Intelligent material sorting method and system in cross-belt full-automatic component feeding process

Abstract

The application provides a material intelligent sorting method and system in a cross-belt full-automatic part feeding process, and relates to the technical field of material sorting. The method comprises the following steps: extracting a stacking level relationship and three-dimensional gap geometric information from a stacking package image of a cross-belt; obtaining a gas hole distribution of a micro gas hole array embedded in a stacked part separation area on the cross-belt; inputting a computational fluid dynamics proxy model based on the gas hole distribution, the stacking level relationship and the three-dimensional gap geometric information to perform online trajectory optimization on the jetting parameters of each gas hole unit in the micro gas hole array, and generating pneumatic separation parameters; and controlling the micro gas hole array to generate pulse air flow for separation when the stacked package reaches the stacked part separation area according to the pneumatic separation parameters. The application can solve the problems of large attitude disturbance and low efficiency in the stacked package separation process, realize real-time monitoring and intelligent optimization control in the separation process, and achieve the technical effects of minimizing the disturbance of the upper package, improving the separation success rate and the stability of the logistics system.

Description

Technical Field

[0001] This application relates to the field of material sorting technology, and in particular to a method and system for intelligent material sorting in a fully automated cross-belt feeding process. Background Technology

[0002] With the continuous improvement of the level of automation in logistics sorting, modern warehousing and distribution centers have put forward higher requirements for high-speed and high-precision stacked package separation capabilities.

[0003] Currently, existing automated sorting systems typically rely on mechanical push-pull devices, vibration mechanisms, or fixed airflow to separate stacked packages. These methods suffer from issues such as coarse movements, significant disturbance to the posture of upper-layer packages, and unstable separation success rates. Consequently, they lack precise identification and dynamic adaptation control mechanisms for the minute gaps between stacked packages, leading to upper-layer packages easily tilting, flipping, or experiencing secondary collisions during the separation process. Furthermore, separation time and energy consumption are difficult to optimize, thus failing to meet the accuracy and stability requirements of high-speed automated sorting.

[0004] In summary, existing technologies suffer from a lack of refined perception and dynamic separation control of the minute contact gaps and spatial hierarchy of stacked packages. This results in large attitude disturbances and low separation efficiency of upper-layer packages during the separation process, further affecting the conveying stability and overall reliability of the logistics automation system. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for intelligent material sorting in the fully automated cross-belt feeding process, in order to solve the technical problems in the prior art, which are caused by the lack of fine perception and dynamic separation control of the tiny contact gaps and spatial hierarchy of stacked packages, resulting in large posture disturbances and low separation efficiency of the upper layer packages during the separation process, which further affects the conveying stability of the logistics automation system and the overall reliability of the sorting operation.

[0006] In view of the above problems, this application provides a method and system for intelligent material sorting in the fully automated feeding process of cross belts.

[0007] In a first aspect, this application provides a method for intelligent material sorting during fully automated cross-belt feeding, implemented through an intelligent material sorting system during fully automated cross-belt feeding. The method includes: acquiring real-time detection images of the cross-belt via a vision module for target recognition; determining stacked package images and extracting the stacking hierarchy and three-dimensional gap geometry information at the stacking interface; acquiring the pore distribution of a micro-pore array embedded below the belt in the stacking separation zone on the cross-belt; inputting the pore distribution, the stacking hierarchy, and the three-dimensional gap geometry information into a pre-trained computational fluid dynamics surrogate model; optimizing the injection parameters of each pore unit in the micro-pore array online to minimize separation time and upper package posture disturbance, thereby generating pneumatic separation parameters; and controlling the micro-pore array to generate pulsed airflow for separation when the stacked packages reach the stacking separation zone, based on the pneumatic separation parameters.

[0008] Preferably, the intelligent material sorting method during the fully automated cross-belt feeding process further includes: segmenting the real-time detection image to obtain the independent mask area of ​​each package; performing intersection-union analysis on the independent mask areas to identify stacked packages; performing occlusion relationship analysis on the stacked packages to output a directed acyclic occlusion map between packages, extracting the stacking order of upper and lower packages to generate the stacking hierarchy relationship; extracting the contact area as the stacking interface based on the directed acyclic occlusion map, performing three-dimensional reconstruction on the stacking interface, and extracting the height curve, width change, and relative surface normal vector of the upper and lower packages to constitute the three-dimensional gap geometry information.

[0009] Preferably, the intelligent material sorting method in the fully automated cross-belt feeding process further includes: the stacked packages are packages with overlapping masks or edge spacing less than a preset threshold.

[0010] Preferably, the intelligent material sorting method in the fully automated cross-belt feeding process further includes: extracting the edge contour features and regional semantic features of each package mask within the stacked packages, and constructing a feature splicing vector for package pairs; classifying occlusion relationships based on the feature splicing vectors, constructing an initial directed graph with each package as a node and the occlusion direction as a directed edge, and obtaining the directed acyclic occlusion graph, wherein the directed acyclic occlusion graph represents the complete stacking hierarchy relationship between packages.

[0011] Preferably, the intelligent material sorting method in the fully automated cross-belt feeding process further includes: encoding the position coordinates and injection direction adjustment range of each pore unit in the pore distribution into a spatial matrix; converting the stacking hierarchy into relative pose constraints of package pairs; parsing the three-dimensional gap geometry information into a gap width sequence and a normal vector field; inputting the spatial matrix, the relative pose constraints, and the gap width sequence and normal vector into the computational fluid dynamics surrogate model, setting the objective function as a weighted sum of separation time and upper package angular velocity, iteratively optimizing the injection parameters of each pore unit, and outputting the pneumatic separation parameters.

[0012] Preferably, the intelligent material sorting method in the fully automated cross-belt feeding process further includes: the computational fluid dynamics proxy model includes a message passing layer and a decoding layer; wherein, the message passing layer is constructed based on a graph convolutional network, using the initial solution of the injection parameters, relative pose constraints, and gap width sequence and normal vector as input nodes, and performs message passing and feature aggregation on the airflow coupling relationship between adjacent air holes through a multi-layer graph convolutional network; the decoding layer outputs the aerodynamic distribution field and separation torque prediction value of the wrapping surface based on the fused node features.

[0013] Preferably, the intelligent material sorting method in the fully automated cross-belt feeding process further includes: retrieving the historical vent jetting parameter matrix with the highest stacking similarity from the separation log database based on the relative pose constraints and the gap width sequence and normal vector as the initial solution for the jetting parameters; constructing an online optimizer based on the objective function; the computational fluid dynamics proxy model receives the initial solution for the jetting parameters, the relative pose constraints, and the gap width sequence and normal vector as input node features, inputs them into the computational fluid dynamics proxy model, iteratively aggregates the airflow coupling between adjacent vent units through a message passing layer, and outputs the aerodynamic distribution field and separation torque prediction value of the wrapping surface at the decoding layer; the aerodynamic distribution field and separation torque prediction value are interacted with the online optimizer, and the online optimizer updates the candidate solutions for the jetting parameters based on the backpropagation of the predicted gradient until the convergence condition is met, and outputs the aerodynamic separation parameters.

[0014] Preferably, the intelligent material sorting method during the fully automatic cross-belt feeding process further includes: switching the air vent unit to negative pressure adsorption mode within a preset time after the pulse air jet is completed, and restoring each air vent unit to standby zero flow state after a preset duration; the preset time is specified by the computational fluid dynamics proxy model based on the predicted time of the package separation instant.

[0015] Preferably, the intelligent material sorting method in the fully automated cross-belt feeding process further includes: performing secondary detection on the separated packages through the vision module to determine whether the stacking is completely decoupled; quantifying the deviation between the actual decoupling result and the predicted value of the computational fluid dynamics proxy model, using the deviation data as incremental training samples, and performing online gradient updates on the computational fluid dynamics proxy model.

[0016] Secondly, this application also provides an intelligent material sorting system for fully automated cross-belt feeding, used to execute the intelligent material sorting method for fully automated cross-belt feeding as described in the first aspect, comprising: an information determination component, used to acquire real-time detection images of the cross-belt through a vision module for target recognition, determine stacked package images, and extract the stacking hierarchy relationship between packages and the three-dimensional gap geometry information at the stacking interface; an air hole acquisition component, used to acquire the air hole distribution of a micro-air hole array embedded under the belt in the stacking separation area on the cross-belt; a parameter generation component, used to input a pre-trained computational fluid dynamics surrogate model based on the air hole distribution, the stacking hierarchy relationship, and the three-dimensional gap geometry information, and to perform online trajectory optimization of the injection parameters of each air hole unit in the micro-air hole array with the goal of minimizing separation time and upper package posture disturbance, thereby generating aerodynamic separation parameters; and an airflow separation component, used to control the micro-air hole array to generate pulsed airflow for separation when the stacked packages reach the stacking separation area according to the aerodynamic separation parameters.

[0017] The technical solution provided in this application has at least the following technical effects or advantages: by achieving the technical goal of real-time monitoring and intelligent optimization control of the stacked package separation process, it achieves the technical effects of minimizing the posture disturbance of the upper package during the separation process, improving the separation success rate and operation efficiency, while ensuring the stability of the logistics conveying system and the reliability of the sorting operation.

[0018] The above description is merely an overview of the technical solution of this application. To enable a clearer understanding of the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the intelligent material sorting method used in the fully automated cross-belt feeding process of this application.

[0021] Figure 2 This is a schematic diagram of the intelligent material sorting system during the fully automated cross-belt feeding process of this application.

[0022] Explanation of reference numerals in the attached diagram: Information determination component 1, pore acquisition component 2, parameter generation component 3, airflow separation component 4. Detailed Implementation

[0023] This application provides a method and system for intelligent material sorting during fully automated cross-belt feeding, solving the technical problem in existing technologies where the lack of refined perception and dynamic separation control of the minute contact gaps and spatial hierarchy of stacked packages leads to large posture disturbances and low separation efficiency of upper-layer packages during separation, further affecting the conveying stability and overall reliability of the logistics automation system. The application achieves the technical goal of real-time monitoring and intelligent optimization control of the stacked package separation process, minimizing posture disturbances of upper-layer packages, improving separation success rate and operational efficiency, while ensuring the stability of the logistics conveying system and the reliability of the sorting operation.

[0024] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0025] Example 1, please refer to the appendix. Figure 1 This application provides a method for intelligent material sorting during fully automated cross-belt feeding, applicable to an intelligent material sorting system during fully automated cross-belt feeding, specifically including the following steps: The system uses a vision module to acquire real-time detection images of the cross straps for target recognition, determines the images of stacked packages, and extracts the stacking hierarchy relationship between packages and the three-dimensional gap geometry information at the stacking interface.

[0026] Furthermore, this application also includes: performing instance segmentation on the real-time detection image to obtain the independent mask region of each package; performing intersection-union analysis on the independent mask regions to identify stacked packages; performing occlusion relationship analysis on the stacked packages to output a directed acyclic occlusion map between packages, extracting the stacking order of upper and lower packages to generate the stacking hierarchy relationship; extracting the contact area as the stacking interface based on the directed acyclic occlusion map, performing three-dimensional reconstruction on the stacking interface, and extracting the height curve, width change, and relative surface normal vector of the upper and lower packages to constitute the three-dimensional gap geometry information.

[0027] Furthermore, this application also includes: the stacked packages are packages with overlapping masks or edge spacing less than a preset threshold.

[0028] Furthermore, this application also includes: extracting the edge contour features and regional semantic features of each package mask within the stacked packages, constructing a feature splicing vector for the package pairs; classifying occlusion relationships based on the feature splicing vectors, constructing an initial directed graph with each package as a node and the occlusion direction as a directed edge, and obtaining the directed acyclic occlusion graph, wherein the directed acyclic occlusion graph represents the complete stacking hierarchy relationship between packages.

[0029] Specifically, instance segmentation of real-time detection images to obtain independent mask regions for each package involves acquiring continuous image data during the cross-belt operation using a vision module, and then using a target instance segmentation network to divide multiple package targets in the image into pixel-level regions, assigning different target region numbers to different packages. Each independent mask region represents a set of binary regions that corresponds one-to-one with a single package target. The pixels in each independent mask region characterize the corresponding package's contour boundaries, spatial occupancy, and local morphological features in the image, thus providing basic target region data for subsequent spatial relationship analysis between packages.

[0030] Furthermore, cross-over-union (CUI) analysis is performed on independent masked regions to identify stacked packages. This involves quantifying the spatial overlap between multiple independent masked regions. By calculating the ratio of the intersection area to the union area of ​​different masked regions, it is determined whether there is spatial overlap or close overlap between packages. The CUI reflects the degree of overlap between different package targets in the image space. When the CUI is greater than a preset threshold or the edge distance is less than a preset interval, it is determined that the corresponding packages are stacked, thus forming a set of stacked packages.

[0031] Furthermore, extracting the edge contour features and regional semantic features of each package mask within the stacked packages to construct a feature concatenation vector for the package pairs involves, after completing package instance segmentation and obtaining independent mask regions, performing geometric morphological analysis on the edge pixels of each package mask to extract edge contour features, including contour length, curvature variation, corner distribution, and shape topology information. Simultaneously, semantic analysis is performed on the pixels within the mask regions to obtain regional texture, color distribution, pattern features, and local surface attributes. Subsequently, for package pairs that may have a stacking relationship, the edge contour features and regional semantic features of each package are concatenated according to predetermined rules to form a high-dimensional vector representing the comprehensive spatial and surface attributes of the package pairs, providing the input feature basis for determining occlusion relationships.

[0032] Furthermore, occlusion relationships are classified based on feature splicing vectors. An initial directed graph is constructed using each package as a node and the occlusion direction as directed edges, resulting in a directed acyclic occlusion (DAO) graph. The DAO graph represents the complete stacking hierarchy between packages. It involves using the constructed package feature splicing vectors as input to a pre-trained classification model to determine the front-to-back occlusion relationship between different package pairs, identifying the occlusion direction, and quantifying the occlusion intensity. Each package node is used as a node element in the graph, and the classified occlusion direction is used as the directed edge between nodes to form the initial directed graph. By removing possible cyclic edges and optimizing the node order, a DAO graph is generated. The DAO graph fully represents the stacking order and spatial hierarchy of stacked packages through the hierarchical structure of nodes and directed edges, providing structured data support for subsequent stacking interface extraction and separation path planning. Based on the occlusion direction, an association structure between nodes and directed edges is established. Nodes in the graph structure are used to represent different package targets, and directed edges are used to represent the occlusion direction and stacking order. By traversing the hierarchical relationship of directed edges, the stacking order between packages is obtained, forming stacking hierarchy data.

[0033] Furthermore, based on the directed acyclic occlusion (DAO) map, the contact area is extracted as the stacking interface. The stacking interface is then reconstructed in 3D, extracting the height curve, width variation, and relative surface normals of the upper and lower packages to form the 3D gap geometry. This involves locating the contact or proximity areas between packages based on the directly overlapping nodes in the DAO map, and defining these areas as the stacking interface. Subsequently, depth images, binocular vision data, or structured light point cloud data are used to restore the spatial surface of the stacking interface, establishing a 3D geometric model of the interface region. Specifically, a binocular stereo vision system combined with a semi-global block matching algorithm is used to obtain a dense disparity map of the stacking interface, which is then converted into a 3D point cloud. Poisson surface reconstruction is then performed on the 3D point cloud to extract a continuous surface model, and the height curve and relative surface normals are extracted along the contact area boundary. The height curve describes the spatial height variation trend of the interface gap along the conveying direction or lateral direction, the width variation represents the scale distribution of the interface opening area, and the relative surface normals describe the spatial orientation difference of the contact surfaces of the upper and lower packages. Together, these constitute the 3D gap geometry for airflow analysis.

[0034] Obtain the pore distribution of the micro-pore array embedded under the belt in the stacked separation area on the cross belt.

[0035] Specifically, obtaining the pore distribution of the micro-pore array embedded beneath the conveyor belt in the stacked package separation zone of the cross-belt conveyor system refers to locating the stacked package separation zone within the cross-belt conveyor system; this is the pre-defined conveyor belt segment where stacked packages are separated upon arrival. The micro-pore array refers to a collection of multiple small pore units embedded beneath the conveyor belt, each capable of independently controlling the airflow direction, flow rate, and pulse timing. The pore distribution represents the spatial arrangement of the micro-pore array, including the coordinate positions of the pores on the conveyor belt plane, the spacing between pores, and the relative positions of adjacent pores, supporting refined airflow control and separation strategy calculations. Obtaining the pore distribution provides fundamental spatial reference data for subsequent aerodynamic separation parameter optimization, jet trajectory planning, and upper package attitude disturbance control.

[0036] Based on the pore distribution, the stacking hierarchy, and the three-dimensional gap geometry information, a pre-trained computational fluid dynamics surrogate model is input. With the goal of minimizing separation time and upper layer encapsulation attitude disturbance, the injection parameters of each pore unit in the micro-pore array are optimized online to generate aerodynamic separation parameters.

[0037] Furthermore, this application also includes: encoding the position coordinates and injection direction adjustment range of each pore unit in the pore distribution into a spatial matrix; converting the stacking hierarchy relationship into relative pose constraints of the wrapping pairs; parsing the three-dimensional gap geometry information into a gap width sequence and a normal vector field; inputting the spatial matrix, the relative pose constraints, and the gap width sequence and normal vector into the computational fluid dynamics surrogate model, setting the objective function as a weighted sum of separation time and upper wrapping angular velocity, iteratively optimizing the injection parameters of each pore unit, and outputting the aerodynamic separation parameters.

[0038] Furthermore, this application also includes: the computational fluid dynamics proxy model includes a message passing layer and a decoding layer; wherein, the message passing layer is constructed based on a graph convolutional network, using the initial solution of the injection parameters, relative pose constraints, and gap width sequence and normal vector as input nodes, and performs message passing and feature aggregation on the airflow coupling relationship between adjacent vents through a multi-layer graph convolutional network; the decoding layer outputs the aerodynamic distribution field and separation torque prediction value of the wrapping surface based on the fused node features.

[0039] Furthermore, this application also includes: retrieving the historical vent jet parameter matrix with the highest stacking similarity from the separation log library based on the relative pose constraints and the gap width sequence and normal vector as the initial solution for the jet parameters; constructing an online optimizer based on the objective function; the computational fluid dynamics proxy model receives the initial solution for the jet parameters, the relative pose constraints, and the gap width sequence and normal vector as input node features, inputs them into the computational fluid dynamics proxy model, iteratively aggregates the airflow coupling between adjacent vent units through a message passing layer, and outputs the aerodynamic distribution field and separation torque prediction value of the wrapping surface at the decoding layer; interacting with the aerodynamic distribution field and separation torque prediction value with the online optimizer, the online optimizer updates the candidate solutions for the jet parameters based on the backpropagation of the predicted gradient until the convergence condition is met, and outputs the aerodynamic separation parameters.

[0040] Specifically, encoding the position coordinates and jet direction adjustment range of each pore unit in the pore distribution into a spatial matrix means that, based on the arrangement structure of the micro-pore array, the two-dimensional or three-dimensional coordinates of each pore unit on the conveyor belt plane and the adjustable jet direction range are digitally represented in matrix form. Each unit of the spatial matrix contains the position parameters of the corresponding pore, the upper and lower limits of the jet angle, and the adjustable degrees of freedom, thereby providing structured input data for subsequent airflow simulation and optimization calculations.

[0041] Furthermore, converting the stacking hierarchy into relative pose constraints of package pairs means defining the relative position and attitude constraints between adjacent stacked packages in space based on the stacking relationship formed in the directed acyclic occlusion graph. This includes the relative center position offset, rotation angle range, and relative orientation of the contact surfaces. The relative pose constraints of package pairs provide constraints for optimizing the vent jet parameters, ensuring that the attitude disturbance of the upper package is minimized during the separation process.

[0042] Furthermore, the three-dimensional gap geometry information is analyzed into a gap width sequence and a normal vector field. This means analyzing the three-dimensional geometric model reconstructed from the stacked interface, extracting the gap width sequence by segmenting the interface along the conveying direction or laterally, representing the size change of the interface opening between the upper and lower packages, and simultaneously calculating the spatial normal vector field of the contact surface of the upper and lower packages to characterize the interface orientation and local surface normal information, thereby providing refined geometric constraints for airflow action analysis and jet optimization.

[0043] Furthermore, the computational fluid dynamics surrogate model includes a message passing layer and a decoding layer, meaning that the surrogate model consists of a message passing layer for feature information exchange and aggregation, and a decoding layer for outputting aerodynamic prediction results.

[0044] Furthermore, the message passing layer is constructed based on a graph convolutional network. It uses the initial solution of injection parameters, relative pose constraints, and gap width sequence and normal vector as input nodes. Through a multi-layer graph convolutional network, it performs message passing and feature aggregation on the airflow coupling relationship between adjacent orifices. This means that each orifice unit is treated as a node in a graph, and the relative pose constraints of the wrapping pair, the gap width sequence of the stacking interface, and the surface normal vector are used as node feature inputs. Information propagation occurs between nodes through the hierarchical structure of the graph convolutional network. Each convolutional layer achieves message passing by weighted summation and nonlinear mapping of the features of neighboring nodes. This allows the airflow interaction relationship between adjacent orifice units to be captured and aggregated in the network, thereby obtaining the coupling feature representation of the entire separated region. The message passing layer contains three layers of graph convolutional networks, each with a node feature dimension of 256, using a linear rectified function as the activation function. The nodes in the graph convolutional network are micro-orifice units. The initial feature vector dimension of each node is 128, containing orifice position coordinates, injection direction angle, flow coefficient, etc., and the edge feature dimension is 64, containing the distance and relative angle between adjacent orifices. The specific expression for message passing is: H (l+1) =σ( ~- - H (l) W (l) ),in =A+I is the adjacency matrix with self-loops added. H is the corresponding degree matrix. (l) For the features of the l-th layer nodes, W (l)The weights are learnable, and σ is the activation function. The objective function is calculated as follows: the aerodynamic distribution field output by the computational fluid dynamics surrogate model is integrated to obtain the resultant force and resultant torque. Combining the physical properties of the package's mass and moment of inertia, the separation time and angular velocity of the upper package are calculated by integrating Newton's second law and the rigid body dynamics equations. The objective function is defined as Loss = α·T sep +β·ω max T sep For the separation time, ω max α represents the maximum angular velocity during the separation process, and β represents preset weighting coefficients. The online optimizer specifically employs the L-BFGS (quasi-Newton method) optimization algorithm, which performs backpropagation and parameter updates by calculating the Jacobian matrix of the objective function with respect to the injection parameters.

[0045] Furthermore, the decoding layer outputs the aerodynamic distribution field and predicted separation torque of the package surface based on the fused node features. This means that the node features aggregated by the message passing layer are used as input, and the high-dimensional features are converted into the magnitude and direction distribution of aerodynamic forces at various points on the package surface, as well as the predicted overall separation torque acting on the stacked packages, through a mapping function in the decoding layer. This provides quantifiable physical feedback information for optimizing the jet parameters of the vent array. The decoding layer is a three-layer multilayer perceptron with 256, 128, and 64 neurons in each layer, respectively. The last layer outputs the aerodynamic distribution field of discrete grid points on the package surface, with a dimension of N×3, where N is the number of grid points and 3 corresponds to the three-dimensional force components, as well as the predicted overall separation torque, also with a dimension of 3. In addition, the decoding layer introduces physical prior constraints based on computational fluid dynamics during training. Specifically, the residual terms of the simplified incompressible fluid Navier-Stokes equations are used as a penalty term in the loss function to ensure that the predicted aerodynamic distribution field satisfies the laws of mass and momentum conservation. For example, treating the pulsed airflow generated by the micro-pore array as a porous medium jet, and introducing the simplified Navier-Stokes equations for incompressible fluids as physical constraints, the continuity equation is: The momentum conservation equation is ρ( Where v is the airflow velocity field, p is the air pressure field, ρ is the air density, μ is the aerodynamic viscosity, and f is the jet source term. Next, a physical constraint residual module is constructed after the decoding layer. Automatic differentiation techniques are used to calculate the partial derivatives of the implicit velocity and pressure fields corresponding to the aerodynamic distribution field output by the decoding layer with respect to the Navier-Stokes equations for incompressible fluids, thus obtaining the mass-conserving residual. Momentum conservation residual R d =∣ρ( Finally, fluid dynamics and physical prior constraints are added to the loss function, and the total loss function is defined as L. total =L data +λ1R m +λ2R d Ldata To predict the mean square error between the aerodynamic forces and the simulated true values, λ1 and λ2 are the weighting coefficients for the physical constraint penalty. This is achieved by minimizing L during the optimization process. total The aerodynamic distribution field output by the forced model satisfies the laws of mass and momentum conservation in fluid mechanics, while also incorporating the empirical formula for jet attenuation distance v(r)∝ The feature weights of the orifice injection direction are initialized.

[0046] Based on relative pose constraints and gap width sequences and normal vectors, the algorithm retrieves the historical orifice ejection parameter matrix with the highest stack similarity from the separation log database as the initial solution for ejection parameters. This involves using orifice ejection parameter data recorded during historical separation operations, and based on the relative pose constraints of the current stack, as well as the gap width sequence and normal vector information of the stack interface, a similarity measurement algorithm is used to retrieve the most matching historical parameter matrix from the log database. The retrieval result is then used as the initial input solution for orifice ejection optimization, accelerating the online optimization process and improving separation performance. Specifically, the current relative pose constraints, gap width sequences, and normal vector fields are flattened into one-dimensional feature vectors. A weighted fusion measurement method using cosine similarity and Euclidean distance is employed to calculate the comprehensive similarity between the current stack features and the features of each historical record in the log database. The orifice ejection parameter matrix corresponding to the historical record with the highest comprehensive similarity is selected as the initial solution.

[0047] Furthermore, constructing an online optimizer based on an objective function refers to using a weighted sum of separation time and upper layer envelope angular velocity as an evaluation index according to a preset optimization objective function to construct an optimization algorithm module for dynamically adjusting the vent jet parameters. The online optimizer can calculate the objective function value of candidate solutions in real time during the iteration process and guide the update direction and step size of the jet parameters to achieve the optimization of the separation effect.

[0048] Furthermore, the computational fluid dynamics proxy model receives the initial solution of the injection parameters, relative pose constraints, and gap width sequence and normal vector as input node features. This input is then fed into the computational fluid dynamics proxy model. Through a message passing layer, the airflow coupling between adjacent vent units is iteratively aggregated. The decoding layer outputs the aerodynamic distribution field and predicted separation torque values ​​of the wrapping surface. This means that the initial solution of the injection parameters and stacking geometric constraint information are input as node features of a graph convolutional network into the message passing layer. Multi-layer graph convolution is used to aggregate the features of the airflow coupling relationship between vent units. Subsequently, the decoding layer maps the aggregated features to the local aerodynamic distribution of the upper wrapping surface and the predicted separation torque values ​​acting on the entire stacked wrapping, providing quantifiable separation performance feedback to the online optimizer.

[0049] Furthermore, the aerodynamic distribution field and predicted separation torque values ​​are interacted with the online optimizer. The online optimizer updates the candidate solutions for injection parameters based on the predicted gradient backpropagation until the convergence condition is met, and outputs the aerodynamic separation parameters. This means that the aerodynamic distribution field and predicted separation torque values ​​output by the surrogate model are used as feedback information for the online optimizer. The derivative of the objective function with respect to the injection parameters is calculated through the gradient backpropagation method, and the candidate solutions for injection parameters are updated according to the iteration rules. The iteration is repeated until the objective function converges, and finally, an aerodynamic separation parameter matrix that can be used to control the micro-orifice array to perform precise separation operations is generated.

[0050] Based on the pneumatic separation parameters, the micro-pore array is controlled to generate pulsed airflow for separation when the stacked package reaches the stack separation zone.

[0051] Furthermore, this application also includes: switching the pore unit to negative pressure adsorption mode within a preset time after the pulsed airflow jet is completed, and restoring each pore unit to standby zero flow state after a preset duration; the preset time is specified by the computational fluid dynamics proxy model based on the predicted time of the package separation instant.

[0052] Furthermore, this application also includes: performing secondary detection on the separated packages through the vision module to determine whether the stacking is completely decoupled; quantifying the deviation between the actual decoupling result and the predicted value of the computational fluid dynamics proxy model, using the deviation data as incremental training samples, and performing online gradient updates on the computational fluid dynamics proxy model.

[0053] Specifically, based on the pneumatic separation parameters, the micro-pore array is controlled to generate pulsed airflow for separation when the stacked packages reach the stack separation zone. This means using the injection parameter matrix of each pore unit obtained through online optimization, including injection time, flow rate, direction and duration, to trigger the micro-pore array through the control system. This generates a high-precision pulsed airflow the instant the package reaches the separation zone, thereby achieving physical separation between the upper and lower packages. This ensures that the upper package experiences minimal attitude disturbance and maintains stable transport during the separation process.

[0054] Furthermore, after the pulse airflow injection is completed, the air vent unit is switched to negative pressure adsorption mode within a preset time. After a preset duration, each air vent unit returns to the standby zero flow state. This means that after the pulse injection is completed, the air vent unit is switched from positive pressure injection state to negative pressure adsorption state by the control unit, so that the bottom surface of the package can be slightly adsorbed to prevent movement and deviation. At the same time, the negative pressure adsorption action is maintained for a preset duration. After the time is over, the air vent unit is restored to the zero flow standby state to reduce energy consumption and ensure that the air vent system under the conveyor belt is in a ready state.

[0055] Furthermore, the preset time is specified by the computational fluid dynamics surrogate model based on the predicted time of the instant of package separation. This means that the surrogate model is used to predict the time required for the instant of separation of stacked packages during the online optimization stage. Based on the prediction results, the start time and duration of negative pressure adsorption after pulse jetting are determined, so as to achieve precise timing matching between separation action and pore control, thereby improving separation reliability and operational stability.

[0056] The secondary inspection of the separated packages by the vision module to determine whether the stacking is completely decoupled refers to the use of a vision module installed on the conveying system to collect image data again after the stacked packages have undergone the separation operation of the micro air hole array. The position, posture and the state of the gap between each package are analyzed. Through instance segmentation, mask region recognition and spatial reconstruction methods, it is determined whether the upper and lower packages have achieved complete physical separation, thereby forming a real-time evaluation result of the separation quality.

[0057] Furthermore, the deviation between the actual decoupling results and the predicted values ​​of the computational fluid dynamics surrogate model is quantified. The deviation data is used as incremental training samples to perform online gradient updates on the computational fluid dynamics surrogate model. This involves comparing the package separation state obtained from the secondary detection with the predicted values ​​of the aerodynamic distribution field and separation torque output by the surrogate model during the jet optimization stage to obtain prediction deviation data, including local separation errors and overall torque differences. The deviation data is then used as incremental samples to input into the surrogate model, and the model parameters are adjusted online through gradient descent or backpropagation methods. This enables the model to adaptively optimize during operation, improving the prediction accuracy and response capability of airflow separation under complex stacking conditions. Specifically, the actual separation state vector is calculated, including the actual separation time, the final pose offset of the upper layer, and the mean square error between the angular velocity and the model's predicted state vector, which is used as the deviation value. During online gradient updates, the loss function uses the aforementioned mean square error plus the physical constraint residual term of fluid dynamics. To ensure the stability of the model's online updates, an empirical replay mechanism is adopted, storing incremental training samples in a fixed-capacity buffer pool. During each update, a batch of historical samples is randomly sampled from the buffer pool and mixed with the current samples for mini-batch gradient descent training. A learning rate decay strategy is set to prevent catastrophic forgetting or oscillation of model parameters.

[0058] In summary, the intelligent material sorting method for the fully automated cross-belt feeding process provided in this application has the following technical effects: by achieving the technical goal of real-time monitoring and intelligent optimization control of the stacked package separation process, it achieves the technical effects of minimizing the posture disturbance of the upper package during the separation process, improving the separation success rate and operation efficiency, while ensuring the stability of the logistics conveying system and the reliability of the sorting operation.

[0059] Example 2: Based on the same inventive concept as the intelligent material sorting method in the fully automated cross-belt feeding process described in the foregoing examples, this application also provides an intelligent material sorting system in the fully automated cross-belt feeding process. Please refer to the appendix. Figure 2 The system includes: an information determination component 1, used to acquire real-time detection images of the cross belts through a vision module for target recognition, determine stacked package images, and extract the stacking hierarchy relationship between packages and the three-dimensional gap geometry information at the stacking interface; an air hole acquisition component 2, used to acquire the air hole distribution of the micro air hole array embedded under the belt in the stacking separation area on the cross belt; a parameter generation component 3, used to input a pre-trained computational fluid dynamics surrogate model based on the air hole distribution, the stacking hierarchy relationship, and the three-dimensional gap geometry information, and to perform online trajectory optimization of the injection parameters of each air hole unit in the micro air hole array with the goal of minimizing separation time and upper package posture disturbance, thereby generating aerodynamic separation parameters; and an airflow separation component 4, used to control the micro air hole array to generate pulsed airflow for separation when the stacked packages reach the stacking separation area according to the aerodynamic separation parameters.

[0060] Furthermore, the intelligent material sorting system during the fully automated cross-belt feeding process is also used for: segmenting real-time detection images to obtain independent mask areas for each package; performing intersection-union analysis on the independent mask areas to identify stacked packages; performing occlusion relationship analysis on the stacked packages to output a directed acyclic occlusion map between packages, extracting the stacking order of upper and lower packages to generate the stacking hierarchy relationship; extracting the contact area as the stacking interface based on the directed acyclic occlusion map, performing three-dimensional reconstruction on the stacking interface, and extracting the height curve, width variation, and relative surface normal vectors of the upper and lower packages to constitute the three-dimensional gap geometry information.

[0061] Furthermore, the intelligent material sorting system during the fully automated cross-belt feeding process is also used for: stacked packages that have overlapping masks or edge spacing less than a preset threshold.

[0062] Furthermore, the intelligent material sorting system in the fully automated cross-belt feeding process is also used to: extract the edge contour features and regional semantic features of each package mask within the stacked packages, and construct feature splicing vectors for package pairs; classify occlusion relationships based on the feature splicing vectors, construct an initial directed graph with each package as a node and the occlusion direction as a directed edge, and obtain the directed acyclic occlusion graph, which represents the complete stacking hierarchy relationship between packages.

[0063] Furthermore, the intelligent material sorting system in the fully automated cross-belt feeding process is also used to: encode the position coordinates and injection direction adjustment range of each pore unit in the pore distribution into a spatial matrix; convert the stacking hierarchy into relative pose constraints of the package pairs; parse the three-dimensional gap geometry information into a gap width sequence and a normal vector field; input the spatial matrix, the relative pose constraints, and the gap width sequence and normal vector into the computational fluid dynamics surrogate model, set the objective function as a weighted sum of separation time and upper package angular velocity, iteratively optimize the injection parameters of each pore unit, and output the pneumatic separation parameters.

[0064] Furthermore, the intelligent material sorting system in the fully automated cross-belt feeding process is also used for: the computational fluid dynamics proxy model includes a message passing layer and a decoding layer; wherein, the message passing layer is constructed based on a graph convolutional network, using the initial solution of the injection parameters, relative pose constraints, and gap width sequence and normal vector as input nodes, and performs message passing and feature aggregation on the airflow coupling relationship between adjacent air holes through a multi-layer graph convolutional network; the decoding layer outputs the aerodynamic distribution field and separation torque prediction value of the wrapping surface based on the fused node features.

[0065] Furthermore, the intelligent material sorting system during the fully automated cross-belt feeding process is also used for: retrieving the historical vent jetting parameter matrix with the highest stacking similarity from the separation log database based on the relative pose constraints and the gap width sequence and normal vector as the initial solution for the jetting parameters; constructing an online optimizer based on the objective function; the computational fluid dynamics proxy model receives the initial solution for the jetting parameters, the relative pose constraints, and the gap width sequence and normal vector as input node features, inputs them into the computational fluid dynamics proxy model, iteratively aggregates the airflow coupling between adjacent vent units through the message passing layer, and outputs the aerodynamic distribution field and separation torque prediction value of the wrapping surface at the decoding layer; interacting with the aerodynamic distribution field and separation torque prediction value with the online optimizer, the online optimizer updates the candidate solutions for the jetting parameters based on the backpropagation of the predicted gradient until the convergence condition is met, and outputs the aerodynamic separation parameters.

[0066] Furthermore, the intelligent material sorting system during the fully automatic cross-belt feeding process is also used to: switch the air vent unit to negative pressure adsorption mode within a preset time after the pulse air jet is completed, and restore each air vent unit to standby zero flow state after a preset duration; the preset time is specified by the computational fluid dynamics proxy model based on the predicted time of the package separation instant.

[0067] Furthermore, the intelligent material sorting system in the fully automated cross-belt feeding process is also used to: perform secondary detection on the separated packages through the vision module to determine whether the stacking is completely decoupled; quantify the deviation between the actual decoupling result and the predicted value of the computational fluid dynamics proxy model, and use the deviation data as incremental training samples to perform online gradient updates on the computational fluid dynamics proxy model.

[0068] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The intelligent material sorting method and specific examples in the fully automatic cross-belt feeding process of the aforementioned embodiment 1 are also applicable to the intelligent material sorting system in the fully automatic cross-belt feeding process of this embodiment. Through the foregoing detailed description of the intelligent material sorting method in the fully automatic cross-belt feeding process, those skilled in the art can clearly understand the intelligent material sorting system in the fully automatic cross-belt feeding process of this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0069] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0070] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for intelligent material sorting during fully automated cross-belt feeding, characterized in that: include: The visual module acquires real-time detection images of the cross straps for target recognition, determines the stacked package images, and extracts the stacking hierarchy relationship between packages and the three-dimensional gap geometry information at the stacking interface. Obtain the pore distribution of the micro-pore array embedded under the belt in the stacked separation area on the cross belt; Based on the pore distribution, the stacking hierarchy, and the three-dimensional gap geometry information, a pre-trained computational fluid dynamics proxy model is input. With the goal of minimizing separation time and upper layer wrapping attitude disturbance, the injection parameters of each pore unit in the micro pore array are optimized online to generate aerodynamic separation parameters. Based on the pneumatic separation parameters, the micro-pore array is controlled to generate pulsed airflow for separation when the stacked package reaches the stack separation zone.

2. The intelligent material sorting method in the fully automated cross-belt feeding process as described in claim 1, characterized in that, Real-time detection images of the cross straps are acquired through a vision module for target recognition, determining the stacked package images and extracting the stacking hierarchy and 3D gap geometry information at the stacking interface, including: Instance segmentation is performed on the real-time detection image to obtain the independent mask region of each package; Perform cross-intersection over union (CUI) analysis on the independent mask regions to identify stacked packages; The stacked packages are analyzed for occlusion relationships, and a directed acyclic occlusion graph between the packages is output. The stacking order of the upper and lower layers of packages is extracted to generate the stacking hierarchy relationship. Based on the directed acyclic occlusion map, the contact area is extracted as the stacking interface. The stacking interface is then reconstructed in three dimensions. The height curve, width variation, and relative surface normal vectors of the upper and lower wrapping surfaces of the gap are extracted to form the three-dimensional gap geometry information.

3. The intelligent material sorting method in the fully automated cross-belt feeding process as described in claim 2, characterized in that, The stacked packages are those with overlapping masks or edge spacing less than a preset threshold.

4. The intelligent material sorting method in the fully automated cross-belt feeding process as described in claim 2, characterized in that, The stacked packages are subjected to occlusion relationship analysis, and a directed acyclic occlusion graph between the packages is output, including: Extract the edge contour features and region semantic features of each package mask within the stacked packages, and construct the feature concatenation vector of the package pairs; Based on the feature splicing vector, occlusion relationship classification is performed. An initial directed graph is constructed with each package as a node and the occlusion direction as a directed edge to obtain the directed acyclic occlusion graph. The directed acyclic occlusion graph represents the complete overlapping hierarchy relationship between packages.

5. The intelligent material sorting method in the fully automated cross-belt feeding process as described in claim 1, characterized in that, Based on the pore distribution, the stacking hierarchy, and the three-dimensional gap geometry information, a pre-trained computational fluid dynamics surrogate model is input. With the goal of minimizing separation time and upper-layer encapsulation attitude perturbation, the injection parameters of each pore unit in the micro-pore array are optimized online to generate aerodynamic separation parameters, including: The position coordinates and injection direction adjustment range of each pore unit in the pore distribution are encoded into a spatial matrix; The stacking hierarchy is converted into relative pose constraints for wrapping pairs; The three-dimensional gap geometry information is analyzed into a gap width sequence and a normal vector field; The spatial matrix, the relative pose constraint, the gap width sequence, and the normal vector are input into the computational fluid dynamics surrogate model. The objective function is set as the weighted sum of the separation time and the upper wrapping angular velocity. The injection parameters of each air hole unit are iteratively optimized, and the aerodynamic separation parameters are output.

6. The intelligent material sorting method in the fully automated cross-belt feeding process as described in claim 5, characterized in that, The computational fluid dynamics proxy model includes a message passing layer and a decoding layer; The message passing layer is constructed based on a graph convolutional network. It takes the initial solution of the injection parameters, the relative pose constraints, the gap width sequence and the normal vector as input nodes. It performs message passing and feature aggregation on the airflow coupling relationship between adjacent air holes through a multi-layer graph convolutional network. The decoding layer outputs the aerodynamic distribution field and separation torque prediction values ​​of the wrapped surface based on the fused node features.

7. The intelligent material sorting method in the fully automated cross-belt feeding process as described in claim 6, characterized in that, The spatial matrix, the relative pose constraints, and the gap width sequence and normal vector are input into the computational fluid dynamics surrogate model. The objective function is set as a weighted sum of separation time and upper wrapping angular velocity. The injection parameters of each vent unit are iteratively optimized, and the aerodynamic separation parameters are output, including: Based on the relative pose constraint and the gap width sequence and normal vector, the historical pore injection parameter matrix with the highest stack similarity is retrieved from the separate log library as the initial solution of the injection parameters; Construct an online optimizer based on the objective function; The computational fluid dynamics proxy model receives the initial solution of the injection parameters, the relative pose constraints, and the gap width sequence and normal vector as input node features, inputs them into the computational fluid dynamics proxy model, iteratively aggregates the airflow coupling between adjacent vent units through the message passing layer, and outputs the aerodynamic distribution field and separation torque prediction value of the wrapping surface at the decoding layer. The aerodynamic distribution field and the predicted separation torque are interacted with an online optimizer. The online optimizer updates the candidate solutions of the injection parameters based on the backpropagation of the predicted gradient until the convergence condition is met, and then outputs the aerodynamic separation parameters.

8. The intelligent material sorting method in the fully automated cross-belt feeding process as described in claim 1, characterized in that, After controlling the micro-pore array to generate pulsed airflow for separation when the stacked package reaches the stack separation zone, the method further includes: Within a preset time after the pulse airflow jet is completed, the pore unit is switched to negative pressure adsorption mode, and after a preset duration, each pore unit returns to standby zero flow state. The preset time is specified by the computational fluid dynamics proxy model based on the predicted time of the instant the package separates.

9. The intelligent material sorting method in the fully automated cross-belt feeding process as described in claim 1, characterized in that, After controlling the micro-pore array to generate pulsed airflow for separation when the stacked package reaches the stack separation zone, the method further includes: The vision module performs secondary detection on the separated packages to determine whether the stacking is completely decoupled. The deviation between the actual decoupling results and the predicted values ​​of the computational fluid dynamics surrogate model is quantified, and the deviation data is used as incremental training samples to perform online gradient updates on the computational fluid dynamics surrogate model.

10. A fully automated material sorting system for cross-belt feeding, characterized in that: The steps for implementing the intelligent material sorting method in the fully automated cross-belt feeding process according to any one of claims 1 to 9 include: The information determination component is used to acquire real-time detection images of cross straps through the vision module for target recognition, determine stacked package images, and extract the stacking hierarchy relationship between packages and the three-dimensional gap geometry information at the stacking interface. A pore acquisition component is used to acquire the pore distribution of a micro-pore array embedded under the belt in the stack separation area on the cross belt; The parameter generation component is used to input a pre-trained computational fluid dynamics proxy model based on the pore distribution, the stacking hierarchy, and the three-dimensional gap geometry information, and to perform online trajectory optimization of the injection parameters of each pore unit in the micro-pore array with the goal of minimizing separation time and upper layer wrapping attitude disturbance, thereby generating aerodynamic separation parameters. An airflow separation component is used to control the micro-pore array to generate pulsed airflow for separation when the stacked package reaches the stack separation zone, based on the pneumatic separation parameters.

Images