AI-based nozzle automatic alignment control system

Abstract

The application relates to the technical field of alignment control, and specifically provides an AI-based automatic alignment control system for a material port, which comprises a three-dimensional digital model construction unit, a point cloud segmentation and target identification unit and a path planning and control instruction generation unit, wherein: the three-dimensional digital model construction unit is used for constructing a three-dimensional digital model; the point cloud segmentation and target identification unit is used for segmenting point clouds in the three-dimensional digital model into different categories, and each category of point clouds represents a different area in the three-dimensional digital model; the point clouds of a bulkhead and a tank port are analyzed, and the point clouds of the area where the bulkhead is located are bulkhead area point clouds, and the point clouds of the area where the tank port is located are tank body area point clouds; the path planning and control instruction generation unit is used for defining the area except the bulkhead area as an obstacle area, and analyzing whether a plurality of obstacle areas can be merged, if the plurality of obstacle areas can be merged, the plurality of obstacle areas are merged; and then an operation track of the bulkhead to the tank port is generated, and the operation track is converted into corresponding control instructions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of alignment control, in particular to an AI-based material port automatic alignment control system. BACKGROUND

[0002] In industrial production, tank truck storage and transportation operations of powder and particle materials (such as cement, fly ash, and mineral powder) are widely used in building materials, chemical industry, energy, and other fields. During the tank truck unloading or loading process, the bulkhead of the conveying equipment needs to be accurately aligned with the tank opening to achieve efficient and leak-free material transmission.

[0003] With the rapid development of industrial automation technology, the existing automatic alignment control system usually constructs a three-dimensional digital model of the work scene through laser scanning, visual imaging, and other means. Then, by setting different object recognition templates, the bulkhead coordinates and tank opening coordinates in the three-dimensional digital model are identified. Then, based on the distance difference between the bulkhead coordinates and the tank opening coordinates, kinematics algorithm is used to analyze the movement path of the bulkhead coordinates. Then, the movement path is converted into control instructions such as motor drive and hydraulic control. The control instructions are executed by the motor and telescopic steel rope actuators to move the bulkhead coordinates, and then the bulkhead coordinates and the tank opening coordinates are overlapped, thereby completing the automatic alignment operation.

[0004] However, due to the complexity of the industrial site environment, there are dense equipment layout, messy material storage, and many similar structural components. If a three-dimensional digital model is constructed by a sensing device at this time, there will be similar pseudo-bulkhead coordinates and pseudo-tank opening coordinates in the three-dimensional digital model, which will cause the recognition system to misjudge and fail to accurately distinguish the real bulkhead coordinates and tank opening coordinates from the numerous coordinate information, resulting in inaccurate judgment of the bulkhead coordinates and tank opening coordinates. In view of this, we propose an AI-based material port automatic alignment control system. SUMMARY

[0005] The purpose of the present application is to solve the problem of complex industrial site environment, dense equipment, messy material, and many similar components, which exist in multiple bulkhead and tank opening areas. The three-dimensional model constructed by the sensing device contains pseudo-coordinates, which causes the recognition system to misjudge and makes it difficult to distinguish the real coordinates. There are also forklifts and other obstacles in the field, and the existing path planning algorithm has weak real-time obstacle avoidance capability, which easily causes the bulkhead to collide, resulting in equipment damage, material leakage and pollution, and production interruption.

[0006] To achieve the above purpose, the present application provides an AI-based material port automatic alignment control system, which includes a three-dimensional digital model construction unit, a point cloud segmentation and target recognition unit, and a path planning and control instruction generation unit, wherein:

[0007] The three-dimensional digital model construction unit is configured to construct a three-dimensional digital model; the point cloud segmentation and target identification unit is configured to segment point clouds in the three-dimensional digital model into different categories, each category of point clouds representing a different region in the three-dimensional digital model; and to sequentially fit point clouds in each region that can be constructed into a circular shape into a fitted circle, and then set a radius threshold, call out the fitted circles with radii within the radius threshold, and then calculate the similarity between the fitted circles, if the similarity between two fitted circles is less than a similarity threshold, determine that the fitted circle with a larger radius is a tank mouth and the fitted circle with a smaller radius is a bulkhead, and the point clouds in the region where the bulkhead is located are bulkhead region point clouds and the point clouds in the region where the tank mouth is located are tank body region point clouds.

[0008] The path planning and control instruction generation unit is configured to define the regions in the point cloud segmentation and target identification unit other than the bulkhead region as obstacle regions, and analyze whether the plurality of obstacle regions can be merged, if the plurality of obstacle regions can be merged, merge the plurality of obstacle regions; and then generate a plurality of running trajectories of the bulkhead to the tank mouth, call out the running trajectory with the smallest running distance as the running trajectory of the bulkhead to the tank mouth, and convert the corresponding control instruction.

[0009] As a further improvement of the technical solution, the three-dimensional digital model construction unit includes a laser camera distance measurement module and a point cloud conversion and modeling module; the laser camera distance measurement module is configured to measure the target distance of laser beams propagating to the surfaces of different objects using a 3D laser camera, and the point cloud conversion and modeling module is configured to construct a three-dimensional digital model according to the distances of the surfaces of different objects.

[0010] As a further improvement of the technical solution, the 3D laser camera in the laser camera distance measurement module includes a transmitting end and a receiving end, and when measuring the target distance, the transmitting end transmits a high-frequency laser beam to the surface of an object at different transmission angles; when the laser beam reaches the surface of the object, the surface of the object will reflect part of the high-frequency laser beam to form reflected light, and the receiving end of the 3D laser camera receives part of the reflected light; and records the time difference of the laser beam from the transmitting end to the receiving end.

[0011] The point cloud conversion and modeling module is configured to perceive the time difference of the laser beam from the transmitting end to the receiving end, multiply the speed of light by the time difference, and then divide by two to calculate the target distance of the transmitting end to the surface of the object; define the transmitting end of the 3D laser camera as the origin of the three-dimensional digital model; and sequentially multiply each target distance by the cosine value of the vertical angle and then by the cosine value of the horizontal angle to obtain the component of the three-dimensional coordinates; multiply the target distance by the cosine value of the vertical angle and then by the sine value of the horizontal angle to obtain the component of the three-dimensional coordinates; and multiply the target distance by the sine value of the vertical angle to obtain the component of the three-dimensional coordinates, thereby converting the target distance into point clouds in the three-dimensional digital model.

[0012] As a further improvement of the technical solution, the point cloud segmentation and target recognition unit comprises a regional point cloud clustering module, a circular feature fitting and center positioning module, and a bulkhead and tank opening feature matching and division module.

[0013] The regional point cloud clustering module is configured to receive a three-dimensional digital model, randomly select a point cloud in the three-dimensional digital model as an initial point cloud, take all point clouds other than the initial point cloud as adjacent point clouds, and sequentially calculate the spatial distance between the initial point cloud and the adjacent point clouds, and analyze the same type of point clouds using the spatial distance.

[0014] The circular feature fitting and center positioning module is configured to set a plurality of center point clouds and edge radii in each region, take the center point clouds as the center, call out the radius point clouds within the edge radii, then construct the covariance matrix of the radius point clouds relative to the center point clouds, eigenvalue decompose the covariance matrix, calculate the curvature, set a curvature threshold, if the curvature is within the curvature threshold, the corresponding point cloud is retained, otherwise the corresponding point cloud is filtered; and then analyze the circles in different regions according to the retained point clouds.

[0015] The bulkhead and tank opening feature matching and division module is configured to set a radius threshold, call out the fitted circles within the radius threshold in each region, sequentially calculate the radius difference between the called out fitted circles in different regions as the similarity, if the similarity < a similarity threshold, then determine that the fitted circle with a larger radius is a tank opening and the fitted circle with a smaller radius is a bulkhead, and the point cloud of the region where the bulkhead is located is the bulkhead region and the point cloud of the region where the tank opening is located is the tank body region.

[0016] As a further improvement of the technical solution, the regional point cloud clustering module is configured to sequentially calculate the square of the axial coordinate difference between the initial point cloud and the adjacent point clouds in 、 and axis coordinates, then add the square of the difference, take the square root of the addition result to obtain the spatial distance between the initial point cloud and the adjacent point cloud, and set a distance threshold, if the spatial distance < the distance threshold, then determine that the initial point cloud data and the adjacent point cloud are the same type of point clouds.

[0017] Then take the initial point cloud as the center, select the point cloud farthest from the initial point cloud in the same type of point clouds again as a secondary point cloud, calculate the spatial distance between the secondary point cloud and the adjacent point clouds, and compare it with the distance threshold again, if the spatial distance < the distance threshold, then again classify the adjacent point clouds corresponding to the secondary point cloud as the same type of point clouds; if the spatial distance between the secondary point cloud and the adjacent point clouds ≥ the distance threshold, then take the secondary point cloud as the edge point cloud of the same type of point clouds, and stop calculating the spatial distance.

[0018] Then, point clouds other than those classified into the same category are selected as the initial point cloud, and the spatial distance between the initial point cloud and adjacent point clouds is calculated. The spatial distance is then repeatedly compared with the distance threshold to segment the point cloud in the 3D digital model into different categories.

[0019] As a further improvement to this technical solution, the circular feature fitting and center positioning module is used to randomly select a number of points from the retained point cloud. The point cloud, based on the selected point cloud, calculate The average coordinates of each point cloud are used as candidate circle centers. The distance from each retained point cloud to the candidate circle center is calculated. If the distances from all retained point clouds to the candidate circle center are the same, it is determined that the candidate circle center can be constructed into a circle, and the candidate circle center is defined as the circle center. Otherwise, a candidate circle center is selected again, and it is determined whether the selected candidate circle center can construct a circle. If there are no retained point clouds where the distances from all retained point clouds to the candidate circle center are the same, the center point cloud and edge radius set in each region of the region point cloud clustering module are adjusted, and the circles and their corresponding circle center coordinates in different regions are analyzed.

[0020] As a further improvement to this technical solution, the path planning and control command generation unit includes an obstacle area definition and restricted area generation submodule, a restricted area merging and analysis submodule, and an optimal running trajectory planning module. The obstacle area definition and restricted area generation submodule defines all areas except the bulk head area as obstacle areas and generates multiple restricted areas. The restricted area merging and analysis submodule calculates the distance between different obstacle areas and the maximum distance between different coordinate axes of the bulk head area point cloud, analyzing whether different obstacle areas can be merged. The optimal running trajectory planning module defines the coordinates in the 3D digital model excluding the restricted areas as the search area, and generates the bulk head point cloud to the can opening point cloud within the search area. Multiple running trajectories are selected, and the trajectory with the shortest running distance is chosen as the running trajectory from the bulk head point cloud to the can opening point cloud, and then converted into control command output.

[0021] As a further improvement to this technical solution, the obstacle area definition and restricted area generation submodule is used to define all areas except the bulk head area as obstacle areas, set a safe distance threshold, and use the safe distance threshold plus the point cloud of each obstacle area to generate multiple restricted areas.

[0022] As a further improvement to this technical solution, the restricted area merging and analysis submodule is used to first calculate the maximum distance between different coordinate axes of the point cloud of the bulk head region, respectively in... , and Along the axes, retrieve the maximum distance between any two point clouds within the bulk head region, i.e., the maximum span along each coordinate axis; then from... , and The maximum value is selected from the maximum spans corresponding to the axes, which is the maximum distance statistic of the point cloud of the bulk head region in different coordinate axis directions;

[0023] Calculate the minimum distance between each point cloud in the obstacle region and the point cloud in the obstacle region, and then find the maximum distance value among the minimum distances; at the same time, calculate the minimum distance between each point cloud in the obstacle region and find the maximum distance value, and finally take the maximum value of the two maximum distance values ​​to obtain the distance between the point clouds in the two obstacle regions.

[0024] If the distance between two obstacle point clouds plus a safety distance threshold is less than the maximum distance between different coordinate axes of the bulk head point cloud, then the two obstacle point clouds are merged to construct a restricted area.

[0025] As a further improvement to this technical solution, the optimal trajectory planning module is used to define the coordinates in the 3D digital model, excluding the restricted area, as the search area, with probability... Directly sample the point cloud at the tank opening, using probability Randomly sample point clouds within the search area and generate multiple running trajectories from the bulk head point cloud to the can mouth point cloud within the search area;

[0026] Calculate the distance between multiple trajectories, and select the trajectory with the smallest distance as the bulk head point cloud. Pointing clouds at the mouth of the jar The trajectory of the operation is recorded and converted into control commands for output.

[0027] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention analyzes different regions in the three-dimensional digital model through the regional point cloud clustering module, so that the circular feature fitting and center positioning module can not only analyze the circles in different regions sequentially to determine the coordinates of the bulk head and the can opening, but also analyze the bulk head region and the can body region in different regions during the process of determining the coordinates of the bulk head and the can opening, avoiding the problem of misjudgment by the recognition system due to the existence of multiple bulk head and can opening regions, and solving the problem of difficulty in distinguishing the real coordinates; it also enables the obstacle region definition and restricted area generation submodule to accurately analyze the obstacle region based on the bulk head region and other regions in the regional point cloud clustering module.

[0029] And the obstacle area definition and no-entry area generation submodule analyzes the obstacle area according to the bulkhead area and other area in the area point cloud clustering module, combines the no-entry area merging analysis submodule to adjust the obstacle area division mode, when the subsequent optimal running track planning module plans the running track, the interference of the bulkhead and the tank body itself on the obstacle judgment is excluded, and the obstacle area safety threshold can be dynamically checked according to the actual size and motion characteristics of the bulkhead, the defects of the existing path planning algorithm are overcome, the collision between the bulkhead and the forklift and other obstacles is avoided, the equipment damage, material leakage pollution and production interruption and other problems caused by the collision are avoided, the running track planning is upgraded from 'blind obstacle avoidance' to 'intelligent planning based on accurate area cognition', the movable boundary of the bulkhead and the absolute no-entry area are determined in advance, and the accuracy, safety and efficiency of the bulkhead movement of the conveying equipment in the tank car unloading or loading link are ensured. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 It is the overall module schematic diagram of the application;

[0031] Figure 2 It is the working principle flow chart of the three-dimensional digital model construction unit of the application;

[0032] Figure 3 It is the working principle flow chart of the point cloud segmentation and target recognition unit of the application;

[0033] Figure 4 It is the working principle flow chart of the path planning and control instruction generation unit of the application;

[0034] Figure 5 It is the detailed working principle schematic diagram of the area point cloud clustering module and the bulkhead tank opening feature matching and division module of the application;

[0035] Figure 6 It is the working principle schematic diagram of the point cloud segmentation and target recognition unit and the path planning and control instruction generation unit of the application;

[0036] Figure 7 It is the 3D laser camera identification schematic diagram of the application.

[0037] The meanings of various labels in the figure are:

[0038] 100, three-dimensional digital model construction unit; 110, laser camera ranging module; 120, point cloud conversion and modeling module; 200, point cloud segmentation and target recognition unit; 210, regional point cloud clustering module; 220, circular feature fitting and center positioning module; 230, bulk head opening feature matching and division module; 300, path planning and control instruction generation unit; 310, obstacle region definition and no-entry zone generation submodule; 320, no-entry zone merging and analysis submodule; 330, optimal running trajectory planning module. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0040] Reference Figures 1-7 As shown, the AI-based opening automatic alignment control system includes a three-dimensional digital model construction unit 100, a point cloud segmentation and target recognition unit 200, and a path planning and control instruction generation unit 300, wherein:

[0041] The three-dimensional digital model construction unit 100 includes a laser camera ranging module 110 and a point cloud conversion and modeling module 120. The laser camera ranging module 110 measures the distance of laser beam propagation to different object surfaces by using a 3D laser camera, and then constructs a three-dimensional digital model reflecting the actual space state. The specific working principle is as follows: the 3D laser camera includes a transmitting end and a receiving end. When measuring the target distance, the transmitting end emits a high-frequency laser beam (usually in the infrared or near-infrared wave band, with a wavelength of 905 nm or 1550 nm) to the object surface (such as bulk head and opening) at different transmission angles (horizontal angle and vertical angle );

[0042] When the laser beam reaches the object surface, the molecules and atoms of the object surface will interact with the laser beam. Part of the laser beam will change the propagation direction due to the specular reflection (such as regular reflection of smooth metal surface) and diffuse reflection (such as irregular reflection of rough bulk head) of the object surface, and will form reflected light from the object surface. The receiving end of the 3D laser camera receives part of the reflected light, and records the time difference between the laser beam from the transmitting end to the receiving end.

[0043] The point cloud conversion and modeling module 120 perceives the time difference between the laser beam from the transmitting end to the receiving end, and calculates the target distance from the transmitting end to the object surface: wherein is the speed of light; and defining the 3D laser camera emitting end as the origin of the three-dimensional digital model, the target distance of each point on the object surface is converted into the point cloud in the three-dimensional digital model through the target distance and the emitting angle (horizontal angle and vertical angle );

[0044] ;

[0045] wherein , and are the axis coordinates, axis coordinates and axis coordinates of the point cloud in the three-dimensional digital model, is the cosine value of the vertical angle , is the cosine value of the horizontal angle , is the sine value of the vertical angle .

[0046] Specifically, after opening the cover above the can body, the can opening area changes from a plane to a concave space. The application measures the distance of the laser beam propagating to the surface of different objects by a 3D laser camera, and then constructs a three-dimensional digital model to realize the mapping from the physical space to the digital model, providing a high-precision spatial data basis for subsequent industrial scenes (such as bulk head, can opening recognition and alignment), and the 3D laser camera avoids the situation that the three-dimensional digital model is not accurate due to complex environment (such as dim light, dust interference, and various object surface materials), ensuring the accuracy and spatial restoration degree of the three-dimensional digital model, so that the three-dimensional digital model can accurately present the background, can body, bulk head and other areas, supporting subsequent trajectory planning, collision avoidance and other functions, ensuring the reliability and efficiency of industrial intelligent application from the data acquisition source, realizing effective conversion of physical space information to digital instructions, and assisting stable operation of automatic production process.

[0047] The point cloud segmentation and target recognition unit 200 includes a regional point cloud clustering module 210, a circular feature fitting and center positioning module 220, and a bulk head and can opening feature matching and division module 230.

[0048] Since the same region (such as the same object surface or continuous structure) is a complete and continuous entity in the physical space, and there is no significant physical fracture or interval on the surface, when constructing a three-dimensional digital model, the 3D laser camera will uniformly measure the point cloud on the object surface according to a certain density, and the point cloud will be closely arranged in space to reflect the continuous form of the object surface.

[0049] Based on the above characteristics, the region point cloud clustering module 210 receives a three-dimensional digital model, randomly selects a point cloud in the three-dimensional digital model as an initial point cloud, sequentially calculates the spatial distance between the initial point cloud and all point clouds other than the initial point cloud (hereinafter referred to as adjacent point clouds), and sets a distance threshold. If the spatial distance is less than the distance threshold, it indicates that the initial point cloud and the adjacent point cloud are in a continuous distribution state in three-dimensional space, their spatial distance is close enough, and they meet the point cloud distribution characteristics of the same physical region (such as the surface of the same object) (i.e. belong to the same continuous surface or part of the structure), so the initial point cloud data and the adjacent point cloud are judged as the same type of point cloud;

[0050] Then, taking the initial point cloud as the center, the point cloud farthest from the initial point cloud in the same type of point cloud is selected again as a secondary point cloud, the spatial distance between the secondary point cloud and the adjacent point cloud is calculated, and the spatial distance is compared with the distance threshold again. If the spatial distance is still less than the distance threshold, the adjacent point cloud corresponding to the secondary point cloud is classified as the same type of point cloud again. If the spatial distance between the secondary point cloud and the adjacent point cloud is greater than or equal to the distance threshold, the secondary point cloud is regarded as an edge point cloud of the same type of point cloud, and the calculation of the spatial distance is stopped.

[0051] Then, the point cloud other than the same type of point cloud is selected again as the initial point cloud, the spatial distance between the initial point cloud and the adjacent point cloud is calculated, and the spatial distance is compared with the distance threshold repeatedly, so as to divide the point clouds in the three-dimensional digital model into different categories, wherein each category of point cloud represents a different region (background region, tank body region, bulk head region, etc.) in the three-dimensional digital model. With the help of point cloud spatial continuity and distance threshold precision judgment, the automatic and fine classification of each point cloud in the three-dimensional digital model is realized, which not only ensures the accurate aggregation of the same type of region point cloud (which conforms to the actual distribution of the continuous surface of the object), but also identifies the region edge (when the distance between the secondary point cloud and the adjacent point cloud exceeds the threshold, the edge point cloud is marked), lays a foundation for the subsequent accurate identification and distinction of the background, tank body, bulk head and other regions in the three-dimensional digital model, and upgrades the analysis of the spatial structure of the three-dimensional digital model from "unordered point cloud" to "structured region cognition" in the industrial scene (such as bulk head alignment and obstacle recognition), which not only improves the accuracy and automation of region classification, but also adapts to complex surface morphology (such as tank body curved surface and irregular bulk head structure) through iterative optimization, and also provides reliable region semantic information for subsequent intelligent control based on three-dimensional digital model (such as bulk head automatic obstacle avoidance and precise alignment), which promotes the accuracy and efficiency of industrial intelligent application from the source of point cloud processing.

[0052] In order to accurately identify the bulk head region and the tank opening region in the three-dimensional digital model, the circular feature fitting and center positioning module 220 sets a plurality of center point clouds , edge radius in each region For the center, call out the edge radius The radius point cloud inside Wherein The number of radius point clouds;

[0053] Construct a radius point cloud Relative to the center point cloud The covariance matrix is:

[0054] ;

[0055] Wherein The mean point of the radius point cloud The mean point The corresponding coordinates are ;

[0056] Specifically, the covariance matrix is used to describe the distribution characteristics (discreteness, correlation, principal direction) of the radius point cloud in the three-dimensional digital model, and the covariance matrix is expanded as follows:

[0057] ;

[0058] Wherein , wherein and are , , Any variable in and is the mean of variable and variable ;

[0059] Eigenvalue decomposition of the covariance matrix , wherein , wherein is a diagonal matrix, the elements on the diagonal are the eigenvalues of , the eigenvalues satisfy ; the elements on the diagonal are the eigenvalues of , and is an eigenvector matrix, and the column vectors are the eigenvectors (representing three orthogonal directions) of ;

[0060] Calculate the curvature: , wherein is the smallest eigenvalue, is the total dispersion intensity; set the curvature threshold , if the curvature is in the curvature threshold If the corresponding point cloud is inside, it will be retained; otherwise, the corresponding point cloud will be filtered out.

[0061] The circular feature fitting and center localization module 220 is also used to randomly select a number of points from the retained point cloud. (Select quantity) Point cloud Based on the selected point cloud ,calculate The average coordinates of the point cloud were used as candidate circle centers. ;

[0062] Calculate all retained point clouds to candidate circle centers distance If all retained point clouds are moved to the candidate circle center The distances are equal, which meets the geometric definition of a circle, therefore, the candidate circle center is determined. It can be constructed as a circle, with the candidate circle center... Define it as the center of the circle; otherwise, select another candidate center. And determine the candidate center. Can it be constructed as a circle? If, within the preserved point cloud, there exist one or more preserved point clouds that extend to the candidate circle's center. When the distances are not equal, the point cloud center points in each region of the region point cloud clustering module 210 are adjusted. and edge radius It can still work stably in complex and ever-changing industrial environments (such as tank deformation and bulk head wear), which enhances the robustness and adaptability of the system, ensures continuous and stable production processes, reduces downtime and adjustment time caused by environmental changes, and indirectly improves production efficiency and capacity. It can also analyze the circles and their corresponding center coordinates in different areas.

[0063] The bulk head tank opening feature matching division module 230 sets a radius threshold value (the radius threshold value can be set by the staff according to the actual bulk head radius, tank opening radius, and can accurately distinguish the bulk head and tank opening area), calls out the fitting circle in each area within the radius threshold value, and sequentially calculates the radius difference between the called out fitting circles in different areas as the similarity, if the similarity < similarity threshold value, since the tank opening is the interface of the material storage container (such as the powder and particle material tank truck tank body), it needs to meet the "large capacity material transmission" demand, so the size is usually larger (convenient for quick loading and unloading, and suitable for a variety of bulk head docking); and the bulk head is the material output end (such as the conveying equipment, filling head), limited by the equipment structure and material flow control demand, the size is relatively smaller (precise control of material output, and adaptation to the large opening of the tank opening), so the fitting circle with large radius is judged as the tank opening, and the fitting circle with small radius is judged as the bulk head, and the point cloud of the area where the bulk head is located is the bulk head area, and the point cloud of the area where the tank opening is located is the tank body area, the bulk head and the tank opening are distinguished by comparing the radius difference, so that the recognition result is highly consistent with the actual production demand, the interface recognition problem caused by equipment wear, deformation and the like in a complex industrial environment is solved, the basic data accuracy of the bulk head alignment and material transmission link is ensured, and the manual checking cost and error are reduced.

[0064] In order to avoid collision between the bulk head and the obstacle area (such as tank body, equipment, etc.) when the bulk head moves, the path planning and control instruction generation unit 300 includes an obstacle area definition and forbidden area generation submodule 310, a forbidden area merging analysis submodule 320 and an optimal running trajectory planning module 330, wherein the obstacle area definition and forbidden area generation submodule 310 defines all areas except the bulk head area as obstacle areas , sets a safety distance threshold value , adds the point cloud of each obstacle area to the safety distance threshold value to generate a plurality of forbidden areas , wherein is the unit normal vector of the obstacle surface, is the safety distance in the normal direction.

[0065] The application further considers that, due to the close distance between different obstacle areas, if fusion processing is not performed during the movement of the bulk head, the bulk head may be misjudged as having a passable space but actually cannot pass safely, or even cause collision risk between the bulk head and the obstacle, therefore the forbidden area merging analysis submodule 320 first calculates the maximum distance between the bulk head area point cloud in different coordinate axes:

[0066] ;

[0067] wherein The maximum distance statistics of the bulkhead region point cloud in different coordinate axis directions, The bulkhead region, The bulkhead region The midpoint cloud, The bulkhead region The maximum span of each point cloud in The maximum span of each point cloud in The maximum distance of the bulkhead region in different coordinate axes in the three-dimensional digital model is obtained by taking the maximum value of the maximum span in three coordinate axis directions, reflecting the overall spatial size of the bulkhead region

[0068] Then the distance between the two obstacle region point clouds is calculated The specific working principle is to first calculate the minimum distance from each point cloud in the obstacle region to the point cloud in the obstacle region , and then find the maximum distance value in the minimum distance; at the same time, the minimum distance of each point cloud in the obstacle region is calculated and the maximum distance value is found, and finally the maximum value of the two maximum distance values is taken to obtain the distance between the two obstacle region point clouds The specific expression is:

[0069] ;

[0070] Wherein , is the point cloud in the first and the second obstacle region, is the minimum upper bound in the obstacle region, is the maximum lower bound in the obstacle region, is the Euclidean distance between the point cloud and the point cloud ;

[0071] If the distance between the two obstacle region point clouds plus a safety distance threshold is less than the maximum distance between the bulkhead point clouds in different coordinate axes , the two obstacle point clouds are fused to construct a forbidden region , so as to optimize the division of the forbidden region and improve the safety and efficiency of subsequent bulkhead motion trajectory planning, avoiding planning failure or collision accidents caused by scattered obstacle regions;

[0072] The optimal running trajectory planning module 330 defines the coordinates in the three-dimensional digital model except the forbidden region as a search region , and the probability is calculated (such as Direct sampling of point cloud at tank opening (In the circular feature fitting and center localization module 220, when fitting a circle to determine the circular opening of the can, the corresponding candidate center is determined.) ), in probability In the search area Internal random sampling point cloud The sampling distribution can be formalized as follows:

[0073] ;

[0074] in Using the Dirac function ensures priority exploration towards the target. For search area The sampling distribution of a point cloud is randomly selected, with each point having an equal probability of being selected. This is achieved by using probability... Prioritize sampling tank opening point cloud Using the Dirac function Prioritize exploration towards the target area, quickly focus on meaningful regions, and avoid blind searching; at the same time, use probability... Random sampling within the search area, combined with global exploration, ensures rapid identification of the target path and discovery of potential better routes in complex industrial environments (such as those with obstacles or changing regional features), thereby improving the efficiency and comprehensiveness of path planning. This enables the generation of a point cloud of bulk head within the search area (the candidate center corresponding to the circle fitted in the circular feature fitting and center localization module 220 when determining the circle of the bulk head). ) to the mouth of the can and point the cloud Multiple running trajectories: ,in Adding clouds to the bulk head, , For the intermediate point cloud of the running trajectory;

[0075] Calculate the running distance between multiple running trajectories: Select the trajectory with the shortest running distance as the bulk head point cloud. Pointing clouds at the mouth of the jar trajectory The converted control instructions are then output to the PLC controller. ,in This is the mechanical transmission ratio. The initial length of the steel rope is obtained by actuator calibration, and the planned path is accurately landed as a motion instruction executable by the device through the actuator calibration parameters, solving the conversion problem from "digital model path" to "physical device motion". The device can still operate stably according to the planned trajectory in complex scenarios such as tank deformation and bulkhead wear, ensuring the reliability of the bulkhead alignment, material transfer and other production processes, reducing problems such as collision and docking failure caused by path deviation, and improving the automation level and stability of industrial production.

[0076] Output control instructions to the PLC controller, which drives the actuator (composed of a motor and a telescopic steel rope, etc.) to drive the bulkhead to move to the tank opening point cloud . .

[0077] Example two: deployment phase of the discharging unit:

[0078] After the bulkhead is in place, the PLC controller drives the actuator to control the discharging unit to lower; during the lowering process, the three-dimensional digital model construction unit 100 updates the three-dimensional digital model in real time to dynamically monitor whether the discharging unit is inclined (to determine whether the material pipe is at a normal operating angle) and whether the material pipe is limited (to determine the lowering distance of the material pipe);

[0079] If the material pipe is found to be deviated too much (beyond the preset deviation threshold), an abnormal handling logic is triggered immediately: stop the current discharging lowering action, control the discharging unit to retract to the initial position first, and then perform the positioning and detection phase again through the point cloud segmentation and target recognition unit 200 and the path planning and control instruction generation unit 300, and perform the bulkhead flow process such as 3D camera shooting and bulkhead walking again to ensure the accuracy of subsequent deployment; if no deviation anomaly occurs, when the material pipe reaches the lower limit and the inclination detection is normal, it means that the discharging unit has been accurately deployed in place.

[0080] Discharging phase: after the material pipe is in place, the system can also automatically start the relevant devices (including dust collection fans, Roots blowers, flow valves, pneumatic valves, etc.) through interlocking logic to start discharging; during the discharging process, the control system dynamically adjusts the opening of the flow valve according to real-time flow data to ensure that the discharging flow always does not exceed the preset threshold, avoiding material overflow or equipment overload caused by excessive flow.

[0081] Operation termination and device shutdown phase: when the weighing system detects that the material weight reaches the target value, the interlocking logic is triggered to automatically shut down all running devices (dust collection fans, Roots blowers, flow valves, pneumatic valves, etc.) and stop discharging.

[0082] Pipe collection and zero stage: after the blanking stops, the control pipe is collected, and the pipe limiting state is monitored in real time; when the pipe reaches the upper limit (collected to the position), the bulk head starts to move to the initial coordinates; after the bulk head returns to the initial position, the whole loading operation is completed.

[0083] Whole process monitoring and abnormal processing: during the whole operation process, the running states of all devices (such as motor speed, valve switch state, fan air pressure, etc.) are continuously monitored; if a device failure or parameter abnormality (such as pipe inclination abnormality, flow threshold value, etc.) is detected, an alarm is triggered and the machine is stopped immediately to prevent the accident from expanding.

[0084] Different operation stages are precisely controlled, monitored and responded in real time through specific technologies, and the core technology corresponds as follows:

[0085] Operation stage Core technology support Tube deployment stage 1. 3D camera recognition technology (analyze circles in different areas, corresponding to the center coordinates of the circle, provide data basis for initial positioning and repositioning after deviation); 2. Servo positioning technology (control the walking of the bulk head, the precision of the tube lowering / raising, including the control of the reset after deviation); 3. Tube posture detection technology (real-time monitoring of tube inclination and deviation, identifying deviation abnormalities); 4. Sensor monitoring technology (monitoring the upper and lower limits of the tube, judging the deployment / raising in place state); Discharging stage 1. Interlocking control technology (realize the "synchronous start and stop" of the equipment, and guarantee the logical coherence of the operation); 2. Sensor monitoring technology (real-time monitoring of flow, weight, and equipment operating parameters, providing data support for flow regulation and abnormality identification); Tube collecting stage 1. Servo positioning technology (control the precision of tube raising and bulk head resetting); 2. Sensor monitoring technology (monitor the upper limit of the tube to confirm the raising in place);

[0086] The above shows and describes the basic principle, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. An AI-based automatic feed inlet alignment control system, characterized in that, It includes a 3D digital model building unit (100), a point cloud segmentation and target recognition unit (200), and a path planning and control command generation unit (300), wherein: The three-dimensional digital model building unit (100) is used to build a three-dimensional digital model; The point cloud segmentation and target recognition unit (200) is used to segment the point cloud in the three-dimensional digital model into different categories, each category of point cloud representing a different region in the three-dimensional digital model; and sequentially fits the point cloud that can be constructed into a circle in each region into a fitted circle, then sets a radius threshold, selects the fitted circles whose radius is within the radius threshold, and then calculates the similarity between the fitted circles. If the similarity between two fitted circles is less than the similarity threshold, then the fitted circle with the larger radius is judged to be the can opening, and the fitted circle with the smaller radius is judged to be the bulk head. At the same time, the point cloud of the region where the bulk head is located is the bulk head region, and the point cloud of the region where the can opening is located is the can body region. The point cloud segmentation and target recognition unit (200) includes a regional point cloud clustering module (210), a circular feature fitting and center localization module (220), and a bulk head can mouth feature matching and segmentation module (230). The regional point cloud clustering module (210) is used to receive the three-dimensional digital model, randomly select a point cloud in the three-dimensional digital model as the initial point cloud, take all point clouds other than the initial point cloud as the adjacent point clouds, and calculate the spatial distance between the initial point cloud and the adjacent point clouds in turn, and use the spatial distance to analyze the same type of point cloud. The circular feature fitting and center localization module (220) is used to set multiple center point clouds and edge radii in each region. With the center point cloud as the center, the radius point cloud within the edge radius is retrieved. Then, the covariance matrix of the radius point cloud relative to the center point cloud is constructed. The covariance matrix is ​​decomposed by eigenvalues, and the curvature is calculated. A curvature threshold is set. If the curvature is within the curvature threshold, the corresponding point cloud is retained; otherwise, the corresponding point cloud is filtered out. Then, the circles in different regions are analyzed based on the retained point clouds. The bulk head can mouth feature matching and division module (230) is used to set a radius threshold, call up the fitted circle within the radius threshold in each region, and calculate the radius difference between the fitted circles in different regions as the similarity. If the similarity is less than the similarity threshold, the fitted circle with the larger radius is judged to be the can mouth and the fitted circle with the smaller radius is the bulk head. At the same time, the point cloud of the region where the bulk head is located is the bulk head region and the point cloud of the region where the can mouth is located is the can body region. The path planning and control command generation unit (300) is used to define the area in the point cloud segmentation and target recognition unit (200) other than the bulk head area as the obstacle area, and analyze whether multiple obstacle areas can be merged. If they can be merged, they are merged. Then, multiple running trajectories from the bulk head to the can opening are generated, the running trajectory with the shortest running distance is selected as the running trajectory from the bulk head to the can opening, and the corresponding control command is converted. The path planning and control command generation unit (300) includes an obstacle area definition and restricted area generation submodule (310), a restricted area merging and analysis submodule (320), and an optimal running trajectory planning module (330). The obstacle area definition and restricted area generation submodule (310) is used to define all areas except the bulk head area as obstacle areas and generate multiple restricted areas. The restricted area merging and analysis submodule (320) is used to calculate the distance between different obstacle areas, the maximum distance between different coordinate axes of the bulk head area point cloud, and analyze whether different obstacle areas can be merged. The optimal running trajectory planning module (330) is used to define the coordinates in the three-dimensional digital model except for the restricted areas as the search area, and generate the bulk head point cloud to the can mouth point cloud within the search area. Multiple running trajectories are selected, and the trajectory with the shortest running distance is chosen as the running trajectory from the bulk head point cloud to the can opening point cloud, and then converted into control command output.

2. The AI-based automatic feed inlet alignment control system according to claim 1, characterized in that: The three-dimensional digital model building unit (100) includes a laser camera ranging module (110) and a point cloud conversion and modeling module (120); the laser camera ranging module (110) is used to measure the target distance of the laser beam propagating to the surface of different objects using a 3D laser camera, and the point cloud conversion and modeling module (120) is used to build a three-dimensional digital model based on the distance to the surface of different objects.

3. The AI-based automatic feed inlet alignment control system according to claim 2, characterized in that: The 3D laser camera in the laser camera ranging module (110) includes a transmitter and a receiver. When measuring the distance to the target, the transmitter emits a high-frequency laser beam at different emission angles toward the surface of the object. When the laser beam reaches the surface of the object, the surface of the object will reflect part of the high-frequency laser beam to form reflected light. The receiver of the 3D laser camera receives part of the reflected light and records the time difference between the laser beam from the transmitter to the receiver. The point cloud conversion and modeling module (120) is used to sense the time difference between the laser beam and the receiver, calculate the target distance from the transmitter to the object surface by multiplying the speed of light by the time difference and then dividing by two; define the 3D laser camera transmitter as the origin of the three-dimensional digital model; and sequentially multiply each target distance by the cosine of the vertical angle and then by the cosine of the horizontal angle to obtain the three-dimensional coordinates. The components are calculated by multiplying the target distance by the cosine of the vertical angle, then by the sine of the horizontal angle, to obtain the three-dimensional coordinates. Quantity; Multiplying the target distance by the sine of the vertical angle yields the three-dimensional coordinates. The components are used to convert the target distance into a point cloud in a three-dimensional digital model.

4. The AI-based automatic feed inlet alignment control system according to claim 3, characterized in that: The regional point cloud clustering module (210) is used to sequentially calculate the initial point cloud and adjacent point clouds in terms of clustering. , and The square of the difference between the axis coordinates is calculated, and then the squares of the differences are added together. The square root of the sum is taken to obtain the spatial distance between the initial point cloud and the adjacent point cloud. A distance threshold is set. If the spatial distance is less than the distance threshold, the initial point cloud data and the adjacent point cloud are determined to be of the same type. Then, using the initial point cloud as the center, the point cloud with the longest spatial distance from the initial point cloud among the same type of point cloud is selected as the secondary point cloud. The spatial distance between the secondary point cloud and its adjacent point clouds is calculated and compared with the distance threshold again. If the spatial distance is still less than the distance threshold, the adjacent point clouds corresponding to the secondary point cloud are reclassified as the same type of point cloud. If the spatial distance between the secondary point cloud and its adjacent point clouds is greater than or equal to the distance threshold, the secondary point cloud is regarded as the edge point cloud of the same type of point cloud, and the calculation of spatial distance is stopped. Then, point clouds other than those classified into the same category are selected as the initial point cloud, and the spatial distance between the initial point cloud and adjacent point clouds is calculated. The spatial distance is then repeatedly compared with the distance threshold to segment the point cloud in the 3D digital model into different categories.

5. The AI-based automatic feed inlet alignment control system according to claim 4, characterized in that: The circular feature fitting and center localization module (220) is used to randomly select a number of points from the retained point cloud. The point cloud, based on the selected point cloud, calculate The average coordinates of each point cloud are used as candidate circle centers. The distance from each retained point cloud to the candidate circle center is calculated. If the distances from all retained point clouds to the candidate circle center are the same, it is determined that the candidate circle center can be constructed into a circle, and the candidate circle center is defined as the circle center. Otherwise, the candidate circle center is selected again, and it is determined whether the selected candidate circle center can be constructed into a circle. If there are no retained point clouds with the same distances from the candidate circle center, the circle center point cloud and edge radius set in each region of the region point cloud clustering module (210) are adjusted, and the circle and the circle center coordinates corresponding to the circle in different regions are analyzed.

6. The AI-based automatic feed inlet alignment control system according to claim 5, characterized in that: The obstacle area definition and restricted area generation submodule (310) is used to define all areas except the bulk head area as obstacle areas, set a safe distance threshold, and use the safe distance threshold plus the point cloud of each obstacle area to generate multiple restricted areas.

7. The AI-based automatic feed inlet alignment control system according to claim 6, characterized in that: The restricted area merging analysis submodule (320) is used to first calculate the maximum distance between different coordinate axes of the point cloud of the bulk head region, respectively in , and Along the axes, retrieve the maximum distance between any two point clouds within the bulk head region, i.e., the maximum span along each coordinate axis; then from... , and The maximum value is selected from the maximum spans corresponding to the axes, which is the maximum distance statistic of the point cloud of the bulk head region in different coordinate axis directions; Calculate the minimum distance between each point cloud in the obstacle region and the point cloud in the obstacle region, and then find the maximum distance value among the minimum distances; at the same time, calculate the minimum distance between each point cloud in the obstacle region and find the maximum distance value, and finally take the maximum value of the two maximum distance values ​​to obtain the distance between the point clouds in the two obstacle regions. If the distance between two obstacle point clouds plus a safety distance threshold is less than the maximum distance between different coordinate axes of the bulk head point cloud, then the two obstacle point clouds are merged to construct a restricted area.

8. The AI-based automatic feed inlet alignment control system according to claim 7, characterized in that: The optimal trajectory planning module (330) is used to define the coordinates in the three-dimensional digital model, excluding the restricted area, as the search area, with probability... Directly sample the point cloud at the tank opening, using probability Randomly sample point clouds within the search area and generate multiple running trajectories from the bulk head point cloud to the can mouth point cloud within the search area; Calculate the distance between multiple trajectories, and select the trajectory with the smallest distance as the bulk head point cloud. Pointing clouds at the mouth of the jar The trajectory of the operation is recorded and converted into control commands for output.

Images