Intelligent production scheduling system and method considering alternative materials and switching costs

Abstract

The application provides a kind of intelligent production scheduling system and method considering alternative materials and switching cost, it is related to production scheduling technical field, the system includes: bipartite graph generation module, according to the order information and alternative material information input by user, generate bipartite graph, including material node and order node, the characteristic information of material node is the material switching cost information of corresponding material, the characteristic information of order node is the order production capacity information of corresponding order, there is edge connection between order node and the order selectable material node, the characteristic information of edge is the consumption amount information of connected order to material;Input coding module is used to generate order coding vector and material coding vector according to bipartite graph;Material encoder is used to carry out deep feature extraction to material coding vector, generate deep material coding vector;Decoder is used to output production scheduling plan according to order coding vector, deep material coding vector and current production scheduling information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of production scheduling, and in particular to an intelligent production scheduling system and method considering alternative materials and switching costs. BACKGROUND

[0002] "Material management and production scheduling" is an important part of fine scheduling in an advanced planning and scheduling (APS) system, and its core task is to select materials and arrange the processing order of orders while considering the selection of alternative materials and the switching of materials during the production process. The decision result has an important influence on production cost and production efficiency. In discrete manufacturing, with the growth of production demand for large-scale personalized orders, material complexity is an important factor restricting the production efficiency of personalized orders. Specifically, on the one hand, production materials have alternative, and different specifications of materials can be selected for each order to produce, but too arbitrary selection of materials will also cause problems such as increased production cost, material shortage or material accumulation, and reduced production efficiency and increased total cost; on the other hand, the continuous use of different materials in the production line will bring additional switching costs: for example, workers need time and energy to think and find the right materials, which are all ineffective losses, reducing production efficiency and quality, and need to be reasonably scheduled to reduce the number of material switching.

[0003] Therefore, it is necessary to develop an intelligent production scheduling system and method considering alternative materials and switching costs, so as to consider alternative materials and material switching costs when making decisions on production scheduling plans, so as to ensure production efficiency and product quality. SUMMARY

[0004] In view of the above problems, the embodiments of the present application provide an intelligent production scheduling system and method considering alternative materials and switching costs, so as to overcome the above problems or at least partially solve the above problems.

[0005] The first aspect of the embodiments of the present application provides an intelligent production scheduling system considering alternative materials and switching costs, the system comprising:

[0006] A bipartite graph generation module generates a bipartite graph according to user input order information and alternative material information, the bipartite graph comprising material nodes and order nodes, the feature information of the material nodes being material switching cost information of the corresponding materials, the feature information of the order nodes being order production quantity information of the corresponding orders, there being an edge connection between the order nodes and the material nodes selectable by the order, the feature information of the edge being consumption quantity information of the connected order-material pair;

[0007] An input encoding module is configured to generate an order encoding vector and a material encoding vector according to the bipartite graph;

[0008] A material encoder is configured to perform deep feature extraction on the material encoding vector to generate a deep material encoding vector;

[0009] A decoder is configured to output a production scheduling plan according to the order encoding vector, the deep material encoding vector, and current production scheduling information.

[0010] The second aspect of the embodiment of the present application further provides an intelligent production scheduling method considering alternative materials and switching costs, which is applied to the intelligent production scheduling system of the first aspect of the embodiment of the present application, and the method comprises the following steps:

[0011] A bipartite graph is generated according to order information and alternative material information input by a user, the bipartite graph comprises material nodes and order nodes, feature information of the material nodes is material switching cost information of corresponding materials, feature information of the order nodes is order production quantity information of corresponding orders, there is an edge connection between the order nodes and the material nodes that can be selected by the orders, and feature information of the edge is consumption quantity information of the connected orders to materials;

[0012] An input encoding module is configured to generate an order encoding vector and a material encoding vector according to the bipartite graph;

[0013] A material encoder is configured to perform deep feature extraction on the material encoding vector to generate a deep material encoding vector;

[0014] A decoder is configured to output a production scheduling plan according to the order encoding vector, the deep material encoding vector, and current production scheduling information.

[0015] The third aspect of the embodiment of the present application further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory, the processor executes the computer program to implement the steps in the intelligent production scheduling method considering alternative materials and switching costs according to the second aspect of the embodiment of the present application.

[0016] The fourth aspect of the embodiment of the present application further provides a computer readable storage medium having a computer program / instruction stored thereon, the computer program / instruction is executed by a processor to implement the steps in the intelligent production scheduling method considering alternative materials and switching costs according to the second aspect of the embodiment of the present application.

[0017] The fifth aspect of the embodiment of the present application further provides a computer program product, when the computer program product runs on an electronic device, the processor executes the steps in the intelligent production scheduling method considering alternative materials and switching costs according to the second aspect of the embodiment of the present application.

[0018] The embodiment of the application provides an intelligent production scheduling system considering alternative materials and switching costs, the system comprises: a bipartite graph generation module, which generates a bipartite graph according to user input order information and alternative material information, the bipartite graph comprises material nodes and order nodes, the feature information of the material node is the material switching cost information of the corresponding material, the feature information of the order node is the order production quantity information of the corresponding order, there is an edge connection between the order node and the material node selectable for the order, and the feature information of the edge is the consumption quantity information of the connected order to the material; an input encoding module is used to generate order encoding vectors and material encoding vectors according to the bipartite graph; a material encoder is used to perform deep feature extraction on the material encoding vectors to generate deep material encoding vectors; and a decoder is used to output a production scheduling plan according to the order encoding vectors, the deep material encoding vectors and current production scheduling information.

[0019] Specific beneficial effects are that the application provides an intelligent production scheduling system considering alternative materials and material switching costs, the system makes decisions (production scheduling plan) on the production order of the order and the selection of alternative materials for each order required by the user given material and order information, so that the production demand is met and the use and switching cost of the material are minimized. The bipartite graph generation module proposes a novel bipartite graph-based production scheduling problem representation, thereby parameterizing the problem instance (user input order information and alternative material information), and only one encoding is required in the processing process, and the scheduling scheme is iteratively output by the encoder. This way is fast, and the output solution is better than the heuristic algorithm designed by human. Based on the generated bipartite graph representation, global information can be effectively utilized to make decisions to generate an optimal scheduling scheme, and the application has the advantages of self-learning, high efficiency and quality, strong generalization, etc. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed to be used in the description of the embodiments of the application will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0021] Figure 1 is a structural schematic diagram of an intelligent production scheduling system considering alternative materials and switching costs provided by the embodiment of the application;

[0022] Figure 2 is a structural schematic diagram of a bipartite graph provided by the embodiment of the application;

[0023] Figure 3 is an execution flow schematic diagram of an intelligent production scheduling system provided by an embodiment of the present application;

[0024] Figure 4 is a step flow schematic diagram of an intelligent production scheduling method provided by an embodiment of the present application;

[0025] Figure 5 is a schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0026] Exemplary embodiments of the present application will be described herein below with reference to the accompanying drawings. Although exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present application can be more thoroughly and completely understood, and so that the scope of the present application can be accurately conveyed to those skilled in the art.

[0027] “Material management and production scheduling” is an important part of detailed scheduling in an advanced planning and scheduling (APS) system, and its core task is to make material selection and order processing sequence arrangement while considering the selection of alternative materials and material switching in the production process. The decision result has an important influence on production cost and production efficiency. In discrete manufacturing, with the growth of production demand for large-scale personalized orders, material complexity is an important factor restricting the production efficiency of personalized orders. Specifically, on the one hand, production materials have alternative, and each order can be produced according to the actual selection of different specifications of materials, but too arbitrary material selection will also cause problems such as increased production cost, material shortage or material accumulation, and reduced production efficiency and increased total cost; on the other hand, the continuous use of different materials in the production line will bring additional switching costs: for example, workers need time and energy to think and find the right material, which are all ineffective losses, reducing production efficiency and quality, and need to be reasonably scheduled to reduce the number of material switching.

[0028] In manufacturing enterprises, production scheduling considering alternative materials and switching costs is often based on a predetermined production sequence and relies on the experience of workers to make small adjustments according to the actual material situation. Specifically, this mode determines the processing order of orders according to the production priority of orders, and the material department supplies materials to the corresponding production line according to the production plan for a period of time in the future. In actual production, line workers select alternative materials used by orders according to experience and material remaining, and make small adjustments to the production sequence to reduce material switching. Although this method can optimize production efficiency to a certain extent, this mode highly dependent on the experience of workers has several drawbacks: 1) The cooperation between the material department and the production department is required to be high, and the material department needs to prepare materials after the production plan is determined, and the production line can start production only after the materials are complete. This production mode is low in efficiency, and is prone to conflict between the production plan and the material complete, which greatly affects the original production plan. 2) The actual production sequence is largely dependent on the experience of workers, and it is difficult to use global production information, and only small adjustments based on the established production plan are allowed, so the optimization effect is very limited, and the actual production situation cannot be well responded.

[0029] In view of the above problems, the embodiments of the present application provide an intelligent production scheduling system and method considering alternative materials and switching costs, to realize the consideration of alternative materials and switching costs when deciding the production scheduling plan, so as to ensure production efficiency and product quality. The intelligent production scheduling system considering alternative materials and switching costs provided by the embodiments of the present application will be described in detail in combination with some embodiments and application scenarios and the accompanying drawings.

[0030] The first aspect of the embodiments of the present application provides an intelligent production scheduling system considering alternative materials and switching costs, referring to Figure 1 , Figure 1 The structure diagram of the intelligent production scheduling system considering alternative materials and switching costs provided by the embodiments of the present application is shown in Figure 1 , which comprises:

[0031] A bipartite graph generation module generates a bipartite graph according to the order information and alternative material information input by a user, the bipartite graph comprises material nodes and order nodes, the feature information of the material nodes is the material switching cost information of the corresponding materials, the feature information of the order nodes is the order production quantity information of the corresponding orders, there is an edge connection between the order nodes and the material nodes selectable by the order, and the feature information of the edge is the consumption quantity information of the connected order to the material.

[0032] An input encoding module is configured to generate an order encoding vector and a material encoding vector according to the bipartite graph.

[0033] The material encoder is configured to perform deep feature extraction on the material encoding vector to generate a deep material encoding vector.

[0034] The decoder is configured to output a production scheduling plan according to the order encoding vector, the deep material encoding vector and current production scheduling information.

[0035] In the embodiment, a novel production scheduling problem representation method based on bipartite graph considering alternative materials and material switching cost is proposed, that is, a bipartite graph generation module is used to parameterize the problem instance (i.e., order information and alternative material information) input by a user. The order information represents the order production quantity and other information of all orders that have not been produced and are being produced. For each order, the material used for production is often not fixed and unique, and one of multiple alternative materials can be selected for production of the order. The alternative material information represents information about the materials that can be used for production of each order, such as material storage quantity, production consumption quantity and other information. In the embodiment, a corresponding bipartite graph representation is generated according to the above information.

[0036] The bipartite graph problem representation proposed in the application is derived from the mixed integer linear programming (MILP) form of the production scheduling problem considering alternative materials and switching cost, and the graph nodes are appropriately merged while retaining the effective information of the problem instance (i.e., the order information and alternative material information input by the user). Specifically, the bipartite graph includes material nodes and order nodes. Referring to Figure 2 , Figure 2 A generation structure diagram of a bipartite graph is shown, as shown in Figure 2 The related variables in the MILP form of the production scheduling problem are merged to finally generate a bipartite graph. The nodes in the bipartite graph are divided into two columns, one column of nodes is a material node, representing all alternative materials considered in the problem instance, for example, material node A represents alternative material A, and material node B represents alternative material B. The other column of nodes is an order node, representing all orders considered in the problem instance, for example, order node 1 represents order 1, and order node 2 represents order 2.

[0037] In actual production, using different materials (for example, originally using material A, and then switching to using material B for order production) will bring additional material switching costs. Therefore, when generating the bipartite graph, information related to the material switching cost is added to the bipartite graph. Specifically, the material switching cost information of the material is represented as the node feature of the corresponding material node, and the production quantity information (such as the total production quantity of the order) of the order is represented as the node feature of the corresponding order node. When a certain order needs a material, there is an edge connection between the corresponding order node and the material node. For example, order 1 can use material A or choose to use material B for production, and there is an edge connection between order 1 corresponding to the order node 1 and the material node A corresponding to the material A, and there is an edge connection between order 1 corresponding to the order node 1 and the material node B corresponding to the material B. The edge feature is the consumption quantity of the order to the material connected.

[0038] In this embodiment, after generating the bipartite graph, the input encoding module is used to generate order encoding vectors and material encoding vectors according to the bipartite graph. Specifically, the input encoding module takes the bipartite graph as input, and uses the original features and topological structure of the bipartite graph to perform message passing, thereby generating corresponding high-dimensional encoding vectors for each material node and order node.

[0039] In a possible implementation, the generating of the order encoding vectors and the material encoding vectors according to the bipartite graph comprises:

[0040] According to the bipartite graph, the material nodes, the order nodes and the edges are normalized and linearly mapped to obtain initial material encoding vectors, initial order encoding vectors and initial edge encoding vectors;

[0041] The initial order encoding vectors are updated by using the initial material encoding vectors and the initial edge encoding vectors through a graph isomorphism neural network with residual connection to obtain the order encoding vectors;

[0042] The initial material encoding vectors are updated by using the order encoding vectors and the initial edge encoding vectors through a graph isomorphism neural network with residual connection to obtain the material encoding vectors.

[0043] In this embodiment, considering the feature heterogeneity of the bipartite graph, the original material nodes, order nodes and edges are first normalized and linearly mapped to 128-dimensional initial material encoding vectors initial order encoding vectors and initial edge encoding vectors Then, based on the topology structure of the input bipartite graph, two-stage information passing is performed to aggregate the information of the material nodes and the order nodes to update the initial encoding vectors (including the initial material encoding vectors, the initial order encoding vectors and the initial edge encoding vectors), to obtain the updated order encoding vectors and the material encoding vectors. In this embodiment, the two-stage update can be performed using a graph isomorphism network (GIN) with residual connection. The GNN can process graph structure data and learn relationship information through message passing between nodes on the graph.

[0044] Specifically, in the first stage, the initial material encoding vectors and the initial edge encoding vectors are used to update the initial order encoding vectors through a graph isomorphism network with residual connection, to obtain the order encoding vectors The initial order encoding vectors are updated according to the following formula:

[0045]

[0046] wherein MLP K is a multi-layer perceptron for updating the order encoding vectors, ∈ is a pre-set hyperparameter; N(k) is a set of one-hop neighbor nodes of the kth order node (i.e., a set composed of all material nodes connected to the kth order node); is the initial material encoding vector of the mth material node, and m belongs to the set of one-hop neighbor nodes of the kth order node. For example, for the kth order node, there are three selectable materials (material 1, material 2 and material 3), i.e., the kth order node is connected to material node 1, material node 2 and material node 3 through edges, and then material node 1, material node 2 and material node 3 constitute N(k). is the initial edge encoding vector corresponding to the edge from the mth material node to the kth order node. The order encoding vectors obtained after the update not only include the feature information of the kth order node (order production quantity information), but also integrate the feature information of the material nodes connected to the kth order node (material switching cost information) and the feature information of the edges connected to the kth order node (consumption quantity information of the connected order to the material).

[0047] In the second stage, the order encoding vectors and the initial edge encoding vectors are used to update the initial material encoding vectors through a graph isomorphism network with residual connection, to obtain the material encoding vectors The initial material coding vector is updated according to the following formula

[0048]

[0049] wherein MLP is a multilayer perceptron for updating the material coding vector, ∈ is a preset hyperparameter, and N(m) is a one-hop neighbor node set of the mth material node (i.e., a set composed of all order nodes connected to the mth material node). M is an initial material coding vector of the mth material node, is an order coding vector of the kth order node, and k belongs to the one-hop neighbor node set N(m) of the mth material node. For example, for the mth material node, three orders (order 1, order 2, and order 3) can be produced, i.e., the mth material node is connected to order node 1, order node 2, and order node 3 through edges, and order node 1, order node 2, and order node 3 constitute N(m). is an initial edge coding vector corresponding to the edge from the kth order node to the mth material node. The updated material coding vector not only includes the feature information (material switching cost information and order production quantity information) of the mth material node, but also integrates the feature information (order production quantity information) of the order nodes connected to the mth material node and the feature information (connection order consumption quantity information) of the connected edges.

[0050] In this embodiment, after obtaining the order coding vector and the material coding vector through the input coding module, a material encoder can be used to perform deep feature extraction on the material coding vector to generate a deep material coding vector. In a possible implementation, the material encoder includes N attention coding layers; each attention coding layer includes two sub-layers: a multi-head attention update layer and a feedforward neural network.

[0051] The deep feature extraction on the material coding vector to generate a deep material coding vector includes:

[0052] In the multi-head attention update layer, a set of query vectors, key vectors, and value vectors are generated for each material coding vector.

[0053] According to the query vectors, the key vectors, and the value vectors, a target multi-head attention function is calculated.

[0054] ​​Step S103, for each of the material encoding vectors, based on the target multi-head attention function, aggregate information of all the material encoding vectors to obtain an aggregated material encoding vector.

[0055] Step S104, after batch normalization processing of the aggregated material encoding vector, obtain an updated material variable vector through the feedforward neural network.

[0056] Step S105, input the updated material encoding vector into the next attention encoding layer until the output result of the last attention encoding layer is obtained, and determine the output result of the last attention encoding layer as the deep material encoding vector.

[0057] In the embodiment, the material encoder performs deep processing on the material encoding vector obtained by the input encoding module through N attention encoding layers. Specifically, each attention encoding layer includes two sub-layers: a multi-head attention update layer (MHA) for performing information aggregation and a feedforward neural network (MLP) for updating, wherein each sub-layer adds a residual connection and batch normalization (BN). The processing process is shown in the following formula:

[0058]

[0059] wherein, is the input of the mth material encoding vector in the lth attention encoding layer (i.e., the output of the (l-1)th attention encoding layer); is the output of the mth material encoding vector in the lth attention encoding layer (i.e., the input of the (l+1)th attention encoding layer). is the aggregated material encoding vector output by the multi-head attention update layer (MHA) in the lth attention encoding layer of the mth material encoding vector; wherein |M| represents the total number of material encoding vectors obtained by the input encoding module.

[0060] Specifically, the input of the lth attention encoding layer is First, perform step S101 to generate a set of query vectors, key vectors and value vectors through the multi-head attention update layer, as shown in the following formula:

[0061]

[0062] is the query vector generated by the mth material encoding vector in the lth attention encoding layer, is the key vector generated by the mth material encoding vector in the lth attention encoding layer, is the value vector generated by the mth material encoding vector in the lth attention encoding layer. The learnable parameters are obtained in a training process (such as the training process of steps S301-S308 hereinafter).

[0063] After the query vector, the key vector, and the value vector are generated, step S102 is performed to calculate the target multi-head attention function. The multi-head attention function is used to collect information from other material nodes and can be regarded as a query vector The weight ratio is weighted. The calculation formula is as follows:

[0064]

[0065] wherein T represents the transpose of a vector, dim k dim represents the dimension of the key vector, represents the attention score of the mth material encoding vector to the nth material encoding vector in the lth attention encoding layer, represents the target multi-head attention function, represents the value vector of the nth material encoding vector in the lth attention encoding layer.

[0066] Then, step S103 is performed to aggregate the information of the remaining material encoding vectors based on the target multi-head attention function for each of the material encoding vectors, that is, to realize weighted summation of the value vectors of the |M| material encoding vectors in the lth attention encoding layer according to the target attention function (that is, to aggregate the information of all material encoding vectors based on the target multi-head attention function), to obtain the aggregated material encoding vector of the mth material encoding vector Specifically, as shown in the following formula:

[0067]

[0068] wherein represents [,...] represents a connection operator, and H represents the number of attention heads.

[0069] Step S104 is performed to further update the aggregated material encoding vector through batch normalization (as shown in formula 3) and then through a feedforward neural network (as shown in formula 4) to obtain an updated material variable vector. Finally, step S105 is performed to input the updated material encoding vector into the next attention encoding layer until the output result of the last attention encoding layer is obtained, and the output result of the last attention encoding layer is determined as the deep material encoding vector

[0070] In this embodiment, after the deep material encoding vector of each material encoding vector is obtained, a decoder is used to obtain the order encoding vector The deep material encoding vector and current production scheduling information, output a production scheduling plan. Wherein, the current production scheduling information represents the order production situation at the current time, including the execution progress information of the orders being executed, etc. The decoder outputs the distribution of the step action and samples to obtain the decision action, i.e. the next used material, based on the material node encoding and order node encoding obtained by the input encoding module and the material encoder, using the information of the current incomplete scheduling scheme (i.e. the current production scheduling information). Then, according to the decision, the current production scheduling information is updated, and based on the updated information, the distribution probability of the next step action is output again, and the decision (the next used material) is obtained, until the use order of all materials is determined, and the final production scheduling plan is generated.

[0071] In a possible implementation, the outputting of the production scheduling plan according to the order encoding vector, the deep material encoding vector and the current production scheduling information comprises:

[0072] In step S201, a decoding vector h is generated according to the order encoding vector, the deep material encoding vector and the current production scheduling information d ; the decoding vector is spliced from a global encoding vector, an order completion condition encoding vector and a material sequence encoding vector.

[0073] In a possible implementation, the global encoding vector is an encoding vector obtained by average pooling all the deep material encoding vectors;

[0074] The order completion condition encoding vector is an encoding vector obtained by multiplying the order encoding vector by the remaining production amount and then performing weighted average;

[0075] The material sequence encoding vector is an encoding vector obtained by passing the current production scheduling information through a long short-term memory network.

[0076] In this embodiment, the decoder first calculates a decoding vector h d which is spliced from three parts: a global encoding vector, an order completion condition encoding vector and a material sequence encoding vector. Specifically, the global encoding vector is the average pooling of all the deep material encoding vectors output by the material encoder, as shown in the following formula:

[0077]

[0078] Wherein, h m represents the global encoding vector, and |M| represents the total number of material encoding vectors obtained by the input encoding module; represents the m-th deep material encoding vector. The global encoding vector contains the aggregation of the information extracted from the bipartite graph, i.e., contains global information. By concatenating the global encoding vector in the decoding vector, the decoder can analyze and calculate based on the global information to obtain more accurate decision results.

[0079] order completion condition encoding vector i.e., the order encoding vector obtained by the input encoding module multiplied by the weighted average of the remaining production amount, as shown in the following formula:

[0080]

[0081] wherein, represents the order completion condition encoding vector obtained in the t-th round of calculation, and |K| represents the total number of order encoding vectors obtained by the input encoding module. represents the remaining production amount corresponding to the k-th order node in the t-th round of calculation. The remaining production amount represents the total amount of products that have not yet determined the production of the material used by the order.

[0082] material order encoding vector, i.e., the encoding vector obtained by the long short-term memory (LSTM) network from the currently selected material (current production scheduling information), represents the sequence composed of the determined material nodes. As shown in the following formula:

[0083]

[0084] wherein, represents the material order encoding vector obtained in the t-th round of calculation, represents the deep encoding vector corresponding to the material used in the t-th round, i.e., In this case, m represents the material v t .

[0085] In step S202, the probability distribution of the current execution action is output according to the decoding vector.

[0086] The decoder generates a decision (i.e., the probability distribution of the current execution action) based on the decoding vector. Specifically, the probability distribution represents the probability distribution of the material to be selected for use at the current time. For example, the probability of using the material corresponding to material node A for production, the probability of using the material corresponding to material node B for production, and the probability of using the material corresponding to material node C for production. According to the order of material use, the corresponding order production sequence can be further determined.

[0087] In a possible implementation, the step S202 of outputting, according to the decoding vector, a probability distribution of a current execution action comprises:

[0088] The step S2021 aggregates the decoding vector h d through a one-to-many attention layer. The step S2022 updates the decoding vector h

[0089] In the embodiment, for the |M| deep material encoding vectors The step S2021 aggregates the decoding vector h d through a one-to-many attention layer. A set of deep query vectors, deep key vectors and deep value vectors are generated. Then, a deep multi-head attention function is calculated according to the deep query vectors, the deep key vectors and the deep value vectors. For the decoding vector h d , information of all the deep material encoding vectors is aggregated based on the deep multi-head attention function to obtain an aggregated decoding vector. Finally, the aggregated decoding vector is processed by batch normalization, and then an updated decoding vector h′ d is obtained through a feedforward neural network.

[0090] The step S2022 makes the updated decoding vector pass through a one-to-many attention function to obtain an attention value of each material node.

[0091] Specifically, for the updated decoding vector h′ d , a set of query vectors q′ d , key vectors and value vectors v′ d are generated.

[0092] q′ d = w′ q h′ d , v′ d = W′ v h′ d ; where W′ q , W′ k and W′ v represent learnable parameters, which are obtained in a training process (such as the training process of steps S301-S308 below).

[0093] The attention value of each material node is calculated according to the following formula:

[0094]

[0095] Where, q′ d For the updated decoded vector h′ d The query vector, For the updated decoded vector h′ d The key vector, dim k Given the dimension of the key vector, the fraction of the masking action is set to -∞; u dm This represents the attention value of the m-th material node.

[0096] Step S2023: After the attention value is masked and normalized exponentially, the probability distribution of the currently executed action is obtained.

[0097] Specifically, a softmax function is used to calculate the probability distribution of the currently executed action. This probability distribution represents the updated decoded vector h′. d With each deep material coding vector The single-head attention weights between them. The probability distribution of the currently executed action is calculated according to the following formula:

[0098]

[0099] Where m = 0, ..., |M|. dm Let a represent the attention value of the m-th material node. t =m indicates that the action performed in round t is to select the m-th material node as the material to be used in this round, s t This represents the state in round t.

[0100] Step S203: The action with the highest probability in the probability distribution is determined as the current action; the current action represents the material currently selected for use.

[0101] Step S204: Update the current production schedule information according to the currently executed action.

[0102] After determining the current action to be performed (the currently selected material), the current production schedule information needs to be updated accordingly (by re-inputting the LSTM model to generate a sequence of already determined material nodes). Correspondingly, in this embodiment, due to the order completion status encoding vector... This represents the order code vector obtained from the input encoding module. Multiplying by a weighted average of the remaining production capacity, and after determining the current action to be performed, the corresponding orders that need to be produced can be determined, and then the remaining production capacity for each order can be calculated. Update the data to obtain the remaining production quantity of the k-th order node for the next round of calculation.

[0103] Step S205, repeat the above steps to regenerate a new decoding vector to determine the next round of execution actions until all orders have been determined production order, output the production scheduling plan.

[0104] After completing the tth round of calculation, determining the current execution action, and updating the corresponding information (current production scheduling information and remaining production volume), re-executing steps S201-S204 to perform the next round of calculation to determine the next execution action (the material used in the next round) and further determine the order of production determined in the next round. Until a complete scheduling plan is obtained (all orders have been determined production order and material used), it is determined that all orders have been arranged for production using materials, and the corresponding production scheduling plan is output. The production scheduling plan is the order of material use and the production order of all orders arranged in the order of calculation rounds.

[0105] The embodiment of the present application aims at a material management and production scheduling system, and designs an intelligent production scheduling system considering alternative materials and switching costs. The present application designs an intelligent production scheduling system based on graph neural networks and multi-head attention mechanisms, takes the bipartite graph representation of problem instances and the current solving state as input, sequentially selects the materials to be accessed, and iteratively constructs material selection decisions and scheduling results. The learning goal of the network is to minimize the sum of the material use cost and switching cost of the output scheduling plan. Based on the available alternative materials and material switching costs, the present application makes decisions on the production order of orders and the selection of alternative materials (calculates the selection of materials and the order of orders), so that the production demand is met and the material use and switching costs are minimized, thereby better reducing production costs and optimizing production efficiency.

[0106] In addition, the present application proposes a network update module of Monte Carlo policy gradient reinforcement learning with a greedy baseline, which updates the related modules (input encoding module, material encoder and decoder) in the system during the training phase. Specifically, in one possible implementation, the system further comprises a network update module; the network update module is configured to train the input encoding module, the material encoder and the decoder according to the following steps:

[0107] Step S301, obtaining a training data set, each training sample data in the training data set being a sample bipartite graph, and the corresponding label being an optimal production scheduling plan.

[0108] Step S302, for each training sample data, the decoder randomly selects a first current execution action from the probability distribution of the output current execution action to generate a first sample production scheduling plan of the training sample data.

[0109] Step S303, for each of the training sample data, the decoder selects the action with the maximum probability as the second current action after outputting the probability distribution of the current action, and generates a second sample production scheduling plan of the training sample data.

[0110] Step S304, according to the first sample production scheduling plan and the label, a first loss function value L(π i ) is calculated. i

[0111] Step S305, according to the second sample production scheduling plan and the label, a second loss function value b(U i ) is calculated.

[0112] Step S306, according to the first loss function value and the second loss function value, a gradient is calculated.

[0113] In one possible implementation, the calculating the gradient according to the first loss function value and the second loss function value comprises:

[0114] According to the first loss function value and the second loss function value, the gradient is calculated according to the following formula:

[0115]

[0116] Wherein, B represents the total number of the training sample data, U i is the i-th training sample data, L(π i | U i ) is the first loss function value, b(U i ) is the second loss function value; π i is the first sample production scheduling plan; p θ (π i | U i ) represents the probability that the optimal scheduling plan of the i-th training sample data U i is the first sample production scheduling plan π i .

[0117] Step S307, according to the gradient, the input encoding module, the material encoder and the decoder are updated.

[0118] Step S308, the training sample data is reselected, and the above steps are executed until a preset training number is reached, and the training is ended.

[0119] ​In a possible implementation, the system further comprises a scheduling scheme post-processing module, which is configured to correct and post-optimize the production scheduling plan according to preset preference rules to obtain a final production scheduling plan.

[0120] Specifically, referring to Figure 3 , Figure 3 An execution flow diagram of an intelligent production scheduling system is shown in FIG. 1, which comprises the following steps: Figure 3As shown, it includes two stages of offline learning and online application, involving an intelligent scheduling module (including: input encoding module, material encoder and decoder), network updating module and scheduling scheme post-processing module. After the user data input is completed, the intelligent scheduling module automatically calls the reinforcement learning training algorithm (corresponding to steps S301-S308), first represents the scheduling problem input by the user as a bipartite graph, all orders and all alternative materials considered in the example are respectively represented as two types of nodes of the bipartite graph, and if the order can use the material for production, there is an edge connection between the corresponding material node and order node. Among them, the switching cost information of the material is represented as the characteristics of the material node, the production information of the order is represented as the node characteristics of the order node, and the demand of the material in the order is represented as the edge characteristics. The bipartite graph representation is input to the input encoding module, the input encoding module performs preliminary encoding and updating on the original order nodes and material nodes in the bipartite graph, and obtains order encoding vectors and material encoding vectors, and then the material encoder performs deep encoding on the updated material encoding vectors. The decoder uses the encoding generated by the input encoding module and the material encoder, and the information of the current incomplete scheduling scheme (i.e. the current production scheduling information) to decide the action of this step, that is, the next used material. The decoder iteratively executes until all materials have been determined to use the order, and a complete scheduling scheme is obtained. During offline training, the network updating module updates the related modules in the system based on the loss value of the scheduling scheme constructed by the intelligent scheduling module based on the Monte Carlo policy gradient reinforcement learning algorithm with a greedy baseline. After the training is completed, the scheduling scheme post-processing module corrects and optimizes the scheduling scheme constructed by the strategy network based on the rules based on the user's preset preferences, further improves the quality of the scheduling scheme, and outputs the final production scheduling plan. On the one hand, based on deep reinforcement learning, a high-quality scheduling strategy can be automatically learned through exploration in a large number of scheduling problem instances, without relying on a large amount of domain knowledge or a pre-solved optimal scheduling scheme; on the other hand, after training is completed using the network updating module, the intelligent production scheduling system can be used for scheduling end to end, wherein the scheduling problem is represented by a bipartite graph, and only one encoding is required during processing. The scheduling scheme is iteratively output by the encoder. This way is fast, and the output solution is of higher quality than traditional human-designed heuristic algorithms; in addition, the intelligent production scheduling system can be used to process scheduling problem instances and distributions that do not appear in training after training, and has good generalization and scalability.

[0121] The second aspect of the embodiment of the application proposes an intelligent production scheduling method considering alternative materials and switching costs, applied to the intelligent production scheduling system of the first aspect of the embodiment of the application, referring to Figure 4 , Figure 4 shows a step flow diagram of an intelligent production scheduling method, as Figure 4As shown, the method comprises:

[0122] In step S401, a bipartite graph is generated according to the order information input by the user and the alternative material information, the bipartite graph comprising material nodes and order nodes, the feature information of the material nodes being the material switching cost information of the corresponding materials, the feature information of the order nodes being the order production quantity information of the corresponding orders, there being an edge connection between the order nodes and the material nodes selectable for the order, the feature information of the edge being the consumption quantity information of the connected order to material.

[0123] In step S402, an order encoding vector and a material encoding vector are generated by using an input encoding module according to the bipartite graph.

[0124] In step S403, a deep material encoding vector is generated by using a material encoder to perform deep feature extraction on the material encoding vector.

[0125] In step S404, a production scheduling plan is output by using a decoder according to the order encoding vector, the deep material encoding vector and the current production scheduling information.

[0126] In a possible implementation, the outputting of the production scheduling plan according to the order encoding vector, the deep material encoding vector and the current production scheduling information comprises:

[0127] generating a decoding vector according to the order encoding vector, the deep material encoding vector and the current production scheduling information, the decoding vector being spliced from a global encoding vector, an order completion condition encoding vector and a material sequence encoding vector;

[0128] outputting a probability distribution of a current execution action according to the decoding vector;

[0129] determining an execution action with the maximum probability in the probability distribution as the current execution action, the current execution action representing a currently selected material to be used;

[0130] updating the current production scheduling information according to the current execution action;

[0131] repeating the above steps to regenerate a new decoding vector to determine a next round of execution action until the production sequence of all orders is determined, and outputting the production scheduling plan.

[0132] In a possible implementation, the global encoding vector is an encoding vector obtained by average pooling of all the deep material encoding vectors;

[0133] the order completion condition encoding vector is an encoding vector obtained by weighted averaging of the order encoding vector multiplied by the remaining production quantity.

[0134] The material order encoding vector is an encoding vector obtained by passing the current production scheduling information through a long short-term memory network.

[0135] In a possible implementation, the outputting of the probability distribution of the current execution action according to the decoding vector comprises:

[0136] The decoding vector is passed through a one-to-many attention layer to aggregate and update deep material encoding vectors, to obtain an updated decoding vector.

[0137] The updated decoding vector is passed through a one-to-many attention function to obtain an attention value of each material node.

[0138] The attention value is passed through a mask and a normalized exponential function to obtain the probability distribution of the current execution action.

[0139] In a possible implementation, the generating of the order encoding vector and the material encoding vector according to the bipartite graph comprises:

[0140] According to the bipartite graph, the material nodes, the order nodes, and the edges are normalized and linearly mapped to obtain initial material encoding vectors, initial order encoding vectors, and initial edge encoding vectors.

[0141] The initial order encoding vectors are updated by using the initial material encoding vectors and the initial edge encoding vectors through a graph isomorphism neural network with a residual connection to obtain the order encoding vectors.

[0142] The initial material encoding vectors are updated by using the order encoding vectors and the initial edge encoding vectors through a graph isomorphism neural network with a residual connection to obtain the material encoding vectors.

[0143] In a possible implementation, the material encoder comprises N attention encoding layers; each of the attention encoding layers comprises two sub-layers: a multi-head attention update layer and a feedforward neural network.

[0144] The deep feature extraction on the material encoding vectors to generate deep material encoding vectors comprises:

[0145] For each of the material encoding vectors, a group of query vectors, key vectors, and value vectors are generated through the multi-head attention update layer.

[0146] A target multi-head attention function is calculated according to the query vectors, the key vectors, and the value vectors.

[0147] For each of the material encoding vectors, based on the target multi-head attention function, information of all the material encoding vectors is aggregated to obtain an aggregated material encoding vector;

[0148] After batch normalization processing is performed on the aggregated material encoding vector, the updated material variable vector is obtained through the feedforward neural network.

[0149] The updated material encoding vector is input into a next attention encoding layer until an output result of a last attention encoding layer is obtained, and the output result of the last attention encoding layer is determined as the deep material encoding vector.

[0150] In a possible implementation, the method further includes:

[0151] The input encoding module, the material encoder, and the decoder are trained according to the following steps:

[0152] A training data set is obtained, each training sample data in the training data set being a sample bipartite graph, and a corresponding label being an optimal production scheduling plan;

[0153] For each of the training sample data, after the decoder outputs a probability distribution of a current execution action, a first current execution action is randomly selected therefrom to generate a first sample production scheduling plan of the training sample data;

[0154] For each of the training sample data, after the decoder outputs a probability distribution of a current execution action, a current execution action with a maximum probability is selected therefrom as a second current execution action to generate a second sample production scheduling plan of the training sample data;

[0155] A first loss function value is calculated according to the first sample production scheduling plan and the label;

[0156] A second loss function value is calculated according to the second sample production scheduling plan and the label;

[0157] A gradient is calculated according to the first loss function value and the second loss function value;

[0158] The input encoding module, the material encoder, and the decoder are updated according to the gradient;

[0159] The training sample data is reselected, and the above steps are performed until a preset training number of times is reached, and the training is ended.

[0160] In a possible implementation, the gradient is calculated according to the first loss function value and the second loss function value, including:

[0161] The gradient is calculated based on the first loss function value and the second loss function value according to the following formula:

[0162] Where B represents the total number of training sample data, U i For the i-th training sample data, L(π) i |U i ) is the value of the first loss function, b(U) i ) represents the value of the second loss function; π i For the production scheduling plan of the first sample; p θ (π i |U i ) represents the i-th training sample data U i The optimal scheduling plan is the production scheduling plan π of the first sample. i The probability of.

[0163] In one possible implementation, the method further includes:

[0164] Based on pre-set preference rules, the production scheduling plan is corrected and optimized to obtain the final production scheduling plan.

[0165] This application also provides an electronic device, see embodiments thereof. Figure 5 , Figure 5 This is a schematic diagram of the electronic device proposed in an embodiment of this application. Figure 5 As shown, the electronic device 100 includes a memory 110 and a processor 120. The memory 110 and the processor 120 are connected via a bus for communication. The memory 110 stores a computer program that can run on the processor 120 to implement the steps in the intelligent production scheduling method that takes into account alternative materials and switching costs as disclosed in the second aspect of the embodiments of this application.

[0166] This application also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps in the intelligent production scheduling method that takes into account alternative materials and switching costs as disclosed in the second aspect of this application.

[0167] This application also provides a computer program product that, when executed on an electronic device, causes a processor to implement the steps of the intelligent production scheduling method that considers alternative materials and switching costs as disclosed in the second aspect of this application.

[0168] The various embodiments in the specification are described in progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts between the embodiments can be mutually referred to.

[0169] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, devices, electronic devices and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing terminal equipment to produce a machine, so that the instructions executed by the computer or other programmable data processing terminal equipment produce a device for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that implements the functions specified in one or more flows and / or blocks.

[0170] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing terminal equipment to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that implements the functions specified in one or more flows and / or blocks.

[0171] These computer program instructions can also be loaded into a computer or other programmable data processing terminal equipment, so that a series of operation steps are performed on the computer or other programmable terminal equipment to produce a computer-implemented process, so that the instructions executed on the computer or other programmable terminal equipment provide a process for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that implements the functions specified in one or more flows and / or blocks.

[0172] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications to the embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present application.

[0173] Finally, it is to be understood that the phraseology or terminology such as "first" and "second" etc. used herein is merely intended to distinguish one entity or operation from another without necessarily requiring or implying any actual such relationship or order between such entities or operations. Moreover, the use of the term "including", "containing" or any other variant thereof is intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or even inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the stated elements.

[0174] The above has carried on the detailed introduction to the intelligent production scheduling system and method considering alternative materials and switching cost provided by the present application, the principle and implementation mode of the present application are described in the text by applying specific examples, the above example explanation is only for helping to understand the method of the present application and its core idea; simultaneously, for the general technical personnel of the field, according to the idea of the present application, there will be changes in the specific implementation mode and application range, and the above is described, the content of the specification should not be understood as the limitation of the present application.

Claims

1. An intelligent production scheduling system considering alternative materials and switching costs, characterized by, The system comprises: A bipartite graph generation module generates a bipartite graph according to user-input order information and alternative material information, the bipartite graph comprising material nodes and order nodes, the feature information of the material nodes being material switching cost information of corresponding materials, the feature information of the order nodes being order production quantity information of corresponding orders, and an edge connection existing between the order nodes and the material nodes selectable for the orders, the feature information of the edge being consumption quantity information of the connected orders on materials; An input encoding module generates order encoding vectors and material encoding vectors according to the bipartite graph; A material encoder performs deep feature extraction on the material encoding vectors to generate deep-layer material encoding vectors; A decoder outputs a production scheduling plan according to the order encoding vectors, the deep-layer material encoding vectors, and current production scheduling information; The output of the production scheduling plan according to the order encoding vectors, the deep-layer material encoding vectors, and the current production scheduling information comprises: Generating a decoding vector according to the order encoding vectors, the deep-layer material encoding vectors, and the current production scheduling information; the decoding vector being spliced from a global encoding vector, an order completion condition encoding vector, and a material sequence encoding vector; Outputting a probability distribution of a current execution action according to the decoding vector; Determining an execution action with the maximum probability in the probability distribution as the current execution action; the current execution action representing a currently selected material to be used; Updating the current production scheduling information according to the current execution action; Repeating the above steps to regenerate a new decoding vector to determine a next round of execution action until production sequences of all orders are determined, and outputting the production scheduling plan; The system further comprises a network updating module; the network updating module is configured to train the input encoding module, the material encoder, and the decoder; The system further comprises a scheduling scheme post-processing module; the scheduling scheme post-processing module is configured to correct and optimize the production scheduling plan according to pre-set preference rules to obtain a final production scheduling plan.

2. The intelligent production scheduling system considering alternative materials and switching costs according to claim 1, wherein, The global encoding vector is an encoding vector obtained by average pooling all the deep-layer material encoding vectors; The order completion condition encoding vector is an encoding vector obtained by multiplying the order encoding vector by the remaining production quantity and then performing weighted average; The material sequence encoding vector is an encoding vector obtained by passing the current production scheduling information through a long short-term memory network.

3. The intelligent production scheduling system considering alternative materials and switching costs according to claim 1, wherein, The output of the probability distribution of the current execution action according to the decoding vector comprises: Making the decoding vector pass through a one-to-many attention layer to aggregate deep-layer material encoding vectors, update the decoding vector, and obtain an updated decoding vector; Making the updated decoding vector pass through a one-to-many attention function to obtain attention values of the material nodes; Making the attention values pass through a mask and a normalization exponential function to obtain the probability distribution of the current execution action.

4. The intelligent production scheduling system considering alternative materials and switching costs according to claim 1, wherein, The generation of the order encoding vectors and the material encoding vectors according to the bipartite graph comprises: According to the bipartite graph, the material nodes, the order nodes and the edges are normalized and linearly mapped to obtain an initial material encoding vector, an initial order encoding vector and an initial edge encoding vector; The initial order encoding vector is updated by a graph isomorphism neural network with a residual connection using the initial material encoding vector and the initial edge encoding vector to obtain the order encoding vector; The initial material encoding vector is updated by a graph isomorphism neural network with a residual connection using the order encoding vector and the initial edge encoding vector to obtain the material encoding vector.

5. The intelligent production scheduling system considering alternative materials and switching costs according to claim 1, wherein, The material encoder comprises N attention encoding layers; each of the attention encoding layers comprises two sub-layers: a multi-head attention update layer and a feedforward neural network; The deep feature extraction of the material encoding vector to generate a deep material encoding vector comprises: For each of the material encoding vectors, a set of query vectors, key vectors and value vectors are generated in the multi-head attention update layer; According to the query vectors, the key vectors and the value vectors, a target multi-head attention function is calculated; For each of the material encoding vectors, the information of all material encoding vectors is aggregated based on the target multi-head attention function to obtain an aggregated material encoding vector; After batch normalization processing of the aggregated material encoding vector, the updated material encoding vector is obtained through the feedforward neural network; The updated material encoding vector is input into the next attention encoding layer until the output result of the last attention encoding layer is obtained, and the output result of the last attention encoding layer is determined as the deep material encoding vector.

6. The intelligent production scheduling system considering alternative materials and switching costs according to claim 1, wherein, The network updating module trains the input encoding module, the material encoder and the decoder according to the following steps: A training data set is obtained, and each training sample data in the training data set is a sample bipartite graph, and the corresponding label is an optimal production scheduling plan; For each of the training sample data, the decoder randomly selects a first current execution action from the probability distribution of the current execution action to generate a first sample production scheduling plan of the training sample data; For each of the training sample data, the decoder selects the execution action with the maximum probability from the probability distribution of the current execution action as a second current execution action to generate a second sample production scheduling plan of the training sample data; A first loss function value is calculated according to the first sample production scheduling plan and the label; A second loss function value is calculated according to the second sample production scheduling plan and the label; Gradients are calculated according to the first loss function value and the second loss function value; The input encoding module, the material encoder and the decoder are updated according to the gradients; The training is ended by reselecting the training sample data and performing the above steps until a preset training number is reached.

7. The intelligent production scheduling system considering alternative materials and switching costs according to claim 6, wherein, The calculation of the gradients according to the first loss function value and the second loss function value comprises: The gradient is calculated according to the first loss function value and the second loss function value according to the following formula: ; wherein B represents a total number of the training sample data, is the i-th training sample data, is the first loss function value, is the second loss function value; is the first sample production scheduling plan, represents a probability that the optimal scheduling plan of the i-th training sample data is the first sample production scheduling plan .

8. An intelligent production scheduling method considering alternative materials and switching costs, characterized by, The method is applied to the intelligent production scheduling system of any one of claims 1-7, and the method comprises: According to the order information and the alternative material information input by the user, a bipartite graph is generated, the bipartite graph comprises material nodes and order nodes, the characteristic information of the material nodes is the material switching cost information of the corresponding materials, the characteristic information of the order nodes is the order production quantity information of the corresponding orders, there is an edge connection between the order nodes and the material nodes that can be selected by the order, and the characteristic information of the edge is the consumption quantity information of the connected order and material; An input encoding module is used to generate order encoding vectors and material encoding vectors according to the bipartite graph; A material encoder is used to perform deep feature extraction on the material encoding vectors to generate deep material encoding vectors; A decoder is used to output a production scheduling plan according to the order encoding vectors, the deep material encoding vectors and the current production scheduling information.

Images