Human-machine collaboration disassembly method for power lithium battery based on digital twinning

Abstract

The application relates to a man-machine cooperation power lithium battery disassembling method based on digital twinning, which comprises the following steps: S1, carrying out digital twinning modeling on objects existing in a physical space to generate a virtual space; S2, firstly, carrying out digital expression on a disassembling task, then inputting the task into a knowledge base for searching, reusing a disassembling sequence of similar components, and for unrecorded components, adopting a genetic algorithm to plan a disassembling sequence, and finally, based on the sequence of all components, splicing a complete disassembling sequence and updating the sequence in the knowledge base; S3, mapping the current physical disassembling state to the virtual space in S1, and then executing the complete disassembling sequence in the virtual space, wherein object data is taken as disturbance data to obtain multiple disassembling paths through an algorithm; and S4, evaluating the disassembling paths, selecting a disassembling path meeting evaluation indexes, and then combining the disassembling sequence to carry out disassembling. The method is short in disassembling time and safe.

Description

Technical Field

[0001] This invention belongs to the field of human-machine collaboration technology and relates to a human-machine collaborative method for disassembling power lithium batteries based on digital twins. Background Technology

[0002] The global electric vehicle market has developed rapidly in recent years, with global sales projected to reach 23 million vehicles by 2030. Assuming a battery life of 10 years and a battery pack weight of 250 kg, an estimated 5.75 million tons of retired batteries will be generated by 2040. However, spent power batteries contain heavy metals, and improper recycling will cause air, soil, and water pollution, which is inconsistent with the concept of sustainable development. Due to their excellent physicochemical properties, lithium-ion batteries have seen rapid production increases, ultimately leading to a growing number of spent lithium-ion batteries reaching the end of their lifespan. These batteries not only contain toxic and harmful substances but also high-value metal elements; proper recycling will promote resource recycling, environmental protection, and sustainable industrial development.

[0003] As shown above, the recycling of used new energy vehicle batteries is an essential requirement for achieving a circular economy, and dismantling is a crucial and indispensable step in tiered utilization, component remanufacturing, and material recycling. Therefore, the efficient dismantling of used power batteries is of paramount importance and a key measure to achieve both economic recovery and environmental benefits.

[0004] However, due to differences in the production, manufacturing, and service processes of new energy vehicles, power lithium batteries have numerous brands, models, and varying retirement statuses. This diversity and uncertainty in the retirement status of retired power batteries leads to limited dismantling information, complex implementation processes, and a high demand for flexible dismantling methods. Currently, companies primarily use manual dismantling, relying on operators' cognitive and decision-making abilities to dismantle different power lithium batteries; however, manual dismantling is inefficient, prone to operator fatigue, and error-prone. Robotic systems, with their high repeatability, precision, and reliability, have been applied to simple, mechanical, and repetitive operations. However, robots have limited cognitive abilities and cannot cope with complex dismantling environments and scenarios, failing to meet the demand for highly flexible dismantling.

[0005] Existing human-robot collaboration is applied to highly standardized and streamlined assembly scenarios, focusing only on simple, mechanical, and repetitive single structural components such as fasteners. This approach is insufficient for disassembling complex products with diverse structures and brands, such as power lithium batteries. Effective disassembly action sequence planning is crucial for guiding the entire disassembly process. Disassembly action sequence planning includes both sequence planning and path planning. Considering the characteristics of human-robot collaborative disassembly action sequence planning, the disassembly time and difficulty differ depending on whether the same component is disassembled by a human or a robot. Therefore, how to perform reasonable sequence and path planning, fully leveraging the advantages of both, is a significant challenge in the field of human-robot collaboration research.

[0006] Reference 1 (Human-robot collaboration disassembly planning for end-of-life product disassembly process[J]. Robotics and Computer-Integrated Manufacturing, 2021, 71: 102-170.) proposes a disassembly planning method based on human-robot collaboration, which utilizes the flexibility and ability of humans to handle complex tasks, as well as the repeatability and accuracy of robots. At the same time, it conducts targeted design of parts based on remanufacturability parameters, which effectively improves disassembly efficiency. However, the lack of transparency of disassembled product information, the uncertainty of the disassembly process, and the complexity and variability of the environment make implementation difficult. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a human-machine collaborative disassembly method for power lithium batteries based on digital twins;

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

[0009] A human-machine collaborative disassembly method for power lithium batteries based on digital twins includes the following steps:

[0010] S1. Create a digital twin model of the objects existing in the actual disassembly scenario, i.e., the physical space, to generate a virtual disassembly scenario, i.e., a virtual space; the objects include the operator, the robot, the parts to be disassembled, and the disassembly tools;

[0011] S2. First, the disassembly task is digitally represented to obtain digital information. Then, the digital information is input into the historical knowledge base (which stores the planned disassembly sequences of components, updated iteratively by a genetic algorithm or obtained by importing data from previous periods) for similarity retrieval. The disassembly sequences of similar components are reused. For components not recorded in the knowledge base, a genetic algorithm is used to plan the disassembly sequence of the component. Finally, based on the disassembly sequences of all components, a complete disassembly sequence is spliced ​​to generate and updated in the knowledge base.

[0012] To obtain a safer and more efficient human-robot collaborative disassembly action sequence, this invention adjusts the fitness function on the genetic algorithm (GA) model, as it is the key to selection. For conventional assembly processes, the optimization objectives are usually product-specific sequence constraints, global goals, and the processing time of each sub-process. However, for the human-robot collaborative disassembly process in this invention, we also consider the number of tool transfers between humans and robots, the duration of idle time, and the disassembly capabilities of humans and robots, in order to better conform to the characteristics of human-robot collaborative disassembly.

[0013] The improvement of the genetic algorithm lies in the fitness function, and the fitness function F of the improved genetic algorithm is shown in equation (1);

[0014] F = W1S c +W2S e +W3S b +W4S t +W5S u (1);

[0015] In the formula, S c S e S b S t and S u These are the capability value, execution time, rest time, number of tool changes, and resource utilization rate for the current subtask; W i The weights of the fitness function,

[0016] In the improved genetic algorithm, mutation recombination and hybrid decoding (i.e. Step 3, Step 5, and Step 6 below) are used to generate new single entities (new feasible decomposition sequences), and repeated iterations are performed. Termination conditions are set (e.g., the number of iterations is set or the fitness function has reached a set value). When the termination condition is reached, a better decomposition sequence is obtained and stored in the historical knowledge base for sequence knowledge reuse.

[0017] The specific steps of the genetic algorithm are as follows:

[0018] Step 1: Initialize the population size n I Based on the constraints of a certain decomposition task and the evaluation index S i i = {c, e, b, t, u}, n is determined manually. I Personal-machine collaborative disassembly sequence, each personal-machine collaborative disassembly sequence represents a chromosome individual ρ;

[0019] Step 2: Preprocessing. Determine the weight W of each indicator based on the global objective. i Let i = {1, 2, 3, 4, 5}, while maintaining the sum of weights.

[0020] Step 3: Fitness Calculation. Calculate the fitness value F for each individual ρ. ρ ;

[0021] Step 4: Determine if the conditions are met. If the individual ρ function value F ρ >F, where F is the set fitness threshold, stops the process, and selects the individual ρ with the highest chromosome selection function value. Otherwise, if the number of offspring n τ>n max n max To determine the maximum number of offspring individuals, replicate the first 40% of individuals ρ. a Become the next offspring n τ+1 A portion, otherwise select the top 40% of individuals ρ a Become n τ+1 Part of;

[0022] Step 5: Crossover operation. Select the remaining individuals ρ b Crossover operation with α%;

[0023] Step 6: Mutation operation. Select the remaining individuals ρ b Perform mutation operation with β%;

[0024] Step 7: Generate the next generation population. Next generation population n τ+1 =ρ a +ρ b ;

[0025] Step 8: Return to Step 3 until the condition is met.

[0026] S3. Map the current physical dismantling state in the physical space to the virtual space in S1 (dismantling is a dynamic process, and the environment update of the virtual space needs to be based on the current dismantling process, which refers to the real-time state in the actual dismantling scenario. The environment (position, motion data, etc.) data update is achieved by sensor technology). The environmental state is updated, and then the complete dismantling sequence generated in S2 is executed in the virtual space. When the dismantling sequence is executed, the sensor collects object data in the physical space as perturbation data to make the virtual space evolve and obtain twin data. Based on the obtained twin data, relevant robot motion evaluation indicators are set, and multiple dismantling paths are obtained through the reinforcement learning DDPG method.

[0027] S4. Evaluate the multiple disassembly paths obtained in S3, select the disassembly path that meets the evaluation criteria, and combine it with the complete disassembly sequence generated in S2, which is considered a reasonable disassembly action sequence. Disassemble the power lithium battery according to the disassembly action sequence.

[0028] As a preferred technical solution:

[0029] As described above, the human-machine collaborative lithium battery dismantling method based on digital twins involves digital twin modeling in step S1. This modeling integrates geometric, physical, behavioral, and rule-based characteristics. The modeling process is as follows: First, a digital object corresponding to the object in the physical space is constructed using 3D modeling software (such as Solidworks), and environmental data is updated through sensor technology. Then, the point cloud in the dismantling environment is classified and segmented using a point cloud segmentation method, and a corresponding XML file storage structure is designed to store the dismantling information obtained after segmentation. Next, the similarity distance between the segmented point cloud and the template point cloud is calculated using the EMD method, and the dismantling scene diagram with attributes is obtained by combining the information stored in XML. Finally, the virtual space is modeled and simulated in the Coppeliasim platform based on the dismantling scene diagram information expression method. To guide the actual dismantling process with the simulation results of step S1, the physical space and virtual space are connected based on the V-rep interface and Socket programming to achieve motion control of the actual robot.

[0030] The specific process of S1 is as follows:

[0031] 1) Environmental data is collected using multi-source data sensors and used to update parameters of the geometric model, such as dimensions, structure, and appearance. For example... Figure 2 As shown, environmental elements can be categorized into static and dynamic environments based on their time-varying nature. Dynamic environments primarily refer to the activity space for workers and robots, including human posture data (standing, bending, etc.), hand position data, and electrical current data. Static environments, on the other hand, mainly refer to the environment composed of static elements during the initialization of the disassembly activity, including position data, size data, and feature data. Using an Intel RealSense D435 depth camera, color information and corresponding depth information within the physical disassembly space can be acquired; for example... Figure 2 The position and posture data of the workers can be obtained by a depth camera;

[0032] 2) Octree-based point cloud segmentation algorithm: The point cloud segmentation algorithm is used to classify and segment the point cloud in the deconstruction environment.

[0033] 3) XML-based methods for expressing environmental information: such as Figure 3 , 4As shown, a corresponding XML file storage structure is designed to store the disassembly information obtained after segmentation. For the static XML file structure, the physical data of each environmental element exists under the "Object_i" tag, involving the environmental element's ID, attribute value, current state value, etc. Task attributes related to the disassembly elements mainly involve object category, object size, object weight, surface roughness, object shape, etc. The current state is a data representation related to the environmental element's current position, occlusion degree, current pose, etc. There are also interrelationships between different environmental elements, mainly manifested in the IDs of other elements associated with the current environmental element, relative distance, relative pose, and assembly relationship corresponding to the disassembly element. The main difference between dynamic and static environments lies in the mathematical representation of dynamic elements such as workers and robots. For the dynamic XML file structure, each disassembly element (i.e., mainly workers and robots) includes its ID, attribute, service, current status, etc. Based on prior knowledge, relevant content is filled under the worker's attribute tag, including gender, height, weight, and moveable range. Service tags indicate whether a worker is currently in operation. In addition, current status tags include the worker's position, speed, attitude, and action. Similar to workers, robots, as another major category of dynamic environmental elements, also involve ID tags, attribute tags, service tags, and current status tags. Robot attributes include specific categories such as gripper opening, maximum load capacity, and maximum continuous working time. Simultaneously, the robot's current status also includes its operating parameters such as current, voltage, power, vibration, electromagnetic fields, and momentum.

[0034] like Figure 5 As shown, after using the point cloud segmentation algorithm, information is extracted from the XML file formed by the physical space by field to obtain the attribute values ​​and state values ​​corresponding to each decomposed element. Then, the Earth Mover's Distance (EMD) method is used to calculate the similarity distance between the segmented point cloud and the scene point cloud in the template point cloud (the scene point cloud is based on the pre-defined physical space layout and represents the layout structure of each element, expressed as the positional relationship of each element in the scene diagram (e.g., the worker is to the left of the robot)). When this value approaches 0, the category of the segmented set is labeled with the category label corresponding to the template. After combining the actions of the worker and the robot, a decomposed scene diagram with semantic relationships can be formed. Combined with the pre-collected attribute information, the structure of the scene diagram can be further improved, and attribute values ​​can be attached next to the instances to obtain a decomposed scene diagram with attributes.

[0035] 4) Scene Graph-Based Information Representation Method: By capturing different regions of the scene graph in the form of a sliding window, the object attributes, relationships between objects, and object actions of that region can be obtained to construct a corresponding 3D digital model in the virtual disassembly unit. Based on the positional attributes between objects, the relative relationships between environmental objects and other elements in space can be further adjusted. Since the initial model is an externally imported 3D model, it can be largely mapped to real environmental objects. By setting embedded functions in the model within the virtual platform, the virtual model can change accordingly with changes in object position and posture; the virtual disassembly environment can be dynamically updated based on changes in the scene graph.

[0036] 5) Virtual space modeling and simulation: Digital twin disassembly simulation is performed based on the Coppeliasim platform, which can guide the actual human-computer collaborative disassembly process;

[0037] 6) Socket-based robot control method: This method establishes a communication connection between the real robot and the computer through socket programming, enabling motion control of the actual robot. The socket-based robot control method uses the vrep interface to establish a connection with the Coppeliasim virtual environment and calls the functions encapsulated in vrep to obtain the robot's joint angles in the simulation environment or set specific joint angles. When the designed robot running path does not pose a collision risk during the simulation, it will output its trajectory points or the robot's six joint angle values, which will be transmitted to the actual robot via socket programming, thus achieving motion control of the actual robot.

[0038] As described above, in the human-machine collaborative power lithium battery disassembly method based on digital twins, the process of digitally representing the disassembly task to obtain digital information in S2 is as follows: First, the component to be disassembled is decomposed into multiple sub-tasks. Then, the point cloud transformer model is used to perform point cloud feature recognition to obtain the disassembly feature matrix of the disassembled component (composed of component name symbols and key dimensions). Then, based on feature matching and reasoning methods, the disassembly constraint matrix and disassembly method matrix are obtained. The disassembly feature matrix, disassembly constraint matrix, and disassembly method matrix are the digital information of the component. Among them, the disassembly constraint matrix and the method matrix are used to obtain the disassembly sequence and disassembly method, and the feature matrix is ​​used for similarity retrieval in the historical knowledge base.

[0039] Based on the composition and structure of lithium batteries, a feasible disassembly sequence is defined. The disassembly features of all components are obtained through the PCT model (i.e., a disassembly feature matrix is ​​constructed for each component). Two components with mutual constraints are used as disassembly reference components. Other components with disassembly constraints are then identified, constructing a disassembly constraint matrix and completing the component attribute encoding method. For example... Figure 6 As shown, based on the composition structure of the lithium battery, the disassembly sequence can be defined as [P1, P2, P3, P4, P5, P6, P7]. P1 to P8 are pre-defined as representing the substrate, while P1′ to P8′ represent the parts to be disassembled. After obtaining the disassembly features of all disassembled components using the PCT model (i.e., constructing the disassembly feature matrix for each component: Mat(a)), components with matching features are selected through pairwise pairing. Typical matching rules include: 1) smooth round hole and smooth round shaft; 2) threaded round hole and threaded round shaft; 3) smooth square hole and smooth square shaft; 4) slots with specific shapes and corresponding wedges, etc. Two parts with mutual constraints are used as disassembly reference parts. Other parts with disassembly constraints are then searched to construct the disassembly constraint matrix: Mat(b) (represented by 0 / 1, where a component with a constraint relationship is 1, otherwise 0). Mat(b) contains two rows of vectors. The top row represents the disassembly constraints between the current part as a reference component and other parts, while the bottom row represents the disassembly constraints between the current part as the component to be disassembled and other parts. For example... Figure 6 As shown, the disassembly method consists of two parts: connection method and disassembly tool. The former mainly refers to bolt and nut connection, screw connection, and snap-fit ​​connection, while the latter can be divided into adjustable wrenches, screwdrivers, pliers, and clamps according to their uses. Both the connection method and the disassembly tool are encoded into numerical form to facilitate the definition of the disassembly method matrix corresponding to the current disassembly subtask: Mat(c). Taking part P_7 as an example, after retrieval based on the input features, the final disassembly method matrix is ​​Mat(c) = [2-B].

[0040] To address the dismantling of power lithium batteries, dismantling sub-tasks must be selectively executed based on the capabilities of workers and robots. Assessing the capabilities of workers and robots is crucial for assigning dismantling sub-tasks. Physical performance evaluation indicators for workers and robots include weight, volume, and operational accuracy. Individual worker capabilities can be defined as grasping range, grasping load, moving speed, field of vision, and reaction time. Robot capabilities are related to grasping load, grasping range, grasping accuracy, moving speed, field of vision, and operation time. To comprehensively consider all evaluation items, the score range for all evaluation items is set to [0,1]. The average value of all evaluation items is summed to obtain the capability score for the worker and robot corresponding to the current dismantling sub-task. Scores greater than 0.5 are assigned to worker tasks, and scores less than 0.5 are assigned to robot tasks.

[0041] The specific process of decomposing the sequence design is as follows: Figure 1 As shown: a) Based on the PCT model, obtain the features of the disassembled parts, take two parts with mutual constraints as the disassembly benchmark parts, find other parts with disassembly constraints, and obtain the disassembly constraint matrix and disassembly method matrix based on feature matching and reasoning methods to obtain the disassembly order; b) According to the disassembly order, perform similarity retrieval of the feature values ​​of each sub-part in the historical knowledge base. When the attribute values ​​of two parts are similar, the planned disassembly sequence can be directly reused; c) For disassembly sub-parts not recorded in the historical knowledge base, the improved genetic algorithm will be used to plan the disassembly sequence of this part. Finally, based on the disassembly order, the complete disassembly sequence is spliced ​​to generate and updated in the knowledge base.

[0042] For similarity retrieval in b above, this invention divides the lithium battery point cloud model into corresponding model sub-blocks according to the sub-task category through spectral clustering, constructs feature tensors using GS (Gaussian curvature), SI (shape index), and SDF, and then jointly constructs a third-order tensor to represent each sub-block of the model, and obtains similar sub-blocks using tensor norm;

[0043]

[0044] Among them, G 0t G is the GS of the t-th point within the Gaussian curvature matrix of model sub-block i. st Let GS be the s-th nearest neighbor of the t-th point, where s = 1, 2, ..., k, t = 1, 2, ..., m, and m is the number of points within model sub-block i. For SI i and SDF i There are similar construction methods, such as:

[0045]

[0046]

[0047] For model sub-block i, by GS i SI i SDF i The third-order tensor T, which is composed of (k+1)×m×3 pages, is... (i) ,Right now:

[0048] T (i) (∶,∶,1)=GS i ;

[0049] T (i) (∶,∶,2)=SI i ;

[0050] T (i) (∶,∶,3)=SDF i ;

[0051] The tensor norm distance of each model sub-block is calculated using the tensor norm formula, and then the cosine similarity evaluation function is used as the distance metric between the two model sub-blocks, as shown in equation (2):

[0052]

[0053] Where X i X is a model sub-block in the historical knowledge base. i ' represents the current model sub-block, ||·|| denotes the tensor norm distance, where cosθ ranges from [0,1]. If this value equals 1, the similarity reaches its maximum. A certain threshold β is set, and the value of β varies depending on different parts. When cosθ>β, the historical subtask decomposition sequence can be reused. Otherwise, it needs to be replanned.

[0054] As described above, in the human-machine collaborative disassembly method for power lithium batteries based on digital twins, multiple disassembly paths are obtained through the reinforcement learning-based DDPG method in step S3. Specifically, the process is as follows: 1) The disassembly sequence planned in step S2 is used as input to DDPG for M training iterations; 2) The coordination method between the worker and the robot during the human-machine collaboration process is fixed, with each training iteration consisting of T steps. For each step, the current state s is recorded. t Perform action a t The reward value r obtained t To obtain the next state s t+1 and with point group transition = (s t ,a t ,r t ,s t+1) are stored in the experience replay pool in the form of ) ; 3) at each step, a small batch (small batch is a proper noun, generally determined according to the computer's computing power and its own dataset, generally set to 32, or 64, 128, can be adjusted according to its own needs and computer capabilities, usually a multiple of 2) transition is randomly sampled from the training pool, the loss is calculated, and the network parameters are updated; 4) it is determined whether the current state has reached the endpoint. If the endpoint is reached, the training ends and the next training begins. Otherwise, training continues until the step length is reached, and finally M decomposition paths are obtained.

[0055] The values ​​of training iterations M and step size T depend on the specific circumstances. In this application, the training iterations M is set to 500 and the step size T is set to 300.

[0056] As described above, the human-machine collaborative disassembly method for power lithium batteries based on digital twins, action a t This is represented by the change in robot joint angle, as shown in equation (6);

[0057] a t =[θ1,θ2,θ3,θ4,θ5,θ6] (6);

[0058] In the formula, θ1, θ2, θ3, θ4, θ5 and θ6 represent the angle values ​​of the robot's six joints, respectively.

[0059] In the human-machine collaborative disassembly method for power lithium batteries based on digital twins as described above, the evaluation indicators in step S3 are: the distance D between the end effector and the component to be disassembled, and the distance D between the end effector and the obstacle. o Safety of robot movement r D, D o and S r The calculation formulas are shown in equations (3) to (5);

[0060]

[0061]

[0062] i∈[0,I] (4);

[0063]

[0064] Among them, (x r ,y r ,z r (x) indicates the position of the end effector. o ,y o ,z o (x) represents the location of the target object. i ,y i ,z iS represents the location of the obstacle, and I represents the number of obstacles; if a collision occurs during the movement, then S... r The value is assigned as -1, otherwise it is assigned as 1.

[0065] The above-mentioned human-machine collaborative disassembly method for power lithium batteries based on digital twins has a reward value r. t The calculation formula is shown in equation (7);

[0066] r t =k1r1+k2r2-k3r3 (7);

[0067] In the formula, r1 represents the positive reward value, used to incentivize the robot to approach the finish line; r2 represents the obstacle avoidance reward value, used to keep the robot as far away from obstacles as possible; r3 represents the time reward value, used to encourage the robot to make effective movements; the reward value r t It consists of three parts: positive reward value r1, obstacle avoidance reward value r2, and time reward value r3, with k1, k2, and k3 as weights;

[0068] r1, r2, and r3 are calculated using equations (8) to (10), respectively.

[0069] r1 = (300 - D) (8);

[0070]

[0071] r3=T t (10);

[0072] In the formula, T represents the distance from the end effector to the i-th obstacle; t This indicates the current step size during training.

[0073] As described above, in the human-machine collaborative dismantling method for power lithium batteries based on digital twins, selecting the dismantling path that meets the evaluation criteria in S4 refers to selecting the dismantling path that meets the reward value r from the multiple dismantling paths obtained in S3. t The longest path.

[0074] Beneficial effects

[0075] (1) The present invention provides a human-machine collaborative power lithium battery disassembly method based on digital twins, which combines the flexibility, cognition and decision-making ability of humans with the high operational intensity of robots, and has great advantages in disassembly time, disassembly cost and disassembly flexibility.

[0076] (2) The present invention provides a human-machine collaborative power lithium battery dismantling method based on digital twins, which realizes digital modeling of human-machine collaborative dismantling. When facing retired products with varying batches, varieties, and retirement conditions, it accurately maps the dismantling process, realizes interactive control between physical and virtual spaces, and optimizes the decision-making process of human-machine collaboration with real-time perceived environmental data. It effectively reduces the reverse operation of human-machine dismantling and debugging process and ensures the work efficiency of human-machine collaborative dismantling. Attached Figure Description

[0077] Figure 1 This invention presents a human-machine collaborative modeling method for disassembling and modeling power lithium batteries based on digital twins.

[0078] Figure 2 This invention describes the process of physical disassembly environment data acquisition and point cloud generation.

[0079] Figure 3 This is the static XML file structure proposed in this invention;

[0080] Figure 4 This invention presents a dynamic XML file structure.

[0081] Figure 5 This invention provides a process for generating a disassembly scene diagram.

[0082] Figure 6 This is the component attribute encoding method proposed in this invention;

[0083] Figure 7 Based on the decomposition results of knowledge reuse;

[0084] Figure 8 This is the pre-programmed disassembly result. Detailed Implementation

[0085] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0086] A human-machine collaborative disassembly method for power lithium batteries based on digital twins, the specific steps of which are as follows:

[0087] S1. Create a digital twin model of the objects existing in the actual disassembly scenario, i.e., the physical space, to generate a virtual disassembly scenario, i.e., a virtual space; the objects include the operator, the robot, the parts to be disassembled, and the disassembly tools;

[0088] The digital twin modeling process is as follows: First, a digital object corresponding to the object in the physical space is constructed based on 3D modeling software, and environmental data is updated through sensor technology. Then, the point cloud in the disassembled environment is classified and segmented using a point cloud segmentation method, and a corresponding XML file storage structure is designed to store the segmented disassembled information. Next, the similarity distance between the segmented point cloud and the template point cloud is calculated using the EMD method, and the information stored in XML is combined to obtain a disassembled scene graph with attributes. Finally, the modeling and simulation of the virtual space is realized in the Coppeliasim platform based on the disassembled scene graph information expression method, and the physical space and virtual space are connected based on the V-rep interface and Socket programming to realize the motion control of the actual robot.

[0089] S2. First, the disassembly task is digitally represented to obtain digital information. Then, the digital information is input into the historical knowledge base for similarity retrieval. The disassembly sequences of similar parts are reused. For parts not recorded in the knowledge base, a genetic algorithm is used to plan the disassembly sequence of the part. Finally, based on the order of all parts, a complete disassembly sequence is spliced ​​to generate and updated in the knowledge base.

[0090] The process of digitally representing the disassembly task to obtain digital information is as follows: First, the task to be disassembled is decomposed into multiple sub-tasks. Then, the point cloud transformer model is used to perform point cloud feature recognition to obtain the disassembly feature matrix of the disassembled component. Then, based on feature matching and reasoning methods, the disassembly constraint matrix and disassembly method matrix are obtained. The disassembly feature matrix, disassembly constraint matrix and disassembly method matrix are the digital information of the component.

[0091] The improvement of the genetic algorithm lies in the fitness function, and the fitness function F of the improved genetic algorithm is shown in equation (1);

[0092] F = W1S c +W2S e +W3S b +W4S t +W5S u (1);

[0093] In the formula, S c S e S b S t and S u These are the capability value, execution time, rest time, number of tool changes, and resource utilization rate for the current subtask; W i The weights of the fitness function,

[0094] For similarity retrieval, this invention divides the lithium battery point cloud model into corresponding model sub-blocks according to the sub-task category through spectral clustering, constructs feature tensors using GS (Gaussian curvature), SI (shape index), and SDF, and then jointly constructs a third-order tensor to represent each sub-block of the model, and obtains similar sub-blocks using tensor norm;

[0095]

[0096] Among them, G 0t G is the GS of the t-th point within the Gaussian curvature matrix of model sub-block i. st Let GS be the s-th nearest neighbor of the t-th point, where s = 1, 2, ..., k, t = 1, 2, ..., m, and m is the number of points within model sub-block i. For SI i and SDF i There are similar construction methods, such as:

[0097]

[0098]

[0099] For model sub-block i, by GS i SI i SDF i The third-order tensor T, which is composed of (k+1)×m×3 pages, is... (i) ,Right now:

[0100] T (i) (∶,∶,1)=GS i ;

[0101] T (i) (∶,∶,2)=SI i ;

[0102] T (i) (∶,∶,3)=SDF i ;

[0103] The tensor norm distance of each model sub-block is calculated using the tensor norm formula, and then the cosine similarity evaluation function is used as the distance metric between the two model sub-blocks, as shown in equation (2):

[0104]

[0105] Where X i X is a model sub-block in the historical knowledge base. i' represents the current model sub-block, ||·|| denotes the tensor norm distance, where cosθ ranges from [0,1]. If this value equals 1, the similarity reaches its maximum. A certain threshold β is set, and the value of β varies depending on different parts. When cosθ>β, the historical subtask decomposition sequence can be reused. Otherwise, it needs to be replanned.

[0106] S3. Map the current physical dismantling state in the physical space to the virtual space in S1 to update the environment state. Then, execute the splicing in S2 in the virtual space to generate a complete dismantling sequence. When executing the dismantling sequence, the sensor collects object data in the physical space as perturbation data to make the virtual space evolve and obtain twin data. Based on the obtained twin data, set relevant robot motion evaluation indicators and obtain multiple dismantling paths through the reinforcement learning DDPG method.

[0107] The established evaluation metrics are: the distance D between the end effector and the part to be disassembled, and the distance D between the end effector and the obstacle. o Safety of robot movement r D, D o and S r The calculation formulas are shown in equations (3) to (5);

[0108]

[0109]

[0110] i∈[0,I] (4);

[0111]

[0112] Among them, (x r ,y r ,z r (x) indicates the position of the end effector. o ,y o ,z o (x) represents the location of the target object. i ,y i ,z i () indicates the location of the obstacle, and I indicates the number of obstacles;

[0113] The specific process of obtaining multiple decomposition paths using the reinforcement learning-based DDPG method is as follows: 1) Use the decomposition sequence planned in S2 as input for DDPG and perform 500 training iterations; 2) Fix the cooperation method between the worker and the robot during human-robot collaboration, with each training loop consisting of 300 steps, and record the current state s for each step. t Perform action a t The reward value r obtainedt To obtain the next state s t+1 and with point group transition = (s t ,a t ,r t ,s t+1 ) are stored in the experience replay pool in the form of ); 3) at each step, a small batch (small batch is a proper noun, generally set to 32, or 64, 128) of transitions are randomly sampled from the training pool, the loss is calculated, and the network parameters are updated; 4) it is determined whether the current state has reached the end point. If the end point is reached, the training ends and the next training begins. Otherwise, training continues until the end point is reached, and finally M decomposition paths are obtained.

[0114] Among them, action a t This is represented by the change in robot joint angle, as shown in equation (6);

[0115] a t =[θ1,θ2,θ3,θ4,θ5,θ6] (6);

[0116] In the formula, θ1, θ2, θ3, θ4, θ5 and θ6 represent the angle values ​​of the robot's six joints, respectively;

[0117] Reward value r t The calculation formula is shown in equation (7);

[0118] r t =k1r1+k2r2-k3r3 (7);

[0119] In the formula, r1 represents the positive reward value, which is used to motivate the robot to approach the finish line; r2 represents the obstacle avoidance reward value, which is used to keep the robot as far away from obstacles as possible; r3 represents the time reward value, which is used to encourage the robot to make effective movements; k1, k2, and k3 are weights.

[0120] r1, r2, and r3 are calculated using equations (8) to (10), respectively.

[0121] r1 = (300 - D) (8);

[0122]

[0123] r3=T t (10);

[0124] In the formula, T represents the distance from the end effector to the i-th obstacle; t This indicates the current step size during training.

[0125] S4. Evaluate the multiple dismantling paths obtained in S3 and select the one that satisfies the reward value r.t After the longest path, the complete disassembly sequence generated in S2 is considered a reasonable disassembly action sequence, and the power lithium battery is disassembled according to the disassembly action sequence.

[0126] The following is a detailed description of the process of disassembling a power lithium battery using the method of the present invention:

[0127] The objects dismantled in the experiment were waste power lithium batteries from a certain manufacturer, designated as S471 standard C box and S472 standard G box.

[0128] Experimental verification was conducted on two key aspects of the parts disassembly process: ① human-machine collaborative disassembly; ② disassembly sequence planning. Two different models were compared: 1) a pre-programmed model; 2) a planning model based on GA and DDPG algorithms under digital twins (i.e., the disassembly method of this application).

[0129] In both stages, 10 parallel groups were set up, with each group attempting to grab or disassemble a specific component. Running time, collision rate, and accuracy data were recorded, and the final result was the average of the 10 parallel groups.

[0130] ① Experimental verification of human-machine collaborative disassembly:

[0131] A disassembly task of 300 S471 standard C-cells and S472 standard G-cells was set up as a disassembly experiment, and the results were verified by testing disassembly time, disassembly deviation, collaboration time, and collision rate. Simultaneously, the geometric point cloud and full-view image dataset of the lithium batteries were acquired using an Intel RealSense depth camera. Hardware resources required for the disassembly experiment: GTX 1080 Ti graphics card, 8GB of video memory.

[0132] The hyperparameter settings for model training during feature decomposition and recognition are shown in Table 1. The model training iteration cycle is 300 epochs, and the learning rate is 10. -4 .

[0133] Table 1 Hyperparameter settings during model training

[0134]

[0135] Collaborative disassembly testing refers to the stage where the robot assists the worker in securing the parts after grasping and transferring them. Table 2 presents the results of the collaborative disassembly test in terms of disassembly time, collision rate, and disassembly accuracy. Considering disassembly time, the method of this invention requires only 13.375 seconds for collaborative disassembly, while the pre-programmed mode requires 20.347 seconds to complete the same disassembly task. This is because the robot's motion path is usually redundant in the pre-programmed method, and due to the inability to guarantee operator safety, the sequence planning is usually sequential and primarily worker-driven, with the robot only providing simple transfer tools. In contrast, the disassembly sequence planned by the method of this application fully utilizes the idle states of both the worker and the robot for parallel operation without human-robot safety interference, effectively shortening the disassembly time. Regarding potential collisions during disassembly, the pre-programmed disassembly method results in a higher collision rate (4.349%), which is significantly lower than the collision rate of the method of this application (1.384%). This is because the robot cannot adaptively adjust its posture according to changes in human posture. Furthermore, the disassembly deviation can be reduced to 0.842 mm when using the method of this application, while the disassembly deviation in the pre-programmed mode reaches 1.788 mm. This is because the robot cannot dynamically adjust its own position and orientation as the position or orientation of the object to be disassembled changes.

[0136] Table 2 Comparison of human-machine collaborative disassembly performance based on different methods

[0137]

[0138] ② Experimental verification of decomposition sequence planning:

[0139] The disassembly sequence planning experiment was used to compare pre-programmed disassembly planning with the knowledge reuse-based disassembly sequence planning of this application (i.e., retrieving and reusing disassembly sequences of similar components from a historical knowledge base) to verify that the disassembly mode proposed in this application can effectively improve disassembly efficiency and task responsiveness. Assuming that the S471 standard C-box is a pre-planned sequence stored in the historical knowledge base, the experiment used the S471 standard G-box for sequence planning. Figure 7 , Figure 8 This paper presents the human-machine disassembly sequence planning and the time consumption of each sequence in the S471 standard G box disassembly task using the knowledge reuse-based disassembly method and the pre-programmed disassembly method of this application. Figure 7 , Figure 8 The specific operation methods are as follows: grabbing the tool, grabbing the disassembled parts, placing the parts to be disassembled, placing the disassembled parts, and disassembling are represented by A, B, C, D, and E respectively, and the numbers following them represent the time taken for each operation.

[0140] The results show that, compared with the time spent on disassembly action sequence planning in the pre-programmed disassembly method for power lithium battery disassembly tasks (291s), the human-machine collaborative power lithium battery disassembly method based on digital twins in this application shortens the disassembly action sequence by nearly 20% (236s). At the same time, the workload of workers in collaborative tasks is significantly reduced. This is because this application effectively reduces the time for replanning new tasks by reusing knowledge from the historical knowledge base through similarity retrieval, thereby shortening the time for disassembly action sequence planning.

Claims

1. A human-machine collaborative disassembly method for power lithium batteries based on digital twins, characterized in that... Includes the following steps: S1. Create a digital twin model of the objects existing in the actual disassembly scenario, i.e., the physical space, to generate a virtual disassembly scenario, i.e., a virtual space; the objects include the operator, the robot, the parts to be disassembled, and the disassembly tools; S2. First, the disassembly task is digitally represented to obtain digital information. Then, the digital information is input into the historical knowledge base for similarity retrieval. The disassembly sequences of similar parts are reused. For parts not recorded in the knowledge base, a genetic algorithm is used to plan the disassembly sequence of the part. Finally, based on the disassembly sequences of all parts, a complete disassembly sequence is spliced ​​to generate and updated in the knowledge base. The improvement of the genetic algorithm lies in the fitness function, and the fitness function F of the improved genetic algorithm is shown in equation (1); F = W1S c + W2S e + W3S b + W4S t + W5S u (1); In the formula, S c , S e , S b , S t , and S u are the capability value, execution time, rest time, tool replacement times, and resource utilization rate under the current subtask, respectively. W i The weights of the fitness function, S3. Map the current physical dismantling state in the physical space to the virtual space in S1 to update the environment state. Then execute the complete dismantling sequence in S2 in the virtual space. When executing the dismantling sequence, the sensor collects object data in the physical space as perturbation data to make the virtual space evolve and obtain twin data. Based on the obtained twin data, set relevant robot motion evaluation indicators and obtain multiple dismantling paths through the reinforcement learning DDPG method. S4. Evaluate the multiple disassembly paths obtained in S3, select the disassembly path that meets the evaluation criteria, and combine it with the complete disassembly sequence in S2, which is considered a reasonable disassembly action sequence. Disassemble the power lithium battery according to the disassembly action sequence.

2. The human-machine collaborative disassembly method for power lithium batteries based on digital twins according to claim 1, characterized in that, The digital twin modeling process in S1 is as follows: First, a digital object corresponding to the object in the physical space is constructed based on 3D modeling software, and environmental data is updated through sensor technology; then, the point cloud in the disassembled environment is classified and segmented using a point cloud segmentation method, and a corresponding XML file storage structure is designed to store the segmented disassembled information; next, the similarity distance between the segmented point cloud and the template point cloud is calculated using the EMD method, and the disassembled scene graph with attributes is obtained by combining the information stored in XML; finally, the virtual space is modeled and simulated in the Coppeliasim platform based on the disassembled scene graph information expression method.

3. The human-machine collaborative disassembly method for power lithium batteries based on digital twins according to claim 2, characterized in that, Using the V-rep interface and Socket programming, the physical space and virtual space are connected to achieve motion control of the actual robot.

4. The human-machine collaborative disassembly method for power lithium batteries based on digital twins according to claim 3, characterized in that, The process of digitally representing the disassembly task to obtain digital information in S2 is as follows: First, the task to be disassembled is decomposed into multiple sub-tasks. Then, the point cloud transformer model is used to perform point cloud feature recognition to obtain the disassembly feature matrix of the disassembled component. Then, based on feature matching and reasoning methods, the disassembly constraint matrix and disassembly method matrix are obtained. The disassembly feature matrix, disassembly constraint matrix and disassembly method matrix are the digital information of the component.

5. The human-machine collaborative disassembly method for power lithium batteries based on digital twins according to claim 4, characterized in that, In S3, multiple decomposition paths are obtained through the reinforcement learning DDPG method. The specific process is as follows: 1) The decomposition sequence planned in S2 is used as the input of DDPG to perform M training cycles. 2) fix the way the worker cooperates with the robot in the human-robot collaboration process, record the current state s for each step under T steps of each training cycle t perform action a t the reward value r obtained t , get the next state s t+1 , and store it in the form of point group transition=(s t , a t , r t , s t+1 ) in the experience replay pool; 3) At each step, a small batch of transitions are randomly sampled from the training pool, the loss is calculated, and the network parameters are updated; 4) It is determined whether the current state has reached the endpoint. If the endpoint is reached, the training ends and the next training begins. Otherwise, training continues until the step ends, and finally, M decomposition paths are obtained.

6. The human-machine collaborative disassembly method for power lithium batteries based on digital twins according to claim 5, characterized in that, Action a t This is represented by the change in robot joint angle, as shown in equation (6); a t =[θ1,θ2,θ3,θ4,θ5,θ6] (6); In the formula, θ1, θ2, θ3, θ4, θ5 and θ6 represent the angle values ​​of the robot's six joints, respectively.

7. The human-machine collaborative disassembly method for power lithium batteries based on digital twins according to claim 6, characterized in that, The evaluation metrics in S3 are: the distance D between the end effector and the part to be disassembled, and the distance D between the end effector and the obstacle. o Safety of robot movement r D, D o and S r The calculation formulas are shown in equations (3) to (5); Among them, (x r ,y r ,z r (x) indicates the position of the end effector. o ,y o ,z o (x) represents the location of the target object. i ,y i ,z i ) represents the location of the obstacle, and I represents the number of obstacles.

8. The human-machine collaborative disassembly method for power lithium batteries based on digital twins according to claim 7, characterized in that, Reward value r t The calculation formula is shown in equation (7); r t =k1r1+k2r2-k3r3 (7); In the formula, r1 represents the positive reward value, which is used to encourage the robot to approach the finish line; r2 represents the obstacle avoidance reward value, which is used to keep the robot as far away from obstacles as possible; r3 represents the time reward value, which is used to encourage the robot to make effective movements; k1, k2, and k3 are weights. r1, r2, and r3 are calculated using equations (8) to (10), respectively. r1＝(300-D) (8)； r3＝T t (10)； In the formula, T represents the distance from the end effector to the i-th obstacle; t This indicates the current step size during training.

9. The human-machine collaborative disassembly method for power lithium batteries based on digital twins according to claim 8, characterized in that, Selecting the decomposition path that meets the evaluation criteria in S4 refers to selecting the path that meets the reward value r from the multiple decomposition paths obtained in S3. t The longest path.

Images