An order-driven intelligent scheduling system based on artificial intelligence

Abstract

This invention discloses an AI-driven intelligent scheduling system for order processing. The system includes modules for data acquisition and graph construction, pressure propagation rate calculation, implicit bottleneck drift prediction, disturbance propagation tree generation, minimum disturbance increment rearrangement solution, and solution issuance. The data acquisition module collects production line status data in real time to construct a spatiotemporal graph. The pressure propagation rate module calculates the probability of node congestion propagation using a spatiotemporal graph convolutional network. The drift prediction module marks predicted drift nodes and generates disturbance signals when pressure exceeds a threshold and downstream buffer decreases. The disturbance tree generation module extracts affected processes based on process dependencies and resource conflicts to construct a disturbance tree. The rearrangement solution module includes a strategy selector and multiple algorithms, dynamically selecting an algorithm based on disturbance characteristics to solve the moving job sequence and issue instructions. This invention achieves proactive bottleneck drift prediction and adaptive local rearrangement, reducing computational overhead and improving stability and response speed.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and artificial intelligence technology, specifically an order-driven intelligent scheduling system based on artificial intelligence. Background Technology

[0002] As the manufacturing industry shifts towards multi-variety, small-batch production, the randomness and volatility of production orders have increased significantly, placing higher demands on the dynamic responsiveness of production scheduling systems. Currently, Advanced Planning and Scheduling (APS) systems based on ERP / MES mainly face the following problems in practical applications:

[0003] 1. Bottleneck drift detection lag

[0004] Bottleneck equipment on a production line is not fixed. Factors such as changes in changeover time, fluctuations in operator proficiency, and changes in material batch characteristics can cause the bottleneck to migrate dynamically during production. Existing systems typically use indicators such as waiting queue length and equipment utilization rate for post-event monitoring. When bottleneck drift is detected, there is already significant backlog or idleness on the production line, resulting in actual production efficiency loss that is difficult to recover in a timely manner.

[0005] 2. The disturbance response mechanism is crude.

[0006] When equipment malfunctions, emergency orders are inserted, or cycle time fluctuates, there are two main approaches: one is to perform a global reordering, recalculating all work orders within the entire planning period. This involves a large amount of calculation and takes a long time (usually several minutes to tens of minutes), making it difficult to meet the real-time requirements of online scheduling; the other is to use a simple right-shift scheduling, pushing subsequent processes back in sequence. This approach can easily lead to a chain reaction of congestion in multiple processes, which in turn reduces overall output efficiency.

[0007] 3. Production scheduling strategy is singular and rigid.

[0008] Many scheduling systems have only one built-in scheduling logic (such as first-come-first-served, shortest processing time priority, or genetic algorithm) and assume that this strategy is applicable to all scenarios. However, in actual production, urgent orders need to be prioritized to meet delivery deadlines, batch changeover scenarios need to be concentrated to reduce cleaning costs, and when equipment utilization is low, large tasks need to be prioritized. A single strategy cannot take multiple optimization objectives into account at the same time, resulting in unsatisfactory scheduling results under certain working conditions. Summary of the Invention

[0009] The purpose of this invention is to provide an order-driven intelligent scheduling system based on artificial intelligence to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: an order-driven intelligent scheduling system based on artificial intelligence, comprising:

[0011] The data acquisition and graph construction module is used to collect status data of each workstation on the production line in real time from the programmable logic controller and manufacturing execution system, and construct a spatiotemporal graph based on the production line topology. The status data includes the current cycle time, buffer material inventory, and equipment waiting time.

[0012] The pressure propagation rate calculation module is used to input the spatiotemporal graph into a pre-trained spatiotemporal graph convolutional network model and calculate the pressure propagation rate of each node, wherein the pressure propagation rate is the probability of congestion spread of the node becoming a new bottleneck in the near future.

[0013] The hidden bottleneck drift prediction module is used to traverse all nodes that have not been marked as initial bottlenecks. If there is a node whose pressure propagation rate is greater than the first threshold, and the buffer material inventory time series corresponding to the direct downstream node of the node satisfies a continuous downward trend, wherein the determination condition for the continuous downward trend is: the linear regression slope of the time series is less than the second threshold and decreases for multiple consecutive periods, then the node is marked as a predicted drift node, and a disturbance source signal containing the node identifier and the predicted drift timestamp is generated.

[0014] The disturbance propagation tree generation module is used to respond to the disturbance source signal. In the current directed acyclic graph scheduling scheme, it traverses the affected processes based on process dependencies and resource time window conflicts. The resource time window conflict is determined by the following method: the new end time of the affected process after the delay exceeds the latest start time of the subsequent process. All affected processes are extracted to form a disturbance propagation tree, while processes outside the tree remain unchanged from the original plan.

[0015] The minimum disturbance increment rearrangement solution module includes: (a) a strategy selector, which selects an algorithm from a variety of preset scheduling algorithms based on the disturbance feature vector, wherein the disturbance feature vector includes pressure propagation rate, downstream buffer slope, proportion of emergency orders, concentration of the same product, current average equipment utilization rate and disturbance tree size; and (b) an algorithm execution unit, which calls the selected algorithm to solve the moving job sequence within the disturbance propagation tree, wherein the moving job sequence satisfies the process dependency relationship and resource constraints within the tree and minimizes the total completion time increment.

[0016] The incremental scheduling scheme distribution module is used to incrementally update the moving work sequence to the current scheduling scheme, generate the incremental scheduling scheme, and convert it into control commands to be distributed to the manufacturing execution system and programmable logic controller.

[0017] As a further preferred embodiment of this technical solution: the specific method for determining resource conflicts in the disturbance propagation tree generation module is as follows: calculate the new end time after the delay of the affected process. Obtain the latest start time of other processes that subsequently use the same production line. ,like If a resource conflict is identified, the subsequent process will be included in the disturbance tree.

[0018] As a further preferred embodiment of this technical solution: the input feature vector of the strategy selector includes at least the normalized values ​​of the following six dimensions: pressure propagation rate, absolute value of downstream buffer slope, proportion of emergency orders, concentration of the same type of product, current average equipment utilization rate, and number of perturbation tree nodes.

[0019] As a further preferred embodiment of this technical solution, the preset multiple production scheduling algorithms include the following six types: default first-in-first-out algorithm, concentrated production algorithm for the same product, critical path + shortest processing time priority algorithm, machine utilization priority algorithm, enhanced genetic algorithm, and combined algorithm.

[0020] As a further preferred embodiment of this technical solution, a closed-loop update module is also included. This module updates the spatiotemporal graph convolutional network by combining offline retraining with incremental fine-tuning: full retraining is performed using historical data after each day or shift; fine-tuning is performed using a small amount of recent data with a low learning rate during online periods; and the loss function is the mean square error between the predicted pressure propagation rate and the actual bottleneck label.

[0021] As a further preferred option of this technical solution: during the system initialization phase, the simulation data of the mechanism model or the pre-trained model obtained by transfer learning is used as the cold start model, and then the closed-loop update is switched after real data is accumulated.

[0022] To achieve the above objectives, the present invention also provides the following technical solution:

[0023] An artificial intelligence-based order-driven intelligent scheduling method, applied to the system described in any one of claims 1 to 6, is characterized by comprising the following steps:

[0024] (1) Collect status data of each workstation on the production line in real time and construct a spatiotemporal diagram;

[0025] (2) Input the spatiotemporal graph into the spatiotemporal graph convolutional network and calculate the pressure propagation rate of each node, wherein the pressure propagation rate is the congestion diffusion probability.

[0026] (3) If the pressure propagation rate of a node is greater than the first threshold and the buffer stock of the direct downstream node shows a continuous downward trend, then the node is predicted to be a bottleneck drift node and a disturbance source signal is generated.

[0027] (4) In the current directed acyclic graph scheduling scheme, the affected processes are traversed along the process dependency edges and resource time window conflict chains. The resource time window conflict chain is determined by whether the new end time after the process delay exceeds the latest start time of the subsequent process, and a disturbance propagation tree is generated.

[0028] (5) Based on the feature vector of the perturbation propagation tree, dynamically select one algorithm from multiple scheduling algorithms and solve the moving job sequence within the perturbation propagation tree;

[0029] (6) Convert the incremental update of the production scheduling plan of the mobile operation sequence into control instructions for execution.

[0030] As a further preferred embodiment of this technical solution, the basis for dynamically selecting the algorithm in step (5) includes: pressure propagation rate, downstream buffer descent slope, proportion of emergency orders, concentration of the same type of product, equipment utilization rate and disturbance tree size.

[0031] As a further preferred embodiment of this technical solution: if the enhanced genetic algorithm is selected in step (5), the chromosome is encoded as a legal topological sort of the process within the perturbation tree, the fitness function comprehensively considers the completion time, the total delay and the total transformation time, the crossover operation adopts crossover that maintains the dependency relationship, and the elite retention strategy is adopted.

[0032] As a further preferred embodiment of this technical solution: the calculation method for the predicted drift timestamp in step (3) is: the current time plus the process flow time between the predicted drift node and the direct downstream node.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] 1. In this invention, the pressure propagation rate of each node is calculated by a spatiotemporal graph convolutional network, and the continuous downward trend of the material inventory in the downstream buffer zone is used as a physical criterion. This can issue an early warning several minutes before the bottleneck actually forms, usually five to fifteen minutes before the bottleneck actually forms. Compared with the post-event monitoring method, this provides more time for the dispatcher to respond and helps to reduce production line waiting and equipment idleness caused by bottleneck drift.

[0035] 2. In this invention, by constructing a perturbation propagation tree, only the processes affected by predicted drift are included in the rescheduling scope. These processes typically account for 20% to 30% of all processes, while keeping the plans for processes outside the tree unchanged. In a simulation environment, this method can reduce the calculation time for a single production scheduling from about 28 seconds for global rescheduling to about 2 to 3 seconds, reducing the amount of calculation by about 90%. This makes online, fast local rescheduling possible and is more suitable for dynamic production environments.

[0036] 3. In this invention, the system incorporates six scheduling algorithms, including first-in-first-out, concentrated production of the same product, critical path plus shortest processing time priority, machine utilization priority, enhanced genetic algorithm, and combined algorithm. The system automatically selects the appropriate algorithm based on characteristics such as pressure propagation rate, downstream buffer slope, and proportion of urgent orders. Under different disturbance scenarios, such as urgent order insertion, batch change of the same product, or low equipment utilization, the system can produce better or closer to the optimal local scheduling results compared to a single algorithm solution.

[0037] 4. In this invention, since the process plans outside the disturbance propagation tree remain unchanged and the processes within the tree are only adjusted to the minimum extent, the overall change in the production scheduling plan is small, avoiding frequent instruction changes caused by global rescheduling. This helps reduce the cost for on-site workers to understand the new plan and reduces the execution chaos that may be caused by large fluctuations in the plan.

[0038] 5. In this invention, the closed-loop update module uses the bottleneck drift data generated after actual execution to update the spatiotemporal graph convolutional network model through a combination of offline retraining and online incremental fine-tuning, so that its prediction accuracy gradually improves with the accumulation of running data. In the system cold start phase, digital twin simulation data or transfer learning models are used for initialization, ensuring that the system can be used in the early stages. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention;

[0040] Figure 2 This is an example diagram of the spatiotemporal graph and adjacency matrix of the present invention;

[0041] Figure 3 This is the perturbation propagation tree generation process of the present invention;

[0042] Figure 4 This is a flowchart illustrating the collaboration between the strategy selector and multiple algorithm libraries in this invention.

[0043] Figure 5 The following is a timing diagram of the six algorithms in local rearrangement. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1: System Overall Architecture and Workflow

[0046] like Figure 1 As shown, this embodiment takes the ointment-making workshop of a detergent manufacturing enterprise as an example to introduce in detail the specific implementation of the system of the present invention. The workshop includes five workstations: ingredient preparation workstation, mixing workstation, emulsification workstation, filling workstation, and packaging workstation. The workstations are connected by conveyor belts to form a unidirectional serial topology. The system is deployed on an edge server, communicates with the PLCs of each workstation through the OPCUA protocol, and integrates with the enterprise MES system through RESTAPI.

[0047] (I) Data Acquisition and Graph Construction

[0048] The processor collects status data from the PLCs at each workstation every 30 seconds. The data collected at a certain moment is as follows:

[0049] workstation Cycle time (seconds / batch) Buffer zone material inventory (batch) Device waiting time (seconds) Ingredients 110 12 3 mix 105 9 2 emulsification 122 15 8 Filling 95 5 1 Package 90 30 0

[0050] like Figure 2 As shown, the processor constructs a space-time graph based on the production line topology. Node set Numbered sequentially from 1 to 5, corresponding to the ingredient preparation, mixing, emulsification, filling, and packaging stations respectively. Directed edges represent the material transport direction: ingredient preparation → mixing → emulsification → filling → packaging. Adjacency matrix. It is a 5×5 square matrix. This indicates that there is a material transfer path from the i-th workstation to the j-th workstation, specifically: The remaining elements are 0.

[0051] Node feature matrix It is a 5×3 real matrix. The i-th row stores the cycle time (seconds / batch), buffer material inventory (batch), and equipment waiting time (seconds) of the i-th workstation at time t.

[0052] Based on the data in the table above, .

[0053] (II) Calculation of pressure transmission rate

[0054] The processor inputs the spatiotemporal graph into a pre-trained spatiotemporal graph convolutional network (ST-GCN) model. ST-GCN first aggregates spatial features for each node: taking the emulsification station node as an example, its incoming edge adjacent node is the hybrid station, the cycle time of the hybrid station is 105 seconds, the waiting time is 2 seconds, and the attention weight calculated by the graph attention mechanism is 0.7. After weighting and aggregating these two features, a spatial aggregated feature vector is obtained. Then, the model performs one-dimensional temporal convolution on the spatial aggregated features of 10 consecutive time windows to extract temporal evolution features. Finally, the pressure propagation rate is output through a fully connected layer and a sigmoid function: 0.21 for ingredient mixing, 0.35 for mixing, 0.83 for emulsification, 0.42 for filling, and 0.18 for packaging.

[0055] The pre-training process of the ST-GCN model is as follows: A time window is extracted from historical production line operation data. As input features, if the node waiting time surges (more than 1.5 times the historical average) and downstream nodes become idle or the buffer continues to decrease within 10 minutes after the observation window ends, then the label y=1; otherwise, y=0. The model uses the Adam optimizer with a learning rate of 0.001, a batch size of 32, and is trained for 100 rounds.

[0056] (III) Prediction of Hidden Bottleneck Shift

[0057] The current default bottleneck of the system is the filling station (because the filling line is the slowest). The processor traverses all nodes except filling and finds that the pressure propagation rate of the emulsification station is 0.83 > the first threshold of 0.7, which meets condition A. Then, it checks the buffer stock time series of the direct downstream node of the emulsification node (filling station): the stock values ​​of the most recent 6 periods are [8,7,6,5,5,4] (unit: batch). The linear regression slope is calculated. ≈−0.8, which is less than the second threshold of -0.1, and is less than the previous period for 6 consecutive periods, satisfying condition B. Therefore, the processor marks the emulsification level as a predicted drift node and generates a disturbance source signal, which includes the node identifier "E001" and the predicted drift timestamp. The predicted drift timestamp = current time 09:00:00 + emulsification → filling flow time 2 minutes = 09:02:00.

[0058] (iv) Generation of the perturbation propagation tree

[0059] like Figure 3 As shown, the processor obtains the current active directed acyclic graph (DAG) scheduling plan, which contains 20 process nodes, covering the entire process from ingredient preparation to packaging. It traverses backwards along the dependency edges, with the emulsification process as the root node:

[0060] First, calculate the earliest start time of each subsequent process in the original plan. Predict the expected delay of drift nodes. =15 minutes (This delay is estimated based on the difference between the predicted drift node's current beat time of 122 seconds and the standard beat time of 100 seconds). Recalculate It was discovered that the subsequent eight processes (including temporary storage, filling, and packaging) All greater than Therefore, these 8 processes are added to the affected set.

[0061] Secondly, resource conflict propagation is performed: for each marked affected process, its occupied production line time window is checked. Taking the emulsification process as an example, its delayed new end time is determined. =09:18:00, another process using the same filling line is subsequently completed. (The original latest start time, which is unrelated to the emulsification process) =09:15:00, due to =09:18:00> =09:15:00, determined to be a resource conflict, process They were also included in the affected set and processed recursively, ultimately resulting in 10 affected processes.

[0062] The above 10 processes, their dependencies, and resource usage relationships are extracted to form a perturbation propagation tree, while the plans for the remaining 10 processes outside the tree remain unchanged.

[0063] (v) Solving for the minimum perturbation increment rearrangement

[0064] like Figure 4 As shown, the minimum perturbation increment rearrangement solution module internally includes a policy selector and a multi-policy algorithm library. The policy selector receives the perturbation feature vector. Pressure transmission rate =0.83, downstream buffer slope =-0.8, percentage of urgent orders within the tree =0.35 (due to 3 orders with delivery times within 4 hours), concentration of similar products =0.2, current average equipment utilization rate =0.42, size of the perturbation tree =10. According to the preset rules ( >0.3), the strategy selector selects "critical path + shortest processing time first algorithm" (CRITICAL_PATH_SPT).

[0065] The algorithm execution unit calls the CRITICAL_PATH_SPT algorithm to perform local rearrangement within the perturbation tree:

[0066] 1. Calculate the earliest start time for each process based on the process dependency graph. and latest start time (Considering the fixed completion time constraints of processes outside the tree), processes on the critical path (total float) =0) includes: emulsification process, filling process, and packaging process.

[0067] 2. Calculate the critical ratio for each process step.

[0068] ,in This refers to the total processing time of this process and all its subsequent processes. The emulsification process corresponding to an urgent order. =0.85 (lower), prioritize production.

[0069] 3. Multi-level sorting: First priority urgency (highest priority), second priority... (Ascending order), third priority processing time (ascending order, SPT), fourth priority creation time (ascending order), after sorting, a new process order is obtained: first process the emulsification, filling, and packaging of urgent orders, and then process other orders.

[0070] 4. Assign the sorted processes to the earliest available production line in sequence to obtain the new start and end times of each process.

[0071] The resulting moving job sequence contains three process nodes whose execution order needs to be adjusted, with a total completion time increment of 14 minutes, while the external process plan remains unchanged.

[0072] (vi) Issuance of incremental production scheduling plan

[0073] The processor incrementally updates the moving job sequence to the current DAG scheduling scheme, generating an incremental scheduling scheme. This scheme only includes the planned changes of 10 processes within the disturbance tree. The processor converts the incremental scheduling scheme into control instructions: for example, modifying the start time of the corresponding work order in the MES, sending the preset value of the timer for "delayed start" or "early start" to the PLC. After the instruction is issued, the production line executes according to the new scheme.

[0074] (vii) Closed-loop update

[0075] After actual execution, the emulsification process was recorded as starting at 09:03:00, which was 3 minutes later than the planned start time of 09:00:00. (Actual label) =1 (because the downstream buffer continues to decrease and this node becomes a bottleneck), the processor will process the sample (input features + predicted value). =0.83 + actual label =1) Store the data in the cache. After the end of each workday, the system adds the accumulated samples of that day to the historical dataset for full offline retraining. Simultaneously, during online sessions, fine-tune the data using the most recent 100 samples at a learning rate of 0.0001 (1 / 10 of the original training learning rate), and the loss function... The ST-GCN weights are updated using gradient descent.

[0076] Cold start phase: During the first week after the system goes live, there is no actual historical data. The system preloads a pre-trained model based on queuing theory simulation (1000 hours of simulation to generate 100,000 samples). After two weeks of accumulating real running data, it switches to the closed-loop update mode described above.

[0077] Example 2: Centralized Batch Production of the Same Product

[0078] When the system detects a backlog of orders for the same product (such as hand sanitizer) and there are no urgent orders ( =0.1) The pressure propagation rate is moderate (0.65), and the downstream buffer decreases gradually (slope -0.05). Strategy selector determination. =0.85>0.6, select the same product group production algorithm (PRODUCT_BATCH).

[0079] The processor groups all hand sanitizer processes within the perturbation tree by product ID and merges them into two large batches: Batch A (total quantity 1300 batches) and Batch B (total quantity 800 batches). It calculates the changeover cost matrix and uses the nearest neighbor algorithm to obtain the optimal order: Batch A is produced first, followed by Batch B. Batch A is arranged to be produced continuously on the serial line of ingredient preparation → mixing → emulsification → filling → packaging, with the total working time accumulated. After Batch A is completed, a 30-minute cleaning time is inserted, and then Batch B is produced. In the end, the total changeover time is reduced from the original 5 times × 30 minutes = 150 minutes to 1 time × 30 minutes = 30 minutes, saving 120 minutes.

[0080] Example 3: Genetic Algorithm Optimization under Complex Coupling Perturbations

[0081] When multiple products are mixed on the production line at the same time, and the perturbation tree size reaches 28 nodes ( =28>20), the strategy selector selects the Enhanced Genetic Algorithm (ENHANCED_GENETIC_ALGORITHM), with the following parameters: population size 20, maximum generations 50, crossover rate 0.8, mutation rate 0.1, elites retained 2, and fitness function... =1, =2, =0.5. After 50 generations of evolution, the total completion time of the optimal solution was reduced from the original 240 minutes to 218 minutes, the total delay was reduced from 45 minutes to 12 minutes, and the calculation time was 2.8 seconds.

[0082] Example 4: Combinatorial Algorithm for Handling Mixed Feature Perturbations

[0083] When the proportion of urgent orders within the disturbed tree is 25% ( When the current equipment utilization rate is only 38%, the strategy selector calls the combined algorithm (COMBINE_ALGORITHM). This algorithm first executes CRITICAL_PATH SPT to handle urgent tasks and fix them at the beginning of the time window; then it executes MACHINE_UTILIZATION during the remaining idle time to fill the idle gaps with non-urgent large tasks. Finally, the urgent orders are completed on time, and the equipment utilization rate is increased from 38% to 52%, which is better than using any algorithm alone.

[0084] Example 5: A Complete Step-by-Step Example from System to Method

[0085] Taking an emergency order scenario as an example, the complete execution method follows the steps described in claim 7:

[0086] Step (1): Collect the status data (cycle time, buffer time, waiting time) of 5 workstations in real time and construct a spatiotemporal graph (node ​​5, edge 4).

[0087] Step (2): Input the spatiotemporal diagram into ST-GCN and output the pressure propagation rate of each node.

[0088] [0.21, 0.35, 0.83, 0.42, 0.18].

[0089] Step (3): Emulsification =0.83>0.7, and the downstream filling buffer continues to decrease (slope -0.8), which is predicted to be a bottleneck drift node, generating a disturbance source signal, timestamp 09:02:00.

[0090] Step (4): In the DAG scheduling scheme, traversing the dependent edges yields 8 direct successors, and traversing the resource conflicts yields 2 additional successors (because...). > Generate a perturbation propagation tree containing 10 nodes.

[0091] Step (5): Based on the feature vector The CRITICAL_PATH_SPT algorithm is selected to solve the moving job sequence (3 process adjustments, total increment 14 minutes) within the tree.

[0092] Step (6): Incrementally update the production scheduling plan and convert it into PLC instructions for execution.

[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An order-driven intelligent scheduling system based on artificial intelligence, characterized in that, include: The data acquisition and graph construction module is used to collect status data of each workstation on the production line in real time from the programmable logic controller and manufacturing execution system, and construct a spatiotemporal graph based on the production line topology. The status data includes the current cycle time, buffer material inventory, and equipment waiting time. The pressure propagation rate calculation module is used to input the spatiotemporal graph into a pre-trained spatiotemporal graph convolutional network model and calculate the pressure propagation rate of each node, wherein the pressure propagation rate is the probability of congestion spread of the node becoming a new bottleneck in the near future. The hidden bottleneck drift prediction module is used to traverse all nodes that have not been marked as initial bottlenecks. If there is a node whose pressure propagation rate is greater than the first threshold, and the buffer material inventory time series corresponding to the direct downstream node of the node satisfies a continuous downward trend, wherein the determination condition for the continuous downward trend is: the linear regression slope of the time series is less than the second threshold and decreases for multiple consecutive periods, then the node is marked as a predicted drift node, and a disturbance source signal containing the node identifier and the predicted drift timestamp is generated. The disturbance propagation tree generation module is used to respond to the disturbance source signal. In the current directed acyclic graph scheduling scheme, it traverses the affected processes based on process dependencies and resource time window conflicts. The resource time window conflict is determined by the following method: the new end time of the affected process after the delay exceeds the latest start time of the subsequent process. All affected processes are extracted to form a disturbance propagation tree, while processes outside the tree remain unchanged from the original plan. The minimum perturbation increment rearrangement solution module contains: (a) A strategy selector for selecting one of a variety of pre-set scheduling algorithms based on a perturbation feature vector, wherein the perturbation feature vector includes pressure propagation rate, downstream buffer slope, proportion of emergency orders, concentration of the same product, current average equipment utilization rate and perturbation tree size; (b) Algorithm execution unit, used to call the selected algorithm to solve the moving job sequence within the disturbance propagation tree, the moving job sequence satisfying the process dependency relationship and resource constraints within the tree, and minimizing the total completion time increment; The incremental scheduling scheme distribution module is used to incrementally update the moving work sequence to the current scheduling scheme, generate the incremental scheduling scheme, and convert it into control commands to be distributed to the manufacturing execution system and programmable logic controller.

2. The system according to claim 1, characterized in that, The specific method for determining resource conflicts in the disturbance propagation tree generation module is as follows: calculate the new end time after the affected process is delayed. Obtain the latest start time of other processes that subsequently use the same production line. ,like If a resource conflict is identified, the subsequent process will be included in the disturbance tree.

3. The system according to claim 1, characterized in that, The input feature vector of the strategy selector includes at least the normalized values ​​of the following six dimensions: pressure propagation rate, absolute value of downstream buffer slope, proportion of emergency orders, concentration of the same product, current average equipment utilization rate, and number of perturbation tree nodes.

4. The system according to claim 1, characterized in that, The pre-set production scheduling algorithms include the following six types: default first-in-first-out algorithm, concentrated production algorithm for the same product, critical path + shortest processing time priority algorithm, machine utilization priority algorithm, enhanced genetic algorithm, and combined algorithm.

5. The system according to claim 1, characterized in that, It also includes a closed-loop update module, which updates the spatiotemporal graph convolutional network by combining offline retraining with incremental fine-tuning: full retraining is performed using historical data after each day or shift; fine-tuning is performed with a small amount of recent data at a low learning rate during online periods; the loss function is the mean square error between the predicted pressure propagation rate and the actual bottleneck label.

6. The system according to claim 5, characterized in that, During the system initialization phase, the simulation data of the mechanism model or the pre-trained model obtained by transfer learning is used as the cold start model, and the closed-loop update is switched after real data is accumulated.

7. An order-driven intelligent scheduling method based on artificial intelligence, applied to the system described in any one of claims 1 to 6, characterized in that, Includes the following steps: (1) Collect status data of each workstation on the production line in real time and construct a spatiotemporal diagram; (2) Input the spatiotemporal graph into the spatiotemporal graph convolutional network and calculate the pressure propagation rate of each node, wherein the pressure propagation rate is the congestion diffusion probability. (3) If the pressure propagation rate of a node is greater than the first threshold and the buffer stock of the direct downstream node shows a continuous downward trend, then the node is predicted to be a bottleneck drift node and a disturbance source signal is generated. (4) In the current directed acyclic graph scheduling scheme, the affected processes are traversed along the process dependency edges and resource time window conflict chains. The resource time window conflict chain is determined by whether the new end time after the process delay exceeds the latest start time of the subsequent process, and a disturbance propagation tree is generated. (5) Based on the feature vector of the perturbation propagation tree, dynamically select one algorithm from multiple scheduling algorithms and solve the moving job sequence within the perturbation propagation tree; (6) Convert the incremental update of the production scheduling plan of the mobile operation sequence into control instructions for execution.

8. The method according to claim 7, characterized in that, The basis for dynamically selecting the algorithm in step (5) includes: pressure propagation rate, downstream buffer descent slope, proportion of emergency orders, concentration of the same product, equipment utilization rate and disturbance tree size.

9. The method according to claim 7, characterized in that, If the enhanced genetic algorithm is selected in step (5), the chromosome is encoded as a legal topological sort of the process within the perturbation tree. The fitness function comprehensively considers the completion time, total delay, and total transformation time. The crossover operation adopts crossover that maintains the dependency relationship and adopts an elite retention strategy.

10. The method according to claim 7, characterized in that, The calculation method for the predicted drift timestamp in step (3) is: the current time plus the process flow time between the predicted drift node and the direct downstream node.

Images