A disassembly task sequence recommendation method based on an LLM-ESR model
Patent Information
- Application Number
- CN202611087423.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-21
- Publication Date
- 2026-10-09
AI Technical Summary
[0007]为解决现有人机协同动力电池拆解过程中难以同时刻画拆解任务语义关系与历史拆解序列协同模式，以及历史拆解经验利用不足导致决策准确性和泛化能力受限的问题，本发明提出一种基于LLM-ESR的人机协同电池拆解任务序列推荐与任务分配方法
[0007]为解决现有人机协同动力电池拆解过程中难以同时刻画拆解任务语义关系与历史拆解序列协同模式，以及历史拆解经验利用不足导致决策准确性和泛化能力受限的问题，本发明提出一种基于LLM-ESR的人机协同电池拆解任务序列推荐与任务分配方法。所述方法首先利用大语言模型对退役动力电池结构文本信息进行编码，获得语义嵌入向量；随后构建双视角模块，从语义视角和协同视角对拆解任务关系进行建模，并通过双向交叉注意力机制实现特征融合；进一步利用具有参数共享机制的GRU4Rec序列编码器学习拆解任务之间的时序依赖关系以及任务与执行主体之间的匹配关系；同时构建检索增强自蒸馏模块，通过检索结构相似的历史退役动力电池样本获取经验知识，并用于模型优化。通过上述技术方案，本发明能够生成合理的拆解任务执行序列，实现拆解任务在人与协作机械臂之间的协同分配，提高复杂退役动力电池拆解场景下的决策准确性、执行效率。
Smart Images
Figure CN122885504A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power battery recycling technology, and in particular to a human-machine collaborative method for recommending and allocating dismantling task sequences for retired power batteries based on a large language model. Background Technology
[0002] With the transformation of the global energy structure, the continuous development of the new energy vehicle industry, and the sustained growth in the number of new energy vehicles, the number of retired power batteries will increase rapidly. If these batteries are not properly managed, they will result in the waste of materials such as lithium and cobalt, and cause environmental pollution. Therefore, how to efficiently recycle and utilize retired power batteries has become a current research hotspot.
[0003] Currently, the main methods for dismantling retired power batteries include manual dismantling, automated dismantling, and human-machine collaborative dismantling. Manual dismantling offers high flexibility and can adapt to complex battery structures, but it carries significant safety risks. Automated dismantling improves efficiency and consistency, but its adaptability and flexibility are limited due to the diverse structures of power batteries. In contrast, human-machine collaborative dismantling combines the flexibility of human labor with the stability of robots and is considered an effective way to achieve efficient dismantling of retired power batteries.
[0004] Existing research has explored the problem of human-machine collaborative dismantling task planning for retired power batteries. For example, a method for human-machine collaborative dismantling planning for retired power batteries has been proposed in the prior art, which optimizes the allocation of dismantling tasks by establishing task constraints. Existing technologies employ game theory and multi-agent deep reinforcement learning methods to achieve human-machine collaborative dismantling task planning for retired power batteries, improving decision-making capabilities in complex scenarios.
[0005] However, existing technologies still have certain limitations in the planning of retired power battery dismantling tasks. On the one hand, existing methods mainly rely on predefined rules, optimization models, or historical interaction data for decision-making, failing to fully utilize the semantic information contained in the battery structure description text, and making it difficult to effectively mine the structural characteristics between different types of retired power batteries and their impact on the dismantling task. On the other hand, most existing methods focus on single-perspective modeling, making it difficult to simultaneously characterize the semantic relationships of the dismantling task and the collaborative patterns in historical dismantling sequences, thus limiting the model's ability to represent complex dismantling scenarios. In addition, in terms of utilizing historical dismantling experience, existing methods have relatively limited retrieval and transfer of dismantling experience from similar retired power batteries, making it difficult to fully leverage the guiding role of historical dismantling knowledge in current task decisions.
[0006] In recent years, Large Language Models (LLMs) have demonstrated superior performance in natural language understanding and reasoning, capable of extracting semantic information from unstructured text and providing decision-making solutions for complex industrial scenarios. Meanwhile, advancements in techniques such as retrieval reinforcement learning and knowledge distillation have offered new solutions for the efficient utilization of historical experience and the improvement of model generalization capabilities. Therefore, how to fully leverage the semantic understanding capabilities of LLMs, integrate structural information from retired power batteries with historical dismantling experience, construct a dismantling task sequence recommendation method, and achieve human-machine collaborative task allocation through the matching relationship between tasks and executors has become a pressing technical problem in the field of retired power battery dismantling. Summary of the Invention
[0007] To address the challenges of simultaneously characterizing the semantic relationships of dismantling tasks and the collaborative patterns of historical dismantling sequences in existing human-machine collaborative power battery dismantling processes, as well as the limitations in decision-making accuracy and generalization ability due to insufficient utilization of historical dismantling experience, this invention proposes a human-machine collaborative battery dismantling task sequence recommendation and task allocation method based on LLM-ESR. The method first encodes the structural textual information of retired power batteries using a large language model to obtain semantic embedding vectors. Then, a dual-view module is constructed to model the relationships between dismantling tasks from both semantic and collaborative perspectives, and feature fusion is achieved through a bidirectional cross-attention mechanism. Furthermore, a GRU4Rec sequence encoder with a parameter-sharing mechanism is used to learn the temporal dependencies between dismantling tasks and the matching relationships between tasks and executing entities. Simultaneously, a retrieval-enhanced self-distillation module is constructed to acquire experiential knowledge by retrieving structurally similar historical retired power battery samples, which is then used for model optimization. Through these technical solutions, this invention can generate reasonable dismantling task execution sequences, achieve collaborative allocation of dismantling tasks between humans and collaborative robotic arms, and improve decision-making accuracy and execution efficiency in complex retired power battery dismantling scenarios.
[0008] The technical solution of this invention is as follows: A method for decomposing task sequences and assigning tasks based on the LLM-ESR model includes the following steps: Step 1: Obtain textual descriptions of the target retired power battery and historical retired power battery dismantling datasets, and encode them into high-dimensional semantic vectors using LLM. The textual descriptions include structural descriptions of the retired power battery and descriptions of the dismantling task sequence and task allocation. Historical retired power battery dismantling data is used for model training, and the target retired power battery is the battery for which the recommended dismantling task sequence and task allocation are to be assigned.
[0009] Step 2: Construct a dual-view module. After the semantic vectors are mapped to dimensions and aligned to features by the adapter, a semantic embedding sequence is formed. At the same time, a co-embedding sequence is generated based on the historical decomposition task sequence.
[0010] Step 3: Input the semantic embedding sequence and the co-embedded sequence into the bidirectional cross-attention mechanism for feature fusion, and model the fused sequence through the GRU4Rec encoder with parameter sharing mechanism to obtain the joint sequence representation of the decomposition task; construct a sequence learning loss function based on pairwise ranking, and use the ranking relationship between the target task and the negative sample task as the supervision signal to train and optimize the dual-view module.
[0011] Step 4: Construct a retrieval enhancement self-distillation module to retrieve battery samples that are semantically similar to the target retired power battery from historical retired power battery samples; perform weighted aggregation on the joint sequence representations corresponding to the retrieved battery samples according to semantic similarity to generate a teacher representation; use the sequence representation corresponding to the target retired power battery as the student representation; and constrain the student representation to approximate the teacher representation by constructing a self-distillation loss function.
[0012] Step 5: After the model training is completed, input the target retired power battery into the optimized LLM-ESR model, and output the corresponding dismantling task sequence recommendation results and task allocation results.
[0013] Furthermore, the LLM described in step 1 uses a parameter freezing method for feature extraction. During model training, only the parameters of the adapter module, dual-view module, and retrieval enhancement self-distillation module are updated to preserve the semantic expressive power of the language model.
[0014] Preferably, the adapter in step 2 includes a first fully connected layer, a nonlinear activation function layer, and a second fully connected layer; wherein, the first fully connected layer is used to perform dimensionality reduction mapping on the high-dimensional semantic vector output by LLM, and the second fully connected layer is used to further adjust the feature distribution so that the mapped semantic embedding sequence and the co-embedded sequence are in a unified feature space.
[0015] A further preferred embodiment, the bidirectional cross-attention mechanism in step 3 includes: (1) an information interaction process from a semantic perspective to a collaborative perspective. This involves mapping the semantic embedding sequence to a query vector. Q Mapping co-embedded sequences to key vectors K Sum value vector V (2) The information interaction process from the collaborative perspective to the semantic perspective. The collaborative embedding sequence is mapped to the query vector. Q' Mapping semantic embedding sequences to key vectors K' Sum value vectorV' By scaling the dot product attention, the attention weight of the collaborative perspective on the semantic perspective is calculated, thereby realizing the reverse guidance of collaborative information on semantic features and thus obtaining a bidirectional fused feature representation.
[0016] Preferably, the GRU4Rec encoder with parameter sharing mechanism in step 3 refers to the use of the same set of GRU network parameters to encode the semantic viewpoint feature sequence and the collaborative viewpoint feature sequence, including updating gate parameters, resetting gate parameters and hidden state parameters, so as to ensure that different viewpoint features are sequence modeled in a unified feature space.
[0017] Preferably, the sequence learning loss function based on pairwise sorting in step 3 is used to constrain the model's prediction probability for the true subsequent decomposition task to be higher than the prediction probability for the negative sample task, and its calculation formula is as follows: L DVM =- log (P(t + a，k │S a，1：k-1 )-P(t - a，k │S a，1：k-1 )) Among them, S a，1：k-1 Indicates the retirement of power batteries a The task sequence consisting of the first k-1 historical dismantling tasks; t + a，k Indicates the actual next dismantling task; t - a，k This represents a task involving negative samples obtained through random sampling.
[0018] Preferably, the semantic similarity in step 4 is calculated using cosine similarity. The cosine similarity between the target decommissioned power battery and historical decommissioned power battery samples is calculated based on the joint sequence representation and then normalized to obtain the corresponding aggregation weights. Finally, the joint sequence representations corresponding to the retrieved similar battery samples are weighted and summed according to the weight values to generate the teacher representation.
[0019] Preferably, the self-distillation loss function in step 4 is used to measure the difference between the student representation and the teacher representation, and the mean squared error loss function is adopted, the calculation formula of which is as follows: L RASD =││Z target -u││ 2 2 Among them, Z target The student representation represents the target retired power battery, while the teacher representation is obtained by weighted aggregation of multiple similar historical retired power battery samples.
[0020] Preferably, during model training, a joint loss function is used to optimize the model, and its calculation formula is as follows: L=L DVM + L RASD in, The hyperparameter used to adjust the self-distillation loss weight is set to a value ranging from 0.1 to 1.0. Attached Figure Description
[0021] The technical solution of the present invention will become clearer and easier to understand through a detailed description of the embodiments of the present invention in conjunction with the accompanying drawings, wherein: Figure 1 This is a diagram illustrating the overall framework of the task sequence recommendation method based on the LLM-ESR model proposed in this invention. Figure 1 In this module, module A is a dual-view module, and module B is a search-enhanced self-distillation module.
[0022] Figure 2 This is a schematic diagram of the dual-view module mechanism in this invention; Figure 2 This includes both semantic and collaborative perspectives.
[0023] Figure 3 This is a schematic diagram of the structure of the GRU4Rec sequence encoder in this invention; Figure 3 The system includes an input layer, a GRU timing processing layer, and an output layer.
[0024] Figure 4 This is a schematic diagram of the retrieval enhancement self-distillation module in this invention; Figure 4 The process includes input, semantic embedding and similar battery screening, feature weighted fusion, and self-distillation loss calculation.
[0025] Figure 5 This is a schematic diagram of the bidirectional cross-attention mechanism in this invention; it includes a semantic embedding sequence, a co-embedding sequence, a query vector, a key vector, a value vector, and a bidirectional feature fusion module. Detailed Implementation
[0026] Taking the NIO ES8 NEV-P50 battery pack as an example, its disassembly tasks include the battery top shell, sealing plate, battery module, high-voltage wiring harness, BMS, thermal pad, cooling system, bolts, and battery box.
[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.
[0028] This invention proposes a human-machine collaborative battery dismantling task sequence recommendation method based on LLM-ESR, which mainly includes modules such as problem description, dual-view module, bidirectional cross-attention fusion, sequence encoding, and retrieval-enhanced self-distillation. Finally, it outputs the optimal dismantling task sequence for the target retired power battery and realizes human-machine collaborative task allocation based on the matching relationship between the task and the executing subject.
[0029] Description of the decision-making problem for the task sequence of dismantling retired power batteries: In this embodiment, a method for recommending dismantling task sequences and allocating tasks is proposed for the human-machine collaborative dismantling process of retired power batteries. To ensure operator safety, a sequential collaborative operation mode is adopted, meaning that the operator and the robotic arm work collaboratively in the same workspace, but do not perform dismantling tasks simultaneously. Only after the robotic arm completes its current dismantling task and exits the workspace can the operator enter the space to perform the corresponding dismantling operation. For retired power batteries, the dismantling process consists of a set of dismantling tasks, denoted as b={t1,t2,…,t…}. n}, where each disassembly task t i This represents a minimum indivisible unit of decomposition. Different decomposition tasks have priority relationships. Therefore, the task sequence and allocation problem can be defined as: determining a task execution sequence s={t} under all constraints. (1) ,t (2) ,…,t (n)}, where t (k) This represents the k-th disassembly task executed in the sequence.
[0030] In historical data, the battery set is B = {b1, b2, b3, ..., b}. │B│}, where |B| represents the number of historical batteries. Each battery b i The corresponding set of disassembly tasks is denoted as T. i ={t 1 i ,t 2 i ,…,t n i}, where n represents the total number of disassembly tasks for the battery. Each battery b i This corresponds to a sequence of task breakdowns and allocation results, and the sequence is represented as follows: This sequence reflects the actual disassembly steps. A historical disassembly dataset is thus constructed. .
[0031] For the target retired power battery Given its structural text description information Disassembly task set Given the historical disassembly dataset D, the goal of this embodiment is to predict the optimal disassembly task sequence and task allocation for the target battery while satisfying the constraints.
[0032] The generation of the decomposition sequence is modeled as a stepwise decision-making process. Given the first K-1 decomposition tasks, the most likely decomposition task t to be executed is predicted. k And its task allocation. Therefore, the conditional probability of the target battery disassembly sequence can be expressed as:
[0033] in This indicates the dismantling task that has been executed so far. , This represents the total number of target battery disassembly tasks. Ultimately, the overall disassembly sequence of the target battery can be represented as:
[0034] in, This represents the set of all feasible disassembly sequences that satisfy the disassembly priority constraint and the safety constraint.
[0035] Building an LLM-ESR framework like Figure 1 As shown, the entire framework mainly consists of two parts: a dual-view module and a retrieval-enhanced self-distillation module. The dual-view module mainly models the decomposed task sequence from the perspectives of semantic information and collaborative information, and achieves multi-dimensional information fusion through feature interaction. The retrieval-enhanced self-distillation module mainly uses the retrieval information of similar samples to perform self-supervised optimization of the model, thereby improving the model's representation ability and the accuracy of sequence recommendation.
[0036] Dual-view module: like Figure 2 As shown, the dual-view module is divided into semantic view modeling and collaborative view modeling. The semantic view aims to characterize the semantic relationships between decomposition tasks. The high-dimensional semantic vectors generated by LLM are input into a lightweight adapter, and dimensionality reduction mapping and feature alignment are achieved through a fully connected network to obtain semantic embedding vectors. Following the execution order of the historical decomposition tasks, the semantic embedding vectors are arranged to form a semantic embedding sequence. This is used to describe the semantic dependencies between tasks. A collaborative perspective is used to learn the temporal dependencies in the historical dismantling task sequence. Based on the dismantling task sequence of historical retired power batteries, a collaborative embedding matrix E is constructed. co Each decomposition task corresponds to a trainable embedding vector. .
[0037] Preferably, principal component analysis is used to reduce the dimensionality of the LLM semantic vectors, and this is used to initialize the co-embedding matrix to maintain consistency between the semantic and co-embedding perspectives and accelerate model convergence. For the historical disassembly sequence S of the battery... a The corresponding cooperative embedding sequence can be obtained by mapping. .
[0038] Bidirectional cross-attention fusion: like Figure 5 As shown, to achieve deep fusion of semantic and collaborative perspectives, this invention employs a bidirectional cross-attention mechanism, including fusion from the semantic perspective to the collaborative perspective and fusion from the collaborative perspective to the semantic perspective. First, the semantic embedding sequence is mapped to a query vector, and the collaborative embedding sequence is mapped to keys and values. The correlation between the two is calculated using scaled dot product attention to obtain a semantically guided collaborative feature representation. The calculation formula is as follows: , .
[0039] Simultaneously, the collaborative embedding sequence is used as the query vector, and the semantic embedding sequence is used as the key and value vector to complete the reverse information interaction, thereby obtaining the collaboratively guided semantic feature representation. The calculation formula is as follows: ,
[0040] The final fused semantic sequence is obtained With co-sequence This provides a unified feature representation for subsequent sequence modeling.
[0041] GRU4Rec sequence encoding: like Figure 3 As shown, the fused semantic sequence and co-sequence The input is a GRU4Rec sequence encoder with a parameter-sharing mechanism to capture the temporal dependencies between disassembly tasks. GRU4Rec is based on a gated recurrent unit, which controls the retention and forgetting of historical information by updating and resetting the gates, thereby effectively modeling long-term and short-term dependencies.
[0042] The sequence representations output by the encoder are denoted as follows: and In the recommendation phase, a logit-level fusion strategy is adopted to concatenate the representation vectors from the two perspectives to calculate the predicted probability of candidate decomposition tasks, and the model is optimized using a pairwise ranking loss function.
[0043] Search Enhanced Self-Distillation Module: like Figure 4As shown, to enhance the generalization ability of the model, this invention constructs a retrieval-enhanced self-distillation module. First, based on the cosine similarity between the semantic vector of the target retired power battery and the semantic vectors of historical batteries in the semantic knowledge base, the Top-K historical battery samples with the highest similarity are selected. Then, the sequence representations obtained from these similar samples in the dual-view module are weighted and aggregated to generate a prior knowledge representation, where the weights are obtained by normalizing the similarity using the softmax function. By minimizing the mean squared error loss between the target battery representation and the prior knowledge representation, a self-distillation loss function is constructed, and this loss function, together with the ranking loss from sequence recommendation, forms a joint loss function to optimize the model.
[0044] By using the optimized LLM-ESR model, the dismantling task sequence decision problem is transformed into the next task prediction problem, thereby outputting the recommended dismantling task sequence for the target retired power battery.
[0045] Experimental setup: The historical retired power battery dismantling data contains 60 sets of dismantling information. Of these, 50 sets are used as the training set for model training, and 10 sets are used as the test set for model performance verification. During model training, the GRU4Rec encoder has 3 layers, the number of similar historical retired power battery samples retrieved in the retrieval enhancement self-distillation module is set to 5, the self-distillation loss weight coefficient α is set to 0.5, the number of attention heads in the bidirectional cross-attention mechanism is set to 4, the learning rate is set to 0.001, and the number of training rounds is set to 100. After training, the structural description information of the target retired power battery is input into the trained LLM-ESR model for inference, resulting in a recommended dismantling task sequence: bolt dismantling, battery casing dismantling, sealing plate dismantling, high-voltage wiring harness dismantling, battery management system (BMS) dismantling, battery module dismantling, thermal pad dismantling, cooling system dismantling, and battery casing dismantling. The corresponding task allocation result is: RHHHHRHRR, where R represents the robot performing the corresponding dismantling task, and H represents the human performing the corresponding dismantling task. Experimental results show that the traditional manual disassembly method takes about 102 minutes to complete the entire disassembly process, while the method proposed in this invention takes about 80.5 minutes, improving the disassembly efficiency by about 21.1%. Beneficial effects
[0046] The LLM-ESR-based task sequence recommendation method proposed in this invention has the following advantages: (1) This invention utilizes LLM to semantically encode the structural description information and disassembly experience information of retired power batteries, fully explores the correlation between battery structural features and disassembly knowledge, and improves the model's semantic understanding ability for different types of retired power batteries.
[0047] (2) The present invention constructs a dual-perspective modeling mechanism to model the disassembly task from the semantic perspective and the collaborative perspective, and realizes the information interaction and feature fusion of the two perspectives through a bidirectional cross-attention mechanism, thereby simultaneously utilizing the semantic information of the battery structure and the historical disassembly behavior pattern to improve the accuracy of the disassembly task sequence recommendation.
[0048] (3) The present invention introduces a search-enhanced self-distillation mechanism, which retrieves historical battery samples that are semantically similar to the target retired power battery, and uses similar samples to generate teacher representations to guide the target sample, so that the model can effectively utilize historical dismantling experience knowledge.
[0049] Note: This invention does not involve biological materials; this is merely a reminder and requires no further action.
Claims
1. A human-computer collaborative method for task sequence decomposition recommendation and task allocation based on the LLM-ESR (Large Language Model - Enhanced Sequence Recommendation) model, characterized in that, Includes the following steps: Step 1: Obtain textual description information from the target retired power battery and historical retired power battery dismantling datasets, and encode it into high-dimensional semantic vectors using LLM. The textual description information includes structural descriptions of the retired power battery and descriptions of the dismantling task sequence and task allocation. Historical retired power battery dismantling data is used for model training, and the target retired power battery is the battery for which the recommended dismantling task sequence and task allocation are to be performed. Step 2: Construct a dual-view module. After dimension mapping and feature alignment of the semantic vectors through an adapter, a semantic embedding sequence is formed; simultaneously, a co-embedding sequence is generated based on the historical dismantling task sequence. Step 3: Input the semantic embedding sequence and the co-embedding sequence into a bidirectional cross-attention mechanism for feature fusion. The fused sequence is then modeled using a GRU4Rec encoder with a parameter sharing mechanism to obtain a joint sequence representation of the dismantling tasks. Simultaneously, human-machine collaborative task allocation is achieved through the matching relationship between tasks and the executing entity. A sequence learning loss function based on pairwise ranking is constructed, using the ranking relationship between the target task and the negative sample task as a supervision signal to train and optimize the dual-view module. Step 4: Construct a retrieval-enhanced self-distillation module to retrieve battery samples semantically similar to the target retired power battery from historical retired power battery samples; weight and aggregate the joint sequence representations corresponding to the retrieved battery samples according to semantic similarity to generate a teacher representation; use the sequence representation corresponding to the target retired power battery as the student representation; and construct a self-distillation loss function to constrain the student representation to approximate the teacher representation. Step 5: After model training is complete, input the target retired power battery into the optimized LLM-ESR model, and output the corresponding dismantling task sequence recommendation results and the human-machine collaborative dismantling task allocation results.
2. The method according to claim 1, characterized in that, The LLM described in step 1 uses parameter freezing for feature extraction. During model training, only the parameters of the adapter module, dual-view module, and retrieval enhancement self-distillation module are updated to preserve the semantic expressive power of the language model.
3. The method according to claim 1, characterized in that: The adapter described in step 2 includes a first fully connected layer, a non-linear activation function layer, and a second fully connected layer. The first fully connected layer is used to perform dimensionality reduction mapping on the high-dimensional semantic vector output by the LLM, and the second fully connected layer is used to further adjust the feature distribution so that the mapped semantic embedding sequence and the co-embedded sequence are in a unified feature space.
4. The method according to claim 1, characterized in that: The bidirectional cross-attention mechanism described in step 3 includes: (1) an information interaction process from a semantic perspective to a collaborative perspective. This involves mapping the semantic embedding sequence to a query vector. Q Mapping co-embedded sequences to key vectors K Sum value vector V (2) The information interaction process from the collaborative perspective to the semantic perspective. The collaborative embedding sequence is mapped to the query vector. Q' Mapping semantic embedding sequences to key vectors K' Sum value vector V' By scaling the dot product attention, the attention weight of the collaborative perspective on the semantic perspective is calculated, thereby realizing the reverse guidance of collaborative information on semantic features and thus obtaining a bidirectional fused feature representation.
5. The method according to claim 1, characterized in that: The GRU4Rec encoder with parameter sharing mechanism mentioned in step 3 refers to encoding the semantic viewpoint feature sequence and the collaborative viewpoint feature sequence using the same set of GRU network parameters, including updating gate parameters, resetting gate parameters, and hidden state parameters, thereby ensuring that different viewpoint features are modeled in a unified feature space. In a further preferred embodiment, the sequence learning loss function based on pairwise sorting mentioned in step 3 is used to constrain the model's prediction probability for the real subsequent decomposition task to be higher than the prediction probability for the negative sample task. Its calculation formula is as follows: L DVM =- log (P(t + a，k │S a，1：k-1 )-P(t - a，k │S a，1：k-1 Among them, S a，1:k-1 Retired power batteries a The task sequence consisting of the first k-1 historical dismantling tasks; t + a,k Indicates the actual next dismantling task; t - a,k This represents a task involving negative samples obtained through random sampling.
6. The method according to claim 1, characterized in that: The semantic similarity in step 4 is calculated using cosine similarity. The cosine similarity between the target decommissioned power battery and historical decommissioned power battery samples is calculated based on the joint sequence representation and then normalized to obtain the corresponding aggregation weights. Finally, the joint sequence representations corresponding to the retrieved similar battery samples are weighted and summed according to the weight values to generate the teacher representation.
7. The method according to claim 1, characterized in that: The self-distillation loss function described in step 4 is used to measure the difference between the student representation and the teacher representation. It employs the mean squared error loss function, and its calculation formula is as follows: L RASD =││Z target -u││ 2 2 Among them, Z target The student representation represents the target retired power battery, while the teacher representation is obtained by weighted aggregation of multiple similar historical retired power battery samples.
8. The method according to claim 1, characterized in that: During model training, a joint loss function is used to optimize the model, and its calculation formula is as follows: L=L DVM + L RASD in, The hyperparameter for adjusting the self-distillation loss weight is set to a value ranging from 0.1 to 1.0.