image in the multi-view image sequence, and the feature descriptors are used to match the feature points of adjacent images to obtain a feature point matching result;

[0014] Based on the feature point matching result, the SFM algorithm is used for sparse reconstruction to obtain a sparse three-dimensional reconstruction result;

[0015] Based on the sparse three-dimensional reconstruction result and the internal and external parameters of the camera, the MVS algorithm is used for dense reconstruction to obtain a 3D point cloud of each pixel point;

[0016] The 3D point cloud is converted into a 3D point cloud model, so as to obtain a three-dimensional point cloud model of the mold to be cleaned. SIFT, namely, Scale-invariant feature transform, is a description used in the field of image processing. The description has scale invariance and can detect key points in an image, and is a local feature descriptor. The SFM algorithm is an offline algorithm for three-dimensional reconstruction based on various collected unordered pictures. The algorithm analyzes and processes the matched image sequence to deduce the three-dimensional structure of the object and the motion trajectory of the camera. MVS (Multi-View Stereo) is a multi-view three-dimensional reconstruction algorithm for matching each pixel point in a photo and reconstructing the three-dimensional coordinates of each pixel point. The internal and external parameters of the camera include lens distortion parameters, an internal parameter matrix, a rotation matrix, a translation vector, a camera coordinate system and a world coordinate system. By estimating the internal and external parameters of the camera, the conversion between the camera coordinate system and the world coordinate system can be realized, and the three-dimensional pose of the space point can be determined.

[0017] In the above-mentioned intelligent control method for mold cleaning, before the step S2, the operation of aligning the mold to be cleaned is further included before the full cleaning of the mold to be cleaned from top to bottom and from left to right, and the operation includes:

[0018] A plurality of images of the mold to be cleaned at different angles are taken, and the pose information of the mold to be cleaned in the coordinate system of the rotating table is obtained by analyzing each image;

[0019] The obtained pose information is compared with the preset target pose, and the deviation angle of the mold to be cleaned is obtained;

[0020] The rotating table is controlled to rotate according to the deviation angle until the mold to be cleaned is aligned. The alignment of the mold to be cleaned before the full cleaning can make the mold to be cleaned cleaned according to the preset path, and the control is more convenient and the cleaning efficiency is improved.

[0021] In the above-mentioned intelligent control method for mold cleaning, in the step S2, the operation of obtaining the pose information of the mold to be cleaned in the coordinate system of the rotating table includes:

[0022] For each image taken, a feature point detection algorithm is used to extract key points in the image and corresponding key point descriptors;

[0023] The feature descriptors of the current image are matched with the feature descriptors obtained by using the SFM algorithm for dense reconstruction in the step S1, and the matching relationship of the feature descriptors is obtained.

[0024] The pose information of the mold to be cleaned in the rotating table coordinate system is obtained by analyzing and calculating the matching relationship of the feature descriptors. The feature point detection algorithm can use the SIFT algorithm.

[0025] In the above intelligent control method for mold cleaning, in the step S2, the operation of optimizing the pose information of the mold to be cleaned by nonlinear optimization or least squares is further included. The pose is optimized to achieve fine adjustment of the pose, so as to further improve the accuracy of the pose estimation.

[0026] In the above intelligent control method for mold cleaning, in the step S2, the operation of comparing the obtained pose information with the preset target pose includes:

[0027] The obtained pose information includes a translation vector T c and a rotation matrix R c ; the preset target pose includes a translation vector T t and a rotation matrix R t ;

[0028] The translation vector deviation T diff is calculated by T c -T t ;

[0029] The distance between the rotation matrices is calculated, which is converted into a rotation angle deviation angle, angle=acos((trace(R c ^T*R t )-1) / 2);

[0030] The translation vector deviation T diff and the rotation angle deviation angle are combined into the control parameters for adjusting the alignment of the mold to be cleaned.

[0031] In the above intelligent control method for mold cleaning, in the step S3, the identification operations of the visual complex area identification algorithm, the visual deep cavity identification algorithm, and the visual hole position identification algorithm all include:

[0032] Obtain the 3D point cloud design model of the mold to be cleaned;

[0033] The 3D point cloud model reconstructed in step S1 is registered with the 3D point cloud design model by ICP algorithm to obtain the mapping relationship between the 3D point cloud design model and the three-dimensional point cloud model;

[0034] The PointNet algorithm is used for point cloud segmentation of the 3D point cloud design model, and the segmentation results of the complex region, the deep cavity region and the hole position region are obtained. Then, according to the mapping relationship between the 3D point cloud design model and the three-dimensional point cloud model, the positions of the complex region, the deep cavity region and the hole position region on the mold to be cleaned are obtained. The ICP algorithm (Iterative Closest Point) is a commonly used iterative closest point matching algorithm, which is commonly used for registration and alignment of point cloud data. In material recognition, the ICP algorithm can be used for registration of two 3D point clouds, that is, the reconstructed mold with cleaning model is aligned with the real-time collected point cloud data. Through iterative closest point matching and optimization, the ICP algorithm can find the optimal transformation matrix to register the two point clouds, thereby obtaining accurate material pose information. In this way, the complex region, the deep cavity region and the hole position can be identified and positioned. The PointNet algorithm is a deep learning algorithm for point cloud processing, which can be used for point cloud segmentation tasks.

[0035] In the above intelligent control method for mold cleaning, in step S4, the high-pressure different-angle spray washing operation for the 3D positions of the complex region, the deep cavity region and the hole position region determined in step S3 includes:

[0036] According to the 3D positions of the complex region, the deep cavity region and the hole position region, the rotating table is controlled to rotate and the spray gun is controlled to aim at the complex region, the deep cavity region or the hole position region, so as to realize the one-by-one spray cleaning of the complex region, the deep cavity region and the hole position region.

[0037] An intelligent control system for mold cleaning, comprising a control unit and a rotating table execution mechanism connected with the control unit for rotating the rotating table by 360°, a spray gun execution mechanism for adjusting the spray position and angle of the spray gun, and an image acquisition mechanism for acquiring omnidirectional image information of the mold to be cleaned. The control unit comprises:

[0038] A model establishment module for visually recognizing the omnidirectional image information acquired by the image acquisition mechanism and establishing a three-dimensional point cloud model of the mold to be cleaned;

[0039] A pose recognition module for visually recognizing the omnidirectional image information acquired by the image acquisition mechanism and acquiring the pose information of the mold to be cleaned in the rotating table coordinate system;

[0040] A complex region recognition module for identifying the 3D positions of the complex region on the mold to be cleaned according to the established three-dimensional point cloud model, the pose information and the pre-stored 3D point cloud design model;

[0041] The deep cavity region identification module is configured to identify the 3D position of the deep cavity region on the mold to be cleaned according to the established three-dimensional point cloud model, the pose information and the pre-stored 3D point cloud design model.

[0042] The hole site region identification module is configured to identify the 3D position of the hole site region on the mold to be cleaned according to the established three-dimensional point cloud model, the pose information and the pre-stored 3D point cloud design model.

[0043] The dirty region identification module is configured to identify whether there is a dirty region on the mold to be cleaned according to the established three-dimensional point cloud model, the pose information and the pre-stored 3D point cloud design model, and identify the 3D position of the dirty region when the dirty region exists.

[0044] The control module is configured to output corresponding control instructions to the rotating table actuator, the spray gun actuator and the image acquisition mechanism according to the 3D positions of the complex region, the deep cavity region, the hole site region and the dirty region and the pre-set control strategy.

[0045] After the mold part is milled on the machine tool, the mold is cleaned by the intelligent control system for mold cleaning. First, the mold to be cleaned is placed on the rotating table, and the rotating table actuator is controlled to rotate the rotating table by 360° to realize omnidirectional image acquisition of the mold to be cleaned by the image acquisition mechanism. The model establishment module in the control unit performs visual recognition on the omnidirectional image information collected by the image acquisition mechanism to establish a three-dimensional point cloud model of the mold to be cleaned. The pose recognition module identifies the pose information of the mold to be cleaned in the rotating table coordinate system. The control module outputs corresponding control instructions according to the pre-set control strategy, i.e. the control steps in the intelligent control method for mold cleaning. First, the control instructions are output to the spray gun actuator to control it to spray and wash the mold to be cleaned in all directions, thereby cleaning most of the dirt such as iron filings. After the all-directional spraying and washing is completed, in order to ensure the cleaning effect of the mold to be cleaned, the complex region, the deep cavity region and the hole site region are identified by the complex region identification module, the deep cavity region identification module and the hole site region identification module. The control module controls the spray gun actuator and the rotating table actuator to act according to the identification results, so that the spray gun sprays and washes these regions one by one in a special way, ensuring that these regions can be fully washed. After the special spraying and washing, the image acquisition mechanism takes omnidirectional image information of the mold to be cleaned again. The dirty region identification module identifies whether there is a dirty region on the mold to be cleaned. When there is still a dirty region, the position of the dirty region is washed again until there is no dirty region on the mold to be cleaned, and the cleaning is completed. The system can realize automatic cleaning of the mold, and can also identify the complex region, the deep cavity region, the hole site region and the dirty region and clean them in a special way, thereby improving the cleaning effect of the mold.

[0046] In the intelligent control method for mold cleaning, the control unit further comprises:

[0047] An angle deviation comparison module is configured to compare the pose information of the mold to be cleaned in the rotating table coordinate system with a preset target pose to obtain a deviation angle of the mold to be cleaned.

[0048] The control module is configured to control the rotating table actuator to work according to the deviation angle to adjust the mold to be cleaned.

[0049] Compared with the prior art, the intelligent control method and system for mold cleaning can realize automatic cleaning of the mold, automatic recognition of hole site regions, complex regions, deep cavity regions and dirty regions, and fixed-point positioning and spraying of these regions, so that the cleaning is more comprehensive and clean, and the cleaning effect of the mold cleaning is effectively improved. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a structural schematic diagram of the present application.

[0051] In the figure, 1, control unit; 11, model establishment module; 12, pose recognition module; 13, complex region recognition module; 14, deep cavity region recognition module; 15, hole site region recognition module; 16, dirty region recognition module; 17, control module; 18, angle deviation comparison module; 2, image acquisition mechanism; 3, spray gun actuator; 4, rotating table actuator. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical scheme and advantages of the present application clearer, the following will further describe the embodiments of the present application in combination with the drawings. The following description of at least one exemplary embodiment is actually only illustrative, but not as any limitation on the present application and its application or use. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0053] After the mold parts are milled on the machine tool, the iron filings and other dirt on the mold parts are cleaned by the intelligent control method for mold cleaning. First, the mold to be cleaned is placed on the rotating table, and then the mold to be cleaned is three-dimensionally reconstructed, and the operation is as follows:

[0054] The full-range image information of the mold to be cleaned is taken by controlling the rotating table to rotate at 360° multiple angles to obtain a multi-view image sequence of the mold to be cleaned.

[0055] The SIFT algorithm is used to detect feature points and corresponding feature descriptors in each image of the multi-view image sequence. Specifically, the SIFT algorithm detects feature points through the following steps: Scale Space Extrema Detection, a Gaussian pyramid is used to construct the scale space of the image, and the image is smoothed at different scales. Then, in each scale space, the values of the pixel points and their surrounding pixel points and adjacent scale space pixel points are compared to find possible key points. Keypoint Localization, the candidate key points found in the Scale Space Extrema Detection step need to be accurately positioned to extract accurate feature descriptors. SIFT algorithm uses a series of image gradient calculation methods such as Difference of Gaussians (DoG) and Hessian matrix to determine the precise location of the key points. Orientation Assignment, each key point is assigned a dominant direction, so that the feature descriptor is invariant to rotation. SIFT algorithm calculates the gradient direction histogram around the key point to select the dominant direction to assign the direction to each key point. Feature Descriptor: In the region around the key point, SIFT algorithm uses the gradient information of the local image block to construct the feature descriptor. Specifically, a fixed size image block is taken around the key point, and the image block is rotated and corrected according to the dominant direction. Then, the image block is divided into small sub-regions (such as 4x4 or 8x8 small blocks), and the amplitude and direction of the gradient are calculated in each sub-region. Finally, these gradient information is combined to form a vector as the feature descriptor of the key point. By generating feature descriptors, SIFT algorithm can represent the local features around each key point in the form of a vector. These descriptors can be used to match images of different views, for tasks such as three-dimensional reconstruction and pose estimation of materials. Feature descriptors can express the shape, texture and other features of the image region around the key point.

[0056] Then, feature point matching is performed between adjacent images using the feature descriptors to obtain the feature point matching result. Feature point matching is to associate the feature points in adjacent images to establish the correspondence between them. This step is to find the same surface feature points of the mold to be cleaned in multiple view images, so as to provide matching point pairs for subsequent three-dimensional reconstruction. Through feature point matching, matching feature point pairs in the image sequence can be obtained, representing feature points on the surface of the material in different view images. These matching point pairs can be used for subsequent stereo geometry calculation and three-dimensional reconstruction, such as three-dimensional point cloud generation, camera pose estimation, etc.

[0057] Based on the feature point matching results, a SFM algorithm is used for sparse reconstruction to obtain a sparse three-dimensional reconstruction result. The SFM algorithm analyzes and processes the matched image sequence to deduce the three-dimensional structure of the mold to be cleaned and the motion trajectory of the camera. Specifically, the sparse reconstruction process usually includes the following steps:

[0058] Feature point matching: For each pair of adjacent images in the image sequence, the corresponding relationship between them is found through feature point matching.

[0059] Camera pose estimation: According to the results of feature point matching, the relative motion between adjacent images is estimated by solving the camera pose (including rotation and translation).

[0060] Three-dimensional point triangulation: According to the principle of multi-view geometry, the three-dimensional coordinates of the feature points are obtained by triangulating them based on the positions of the feature points observed by multiple cameras and the camera parameters.

[0061] Pose and structure optimization: Using a nonlinear optimization algorithm, the camera pose and three-dimensional point positions are optimized by minimizing the reprojection error to improve the accuracy of the reconstruction.

[0062] Sparse reconstruction: In the optimization process, a threshold or other strategy is set to only retain the three-dimensional coordinates of the key points related to the reconstruction of the mold to be cleaned, thereby realizing sparse reconstruction.

[0063] After that, based on the sparse three-dimensional reconstruction result and the internal and external parameters of the camera, the MVS algorithm is used for dense reconstruction. The specific operation is as follows: image alignment: align the input multiple images to a reference coordinate system. This can be achieved by estimating the relative pose transformation (such as rotation and translation) between each image and the camera internal and external parameters.

[0064] Disparity map calculation: For each pair of adjacent images, calculate their disparity map. Disparity refers to the horizontal displacement of corresponding pixels at different viewing angles, which can be used to estimate the depth of the object surface.

[0065] Disparity map optimization: By smoothing and optimizing the disparity map, noise and inconsistencies can be reduced, and the reconstruction quality can be improved. Common optimization methods include disparity map filtering and global energy minimization.

[0066] Three-dimensional reconstruction: Using the disparity information calculated from the disparity map, combined with the camera internal and external parameters and the sparse three-dimensional reconstruction result, the dense reconstruction of the three-dimensional point cloud is performed. Interpolation methods or voxel grid methods can be used to generate dense three-dimensional point clouds.

[0067] Surface reconstruction: The dense 3D point cloud is used to reconstruct the surface of the object, generating a 3D surface model. Common methods include triangulation and surface fitting.

[0068] Quality assessment and post-processing: The reconstruction results are assessed and post-processed, such as removing invalid points, filling holes, and correcting lighting, etc.

[0069] Further, the 3D point cloud of each pixel point is obtained; the 3D point cloud refers to a set of discrete points in three-dimensional space, each point has three coordinate values (x, y, z), indicating its position in space. In multi-view three-dimensional reconstruction of materials, by using MVS algorithm for dense reconstruction, the three-dimensional coordinates of each pixel point can be obtained, thereby forming a complete 3D point cloud.

[0070] Finally, the 3D point cloud is converted into a 3D point cloud model, thereby obtaining the three-dimensional point cloud model of the mold to be cleaned. Specifically, a data structure suitable for storing and representing the 3D point cloud is selected. Common data structures include point cloud index, point cloud mesh, etc. Point cloud connection: Connect discrete 3D points to form continuous geometric shapes. This can be achieved through triangulation, surface reconstruction, etc. The relationship between points is established to obtain a smooth surface. Point cloud attributes: In addition to coordinate information, 3D point cloud models usually contain other attributes such as color, normal, intensity, etc. These attributes can be obtained through sensors or calculated through post-processing algorithms. Associating these attributes with the coordinates of the points can enhance the visualization and analysis capabilities of the point cloud model. Data format: Save the 3D point cloud model in a specific data format for storage, transmission, and processing. Common point cloud data formats include PLY, XYZ, LAS, OBJ, etc., which have different characteristics and application scenarios.

[0071] After obtaining the three-dimensional point cloud model of the mold to be cleaned, the mold to be cleaned is placed on the rotating table, and the mold to be cleaned is placed on the rotating table. The operation includes:

[0072] Take multiple images of the mold to be cleaned from different angles, for each image taken, use a feature point detection algorithm to extract key points and corresponding key point descriptors in the image; match the feature descriptors of the current image with the feature descriptors obtained by using the SFM algorithm for dense reconstruction to obtain the matching relationship of the feature descriptors; by analyzing and calculating the matching relationship of the feature descriptors, the pose information of the mold to be cleaned in the rotating table coordinate system is obtained, as an optimization, the pose information of the mold to be cleaned is optimized through nonlinear optimization or least squares;

[0073] Then compare the obtained pose information with the preset target pose to obtain the deviation angle of the mold to be cleaned; specifically, the obtained pose information includes a translation vector T cand a rotation matrix R c ; the preset target pose includes a translation vector T t and a rotation matrix R t ; a translation vector deviation is calculated to obtain a translation vector deviation T diff = T c -T t ; a distance between the rotation matrices is calculated and converted into a rotation angle deviation angle, angle = acos((trace(R c ^T*R t )-1) / 2); the translation vector deviation T diff and the rotation angle deviation angle are combined into control parameters for adjusting the alignment of the mold to be cleaned.

[0074] Then, the rotating table is controlled to rotate according to the deviation angle until the mold to be cleaned is aligned. After alignment, the mold to be cleaned can be cleaned in all aspects according to the preset path, and the control is more convenient and the cleaning efficiency is improved.

[0075] After the full-surface spray cleaning is completed, in order to ensure the cleaning effect of the mold to be cleaned, the complex region, the deep cavity region and the hole region are identified. The specific identification process is as follows: based on the established three-dimensional point cloud model, the complex region, the deep cavity region and the hole region are identified according to the visual complex region identification algorithm, the visual deep cavity identification algorithm and the visual hole identification algorithm respectively, and the pose information of the mold to be cleaned is identified through the visual identification technology, the 3D positions of the complex region, the deep cavity region and the hole region are determined based on the identified pose information, and the identification operations of the visual complex region identification algorithm, the visual deep cavity identification algorithm and the visual hole identification algorithm all include: obtaining the 3D point cloud design model of the mold to be cleaned; performing 3D point cloud registration on the three-dimensional point cloud model reconstructed by the ICP algorithm and the 3D point cloud design model to obtain the mapping relationship between the 3D point cloud design model and the three-dimensional point cloud model; performing point cloud segmentation on the 3D point cloud design model using the PointNet algorithm to obtain the segmentation results of the complex region, the deep cavity region and the hole region, and then obtaining the positions of the complex region, the deep cavity region and the hole region on the mold to be cleaned according to the mapping relationship between the 3D point cloud design model and the three-dimensional point cloud model.

[0076] The operations of the ICP algorithm include: initialization: first select an initial transformation matrix to perform initial transformation on one of the point clouds. Usually use some prior information or initial estimate to initialize, such as through rough initial alignment or pose estimation provided by the sensor. Nearest point matching: for each point to be registered, find the nearest point in the 3D point cloud design model. This can be achieved by calculating the distance (such as Euclidean distance) between points. Finally, a set of point pairs is obtained, representing the corresponding nearest point pairs in the source point cloud and the target point cloud. Calculate registration transformation: using the nearest point pairs, optimize the initial transformation by calculating the transformation matrix between the two point clouds. Common methods are to minimize the distance error between point pairs, such as using least squares or iterative optimization algorithms (such as Gauss-Newton method or Levenberg-Marquardt algorithm) to solve the optimal rotation matrix and translation vector. Update transformation matrix: according to the optimized transformation matrix, transform the source point cloud so that the source point cloud is closer to the target point cloud. Then, repeat the nearest point matching and calculation of registration transformation steps until the convergence condition is reached (such as the number of iterations reaches the upper limit or the change of the transformation matrix is less than a certain threshold). Output registration result: finally, the ICP algorithm outputs the optimized transformation matrix, which represents how to register the source point cloud with the target point cloud so that they are as close as possible in space.

[0077] Using the PointNet algorithm for 3D point cloud segmentation of 3D point cloud design models can be achieved through the following steps:

[0078] Data preprocessing: preprocess the collected point cloud data, including denoising, filtering and normalization operations. This can eliminate noise and unnecessary information, and convert the point cloud data to a unified coordinate system for subsequent processing.

[0079] Network design: design a PointNet network model that accepts point clouds as input and outputs the segmentation label or prediction probability of each point. The PointNet network usually consists of multiple layers, including shared fully connected layers, feature conversion layers and global feature extraction layers, etc.

[0080] Feature extraction: input the point cloud data into the PointNet network, which will extract features for each point. The PointNet algorithm captures the structural information of the point cloud by learning the local and global features of each point. These features can include the coordinates, colors, normals, etc. of the points.

[0081] Segmentation prediction: based on the extracted features, the PointNet network can make segmentation predictions for each point. This can be achieved by adding fully connected layers or convolutional layers for classification or segmentation operations. The output segmentation result can be the class label or corresponding segmentation probability of each point.

[0082] Post-processing: Post-processing is performed on the segmentation result, and some algorithms are used to further extract the position information of complex regions, deep cavity local areas and hole sites. Some algorithms include clustering algorithms, morphological operations, region growing and other techniques to identify and locate the regions of interest according to the segmentation result.

[0083] After identifying the 3D positions of the complex regions, deep cavity regions and hole site regions, the rotating table is controlled to rotate and the spray gun is controlled to aim at the complex regions, deep cavity regions or hole site regions according to the 3D positions of the complex regions, deep cavity regions and hole site regions, and the water pressure is increased to realize the one-by-one spray flushing cleaning of the complex regions, deep cavity regions and hole site regions. The complex regions, deep cavity regions and hole site regions are subjected to one-by-one special point positioning for spray flushing and spray washing, so as to ensure that these regions can be fully washed, and after special washing, the to-be-cleaned mold is subjected to visual identification again to judge whether there is a dirty area, if there is a dirty area, the 3D position of the dirty area is subjected to high-pressure washing until there is no dirty area, and then the cleaning is completed. The method realizes automatic cleaning of the mold through visual identification technology, can identify complex regions, deep cavity regions, hole site regions and dirty regions and perform special point cleaning, and improves the cleaning effect of the mold.

[0084] The intelligent control method for mold cleaning is realized based on the intelligent control system shown in Figure 1 The intelligent control system for mold cleaning includes a control unit 1, a rotating table execution mechanism 4 for rotating the rotating table by 360°, a spray gun execution mechanism 3 for adjusting the spray position and angle of the spray gun, and an image acquisition mechanism 2 for acquiring full-view image information of the to-be-cleaned mold, and the control unit 1 includes:

[0085] A model establishing module 11 is configured to perform visual identification on the full-view image information acquired by the image acquisition mechanism 2, and establish a three-dimensional point cloud model of the to-be-cleaned mold;

[0086] A pose recognition module 12 is configured to perform visual identification on the full-view image information acquired by the image acquisition mechanism 2, and acquire pose information of the to-be-cleaned mold in the rotating table coordinate system;

[0087] A complex region recognition module 13 is configured to identify the 3D position of the complex region on the to-be-cleaned mold according to the established three-dimensional point cloud model, pose information and pre-stored 3D point cloud design model;

[0088] A deep cavity region recognition module 14 is configured to identify the 3D position of the deep cavity region on the to-be-cleaned mold according to the established three-dimensional point cloud model, pose information and pre-stored 3D point cloud design model;

[0089] The hole site region identification module 15 is configured to identify the 3D position of the hole site region on the mold to be cleaned according to the established 3D point cloud model, the pose information and the pre-stored 3D point cloud design model.

[0090] The dirty region identification module 16 is configured to identify whether there is a dirty region on the mold to be cleaned according to the established 3D point cloud model, the pose information and the pre-stored 3D point cloud design model, and identify the 3D position of the dirty region when the dirty region exists.

[0091] The angle deviation comparison module 18 is configured to compare the pose information of the mold to be cleaned in the rotating table coordinate system with a preset target pose, and obtain a deviation angle of the mold to be cleaned.

[0092] The control module 17 is configured to control the rotating table actuator 4 to work to correct the mold to be cleaned according to the deviation angle, and output corresponding control instructions to the rotating table actuator 4, the spray gun actuator 3 and the image acquisition mechanism 2 according to the 3D positions of the complex region, the deep cavity region, the hole site region and the dirty region and a preset control strategy. The preset control strategy refers to first controlling the image acquisition mechanism 2 to take images in all directions, for establishing a 3D model of the mold to be cleaned, then performing all-around cleaning on the mold to be cleaned, and then identifying the complex region, the deep cavity region, the hole site region and the dirty region, i.e., the control steps in the control method.

[0093] The image acquisition mechanism 2 includes a camera, which is configured to take images of the mold to be cleaned in all directions. The number of cameras can be one or more. When the number of cameras is one, the position of the camera is adjusted by a moving mechanism to realize taking images of the mold to be cleaned in all directions. When the number of cameras is more than one, the cameras can be arranged around the mold to be cleaned to realize taking images in all directions.

[0094] The specific embodiments described herein are merely illustrative of the spirit of the present application. Those skilled in the art of the present application can make various modifications or supplements to the described specific embodiments or replace them with similar ways, without departing from the spirit of the present application or exceeding the scope defined by the appended claims.

Claims

1. A method for intelligent control of mold cleaning, characterized in that, Comprise the following steps: S1, three-dimensional reconstruction of the mold to be cleaned is carried out by visual recognition technology, the operation includes: placing the mold to be cleaned on the rotating table, collecting the full range of image information of the mold to be cleaned by controlling the rotating table to rotate at 360 degrees, and obtaining a multi-view image sequence of the mold to be cleaned; SIFT algorithm is used to detect the feature points and corresponding feature descriptors of each image in the multi-view image sequence, and the feature descriptors are used to match the feature points of adjacent images to obtain a feature point matching result; Based on the feature point matching result, SFM algorithm is used for sparse reconstruction to obtain a sparse three-dimensional reconstruction result; Based on the sparse three-dimensional reconstruction result and the internal and external parameters of the camera, MVS algorithm is used for dense reconstruction to obtain the 3D point cloud of each pixel point; The 3D point cloud is converted into a 3D point cloud model, so as to obtain a three-dimensional point cloud model of the mold to be cleaned; S2, the mold to be cleaned is cleaned from top to bottom and from left to right; S3, after cleaning, based on the three-dimensional point cloud model obtained in step S1, the complex region, the deep cavity region and the hole region are identified according to the visual complex region identification algorithm, the visual deep cavity identification algorithm and the visual hole identification algorithm respectively, and the pose information of the mold to be cleaned is identified by visual recognition technology, and the 3D positions of the complex region, the deep cavity region and the hole region are determined based on the identified pose information; S4, the water pressure is enhanced, and the 3D positions of the complex region, the deep cavity region and the hole region determined in step S3 are sprayed and washed at high pressure and different angles; S5, the mold to be cleaned is again identified by visual recognition technology, and it is judged whether there is a dirty area in the mold to be cleaned, if there is a dirty area, the 3D position of the dirty area is sprayed and washed at high pressure until there is no dirty area.

2. The intelligent control method for mold cleaning according to claim 1, wherein, In the step S2, before the mold to be cleaned is cleaned from top to bottom and from left to right, the step of adjusting the mold to be cleaned is further included, and the operation includes: Multiple images of the mold to be cleaned at different angles are taken, and the pose information of the mold to be cleaned in the rotating table coordinate system is obtained by analyzing each image; The obtained pose information is compared with the preset target pose to obtain the deviation angle of the mold to be cleaned; The rotating table is controlled to rotate according to the deviation angle until the mold to be cleaned is adjusted.

3. The intelligent control method for mold cleaning according to claim 2, wherein, In the step S2, the operation of obtaining the pose information of the mold to be cleaned in the rotating table coordinate system includes: For each image taken, a feature point detection algorithm is used to extract key points and corresponding key point descriptors in the image; The feature descriptors of the current image are matched with the feature descriptors obtained by using the SFM algorithm for dense reconstruction in step S1 to obtain the matching relationship of the feature descriptors; By analyzing and calculating the matching relationship of the feature descriptors, the pose information of the mold to be cleaned in the rotating table coordinate system in the current image is obtained.

4. The intelligent control method for mold cleaning according to claim 3, wherein, In the step S2, the operation of optimizing the pose information of the mold to be cleaned by nonlinear optimization or least squares is further included.

5. The intelligent control method for mold cleaning according to claim 2 or 3 or 4, characterized in that, In the step S2, the operation of comparing the obtained pose information with the preset target pose includes: The obtained pose information includes a translation vector T c and a rotation matrix R c ; the preset target pose includes a translation vector T t and a rotation matrix R t ; Compute translation vector deviation, obtain translation vector deviation T diff = T c -T t ; The distance between rotation matrices is computed, converted to a rotation angle deviation angle, angle = acos((trace(R c T * R t -1) / 2); Translation vector deviation T diff and the rotation angle deviation angle to form a control parameter for adjusting the alignment of the mold to be cleaned.

6. The intelligent control method for mold cleaning according to any one of claims 1 to 4, characterized in that, In the step S3, the identification operations of the visual complex region identification algorithm, the visual deep cavity identification algorithm, and the visual hole site identification algorithm all include: Obtaining a 3D point cloud design model of the mold to be cleaned; Performing 3D point cloud registration on the three-dimensional point cloud model reconstructed in step S1 and the 3D point cloud design model through an ICP algorithm to obtain a mapping relationship between the 3D point cloud design model and the three-dimensional point cloud model; Performing point cloud segmentation on the 3D point cloud design model using a PointNet algorithm to obtain segmentation results of the complex region, the deep cavity region, and the hole site region, and then obtaining the positions of the complex region, the deep cavity region, and the hole site region on the mold to be cleaned according to the mapping relationship between the 3D point cloud design model and the three-dimensional point cloud model.

7. The intelligent control method for mold cleaning according to any one of claims 1 to 4, characterized in that, In the step S4, the high-pressure different-angle spray cleaning operation for the 3D positions of the complex region, the deep cavity region, and the hole site region determined in step S3 includes: Controlling the rotating table to rotate and controlling the spray gun to aim at the complex region, the deep cavity region, or the hole site region according to the 3D positions of the complex region, the deep cavity region, and the hole site region to realize the one-by-one spray cleaning of the complex region, the deep cavity region, and the hole site region.

8. An intelligent control system for mold cleaning, characterized by, The intelligent control system includes a control unit (1), a rotating table execution mechanism (4) for rotating the rotating table by 360°, a spray gun execution mechanism (3) for adjusting the spray position and angle of the spray gun, and an image acquisition mechanism (2) for acquiring all-around image information of the mold to be cleaned, and the control unit (1) includes: A model establishment module (11) for performing visual identification on the all-around image information acquired by the image acquisition mechanism (2) and establishing a three-dimensional point cloud model of the mold to be cleaned; A pose recognition module (12) for performing visual identification on the all-around image information acquired by the image acquisition mechanism (2) and acquiring pose information of the mold to be cleaned in the rotating table coordinate system; A complex region identification module (13) for identifying the 3D positions of the complex region on the mold to be cleaned according to the established three-dimensional point cloud model, the pose information, and the pre-stored 3D point cloud design model; A deep cavity region identification module (14) for identifying the 3D positions of the deep cavity region on the mold to be cleaned according to the established three-dimensional point cloud model, the pose information, and the pre-stored 3D point cloud design model; A hole site region identification module (15) for identifying the 3D positions of the hole site region on the mold to be cleaned according to the established three-dimensional point cloud model, the pose information, and the pre-stored 3D point cloud design model; A dirty region identification module (16) for identifying whether there is a dirty region on the mold to be cleaned according to the established three-dimensional point cloud model, the pose information, and the pre-stored 3D point cloud design model, and identifying the 3D positions of the dirty region when the dirty region exists; A control module (17) for outputting corresponding control instructions to the rotating table execution mechanism (4), the spray gun execution mechanism (3), and the image acquisition mechanism (2) according to the 3D positions of the complex region, the deep cavity region, the hole site region, and the dirty region and a pre-set control strategy.

9. The intelligent control system for mold cleaning of claim 8, wherein, The control unit (1) further includes: An angle deviation comparison module (18) is configured to compare the pose information of the mold to be cleaned in the rotating table coordinate system with a preset target pose to obtain a deviation angle of the mold to be cleaned. The control module (17) is configured to control the rotating table actuator (4) to work according to the deviation angle to correct the mold to be cleaned.

Images