A Retired Battery Disassembly Scheduling Method Based on Reinforcement Learning

Abstract

The present invention relates to the technical field of resource recycling, and discloses a disassembly scheduling method for retired batteries based on reinforcement learning. The method includes: obtaining the machine state of the disassembly machine, constructing a graph matrix encoding based on the battery disassembly process steps, constructing an MGCN unit for extracting the connectivity and topological structure between different parts of the battery based on the Conv2d k1 block and the Conv2d k3 block, and constructing a multi-graph convolutional network according to the MGCN unit; inputting the machine state and the graph matrix encoding into the multi-graph convolutional network to obtain disassembly weights; obtaining an optimized disassembly path by combining depth search according to the disassembly weights, and completing the disassembly of the retired battery according to the optimized disassembly path, which solves the problem of low efficiency of existing models such as existing reinforcement learning when applied to battery disassembly scheduling.

Description

Technical Field

[0001] The present invention relates to the technical field of resource recycling, and particularly relates to a disassembly scheduling method for retired batteries based on reinforcement learning. Background Art

[0002] With the rapid development of the global electric vehicle industry, the production and usage of power batteries have increased sharply, and the problems of disassembly and recycling of retired batteries have become increasingly prominent. How to efficiently and safely disassemble retired batteries and maximize the recovery of valuable materials has become a technical problem that urgently needs to be solved in the battery recycling industry.

[0003] Currently, in the process of power battery disassembly, models such as reinforcement learning are combined to optimize the battery disassembly path. However, in the process of optimizing the disassembly path by existing models such as reinforcement learning, the relevance of in-depth disassembly is not considered, and the disassembly process is not intelligent and efficient enough to effectively achieve intelligent disassembly, and there is still room for optimization.

[0004] When the reinforcement learning model is applied to optimize the disassembly scheduling, there are problems such as the incentive function design being unable to effectively reflect the multi-stage and multi-interaction dynamics of the disassembly process, insufficient multi-scale feature extraction, and limitations in the optimization path and feedback mechanism. Summary of the Invention

[0005] The present invention provides a disassembly scheduling method for retired batteries based on reinforcement learning to solve the problem of low efficiency when existing models such as reinforcement learning are applied to battery disassembly scheduling.

[0006] To achieve the above object, the present invention is realized through the following technical solutions:

[0007] The present invention provides a disassembly scheduling method for retired batteries based on reinforcement learning, including the following steps:

[0008] Step 1: Obtain the machine state of the disassembly machine, construct a graph matrix encoding based on the battery disassembly process steps, construct an MGCN unit for extracting the connectivity and topological structure between different parts of the battery based on the Conv2d k1 block and the Conv2d k3 block, and construct a multi-graph convolutional network according to the MGCN unit;

[0009] Step 2: Input the machine state and the graph matrix encoding into the multi-graph convolutional network to obtain the disassembly weights;

[0010] Step 3: Obtain an optimized disassembly path by combining the disassembly weights with depth search, and complete the disassembly of the retired battery according to the optimized disassembly path.

[0011] Furthermore, the MGCN unit is a multi-source information fusion feature extraction unit with three inputs and one output. The MGCN unit includes an input layer, a fusion extraction layer, a weighting layer, and an output layer;

[0012] The input layer includes three Conv2d k1 blocks that respectively extract local features of three inputs;

[0013] The fusion extraction layer includes a fusion block and a Conv2d k3 block that extracts spatial relationships;

[0014] The weighting layer includes an SE block;

[0015] The output layer includes a Conv2d k1 block that performs linear transformation and dimension compression and an output block.

[0016] Furthermore, constructing a multi-graph convolutional network based on the MGCN unit includes: constructing an action module and a value module based on the MGCN unit, and constructing a multi-graph convolutional network based on the action module and the value module;

[0017] The action module extracts multi-scale features based on the machine state and the graph matrix encoding to obtain disassembly weights;

[0018] The value module obtains the multi-scale features extracted by the action module and optimizes the disassembly weights in combination with an activation function.

[0019] Furthermore, the action module sequentially includes an MGCN unit for extracting machine state and graph matrix encoding features, a detail extraction Conv2d k3 block for further extracting surface and internal detail features of the machine state and the graph matrix encoding, an MGCN unit for strengthening the machine state and graph matrix encoding features, a detail strengthening Conv2d k3 block for further strengthening the surface and internal detail features of the machine state and the graph matrix encoding, and a Conv2d k1 block for compressing and integrating all the strengthened features.

[0020] Furthermore, the value module is respectively connected to the detail extraction Conv2d k3 block and the detail strengthening Conv2d k3 block of the action module, and integrates and evaluates the obtained multi-scale features.

[0021] Furthermore, the value module includes an acquisition unit for acquiring multi-scale features, an integration unit for integrating the multi-scale features, and an evaluation unit for evaluating the integrated multi-scale features;

[0022] The acquisition unit includes a first channel, a second channel, and a Conv2d k1 block. Both the first channel and the second channel are composed of two Conv2d k3 blocks and are connected to the action module and the integration module. The Conv2d k1 block is set between the first Conv2d k3 block in the connection direction of the first channel, the second channel and the action module;

[0023] The integration unit includes 1 Conv2d k1 block;

[0024] The evaluation unit includes 1 fully connected block and an MGCN unit.

[0025] Furthermore, the value module obtains the multi-scale features extracted by the action module and optimizes the disassembly weights in combination with the activation function, which includes: the value module makes a value prediction on the disassembly weights to be output by the action module based on the multi-scale features, the activation function performs a feedback calculation in combination with the machine state, and uses the difference between the feedback calculation and the value prediction as the objective function to optimize the disassembly weights.

[0026] Furthermore, the machine state includes the average completion time, process waiting time, delay penalty, assembly and disassembly time, and station processing capacity;

[0027] The activation function is represented by the following formula:

[0028] ;

[0029] Among them, represents the activation function, represents the average completion time, represents the process waiting time, represents the delay penalty, represents the assembly and disassembly time, represents the station processing capacity, all represent weight coefficients;

[0030] The average completion time is represented by the following formula:

[0031] ;

[0032] Among them, represents the th time required to complete the disassembly of the battery pack, represents the maximum allowed completion time;

[0033] The process waiting time is represented by the following formula:

[0034] ;

[0035] Among them, Indicates the waiting time for the disassembly process of the th battery pack, and represents the maximum allowable waiting time;

[0036] The delay penalty is expressed by the following formula:

[0037] ;

[0038] where represents the total delay time of disassembling the battery;

[0039] The assembly and disassembly time is expressed by the following formula:

[0040] ;

[0041] where represents the time required to disassemble the battery pack from the machine, and represents the time required to assemble the battery pack onto the machine;

[0042] The processing capacity of the station is expressed by the following formula:

[0043] ;

[0044] where represents the maximum processing capacity of the station, i.e., the time required to disassemble one battery pack, and represents the th battery pack at the th process by the machine, and represents the total number of battery packs to be disassembled.

[0045] Furthermore, constructing a graph matrix encoding based on the battery disassembly process steps includes: constructing a disassembly path according to the battery disassembly process steps in combination with the first predetermined rule, using the numbers of the nodes of the disassembly path as the rows and columns of the graph matrix encoding, and filling the rows and columns of the graph matrix encoding in combination with the node relationships of the disassembly path and the second predetermined rule;

[0046] The first predetermined rule includes: numbering the main processes and key nodes of the battery disassembly process steps, setting the nodes that need to satisfy the multi-process condition paths as AND nodes, and setting the path nodes of whether to perform disassembly after battery performance detection as OR nodes;

[0047] The second predetermined rule includes: defining 1 for the node relationship of AND node or connection relationship, and defining 2 for the node relationship of OR node.

[0048] Further, the method for obtaining an optimized disassembly path by combining disassembly weights with depth search includes: traversing the disassembly path under predetermined constraints according to the disassembly weights and depth search to calculate statistical values, and determining the optimized disassembly path based on the statistical values in combination with the battery pack detection information;

[0049] The predetermined constraints are set based on the states of the machine and the battery, the process sequence, single processes, assembly and disassembly times, and the inspection station nodes;

[0050] The statistical values are calculated by the following formula:

[0051] ;

[0052] where, represents the statistical value of the th node in the th branch when traversing to the th node, represents the sum of all possible out-degrees triggered from this node, represents the in-degree of the th node in the th branch, represents the out-degree of the th node in the th branch, represents the in-degree of the th node in the th branch.

[0053] Beneficial effects:

[0054] A method for disassembling and scheduling retired batteries based on reinforcement learning provided by the present invention. By designing a multi-graph convolutional network, obtaining battery information, interacting with and learning from the environment, and implementing according to the disassembly situation of the battery, the present application realizes autonomous adjustment of the disassembly path and process flow, and achieves a more efficient and flexible disassembly;

[0055] By reasonably designing the incentive function, resource allocation and scheduling are optimized, avoiding resource waste and bottleneck effects during the disassembly process, covering multiple key performance indicators such as average completion time, process waiting time, total delay penalty, assembly and disassembly times, and inspection station processing capacity. This design of the comprehensive incentive function can effectively optimize multiple objectives simultaneously, improve the overall disassembly efficiency, rather than just focusing on a single objective, and improve the overall efficiency and stability of the production line;

[0056] By designing negative incentives for different indicators, unnecessary waiting time, delays, and unreasonable resource allocations can be reduced, thereby improving the efficiency and rationality of the disassembly process. These incentive functions can directly affect the scheduling system and drive the system towards the optimization goal.

[0057] By introducing MGCN units, the multi-graph convolutional network enhances the ability to capture the complex relationships and interactions between different parts of power batteries. Traditional methods may not be able to effectively integrate multi-scale information, while the present invention optimizes the multi-scale feature extraction path through MGCN units and can accurately capture different levels of details of battery components during the disassembly process.

[0058] Through the SE module, the present invention can intelligently adjust the importance of feature maps, enabling the system to dynamically adjust the disassembly strategy and improving the adaptability and decision-making accuracy of the system in complex environments.

[0059] A feedback mechanism is designed to enable the model to adjust decisions according to real-time feedback during the disassembly process, enhancing the flexibility and efficiency of the disassembly process. This dynamic optimization method enables the disassembly process to cope with environmental changes, optimize the disassembly path, and thus improve the overall disassembly efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a schematic diagram of the network structure of the MGCN unit in the multi-graph convolutional network in the embodiment of the present invention;

[0061] Figure 2 It is a schematic diagram of the network structure of the multi-graph convolutional network in the embodiment of the present invention, where A represents the obtained optimized disassembly path;

[0062] Figure 3 It is a flow chart of the disassembly process steps of three different batteries in the embodiment of the present invention;

[0063] Figure 4 It is a schematic diagram of the graph matrix encoding constructed in the embodiment of the present invention;

[0064] Figure 5 It is a flow chart of the optimized disassembly process steps of the disassembly process steps of three different batteries in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] The technical solutions of the present invention will be described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0066] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the art to which the present invention pertains. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "a" or "one" do not denote a quantity limitation, but mean that there is at least one. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "up", "down", "left" and "right" are only used to indicate relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationships also change accordingly.

[0067] Please refer to Figure 1-2 , the embodiments of the present application provide a method for scheduling the disassembly of retired batteries based on reinforcement learning, including:

[0068] Step 1: Obtain the machine state of the disassembly machine, construct a graph matrix encoding based on the battery disassembly process steps, construct an MGCN unit for extracting the connectivity and topological structure between different parts of the battery based on the Conv2d k1 block and the Conv2d k3 block, and construct a multi-graph convolutional network according to the MGCN unit;

[0069] Please refer to Figure 3-4 , constructing a graph matrix encoding based on the battery disassembly process steps includes: constructing a disassembly path according to the battery disassembly process steps in combination with a first predetermined rule, and constructing a graph matrix encoding based on the disassembly path;

[0070] The first predetermined rule includes: numbering the main processes and key nodes of the battery disassembly process steps, setting the nodes that need to satisfy the multi-process condition paths as AND nodes, and setting the path nodes for whether to perform disassembly after battery performance detection as OR nodes. The construction of the matrix encoding more effectively expresses the OR nodes and improves the decision-making of node selection and the efficiency of the disassembly path.

[0071] Constructing a graph matrix encoding based on the disassembly path includes: using the numbers of the nodes of the disassembly path as the rows and columns of the matrix encoding, and filling the rows and columns of the matrix encoding in combination with a second predetermined rule based on the node relationships of the disassembly path;

[0072] The second predetermined rule includes: defining 1 when the node relationship is an AND node or a connection relationship, and defining 2 when the node relationship is an OR node.

[0073] Among them, the MGCN unit is a multi-source information fusion feature extraction unit with three inputs and one output. The MGCN unit includes an input layer, a fusion extraction layer, a weighting layer and an output layer;

[0074] The input layer includes three Conv2d k1 blocks that respectively extract local features of three types of inputs;

[0075] The fusion extraction layer includes a fusion block and a Conv2d k3 block that extracts spatial relationships;

[0076] The weighting layer includes an SE block;

[0077] The output layer includes a Conv2d k1 block that performs linear transformation and dimension compression and an output block.

[0078] Constructing a multi-graph convolutional network based on the MGCN unit specifically includes:

[0079] Constructing an action module and a value module based on the MGCN unit, and constructing a multi-graph convolutional network based on the action module and the value module;

[0080] The action module extracts multi-scale features based on the machine state and the graph matrix encoding to obtain disassembly weights;

[0081] The value module obtains the multi-scale features extracted by the action module and combines an activation function to optimize the disassembly weights, that is, the value module performs value prediction on the disassembly weights to be output by the action module based on the multi-scale features. The activation function collects the machine state for feedback calculation, and uses the difference between the feedback calculation and the value prediction as the objective function to optimize the disassembly weights;

[0082] Specifically, the action module sequentially includes an MGCN unit for extracting machine state and graph matrix encoding features, a detail extraction Conv2d k3 block for further extracting surface and internal detail features of the machine state and the graph matrix encoding, an MGCN unit for strengthening the machine state and graph matrix encoding features, a detail strengthening Conv2d k3 block for further strengthening surface and internal detail features of the machine state and the graph matrix encoding, and a Conv2d k1 block for compressing and integrating all the strengthened features.

[0083] The value module is respectively connected to the detail extraction Conv2d k3 block and the detail strengthening Conv2d k3 block of the action module, and integrates and evaluates the obtained multi-scale features;

[0084] The value module includes an acquisition unit for obtaining multi-scale features, an integration unit for integrating the multi-scale features, and an evaluation unit for evaluating the integrated multi-scale features, where the acquisition unit is connected to the detail extraction Conv2d k3 block and the detail strengthening Conv2d k3 block of the action module;

[0085] The acquisition unit includes a first channel, a second channel, and a Conv2d k1 block. Both the first channel and the second channel are composed of two Conv2d k3 blocks and are connected to the action module and the integration module. The Conv2d k1 block is arranged between the first Conv2d k3 block in the connection direction of the first channel, the second channel and the action module;

[0086] The integration unit includes 1 Conv2d k1 block;

[0087] The evaluation unit includes 1 fully connected block and an MGCN unit.

[0088] The value module performs value prediction on the disassembly weights to be output by the action module based on multi-scale features. The incentive function combines the machine state for feedback calculation, and uses the difference between the feedback calculation and the value prediction as the objective function to optimize the disassembly weights;

[0089] The machine state includes the average completion time, process waiting time, delay penalty, assembly and disassembly time, and station processing capacity;

[0090] The incentive function is expressed by the following formula:

[0091] ;

[0092] Wherein, represents the incentive function, represents the average completion time, represents the process waiting time, represents the delay penalty, represents the assembly and disassembly time, represents the station processing capacity, all represent weight coefficients. In this embodiment, the weight coefficients of each parameter are all set to 0.2;

[0093] The average completion time is expressed by the following formula:

[0094] ;

[0095] Wherein, represents the time required to complete the disassembly of the th battery pack, represents the maximum allowed completion time;

[0096] The process waiting time is expressed by the following formula:

[0097] ;

[0098] Wherein, represents the process waiting time for the disassembly of the th battery pack, Indicates the longest allowable waiting time;

[0099] The delay penalty is expressed by the following formula:

[0100] ;

[0101] where, Indicates the total delay time of disassembling the battery;

[0102] The assembly and disassembly time is expressed by the following formula:

[0103] ;

[0104] where, Indicates the time required to disassemble the battery pack from the machine, Indicates the time required to assemble the battery pack onto the machine;

[0105] The station processing capacity is expressed by the following formula:

[0106] ;

[0107] where, Indicates the maximum processing capacity of the station, that is, the time required to disassemble a battery pack, Indicates the th battery pack at the th process by the machine processing time required, Indicates the total number of battery packs to be disassembled.

[0108] Step 2: Encode the machine state and the graph matrix and input them into the multi-graph convolutional network to obtain the disassembly weights;

[0109] Step 3: Combine the disassembly weights with depth search to obtain the optimized disassembly path, and complete the disassembly of the retired battery according to the optimized disassembly path.

[0110] Specifically, please refer to Figure 5 , traverse the disassembly path according to the disassembly weights combined with depth search to calculate the statistical value under the predetermined constraints, and determine the optimized disassembly path based on the statistical value combined with the battery pack detection information. Among them, the battery information includes module information and cell information, and the battery pack detection information includes the quality of the module and the cell, that is, determine the optimized disassembly path based on the statistical value combined with the modules and cells disassembled from the battery pack.

[0111] The predetermined constraints are set based on the states of the machine and the battery, the process sequence, the single process, the assembly and disassembly time, and the inspection station nodes;

[0112] The state constraints of the machines and batteries are represented by the following formula:

[0113] ;

[0114] where, means that all machines and batteries are available at time and are not bound to any machine;

[0115] The process sequence constraints are represented by the following formula:

[0116] ;

[0117] where, represents the time required for the th battery pack in the th process;

[0118] represents the time required for the th battery pack in the th process;

[0119] represents the time required for the th battery pack to be processed by the th machine in the th process;

[0120] represents the time required for the battery pack to be disassembled from the th machine, represents the time required for the battery pack to be assembled onto the th machine. This formula indicates that the battery processes are executed in sequence;

[0121] The single-process constraints are represented by the following formula:

[0122] ;

[0123] where, represents the total number of processes for the th battery pack, represents the number of machines in the workshop;

[0124] represents that at time , if the th battery pack's th process is processed by the th machine, then , otherwise 0;

[0125] This formula indicates that a battery can only undergo one process at any given time;

[0126] ;

[0127] This formula indicates that each machine can only handle one process at any time;

[0128] The assembly and disassembly time constraints are expressed by the following formula:

[0129] ;

[0130] in, exist Moment, Battery Pack The process is Machine processing, , otherwise 0;

[0131] This formula indicates that each battery must consider the time constraint during assembly and disassembly;

[0132] The inspection station node constraints are expressed by the following formula:

[0133] ;

[0134] in, Represents the total number of retired batteries, It represents the processing capacity of the inspection station, that is, the number of batteries that can be inspected per unit time; this formula means that the module and cell of each battery are the process bottleneck at the inspection station, and its processing capacity needs to be given priority.

[0135] The statistical value is calculated by the following formula:

[0136] ;

[0137] in, Indicates traversal to Node, The first branch The statistics of nodes, Indicates that from The sum of all possible out-degrees triggered by the node, Indicates The first branch The in-degree of a node, Indicates The first branch The out-degree of a node, Indicates The first branch The in-degree of a node.

[0138] See also Figure 3 , for the node 6 of battery 1, the out-degree of node 6 is -3. There are a total of nodes in a branch. When , the branches are counted as nodes 9 and 10. The sum of the in-degree and out-degree of node 9 is equal to 0, while the in-degree of node 10 is 1, and the statistical value is -2 at this time. When , the branches are counted as nodes 9, 10, 11, and 15. The sum of the in-degree and out-degree of nodes 9, 10, and 11 is equal to 0, while the in-degree of node 15 is 3, and the statistical value is 0 at this time.

[0139] Finally, in this embodiment, a multi-graph convolutional network constructed by using a method for disassembling and scheduling retired batteries based on reinforcement learning according to the present invention is used to make decisions on disassembly paths together with manual scheduling, genetic algorithms, and particle swarm optimization algorithms. Please refer to Table 1 for details.

[0140] Table 1: Average Completion Time, Process Waiting Time, Delay Penalty, and Total Objective

[0141]

[0142] As can be seen from Table 1, the multi-graph convolutional network is superior to traditional manual scheduling, genetic algorithms, and particle swarm optimization algorithms in the scheduling of retired battery disassembly. In terms of indicators such as average completion time, process waiting time, and delay penalty, the multi-graph convolutional network performs the most prominently. In particular, the average completion time is reduced to 3287 seconds, the process waiting time is reduced to 482 seconds, the delay penalty is reduced to 65 seconds, and the comprehensive total objective value reaches 3834 seconds. Each value is the lowest among all methods, clearly demonstrating that the method of the present invention has excellent scheduling optimization capabilities.

[0143] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should fall within the protection scope determined by the claims.

Claims

1. A method for disassembling and scheduling retired batteries based on reinforcement learning, characterized in that: The steps include: Step 1: Obtain the machine state of the disassembly machine, construct a graph matrix encoding based on the battery disassembly process steps, construct an MGCN unit based on the Conv2dk1 block and the Conv2d k3 block to extract the connectivity and topology between different parts of the battery, and construct a multi-graph convolutional network based on the MGCN unit; The MGCN unit is a multi-source information fusion feature extraction unit with three inputs and one output, and the MGCN unit includes an input layer, a fusion extraction layer, a weighted layer and an output layer; The input layer includes three Conv2d k1 blocks for extracting local features of three inputs respectively; The fusion extraction layer includes a fusion block and a Conv2d k3 block for extracting spatial relations; The weighted layer comprises a SE block; The output layer includes a Conv2d k1 block and an output block for performing linear transformation and dimensionality compression; The constructing of a multi-graph convolutional network according to the MGCN unit includes: constructing an action module and a value module according to the MGCN unit, and constructing a multi-graph convolutional network based on the action module and the value module; The action module extracts multi-scale features based on machine state and graph matrix encoding to obtain disassembly weights; The value module obtains the multi-scale features extracted by the action module and optimizes the disassembly weights in combination with the excitation function; The activation function is expressed by the following formula: R=λ1·r1+λ2·r2+λ3·r3+λ4·r4+λ5·r5; Among them, R represents the incentive function, r1 represents the average completion time, r2 represents the process waiting time, r3 represents the delay penalty, r4 represents the assembly and disassembly time, r5 represents the station processing capacity, and λ1,λ2,λ3,λ4,λ5 all represent weight coefficients; Step 2: Input the machine state and graph matrix encoding into the multi-graph convolutional network to obtain the disassembly weight; Step 3: Obtain the optimized disassembly path based on the disassembly weight and deep search, and complete the disassembly of the retired batteries according to the optimized disassembly path.

2. The method for dismantling retired batteries based on reinforcement learning according to claim 1 is characterized in that: The action module sequentially includes an MGCN unit for extracting machine state and graph matrix encoding features, a detail extraction Conv2d k3 block for further extracting machine state and graph matrix encoding surface and internal detail features, an MGCN unit for strengthening machine state and graph matrix encoding features, a detail strengthening Conv2d k3 block for further strengthening machine state and graph matrix encoding surface and internal detail features, and a Conv2d k1 block for compressing and integrating all strengthened features.

3. The method for dismantling and scheduling retired batteries based on reinforcement learning according to claim 2 is characterized in that: The value module is connected to the detail extraction Conv2d k3 block and the detail enhancement Conv2d k3 block of the action module respectively, and integrates and evaluates the acquired multi-scale features.

4. The method for dismantling and scheduling retired batteries based on reinforcement learning according to claim 3 is characterized in that: The value module includes an acquisition unit for acquiring multi-scale features, an integration unit for integrating the multi-scale features, and an evaluation unit for evaluating the integrated multi-scale features; The acquisition unit includes a first channel, a second channel and a Conv2d k1 block, the first channel and the second channel are each composed of two Conv2d k3 blocks, and are both connected to the action module and the integration module, and the Conv2d k1 block is arranged between the first Conv2d k3 block in the direction where the first channel, the second channel and the action module are connected; The integration unit includes 1 Conv2d k1 block; The evaluation unit includes 1 fully connected block and one MGCN unit.

5. The method for dismantling and scheduling retired batteries based on reinforcement learning according to any one of claims 1 to 4, characterized in that: The value module obtains the multi-scale features extracted by the action module and optimizes the disassembly weight in combination with the incentive function, including: the value module predicts the value of the disassembly weight output by the action module based on the multi-scale features, the incentive function performs feedback calculation in combination with the machine state, and uses the difference between the feedback calculation and the value prediction as the objective function to optimize the disassembly weight.

6. The method for dismantling and scheduling retired batteries based on reinforcement learning according to claim 5 is characterized in that: The machine status includes average completion time, process waiting time, delay penalty, assembly and disassembly time, and station processing capacity; The average completion time is expressed by the following formula: Among them, T i represents the time required to complete the disassembly of the ith battery pack, T max Indicates the maximum allowed completion time; The process waiting time is expressed by the following formula: Among them, w i represents the waiting time for the disassembly of the ith battery pack, w max Indicates the maximum allowed waiting time; The delay penalty is expressed by the following formula: r3＝-L i ； Among them, L i Indicates the total delay time for disassembling the battery; The assembly and disassembly time is expressed by the following formula: r4＝-(α k +β k )； Among them, α k represents the time required to remove the battery pack from the k machine, β k Indicates the time required for the battery pack to be delivered to the k machine; The processing capacity of the workstation is expressed by the following formula: Among them, C test represents the maximum processing capacity of the workstation, that is, the time required to disassemble a battery pack, i,j,k(t) represents the time required for the i-th battery pack to be processed by the k-th machine in the j-th process, and N represents the total number of battery packs to be disassembled.

7. The method for dismantling and scheduling retired batteries based on reinforcement learning according to any one of claims 1 to 4, characterized in that: Constructing a graph matrix code based on the battery disassembly process steps includes: constructing a disassembly path according to the battery disassembly process steps in combination with a first predetermined rule, using the numbers of the nodes of the disassembly path as rows and columns of the graph matrix code, and filling the rows and columns of the graph matrix code based on the node relationship of the disassembly path in combination with a second predetermined rule; The first predetermined rule includes: numbering the main processes and key nodes of the battery disassembly process steps, setting the nodes of the paths that need to meet multiple process conditions at the same time as AND nodes, and setting the path nodes of whether to disassemble the battery after performance testing as OR nodes; The second predetermined rule includes: if the node relationship is an AND node or a connection relationship, it is defined as 1; if the node relationship is an OR node, it is defined as 2.

8. The method for dismantling and scheduling retired batteries based on reinforcement learning according to claim 7 is characterized in that: The obtaining of the optimized disassembly path according to the disassembly weight combined with the deep search comprises: traversing the disassembly path under predetermined constraints to calculate statistical values ​​according to the disassembly weight combined with the deep search, and determining the optimized disassembly path based on the statistical values ​​combined with the battery pack detection information; The predetermined constraints are based on the status of the machine and the battery, the process sequence, the single process, the assembly and disassembly time, and the detection station node setting; The statistical value is calculated by the following formula: Among them, S(iJ) represents the statistical value of the jth node in the i-th branch when traversing to the OR node, ToutD represents the sum of all possible out-degrees triggered from the OR node, inD(i,j-1) represents the in-degree of the j-1th node in the i-th branch, outD(i,j-1) represents the out-degree of the j-1th node in the i-th branch, and inD(i,j) represents the in-degree of the j-1th node in the i-th branch.

Images