A Scheduling Method for MES System in the Lithium Battery Industry Based on Deep Learning

Abstract

The present invention discloses a scheduling method for the MES system in the lithium battery industry based on deep learning, which includes the following steps: S1, collecting and preprocessing lithium battery production data; S2, constructing a twin-tower Siamese network to encode the standardized sequence and generate a similarity matrix; S3, using Transformer-XL to model long-term dependencies and extract cross-cycle scheduling relationships; S4, constructing the scheduling relationships into an associated weight tensor, inputting it into a neural Turing machine, and generating an original scheduling instruction stream through read and write operations; S5, performing conflict detection and dependency rearrangement based on the controller memory comparison mechanism, and iteratively updating the state of the neural Turing machine; S6, feeding back candidate solutions to the MES execution layer and updating the model parameters. The present invention integrates multiple model structures, realizes intelligent modeling and dynamic optimization of lithium battery scheduling, and effectively improves the scheduling efficiency, response speed and system adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery production, and particularly to a production scheduling method for a lithium battery industry MES system based on deep learning. Background Technique

[0002] The Manufacturing Execution System (MES) has been widely applied in major industrial manufacturing industries. MES is mainly used to connect the planning layer and the control layer of an enterprise, and undertakes tasks such as the collection, monitoring, and command and dispatch of production process data. It is a key system for realizing workshop-level manufacturing management. In the lithium battery industry, due to its characteristics such as complex processes, numerous procedures, intensive equipment, and short product cycles, the importance of the MES system is particularly prominent. An efficient MES system can not only improve production capacity utilization rate, but also effectively reduce production bottlenecks and improve resource allocation efficiency. However, the production scheduling methods of traditional MES systems still face many challenges in practical applications.

[0003] Currently, the mainstream production scheduling methods still mainly rely on rule-based heuristic algorithms and traditional optimization algorithms. These methods usually rely on a large number of pre-set artificial rules and assume that the production environment is static and stable. Facing the frequently fluctuating customer demands, dynamically changing equipment states, and complex multi-process flows in the lithium battery industry, this static production scheduling model seems inadequate. For example, when the material supply is interrupted, the equipment breaks down suddenly, or the customer orders increase suddenly, the traditional methods often cannot respond quickly, resulting in a decrease in resource utilization rate and even the overall failure of the production scheduling plan. In addition, it is difficult for these methods to effectively model the non-linear patterns, periodic characteristics, and cross-time-dependent relationships in production data, thereby limiting the accuracy and flexibility of the production scheduling strategy.

[0004] In recent years, with the development of deep learning technology, more and more research has begun to explore its application in intelligent manufacturing, especially showing strong modeling capabilities in the field of production scheduling optimization. Models such as Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Long Short-Term Memory Network (LSTM) have been introduced into production prediction and scheduling tasks and achieved certain results. However, these models still have problems such as insufficient long-term dependence capture ability, limited input data dimension, and imperfect model memory mechanism. For example, standard RNN and LSTM are prone to gradient disappearance or gradient explosion when facing long sequence data, and it is difficult to capture the production scheduling dependence information with a long time span; while the traditional attention mechanism, although alleviating this problem to a certain extent, cannot completely solve the problems such as insufficient time information expression and long-term state forgetting.

[0005] In addition, most current deep learning models are still mainly applied in the production scheduling task in a single structure, failing to fully utilize the collaborative advantages of multiple models. For example, the Siamese network has a natural advantage in extracting the similarity between time series, but it is usually independently applied to anomaly detection or sample matching tasks and has not been deeply integrated with the production scheduling decision-making process; although the Transformer-XL model performs excellently in long sequence modeling, it is mainly applied in the field of natural language processing and has not formed a structured adaptation mechanism for manufacturing scheduling data; the Neural Turing Machine (NTM) has the ability to simulate the read-write storage of a Turing machine and is very suitable for complex logic and state tracking tasks, but there is still a lack of systematic research on issues such as its coupling method with manufacturing data and input structure design.

[0006] Another core difficulty faced by current deep learning-based production scheduling methods lies in the lack of an effective data fusion mechanism. In the production process of lithium batteries, information sources such as equipment status, process switching, inventory data, and historical orders are diverse and have complex time series. How to achieve effective fusion, associated modeling, and dynamic feedback of data from different sources is the key to improving the accuracy of the production scheduling algorithm. The traditional splicing-based data integration method is difficult to capture the interaction relationships between multi-dimensional heterogeneous information and cannot adaptively adjust the information weights in a dynamic environment, resulting in the production scheduling plan lacking context relevance and timeliness.

[0007] In addition, there is also a lag in the execution feedback mechanism of existing production scheduling methods. Most models adopt the method of offline training and static deployment and lack the ability to interact with the production execution layer in real time. Once an anomaly occurs at the production site, the model parameters cannot be updated in a timely manner, and the production scheduling strategy cannot be adjusted quickly, thus affecting the stable operation of the entire production chain. Especially in an environment such as the lithium battery industry where high requirements are placed on response speed and flexible production scheduling, the low timeliness and weak adaptability of existing methods severely restrict the improvement of the intelligent level.

[0008] Therefore, how to provide a production scheduling method for the MES system in the lithium battery industry based on deep learning is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0009] An object of the present invention is to propose a production scheduling method for the MES system in the lithium battery industry based on deep learning. The present invention fully integrates the similarity feature extraction ability of the Siamese network, the long-term dependence modeling ability of the Transformer-XL model, and the read-write state reasoning mechanism of the Neural Turing Machine, and details a production scheduling instruction generation method for multi-processes, dynamic resources, and cross-cycle task dependencies in the lithium battery manufacturing scenario, having the advantages of high modeling accuracy, fast response speed, strong adaptability, and high production scheduling flexibility.

[0010] A scheduling method for a MES system in the lithium battery industry based on deep learning according to an embodiment of the present invention includes the following steps:

[0011] S1. Collect lithium battery production data and perform preprocessing to obtain a standardized sequence;

[0012] S2. Construct a Siamese network based on a two-tower structure, perform dual-branch encoding on the standardized sequence, extract similarity features of the standardized sequence through a shared-weight subnetwork, and generate a similarity matrix;

[0013] S3. Concatenate and fuse the similarity matrix with the standardized sequence to form a feature fusion sequence, and use the Transformer-XL model to perform long-term dependence modeling on the feature fusion sequence, and extract cross-cycle scheduling dependence relationships through relative position embedding and an expandable memory cache mechanism;

[0014] S4. Convert the scheduling dependence relationship into a scheduling tensor containing associated weights, and input the scheduling tensor into a neural Turing machine. The neural Turing machine includes a readable and writable external storage matrix and a controller, and generates an original scheduling instruction stream through read and write operations on the scheduling tensor;

[0015] S5. Based on the memory comparison mechanism inside the controller, perform conflict detection and dependence relationship rearrangement on the original scheduling instruction stream, and generate a candidate scheduling plan set by iteratively updating the state of the neural Turing machine;

[0016] S6. Feed the candidate scheduling plan set back to the execution layer of the MES system for production plan execution, periodically collect feedback data, and iteratively update the parameters of the Siamese network and the Transformer-XL model.

[0017] Optionally, the production data includes equipment status, production requirements, material inventory, production progress, production processes, and historical production data.

[0018] Optionally, the preprocessing includes unified timestamp alignment, discrete value normalization, missing item interpolation and noise value elimination.

[0019] Optionally, the S2 specifically includes:

[0020] S21. Denote the standardized sequence as where represents the feature vector at time step t, t is the current time step index, satisfying , T is the total number of time steps, perform a two-way input on the sequence X, and input it into the left tower encoder and the right tower encoder respectively;

[0021] S22. Both the left tower encoder and the right tower encoder consist of a convolutional sub-network and a fully connected layer with shared weights, which extract features from the input sequence. Let the output of the left tower encoder be , and the output of the right tower encoder be , where , are the feature vectors output by the left tower encoder and the right tower encoder after encoding, t is the current time step index, and it satisfies ;

[0022] S23. Calculate the similarity between the left tower output feature and the right tower output feature to generate a similarity matrix S:

[0023] ;

[0024] where, represents the similarity degree between time step i and time step j in the similarity matrix, represents the natural exponential function with base e, represents the inner product of the left tower and right tower feature vectors, represents the square of the Euclidean distance, W represents the learnable bilinear mapping weight matrix, represents the bilinear similarity mapping under the learnable weight matrix, , are preset non - negative hyperparameters, and T is the total number of time steps.

[0025] Optionally, the Siamese network adopts a two - tower structure design, including two feature encoding sub - networks with shared parameters. Each sub - network consists of two one - dimensional convolutional layers, a batch normalization layer, and a ReLU activation function, which are used to perform parallel feature extraction on the input standardized production data sequence, and calculate the Euclidean distance, inner product, and weighted bilinear similarity between feature vectors at any time step in the output stage, and finally generate a two - dimensional similarity matrix for characterizing the similarity between different time steps.

[0026] Optionally, S3 specifically includes:

[0027] S31. Concatenate and fuse the standardized sequence X and the similarity matrix S to construct a feature fusion sequence , where T is the total number of time steps, and the feature fusion vector is represented as , represents the feature vector at time step t, represents the similarity vector of the t - th row, represents the vector concatenation operation;

[0028] S32. Use the Transformer-XL model to perform long-term dependence modeling on the feature fusion sequence. Input the fusion vector sequence F into the Transformer-XL model, and adopt the relative position embedding and scalable memory cache mechanism to model the temporal correlation, and construct the cross-cycle state matrix Z:

[0029] ;

[0030] Among them, represents the state at time step t, T is the total number of time steps, represents the query vector, represents the key vector, represents the value vector, represents the relative position vector, , , is a learnable projection matrix, represents the fused historical state vector, b is the bias term, represents the non-linear transformation function, is the normalized attention weight of the t-th query and the i-th key, is a very small constant to avoid division by zero error, represents the Hadamard element-wise product, represents the L2 norm, m represents the number of vectors stored in the historical memory cache, represents the sine function, represents the cosine function;

[0031] S33. Non-linearly combine each with the corresponding dimension state in the historical state cache matrix to generate the scheduling dependence sequence D:

[0032] ;

[0033] Among them, represents the scheduling dependence vector at time step t, represents the activation function with residual connection, represents the square of the Euclidean distance between the current state and the historical state, represents the natural exponential function with base e, represents the Sigmoid function, , are weight matrices, is the bias term, is the scaling coefficient, represents the element-wise maximum operation.

[0034] Optionally, the feature fusion sequence F is formed by concatenating the row vectors corresponding to each time step in the similarity matrix with the feature vectors at the corresponding time steps in the normalization sequence in a dimension-wise manner, forming a fusion vector sequence with a dimension equal to the sum of the original feature dimension and the similarity vector dimension, maintaining strict alignment in the time domain, and enabling each fusion vector to contain both the original production state at the current moment and the global similarity structure information, so as to enhance the ability of the Transformer-XL model to jointly model local and global features when modeling cross-cycle scheduling dependencies.

[0035] Optionally, the specific steps of S4 include:

[0036] S41. Convert the scheduling dependency sequence D into a scheduling tensor P containing association weights, and the construction method is as follows:

[0037] Construct a corresponding dependency combination vector for any two time steps i and j and stack all the dependency combination vectors after embedding mapping to form a third-order scheduling tensor where represents the ternary relationship combination function, represents the scheduling dependency vector at time step i, represents the scheduling dependency vector at time step j, , representing the time step distance, and T is the total number of time steps;

[0038] S42. Input the scheduling tensor P into a neural Turing machine, which consists of an external storage matrix M and a set of controller state vectors H. Let the controller state at the current time step be , and generate a scheduling instruction stream Y through read and write operations:

[0039] ;

[0040] where represents the scheduling instruction at time step t, represents the controller state vector at time step t, , , represent trainable weight matrices, is the bias vector, represents the attention weight of the combination index (i, j) at the t-th moment in the scheduling tensor, represents the dependency combination vector in the scheduling tensor , is the read weight of the controller for the storage unit index k, N represents the number of storage units, represents the content of the k-th storage unit, represents the normalized activation mapping function;

[0041] S43. According to the scheduling instruction at each moment perform an internal update on the controller state vector and execute the update of the external storage matrix M at time step t + 1, and the update process is determined by the erasure vector , the addition vector and the write address weight jointly:

[0042] ;

[0043] Among them, represents the content of the k-th storage unit at time step t + 1, represents the content of the k-th storage unit at time step t, represents the Hadamard element-wise product, represents the tensor product, , represents the linear transformation matrix of the erasure vector and the addition vector, represents the interaction mapping matrix of the controller state, represents the bilateral attention coefficient function, is the Sigmoid function, is a non-zero constant for normalization.

[0044] Optionally, the definition of the ternary relation combination function is:

[0045] ;

[0046] Among them, , respectively represent the scheduling dependency vectors at time steps i and j, represents the time step distance, is a vector with all elements being 1, is the element-wise square difference, represents the Hadamard product, represents the vector concatenation operation, is the linear transformation matrix, is the bias vector, represents the mapping function with non-linear activation.

[0047] Optionally, S5 specifically includes:

[0048] S51. Receive the scheduling instruction stream sequence , and for each scheduling instruction at each moment, combine it with the controller state vector Compare them and construct a set of memory difference vectors by combining the state differences in the time series. , where T is the total number of time steps, and each difference vector is obtained by the linear combination of the current instruction vector and the state difference, and is used to characterize the content offset degree between scheduling instructions;

[0049] S52. Perform a conflict detection operation on the scheduling instruction stream Y according to the set of memory difference vectors to construct a scheduling conflict graph G=(V,E), where the vertex set V corresponds to the scheduling instructions at each moment, and the edges in the edge set E indicate that there are resource conflicts, timing conflicts or material overlaps between any two instructions. By traversing each edge, extract the conflict path set C for dependency rearrangement processing;

[0050] S53. Based on the conflict path set C and the state difference vector set , adopt an iterative state evolution strategy to rearrange the dependency relationship of the positions of conflicting instruction pairs in the execution order. In each round of iteration, reorder the instruction stream according to the resource priority, production process order, and task urgency. The updated state vector is used to reconstruct the internal memory content of the controller and complete a round of neural Turing machine state refresh;

[0051] S54. Generate a set of candidate scheduling plans through iterative state evolution and conflict rearrangement operations. Each scheduling plan represents a conflict-free scheduling instruction sequence that satisfies the current resource constraints and production order, and upload the set of candidate scheduling plans in real time for obtaining feedback from the execution layer of the MES system.

[0052] The beneficial effects of the present invention are as follows:

[0053] First, the present invention effectively extracts the similarity features between each time step in the lithium battery production data by constructing a Siamese network based on a two-tower structure, solves the problem that traditional scheduling methods cannot identify complex timing patterns and periodic repetitive behaviors, and realizes high-dimensional modeling and feature expression of the implicit similar structure in the production process, providing a more discriminative input basis for the subsequent scheduling logic.

[0054] Second, use the Transformer-XL model to perform long-term dependency modeling on the fused timing features, combine the relative position embedding and the scalable memory cache mechanism, successfully capture the scheduling relationships and evolution laws across cycles, significantly improve the expression ability of the scheduling model for long-distance timing dependencies, make up for the deficiencies of traditional RNN and LSTM in processing long sequence information such as memory decay and state loss, and enhance the adaptability and robustness of the model in the face of complex situations such as sudden order changes and production capacity constraint adjustments.

[0055] Finally, by means of the readable and writable external memory mechanism introduced by the neural Turing machine structure, the dynamic reading and writing operations of the scheduling tensor are completed under the drive of the controller, realizing the generation of the scheduling instruction stream under high-complexity conditions. Through the controller memory comparison mechanism, conflict detection and dependency rearrangement are carried out, effectively improving the logical consistency and scheduling accuracy of the instruction output. This mechanism, combined with the feedback link of the MES execution layer, realizes the dynamic update of the model, has strong real-time performance and closed-loop learning ability, and comprehensively enhances the decision-making efficiency and execution flexibility of the MES system in the lithium battery industry for intelligent scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0057] Figure 1 is a flowchart of a scheduling method for a lithium battery industry MES system based on deep learning proposed by the present invention;

[0058] Figure 2 is a schematic diagram of the overall structure of the Siamese network of a scheduling method for a lithium battery industry MES system based on deep learning proposed by the present invention;

[0059] Figure 3 is a schematic diagram of the scheduling dependency and feature fusion process of a scheduling method for a lithium battery industry MES system based on deep learning proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0061] Refer to Figures 1-3 , a scheduling method for a lithium battery industry MES system based on deep learning, includes the following steps:

[0062] S1. Collect lithium battery production data and perform preprocessing to obtain a standardized sequence;

[0063] S2. Construct a Siamese network based on a two-tower structure, perform double-branch encoding on the standardized sequence, extract the similarity features of the standardized sequence through a shared-weight sub-network, and generate a similarity matrix;

[0064] S3. Concatenate and fuse the similarity matrix with the standardized sequence to form a feature fusion sequence, and use the Transformer-XL model to perform long-term dependency modeling on the feature fusion sequence, and extract the scheduling dependency relationship across cycles through relative position embedding and an extensible memory caching mechanism;

[0065] S4. Convert the production scheduling dependency relationship into a production scheduling tensor containing correlation weights, and input the production scheduling tensor into a neural Turing machine. The neural Turing machine includes a readable and writable external storage matrix and a controller, and generates an original production scheduling instruction stream through read and write operations on the production scheduling tensor;

[0066] S5. Based on the memory comparison mechanism inside the controller, perform conflict detection and dependency relationship rearrangement on the original production scheduling instruction stream, and generate a set of candidate production scheduling plans through iterative update of the neural Turing machine state;

[0067] S6. Feed the set of candidate production scheduling plans back to the execution layer of the MES system for production plan execution, periodically collect feedback data, and perform iterative update on the parameters of the Siamese network and the Transformer-XL model.

[0068] By constructing a complete production scheduling process, the present invention integrates data preprocessing, similarity modeling, long-term dependency analysis, production scheduling instruction generation, and feedback iteration mechanism to form a deep learning production scheduling method for the lithium battery manufacturing scenario, which solves the problems of slow production scheduling response, difficult resource coordination, and lack of dynamic update ability under complex process flows, and significantly improves the intelligent production scheduling level of the MES system.

[0069] In this embodiment, the production data includes equipment status, production demand, material inventory, production progress, production processes, and historical production data.

[0070] By integrating multi-source data such as equipment status, production demand, material inventory, production progress, production processes, and historical records in lithium battery production data, the present invention ensures the integrity and timeliness of production scheduling input, makes production scheduling decisions closer to actual on-site conditions, and enhances the usability and adaptability of the model.

[0071] In this embodiment, the preprocessing includes unified timestamp alignment, discrete value normalization, missing item interpolation and noise value removal.

[0072] By means of the preprocessing methods of unified timestamp alignment, discrete value normalization, missing item interpolation and noise value removal, the present invention effectively cleans abnormal and inconsistent problems in the original data, ensures the quality of the subsequent model input data, and improves the stability and accuracy of feature modeling and training.

[0073] In this embodiment, the specific content of S2 includes:

[0074] S21. Denote the standardized sequence as , where represents the feature vector at time step t, t is the current time step index, and satisfies , where \(T\) is the total number of time steps, the sequence \(X\) is input in a dual-path manner and fed into the left tower encoder and the right tower encoder respectively;

[0075] S22. Both the left tower encoder and the right tower encoder are composed of a convolutional sub-network with shared weights and a fully connected layer, which are used to extract features from the input sequence. Let the output of the left tower encoder be , and the output of the right tower encoder be , where , are the feature vectors output by the left tower encoder and the right tower encoder after encoding, and \(t\) is the current time step index, satisfying ;

[0076] S23. Calculate the similarity between the left tower output feature and the right tower output feature to generate a similarity matrix \(S\):

[0077] ;

[0078] where, represents the similarity degree between time step \(i\) and time step \(j\) in the similarity matrix, represents the natural exponential function with base \(e\), represents the inner product of the left tower and right tower feature vectors, represents the square of the Euclidean distance, \(W\) represents the learnable bilinear mapping weight matrix, represents the bilinear similarity mapping under the learnable weight matrix, , are preset non-negative hyperparameters, and \(T\) is the total number of time steps.

[0079] The present invention constructs a Siamese network based on shared weights, uses a two-tower structure to extract similarities from the standardized sequence, realizes dynamic matching modeling between time steps, generates a high-dimensional similarity matrix representing the global similarity structure, and provides an accurate basis for subsequent long-term dependence analysis.

[0080] In this embodiment, the Siamese network adopts a two-tower structure design, including two feature encoding sub-networks with shared parameters. Each sub-network consists of two one-dimensional convolutional layers, a batch normalization layer, and a ReLU activation function, which are used to perform parallel feature extraction on the input standardized production data sequence, and calculate the Euclidean distance, inner product, and weighted bilinear similarity between any time step feature vectors at the output stage, and finally generate a two-dimensional similarity matrix for representing the similarity between different time steps.

[0081] The Siamese network two-tower structure proposed by the present invention combines convolutional coding and non-linear activation, enhancing the model's ability to extract fine-grained features in the input sequence. At the same time, in the similarity calculation, it combines Euclidean distance, vector inner product, and bilinear mapping, improving the expressive ability of time series similarity modeling.

[0082] In this embodiment, S3 specifically includes:

[0083] S31. Concatenate and fuse the normalized sequence X and the similarity matrix S to construct a feature fusion sequence , where T is the total number of time steps, and the feature fusion vector is expressed as , represents the feature vector at time step t, represents the similarity vector of the t-th row, represents the vector concatenation operation;

[0084] S32. Use the Transformer-XL model to perform long-term dependence modeling on the feature fusion sequence. Input the fusion vector sequence F into the Transformer-XL model, and use relative position embedding and an expandable memory cache mechanism to model time correlation, constructing a cross-cycle state matrix Z:

[0085] ;

[0086] Among them, represents the state at time step t, T is the total number of time steps, represents the query vector, represents the key vector, represents the value vector, represents the relative position vector, , , is a learnable projection matrix, represents the fused historical state vector, b is the bias term, represents the non-linear transformation function, is the normalized attention weight of the t-th query and the i-th key, is a very small constant to avoid division by zero error, represents the Hadamard element-wise product, represents the L2 norm, m represents the number of vectors stored in the historical memory cache, represents the sine function, represents the cosine function;

[0087] S33. For each and the historical state cache matrix Perform a non-linear combination on the corresponding dimensional states to generate a scheduling dependency sequence D:

[0088] ;

[0089] Among them, represents the scheduling dependency vector at time step t, represents an activation function with residual connections, represents the square of the Euclidean distance between the current state and the historical state, represents the natural exponential function with base e, represents the Sigmoid function, 、 are weight matrices, is a bias term, is a scaling factor, represents an element-wise maximum operation.

[0090] In the present invention, by splicing and fusing the similarity matrix with the standardized sequence, inputting it into the Transformer-XL model for long-term dependency modeling, and combining the relative position embedding and the historical cache mechanism, the implicit associations and long-distance dependencies between cross-cycle processes in lithium battery production are accurately characterized, effectively improving the temporal rationality of the scheduling plan.

[0091] In this embodiment, the feature fusion sequence F is formed by fusing the row vectors corresponding to each time step in the similarity matrix with the feature vectors of the corresponding time steps in the standardized sequence in a dimension-wise splicing manner, forming a fusion vector sequence with a dimension equal to the sum of the original feature dimension and the similarity vector dimension, maintaining strict alignment in the time domain, and enabling each fusion vector to contain both the original production state at the current moment and the global similarity structure information, so as to enhance the ability of the Transformer-XL model to jointly model local and global features when modeling cross-cycle scheduling dependency relationships.

[0092] The feature fusion method designed in the present invention strictly aligns time steps, fuses the current production state with the global similarity structure, enabling each input at each moment to have the ability to express local-global mixed features, and enhancing the modeling ability and prediction accuracy of the Transformer-XL model for complex scheduling patterns.

[0093] In this embodiment, S4 specifically includes:

[0094] S41. Convert the scheduling dependency sequence D into a scheduling tensor P containing association weights, and the construction method is as follows:

[0095] Construct a corresponding dependency combination vector for any two time steps i and j , and all dependency combination vectors After being embedded and mapped, they are stacked to form a third-order scheduling tensor , where represents a ternary relation combination function, represents the scheduling dependence vector at time step i, represents the scheduling dependence vector at time step j, , representing the time step distance, and T is the total number of time steps;

[0096] S42. Input the scheduling tensor P into a neural Turing machine, which is composed of an external storage matrix M and a set of controller state vectors H. Let the current time step controller state be , and generate a scheduling instruction stream Y through read and write operations:

[0097] ;

[0098] where represents the scheduling instruction at time step t, represents the controller state vector at time step t, , , represent trainable weight matrices, is a bias vector, represents the attention weight of the combination index (i, j) at the t-th moment in the scheduling tensor, represents the dependence combination vector in the scheduling tensor , is the read weight of the controller for the storage unit index k, and N represents the number of storage units, represents the content of the k-th storage unit, represents the normalized activation mapping function;

[0099] S43. According to the scheduling instruction at each moment, perform an internal update on the controller state vector , and perform an update on the external storage matrix M at time step t + 1. The update process is jointly determined by the erasure vector , the addition vector and the write address weight :

[0100] ;

[0101] where represents the content of the k-th storage unit at time step t + 1, represents the content of the k-th storage unit at time step t, represents the Hadamard element-wise product, represents the tensor product, , represents the linear transformation matrix of the erasure vector and the addition vector, represents the interaction mapping matrix of the controller state, represents the bilateral attention coefficient function, is the Sigmoid function, is a non-zero constant for normalization.

[0102] In this embodiment, the ternary relation combination function is defined as:

[0103] ;

[0104] where, , respectively represent the scheduling dependency vectors at time steps i and j, represents the time step distance, is a vector with all elements being 1, is the element-wise squared difference, represents the Hadamard product, represents the vector concatenation operation, is the linear transformation matrix, is the bias vector, represents the mapping function with non-linear activation.

[0105] The ternary relation combination function proposed by the present invention integrates the scheduling dependency vector, the time step difference and the non-linear relationship between vectors, and improves the expression ability of the scheduling tensor construction through a structured mapping method, ensuring the logical integrity and relevance of the input of the neural Turing machine, and further improving the context consistency of instruction generation.

[0106] In this embodiment, S5 specifically includes:

[0107] S51. Receive the scheduling instruction stream sequence , and compare the scheduling instruction at each moment with the controller state vector , and construct a memory difference vector set in combination with the state difference in the time series, where T is the total number of time steps, and each difference vector is obtained by the linear combination of the current instruction vector and the state difference, and is used to characterize the content offset degree between scheduling instructions;

[0108] S52. According to the memory difference vector set Perform a conflict detection operation on the production scheduling instruction stream Y, and construct a production scheduling conflict graph G=(V,E), where the vertex set V corresponds to the production scheduling instructions at each moment, and the edges in the edge set E indicate that there are resource conflicts, timing conflicts, or material overlaps between any two instructions. By traversing each edge, extract the conflict path set C for dependency rearrangement processing;

[0109] S53. Based on the conflict path set C and the state difference vector set , adopt an iterative state evolution strategy to rearrange the dependency relationship of the positions of conflicting instruction pairs in the execution order. In each round of iteration, reorder the instruction stream according to the resource priority, production process order, and task urgency. The updated state vector is used to reconstruct the internal memory content of the controller and complete a round of neural Turing machine state refresh;

[0110] S54. Generate a candidate production scheduling plan set through iterative state evolution and conflict rearrangement operations, where each production scheduling plan represents a conflict-free production scheduling instruction sequence that meets the current resource constraints and production order, and upload the candidate production scheduling plan set in real time for obtaining feedback from the execution layer of the MES system.

[0111] The present invention realizes the automatic correction of resource conflicts and execution dependencies by introducing a conflict detection method based on the controller memory comparison mechanism, combining the production scheduling conflict graph modeling and the state evolution rearrangement strategy, and finally outputs diverse candidate production scheduling plans that meet the process logic, resource limitations, and production rhythm, significantly improving the reliability, flexibility, and practicality of the production scheduling decision-making.

[0112] Embodiment 1:

[0113] To verify the feasibility of the present invention in implementation, apply the present invention to a certain lithium battery intelligent manufacturing production line, and conduct production scheduling optimization practice for the problems of lagging plan response, frequent equipment resource allocation conflicts, and inability to dynamically respond to process changes in its production scheduling link. The production line processes approximately 95 tons of raw materials per day, involving 19 types of raw materials, 7 major production sections, and 27 specific processes, and the total number of equipment reaches 114. The production tasks highly depend on multi-process collaborative operations, and at the peak, it will receive no less than 32 batches of order demands per day, with a tight production rhythm, and extremely high requirements for the response speed, accuracy, and flexible scheduling ability of the production scheduling system.

[0114] During the actual operation process, the rule-based scheduling method used by traditional MES systems shows significant limitations. Data monitoring found that, on average, the cumulative idling time of equipment due to scheduling conflicts reaches 276 hours per month, the material waiting time exceeds 93 hours, and the production capacity utilization rate has long hovered below 82%. Especially in the case of sudden orders or temporary equipment maintenance, the original system cannot perform rapid rescheduling based on the global resource status, resulting in the delay of delivery of some high-priority orders and affecting the stability of the overall delivery period.

[0115] Based on the above problems, the method proposed in the present invention is embedded in the existing MES system. Using the original production data (equipment status, order requirements, inventory situation, etc.) as input, a Siamese network is used to construct a similarity structure between time steps, and then it is integrated into the Transformer-XL network to model long-term dependencies. Finally, the scheduling instruction generation is completed through the neural Turing machine module. At the initial stage of system deployment, 18,672 historical order data for 3 months were selected for pre-training. After standardization, each order corresponds to an average of 46 time-step sequence points. After using GPU parallel acceleration for training, the model converges stably at the 18th round, and the training takes about 7 hours.

[0116] The actual online test period is set to 30 days. During this period, the system processed a total of 896 batches of orders, and the average daily order processing volume increased to 29.9 batches. The response time of the scheduling model for processing sudden orders (such as inserting an emergency process flow within 2 hours) decreased from 21 minutes in the original method to 3.7 minutes, and the correct rate of automatic instruction generation increased to 98.3%. When the system faces equipment maintenance interference, the average time to restore an executable scheduling plan decreased from 12 minutes to 4.2 minutes. At the same time, the average idling time of the equipment decreased from the original 9.2 hours per day to 2.6 hours, the material waiting time was shortened to within 1.3 hours per day, and the production capacity utilization rate steadily increased to 91.7%.

[0117] In addition, the introduction of the feedback mechanism enables the model to automatically complete weight fine-tuning according to the actual progress data returned by the MES execution layer, and perform model micro-updates at night every day. On average, each round of parameter correction involves 460,000 weight nodes. The dynamic nature of the model structure ensures that, in the face of changes in production conditions on different production days, it can still generate a reasonable set of candidate scheduling plans with process logic consistency, further improving the stability and generalization ability of the scheduling strategy.

[0118] The following table shows the comparison of the main scheduling performance before and after the deployment of the present invention:

[0119] Table 1 Comparison of Core Indicators of Deep Learning Scheduling Method before and after Deployment

[0120] ;

[0121] As can be seen from the above implementation, the method proposed by the present invention can effectively improve the scheduling intelligence level in the lithium battery MES production scheduling scenario and has significant optimization value. It shows strong adaptability in key links such as multi-process collaboration, resource conflict avoidance, and production plan reconstruction, not only improving the overall production efficiency of the factory but also providing a technical route for subsequent promotion and application in other complex manufacturing scenarios.

[0122] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A scheduling method for the MES system in the lithium battery industry based on deep learning, characterized in that, It includes the following steps: S1. Collect the production data of lithium batteries and perform preprocessing to obtain a standardized sequence; S2. Construct a Siamese network based on a two-tower structure, perform dual-branch encoding on the standardized sequence, extract the similarity features of the standardized sequence through a shared-weight subnetwork, and generate a similarity matrix; S3. Concatenate and fuse the similarity matrix with the standardized sequence to form a feature fusion sequence, and use the Transformer-XL model to perform long-term dependence modeling on the feature fusion sequence, and extract the production scheduling dependence relationship across cycles through relative position embedding and an extensible memory caching mechanism; S4. Convert the production scheduling dependence relationship into a production scheduling tensor containing associated weights, and input the production scheduling tensor into a neural Turing machine. The neural Turing machine includes a readable and writable external storage matrix and a controller, and generates an original production scheduling instruction stream through read and write operations on the production scheduling tensor; S5. Based on the memory comparison mechanism inside the controller, perform conflict detection and dependence relationship rearrangement on the original production scheduling instruction stream, and generate a set of candidate production scheduling plans by iteratively updating the state of the neural Turing machine; S6. Feed the set of candidate production scheduling plans back to the execution layer of the MES system for production plan execution, periodically collect feedback data, and iteratively update the parameters of the Siamese network and the Transformer-XL model.

2. The scheduling method of the MES system for the lithium battery industry based on deep learning according to claim 1, characterized in that, The production data includes equipment status, production demand, material inventory, production progress, production processes, and historical production data.

3. A scheduling method for a MES system in the lithium battery industry based on deep learning according to claim 1, characterized in that, The preprocessing includes unified timestamp alignment, discrete value normalization, missing item interpolation and noise value elimination.

4. A scheduling method for a MES system in the lithium battery industry based on deep learning according to claim 1, characterized in that, The specific content of S2 includes: S21. Denote the standardized sequence as , where represents the feature vector at time step t, t represents the current time step, and it satisfies . T is the total number of time steps. The sequence X is input in a dual-path manner and is respectively input into the left tower encoder and the right tower encoder; S22. Both the left tower encoder and the right tower encoder are composed of a convolutional sub-network and a fully connected layer with shared weights, which extract features from the input sequence. Let the output of the left tower encoder be , and the output of the right tower encoder be , where , are the feature vectors output by the left tower encoder and the right tower encoder after encoding, t represents the current time step, and satisfies ; S23. Output the features of the left tower and the output features of the right tower to perform similarity calculation and generate a similarity matrix: ； Among them, represents the similarity degree between time step i and time step j in the similarity matrix, represents the natural exponential function with base e, represents the inner product of the feature vectors of the left tower and the right tower, represents the square of the Euclidean distance, and W represents the learnable bilinear mapping weight matrix, represents the bilinear similarity mapping under the learnable weight matrix, and are preset non - negative hyperparameters, and T is the total number of time steps.

5. A scheduling method for a lithium battery industry MES system based on deep learning according to claim 4, characterized in that, The Siamese network adopts a two-tower structure design, including two feature encoding subnets with shared parameters. Each subnet consists of two layers of one-dimensional convolutional layers, batch normalization layers, and ReLU activation functions, and is used to perform parallel feature extraction on the input standardized production data sequence, and calculate the Euclidean distance, inner product, and weighted bilinear similarity between feature vectors at any time step in the output stage, and finally generate a two-dimensional similarity matrix for characterizing the similarity between different time steps.

6. A scheduling method for a MES system in the lithium battery industry based on deep learning according to claim 1, characterized in that, The specific content of S3 includes: S31. Concatenate and fuse the standardized sequence X with the similarity matrix S to construct a feature fusion sequence , where T is the total number of time steps, and the feature fusion vector is expressed as , represents the feature vector at time step t, represents the similarity vector at time step t, represents the vector concatenation operation; S32. Use the Transformer-XL model to perform long-term dependence modeling on the feature fusion sequence. Input the feature fusion sequence F into the Transformer-XL model, and use relative position embedding and an extensible memory caching mechanism to model time correlation to construct a cross-cycle state matrix Z; ; in, represents the state at time step t, T is the total number of time steps, represents the query vector, represents the key vector, represents a value vector, represents the relative position vector, , , is the learnable projection matrix, represents the fused historical state vector, b is the bias term, represents the nonlinear transformation function, is the normalized attention weight of the t-th query and the i-th key, is a very small constant to avoid division by zero errors, represents the Hadamard element-wise product, represents the L2 norm, m represents the number of vectors stored in the history memory cache, represents the sine function, represents the cosine function; S33. Combine each with the corresponding dimension state in the historical status cache matrix to generate a production scheduling dependency sequence D: ; Among them, represents the production scheduling dependence vector at time step t, represents the activation function with residual connection, represents the square of the Euclidean distance between the current state and the historical state, represents the natural exponential function with base e, represents the Sigmoid function, 、 are weight matrices, is the bias term, is the scaling coefficient, represents the element-wise maximum operation.

7. A scheduling method for a lithium battery industry MES system based on deep learning according to claim 6, characterized in that, The feature fusion sequence F is formed by fusing the row vectors corresponding to each time step in the similarity matrix with the feature vectors at the corresponding time steps in the standardized sequence in a dimension-wise concatenation manner, forming a fusion vector sequence with a dimension equal to the sum of the original feature dimension and the similarity vector dimension, maintaining strict alignment in the time domain, and enabling each fusion vector to contain both the original production state at the current moment and the global similarity structure information, so as to enhance the ability of the Transformer-XL model to jointly model local and global features when modeling cross-cycle production scheduling dependence relationships.

8. A scheduling method for a MES system in the lithium battery industry based on deep learning according to claim 1, characterized in that, The specific content of S4 includes: S41. Convert the production scheduling dependency sequence D into a production scheduling tensor P containing associated weights, which is constructed as follows: Construct the corresponding dependency combination vectors for any two time steps i and j , and stack all the dependency combination vectors after embedding mapping to form a third-order scheduling tensor , where represents the ternary relation combination function, represents the scheduling dependency vector at time step i, represents the scheduling dependency vector at time step j, , represents the time step distance, and T is the total number of time steps; S42. Input the scheduling tensor P into a neural Turing machine, which consists of an external storage matrix M and a set of controller state vectors H. Let the controller state at the current time step be , and generate a scheduling instruction stream Y through read and write operations: ; Among them, represents the production scheduling instruction at time step t, represents the controller state vector at time step t, , , represent trainable weight matrices, is the bias vector, represents the attention weight for the combination index (i, j) at the t-th moment in the production scheduling tensor, represents the dependency combination vector in the production scheduling tensor , is the read weight of the controller for the storage unit index k, and N represents the number of storage units, represents the content of the k-th storage unit, represents the normalized activation mapping function; S43. According to the production scheduling instructions at each moment internally update the controller state vector and perform the update of the external storage matrix M at time step t + 1, and the update process is determined by the erasure vector , the addition vector and the write address weight jointly: ; wherein, represents the content of the k-th storage unit at time step t + 1, represents the content of the k-th storage unit at time step t, represents the Hadamard element-wise product, represents the tensor product, , represents the linear transformation matrix of the erasure vector and the addition vector, represents the interaction mapping matrix of the controller state, represents the bilateral attention coefficient function, is the Sigmoid function, is a non-zero constant for normalization.

9. A scheduling method for a MES system in the lithium battery industry based on deep learning according to claim 8, characterized in that, The ternary relation combination function is defined as: ; Among them, , represent the scheduling dependence vectors at time steps i and j respectively, represents the time step distance, is a vector with all elements being 1, is the element-wise squared difference, represents the Hadamard product, represents the vector concatenation operation, is the linear transformation matrix, is the bias vector, represents the mapping function with a non-linear activation.

10. A scheduling method for a lithium battery industry MES system based on deep learning according to claim 1, characterized in that, The specific steps of S5 are as follows: S51. Receive the production scheduling instruction stream sequence , and compare the production scheduling instruction at each moment with the controller state vector , and construct a set of memory difference vectors by combining the state differences in the time series, where T is the total number of time steps, and each difference vector is obtained by the linear combination of the current instruction vector and the state difference, and is used to characterize the content offset degree between production scheduling instructions; S52. According to the memory difference vector set Perform a conflict detection operation on the production scheduling instruction stream Y to construct a production scheduling conflict graph G=(V,E), where the vertex set V corresponds to the production scheduling instructions at each moment, and the edges in the edge set E indicate that there are resource conflicts, timing conflicts, or material overlaps between any two instructions. By traversing each edge, extract the conflict path set C for dependency rearrangement processing; S53. Based on the conflict path set C and the state difference vector set , an iterative state evolution strategy is adopted to rearrange the dependency relationship of the positions of conflicting instruction pairs in the execution order. In each round of iteration, the instruction stream is reordered according to resource priority, production process order, and task urgency. The updated state vector is used to reconstruct the internal memory content of the controller and complete a round of neural Turing machine state refresh; S54. Generate a set of candidate production scheduling plans through iterative state evolution and conflict rearrangement operations. Each production scheduling plan represents a conflict-free production scheduling instruction sequence that satisfies the current resource constraints and production order, and upload the set of candidate production scheduling plans in real time for obtaining feedback from the execution layer of the MES system.

Images