Power battery disassembly path optimization method based on autoregression generation strategy

Abstract

The invention relates to the technical field of battery disassembly and recovery, and discloses a power battery disassembly path optimization method based on an autoregression generation strategy. The method comprises the following steps: acquiring multi-source data of a to-be-disassembled battery, normalizing the multi-source data, inputting the normalized multi-source data into a multi-decoding-layer Transform model, performing dimension reduction processing on the battery data by the multi-decoding-layer Transform model, performing splicing fusion processing to obtain fusion characteristics, acquiring a task ID and a score of the to-be-disassembled battery according to the fusion characteristics, and storing the task ID and the score of the to-be-disassembled battery. The next step of disassembly of the battery is completed according to the score and the task ID; and based on an autoregressive generation strategy, repeatedly taking the task ID and the fusion feature obtained this time as the input of the multi-decoding-layer Transformer model, and obtaining the score and the task ID for the next disassembly of the battery until the disassembly of the battery is completed. The problem that an optimal path cannot be effectively planned through optimization of an existing battery disassembling path method is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of battery disassembly and recycling, and particularly to a method for optimizing the disassembly path of power batteries based on an autoregressive generation strategy. Background Art

[0002] With the rapid development of the electric vehicle market, the problem of retired power batteries is becoming increasingly serious. The disassembly and recycling of retired batteries are not only crucial for environmental protection, but also of great significance for recycling valuable materials, reducing resource waste and costs. Traditional battery disassembly processes usually rely on manual experience and fixed disassembly path arrangements, lacking flexibility and intelligence, resulting in low disassembly efficiency, low equipment utilization rate, and difficulty in meeting the optimization requirements under different battery types, disassembly tasks and environmental conditions. Therefore, how to design an efficient, flexible and schedulable disassembly path optimization method has become an urgent problem to be solved in the field of battery recycling.

[0003] Traditional methods for optimizing battery disassembly paths mainly rely on heuristic algorithms, such as genetic algorithms, particle swarm optimization algorithms, etc. However, methods such as genetic algorithms and particle swarm optimization algorithms still have several problems, being prone to falling into local optimal solutions. Especially when the complexity of disassembly tasks is relatively high, it may lead to low search efficiency. When dealing with multi-source data in disassembly tasks, the optimization results are unstable, unable to comprehensively consider the complex dependencies between tasks, and ignoring how to effectively integrate and optimize static disassembly data and dynamic task scheduling data. Summary of the Invention

[0004] The present invention provides a method for optimizing the disassembly path of power batteries based on an autoregressive generation strategy to solve the problem that the existing battery disassembly path optimization method cannot effectively plan the optimal path.

[0005] To achieve the above object, the present invention is realized through the following technical solutions: In a first aspect, the present invention provides a method for optimizing the disassembly path of power batteries based on an autoregressive generation strategy, including the following steps: Step 1: Obtain multi-source data of the battery to be disassembled, and perform normalization processing on the multi-source data to obtain battery data; Step 2: Input the battery data into a multi-decoding layer Transformer model, perform dimensionality reduction processing on the battery data through the multi-decoding layer Transformer model to obtain multiple pieces of dimensionality-reduced data, and perform splicing and fusion processing on the multiple pieces of dimensionality-reduced data to obtain a fusion feature. The multi-decoding layer Transformer model obtains the task ID and score of the battery to be disassembled according to the fusion feature, and completes the next disassembly of the battery according to the score and task ID; Step 3: Based on the autoregressive generation strategy, repeatedly use the task ID and fusion features obtained this time as the input of the multi-decoding layer Transformer model to obtain the score and task ID for the next disassembly step of the battery until the disassembly of the battery is completed.

[0006] Regarding the task ID and score output by the multi-decoding layer Transformer model, it is determined whether to perform the next disassembly step according to the task ID based on the comparison between the score and a preset threshold. If it exceeds the preset threshold, then compare the task ID with the target scheduling data file collected additionally to determine that the task ID belongs to the battery disassembly action that should be executed in the target scheduling file, and perform the disassembly of the battery based on the determined battery disassembly action. If it does not exceed the preset threshold, then obtain the task ID and score again.

[0007] Further, in Step 1, the obtaining of the multi-source data of the battery to be disassembled includes: obtaining the multi-source data of the battery to be disassembled from a specified folder; The multi-source data is composed of the data content in the specified folder. The data content includes the fromto data in the fromto folder and the machining data in the machining folder. The fromto data includes the positions of the starting module and the ending module, and the machining data includes the module ID, device ID, disassembly time, and disposal cost.

[0008] Further, the normalization processing includes alignment and MinMax normalization processing. The alignment is to align the fromto data and the machining data based on the number of data rows; The MinMax normalization processing is represented by the following formula: ; where, is the multi-source data after normalization, is any input variable to be MinMax-normalized in the multi-source data, is the minimum value in the column data where the input variable is located, is the maximum value in the column data where the input variable is located.

[0009] Through strict alignment and normalization processing, it is ensured that data of different scales can be uniformly converted to the same magnitude, avoiding the problem of gradient mismatch in model training. This process not only improves the training efficiency of the model but also enhances the stability of the final path optimization result.

[0010] Further, in Step 2, the dimensionality reduction processing includes Emberdding processing for the fromto data and fully connected linear mapping for the machining data; The dimensionality reduction processing of the fromto data includes: mapping the fromto data to a continuous vector space to obtain embedding vectors, and mapping the embedding vectors to a feature space with half the dimension of the model's hidden vectors through a fully connected layer; The dimensionality reduction processing of the machining data includes: linearly mapping the machining data through a fully connected layer.

[0011] Furthermore, the process of splicing and fusing multiple pieces of dimensionality-reduced data to obtain fused features includes: splicing the dimensionality-reduced fromto data and machining data in terms of dimensional features, and then further mapping through a fully connected layer to obtain fused features.

[0012] Through the feature processing strategy of separate mapping and subsequent fusion, the present invention greatly improves the effect of information fusion in the feature expression of disassembly paths and operation data. The combination of multi-modal data enables the system to comprehensively understand various complex factors in the disassembly process, thereby improving the accuracy and efficiency of disassembly path optimization.

[0013] Furthermore, the multi-decoder layer Transformer model includes: a Transformer model composed of several decoder layers; Each decoder layer sequentially includes a self-attention module, a cross-attention module, and a feed-forward neural network, and a Dropout operation is added after each sub-module of the self-attention module, the cross-attention module, and the feed-forward neural network.

[0014] Through the above operations, dynamic and static input information from different data sources is effectively processed. This structure enhances the expressive power of the model, especially in complex disassembly paths and equipment scheduling, and can capture the complex relationships between tasks.

[0015] Furthermore, the multi-decoder layer Transformer model constructs a loss function based on cross-entropy loss, uniqueness penalty, and ranking consistency penalty; The cross-entropy loss includes: the gap between the task ID output by the multi-decoder layer Transformer model and the preset true target sequence; The uniqueness penalty includes: penalizing the repeated task IDs in the generated sequence; The ranking consistency penalty includes: penalizing the task IDs with inconsistent scheduling orders.

[0016] Furthermore, in step 3, the multi-decoder layer Transformer model combines sine and cosine functions to perform positional encoding on the fused features to obtain the positional information of each data point in the multi-source data and gradually generate task IDs; The position encoding of the fused features by combining sine and cosine functions includes: generating encoding formulas using sine and cosine functions respectively, performing position encoding on the even indices of each position and embedding dimension in the fused features using the sine encoding formula, and performing position encoding on the odd indices of each position and embedding dimension using the sine encoding formula; The sine encoding formula is represented by the following formula: ; where, represents the fused features encoded by the sine encoding formula, represents the position, represents the even index of the embedding dimension, represents the dimension of the input vector; The cosine encoding formula is represented by the following formula: ; where, represents the fused features encoded by the cosine encoding formula, represents the odd index of the embedding dimension.

[0017] Furthermore, the autoregressive generation strategy includes: initial input setting, step-by-step generation mechanism, repeated task masking, and prediction update to generate a sequence; The initial input setting includes: setting dynamic inputs for the multi-decoding layer Transformer model and setting the initial input of the dynamic inputs to a zero vector; The step-by-step generation mechanism includes: using the task ID output by the multi-decoding layer Transformer model at the current step as the input; The repeated task masking includes: adjusting the corresponding values of the logits vector output by the multi-decoding layer Transformer model at the current step in the corresponding set of generated task IDs to negative infinity for masking; The prediction update to generate a sequence includes: selecting the task ID corresponding to the logits vector with the highest probability value in the logits vector after the repeated task masking process as the output at the current step.

[0018] Among them, the logits vector is the original output of the last linear layer of the multi-decoding layer Transformer model, which is an unnormalized real number vector. After the logits vector is normalized, it is converted into a probability distribution, and the final output task ID is obtained based on the probability distribution.

[0019] Beneficial effects: A method for optimizing the disassembly path of power batteries based on an autoregressive generation strategy provided by the present invention combines text information, machine data, and location information, and uses a multi-decoding layer Transformer model combined with an autoregressive generation mechanism to optimize the disassembly path of power batteries, and can accurately predict the order of disassembly tasks. This accurate prediction of task order effectively improves the efficiency and rationality of the disassembly process. By gradually generating task IDs, the adaptability and reliability of the system are enhanced.

[0020] By combining static multi-source data and dynamic task IDs, the optimization of task sequences is carried out. This method not only improves the utilization efficiency of data but also can dynamically adapt to changes in disassembly tasks, enhancing the real-time performance and flexibility of the system.

[0021] The absolute position encoding, data preprocessing, and feature fusion technologies designed in the present invention are particularly applicable to the field of battery disassembly and recycling, and can effectively cope with complex industrial environments and production requirements. In addition, by using fixed sine and cosine position encodings, the problem that parameters may fall into local optima during training is avoided, and the stability and efficiency of the model in actual industrial applications are improved. Brief Description of the Drawings

[0022] Figure 1 It is a flowchart of a method for optimizing the disassembly path of power batteries based on an autoregressive generation strategy according to an embodiment of the present invention. Detailed Embodiments

[0023] The technical solutions of the present invention will be described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0024] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the art to which the present invention belongs. The "first", "second", and similar terms used in the present invention do not denote any order, quantity, or importance, but are only used to distinguish different components. Similarly, terms such as "a" or "one" do not denote a quantity limitation, but indicate the existence of at least one. "Connection" or "connected" and similar terms are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right", etc. are only used to represent relative position relationships. When the absolute position of the object being described changes, the relative position relationship also changes accordingly.

[0025] Please refer to Figure 1, an embodiment of the present application provides a method for optimizing the disassembly path of a power battery based on an autoregressive generation strategy, including the following steps: Step 1: Obtain multi-source data of the battery to be disassembled, and perform normalization processing on the multi-source data to obtain battery data; Obtaining the multi-source data of the battery to be disassembled includes: obtaining the multi-source data of the battery to be disassembled from a specified folder; The multi-source data is composed of the data content in the specified folder. The data content includes the fromto data in the fromto folder and the machining data in the machining folder. The fromto data includes the positions of the starting module and the ending module, and the machining data includes the module ID, device ID, disassembly time, and disposal cost.

[0026] Regarding the fromto data, each row includes two numbers. The first number represents the starting module, and the second number represents the ending module. Regarding the machining data, each row includes four numbers. The first number represents the module ID, the second number represents the device ID, the third number represents the disassembly time, and the fourth number represents the disposal cost.

[0027] The normalization processing includes alignment and MinMax normalization processing. The alignment is based on the number of data rows, and the fromto data and machining data are aligned; The MinMax normalization processing is represented by the following formula: ; Among them, is the normalized multi-source data, is any input variable to be MinMax-normalized in the multi-source data, is the minimum value in the column data where the input variable is located, is the maximum value in the column data where the input variable is located. For example, is 1, is 5. For a data value of 3, its normalized value is 0.5.

[0028] Step 2: Input the battery data into the multi-decoding layer Transformer model, perform dimensionality reduction processing on the battery data through the multi-decoding layer Transformer model, obtain multiple pieces of dimensionality-reduced data, and perform splicing and fusion processing on the multiple pieces of dimensionality-reduced data to obtain a fusion feature. The multi-decoding layer Transformer model obtains the task ID and score of the battery to be disassembled according to the fusion feature, and completes the next disassembly of the battery according to the score and task ID; In Figure 1Among them, the machine data is machining data, the from-to text is fromto data, and the position is position data, which is the index position of the numbers in the machining data and the fromto data. The dynamic input represents the loop operation of the autoregressive generation strategy; Operation represents the operation, that is, whether to execute after comparing the score with the preset threshold, and the battery disassembly action determined by comparing the task ID with the target scheduling data; The dimensionality reduction processing includes the Emberdding processing for the fromto data and the fully connected linear mapping for the machining data; The dimensionality reduction processing of the fromto data includes: mapping the fromto data to a continuous vector space to obtain the embedding vector, and mapping the embedding vector to the feature space with half of the dimensionality of the model's hidden vector through a fully connected layer. The mathematical expression of the mapping process is: ; Among them, is the weight matrix, is the bias vector, represents the input fromto data, represents the mapped embedding vector; The dimensionality reduction processing of the machining data includes: linearly mapping the machining data through a fully connected layer. The mathematical expression of the linear mapping is: ; Among them, and are the corresponding mapping parameters respectively, represents the input machining data, represents the corresponding low-dimensional data after the linear mapping of the machining data.

[0029] Performing splicing and fusion processing on multiple pieces of dimensionality-reduced data to obtain the fusion feature includes: splicing the dimensionality-reduced fromto data and the machining data in terms of dimensional features, and then further mapping through a fully connected layer to obtain the fusion feature.

[0030] In this embodiment, the input of the multi-decoder layer Transformer model also includes the position data, which is the index position of the numbers in the machining data and the fromto data. This is well-known common knowledge to those skilled in the art and will not be elaborated here.

[0031] The multi-decoder layer Transformer model includes: a Transformer model composed of several decoder layers; The decoding layer sequentially includes a self-attention module, a cross-attention module, and a feed-forward neural network.

[0032] The self-attention module performs a linear transformation on the input vector to obtain queries, keys, and values, and in multiple parallel heads, calculates attention scores respectively. After the outputs of all heads are concatenated, through a linear mapping, and then a residual connection is made with the input for normalized fusion; The cross-attention module uses the output features of the self-attention module as queries, maps the pre-generated static content into keys and values, performs cross-modal attention calculation, and after calculation, makes a residual connection with the output features of the self-attention module, and finally performs normalized fusion to output the output features of the cross-attention module; The feed-forward neural network consists of two layers of linear transformation and a non-linear activation function in the middle, and also performs a normalized fusion step after calculation; To prevent overfitting, Dropout is used after the above self-attention module, cross-attention module, and feed-forward neural network.

[0033] The multi-decoding layer Transformer model constructs a loss function based on cross-entropy loss, uniqueness penalty, and ranking consistency penalty; The cross-entropy loss includes: the gap between the task IDs output by the multi-decoding layer Transformer model and the preset true target sequence, which is represented by the following formula: ; Among them, represents the value of the cross-entropy loss, which is used to quantify the error between the model prediction distribution and the true target distribution. The symbol represents the batch size, that is, the number of samples input into the model at one time, and represents the length of the task ID sequence in each sample. The indices and correspond to the index of the sample within the batch and the index of the position in the sequence respectively. The variable represents the true task label of the rd sample at the position , and is the probability that the model predicts that the task at this position belongs to a specific category. The natural logarithm function is used to transform the probability distribution, and the overall external normalization factor ensures that the loss value will not produce unreasonable deviations due to changes in the batch size and sequence length, thereby promoting equal-weight processing of each task ID position during model training.

[0034] The uniqueness penalty includes: punishing repeated task IDs in the generated sequence, which is represented by the following formula: ; Among them, represents the value of the uniqueness penalty, which reflects the situation of the model generating duplicate tasks. The variable still represents the batch size, while is the time index in the task sequence, ranging from 1 to . Among them, represents the th sample at the moment generates the task ID, which is compared with the set of tasks generated in the previous stage. The indicator function returns 1 when the condition is met (that is, the current task ID already exists in the previously generated set), otherwise it returns 0. After summing over all samples and sequence positions, it is then normalized by to measure the overall proportion of duplicate task IDs. Through this penalty term, the possibility of the model generating duplicate tasks can be effectively reduced, thus ensuring the uniqueness of the task scheduling sequence; The sorting consistency penalty includes: penalizing the task IDs with inconsistent scheduling orders, which is expressed by the following formula: ; Among them, quantifies the deviation between the model-predicted task sequence and the true task sequence at each time position. The variable still represents the batch size, is the position index of the task sequence. The symbol represents the th sample's task ID predicted by the model at the position , while represents the actual true task ID at this position. By calculating the absolute error between the predicted value and the true value, this loss term can truly reflect the sorting difference between the generated sequence and the target sequence. After outer summation and normalization, this loss term constrains the model's sorting, thus promoting the generated task sequence to be as consistent as possible with the true order; The above losses are weighted and combined to form a loss function: ; Among them, and are the weight coefficients of the corresponding penalty terms.

[0035] Step 3: Based on the autoregressive generation strategy, repeatedly use the task ID and the fusion feature obtained this time as the input of the Transformer model to obtain the score and task ID for the next disassembly of the battery until the disassembly of the battery is completed.

[0036] The task ID-based autoregressive generation strategy includes: initial input setting, step-by-step generation mechanism, repeated task masking, and prediction update to generate a sequence; The initial input setting includes: setting dynamic inputs for the multi-decoder layer Transformer model and setting the initial input of the dynamic input to a zero vector; The step-by-step generation mechanism includes: using the task ID output by the multi-decoder layer Transformer model at the current step as the input; The repeated task masking includes: adjusting the corresponding values of the logits vector output by the multi-decoder layer Transformer model at the current step in the corresponding set of generated task IDs to negative infinity for masking; The logits vector is the raw output of the last linear layer of the multi-decoder layer Transformer model, which is an unnormalized real number vector. After the logits vector is normalized, it is converted into a probability distribution, and the final output task ID is obtained based on the probability distribution; The prediction update to generate a sequence includes: in the logits vector after the repeated task masking process, selecting the task ID corresponding to the logits vector with the highest probability value as the output of the current step.

[0037] The multi-decoder layer Transformer model combines sine and cosine functions to perform position encoding on the fused features to obtain the position information of each data point in the multi-source data; Combining sine and cosine functions to perform position encoding on the fused features includes: using sine and cosine functions to generate encoding formulas respectively, performing position encoding on the even indices of each position and embedding dimension in the fused features using the sine encoding formula, and performing position encoding on the odd indices of each position and embedding dimension using the sine encoding formula; The sine encoding formula is represented by the following formula: ; where represents the fused features encoded by the sine encoding formula, represents the position, represents the even index of the embedding dimension, represents the dimension of the input vector; The cosine encoding formula is represented by the following formula: ; where represents the fused features encoded by the cosine encoding formula, represents the odd index of the embedding dimension.

[0038] Based on the above steps, the traditional genetic algorithm, particle swarm optimization algorithm, Transformer model are compared with the present application, and the battery disassembly path optimization task is trained and tested. The goal is to minimize the total time and cost during the disassembly process while following the task sequence constraints. Please refer to Table 1.

[0039] Table 1: Comparison Results of the Present Application and Various Models

[0040] Specifically, this method significantly reduces the total time required for disassembly, reducing by approximately 18%, 15%, and 11% compared to the genetic algorithm, particle swarm optimization, and basic Transformer respectively. This result indicates that by optimizing the task scheduling order through the autoregressive generation model, this method effectively reduces the waiting time and time waste between tasks. In terms of disassembly cost, this method reduces by approximately 20%, 17%, and 15% compared to the genetic algorithm, particle swarm optimization, and basic Transformer respectively. This result shows that this method can effectively reduce equipment usage and resource consumption during the disassembly process, thus improving the economy of disassembly. Most importantly, the total objective function value of this method is 378, reducing by approximately 20%, 17%, and 14% compared to the genetic algorithm, particle swarm optimization, and basic Transformer respectively, demonstrating its superiority in comprehensively considering disassembly time and cost.

[0041] The advantages of this method are mainly reflected in several aspects. First, the autoregressive generation strategy is one of the core features of this method. By gradually generating task IDs, in each step of the generation process, the generated tasks are dynamically masked, avoiding the repeated generation of tasks and ensuring the uniqueness of the task sequence. This strategy significantly improves the quality of task scheduling and avoids the problem of repeated tasks that may occur in traditional methods. Second, this method fully integrates static disassembly data and dynamic task scheduling data through deep feature fusion, and uses the Transformer autoregressive generation model to predict the task order. Different from traditional methods that only optimize the path through heuristic search, the deep learning model can automatically learn the complex relationships between tasks and optimize the task order. Third, cross-modal feature fusion is a highlight of this method. By effectively combining information from static and dynamic data sources, this method not only retains the key information of each data source but also enhances the model's understanding of the complexity of the disassembly path, thus improving the effect of the scheduling strategy. Finally, this method ensures the uniqueness, rationality, and accuracy of the task sequence by introducing fine optimization of the objective function, combining cross-entropy loss, uniqueness penalty, and sorting consistency penalty, thereby further improving the quality of task scheduling.

[0042] In summary, the optimized system for battery disassembly path based on the autoregressive generation model proposed in the present invention demonstrates significant advantages in disassembly path optimization compared with traditional genetic algorithms, particle swarm optimization, and basic Transformer methods. The autoregressive generation strategy, deep feature fusion, cross-modal feature processing, and refined objective function design enable this method to effectively generate task ID sequences with high efficiency, uniqueness, and rationality, thus promoting the overall optimization of the battery disassembly path. Therefore, this method provides a new and efficient optimization means for the field of battery disassembly and recycling and has high application prospects.

[0043] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative efforts. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art shall fall within the protection scope determined by the claims.

Claims

1. A power battery disassembly path optimization method based on autoregressive generation strategy, characterized in that: The steps include: Step 1: Obtain multi-source data of the battery to be disassembled, and normalize the multi-source data to obtain battery data; Step 2: Input the battery data into the multi-decoding layer Transformer model, perform dimensionality reduction processing on the battery data through the multi-decoding layer Transformer model, obtain multiple copies of dimensionality reduction data, and concatenate and fuse the multiple copies of dimensionality reduction data to obtain fusion features. The multi-decoding layer Transformer model obtains the task ID and score of the battery to be disassembled based on the fusion features, and completes the next step of battery disassembly based on the score and task ID; Step 3: Based on the autoregressive generation strategy, the task ID and fusion features obtained this time are repeatedly used as the input of the multi-decoding layer Transformer model to obtain the score and task ID for the next step of battery disassembly until the battery is disassembled.

2. The power battery disassembly path optimization method based on the autoregressive generation strategy according to claim 1 is characterized in that: In step 1, the obtaining of multi-source data of the battery to be disassembled includes: obtaining multi-source data of the battery to be disassembled from a designated folder; The multi-source data is composed of data contents in a designated folder, including fromto data in a fromto folder and machining data in a machining folder, the fromto data including the positions of a starting module and an ending module, and the machining data including a module ID, a device ID, a disassembly time and a disposal cost.

3. The power battery disassembly path optimization method based on autoregressive generation strategy according to claim 2 is characterized in that: The normalization process includes alignment and MinMax normalization process, wherein the alignment is based on the number of data rows, and the fromto data and machining data are aligned; The MinMax normalization process is expressed by the following formula: ； in, is the normalized multi-source data, is any input variable in the multi-source data to be processed by MinMax normalization. is the minimum value in the column data where the input variable is located. It is the maximum value in the column data where the input variable is located.

4. The power battery disassembly path optimization method based on autoregressive generation strategy according to claim 2 is characterized in that: In step 2, the dimensionality reduction process includes Emberdding process for fromto data and fully connected linear mapping for machining data; The dimensionality reduction processing of the fromto data includes: mapping the fromto data to a continuous vector space to obtain an embedded vector, and mapping the embedded vector to a feature space with half the dimension of the model's latent vector through a fully connected layer; The dimension reduction process of machining data includes linear mapping of machining data through a fully connected layer.

5. The power battery disassembly path optimization method based on autoregressive generation strategy according to claim 4 is characterized in that: The method of performing splicing and fusing multiple copies of dimensionality reduction data to obtain fusion features includes: splicing the fromto data after dimensionality reduction processing with the machining data on dimensional features, and then further mapping through a fully connected layer to obtain fusion features.

6. The power battery disassembly path optimization method based on autoregressive generation strategy according to claim 1 is characterized in that: The multi-decoding layer Transformer model includes: a Transformer model composed of a plurality of decoding layers; The decoding layer includes a self-attention module, a cross-attention module and a feedforward neural network in sequence, and a Dropout operation is added after each submodule of the self-attention module, the cross-attention module and the feedforward neural network.

7. The power battery disassembly path optimization method based on autoregressive generation strategy according to claim 6 is characterized in that: The multi-decoding layer Transformer model constructs a loss function based on cross entropy loss, uniqueness penalty, and sort consistency penalty; The cross entropy loss includes: the gap between the task ID output by the multi-decoding layer Transformer model and the preset real target sequence; The uniqueness penalty includes: penalizing duplicate task IDs in the generated sequence; The sort consistency penalty includes: punishing task IDs with inconsistent scheduling order.

8. The power battery disassembly path optimization method based on autoregressive generation strategy according to claim 6 is characterized in that: In step 3, the multi-decoding layer Transformer model combines sine and cosine functions to positionally encode the fused features to obtain the position information of each data point in the multi-source data and gradually generate the task ID; The method of combining sine and cosine functions to perform position encoding on the fused features includes: using sine and cosine functions to generate encoding formulas respectively, using the sine encoding formula to perform position encoding on the even index of each position and embedding dimension in the fused features, and using the sine encoding formula to perform position encoding on the odd index of each position and embedding dimension.

9. The power battery disassembly path optimization method based on the autoregressive generation strategy according to any one of claims 1 to 8, characterized in that: The autoregressive generation strategy includes: initial input setting, step-by-step generation mechanism, repeated task shielding, and prediction update generation sequence; The initial input setting includes: setting a dynamic input for a multi-decoding layer Transformer model, and setting an initial input of the dynamic input to a zero vector; The step-by-step generation mechanism includes: taking the task ID output by the current step of the multi-decoding layer Transformer model as input; The repeated task shielding includes: adjusting the corresponding value of the logits vector output by the current step of the multi-decoding layer Transformer model in the corresponding generated task ID set to negative infinity for shielding; The prediction update generation sequence includes: selecting the task ID corresponding to the logits vector with the highest probability value from the logits vector after repeated task masking as the output of the current step.

Images