Process route-based production collaborative control method and system

Abstract

This invention provides a production collaborative control method and system based on process routes, belonging to the field of control. It includes constructing a graph model with production units and workpieces as nodes, where edges represent process flow and resource dependencies; encoding the real-time status of nodes and process route features to generate initial node features; calculating the proportion of completed processes for workpiece nodes as a completion scalar; calculating the original attention score based on the initial features and edge relationships; for workpiece nodes, correcting the score using an asymmetric decay function adjusted by completion based on the upstream and downstream positions of neighboring nodes in the process route to obtain attention coefficients, and aggregating neighbor features accordingly; inputting the aggregated features of all nodes into a decoder, where a resource selection module calculates the execution probability; a process sequencing module determines the execution order; and finally, outputting production collaborative control instructions. This invention solves the problem of existing technologies being unable to identify the different importance of process sequencing in decision-making under complex conditions.

Description

Technical Field

[0001] This invention belongs to the field of control, and in particular relates to a production collaborative control method and system based on process route. Background Technology

[0002] In modern discrete manufacturing, production scheduling methods often employ mathematical programming-based models or rule-based heuristic algorithms. Mathematical programming methods struggle to adapt to the demands of large-scale, fast-paced production; heuristic algorithms, based on local information, lack a global perspective and are ill-suited to handling unconventional situations such as sudden equipment failures, urgent order insertions, and material delays, resulting in poor adaptability of scheduling schemes. Intelligent scheduling methods based on artificial intelligence technologies, such as deep reinforcement learning and graph neural networks (GNNs), abstract the production system into a graph structure, enabling the detection of workpieces, equipment, and their process flows and resource dependencies. By propagating and aggregating information on the graph, feature representations, including system states, can be learned, allowing for scheduling decisions. However, existing technologies do not adequately integrate process route information with real-time equipment load and workpiece progress status, hindering the formation of node state awareness. During graph information aggregation, existing GNN models employ a homogeneous aggregation approach, processing neighboring node information indiscriminately and failing to recognize the different importance of upstream and downstream processes for a workpiece node in decision-making. This indiscriminate aggregation weakens the model's understanding of process flow directionality, limiting the global optimality of scheduling decisions. Summary of the Invention

[0003] This invention proposes a production collaborative control method based on process route to address the problem that existing technologies do not sufficiently integrate process route information with real-time equipment load and workpiece progress status, and cannot identify the different importance of upstream and downstream processes in decision-making for a workpiece node. The method includes: Construct a graph model, setting production units and workpieces as nodes, and process flow and resource dependencies as edges, and obtain the real-time status and process route characteristics of each node. The real-time state and process route features are encoded and fused through a cross-attention mechanism to generate initial features for each node, and the proportion of completed processes is calculated for each workpiece node as a completion scalar. The original attention score is calculated based on the initial features and edge features; and when the target node is a workpiece node, the attention coefficient is obtained by correcting the original attention score using the completion scalar based on the upstream and downstream positions of neighboring nodes in the process route of the workpiece node; the attention coefficient is used to perform a weighted summation of the features of neighboring nodes to generate the aggregated features of the target node. The aggregated features of all nodes are input into the decoder for decoding decision-making. The decoding decision-making includes: calculating the execution probability of each workpiece on the available production unit through the resource selection module; and when multiple workpieces compete for the same production unit, the process sequencing module calculates the sequencing value based on the aggregated features of the workpieces and the remaining process time, determines the execution order, and outputs production collaborative control instructions.

[0004] Furthermore, the present invention also relates to a production collaborative control system based on a process route, comprising the following modules: The acquisition module is used to construct a graph model, setting production units and workpieces as nodes, process flow and resource dependencies as edges, and acquiring the real-time status and process route characteristics of each node. The calculation module is used to encode and fuse the real-time state and process route features through a cross-attention mechanism to generate the initial features of each node, and to calculate the proportion of completed processes for each workpiece node as a completion scalar. The generation module is used to calculate the original attention score based on the initial features and edge features; and when the target node is a workpiece node, it obtains the attention coefficient by correcting the original attention score using the completion scalar according to the upstream and downstream positions of the neighboring nodes in the process route of the workpiece node; and uses the attention coefficient to perform a weighted summation of the features of the neighboring nodes to generate the aggregated features of the target node. The output module is used to input the aggregated features of all nodes into the decoder for decoding decisions. The decoding decisions include: calculating the execution probability of each workpiece on the available production unit through the resource selection module; and when multiple workpieces compete for the same production unit, the process sequencing module calculates the sequencing value based on the aggregated features of the workpieces and the remaining process time, determines the execution order, and outputs production collaborative control instructions.

[0005] This invention constructs a production system graph model, which can consider the relationships between production units, workpieces, process flows, and resource dependencies from a global perspective. It integrates the real-time status of nodes with process route characteristics to represent the state of the production site. Utilizing an asymmetric decay function adjusted by workpiece completion, it can distinguish the upstream and downstream positions of neighboring nodes in the process route, allowing the model to focus on node information that significantly impacts future decisions. Through decisions on resource selection and process sequencing, it achieves rational allocation of production resources and conflict resolution, optimizes production scheduling schemes, and improves overall production collaboration efficiency. Attached Figure Description

[0006] Figure 1 A flowchart of the first embodiment; Figure 2 A schematic diagram of node feature encoding; Figure 3This is a schematic diagram illustrating the cross-attention mechanism and completion calculation. Figure 4 This is a schematic diagram of an asymmetric decay function; Figure 5 This is a schematic diagram of the process sequencing module. Detailed Implementation

[0007] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0008] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0009] In the first embodiment, the present invention proposes a production collaborative control method based on process route, such as... Figure 1 ,include: S1. Construct a graph model, setting production units and workpieces as nodes, process flow and resource dependencies as edges, and obtain the real-time status and process route characteristics of each node. Construct a heterogeneous graph where production unit nodes are such as CNC machine tools and assembly tables, and workpiece nodes are such as parts to be processed. Process flow relationships connect a workpiece node to a production unit node required for a process, assigning a process sequence number attribute to indicate which process the production unit is responsible for. Resource dependency relationships connect a workpiece node to a currently occupied production unit node. The real-time status of a production unit node includes idle, running, and faulty. The real-time status of a workpiece node includes waiting for processing, processing, and completed. For workpiece nodes, the process route features are the total number of processes, the current process number, and the remaining total processing time; for production unit nodes, they are the processing type, efficiency, and accuracy level. Encode the real-time status as a one-hot vector and the process route features as a numerical vector. For example, if a production unit's process route processing type is milling, its efficiency is 1.2, and its accuracy level is 0.9, then its process route feature vector could be [1, 1.2, 0.9].

[0010] In an optional embodiment, the construction of the graph model includes: The real-time status of the production unit is encoded as a unique hot vector, representing the idle, processing, or fault status. The process route features of the workpiece node are encoded as a numerical vector containing the total number of processes, the current process number, and the standard processing time of the current process.

[0011] Assuming a production unit exists in three states: idle, processing, and faulty, it can be represented using a three-dimensional one-hot vector. For example, the idle state is encoded as [1,0,0], the processing state as [0,1,0], and the faulty state as [0,0,1]. This vector reflects the availability of the production unit at the time of scheduling decision, such as... Figure 2 For each workpiece to be processed, the process information is transformed into a numerical vector. For example, if a workpiece has a total of 8 processes, and is currently in the 3rd process with a standard processing time of 30 minutes, then the process route feature vector can be represented as [8, 3, 30]. This vector contains the workpiece's global progress, current local progress, and current task load information. The graph neural network model consists of an initial feature embedding module, a multi-layer graph attention network encoder, and two policy decoder heads. The training set consists of job shop scheduling problem instances generated through simulation. Each instance contains a set of workpieces and process routes, processing time, and production unit information. The model is trained using a policy gradient reinforcement learning algorithm, such as REINFORCE, to minimize the maximum completion time. The model input is a heterogeneous graph constructed at each decision moment, where nodes represent workpieces and production units. The initial features of each node are a one-hot encoded state vector and a numericalized process route vector. The model output consists of two parts: a resource selection probability distribution and a process ranking value, which together determine the next workpiece pairing and its order.

[0012] S2, by encoding and fusing the real-time state and process route features through a cross-attention mechanism, the initial features of each node are generated, and the proportion of completed processes for each workpiece node is calculated as a completion scalar. The real-time status features of the node are used as the query vector Q, and the vector process route features are used as the key vector K and value vector V. The fused features are obtained by scaling dot product attention calculation. The fused features are then concatenated with the original real-time status and process route features to obtain the initial feature vector of each node. The completion scalar is obtained by dividing the number of completed processes of the workpiece by the total number of processes.

[0013] To fuse two types of heterogeneous information from nodes—state features and process route features—to obtain an informative initial node representation, in an optional embodiment, the real-time state and process route features are fused using a cross-attention mechanism to generate initial features for each node, and the proportion of completed processes for each workpiece node is calculated as a completion scalar, including: For each node, the state features of the node are used as the query vector in the cross-attention mechanism, and the process route features are used as the key vector and value vector to generate the fused initial feature vector. The completion scalar is calculated based on the current process number and the total number of processes of the workpiece.

[0014] Taking a workpiece node as an example, its state feature, such as whether it is waiting for processing, is used as the query vector Q. The workpiece's process route feature vector, such as a vector containing the total number of processes and the current process number, is used as both the key vector K and the value vector V. By calculating the similarity score between the query vector and the key vector, the part of the process route feature most relevant to the current state can be identified. The similarity score is normalized using the Softmax function, and a unified initial feature vector is generated through weighted summation. Simultaneously, a completion scalar is calculated; this scalar is a value between 0 and 1, obtained by dividing the currently completed process number by the total number of processes. For example, if a workpiece has a total of 10 processes and is currently in the 4th process, then the completion scalar p is 0.3. Figure 3 .

[0015] S3, calculate the original attention score based on the initial features and edge features; and when the target node is a workpiece node, obtain the attention coefficient by correcting the original attention score using the completion scalar according to the upstream and downstream positions of neighboring nodes in the process route of the workpiece node; use the attention coefficient to perform a weighted summation of the features of neighboring nodes to generate the aggregated features of the target node; The original attention score of any two connected nodes is calculated using a single-layer feedforward network in a graph attention network. In this invention, the original attention score is calculated based on the initial features and edge features to obtain the association strength between the target node and its neighboring nodes in the current production state. Specifically, the initial features represent the real-time status and process route information of the node itself, while the edge features are used to characterize the process flow relationship or resource dependency relationship between nodes, thus together constituting the context information for interaction between nodes.

[0016] By inputting the initial features of the target node, the initial features of its neighboring nodes, and the corresponding edge features into the attention calculation unit, a raw attention score is obtained. This score comprehensively reflects the importance of node attribute differences and node relationship characteristics to production collaboration decisions. Furthermore, by introducing the raw attention score, the influence weights of different neighboring nodes on the target node's decision can be adaptively distinguished in the graph model, avoiding equal weighting of all neighboring nodes. This allows the model to focus more on node relationships that have a higher impact on production collaboration under the current process stage and resource conditions.

[0017] In actual production, the focus of attention on upstream and downstream process nodes differs at different processing stages. When a workpiece is in the early stages of processing, its production collaboration decisions rely more on the preceding processes and related resource status; however, as workpiece processing gradually completes, the focus shifts to the arrangement of subsequent processes and the availability of the target production unit. The raw attention score calculated solely based on node characteristics and relationships is insufficient to fully reflect the impact of the workpiece's processing stage on collaborative decisions. Therefore, this invention uses a completion scalar calculated from the workpiece's current process number and the total number of processes to adaptively correct the raw attention score based on stage perception. By combining the upstream and downstream positional relationships of neighboring nodes in the process route, attention allocation can be adjusted according to the workpiece's completion level, thereby strengthening the focus on upstream related nodes in the early stages of processing and gradually increasing the influence weight on downstream related nodes in the later stages. For the target workpiece node, the neighboring production unit nodes are marked as upstream or downstream based on their position on the process route relative to the current process. If it is a downstream node, the attenuation coefficient is calculated using a negative exponential function with the difference between 1 and the completion scalar as the exponent. If it is an upstream node, the attenuation coefficient is calculated using a negative exponential function with the completion scalar as the exponent. The original attention score is multiplied by the corresponding attenuation coefficient to obtain the corrected attention score, and the corrected score is normalized to obtain the attention coefficient. The target node itself is also considered as one of its neighboring nodes, and no asymmetric attenuation correction is applied.

[0018] For each target node, the feature vectors of all neighboring nodes are multiplied by their corresponding attention coefficients, the weighted feature vectors are summed, and the summation result is passed through an activation function such as ELU to obtain the aggregated features of the target node.

[0019] More specifically, the step of correcting the original attention score using an asymmetric decay function adjusted by the completion scalar includes: When the neighboring node is the upstream process node of the target workpiece node, the process is determined by the completion scalar p and the upstream attenuation coefficient. Determined decay function Make corrections; When the neighboring node is a downstream process node of the target workpiece node, the process is determined by the completion scalar p and the downstream attenuation coefficient. Determined decay function Make corrections.

[0020] Calculate the initial attention score between the target workpiece node and its neighboring nodes. Adjust this score based on the type of the neighboring nodes. Assume the completion scalar of the target workpiece is p, for example, p = 0.7, indicating that the workpiece is 70% complete.

[0021] In the first case, if the neighboring node is an upstream process node of the workpiece (i.e., a completed process), its influence should decrease as the workpiece processing progresses. In this case, an upstream decay function is used, multiplying the original attention score by... Since p is large, the correction factor will reduce the attention weight of the upstream node. In the second case, if the neighboring node is a downstream process node (i.e., the process to be processed), its influence should increase as the workpiece approaches that process. In this case, a downstream attenuation function is used, multiplying the original attention score by... Because 1-p is small, the correction factor has minimal attenuation of attention weight on downstream nodes, and even almost no attenuation when p approaches 1, thereby strengthening the influence of upcoming processes, such as... Figure 4 .

[0022] In an optional embodiment, the step of using the attention coefficient to perform a weighted summation of the features of neighboring nodes to generate aggregated features of the target node includes: After applying a learnable linear transformation to the features of each neighboring node, the attention coefficients are used for weighted summation.

[0023] For each neighboring node of the target node, the current feature vector undergoes a shared linear transformation layer, which is a weight matrix W. For example, the 64-dimensional feature vector of a neighboring node... It will be multiplied by a weight matrix W with dimensions of 128×64 to obtain a 128-dimensional high-order feature representation. Map the features of all neighboring nodes to the same feature space.

[0024] The model has calculated the attention coefficient of the target node for each of its neighboring nodes, and this coefficient has been corrected using an asymmetric decay function. Assume the attention coefficient of target node i for neighboring node j is... New features of target node i By transforming the features of all neighboring nodes Use the corresponding attention coefficient We obtain the result by weighted summation, i.e. Through this process, the target node can selectively gather important information from its neighbors, thereby updating its own feature representation.

[0025] S4, input the aggregated features of all nodes into the decoder for decoding decision. The decoding decision includes: calculating the execution probability of each workpiece on the available production unit through the resource selection module; and when multiple workpieces compete for the same production unit, the process sequencing module calculates the sequencing value based on the aggregated features of the workpiece and the remaining process time, determines the execution order, and outputs production collaborative control instructions.

[0026] The resource selection module concatenates the aggregated features of a workpiece with the aggregated features of each available production unit, inputs this into a multilayer perceptron network, and outputs a matching score. Then, it applies the Softmax function to the matching scores of all available units to calculate the execution probability. The process sequencing module concatenates the aggregated features of each workpiece competing for the same unit with the remaining total process time, inputs this into another multilayer perceptron network, and outputs a ranking value. The execution priority is determined according to the ranking value from high to low. The output collaborative control instruction format is to assign the specified workpiece to the specified production unit to execute the specified process at the specified time.

[0027] In order to select the most suitable processing equipment for the workpieces to be scheduled, in an optional embodiment, the calculation of the execution probability of each workpiece on the available production unit by the resource selection module includes: The aggregated features of the workpiece to be scheduled are concatenated with the aggregated features of each available production unit to obtain a candidate feature vector; The candidate feature vectors are input into a multilayer perceptron with shared parameters, and a scalar score is output. The scalar scores of all available production units are normalized using the Softmax function to obtain the execution probability distribution of the workpiece on each available production unit.

[0028] Specifically, suppose there is an artifact to be scheduled. The current process can be carried out in the production unit. and Upward processing. Extracting the workpiece. Aggregated feature vectors and production units and Aggregated feature vectors and The workpiece features are concatenated with the features of each available device to form a candidate feature vector, for example... and If the feature vector dimension is 128, the concatenated vector dimension is 256.

[0029] The concatenated candidate feature vectors are input one by one into a multilayer perceptron network with shared parameters. This network structure consists of several fully connected layers and activation functions, and its output is a single scalar value representing the matching score between the workpiece and the device combination. For example, the input... A score of 2.5 was obtained. A score of 1.8 was obtained. Shared parameters mean that the exact same network is used for evaluation regardless of the combination of workpiece and equipment, ensuring consistency in the evaluation criteria. All scores obtained, i.e., [2.5, 1.8], are normalized using the Softmax function. The Softmax function converts the original scores into a probability distribution, for example, [0.67, 0.33], representing the workpiece... Assigned to The probability is 67%, allocated to The probability is 33%.

[0030] A shared-parameter multilayer perceptron is a feedforward neural network, typically consisting of an input layer, one or two hidden layers, and an output layer. For example, the input layer receives a concatenated 256-dimensional candidate feature vector, the hidden layers can have 128 neurons using ReLU as the activation function, and the output layer is a single neuron that outputs a scalar score. Because it's a shared-parameter network, all job-to-production-unit pairings are evaluated using the same set of network weights. This network is part of the overall scheduling policy model, and its training set is identical to the main model, consisting entirely of scheduling problem instances. Training is performed using an end-to-end reinforcement learning framework. The policy loss, calculated from the scheduling objective such as minimizing the maximum completion time, updates the parameters of the multilayer perceptron through backpropagation, enabling it to learn to assign higher scores to better job-to-equipment allocation schemes. The input is the concatenated feature vector of the job to be scheduled and the available production units, and the output is the fitness score for that pairing.

[0031] To address the issue of determining which workpiece should be processed first when multiple workpieces are assigned to the same production unit, in an optional embodiment, the process sequencing module calculates a sequencing value based on the workpiece's aggregation characteristics and remaining process time to determine the execution order, including: For each workpiece competing in the same production unit, the sum of the standard processing times of all unfinished operations of the workpiece is calculated as the remaining process time; The aggregated feature vector of the workpiece is concatenated with the remaining process time; The concatenated feature vectors are input into the ranking evaluation network to output ranking values, and the workpiece with the highest ranking value is selected as the priority workpiece to be executed.

[0032] Assuming the workpiece and Everyone is waiting for the production unit. Workpiece There are 3 processes remaining, with standard processing times of 10, 20, and 15 minutes respectively. Therefore, the remaining processing time is 45 minutes. (Workpiece) There are two remaining processes, with standard processing times of 30 and 25 minutes respectively, leaving a total processing time of 55 minutes. These values ​​are extracted as important heuristic information.

[0033] The aggregated feature vector obtained from the graph model for each workpiece is normalized and then concatenated with the calculated scalar value of the remaining process time. The concatenated feature vector not only contains the contextual information learned by the graph model but also incorporates heuristic information about the total number of future tasks. These concatenated feature vectors are then fed into a ranking evaluation network, typically another multilayer perceptron, which outputs a ranking value for each workpiece. For example, The sorting value is 8.2. The sorting value is 6.5. Select the workpiece with the highest sorting value. This will be the next workpiece to be processed in that production unit.

[0034] The ranking evaluation network is a multilayer perceptron with a structure similar to the resource selection module. The network structure can be planned as an input layer matching the concatenated feature dimensions, several hidden layers using the ReLU activation function, and an output layer producing a single ranking value. For example, the input layer has a dimension of 129, the hidden layers have 64 neurons, and the output layer has 1 neuron. This network is trained together with the entire scheduling agent model, using the same set of scheduling problem instances as the training set. The input is a concatenated vector of aggregated feature vectors of jobs competing for the same production unit and the remaining process time; the output is a scalar ranking value representing the execution priority of that job, such as... Figure 5 .

[0035] In the second embodiment, the present invention also proposes a production collaborative control system based on process route, comprising the following modules: The acquisition module is used to construct a graph model, setting production units and workpieces as nodes, process flow and resource dependencies as edges, and acquiring the real-time status and process route characteristics of each node. The calculation module is used to encode and fuse the real-time state and process route features through a cross-attention mechanism to generate the initial features of each node, and to calculate the proportion of completed processes for each workpiece node as a completion scalar. The generation module is used to calculate the original attention score based on the initial features and edge features; and when the target node is a workpiece node, it obtains the attention coefficient by correcting the original attention score using the completion scalar according to the upstream and downstream positions of the neighboring nodes in the process route of the workpiece node; and uses the attention coefficient to perform a weighted summation of the features of the neighboring nodes to generate the aggregated features of the target node. The output module is used to input the aggregated features of all nodes into the decoder for decoding decisions. The decoding decisions include: calculating the execution probability of each workpiece on the available production unit through the resource selection module; and when multiple workpieces compete for the same production unit, the process sequencing module calculates the sequencing value based on the aggregated features of the workpieces and the remaining process time, determines the execution order, and outputs production collaborative control instructions.

[0036] In an optional embodiment, the construction of the graph model includes: The real-time status of the production unit is encoded as a unique hot vector, representing the idle, processing, or fault status. The process route features of the workpiece node are encoded as a numerical vector containing the total number of processes, the current process number, and the standard processing time of the current process.

[0037] In an optional embodiment, the step of encoding and fusing the real-time state and process route features through a cross-attention mechanism to generate initial features for each node, and calculating the proportion of completed processes for each workpiece node as a completion scalar, includes: For each node, the state features of the node are used as the query vector in the cross-attention mechanism, and the process route features are used as the key vector and value vector to generate the fused initial feature vector. The completion scalar is calculated based on the current process number and the total number of processes of the workpiece.

[0038] In an optional embodiment, correcting the original attention score using an asymmetric decay function adjusted by the completion scalar includes: When the neighboring node is the upstream process node of the target workpiece node, the process is determined by the completion scalar p and the upstream attenuation coefficient. Determined decay function Make corrections; When the neighboring node is a downstream process node of the target workpiece node, the process is determined by the completion scalar p and the downstream attenuation coefficient. Determined decay function Make corrections.

[0039] In an optional embodiment, the step of using the attention coefficient to perform a weighted summation of the features of neighboring nodes to generate aggregated features of the target node includes: After applying a learnable linear transformation to the features of each neighboring node, the attention coefficients are used for weighted summation.

[0040] In an optional embodiment, the calculation of the execution probability of each workpiece on the available production unit via the resource selection module includes: The aggregated features of the workpiece to be scheduled are concatenated with the aggregated features of each available production unit to obtain a candidate feature vector; The candidate feature vectors are input into a multilayer perceptron with shared parameters, and a scalar score is output. The scalar scores of all available production units are normalized using the Softmax function to obtain the execution probability distribution of the workpiece on each available production unit.

[0041] In an optional embodiment, the step of determining the execution order by the process sequencing module calculating a sequencing value based on the aggregation characteristics of the workpiece and the remaining process time includes: For each workpiece competing in the same production unit, the sum of the standard processing times of all unfinished operations of the workpiece is calculated as the remaining process time; The aggregated feature vector of the workpiece is concatenated with the remaining process time; The concatenated feature vectors are input into the ranking evaluation network to output ranking values, and the workpiece with the highest ranking value is selected as the priority workpiece to be executed.

[0042] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0043] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A production collaborative control method based on process route, characterized in that, include: Construct a graph model, setting production units and workpieces as nodes, and process flow and resource dependencies as edges, and obtain the real-time status and process route characteristics of each node. The real-time state and process route features are encoded and fused through a cross-attention mechanism to generate initial features for each node, and the proportion of completed processes is calculated for each workpiece node as a completion scalar. The original attention score is calculated based on the initial features and edge features; and when the target node is a workpiece node, the attention coefficient is obtained by correcting the original attention score using the completion scalar based on the upstream and downstream positions of neighboring nodes in the process route of the workpiece node. The attention coefficients are used to perform a weighted summation of the features of neighboring nodes to generate aggregated features of the target node. The aggregated features of all nodes are input into the decoder for decoding decision-making. The decoding decision-making includes: calculating the execution probability of each workpiece on the available production unit through the resource selection module; When multiple workpieces compete for the same production unit, the process sequencing module calculates a sequencing value based on the aggregation characteristics of the workpieces and the remaining process time, determines the execution order, and outputs production coordination control instructions.

2. The method according to claim 1, characterized in that, The construction graph model includes: The real-time status of the production unit is encoded as a unique hot vector, representing the idle, processing, or fault status. The process route features of the workpiece node are encoded as a numerical vector containing the total number of processes, the current process number, and the standard processing time of the current process.

3. The method according to claim 1 or 2, characterized in that, The process involves encoding and fusing the real-time state and process route features through a cross-attention mechanism to generate initial features for each node, and calculating the proportion of completed processes for each workpiece node as a completion scalar, including: For each node, the state features of the node are used as the query vector in the cross-attention mechanism, and the process route features are used as the key vector and value vector to generate the fused initial feature vector. The completion scalar is calculated based on the current process number and the total number of processes of the workpiece.

4. The method according to claim 1 or 2, characterized in that, The step of correcting the original attention score using an asymmetric decay function adjusted by the completion scalar includes: When the neighboring node is the upstream process node of the target workpiece node, the process is determined by the completion scalar p and the upstream attenuation coefficient. Determined decay function Make corrections; When the neighboring node is a downstream process node of the target workpiece node, the process is determined by the completion scalar p and the downstream attenuation coefficient. Determined decay function Make corrections.

5. The method according to claim 1 or 2, characterized in that, The step of using the attention coefficient to perform a weighted summation of the features of neighboring nodes to generate the aggregated features of the target node includes: After applying a learnable linear transformation to the features of each neighboring node, the attention coefficients are used for weighted summation.

6. The method according to claim 1 or 2, characterized in that, The calculation of the execution probability of each workpiece on the available production unit through the resource selection module includes: The aggregated features of the workpiece to be scheduled are concatenated with the aggregated features of each available production unit to obtain a candidate feature vector; The candidate feature vectors are input into a multilayer perceptron with shared parameters, and a scalar score is output. The scalar scores of all available production units are normalized using the Softmax function to obtain the execution probability distribution of the workpiece on each available production unit.

7. The method according to claim 1, characterized in that, The step of determining the execution order by calculating a sorting value based on the aggregation characteristics of the workpiece and the remaining process time by the process sorting module includes: For each workpiece competing in the same production unit, the sum of the standard processing times of all unfinished operations of the workpiece is calculated as the remaining process time; The aggregated feature vector of the workpiece is concatenated with the remaining process time; The concatenated feature vectors are input into the ranking evaluation network to output ranking values, and the workpiece with the highest ranking value is selected as the priority workpiece to be executed.

8. A production collaborative control system based on a process route, characterized in that, Includes the following modules: The acquisition module is used to construct a graph model, setting production units and workpieces as nodes, process flow and resource dependencies as edges, and acquiring the real-time status and process route characteristics of each node. The calculation module is used to encode and fuse the real-time state and process route features through a cross-attention mechanism to generate the initial features of each node, and to calculate the proportion of completed processes for each workpiece node as a completion scalar. The generation module is used to calculate the original attention score based on the initial features and edge features; and when the target node is a workpiece node, the attention coefficient is obtained by correcting the original attention score using the completion scalar according to the upstream and downstream positions of the neighboring nodes in the process route of the workpiece node. The attention coefficients are used to perform a weighted summation of the features of neighboring nodes to generate aggregated features of the target node. The output module is used to input the aggregated features of all nodes into the decoder for decoding decisions. The decoding decisions include: calculating the execution probability of each workpiece on the available production unit through the resource selection module. When multiple workpieces compete for the same production unit, the process sequencing module calculates a sequencing value based on the aggregation characteristics of the workpieces and the remaining process time, determines the execution order, and outputs production coordination control instructions.

9. The system according to claim 8, characterized in that, The construction graph model includes: The real-time status of the production unit is encoded as a unique hot vector, representing the idle, processing, or fault status. The process route features of the workpiece node are encoded as a numerical vector containing the total number of processes, the current process number, and the standard processing time of the current process.

10. The system according to claim 8 or 9, characterized in that, The process involves encoding and fusing the real-time state and process route features through a cross-attention mechanism to generate initial features for each node, and calculating the proportion of completed processes for each workpiece node as a completion scalar, including: For each node, the state features of the node are used as the query vector in the cross-attention mechanism, and the process route features are used as the key vector and value vector to generate the fused initial feature vector. The completion scalar is calculated based on the current process number and the total number of processes of the workpiece.

Images