Production scheduling system scheduling decision determination method and device and electronic equipment

Abstract

The application provides a scheduling decision determination method and device of a production scheduling system and electronic equipment, relates to the technical field of intelligent manufacturing and industrial scheduling, and comprises the following steps: collecting equipment state information, job state information and order state information of the production scheduling system in a working process; determining a multi-dimensional production scheduling state matrix and scheduling sequence information according to the equipment state information, the job state information and the order state information; determining a production scheduling state according to the multi-dimensional production scheduling state matrix, a first neural network, scheduling sequence information, and local structure features and global structure features determined by a second neural network; determining a production scheduling system scheduling decision according to the production scheduling state and a deep reinforcement learning scheduling decision algorithm; and the deep reinforcement learning scheduling decision algorithm is an algorithm for generating the production scheduling system scheduling decision based on the production scheduling state. The technical scheme of the application improves the accuracy of the production scheduling system scheduling decision.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and industrial scheduling technology, and in particular to a method, apparatus and electronic equipment for determining scheduling decisions in a production scheduling system. Background Technology

[0002] With the continuous development of intelligent manufacturing and industrial information technology, the application of production scheduling systems in the manufacturing industry has become increasingly widespread, and their importance is becoming increasingly prominent. A reasonable production scheduling strategy can effectively improve equipment utilization, shorten production cycles, and reduce production costs. This is of great significance for complex production scenarios such as semiconductor manufacturing and multi-variety, small-batch production. However, actual production scheduling systems are typically characterized by complex processes, resource constraints, and dynamic changes in the production environment, making scheduling problems complex and full of uncertainty.

[0003] Currently, in traditional production scheduling technologies, decisions are typically made based on rule-based or mathematical optimization methods, such as priority-based rule scheduling or optimization algorithms based on deterministic models. While these methods have some effectiveness in specific scenarios, they are highly dependent on system modeling and struggle to adapt to dynamic changes such as order fluctuations, equipment failures, and uncertain processing times, resulting in insufficient stability and reliability of scheduling decisions. Furthermore, in large-scale and complex production environments, they often suffer from high computational complexity and insufficient real-time performance.

[0004] Therefore, how to enhance the credibility, stability, and engineering availability of scheduling decisions remains a pressing technical problem that needs to be solved in the field of intelligent scheduling. Summary of the Invention

[0005] This invention provides a method, apparatus, and electronic device for determining scheduling decisions in a production scheduling system. It addresses the shortcomings of existing methods based on rules or mathematical optimization, such as priority rule scheduling or optimization algorithms based on deterministic models. While these methods may be effective in specific scenarios, they heavily rely on system modeling and struggle to adapt to dynamic changes such as order fluctuations, equipment failures, and uncertain processing times, leading to insufficient stability and reliability in scheduling decisions. The invention achieves this by using a neural network to extract features from scheduling sequence information and determine the production scheduling state during the complex operation of a production scheduling system. Then, a deep reinforcement learning scheduling decision algorithm and the production scheduling state are used to generate scheduling decisions for the production scheduling system. Based on these decisions, the system is dynamically scheduled, improving the reliability, stability, and engineering usability of scheduling decisions, as well as enhancing dynamic scheduling performance.

[0006] This invention provides a method for determining scheduling decisions in a production scheduling system, comprising the following steps.

[0007] Collect status information of the production scheduling system during operation; the status information includes equipment status information, operation status information and order status information; Determine the multidimensional production scheduling status matrix and scheduling sequence information based on equipment status information, operation status information, and order status information; The local and global structural features are determined based on the multidimensional production scheduling state matrix, the first neural network, the scheduling sequence information, and the second neural network. The first neural network is a network that extracts the local structural features of adjacent scheduling objects in the multidimensional production scheduling state matrix. The second neural network is a network that uses a self-attention mechanism to extract the global structural features of the dependencies between scheduling objects in the scheduling sequence information. Determine the production scheduling status based on local and global structural characteristics; The production scheduling system makes scheduling decisions based on the production scheduling status and a deep reinforcement learning scheduling decision algorithm. The production scheduling system makes decisions on instructing job releases, processing sequences, and resource allocations during the production scheduling process. The deep reinforcement learning scheduling decision algorithm is an algorithm that generates production scheduling system scheduling decisions based on the production scheduling status.

[0008] According to the present invention, a method for determining scheduling decisions in a production scheduling system determines a multi-dimensional production scheduling state matrix and scheduling sequence information based on equipment status information, job status information, and order status information, including: Construct a multi-dimensional production scheduling status matrix based on equipment status information, operation status information, and order status information; The multidimensional production scheduling state matrix is ​​serialized to obtain scheduling sequence information.

[0009] According to the present invention, a method for determining scheduling decisions in a production scheduling system determines local structural features and global structural features based on a multi-dimensional production scheduling state matrix, a first neural network, scheduling sequence information, and a second neural network, including: The multidimensional production scheduling state matrix is ​​input into the first neural network to obtain the local structural features output by the first neural network. The scheduling sequence information is input into the second neural network to obtain the global structural features output by the second neural network.

[0010] According to the scheduling decision determination method of the production scheduling system provided by the present invention, a multi-dimensional production scheduling state matrix is ​​input into a first neural network to obtain the local structural features output by the first neural network, including: The multidimensional production scheduling state matrix is ​​input into the feature extraction unit of the first neural network to obtain the two-dimensional features output by the feature extraction unit; Two-dimensional features are input into the convolution processing unit of the first neural network to obtain the convolutional structure features output by the convolution processing unit; The convolutional structure features are input into the feature aggregation unit of the first neural network to obtain the local structural features output by the feature aggregation unit.

[0011] According to the method for determining scheduling decisions in a production scheduling system provided by the present invention, scheduling sequence information is input into a second neural network to obtain the global structural features output by the second neural network, including: The scheduling sequence information is input into the feature embedding unit of the second neural network to obtain the high-dimensional feature information input to the feature embedding unit. The high-dimensional feature information is input into the position encoding unit of the second neural network to obtain the embedded feature information output by the position encoding unit; The embedded feature information is input into the self-attention modeling unit of the second neural network to obtain the global scheduling dependency output by the self-attention modeling unit; The global scheduling dependency is input into the feedforward fully connected unit of the second neural network to obtain the global structural features output by the feedforward fully connected unit.

[0012] According to the present invention, a method for determining scheduling decisions in a production scheduling system determines the production scheduling state based on local structural features and global structural features, including: The production scheduling status is obtained by fusing local and global structural features; the fusing process includes at least one of feature concatenation, feature mapping, and feature weighting.

[0013] According to the method for determining scheduling decisions in a production scheduling system provided by the present invention, after determining the scheduling decision of the production scheduling system based on the production scheduling state and a deep reinforcement learning scheduling decision algorithm, the method further includes: The scheduling interpretation result is determined based on the scheduling sequence information and the second neural network; Based on the scheduling interpretation results, the scheduling decision of the production scheduling system is adjusted, the result of the adjustment is determined, and based on the result of the adjustment, the process of determining the scheduling decision of the production scheduling system based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm continues.

[0014] According to the scheduling decision determination method of the production scheduling system provided by the present invention, the scheduling interpretation result includes the scheduling object-level attention interpretation result and the scheduling feature-level contribution interpretation result; the scheduling interpretation result is determined based on the scheduling sequence information and the second neural network, including: Obtain the self-attention weight matrix of the second neural network during the scheduling decision-making process based on the scheduling sequence information; The attention interpretation result at the scheduling object level is determined based on the self-attention weight matrix; The contribution of the scheduling sequence information is decomposed to obtain the interpretation results of the scheduling feature-level contribution.

[0015] The present invention also provides a scheduling decision determination device for a production scheduling system, comprising the following modules: The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a scheduling decision determination method as described in any of the above-described production scheduling systems.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a scheduling decision determination method as described above for any of the production scheduling systems.

[0017] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements a scheduling decision determination method for any of the production scheduling systems described above.

[0018] This invention provides a method, apparatus, and electronic device for determining scheduling decisions in a production scheduling system. The method involves collecting state information of the production scheduling system during its operation, including equipment state information, job state information, and order state information. A multi-dimensional production scheduling state matrix and scheduling sequence information are determined based on these information. Local and global structural features are determined based on the multi-dimensional production scheduling state matrix, a first neural network, the scheduling sequence information, and a second neural network. The first neural network extracts the local structural features of adjacent scheduling objects in the multi-dimensional production scheduling state matrix. The second neural network uses a self-attention mechanism to extract global structural features of the dependencies between scheduling objects in the scheduling sequence information. The production scheduling state is determined based on the local and global structural features. Finally, the production scheduling system's scheduling decision is determined based on the production scheduling state and a deep reinforcement learning scheduling decision algorithm. The production scheduling system's scheduling decision involves instructing job releases, processing sequences, and resource allocation during the production scheduling system's operation. The deep reinforcement learning scheduling decision algorithm is an algorithm that generates production scheduling system scheduling decisions based on the production scheduling state. The technical solution of this invention addresses the shortcomings of existing methods based on rules or mathematical optimization, such as priority rule scheduling or optimization algorithms based on deterministic models. While these methods may have some effectiveness in specific scenarios, they are highly dependent on system modeling and struggle to adapt to dynamic changes such as order fluctuations, equipment failures, and uncertain processing times, resulting in insufficient stability and reliability of scheduling decisions. The invention achieves this by using a neural network to extract features from scheduling sequence information and determine the production scheduling state during the operation of a complex production scheduling system. Then, a deep reinforcement learning scheduling decision algorithm and the production scheduling state are used to generate a production scheduling system decision. Based on this decision, the production scheduling system is dynamically scheduled, improving the reliability, stability, and engineering usability of scheduling decisions, as well as enhancing dynamic scheduling performance. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is one of the flowcharts illustrating the scheduling decision determination method for the production scheduling system provided by the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of the first neural network provided by the present invention.

[0022] Figure 3 This is a schematic diagram of the operation process of the second neural network provided by the present invention.

[0023] Figure 4 This is a schematic diagram of the workflow of the deep reinforcement learning scheduling decision algorithm provided by the present invention.

[0024] Figure 5 This is a schematic diagram of the workflow for multimodal interpretable analysis provided by the present invention.

[0025] Figure 6 This is the second flowchart illustrating the scheduling decision determination method for the production scheduling system provided by the present invention.

[0026] Figure 7 This is a schematic diagram of the scheduling decision determination device of the production scheduling system provided by the present invention.

[0027] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] The following is combined with Figure 1 The present invention describes a method for determining scheduling decisions in a production scheduling system. This method is applicable to dynamic scheduling scenarios based on interpretable feedback and deep reinforcement learning. The execution subject of this method can be an electronic device or a scheduling decision determination device for a production scheduling system installed in the electronic device. This scheduling decision determination device for a production scheduling system can be implemented through software, hardware, or a combination of both. Figure 1 This is one of the flowcharts illustrating the scheduling decision determination method for the production scheduling system provided by the present invention, such as... Figure 1 As shown, the method includes the following steps 101, 102, 103, 104 and 105.

[0030] Step 101: Collect the status information of the production scheduling system during the operation process; the status information includes equipment status information, operation status information and order status information.

[0031] In this step, the equipment status information includes the current working status of the production equipment in the production scheduling system, the remaining processing time, the load status of the production equipment, the availability identifier of the production equipment, and the process information of the production equipment during its operation. This embodiment does not limit this.

[0032] The job status information includes the current processing step, the number of completed steps, the number of steps to be processed, the current waiting time, and the historical cumulative waiting time, etc. This embodiment does not limit these.

[0033] Order status information includes order type, order delivery date, remaining delivery time, delivery slack, and order priority information, etc. This embodiment does not limit these.

[0034] Status information may also include in-process status information, etc., but this embodiment does not limit this.

[0035] Specifically, when collecting status information of the production scheduling system during operation, the equipment status information can be collected in real time by the Manufacturing Execution System (MES), the equipment control system, or a pre-determined production database. The operation status information, order status information, and work-in-process status information can be obtained from the historical production records of the production scheduling system. This embodiment does not limit this.

[0036] In one specific embodiment, the process of collecting status information of the production scheduling system during its operation can be carried out through real-time collection, periodic collection, or time-triggered collection. This embodiment does not limit the method.

[0037] Step 102: Determine the multi-dimensional production scheduling status matrix and scheduling sequence information based on equipment status information, operation status information, and order status information.

[0038] In one specific embodiment, after obtaining the status information, the status information is uniformly abstracted into different entities, that is, the scheduling object is extracted, and the scheduling object, such as the device object, job object and order object, is used as the scheduling object dimension, and the status information of each device, job and order is mapped to the corresponding scheduling feature dimension.

[0039] The number of scheduling objects and scheduling feature dimensions is at least one, and can be flexibly configured according to the specific production scenario of the production scheduling system. This embodiment does not limit this.

[0040] For example, the scheduling object dimension is arranged according to a preset object order, which may include, for example, equipment objects, job objects, and order objects in sequence; the scheduling feature dimension may include, for example, remaining processing time feature, delivery slack feature, waiting time feature, queue length feature, and status identifier feature, etc., but this embodiment does not limit this.

[0041] In one specific embodiment, to facilitate subsequent processing by the first neural network and the second neural network, the device status information can be further normalized or standardized. For example, the remaining processing time, waiting time, and delivery date-related features can be mapped to a unified numerical range to eliminate the impact of differences in the dimensions of different features on scheduling decisions.

[0042] In one specific embodiment, determining a multidimensional production scheduling state matrix and scheduling sequence information based on equipment status information, job status information, and order status information includes: constructing a multidimensional production scheduling state matrix based on equipment status information, job status information, and order status information; and performing serialization processing on the multidimensional production scheduling state matrix to obtain scheduling sequence information.

[0043] Among them, the multi-dimensional production scheduling state matrix can comprehensively reflect the overall operating status of the production scheduling system at the current scheduling moment.

[0044] Specifically, when constructing the multidimensional production scheduling state matrix, the row direction of the multidimensional production scheduling state matrix corresponds to the scheduling object dimension, with each row representing a device object, job object, or order object; the column direction of the multidimensional production scheduling state matrix corresponds to the scheduling feature dimension, with each column representing a type of status information. This embodiment does not limit this aspect.

[0045] After constructing the multidimensional production scheduling state matrix, it can be further serialized to obtain scheduling sequence information.

[0046] Step 103: Determine the local structural features and global structural features based on the multidimensional production scheduling state matrix, the first neural network, the scheduling sequence information, and the second neural network; wherein, the first neural network is a network that extracts the local structural features of adjacent scheduling objects in the multidimensional production scheduling state matrix; the second neural network is a network that uses a self-attention mechanism to extract the global structural features of the dependencies between scheduling objects in the scheduling sequence information.

[0047] In this step, the first neural network may be, for example, a convolutional neural network (CNN), and the second neural network may be, for example, a Transformer network (a deep learning model architecture based on self-attention mechanism used to process sequential data). This embodiment does not limit this.

[0048] The first neural network employs a multi-layer convolutional neural network with a structure similar to that of the LeNet network (a classic convolutional neural network consisting of multiple layers, including convolutional layers, pooling layers, and fully connected layers) to extract local features from the scheduling sequence information after the multi-dimensional production scheduling state matrix has been serialized.

[0049] Specifically, local and global structural features are determined based on the multidimensional production scheduling state matrix, the first neural network, scheduling sequence information, and the second neural network. The first neural network is a network that extracts the local structural features of adjacent scheduling objects in the multidimensional production scheduling state matrix. The second neural network is a network that uses a self-attention mechanism to extract the global structural features of the dependencies between scheduling objects in the scheduling sequence information.

[0050] In one specific embodiment, determining local structural features and global structural features based on a multidimensional production scheduling state matrix, a first neural network, scheduling sequence information, and a second neural network includes: inputting the multidimensional production scheduling state matrix into the first neural network to obtain the local structural features output by the first neural network; and inputting the scheduling sequence information into the second neural network to obtain the global structural features output by the second neural network.

[0051] Specifically, the multidimensional production scheduling state matrix is ​​input into the first neural network to obtain the local structural features output by the first neural network; the scheduling sequence information is input into the second neural network to obtain the global structural features output by the second neural network.

[0052] In one specific embodiment, inputting a multidimensional production scheduling state matrix into a first neural network to obtain local structural features output by the first neural network includes: inputting the multidimensional production scheduling state matrix into a feature extraction unit of the first neural network to obtain two-dimensional features output by the feature extraction unit; inputting the two-dimensional features into a convolution processing unit of the first neural network to obtain convolutional structural features output by the convolution processing unit; and inputting the convolutional structural features into a feature aggregation unit of the first neural network to obtain local structural features output by the feature aggregation unit.

[0053] In this step, in this step, Figure 2 This is a schematic diagram of the structure of the first neural network provided by the present invention, as shown below. Figure 2As shown, the first neural network includes convolutional layers, average pooling layers, and fully connected layers.

[0054] The feature extraction unit includes a first feature extraction convolutional layer C1, a second feature extraction convolutional layer C4, and a third feature extraction convolutional layer C4; the feature aggregation unit includes a first fully connected layer FC1 and a second fully connected layer FC2; the convolution processing unit includes three average pooling layers, which are not limited in this embodiment.

[0055] The scheduling sequence information is represented as follows: , This indicates that the multidimensional production scheduling state matrix is ​​serialized to obtain scheduling sequence information. Represents a set, This indicates the number of scheduling objects in the scheduling object. This represents the scheduling feature dimension corresponding to each scheduling object.

[0056] Specifically, the multidimensional production scheduling state matrix is ​​input into the feature extraction unit of the first neural network, and the feature extraction unit processes the multidimensional production scheduling state matrix. The mapping is performed to obtain the two-dimensional features output by the feature extraction unit. Then, the two-dimensional features The input to the convolutional processing unit of the first neural network extracts convolutional structure features layer by layer through multiple feature extraction convolutional layers (C1, C2, C3, C4). For a given feature... The feature extraction convolutional layer of the layer extracts the features of the convolutional structure as shown in formula (1).

[0057] (1) In formula (1), This represents a two-dimensional convolution operation. express Layer convolution kernel parameters, Indicates the bias term. Indicates the first The convolutional structure characteristics of the layers, Indicates the first The convolutional structure characteristics of the layers.

[0058] The advantage of this setup is that by performing convolution operations on the convolutional layers through multiple feature extraction convolutional processing units, the first neural network can learn the local correlations and state features between adjacent scheduling objects in the multidimensional production scheduling state matrix. These local correlations and state features are used to reflect local resource competition relationships and local congestion states.

[0059] After each feature extraction convolutional layer, a smooth pooling layer is introduced, and the calculation of the smooth pooling layer is shown in Equation (2).

[0060] (2) In formula (2), This represents the local region corresponding to the pooling window between scheduling objects 𝑗 and 𝑖. Indicates the first The convolutional structure features between layer scheduling objects A and B. Indicates and Different feature extraction convolutional layers; 𝑖 and 𝑗 both represent the numbers of different scheduling objects. Each represents a number of a different scheduling object. Includes 𝑖 and 𝑗, Indicates the first Convolutional structure features between layer scheduling objects 𝑗 and 𝑖.

[0061] For the fully connected layer of the feature aggregation unit, the mapping relationship of the fully connected layer is shown in Equation (3).

[0062] (3) In formula (3), Indicates the first The output of a fully connected layer Indicates the first The weights of each fully connected layer Indicates the first The output of a fully connected layer Indicates the first The bias of a fully connected layer (.) denotes a non-mapping function. After obtaining the outputs of each fully connected layer, feature fusion is performed on the outputs of each fully connected layer to finally obtain local structural features. The calculation is shown in formula (4).

[0063] (4) In formula (4), Indicates to The result of feature fusion of the outputs of fully connected layers.

[0064] In one specific embodiment, the scheduling sequence information is input into a second neural network to obtain the global structural features output by the second neural network. This includes: inputting the scheduling sequence information into the feature embedding unit of the second neural network to obtain high-dimensional feature information input into the feature embedding unit; inputting the high-dimensional feature information into the position encoding unit of the second neural network to obtain embedded feature information output by the position encoding unit; inputting the embedded feature information into the self-attention modeling unit of the second neural network to obtain the global scheduling dependency output by the self-attention modeling unit; and inputting the global scheduling dependency into the feedforward fully connected unit of the second neural network to obtain the global structural features output by the feedforward fully connected unit.

[0065] In this step, the second neural network specifically performs global dependency modeling on the scheduling sequence information after serializing the multidimensional production scheduling state matrix, in order to extract the system-level correlation between scheduling objects and generate a network with global structural features.

[0066] Among them, scheduling sequence information , Indicates scheduling sequence information, Indicates the number of scheduled objects. Indicates the first Status information of each scheduling object.

[0067] Specifically, the scheduling sequence information is input into the feature embedding unit of the second neural network. Based on the feature embedding unit, the scheduling sequence information is subjected to embedding mapping processing to obtain the high-dimensional feature information input to the feature embedding unit. ,in, Indicates the first High-dimensional feature information of each scheduling object The feature embedding unit uses a feature embedding mapping function, which maps the original scheduling sequence data to a unified representation space. This process yields high-dimensional feature information. Next, the high-dimensional feature information is input into the position encoding unit of the second neural network. This position encoding unit then processes the information into positional encoding or object encoding to distinguish the order of different scheduling objects and their roles in the production scheduling system. This yields the embedded feature information output by the position encoding unit. ,in, This represents the position encoding vector or object encoding vector corresponding to the nth scheduling object. Simultaneously, based on embedded feature information... Determine the sequence of scheduling objects, the sequence of scheduling objects , Represents a sequence of scheduling objects. Indicates the first Embedded feature information of each scheduling object The number of scheduling objects is represented by this feature. By constructing different types of scheduling objects into a unified sequence, multi-source scheduling information can be jointly modeled within the same feature space. Specifically, within the Transformer encoder of the second neural network, a self-attention modeling mechanism is used to calculate the correlation weights between scheduling objects, enabling each scheduling object to comprehensively consider the state information of other scheduling objects when updating its own feature representation. The core idea is that different scheduling objects are not independent of each other, but rather have complex relationships such as equipment load transfer, job queuing impact, and order delivery constraints.

[0068] For example, the correlation weights between scheduling objects The calculation is shown in formula (5).

[0069] (5) In formula (5), This represents the relevance weight of scheduling object A to scheduling object B. The relevance weight indicates the degree of influence of scheduling object A on scheduling object B. This represents the correlation calculation function, which can be set according to the actual application. This embodiment does not limit this.

[0070] In formula (5), This represents the natural exponential function. Represents the scheduling object .

[0071] After determining the relevance weights, the embedded feature information is processed based on these relevance weights. Perform global weighted aggregation. This represents the embedded feature information of the scheduling object 𝑗; through global weighted aggregation, the scheduling object features, including global dependency information, are obtained. The calculation is shown in formula (6).

[0072] (6) In formula (6), Represents the scheduling object The advantage of setting the characteristics of the scheduling objects is that each scheduling object can pay attention to other objects that have an important impact on the current scheduling decision during the feature update process, thereby realizing the modeling of global scheduling dependencies across devices, processes and orders.

[0073] In one specific embodiment, further, to maintain the integrity of the embedded feature information of the scheduling object, the features of the scheduling object are... and embedded feature information The final object feature information is obtained by fusing through residual connections and combined with layer normalization. Object feature information The calculation is shown in formula (7).

[0074] (7) In formula (7), The layer normalization function is represented. Represents the scheduling object Object characteristic information.

[0075] Based on this, the object feature information of each scheduled object after layer normalization is further transformed and reconstructed through feedforward mapping to enhance the feature expression capability of the object feature information. The feedforward mapping process is shown in formula (8).

[0076] (8) In formula (8), This represents the mapping function of the feedforward neural network. Indicates the scheduling object Object feature information The scheduling object obtained after feedforward mapping The target object's characteristic information.

[0077] Finally, the target object feature information Input the self-attention modeling unit of the second neural network to obtain the global scheduling dependency output by the self-attention modeling unit; input the global scheduling dependency into the feedforward fully connected unit of the second neural network to obtain the global structural features output by the feedforward fully connected unit.

[0078] In one specific embodiment, when determining the scheduling object Target object feature information Next, the target object feature information of all scheduled objects is weighted and aggregated to generate global structural features. Global structural features The calculation is shown in formula (9).

[0079] (9) In formula (9), Represents the scheduling object The weighting coefficients (global scheduling dependencies). This represents global structural features.

[0080] For example, Figure 3 This is a schematic diagram of the workflow of the second neural network provided by the present invention. Figure 3As shown, it includes steps 301, 302, 303, 304, 305, 306 and 307.

[0081] Step 301: Serialize the multidimensional production scheduling state matrix to obtain scheduling sequence information.

[0082] Step 302: Input the scheduling sequence information into the feature embedding unit of the second neural network to obtain the high-dimensional feature information input to the feature embedding unit.

[0083] In this step, the second neural network may be, for example, a Transformer network structure, and the feature embedding unit is used to linearly map the state features of the scheduling sequence information to obtain a high-dimensional feature representation. This embodiment does not limit this.

[0084] Specifically, the scheduling sequence information is input into the feature embedding unit of the second neural network. Based on the feature embedding unit, the scheduling sequence information is subjected to embedding mapping processing to obtain the high-dimensional feature information input to the feature embedding unit. ,in, Indicates the first High-dimensional feature information of each scheduling object The feature embedding mapping function used by the feature embedding unit maps the original scheduling sequence data to a unified representation space.

[0085] Step 303: Input the high-dimensional feature information into the position encoding unit of the second neural network to obtain the embedded feature information output by the position encoding unit.

[0086] In this step, the location coding unit is used to introduce location information or entity identification information for different scheduling objects, and this embodiment does not limit this.

[0087] Specifically, in obtaining high-dimensional feature information Next, the high-dimensional feature information is input into the position encoding unit of the second neural network. This position encoding unit then processes the information into positional encoding or object encoding to distinguish the order of different scheduling objects and their roles in the production scheduling system. This yields the embedded feature information output by the position encoding unit. ,in, This represents the location encoding vector or object encoding vector corresponding to the nth scheduling object.

[0088] Step 304: Input the sequence of scheduled objects into the Transformer encoder of the second neural network to obtain the features of the scheduled objects output by the Transformer encoder of the second neural network.

[0089] Specifically, in determining the embedded feature information Then, based on the embedded feature information Determine the sequence of scheduling objects, the sequence of scheduling objects , Represents a sequence of scheduling objects. Indicates the first Embedded feature information of each scheduling object This represents the number of scheduled objects. Specifically, within the Transformer encoder of the second neural network, a self-attention modeling mechanism is used to calculate the correlation weights between scheduled objects. After determining the correlation weights, the embedded feature information is then processed based on these weights. Perform global weighted aggregation. This represents the embedded feature information of the scheduling object 𝑗; through global weighted aggregation, the scheduling object features, including global dependency information, are obtained. The calculation is shown in formula (6).

[0090] Step 305: Determine the object feature information based on the scheduling object features and embedded feature information.

[0091] Specifically, to maintain the integrity of the embedded feature information of the scheduling object, the features of the scheduling object are... and embedded feature information The final object feature information is obtained by fusing through residual connections and combined with layer normalization. Object feature information The calculation is shown in formula (7).

[0092] Step 306: Determine global scheduling dependencies based on object characteristic information.

[0093] Specifically, after obtaining the object feature information, the object feature information of each scheduled object after layer normalization is nonlinearly transformed and reconstructed through feedforward mapping to enhance the feature representation capability of the object feature information. The feedforward mapping process is shown in formula (8). In formula (8), This represents the mapping function of the feedforward neural network. Indicates the scheduling object Object feature information The scheduling object obtained after feedforward mapping The target object feature information. Finally, the target object feature information... Input the self-attention modeling unit of the second neural network to obtain the global scheduling dependency output by the self-attention modeling unit.

[0094] Among them, the global scheduling dependency relationship, namely the weight coefficient of the scheduling object, is used to measure the importance of different scheduling objects to the current production scheduling system's scheduling decision.

[0095] Step 307: Perform feature aggregation on the features of the scheduling object based on the global scheduling dependency relationship to obtain the global structural features.

[0096] Specifically, the global scheduling dependencies (weight coefficients) and the scheduling objects are... Target object feature information The input is the feedforward fully connected unit of the second neural network, and weighted aggregation is performed based on the feedforward fully connected unit to generate global structural features. Global structural features The calculation is shown in formula (9).

[0097] The advantage of this setup is that the obtained global structural features are used to comprehensively reflect the overall operating status of the production scheduling system at the current scheduling moment. Moreover, the global structural features can capture the long-range dependencies between scheduling objects and system-level scheduling constraints, enabling subsequent deep reinforcement learning scheduling decision algorithms to have the ability to perceive the global congestion situation, the impact of order priority, and the cross-device resource competition relationship when making scheduling decisions for the production scheduling system. This significantly improves the global consistency and stability of dynamic scheduling decisions.

[0098] Step 104: Determine the production scheduling status based on local and global structural features.

[0099] Specifically, after obtaining the local and global structural features, feature fusion is performed on the local and global structural features to obtain the unified production scheduling state of the production scheduling system.

[0100] The advantage of this setup is that it reflects the local operating status and overall scheduling situation of the production scheduling system based on the production scheduling status.

[0101] In one specific embodiment, determining the production scheduling state based on local structural features and global structural features includes: performing a fusion process on the local structural features and global structural features to obtain the production scheduling state; wherein the fusion process includes at least one of feature concatenation, feature mapping, and feature weighting.

[0102] Specifically, local and global structural features are fused to obtain structural fusion features, which are then mapped and compressed to generate a unified production scheduling state. The fusion process includes at least one of feature concatenation, feature mapping, and feature weighting.

[0103] The advantage of this setup is that it avoids deviations caused by scheduling decisions based solely on local or global structural features, thus allowing the subsequent determination of production scheduling system scheduling decisions to consider the dependencies between local and global structural features simultaneously.

[0104] For example, due to local structural features and global structural features The feature dimensions and semantic spaces may differ, therefore, for local structural features... and global structural features First, feature mapping is performed to map local structural features. and global structural features Mapping to a unified feature space, for local structural features Mapping is performed to obtain the target's local structural features global structural features Mapping is performed to obtain the target global structural features ;in, This represents the feature mapping function used for local structural feature mapping, which is used to align features from different sources within the same representation space. Both represent feature mapping functions used for global structural feature mapping, which are used to align features from different sources within the same representation space.

[0105] In one specific embodiment, the target local structural features are obtained by feature mapping from local structural features and global structural features using feature concatenation. and target global structural features The fusion process is performed to obtain structural fusion features. For example, it could be formula (10).

[0106] (10) The advantage of this setup is that, through the feature splicing fusion process, the structural fusion features simultaneously include the target's local structural features. and target global structural features .

[0107] In one specific embodiment, a weighted fusion mechanism may also be introduced to map the target local structural features to the local structural features obtained after feature mapping of local and global structural features. and target global structural features Feature weighting is performed to obtain structural fusion features. For example, it could be formula (11).

[0108] (11) In formula (11), Represents the fusion weight coefficient, and Used to balance the local structural features of the target and target global structural features The importance of fusion representation.

[0109] Furthermore, after obtaining structural fusion characteristics Subsequently, further analysis of structural fusion characteristics was conducted. Mapping and compression processes are performed to generate a unified production scheduling state. Production scheduling status The calculation is shown in formula (12).

[0110] (12) In formula (12), This represents a mapping and compression operation function, which is not limited in this embodiment.

[0111] In formula (12), It represents the production scheduling status, which is used to comprehensively reflect the local operating status and global scheduling situation of the production scheduling system at the current scheduling moment, and serves as the input for subsequent deep reinforcement learning scheduling decision algorithms.

[0112] The advantage of this setup is that, through operations such as mapping and compression, redundant features can be reduced and key scheduling information can be highlighted, thereby improving the computational efficiency and stability of the scheduling decision-making process.

[0113] In summary, feature fusion processing can simultaneously consider local congestion and the scheduling constraints of the production scheduling system during the decision-making process, thus avoiding the problem of unbalanced scheduling strategies. Furthermore, the use of an adjustable fusion weight method can adaptively adjust the degree of attention to local and global factors when dealing with different scheduling scenarios. When the overall production scheduling system load reaches a high level or order delivery deadlines are approaching, the weight of global scheduling factors should be increased; if local equipment experiences congestion or abnormalities, the weight of local scheduling features should be increased. Therefore, the production scheduling state determined after feature fusion processing is a more comprehensive representation, significantly enhancing the robustness of dynamic scheduling methods in complex production environments.

[0114] Step 105: Determine the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm; wherein, the production scheduling system scheduling decision is the decision to instruct the release of operations, processing sequence and resource allocation during the operation of the production scheduling system; the deep reinforcement learning scheduling decision algorithm is an algorithm that generates the production scheduling system scheduling decision based on the production scheduling status.

[0115] In this step, the deep reinforcement learning scheduling decision algorithm can be, for example, the Asynchronous Advantage Actor-Critic (A3C), and this embodiment does not limit it.

[0116] The scheduling decisions of the production scheduling system may include, for example, scheduling control parameters and scheduling actions, which are not limited in this embodiment.

[0117] Specifically, after determining the production scheduling status, A3C will generate production scheduling system scheduling decisions based on the production scheduling status, including scheduling control parameters and scheduling actions, and update and optimize the production scheduling system scheduling decisions in real time to achieve dynamic generation of production scheduling system scheduling decisions; the production scheduling system scheduling decisions guide the production scheduling system to perform operations such as publishing, task allocation or resource allocation at the current scheduling moment.

[0118] The advantage of this setup is that by introducing the A3C reinforcement learning framework, it can perform adaptive learning in complex and constantly changing production environments, thereby avoiding dependence on fixed rules and solving the performance bottleneck of static scheduling strategies.

[0119] In one specific embodiment, A3C includes an Actor network and a Critic network; wherein, the Actor network is used to model scheduling strategies.

[0120] Specifically, the Actor network is represented by a unified production scheduling state. As input, the probability distribution of scheduling actions or continuous scheduling control parameters are output through the policy function, which is exemplified as shown in formula (13).

[0121] (13) In formula (13), Represents the policy function. This represents the scheduling action output by at the scheduling time. This represents the policy parameters of the Actor network. This represents the production scheduling status of at scheduling time , through the policy function. It can generate production scheduling system decisions under different production scheduling states, and thus gradually learn better production scheduling system decisions during long-term operation.

[0122] In one specific embodiment, the Critic network is used to evaluate the state value of the current production scheduling state, and can measure the long-term benefits that can be obtained by making production scheduling system scheduling decisions in this production scheduling state.

[0123] Specifically, the Critic network uses a unified production scheduling status at scheduling time 𝑡 As input, the corresponding state value estimation information is output, as shown in formula (14).

[0124] (14) In formula (14), The parameters of the Critic network are represented by the state value estimation. The Critic network can provide an evaluation basis for updating the scheduling decisions of the production scheduling system of the Actor network.

[0125] In one specific embodiment, the advantage function and strategy are further updated. Specifically, an advantage function is introduced to measure the merits of the current scheduling action relative to the average strategy. It is represented by formula (15).

[0126] (15) In formula (15), This represents the dominance function at scheduling time 𝑡; This indicates the cumulative reward or incentive signal at the scheduling time. This indicates the production scheduling status of the Critic network at scheduling time 𝑡. State value estimation information.

[0127] Then, based on the advantage function The policy function of the Actor network is updated to obtain the objective function. The calculation is shown in formula (16).

[0128] (16) In formula (16), Describe the objective function. Represents the policy function. This represents the dominance function at scheduling time 𝑡.

[0129] Meanwhile, the Critic network updates by minimizing the state value estimation error, and its exemplary loss function is as follows: It can be expressed as formula (17).

[0130] (17) In formula (17), This indicates the cumulative reward or incentive signal at the scheduling time. This indicates the production scheduling status of the Critic network at scheduling time 𝑡. State value estimation information, This represents the loss function.

[0131] Based on the objective function and loss function This embodiment does not limit the updating and optimization of the Actor network.

[0132] In one specific embodiment, when updating the scheduling decision of the production scheduling system, A3C adopts an asynchronous update mechanism. By adopting the asynchronous update method, the training efficiency of the production scheduling system's scheduling decision can be significantly improved, and the generalization ability of the production scheduling system's scheduling decision in different production scenarios can also be enhanced.

[0133] Specifically, the production scheduling system generates scheduling decisions through an Actor network, evaluates these decisions through a Critic network, and continuously optimizes them using a dominance function.

[0134] The advantage of this setup is that by introducing A3C, stable and efficient production scheduling system scheduling decisions can be achieved in dynamic production scheduling scenarios with multiple constraints, significantly improving the adaptability of the production scheduling system scheduling decisions and optimizing the overall performance.

[0135] In this invention, A3C is based on the specific structural form, objective function, loss function and update method of the Actor network and Critic network, which are not limited in this embodiment.

[0136] In one specific embodiment, Figure 4 This is a schematic diagram of the workflow of the deep reinforcement learning scheduling decision algorithm provided by the present invention, as shown below. Figure 4 As shown, it includes steps 401, 402, 403, 404 and 405.

[0137] Step 401: Determine the production scheduling status.

[0138] Specifically, the production scheduling status is determined based on local and global structural characteristics.

[0139] Step 402: Input the production scheduling status into the Actor network in the deep reinforcement learning scheduling decision algorithm to obtain the production scheduling system scheduling decision output by the Actor network.

[0140] Specifically, the Actor network is represented by a unified production scheduling state. As input, the production scheduling system outputs a scheduling decision through a strategy function. This decision includes the probability distribution of scheduling actions or continuous scheduling control parameters, as exemplified in formula (13). In formula (13), Represents the policy function. This represents the scheduling action output by at the scheduling time. This represents the policy parameters of the Actor network. This represents the production scheduling status of at scheduling time , through the policy function. It can generate production scheduling system decisions under different production scheduling states, and thus gradually learn better production scheduling system decisions during long-term operation.

[0141] Step 403: Input the production scheduling status into the Critic network in the deep reinforcement learning scheduling decision algorithm to obtain the state value estimation information output by the Critic network.

[0142] Specifically, the Critic network is used to evaluate the state value of the current production scheduling state, and can measure the long-term benefits that can be obtained by making production scheduling system scheduling decisions in this production scheduling state.

[0143] Specifically, the Critic network uses a unified production scheduling status at scheduling time 𝑡 As input, the corresponding state value estimation information is output, as shown in formula (14). In formula (14), The parameters of the Critic network are represented by the state value estimation. The Critic network can provide an evaluation basis for updating the scheduling decisions of the production scheduling system of the Actor network.

[0144] In one specific embodiment, steps 402 and 403 are executed synchronously, and this embodiment does not limit this.

[0145] Step 404: Introduce the advantage function to evaluate the current scheduling action and determine the objective function and loss function.

[0146] Specifically, a dominance function is introduced to measure the merits of the current scheduling action relative to the average policy. This is represented by formula (15). In formula (15), This represents the dominance function at scheduling time 𝑡; This indicates the cumulative reward or incentive signal at the scheduling time. This indicates the production scheduling status of the Critic network at scheduling time 𝑡. The state value estimation information. Then, based on the advantage function. The policy function of the Actor network is updated to obtain the objective function. The calculation is shown in formula (16). In formula (16), Describe the objective function. Represents the policy function. This represents the advantage function at scheduling time 𝑡. Meanwhile, the Critic network updates by minimizing the state value estimation error; its exemplary loss function is shown below. This can be represented by formula (17). In formula (17), This indicates the cumulative reward or incentive signal at the scheduling time. This indicates the production scheduling status of the Critic network at scheduling time 𝑡. State value estimation information, This represents the loss function.

[0147] Step 405: Update the Actor network asynchronously based on the objective function and loss function.

[0148] Specifically, in determining the objective function and loss function Then, based on the objective function and loss function Asynchronous policy updates and optimizations are performed on the Actor network.

[0149] The advantage of this setup is that it optimizes the Actor network and improves the accuracy and stability of the scheduling decisions made by the production scheduling system output by the Actor network.

[0150] In one specific embodiment, the method further includes: after determining the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm, parsing and transforming the production scheduling system scheduling decision, mapping the production scheduling system scheduling decision into scheduling control parameters that can be directly used in the production scheduling system, and directly performing scheduling control on the production scheduling system based on the scheduling control parameters.

[0151] In one specific embodiment, the method further includes: after generating the scheduling control parameters, performing constraint verification on the scheduling control parameters to ensure that the scheduling control parameters comply with the process constraints, safety constraints, and operational boundary conditions of the production scheduling system. When the scheduling control parameters exceed the preset constraint range, the scheduling control parameters are corrected, and the corrected scheduling instructions are determined to ensure the executability of the corrected scheduling instructions and the stability of the production process.

[0152] In one specific embodiment, after determining the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm, the method further includes: determining the scheduling interpretation result based on the scheduling sequence information and the second neural network; adjusting the production scheduling system scheduling decision according to the scheduling interpretation result, determining the strategy adjustment result, and continuing to execute the step of determining the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm based on the strategy adjustment result.

[0153] Specifically, after determining the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm, the scheduling interpretation result is determined based on the scheduling sequence information and the second neural network; the scheduling system scheduling decision is adjusted according to the scheduling interpretation result, the strategy adjustment result is determined, and based on the strategy adjustment result, the steps of determining the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm are continued.

[0154] In one specific embodiment, the scheduling interpretation result includes a scheduling object-level attention interpretation result and a scheduling feature-level contribution interpretation result. Determining the scheduling interpretation result based on the scheduling sequence information and the second neural network includes: obtaining the self-attention weight matrix of the second neural network in the scheduling decision process of the scheduling sequence information; determining the scheduling object-level attention interpretation result based on the self-attention weight matrix; and performing contribution decomposition on the scheduling sequence information to obtain the scheduling feature-level contribution interpretation result.

[0155] In this step, the self-attention weight matrix refers to the weights of the correlation terms between scheduling object 𝑗 and scheduling object 𝑖. ,in, This indicates the degree of influence of scheduling object 𝑗 on scheduling object 𝑖 during the scheduling decision-making process.

[0156] Specifically, the self-attention weight matrix of the second neural network in the process of making scheduling decisions on scheduling sequence information is obtained; the attention interpretation result at the scheduling object level is determined based on the self-attention weight matrix; and the contribution decomposition of the scheduling sequence information is performed to obtain the contribution interpretation result at the scheduling feature level.

[0157] In one specific embodiment, the method further includes aggregating the attention weight matrix along the row or column direction to obtain a comprehensive attention index for each scheduling object (i.e., the attention interpretation result at the scheduling object level). The calculation is shown in formula (18).

[0158] (18) In formula (18), This indicates the level of attention given to the nth scheduling object in the current production scheduling system's scheduling decision.

[0159] The advantage of this setup is that, through the attention weight interpretation mechanism, it is possible to identify the key equipment, operations, and order objects that are of primary concern during the scheduling decision-making process, thereby achieving interpretable analysis at the scheduling object level.

[0160] In one specific embodiment, the scheduling sequence information is decomposed into contribution values ​​to obtain the scheduling feature-level contribution interpretation results.

[0161] In this step, SHAP (SHapley Additive exPlanations, a tool for interpreting machine learning models) is used to quantify the contribution of scheduling sequence information. This embodiment does not limit this step.

[0162] Specifically, firstly, the scheduling sequence information is transformed to obtain a scheduling state feature vector. ,in, This represents the number of scheduling state features in the scheduling sequence information. Then, the scheduling state feature vector... Decompose the contribution value to obtain the scheduling decision output contribution value. Then, based on the contribution values ​​of all scheduling state features, the scheduling feature-level contribution interpretation result is determined. The calculation is shown in formula (19).

[0163] (19) In formula (19), Indicates the baseline contribution value. Indicates the first The contribution of each scheduling state feature to the scheduling decision of the production scheduling system.

[0164] The advantage of this setup is that statistical analysis of the contribution values ​​corresponding to each scheduling state characteristic can clarify the degree of positive or negative influence of different scheduling state characteristics on the scheduling decisions of the production scheduling system, thereby achieving interpretable analysis at the scheduling characteristic level.

[0165] In one specific embodiment, after obtaining the scheduling object-level attention interpretation result and the scheduling feature-level contribution interpretation result, the method further includes: fusing the scheduling object-level attention interpretation result and the scheduling feature-level contribution interpretation result to construct a scheduling reliability evaluation index.

[0166] Specifically, scheduling reliability assessment indicators The calculation is shown in formula (20).

[0167] (20) In formula (20), This represents the opportunity to represent the scheduling state characteristics associated with the nth scheduling object.

[0168] The advantage of this setup is that, based on the scheduling reliability assessment index, it can comprehensively reflect the degree of dependence of the deep reinforcement learning scheduling decision-making algorithm on key scheduling objects and key scheduling features during the scheduling decision-making process, providing a quantitative reference for the analysis and correction of subsequent production scheduling system scheduling decisions.

[0169] In summary, multimodal interpretability analysis based on attention weight matrices and contribution values ​​can explain the scheduling decision-making process of deep reinforcement learning scheduling algorithms from both the scheduling object level and the scheduling feature level. This transformation turns the scheduling decision-making process from an inexplicable black box problem into a transparent decision-making process with a traceable basis, effectively improving the credibility, controllability, and engineering application value of scheduling decisions in production scheduling systems.

[0170] In one specific embodiment, after obtaining the scheduling reliability assessment index, it is determined whether the scheduling reliability assessment index is within a preset reasonable range. When it is determined that the scheduling reliability assessment index is within the preset reasonable range, scheduling control is performed according to the original production scheduling system scheduling decision. When it is determined that the scheduling reliability assessment index is not within the preset reasonable range, a strategy correction process is triggered.

[0171] Specifically, when it is determined that the scheduling reliability assessment index is not within the preset reasonable range, the strategy correction process is triggered to adaptively adjust the production scheduling status. After the adjustment, the steps of determining the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm continue to be executed.

[0172] The advantage of this setup is that it creates a closed-loop optimization mechanism in the scheduling decision-making process, which not only ensures scheduling performance but also improves the security, stability, and controllability of the production scheduling system's scheduling decisions. It is especially suitable for complex production scheduling scenarios with stringent requirements for scheduling reliability.

[0173] In one specific embodiment, Figure 5 This is a schematic diagram of the workflow for multimodal interpretable analysis provided by the present invention, such as... Figure 5 As shown, it includes steps 501, 502, 503, 504, and 505.

[0174] Step 501: Determine the scheduling decision of the production scheduling system.

[0175] Step 502: Obtain the self-attention weight matrix of the second neural network in the process of determining the scheduling decision of the production scheduling system based on the scheduling sequence information.

[0176] Specifically, in this step, the self-attention weight matrix refers to the weights of the correlation terms between scheduling object 𝑗 and scheduling object 𝑖. ,in, This indicates the degree of influence of scheduling object 𝑗 on scheduling object 𝑖 during the scheduling decision-making process.

[0177] Step 503: Determine the scheduling object-level attention interpretation result based on the self-attention weight matrix.

[0178] Specifically, the attention interpretation result at the scheduling object level is determined based on the self-attention weight matrix.

[0179] In one specific embodiment, the method further includes aggregating the attention weight matrix along the row or column direction to obtain a comprehensive attention index for each scheduling object. The calculation is shown in formula (18).

[0180] (18) In formula (18), This indicates the level of attention given to the nth scheduling object in the current production scheduling system's scheduling decision.

[0181] The advantage of this setup is that, through the attention weight interpretation mechanism, it is possible to identify the key equipment, operations, and order objects that the production scheduling system focuses on during the scheduling decision-making process, thereby achieving interpretable analysis at the scheduling object level.

[0182] Step 504: Decompose the scheduling sequence information by contribution to obtain the scheduling feature-level contribution interpretation results.

[0183] In this step, SHAP (SHapley Additive exPlanations, a tool for interpreting machine learning models) is used to quantify the contribution of scheduling sequence information. This embodiment does not limit this step.

[0184] Specifically, firstly, the scheduling sequence information is transformed to obtain a scheduling state feature vector. ,in, This represents the number of scheduling state features in the scheduling sequence information. Then, the scheduling state feature vector... Decompose the contribution value to obtain the scheduling decision output contribution value. Then, based on the contribution values ​​of all scheduling state features, the scheduling feature-level contribution interpretation result is determined. The calculation is shown in formula (19). In formula (19), Indicates the baseline contribution value. Indicates the first The contribution of each scheduling state feature to the scheduling decision of the production scheduling system.

[0185] The advantage of this setup is that statistical analysis of the contribution values ​​corresponding to each scheduling state characteristic can clarify the degree of positive or negative influence of different scheduling state characteristics on the scheduling decisions of the production scheduling system, thereby achieving interpretable analysis at the scheduling characteristic level.

[0186] Step 505: Integrate the interpretation results of the attention at the scheduling object level and the interpretation results of the contribution at the scheduling feature level to construct a scheduling reliability evaluation index.

[0187] Specifically, after obtaining the attention interpretation results at the scheduling object level and the contribution interpretation results at the scheduling feature level, the process also includes: fusing the attention interpretation results at the scheduling object level and the contribution interpretation results at the scheduling feature level to construct a scheduling reliability evaluation index.

[0188] Specifically, scheduling reliability assessment indicators The calculation is shown in formula (20). In formula (20), This represents the opportunity to represent the scheduling state characteristics associated with the nth scheduling object.

[0189] The advantage of this setup is that, based on the scheduling reliability assessment index, it can comprehensively reflect the degree of dependence of the deep reinforcement learning scheduling decision-making algorithm on key scheduling objects and key scheduling features during the scheduling decision-making process, providing a quantitative reference for the analysis and correction of subsequent production scheduling system scheduling decisions.

[0190] In one specific embodiment, Figure 6 This is the second flowchart illustrating the scheduling decision determination method for the production scheduling system provided by this invention, as shown below. Figure 6 As shown, the scheduling decision determination method of the production scheduling system includes steps 601, 602, 603, 604, 605, 606 and 607.

[0191] Step 601: Construct a multi-dimensional production scheduling status matrix based on equipment status information, operation status information, and order status information.

[0192] Specifically, the system collects status information during the production scheduling process; this status information includes equipment status information, job status information, and order status information. A multi-dimensional production scheduling status matrix is ​​constructed based on these information. In constructing this matrix, the rows correspond to the scheduling object dimension, with each row representing a single equipment object, job object, or order object. The columns correspond to the scheduling feature dimension, with each column representing a type of status information; this embodiment does not impose any limitations on this aspect.

[0193] Step 602: Input the multidimensional production scheduling state matrix into the first neural network to obtain the local structural features output by the first neural network.

[0194] Specifically: the multidimensional production scheduling state matrix is ​​input into the feature extraction unit of the first neural network to obtain the two-dimensional features output by the feature extraction unit; the two-dimensional features are input into the convolution processing unit of the first neural network to obtain the convolutional structure features output by the convolution processing unit; the convolutional structure features are input into the feature aggregation unit of the first neural network to obtain the local structure features output by the feature aggregation unit.

[0195] Step 603: Input the scheduling sequence information into the second neural network to obtain the global structural features output by the second neural network.

[0196] Specifically, the scheduling sequence information is input into the feature embedding unit of the second neural network to obtain the high-dimensional feature information input to the feature embedding unit; the high-dimensional feature information is input into the position encoding unit of the second neural network to obtain the embedded feature information output by the position encoding unit; the embedded feature information is input into the self-attention modeling unit of the second neural network to obtain the global scheduling dependency output by the self-attention modeling unit; the global scheduling dependency is input into the feedforward fully connected unit of the second neural network to obtain the global structural features output by the feedforward fully connected unit.

[0197] Step 604: Perform fusion processing on local structural features and global structural features to obtain the production scheduling status.

[0198] Specifically, after obtaining local and global structural features, the local and global structural features are fused to obtain structural fusion features. Then, the structural fusion features are mapped and compressed to generate a unified production scheduling state. The fusion process includes at least one of feature concatenation, feature mapping, and feature weighting.

[0199] Step 605: Determine the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm.

[0200] Specifically, after determining the production scheduling status, A3C will generate production scheduling system scheduling decisions based on the production scheduling status, including scheduling control parameters and scheduling actions, and update and optimize the production scheduling system scheduling decisions in real time to achieve dynamic generation of production scheduling system scheduling decisions; the production scheduling system scheduling decisions guide the production scheduling system to perform operations such as publishing, task allocation or resource allocation at the current scheduling moment.

[0201] The advantage of this setup is that by introducing the A3C reinforcement learning framework, it can perform adaptive learning in complex and constantly changing production environments, thereby avoiding dependence on fixed rules and solving the performance bottleneck of static scheduling strategies.

[0202] Step 606: Perform multimodal interpretability analysis on the scheduling decisions of the production scheduling system to obtain the interpretation results of the attention at the scheduling object level and the interpretation results of the contribution at the scheduling feature level.

[0203] Specifically, after determining the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm, the scheduling interpretation result is determined based on the scheduling sequence information and the second neural network. The scheduling interpretation result includes the scheduling object-level attention interpretation result and the scheduling feature-level contribution interpretation result. Based on the scheduling interpretation result, the production scheduling system scheduling decision is adjusted, the strategy adjustment result is determined, and based on the strategy adjustment result, the steps of determining the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm are continued.

[0204] Step 607: Integrate the interpretation results of the attention at the scheduling object level and the interpretation results of the contribution at the scheduling feature level to construct a scheduling reliability evaluation index.

[0205] Specifically, after obtaining the scheduling reliability assessment index, it is determined whether the scheduling reliability assessment index is within a preset reasonable range. When it is determined that the scheduling reliability assessment index is within the preset reasonable range, scheduling control is performed according to the original output production scheduling system scheduling decision. When it is determined that the scheduling reliability assessment index is not within the preset reasonable range, the strategy correction process is triggered to adaptively adjust the production scheduling status. After the adjustment, the steps of determining the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm continue to be executed.

[0206] This invention provides a method for determining scheduling decisions in a production scheduling system. The method involves collecting state information of the production scheduling system during its operation, including equipment state information, job state information, and order state information. A multi-dimensional production scheduling state matrix and scheduling sequence information are determined based on these information. Local and global structural features are determined based on the multi-dimensional production scheduling state matrix, a first neural network, the scheduling sequence information, and a second neural network. The first neural network extracts the local structural features of adjacent scheduling objects in the multi-dimensional production scheduling state matrix. The second neural network uses a self-attention mechanism to extract global structural features of the dependencies between scheduling objects in the scheduling sequence information. The production scheduling state is determined based on the local and global structural features. Finally, the scheduling decision of the production scheduling system is determined based on the production scheduling state and a deep reinforcement learning scheduling decision algorithm. The production scheduling decision involves instructing job releases, processing sequences, and resource allocation during the operation of the production scheduling system. The deep reinforcement learning scheduling decision algorithm is an algorithm that generates production scheduling system scheduling decisions based on the production scheduling state. The technical solution of this invention addresses the shortcomings of existing methods based on rules or mathematical optimization, such as priority rule scheduling or optimization algorithms based on deterministic models. While these methods may have some effectiveness in specific scenarios, they are highly dependent on system modeling and struggle to adapt to dynamic changes such as order fluctuations, equipment failures, and uncertain processing times, resulting in insufficient stability and reliability of scheduling decisions. The invention achieves this by using a neural network to extract features from scheduling sequence information and determine the production scheduling state during the operation of a complex production scheduling system. Then, a deep reinforcement learning scheduling decision algorithm and the production scheduling state are used to generate a production scheduling system decision. Based on this decision, the production scheduling system is dynamically scheduled, improving the reliability, stability, and engineering usability of scheduling decisions, as well as enhancing dynamic scheduling performance.

[0207] The scheduling decision determination device for the production scheduling system provided by the present invention is described below. The scheduling decision determination device for the production scheduling system described below and the scheduling decision determination method for the production scheduling system described above can be referred to in correspondence.

[0208] Figure 7 This is a schematic diagram of the scheduling decision determination device of the production scheduling system provided by the present invention, with reference to... Figure 7 As shown, the scheduling decision determination device 700 of the production scheduling system includes: an information acquisition module 701, an information determination module 702, a feature determination module 703, a status determination module 704, and a decision determination module 705; wherein, The information acquisition module 701 is used to collect the status information of the production scheduling system during the operation process; the status information includes equipment status information, operation status information and order status information.

[0209] The information determination module 702 is used to determine the multi-dimensional production scheduling status matrix and scheduling sequence information based on equipment status information, operation status information and order status information.

[0210] The feature determination module 703 is used to determine local structural features and global structural features based on the multidimensional production scheduling state matrix, the first neural network, the scheduling sequence information, and the second neural network. The first neural network is a network that extracts the local structural features of adjacent scheduling objects in the multidimensional production scheduling state matrix. The second neural network is a network that uses a self-attention mechanism to extract the global structural features of the dependencies between scheduling objects in the scheduling sequence information.

[0211] The state determination module 704 is used to determine the production scheduling state based on local and global structural features.

[0212] The decision determination module 705 is used to determine the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm. The production scheduling system scheduling decision is the decision made by the production scheduling system to indicate the release of operations, processing sequence and resource allocation during the operation process. The deep reinforcement learning scheduling decision algorithm is an algorithm that generates the production scheduling system scheduling decision based on the production scheduling status.

[0213] In one example embodiment, the information determination module 702 is specifically used to: construct a multi-dimensional production scheduling state matrix based on equipment status information, job status information and order status information; and perform serialization processing on the multi-dimensional production scheduling state matrix to obtain scheduling sequence information.

[0214] In one example embodiment, the feature determination module 703 is specifically used to: input the multidimensional production scheduling state matrix into the first neural network to obtain the local structural features output by the first neural network; and input the scheduling sequence information into the second neural network to obtain the global structural features output by the second neural network.

[0215] In one example embodiment, the feature determination module 703 inputs a multidimensional production scheduling state matrix into a first neural network to obtain local structural features output by the first neural network. Specifically, it is used to: input the multidimensional production scheduling state matrix into the feature extraction unit of the first neural network to obtain two-dimensional features output by the feature extraction unit; input the two-dimensional features into the convolution processing unit of the first neural network to obtain convolutional structural features output by the convolution processing unit; and input the convolutional structural features into the feature aggregation unit of the first neural network to obtain local structural features output by the feature aggregation unit.

[0216] In one example embodiment, the feature determination module 703 inputs scheduling sequence information into a second neural network to obtain global structural features output by the second neural network. Specifically, it is used to: input scheduling sequence information into the feature embedding unit of the second neural network to obtain high-dimensional feature information input by the feature embedding unit; input the high-dimensional feature information into the position encoding unit of the second neural network to obtain embedded feature information output by the position encoding unit; input the embedded feature information into the self-attention modeling unit of the second neural network to obtain global scheduling dependencies output by the self-attention modeling unit; and input the global scheduling dependencies into the feedforward fully connected unit of the second neural network to obtain global structural features output by the feedforward fully connected unit.

[0217] In one example embodiment, the state determination module 704 is specifically used to: perform fusion processing on local structural features and global structural features to obtain the production scheduling state; wherein, the fusion processing includes at least one of feature splicing, feature mapping and feature weighting.

[0218] In one example embodiment, the apparatus further includes a strategy adjustment module. The strategy adjustment module is configured to: after determining the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm, determine the scheduling interpretation result based on the scheduling sequence information and the second neural network; adjust the production scheduling system scheduling decision based on the scheduling interpretation result, determine the strategy adjustment result, and based on the strategy adjustment result, continue executing the step of determining the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm.

[0219] In one example embodiment, the scheduling interpretation results include scheduling object-level attention interpretation results and scheduling feature-level contribution interpretation results.

[0220] In one example embodiment, the policy adjustment module determines the scheduling interpretation result based on the scheduling sequence information and the second neural network. Specifically, it is used to: obtain the self-attention weight matrix of the second neural network in the process of making scheduling decisions on the scheduling sequence information; determine the scheduling object-level attention interpretation result based on the self-attention weight matrix; and perform contribution decomposition on the scheduling sequence information to obtain the scheduling feature-level contribution interpretation result.

[0221] The apparatus of this embodiment can be used to execute the method of any embodiment in the side embodiment of the scheduling decision determination method of the production scheduling system. Its specific implementation process and technical effects are similar to those in the side embodiment of the scheduling decision determination method of the production scheduling system. For details, please refer to the detailed description in the side embodiment of the scheduling decision determination method of the production scheduling system, which will not be repeated here.

[0222] Figure 8This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 8 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840. The processor 810, communications interface 820, and memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions from the memory 830 to execute a scheduling decision determination method for the production scheduling system. This method includes: collecting status information of the production scheduling system during its operation; wherein the status information includes equipment status information, job status information, and order status information. Determine the multidimensional production scheduling status matrix and scheduling sequence information based on equipment status information, operation status information, and order status information; The local and global structural features are determined based on the multidimensional production scheduling state matrix, the first neural network, the scheduling sequence information, and the second neural network. The first neural network is a network that extracts the local structural features of adjacent scheduling objects in the multidimensional production scheduling state matrix. The second neural network is a network that uses a self-attention mechanism to extract the global structural features of the dependencies between scheduling objects in the scheduling sequence information. Determine the production scheduling status based on local and global structural characteristics; The production scheduling system makes scheduling decisions based on the production scheduling status and a deep reinforcement learning scheduling decision algorithm. The production scheduling system makes decisions on instructing job releases, processing sequences, and resource allocations during the production scheduling process. The deep reinforcement learning scheduling decision algorithm is an algorithm that generates production scheduling system scheduling decisions based on the production scheduling status.

[0223] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0224] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the scheduling decision determination method of the production scheduling system provided by the above methods. The method includes: collecting the status information of the production scheduling system during the working process; wherein the status information includes equipment status information, job status information and order status information. Determine the multidimensional production scheduling status matrix and scheduling sequence information based on equipment status information, operation status information, and order status information; The local and global structural features are determined based on the multidimensional production scheduling state matrix, the first neural network, the scheduling sequence information, and the second neural network. The first neural network is a network that extracts the local structural features of adjacent scheduling objects in the multidimensional production scheduling state matrix. The second neural network is a network that uses a self-attention mechanism to extract the global structural features of the dependencies between scheduling objects in the scheduling sequence information. Determine the production scheduling status based on local and global structural characteristics; The production scheduling system makes scheduling decisions based on the production scheduling status and a deep reinforcement learning scheduling decision algorithm. The production scheduling system makes decisions on instructing job releases, processing sequences, and resource allocations during the production scheduling process. The deep reinforcement learning scheduling decision algorithm is an algorithm that generates production scheduling system scheduling decisions based on the production scheduling status.

[0225] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program implements a scheduling decision determination method for a production scheduling system provided by the above methods. The method includes: collecting status information of the production scheduling system during its operation; wherein the status information includes equipment status information, job status information, and order status information. Determine the multidimensional production scheduling status matrix and scheduling sequence information based on equipment status information, operation status information, and order status information; The local and global structural features are determined based on the multidimensional production scheduling state matrix, the first neural network, the scheduling sequence information, and the second neural network. The first neural network is a network that extracts the local structural features of adjacent scheduling objects in the multidimensional production scheduling state matrix. The second neural network is a network that uses a self-attention mechanism to extract the global structural features of the dependencies between scheduling objects in the scheduling sequence information. Determine the production scheduling status based on local and global structural characteristics; The production scheduling system makes scheduling decisions based on the production scheduling status and a deep reinforcement learning scheduling decision algorithm. The production scheduling system makes decisions on instructing job releases, processing sequences, and resource allocations during the production scheduling process. The deep reinforcement learning scheduling decision algorithm is an algorithm that generates production scheduling system scheduling decisions based on the production scheduling status.

[0226] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0227] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0228] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for determining scheduling decisions in a production scheduling system, characterized in that, include: Collect status information of the production scheduling system during its operation; wherein, the status information includes equipment status information, job status information, and order status information; A multidimensional production scheduling status matrix and scheduling sequence information are determined based on the equipment status information, the operation status information, and the order status information. Local structural features and global structural features are determined based on the multidimensional production scheduling state matrix, the first neural network, the scheduling sequence information, and the second neural network; wherein, the first neural network is a network that extracts the local structural features of adjacent scheduling objects in the multidimensional production scheduling state matrix; the second neural network is a network that uses a self-attention mechanism to extract the global structural features of the dependencies between scheduling objects in the scheduling sequence information. The production scheduling status is determined based on the local structural features and the global structural features; The production scheduling system makes scheduling decisions based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm. The production scheduling system decision refers to the decision made by the production scheduling system during operation regarding job release, processing sequence, and resource allocation. The deep reinforcement learning scheduling decision algorithm is an algorithm that generates the production scheduling system decision based on the production scheduling status.

2. The scheduling decision determination method for the production scheduling system according to claim 1, characterized in that, The step of determining the multi-dimensional production scheduling status matrix and scheduling sequence information based on the equipment status information, the operation status information, and the order status information includes: The multidimensional production scheduling status matrix is ​​constructed based on the equipment status information, the operation status information, and the order status information; The multidimensional production scheduling state matrix is ​​serialized to obtain the scheduling sequence information.

3. The scheduling decision determination method for the production scheduling system according to claim 1, characterized in that, The step of determining local structural features and global structural features based on the multidimensional production scheduling state matrix, the first neural network, the scheduling sequence information, and the second neural network includes: The multidimensional production scheduling state matrix is ​​input into the first neural network to obtain the local structural features output by the first neural network. The scheduling sequence information is input into the second neural network to obtain the global structural features output by the second neural network.

4. The scheduling decision determination method for the production scheduling system according to claim 3, characterized in that, The step of inputting the multidimensional production scheduling state matrix into the first neural network to obtain the local structural features output by the first neural network includes: The multidimensional production scheduling state matrix is ​​input into the feature extraction unit of the first neural network to obtain the two-dimensional features output by the feature extraction unit; The two-dimensional features are input into the convolution processing unit of the first neural network to obtain the convolutional structure features output by the convolution processing unit; The convolutional structure features are input into the feature aggregation unit of the first neural network to obtain the local structural features output by the feature aggregation unit.

5. The scheduling decision determination method for the production scheduling system according to claim 3, characterized in that, The step of inputting the scheduling sequence information into the second neural network to obtain the global structural features output by the second neural network includes: The scheduling sequence information is input into the feature embedding unit of the second neural network to obtain the high-dimensional feature information input to the feature embedding unit. The high-dimensional feature information is input into the position encoding unit of the second neural network to obtain the embedded feature information output by the position encoding unit; The embedded feature information is input into the self-attention modeling unit of the second neural network to obtain the global scheduling dependency output by the self-attention modeling unit; The global scheduling dependency is input into the feedforward fully connected unit of the second neural network to obtain the global structural features output by the feedforward fully connected unit.

6. The scheduling decision determination method for a production scheduling system according to any one of claims 1-5, characterized in that, Determining the production scheduling status based on the local structural features and the global structural features includes: The local structural features and the global structural features are fused to obtain the production scheduling status; wherein the fusion process includes at least one of feature concatenation, feature mapping and feature weighting.

7. The scheduling decision determination method for a production scheduling system according to any one of claims 1-5, characterized in that, After determining the production scheduling system scheduling decision based on the production scheduling state and the deep reinforcement learning scheduling decision algorithm, the method further includes: The scheduling interpretation result is determined based on the scheduling sequence information and the second neural network; Based on the scheduling interpretation results, the scheduling decision of the production scheduling system is adjusted according to the strategy, the strategy adjustment result is determined, and based on the strategy adjustment result, the step of determining the scheduling decision of the production scheduling system according to the production scheduling status and the deep reinforcement learning scheduling decision algorithm is continued.

8. The scheduling decision determination method for a production scheduling system according to claim 7, characterized in that, The scheduling interpretation result includes a scheduling object-level attention interpretation result and a scheduling feature-level contribution interpretation result; determining the scheduling interpretation result based on the scheduling sequence information and the second neural network includes: Obtain the self-attention weight matrix of the second neural network during the scheduling decision-making process of the scheduling sequence information; The scheduling object-level attention interpretation result is determined based on the self-attention weight matrix; The scheduling sequence information is decomposed into contribution values ​​to obtain the scheduling feature-level contribution explanation results.

9. A scheduling decision-making device for a production scheduling system, characterized in that, include: The information acquisition module is used to collect status information of the production scheduling system during its operation; wherein, the status information includes equipment status information, operation status information and order status information; The information determination module is used to determine a multi-dimensional production scheduling status matrix and scheduling sequence information based on the equipment status information, the operation status information and the order status information; The feature determination module is used to determine local structural features and global structural features based on the multidimensional production scheduling state matrix, the first neural network, the scheduling sequence information, and the second neural network; wherein, the first neural network is a network that extracts local structural features of adjacent scheduling objects in the multidimensional production scheduling state matrix; the second neural network is a network that uses a self-attention mechanism to extract global structural features of the dependencies between scheduling objects in the scheduling sequence information. A state determination module is used to determine the production scheduling state based on the local structural features and the global structural features; The decision-making module is used to determine the production scheduling system scheduling decision based on the production scheduling status and the deep reinforcement learning scheduling decision algorithm; wherein, the production scheduling system scheduling decision is a decision made by the production scheduling system to indicate job release, processing sequence and resource allocation during the operation process; the deep reinforcement learning scheduling decision algorithm is an algorithm that generates the production scheduling system scheduling decision based on the production scheduling status.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the scheduling decision determination method of the production scheduling system as described in any one of claims 1 to 8.

Images