AI-based intelligent dismantling sequence planning method for power batteries

Abstract

This invention discloses an artificial intelligence-based intelligent dismantling sequence planning method for power batteries, belonging to the field of power battery dismantling technology. The method includes initializing a dismantling sequence simulation environment, loading a dismantling action constraint relationship network, a dismantling tool action model, and a dismantling safety rule base; employing a reinforcement learning agent to drive Monte Carlo tree search, simulating component dismantling order to generate multiple candidate sequences, and simultaneously collecting simulation data such as tool selection, time consumption, safety risk score, and resource consumption assessment for each step; constructing a multi-objective evaluation model with four dimensions: dismantling efficiency, safety, resource cost, and component integrity recovery rate, quantifying and fusing data, and calculating a comprehensive score; selecting the highest-scoring sequence as the recommended sequence, and outputting an executable process guidance document with detailed steps, a tool list, and risk warnings. This invention overcomes the limitations of human experience, covers multiple paths under complex constraints, quantifies and balances multi-objective conflicts, and improves the scientific nature of power battery dismantling.

Description

Technical Field

[0001] This invention belongs to the field of power battery dismantling technology, specifically an intelligent dismantling sequence planning method for power batteries based on artificial intelligence. Background Technology

[0002] In the recycling of power batteries, the planning of the dismantling sequence directly affects dismantling efficiency, safety, and resource recovery benefits. Existing technologies mostly rely on manual experience or preset fixed dismantling sequences, making it difficult to systematically explore globally feasible sequences considering the multi-component coupling structure of the battery pack, differences in tool compatibility, and potential risk variables.

[0003] Existing technical solutions have shortcomings. Human experience relies on individual cognition, which can easily overlook better disassembly paths and is difficult to adapt to the structural variations of different battery models. Fixed rules are only sorted by component location or type, without integrating dynamic verification of tool action characteristics and safety rules, resulting in sequences that often have problems such as tool mismatch and high-risk steps being placed at the beginning. At the same time, existing evaluations often focus on single indicators such as disassembly time, ignoring the quantification of safety risks, the accumulation of resource consumption, and the component integrity recovery rate. This may lead to safety accidents, excessive tool wear, or damage to high-value components when the recommended sequences are actually executed.

[0004] This invention aims to address the shortcomings of existing technologies in generating disassembly sequences under complex constraints, such as a lack of intelligent exploration capabilities and a lack of quantitative fusion mechanisms for multi-objective evaluation. It requires dynamically generating candidate sequences covering multiple possibilities within a simulation environment, simultaneously collecting multi-dimensional data on tools, time, risk, and resources, and establishing a comprehensive evaluation system that balances efficiency, safety, cost, and recycling to address the high complexity and multi-objective conflicts in battery disassembly. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; To this end, the present invention proposes an intelligent dismantling sequence planning method for power batteries based on artificial intelligence, including: Initialize the disassembly sequence simulation environment by loading the pre-built disassembly action constraint relationship network, the preset disassembly tool action model, and the disassembly safety rule library into the disassembly sequence simulation environment; In the disassembly sequence simulation environment, a reinforcement learning agent is used to perform Monte Carlo tree search. By simulating different component disassembly sequences, multiple candidate disassembly sequences and their corresponding simulation process data are generated. The simulation process data includes the tool selection, time consumption, safety risk score, and resource consumption assessment for each disassembly step. A multi-objective evaluation model is constructed to comprehensively evaluate the simulation process data of each candidate disassembly sequence and calculate the comprehensive score of each candidate disassembly sequence. The evaluation dimensions of the multi-objective evaluation model include disassembly efficiency, safety, resource cost, and component integrity recovery rate. The candidate disassembly sequence with the highest comprehensive score is selected as the recommended disassembly sequence, and the detailed steps, tool list, and risk warnings of the recommended disassembly sequence are output as an executable disassembly process guidance document.

[0006] Furthermore, the method for constructing the decomposition action constraint relationship network includes: The external image data and internal structure scan data of the power battery pack to be disassembled are acquired by a visual acquisition device, and the data are fused to generate a three-dimensional structural model data of the battery pack; wherein, the power battery pack to be disassembled includes battery modules, connectors, housing and wiring harnesses; The battery pack 3D structural model data is identified and labeled using a pre-set battery type knowledge base. The labeling content includes the type, material, connection relationship and potential hazard level of the component, and generates a battery pack structure knowledge graph with semantic tags. Based on the knowledge graph of battery pack structure with semantic tags, a disassembly action constraint relationship network is constructed. In the disassembly action constraint relationship network, the nodes represent disassembleable components, and the edges represent the physical connections and disassembly order dependencies between components.

[0007] Furthermore, the step of using a preset battery type knowledge base to identify and label components in the three-dimensional structural model data of the battery pack includes: The battery pack 3D structural model data is matched with the standard battery pack 3D template in the battery type knowledge base by point cloud registration and feature comparison. Identify the regions in the three-dimensional structural model data of the battery pack that match the features of the standard template, label the regions as standard components, and inherit the type, material, and potential hazard level information of the standard components in the standard template; For feature regions that fail to match the standard template, their component type and material properties are inferred through geometric shape analysis, material reflection characteristic analysis, and combined with common component features in the knowledge base, and then labeled. Analyze the spatial adjacency relationships and connector distribution among all labeled components, and label the physical connection relationships between components, including bolted connections, welding, plug-in connections, and adhesive connections. The labeling information of all components is associated and stored to generate the battery pack structure knowledge graph with semantic tags.

[0008] Furthermore, the construction of a disassembly action constraint relationship network based on the battery pack structure knowledge graph with semantic tags includes: Each component in the battery pack structure knowledge graph with semantic tags is used as a network node; Analyze the physical connections between components in the battery pack structure knowledge graph. If there is a direct physical connection between two components, then establish an undirected edge between the nodes representing these two components. Based on the principle of physical feasibility of disassembly operations, a disassembly sequence dependency rule is defined; if the disassembly of component A is a prerequisite for contacting or disassembling component B, then a directed edge from A to B is established between node A and node B. Each edge in the network is assigned a weight attribute, which is calculated based on the type of connection, the connection strength, and the hazard of the materials involved. Integrate all nodes, edges, and weight attributes to form the decomposition action constraint relationship network.

[0009] Furthermore, the initialization of the disassembly sequence simulation environment includes loading the pre-constructed disassembly action constraint relationship network, the preset disassembly tool action model, and the disassembly safety rule library into the disassembly sequence simulation environment, including: A virtual three-dimensional space is created as the basic scene of the disassembly sequence simulation environment, and the three-dimensional structural model data of the battery pack is imported into the basic scene; In the disassembly sequence simulation environment, a disassembly tool action model is established for each type of disassembly tool. The disassembly tool action model defines the tool's operating range, force type, effectiveness for specific connection relationships, and operation time parameters. The disassembly safety rule base is loaded into the simulation environment. The disassembly safety rule base contains safety operating procedures, risk triggering conditions and response strategy logic for components of different hazard levels and different operating steps. The disassembly action constraint relationship network is mapped onto the basic scene of the simulation environment, so that it is associated with the spatial position of the three-dimensional model, forming an interactive simulation environment with spatial constraints.

[0010] Furthermore, in the disassembly sequence simulation environment, the Monte Carlo tree search performed using a reinforcement learning agent includes: The state space of a reinforcement learning agent is defined as a subgraph of the set of currently disassembled parts, the set of remaining parts, and the constraint relationship network of the current disassembly action; The action space of the reinforcement learning agent is defined as selecting a component from the current set of detachable components as the next disassembly target; the current set of detachable components is determined according to the disassembly action constraint relationship network. Starting from the initial state, the reinforcement learning agent performs Monte Carlo tree search simulations of multiple decomposition sequences in the simulation environment by selecting actions, simulating the execution of decomposition steps, and updating the state and environmental feedback in a loop. Each simulation process records the generated disassembly step sequence, and the simulation environment calculates and generates the tool selection, time consumption, safety risk score and resource consumption assessment data in real time during the simulation process, which together constitute the simulation process data; When the simulation reaches the termination state or the preset simulation depth, the search and simulation of a candidate disassembly sequence is completed.

[0011] Furthermore, the construction of a multi-objective evaluation model to comprehensively evaluate the simulation process data of each candidate decomposition sequence includes: Define a disassembly efficiency evaluation sub-model, which calculates an efficiency score based on the total time consumption and number of steps of the candidate disassembly sequence; A safety evaluation sub-model is defined, which evaluates the cumulative risk and peak risk level of the entire sequence based on the historical safety risk score in the simulation process data, and calculates the safety score. Define a resource cost evaluation sub-model, which calculates the resource cost score based on the tool usage cost, energy consumption assessment and auxiliary material consumption assessment in the simulation process data. Define a component integrity recovery rate evaluation sub-model. The component integrity recovery rate evaluation sub-model evaluates the integrity rate of key recyclable components based on the component damage situation predicted during the simulation process and calculates the recovery rate score. A preset weight coefficient is assigned to the score of each evaluation sub-model, and the weighted sum is used to obtain the comprehensive score of each candidate decomposition sequence.

[0012] Furthermore, the step of outputting the detailed steps, tool list, and risk warnings of the recommended disassembly sequence into an executable disassembly process guidance document includes: For each disassembly step in the recommended disassembly sequence, generate a detailed process card in order; Each process card includes the identification of the component to be disassembled, the model and parameters of the required disassembly tools, a detailed description of the operating method, the estimated time, safety precautions, and checkpoints indicating the completion status of the previous steps. All process cards, tool lists, general diagrams, and dynamic replanning trigger conditions are integrated and arranged into a structured document or interactive electronic manual to form the executable disassembly process guidance document.

[0013] Furthermore, the method also includes: During actual disassembly, data from the actual disassembly process is continuously collected by sensors while executing the disassembly process guidance document generated according to the recommended disassembly sequence. The actual disassembly process data is compared with the simulation process data of the corresponding steps in real time. If the comparison deviation exceeds the preset fault tolerance threshold, the dynamic replanning process of the disassembly sequence is triggered.

[0014] Furthermore, the dynamic replanning process for triggering the disassembly sequence includes: Suspend the implementation of the current dismantling process guidelines; Based on the results of the disassembly steps that have been actually completed, update the three-dimensional structural model data of the battery pack and the knowledge graph of the battery pack structure with semantic tags to form the updated knowledge graph of the current remaining part. Based on the updated knowledge graph, the remaining disassembly action constraint relationship network is reconstructed and loaded into the updated simulation environment. Re-execute the Monte Carlo tree search, candidate sequence generation, and multi-objective evaluation process to generate new recommended decomposition sequences; Based on the new recommended disassembly sequence, a disassembly process guidance document is generated for subsequent steps, prompting operators to continue.

[0015] Compared with the prior art, the beneficial effects of the present invention are: In the disassembly sequence simulation environment, a reinforcement learning agent is used to drive Monte Carlo tree search to generate candidate sequences. The reinforcement learning agent learns disassembly strategies through interaction with the environment, while the Monte Carlo tree search explores the potential benefits of different disassembly orders through random simulation and backtracking statistics. Combined within a pre-loaded framework of disassembly action constraint relationships, tool action models, and a safety rule base, the simulation of component disassembly sequences simultaneously records data such as tool selection suitability, single-step and cumulative time consumption, real-time safety risk scores calculated based on the safety rule base, and resource consumption assessment data including tool energy consumption and auxiliary material usage. This dynamic generation method overcomes the limitations of human experience, covers more feasible paths, and combines dynamic screening steps with tool and safety rules to make candidate sequences more closely match actual constraints. The simulation data fully reflects the execution details, providing rich evidence for subsequent optimization.

[0016] A multi-objective evaluation model is constructed, encompassing four dimensions: dismantling efficiency, safety, resource cost, and component integrity recovery rate. The dismantling efficiency dimension is directly linked to simulation time consumption data; the safety dimension weights and accumulates the risk scores of each step; the resource cost dimension integrates the assessment values ​​of tool wear and auxiliary material usage; and the component integrity recovery rate dimension calculates the percentage of recyclable intact components using simulation data on dismantling stress and deformation. The data from these four dimensions are quantified and fused to form a comprehensive score. This model overcomes the bias of single-indicator evaluation, comprehensively balances multiple objective conflicts, and avoids extreme choices that prioritize efficiency over safety or recovery over resources. It achieves a better balance across multiple dimensions in the recommended sequence, improving the feasibility and overall benefits of actual dismantling. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the intelligent dismantling sequence planning method for power batteries based on artificial intelligence as described in this invention. Figure 2 A flowchart for constructing the network of action constraint relationships; Figure 3 A graph showing the comparison between the simulation and actual time consumption of the intelligent disassembly sequence of power batteries and the monitoring of the fault tolerance threshold; Figure 4 A chart showing the correlation between dismantling process time and safety risk trends; Figure 5 A heatmap showing the multi-dimensional deviation rate during the disassembly of power batteries. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1The overall implementation scheme of the intelligent dismantling sequence planning method for power batteries based on artificial intelligence includes the following steps: initializing the dismantling sequence simulation environment, loading the pre-constructed dismantling action constraint relationship network, the preset dismantling tool action model, and the dismantling safety rule library into the dismantling sequence simulation environment; in the dismantling sequence simulation environment, using a reinforcement learning agent to perform Monte Carlo tree search, generating multiple candidate dismantling sequences and their corresponding simulation process data by simulating different component dismantling sequences, the simulation process data including tool selection, time consumption, safety risk score, and resource consumption assessment for each dismantling step; constructing a multi-objective evaluation model to comprehensively evaluate the simulation process data of each candidate dismantling sequence, calculating the comprehensive score of each candidate dismantling sequence, the evaluation dimensions of the multi-objective evaluation model including dismantling efficiency, safety, resource cost, and component integrity recovery rate; selecting the candidate dismantling sequence with the highest comprehensive score as the recommended dismantling sequence, and outputting the detailed steps, tool list, and risk warnings of the recommended dismantling sequence as an executable dismantling process guidance document.

[0020] See Figure 2 In one embodiment of the present invention, the method for constructing a disassembly action constraint relationship network acquires the appearance image data and internal structure scan data of the power battery pack to be disassembled through a visual acquisition device, and performs data fusion processing to generate three-dimensional structural model data of the battery pack. The power battery pack to be disassembled includes battery modules, connectors, shells, and wiring harnesses. The three-dimensional structural model data of the battery pack is used to identify and label components using a preset battery type knowledge base. The labeling content includes the type, material, connection relationship, and potential hazard level of the components, generating a battery pack structure knowledge graph with semantic tags. Based on the battery pack structure knowledge graph with semantic tags, a disassembly action constraint relationship network is constructed. The nodes in the disassembly action constraint relationship network represent disassembleable components, and the edges represent the physical connection and disassembly order dependency relationship between components. The component identification and labeling process involves point cloud registration and feature comparison between the battery pack 3D structural model data and the standard battery pack 3D template in the battery type knowledge base; identifying regions in the battery pack 3D structural model data that match the features of the standard template, labeling these regions as standard components, and inheriting the type, material, and potential hazard level information of the standard components in the standard template; for feature regions that do not match the standard template, inferring their component type and material properties through geometric shape analysis, material reflection characteristic analysis, and combining general component features in the knowledge base, and labeling them; analyzing the spatial adjacency relationships and connector distribution among all labeled components, labeling the physical connection relationships between components, including bolted connections, welding, plugging, and bonding; and associating and storing the labeling information of all components to generate the battery pack structure knowledge graph with semantic tags.

[0021] The construction of the disassembly action constraint network uses each component in the battery pack structure knowledge graph with semantic tags as a network node; analyzes the physical connection relationships between components in the battery pack structure knowledge graph, and establishes an undirected edge between the nodes representing the two components if there is a direct physical connection between them; defines a disassembly sequence dependency rule based on the physical feasibility principle of the disassembly operation, and establishes a directed edge from node A to node B if the disassembly of component A is a prerequisite for contacting or disassembling component B; assigns a weight attribute to each edge in the network, and the weight attribute is calculated based on the type of connection relationship, connection strength, and the hazard of the materials involved; integrates all nodes, edges, and weight attributes to form the disassembly action constraint network.

[0022] In practical implementation, the method for constructing the disassembly action constraint relationship network is applied to an exemplary power battery pack to be disassembled. This power battery pack is a ternary lithium-ion battery pack used in vehicles, containing twelve battery modules, busbars and high-voltage connectors, an aluminum alloy upper and lower shell, and high-voltage and signal wiring harnesses. A high-resolution appearance image data of the six surfaces of the power battery pack to be disassembled is acquired using a visual acquisition device consisting of a high-precision 3D laser scanner and a multispectral industrial camera. Internal structural point cloud scan data is also obtained by scanning through the shell. The texture information of the appearance image data and the spatial coordinate information of the internal structural point cloud scan data are fused to generate a 3D structural model data of the battery pack with color and material texture. A pre-set battery type knowledge base is used to identify and label the components of the 3D structural model data of the battery pack. This knowledge base stores standard 3D templates and semantic information for various mainstream power battery pack models. The 3D structural model data of the battery pack is then matched with the corresponding standard battery pack 3D templates in the battery type knowledge base using point cloud registration and feature comparison. The registration process uses an iterative nearest-point algorithm to align the coordinate spaces. Identify areas in the 3D structural model data of the battery pack that match the features of the standard template, such as battery module units with consistent size and arrangement, and standard bolt holes. Label these areas as standard components and inherit the type labels of "battery module" and "bolt", the material properties of "lithium-ion battery" and "steel", and the potential hazard level information of "high risk of electric shock" and "low mechanical risk" from the standard template.

[0023] For feature areas that do not match the standard template, their geometric features are analyzed, such as the geometric shape analysis of long, soft fillers, to determine that they are buffer pads. By analyzing the material reflectivity, such as the reflectivity of a high-gloss insulating sheath, and combining this with the general component features of "high-voltage harnesses" in the knowledge base, the component type is inferred to be a high-voltage harness, and the material properties are labeled as "silicone and copper." The spatial adjacency relationships and connector distributions between all labeled components are analyzed to label the physical connections between components, such as the "bolted connection" between the battery module and the busbar, and the "snap-fit ​​and adhesive" relationship between the top cover and the bottom shell. Finally, the types, materials, potential hazard levels, and physical connections between all components are associated and stored to generate a battery pack structure knowledge graph with semantic tags. Based on the semantically tagged battery pack structure knowledge graph, a disassembly action constraint relationship network is constructed. Each component in the battery pack structure knowledge graph is used as a network node, for example, generating "top cover_node," "battery module A_node," and "high-voltage harness X_node." Analyze the physical connections between components in the battery pack structure knowledge graph. If there is a direct physical connection between two components, an undirected edge is created between the nodes representing these two components. For example, an undirected edge is created between "Top Cover_Node" and "Lower Shell_Node" to represent their snap-fit ​​connection. Define disassembly sequence dependency rules based on the physical feasibility principle of disassembly operations. If the removal of component A is a prerequisite for contacting or removing component B, a directed edge is created between node A and node B, pointing from A to B. For example, if the top cover must be removed to access the internal battery module, a directed edge is created from "Top Cover_Node" to "Battery Module A_Node". Assign a weight attribute to each edge in the network. The weight attribute is calculated based on the type of connection, the connection strength coefficient, and the hazard coefficient of the materials involved, specifically quantified using the following formula:

[0024] in, Represents the connection node With nodes The final weight of the edge, This represents the baseline operational difficulty coefficient assigned based on the type of connection (e.g., welding, bonding, bolting). This represents the strength coefficient calculated based on connection strength (such as torque value, bond area). This represents a hazard factor calculated based on the highest hazard level of the components involved in the connection. , , These are preset adjustment coefficients for operational difficulty, connection strength, and risk. All nodes, edges, and their calculated weight attributes are integrated to form a network of disassembly action constraints for subsequent simulations.

[0025] In one embodiment of the present invention, a virtual three-dimensional space is created as the base scene of the initial disassembly sequence simulation environment, and the three-dimensional structural model data of the battery pack is imported into the base scene. In the disassembly sequence simulation environment, a disassembly tool action model is established for each type of disassembly tool. The disassembly tool action model defines the tool's operating range, force type, validity for specific connection relationships, and operation time parameters. The disassembly safety rule library is loaded into the simulation environment. The disassembly safety rule library contains safety operating procedures, risk triggering conditions, and response strategy logic for components with different risk levels and different operating steps. The disassembly action constraint relationship network is mapped to the base scene of the simulation environment, so that it is associated with the spatial position of the three-dimensional model, forming an interactive simulation environment with spatial constraints.

[0026] In practical implementation, the integrated simulation software platform possesses a physics engine and graphics rendering capabilities. A virtual 3D space is created as the basic scene for the disassembly sequence simulation environment. This basic scene defines the gravity direction, lighting conditions, and basic collision detection rules. Pre-generated 3D structural model data of the battery pack, acquired and processed through visual analysis, is imported into the basic scene. This data is loaded in the form of a mesh model and material textures, accurately reproducing the geometry and surface texture of the battery pack's outer shell, modules, wiring harnesses, and other components, giving it a visualized 3D entity representation in the virtual space.

[0027] In the disassembly sequence simulation environment, disassembly tool motion models are established for each type of disassembly tool. For example, corresponding disassembly tool motion models are established for "torque wrench," "insulated screwdriver," "laser cutting head," and "robotic arm suction cup." The disassembly tool motion model defines the tool's operating range, such as the effective suction radius and maximum load of the robotic arm suction cup; defines the type of force applied, such as the torque wrench applying rotational torque and the laser cutting head applying thermal cutting energy; defines the tool's effectiveness for specific connection relationships, such as the insulated screwdriver being effective only for bolted and screw connections, but ineffective for welded connections; and defines operation time parameters, which are derived through a calculation model. For example, the simulation time for disassembling a bolted connection is... The calculation formula is as follows:

[0028] in: This represents the baseline operating time for this type of connection under standard tightening torque. This represents the actual tightening torque coefficient of the current connection. The standard tightening torque coefficient is represented by κ, which is a proportional factor indicating the sensitivity of adjustment time to torque deviation. In practical implementation, the disassembly safety rule base is loaded into the simulation environment. The disassembly safety rule base contains safety operating procedures, risk triggering conditions, and response strategy logic for components with different hazard levels and different operating steps. For example, the rule base stipulates that when the simulation operation involves a component marked as "high risk of electric shock," the "voltage detection" and "discharge" steps must first be simulated in the simulation. If physical contact is performed directly without these steps, an "electric shock risk" event is triggered and a high-risk score is recorded. The rule base also defines the logical conditions for triggering a "short circuit risk" event when the wrong tool is selected, such as using metal tools to directly operate live components.

[0029] In some embodiments, the disassembly action constraint relationship network is mapped to the basic scene of the simulation environment. The mapping process binds each node in the disassembly action constraint relationship network to its corresponding 3D model component based on the spatial coordinates of each component in the 3D structural model data of the battery pack. This ensures that the connections and dependencies between nodes are consistent with the spatial occlusion and assembly relationships of the 3D model components. It can be understood that the interactive simulation environment formed after mapping requires the reinforcement learning agent to not only logically adhere to the sequential dependencies defined by the disassembly action constraint relationship network when simulating disassembly actions, but also to handle the spatial interference problem between tools and components in 3D space. Optionally, the simulation environment will calculate and provide feedback in real time on the rationality of tool selection, time consumption, triggered safety risk scores, and resource consumption assessment data based on the tool energy consumption and wear model for each simulated disassembly step, based on the disassembly tool action model and disassembly safety rule base. In some embodiments, the basic scene of the virtual 3D space can also be configured with multiple fixed observation viewpoints and measurement coordinate systems, facilitating observation of the disassembly process from different angles and accurate measurement of parameters such as component displacement and tool angle during the simulation. It can be understood that a complete simulation environment initialization is the foundation for subsequent reliable Monte Carlo tree search and simulation.

[0030] In one embodiment of the present invention, a reinforcement learning agent is used for Monte Carlo tree search. The state space of the reinforcement learning agent is defined as a subgraph of the current set of disassembled parts, the set of remaining parts, and the constraint relationship network of the current disassembly action. The action space of the reinforcement learning agent is defined as selecting a part from the current set of disassembleable parts as the next disassembly target. The current set of disassembleable parts is determined according to the disassembly action constraint relationship network. Starting from the initial state, the reinforcement learning agent performs multiple Monte Carlo tree search simulations of disassembly sequences in the simulation environment by selecting actions, simulating the execution of disassembly steps, updating the state and environmental feedback in a loop. Each simulation process records the generated disassembly step sequence, and the simulation environment calculates and generates the tool selection, time consumption, safety risk score, and resource consumption assessment data in real time during the simulation process, which together constitute the simulation process data. When the simulation reaches the termination state or reaches the preset simulation depth, the search and simulation of a candidate disassembly sequence is completed.

[0031] In practice, the Monte Carlo tree search process using a reinforcement learning agent unfolds within an initialized disassembly sequence simulation environment. The state space of the reinforcement learning agent is defined as a data structure containing three pieces of information: the first is the set of currently disassembled parts, which records a list of unique identifiers for parts removed from the simulation environment; the second is the set of remaining parts, which records a list of unique identifiers for parts not yet disassembled; and the third is a subgraph of the current disassembly action constraint network, generated by removing the nodes corresponding to disassembled parts and their associated edges from the original disassembly action constraint network, visually representing the connections and dependencies between the remaining parts. The action space of the reinforcement learning agent is defined as selecting a part from the current set of disassembleable parts as the next disassembly target. The current set of disassembleable parts is determined by querying the disassembly action constraint network subgraph in the current state, specifically identifying the parts represented by all nodes with an in-degree of zero in this subgraph, which have no undisassembled preceding dependent parts in the current simulation state.

[0032] Starting from an initial state, the reinforcement learning agent performs a Monte Carlo tree search simulation of multiple decomposition sequences in a simulated environment, iteratively selecting actions, simulating decomposition steps, and updating the state and environmental feedback. Each action selection is performed within the framework of Monte Carlo tree search, involving tree traversal, node expansion, and backtracking evaluation. For each node representing a state in the tree, its child nodes correspond to all legal decomposition actions starting from that state. The reinforcement learning agent selects actions according to a policy that balances the utilization of high-average-reward actions with the exploration of underexplored actions. The policy's evaluation score... Calculation formula:

[0033] in: This represents the average cumulative reward from the historical simulations starting from this action node. This represents the total number of visits to its parent state node. λ represents the number of times the action node itself has been visited, and λ is a hyperparameter that controls the exploration weight.

[0034] In practical implementation, when a reinforcement learning agent selects a disassembly action in the simulation, the simulation environment, based on the disassembly tool action model and the disassembly safety rule base, simulates the specific disassembly operation of the selected component. The operation process considers tool compatibility, spatial interference, and safety rules. After the simulation is completed, the simulation environment updates its state, moving the disassembled component from the remaining set to the disassembled set and updating the disassembly action constraint relationship network subgraph. Simultaneously, the environment provides an immediate reward signal for this step, which integrates the reciprocal of the step's time consumption, the negative value of the safety risk score, and the negative value of the resource consumption assessment. Each simulation process records the generated disassembly step sequence, and during the simulation, the simulation environment calculates and generates tool selection, time consumption, safety risk score, and resource consumption assessment data in real time, collectively constituting the simulation process data. The tool selection data records the tool model and usage parameters automatically matched for each disassembly step in the simulation; the time consumption data records the virtual time consumed by each step of the simulation, which is calculated and accumulated based on the operation time consumption parameters defined in the disassembly tool action model; the safety risk score data records the risk score accumulated for each step of the operation due to violations of rules in the disassembly safety rule base; and the resource consumption assessment data records the virtual energy consumption and wear-and-tear cost caused by tool use. In some embodiments, the data recording of the simulation process is stored in a time-series database to ensure that the data of each simulation step strictly corresponds to the step order of the disassembly sequence. When the simulation reaches the termination state or reaches the preset simulation depth, the search and simulation of a candidate disassembly sequence is completed. The termination state is defined as the state where the set of remaining parts is empty. The preset simulation depth can be set to a sufficiently large upper limit of the number of steps to prevent the simulation from looping infinitely. It can be understood that through a large number of such simulations, Monte Carlo tree search can explore a variety of different disassembly sequence branches. Optionally, during the simulation, for disassembly actions involving high-voltage or flammable components, the simulation environment will provide enhanced risk warnings and significantly increase the safety risk score for that step based on a disassembly safety rule base. This will affect the immediate reward obtained by the reinforcement learning agent, thereby guiding it to learn to avoid high-risk operation sequences. In some embodiments, the root node of the Monte Carlo tree search is initialized at the start of each planned task and continues to simulate until a preset total computation time or number of simulations is reached, thereby generating a set containing multiple candidate disassembly sequences and their complete simulation process data. It can be understood that the reinforcement learning agent explores and learns through Monte Carlo tree search with the goal of discovering a better disassembly sequence under comprehensive evaluation.

[0035] See Figure 3 This study presents a comparison of the time consumption of simulated and actual disassembly data across 12 disassembly steps, along with monitoring of fault tolerance thresholds. The blue curve represents the disassembly time of the simulated data, while the red curve represents the actual disassembly time. The green dashed lines represent the upper and lower fault tolerance thresholds, respectively. The red-filled area indicates the risk zone where the actual time exceeds the upper threshold, and the yellow-filled area indicates the abnormal zone where the actual time is below the lower threshold. Specifically, in step 1, the actual disassembly time is below the lower fault tolerance threshold, falling within the yellow area exceeding the lower threshold, reflecting that the actual operational efficiency in the initial stage is higher than the simulation expectation. From steps 2 to 6, the actual time continuously increases, reaching a peak in step 6, and the actual time in this stage remains within the red area exceeding the upper threshold, indicating that the actual disassembly difficulty in this stage exceeds the simulation prediction, requiring a dynamic replanning warning. In steps 7 to 10, the data fluctuations return to the fault tolerance threshold range, while in step 12, the actual time falls back to near the lower threshold. Through time-series time comparison and threshold monitoring, dynamic verification between simulation and actual data is achieved during the disassembly process, providing crucial time-series data support for the dynamic replanning of the disassembly sequence.

[0036] In one embodiment of the present invention, a multi-objective evaluation model is constructed, defining a disassembly efficiency evaluation sub-model, which calculates an efficiency score based on the total time consumption and number of steps of the candidate disassembly sequence; defining a safety evaluation sub-model, which assesses the cumulative risk and peak risk level of the entire sequence based on the historical safety risk score records in the simulation process data, and calculates a safety score; defining a resource cost evaluation sub-model, which calculates a resource cost score based on tool usage costs, energy consumption assessment, and auxiliary material consumption assessment in the simulation process data; and defining a component integrity recovery rate evaluation sub-model, which assesses the integrity rate of key recyclable components based on the component damage predicted during the simulation process, and calculates a recovery rate score; and assigning preset weight coefficients to the scores of each evaluation sub-model, and weighted summing to obtain the comprehensive score of each candidate disassembly sequence. The output disassembly process guidance document will recommend each disassembly step in the disassembly sequence and generate detailed process cards in order. Each process card includes the identification of the component to be disassembled, the model and parameters of the required disassembly tools, a description of the specific operating method, the estimated time, safety precautions, and checkpoints for the completion status of the previous steps. All process cards, tool lists, general diagrams, and descriptions of dynamic replanning trigger conditions will be integrated and arranged into a structured document or interactive electronic manual to form the executable disassembly process guidance document.

[0037] In practical implementation, a multi-objective evaluation model is constructed to comprehensively evaluate the simulation process data of each candidate disassembly sequence. Taking the disassembly plan of a certain type of power battery pack containing twenty main components as an example, Monte Carlo tree search simulation generates three candidate disassembly sequences with differences. The multi-objective evaluation model will quantitatively score these three candidate disassembly sequences. A disassembly efficiency evaluation sub-model is defined, which calculates the efficiency score based on the total time consumption and the number of steps of the candidate disassembly sequence. The total time consumption is directly derived from the sum of the time consumption of all steps in the simulation process data, and the number of steps is the total number of steps contained in the disassembly sequence. The efficiency score is calculated based on these factors. The calculation uses the following formula:

[0038] in: This represents the total time consumed. Represents the number of steps. It is the efficiency normalization coefficient.

[0039] Define a safety evaluation sub-model. Based on the historical safety risk scores in the simulation process data, this sub-model assesses the cumulative risk and peak risk level of the entire sequence, calculates the safety score, where the cumulative risk value is the sum of the safety risk scores for all steps, and the peak risk level is the maximum value among all safety risk scores. The calculation formula is:

[0040] in: Represents the cumulative risk value. Represents the peak risk level. It is the safety normalization coefficient.

[0041] Define a resource cost evaluation sub-model. This sub-model calculates a resource cost score based on tool usage costs, energy consumption assessments, and auxiliary material consumption assessments from simulation data. Tool usage costs are calculated based on the tool model, its unit time usage rate, and usage time. Energy consumption assessments are based on tool power and usage time. Auxiliary material consumption, such as wiping cloths and insulating tape, is charged based on usage records from the simulation. The resource cost score is then calculated. The calculation formula is:

[0042] in: , , These represent tool usage costs, energy costs, and material consumption costs, respectively. It is the cost normalization coefficient.

[0043] A component integrity recovery rate evaluation sub-model is defined. Based on the component damage predicted during the simulation, this sub-model assesses the integrity rate of key recyclable components and calculates the recovery rate score. The simulation process predicts and marks the component scratches, deformations, and structural damage that may occur at each disassembly step, and the recovery rate score is calculated accordingly. The calculation formula is:

[0044] in: This represents the number of critical recyclable parts predicted to be intact after the simulation. This represents the total number of critical recyclable components.

[0045] Each evaluation sub-model's score is assigned a pre-defined weight coefficient, and the scores are weighted and summed to obtain the comprehensive score for each candidate decomposition sequence. (Comprehensive Score) The calculation formula is:

[0046] in: , , , These are the efficiency scores of the sub-models used for efficiency evaluation. Safety evaluation sub-model safety score Resource cost evaluation sub-model resource cost score Component integrity recovery rate evaluation sub-model recovery rate score The weight coefficients, and satisfying An example comparison of evaluation results is shown in Table 1 below: Table 1: Example of Comparison of Multi-Objective Evaluation Results for Candidate Decomposition Sequences

[0047] In the specific implementation, the candidate disassembly sequence with the highest comprehensive score was selected as the recommended disassembly sequence. According to the comparison results in Table 1, candidate sequence B received the highest comprehensive score due to its superior performance in safety and cost control, and was selected as the recommended disassembly sequence. The detailed steps, tool list, and risk warnings of the recommended disassembly sequence were output as an executable disassembly process guidance document. The output process first generated detailed process cards for each disassembly step in the recommended disassembly sequence in sequence. Each process card included the identifier of the part to be disassembled, such as "Part ID: BM-03"; the model and parameters of the required disassembly tool, such as "Tool: Insulated torque wrench, preset torque: 15 N·m"; a specific description of the operation method, such as "Vertically align the bolt, slowly apply torque to the preset value, and rotate counterclockwise three and a half turns"; the estimated time, such as "Estimated time: 45 seconds"; safety precautions, such as "Note: Before operation, confirm that the multimeter voltage is below 5V"; and checkpoints for the completion status of the previous steps, such as "Checkpoint: Confirm that the top cover (Part ID: TOP-01) has been removed". In practice, all process cards, tool lists, general schematic diagrams, and dynamic replanning trigger conditions are integrated and compiled into a structured document or interactive electronic manual, forming an executable disassembly process guidance document. The tool list summarizes the models, quantities, and uses of all tools used in the process cards. The general schematic diagram marks the location and number of the component corresponding to each disassembly step. The dynamic replanning trigger conditions specify under what deviations the replanning process needs to be paused and restarted. In some embodiments, the process cards are generated using an automatic layout algorithm to ensure clear correspondence between text and graphics. Optionally, the interactive electronic manual allows operators to click on components on the 3D model to access the corresponding process card. In some embodiments, weighting coefficients... , , , The configuration can be tailored to the specific disassembly task. For example, in scenarios where safety is emphasized during the disassembly of retired batteries, the safety score of the safety evaluation sub-model can be improved. Weighting coefficients The assignment of values ​​indicates that the multi-objective evaluation model comprehensively weighs efficiency, safety, cost, and recycling dimensions through quantitative methods. Optionally, the prediction of component damage in the component integrity recovery rate evaluation sub-model is calculated based on the component material properties and tool force model defined in the simulation physics engine. The final generated disassembly process guidance document provides operators with standardized and visualized operational guidelines.

[0048] See Figure 4Using the disassembly process steps as the horizontal axis, the dynamic changes of two-dimensional quantitative indicators are presented simultaneously: the left vertical axis represents the estimated time (seconds) for a single process, and the right vertical axis represents the safety risk score for a single process (1-9 points, with higher scores indicating higher risk levels). From the perspective of process sequence characteristics, steps 2, 4, 7, and 10 are key risk nodes. Step 2 reaches its peak safety risk score, corresponding to the disassembly of high-risk components. Step 10's safety risk score and estimated time both reach their global peak, reflecting that this step involves the final disassembly of high-voltage wiring harnesses or modules, combining high operational complexity with high safety risks. In terms of time consumption, steps 4 and 7 show significant peaks in estimated time consumption, corresponding to the disassembly process of housings with dense bolt connections; step 5 represents the lowest point in time consumption across the entire sequence, corresponding to the disassembly of low-voltage wiring harnesses that only require simple plugging and unplugging. The fluctuations in safety risk scores show a local positive correlation with the expected time consumption. In particular, in highly complex processes, the increase in operation time is directly related to the extension of risk exposure time. This trend provides key time-series risk basis for the dynamic replanning of subsequent disassembly sequences.

[0049] In one embodiment of the present invention, during actual disassembly based on the disassembly process guidance document generated according to the recommended disassembly sequence, sensors continuously collect actual disassembly process data; the actual disassembly process data is compared in real time with the simulation process data of the corresponding steps; if the comparison deviation exceeds a preset fault tolerance threshold, a dynamic replanning process for the disassembly sequence is triggered. The dynamic replanning process suspends the execution of the current disassembly process guidance document; based on the results of the actually completed disassembly steps, the three-dimensional structural model data of the battery pack and the battery pack structure knowledge graph with semantic tags are updated to form an updated knowledge graph for the remaining part; based on the updated knowledge graph, the disassembly action constraint relationship network for the remaining part is reconstructed and loaded into the updated simulation environment; the Monte Carlo tree search, candidate sequence generation, and multi-objective evaluation process are re-executed to generate a new recommended disassembly sequence; based on the new recommended disassembly sequence, a disassembly process guidance document for subsequent steps is generated, and the operator is prompted to continue execution.

[0050] In practice, when performing actual disassembly according to the disassembly process guidance document generated based on the recommended disassembly sequence, operators follow the process cards in the disassembly process guidance document step by step. Simultaneously, sensors installed on the disassembly workbench and tools continuously collect data on the actual disassembly process. This data includes tool force data collected by a 3D force sensor, component separation image sequences collected by a vision sensor, the actual operation time recorded by a timer, and ambient temperature, humidity, and gas concentration data collected by environmental sensors. The actual disassembly process data is compared in real time with the simulation process data for each step. The comparison includes the deviation between the operation time of each step and the simulation's predicted time, the deviation between the actual tool force and the simulation's expected force, and the visual feature differences between the actual component separation shape and the simulation's predicted shape. For each operation step, a comprehensive deviation value is calculated. :

[0051] in: This represents the overall deviation value of the current step. This represents the number of data dimensions being compared. and Representing the first Actual measured values ​​and simulated expected values ​​for each dimension It is the first Preset weighting coefficients for each dimension. If the overall deviation value... If the preset fault tolerance threshold is exceeded, the dynamic replanning process of the disassembly sequence is triggered. The fault tolerance threshold is set differently according to the safety risk level of the operation type. For example, the fault tolerance threshold for the high-voltage harness disassembly step is lower than that for the ordinary housing disassembly step.

[0052] In some embodiments, the dynamic replanning process that triggers the disassembly sequence first suspends the execution of the current disassembly process guidance document, issues a pause command to the operator through a human-machine interface, and prompts a deviation alarm. Based on the results of the disassembly steps that have been actually completed, the 3D structural model data of the battery pack and the battery pack structure knowledge graph with semantic tags are updated to form an updated knowledge graph of the remaining parts. The update process utilizes the latest collected actual disassembly process data, especially the component separation image sequence collected by the visual sensor, and corrects the state of the disassembled components in the original model through 3D reconstruction technology, and confirms whether the actual position and orientation of the remaining components are consistent with the original knowledge graph. If unexpected connection relationships or component damage are found, they are marked in the updated knowledge graph. Optionally, in a specific scenario, when disassembling a component marked as bolted, if the actual force is consistently higher than the simulation expectation and the visual features show that the component has not been separated, the overall deviation value is considered. If a threshold is exceeded, the system determines that there may be unlabeled weld points, triggering dynamic reprogramming. Based on the updated knowledge graph, the remaining disassembly action constraint relationship network is reconstructed and loaded into the updated simulation environment. The reconstruction process follows the same rules as the initial construction, but the network nodes only contain the remaining parts and incorporate newly discovered connection constraints. The Monte Carlo tree search, candidate sequence generation, and multi-objective evaluation process are re-executed to generate a new recommended disassembly sequence. The initial state of this search is the state corresponding to the currently completed disassembly steps. The simulation environment simulates based on the updated model and rules. Based on the new recommended disassembly sequence, a disassembly process guidance document for subsequent steps is generated, prompting the operator to continue. The newly generated disassembly process guidance document starts after the currently interrupted step and includes specialized tool selection and operation instructions for newly discovered situations (such as hidden weld points). It can be understood that the dynamic reprogramming process enables the system to cope with unexpected situations in actual disassembly that do not conform to the initial model. In some embodiments, the comparison between actual disassembly process data and simulation process data is continuous, and the system maintains a real-time updated deviation sequence. Understandably, the preset fault tolerance threshold needs to be calibrated based on a large amount of historical decomposition data. Optionally, during the dynamic replanning process, the system will record the type and reason for the deviation that triggers the replanning, which will be used to subsequently optimize the construction accuracy of the initial knowledge graph and the reliability of the simulation model.

[0053] See Figure 5 The figure presents the distribution of deviation rates for different disassembly steps across various deviation data dimensions. Specifically, the figure uses the disassembly steps (shell disassembly, wire harness separation, connector disassembly, module disassembly, cell extraction, and high-voltage component disassembly) as the horizontal axis and the deviation data dimensions (operation time, tool force, visual features, and environmental data) as the vertical axis. It quantifies the deviation rate of each step-dimensional combination by combining color gradients and numerical annotations. Among the various parameters, the deviation rate of tool force was highest in the high-voltage component disassembly step, reaching 33.33%, while the deviation rate of tool force in the wire harness separation step also reached 20.0%. This indicates that there is a significant difference between the actual disassembly force of the high-voltage component and the wire harness and the simulation expectation, which may be due to the lack of preset connection strength or component structure variations. The deviation rate of operation time in the high-voltage component disassembly step was 15.0%, while the deviation rates of operation time in the module disassembly and cell extraction steps were both 10.0%, reflecting that the actual time consumption of these complex steps fluctuates greatly. The deviation rate of the visual feature dimension was generally in the range of 1.05% to 10.11%, with the highest deviation rate in the high-voltage component disassembly step, indicating that there is a certain difference between the actual performance of the component separation form and the simulation prediction. The deviation rate of the environmental data dimension was generally less than 2%, and the fluctuation between each step was minimal, indicating that the actual data of environmental factors such as temperature, humidity, and gas concentration were highly consistent with the simulation expectation and had low interference with the disassembly process.

[0054] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A power battery intelligent dismantling sequence planning method based on artificial intelligence, characterized in that, include: Initialize the disassembly sequence simulation environment by loading the pre-built disassembly action constraint relationship network, the preset disassembly tool action model, and the disassembly safety rule library into the disassembly sequence simulation environment; In the disassembly sequence simulation environment, a reinforcement learning agent is used to perform Monte Carlo tree search. By simulating different component disassembly sequences, multiple candidate disassembly sequences and their corresponding simulation process data are generated. The simulation process data includes the tool selection, time consumption, safety risk score, and resource consumption assessment for each disassembly step. A multi-objective evaluation model is constructed to comprehensively evaluate the simulation process data of each candidate disassembly sequence and calculate the comprehensive score of each candidate disassembly sequence. The evaluation dimensions of the multi-objective evaluation model include disassembly efficiency, safety, resource cost, and component integrity recovery rate. The candidate disassembly sequence with the highest comprehensive score is selected as the recommended disassembly sequence, and the detailed steps, tool list, and risk warnings of the recommended disassembly sequence are output as an executable disassembly process guidance document.

2. The intelligent dismantling sequence planning method for power batteries based on artificial intelligence as described in claim 1, characterized in that, The method for constructing the decomposition action constraint relationship network includes: The external image data and internal structure scan data of the power battery pack to be disassembled are acquired by a visual acquisition device, and the data are fused to generate a three-dimensional structural model data of the battery pack; wherein, the power battery pack to be disassembled includes battery modules, connectors, housing and wiring harnesses; The battery pack 3D structural model data is identified and labeled using a pre-set battery type knowledge base. The labeling content includes the type, material, connection relationship and potential hazard level of the component, and generates a battery pack structure knowledge graph with semantic tags. Based on the knowledge graph of battery pack structure with semantic tags, a disassembly action constraint relationship network is constructed. In the disassembly action constraint relationship network, the nodes represent disassembleable components, and the edges represent the physical connections and disassembly order dependencies between components.

3. The intelligent dismantling sequence planning method for power batteries based on artificial intelligence as described in claim 2, characterized in that, The step of using a pre-set battery type knowledge base to identify and label components in the 3D structural model data of the battery pack includes: The battery pack 3D structural model data is matched with the standard battery pack 3D template in the battery type knowledge base by point cloud registration and feature comparison. Identify the regions in the three-dimensional structural model data of the battery pack that match the features of the standard template, label the regions as standard components, and inherit the type, material, and potential hazard level information of the standard components in the standard template; For feature regions that fail to match the standard template, their component type and material properties are inferred through geometric shape analysis, material reflection characteristic analysis, and combined with common component features in the knowledge base, and then labeled. Analyze the spatial adjacency relationships and connector distribution among all labeled components, and label the physical connection relationships between components, including bolted connections, welding, plug-in connections, and adhesive connections. The labeling information of all components is associated and stored to generate the battery pack structure knowledge graph with semantic tags.

4. The intelligent dismantling sequence planning method for power batteries based on artificial intelligence as described in claim 2, characterized in that, The construction of a disassembly action constraint relationship network based on the battery pack structure knowledge graph with semantic tags includes: Each component in the battery pack structure knowledge graph with semantic tags is used as a network node; Analyze the physical connections between components in the battery pack structure knowledge graph. If there is a direct physical connection between two components, then establish an undirected edge between the nodes representing these two components. Based on the principle of physical feasibility of disassembly operations, a disassembly sequence dependency rule is defined; if the disassembly of component A is a prerequisite for contacting or disassembling component B, then a directed edge from A to B is established between node A and node B. Each edge in the network is assigned a weight attribute, which is calculated based on the type of connection, the connection strength, and the hazard of the materials involved. Integrate all nodes, edges, and weight attributes to form the decomposition action constraint relationship network.

5. The intelligent dismantling sequence planning method for power batteries based on artificial intelligence as described in claim 1, characterized in that, The initialization of the disassembly sequence simulation environment includes loading a pre-constructed network of disassembly action constraints, a preset disassembly tool action model, and a disassembly safety rule base into the disassembly sequence simulation environment, including: A virtual three-dimensional space is created as the basic scene of the disassembly sequence simulation environment, and the three-dimensional structural model data of the battery pack is imported into the basic scene; In the disassembly sequence simulation environment, a disassembly tool action model is established for each type of disassembly tool. The disassembly tool action model defines the tool's operating range, force type, effectiveness for specific connection relationships, and operation time parameters. The disassembly safety rule base is loaded into the simulation environment. The disassembly safety rule base contains safety operating procedures, risk triggering conditions and response strategy logic for components of different hazard levels and different operating steps. The disassembly action constraint relationship network is mapped onto the basic scene of the simulation environment, so that it is associated with the spatial position of the three-dimensional model, forming an interactive simulation environment with spatial constraints.

6. The intelligent dismantling sequence planning method for power batteries based on artificial intelligence as described in claim 1, characterized in that, In the disassembly sequence simulation environment, the Monte Carlo tree search is performed using a reinforcement learning agent, including: The state space of a reinforcement learning agent is defined as a subgraph of the set of currently disassembled parts, the set of remaining parts, and the constraint relationship network of the current disassembly action; The action space of the reinforcement learning agent is defined as selecting a component from the current set of detachable components as the next disassembly target; the current set of detachable components is determined according to the disassembly action constraint relationship network. Starting from the initial state, the reinforcement learning agent performs Monte Carlo tree search simulations of multiple decomposition sequences in the simulation environment by selecting actions, simulating the execution of decomposition steps, and updating the state and environmental feedback in a loop. Each simulation process records the generated disassembly step sequence, and the simulation environment calculates and generates the tool selection, time consumption, safety risk score and resource consumption assessment data in real time during the simulation process, which together constitute the simulation process data; When the simulation reaches the termination state or the preset simulation depth, the search and simulation of a candidate disassembly sequence is completed.

7. The intelligent dismantling sequence planning method for power batteries based on artificial intelligence as described in claim 1, characterized in that, The construction of a multi-objective evaluation model comprehensively evaluates the simulation process data of each candidate decomposition sequence, including: Define a disassembly efficiency evaluation sub-model, which calculates an efficiency score based on the total time consumption and number of steps of the candidate disassembly sequence; A safety evaluation sub-model is defined, which evaluates the cumulative risk and peak risk level of the entire sequence based on the historical safety risk score in the simulation process data, and calculates the safety score. Define a resource cost evaluation sub-model, which calculates the resource cost score based on the tool usage cost, energy consumption assessment and auxiliary material consumption assessment in the simulation process data. Define a component integrity recovery rate evaluation sub-model. The component integrity recovery rate evaluation sub-model evaluates the integrity rate of key recyclable components based on the component damage situation predicted during the simulation process and calculates the recovery rate score. A preset weight coefficient is assigned to the score of each evaluation sub-model, and the weighted sum is used to obtain the comprehensive score of each candidate decomposition sequence.

8. The intelligent dismantling sequence planning method for power batteries based on artificial intelligence as described in claim 1, characterized in that, The step of outputting the detailed steps, tool list, and risk warnings of the recommended disassembly sequence into an executable disassembly process guidance document includes: For each disassembly step in the recommended disassembly sequence, generate a detailed process card in order; Each process card includes the identification of the component to be disassembled, the model and parameters of the required disassembly tools, a detailed description of the operating method, the estimated time, safety precautions, and checkpoints indicating the completion status of the previous steps. All process cards, tool lists, general diagrams, and dynamic replanning trigger conditions are integrated and arranged into a structured document or interactive electronic manual to form the executable disassembly process guidance document.

9. The intelligent dismantling sequence planning method for power batteries based on artificial intelligence as described in claim 1, characterized in that, The method further includes: During actual disassembly, data from the actual disassembly process is continuously collected by sensors while executing the disassembly process guidance document generated according to the recommended disassembly sequence. The actual disassembly process data is compared with the simulation process data of the corresponding steps in real time. If the comparison deviation exceeds the preset fault tolerance threshold, the dynamic replanning process of the disassembly sequence is triggered.

10. The intelligent dismantling sequence planning method for power batteries based on artificial intelligence as described in claim 9, characterized in that, The dynamic replanning process for triggering the disassembly sequence includes: Suspend the implementation of the current dismantling process guidelines; Based on the results of the disassembly steps that have been actually completed, update the three-dimensional structural model data of the battery pack and the knowledge graph of the battery pack structure with semantic tags to form the updated knowledge graph of the current remaining part. Based on the updated knowledge graph, the remaining disassembly action constraint relationship network is reconstructed and loaded into the updated simulation environment. Re-execute the Monte Carlo tree search, candidate sequence generation, and multi-objective evaluation process to generate new recommended decomposition sequences; Based on the new recommended disassembly sequence, a disassembly process guidance document is generated for subsequent steps, prompting operators to continue.

Images