Warping workshop intelligent scheduling management method and system

Abstract

The invention belongs to the field of production control, and particularly discloses an intelligent scheduling management method and system for a warping workshop, and the method comprises the steps: firstly obtaining cheese task data, warping machine state data and other warping workshop data; establishing a bipartite graph taking a cone yarn task node and a warping machine node as vertexes, establishing edges based on constraints such as color number matching and generating feature vectors; inputting the bipartite graph into a pre-training graph neural network model, and outputting a matching score of an edge through coding aggregation; after edges are filtered according to rigid constraints, the global matching problem of the weighted bipartite graph is solved to obtain an optimal allocation scheme; and finally, issuing and executing the scheme-to-equipment control instruction. According to the method, global optimal scheduling is realized by integrating multi-dimensional constraints, the production efficiency and the product quality are improved, dynamic change of the workshop is adapted, and intelligent management of the warping workshop is improved.

Description

Technical Field

[0001] This invention belongs to the field of production control, specifically relating to an intelligent scheduling and management method and system for warping workshops. Background Technology

[0002] In the warping workshop of the textile industry, warping scheduling is a key link affecting production efficiency and product quality. Traditional warping scheduling uses linear matching logic, allocating yarn packages only based on the idle state and speed matching degree of the warping machine, without considering the multi-dimensional complexity of the warping task, resulting in difficulty in ensuring quality consistency and limiting production efficiency.

[0003] Specifically, warping tasks are affected by a combination of factors: yarn batches need to be uniform to avoid dyeing differences, color codes need to be consistent to prevent color variations in finished products, tension needs to be within the warping machine's allowable range to prevent yarn breakage or uneven winding, and the load on the conveyor line path (to avoid congestion and timeouts) and task priority (to meet rush production needs) must be considered simultaneously. Traditional scheduling ignores these factors, which can easily lead to problems such as reallocation, equipment downtime, and excessive color differences, increasing rework costs and production delays.

[0004] In addition, the workshop status is dynamic, with the position of yarn packages, the operating status of warping machines, and the path load changing over time. Traditional static scheduling cannot adapt in real time, further reducing the rationality of scheduling. Summary of the Invention

[0005] To address this, the present invention provides an intelligent scheduling and management method and system for warping workshops, thereby solving the aforementioned technical problems.

[0006] This invention provides an intelligent scheduling and management method for a warping workshop, comprising the following steps: S1, acquiring warping workshop data, wherein the warping workshop data includes yarn package task data, warping machine status data, workshop topology and path information, and historical scheduling records.

[0007] S2. Establish a bipartite graph with the yarn bobbin task node and the warping machine node as vertices. Based on color matching, path reachability and tension constraints, establish directed edges and generate feature vectors for nodes and edges.

[0008] S3, the bipartite graph is input into a pre-trained graph neural network model, wherein the graph neural network model encodes and aggregates node and edge information through a node embedding layer and multiple message passing layers, and finally outputs a matching score representing the overall matching quality for each directed edge through an edge scoring head.

[0009] S4. Based on preset hard constraints, all the directed edges are filtered out, and edges that do not meet the requirements of color compatibility, tension range, or path reachability are removed.

[0010] S5 takes the filtered directed edges and their matching scores as input, constructs and solves a weighted bipartite graph global matching problem with the goal of maximizing the total matching score, and thus obtains the optimal allocation scheme of yarn bobbin tasks to warping machines.

[0011] S6, the allocation scheme is converted into specific equipment control instructions and sent to the workshop execution system for execution.

[0012] Preferably, the warping workshop data is preprocessed, including setting a sliding time window to collect time-series data, standardizing continuous numerical features, numerically encoding or embedding discrete classification features, and filling in missing data.

[0013] Preferably, the yarn bobbin task node includes a standardized speed sequence summary, remaining bobbin length, color number embedding vector, tension level, priority, and current path load characteristics; the warping machine node includes current machine speed, target color number embedding vector, allowable tension range, status code, priority, and expected idle time characteristics; the directed edge includes speed matching degree, path transmission delay estimation, historical pairing success rate, color number compatibility score, and tension penalty value characteristics.

[0014] Preferably, the bipartite graph is a time-series dynamic graph. By introducing time coding or using a sliding time window, the time-series changes of the yarn winding task and the warping machine are integrated into the feature representation of nodes and edges to capture the dynamic evolution of the workshop state.

[0015] Preferably, the graph neural network model adopts a bidirectional message passing mechanism, including forward message passing from the yarn package task node to the warping machine node, and reverse message passing from the warping machine node to the yarn package task node; the message passing integrates the hidden state features of connected nodes and the features of edges, and uses an attention mechanism to perform weighted aggregation of messages from different neighboring nodes.

[0016] Preferably, the graph neural network model is trained using a multi-objective joint loss function for supervised training; the joint loss function includes at least: a supervised learning loss based on historical successful pairing data, a ranking comparison loss to ensure that high-quality pairings score higher than low-quality pairings, and a constraint-aware loss to penalize pairings that violate tension or path capacity constraints.

[0017] Preferably, the global matching problem of the weighted bipartite graph is solved by using the Hungarian algorithm or the minimum cost flow algorithm based on maximum flow, in order to obtain the global optimal matching under the constraints that each yarn bobbin task is assigned at most once and the number of tasks received by each warping machine does not exceed its task slot capacity.

[0018] Preferably, during execution, the system status is monitored in real time and compared with the prediction of the graph neural network. When the deviation exceeds the threshold, a local or global replanning process is triggered, and steps S2 to S5 are re-executed to generate a new scheduling scheme.

[0019] Preferably, the replanning process is a local replanning. When an anomaly is detected, edge scoring and matching decisions are re-performed only for the subgraph formed by the yarn bobbin task nodes and warping machine nodes affected by the anomaly, while the scheduling scheme of the unaffected parts of the system remains unchanged.

[0020] In another aspect, this application also provides an intelligent scheduling and management system for a warping workshop, comprising: a warping workshop data acquisition module for acquiring warping workshop data, wherein the warping workshop data includes yarn package task data, warping machine status data, workshop topology and path information, and historical scheduling records.

[0021] The bipartite graph construction module is used to build a bipartite graph with the yarn package task node and the warping machine node as vertices. It establishes directed edges based on color matching, path reachability and tension constraints, and generates feature vectors for nodes and edges.

[0022] The matching score module is used to input the bipartite graph into a pre-trained graph neural network model, wherein the graph neural network model encodes and aggregates node and edge information through a node embedding layer and multiple message passing layers, and finally outputs a matching score representing the overall matching quality for each directed edge through an edge scoring head.

[0023] The directed edge filtering module is used to filter all the directed edges based on preset hard constraints, and remove edges that do not meet the requirements of color compatibility, tension range, or path reachability.

[0024] The allocation scheme acquisition module is used to take the filtered directed edges and their matching scores as input, construct and solve a weighted bipartite graph global matching problem with the goal of maximizing the total matching score, so as to obtain the optimal allocation scheme of yarn bobbin tasks to warping machines.

[0025] The execution module is used to convert the allocation scheme into specific equipment control instructions and send them to the workshop execution system for execution.

[0026] This invention utilizes temporal dynamic bipartite graph modeling with a graph neural network (GNN) at its core to integrate multi-dimensional constraints such as yarn batch, color number, and tension, achieving globally optimal matching between yarn packages and warping machines. This effectively solves the problems of traditional scheduling neglecting quality consistency and multi-factor coupling conflicts, reducing the risk of abnormalities such as color difference and yarn breakage, and minimizing rework and downtime costs. It employs the Hungarian algorithm or minimum cost flow algorithm to solve the global matching, balancing optimality in small-scale scenarios with real-time performance in large-scale scenarios. By combining real-time monitoring and local replanning, it dynamically adapts to changes in workshop status, avoiding scheduling lag, and improving the production efficiency and intelligent scheduling level of the warping workshop while ensuring product quality. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0028] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0029] Figure 1 A flowchart of an intelligent scheduling and management method for a warping workshop provided in an embodiment of the present invention.

[0030] Figure 2 This is a schematic diagram of the architecture of the GNN model in an embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram of a two-way message passing process provided in an embodiment of the present invention.

[0032] Figure 4 This is a schematic diagram of a local replanning process provided in an embodiment of the present invention.

[0033] Figure 5 This is a schematic diagram of an intelligent scheduling and management system for a warping workshop provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] like Figure 1 As shown, this embodiment of the invention discloses an intelligent scheduling and management method 100 for a warping workshop, which includes the following steps: S1, acquiring warping workshop data, wherein the warping workshop data includes yarn package task data, warping machine status data, workshop topology and path information, and historical scheduling records.

[0036] S2. Establish a bipartite graph with the yarn bobbin task node and the warping machine node as vertices. Based on color matching, path reachability and tension constraints, establish directed edges and generate feature vectors for nodes and edges.

[0037] S3, the bipartite graph is input into a pre-trained graph neural network model, wherein the graph neural network model encodes and aggregates node and edge information through a node embedding layer and multiple message passing layers, and finally outputs a matching score representing the overall matching quality for each directed edge through an edge scoring head.

[0038] S4. Based on preset hard constraints, all the directed edges are filtered out, and edges that do not meet the requirements of color compatibility, tension range, or path reachability are removed.

[0039] S5 takes the filtered directed edges and their matching scores as input, constructs and solves a weighted bipartite graph global matching problem with the goal of maximizing the total matching score, and thus obtains the optimal allocation scheme of yarn bobbin tasks to warping machines.

[0040] S6, the allocation scheme is converted into specific equipment control instructions and sent to the workshop execution system for execution.

[0041] In some embodiments, for S1, the yarn bobbin task data, for example, is extracted primarily from the real-time database of the workshop MES and the conveyor line PLC, covering all yarn bobbin tasks currently awaiting assignment or about to be loaded onto the conveyor line.

[0042] Specifically, this may include basic identification information, such as batch number (e.g., A01, Class A dyeing batch), color number (e.g., #FF0033), yarn package number (unique code, e.g., TS-001), and current conveyor line ID (e.g., Conveyor-03, identifying the current conveyor line of the yarn package).

[0043] The timing and physical characteristics include, for example, arrival timestamps (the system record time when the yarn enters the current conveyor line, accurate to the second), current sampling speed historical sequence (the yarn conveying speed data collected by the conveyor line PLC every 5 seconds within the last 180-300 seconds, forming a timing sequence), and remaining yarn length (unit: meters, which is synchronized to the PLC from the laser length measuring instrument data before the yarn is hung, and updated in real time).

[0044] Process and priority characteristics, such as tension level (e.g., 1-5, with 1 being the lowest and 5 being the highest, tension detection data from the previous winding process machine, synchronized to MES) and priority labels (e.g., "urgent", "normal", "low priority", set in ERP according to the order delivery date, synchronized to the MES yarn task table).

[0045] For example, the status data of the warping machine is collected from the background database of the local PLC control system, MES and machine HMI (human-machine interface) of the warping machine, so as to reflect the current operating status and process requirements of each warping machine in real time.

[0046] Specifically, this may include basic and status information, such as the warping machine ID (unique code, such as "WJ-08") and the current task status (enumerated values: idle / start / running / tail yarn / maintenance, determined by the warping machine PLC based on the equipment sensor signals, such as "running" status corresponding to machine spindle speed > 0 and no fault alarm).

[0047] Process parameter information, such as target color number / formula (the color number of the yarn package and the process formula number required by the current order to be processed, such as #FF0033-Recipe05), current machine speed (real-time speed of the warping machine spindle, unit: meters / minute, collected by the PLC speed sensor), and allowable tension range (such as "20-25N", preset in MES according to the process formula and synchronized to the warping machine PLC).

[0048] Scheduling-related information, such as priority (corresponding to the priority of yarn winding tasks, such as setting the warping machine associated with an urgent order to the highest priority), and estimated remaining processing time (unit: minutes, calculated by the PLC based on the length already processed for the current task and the current machine speed, formula: remaining processing time = (total processing length - processed length) / current machine speed).

[0049] For example, workshop topology and path information can be collected from the databases of the workshop WMS (Warehouse Management System) and the conveyor line PLC to reflect the physical connection and real-time load status of the workshop conveyor network.

[0050] Specifically, this may include topology information, such as the connectivity of the transmission lines (e.g., the output of "Conveyor-03" connects to "Conveyor-05" and "Conveyor-07" to form a passable path) and empirical values ​​of path transmission delay (e.g., the average transmission time of "Conveyor-03→Conveyor-05" is 2 minutes, which is based on historical transmission data statistics).

[0051] Real-time load information, such as the current load of a path (unit: yarn, referring to the number of yarn bobbins currently in transit on a certain conveyor line or a certain path, such as "Conveyor-03" currently has 12 yarn bobbins in transit).

[0052] Historical scheduling records, for example, can be collected from the MES scheduling log and the anomaly record module of the quality traceability system.

[0053] Specifically, this may include allocation relationship records, such as historical allocations (e.g., "TS-023" yarn package is allocated to warping machine 1 (ID "WJ-08")), and result labels (enumerated values: success / reassignment / caused downtime / color difference problem, labeled by MES according to task completion status and quality inspection results).

[0054] Performance-related records, such as output (the actual length produced by the warping machine corresponding to this assignment, in meters) and delay time (if the task is not completed on time, delay time = actual completion time - planned completion time, in minutes).

[0055] Preferably, the synchronization frequency is set differently according to the real-time requirements of the data. For data with high real-time requirements, such as warping machine status and conveyor line path load, the synchronization frequency is 1-10 seconds; for data with low real-time requirements, such as basic information on yarn bobbin tasks and historical scheduling records, the synchronization frequency is 1-5 minutes. A unified storage of data from multiple systems is achieved through an intermediate database. For example, an associated index can be established based on timestamps and device IDs / yarn bobbin IDs to ensure that full data of the same time dimension and the same object (such as a specific yarn bobbin or warping machine) can be quickly retrieved during subsequent preprocessing, avoiding data fragmentation.

[0056] Preferably, the warping workshop data is preprocessed, including setting a sliding time window to collect time-series data, standardizing continuous numerical features, numerically encoding or embedding discrete classification features, and filling in missing data.

[0057] Specifically, based on the principle of capturing dynamic changes in the workshop within a short period of time while balancing data timeliness and computational efficiency, the duration of a sliding time window is set, for example. The value range is 180-300 seconds. It can be understood that in actual applications, it can be flexibly adjusted according to the yarn conveying cycle in the workshop and the switching frequency of the warping machine (e.g., 180 seconds for a short conveying cycle and 300 seconds for a long cycle). This invention does not impose any restrictions.

[0058] The window sliding step size is synchronized with the data acquisition frequency, for example, set to 5-10 seconds (matching the PLC's real-time sampling frequency), ensuring that each window slide covers the latest time-series data without data redundancy. In one embodiment, a statistical summary is performed on the speed sequence within the window, calculating the speed mean (reflecting the average yarn conveying speed within that time period, avoiding interference from single-point outliers) and speed slope (calculated through linear fitting, reflecting the speed change trend; a positive slope indicates increasing speed, and a negative slope indicates decreasing speed). The mean and slope are used as the time-series feature summary for subsequent model inputs, replacing the original long-series data and reducing computational complexity.

[0059] In one embodiment, continuous numerical features are standardized. For example, five types of continuous numerical features are selected: speed (including the average speed within the window), remaining cylinder length, path transmission delay, current path load (number of cylinders in transit), and estimated remaining processing time. Standardization eliminates the influence of dimensions. Exemplarily, z-score standardization or min-max standardization is used, which will not be elaborated further in this embodiment.

[0060] In one embodiment, discrete classification features are numerically encoded or embedded, for example, discrete features with clear priority or hierarchical relationships, such as priority, tension level, and warping machine task status, are numerically encoded.

[0061] Specifically, for priority coding, for example, it is mapped to numerical values ​​according to "urgency level", such as "urgent (rush orders) = 3, normal = 2, low priority = 1".

[0062] For tension level coding, for example, it is mapped to numerical values ​​according to "tension magnitude", such as "high tension = 5, medium-high tension = 4, medium tension = 3, medium-low tension = 2, low tension = 1". When matched with the warping machine's "allowable tension range" (a continuous feature that has been standardized), the tension difference can be quickly calculated (e.g., there is a tension conflict between the yarn package with code 5 and the warping machine with an allowable range of 3-4).

[0063] For the task status coding of warping machines, for example, the "whether a new task can be accepted" is mapped to a numerical value, such as "Idle = 1 (can be accepted immediately), tail yarn = 2 (will be idle soon), start = 3 (ready to run, cannot be accepted temporarily), running = 4 (working), maintenance = 5 (cannot be accepted)". After coding, the model can quickly filter warping machine nodes in the "cannot be accepted" state.

[0064] For discrete features that have no hierarchical relationship and only need to distinguish categories, such as color codes (different color codes have no priority difference, but must strictly match the target color code of the warping machine to avoid color difference), embedding representation is used.

[0065] Specifically, for example, all color codes are first one-hot encoded. If there are 10 color codes in the workshop, then color code A corresponds to the vector [1,0,0,...,0], and color code B corresponds to [0,1,0,...,0], ensuring the independence of different color codes.

[0066] Next, a low-dimensional embedding mapping is performed. Since the dimension of the one-hot vector is consistent with the number of color numbers (the more color numbers, the higher the dimension, which can easily lead to the curse of dimensionality), it is converted into a low-dimensional embedding vector of 64-256 dimensions through linear mapping (preferably, matching the embedding dimension of GNN nodes).

[0067] Preferably, a layered filling strategy is adopted to address the potential short-term data shortages or long-term data gaps that may occur during the data collection process, ensuring data integrity and accuracy.

[0068] Specifically, for short-term missing data, such as scenarios with temporary data interruptions (e.g., sensors briefly going offline) or missing data for a duration of ≤1 sliding time window, the missing data is filled in using previous values / mean values.

[0069] Specifically, previous value imputation includes, if the missing data is a time-series continuous feature (such as speed or path delay), directly using the value of the sampling point before the missing time to imput it (e.g., if the speed data is missing at the 10th second, the speed value at the 9th second is used to replace it), thus preserving the continuity of the time-series data and avoiding trend breakage.

[0070] Mean imputation includes using the mean of data from the same batch / type to imput the missing data if the missing data is a non-temporally continuous feature (such as remaining bobbin length or path load) or an ordered discrete feature (such as tension level). For example, if the remaining bobbin length data of a certain yarn is missing, the mean of the remaining bobbin length of other yarns in the same batch (same dyeing batch) is calculated as the imputation value; if the tension level of a certain warping machine is missing, the mode of the current tension level of the same model of warping machine (the level with the highest frequency) is taken as the imputation value.

[0071] For scenarios with long-term missing data, such as those where the data is missing for more than one sliding time window, the interference of missing data on the model can be reduced by imputing with historical means and adjusting weights.

[0072] Specifically, historical mean imputation involves extracting the historical average of similar objects and process conditions from the MES historical database as the imputation value. For example, for a warping machine, the historical average of "same type of object and process conditions" is extracted from the MES historical database as the imputation value. Uncertainty weight adjustment involves adding an uncertainty weight label (value 0.5-0.8, normal data weight is 1.0) next to the imputed data. The weight is negatively correlated with the duration of missing data (the longer the missing time, the lower the weight). When the GNN model calculates node / edge features, it will reduce the contribution of missing data according to this weight (e.g., the imputation machine speed with a weight of 0.5 only accounts for 50% of the influence weight in the speed matching degree calculation), avoiding erroneous data guiding the model's decision.

[0073] In some embodiments, for S2, the yarn bobbin task node ( The filtering scope includes yarn bobbins currently awaiting allocation on the conveyor line and those about to be assigned (they can enter the allocation process within one sliding time window in the future). Each node corresponds to one independent yarn bobbin, with the yarn bobbin number as a unique identifier to avoid duplication or omission.

[0074] Warping machine nodes ( The filtering scope includes warping machines that are currently available to accept new tasks and those that will soon be idle (estimated remaining processing time ≤ sliding time window duration). Each node corresponds to the current task slot of one warping machine (each warping machine has only one task slot, that is, it can accept a maximum of one yarn package task at a time). The warping machine ID (such as "WJ-08") is used as a unique identifier to exclude unavailable warping machines that are under maintenance, running, and have long remaining processing time.

[0075] For directed edges, obviously incompatible pairs are first excluded to reduce the amount of subsequent calculations. Only the yarn-warping machine pairs that meet the three basic conditions of color matching, path reachability, and tension constraint are retained to construct directed edges (the direction is yarn task node → warping machine node, indicating that the yarn can be transported to the warping machine).

[0076] Specifically, for color matching constraints, such as comparing the color number of the yarn package with the target color number of the warping machine, only the pairing of completely identical or similar color numbers that are permissible by the process (such as some orders that allow the mixing of light and dark colors of the same color family) is retained, and the pairing of color numbers with large differences and easy to produce color difference is directly excluded (such as the warping machine having no corresponding edge for red yarn package and blue target color number).

[0077] For path reachability constraints, for example, based on workshop topology and path information, it is determined whether the current conveyor line of the yarn package can reach the corresponding upper conveyor line of the warping machine through the existing conveyor network. For example, by querying the preset conveyor line connectivity table, if there is at least one unblocked and fault-free path (the current load of the path is ≤90% of the rated capacity to avoid congestion leading to timeout), it is determined to be reachable and the pair is retained; otherwise (such as no connected path or the path is completely congested), it is excluded.

[0078] For tension constraints, for example, comparing the tension level of the yarn package with the allowable tension range of the warping machine, only retaining the pairings where the tension level of the yarn package falls within the allowable tension range of the warping machine or the tension difference is ≤1 level (e.g., when the allowable tension range of the warping machine is 3-4 levels, and the tension level of the yarn package is 2 or 5 levels, the tension difference is 1 level, and the process allows for short-term adaptation), and excluding the pairings where the tension difference is ≥2 levels, which are likely to cause yarn breakage or uneven winding (e.g., there is no corresponding edge between a 5-level high-tension yarn package and a 2-level low-tension warping machine).

[0079] Preferably, the physical attributes, process requirements, and real-time status of the yarn bobbin task are transformed into numerical feature vectors with dimensions adapted to the GNN embedding dimensions (e.g., 64-256 dimensions, with feature terms extended to the corresponding dimensions through linear mapping). Specifically, the yarn bobbin task node includes a standardized speed sequence summary, remaining bobbin length, color number embedding vector, tension level, priority, and current path load features.

[0080] Specifically, for the standardized speed sequence summary, it uses, for example, the average speed within a sliding time window (normalized by z-score) and the speed slope (calculated by linear fitting, such as the slope of 0.0067 m / (min・s) when the speed increases from 20 m / min to 22 m / min in the last 300 seconds, and the value is taken after standardization) to reflect the average level and trend of the yarn conveying speed.

[0081] The remaining bobbin length, for example, is the remaining bobbin length value after being normalized to the [0,1] interval by min-max (e.g., if the full bobbin length is 1000m, the remaining 500m is normalized to 0.5), which reflects the amount of bobbin yarn that can be processed.

[0082] For color code embedding vectors, such as low-dimensional embedding vectors (64-256 dimensions) that map the color code of yarn through one-hot encoding, color code category information is preserved and the curse of dimensionality is avoided.

[0083] For tension levels, such as numerically encoded tension level values ​​(e.g., low tension = 1, medium tension = 3, high tension = 5), it is convenient to compare with the allowable tension range of the warping machine.

[0084] Priority, for example, is a numerically encoded priority value (low priority = 1, normal = 2, urgent = 3), which reflects the urgency of the task.

[0085] For the current path load, such as the real-time load (number of bobbins in transit) of the conveyor line where the bobbins are currently located, after min-max standardization (e.g., if the conveyor line has a rated capacity of 20 bobbins and there are currently 10 bobbins, the standardized value is 0.5), it reflects the congestion risk of the current conveying environment of the bobbins.

[0086] Preferably, the operating status, process requirements, and scheduling needs of the warping machine are transformed into numerical feature vectors. Specifically, the warping machine node includes features such as current machine speed, target color number embedding vector, allowable tension range, status code, priority, and expected idle time.

[0087] Specifically, the current machine speed, such as the real-time spindle speed of the warping machine after z-score standardization (calculated based on the historical average and standard deviation of machine speeds), reflects the current processing speed level.

[0088] For the target color number embedding vector, such as a low-dimensional vector that is homologous to the yarn color number embedding vector (ensuring consistency of the color number feature space), it is convenient to calculate the color number compatibility.

[0089] For the allowable tension range, for example, the "minimum" and "maximum" of the allowable tension range are encoded separately (e.g., if 2-4 levels are allowed, it is encoded as [2,4], and then normalized and extended to the embedding dimension), preserving the tension adaptation interval information.

[0090] For status coding, such as the warping machine task status after numerical coding (idle=1, tail yarn=2, start=3, running=4, maintenance=5), it directly indicates whether the warping machine can take on a new task.

[0091] For priority, a value consistent with the coding rules for yarn winding task priority (e.g., warping machine priority = 3 associated with an emergency order) is used to ensure that the priority dimension can be compared across nodes.

[0092] The estimated idle time, such as the time from when the warping machine finishes the current task to when it can take on a new task (e.g., 10 minutes of remaining processing time plus 2 minutes of downtime for material change, resulting in an estimated idle time of 12 minutes), is normalized to the [0,1] interval by min-max to reflect the time availability of the warping machine.

[0093] Preferably, for the compatibility, historical performance, and transmission cost of the yarn package-warping machine pairing, an edge feature vector is constructed to quantify the quality of the relationship between the two. Specifically, the directed edge includes features such as speed matching degree, path transmission delay estimation, historical pairing success rate, color code compatibility score, and tension penalty value.

[0094] Specifically, for speed matching degree, it can be calculated by, for example, by the average speed of the standardized yarn package / the current speed of the standardized warping machine (if the ratio is in the range of 0.9-1.1, the matching degree = 1.0; the greater the deviation of the ratio, the lower the matching degree is linearly, such as when the ratio is 0.8, the matching degree = 0.8), reflecting the speed synergy (small speed differences can reduce the risk of yarn breakage).

[0095] For path transmission delay estimation, for example, based on workshop topology and path information, the shortest path transmission time of the yarn package from the current position to the warping machine is calculated (e.g., via 2 conveyor lines, the total delay = the sum of empirical values ​​of the path delay of each segment), and then normalized to the [0,1] interval by min-max. The shorter the delay, the higher the score.

[0096] For historical pairing success rate, for example, by querying historical scheduling records, the percentage of successful allocations of this type of yarn (same batch, same color number, same tension level) to this warping machine is calculated out of the total number of allocations (e.g., 9 out of 10 allocations were successful, success rate = 0.9), which directly reflects the historical reliability of pairing.

[0097] For color compatibility scores, for example, if the yarn color is exactly the same as the target color of the warping machine, the score = 1.0; if it is a similar color that is permissible by the process, the score = 0.8-0.9 (adjusted according to the color difference value, the smaller the color difference value, the higher the score); if there is no compatibility relationship, the score = 0 (already excluded in the edge screening stage), thus quantifying the color compatibility risk.

[0098] For tension penalty values, for example, if the tension level of the yarn package falls within the allowable tension range of the warping machine, the penalty value = 0; if the tension difference = level 1, the penalty value = 0.1 (minor penalty); if the tension difference ≥ level 2 (excluded in the side screening stage, this is for redundant verification), the penalty value = 1.0 (severe penalty). The penalty mechanism avoids tension conflicts.

[0099] In some alternative embodiments, to address the issue that static graphs cannot reflect dynamic problems such as changes in yarn package position, warping machine status updates, and path load fluctuations, a time-series dynamic graph is constructed. Specifically, the bipartite graph model is a time-series dynamic graph. By introducing time coding or using a sliding time window, the time-series change characteristics of yarn package tasks and warping machines are integrated into the feature representation of nodes and edges to capture the dynamic evolution of the workshop state.

[0100] Specifically, in one embodiment, the vertex filtering-edge filtering-feature vector generation process is periodically re-executed with a sliding time window duration (e.g., 180-300 seconds) to achieve dynamic updates of the graph structure. For example, for vertex updates, tasks such as removing yarn bobbin nodes that have been assigned and warping machine nodes that have entered maintenance status are performed; new yarn bobbin nodes that have just entered the conveyor line and warping machine nodes that are about to be idle are added to ensure that vertices always cover the currently active scheduling objects.

[0101] For edge updates, directed edges are re-selected based on the latest color code requirements, path load, and tension status (e.g., if a path changes from congested to unobstructed, a corresponding edge is added; if the tension range of a warping machine is adjusted, edges that do not meet the requirements are deleted) to avoid invalid edges based on outdated information.

[0102] For feature vector updates, for example, the feature values ​​of each node and edge are recalculated (such as a decrease in the remaining bobbin length, a shortening of the expected idle time of the warping machine, and changes in path transmission delay) to ensure that the feature vector can reflect the object status in real time.

[0103] In one embodiment, time encoding is added to the node and edge feature vectors. For example, the specific implementation includes relative timestamp encoding, which calculates the relative time of the yarn arriving at the current conveyor line (e.g., arriving 10 minutes ago, relative time = -600 seconds) and the relative time of the warping machine's expected idle time (e.g., idle after 12 minutes, relative time = 720 seconds) based on the current scheduling decision time. The relative time is then mapped to a periodic vector (e.g., sin(2πt / T), cos(2πt / T), where T is the time period, such as 3600 seconds) using the sin / cos function and embedded into the node feature vector.

[0104] Therefore, through the above time-series extension method, the bipartite graph can be updated in real time to follow the changes in the workshop status, ensuring that the subsequent GNN model can output matching scores that fit the reality based on the dynamic graph structure, avoiding the decision lag problem caused by static graphs.

[0105] In some embodiments, for S3, such as Figure 2 As shown, the architecture of the GNN model includes a node embedding layer, a multi-message passing layer, and an edge scoring head. The node embedding layer is used to unify feature dimensions and normalize, the multi-message passing layer is used for bidirectional information flow and attention aggregation (exemplarily, layers 2-4), and the edge scoring head is used to output a comprehensive matching score.

[0106] Specifically, in one embodiment, the node embedding layer will embed the feature vector of the yarn bobbin task node. With warping machine node feature vector Mapping to a hidden state space of the same dimension (e.g., embedding dimensions of 64-256, which can be flexibly adjusted according to the scale of workshop data and computing power) eliminates the differences in the original feature dimensions.

[0107] For a unified feature dimension, for example, the 6-dimensional original features of the yarn bobbin node (standardized speed sequence summary, remaining bobbin length, etc.) and the 6-dimensional original features of the warping machine node (current machine speed, target color embedding, etc.) are extended to a preset embedding dimension through a linear mapping layer (e.g., 6-dimensional → 64-dimensional). During the mapping process, the scattered feature information is integrated into a high-dimensional vector through a learnable weight matrix.

[0108] For normalization, for example, performing LayerNorm normalization on the node feature vectors after linear mapping can be simplified to the following formula: ;in, The characteristic mean, For characteristic variance, For learnable parameters, To prevent tiny values ​​with a denominator of 0, and to avoid some feature values ​​from dominating model learning (e.g., the remaining cylinder length is standardized to take values ​​of 0-1, and the color number embedding vector is taken to take values ​​of -1-1, which can be in the same order of magnitude after normalization).

[0109] After the above processing, the output yarn bobbin node is initially hidden. Initial hidden state of warping machine nodes Both have the same dimensions (e.g., 64 dimensions) and can directly participate in subsequent message transmission.

[0110] In one embodiment, the message passing layer is the core of GNN in capturing the relationships between nodes. In one embodiment, a bidirectional message passing mechanism is adopted (yarn bobbin → warping machine, warping machine → yarn bobbin), and neighbor messages are weighted and aggregated through an attention mechanism. Each layer updates features based on the hidden state of the nodes in the previous layer.

[0111] Specifically, such as Figure 3 As shown, the bidirectional message passing process includes the following steps: S301, setting message passing layer parameters; where, the number of layers is selected, for example, according to the scale of workshop nodes (such as 50-200 yarn bobbin nodes, 10-50 warping machine nodes) it is set to 2-4 layers. Too few layers (<2 layers) cannot fully integrate cross-node information, and too many layers (>4 layers) can easily lead to gradient vanishing and calculation delay. 2-4 layers can balance feature fusion effect and engineering practicality.

[0112] Attention mechanism types, such as graph attention, allow the model to automatically learn the importance weights of different neighboring nodes, avoiding the obscuring of key node information by traditional sum / mean aggregation (e.g., the impact of high-priority yarn on the warping machine should be greater than that of low-priority yarn).

[0113] S302, Forward message passing; yarn package node → warping machine node, the warping machine learns the yarn package features; specifically, taking the l-th layer → l+1-th layer as an example, the first step: generate the message vector from the yarn package to the warping machine for each directed edge from the yarn package node S to the warping machine node M, based on the feature vector of this edge. (Speed ​​matching degree, historical success rate, etc.), Hidden state of the l-th layer of the yarn bobbin node. Hidden state of the first layer of the warping machine node Message vectors are generated through message functions: The symbol "||" indicates feature concatenation (e.g., 64-dimensional + 64-dimensional + 64-dimensional = 192-dimensional). The linear mapping function (192-dimensional → 64-dimensional) compresses the concatenated high-dimensional features to the embedding dimension, while incorporating the association information of the edge features (e.g., if the speed matching degree is high, the corresponding weight in the message vector is greater).

[0114] Step 2: Attention Weight Calculation. For all neighboring yarn bobbins (i.e., yarn bobbins connected to M by an edge) of warping machine node M, calculate the attention weight for each neighbor. It can be simplified to the formula: ;in, This is the attention weight matrix (from 128 dimensions to 1 dimension). For activation function, d is the dimension normalization factor (where d is the embedding dimension). Ensure the sum of the attention weights of all neighbors of warping machine M is 1. High-priority, high-speed matching bobbin nodes will receive higher priority. (e.g., 0.8), while low-priority yarns have a lower weight (e.g., 0.1).

[0115] Step 3: Update the hidden state of the warping machine node by summing the message vectors of all neighbors of the warping machine node M according to the attention weights, and obtain the aggregated message. Then, it interacts with its own l-th hidden state. Merge and update to the (l+1)th hidden state. It can be simplified to the formula: ;in It is a linear transformation matrix. For bias terms, The activation function introduces non-linearity, allowing the warping machine node features to simultaneously contain key information about its own state and neighboring yarns (such as the warping machine being able to sense the allocation needs of high-priority yarns).

[0116] S303, reverse message passing, warping machine node → yarn cone node, yarn cone learns the characteristics of the warping machine, the process is symmetrical with the forward passing, allowing the yarn cone node to perceive the state constraints of the warping machine (such as whether the warping machine is about to be idle, whether the allowable tension range is suitable), avoiding excessive assignment of yarn cones to a certain warping machine.

[0117] Specifically, the message vector from the warping machine to the yarn package is generated: Among them, the feature of reused edges To ensure consistency of related information.

[0118] Calculate the attention weight of the yarn bobbin node to the neighboring warping machine: .

[0119] If the weight of an idle warping machine is higher than that of a running warping machine; update the hidden status of the yarn bobbin nodes: The meanings of the above parameters correspond to the formula parameters in step S202, and will not be repeated here.

[0120] S304, multi-layer iterative update, repeating the bidirectional message passing process 2-4 layers above. Each layer optimizes the feature representation based on the hidden state of the previous layer. As the number of layers increases, the node's hidden state gradually integrates multi-level association information of itself, its direct neighbors, and its indirect neighbors (e.g., yarn bobbins can not only sense the directly connected warping machine, but also sense the load status of other adjacent yarn bobbins through the warping machine). Finally, the node's hidden state at the Lth level (e.g., the 4th layer) is output. (Spindle yarn task) and (Warping machine).

[0121] In one embodiment, the core of the edge scoring head is to calculate the comprehensive matching score score(S,M) of each directed edge S→M based on the node hidden state and edge features of the final layer. The score range is mapped to [0,1] (1 represents the best match, 0 represents the worst match). For example, the specific implementation may include: feature concatenation, including converting the final hidden state of the yarn bobbin node... The final hidden state of the warping machine node and the edge feature vector. The concatenation results in a high-dimensional feature vector, such as 64-dimensional + 64-dimensional + 64-dimensional = 192-dimensional.

[0122] Linear transformation and activation include compression and nonlinear transformation of the concatenated features through a two-layer fully connected network. For example, the first layer maps the 192-dimensional features to 32-dimensional features and activates them using ReLU; the second layer maps the 32-dimensional features to 1-dimensional features, and finally, the output value is constrained to the [0,1] interval by the Sigmoid activation function to obtain the final matching score. .

[0123] The scores comprehensively reflect multiple dimensions of indicators such as speed matching, color compatibility, tension adaptability, historical success rate, and priority matching (e.g., a score of 0.9 indicates that the pairing is optimal in terms of quality, efficiency, and priority, while a score of 0.2 indicates that there is a significant constraint conflict).

[0124] In an optional embodiment, the graph neural network model is trained under supervised training using a multi-objective joint loss function; the joint loss function includes at least: a supervised learning loss based on historical successful pairing data, a ranking comparison loss to ensure that high-quality pairings score higher than low-quality pairings, and a constraint-aware loss to penalize pairings that violate tension or path capacity constraints.

[0125] Specifically, the first step is to prepare training data, including constructing edge-level training samples. Specifically, the historical scheduling time is traced back in time series (e.g., every 5 minutes for the past 6 months), and the corresponding bipartite graph (including node features, edge features, and edge existence) is reconstructed for each time point to simulate the workshop state at that time.

[0126] By sampling positive samples and undersampling negative samples, the ratio of positive to negative samples is ensured to be close to 1:1, avoiding model bias towards the majority class. Positive samples are edges that were actually assigned in history and whose result label is successful. For example, after S1→M1 was actually assigned, there was no yarn breakage or color difference, and the label is y=1. Each positive sample corresponds to one high-scoring edge. Negative samples are edges that were not assigned in history or whose result label after assignment was redistribution / stoppage / color difference. For example, S2→M1 was not assigned, and S3→M2 was assigned and then stopped, and the label is y=0. Each negative sample corresponds to one low-scoring edge.

[0127] In one embodiment, a hybrid loss function combining supervision, comparison, and constraints is used, which simplifies to the following formula: ; These are weighting coefficients, summing to 1 (e.g., 0.3, 0.4, 0.2, 0.1). The specific implementation of each loss term is as follows: Supervised Learning Loss The model fits historical successful pairings; specifically, based on the labels of historical positive / negative samples, the cross-entropy loss is used to measure the difference between the model's predicted score and the true label, which can be simplified to the formula: Where N is the number of edge instances within a batch. This is the model's prediction score for the i-th edge. The loss value is minimized when the model's prediction score for positive samples is close to 1 and for negative samples is close to 0, ensuring that the model learns the feature patterns of historical successful pairings (such as edges with high speed matching degree and high historical success rate are more likely to be predicted as high scores).

[0128] Ranking and comparison loss To ensure the reasonableness of the score ranking, specifically, for multiple candidate yarn packages on the same warping machine or multiple candidate warping machines for the same yarn package, a pairwise hinge loss is used to force high-quality pairs to score higher than low-quality pairs, which is simplified to the formula: Where K is the number of high-quality to low-quality pairings (P) within a batch. Scoring is given for high-quality pairings (such as successfully assigned edges). For poor-quality pairings (such as unassigned edges), the margin is a preset threshold (e.g., 0.2, ensuring the score difference between good and poor-quality pairings is at least 0.2). When When the gap decreases, the loss is 0; otherwise, the loss increases as the gap decreases, to avoid the unreasonable situation where the score of a high-quality pair is lower than that of a low-quality pair (e.g., the score of a successfully assigned edge is 0.6, and the score of an unassigned edge is 0.7).

[0129] Constrained perceived loss To avoid hard constraint conflicts, specifically, for pairings that violate the tension range or exceed the path capacity, a penalty term is introduced to force the model to learn the executable solution space, which simplifies to the formula: ;in This is an indicator function (set to 1 if the i-th edge violates the constraint, otherwise set to 0). `penalty` is the penalty coefficient (ensuring sufficient penalty strength), and `penalty` is a fixed penalty value (e.g., 1.0). For example, if the yarn tension level on a certain edge exceeds the allowable range of the warping machine (violating the constraint), then the constraint loss for that edge is... This significantly increases the total loss, prompting the model to actively reduce the score for such illegal pairings (e.g., from 0.5 to 0.1).

[0130] Regularization loss To prevent overfitting, specifically, L2 regularization is used to constrain model parameters (such as the weight matrices of linear mapping layers and attention layers), as shown in the formula: Where W is the set of all learnable parameters of the model. The L2 norm of the parameter is used. This loss term increases with the absolute value of the parameter, avoiding the model's over-reliance on partial features (such as focusing only on priority while ignoring tension constraints) and improving the model's generalization ability (such as being able to score reasonably when faced with new color yarns).

[0131] In some embodiments, for S4, obviously incompatible pairs are eliminated in advance to reduce the computational load of subsequent global matching. Based on the hard constraints of preset color number compatibility, tension range, and path reachability, hierarchical filtering rules and verification mechanisms are established to ensure that the remaining directed edges meet the basic process and physical constraints of workshop production.

[0132] Specifically, before performing filtering, the baseline parameters required for constraint determination are extracted to ensure that the rules are quantifiable and executable. All parameters are from the same source as the coarse filtering parameters of the S2 step edge screening, but the determination accuracy is higher (S2 is "coarse filtering" to reduce the number of initial edges, and S4 is "fine filtering" to eliminate potential violation edges). For example, the specific parameter sources and settings include: color compatibility baseline: read the color compatibility list of the current production order from the MES process parameters, and clarify the "fully compatible color numbers" (such as the order requires color number #FF0033, only #FF0033 is fully compatible) and the "process-allowed approximately compatible color numbers" (such as some blended orders allow #FF0033 and #FF0044 (one shade lighter than the same color family) to be compatible, and the color difference value must be marked as ≤ΔE2.0, where ΔE is the internationally accepted color difference value standard). Color number combinations that do not meet the list are directly determined to be incompatible.

[0133] Tension range benchmark: Read the "real-time allowable tension range" of each warping machine from the PLC control system of the warping machine (different from the "static allowable range" of S2, this is the dynamic range that takes into account the current operating status of the machine. For example, if the tension accuracy decreases slightly after the warping machine has been running for 1 hour, the allowable range is finely adjusted from 20-25N to 19-26N). At the same time, extract the "actual tension detection value" of each yarn bobbin from the yarn bobbin task table (not the "tension level" of S2, but the real-time value collected by the laser tension meter, in N) to ensure that the tension comparison is based on accurate values ​​rather than levels.

[0134] Path reachability benchmark: Obtain the "current load rate" (number of yarns in transit / rated capacity, e.g., Conveyor-03 has a rated capacity of 20 yarns, currently has 18 yarns in transit, load rate is 90%), "equipment operating status" (normal / fault / maintenance), and "transmission delay threshold" (e.g., the longest allowable transmission delay from the current conveyor line to the hanging point on the warping machine, set according to the expected idle time of the warping machine, e.g., if the warping machine will be idle in 10 minutes, the transmission delay threshold is set to 8 minutes, with a 2-minute allowance for hanging preparation time). Paths that do not meet the physical transmission restrictions are determined to be unreachable.

[0135] Based on the above benchmark parameters, three hard constraint filters are implemented in layers. Specifically, the first layer is color code compatibility filtering, which eliminates edges with color difference risks. For example, a dual judgment logic of exact matching and color difference value verification is adopted to remove directed edges that do not meet the compatibility requirements. The specific steps are as follows: Step 1: Fully compatible color code matching: Compare the "actual color code" of the yarn package node with the "target color code" of the warping machine node. If the two are completely consistent (e.g., yarn package color code #FF0033, warping machine target color code #FF0033), it is directly judged as compatible and the edge is retained; if they are inconsistent, proceed to the second step of approximate compatibility verification.

[0136] Step 2: Approximate compatibility color difference value verification: If the color code of the yarn package is inconsistent with that of the warping machine, check the "Color Code Compatibility List" to confirm whether it falls within the "approximate compatibility allowed by the process".

[0137] If it is not in the list, remove that side directly (e.g., if yarn package #FF0033 and warping machine #00FF00 are not in the list, they are considered incompatible).

[0138] If the yarn is in the list, extract the "historical color difference value record" between the yarn package and the target color number of the warping machine (e.g., the average ΔE1.5 of the same color number combination in the past 3 months). At the same time, collect the "real-time color difference value" of the yarn package (e.g., ΔE1.8) through the workshop color difference meter. If the real-time color difference value is ≤ the threshold marked in the list (e.g., ΔE2.0), it is determined to be compatible and the edge is retained; if the real-time color difference value is > the threshold (e.g., ΔE2.5), it is determined to be incompatible and the edge is removed.

[0139] Optionally, for multi-cylinder warping tasks (which require multiple yarn bobbins to be hung simultaneously), the color difference value between all candidate yarn bobbins on the same warping machine is additionally checked. The color difference value between yarn bobbins is required to be ≤ΔE1.0 to avoid uneven appearance of the finished product due to color difference between yarn bobbins. If the color difference value of any two yarn bobbins exceeds the standard, all relevant edges corresponding to that warping machine are removed.

[0140] The second layer: tension range filtering, eliminating edges with risks of yarn breakage / uneven winding. For example, to ensure stable tension during the warping process and avoid yarn breakage or uneven winding density, a filtering logic based on precise tension values ​​rather than grades is established for numerical range matching and dynamic allowance verification. Directed edges with excessive tension are eliminated. The specific steps are as follows: Step 1: Tension numerical range matching: Extract the actual tension detection value of the yarn package (e.g., 22N) and the real-time allowable tension range of the warping machine (e.g., 19-26N). If the actual tension value of the yarn package falls within the allowable tension range of the warping machine, it is directly determined to be tension compatible and the edge is retained; if it exceeds the range (e.g., yarn package tension of 28N, exceeding the 26N upper limit), proceed to the second step of dynamic allowance verification.

[0141] Step 2: Dynamic margin verification: Considering the slight fluctuations in tension during the operation of the warping machine (such as ±1N), a "dynamic margin threshold" is set (usually 5% of the upper limit of the allowable range, such as an upper limit of the allowable range of 26N, a dynamic margin of 1.3N, and an actual acceptable upper limit of 27.3N).

[0142] If the actual tension value of the yarn package is ≤ the upper limit of the allowable range + dynamic allowance (e.g., 28N>27.3N, exceeding the allowance), or ≥ the lower limit of the allowable range - dynamic allowance (e.g., yarn package tension 17N<19-1.3=17.7N, exceeding the allowance), it is determined to be tension incompatible and the side is removed.

[0143] If the actual tension value of the yarn package is within the "allowable range ± dynamic allowance" (e.g., yarn tension 27N ≤ 27.3N), and the warping machine PLC feedback "current tension control accuracy meets the standard" (e.g., tension fluctuation ≤ ±0.5N), it is judged as compatible and the side is retained; if the tension control accuracy of the warping machine does not meet the standard (fluctuation > ±1N), even if the value is within the allowance, the side is still removed (to avoid fluctuation causing instantaneous out-of-range).

[0144] The third layer: Path reachability filtering, eliminating edges with the risk of delivery timeout / congestion, ensuring that the yarn package can reach the hanging point before the warping machine is idle and avoiding conveyor line congestion. Based on the real-time path status, a triple filtering logic of path connectivity, load rate, and delay is established to remove directed edges that are unreachable or have timeouts. The specific steps are as follows: Step 1: Path connectivity verification: Through the workshop topology and path information, query the "shortest path" from the current conveyor line of the yarn package (e.g., Conveyor-01) to the corresponding hanging conveyor line of the warping machine (e.g., Conveyor-08) (preferably, based on Dijkstra's algorithm, the path with the shortest transmission delay is selected first). If any section of the shortest path is in a "fault" or "maintenance" state (e.g., Conveyor-05 is faulty in the path) and there are no other alternative paths (e.g., no backup path Conveyor-01→Conveyor-06→Conveyor-08), it is determined that the path is not connected and the edge is removed; if there is at least one fully normal path, proceed to Step 2: Load rate verification.

[0145] Step 2: Path load rate verification: Calculate the "average load rate" of all conveyor lines in the shortest path (e.g., if the path includes Conveyor-01 (load rate 70%) and Conveyor-06 (load rate 85%), the average load rate is 77.5%). At the same time, check if there are any "congested conveyor lines with a load rate ≥ 95%" (e.g., Conveyor-07 has a load rate of 98%, which is considered congested).

[0146] If the average load rate is >90% or there is a congested transmission line, the path is considered to be overloaded. Even if the path is connected, it will easily lead to increased transmission delay, and the edge will be removed.

[0147] If the average load rate is ≤90% and there are no congested conveyor lines, proceed to the third step: transmission delay verification.

[0148] Step 3: Transmission delay verification: Based on the empirical value of path transmission delay (e.g., the empirical delay of Conveyor-01→Conveyor-06 is 2 minutes, the empirical delay of Conveyor-06→Conveyor-08 is 3 minutes, and the total empirical delay is 5 minutes), the delay is corrected in combination with the current load rate (for every 10% increase in load rate, the delay increases by 10%, such as an average load rate of 80%, the delay is corrected to 5 minutes × (1 + 0%) = 5 minutes; a load rate of 90%, the delay is corrected to 5 minutes × (1 + 10%) = 5.5 minutes), thus obtaining the "corrected transmission delay".

[0149] If the corrected transmission delay is less than or equal to the warping machine's "transmission delay threshold" (e.g., if the warping machine is idle after 10 minutes, the threshold is 8 minutes, and 5.5 minutes ≤ 8 minutes), the path is considered reachable, and the edge is retained.

[0150] If the corrected transmission delay is greater than the threshold (e.g., the corrected delay of 9 minutes is greater than 8 minutes), it is determined to be a transmission timeout (the warping machine has already started the next task when the yarn arrives), and that side is removed.

[0151] Therefore, through the above process, the S4 stage can effectively eliminate incompatible edges, significantly reducing the computational complexity of global matching in the subsequent S5 stage, while ensuring that all remaining edges meet the basic production constraints, laying the foundation for generating a feasible scheduling scheme.

[0152] In some embodiments, for S5, based on the bipartite graph filtered in S4, the yarn package-warping machine assignment is transformed into a mathematically solvable weighted bipartite graph matching problem. Among them, the vertex set: retains the set of yarn spinning task nodes defined in S2. (The number of yarn packages to be allocated after filtering is denoted as n), and the set of warping machine nodes. (The number of warping machines that can accept tasks after filtering is denoted as m), ensuring that the vertices contain only valid scheduling objects.

[0153] Edge set: Only the directed edges E' after S4 filtering (edges that satisfy color number, tension, and path constraints) are retained. The corresponding candidate scheme for assigning the yarn package s to the warping machine m.

[0154] Weights: The weight of each edge (s,m) is the matching score score(s,m) output by the GNN model in S3 (values ​​0-1). The higher the score, the better the overall fit (quality, efficiency, priority) of the pair, which is the basis for maximizing the total matching score.

[0155] Constraints transform actual production limitations into mathematical constraints, preventing invalid solutions such as "one yarn, multiple warping machines" or "one machine exceeding its capacity." Specific constraints include: yarn bobbin allocation constraints, for example, each yarn bobbin task can be assigned to at most one warping machine, that is, for any yarn bobbin s, at most one edge among all associated edges (s,m) is selected (denoted by the allocation variable). This indicates that the allocation was successful. (Indicating unassigned), the mathematical expression is: .

[0156] Warping machine capacity constraints, for example, mean that the number of tasks received by each warping machine cannot exceed the capacity of its task slot. Warping machines are usually single-task slots (can only handle one yarn package task at a time). That is, for any warping machine m, at most one edge among all associated edges (s,m) is selected. The mathematical expression is: .

[0157] To achieve the highest overall global matching score, construct the objective function: That is, by choosing the appropriate (0 or 1) maximizes the sum of the weights of all selected edges, ensuring that the final allocation scheme achieves optimal fit overall, rather than local optimum (e.g., avoiding selecting high-scoring yarn for a warping machine, which would result in multiple yarns not being matched with warping machines).

[0158] Choose an appropriate solution based on the scale of the yarn package-warping machine (i.e., the size of n and m), balancing "solution accuracy" and "computational efficiency." Specifically, for small-scale scenarios ( and For example, when the number of yarn packages to be allocated and the number of warping machines that can be accepted are small (such as local scheduling in the workshop or emergency order replenishment scenarios), the Hungarian algorithm is used to solve the global optimum. Specifically, (1) Algorithm preprocessing: Constructing the cost matrix: Since the standard Hungarian algorithm is usually used for "minimum cost matching", it is necessary to transform "maximizing the total score" into "minimizing the total loss": Calculate the maximum score threshold max_score (take the maximum value of all edge score(s,m), such as 0.95); Construct The loss matrix C, matrix elements (If (s,m) is not filtered afterward, Set it to max_score+1 to indicate that this pairing is not optional. At this point, "minimize the total loss" "This is equivalent to "maximizing the total score Z".

[0159] (2) Execute the Hungarian algorithm: Step 1, matrix row reduction: Subtract the minimum value of each row of the loss matrix so that each row has at least one 0 element (representing the relatively best candidate pair in the row).

[0160] Step 2, Matrix column reduction: Subtract the minimum value of each column of the row-reduced matrix to ensure that each column has at least one 0 element, further simplifying the matching selection.

[0161] Step 3, Cover 0 elements: Cover all 0 elements with the fewest horizontal and vertical lines. If the number of covering lines is equal to min(n,m) (the smaller value between the number of yarn packages and the number of warping machines), a match can be found directly; otherwise, proceed to step 4.

[0162] Step 4, Adjust the matrix: Calculate the minimum value of the uncovered elements, subtract the minimum value from the uncovered rows, and add the minimum value to the covered columns. Repeat step 3 until the number of covering lines equals min(n,m).

[0163] Step 5, Find the optimal match: Select a pair from the position containing 0 elements, ensuring that each yarn package is selected only once by the warping machine. This minimizes the total loss, i.e., maximizes the total score. Output. The set of pairs.

[0164] The Hungarian algorithm has a time complexity of O(k^3 (k=min(n,m))). The computation time can be controlled within 1 second (meeting the real-time scheduling needs of the workshop) and can guarantee the global optimal solution. It is suitable for small-scale scenarios with high accuracy requirements (such as high-priority order scheduling).

[0165] For large-scale scenarios ( or When there are many yarn packages to be allocated or warping machines that can be accepted (such as batch scheduling in the workshop at all times), the computational workload of the Hungarian algorithm increases dramatically (such as the complexity reaching 125,000 when k=50). This embodiment adopts the minimum cost flow algorithm based on maximum flow, which improves computational efficiency while ensuring "approximate optimality". The specific steps include: (1) Constructing a flow network model: Transforming the bipartite graph into a flow network of "source point-yarn package-warping machine-sink point" to realize the mapping between "cost" and "flow". Specifically, node addition: Add source point S (the starting point of the flow) and sink point T (the ending point of the flow). The source point connects all yarn package nodes, and all warping machine nodes connect to the sink point. Edge and capacity and cost settings: Edge from source point S to yarn package s: capacity is set to 1 (each yarn package can flow out a maximum of 1 unit flow, corresponding to 1 allocation), cost is set to 0 (no additional cost); Edge from yarn package s to warping machine m (only the filtered edge is retained): capacity is set to 1 (each pair can flow out a maximum of 1 unit flow), cost is set to (Consistent with the Hungarian algorithm, converting scores into costs); Edge from warping machine m to sink T: capacity set to 1 (each warping machine receives at most 1 unit flow, corresponding to 1 acceptance), cost set to 0.

[0166] (2) Perform minimum cost maximum flow solution: The goal is to find the maximum flow from S to T (i.e. the maximum number of yarn bobbins to warping machine pairs that can be allocated) and the minimum total cost (i.e. the maximum total score) under the maximum flow. Specifically, the successive shortest augmenting path algorithm is adopted. The specific steps include: Step 1, initialization: set the flow of all edges in the flow network to 0, the total cost to 0, and the maximum flow to 0.

[0167] Step 2, find the shortest path: Use the Bellman-Ford algorithm or the SPFA algorithm (suitable for scenarios with negative cost edges) to find the minimum cost path from S to T (the path with the minimum sum of costs for all edges).

[0168] Step 3, augmenting flow: In the shortest path found, calculate the maximum augmenting flow (here all edge capacity is 1, so the augmenting flow is 1), add 1 to the flow of each edge on the path, add the path cost to the total cost, and add 1 to the maximum flow.

[0169] Step 4, Repeat the iteration: Repeat steps 2-3 until no augmenting path from S to T can be found (i.e., the maximum flow is reached). At this point, the total cost is minimized, and the corresponding flow path is the optimal match (the path from source point → yarn bobbin → warping machine → sink point corresponds to...). ).

[0170] The algorithm has a time complexity of O(n). (F is the maximum flow, E is the number of edges, and V is the number of nodes). When n=100 and m=50, the maximum flow is usually ≤50, the calculation time can be controlled within 3 seconds (meeting real-time requirements), and it can handle a large number of nodes, making it suitable for full-scale scheduling scenarios in the workshop. Although it cannot theoretically guarantee absolute global optimality, in practical applications, the score difference from the global optimal solution is usually ≤5%, and the accuracy can meet production needs.

[0171] Extract from algorithm output All pairings form a “Bottle-Warping Machine Allocation List”, which, for example, includes: yarn number, warping machine ID, matching score, corresponding transport path (the shortest path for the pairing extracted from the S2 workshop topology and path information), estimated transmission delay (the corrected delay extracted from the S4 path reachability check), and estimated hanging time (current time and estimated transmission delay).

[0172] In some embodiments, for S6, the allocation result of each pair of "yarn package-warping machine" is decomposed into conveyor control instructions, warping machine preparation instructions, and hanging coordination instructions to ensure coverage of the entire process of yarn package from the current position to the warping machine processing. The instruction format is consistent with the original production control of the production line, and will not be described in detail here.

[0173] Preferably, in one embodiment, in order to detect deviations in a timely manner and provide early warnings of anomalies, two types of monitoring indicators are set up and compared in real time with the predicted values ​​of the GNN model in S3.

[0174] Specifically, key timing indicators include: monitoring "actual yarn package transmission delay" (actual time of yarn package leaving the current conveyor line - predicted time) and "actual warping machine idle time" (actual time of warping machine completing the current task - predicted time). The comparison threshold is set to 20%, that is, when the deviation between the actual value and the predicted value exceeds 20%, an early warning is triggered (e.g., the predicted transmission delay is 5 minutes, the actual delay is 6.1 minutes, and the deviation is 22%).

[0175] Key status indicators: Monitor "actual tension value of yarn package" (compared with the tension matching value predicted by GNN), "actual load rate of conveyor line" (compared with the predicted path load rate), and "actual speed of warping machine" (compared with the predicted processing speed). The comparison threshold is set to 10% (e.g., predicted tension is 22N, actual tension is 24.5N, deviation is 11.4%).

[0176] When any indicator deviation exceeds the threshold, or a fault signal is received (such as conveyor line shutdown or warping machine tension loss), the abnormal response process is immediately triggered. The first step is to determine the scope of replanning (local / global), including the following local replanning determination conditions: the abnormality only affects a "single bobbin" or a "local equipment cluster" (such as a bobbin with an abnormal speed during Conveyor-05 transmission, which only affects the bobbin and the corresponding warping machine 1 (ID WJ-08); or a Conveyor-05 fault, which only affects the 3 bobbins that need to pass through the conveyor line), and does not affect the normal operation of other conveyor lines and warping machines, it is determined as a "local abnormality" and local replanning is initiated.

[0177] Global replanning criteria: If an anomaly affects "global resources" (such as fluctuations in the workshop's main power supply causing multiple conveyor lines to stop simultaneously, or the urgent insertion of high-priority orders causing global priority adjustments), and more than 50% of the allocation and pairing are affected, it is judged as a "global anomaly" and global replanning is initiated (S2-S5 are re-executed).

[0178] Among them, such as Figure 4 As shown, the process of local replanning, for example, specifically includes: (1) S401, construction of affected subgraph: subgraph vertex filtering: only retain the yarn bobbin nodes affected by the abnormality (such as yarn bobbins with abnormal speed, yarn bobbins that cannot be transmitted due to path failure) and "affected warping machine nodes" (such as the original warping machine corresponding to the yarn bobbin, and the spare warping machine that can accept the affected yarn bobbin), and exclude the nodes that are not affected (such as other yarn bobbins that are transmitted normally, and warping machines that have been hung normally).

[0179] Subgraph edge filtering: Based on the graph modeling logic of S2, only the directed edges of the affected nodes are reconstructed. The filtering conditions are the same as those of S4 (color compatibility, tension adaptation, path reachability), but the "backup path" or "backup warping machine" is selected first (e.g., if Conveyor-05 fails, the path Conveyor-06→Conveyor-08 is replanned for the affected yarn packages).

[0180] (2) S402, Subgraph edge scoring and matching: Edge score update: Input the constructed subgraph into the pre-trained GNN model in S3, recalculate the matching score of all candidate edges in the subgraph, and at this time the model will automatically increase the adaptation weight of "backup resources" (such as the score of backup warping machine 2 (ID is WJ-09) will be adjusted according to the current idle state).

[0181] Subgraph matching solution: The S5 algorithm is used (Hungarian algorithm for small-scale subgraphs and minimum cost flow algorithm for large-scale subgraphs) to solve the local optimal matching that "maximizes the total matching score" within the subgraph and generate a "local adjustment scheme" (such as reassigning the yarn TS-20240510-001 originally assigned to warping machine 1 (ID WJ-08) to warping machine 2 (ID WJ-09)).

[0182] (3) S403, Adjustment command issuance and execution connection: New control commands are generated only for the equipment involved in the local adjustment scheme (such as generating a preparation command for warping machine 2 (ID is WJ-09) and a transmission command for Conveyor-06). Unaffected equipment continues to execute the original command to avoid global command restart.

[0183] Set an adjustment connection time window (e.g., issue a new instruction within 2 minutes after the original instruction is interrupted) to ensure that the processing of the affected yarn packages is not interrupted (e.g., after the yarn package stops at Conveyor-05, it is switched to Conveyor-06 for transmission within 2 minutes, with no significant waiting time).

[0184] Therefore, through the aforementioned local replanning process, abnormal problems can be quickly resolved without interfering with the normal global scheduling, ensuring the continuity and stability of the warping workshop scheduling.

[0185] Figure 5 An intelligent scheduling and management system 500 for a warping workshop is shown. This device embodiment is similar to... Figure 1 Corresponding to the method embodiment shown, it specifically includes: a warping workshop data acquisition module 501, used to acquire warping workshop data, the warping workshop data including yarn bobbin task data, warping machine status data, workshop topology and path information, and historical scheduling records.

[0186] The bipartite graph construction module 502 is used to build a bipartite graph with the yarn package task node and the warping machine node as vertices. It establishes directed edges based on color matching, path reachability and tension constraints, and generates feature vectors for nodes and edges.

[0187] The matching score module 503 is used to input the bipartite graph into a pre-trained graph neural network model, wherein the graph neural network model encodes and aggregates node and edge information through a node embedding layer and multiple message passing layers, and finally outputs a matching score representing the overall matching quality for each directed edge through an edge scoring head.

[0188] The directed edge filtering module 504 is used to filter all the directed edges based on preset hard constraints, and remove edges that do not meet the requirements of color compatibility, tension range or path reachability.

[0189] The allocation scheme acquisition module 505 is used to take the filtered directed edges and their matching scores as input, construct and solve a weighted bipartite graph global matching problem with the goal of maximizing the total matching score, so as to obtain the optimal allocation scheme of yarn bobbin tasks to warping machines.

[0190] The execution module 506 is used to convert the allocation scheme into specific equipment control instructions and send them to the workshop execution system for execution.

[0191] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent scheduling and management of a warping workshop, characterized in that, Includes the following steps: S1. Acquire warping workshop data, including yarn package task data, warping machine status data, workshop topology and path information, and historical scheduling records; S2. Establish a bipartite graph with yarn package task nodes and warping machine nodes as vertices, establish directed edges based on color matching, path reachability, and tension constraints, and generate feature vectors for nodes and edges; S3. Input the bipartite graph into a pre-trained graph neural network model, wherein the graph neural network model encodes and aggregates node and edge information through a node embedding layer and multiple message passing layers, and finally outputs a matching score representing the overall matching quality for each directed edge through an edge scoring head; S4. Filter all the directed edges based on preset hard constraints, and remove edges that do not meet color compatibility, tension range, or path reachability requirements. S5. Using the filtered directed edges and their matching scores as input, construct and solve a weighted bipartite graph global matching problem with the goal of maximizing the total matching score, thereby obtaining the optimal allocation scheme of yarn bobbin tasks to the warping machine; S6. Convert the allocation scheme into specific equipment control instructions and send them to the workshop execution system for execution.

2. The intelligent scheduling and management method for a warping workshop according to claim 1, characterized in that, It also includes preprocessing the warping workshop data, including setting a sliding time window to collect time-series data, standardizing continuous numerical features, numerically encoding or embedding discrete categorical features, and filling in missing data.

3. The intelligent scheduling and management method for a warping workshop according to claim 1, characterized in that, The yarn bobbin task node includes a standardized speed sequence summary, remaining bobbin length, color number embedding vector, tension level, priority, and current path load characteristics; the warping machine node includes current machine speed, target color number embedding vector, allowable tension range, status code, priority, and expected idle time characteristics; the directed edge includes speed matching degree, path transmission delay estimation, historical pairing success rate, color number compatibility score, and tension penalty value characteristics.

4. The intelligent scheduling and management method for a warping workshop according to claim 1, characterized in that, include: The bipartite graph is a time-series dynamic graph. By introducing time coding or using a sliding time window, the time-series changes of the yarn winding task and the warping machine are integrated into the feature representation of nodes and edges to capture the dynamic evolution of the workshop state.

5. The intelligent scheduling and management method for a warping workshop according to claim 1, characterized in that, The graph neural network model adopts a bidirectional message passing mechanism, including forward message passing from the yarn package task node to the warping machine node, and reverse message passing from the warping machine node to the yarn package task node. The message passing integrates the hidden state features of connected nodes and the features of edges, and uses an attention mechanism to perform weighted aggregation of messages from different neighboring nodes.

6. The intelligent scheduling and management method for a warping workshop according to claim 1, characterized in that, The graph neural network model is trained using a multi-objective joint loss function for supervised training. The joint loss function includes at least: a supervised learning loss based on historical successful pairing data, a ranking comparison loss to ensure that high-quality pairings score higher than low-quality pairings, and a constraint-aware loss to penalize pairings that violate tension or path capacity constraints.

7. The intelligent scheduling and management method for a warping workshop according to claim 1, characterized in that, To solve the global matching problem of weighted bipartite graphs, the Hungarian algorithm or the minimum cost flow algorithm based on maximum flow is adopted to find the globally optimal matching under the constraints that each yarn bobbin task is assigned at most once and the number of tasks received by each warping machine does not exceed its task slot capacity.

8. The intelligent scheduling and management method for a warping workshop according to claim 1, characterized in that, During execution, the system status is monitored in real time and compared with the prediction of the graph neural network. When the deviation exceeds the threshold, a local or global replanning process is triggered, and steps S2 to S5 are re-executed to generate a new scheduling scheme.

9. The intelligent scheduling and management method for a warping workshop according to claim 8, characterized in that, The replanning process is a local replanning. When an anomaly is detected, edge scoring and matching decisions are re-performed only for the subgraph formed by the yarn bobbin task nodes and warping machine nodes affected by the anomaly, while the scheduling scheme of the unaffected parts of the system remains unchanged.

10. An intelligent scheduling and management system for a warping workshop, characterized in that, include: The warping workshop data acquisition module is used to acquire warping workshop data, which includes yarn package task data, warping machine status data, workshop topology and path information, and historical scheduling records; the bipartite graph construction module is used to build a bipartite graph with yarn package task nodes and warping machine nodes as vertices, establish directed edges based on color matching, path reachability and tension constraints, and generate feature vectors for nodes and edges; The matching score module is used to input the bipartite graph into a pre-trained graph neural network model, wherein the graph neural network model encodes and aggregates node and edge information through a node embedding layer and multiple message passing layers, and finally outputs a matching score representing the overall matching quality for each directed edge through an edge scoring head. The directed edge filtering module is used to filter all the directed edges based on preset hard constraints, and remove edges that do not meet the requirements of color compatibility, tension range or path reachability. The allocation scheme acquisition module is used to take the filtered directed edges and their matching scores as input, construct and solve a weighted bipartite graph global matching problem with the goal of maximizing the total matching score, so as to obtain the optimal allocation scheme of yarn bobbin tasks to warping machines; the execution module is used to convert the allocation scheme into specific equipment control instructions and send them to the workshop execution system for execution.

Images