Battery pack special-shaped part accurate positioning method and system fusing disassembly topological relation

Abstract

The invention discloses a battery pack special-shaped part accurate positioning method and system fusing a disassembly topological relation, and belongs to the technical field of power battery disassembly positioning. The method comprises the following steps: acquiring a global point cloud of a battery pack by using a global camera, and registering with a preset CAD model to obtain a transformation matrix; a graph neural network assembly relation scene graph is constructed based on the matrix and a CAD model analysis result, and semantic association recognition of the parts is achieved; performing difference detection by comparing the registered CAD model with the global point cloud, and marking the state of the part; and planning an optimal recognition view angle of the local camera according to a sheltered or uncertain state, controlling the local camera to obtain a local point cloud and performing fine positioning, and updating pose information in the scene graph. The method is mainly used for accurate positioning of special-shaped parts in the power battery disassembling process, data support is provided for robot disassembling operation, the positioning robustness under the severe working condition is improved, and meanwhile the positioning problem caused by shielding is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of power battery disassembly and positioning technology, specifically to the field of precise positioning technology for irregularly shaped battery pack components guided by machine vision, graph neural networks and CAD models. Background Technology

[0002] In the disassembly of power batteries, the precise positioning of irregularly shaped components in the battery pack is crucial to ensuring accurate and efficient disassembly operations. Currently, most existing technologies employ a bottom-up, segmentation-then-registration approach, which involves extracting the 3D point cloud of the component and aligning it with the model's point cloud to estimate the component's pose. However, in real-world disassembly scenarios, this method encounters numerous challenges.

[0003] Power batteries contain numerous highly reflective metal components. At certain angles, light reflection can blur the details of these components, making them difficult to distinguish. Meanwhile, retired battery packs, often stored outdoors, are prone to dirt, corrosion, or damage during transport, all of which complicate the positioning of irregularly shaped parts.

[0004] The internal structure of the battery pack is very complex, and components such as flexible wiring harnesses and cooling pipes can easily obstruct key components to be disassembled. The global camera is too far away to capture fine details; while in passive recognition mode, the local camera may not be able to avoid obstructions due to the angle of view. If a segmented shooting method is used, it will increase the time required for the task.

[0005] Furthermore, most existing methods rely on spatial proximity to combine parts into assemblies, lacking a deep understanding of the assembly's function and structure. When the target instance has complex boundaries, the parts are too close together, or the environment is cluttered, it becomes difficult to distinguish between parts with different functions but similar appearances, thus affecting the accuracy of positioning.

[0006] In summary, these issues make it difficult for existing technologies to achieve stable and high-precision positioning of irregularly shaped battery pack components in real-world disassembly scenarios, thus failing to meet the needs of actual disassembly operations. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for precise positioning of irregularly shaped battery pack components by integrating disassembly topology relationships, in order to solve the problems of insufficient robustness of existing technologies under harsh battery pack disassembly conditions and the inability of passive vision to cope with occlusion leading to positioning failure.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] In a first aspect, the present invention provides a method for precise positioning of irregularly shaped components in a battery pack by integrating disassembly and topological relationships, comprising the following steps:

[0010] The global point cloud data of the battery pack to be disassembled is acquired using a global camera, and the global point cloud data is registered with a preset battery pack CAD model to obtain a transformation matrix.

[0011] Based on the transformation matrix and the CAD model analysis results, a graph neural network assembly relationship scene graph is constructed to obtain a scene graph containing semantic relationships of parts;

[0012] Based on the scene diagram and the registered CAD model, difference detection is performed with the global point cloud data to obtain the component status marking results;

[0013] Using the occlusion or uncertainty states in the state labeling results, the optimal recognition angle of the local camera is planned;

[0014] The local camera is controlled to move to the optimal recognition viewpoint to acquire local point cloud data. The local point cloud data is used for fine localization, the pose information of the parts in the scene map is updated, and the six-dimensional pose data of all target parts is output.

[0015] In one possible implementation, during registration, feature descriptors are extracted using the global point cloud and the point cloud sampled from the CAD model, and coarse registration is performed using a random sampling consistency algorithm to obtain an initial transformation matrix.

[0016] Based on the initial transformation matrix, the iterative nearest point algorithm is used for fine registration to obtain the final transformation matrix;

[0017] The global camera and robot base coordinate system are globally calibrated to obtain the first hand-eye calibration matrix;

[0018] The local camera and the robot end flange coordinate system are calibrated using hand-eye calibration to obtain a second hand-eye calibration matrix.

[0019] By using the robot's forward kinematics and the first and second hand-eye calibration matrices, the global point cloud data and local point cloud data are unified into the robot's base coordinate system.

[0020] In one possible implementation, when constructing the scene diagram, the assembly file of the CAD model is parsed, each independent component is defined as a node, and the physical contact and assembly constraints between components are defined as edges.

[0021] Based on the nodes and edges, a graph neural network model is constructed, and the relationships between components are learned and semantic functions are inferred through the graph neural network model;

[0022] A query is initiated to the graph neural network model to obtain a detachable component subgraph containing the target component and its associated parts. The detachable component subgraph is formed based on the structural and functional associations of the components.

[0023] In one possible implementation, during difference detection, the registered CAD model is compared with the global point cloud data in three dimensions to identify regions that exist in the CAD model but are missing in the global point cloud data, and abnormal regions that do not exist in the CAD model but exist in the global point cloud data.

[0024] Based on the identification results of the missing and abnormal regions, the status of the corresponding components in the scene diagram is marked as verified, uncertain occlusion, or abnormally damaged, respectively, thus completing the status update of the scene diagram.

[0025] In one possible implementation, when planning the optimal recognition viewpoint, a virtual hemisphere is generated around the expected location of the component whose state is marked as uncertain due to occlusion, and multiple candidate viewpoints are obtained by sampling through the virtual hemisphere;

[0026] By combining the occlusion information in the global point cloud data and the robot motion constraints, the spatial occlusion rate of each candidate viewpoint is calculated;

[0027] Based on the spatial occlusion rate, the candidate viewpoint with the lowest occlusion rate and the highest information gain is selected as the optimal recognition viewpoint.

[0028] In one possible implementation, during fine localization, an instance segmentation network is used to extract the point cloud of the target component from the local point cloud data; the point cloud of the target component is processed using an iterative nearest point algorithm or a six-dimensional pose estimator to obtain the true pose of the target component.

[0029] Update the actual pose to the scene graph and query the state of the detachable component subgraph again;

[0030] If there are still components with uncertain occlusion states, repeat the steps of optimal identification view planning and fine positioning until all components in the detachable component sub-graph are marked as verified.

[0031] In one possible implementation, when extracting feature descriptors, a point cloud feature extraction algorithm is used to calculate the features of the global point cloud and the sampled point cloud of the CAD model;

[0032] During global calibration, the relative positional relationship between the global camera coordinate system and the robot base coordinate system is determined by the global calibration algorithm, and the first hand-eye calibration matrix is ​​obtained.

[0033] During hand-eye calibration, the relative positional relationship between the local camera coordinate system and the robot end flange coordinate system is determined by the hand-eye calibration algorithm, thus obtaining the second hand-eye calibration matrix.

[0034] In one possible implementation, when parsing the assembly file of the CAD model, the geometric information of the parts and the assembly constraint relationships are extracted.

[0035] When constructing a graph neural network model, the model learns semantic relationships such as fixation and connection between components to achieve automatic prediction of disassembly interference and component function.

[0036] When initiating a graph query, based on the node information of the target component, all components that are associated with it by edges are retrieved.

[0037] In one possible implementation, when generating the virtual hemisphere, the expected position of the component in the uncertain occlusion state is used as the center of the sphere, and a reasonable radius is set to construct the virtual hemisphere to ensure that the candidate viewpoint covers the unobstructed observation range.

[0038] When calculating the spatial occlusion rate, the three-dimensional coordinates of the occluders in the global point cloud and the line-of-sight vectors of the candidate viewpoints are combined to determine whether the line of sight is blocked, and the degree of occlusion is quantified to obtain the spatial occlusion rate.

[0039] When selecting the best recognition perspective, it is necessary to simultaneously verify whether the candidate viewpoint is within the effective working distance of the local camera and whether the robot can reach it, to ensure the feasibility of the perspective.

[0040] In one possible implementation, when using an instance segmentation network, a lightweight segmentation algorithm is selected to process the local point cloud data and accurately extract the point cloud region of the target component.

[0041] When updating the scene diagram, the pose information and status markers of the components are modified simultaneously to ensure that the scene diagram is consistent with the actual situation;

[0042] When outputting six-dimensional pose data, the position and orientation information of all verified target components and their associated fasteners are integrated to generate a structured pose list.

[0043] Secondly, the present invention provides a precise positioning system for irregularly shaped battery pack components based on integrated disassembly and topology relationships, comprising: a global eye-out 3D camera, a local eye-hand 3D camera, a robot system, a central processing unit, a CAD model storage module, and a graph neural network model module. The global eye-out 3D camera is fixed on the gantry of the robot workstation and is signal-connected to the central processing unit. The local eye-hand 3D camera is mounted on the end effector flange of the robot system and is signal-connected to the central processing unit. The robot system is signal-connected to the central processing unit. The CAD model storage module is signal-connected to the central processing unit. The graph neural network model module is signal-connected to the central processing unit. The central processing unit is used to execute the precise positioning method for irregularly shaped battery pack components based on integrated disassembly and topology relationships.

[0044] Compared with existing technologies, the advantages of this invention are as follows: Compared with existing positioning methods that rely on bottom-up segmentation and registration, this invention significantly improves the practicality and reliability of positioning irregularly shaped parts in battery packs through a combination of CAD model guidance and graph neural network inference. The top-down verification strategy guided by the CAD model eliminates the need to search for parts in a cluttered global image; instead, it verifies the existence of parts at expected locations based on the registered model. This greatly reduces sensitivity to image quality, enabling stable positioning even with highly reflective metal parts and damaged retired battery packs, thus solving the problem of easy failure of existing technologies under harsh conditions.

[0045] By constructing a 3D assembly relationship scene graph using graph neural networks, components are defined as nodes and assembly constraints as edges. Semantic reasoning is used to understand the functional relationships between parts, rather than relying on spatial proximity to combine components. This ensures the accuracy of identifying detachable components, preventing the inclusion of irrelevant parts or the omission of necessary fasteners, thus overcoming the shortcomings of existing technologies in understanding structural functions. An active vision strategy is introduced: when difference detection identifies occlusion or uncertainty, the system automatically plans the optimal recognition angle and controls the local camera to move to an unobstructed position for scanning. This overcomes the limitation of existing passive vision systems, which fail when encountering occlusion, and ensures the observability of critical components.

[0046] By adopting a hierarchical strategy of global-local collaboration, the global camera quickly completes coarse localization, while the local camera performs high-precision scanning only in key areas guided by the GNN. This concentrates computational and time resources on the core task, avoiding the problem of insufficient precision of the global camera and overcoming the disadvantages of limited field of view and low efficiency of the local camera, thus achieving the optimal balance between localization efficiency and accuracy. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is an overall flowchart of the precise positioning of irregularly shaped battery pack components based on the integrated disassembly topology of this invention.

[0049] Figure 2 This is a schematic diagram of the system hardware architecture according to an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of global scanning and CAD model registration according to an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram illustrating the construction and reasoning of the graph neural network assembly relationship scene graph in an embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram illustrating fine localization using local point cloud data, as an embodiment of the present invention.

[0053] Figure 6 This is a flowchart illustrating the precise positioning method for irregularly shaped battery pack components according to an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0055] Example:

[0056] It should be noted that the terms "comprising" and "having" and any variations thereof in the embodiments of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0057] See Figure 6 and combined Figure 1 This invention provides a method for precise positioning of irregularly shaped battery pack components based on integrated disassembly topology, comprising the following steps:

[0058] Step 110: Use a global camera to acquire global point cloud data of the battery pack to be disassembled, and register the global point cloud data with a preset battery pack CAD model to obtain a transformation matrix.

[0059] The registration process includes the following steps: extracting feature descriptors from the global point cloud and the point cloud sampled from the CAD model; performing coarse registration using a random sampling consensus algorithm to obtain an initial transformation matrix; performing fine registration using an iterative nearest-point algorithm based on the initial transformation matrix to obtain a final transformation matrix; and performing global calibration of the global camera and the robot base coordinate system to obtain a first hand-eye calibration matrix. base_ H _globalCam Hand-eye calibration is performed on the local camera and the robot end-effector coordinate system to obtain a second hand-eye calibration matrix. end_ H _localCam ; Using the robot's forward kinematics and the first hand-eye calibration matrix base_ H _globalCam Second hand eye calibration matrix end_ H _localCam Unifying global and local point cloud data into the robot's base coordinate system, i.e. robotBase _H_ pointCloud1 = robotBase_ H _globalCam * globalCam_ H _pointCloud1 and robotBase _H_ pointCloud2 = robotBase_ H _robotEnd * robotEnd_ H _localCam * localCam _H_ pointCloud2 .

[0060] Specifically, the feature descriptor is a vector representing the local geometric features of the point cloud; coarse registration is a registration method that quickly obtains the initial spatial alignment relationship; fine registration is a method that optimizes to obtain the precise alignment relationship; the first hand-eye calibration matrix is ​​a matrix representing the relationship between the global camera coordinate system and the robot base coordinate system, which can be a homogeneous matrix obtained based on the global calibration algorithm; the second hand-eye calibration matrix is ​​a matrix representing the relationship between the local camera coordinate system and the robot end flange coordinate system, which can be a transformation matrix calculated by the hand-eye calibration algorithm.

[0061] See Figure 3The central processing unit calls the point cloud feature extraction algorithm to sample point cloud feature descriptors for both the global point cloud and the CAD model. It then runs a random sampling consistency algorithm to find the initial spatial correspondence between the two types of point clouds, obtaining an initial transformation matrix. Using the initial transformation matrix as the initial value, it executes an iterative nearest-point algorithm to minimize the distance error between point clouds, obtaining the final transformation matrix. Next, it initiates a global calibration process, collecting relative position data between the global camera and the robot base through a calibration board, and runs a global calibration algorithm to obtain the first hand-eye calibration matrix. Finally, it initiates a hand-eye calibration process, moving the local camera to multiple poses using the robot, collecting calibration data, and calculating the second hand-eye calibration matrix. Finally, using the robot's forward kinematics, it uniformly transforms the global and local point cloud data to the robot base coordinate system using the first and second hand-eye calibration matrices, respectively.

[0062] Further, when extracting feature descriptors, the process includes: using a point cloud feature extraction algorithm to calculate the features of the global point cloud and the CAD model sampled point cloud; during global calibration, determining the relative positional relationship between the global camera coordinate system and the robot base coordinate system using a global calibration algorithm to obtain a first hand-eye calibration matrix; and during hand-eye calibration, determining the relative positional relationship between the local camera coordinate system and the robot end flange coordinate system using a hand-eye calibration algorithm to obtain a second hand-eye calibration matrix.

[0063] Specifically, the point cloud feature extraction algorithm can be a fast point feature histogram algorithm, which is an algorithm for calculating the geometric features of the point cloud; the global calibration algorithm is an algorithm for determining the relative position of the global camera and the robot base, and can be a calibration algorithm based on a calibration board; the hand-eye calibration algorithm is an algorithm for determining the relative position of the local camera and the end flange, and can be the Tsai-Lenz calibration algorithm.

[0064] Step 120: Based on the transformation matrix and the CAD model analysis results, construct a graph neural network assembly relationship scene graph to obtain a scene graph containing semantic relationships of parts.

[0065] The process of constructing the scene graph includes the following steps: parsing the assembly file of the CAD model, defining each independent component as a node, and defining the physical contact and assembly constraints between components as edges; constructing a graph neural network model based on the nodes and edges, learning the relationships between components and inferring semantic functions through the graph neural network model; and initiating a query to the graph neural network model to obtain a detachable component subgraph containing the target component and its associated parts, wherein the detachable component subgraph is formed based on the structural and functional associations of the components.

[0066] Specifically, the assembly file is a file that records the information of the parts in the CAD model, and can be a STEP format file; the node is a unit in the graph neural network that represents an independent part, and can be a node representing a BMS, screw, or copper busbar; the edge is a connection unit that represents the relationship between parts, and can be an edge representing the fixing relationship between the screw and the BMS; the detachable component sub-graph is a local graphic containing the target part and functionally related components, and can be a sub-graph containing the BMS and 6 fixing screws.

[0067] See Figure 4 The central processing unit reads the assembly file of the CAD model, parses it to obtain the geometric information and assembly constraints of each independent component, defines each component as a node, and defines assembly constraints such as physical contact, coaxiality, and surface mating as edges. Based on the dataset of nodes and edges, a graph neural network model is constructed. The model is input with training data labeled with semantic relationships. The model learns the association rules between components through iterative training, and then infers semantic functions such as "screw fixing BMS" and "copper busbar connecting module". According to the disassembly requirements, a query command is sent to the graph neural network model, such as "get the disassembled components of BMS". The model retrieves all components that are associated with the BMS node by edges and returns a subgraph of disassembled components containing the BMS and its 6 fixing screws.

[0068] Furthermore, when parsing the assembly file of the CAD model, the geometric information and assembly constraint relationships of the parts are extracted; when constructing the graph neural network model, the semantic relationships such as fixation and connection between parts are learned through the model to realize automatic prediction of disassembly interference and component function; when initiating a graph query, based on the node information of the target part, all parts that have edges associated with it are retrieved.

[0069] Specifically, geometric information refers to the dimensions and shape of components, such as the diameter of a screw or the length, width, and height data of the BMS; assembly constraints are the assembly restrictions between components, such as the coaxiality constraints between the screw and the BMS; semantic relationships are the functional associations between components, such as the fixing relationships between the screw and the BMS; and graph queries are operations for retrieving associated components, such as instructions to retrieve all associated screws based on BMS nodes.

[0070] Step 130: Perform difference detection based on the scene diagram and the registered CAD model and global point cloud data to obtain the component status marking results.

[0071] The difference detection process includes the following steps: comparing the registered CAD model with the global point cloud data in three dimensions to identify regions that exist in the CAD model but are missing in the global point cloud data, and abnormal regions that are absent in the CAD model but exist in the global point cloud data; based on the identification results of the missing and abnormal regions, marking the status of the corresponding components in the scene map as verified, uncertain occlusion, or abnormally damaged, respectively, and completing the status update of the scene map.

[0072] Specifically, missing areas are areas where the CAD model has corresponding parts but the global point cloud does not have corresponding points, such as the area where a screw is located that is obscured by a wire harness; abnormal areas are areas where the CAD model does not have corresponding parts but the global point cloud has points, such as the area where a dirt point cloud cluster with a diameter of 5cm is located; the status label result is an identifier of the actual status of the parts, which can be a label indicating that it has been verified, is uncertain due to obscuration, or is abnormally damaged.

[0073] The central processing unit calls the registered CAD model and global point cloud data, and starts the 3D comparison algorithm. It traverses each component in the CAD model, and checks whether there is a matching point cloud in the global point cloud data within a radius of 3mm, centered on its expected pose. If there is no matching point cloud, it is marked as a missing area. At the same time, it traverses the global point cloud data, identifies point cloud clusters in the CAD model that do not have corresponding components, and marks them as abnormal areas. According to the recognition results, if there is a point cloud match within a radius of 3mm at the expected position of the component, the corresponding node status in the scene graph is marked as verified. If it is a missing area and there is point cloud occlusion around it, it is marked as occlusion uncertain. If it is an abnormal area and the corresponding component has an abnormal shape, it is marked as abnormally damaged, and the status update of the scene graph is completed.

[0074] Step 140: Using the occlusion or uncertain state in the state labeling results, plan the optimal recognition angle for the local camera.

[0075] The process of planning the optimal recognition viewpoint includes the following steps: generating a virtual hemisphere around the expected location of the component whose state is marked as uncertain due to occlusion; obtaining multiple candidate viewpoints by sampling through the virtual hemisphere; calculating the spatial occlusion rate of each candidate viewpoint by combining the occlusion information in the global point cloud data and the robot motion constraints; and selecting the candidate viewpoint with the lowest occlusion rate and the highest information gain based on the spatial occlusion rate, thus determining it as the optimal recognition viewpoint.

[0076] Specifically, the virtual hemisphere is a hemispherical region constructed around the expected position of the occluded uncertain component. It can be a hemisphere with a radius of 5cm centered at that position. The candidate viewpoint is a potential shooting position sampled from the virtual hemisphere. It can be 20 evenly distributed sampling points. The spatial occlusion rate is an indicator that measures the degree of obstruction of the line of sight. It can be a value from 0 to 100%. The optimal recognition angle is the candidate viewpoint that meets the shooting requirements.

[0077] For components in uncertain occlusion states, the central processing unit constructs a virtual hemisphere with a radius of 5cm, centered on the expected location, to ensure coverage of the unobstructed observation range. Twenty candidate viewpoints are uniformly sampled on the surface of the virtual hemisphere, and the 3D coordinates of each viewpoint are recorded. The 3D coordinates of occluders in the global point cloud are acquired, and the line-of-sight vector from each candidate viewpoint to the target component is calculated. It is determined whether the line-of-sight vector intersects with the coordinates of the occluder, quantifying the degree of occlusion to obtain the spatial occlusion rate. Simultaneously, it verifies whether the candidate viewpoint is within the effective working distance of the local camera (0.2 to 0.7 meters), using a robot kinematics model to determine whether the robot can move to that viewpoint without collision. The viewpoint with the highest score—that is, the lowest occlusion rate, the highest information gain, and simultaneously meeting the working distance and robot accessibility requirements—is selected as the optimal recognition viewpoint.

[0078] Furthermore, when generating the virtual hemisphere, the expected position of the component in the uncertain occlusion state is used as the center of the sphere, and a reasonable radius is set to construct the virtual hemisphere to ensure that the candidate viewpoint covers the unobstructed observation range. When calculating the spatial occlusion rate, the three-dimensional coordinates of the occluder in the global point cloud and the line-of-sight vector of the candidate viewpoint are combined to determine whether the line of sight is obstructed, and the degree of occlusion is quantified to obtain the spatial occlusion rate. When selecting the best recognition angle, it is simultaneously verified whether the candidate viewpoint is within the effective working distance of the local camera and whether the robot can reach it to ensure the feasibility of the viewpoint.

[0079] Specifically, the reasonable radius is the radius of the hemisphere covering the unobstructed observation range, which can be 8 cm; the effective working distance is the distance range from which the local camera can clearly capture point clouds, which can be 0.2 to 0.7 m; robot reachability is the state in which the robot moves to the viewpoint without collision, which can be the state verified based on the robot's kinematic model; and the spatial occlusion rate can be a percentage value calculated based on the line-of-sight occlusion length.

[0080] Step 150: Control the local camera to move to the optimal recognition viewpoint to obtain local point cloud data, perform fine positioning using the local point cloud data, update the pose information of the parts in the scene map, and output the six-dimensional pose data of all target parts.

[0081] The fine localization process includes the following steps: extracting the point cloud of the target component from the local point cloud data using an instance segmentation network; processing the point cloud of the target component using an iterative nearest point algorithm or a six-dimensional pose estimator to obtain the true pose of the target component; updating the true pose to the scene graph and re-querying the state of the disassembled component subgraph; if there are still components with uncertain occlusion states, repeating the optimal recognition view planning and fine localization steps until all components in the disassembled component subgraph are marked as verified.

[0082] Specifically, the instance segmentation network is a network model that separates the point cloud of the target component, which can be a PointNet++ network; the true pose is the accurate six-dimensional pose data of the target component, which can be position and orientation data with an error of less than 0.1 mm; the state query is the operation of retrieving the state of the subgraph of the decomposable component, which can be an instruction to query whether all nodes are marked as verified.

[0083] The robot, equipped with a handheld 3D camera and a local eye, moves to the optimal recognition viewpoint, collecting approximately 50,000 local point cloud data points and transmitting them to the central processing unit (CPU). The CPU then invokes an instance segmentation network to accurately segment the point cloud of the target component from the local point cloud data, eliminating background and occlusion interference. It runs an iterative nearest-point algorithm or a six-dimensional pose estimator to process the point cloud of the target component, obtaining the true six-dimensional pose. The true pose is written to the scene graph, updating the corresponding component's state to "verified." A state query is initiated to the graph neural network model to retrieve the states of all nodes in the subgraph of the decomposable components. If any components still have uncertain states due to occlusion, the robot returns to the optimal recognition viewpoint planning step, repeating the viewpoint planning, local point cloud acquisition, and fine localization operations until all components are marked as "verified."

[0084] Furthermore, when using an instance segmentation network, a lightweight segmentation algorithm is selected to process the local point cloud data and accurately extract the point cloud regions of the target components; when updating the scene graph, the pose information and state markers of the components are modified simultaneously to ensure that the scene graph is consistent with the actual situation; when outputting six-dimensional pose data, the position and orientation information of all verified target components and their associated fasteners are integrated to generate a structured pose list.

[0085] Specifically, lightweight segmentation algorithms are computationally inefficient and highly efficient segmentation algorithms, which can be simplified versions optimized for point cloud data; local point cloud data refers to small-area, high-precision point clouds acquired by local cameras; the point cloud region of the target component is a set of point clouds containing only the target component, which can be the point cloud region of the head of an internal hex screw. The scene graph is an assembly relationship model built based on graph neural networks, which can be a network containing nodes and associated edges such as BMS, screws, and copper busbars; pose information is the six-dimensional position and orientation data of the component, which can be the specific values ​​of position (xyz) and orientation (rx ry rz); state markers are identifiers representing the actual state of the component, which can be verified, occluded and uncertain, or abnormally damaged; the actual situation is the real state and orientation of the component in a real disassembly scenario. Six-dimensional pose data refers to the three-dimensional position and three-dimensional attitude data of the component; the verified target component is the core disassembly component with the status marked as verified, which can be the BMS; the associated fastener is the component that has a fixed relationship with the target component, which can be the 6 screws that fix the BMS; the structured pose list is a set of pose data organized in a unified format, which can be tabular data containing component name, position coordinates, and attitude angle.

[0086] See Figure 2 This invention also provides a precise positioning system for irregularly shaped battery pack components based on integrated disassembly and topology relationships, comprising: a global eye-out 3D camera, a local eye-hand 3D camera, a robot system, a central processing unit, a CAD model storage module, and a graph neural network model module. The global eye-out 3D camera is fixed on the robot workstation gantry and signal-connected to the central processing unit. The local eye-hand 3D camera is mounted on the end effector flange of the robot system and signal-connected to the central processing unit. The robot system is signal-connected to the central processing unit. The CAD model storage module is signal-connected to the central processing unit. The graph neural network model module is signal-connected to the central processing unit. The central processing unit is used to execute the precise positioning method for irregularly shaped battery pack components based on integrated disassembly and topology relationships.

[0087] Specifically, the global eye-on 3D camera is a device that permanently acquires global point clouds, and can be a Mech-Mind high-precision industrial-grade 3D camera; the local eye-on 3D camera is a high-precision acquisition device installed at the end of the robot, and can be a Mech-Mind 3D camera; the robot system can be the Yuejiang CR10A 6-axis robot; the central processing unit can be an industrial PC equipped with an NVIDIA RTX 4080 GPU; the CAD model storage module can be a hard drive that stores STEP format files; and the graph neural network model module can be a functional module that runs GCN or DGCNN models.

[0088] The global eye-on 3D camera is fixed to the robot workstation gantry via a mechanical bracket and connects to the central processing unit (CPU) via gigabit Ethernet. The local eye-on 3D camera is bolted to the end effector flange of the robot and communicates with the CPU via the robot's built-in signal lines. The robot system connects to the CPU via a CAN bus to receive control commands and provide pose data. The CAD model storage module connects to the CPU via a SATA interface, and the graph neural network model module interacts with the CPU via a PCIe interface.

[0089] After system startup, the central processing unit (CPU) sends instructions to the CAD model storage module to read the CAD model corresponding to the battery pack to be disassembled. The CPU controls the robot system to move to the preset global scanning position, activates the global eye-on-the-hand 3D camera to collect global point cloud data, and transmits the data to the CPU for registration with the CAD model to obtain the transformation matrix. The CPU calls the graph neural network model module to construct an assembly relationship scene diagram based on the transformation matrix and the CAD model analysis results. Difference detection is performed through 3D comparison to mark the status of parts. For parts with uncertain occlusion states, the optimal recognition viewpoint is planned, and the robot system carrying the local eye-on-the-hand 3D camera moves to that viewpoint to collect local point cloud data. The CPU performs fine localization on the local point cloud data, updates the pose information in the scene diagram, and repeats the viewpoint planning and fine localization steps until all target parts are marked as verified. Finally, six-dimensional pose data is output for robot disassembly operations.

[0090] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0091] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for precise positioning of irregularly shaped components in a battery pack by integrating disassembly and topological relationships, characterized in that, Includes the following steps: The global point cloud data of the battery pack to be disassembled is acquired using a global camera, and the global point cloud data is registered with a preset battery pack CAD model to obtain a transformation matrix. Based on the transformation matrix and the CAD model analysis results, a graph neural network assembly relationship scene graph is constructed to obtain a scene graph containing semantic relationships of parts; Based on the scene diagram and the registered CAD model, difference detection is performed with the global point cloud data to obtain the component status marking results; Using the occlusion or uncertainty states in the state labeling results, the optimal recognition angle of the local camera is planned; The local camera is controlled to move to the optimal recognition viewpoint to acquire local point cloud data. The local point cloud data is used for fine localization, the pose information of the parts in the scene map is updated, and the six-dimensional pose data of all target parts is output.

2. The method according to claim 1, characterized in that, During registration, feature descriptors are extracted using the global point cloud and the point cloud sampled from the CAD model. Coarse registration is then performed using a random sampling consistency algorithm to obtain the initial transformation matrix. Based on the initial transformation matrix, the iterative nearest point algorithm is used for fine registration to obtain the final transformation matrix; The global camera and robot base coordinate system are globally calibrated to obtain the first hand-eye calibration matrix; The local camera and the robot end flange coordinate system are calibrated using hand-eye calibration to obtain a second hand-eye calibration matrix. By using the robot's forward kinematics and the first and second hand-eye calibration matrices, the global point cloud data and local point cloud data are unified into the robot's base coordinate system.

3. The method according to claim 1, characterized in that, When constructing the scene diagram, the assembly file of the CAD model is parsed, each independent component is defined as a node, and the physical contact and assembly constraints between components are defined as edges; Based on the nodes and edges, a graph neural network model is constructed, and the relationships between components are learned and semantic functions are inferred through the graph neural network model; A query is initiated to the graph neural network model to obtain a detachable component subgraph containing the target component and its associated parts. The detachable component subgraph is formed based on the structural and functional associations of the components.

4. The method according to claim 1, characterized in that, During the difference detection, the registered CAD model is compared with the global point cloud data in three dimensions to identify regions that exist in the CAD model but are missing in the global point cloud data, and abnormal regions that do not exist in the CAD model but exist in the global point cloud data. Based on the identification results of the missing and abnormal regions, the status of the corresponding components in the scene diagram is marked as verified, uncertain occlusion, or abnormally damaged, respectively, thus completing the status update of the scene diagram.

5. The method according to claim 1, characterized in that, When planning the optimal recognition perspective, a virtual hemisphere is generated around the expected location of the component whose status is marked as uncertain due to occlusion, and multiple candidate viewpoints are obtained by sampling through the virtual hemisphere. By combining the occlusion information in the global point cloud data and the robot motion constraints, the spatial occlusion rate of each candidate viewpoint is calculated; Based on the spatial occlusion rate, the candidate viewpoint with the lowest occlusion rate and the highest information gain is selected as the optimal recognition viewpoint.

6. The method according to claim 1, characterized in that, During fine localization, the local point cloud data is processed by an instance segmentation network to extract the point cloud of the target component; the point cloud of the target component is processed by an iterative nearest point algorithm or a six-dimensional pose estimator to obtain the true pose of the target component. Update the actual pose to the scene graph and query the state of the detachable component subgraph again; If there are still components with uncertain occlusion states, repeat the steps of optimal identification view planning and fine positioning until all components in the detachable component sub-graph are marked as verified.

7. The method according to claim 2, characterized in that, When extracting feature descriptors, a point cloud feature extraction algorithm is used to calculate the features of the global point cloud and the sampled point cloud of the CAD model; During global calibration, the relative positional relationship between the global camera coordinate system and the robot base coordinate system is determined by the global calibration algorithm, and the first hand-eye calibration matrix is ​​obtained. During hand-eye calibration, the relative positional relationship between the local camera coordinate system and the robot end flange coordinate system is determined by the hand-eye calibration algorithm, thus obtaining the second hand-eye calibration matrix.

8. The method according to claim 3, characterized in that, When parsing assembly files of a CAD model, extract the geometric information of the components and the assembly constraint relationships; When constructing a graph neural network model, the model learns semantic relationships such as fixation and connection between components to achieve automatic prediction of disassembly interference and component function. When initiating a graph query, based on the node information of the target component, all components that are associated with it by edges are retrieved.

9. The method according to claim 5, characterized in that, When generating a virtual hemisphere, the expected position of the component in the uncertain state of occlusion is used as the center of the sphere, and a reasonable radius is set to construct the virtual hemisphere to ensure that the candidate viewpoint covers the unobstructed observation range; When calculating the spatial occlusion rate, the three-dimensional coordinates of the occluders in the global point cloud and the line-of-sight vectors of the candidate viewpoints are combined to determine whether the line of sight is blocked, and the degree of occlusion is quantified to obtain the spatial occlusion rate. When selecting the best recognition perspective, it is necessary to simultaneously verify whether the candidate viewpoint is within the effective working distance of the local camera and whether the robot can reach it, to ensure the feasibility of the perspective.

10. A precise positioning system for irregularly shaped battery pack components that integrates disassembly and topological relationships, characterized in that, include: The system comprises a global eye-out 3D camera, a local eye-hand 3D camera, a robot system, a central processing unit (CPU), a CAD model storage module, and a graph neural network model module. The global eye-out 3D camera is fixed on the robot workstation gantry and is signal-connected to the CPU. The local eye-hand 3D camera is mounted on the end effector flange of the robot system and is signal-connected to the CPU. The robot system is signal-connected to the CPU. The CAD model storage module is signal-connected to the CPU. The graph neural network model module is signal-connected to the CPU. The CPU is used to execute the precise positioning method for irregularly shaped battery pack components based on the fusion and disassembly topology relationship as described in any one of claims 1 to 9.

Images