Lithium battery recycling and disassembling method based on AI+3D vision

Abstract

The invention discloses a lithium battery recycling and disassembling method based on AI + 3D vision, and relates to the technical field of battery recycling, and the method comprises the steps of S1, multispectral polarization image acquisition, S2, polarization parameter dynamic calibration, S3, image fusion, S4, deep learning denoising optimization, S5, 3D point cloud reconstruction and feature extraction, and S6, robot motion planning and execution. According to the method, a dynamic situation field model is combined with an impedance control principle, so that the robot can sense obstacle distribution in real time in a dynamic environment and autonomously adjust a motion strategy, and meanwhile, the contact force between an end effector and a target object is accurately controlled; a rejection and attraction field is modeled in real time, so that the sudden mechanical interference risk in the disassembling process is effectively avoided; the impedance coupling mechanism ensures that the robot maintains stable interaction force in contact operation through force feedback closed-loop adjustment, and battery damage or tool failure caused by sudden change of external force is avoided.

Description

Technical Field

[0001] The present invention belongs to the field of battery recycling technology, and in particular relates to a lithium battery recycling and disassembly method based on AI+3D vision. Background Art

[0002] Large quantities of discarded batteries waste precious resources and energy, while also polluting the environment. Used lithium-ion batteries are rich in recyclable resources, including precious metals like aluminum and copper, as well as graphite and carbon-based materials. Improper disposal of these used lithium batteries not only results in a massive waste of resources but also causes serious environmental pollution.

[0003] In the field of lithium battery recycling, robotic automation requires high-precision contact force control in complex dynamic environments. Traditional path planning methods are mostly based on static environment assumptions or single-modal feedback (such as pure vision or pure force perception), which makes it difficult to effectively coordinate environmental obstacle avoidance and contact force regulation. Especially when dealing with highly reflective and multi-obstacle battery packs, existing technologies often cause excessive fluctuations in the contact force between the end effector and the target object due to force control lag or path rigidity, which in turn causes damage to the battery structure (such as leakage due to excessively deep milling of the terminal) or abnormal collisions between the tool and the workpiece, seriously restricting the safety and reliability of the disassembly process. To address the above problems, the following solutions are proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide a lithium battery recycling and disassembly method based on AI+3D vision. By combining the dynamic potential field model with the impedance control principle, the robot can perceive the distribution of obstacles in a dynamic environment in real time and autonomously adjust its motion strategy, while accurately controlling the contact force between the end effector and the target object, solving the problem that existing visual systems are difficult to effectively coordinate environmental obstacle avoidance and contact force regulation.

[0005] To solve the above technical problems, the present invention is achieved through the following technical solutions:

[0006] The present invention provides a lithium battery recycling and disassembly method based on AI+3D vision, comprising:

[0007] Step S1, multispectral polarization image acquisition: Use a multispectral polarization camera to capture images of highly reflective areas from four orthogonal angles, and simultaneously record ambient light and spatial posture data;

[0008] Step S2, dynamic calibration of polarization parameters: calculating the degree of polarization based on the Fresnel reflection model and fitting the surface normal vector mapping diagram;

[0009] Step S3, image fusion: dynamically calculate the fusion weights through the weight distribution network, and generate the anti-reflection image by weighting;

[0010] Step S4, deep learning denoising optimization: input the fused image into the deep denoising network N θ , adopts the residual dense block structure and adversarial loss function to output high signal-to-noise ratio images, further removes image noise and artifacts through the adversarial generative network, enhances detail performance, and outputs clear images with high resolution and low noise, ensuring the accuracy and reliability of 3D point cloud reconstruction;

[0011] Step S5, 3D point cloud reconstruction and feature extraction: Combine structured light binocular matching to generate submillimeter point clouds and extract key geometric features;

[0012] Step S6, robot motion planning and execution: planning the disassembly path based on the feature data, adjusting the contact force in real time in combination with impedance control, and realizing lithium battery disassembly path planning;

[0013] Step S7, update and optimization: Update model parameters through online learning to achieve adaptive evolution of the anti-reflection algorithm. Continuously optimize the anti-reflection model through online learning to adapt to different battery models and working conditions, improve the long-term stability and generalization capability of the system, and achieve intelligent upgrades.

[0014] Furthermore, the step S1, multispectral polarization image acquisition, specifically includes the following steps:

[0015] Step S11: Using a multispectral polarization camera, fix the position around the lithium battery pack and collect original images {I1, I2, I3, I4} of the high-reflective area in four orthogonal polarization directions;

[0016] Step S12: synchronously record the ambient light intensity L and the angle θ between the camera and the target plane, and store them as metadata;

[0017] This design uses a multispectral polarization camera to capture images of highly reflective areas of lithium batteries from different angles, obtaining raw data and providing multi-angle information for subsequent anti-reflection processing; it also records ambient light and camera position to ensure data consistency.

[0018] Furthermore, the step S2, dynamic calibration of polarization parameters, specifically includes the following steps:

[0019] Step S21: Calculate the polarization degree ρ of each pixel according to the Fresnel reflection model k , the formula is:

[0020]

[0021] Where, I max is the maximum intensity value of the same pixel at all polarization angles, I min is the minimum intensity value of the same pixel at all polarization angles, and ε is a minimum constant;

[0022] Step S22: Fitting the polarization angle offset φ of each region by the least squares method k , establish a normal vector map of the reflective surface;

[0023] This design calculates the polarization degree and normal vector based on the Fresnel reflection model, quantifies the reflective characteristics, and provides precise physical parameters for subsequent image fusion, ensuring the scientific nature and adaptability of the anti-reflection algorithm.

[0024] Furthermore, the image fusion step S3 specifically includes the following steps:

[0025] Step S31: Input the original image of step S1 into the pre-trained weight distribution network and output the fusion weight α of each polarization angle k ;

[0026] Step S32: Perform weighted fusion to obtain a preliminary anti-reflection image I fusion , the formula is:

[0027]

[0028] Where, α k is the multimodal fusion weight, I k is the original image of the kth polarization angle, ρ k is the polarization parameter, φ k is the polarization angle offset, λ k is the wavelength weight coefficient;

[0029] This design uses deep learning to dynamically assign weights to images with different polarizations, fuses multi-angle information, preliminarily eliminates reflective interference, improves image quality, and lays the foundation for subsequent high-precision 3D reconstruction.

[0030] Furthermore, in step S4, the adversarial loss function in deep learning denoising optimization is defined as:

[0031] L=λ adv L adv +λ per L per +λ rec ||I out -I gt ||1;

[0032] Where L is the total loss function, L adv To combat the loss, L per is the perceptual loss, λ adv ,λ per ,λ rec is the weight coefficient of each loss item, I out Images optimized for the web, I gt is the calibrated reflection-free ground truth image.

[0033] Furthermore, the step S5, 3D point cloud reconstruction and feature extraction specifically includes the following steps:

[0034] Step S51: Using the structured light projector and the I output from step S4 out Perform binocular matching to generate 3D point clouds with sub-millimeter accuracy;

[0035] Step S52: Extract key features from the point cloud, specifically:

[0036] The coordinates of the screw hole center (x, y, z), the curvature of the busbar edge κ, and the end plate plane equation ax+by+cz+d=0;

[0037] This design generates high-precision 3D point clouds based on optimized images, extracts key geometric features (such as screw position, edge curvature, etc.), and provides the robot with accurate disassembly guidance data.

[0038] Furthermore, the step S6, robot motion planning and execution specifically includes the following steps:

[0039] Step S61: Construct a dynamic potential field map based on the 3D point cloud features extracted in step 5, specifically:

[0040] Set the center of the screw hole (x, y, z) as the repulsive potential field source and the potential field radius;

[0041] The busbar curvature mutation point κ is greater than 0.3mm -1 Marked as a high-risk area, the potential field strength is increased by 50%;

[0042] The end plate plane equation ax+by+cz+d=0 is used as a global constraint to limit the robot motion within ±3mm in the direction of the normal vector.

[0043] Step S62: Define the potential field formula:

[0044]

[0045] Where q is the joint angle vector of the six-axis robot, is the coordinate of the nearest point of the ith obstacle in the robot base coordinate system, w i is the potential field weight of the i-th obstacle, β is the potential field attenuation coefficient, F ext is the real-time contact force measured by the force sensor at the end of the robot, F desired is the desired contact force, λ is the impedance term weight, η is the target attraction term weight, q end is the end joint angle of the current path segment, q goal Decompose the joint angles of the target pose;

[0046] The potential field formula P(q) is input into the optimization solver of the ROS-MoveIt! framework, and the sequential quadratic programming algorithm is used to iteratively solve the optimal joint angle path q(t):

[0047]

[0048] Where T is the total time of path execution, is the joint angular velocity vector, τ j is the driving torque of the jth joint, τ max is the upper limit of joint torque;

[0049] The path is discretized into 1000 waypoints with a time resolution of Δt = 10 ms;

[0050] Step S63: During the path execution phase, the impedance term weight λ is adjusted through real-time feedback, specifically:

[0051] When |F ext -F desired When |>0.5N, adaptive adjustment is activated, and the formula is:

[0052] λ new =λ base ×(1+γ×tanh(|F ext -F desired |));

[0053] Where λ new is the adjusted impedance weight, λ base is the basic impedance weight, γ is the gain coefficient;

[0054] When a sudden collision is detected, the potential field expansion is triggered immediately: the potential field radius r is instantly expanded to 20mm, and the path is replanned;

[0055] Step S64: The outer loop receives the screw hole pose (x′, y′, z′) updated by 3D vision every 50ms, fuses historical data through the Kalman filter, and corrects the path deviation:

[0056] q corrected =q planned +K×(q vision -q planned );

[0057] Where q corrected is the corrected joint angle vector, q planned is the theoretical joint angle in the planned path, q vision is the target pose joint angle updated based on 3D vision, K is the Kalman filter gain matrix;

[0058] The inner loop adopts the admittance control model, according to Fext Dynamically adjust the end pose, the formula is:

[0059] Δx=M -1 ×(F ext -F desired )×Δt 2 ;

[0060] Where Δx is the posture adjustment of the end effector, M is the virtual mass matrix, and Δt is the control period;

[0061] Step S65: If convergence is still not achieved after three consecutive path corrections, the operation is suspended and the self-diagnosis module is activated: the target area is rescanned using 3D vision to update the potential field map; the obstacle avoidance path is generated using the RRT*-Connect algorithm, with a local replanning time of < 200ms;

[0062] Synchronously detect the height of the pole after milling. If the height is lower than the threshold, an audible and visual alarm will be triggered and the current action will be terminated;

[0063] This design effectively avoids the risk of sudden mechanical interference during the disassembly process by real-time modeling of repulsion and attraction fields, such as obstacle displacement caused by battery module misalignment or loose screws; through force feedback closed-loop adjustment, it ensures that the robot maintains a stable interaction force during contact operations, avoiding battery damage or tool failure due to sudden changes in external forces.

[0064] The present invention has the following beneficial effects:

[0065] 1. The present invention improves the autonomy and safety of robots in performing complex disassembly tasks by deeply integrating environmental perception and force interaction control. By combining the dynamic potential field model with the impedance control principle, the robot can perceive the distribution of obstacles in a dynamic environment in real time and autonomously adjust its motion strategy, while accurately controlling the contact force between the end effector and the target object. By real-time modeling of repulsion and attraction fields, the risk of sudden mechanical interference during the disassembly process, such as obstacle displacement caused by battery module misalignment or loose screws, is effectively avoided. The impedance coupling mechanism ensures that the robot maintains a stable interaction force during contact operations through force feedback closed-loop regulation, avoiding battery damage or tool failure caused by sudden changes in external forces. In addition, the algorithm realizes real-time correction of path deviations through multimodal feedback fusion (vision and force perception), enhancing the system's robustness to interference factors such as lighting changes and mechanical vibrations. This design not only optimizes the robot's motion efficiency, but also reduces the collision risks and operational errors common in traditional rigid control strategies.

[0066] 2. The present invention effectively overcomes the imaging interference problem of highly reflective areas on the surface of lithium batteries by integrating multimodal polarized light imaging technology with a deep learning denoising algorithm. The dynamic calibration of polarization angles and the multispectral fusion mechanism can accurately analyze the detailed information of areas with different reflective characteristics, suppressing the negative impact of specular reflection and stray light on image quality. The deep denoising network further eliminates noise and enhances feature resolution, so that subsequent 3D point cloud reconstruction can more accurately reflect the structural characteristics of the battery pack. This design provides reliable visual input for robot disassembly, ensuring that the geometric and texture information of key components can still be stably acquired under complex working conditions.

[0067] 3. The present invention's AI-based dynamic weight allocation and flexible parameter adaptation technology enables the system to adapt to various brands and models of lithium battery packs without relying on preset templates or manual intervention. Through online learning and multimodal data fusion, the algorithm can autonomously identify structural differences in battery packs and adjust visual positioning and motion planning strategies in real time. This design reduces the need for production line modifications and lowers equipment switching costs, while supporting rapid response to the disassembly needs of new battery products and improving the flexibility and scalability of the overall production line.

[0068] 4. Through sub-millimeter precision 3D point cloud reconstruction and robotic impedance control technology, the present invention can accurately identify the spatial position of key components such as screws and busbars, and plan collision-free motion trajectories; high-precision milling and dynamic contact force control mechanisms can avoid the risk of battery damage or electrolyte leakage due to operational deviations; in addition, real-time monitoring and alarm functions can promptly identify abnormal operating conditions (such as insufficient pole height), actively suspend operations and issue warnings, and effectively prevent safety hazards such as short circuits and leakage.

[0069] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0071] Figure 1 This is a flow chart of a lithium battery recycling and disassembly method based on AI+3D vision in the present invention. DETAILED DESCRIPTION

[0072] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0073] See also Figure 1 As shown, the present invention is a lithium battery recycling and disassembly method based on AI+3D vision, comprising:

[0074] Step S1: Multispectral polarization image acquisition:

[0075] Step S11: Use a multispectral polarization camera (wavelength range 400-1000nm) to fix the position around the lithium battery pack and collect original images {I1, I2, I3, I4} of the highly reflective areas (screw holes, busbar, end plate) in four orthogonal polarization directions (0°, 45°, 90°, 135°);

[0076] Step S12: synchronously record the ambient light intensity L and the angle θ between the camera and the target plane (obtained by a laser rangefinder), and store them as metadata.

[0077] Step S2: Dynamic calibration of polarization parameters:

[0078] Step S21: Calculate the polarization degree ρ of each pixel according to the Fresnel reflection model k , the formula is:

[0079]

[0080] Where, I max is the maximum intensity value of the same pixel at all polarization angles, I min is the minimum intensity value of the same pixel at all polarization angles, and ε is a minimum constant;

[0081] Step S22: Fitting the polarization angle offset φ of each region by the least squares method k , establish a normal vector map of the reflective surface.

[0082] Step S3: Image fusion:

[0083] Step S31: Input the original image of step S1 into the pre-trained weight distribution network (U-Net structure) and output the fusion weight α of each polarization angle k ;

[0084] Step S32: Perform weighted fusion to obtain a preliminary anti-reflection image I fusion , the formula is:

[0085]

[0086] Where, α k is the multimodal fusion weight, I k is the original image of the kth polarization angle, ρ k is the polarization parameter, φ k is the polarization angle offset, λ k is the wavelength weight coefficient.

[0087] Step S4: Deep learning denoising optimization:

[0088] Will I fusion Input deep denoising network N θ , the network adopts a residual dense block structure and adversarial loss function to output a high signal-to-noise ratio image I out , the resolution is increased to 2048×2048, and the noise PSNR is >40dB;

[0089] The adversarial loss function is defined as:

[0090] L=λ adv L adv +λ per L per +λ rec ||I out -I gt ||1;

[0091] Where L is the total loss function, L adv To combat the loss, L per is the perceptual loss, λ adv ,λ per ,λ rec is the weight coefficient of each loss item, I out Images optimized for the web, I gt is the calibrated reflection-free ground truth image.

[0092] Step S5: 3D point cloud reconstruction and feature extraction:

[0093] Step S51: Using the structured light projector and the I output from step S4 out Perform binocular matching to generate 3D point clouds with sub-millimeter accuracy;

[0094] Step S52: Extract key features from the point cloud, specifically:

[0095] The coordinates of the screw hole center (x, y, z), the busbar edge curvature κ, and the end plate plane equation ax+by+cz+d=0.

[0096] Step S6: Robot motion planning and execution:

[0097] Step S61: Input the 3D point cloud features extracted in step 5 (the screw hole center coordinates (x, y, z), the busbar edge curvature κ, and the end plate plane equation ax+by+cz+d=0) into the ROS motion planner;

[0098] Construct a dynamic potential field map based on characteristic data:

[0099] Repulsive potential field source: The center of each screw hole is set as the repulsive center, the potential field radius r = 10mm, the potential field strength weight , where β = 0.5 mm -1 ;

[0100] Dangerous area marking: for busbar curvature mutation point (κ>0.5mm -1 ) Apply an enhanced potential field, and the weight is increased by 50%;

[0101] Global constraint: Using the end plate plane equation as the motion boundary, the robot end is restricted to move within ±3mm in the normal vector direction;

[0102] Step S62: Define the dynamic potential field-impedance coupling objective function:

[0103]

[0104] Parameter settings: λ = 0.8, η = 1.2, determined by offline calibration experiments;

[0105] Use the SQP optimizer of the ROS-MoveIt! framework to solve the joint velocity With the torque constraint (||τ j ||≤τ max ) is discretized into 1000 waypoints with a time resolution of Δt = 10ms;

[0106] Step S63: Dynamic adjustment of contact force: The end effector (electric screwdriver, vacuum suction cup) is equipped with a six-dimensional force sensor to collect the contact force F in real time. ext ;

[0107] When |F ext When -5N|>0.5N, the adaptive impedance weight is activated:

[0108] λ new =λ base ×(1+2.0×tanh(|F ext -5|));

[0109] Ensure that the contact force is stable at 5N±0.2N;

[0110] Sudden collision response: If F is detected ext>10N, the repulsive potential field radius is instantly expanded to 20mm, and local path replanning is triggered;

[0111] Step S64: Visual feedback outer loop: receiving the screw hole pose (x′, y′, z′) updated by 3D vision every 50ms;

[0112] The Kalman filter is used to fuse historical path data and calculate the posture deviation Δq=q vision -q planned ;

[0113] Correction path:

[0114] q corrected =q planned +K×(q vision -q planned );

[0115] The filter gain matrix K = diag(0.7, 0.7, 0.7, 0.3, 0.3, 0.3, 0.3) ensures Lyapunov stability.

[0116] Force feedback inner loop: Adopting admittance control model, according to F ext Dynamically adjust the end pose:

[0117] Δx=M -1 ×(F ext -5)×(0.01) 2 ;

[0118] Virtual mass matrix M = diag(0.1, 0.1, 0.1, 0.05, 0.05, 0.05) kg;

[0119] Step S65: Path convergence detection: If the path does not converge after three consecutive corrections (||q corrected -q goal ||>1mm), suspend operation and start self-diagnosis;

[0120] Local replanning: rescan the target area through 3D vision and update the obstacle positions in the potential field map;

[0121] The RRT*-Connect algorithm is used to generate obstacle avoidance paths, with planning time limited to 200ms.

[0122] Pole safety monitoring: Synchronously detect the pole height after milling. If the height is lower than 2.0mm, an audible and visual alarm will be triggered and the current action will be terminated.

[0123] Step S7, update and optimization: collect mechanical and optical data during the disassembly process, and update the AI anti-reflection model parameter θ through the online learning module (LSTM network) to achieve algorithm self-evolution.

[0124] A specific application of this embodiment is:

[0125] Background of the embodiment:

[0126] A lithium battery recycling plant automated the disassembly of a retired new energy vehicle lithium battery pack (brand: Company A, model: NCM-120). The pack contained six modules, each consisting of 12 cells connected by busbars. The modules had highly reflective screw holes and end plates on the surface.

[0127] Detailed implementation steps

[0128] 1. Multimodal polarization image acquisition:

[0129] Equipment configuration: A multispectral polarization camera (wavelength range 400-1000nm, resolution 4096×2160) is used, fixed 50cm above the robot end effector and at a 45° angle to the battery pack plane.

[0130] Data acquisition: The camera captures images of the battery pack surface at four polarization directions: 0°, 45°, 90°, and 135° (exposure time 10ms, aperture f / 5.6), obtaining the original image set {I1, I2, I3, I4}.

[0131] The ambient light intensity (800 lux measured by the light sensor) and the angle between the camera and the battery pack plane (45°±0.5° measured by the laser rangefinder) are simultaneously recorded and stored as metadata.

[0132] 2. Dynamic calibration of polarization parameters:

[0133] Polarization degree calculation: Select the screw hole area (ROI: 200×200 pixels) and extract the maximum intensity value I in each polarization direction max With the minimum intensity value I min , according to the formula Calculate the local polarization degree ρ k ;

[0134] Calibration results: Average ρ in the screw hole area k =0.78, busbar area ρ k =0.65;

[0135] Normal vector fitting: Based on the least squares method, normal vector fitting is performed on the reflective area in the 45° polarization image to generate a surface normal vector map for subsequent 3D reconstruction.

[0136] 3. Image fusion:

[0137] Weight distribution: The original image is input into the pre-trained U-Net weight distribution network (input size 1024×1024, output channel number 4), and the fusion weights α1=0.32, α2=0.28, α3=0.25, α4=0.15 are dynamically generated;

[0138] Image synthesis: Four polarization images are weighted and fused according to the formula to obtain the anti-reflection fusion image I fusion , the resolution is increased to 2048×2048.

[0139] 4. Deep learning denoising optimization:

[0140] Network Processing: I fusion Input deep denoising network N θ , (structure: 5 residual dense blocks, number of channels 64), output high signal-to-noise ratio image I out (PSNR=42.3dB);

[0141] Processing time: 85ms for a single frame (NVIDIA Jetson AGX Xavier platform);

[0142] 5.3D point cloud reconstruction and feature extraction:

[0143] Point cloud generation: using a structured light projector (wavelength 532nm, fringe density 50 lines / mm) and denoised image I out Perform binocular matching to generate sub-millimeter point clouds (accuracy 0.05mm);

[0144] Feature extraction: Extract the coordinates of the screw hole center (x = 125.3 mm, y = 67.8 mm, z = 10.2 mm) and the busbar edge curvature κ = 0.12 mm -1 , the end plate plane equation is 0.87x-0.15y+0.48z-32.6=0;

[0145] 6. Robot motion planning and execution:

[0146] Path planning: Feature data was fed into the ROS motion planner (MoveIt! framework) to generate a disassembly path for a six-axis collaborative robot (model: UR10e), with a planning time of 1.2 seconds.

[0147] Screw removal: The electric screwdriver (torque 3.5 N·m) was aligned with the screw hole, and the contact force was adjusted to 5.1 N based on impedance control to remove six sets of screws (each screw took 4 seconds).

[0148] Module grabbing: The module is grabbed by a vacuum suction cup (negative pressure -80kPa) and moved to the disassembly table. The pole height is detected in real time (current value 2.3mm, safety threshold 2.0mm), and no alarm is triggered.

[0149] Specifically:

[0150] Step S61: The 3D point cloud features extracted in step 5 (the coordinates of the center of the screw hole (x = 125.3 mm, y = 67.8 mm, z = 10.2 mm), the busbar edge curvature κ = 0.12 mm) are converted into -1 The end plate plane equation (0.87x-0.15y+0.48z-32.6=0) is input into the ROS motion planner;

[0151] Construct a dynamic potential field map based on characteristic data:

[0152] Repulsive potential field source: The center of each screw hole is set as the repulsive center, the potential field radius r = 10mm, and the potential field strength weight w i =exp(-β×||qq obsi ||), where β = 0.5 mm -1 ;

[0153] Dangerous area marking: for busbar curvature mutation point (κ>0.5mm -1 ) Apply an enhanced potential field, and the weight is increased by 50%;

[0154] Global constraint: Using the end plate plane equation as the motion boundary, the robot end is restricted to move within ±3mm in the normal vector direction;

[0155] Step S62: Define the dynamic potential field-impedance coupling objective function:

[0156]

[0157] Parameter settings: λ = 0.8, η = 1.2, determined by offline calibration experiments;

[0158] Use the SQP optimizer of the ROS-MoveIt! framework to solve the joint velocity With the torque constraint (||τ j ||≤τ max ) is discretized into 1000 waypoints with a time resolution of Δt = 10ms;

[0159] Step S63: Dynamic adjustment of contact force: The end effector (electric screwdriver, vacuum suction cup) is equipped with a six-dimensional force sensor to collect the contact force F in real time. ext ;

[0160] When |F ext When -5N|>0.5N, the adaptive impedance weight is activated:

[0161] λ new =λ base×(1+2.0×tanh(|F ext -5|));

[0162] Ensure that the contact force is stable at 5N±0.2N;

[0163] Sudden collision response: If F is detected ext >10N, the repulsive potential field radius is instantly expanded to 20mm, and local path replanning is triggered;

[0164] Step S64: Visual feedback outer loop: receiving the screw hole pose (x′, y′, z′) updated by 3D vision every 50ms;

[0165] The Kalman filter is used to fuse historical path data and calculate the posture deviation Δq=q vision -q planned ;

[0166] Correction path:

[0167] q corrected =q planned +K×(q vision -q planned );

[0168] The filter gain matrix K = diag(0.7, 0.7, 0.7, 0.3, 0.3, 0.3, 0.3) ensures Lyapunov stability.

[0169] Force feedback inner loop: Adopting admittance control model, according to F ext Dynamically adjust the end pose:

[0170] Δx=M -1 ×(F ext -5)×(0.01) 2 ;

[0171] Virtual mass matrix M = diag(0.1, 0.1, 0.1, 0.05, 0.05, 0.05) kg;

[0172] Step S65: Path convergence detection: If the path does not converge after three consecutive corrections (||q corrected -q goal ||>1mm), suspend operation and start self-diagnosis;

[0173] Local replanning: rescan the target area through 3D vision and update the obstacle positions in the potential field map;

[0174] The RRT*-Connect algorithm is used to generate obstacle avoidance paths, with a planning time limit of 200ms. Pole safety monitoring: The height of the pole after milling is synchronously detected. If the height is less than 2.0mm, an audible and visual alarm is triggered and the current action is terminated.

[0175] 7. Update optimization:

[0176] Data transmission: Collect images and force data (sampling rate 100Hz) during the disassembly process and upload them to the cloud server;

[0177] Model update: The online learning module (LSTM network, 128 hidden nodes) analyzes the data and updates the denoising network parameters θ. After iteration, the PSNR is improved to 43.1dB.

[0178] Implementation effect:

[0179] Imaging quality: Screw hole reflection suppression rate reaches 92%, busbar feature extraction error is <0.08mm;

[0180] Disassembly efficiency: The time required to disassemble a single module is reduced from 15 minutes by manual operation to 5 minutes;

[0181] Safety: Collision detection response time is <50ms, no misoperation occurs when the pole height is lower than 2.0mm, zero false alarms are detected for pole height, and the risk of leakage is reduced to 0.01%.

[0182] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0183] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A lithium battery recycling and disassembly method based on AI+3D vision, characterized in that: The disassembly method comprises the following steps: Step S1, multispectral polarization image acquisition: Use a multispectral polarization camera to capture images of highly reflective areas from four orthogonal angles, and simultaneously record ambient light and spatial posture data; Step S2, dynamic calibration of polarization parameters: calculating the degree of polarization based on the Fresnel reflection model and fitting the surface normal vector mapping diagram; Step S3, image fusion: dynamically calculate the fusion weights through the weight distribution network, and generate the anti-reflection image by weighting; Step S4, deep learning denoising optimization: input the fused image into the deep denoising network N θ , using residual dense block structure and adversarial loss function to output high signal-to-noise ratio images; Step S5, 3D point cloud reconstruction and feature extraction: Combine structured light binocular matching to generate submillimeter point clouds and extract key geometric features; Step S6, robot motion planning and execution: planning the disassembly path based on the feature data, adjusting the contact force in real time in combination with impedance control, and realizing lithium battery disassembly path planning; Step S7, update optimization: update the model parameters through online learning to achieve adaptive evolution of the anti-reflection algorithm.

2. The lithium battery recycling and disassembly method based on AI+3D vision according to claim 1 is characterized in that: The step S1, multispectral polarization image acquisition, specifically includes the following steps: Step S11: Using a multispectral polarization camera, fix the position around the lithium battery pack and collect original images {I1, I2, I3, I4} of the high-reflective area in four orthogonal polarization directions; Step S12: synchronously record the ambient light intensity L and the angle θ between the camera and the target plane, and store them as metadata.

3. The lithium battery recycling and disassembly method based on AI+3D vision according to claim 1 is characterized in that: The step S2, dynamic calibration of polarization parameters, specifically includes the following steps: Step S21: Calculate the polarization degree ρ of each pixel according to the Fresnel reflection model k , the formula is: Where, I max is the maximum intensity value of the same pixel at all polarization angles, I min is the minimum intensity value of the same pixel at all polarization angles, and ε is a minimum constant; Step S22: Fitting the polarization angle offset φ of each region by the least squares method k , establish a normal vector map of the reflective surface.

4. The lithium battery recycling and disassembly method based on AI+3D vision according to claim 1 is characterized in that: The image fusion step S3 specifically includes the following steps: Step S31: Input the original image of step S1 into the pre-trained weight distribution network and output the fusion weight α of each polarization angle k ; Step S32: Perform weighted fusion to obtain a preliminary anti-reflection image I fusion , the formula is: Where, α k is the multimodal fusion weight, I k is the original image of the kth polarization angle, ρ k is the polarization parameter, φ k is the polarization angle offset, λ k is the wavelength weight coefficient.

5. The lithium battery recycling and disassembly method based on AI+3D vision according to claim 1 is characterized in that: In step S4, the adversarial loss function in deep learning denoising optimization is defined as: L=λ adv L adv +λ per L per +λ rec ||I out -I gt ||1; Where L is the total loss function, L adv To combat the loss, L per is the perceptual loss, λ adv ,λ per ,λ rec is the weight coefficient of each loss item, I out Images optimized for the web, I gt is the calibrated reflection-free ground truth image.

6. The lithium battery recycling and disassembly method based on AI+3D vision according to claim 1, characterized in that: The step S5, 3D point cloud reconstruction and feature extraction, specifically includes the following steps: Step S51: Using the structured light projector and the I output from step S4 out Perform binocular matching to generate 3D point clouds with sub-millimeter accuracy; Step S52: Extract key features from the point cloud, specifically: The coordinates of the screw hole center (x, y, z), the busbar edge curvature κ, and the end plate plane equation ax+by+cz+d=0.

7. The lithium battery recycling and disassembly method based on AI+3D vision according to claim 1 is characterized in that: Said step S6, robot motion planning and execution specifically comprises the following steps: Step S61: Construct a dynamic potential field map based on the 3D point cloud features extracted in step 5, specifically: Set the center of the screw hole (x, y, z) as the repulsive potential field source and the potential field radius; The busbar curvature mutation point κ is greater than the threshold and marked as a high-risk area, and the potential field strength is increased by 50%; The end plate plane equation ax+by+cz+d=0 serves as a global constraint to limit the robot motion. Step S62: Define the potential field formula: Where q is the joint angle vector of the six-axis robot, is the coordinate of the nearest point of the ith obstacle in the robot base coordinate system, w i is the potential field weight of the i-th obstacle, β is the potential field attenuation coefficient, F ext is the real-time contact force measured by the force sensor at the end of the robot, F desired is the desired contact force, λ is the impedance term weight, η is the target attraction term weight, q end is the end joint angle of the current path segment, q goal Decompose the joint angles of the target pose; The potential field formula P(q) is input into the optimization solver of the ROS-MoveIt! framework, and the sequential quadratic programming algorithm is used to iteratively solve the optimal joint angle path q(t): Where T is the total time of path execution, is the joint angular velocity vector, τ j is the driving torque of the jth joint, τ max is the upper limit of joint torque; Step S63: During the path execution phase, the impedance term weight λ is adjusted through real-time feedback, specifically: When |F ext -F desired When | is greater than the threshold, adaptive adjustment is activated, and the formula is: l new =λ base ×(1+γ×tanh(F ext -F desired |)); Where λ new is the adjusted impedance weight, λ base is the basic impedance weight, γ is the gain coefficient; If a sudden collision is detected, the potential field expansion is triggered immediately: the potential field radius r is instantly expanded, and the path is replanned; Step S64: The outer loop intermittently receives the screw hole pose (x′, y′, z′) updated by 3D vision, fuses historical data through the Kalman filter, and corrects the path deviation: q corrected =q planned +K×(q vision -q planned ); Where q corrected is the corrected joint angle vector, q planned is the theoretical joint angle in the planned path, q vision is the target pose joint angle updated based on 3D vision, K is the Kalman filter gain matrix; The inner loop adopts the admittance control model, according to F ext Dynamically adjust the end pose, the formula is: Δx=M -1 ×(F ext -F desired )×Δt 2 ; Where Δx is the posture adjustment of the end effector, M is the virtual mass matrix, and Δt is the control period; Step S65: If convergence is still not achieved after three consecutive path corrections, the operation is suspended and the self-diagnosis module is activated: the target area is rescanned using 3D vision to update the potential field map; and the obstacle avoidance path is generated using the RRT*-Connect algorithm; The height of the pole after milling is detected synchronously. If the height is lower than the threshold, an audible and visual alarm is triggered and the current action is terminated.

Images