A method and related device for optimizing disassembly sequence of retired power battery of an automobile

Abstract

The application provides a kind of automobile decommissioning power battery disassembly sequence optimization method and related device, it is related to data processing technical field.By introducing the disassembly risk priori knowledge provided by battery manufacturer, a multi-objective disassembly sequence planning mathematical model integrated with risk perception is established, and a disassembly sequence planning algorithm based on graph attention network and constraint perception pointer network is adopted, realizing the end-to-end real-time planning of disassembly sequence for different categories of battery packs.Not only can effectively utilize the prior disassembly knowledge of official manual, ensure the efficiency and safety of disassembly process, but also can adapt to complex battery pack structure and dynamically changing disassembly scene, providing a new solution for the disassembly of decommissioned power battery.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a method for optimizing a disassembly sequence of a retired power battery of an automobile and related devices. BACKGROUND

[0002] With the rapid development of the new energy automobile industry, the recycling and disassembly of retired power batteries have become increasingly prominent. Retired power batteries contain various harmful substances, such as heavy metals, electrolyte, etc. If not properly handled, it will cause serious pollution to the environment, and also waste valuable metal resources. Therefore, efficiently and safely disassembling retired power batteries and realizing resource recycling have become an important part of the sustainable development of the new energy automobile industry.

[0003] However, the disassembly process of retired power batteries faces many challenges. First, the battery pack structure is complex, and different types of battery packs differ in component composition, connection method, etc., which makes the disassembly sequence planning extremely complex. Traditional disassembly methods often rely on manual experience, which is difficult to adapt to diversified battery pack structures and is prone to disassembly errors, resulting in low disassembly efficiency and even possible safety accidents. Second, there is a high risk in the disassembly process of retired power batteries. The chemical substances inside the battery may leak, burn or even explode due to improper operation, posing a serious threat to the personal safety of the operator.

[0004] Currently, although battery manufacturers provide disassembly task instructions and operation manuals, relying solely on these static guidance documents cannot meet the individual differences of battery packs and the dynamic changes in the disassembly process. In the face of unexpected situations or abnormal battery pack structures, it is necessary to quickly adjust the disassembly sequence to cope with the complex and variable disassembly environment. In the prior art, although some research attempts to plan the disassembly sequence through optimization algorithms, these methods mostly focus on disassembly efficiency, ignoring the risk factors in the disassembly process. Moreover, these methods often show poor adaptability and flexibility when faced with complex battery pack structures and dynamic disassembly scenarios.

[0005] Therefore, there is an urgent need for a method that can consider disassembly efficiency and disassembly risk comprehensively and plan disassembly sequences in real time to meet the actual needs of retired power battery disassembly. SUMMARY

[0006] In order to meet the actual needs of retired power battery disassembly, the present application provides a method for optimizing a disassembly sequence of a retired power battery of an automobile and related devices.

[0007] In a first aspect, the present application provides a method for optimizing a disassembly sequence of a retired power battery of an automobile, which adopts the following technical solution:

[0008] A method for optimizing a disassembly sequence of a retired automotive power battery, comprising:

[0009] According to the disassembly task instructions and operation manuals provided by the manufacturer, the battery pack part number is taken as a node, and an extended part precedence graph is generated to generate a precedence relationship matrix;

[0010] An optimization model is established including a disassembly direction transition number expression and a disassembly tool replacement number expression;

[0011] A pointer network model based on a graph attention network and a dynamic constraint mask attention mechanism is used to input the precedence relationship matrix and the part feature matrix and obtain an output result;

[0012] The entropy weight approximation ideal solution sorting method is used to select a compromise optimal disassembly sequence.

[0013] Optionally, the step of establishing an optimization model including a disassembly direction transition number expression and a disassembly tool replacement number expression comprises:

[0014] establishing a disassembly process direction transition function , a disassembly process tool replacement function , and a disassembly process risk assessment function ;

[0015] According to the disassembly process direction transition function, the disassembly process tool replacement function, and the disassembly process risk assessment function, an optimization model based on a disassembly strategy is generated, wherein the retired power battery is disassembled from parts, the disassembly sequence of the retired power battery, satisfies that each part is disassembled only once and meets the precedence constraint relationship between parts, represents the part disassembled in the first disassembly sequence.

[0016] Optionally, the disassembly process direction transition function comprises:

[0017]

[0018] wherein three directions are defined as , , and using one-hot encoding to represent the operation direction in the disassembly process as up, down, left, right, and front and back.

[0019] Optionally, the disassembly process tool replacement function :

[0020]

[0021] where one-hot encoding is used Three tools are defined as , and to represent the wrench, pliers and screwdriver in the disassembly process.

[0022] Optionally, the disassembly process risk assessment function includes:

[0023]

[0024] where s represents the part risk coefficient.

[0025] Optionally, the optimization model is represented as:

[0026]

[0027] where, , and each element in the weight vector has been normalized and added to 1.

[0028] Optionally, in the pointer network model based on the graph attention network and the dynamic constraint mask attention mechanism, it further includes:

[0029] The double-layer encoder processes the part features, first uses the gated recurrent unit network to aggregate the sequential information of the features, and then enters the graph attention network to embed the part features into a high-dimensional space. The decoder generates the disassembly sequence by using the attention mechanism with a dynamic mask mechanism.

[0030] Optionally, before the step of inputting the priority relation matrix and the part feature matrix and obtaining the output result, it further includes:

[0031] The model parameters are optimized by the center self-criticism reinforcement learning algorithm, and iterative learning is performed based on the policy gradient update rule.

[0032] Optionally, the step of optimizing the model parameters by the center self-criticism reinforcement learning algorithm and performing iterative learning based on the policy gradient update rule includes:

[0033] The disassembly sequence planning multi-objective mathematical model is standardized into a symbol of a reinforcement learning problem, the state space, the action space, each state is defined as the set of all parts that have been disassembled at time t, that is, , and the action is defined as the selection of the part to be disassembled at the next time step t+1 ;

[0034] Policy is defined as , which represents a set of part states that have been disassembled given time t , the policy returns the probability distribution of the next disassembled part, where represents the trainable parameters of the neural network;

[0035] The expected reward is defined as the negative value of the subproblem objective function:

[0036]

[0037] where is the space of part sets, is the space of all possible permutations of , and is the distribution of predicted by the neural network;

[0038] The objective function is defined as the expected reward, i.e.

[0039]

[0040] The gradient of the objective function is calculated as follows:

[0041]

[0042] where is the batch size of training, is the number of disassembly sequence solutions generated for each subproblem, is the objective function value of the th disassembly sequence of the th subproblem, is the baseline, is the th part feature input of the th subproblem , the output is the probability of the disassembly sequence , and the training process optimizes the parameters using the gradient ascent method:

[0043]

[0044] where is the learning rate, a positive scalar that controls the step size of parameter adjustment at each update, and the learning rate determines the magnitude of the algorithm's progress in the gradient direction of the loss function. The selection of the policy gradient baseline uses Centralself-critic, and the calculation formula of the baseline is as follows:

[0045]

[0046] wherein represents the reward obtained by using the greedy strategy to obtain the disassembly sequence, represents the reward obtained by using the strategy obtained by random sampling.

[0047] In a second aspect, the application provides a disassembly sequence optimization device for retired automotive power batteries, comprising:

[0048] A matrix generation module is configured to generate an extended part precedence graph by taking battery pack part numbers as nodes and generating the extended part precedence graph according to the disassembly task specification and the operation manual provided by the manufacturer, so as to generate a precedence relationship matrix.

[0049] A model construction module is configured to establish an optimization model containing a disassembly direction transition number expression and a disassembly tool replacement number expression.

[0050] A data input module is configured to input the precedence relationship matrix and a part feature matrix by using a pointer network model based on a graph attention network and a dynamic constraint mask attention mechanism, and to obtain an output result.

[0051] A sequence determination module is configured to select a compromise optimal disassembly sequence by using an entropy weight approximation ideal solution ordering method on the output result.

[0052] In summary, the application introduces the disassembly risk prior knowledge provided by the battery manufacturer, establishes a multi-objective disassembly sequence planning mathematical model integrating risk perception, and uses a disassembly sequence planning algorithm based on a graph attention network and a constraint perception pointer network to realize end-to-end real-time planning of disassembly sequences for different types of battery packs. Not only can the prior disassembly knowledge of the official manual be effectively utilized to ensure the efficiency and safety of the disassembly process, but also the complex battery pack structure and the dynamically changing disassembly scene can be adapted, thereby providing a brand-new solution for the disassembly of retired power batteries. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is a flowchart of a first embodiment of the disassembly sequence optimization method for retired automotive power batteries of the application;

[0054] Figure 2 is a model architecture diagram of the disassembly sequence optimization method for retired automotive power batteries of the application;

[0055] Figure 3 is an input feature vector diagram of the sub-problem regularization of the disassembly sequence optimization method for retired automotive power batteries of the application;

[0056] Figure 4is a structural block diagram of a first embodiment of a vehicle retired power battery disassembly sequence optimization device of the present application. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below with reference to the drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0058] The embodiment of the present application provides a vehicle retired power battery disassembly sequence optimization method, referring to Figure 1 , Figure 1 is a flowchart of a first embodiment of a vehicle retired power battery disassembly sequence optimization method of the present application.

[0059] In the embodiment, the vehicle retired power battery disassembly sequence optimization method comprises the following steps:

[0060] Step S10: According to the disassembly task specification and operation manual provided by the manufacturer, the battery pack part number is taken as the node and the extended part precedence graph is generated to generate the precedence relationship matrix.

[0061] It should be noted that for the current widely existing retired power battery, the manufacturer usually provides an official disassembly task specification and detailed operation manual to provide a relatively safe and guiding path for manual disassembly process. The guidebook lists the disassembly steps of the battery pack in detail and systematically numbers each component to ensure the standardization and safety of the disassembly process. Based on these numbers, this study takes them as the nodes of the part precedence graph, thereby combining the disassembly guide provided by the manufacturer with the actual disassembly requirements, laying a foundation for the subsequent multi-objective disassembly sequence planning model.

[0062] For the problem of high risk of power battery disassembly, the embodiment introduces the risk precedence constraint as a core dimension. By directly linking the component disassembly sequence with potential hazards, a "risk-driven disassembly topology structure" is formed. For example, when disassembling the positive electrode assembly, the logical node of forcibly pre-positioning the disconnection of the positive electrode busbar is used to completely block the conduction path of the electric shock risk, realizing the paradigm shift from "passive risk avoidance" to "active safety design".

[0063] Thus, in the process of constructing the part precedence graph, first, according to the CAD drawing or manual operation experience, the physical constraint relationship between parts is accurately identified. Based on these constraint relationships, and in accordance with the part numbering rules mentioned earlier, the parts that are mutually constrained are connected by a directed arrow, with the arrow direction pointing from the constraint-imposed party to the constrained party, thereby constructing a standard part precedence graph. On this basis, a risk precedence constraint mechanism is further introduced, and an extended part precedence graph is finally formed. This innovative extended relationship graph not only ensures the logicality and systematicness of the disassembly process, but also significantly improves the safety and reliability of the disassembly process, providing more scientific guidance for subsequent disassembly planning and operation.

[0064] Step S20: Establish an optimization model containing the disassembly direction change frequency expression and the disassembly tool replacement frequency expression.

[0065] It should be noted that the step of establishing an optimization model containing the disassembly direction change frequency expression and the disassembly tool replacement frequency expression includes: establishing a disassembly process direction change function , a disassembly process tool replacement function , and a disassembly process risk assessment function .

[0066] According to the disassembly process direction change function, the disassembly process tool replacement function, and the disassembly process risk assessment function, an optimization model based on the decomposition strategy is generated, wherein the disassembly of the retired power battery is composed of parts, is the disassembly sequence of the retired power battery, satisfies that each part is disassembled only once and meets the precedence constraint relationship between parts, represents the part disassembled in the th position in the disassembly sequence.

[0067] The disassembly process direction change function includes:

[0068]

[0069] wherein three directions are defined as , , and using one-hot encoding to represent the operation direction in the disassembly process as up, left, right, and front.

[0070] The disassembly process tool replacement function :

[0071]

[0072] wherein three directions are defined as , , and using one-hot encoding Three tools are defined as , and to represent the tools wrench, pliers, screwdriver in the disassembly process.

[0073] It should be noted that the internal structure of the power battery adopts a crisscross arrangement structure, and there is a high coupling characteristic between adjacent components. In view of this, if a non-continuous disassembly operation is implemented, the safety risk of the disassembly operation will be significantly increased. Therefore, the battery manufacturer usually provides a standardized disassembly process as a reference. In addition, different parts have different risk levels due to the difference in structure and function, so a part risk coefficient This index is specially introduced. Its definition rule is mainly based on potential safety accidents. The higher the risk coefficient value, the greater the safety risk that the component may bring in the disassembly process, and special attention needs to be paid in the disassembly process. The parts to be disassembled immediately before are as much as possible the same as the official guidance. Based on the risk coefficient of the disassembled parts, in order to minimize the disassembly operation risk, the consistency degree of the disassembly sequence with the official manual is taken as the risk evaluation standard, and a disassembly process risk evaluation function is established.

[0074]

[0075] Wherein, s represents the part risk coefficient.

[0076] Based on the decomposition strategy, a multi-objective optimization mathematical model is solved, weights are set for each target, the multi-objective problem is decomposed into single-objective sub-problems under different weights, the optimal solution of each sub-problem is solved, and finally the problem is obtained by merging and non-dominated sorting. The Pareto front. Assuming that the three target preferences of the disassembly sequence planning problem are , each element in the weight vector has been normalized and added to 1, where a larger value indicates higher importance. The optimization model is represented as:

[0077] .

[0078] Step S30: A pointer network model based on a graph attention network and a dynamic constraint mask attention mechanism is adopted, the priority relationship matrix and the part feature matrix are input, and an output result is obtained.

[0079] It should be noted that the traditional iterative search algorithm has significant limitations in dealing with the battery disassembly sequence planning problem. Such algorithms are difficult to meet the real-time disassembly sequence planning requirements, and once the input characteristics change slightly, the algorithm needs to be redesigned or iterated to solve, resulting in poor generalization. In contrast, the pointer network, as a network model specifically designed to solve sequence output, has significant advantages. Once trained, the pointer network can quickly solve the result in a short time and exhibit good generalization performance for different problems. Therefore, the pointer network has high applicability to the battery disassembly sequence planning problem.

[0080] In a specific implementation, a multi-objective optimization mathematical model is solved based on a decomposition strategy, a weight is set for each target, the multi-objective problem is decomposed into single-objective sub-problems under different weights, the optimal solution of each sub-problem is solved, and finally the non-dominated sorting is combined to obtain the Pareto front of the problem. Suppose that the three target preferences of the disassembly sequence planning problem are , each element in the weight vector has been normalized and added to 1, where a larger value indicates higher importance. Therefore, the weight vector The scalarized objective function of the disassembly sequence under the weight vector is represented as follows:

[0081]

[0082] where represents a given retired traction battery part feature matrix, a target weight vector, and a part priority constraint relationship state, represents the objective function value of the disassembly sequence under the given feature matrix . The goal of pointer network training is to learn a random strategy , under the condition of a given retired traction battery part feature matrix, a target weight vector, and a part priority constraint relationship state , output a higher probability for a disassembly sequence with a smaller target function value, and the pointer network decomposes the probability in a chain rule as follows:

[0083]

[0084] where each factor on the right side of the equation is obtained from the output of the layer of the pointer network.

[0085] It should be noted that, in order to ensure that the output of the pointer network meets the priority relationship of the extended part precedence graph, further improvement is needed. Therefore, the embodiment constructs a pointer network model introducing a graph attention network and a dynamic constraint mask attention mechanism. The graph attention network can efficiently process graph structure data and can assign different weights to different nodes in the neighborhood, thereby better adapting to different sizes of the neighborhood. The dynamic constraint mask attention mechanism can update the evaluation score of the attention mechanism during the decoding process, ensuring that the generated disassembly sequence strictly follows the disassembly order constraint. Through these improvements, the embodiment not only improves the applicability of the pointer network in battery disassembly sequence planning, but also significantly enhances the real-time performance and generalization ability of the model. Figure 2 Model architecture of the embodiment.

[0086] The embodiment adopts GRU-GAT double-layer encoding to promote effective learning of part features and ensure full use of information. Since the risk assessment function of the disassembly sequence multi-objective planning mathematical model is related to the order of input, and the GRU has similar functions to the LSTM with fewer training parameters, the GRU network is selected to aggregate the input feature information and the order information of the input features, and the features of the parts are stored in a vector with a fixed size. The graph attention network (GAT) is used to process the part precedence relationship matrix to extract the precedence relationship information between parts.

[0087] The encoder input includes part feature information (wherein the number of parts, the feature dimension of each part) and the part precedence relationship matrix .

[0088] To solve the sub-problems decomposed under different weights using a single model, the weight of the sub-problem needs to be appended to the part feature information after encoding. The specific definition rules of the encoder GRU network input are as follows:

[0089] a. The feature vector of the part and the weight vector of the target are spliced together to form the input vector. The order of the target corresponding to the elements in the feature vector should be the same as the order of the weight vector;

[0090] b. In the feature vector, the feature dimension associated with each target remains consistent. If the feature dimension of a part is smaller than that of other parts, the dimension of the smaller feature needs to be padded according to the maximum feature dimension. The padding value is set to 1.

[0091] For example Figure 3The part features include disassembly tools, disassembly directions and risk coefficients. The disassembly tools and disassembly directions are three-dimensional one-hot encodings, and the risk coefficient is a real number, so it is necessary to fill 1 to become a three-dimensional vector. The weight vector attached to the end of the part features of the same sub-problem is the same. Through the construction of rules, the order and position of the elements in the input vector are given practical significance, increasing the amount of information provided by the input vector to the model. At the same time, the weight of the target is also included in the input vector, which ensures that the information sent to all subsequent modules contains the target weight.

[0092] The GAT network processes the part feature matrix to obtain the high-dimensional features of each part As shown in the formula:

[0093]

[0094] Wherein represents the high-dimensional features of the part .

[0095] After GRU processing, the fusion features of the part are obtained as the final output of the encoder.

[0096] In specific implementation, the decoder is described as follows:

[0097] In order to minimize the consumption of training models, the embodiment adopts an RNN model decoder of a gated recurrent unit (GRU), stores all node embedding information input into the decoder (i.e. mixed feature information of all parts), and uses the node embedding information of the last disassembled part , and the decoder hidden state of the previous time , the decoder can output a decoding vector containing information about all parts disassembled so far, i.e. the hidden state of the current time , which is used as the input of the attention mechanism at time.

[0098] It should be noted that the attention mechanism includes:

[0099] According to the fusion features and the decoder hidden state , the next time step disassembly part probability vector is outputted. The working principle of the attention mechanism with dynamic priority constraint mask is as follows:

[0100] Initialize the hidden state of the decoder (usually use the last state of the encoder or other initialization methods), and perform the following steps for each time step (from 1 to N):

[0101] Step 1: Calculate attention scores (unmasked) according to current decoder hidden state and fused features of encoder

[0102]

[0103] where, is the transpose of query vector, is the fused features of encoder output is the weight of each time step hidden state of decoder output, all of which are neural network learnable parameters. Step 2: To meet the priority constraint relationship between parts in the generated expansion part priority graph, obtain the dynamic mask of the current time step through the dynamic mask generation mechanism

[0104] Add the mask to the attention score:

[0105]

[0106] Step 3: Get the probability distribution of the next disassembly part through layer .

[0107]

[0108] Step 4: According to the probability distribution , select the part to be disassembled , and use sampling and greedy methods respectively during training, and use greedy strategy during testing.

[0109] Step 5: Update the disassembly state, set the disassembly state of the part to 1, that is , and update the undismantled set .

[0110] Step 6: Embed the selected part into the input of the decoder GRU and update the hidden state.

[0111]

[0112] Return to step 1 until all parts are disassembled.

[0113] The dynamic mask generation process of each time step of the embodiment is as follows: let the priority relationship matrix be , where represents the disassembly part ​​The parts need to be disassembled first. That is, there exists a path from arrive A directed edge represents yes The preceding components. At the start of disassembly, all components are not yet disassembled. We maintain a disassembly state vector. Initially Indicates parts Not disassembled; after disassembly, the value is set to 1. At time step... Define the set of currently undisassembled parts as For each part The number of unremoved front parts is calculated as shown in the following formula.

[0114]

[0115] Note the parts here. It is a part The pre-installed components, therefore when and (Component When not disassembled, additional parts will be added. Number of front-end components that were not disassembled Therefore, the set of detachable parts is:

[0116]

[0117] Dynamic mask It is a vector of length equal to the number of parts, defined as follows:

[0118] .

[0119] It should be noted that, given the difficulty in obtaining optimal solution labels for battery pack disassembly sequence planning, traditional supervised learning methods face the dilemma of enormous upfront workload and difficulty in adapting to solving unknown optimal solutions due to their reliance on large amounts of labeled data. Therefore, an unsupervised learning method is adopted, combined with a center-based self-criticism reinforcement learning algorithm, to train the neural network model.

[0120] The actor-critic model is developed from the policy gradient method and is designed as a hybrid model to utilize both policy-based and value-based reinforcement learning methods. Generally, there are two main components in the actor-critic model. The actor is used to handle the generation of the policy, and then decides what action the agent should take next. The actor in this embodiment, i.e., the improved pointer network proposed, provides the complete disassembly sequence. The critic uses a value-based method to evaluate the action taken by the agent. In the disassembly sequence planning problem, the actor gives the disassembly sequence, and the critic will receive the sequence and give the reward value of the sequence. This reward value can be used as an approximate reward value when calculating the loss function to update the weights of the two networks.

[0121] The self-critic method only uses a single artificial neural network model, rather than the usual two actor and critic models. In this architecture, the artificial neural network model acts as both the actor and the critic. When acting as the actor, the agent will randomly select the next city to visit from the probability distribution generated by the model. The model in this case will return the generated itinerary and the probability distribution of selecting the next city. When the model needs to act as the critic, the agent no longer randomly selects the next city to visit, but selects the city with the highest probability and generates the reward value corresponding to the disassembly sequence.

[0122] The strategy generates disassembly sequences in two different ways, using the same artificial neural network, one is randomly generated, and the other is the optimal disassembly sequence under the greedy strategy. Based on this way, the difference between the reward of the disassembly sequence randomly sampled and the center self-critic baseline is calculated, and the gradient is used to update the network parameters. Compared with using the exponentially moving average of the reward, the center self-critic baseline accelerates the convergence speed. The specific steps are as follows:

[0123] In the pointer network model based on the graph attention network and the dynamic constraint mask attention mechanism, further comprising: using a double-layer encoder to process the part feature, first using a gated recurrent unit network to aggregate the sequential information of the feature, and then entering the graph attention network to embed the part feature into a high-dimensional space; the decoder uses an attention mechanism with a dynamic mask mechanism to generate the disassembly sequence.

[0124] In specific implementation, before the step of inputting the priority relation matrix and the part feature matrix and obtaining the output result, further comprising: optimizing the model parameters through the center self-critic reinforcement learning algorithm, and performing iterative learning based on the policy gradient update rule.

[0125] The step of optimizing the model parameters through the center self-critic reinforcement learning algorithm and performing iterative learning based on the policy gradient update rule comprises: standardizing the disassembly sequence planning multi-objective mathematical model into a symbol of a reinforcement learning problem, for the state space, For the action space, each state is defined as the set of all parts that have been disassembled up to time t, i.e. , the action is defined as the selection of a part to disassemble at the next time step t + 1 ;

[0126] The policy is defined as , which represents a given set of part states that have been disassembled up to time t , the policy returns a probability distribution over the next part to disassemble, where denotes the trainable parameters of the neural network

[0127] The expected reward is defined as the negative value of the subproblem objective function:

[0128]

[0129] where is the space of part sets, is the space of all possible permutations of the parts, is the distribution over predicted by the neural network ;

[0130] The objective function is defined as the expected reward, i.e.

[0131]

[0132] The gradient of the objective function is then computed as follows:

[0133]

[0134] where, is the batch size for training, is the number of disassembly sequence solutions generated for each subproblem, is the objective function value for the th disassembly sequence of the th subproblem, is the baseline, is the th part feature input for the th subproblem , the output is the probability of the disassembly sequence , the training process then optimizes the parameters using gradient ascent method:

[0135]

[0136] where, is the learning rate, a positive scalar that controls the step size of the parameter adjustment at each update, the learning rate determines how much the algorithm moves in the direction of the gradient of the loss function, the selection of the policy gradient baseline adopts Central self-critic, the calculation formula of the baseline is as follows:

[0137]

[0138] wherein represents the reward obtained by using the greedy strategy to obtain the disassembly sequence, represents the reward of the disassembly sequence obtained by using the strategy randomly sampled.

[0139] Step S40: The output result is selected by using the entropy weight approximation ideal solution sorting method to select a compromise optimal disassembly sequence.

[0140] In specific implementation, the trained model can obtain the Pareto frontier optimal solution in a short time after being given the input part input feature information and the part priority constraint matrix, and therefore the entropy weight TOPSIS evaluation and decision method is selected to select a compromise optimal solution as the final disassembly sequence scheme.

[0141] Each solution in the Pareto frontier has three index values, i.e., the number of disassembly operation direction changes, the number of disassembly tool changes, and the risk level. In order to eliminate the dimensional differences between different indexes, the range standardization method is used to normalize the data, so that the index data is unified to the same scale, so as to ensure the accuracy and reliability of data analysis.

[0142]

[0143] wherein, represents the jth index value corresponding to the ith Pareto frontier solution.

[0144] The entropy weight method is an objective index weight assignment method, which can calculate the weight of each index by analyzing the observation value of each index. The steps are as follows:

[0145] Step 1: assuming that the battery pack has n Pareto frontier solutions, m part indexes, and the original data matrix is obtained according to the above index quantification method

[0146] Step 2: calculate the contribution degree matrix of each part state index map the value of each index to 0-1:

[0147] ​​​​​

[0148] Step 3: Calculate the information entropy of the first index, which represents the total contribution of the index:

[0149]

[0150] Step 4: Calculate the normalized difference coefficient:

[0151]

[0152] Step 5: Calculate the index weight:

[0153]

[0154] Step 6: To extract the best solution, continue to use the TOPSIS model to comprehensively score the fitness value of the Pareto frontier solution, and select the highest score as the best solution. The distance between the first solution and the positive ideal solution is calculated as shown in the formula, and the distance between the first solution and the negative ideal solution is calculated as shown below.

[0155]

[0156]

[0157] where, and are the minimum and maximum values of the first index (in the standardized matrix).

[0158] Step 7: Calculate the relative closeness, which is the part recycling benefit and sort it as follows. The larger the relative closeness, the closer the part is to the ideal solution and the farther it is from the negative ideal solution, so the recycling benefit is higher. Finally, according to the size, sort the recycling benefits of each part to determine the order of the recycling parts.

[0159] .

[0160] It should be noted that the embodiment is aimed at the problem of increased disassembly risk caused by the complex internal structure of power batteries, high coupling degree, and non-continuity of disassembly operation. A multi-objective disassembly sequence planning method integrating risk perception is proposed. Since the battery pack is a dangerous recycling item, the disassembly process not only needs to pursue maximum efficiency, but also needs to ensure safety. The embodiment innovatively constructs a multi-objective disassembly sequence planning mathematical model integrating risk perception, and designs a battery disassembly sequence real-time planning algorithm based on graph attention network and constraint perception pointer network to meet the potential needs of real-time disassembly sequence planning in the actual recycling and disassembly scene of retired power batteries.

[0161] Compared with traditional disassembly sequence planning technology, the method not only introduces disassembly risk priority constraints and multi-objective optimization models, effectively reducing the safety risks caused by improper disassembly sequence, but also guarantees disassembly efficiency. In addition, the method emphasizes the advantages of real-time solving. Through the algorithm design based on graph attention network (GAT) and constraint perception pointer network, the method can realize end-to-end real-time planning and quickly output non-dominated disassembly sequence solutions. This real-time solving capability ensures that in a dynamic disassembly scene, the model can adjust and optimize the disassembly sequence in real time according to the current disassembly state and constraint conditions, thereby better adapting to complex disassembly environments and improving the flexibility and response speed of the disassembly process.

[0162] By introducing the disassembly risk prior knowledge provided by the battery manufacturer, the embodiment establishes a multi-objective disassembly sequence planning mathematical model integrating risk perception, and adopts a disassembly sequence planning algorithm based on graph attention network and constraint perception pointer network to realize end-to-end real-time planning of disassembly sequences for different types of battery packs. Not only can the prior disassembly knowledge of official manuals be effectively utilized to ensure the efficiency and safety of the disassembly process, but also the complex battery pack structure and dynamic disassembly scene can be adapted to, providing a new solution for the disassembly of retired power batteries.

[0163] Reference Figure 4 , Figure 4 The structure block diagram of the first embodiment of the automobile retired power battery disassembly sequence optimization device of the application is shown in FIG. 1.

[0164] As shown in FIG. 2, the automobile retired power battery disassembly sequence optimization device proposed in the embodiment of the application comprises: Figure 4

[0165] The matrix generation module 10 is used to generate an extended part precedence graph by taking the battery pack part number as a node according to the disassembly task specification and operation manual provided by the manufacturer, so as to generate a precedence relationship matrix;

[0166] The model construction module 20 is used to establish an optimization model containing a disassembly direction transition number expression and a disassembly tool replacement number expression.​

[0167] The data input module 30 is configured to input the priority relationship matrix and the part feature matrix into a pointer network model based on a graph attention network and a dynamic constraint mask attention mechanism, and obtain an output result.

[0168] The sequence determination module 40 is configured to select a compromise optimal disassembly sequence by using an entropy weight technique for order preference by similarity to ideal solution (TOPSIS) on the output result.

[0169] It should be understood that the above is only for illustration, and does not constitute any limitation on the technical solutions of the present application. In specific applications, those skilled in the art can set up as needed, and the present application does not limit this.

[0170] The embodiment introduces the disassembly risk priori knowledge provided by the battery manufacturer, establishes a multi-objective disassembly sequence planning mathematical model integrated with risk perception, and uses a disassembly sequence planning algorithm based on a graph attention network and a constraint perception pointer network to realize end-to-end real-time planning of disassembly sequences of different types of battery packs. Not only can the priori disassembly knowledge of the official manual be effectively utilized to ensure the efficiency and safety of the disassembly process, but also the complex battery pack structure and the dynamically changing disassembly scene can be adapted to provide a new solution for the disassembly of retired power batteries.

[0171] It should be noted that the above-described workflow is only illustrative and does not limit the scope of protection of the present application. In actual applications, those skilled in the art can select part or all of them to achieve the purpose of the embodiment scheme according to actual needs, which is not limited herein.

[0172] In addition, technical details not described in detail in the embodiment can be referred to the method for optimizing the disassembly sequence of the retired power battery of the automobile provided by any embodiment of the present application, which will not be described here.

[0173] In addition, it should be noted that in this document, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or system. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or system including the element.

[0174] The above sequence numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0175] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, also can be through hardware, but in many cases the former is the better embodiment. Based on such understanding, the technical solutions of the present application essentially or say the part of the contribution to the prior art can be embodied in the form of software products, the computer software product is stored in a storage medium (such as read-only memory (ReadOnly Memory, ROM) / RAM, disk, optical disc), including a number of instructions to make a terminal device (may be a mobile phone, computer, server, or network equipment, etc.) executes the method of each embodiment of the present application. The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, any equivalent structure or equivalent process transformation made by using the contents of the present application specification and drawings, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for optimizing the disassembly sequence of retired automotive traction battery packs, characterized in that, The method comprises the steps of: According to the disassembly task instructions and operation manual provided by the manufacturer, the battery pack part number is taken as a node, and an extended part precedence graph is generated to generate a precedence relationship matrix; An optimization model containing a disassembly direction conversion frequency expression and a disassembly tool replacement frequency expression is established; A pointer network model based on a graph attention network and a dynamic constraint mask attention mechanism is used to input the precedence relationship matrix and a part feature matrix and obtain an output result; The entropy weight TOPSIS method is used to select a compromise optimal disassembly sequence from the output result. The step of establishing the optimization model containing the disassembly direction conversion frequency expression and the disassembly tool replacement frequency expression comprises: Establishing a disassembly process direction change function , a disassembly process tool change function , and a disassembly process risk assessment function ; According to the disassembly process direction change function, the disassembly process tool replacement function and the disassembly process risk assessment function, an optimization model based on the disassembly strategy is generated, wherein the disassembly of the retired power battery is composed of one part, a disassembly sequence of the retired power battery, satisfies that each part is disassembled only once and meets the priority constraint relationship between the parts, is represented as the part disassembled in the first disassembly sequence. The disassembling process direction transformation function comprises: wherein a one-hot encoding is used Three directions are defined as , and to represent the operation direction in the disassembly process as up and down, left and right, front and back; The disassembling process tool changing function : wherein a one-hot encoding is used Three tools are defined as , and to represent the tools wrench, pliers, screwdriver in the disassembly process.

2. The method of claim 1, wherein, Dismantling process risk assessment function comprises: Where s represents a part risk coefficient.

3. The method of claim 1, wherein, The optimization model is represented as: wherein, and the weight vector each element in which has been normalized and added up to 1.

4. The method of claim 1, wherein, In the pointer network model based on the graph attention network and the dynamic constraint mask attention mechanism, the following steps are further included: A double-layer encoder is used to process part features, a gated recurrent unit network is used to aggregate sequential information, and then the part features are embedded into a high-dimensional space through a graph attention network, and an attention mechanism with a dynamic mask mechanism is used in the decoder to generate a disassembly sequence.

5. The method of claim 1, wherein, Before the step of inputting the precedence relationship matrix and the part feature matrix and obtaining the output result, the following steps are further included: The model parameters are optimized through a central self-criticism reinforcement learning algorithm, and iterative learning is performed based on a policy gradient update rule.

6. The method of claim 5, wherein, The step of optimizing the model parameters through the central self-criticism reinforcement learning algorithm and performing iterative learning based on the policy gradient update rule comprises: The disassembly sequence planning multi-objective mathematical model is standardized as a symbol of reinforcement learning problem, for the state space, for the action space, each state is defined as the set of all parts that have been disassembled at time t, i.e. , action is defined as the selection of the part to be disassembled at the next time step t+1 ; The policy is defined as , representing a set of part states that have been disassembled given the time instant The policy returns a probability distribution over the next disassembled part, where denotes the trainable parameters of the neural network; The expected reward is defined as the negative value of the sub-problem objective function: where is the space of part sets, is all possible permutations of the space, is the distribution of predicted by the neural network; Objective function is defined as the expected reward, i.e. The gradient of the objective function is calculated as follows: where, is the batch size of training, is the number of disassembly sequence solutions generated for each sub-problem, is the objective function value of the th disassembly sequence for the th sub-problem, is the baseline, is the objective function value of the th disassembly sequence for the th sub-problem given the part feature input The probability of outputting the disassembly sequence is then optimized using the gradient ascent method. wherein, is a learning rate, a positive scalar that controls the step size of the parameter adjustment at each update, the learning rate determines the magnitude of the steps the algorithm takes in the direction of the gradient of the loss function, the selection of the policy gradient baseline employs Central self-critic, the baseline is calculated as: wherein represents the reward obtained by using the greedy strategy to obtain the disassembly sequence, represents the reward obtained by using the strategy obtained by randomly sampling the disassembly sequence.

7. An automobile retired power battery disassembly sequence optimization device, characterized in that, The method of claim 1 is executed, comprising: A matrix generation module is configured to generate a precedence relationship matrix by taking a battery pack part number as a node and generating an extended part precedence graph according to disassembly task instructions and operation manual provided by a manufacturer. A model construction module is configured to establish an optimization model containing a disassembly direction conversion frequency expression and a disassembly tool replacement frequency expression. A data input module is configured to input the precedence relationship matrix and a part feature matrix into a pointer network model based on a graph attention network and a dynamic constraint mask attention mechanism, and obtain an output result. A sequence determination module is configured to select a compromise optimal disassembly sequence from the output result using an entropy weight TOPSIS method.

Images