Battery disassembly flexible scheduling method and system based on MOE

Abstract

The application relates to the technical field of intelligent manufacturing and industrial scheduling, and discloses a battery disassembly flexible scheduling method and system based on MOE. A feature processing module with multiple branches of input and single output is constructed based on position coding and embedding table, a multi-layer perception mechanism, and multi-layer self-attention coding; a decision backbone network is constructed based on dynamic routing, an expert network, and weighted attention; an output channel is constructed based on multi-layer self-attention decoding and a multi-layer perception mechanism; and a scheduling model is constructed based on the feature processing module, the decision backbone network, and the output channel; the disassembly process data of the retired power battery is acquired, is input into the scheduling model to acquire a corresponding operation sequence, and the disassembly of the retired power battery is completed based on the operation sequence, so that the problem that the existing retired power battery recycling and disassembly scheme cannot realize accurate scheduling is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing and industrial scheduling, and particularly relates to a battery disassembly flexible scheduling method and system based on MOE. BACKGROUND

[0002] With the rapid development of the new energy vehicle industry, the recycling and disassembly of retired power batteries have increasingly become a key link for resource recycling and environmental protection. The battery disassembly process usually includes multiple steps such as battery module identification, process judgment, structure disassembly, and resource recycling, involving complex process flow, various operation types, and heterogeneous resource states. For this scenario, how to achieve efficient and flexible scheduling optimization while meeting multiple target requirements such as efficiency, energy consumption, and safety is a core technical problem that needs to be solved at present.

[0003] Existing technical solutions include heuristic and meta-heuristic methods, reinforcement learning methods, and graph neural network and attention mechanism methods. For the above methods, there are many problems such as difficulty in efficiently integrating multi-source heterogeneous scheduling information, weak multi-objective optimization capability, focusing on single or static weight targets, lack of dynamic expert strategy and online adaptive feedback mechanism, low model modularization degree, and inability to flexibly adjust the strategy structure according to task changes. SUMMARY

[0004] The present application provides a battery disassembly flexible scheduling method and system based on MOE to solve the problem that existing retired power battery recycling and disassembly solutions cannot achieve precise scheduling.

[0005] To achieve the above purpose, the present application realizes the following technical solutions:

[0006] In a first aspect, the present application provides a battery disassembly flexible scheduling method based on MOE (Mixture-of-Experts), comprising the following steps:

[0007] Step 1: Based on position coding and embedding table, multi-layer perception mechanism combined with multi-layer self-attention coding, a multi-branch input single-output feature processing module is constructed, based on dynamic routing, expert network and weighted attention, a decision backbone network is constructed, based on multi-layer self-attention decoding and multi-layer perception mechanism, an output channel is constructed, based on the feature processing module, the decision backbone network and the output channel, a scheduling model is constructed;

[0008] Among them, the expert network is a decision network structure constructed based on the mixed expert gate control;

[0009] Step 2: Obtain the disassembly process data of the retired power battery, input the scheduling model to obtain the corresponding operation sequence, and complete the disassembly of the retired power battery based on the operation sequence.

[0010] Through the above design, efficient alignment and expression of multi-modal features are realized, the adaptability of the scheduling system in the multi-objective optimization scene is improved, and the self-adaptive optimization capability based on real-time feedback is also possessed, which can dynamically adjust the expert weight and routing parameter according to the actual execution effect, thereby effectively improving the scheduling efficiency and resource utilization, and reducing the energy consumption and default risk.

[0011] Further, the job order sequence output by the scheduling model is used as input data in the next round of disassembly of the retired power battery together with the disassembly process data.

[0012] Through the above design, the feedback collection and strategy updating mechanism is introduced, which can dynamically adjust the expert weight and routing parameter according to the real-time scheduling effect (such as completion time, default rate, energy consumption, etc.), thereby realizing closed-loop self-optimization, enhancing the long-term stability and adaptability of the system, and being suitable for continuous deployment in dynamic production environment.

[0013] Further, the feature processing module adopts position encoding and embedding table and multi-layer perception mechanism for mapping processing based on the data categories of the input disassembly process data and the job order sequence.

[0014] Further, the disassembly process data includes job process parameters, machine resources, disassembly time and disassembly energy consumption, and the data categories include discrete data and continuous data.

[0015] The feature processing module adopts position encoding and embedding table for mapping processing of the job order sequence and the job process parameters.

[0016] The feature processing module adopts multi-layer perception mechanism for mapping processing of the machine resources, disassembly time and disassembly energy consumption.

[0017] Through the above design, the feature processing module is constructed by position encoding, embedding table and multi-layer perception mechanism, and the data categories of each data in the disassembly process data are processed accordingly, ensuring that all input features are mapped to a unified high-dimensional feature space, and fundamentally solving the problem of insufficient feature modeling dimension and weak fusion effect in traditional methods.

[0018] The feature processing module integrates multiple source elements such as jobs, resources and environment, adopts a unified feature embedding and encoding method, combines dynamic activation and attention fusion of the expert network based on the dynamic routing mechanism, effectively considers multiple scheduling objectives such as job delay, deadline, energy consumption, etc., and significantly improves the flexibility and multi-objective collaborative optimization capability of scheduling decision.

[0019] Further, the feature processing module adopts multi-layer self-attention encoding to output first fusion features for the mapping-processed disassembly process data and job order sequence.

[0020] Further, the decision backbone network determines the expert values corresponding to the activated expert pairs in the constructed expert network set based on the first fusion feature, dynamic routing and the expert network, maps the plurality of expert values to obtain a plurality of value vectors and a key vector, linearly transforms the first fusion feature to obtain a query vector, performs attention weighting calculation based on the value vector, the key vector and the query vector to obtain a second fusion feature, and performs residual addition and first normalization processing on the expert score vector and the second fusion feature to obtain a third fusion feature.

[0021] The expert network set comprises a plurality of experts and corresponding hyperparameters, and each expert designs a battery disassembly scheme for different scheduling targets.

[0022] Through the above design, each expert is designed for different optimization targets, and the outputs thereof are fused through an attention mechanism, so that the scheduling efficiency can be ensured while the energy consumption control and the default risk avoidance are taken into account, and the problem that the traditional method is difficult to balance multiple scheduling targets can be effectively overcome.

[0023] Further, the calculation of the expert values corresponding to the plurality of experts comprises: linearly transforming the first fusion feature combined with the dynamic routing to obtain an expert score vector, obtaining an expert activation probability based on the expert score vector combined with the second normalization, determining a plurality of activated experts based on the expert activation probability and the expert network set, and determining the expert values based on the plurality of activated experts, the corresponding expert activation probability and the first fusion feature.

[0024] Further, the output channel decodes the third fusion feature to obtain a decoding feature through self-attention decoding, obtains a scheduling score through a multi-layer perception mechanism based on the decoding feature, and determines a job order sequence based on the scheduling score combined with an Argmax or a sampling strategy.

[0025] Further, the scheduling model constructs a multi-objective reward function based on the total job completion time, the deadline default degree and the energy consumption, and reversely updates the network parameters of the expert network based on the multi-objective reward function combined with a policy gradient method.

[0026] In a second aspect, the present application further provides a battery disassembly flexible scheduling system based on MOE, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the method according to any one of the above aspects when executing the computer program.

[0027] Advantages:

[0028] The MOE-based flexible scheduling method and system for battery disassembly provided by the application can realize unified embedding and coding of multi-source heterogeneous characteristics in a battery disassembly operation process, automatically generate an efficient scheduling sequence through dynamic routing selection and expert network output fusion, and continuously optimize the scheduling strategy in combination with an online feedback mechanism;

[0029] In the decision backbone network, by introducing a dynamic routing-based expert network, the optimal expert can be selected in the expert network according to the characteristics of the scheduling task, different scheduling targets (such as minimizing energy consumption, reducing delay, and improving resource utilization) can be targetedly optimized, and the generalization ability and flexibility of the scheduling strategy in different scenarios are improved.

[0030] Compared with the prior art, the application has significant improvement in multi-objective scheduling optimization, task adaptability, feature fusion expression, and continuous optimization capability, and is suitable for intelligent manufacturing and battery disassembly scenarios with dynamic complexity and multi-objective constraints, and has high engineering application value and promotion potential. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 The network structure of the scheduling model of the embodiment of the application is shown. DETAILED DESCRIPTION

[0032] The technical solutions of the application will be described clearly and completely below. Obviously, the described embodiments are only some of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0033] Unless otherwise defined, the technical terms or scientific terms used in the application should be understood as the usual meanings understood by those skilled in the art to which the application belongs. The terms “first”, “second”, and similar terms used in the application do not represent any order, quantity, or importance, but are only used to distinguish different components. Similarly, “one” or “a” and similar terms do not represent a quantity limit, but represent the existence of at least one. The terms “connected” or “connected” and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms “up”, “down”, “left”, “right”, and the like only represent relative positional relationships, which change when the absolute position of the described object changes.

[0034] The embodiment of the application provides a MOE-based flexible scheduling method for battery disassembly, which comprises the following steps:

[0035] Step 1: based on the position coding and embedding table, the multi-layer perception mechanism is combined with the multi-layer self-attention coding to construct a multi-branch input single-output feature processing module, based on the dynamic routing, the expert network and the weighted attention to construct the decision backbone network, based on the multi-layer self-attention decoding and the multi-layer perception mechanism to construct the output channel, based on the feature processing module, the decision backbone network and the output channel to construct the scheduling model;

[0036] Step 2: obtain the disassembly process data of the retired power battery, input the scheduling model to obtain the corresponding operation sequence, and complete the disassembly of the retired power battery based on the operation sequence.

[0037] Please refer to Figure 1 After the disassembly process data is input into the scheduling model, the feature processing module uses the position coding and embedding table and the multi-layer perception mechanism to map and process the data categories of the input disassembly process data and the operation sequence. Specifically, the disassembly process data includes operation process parameters, machine resources, disassembly time and disassembly energy consumption. The feature processing module uses the position coding and embedding table to map and process the discrete data of the operation sequence and the operation process parameters. The feature processing module uses the multi-layer perception mechanism to map and process the continuous data of the machine resources, disassembly time and disassembly energy consumption, so as to realize the extraction of the high-dimensional vector of the deep expression of the disassembly process data.

[0038] After mapping and processing, the feature processing module splices the high-dimensional vectors corresponding to the mapped and processed disassembly process data and the operation sequence to obtain high-dimensional fusion features, and outputs the first fusion features based on the high-dimensional fusion features using the multi-layer self-attention coding, which realizes the efficient alignment and deep fusion of multi-source heterogeneous features, and provides an accurate and solid feature basis for the subsequent combination of the dynamic routing expert network;

[0039] The decision backbone network determines the expert values corresponding to a plurality of activated experts in the constructed expert network set based on the first fusion features combined with the dynamic routing and the expert network.

[0040] The calculation of the expert values corresponding to the plurality of experts includes: obtaining an expert score vector based on the linear transformation of the first fusion features combined with the dynamic routing, obtaining an expert activation probability based on the expert score vector combined with the second normalization (Softmax normalization), the expert activation probability reflects the relative weight of each expert being activated, and the first fusion features combined with the weighted calculation are used to obtain the expert values corresponding to the plurality of activated experts and their corresponding weights, and then the subsequent attention feature fusion process is performed to obtain the second fusion features and the third fusion features. The output expert value is represented by the following formula:

[0041] ;

[0042] wherein, represents a set composed of several expert values outputted; represents the weight of the th activated expert; represents the output result of the expert network, i.e. the expert value.

[0043] wherein, the expert network set includes several experts and corresponding hyperparameters, and each expert designs a battery disassembly scheme for different scheduling targets;

[0044] Through the above dynamic expert selection and activation mechanism based on dynamic routing, the most suitable expert for the current scheduling scene can be activated adaptively according to the input first fusion feature, and the specific advantages of different experts can be fully exerted, so as to effectively improve the flexibility and optimization effect of the multi-objective scheduling problem.

[0045] Based on the mapping of several expert values, several value vectors and key vectors are obtained, the first fusion feature is linearly transformed to obtain a query vector, attention weighting calculation is performed based on the value vector, the key vector and the query vector to obtain a second fusion feature, and the expert score vector and the second fusion feature are added in residual and first normalized to obtain a third fusion feature;

[0046] Specifically, the key vectors and the value vectors of all experts are stacked to form complete key matrices and value matrices, then the scaling dot product attention score of the query vector and the key matrix is calculated, the second fusion feature is obtained by weighting fusion according to the attention score and the value matrix, and the third fusion feature is obtained by adding the expert score vector and the second fusion feature in residual and first normalization (LayerNorm normalization);

[0047] Finally, the output channel adopts self-attention decoding to decode the third fusion feature to obtain a decoding feature, and a scheduling score is obtained by combining the decoding feature with a multi-layer perception mechanism, and a job order sequence is determined based on the scheduling score combined with an Argmax or a sampling strategy;

[0048] The scheduling model constructs a multi-objective reward function based on the total job completion time, the deadline violation degree and the energy consumption, and the reward function is represented by the following formula:

[0049] ;

[0050] wherein, represents the reward value outputted; represents the total job completion time; represents the deadline violation degree; represents the energy consumption, respectively, are used to balance the importance of each indicator.

[0051] The network parameters of the expert network are updated in reverse based on the reward value output by the reward function combined with the policy gradient method. The specific loss function defined can be represented as:

[0052]

[0053] wherein, represents the probability of the model generating the selected job sequence according to the current policy, represents the current selected job order index sequence. Here, only the top 2 jobs with the highest probability are selected as the candidate order to generate the final scheduling sequence. In this way, while ensuring that high-probability jobs are prioritized, the search space is controlled and the computational complexity is reduced. represents the selected probability distribution generated by the model for each job during decoding, also known as the scheduling distribution or route probability. It reflects the model's prediction of the likelihood of each job being arranged in the sequence under the current state, and is the basis for calculating the gradient in the policy gradient method;

[0054] For the model training of the scheduling model, the Adam optimizer is selected to accelerate the training, and the learning rate is initialized to 0.001, with the ability to adaptively adjust. In addition, an experience replay pool is initialized, and its size is set to 10000, from which 64 experience samples are randomly sampled for each training batch. This allows the scheduling model to obtain diverse training data during the learning process and avoid overfitting;

[0055] During the training process, the epsilon-greedy exploration strategy is used to balance exploration and exploitation. The initial exploration rate (epsilon) is set to 1.0, allowing the model to explore more to help it learn a variety of possible strategies. As the training progresses, epsilon gradually decays to 0.01, ensuring that the scheduling model gradually relies more on learned knowledge during the learning process. The decay of the exploration rate uses an exponential decay strategy, i.e., every time a training cycle is completed, epsilon decreases by 5%. This strategy helps the scheduling model avoid falling into a local optimal solution too early and maintains a certain exploration space throughout the training process;

[0056] ​To break the time correlation of the training data and enhance the stability of the training, an experience replay mechanism is adopted. In each training step, the model stores the current state, action, reward, and next state into the experience replay pool. The size of the replay pool is set to 10000, ensuring that enough experiences are stored for training. In each training, 64 samples are randomly selected from the replay pool for training, which can increase the diversity of training, avoid over-reliance on the time sequence of data, and improve the stability and convergence speed of training.

[0057] During training, the incentive function mainly considers three aspects: job delay, energy consumption, and deadline violation. For example, reducing job delay will increase the reward by 10 points each time, and reducing 1 kWh of energy consumption will increase the reward by 5 points. For the reward of each task, these incentive items are weighted and summed to form the total reward. Then, the Q-learning algorithm is used to update the Q-value function. The learning rate is set to 0.01, and the discount factor is 0.9, which ensures that the system can gradually optimize the strategy according to the feedback and adjust the Q-value according to the Bellman equation. Each time the Q-value function is updated, the system ensures that it gradually approaches the optimal strategy, effectively optimizing task scheduling.

[0058] In the main training loop, the model iterates multiple rounds through continuous interaction with the environment, gradually improving scheduling accuracy. In each round, a action is selected based on the current strategy, the action is executed and the reward is calculated, and the current state, action, reward, and next state are stored in the experience replay pool. After each training period, the strategy is updated and the Q-value is calculated using Q-learning to optimize the policy network. To further improve the performance of the scheduling model, the exploration rate is gradually reduced, making the model rely more on learned knowledge for decision-making. The training process will be repeated for multiple epochs (e.g., 500 epochs), and the model performance will be evaluated after each period, such as through cumulative returns, average delay, and other indicators for comprehensive evaluation. Based on the evaluation results, the policy network, learning rate, and exploration rate are adjusted to ensure that the scheduling model converges effectively and achieves the best scheduling effect.

[0059] Through these series of training steps, the system can gradually optimize the scheduling decisions for battery disassembly tasks, and gradually improve the scheduling accuracy and resource utilization in handling multi-objective optimization problems such as job delay, deadline violation, and energy consumption.

[0060] The scheduling model effectively reduces the average job completion time, reduces the deadline violation rate, and reduces the overall energy consumption. By introducing a dynamic routing expert network, the scheduling decision accuracy and response speed are improved while ensuring reasoning efficiency. Based on global fused features and expert knowledge, efficient sequence modeling and decision output are achieved, and intelligent and flexible scheduling plans are automatically generated, further improving the scheduling performance and generalization ability of the system.

[0061] Step 3: The job order sequence output by the scheduling model is used as input data in the next round of disassembly of the retired power battery, together with the disassembly process data, to reacquire a new job order sequence.

[0062] Finally, to verify the battery disassembly flexible scheduling method and scheduling model based on MOE proposed in the present application, a comparison is made with the traditional optimization method and the traditional reinforcement learning method, focusing on the disassembly time, disassembly cost, job delay, and system stability. For details, please refer to Table 1.

[0063] The traditional optimization method includes genetic algorithm (GA) and particle swarm optimization (PSO).

[0064] Table 1: Comparison table of the present application and traditional optimization method and traditional reinforcement learning method.

[0065]

[0066] Based on Table 1, in terms of disassembly time, the present application reduces about 40% compared with the traditional method. Through dynamic adjustment of job order and reasonable resource scheduling, the system can significantly reduce idle time and waiting time, thereby accelerating the overall disassembly process.

[0067] In terms of disassembly cost, the present application reduces the disassembly cost by about 28% through the optimization strategy of reinforcement learning and expert network. In contrast, the traditional method lacks fine resource control and scheduling strategy, resulting in high energy consumption and resource consumption.

[0068] In terms of job delay, the present application significantly reduces job delay by about 65% through intelligent scheduling. Especially in complex task scheduling, the system can adjust the strategy in real time according to environmental changes to ensure that the job is completed on time.

[0069] From the test results, it can be seen that the battery disassembly flexible scheduling method based on MOE surpasses the traditional method in many aspects. Not only does it significantly improve the disassembly efficiency and reduce the execution time, but also optimizes the energy consumption through precise resource scheduling and reduces the disassembly cost. At the same time, the job delay is greatly reduced, making the task more efficiently completed.

[0070] Compared with the traditional method, the battery disassembly flexible scheduling method based on MOE proposed in the present application can better handle complex multi-objective optimization problems. Through the expert network and hybrid attention mechanism, it flexibly adapts to environmental changes and makes real-time adjustments. This feature makes the system have great potential for application in actual production environment, especially suitable for dynamic and variable battery disassembly tasks.

[0071] The preferred embodiments of the present application have been described above in detail. It should be understood that modifications and variations to the present application can be affected by those skilled in the art without departing from the scope of the application. Accordingly, it is intended that all of the subject matter of the above description and the claims be interpreted to encompass all such modifications and changes.

Claims

1. A MOE-based flexible scheduling method for battery disassembly, characterized in that, The method comprises the following steps: Step 1: constructing a multi-branch input single-output feature processing module based on position encoding and embedding table, multi-layer perception mechanism and multi-layer self-attention encoding, constructing a decision backbone network based on dynamic routing, expert network and weighted attention, constructing an output channel based on multi-layer self-attention decoding and multi-layer perception mechanism, and constructing a scheduling model based on the feature processing module, the decision backbone network and the output channel; Step 2: obtaining disassembly process data of the retired power battery, inputting the scheduling model to obtain a corresponding operation sequence, and completing the disassembly of the retired power battery based on the operation sequence; The operation sequence output by the scheduling model is used as input data in the next round of disassembly of the retired power battery together with the disassembly process data; The feature processing module adopts position encoding and embedding table and multi-layer perception mechanism for mapping processing based on the data categories of the input disassembly process data and operation sequence; The disassembly process data comprises operation process parameters, machine resources, disassembly time and disassembly energy consumption, and the data categories comprise discrete data and continuous data; The feature processing module adopts position encoding and embedding table for mapping processing of the operation sequence and operation process parameters; The feature processing module adopts multi-layer perception mechanism for mapping processing of the machine resources, disassembly time and disassembly energy consumption; The feature processing module adopts multi-layer self-attention encoding for outputting first fusion features of the mapping-processed disassembly process data and operation sequence; The decision backbone network determines the expert values corresponding to a plurality of activated experts in the constructed expert network set based on the first fusion features, dynamic routing and expert network, maps the plurality of expert values to obtain a plurality of value vectors and key vectors, linearly transforms the first fusion features to obtain a query vector, performs attention weighting calculation based on the value vectors, key vectors and query vector to obtain second fusion features, and performs residual addition and first normalization processing on the expert score vector and the second fusion features to obtain third fusion features; The expert network set comprises a plurality of experts and corresponding hyperparameters, and each expert designs a battery disassembly scheme for different scheduling targets.

2. The MOE-based flexible scheduling method for battery teardown according to claim 1, wherein, The calculation of the expert values corresponding to the plurality of experts comprises: linearly transforming the first fusion features based on the dynamic routing to obtain an expert score vector, obtaining an expert activation probability based on the expert score vector and the second normalization, determining a plurality of activated experts based on the expert activation probability and the expert network set, and determining the expert values based on the plurality of activated experts, the corresponding expert activation probability and the first fusion features.

3. The MOE-based flexible scheduling method for battery teardown according to claim 1, wherein, The output channel decodes the third fusion features to obtain decoding features by using self-attention decoding, obtains scheduling scores by using multi-layer perception mechanism on the decoding features, and determines the operation sequence based on the scheduling scores and Argmax or sampling strategy.

4. The MOE-based battery disassembly flexible scheduling method according to any one of claims 1-3, characterized in that, The scheduling model constructs a multi-objective reward function based on total operation completion time, deadline violation degree and energy consumption, and reversely updates the network parameters of the expert network based on the multi-objective reward function and the policy gradient method. 5.A MOE-based flexible scheduling system for battery disassembly, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The processor implements the steps of the method of any one of claims 1-4 when executing the computer program.

Images