Production line scheduling optimization method based on real-time data feedback

Abstract

The invention discloses a production line scheduling optimization method based on real-time data feedback, and relates to the technical field of production scheduling, and the method comprises the steps: collecting multi-source heterogeneous data of a production line in real time, carrying out the time sequence alignment and difference normalization, and generating a deviation state vector; based on the scheduling feedback intensity index, constructing a lightweight deep reasoning model, performing dynamic mapping, generating a corresponding scheduling intervention level, performing multi-dimensional disturbance influence modeling on the deviation state vector through a graph neural network, and constructing a disturbance causal map; on the basis of the scheduling intervention level and the disturbance causal atlas, key nodes and propagation paths are identified, rule reasoning is carried out in combination with a scheduling rule, and multiple scheduling adjustment actions are generated; and performing multi-stream coupling disturbance cost prediction on the plurality of scheduling adjustment actions through discrete event simulation, and selecting a candidate scheduling adjustment action with the minimum disturbance cost and the optimal comprehensive benefit as a production line scheduling optimization scheme. According to the method, high-robustness and self-adaptive closed-loop scheduling optimization is achieved.

Description

Technical Field

[0001] This invention relates to the field of production scheduling technology, and in particular to a production line scheduling optimization method based on real-time data feedback. Background Technology

[0002] Real-time data feedback-based production line scheduling optimization technology plays a crucial role in today's intelligent manufacturing and industrial automation fields. By integrating multi-source heterogeneous sensor data with dynamic decision-making mechanisms, it enables rapid perception and response to production disturbances. Existing technologies typically rely on equipment status monitoring, process sequence recording, and queue information collection to construct a real-time execution status dataset. Based on this dataset, deviation analysis and scheduling adjustments are performed to maintain the stability and efficiency of the production line. These methods generally employ sliding window statistics, rule matching mechanisms, or simulation evaluation methods to support the quantitative characterization of deviation behavior and the generation of intervention strategies. They have been widely applied in flexible manufacturing and discrete assembly scenarios.

[0003] In the field of production line scheduling optimization based on real-time data feedback, traditional production line scheduling optimization methods have two limitations: First, the scheduling triggering mechanism mostly relies on a single-dimensional state threshold judgment, lacking a joint characterization of the cumulative effect of deviation state and dynamic evolution trend, making it difficult to accurately depict the stage characteristics of disturbance development; Second, the scheduling action generation process is usually independent of the disturbance propagation mechanism, and no explicit model of causal relationship between processes has been established, making it difficult for adjustment strategies to accurately locate key bottleneck nodes and predict the multi-flow coupling effect. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a production line scheduling optimization method based on real-time data feedback to solve the problems of scheduling triggers lacking joint perception of deviation accumulation and dynamic trends, and scheduling actions not integrating the causal mechanism of process disturbances.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a production line scheduling optimization method based on real-time data feedback. The method includes: real-time acquisition of multi-source heterogeneous data from the production line; time-series alignment and differential normalization to generate a deviation state vector; performing a squared-weighted mapping on the deviation state vector to generate an execution deviation energy value, and calculating the cumulative increment and rate of change within a sliding time window to generate a scheduling feedback intensity index; constructing a lightweight deep inference model based on the scheduling feedback intensity index, performing dynamic mapping to generate corresponding scheduling intervention levels, modeling the multi-dimensional disturbance impact of the deviation state vector using a graph neural network, and constructing a disturbance causal graph; identifying key nodes and propagation paths based on the scheduling intervention level and the disturbance causal graph, and performing rule inference based on scheduling rules to generate multiple scheduling adjustment actions; predicting the multi-flow coupling disturbance cost of multiple scheduling adjustment actions through discrete event simulation, and selecting the candidate scheduling adjustment action with the minimum disturbance cost and optimal overall benefit as the production line scheduling optimization scheme.

[0008] As a preferred embodiment of the production line scheduling optimization method based on real-time data feedback described in this invention, the specific steps for generating the real-time execution status dataset are as follows:

[0009] The production line's multi-source heterogeneous data includes equipment operating status codes, equipment load values, process start time, process completion time, and process queue length;

[0010] Perform time-series alignment on multi-source heterogeneous data from the production line and organize it into a real-time execution status dataset.

[0011] As a preferred embodiment of the production line scheduling optimization method based on real-time data feedback described in this invention, the specific steps for generating the deviation state vector are as follows:

[0012] The mean of the historical execution status dataset is obtained by sliding a time window, and the difference is calculated between it and the real-time execution status dataset to obtain the deviation of the execution status dataset.

[0013] The deviation of the execution state dataset is normalized and a four-dimensional ordered vector is constructed to generate the deviation state vector.

[0014] As a preferred embodiment of the production line scheduling optimization method based on real-time data feedback described in this invention, the specific steps for generating the execution deviation energy value are as follows:

[0015] The deviation state vector is obtained by linearly scaling the deviation state vector using a positive definite diagonal matrix.

[0016] Calculate the sum of squares of each component in the scaled deviation state vector to generate the execution deviation energy value.

[0017] As a preferred embodiment of the production line scheduling optimization method based on real-time data feedback described in this invention, the specific steps for calculating the cumulative increment within the sliding time window are as follows:

[0018] The execution deviation energy values ​​are stored in the order of sampling time to form a time series of execution deviation energy values;

[0019] Extract continuous historical data segments from the time series of execution deviations from energy values ​​to obtain the historical execution deviation energy value sequence;

[0020] Extract the execution deviation energy value at the start of the window from the historical execution deviation energy value sequence, and calculate the sum of the differences between the execution deviation energy value at each sampling time within the sliding time window and the execution deviation energy value at each sampling time to generate the cumulative increment of the execution deviation energy value.

[0021] As a preferred embodiment of the production line scheduling optimization method based on real-time data feedback described in this invention, the specific steps for generating the scheduling feedback intensity index are as follows:

[0022] Calculate the first-order difference mean of the execution deviation energy values ​​at adjacent sampling times in the historical execution deviation energy value sequence, and generate the execution deviation energy value change rate;

[0023] The cumulative increment of execution deviation energy value and the rate of change of execution deviation energy value are linearly combined to generate a scheduling feedback intensity index.

[0024] As a preferred embodiment of the production line scheduling optimization method based on real-time data feedback described in this invention, the specific steps for generating the corresponding scheduling intervention level are as follows:

[0025] The scheduling feedback intensity index is input into a lightweight deep inference model and nonlinear mapping is performed to generate mapping values ​​corresponding to multiple scheduling intervention levels.

[0026] The mapping values ​​corresponding to multiple scheduling intervention levels are subjected to probability normalization to generate a probability distribution of scheduling intervention levels.

[0027] Based on the probability distribution of scheduling intervention levels, the probability values ​​corresponding to each scheduling intervention level are compared, and the one with the largest probability value is selected as the corresponding scheduling intervention level.

[0028] As a preferred embodiment of the production line scheduling optimization method based on real-time data feedback described in this invention, the specific steps for constructing the disturbance causal graph are as follows:

[0029] A process-dependent directed graph is constructed based on the production line process route. State component data is extracted from the deviation state vector and combined into the node initial feature vector.

[0030] The process-dependent directed graph and the initial feature vector of the nodes are input into the graph convolutional network, and multiple rounds of neighborhood information aggregation are performed to output the node perturbation embedding vector.

[0031] Based on the node perturbation embedding vector, the attention score of the process-dependent directed graph is calculated and filtered, and a weighted adjacency matrix is ​​constructed as the perturbation causal graph.

[0032] As a preferred embodiment of the production line scheduling optimization method based on real-time data feedback described in this invention, the specific steps for generating multiple scheduling adjustment actions are as follows:

[0033] The disturbance response process is triggered based on the scheduling intervention level. When the scheduling intervention level is level 1, no scheduling adjustment is triggered. When the scheduling intervention level is level 1 or above, the disturbance response process is initiated.

[0034] In the perturbation causal graph, the propagation path is obtained through depth-first search, and the perturbation impact degree is calculated. Nodes with an impact degree greater than the preset threshold are marked as key nodes.

[0035] The attributes of key nodes are compared one by one with the preconditions in the scheduling rule base. The scheduling rules that meet the conditions are matched, and the scheduling actions corresponding to the matching scheduling rules are extracted to generate multiple scheduling adjustment actions.

[0036] As a preferred embodiment of the production line scheduling optimization method based on real-time data feedback described in this invention, the specific steps of the production line scheduling optimization scheme are as follows:

[0037] A discrete event simulation environment is constructed, and the real-time execution state dataset, multiple scheduling adjustment actions and disturbance causal graph are loaded as the initial configuration. Multi-dimensional coupling and correlation rules are established based on the disturbance causal graph, and the initial event sequence is generated and stored in the event scheduling queue.

[0038] Retrieve events from the event scheduling queue and update the global clock. Perform state updates and drive the state variables of associated nodes to synchronize and link according to the coupling and association rules. Generate subsequent events and insert them into the event scheduling queue. Repeat the process until the simulation ends to obtain the state evolution trajectory of the entire process.

[0039] Based on the entire process state evolution trajectory, the time cost, resource cost, and stability cost of each scheduling adjustment action are calculated as the sum of the products of the corresponding weight coefficients to generate the disturbance cost;

[0040] Based on the entire process state evolution trajectory, the sum of the products of the capacity utilization improvement rate, work-in-process inventory reduction rate and order on-time delivery rate of each scheduling adjustment action with their corresponding weight coefficients is calculated to generate comprehensive benefits;

[0041] Iterate through all scheduling adjustment actions to select the one with the minimum disturbance cost, and then filter the scheduling adjustment actions with the minimum disturbance cost to form a set of scheduling adjustment actions with the minimum disturbance cost.

[0042] The scheduling adjustment action with the largest comprehensive benefit value is selected from the set of scheduling adjustment actions with the minimum disturbance cost as the production line scheduling optimization scheme.

[0043] The beneficial effects of this invention are as follows: by generating a scheduling feedback intensity index, a joint quantitative characterization of the cumulative increment and rate of change of the energy of the deviation state within the sliding time window is achieved, which effectively improves the foresight of the intervention timing and the adaptability of the intensity setting; by constructing a disturbance causal graph and combining discrete event simulation to predict the cost of multi-flow coupled disturbances, the scheduling adjustment action is accurately coordinated in key node positioning, propagation path tracking and multi-dimensional benefit-cost trade-offs, and finally achieves highly robust and adaptive closed-loop scheduling optimization. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of a production line scheduling optimization method based on real-time data feedback.

[0046] Figure 2 A flowchart for generating the scheduling feedback strength index.

[0047] Figure 3 A flowchart illustrating the interaction between the scheduling intervention level and the causal graph of the disturbance.

[0048] Figure 4 A flowchart for generating a production line scheduling optimization plan. Detailed Implementation

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0051] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0052] Reference Figures 1-4 This is one embodiment of the present invention, which provides a production line scheduling optimization method based on real-time data feedback, including the following steps:

[0053] S1: Real-time acquisition of multi-source heterogeneous data from the production line, time-series alignment and differential normalization, and generation of deviation state vector;

[0054] S1.1: The multi-source heterogeneous data of the production line includes equipment operating status codes, equipment load values, process start time, process completion time, and process queue length;

[0055] Specifically, during production line operation, the equipment's operating status code is read in real time through the interface between the built-in sensors and the controller. The operating status code is a discrete code representing the current working mode of the equipment, including four status types: idle, running, paused, and fault. The equipment load value is obtained through the PLC load feedback signal. The equipment load value is a continuous real number representing the current resource utilization rate of the equipment. The process start time is recorded by the process trigger sensor, which is the timestamp of the first workpiece entering the processing area. The process completion time is recorded synchronously by the process completion sensor, which is the timestamp of the last workpiece leaving the processing area. The process queue length is obtained through the process entry buffer counter. The process queue length is a non-negative integer representing the number of work-in-process items currently waiting to enter the process. The equipment operating status code, equipment load value, process start time, process completion time, and process queue length are all collected synchronously with a unified sampling period and time-stamped using timestamps as keys. The equipment operating status code, equipment load value, process start time, process completion time, and process queue length are aggregated into a single data structure to generate multi-source heterogeneous data for the production line.

[0056] S1.2: Perform time-series alignment on multi-source heterogeneous data from the production line and organize it into a real-time execution status dataset;

[0057] Specifically, the timestamps corresponding to equipment operating status codes, equipment load values, process start times, process completion times, and process queue lengths are extracted from the multi-source heterogeneous data of the production line. Using a unified sampling timestamp sequence as the alignment benchmark, the nearest neighbor interpolation method is used to generate aligned equipment operating status codes for the equipment operating status codes, the linear interpolation method is used to generate aligned equipment load values ​​for the equipment load values, and the zero-order hold interpolation method is used to generate aligned process queue lengths for the process queue lengths. The aligned equipment operating status codes, aligned equipment load values, process start times, process completion times, and aligned process queue lengths are sorted according to the process start time to construct a structured time-series sequence indexed by time, generating a real-time execution status dataset.

[0058] It should be noted that the nearest neighbor interpolation method refers to selecting the data value of the existing sampling point closest to the target time point as the interpolation result when interpolating data at the target time point.

[0059] The zero-order hold-before interpolation method refers to selecting the data value of the nearest existing sampling point that is less than or equal to the target time point as the interpolation result, and keeping the data value unchanged between adjacent sampling points.

[0060] S1.3: Obtain the mean of the historical execution status dataset through a sliding time window, and perform a difference operation with the real-time execution status dataset to obtain the deviation of the execution status dataset;

[0061] Specifically, the actual processing time of a process is calculated as the difference between the process completion time and the process start time. The arithmetic mean of each dimension of the historical execution status dataset within the sliding time window is calculated to form a historical mean vector. The difference between the equipment load value and the historical mean of the equipment load value in the real-time execution status dataset, the difference between the actual processing time of a process and the historical mean of the actual processing time of a process, and the difference between the process queue length and the historical mean of the process queue length are combined to form the execution status dataset deviation.

[0062] S1.4: Normalize the deviation of the execution state dataset and construct a four-dimensional ordered vector to generate the deviation state vector.

[0063] Specifically, the deviations of equipment load values, actual processing time of processes, and process queue length are extracted from the deviation dataset of the execution status dataset. The standard deviation of the equipment load values ​​within the sliding time window is calculated, and the ratio of the equipment load value deviation to the standard deviation of the equipment load values ​​is obtained to obtain the normalized equipment load value deviation. The standard deviation of the actual processing time of processes within the sliding time window is calculated, and the quotient of the actual processing time deviation to the standard deviation of the actual processing time is obtained to obtain the normalized actual processing time deviation. The standard deviation of the process queue length within the sliding time window is calculated, and the quotient of the process queue length deviation to the standard deviation of the process queue length is obtained to obtain the normalized process queue length deviation. The deviation of the equipment operation status code remains unchanged and is not normalized. The normalized equipment load value deviation, normalized actual processing time deviation, normalized process queue length deviation, and equipment operation status code deviation are arranged in a fixed order of equipment load, actual processing time, process queue length, and equipment operation status code to construct a four-dimensional ordered vector, generating the deviation status vector.

[0064] S2: Perform a squared weighted mapping on the deviation state vector to generate the execution deviation energy value, and calculate the cumulative increment and rate of change within the sliding time window to generate the scheduling feedback intensity index;

[0065] S2.1: The deviation state vector is linearly scaled using a positive definite diagonal matrix to obtain the scaled deviation state vector;

[0066] Specifically, the preset scaling factors for equipment load values, actual process processing time, process queue length, and equipment status codes are sequentially placed in the first to fourth positions of the main diagonal of a fourth-order diagonal matrix, with off-diagonal elements set to zero, forming a positive definite diagonal matrix. The deviation state vector is then linearly scaled against the positive definite diagonal matrix, so that the normalized deviations in equipment load values, actual process processing time, process queue length, and equipment status codes are weighted and amplified according to their respective scaling factors to reflect the relative importance of each dimension to production stability. The result is the scaled deviation state vector.

[0067] It should be noted that the preset equipment operation status code scaling factor, preset equipment load value scaling factor, preset actual process processing time scaling factor, and preset process queue length scaling factor are constant parameters pre-configured based on the weight of the impact of deviations in each dimension of the production line's historical operation data on the overall stability of the production process. The preset equipment operation status code scaling factor is a real number greater than zero and less than or equal to 1.2, with a typical value of 1.0; the preset equipment load value scaling factor is a real number greater than zero and less than or equal to 1.0, with a typical value of 0.8; the preset actual process processing time scaling factor is a real number greater than zero and less than or equal to 0.9, with a typical value of 0.6; and the preset process queue length scaling factor is a real number greater than zero and less than or equal to 0.8, with a typical value of 0.5.

[0068] S2.2: Calculate the sum of squares of each component in the scaled deviation state vector to generate the execution deviation energy value;

[0069] Specifically, the squares of the scaled equipment load value deviation, the scaled actual processing time deviation, the scaled process queue length deviation, and the integer scaled equipment running status code deviation in the scaled deviation state vector are calculated, and the four squared values ​​are added together to generate the execution deviation energy value.

[0070] S2.3: Store the execution deviation energy values ​​in the order of sampling time to form a time series of execution deviation energy values;

[0071] Specifically, the execution deviation energy value generated at the current sampling time is bound to the timestamp of the current sampling time to form a time-stamped execution deviation energy value data point; the time-stamped execution deviation energy value data point is appended to the end of the execution deviation energy value storage queue; when the number of data points in the execution deviation energy value storage queue exceeds the preset maximum length, the data point stored first in the execution deviation energy value storage queue is deleted; all data points in the execution deviation energy value storage queue are kept in ascending order of timestamp; all time-stamped execution deviation energy value data points in the execution deviation energy value storage queue are organized into a time series list according to the timestamp order to form the time series of execution deviation energy values.

[0072] It should be noted that the maximum length is an integer determined by the ratio of the duration of the sliding time window to the data sampling period, and the value range is greater than or equal to 60 and less than or equal to 3600.

[0073] S2.4: Extract continuous historical data segments from the time series of execution deviation energy values ​​to obtain the historical execution deviation energy value sequence;

[0074] Specifically, the timestamp of the latest data point is extracted from the time series of execution deviation energy values ​​and used as the end time of the sliding time window; the start time of the sliding time window is calculated according to the preset duration of the sliding time window; the time series of execution deviation energy values ​​is traversed, and all execution deviation energy values ​​whose timestamps are greater than or equal to the start time of the sliding time window and less than or equal to the end time of the sliding time window are filtered; the filtered execution deviation energy values ​​are sorted in ascending order by timestamp to obtain the historical execution deviation energy value sequence.

[0075] It should be noted that the duration is a constant parameter pre-configured based on the typical process cycle and time delay tolerance limit of the production line, with a value range of greater than or equal to 30 seconds and less than or equal to 3600 seconds.

[0076] S2.5: Extract the execution deviation energy value at the start of the window from the historical execution deviation energy value sequence, and calculate the sum of the differences between the execution deviation energy value at each sampling time within the sliding time window and the execution deviation energy value at each sampling time, to generate the cumulative increment of the execution deviation energy value;

[0077] Specifically, the execution deviation energy value with a timestamp equal to the start time of the sliding time window is selected from the historical execution deviation energy value sequence and used as the execution deviation energy value at the start time of the window; all execution deviation energy values ​​with timestamps greater than the start time of the sliding time window and less than or equal to the end time of the sliding time window are traversed in the historical execution deviation energy value sequence; for each traversed execution deviation energy value, the difference between the execution deviation energy value at each sampling time within the sliding time window and the execution deviation energy value at the start time of the window is calculated; all differences are summed to generate the cumulative increment of the execution deviation energy value.

[0078] S2.6: Calculate the first-order difference mean of the execution deviation energy values ​​at adjacent sampling times in the historical execution deviation energy value sequence, and generate the execution deviation energy value change rate;

[0079] Specifically, all execution deviation energy values ​​are extracted from the historical execution deviation energy value sequence in ascending order of timestamps to obtain an ordered execution deviation energy value list; each adjacent execution deviation energy value in the ordered execution deviation energy value list is traversed, and the difference between the next execution deviation energy value and the previous execution deviation energy value is calculated to obtain the adjacent difference; all adjacent differences are summed to obtain the total adjacent difference; the total number of adjacent differences is counted; the quotient of the sum of adjacent differences and the total number of adjacent differences is calculated to generate the execution deviation energy value change rate.

[0080] The expression for calculating the rate of change of the deviation energy value is:

[0081] ;

[0082] in, This indicates the rate of change of the deviation energy value. This represents the total number of execution deviations from energy values ​​in the historical execution deviation energy value sequence; This indicates the first value in the historical execution deviation energy value sequence. The execution deviation energy value at each sampling time; This indicates the first value in the historical execution deviation energy value sequence. +1 sampling time step execution deviation energy value; This represents the total number of adjacent differences.

[0083] S2.7: Linearly combine the cumulative increment of execution deviation energy value with the rate of change of execution deviation energy value to generate a scheduling feedback intensity index.

[0084] Specifically, the preset cumulative increment weight coefficient and the preset rate of change weight coefficient are read from the initialization configuration data area; the product of the preset cumulative increment weight coefficient and the cumulative increment of the execution deviation energy value is calculated; the product of the preset rate of change weight coefficient and the rate of change of the execution deviation energy value is calculated; and the two products are summed to generate the scheduling feedback intensity index.

[0085] It should be noted that the initialization configuration data area refers to a parameter storage area that is pre-divided and permanently stored in a non-volatile storage medium, used to store all the preset constant parameters required for the scheduling optimization process.

[0086] The cumulative increment weight coefficient is a constant parameter calibrated based on the contribution of the cumulative increment of the execution deviation energy value in historical scheduling intervention to the deviation suppression effect. Its value range is a real number greater than zero and less than or equal to 1.0, with a typical value of 0.6.

[0087] The rate of change weighting coefficient is a constant parameter calibrated based on the contribution of the rate of change of deviation energy value in historical scheduling intervention to the accuracy of deviation trend prediction. Its value range is a real number greater than zero and less than or equal to 1.0, with a typical value of 0.4.

[0088] S3: Based on the scheduling feedback intensity index, a lightweight deep inference model is constructed, dynamic mapping is performed, corresponding scheduling intervention levels are generated, and multi-dimensional disturbance impact modeling is performed on the deviation state vector through graph neural network, and a disturbance causal graph is constructed.

[0089] S3.1: The network structure of a lightweight deep inference model includes an input layer, a first hidden layer, a second hidden layer, and an output layer;

[0090] Specifically, the lightweight deep inference model includes an input layer, a first hidden layer, a second hidden layer, and an output layer. The first hidden layer contains thirty-two neurons and is connected to the input layer via a fully connected method. The second hidden layer contains sixteen neurons and is connected to the first hidden layer via a fully connected method. The output layer contains four neurons and is connected to the second hidden layer via a fully connected method.

[0091] S3.2: Configure the input layer to the first hidden layer, configure the weight parameters and bias parameters, and set the non-linear rectification function; configure the first hidden layer to the second hidden layer, configure the weight parameters and bias parameters, and set the ReLU activation function; configure the second hidden layer to the output layer, configure the weight parameters and bias parameters, and set the Softmax normalization function to complete the construction of the lightweight deep inference model.

[0092] Specifically, the weight and bias parameters from the input layer to the first hidden layer are loaded from the initialization configuration data area, and the nonlinear rectification function from the input layer to the first hidden layer is set; the weight and bias parameters from the first hidden layer to the second hidden layer are loaded from the initialization configuration data area, and the ReLU activation function from the first hidden layer to the second hidden layer is set; the weight and bias parameters from the second hidden layer to the output layer are loaded from the initialization configuration data area, and the Softmax normalization function from the second hidden layer to the output layer is set, thus completing the construction of the lightweight deep inference model.

[0093] S3.3: Input the scheduling feedback intensity index into the lightweight deep inference model, perform nonlinear mapping processing, and generate mapping values ​​corresponding to multiple scheduling intervention levels;

[0094] Specifically, the scheduling feedback intensity index is input into the input layer of the lightweight deep inference model. The input layer performs a linear transformation on the scheduling feedback intensity index and the weight parameters from the input layer to the first hidden layer, and adds a bias parameter to obtain the input value of the first hidden layer. The first hidden layer applies a nonlinear rectified function to the first hidden layer input value to generate a first hidden layer feature vector. The first hidden layer feature vector is input into the second hidden layer. The second hidden layer performs a linear transformation on the first hidden layer feature vector and the weight parameters from the first hidden layer to the second hidden layer, and adds a bias parameter to obtain the input value of the second hidden layer. The second hidden layer applies a ReLU activation function to the second hidden layer input value to generate a second hidden layer feature vector. The second hidden layer feature vector is input into the output layer. The output layer performs a fully connected mapping on the second hidden layer feature vector and the weight parameters from the second hidden layer to the output layer, and adds a bias parameter to generate mapping values ​​corresponding to multiple scheduling intervention levels.

[0095] S3.4: Perform probability normalization on the mapping values ​​corresponding to multiple scheduling intervention levels to generate a probability distribution of scheduling intervention levels;

[0096] Specifically, the mapping values ​​corresponding to multiple scheduling intervention levels are exponentially calculated to obtain the exponential value corresponding to each scheduling intervention level; the exponential values ​​corresponding to all scheduling intervention levels are summed to obtain the total exponential value; the ratio of the exponential value corresponding to each scheduling intervention level to the total exponential value is calculated to obtain the probability value corresponding to each scheduling intervention level; the probability values ​​corresponding to all scheduling intervention levels form the probability distribution of the scheduling intervention level.

[0097] S3.5: Based on the probability distribution of the scheduling intervention level, compare the probability values ​​corresponding to each scheduling intervention level, and select the one with the largest probability value as the corresponding scheduling intervention level.

[0098] Specifically, iterate through the probability distribution of all scheduling intervention levels and compare the probability values ​​of adjacent scheduling intervention levels. Keep the larger one and continue to compare it with the probability value of the next scheduling intervention level until all scheduling intervention levels have been compared. The scheduling intervention level corresponding to the largest probability value is then taken as the corresponding scheduling intervention level.

[0099] S3.6: Construct a process-dependent directed graph based on the production line process route, extract state component data from the deviation state vector, and combine them into the node initial feature vector;

[0100] Specifically, a process-dependent directed graph is constructed based on the production line process route. Each process and each piece of equipment is defined as a node in the graph neural network. The processing order between processes, the processing binding relationship between equipment and processes, and the resource competition relationship caused by shared equipment are defined as directed edges. Relevant components are extracted from the deviation state vector, including the deviation of equipment load value, deviation of actual processing time of process, deviation of process queue length, and deviation of equipment running status code corresponding to each node. The extracted relevant components are combined in order to generate the initial feature vector of the corresponding node.

[0101] S3.7: Input the process-dependent directed graph and the initial feature vector of the nodes into the graph convolutional network, perform multiple rounds of neighborhood information aggregation, and output the node perturbation embedding vector;

[0102] Specifically, the process-dependent directed graph and the initial feature vectors of the nodes are input into the graph convolutional network. The aggregation weight matrices for the first, second, and third rounds are read from the initialization configuration data area. In the first round, the arithmetic mean of the initial feature vectors of the direct successor nodes of each node in the process-dependent directed graph is calculated. The arithmetic mean is concatenated with the initial feature vectors of the nodes to obtain the concatenated feature vector. The concatenated feature vector is linearly transformed with the aggregation weight matrix of the first round, and the result of the linear transformation is nonlinearly rectified to generate the updated feature vector of the first round. The second and third rounds use the updated feature vectors of the first and second rounds as inputs, respectively, and repeat the above neighborhood aggregation, concatenation, linear transformation, and nonlinear rectification operations. The updated feature vector generated in the third round is used as the node perturbation embedding vector.

[0103] S3.8: Based on the node perturbation embedding vector, calculate and filter the attention score in the process-dependent directed graph, and construct a weighted adjacency matrix as the perturbation causal graph.

[0104] Specifically, based on the node perturbation embedding vector, all ordered node pairs in the process-dependent directed graph are traversed. For each ordered node pair, the dot product of the node perturbation embedding vector of the source node and the node perturbation embedding vector of the target node is calculated to generate the attention score of the ordered node pair. The attention scores of all ordered node pairs are compared with the attention threshold, and ordered node pairs with attention scores greater than the attention threshold are selected. A zero matrix with the same number of nodes as the process-dependent directed graph is initialized as a weighted adjacency matrix. For each selected ordered node pair, the attention score of the ordered node pair is filled into the corresponding source node row and target node column in the weighted adjacency matrix. The filled weighted adjacency matrix is ​​used as the perturbation causal graph.

[0105] It should be noted that the attention threshold is set according to the requirements for sensitivity to disturbance propagation, and the value is usually a real number between 0.5 and 0.9, for example, determined by statistical analysis of attention scores of effective causal associations in historical disturbance events.

[0106] S4: Based on the scheduling intervention level and disturbance causal graph, identify key nodes and propagation paths, and combine scheduling rules to perform rule reasoning to generate multiple scheduling adjustment actions;

[0107] S4.1: The disturbance response process is triggered based on the scheduling intervention level. When the scheduling intervention level is level 1, no scheduling adjustment is triggered. When the scheduling intervention level is level 1 or above, the disturbance response process is initiated.

[0108] Specifically, the system reads the numerical value of the scheduling intervention level and determines whether the scheduling intervention level is equal to level one. If the scheduling intervention level is equal to level one, the current scheduling adjustment process is terminated and no scheduling adjustment action is generated. If the scheduling intervention level is equal to level two, level three, or level four, the disturbance response process is initiated, and the system enters the critical node identification and propagation path analysis stage.

[0109] S4.2: In the perturbation causal graph, the propagation path is obtained through depth-first search, and the perturbation impact degree is calculated. Nodes with an impact degree greater than the preset threshold are marked as key nodes.

[0110] Specifically, in the perturbation causal graph, a depth-first search is performed starting from the perturbation source node, traversing all reachable nodes and recording the directed propagation paths from the perturbation source node to each reachable node; for each node on each propagation path, the magnitude of the deviation state vector corresponding to each node and the maximum value of the attention scores of all incoming edges of each node are extracted, and the product of the magnitude of the deviation state vector and the maximum value of the attention scores of the incoming edges is calculated to generate the perturbation impact degree of each node; the perturbation impact degree of each node is compared with the impact threshold, and nodes with perturbation impact degrees greater than the impact threshold are marked as key nodes.

[0111] It should be noted that the impact threshold is set according to the requirements for sensitivity to disturbance propagation, and the value is usually a real number between 0.5 and 0.9, for example, determined by statistical analysis of the impact of disturbances through effective causal relationships in historical disturbance events.

[0112] S4.3: Compare the attributes of key nodes with the preconditions in the scheduling rule base one by one, match the scheduling rules that meet the conditions, extract the scheduling actions corresponding to the matching scheduling rules, and generate multiple scheduling adjustment actions.

[0113] Specifically, the process involves iterating through each scheduling rule in the scheduling rule base and extracting the preconditions for the current scheduling rule; obtaining the attributes of key nodes, including key node type, disturbance impact value, and current production status; comparing the attributes of key nodes with the preconditions of the current scheduling rule item by item to determine whether the key node type matches the key node type specified in the preconditions, whether the disturbance impact value is within the disturbance impact range specified in the preconditions, and whether the current production status meets the current production status specified in the preconditions; if all comparison items are satisfied, the current scheduling rule is determined to be a matching scheduling rule; extracting the scheduling actions corresponding to the matching scheduling rule; repeating the above iteration, extraction, comparison, determination, and extraction operations until all scheduling rules in the scheduling rule base have been processed; and aggregating all extracted scheduling actions to generate multiple scheduling adjustment actions.

[0114] It should be noted that the scheduling rule base refers to a pre-configured collection of multiple scheduling rules. Each scheduling rule contains two parts: preconditions and scheduling actions. The preconditions describe the production state characteristics that need to be met to trigger the scheduling rule, including the type of critical node, the range of disturbance impact, and the current production state. The scheduling actions describe the specific operation instructions that should be executed when the preconditions are met, including process rescheduling instructions, equipment switching instructions, or buffer adjustment instructions. The scheduling rules in the scheduling rule base are stored in a structured manner and support reading and matching conditions one by one.

[0115] S5: Through discrete event simulation, multi-flow coupling disturbance cost prediction is performed on multiple scheduling adjustment actions. The candidate scheduling adjustment action with the minimum disturbance cost and the best overall benefit is selected as the production line scheduling optimization scheme.

[0116] S5.1: Construct a discrete event simulation environment, load the real-time execution state dataset, multiple scheduling adjustment actions and disturbance causal graph as the initial configuration, establish multi-dimensional coupling association rules based on the disturbance causal graph, generate the initial event sequence and store it in the event scheduling queue;

[0117] Specifically, the event types in the discrete event simulation environment are defined as process start events, process completion events, and equipment state change events. State variables are defined as equipment load values, process queue length, and actual process processing time. The real-time execution state dataset, multiple scheduling adjustment actions, and a disturbance causal graph are loaded as the initial configuration of the discrete event simulation environment. Based on the causal correlation strength between nodes in the disturbance causal graph, multi-dimensional coupling association rules are established between equipment load values, process queue length, and actual process processing time. These multi-dimensional coupling association rules are used to update the state variables of associated nodes during event processing. An initial event sequence is generated based on the real-time execution state dataset and multiple scheduling adjustment actions in the initial configuration. The initial event sequence includes process start events, process completion events, and equipment state change events ordered by occurrence time. The initial event sequence is stored in the event scheduling queue.

[0118] S5.2: Retrieve events from the event scheduling queue and update the global clock, perform state updates and drive the state variables of associated nodes to synchronize and link according to the coupling association rules, generate subsequent events and insert them into the event scheduling queue, repeat the execution until the simulation ends, and obtain the state evolution trajectory of the whole process.

[0119] Specifically, the event with the smallest timestamp is retrieved from the event scheduling queue, and the global clock in the discrete event simulation environment is updated to the timestamp of the event with the smallest timestamp. Based on the type of the event with the smallest timestamp and the associated process or equipment identifier, the corresponding equipment load value, process queue length, or actual process processing time is updated. Through multi-dimensional coupling association rules, other nodes with causal relationships to the currently updated state variables are found, and the equipment load value, process queue length, or actual process processing time of the associated nodes is updated synchronously. Based on the updated state variables and preset production logic, it is determined whether a new process start event, process completion event, or equipment state change event has been triggered. If triggered, a corresponding new event is generated, and the occurrence timestamp of the new event is calculated according to the production logic. The new event is inserted into the event scheduling queue. The operation of retrieving events from the event scheduling queue and inserting new events is repeated until the event scheduling queue is empty. The sequence of equipment load values, process queue lengths, and actual process processing times recorded in chronological order during the simulation for each scheduling adjustment action is obtained as the entire process state evolution trajectory.

[0120] It should be noted that the multi-dimensional coupling association rule is a mapping relationship established based on the causal association strength between nodes in the perturbation causal graph. It is used to realize the linkage update between three state variables, namely, equipment load value, process queue length and actual process processing time, in the discrete event simulation process. The multi-dimensional coupling association rule takes the currently updated state variable as input, determines the affected associated nodes based on the directed edges from the node where the state variable is located to other nodes in the perturbation causal graph and the corresponding attention scores, and calculates the adjustment amount of the state variable of each associated node according to the attention scores and the propagation function.

[0121] Production logic refers to a set of deterministic rules used to describe the operation of a production line in a discrete event simulation environment. It is based on the physical constraints and pre-set process flow of the actual manufacturing process and includes: a process can only start after the preceding process is completed, equipment can only process one process at a time, the process processing time is determined by the process standard, and the queue follows the first-in, first-out principle.

[0122] S5.3: Based on the entire process state evolution trajectory, calculate the sum of the products of the time cost, resource cost, and stability cost of each scheduling adjustment action with the corresponding weight coefficients to generate the disturbance cost;

[0123] Specifically, the sequence of process completion times corresponding to each scheduling adjustment action is extracted from the entire process state evolution trajectory. Combined with the average process completion times in the historical execution state dataset within the sliding time window, the average absolute deviation of the process completion time from the historical average process completion time is calculated as the time cost. The sequence of equipment load values ​​corresponding to each scheduling adjustment action is extracted from the entire process state evolution trajectory. The product of the cumulative duration of the equipment load value exceeding the rated load threshold and the overload magnitude is calculated as the resource cost. The sequence of process queue lengths corresponding to each scheduling adjustment action is extracted from the entire process state evolution trajectory. The standard deviation of the process queue length is calculated as the stability cost. The product of time cost and the first weight coefficient, the product of resource cost and the second weight coefficient, and the product of stability cost and the third weight coefficient are calculated. The products of time cost and the first weight coefficient, resource cost and the second weight coefficient, and stability cost and the third weight coefficient are summed to generate the disturbance cost.

[0124] It should be noted that the rated load threshold is preset according to the specifications provided by the equipment manufacturer and the safe operation requirements of the production line, and the value range is 80% to 95% of the maximum allowable load of the equipment.

[0125] S5.4: Based on the entire process state evolution trajectory, calculate the sum of the products of the capacity utilization improvement rate, work-in-process inventory reduction rate, and order on-time delivery rate of each scheduling adjustment action with their corresponding weight coefficients to generate comprehensive benefits;

[0126] Specifically, the equipment load value sequence corresponding to each scheduling adjustment action is extracted from the entire process state evolution trajectory. The ratio of the cumulative time during which the equipment load value is below the rated load threshold to the total simulation time is calculated as the capacity utilization rate improvement rate. The process queue length sequence corresponding to each scheduling adjustment action is extracted. The ratio of the difference between the total work-in-process inventory under the current scheduling adjustment action and the total work-in-process inventory under the benchmark scheduling scheme to the total work-in-process inventory under the benchmark scheduling scheme is calculated as the work-in-process inventory reduction rate. The process completion time sequence corresponding to each scheduling adjustment action is extracted. The ratio of the number of orders whose actual completion time is no later than the order's required delivery time to the total number of orders is calculated as the order on-time delivery rate. The product of the capacity utilization rate improvement rate and the fourth weight coefficient, the product of the work-in-process inventory reduction rate and the fifth weight coefficient, and the product of the order on-time delivery rate and the sixth weight coefficient are calculated. The above three products are summed to generate the comprehensive benefit.

[0127] S5.5: Traverse all scheduling adjustment actions to select the minimum disturbance cost, filter the scheduling adjustment actions with the minimum disturbance cost, and form a set of scheduling adjustment actions with the minimum disturbance cost;

[0128] Specifically, iterate through all the disturbance costs corresponding to the scheduling adjustment actions, and initialize the minimum disturbance cost as the disturbance cost of the first scheduling adjustment action. Then, compare the disturbance cost of each subsequent scheduling adjustment action with the current minimum disturbance cost. If the disturbance cost of the current scheduling adjustment action is less than the minimum disturbance cost, update the minimum disturbance cost to the disturbance cost of the current scheduling adjustment action. After completing the traversal of all scheduling adjustment actions, iterate through all scheduling adjustment actions again and determine whether the disturbance cost of each scheduling adjustment action is equal to the minimum disturbance cost. If they are equal, add the scheduling adjustment action to the set of scheduling adjustment actions with the minimum disturbance cost. After the traversal is completed, the set of scheduling adjustment actions with the minimum disturbance cost is formed.

[0129] S5.6: Select the scheduling adjustment action with the largest comprehensive benefit value from the set of scheduling adjustment actions with the minimum disturbance cost, and use it as the production line scheduling optimization scheme.

[0130] Specifically, iterate through all scheduling adjustment actions in the set of minimum disturbance cost scheduling adjustment actions, and obtain the comprehensive benefit value corresponding to each scheduling adjustment action; initialize the maximum comprehensive benefit value to the comprehensive benefit value of the first scheduling adjustment action in the set of minimum disturbance cost scheduling adjustment actions; compare the comprehensive benefit value of each subsequent scheduling adjustment action with the current maximum comprehensive benefit value. If the comprehensive benefit value of the current scheduling adjustment action is greater than the maximum comprehensive benefit value, then update the maximum comprehensive benefit value to the comprehensive benefit value of the current scheduling adjustment action; after completing the traversal of all scheduling adjustment actions, iterate through the set of minimum disturbance cost scheduling adjustment actions again, and determine whether the comprehensive benefit value of each scheduling adjustment action is equal to the maximum comprehensive benefit value. If they are equal, then the corresponding scheduling adjustment action is taken as the production line scheduling optimization scheme and the traversal is terminated.

[0131] In summary, this invention achieves a joint quantitative characterization of the cumulative increment and rate of change of energy in the deviated state within the sliding time window by generating a scheduling feedback intensity index, effectively improving the foresight of intervention timing and the adaptability of intensity setting; and by constructing a disturbance causal graph and combining discrete event simulation to predict the cost of multi-flow coupled disturbances, it achieves precise coordination of scheduling adjustment actions in key node positioning, propagation path tracking, and multi-dimensional benefit-cost trade-offs, ultimately achieving highly robust and adaptive closed-loop scheduling optimization.

[0132] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A production line scheduling optimization method based on real-time data feedback, characterized in that: include, Real-time acquisition of multi-source heterogeneous data from the production line, time-series alignment and differential normalization, and generation of deviation state vectors; The deviation state vector is squared and weighted to generate the execution deviation energy value. The cumulative increment and rate of change are calculated within the sliding time window to generate the scheduling feedback intensity index. Based on the scheduling feedback intensity index, a lightweight deep inference model is constructed, dynamic mapping is performed, and corresponding scheduling intervention levels are generated. A graph neural network is used to model the multi-dimensional disturbance impact of the deviation state vector and construct a disturbance causal graph. Based on the scheduling intervention level and disturbance causal graph, key nodes and propagation paths are identified, and rule reasoning is performed in conjunction with scheduling rules to generate multiple scheduling adjustment actions; By using discrete event simulation, the multi-flow coupling disturbance cost of multiple scheduling adjustment actions is predicted. The candidate scheduling adjustment action with the minimum disturbance cost and the best overall benefit is selected as the production line scheduling optimization scheme.

2. The production line scheduling optimization method based on real-time data feedback as described in claim 1, characterized in that: The multi-source heterogeneous data of the production line includes equipment operating status codes, equipment load values, process start time, process completion time, and process queue length; Perform time-series alignment on multi-source heterogeneous data from the production line and organize it into a real-time execution status dataset.

3. The production line scheduling optimization method based on real-time data feedback as described in claim 2, characterized in that: The specific steps for generating the deviation state vector are as follows: The mean of the historical execution status dataset is obtained by sliding a time window, and the difference is calculated between it and the real-time execution status dataset to obtain the deviation of the execution status dataset. The deviation of the execution state dataset is normalized and a four-dimensional ordered vector is constructed to generate the deviation state vector.

4. The production line scheduling optimization method based on real-time data feedback as described in claim 3, characterized in that: The specific steps for generating the execution deviation energy value are as follows: The deviation state vector is obtained by linearly scaling the deviation state vector using a positive definite diagonal matrix. Calculate the sum of squares of each component in the scaled deviation state vector to generate the execution deviation energy value.

5. The production line scheduling optimization method based on real-time data feedback as described in claim 4, characterized in that: The specific steps for calculating the cumulative increment within the sliding time window are as follows: The execution deviation energy values ​​are stored in the order of sampling time to form a time series of execution deviation energy values; Extract continuous historical data segments from the time series of execution deviations from energy values ​​to obtain the historical execution deviation energy value sequence; Extract the execution deviation energy value at the start of the window from the historical execution deviation energy value sequence, and calculate the sum of the differences between the execution deviation energy value at each sampling time within the sliding time window and the execution deviation energy value at each sampling time, to generate the cumulative increment of the execution deviation energy value.

6. The production line scheduling optimization method based on real-time data feedback as described in claim 4 or 5, characterized in that: The specific steps for generating the scheduling feedback strength index are as follows: Calculate the first-order difference mean of the execution deviation energy values ​​at adjacent sampling times in the historical execution deviation energy value sequence, and generate the execution deviation energy value change rate; The cumulative increment of execution deviation energy value and the rate of change of execution deviation energy value are linearly combined to generate a scheduling feedback intensity index.

7. The production line scheduling optimization method based on real-time data feedback as described in claim 6, characterized in that: The specific steps for generating the corresponding scheduling intervention level are as follows: The scheduling feedback intensity index is input into a lightweight deep inference model and nonlinear mapping is performed to generate mapping values ​​corresponding to multiple scheduling intervention levels. The mapping values ​​corresponding to multiple scheduling intervention levels are subjected to probability normalization to generate a probability distribution of scheduling intervention levels. Based on the probability distribution of scheduling intervention levels, the probability values ​​corresponding to each scheduling intervention level are compared, and the one with the largest probability value is selected as the corresponding scheduling intervention level.

8. The production line scheduling optimization method based on real-time data feedback as described in claim 1 or 3, characterized in that: The specific steps for constructing the perturbation causal graph are as follows: A process-dependent directed graph is constructed based on the production line process route. State component data is extracted from the deviation state vector and combined into the node initial feature vector. The process-dependent directed graph and the initial feature vector of the nodes are input into the graph convolutional network, and multiple rounds of neighborhood information aggregation are performed to output the node perturbation embedding vector. Based on the node perturbation embedding vector, the attention score of the process-dependent directed graph is calculated and filtered, and a weighted adjacency matrix is ​​constructed as the perturbation causal graph.

9. The production line scheduling optimization method based on real-time data feedback as described in claim 8, characterized in that: The specific steps for generating multiple scheduling adjustment actions are as follows. The disturbance response process is triggered based on the scheduling intervention level. When the scheduling intervention level is level 1, no scheduling adjustment is triggered. When the scheduling intervention level is level 1 or above, the disturbance response process is initiated. In the perturbation causal graph, the propagation path is obtained through depth-first search, and the perturbation impact degree is calculated. Nodes with an impact degree greater than the preset threshold are marked as key nodes. The attributes of key nodes are compared one by one with the preconditions in the scheduling rule base. The scheduling rules that meet the conditions are matched, and the scheduling actions corresponding to the matching scheduling rules are extracted to generate multiple scheduling adjustment actions.

10. The production line scheduling optimization method based on real-time data feedback as described in claim 9, characterized in that: The specific steps of this production line scheduling optimization scheme are as follows. A discrete event simulation environment is constructed, and the real-time execution state dataset, multiple scheduling adjustment actions and disturbance causal graph are loaded as the initial configuration. Multi-dimensional coupling and correlation rules are established based on the disturbance causal graph, and the initial event sequence is generated and stored in the event scheduling queue. Retrieve events from the event scheduling queue and update the global clock. Perform state updates and drive the state variables of associated nodes to synchronize and link according to the coupling and association rules. Generate subsequent events and insert them into the event scheduling queue. Repeat the process until the simulation ends to obtain the state evolution trajectory of the entire process. Based on the entire process state evolution trajectory, the time cost, resource cost, and stability cost of each scheduling adjustment action are calculated as the sum of the products of the corresponding weight coefficients to generate the disturbance cost; Based on the entire process state evolution trajectory, the sum of the products of the capacity utilization improvement rate, work-in-process inventory reduction rate and order on-time delivery rate of each scheduling adjustment action with their corresponding weight coefficients is calculated to generate comprehensive benefits; Iterate through all scheduling adjustment actions to select the one with the minimum disturbance cost, and then filter the scheduling adjustment actions with the minimum disturbance cost to form a set of scheduling adjustment actions with the minimum disturbance cost. The scheduling adjustment action with the largest comprehensive benefit value is selected from the set of scheduling adjustment actions with the minimum disturbance cost as the production line scheduling optimization scheme.

Images