Disassembling method and system of battery pack disassembling robot based on image recognition

Abstract

The invention discloses a disassembling method and system of a battery pack disassembling robot based on image recognition, and relates to the technical field of image recognition, and the method comprises the steps: determining a corresponding abnormal region based on the image recognition of each battery pack image; determining corresponding abnormal features according to the image recognition of each abnormal area, and constructing a three-dimensional maintenance model of the battery pack based on a plurality of abnormal features, the overall form of the battery pack and the surrounding environment of the battery pack; a plurality of abnormal events are determined based on recognition of the three-dimensional maintenance model, and a disassembling path of a disassembling mechanical arm relative to the battery pack is determined according to the abnormal events, the corresponding abnormal areas and the disassembling mechanical arm of the battery pack disassembling robot. The intelligent disassembling system of the disassembling mechanical arm is determined based on the disassembling path, the position of the abnormal area and the posture of the battery pack, image recognition of all the battery pack images is achieved, all the abnormal areas are further controlled, and the accuracy of the intelligent disassembling system of the disassembling mechanical arm is improved.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a disassembly method and system for a battery pack disassembly robot based on image recognition. Background Technology

[0002] With the development of technology, battery pack disassembly robots are gradually being applied to battery pack maintenance production lines. These robots perform multi-level disassembly of battery packs in the maintenance production line. A battery pack disassembly robot is an industrial robot system specifically designed for disassembling used or damaged battery packs. It has a corresponding disassembly robotic arm. In the existing technology, the battery pack disassembly robot has a built-in camera and takes pictures of the battery pack from different angles to collect multiple images of the battery pack. Based on the graphic recognition of multiple images, it determines the corresponding abnormal features and disassembles each abnormal feature along a first path. However, it ignores the influence of the disassembly path and the position of the abnormal area, which leads to a decrease in the accuracy of the intelligent disassembly system of the disassembly robotic arm. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a disassembly method and system for a battery pack disassembly robot based on image recognition.

[0004] This invention provides a disassembly method for a battery pack disassembly robot based on image recognition, comprising: During the disassembly of the battery pack by the battery pack disassembly robot, the robot uses multiple built-in cameras to perform visual inspection of the battery pack, so as to collect multiple battery pack images in different dimensions, and determine the corresponding abnormal areas based on the image recognition of each battery pack image. In multiple abnormal regions, the corresponding abnormal features are determined based on the image recognition of each abnormal region. Based on multiple abnormal features, the overall shape of the battery pack, and the surrounding environment of the battery pack, a three-dimensional maintenance model of the battery pack is constructed. The three-dimensional maintenance model is a multi-layered, semantically rich scene graph, which includes a geometric layer, a semantic layer, a state layer, and a relational layer. Based on the identification of the three-dimensional maintenance model, multiple abnormal events are identified. Based on each abnormal event, the corresponding abnormal area, and the disassembly robot arm of the battery pack disassembly robot, the disassembly path of the disassembly robot arm relative to the battery pack is determined. Based on the disassembly path, the position of the abnormal area, and the attitude of the battery pack, the intelligent disassembly system of the disassembly robot arm is determined. Multiple disassembly nodes are identified based on the disassembly path, and corresponding intelligent disassembly behaviors are determined based on these multiple disassembly nodes, the intelligent disassembly system, and the disassembly robotic arm. The intelligence level of each intelligent disassembly behavior is determined based on the individual intelligent disassembly actions, the battery pack's posture, and the corresponding disassembly vibration data. A dynamic control system for the battery pack disassembly robot is then determined based on the intelligence level of each intelligent disassembly behavior, the corresponding behavioral trajectory, and the real-time disassembly video of the battery pack. The intelligence level includes excellent, normal, warning, or dangerous. The dynamic control system comprises high-level supervisory control and low-level real-time control. High-level supervisory control corresponds to the first level of dynamic control. High-level supervisory control and low-level real-time control are coordinated through an arbitrator or state machine.

[0005] This invention provides a disassembly system for a battery pack disassembly robot based on image recognition. The disassembly system is applied to the aforementioned disassembly method for a battery pack disassembly robot based on image recognition. The disassembly system includes: The first image recognition module is used to perform visual inspection of the battery pack based on multiple cameras built into the battery pack disassembly robot during the disassembly process of the battery pack disassembly robot, so as to collect multiple battery pack images in different dimensions, and determine the corresponding abnormal areas based on the image recognition of each battery pack image. The second image recognition module is used to determine the corresponding abnormal features in multiple abnormal areas based on the image recognition of each abnormal area, and to construct a three-dimensional maintenance model of the battery pack based on multiple abnormal features, the overall shape of the battery pack and the surrounding environment of the battery pack. The intelligent disassembly system module is used to identify multiple abnormal events based on the recognition of the three-dimensional maintenance model. Based on each abnormal event, the corresponding abnormal area, and the disassembly robot arm of the battery pack disassembly robot, the disassembly path of the disassembly robot arm relative to the battery pack is determined. Based on the disassembly path, the position of the abnormal area, and the attitude of the battery pack, the intelligent disassembly system of the disassembly robot arm is determined. The intelligent disassembly behavior module is used to determine multiple disassembly nodes based on the identification of the disassembly path, and to determine the corresponding intelligent disassembly behavior based on the multiple disassembly nodes, the intelligent disassembly system and the disassembly robotic arm; The dynamic control system module is used to determine the intelligent disassembly level of each intelligent disassembly behavior based on the attitude of the battery pack and the corresponding disassembly vibration data; and to determine the dynamic control system of the battery pack disassembly robot based on the intelligent disassembly level of each intelligent disassembly behavior, the corresponding behavior trajectory and the real-time disassembly video of the battery pack.

[0006] Compared with the prior art, the beneficial effects of the present invention are: In this embodiment of the invention, the method described herein, during the disassembly of a battery pack by a battery pack disassembly robot, performs visual inspection of the battery pack using multiple cameras built into the robot to collect multiple battery pack images from different dimensions. Based on image recognition of each battery pack image, corresponding abnormal areas are identified. Within these abnormal areas, corresponding abnormal features are determined based on image recognition of each abnormal area. A three-dimensional maintenance model of the battery pack is constructed based on these abnormal features, the overall shape of the battery pack, and its surrounding environment. Multiple abnormal events are identified based on the recognition of the three-dimensional maintenance model. The disassembly path of the disassembly robot relative to the battery pack is determined based on each abnormal event, the corresponding abnormal area, and the disassembly robot's robotic arm. Based on this disassembly path, the position of the abnormal area, and the battery pack's posture, an intelligent disassembly system for the disassembly robot is determined. This method introduces image recognition of the battery pack images and further identifies abnormal areas, managing multiple abnormal events. It incorporates a holistic consideration of the disassembly path, the position of the abnormal area, and the battery pack's posture, improving the accuracy of the intelligent disassembly system for the disassembly robot.

[0007] Therefore, multiple disassembly nodes are identified based on the disassembly path. Based on these nodes, the intelligent disassembly system, and the robotic arm, corresponding intelligent disassembly behaviors are determined. The intelligent disassembly level of each behavior is determined based on the battery pack's posture and corresponding vibration data. A dynamic control system for the battery pack disassembly robot is established based on the intelligent disassembly level of each behavior, its corresponding trajectory, and real-time disassembly video of the battery pack. This introduces intelligent disassembly behaviors, achieving a holistic consideration of the intelligent disassembly level, corresponding trajectory, and real-time disassembly video of the battery pack, thus improving the accuracy of the dynamic control system for the battery pack disassembly robot. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the disassembly method of the battery pack disassembly robot based on image recognition in an embodiment of the present invention. Figure 2 This is a flowchart illustrating step S11 of the disassembly method of the battery pack disassembly robot based on image recognition in an embodiment of the present invention. Figure 3 This is a flowchart illustrating step S12 of the disassembly method of the battery pack disassembly robot based on image recognition in an embodiment of the present invention. Figure 4 This is a flowchart illustrating step S13 of the disassembly method of the battery pack disassembly robot based on image recognition in an embodiment of the present invention. Figure 5This is a flowchart illustrating step S14 of the disassembly method of the battery pack disassembly robot based on image recognition in an embodiment of the present invention. Figure 6 This is a flowchart illustrating step S15 of the disassembly method of the battery pack disassembly robot based on image recognition in an embodiment of the present invention. Figure 7 This is a schematic diagram of the structural composition of the battery pack disassembly robot based on image recognition in an embodiment of the present invention. Detailed Implementation

[0009] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0010] Please see Figures 1 to 7 A disassembly method for a battery pack disassembly robot based on image recognition, applied to image recognition scenarios; the disassembly method for a battery pack disassembly robot based on image recognition includes: Step S11: During the disassembly of the battery pack by the battery pack disassembly robot, the battery pack is visually inspected by multiple cameras built into the battery pack disassembly robot to collect multiple battery pack images in different dimensions, and the corresponding abnormal areas are determined based on the image recognition of each battery pack image. Step S12: In multiple abnormal areas, determine the corresponding abnormal features based on the image recognition of each abnormal area, and construct a three-dimensional maintenance model of the battery pack based on multiple abnormal features, the overall shape of the battery pack and the surrounding environment of the battery pack. Step S13: Based on the identification of the three-dimensional maintenance model, identify multiple abnormal events. Based on each abnormal event, the corresponding abnormal area, and the disassembly robot arm of the battery pack disassembly robot, determine the disassembly path of the disassembly robot arm relative to the battery pack. Based on the disassembly path, the position of the abnormal area, and the attitude of the battery pack, determine the intelligent disassembly system of the disassembly robot arm. Step S14: Based on the identification of the disassembly path, determine multiple disassembly nodes, and based on the multiple disassembly nodes, the intelligent disassembly system, and the disassembly robotic arm, determine the corresponding intelligent disassembly behavior; Step S15: Determine the intelligent disassembly level of each intelligent disassembly behavior based on the attitude of the battery pack and the corresponding disassembly vibration data; determine the dynamic control system of the battery pack disassembly robot based on the intelligent disassembly level of each intelligent disassembly behavior, the corresponding behavior trajectory and the real-time disassembly video of the battery pack.

[0011] refer to Figure 2 In step S11, the specific steps are as follows: S111: When the battery pack is in the disassembly station, the battery pack disassembly robot disassembles the battery pack. Based on the current posture of the battery pack and the battery pack and the battery pack disassembly robot, the corresponding disassembly space system is determined. In the disassembly space system, the multiple cameras built into the battery pack disassembly robot are triggered to perform visual inspection of the battery pack. At this time, each side wall in the battery pack is visually inspected and multiple battery pack images in different dimensions are collected. S112: In multiple battery pack images, the corresponding component features are determined based on image recognition of each battery pack image, and the corresponding abnormal regions are determined according to the feature position, corresponding feature shape and corresponding abnormality coefficient of each component feature.

[0012] In the embodiments of this application, when the battery pack is at the disassembly station, the battery pack disassembly robot disassembles the battery pack. Based on the current posture of the battery pack and the battery pack and the disassembly robot, a corresponding disassembly space system is determined. In this disassembly space system, multiple cameras built into the battery pack disassembly robot are triggered to perform visual inspection of the battery pack. At this time, each side wall in the battery pack is visually inspected, and multiple battery pack images of different dimensions are collected. This takes into account the current posture of the battery pack and the overall consideration of the battery pack and the battery pack disassembly robot, ensuring the accuracy of the corresponding disassembly space system.

[0013] At this point, the system needs to construct a dynamic three-dimensional Cartesian coordinate system centered on the robot, i.e., disassemble the spatial system. This process typically involves establishing a base coordinate system with the center of the robot's base as the origin, and all subsequent pose descriptions and path planning will be performed within this coordinate system.

[0014] To accurately perceive the pose of the target object, the system uses a six-dimensional torque sensor or visual marker on the robot's end effector, combined with joint encoder data, to calculate the three-dimensional coordinates and orientation of the battery pack in the base coordinate system in real time. Meanwhile, hand-eye calibration is a key technology for establishing this spatial system. It achieves accurate conversion from two-dimensional image pixels to three-dimensional physical coordinates by solving the homogeneous transformation matrix between the camera coordinate system and the robot's end effector or base coordinate system.

[0015] Within a unified spatial framework, multiple cameras perform collaborative visual inspection tasks. These cameras are strategically positioned in different locations and perform different functions: global cameras are typically mounted above the work area to monitor the overall outline and orientation of the battery pack macroscopically; while local or hand-eye cameras are mounted at the robot's end effector, moving with the robotic arm to focus on close-range, high-precision detail inspection.

[0016] To achieve full-dimensional data acquisition, the system will acquire complementary information from spatial dimensions (different perspectives) and spectral dimensions (visible light, infrared, 3D structured light, etc.). The robot will also execute a full sidewall coverage strategy, controlling the end effector to perform a series of pose changes around the battery pack, moving sequentially to the front, side, and top surfaces to ensure that the key features of each sidewall are clearly captured.

[0017] Specifically, these technical details are fully demonstrated in the dynamic disassembly scenario of the robot disassembling the B-type battery pack. At the start of disassembly, the robot precisely transports the B-type battery pack to the disassembly station and closes and fixes it with the end effector. At this time, the data from the torque sensor and joint encoder are collected in real time, and the controller quickly calculates the precise pose of the battery pack in the robot's base coordinate system, for example, the position is (500mm, 300mm, 200mm) and the attitude is (0°, 5°, 0°). This pose data becomes the spatial origin of the current disassembly task. Since the hand-eye calibration has been completed, the system clearly knows the precise positional relationship of each camera relative to the robot.

[0018] The multi-camera collaborative detection program is triggered; the global camera installed directly above the workstation takes a top view, and the image recognition algorithm quickly locates the four corners of the battery pack and verifies them with the previously calculated pose data to ensure that the battery pack has not been accidentally shifted.

[0019] The robot began executing the sidewall inspection sequence: the robotic arm moved to the front of the battery pack, and the end effector camera acquired high-resolution RGB images at approximately 40cm to identify screws and maintenance switches on the panel; then the robotic arm moved to an upward angle, and a 3D structured light camera viewed the top surface at a 45-degree angle, generating a high-precision 3D point cloud through a projected grating pattern to detect minute deformations of the top cover; the robotic arm then moved to the right, and an infrared thermal imaging camera aimed at the side to inspect it, successfully capturing an abnormally high-temperature area, which may indicate an internal risk of thermal runaway.

[0020] Furthermore, in multiple battery pack images, the corresponding component features are determined based on image recognition of each battery pack image. The corresponding abnormal regions are determined according to the feature position, corresponding feature shape, and corresponding abnormal coefficient of each component feature. This comprehensive consideration of the feature position, corresponding feature shape, and corresponding abnormal coefficient of each component feature ensures the accuracy of the corresponding abnormal regions.

[0021] At this point, the goal of determining component features based on image recognition is to accurately identify all operationally significant components from multi-dimensional images and extract their descriptive information. This is usually achieved using deep learning-based object detection and semantic segmentation networks (such as the YOLO series, Mask R-CNN, etc.).

[0022] Object detection algorithms are responsible for locating components in an image and marking their position and category with bounding boxes. Semantic segmentation goes a step further, performing pixel-level classification to accurately delineate the outline of each component, which is crucial for assessing the morphological integrity of the component. After identifying the component, the system extracts a set of standardized descriptors, namely component features, for each instance. These features include: morphological vectors describing the geometric appearance and surface state of the component (such as length, width, roundness, texture roughness, etc.) combined with the three-dimensional spatial coordinates after S111 space system transformation, as well as color attributes such as dominant hue and saturation calculated in HSV or Lab color space, used to identify corrosion or contamination.

[0023] The system calculates a quantified anomaly coefficient for each component feature using a multimodal fusion evaluation model. This model is typically a weighted function that integrates morphological deviation scores, positional deviation scores, and texture anomaly scores. The morphological deviation score is obtained by comparing the extracted features with standard morphological templates in the database. The positional deviation score compares the actual position with the theoretical position in the standard battery pack CAD model. The texture or color anomaly score is evaluated using a pre-trained anomaly detection model. The system marks abnormal regions using threshold judgment. When the calculated anomaly coefficient exceeds a preset threshold, the entire region of the component or the aggregated continuous surface region is marked as an anomaly region, and a structured list containing the location, size, anomaly type, and anomaly coefficient is output.

[0024] Specifically, S112 begins processing the multi-dimensional data collected by S111; in the component feature determination stage, the system processes the front RGB image, the target detection model identifies 10 Phillips head screws and 1 high-voltage connector, and the semantic segmentation model accurately outlines their contours; the system then extracts the morphological features of each screw in the robot base coordinate system, such as its three-dimensional spatial position, radius, and roundness, as well as the position and color of the connector; the system processes the 3D point cloud of the top surface, and the point cloud processing algorithm analysis reveals a local protrusion in the center of the top cover with a diameter of about 5cm and a height of about 2mm. The position, height, and volume of this unexpected geometric feature are extracted; when processing the infrared thermal image on the right side, the temperature analysis algorithm detects a hot spot exceeding 65℃, and its centroid position, maximum temperature, and area are extracted as thermal features.

[0025] For the 10 screws, the system compared their shapes with the standard template and found that screw #3 had a slight stripped thread at the head, resulting in an increased shape deviation score and an anomaly coefficient of 0.75, exceeding the threshold of 0.7. Therefore, the location of screw #3 was marked as a stripped screw abnormal area. For the bulge on the top cover, its height and volume far exceeded the normal tolerance, resulting in an extremely high shape deviation score and an anomaly coefficient of 0.95. This area was marked as a top cover bulge abnormal area. For the hot spot on the side, the temperature of 65°C far exceeded the safety threshold. Its temperature distribution abnormal score and positional deviation score were both very high, with an anomaly coefficient as high as 0.98. Therefore, it was marked as an internal thermal runaway risk abnormal area. S112 output three clear abnormal areas: a stripped screw, a bulging top cover, and an overheated sidewall, each with precise spatial coordinates and a quantified risk level.

[0026] refer to Figure 3 In step S12, the specific steps are as follows: S121: Monitor multiple abnormal areas in real time, perform image recognition on each abnormal area, determine the corresponding abnormal features based on the image recognition of each abnormal area, and mark the feature location of the abnormal feature. S122: Collect the overall shape of the battery pack, determine the surrounding environment of the battery pack based on the detection of the surrounding area of ​​the battery pack, determine the first maintenance model based on the feature location of each abnormal feature and the overall shape of the battery pack, determine the second maintenance model based on the feature location of each abnormal feature and the surrounding environment of the battery pack, and construct a three-dimensional maintenance model of the battery pack based on the synthesis of the first maintenance model and the second maintenance model.

[0027] In the embodiments of this application, multiple abnormal regions are monitored in real time, and image recognition is performed on each abnormal region. Based on the image recognition of each abnormal region, the corresponding abnormal features are determined, and the feature position of the abnormal feature is marked. This approach takes into account the overall consideration of image recognition of each abnormal region and ensures the accuracy of the corresponding abnormal features.

[0028] At this point, the system continuously tracks multiple abnormal regions through a real-time monitoring loop. This is not a simple repetitive photo capture, but a closed-loop process of prediction-tracking-verification. The system instantiates a tracker, such as a Kalman filter or a deep learning tracker, for each abnormal region identified in S112. In each control cycle, the tracker predicts the location of the abnormal region in the current image based on the robot's kinematic model and the state at the previous moment. When a new image is acquired, the system performs target detection near the predicted location and correlates the detection results with the prediction. If a match is successful, the tracker state is updated with the new observations to correct the prediction error and ensure that the target is not lost even during brief occlusion or rapid movement.

[0029] The system performs more targeted and in-depth image recognition on each tracked anomaly area. Thanks to the accurate predicted location provided by the tracker, the system does not need to perform full image analysis. It only needs to crop out a small region of interest (ROI) containing the anomaly area, thus highly concentrating computing resources. Within the ROI, the system runs more specialized and refined recognition models: for geometric anomalies, high-resolution edge detection algorithms may be used to quantify the slippage depth or calculate the precise size of the crack; for state-related anomalies, spectral curves in specific bands will be analyzed or classification networks will be used to determine the type of leaked liquid. In addition, through the recognition results of multiple consecutive frames, the system can also perform temporal feature analysis to determine the dynamic trend of the anomaly, such as whether the hot spot temperature is continuously rising or whether the crack is expanding. This is an important basis for assessing the risk level.

[0030] Within the ROI, the system uses algorithms such as corner detection, centroid calculation, or contour fitting to improve the accuracy of feature location from pixel level to sub-pixel level. Combining the disassembly space system established in S111 and the precise hand-eye calibration matrix, the system uses this sub-pixel level two-dimensional image coordinates, along with the current camera's intrinsic and extrinsic parameters and depth information, to calculate the three-dimensional coordinates of the abnormal feature in the robot's base coordinate system in real time. This location information is appended with a high-precision timestamp, forming a dynamically updated data point, ensuring that the three-dimensional maintenance model is synchronized with the real world in time and space.

[0031] Specifically, the system has identified two abnormal areas: stripped screw #3 and a hot spot on the side. S121 begins real-time monitoring of these areas. While monitoring the stripped screw #3, the robot controls the robotic arm carrying the end-effector camera to approach it. During the movement, the Kalman filter continuously predicts the screw's position in the next frame image and successfully finds and updates the tracker in the new image. Even if the image is slightly blurred due to motion, the tracking remains stable. When the robotic arm is 10cm away from the screw, the system initiates high-resolution recognition within the ROI. The edge detection algorithm accurately delineates the edge of the cross groove and calculates the edge defect area caused by the stripping to be 0.15 square millimeters, which is a more accurate quantitative feature than the stripping in S112. The system uses the centroid algorithm to locate the screw center with sub-pixel precision and calculates its real-time three-dimensional coordinates in the robot base coordinate system as (502.3mm, 305.8mm, 198.2mm), while also adding a timestamp.

[0032] Meanwhile, a global infrared thermal imaging camera continuously scans the side of the battery pack at a frequency of 30Hz to monitor side hot spots; the tracker stably locks onto the hot spot area, and in each frame of the thermal map, the system calculates the highest temperature, average temperature, and area of ​​that area; by analyzing five consecutive frames of data, the system found that the highest temperature rose from 65.1℃ to 65.3℃, showing a slow upward trend, and this rate of temperature change became a new and key dynamic anomaly feature; the system also converts the centroid position of the hot spot area to the robot base coordinate system in real time through camera calibration parameters, and its coordinates are continuously updated with the possible slight displacement of the battery pack.

[0033] Furthermore, the overall shape of the battery pack is collected, and the surrounding environment of the battery pack is determined based on the detection of the surrounding area. A first-level maintenance model is determined based on the feature location of each abnormal feature and the overall shape of the battery pack. A second-level maintenance model is determined based on the feature location of each abnormal feature and the surrounding environment of the battery pack. A three-dimensional maintenance model of the battery pack is constructed based on the synthesis of the first-level and second-level maintenance models. This model takes into account the feature location of each abnormal feature and the overall consideration of the surrounding environment of the battery pack, ensuring the accuracy of the second-level maintenance model.

[0034] At this point, the system needs to collect the overall shape of the battery pack and its surrounding environment, which is the data foundation for model building. The overall shape of the battery pack is mainly obtained through three-dimensional vision sensors, such as using structured light scanning, laser contour scanning or multi-view stereo vision technology to generate high-density three-dimensional point clouds. At the same time, in order to build an occupation grid map or three-dimensional point cloud map of the workspace, the system will use 3D LiDAR or depth camera to perform a large-scale scan of the surrounding environment. This process not only needs to identify static objects, but also needs to detect and track dynamic objects.

[0035] The system constructs the first maintenance model based on the collected data, namely the internal-morphological model. This model uses the overall morphological point cloud of the battery pack as the geometric skeleton. After point cloud preprocessing and meshing algorithms, a continuous triangular mesh model is generated as the digital shell of the battery pack. The system attaches the anomaly features with precise three-dimensional coordinates updated in real time in S121 as semantic tags to this mesh model, that is, it associates the data of the anomaly features with the nearest vertex or facet in the mesh model. Therefore, the first model is a high-fidelity organ model with pathological markings, which accurately describes the geometry of the battery pack and clearly marks the location and attributes of all lesions.

[0036] The system identifies a second maintenance model, namely the external-environment model. This model uses a 3D map of the surrounding environment as a spatial framework and identifies different object instances through segmentation and clustering to construct a 3D scene map. In this scene map, the battery pack is precisely located as an independent object, and the 3D coordinates of the abnormal features obtained in S121 are also associated with the battery pack object as object attributes. Therefore, the second model is a macroscopic sandbox with the world coordinate system as a reference. It not only knows that there is a problem with the battery pack, but also knows the spatial relationship between the problem point and the robot, walls or other objects.

[0037] The system synthesizes these two models into the final 3D maintenance model of the battery pack. Since both are based on a unified disassembly space system, their synthesis is a data alignment and superposition in the coordinate system. The final 3D maintenance model is a multi-layered, semantically rich scene graph, which includes a geometric layer (precise 3D mesh), a semantic layer (object category labels), a state layer (dynamic attributes of abnormal features), and a relationship layer (describing the spatial relationships between objects). This comprehensive model is the only authoritative data source for subsequent path planning, risk assessment, and decision-making.

[0038] Specifically, during the data acquisition phase, the robot uses a 3D structured light camera at its end to perform a complete scan of the B battery pack, obtaining a high-precision point cloud containing details of the top cover bulge. At the same time, a 3D LiDAR above the workstation scans the entire work area, generating a point cloud map that includes the robot, battery pack, mobile tool cart, and waste bin.

[0039] When constructing the first maintenance model, the system meshes the point cloud data of the battery pack to form a fine 3D digital shell; the precise three-dimensional coordinates and stripped area of ​​the No. 3 screw (updated in real time by S121) of the stripped screw, as well as the coordinates of the hot spot on the side and the attribute of 65.3℃ and rising, are attached as semantic tags to the corresponding positions of the 3D shell; at this point, the model clearly shows a battery pack with pathological markings.

[0040] When constructing the first maintenance model, the system meshes the point cloud data of the battery pack to form a fine 3D digital shell; the precise three-dimensional coordinates and stripped area of ​​the No. 3 screw (updated in real time by S121) of the stripped screw, as well as the coordinates of the hot spot on the side and the attribute of 65.3℃ and rising, are attached as semantic tags to the corresponding positions of the 3D shell; at this point, the model clearly shows a battery pack with pathological markings.

[0041] The system merges the two models in a unified robot base coordinate system to construct a three-dimensional maintenance model. This final model not only knows that the battery pack itself has a bulge, a stripped screw, and an overheated hot spot, but it also understands that the hot spot that is heating up is very close to the waste bin containing flammable materials. Based on this, the model can issue a high-level instruction to subsequent steps: when planning the dismantling path, the risk of the hot spot must be dealt with first, and the movement trajectory of any robotic arm must be far away from the waste bin to prevent thermal runaway from triggering a chain reaction.

[0042] refer to Figure 4 In step S13, the specific steps are as follows: S131: Real-time monitoring of the three-dimensional maintenance model, determining the corresponding recognition mode based on the matching between the three-dimensional maintenance model and the battery pack disassembly robot, and triggering the recognition of the three-dimensional maintenance model along the recognition mode to identify multiple abnormal events; S132: In multiple abnormal events, the corresponding abnormal area is determined according to the identification of each abnormal event, and the disassembly robot arm of the battery pack disassembly robot is marked. Based on each abnormal event, the corresponding abnormal area and the disassembly robot arm of the battery pack disassembly robot, the disassembly path of the disassembly robot arm relative to the battery pack is determined. S133: Collect the attitude of the battery pack, determine the first intelligent disassembly parameters of the disassembly robot arm based on the disassembly path and the attitude of the battery pack, determine the second intelligent disassembly parameters of the disassembly robot arm based on the location of the abnormal area and the attitude of the battery pack, and determine the intelligent disassembly system of the disassembly robot arm according to the mapping relationship between the first intelligent disassembly parameters, the second intelligent disassembly parameters and the intelligent disassembly system.

[0043] In the embodiments of this application, the three-dimensional maintenance model is monitored in real time, and the corresponding recognition mode is determined based on the matching between the three-dimensional maintenance model and the battery pack disassembly robot. The recognition of the three-dimensional maintenance model is triggered along the recognition mode to identify multiple abnormal events. This approach is compatible with the overall consideration of matching between the three-dimensional maintenance model and the battery pack disassembly robot, ensuring the accuracy of the corresponding recognition mode.

[0044] At this point, the system proactively detects state changes by monitoring the three-dimensional maintenance model in real time. This monitoring is not passive observation, but rather active access to the model database at high frequency control cycles to check the status of all key data items. The system not only cares about the current value, but also calculates the trend of change through change detection algorithms, such as the temperature change rate of hot spots or the area change rate of cracks. The system has a large number of built-in triggers, which include not only simple numerical thresholds, but also more complex logical combinations. Once a trigger is met, it will send a high-level situational signal to the pattern matching module, indicating that the decision-making level needs to intervene.

[0045] The system has a pre-built pattern library. Each pattern is a set of expert knowledge for a specific scenario, such as safety-first pattern, efficiency-first pattern, or fine operation pattern. The matching process is a comprehensive evaluation of multiple factors. The system will select the pattern with the highest score as the current recognition pattern based on the highest risk type in the model, the capabilities of the robot's current end tool, and the overall task instructions issued by the upper layer, through weighted scoring or decision trees.

[0046] Each recognition pattern is bound to a specific inference engine or rule set, which is responsible for extracting anomalous events from the model's raw data. It aggregates relevant data points in the model and performs contextual analysis in conjunction with scene relationships, such as determining whether a high-temperature point is close to an electrolyte leakage area to assess the risk of combustion. The final generated anomalous event is a structured data object containing the event type, a list of all anomalous features that triggered the event, a priority determined by the pattern definition and the severity of the features, and a suggested handling strategy.

[0047] Specifically, its three-dimensional maintenance model already includes top cover bulge, No. 3 screw stripping, and side hot spot; during the real-time monitoring phase, the system detected the temperature of the side hot spot at second T as 65.3°C, with a temperature change rate of +0.2°C / min. This triggered the system's built-in IF(Temp>65°C)AND(dT / dt>0.1°C / min) logic, sending a situation signal indicating a deteriorating high temperature trend to the system.

[0048] Since the worsening high temperature trend is a high-risk signal, the score of the safety-first mode has been significantly improved. The system also confirmed that the robot's end effector could be replaced with a cooling device. After comprehensive evaluation, the safety-first mode was finally activated. Driven by this mode, its bound inference engine started to work and performed in-depth analysis of the model data. The engine used the data of the side hotspots as input and generated the first abnormal event: Event_001, with the type Thermal_Runaway_Risk (thermal runaway risk), the priority Critical (urgent), and the suggested strategy Immediate_Cooling (immediate cooling).

[0049] The engine detected a structural defect in the top cover bulge, which is considered high-risk even in safety-first mode, generating a second event: Event_002, of type Structural_Integrity_Compromised, with a priority of High, and a recommended strategy of Controlled_Depressurization; the engine evaluated that screw #3 was stripped, which is considered a low-risk mechanical problem in the current mode, generating a third event: Event_003, with a priority of Low; S131 outputs an ordered list containing the three abnormal events, with a priority of Event_001>Event_002>Event_003.

[0050] Furthermore, in multiple abnormal events, the corresponding abnormal areas are identified based on the identification of each abnormal event, and the disassembly robot arm of the battery pack disassembly robot is marked. Based on each abnormal event, the corresponding abnormal area, and the disassembly robot arm of the battery pack disassembly robot, the disassembly path of the disassembly robot arm relative to the battery pack is determined. This takes into account the overall consideration of each abnormal event, the corresponding abnormal area, and the disassembly robot arm of the battery pack disassembly robot, ensuring the accuracy of the disassembly path of the disassembly robot arm relative to the battery pack.

[0051] At this point, the system determines the abnormal region in physical space based on the abnormal event. This step maps the logical event back to the specific region. The system parses the trigger feature list in each abnormal event object and queries the three-dimensional geometric region corresponding to these features in the three-dimensional maintenance model. In some cases, an event may be triggered by multiple adjacent features. The system will use a clustering algorithm to aggregate them into a single, continuous abnormal region. In addition, for safety, the system will also spatially expand the region according to the type and risk level of the event to create a virtual safety boundary or restricted area, such as creating a thermal impact buffer for a high-temperature event.

[0052] The system needs to clearly define the disassembly robot arm to perform the task. In a multi-arm collaborative system, the system will assign the most suitable robot arm based on the location of each abnormal area, the required operation type, and the current load and tool configuration of each robot arm. In a single-arm system, the system will mark the currently unique robot arm by default. The system loads the kinematic model of the marked robot arm, including its DH parameters, joint constraints, and the size and shape of the end effector. These data are the basic constraints for subsequent path planning.

[0053] The system needs to clearly define the disassembly robot arm to perform the task. In a multi-arm collaborative system, the system will assign the most suitable robot arm based on the location of each abnormal area, the required operation type, and the current load and tool configuration of each robot arm. In a single-arm system, the system will mark the currently unique robot arm by default. The system loads the kinematic model of the marked robot arm, including its DH parameters, joint constraints, and the size and shape of the end effector. These data are the basic constraints for subsequent path planning.

[0054] Specifically, S131 has output three abnormal events with decreasing priority: Event_001 (risk of thermal runaway), Event_002 (damage to structural integrity), and Event_003 (screw stripping). During the abnormal area determination stage, the system resolved that the triggering feature of Event_001 is the side hot spot B and located its three-dimensional center point. Since this is a critical event, the system created a spherical high-temperature restricted area with a radius of 8cm centered on this point. The system located the triggering feature of Event_002, the top cover bulge A, to its surface area and marked it as a structurally vulnerable area. The system located the three-dimensional center point of the triggering feature of Event_003, screw stripping C.

[0055] During the marking and disassembly of the robotic arm phase, the system confirms that only the robot is currently available, and then loads the kinematic model of its six-DOF robotic arm and the parameters of the multi-functional tool head currently installed at the end effector; during the determination of the disassembly path phase, the system determines the access order according to priority as follows: high temperature restricted area > structurally vulnerable area > screw C.

[0056] The system uses the RRT* algorithm to plan the path in segments: planning a path from the current location to the safety observation point at the edge of the high-temperature restricted area; then planning a path from that observation point to the operation preparation point near the structurally vulnerable area, ensuring that it does not pass over the restricted area; planning a very fine path to accurately reach the final operation point on the bulge surface; and finally planning the approach and operation path from the bulge to screw C.

[0057] The system smoothly connects these path segments and performs global optimization, ultimately generating a complete and ordered pose sequence. This path not only specifies the precise location the robot must reach, but also enforces a safe sequence of handling thermal risks first, then structural risks, and finally mechanical problems, while intelligently avoiding all known and potential dangers throughout the process.

[0058] Therefore, the battery pack's orientation is collected, and the first intelligent disassembly parameters for the disassembly robot arm are determined based on the disassembly path and the battery pack's orientation. The second intelligent disassembly parameters for the disassembly robot arm are determined based on the location of the abnormal area and the battery pack's orientation. The intelligent disassembly system for the disassembly robot arm is determined according to the mapping relationship between the first intelligent disassembly parameters, the second intelligent disassembly parameters, and the intelligent disassembly system. This overall consideration ensures the accuracy of the intelligent disassembly system for the disassembly robot arm. At the same time, image recognition of the battery pack image is introduced, and further image recognition of abnormal areas is performed to control multiple abnormal events. This overall consideration of the disassembly path, the location of the abnormal area, and the battery pack's orientation improves the accuracy of the intelligent disassembly system for the disassembly robot arm.

[0059] At this point, the system needs to acquire the real-time attitude of the battery pack, which is the reference frame for all parameter calculations, ensuring that all actions are relative to the current real state of the target. The attitude data mainly comes from the robot's forward kinematics (calculating the end effector pose by reading the joint encoder values) and the external vision system (for independent verification), and must be continuously updated at a high frequency to cope with possible minute attitude changes of the battery pack.

[0060] The system determines the first intelligent disassembly parameters, which define how the robotic arm moves along the path planned by S132. Its core task is to convert the pose point sequence in Cartesian space into the target angle sequence of each joint of the robot through inverse kinematics calculation. The system interpolates between adjacent joint angle points to generate a smooth and continuous joint space trajectory (such as using an S-shaped velocity curve), and calculates the output torque required by each joint to track the trajectory based on the dynamic model of the robotic arm, so as to compensate for the effects of gravity, inertia and other factors in advance.

[0061] The system determines the second intelligent disassembly parameters, which define the specific operations to be performed after the robotic arm reaches the target area. These parameters include end effector parameters that are precisely set according to the characteristics of the abnormal area, such as force control parameters (target contact force, force control mode), motion parameters (speed, torque) and clamping force. More importantly, it also includes expert knowledge directly related to the disassembly process, namely process parameters, such as defining specific operation types (probing, drilling, loosening) and the timing logic of the actions (such as applying force first and then rotating).

[0062] The system integrates all parameters into an executable intelligent disassembly system through a predefined mapping relationship. This mapping relationship specifies how to combine the first and second parameters into a complete instruction block. The system traverses the entire path generated by S132 and generates a complete instruction sequence for each critical path segment. The final product is a structured data packet, which is sent to the robot's underlying motion controller to drive the robot to accurately complete the entire intelligent disassembly task.

[0063] Specifically, S132 has planned a path that prioritizes the processing of side hotspots, with the path ending at Pose_Final_Hotspot. During the battery pack attitude acquisition phase, the robot calculates the current precise attitude of battery pack B through joint encoders, such as a small pitch angle of 5.2°. This data becomes the key input for all subsequent coordinate transformations.

[0064] In the first stage of determining the intelligent disassembly parameters, the system acquires the path pose sequence from the starting point to the hot spot. Through inverse kinematics calculation, each pose point is converted into a target angle sequence of 6 joints. For example, the endpoint pose is calculated as a set of specific joint angles. The system uses S-shaped velocity curves to interpolate these angle points, generate a smooth joint trajectory, and calculate the additional torque that each joint needs to output during the acceleration phase.

[0065] In the stage of determining the second intelligent disassembly parameters, the system identifies that the path endpoint corresponds to the side hot spot abnormal area; according to the preset process rules for thermal risks, the system sets the end effector parameters: the operation type is Apply_Coolant (apply coolant), the tool is switched to the cooling nozzle, the spray pressure is set to 0.3MPa, the coolant flow rate is 50ml / min, and the timing logic of continuous spraying for 15 seconds after reaching the target point is defined.

[0066] During the stage of determining the intelligent disassembly system, the system calls the mapping template to package the first parameter (smooth joint trajectory) and the second parameter (detailed cooling operation instructions) into an instruction block. This instruction block is sent to the robot's controller. After receiving it, the robot moves precisely and smoothly to the hot spot and performs a precise cooling operation for 15 seconds. This complete and parameterized instruction package is the intelligent disassembly system for this abnormal event.

[0067] refer to Figure 5 In step S14, the specific steps are as follows: S141: Collect disassembly paths, determine multiple disassembly areas based on the distribution location of the disassembly paths, the corresponding path shape and the three-dimensional maintenance model of the battery pack, and determine the corresponding disassembly nodes according to the regional location of each disassembly area, the corresponding disassembly components and the disassembly robotic arm, so as to collect multiple disassembly nodes. S142: Determine the first disassembly behavior based on multiple disassembly nodes and the intelligent disassembly system; determine the second disassembly behavior based on multiple disassembly nodes and the disassembly robotic arm; and determine the corresponding intelligent disassembly behavior based on the control system of the first disassembly behavior, the second disassembly behavior, and the disassembly robotic arm.

[0068] In the embodiments of this application, disassembly paths are collected, and multiple disassembly areas are determined based on the distribution location of the disassembly paths, the corresponding path shapes, and the three-dimensional maintenance model of the battery pack. According to the regional location of each disassembly area, the corresponding disassembly components, and the disassembly robotic arm, the corresponding disassembly nodes are determined to collect multiple disassembly nodes. This approach takes into account the overall consideration of the regional location of each disassembly area, the corresponding disassembly components, and the disassembly robotic arm, ensuring the accuracy of the corresponding disassembly nodes.

[0069] At this point, the system directly obtains the complete path output by S132 and performs geometric morphology analysis on it, including calculating the curvature of each point on the path to identify straight and turning segments, calculating the proximity of path points to objects in the model, and identifying acceleration and deceleration segments by referencing the speed curve. Based on these analyses, the system divides the path into several logically independent decomposition areas, such as a navigation area that is far from the target and mainly involves high-speed translation; an approach area that slowly approaches the target; an operation area that physically interacts with the target or performs precise detection; and an avoidance area that is generated to avoid obstacles.

[0070] The system generates nodes at key locations based on the attributes of each region. For example, it generates boundary nodes at the start and end points of each region as state transition markers, inserts intermediate nodes within long-distance regions to ensure tracking accuracy, and generates core operation nodes at the end of the operation region. Each generated node is a rich data structure whose attributes are determined by fusing multi-source information, including the 3D coordinates obtained from the path, the region or target component determined by querying the 3D maintenance model, and the joint angles and configurations of the robotic arm when it reaches the node, calculated through forward kinematics. The system organizes all nodes with complete attribute tags into an ordered list according to the path sequence, translating a continuous path into a discrete, segmented itinerary with a clear task objective.

[0071] Specifically, S132 has planned a path that prioritizes the handling of side hot spots. During the disassembly area determination phase, the robot acquires this complete path. Through morphological analysis, the system discovers that the first 80% of the path is a long-distance straight-line translation, while the last 20% is a slow curve approaching the side of the battery pack. Through proximity analysis, the system finds that at 80% of the path, the distance to the battery pack begins to decrease rapidly. Based on this, the system divides the area from the starting point to 80% of the path into a navigation area, characterized by high speed and long distance; the area from 80% to 95% into an approach area, characterized by deceleration and approach; and the last 5% into an operation area, characterized by low speed and high precision, directly facing the side hot spots.

[0072] During the node determination phase, the system generates four nodes at the boundaries and key points of the aforementioned areas: Node_Nav_Start (start of navigation area) at the path start point; Node_Approach_Start (end of navigation area, start of approach area) at 80% of the path; Node_Operation_Start (end of approach area, start of operation area) at 95% of the path; and Node_Operation_Final (core task trigger node) at the path end point.

[0073] During the node attribute marking phase, for Node_Nav_Start and Node_Approach_Start, since they are only transition points, their disassembled component fields are empty; while for Node_Operation_Final, the system determines its location as 10cm directly in front of the hot spot center by querying the 3D maintenance model, and the disassembled component is the hot spot area on the side of the B battery pack; at the same time, the system calculates the joint configuration of the robotic arm when it reaches this point through inverse kinematics, and confirms that the posture is stable under this configuration, which is suitable for cooling operations.

[0074] Furthermore, the first disassembly behavior is determined based on multiple disassembly nodes and the intelligent disassembly system, and the second disassembly behavior is determined based on multiple disassembly nodes and the disassembly robot arm. The corresponding intelligent disassembly behavior is determined according to the control system of the first disassembly behavior, the second disassembly behavior, and the disassembly robot arm. This approach takes into account the overall considerations of the first disassembly behavior, the second disassembly behavior, and the control system of the disassembly robot arm, ensuring the accuracy of the corresponding intelligent disassembly behavior.

[0075] At this point, the nodes provided by S141 are discrete key pose points. The core task of the first decomposition behavior is to perform high-order interpolation between these nodes to generate a smooth joint space trajectory that is continuous in terms of position, velocity, acceleration, and even jerk. The generated smooth trajectory is then decomposed into a series of high-frequency control commands and sent to the servo drivers of each joint of the robotic arm. The closed-loop controller inside the driver will adjust the motor output torque in real time according to the encoder feedback to ensure that the joint accurately tracks the target trajectory.

[0076] The second disassembly behavior is event-driven. When the main controller's state machine detects that the robotic arm has reached a specific operation node, it will trigger an event signal. This event will activate an operation subroutine to extract the second intelligent disassembly parameters associated with that node from the intelligent disassembly system of S13, such as torque, speed, and clamping force. These parameters are then converted into specific control commands for the end effector or related sensors, such as setting speed and torque limits for the electric screwdriver, or activating the force / position hybrid controller to maintain a preset contact force during the interaction.

[0077] The system uses a unified control framework to orchestrate macroscopic motion and microscopic operations in a time sequence to determine the final intelligent disassembly behavior. This behavior synthesis is achieved through a finite state machine in a hierarchical control system. The state machine enters the NAVIGATING state, activating the first level of behavior and driving the robotic arm to move along the node sequence. When a node is reached, the state machine evaluates whether it is an operation node. If so, it pauses the first level of behavior, switches to the EXECUTING_ACTION state, and activates the second level of behavior to perform the specific operation. After the operation is completed, the state machine resumes the first level of behavior and continues to the next node. This complete execution flow, orchestrated by the state machine and seamlessly switching between macroscopic motion and microscopic operations, is the intelligent disassembly behavior. It ensures that the robot can not only reach the target location but also do it correctly, in the right place.

[0078] Specifically, S141 has generated a node sequence for processing the side hot spot, and S13 has generated a corresponding intelligent disassembly system. In the stage of determining the first disassembly behavior, the robot's main controller acquires the node sequence, performs spline interpolation between these nodes, and generates a smooth joint space trajectory. The controller decomposes the trajectory into high-frequency instructions and sends them to the servo drivers of each joint, driving the robotic arm to start moving smoothly from the starting point to the front of the hot spot.

[0079] In the second disassembly phase, when the robotic arm is about to reach the end operation node, the motion controller sends an approach target signal. After receiving the signal, the top-level state machine switches to the EXECUTING_ACTION state and extracts the second parameter for this node from the intelligent disassembly system of S13: Apply_Coolant, which lasts for 15 seconds. The controller then sends an opening command to the solenoid valve of the end effector and starts a 15-second timer to begin the cooling operation.

[0080] During the intelligent disassembly behavior determination phase, the robot's layered state machines begin to work collaboratively; the system enters the NAVIGATING state and executes the first behavior, with the robotic arm moving towards the target point; upon arrival, the system switches to the EXECUTING_ACTION state and executes the second behavior, opening the coolant valve to begin spraying; after 15 seconds, the timer is triggered, and the system enters the ACTION_COMPLETED state, closing the valve; the system then re-enters the NAVIGATING state and executes the first behavior, moving the robotic arm to the preset safe evacuation point.

[0081] refer to Figure 6 In step S15, the specific steps are as follows: S151: Real-time monitoring of each intelligent disassembly behavior, and determination of the disassembly vibration data of the battery pack based on the dynamic monitoring of the battery pack. The first intelligent disassembly coefficient is determined based on each intelligent disassembly behavior and the corresponding disassembly vibration data. The second intelligent disassembly coefficient is determined based on each intelligent disassembly behavior and the attitude of the battery pack. The intelligent disassembly level of the intelligent disassembly behavior is determined based on the mapping relationship between the first intelligent disassembly coefficient, the second intelligent disassembly coefficient and the intelligent disassembly level. S152: Mark the behavior trajectory of each intelligent disassembly behavior and collect real-time disassembly video of the battery pack. Determine the first dynamic control content of the battery pack disassembly robot based on the intelligent disassembly level of each intelligent disassembly behavior and the real-time disassembly video of the battery pack. S153: Determine the second level of dynamic control content for the battery pack disassembly robot based on the behavior trajectory of each intelligent disassembly behavior and the real-time disassembly video of the battery pack; determine the dynamic control system of the battery pack disassembly robot based on the first level of dynamic control content and the second level of dynamic control content.

[0082] In the embodiments of this application, each intelligent disassembly behavior is monitored in real time, and the disassembly vibration data of the battery pack is determined based on the dynamic monitoring of the battery pack. A first intelligent disassembly coefficient is determined based on each intelligent disassembly behavior and the corresponding disassembly vibration data. A second intelligent disassembly coefficient is determined based on each intelligent disassembly behavior and the attitude of the battery pack. The intelligent disassembly level of the intelligent disassembly behavior is determined based on the mapping relationship between the first intelligent disassembly coefficient, the second intelligent disassembly coefficient, and the intelligent disassembly level. This approach takes into account the overall consideration of the mapping relationship between the first intelligent disassembly coefficient, the second intelligent disassembly coefficient, and the intelligent disassembly level, ensuring the accuracy of the intelligent disassembly level of the intelligent disassembly behavior.

[0083] At this point, the system establishes the data foundation for evaluation through real-time monitoring and data acquisition, which requires high synchronization and high precision. The system obtains metadata of the currently executing intelligent disassembly behavior in real time through the internal interface of the robot controller, including behavior ID, type and current stage. At the same time, in order to capture the dynamic response of the battery pack during disassembly, the system relies on high-sensitivity sensors, such as high-frequency accelerometers or IMUs installed on the robot wrist or tooling fixtures, to collect vibration data, and obtains the six-dimensional pose of the battery pack in real time through robot forward kinematics calculation or external 3D vision system.

[0084] The system determines the first intelligent decomposition coefficient, namely the process quality coefficient, to quantify and evaluate the smoothness of the current behavior execution process and the quality of physical interaction. The system associates and matches the real-time collected vibration data with the current behavior type, and each behavior has a predefined ideal vibration characteristic model. By processing the vibration time-domain signal in real time, the system extracts time-domain features such as root mean square and peak value, and analyzes its spectral distribution through fast Fourier transform. The system compares the extracted real-time features with the ideal model and calculates a normalized coefficient through a weighted evaluation function or anomaly detection algorithm. The closer the coefficient is to 1, the more abnormal the vibration and the worse the process quality.

[0085] The system determines a second intelligent disassembly coefficient, namely the state stability coefficient, to quantify and evaluate the spatial stability of the disassembly target throughout the process. The system associates the real-time collected battery pack attitude data with the current behavior and compares the current attitude with a reference attitude to calculate the deviation. The system calculates a normalized coefficient based on the magnitude and rate of change of the attitude deviation. The closer the coefficient is to 1, the more unstable the attitude is, and there may be a risk of loosening of the fixation or violent release of internal stress.

[0086] The system combines the evaluations from two dimensions into an intuitive and action-guided intelligent decomposition level through a decision fusion step. The system has a built-in decision fusion model that defines how to map from the two coefficients to the final level. This model can be an IF-THEN rule base, a fuzzy logic controller, or a small neural network. The final output intelligent decomposition level is a discrete label with clear guidance. The intelligent decomposition level includes excellent, normal, warning, or danger. This level is the direct trigger for S152 to perform dynamic adjustment.

[0087] Specifically, the robot is performing an unscrewing action, loosening a severely corroded screw. During the real-time monitoring and data acquisition phase, the system monitors the type and stage of the current action. The wrist IMU collects vibration data at a frequency of 1 kHz, while the 3D vision system calculates the battery pack attitude at a frequency of 30 Hz. In the stage of determining the first intelligent disassembly coefficient, at the moment of loosening, the vibration signal collected by the IMU shows a very high instantaneous peak. The system's FFT analysis shows that an energy peak far exceeding the normal background noise appears near 500 Hz, which is consistent with the typical characteristics of metal viscous slip.

[0088] Based on this, the system calculated the first intelligent disassembly coefficient to be 0.85, indicating that a serious physical anomaly occurred in the process. In the stage of determining the second intelligent disassembly coefficient, at the moment when the maximum torque was applied, the 3D vision system detected that the battery pack underwent an instantaneous twist of 0.5 degrees around the Z-axis, which was much greater than the preset safety threshold of 0.1 degrees. Based on this, the system calculated the second intelligent disassembly coefficient to be 0.6, indicating that the stability of the target was challenged.

[0089] During the stage of determining the intelligent disassembly level, there is a rule in the system's rule base: IF (first coefficient > 0.8) OR (second coefficient > 0.7) THEN level = dangerous. Since the first coefficient 0.85 exceeds the threshold, the system immediately assesses the intelligent disassembly level of the current behavior as dangerous. This dangerous level is immediately sent to S152, triggering dynamic control. The system may immediately command the robot to stop applying torque and slightly reverse the screw to release stress and prevent the screw from breaking completely.

[0090] Furthermore, the behavioral trajectories of each intelligent disassembly action are marked, and real-time disassembly videos of the battery pack are collected. Based on the intelligent disassembly level of each intelligent disassembly action and the real-time disassembly videos of the battery pack, the first level of dynamic control content of the battery pack disassembly robot is determined. This takes into account both the intelligent disassembly level of each intelligent disassembly action and the real-time disassembly videos of the battery pack, ensuring the accuracy of the first level of dynamic control content of the battery pack disassembly robot.

[0091] At this time, the system maintains a behavior execution context in memory. When an intelligent disassembly behavior is activated, its corresponding disassembly node sequence is loaded and marked as being executed, forming a clear logical view of executed-being executed-to be executed. This trajectory represents the current plan and intent. At the same time, through high-definition industrial cameras deployed at the workstation, the system continuously collects objective video streams of the battery pack disassembly process, providing external reality evidence for decision-making that is not affected by the internal state of the robot.

[0092] The system determines the first level of dynamic control content through multimodal fusion evaluation. The system uses the intelligent disassembly level assessed by S151 as the most critical decision input. At the same time, it uses computer vision to analyze real-time disassembly video to verify or supplement the level judgment, including detecting whether new anomalies that are not marked in the model appear in the video, such as cracks, leaks or smoke, and confirming whether the actual state of the objects in the video is consistent with the state believed by the controller.

[0093] Based on the results of this fusion assessment, the system selects and generates specific first-level dynamic control content from a predefined control strategy library. This content consists of discontinuous, strategic instructions designed to change the course of the current behavior, including: continuing execution when the level is excellent or normal; reconfiguring parameters when the level is warning, such as reducing speed or torque; immediately pausing and rolling back the behavior when the level is dangerous or a sudden anomaly is detected; and if the problem cannot be resolved after rolling back, then behavior replacement and replanning are performed, requesting a new alternative path or behavior sequence from a higher-level task planner.

[0094] Specifically, the robot is performing an unscrewing action to loosen a rusty screw, which S151 has assessed as dangerous. During the stage of marking the behavior trajectory and collecting real-time video, the system marks the trajectory of the current unscrewing behavior and clarifies that the robotic arm is at the endpoint and is performing the loosening action. At the same time, the high-definition camera at the workstation is clearly capturing images of the screw and the surrounding area. During the stage of determining the first level of dynamic control, the system receives the danger level from S151 as the highest priority alarm.

[0095] Meanwhile, the video analysis module processes the current frame, and its anomaly segmentation model detects a bright area at the root of the screw whose features highly match those of a metal crack. After integrating the two strong negative signals of danger level and crack detection, the system determines that the current loosening behavior is highly likely to cause the screw to break. Therefore, the system selects the highest level of intervention from the strategy library and immediately sends the first level of dynamic control to the robot's controller: urgently stop all movement, execute the rollback procedure to the previous safe node, and send a replanning request to S13, along with the constraint that there is a crack at the root of the screw and that the use of high torque is prohibited.

[0096] Therefore, the second level of dynamic control for the battery pack disassembly robot is determined based on the behavioral trajectories of each intelligent disassembly behavior and the real-time disassembly video of the battery pack. The dynamic control system of the battery pack disassembly robot is then determined based on the first and second levels of dynamic control, incorporating a holistic consideration of both levels to ensure the accuracy of the system. Furthermore, the introduction of intelligent disassembly behaviors allows for a comprehensive consideration of the intelligent disassembly level of each behavior, its corresponding behavioral trajectory, and the real-time disassembly video of the battery pack, thereby improving the accuracy of the dynamic control system.

[0097] At this point, in each frame of video, the system uses computer vision algorithms to lock and track the visual features of the target component in real time. By comparing the current coordinates of the features in the image with the expected coordinates, the image spatial error vector is calculated. This image error is mapped to the compensation speed or pose fine-tuning amount that the robot end needs to perform through the image Jacobian matrix. The final generated second dynamic control content is this real-time, continuous compensation vector, which is added to the original motion command generated by S142 in real time as a superposition signal to achieve precise correction.

[0098] The system employs a hierarchical control architecture to define the final dynamic control system, which comprises two layers: high-level supervisory control and low-level real-time control. High-level supervisory control, corresponding to the first layer of dynamic control, is a slow loop (second-level), responsible for decision-making and strategy changes, outputting discrete, event-driven instructions. Low-level real-time control, corresponding to the second layer of dynamic control, is a fast loop (millisecond-level), responsible for precise trajectory tracking and real-time fine-tuning, outputting continuous, high-frequency compensation signals. The high-level supervisory control and low-level real-time control coordinate through an arbitrator or state machine: in the absence of high-level intervention, the low-level real-time control loop continues to operate; when the high-level issues an instruction, the arbitrator suspends the low-level control, executes the new strategy, and then reactivates it. Specifically, the robot has received the instruction from S152 to re-tighten the screw that is at risk of stripping in a slow, low-torque mode; in the second stage of determining the dynamic control content, the robot's nominal trajectory is to move towards the center of the screw, while the high-definition camera captures video in real time; the visual servo loop locks the center point of the cross-head screw cap in real time, and because the stripping causes a slight offset in the center of the screw cap, the system calculates an image error.

[0099] Using the image Jacobian matrix and control law, the system calculates that the robot end effector needs to perform a small pose compensation, such as (moving -0.1mm along the X-axis and +0.07mm along the Y-axis). This fine-tuning vector is the second layer of dynamic control. It is superimposed on the robot's forward motion command at a high frequency, so that the screwdriver moves laterally in real time during the forward movement, always aligning with the center of the stripped screw.

[0100] During the stage of establishing the dynamic control system, the robot's control system now operates on a complete hierarchical architecture. The higher level (the first level of dynamic control) is executing the slow, low-torque strategy issued by S152, setting the macroscopic boundaries of operation, such as the upper limits of rotational speed and torque. The lower level (the second level of dynamic control) is executing the real-time visual servoing of S153, performing high-precision path correction within these macroscopic boundaries. The two work together: if S151 detects abnormal vibration again, S152 may intervene again, issuing a command to further reduce torque. The higher level command will immediately update the upper limits of the parameters of the lower level control, while the real-time correction of the lower level will continue to work seamlessly, ensuring accurate operation even under the new, more conservative parameters.

[0101] Please see Figure 7 , Figure 7 This is a schematic diagram of the structural composition of the battery pack disassembly robot based on image recognition in an embodiment of the present invention; the battery pack disassembly robot based on image recognition includes: The first image recognition module 21 is used to perform visual inspection of the battery pack based on multiple cameras built into the battery pack disassembly robot during the disassembly process of the battery pack disassembly robot, so as to collect multiple battery pack images in different dimensions, and determine the corresponding abnormal areas based on the image recognition of each battery pack image. The second image recognition module 22 is used to determine the corresponding abnormal features in multiple abnormal areas based on the image recognition of each abnormal area, and to construct a three-dimensional maintenance model of the battery pack based on multiple abnormal features, the overall shape of the battery pack and the surrounding environment of the battery pack. The intelligent disassembly system module 23 is used to identify multiple abnormal events based on the recognition of the three-dimensional maintenance model, determine the disassembly path of the disassembly robot relative to the battery pack based on each abnormal event, the corresponding abnormal area and the disassembly robot arm of the battery pack disassembly robot, and determine the intelligent disassembly system of the disassembly robot arm based on the disassembly path, the position of the abnormal area and the attitude of the battery pack. The intelligent disassembly behavior module 24 is used to determine multiple disassembly nodes based on the identification of the disassembly path, and to determine the corresponding intelligent disassembly behavior based on the multiple disassembly nodes, the intelligent disassembly system and the disassembly robotic arm; The dynamic control system module 25 is used to determine the intelligent disassembly level of each intelligent disassembly behavior based on the attitude of the battery pack and the corresponding disassembly vibration data; and to determine the dynamic control system of the battery pack disassembly robot based on the intelligent disassembly level of each intelligent disassembly behavior, the corresponding behavior trajectory and the real-time disassembly video of the battery pack.

[0102] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A disassembly method for a battery pack disassembly robot based on image recognition, characterized in that, include: During the disassembly of the battery pack by the battery pack disassembly robot, the robot uses multiple built-in cameras to perform visual inspection of the battery pack, so as to collect multiple battery pack images in different dimensions, and determine the corresponding abnormal areas based on the image recognition of each battery pack image. In multiple abnormal regions, the corresponding abnormal features are determined based on the image recognition of each abnormal region. Based on multiple abnormal features, the overall shape of the battery pack, and the surrounding environment of the battery pack, a three-dimensional maintenance model of the battery pack is constructed. The three-dimensional maintenance model is a multi-layered, semantically rich scene graph, which includes a geometric layer, a semantic layer, a state layer, and a relational layer. Based on the identification of the three-dimensional maintenance model, multiple abnormal events are identified. Based on each abnormal event, the corresponding abnormal area, and the disassembly robot arm of the battery pack disassembly robot, the disassembly path of the disassembly robot arm relative to the battery pack is determined. Based on the disassembly path, the position of the abnormal area, and the attitude of the battery pack, the intelligent disassembly system of the disassembly robot arm is determined. Multiple disassembly nodes are identified based on the disassembly path, and corresponding intelligent disassembly behaviors are determined based on these multiple disassembly nodes, the intelligent disassembly system, and the disassembly robotic arm. The level of intelligent disassembly for each intelligent disassembly behavior is determined based on the individual intelligent disassembly behaviors, the battery pack's posture, and the corresponding disassembly vibration data. The dynamic control system of the battery pack disassembly robot is determined based on the intelligent disassembly level of each intelligent disassembly behavior, the corresponding behavior trajectory, and the real-time disassembly video of the battery pack. The intelligent disassembly level includes excellent, normal, warning, or dangerous. The dynamic control system includes high-level supervision control and low-level real-time control. The high-level supervision control corresponds to the first level of dynamic control. The high-level supervision control and the low-level real-time control are coordinated through an arbitrator or state machine.

2. The disassembly method of the battery pack disassembly robot based on image recognition according to claim 1, characterized in that, During the disassembly of the battery pack by the battery pack disassembly robot, visual inspection of the battery pack is performed based on multiple cameras built into the robot to acquire multiple battery pack images from different dimensions. Based on image recognition of each battery pack image, corresponding abnormal areas are determined, including: When the battery pack is in the disassembly station, the battery pack disassembly robot disassembles the battery pack. Based on the current posture of the battery pack and the battery pack and the disassembly robot, the corresponding disassembly space system is determined. In this disassembly space system, the multiple cameras built into the battery pack disassembly robot are triggered to perform visual inspection of the battery pack. At this time, each side wall in the battery pack is visually inspected, and multiple battery pack images in different dimensions are collected. In multiple battery pack images, the corresponding component features are determined based on image recognition of each battery pack image, and the corresponding abnormal regions are determined based on the feature location, corresponding feature shape and corresponding anomaly coefficient of each component feature.

3. The disassembly method of the battery pack disassembly robot based on image recognition according to claim 1, characterized in that, In multiple abnormal regions, corresponding abnormal features are determined based on image recognition of each abnormal region. A three-dimensional maintenance model of the battery pack is constructed based on these abnormal features, the overall shape of the battery pack, and its surrounding environment. This includes: Real-time monitoring of multiple abnormal areas, image recognition of each abnormal area, determination of corresponding abnormal features based on image recognition of each abnormal area, and marking of the feature location of the abnormal feature. The overall shape of the battery pack is collected, and the surrounding environment of the battery pack is determined based on the detection of the surrounding area. The first maintenance model is determined based on the feature location of each abnormal feature and the overall shape of the battery pack. The second maintenance model is determined based on the feature location of each abnormal feature and the surrounding environment of the battery pack. A three-dimensional maintenance model of the battery pack is constructed based on the synthesis of the first and second maintenance models.

4. The disassembly method of the battery pack disassembly robot based on image recognition according to claim 1, characterized in that, The method identifies multiple abnormal events based on a three-dimensional maintenance model. Based on each abnormal event, the corresponding abnormal region, and the disassembly robot's arm, a disassembly path is determined relative to the battery pack. Based on this disassembly path, the location of the abnormal region, and the battery pack's orientation, an intelligent disassembly system for the disassembly robot is determined, including: The three-dimensional maintenance model is monitored in real time. Based on the matching between the three-dimensional maintenance model and the battery pack disassembly robot, the corresponding recognition mode is determined, and the recognition of the three-dimensional maintenance model is triggered along the recognition mode to identify multiple abnormal events. In the event of multiple abnormal events, the corresponding abnormal area is identified based on the identification of each abnormal event, and the disassembly robot arm of the battery pack disassembly robot is marked. Based on each abnormal event, the corresponding abnormal area, and the disassembly robot arm of the battery pack disassembly robot, the disassembly path of the disassembly robot arm relative to the battery pack is determined.

5. The disassembly method of the battery pack disassembly robot based on image recognition according to claim 4, characterized in that, The method of identifying multiple abnormal events based on a three-dimensional maintenance model, determining the disassembly path of the disassembly robot relative to the battery pack based on each abnormal event, the corresponding abnormal area, and the disassembly robot's disassembly arm, and determining the intelligent disassembly system of the disassembly robot based on the disassembly path, the position of the abnormal area, and the attitude of the battery pack, further includes: The battery pack's orientation is collected, and the first intelligent disassembly parameters of the disassembly robot arm are determined based on the disassembly path and the battery pack's orientation. The second intelligent disassembly parameters of the disassembly robot arm are determined based on the location of the abnormal area and the battery pack's orientation. The intelligent disassembly system of the disassembly robot arm is determined according to the mapping relationship between the first intelligent disassembly parameters, the second intelligent disassembly parameters, and the intelligent disassembly system.

6. The disassembly method of the battery pack disassembly robot based on image recognition according to claim 1, characterized in that, The process of identifying multiple disassembly nodes based on the disassembly path, and determining corresponding intelligent disassembly behaviors based on these multiple disassembly nodes, the intelligent disassembly system, and the disassembly robotic arm, includes: The disassembly path is collected, and multiple disassembly areas are determined based on the distribution location of the disassembly path, the corresponding path shape, and the three-dimensional maintenance model of the battery pack. The corresponding disassembly nodes are determined according to the regional location of each disassembly area, the corresponding disassembly components, and the disassembly robotic arm, so as to collect multiple disassembly nodes.

7. The disassembly method of the battery pack disassembly robot based on image recognition according to claim 6, characterized in that, The process of determining multiple disassembly nodes based on the identification of the disassembly path, and determining the corresponding intelligent disassembly behavior based on the multiple disassembly nodes, the intelligent disassembly system, and the disassembly robotic arm, further includes: The first disassembly behavior is determined based on multiple disassembly nodes and the intelligent disassembly system. The second disassembly behavior is determined based on multiple disassembly nodes and the disassembly robotic arm. The corresponding intelligent disassembly behavior is determined based on the control system of the first disassembly behavior, the second disassembly behavior, and the disassembly robotic arm.

8. The disassembly method of the battery pack disassembly robot based on image recognition according to claim 1, characterized in that, The intelligent disassembly level of the intelligent disassembly behavior is determined based on each intelligent disassembly behavior, the battery pack's posture, and the corresponding disassembly vibration data. A dynamic control system for the battery pack disassembly robot is determined based on the intelligence level of each intelligent disassembly behavior, the corresponding behavioral trajectory, and real-time disassembly video of the battery pack. This system includes: Real-time monitoring of each intelligent disassembly behavior, and determination of the disassembly vibration data of the battery pack based on dynamic monitoring of the battery pack. A first intelligent disassembly coefficient is determined based on each intelligent disassembly behavior and the corresponding disassembly vibration data. A second intelligent disassembly coefficient is determined based on each intelligent disassembly behavior and the attitude of the battery pack. The intelligent disassembly level of the intelligent disassembly behavior is determined based on the mapping relationship between the first intelligent disassembly coefficient, the second intelligent disassembly coefficient and the intelligent disassembly level.

9. The disassembly method of the battery pack disassembly robot based on image recognition according to claim 8, characterized in that, The intelligent disassembly level of the intelligent disassembly behavior is determined based on each intelligent disassembly behavior, the battery pack's posture, and the corresponding disassembly vibration data. The dynamic control system for the battery pack disassembly robot is determined based on the intelligent disassembly level of each intelligent disassembly behavior, the corresponding behavior trajectory, and real-time disassembly video of the battery pack. This system also includes: The behavior trajectory of each intelligent disassembly behavior is marked, and real-time disassembly video of the battery pack is collected. Based on the intelligent disassembly level of each intelligent disassembly behavior and the real-time disassembly video of the battery pack, the first dynamic control content of the battery pack disassembly robot is determined. The second level of dynamic control for the battery pack disassembly robot is determined based on the behavioral trajectories of each intelligent disassembly action and the real-time disassembly video of the battery pack. The dynamic control system for the battery pack disassembly robot is then determined based on the first and second levels of dynamic control.

10. A disassembly system for a battery pack disassembly robot based on image recognition, characterized in that, The image recognition-based battery pack disassembly robot disassembly system is applied to the disassembly method of the image recognition-based battery pack disassembly robot as described in any one of claims 1-9, wherein the image recognition-based battery pack disassembly robot disassembly system comprises: The first image recognition module is used to perform visual inspection of the battery pack based on multiple cameras built into the battery pack disassembly robot during the disassembly process of the battery pack disassembly robot, so as to collect multiple battery pack images in different dimensions, and determine the corresponding abnormal areas based on the image recognition of each battery pack image. The second image recognition module is used to determine the corresponding abnormal features in multiple abnormal areas based on the image recognition of each abnormal area, and to construct a three-dimensional maintenance model of the battery pack based on multiple abnormal features, the overall shape of the battery pack and the surrounding environment of the battery pack. The intelligent disassembly system module is used to identify multiple abnormal events based on the recognition of the three-dimensional maintenance model. Based on each abnormal event, the corresponding abnormal area, and the disassembly robot arm of the battery pack disassembly robot, the disassembly path of the disassembly robot arm relative to the battery pack is determined. Based on the disassembly path, the position of the abnormal area, and the attitude of the battery pack, the intelligent disassembly system of the disassembly robot arm is determined. The intelligent disassembly behavior module is used to determine multiple disassembly nodes based on the identification of the disassembly path, and to determine the corresponding intelligent disassembly behavior based on the multiple disassembly nodes, the intelligent disassembly system and the disassembly robotic arm; The dynamic control system module is used to determine the intelligent disassembly level of each intelligent disassembly behavior based on the attitude of the battery pack and the corresponding disassembly vibration data; and to determine the dynamic control system of the battery pack disassembly robot based on the intelligent disassembly level of each intelligent disassembly behavior, the corresponding behavior trajectory and the real-time disassembly video of the battery pack.

Images