Abnormal propagation graph-based clothing production line self-healing scheduling method, device and medium

Abstract

This invention relates to the field of intelligent manufacturing and production scheduling technology in the apparel industry. It discloses a self-healing scheduling method, equipment, and medium for apparel production lines based on anomaly propagation graphs. The method operates on a multi-agent hierarchical decision-making architecture and includes: fusing multi-source production data for anomaly perception; combining manufacturing context with a large-scale apparel industry model for semantic diagnosis; constructing a process dependency graph, calculating edge weights based on buffer margins, and generating a ripple subgraph layer by layer along directed edges to achieve quantitative assessment of the ripple range; obtaining the ripple subgraph and order priorities to combine and generate candidate scheduling schemes, injecting the candidate schemes into a virtual production line simulation model for three-dimensional scoring to select the optimal scheme; issuing and executing the optimal scheme, constructing the processing process as a case study and storing it in a database, and updating the edge weights of the anomaly propagation graph in reverse. This invention solves the problem of subjective evaluation bias, reduces the probability of secondary conflicts, and achieves continuous self-evolution of experience accumulation and system accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and production scheduling technology for clothing, specifically to a self-healing scheduling method, equipment, and medium for clothing production lines based on anomaly propagation graphs. Background Technology

[0002] In the personalized customization production model of clothing, the process path of an order often varies due to differences in style, fabric, and craftsmanship requirements, resulting in fragmented orders and diversified paths on the production line. To maintain the operation of the production process, manufacturing companies typically deploy manufacturing execution systems and advanced planning and scheduling systems for production line scheduling management, which are used to coordinate and distribute various processing tasks.

[0003] In actual production operations, production anomalies such as equipment failures, material delays, or staff absences occur frequently. Existing production scheduling systems often rely on manual monitoring and response mechanisms to handle these anomalies. Specifically, after an anomaly occurs, it is typically reported by on-site operators via terminal devices, or an alarm is issued by the basic monitoring system. Upon receiving the anomaly information, the production scheduler analyzes the current production line status based on their experience, assesses the potential impact of the anomaly, and formulates corresponding adjustment plans for the production line to implement.

[0004] Among the previously disclosed technical solutions, CN121563061A discloses a real-time response method and system for production anomalies based on graph structures. This method constructs a directed graph based on a product structure tree and resource model, describing the scope of anomaly impact based on the dependencies between components, processes, and resources. After detecting a production anomaly, a graph traversal algorithm is used to propagate the impact from the anomaly node to its upstream and downstream components, determining the set of affected nodes and performing local scheduling recalculation for the affected area. The core of this solution lies in using graph structures to achieve production anomaly impact scope analysis and local scheduling response, primarily targeting rapid anomaly recovery in general manufacturing scenarios. CN120338430B discloses a digital collaborative management method and system for MES based on deep learning. By collecting multi-source heterogeneous production data, it constructs a production line operation dependency graph using node feature embedding and edge relationship learning. Combined with anomaly detection, root cause localization, digital twin simulation, and iterative scheduling models, it achieves anomaly analysis and collaborative optimization in the production process. The core of this solution lies in data-driven production state modeling and multi-objective scheduling optimization, targeting digital collaborative management at the MES system level.

[0005] However, existing anomaly scheduling methods suffer from low accuracy in assessing the scope of anomaly impact in practical applications. Faced with complex process dependencies in fragmented order scenarios, current technologies lack topological modeling methods for material flow and workstation sharing between processes. They fail to incorporate the strength of dependencies and buffer states between processes into quantitative calculations, making it impossible to quantitatively analyze the multi-level propagation effects of anomalies along process paths when relying solely on the scheduler's personal experience. This experience-dependent assessment method is prone to subjective judgment bias, preventing the management system from accurately defining the actual impact of anomalies on downstream processes and related orders. Consequently, subsequent production rescheduling plans lack objective data support. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a self-healing scheduling method, equipment, and medium for garment production lines based on anomaly propagation graphs. This solves the problems of delayed anomaly perception, difficulty in quantifying and assessing the scope of impact, lack of prior verification of scheduling schemes, and inability to accumulate and reuse anomaly handling experience in existing personalized garment production lines.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides a self-healing scheduling method for a garment production line based on anomaly propagation graphs, which operates on an underlying multi-agent hierarchical decision-making architecture. The method includes:

[0008] Multi-source anomaly perception steps: The sensing agent collects and integrates the production data continuously generated by the garment production line in real time, extracts statistical features and verifies them, and outputs a list of abnormal events;

[0009] Anomaly semantic diagnosis steps: The diagnostic agent receives the list of abnormal events, extracts the manufacturing context information of the workstation corresponding to the abnormal event in the list of abnormal events, concatenates the abnormal event information and the manufacturing context information into a structured description text, and inputs it into a pre-deployed large model of the apparel industry for semantic diagnosis, and outputs a diagnostic result containing the estimated recovery time range.

[0010] Anomaly propagation graph construction and impact assessment steps: Extract the production line process dependency topology and build a process dependency graph, calculate the edge weights of directed edges in the process dependency graph; based on the diagnosis results, take the process where the anomaly occurred as the source node, propagate it layer by layer along the directed edges in the process dependency graph, deduce the impact probability and impact time window of each downstream node, prune the nodes whose impact probability is lower than the pruning threshold and their downstream subtrees to obtain the impact subgraph;

[0011] Candidate scheduling scheme combination generation steps: The scheduling agent obtains the ripple subgraph and order priority data, retrieves historical cases in the strategy knowledge base, and when the case reuse condition is not met, it selects applicable action subsets from the preset strategy action library and arranges and combines them to generate a candidate scheme set. The apparel industry big model sorts the candidate scheme set according to the order priority data.

[0012] Digital twin pre-simulation verification steps: The sorted candidate solutions are sequentially injected into the virtual production line simulation model containing the production line digital mapping, discrete event simulation is performed, and the comprehensive score of each candidate solution is calculated from three dimensions: completion time offset, equipment utilization rate and secondary conflict risk. The candidate solution with the highest comprehensive score is selected as the optimal solution.

[0013] The execution steps include: the execution agent breaks down the optimal solution into specific execution instructions and sends them to the corresponding workstation terminals; after the solution is executed, the current anomaly handling process is constructed as a triplet case and stored in the policy knowledge base, and the edge weight parameters of the anomaly propagation graph are updated in reverse according to the actual observed propagation intensity.

[0014] Preferably, in the multi-source anomaly detection step, the production data includes equipment IoT data, quality inspection data, material arrival data, and personnel attendance data; the collected time-series data is traversed using a preset sliding window, and the mean offset and variance change rate of the signal sequence are calculated within each time sliding window. The signal sequence within the window is compared with a statistical threshold calculated based on the mean and standard deviation, and data segments exceeding the statistical threshold are filtered out; the data segments exceeding the statistical threshold are input into a lightweight time-series detection model for secondary confirmation, anomaly confidence is output, and an anomaly event list is generated. Each record in the anomaly event list is a quadruple containing an anomaly type code, an occurrence workstation identifier, an occurrence time, and the anomaly confidence.

[0015] Preferably, in the anomaly semantic diagnosis step, the manufacturing context information includes the work order number, process type, the position sequence number of the corresponding process in the order process path, and the buffer work-in-process quantity of upstream and downstream processes; extract the field template of the structured description, and concatenate the quadruple containing the anomaly event information with the manufacturing context information according to the field template of the structured description to generate the structured description text; combine the temperature parameter inference of the large model to control the garment industry large model to perform semantic diagnosis of anomalies based on knowledge of the garment production field, and the diagnosis result also includes the probability distribution vector of the anomaly root cause category output by the garment industry large model.

[0016] Preferably, in the anomaly propagation graph construction and ripple assessment step, the process dependency graph is constructed using all processes on the production line as the node set and the material flow relationship and workstation sharing relationship between processes as the directed edge set; real-time data of buffer margin between each process is extracted to calculate the edge weight of the directed edge, the edge weight being equal to the product of the inter-process dependency strength and the negative exponential buffer decay coefficient, wherein the negative exponential buffer decay coefficient is calculated based on the buffer decay coefficient and the ratio of the current buffer work-in-process quantity to the standard buffer quantity; the upper limit of the estimated recovery time interval is set to... The initial value of the source node's influence time window is defined. For affected processes with multiple predecessor nodes, the influence probability is calculated by combining the independent events of the influence probabilities of each predecessor node. The influence probability is equal to the difference between the product of the probability that each predecessor node did not experience an influence and the product of the probability that each predecessor node did not experience an influence and the corresponding directed edge weight. At the same time, the influence time window of each node is calculated. The influence time window of each affected node is equal to the maximum value of the sum of the influence time windows of its predecessor nodes and the standard flow time between processes.

[0017] Preferably, in the candidate scheduling scheme combination generation step, the pre-set strategy action library includes five basic scheduling actions: process rerouting actions, workstation mutual assistance actions, task resegmentation actions, buffer release actions, and outsourcing switching actions. Based on the process type, equipment specialization, workstation substitutability, and personnel skill interoperability attributes of the affected processes, a subset of actions meeting the applicable preconditions is selected from the five basic scheduling actions and arranged in combination to generate the candidate scheme set. The priority score of the affected orders is calculated, where the priority score equals the product of the normalized value of the remaining days until delivery and a first weight coefficient, plus the sum of the product of the normalized value of the customer level and a second weight coefficient. The apparel industry big data model is controlled to sort the candidate scheme set according to the priority score and optimize the time parameters within the scheme.

[0018] Preferably, in the digital twin pre-simulation verification step, the process allocation and resource configuration are adjusted according to the parameters set by the candidate schemes, and the simulation is performed within the virtual production line simulation model until all affected orders are completed or the simulation time limit is reached; the simulation data is extracted, and the completion time offset score, equipment utilization rate score, and secondary conflict risk score are calculated respectively. The scores of the three dimensions are multiplied by the corresponding dimension weight coefficients and then weighted and summed to obtain the comprehensive score; wherein, the secondary conflict risk score is equal to the difference between one and the ratio; if the comprehensive score of all candidate schemes is lower than a preset threshold, the action combination space is expanded to regenerate candidate schemes, and a manual intervention request is generated after the number of rollbacks reaches the preset rollback limit.

[0019] Preferably, in the execution and case knowledge accumulation steps, the same three-dimensional scoring function as the digital twin pre-simulation verification is used to calculate the actual recovery effect score by replacing the simulation data with actual execution data; a triplet case containing anomaly feature vectors, a structured description of the scheduling scheme, and the actual recovery effect score is constructed, and the triplet case is stored in the strategy knowledge base; the actual observed propagation intensity is derived by back-calculating the ratio of the actual affected time window to the predicted time window, and a new dependency intensity is calculated using an exponential moving average method. The new dependency intensity is equal to the product of the old dependency intensity before the update and the first proportional coefficient, plus the product of the actual observed propagation intensity and the learning rate. The first proportional coefficient is equal to the difference between the first and the learning rate. The edge weight parameters of the anomaly propagation graph are then updated with the new dependency intensity.

[0020] A second aspect of the present invention provides a garment production line self-healing scheduling device based on anomaly propagation graph, comprising a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement all the steps of the garment production line self-healing scheduling method based on anomaly propagation graph provided in the first aspect of the present invention.

[0021] A third aspect of the present invention provides a non-volatile computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements all the steps of the garment production line self-healing scheduling method based on anomaly propagation graph provided in the first aspect of the present invention.

[0022] This invention provides a self-healing scheduling method, equipment, and medium for garment production lines based on anomaly propagation graphs. It offers the following advantages:

[0023] 1. This invention constructs an anomaly propagation graph by using process steps as nodes and material flow and workstation sharing relationships as directed weighted edges. It determines edge weights by combining dependency strength and attenuation coefficients calculated based on buffer margins. Then, it uses the process step in which the anomaly occurs as the source node and deduces the probability of downstream processes being affected and the time window of the impact along the directed edges to generate a ripple subgraph. This enables a quantitative assessment of the impact range of multi-level linkage propagation of abnormal events along complex process paths in the garment customization production line, and solves the problem of subjective judgment bias caused by relying on manual experience assessment.

[0024] 2. This invention injects the generated candidate solutions into a virtual production line simulation model containing digital mapping to perform simulations. It extracts simulation data from three dimensions—completion time offset, equipment utilization rate, and secondary conflict risk—to calculate a comprehensive score to select the optimal solution. Furthermore, when the solution score is lower than a preset threshold, a rollback mechanism that expands the action combination space is triggered. This achieves pre-verification of the rescheduling solution before actual issuance, reduces the probability of secondary conflicts after execution, and improves the first-time execution success rate of the rescheduling solution.

[0025] 3. This invention constructs the actual anomaly handling process after execution, along with the actual effect score, into a triplet case and stores it in the strategy knowledge base for subsequent reuse. At the same time, it infers the actual propagation intensity based on the ratio of the actual affected time window to the predicted time window, and uses the exponential moving average method to calculate the new dependency intensity to dynamically update the edge weight parameters of the anomaly propagation graph. This achieves the structured accumulation of historical anomaly handling experience, which makes the accuracy of impact assessment continuously improve with system operation and reduces the computational resource consumption when handling similar anomalies repeatedly. Attached Figure Description

[0026] Figure 1 This is an overall flowchart of the method of the present invention;

[0027] Figure 2 A comparison chart of F1 scores for the impact assessment of three types of abnormal events;

[0028] Figure 3 A diagram illustrating the improvement in scheduling scheme quality achieved through digital twin simulation.

[0029] Figure 4 A trend chart showing the improvement in system efficiency based on case knowledge accumulation. Detailed Implementation

[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] See attached document Figure 1-4The self-healing scheduling method for garment production lines based on anomaly propagation graphs relies on a multi-agent hierarchical decision-making architecture. In terms of physical deployment, this hierarchical architecture resides on edge computing nodes or cloud servers with data processing and storage capabilities. The servers establish bidirectional communication connections with various sensors, programmable logic controllers (PLCs), and operating terminals on the garment production line via an industrial local area network (LAN), and interact with the production line's existing manufacturing execution system (MES) and advanced planning and scheduling system. Logically, the hierarchical decision-making architecture is divided into distributed sensing agents, diagnostic agents, scheduling agents, and execution agents. Each agent runs in the computing platform as a background service or an independent thread, synchronizing its state and interacting with data through internal message queues, shared memory, or system buses to achieve instruction transmission and coordinated execution.

[0032] Based on the above multi-agent hierarchical decision-making architecture, the production line self-healing scheduling method provided in this embodiment includes the following steps:

[0033] Step S1, Multi-source Anomaly Detection. The underlying equipment of the garment production line continuously generates production data, which is collected and integrated in real time by the sensing agent. Production data includes equipment IoT data, quality inspection data, material arrival data, and personnel attendance data.

[0034] The perceptual agent uses a pre-defined time sliding window to traverse the aforementioned time-series data, extracts statistical features, and introduces an anomaly detection model to review data segments that deviate from the normal pattern, outputting a list of abnormal events. Each record in the list of abnormal events contains an anomaly type code, the workstation identifier where it occurred, the time of occurrence, and a confidence level, represented as a quadruple.

[0035] Step S2, Anomaly Semantic Diagnosis. The four-tuple output by the perception agent is input into the diagnostic agent. The diagnostic agent extracts the manufacturing context information of the workstation corresponding to the anomaly. The context information specifically includes the work order number, process type, the position sequence number of the corresponding process in the order's process path, and the quantity of buffered work-in-process in upstream and downstream processes. The anomaly event information and context information are concatenated into a structured descriptive text, which is then input into the pre-deployed apparel industry big data model. Based on knowledge of the apparel production domain, the apparel industry big data model performs semantic diagnosis on the underlying causes of the current anomaly, outputs probability distribution vectors for multiple predefined anomaly root cause categories, and simultaneously predicts and estimates the recovery time range.

[0036] Step S3: Anomaly Propagation Graph Construction and Impact Assessment. After obtaining the diagnostic results, extract the production line process dependency topology. Using all processes within the production line as the node set, and the material flow relationships and workstation sharing relationships between processes as the directed edge set, construct a process dependency graph. Extract real-time data on buffer margins between each process, calculate the edge weights of the directed edges; the edge weights are determined by the dependency strength combined with the negative exponential buffer decay coefficient. Using the process where the anomaly occurred as the source node, propagate the anomaly layer by layer along the directed edges in the process dependency graph, extrapolating the impact probability and impact time window for downstream nodes. Prune nodes with impact probabilities below the pruning threshold and their downstream subtrees to obtain the impact subgraph. Within the scope of the impact subgraph, statistically analyze the set of affected orders and the estimated delay time.

[0037] Step S4: Generation of candidate scheduling schemes. After obtaining the affected subgraph and order priority data including delivery urgency and customer level, the scheduling agent calculates the feature vector of the current anomaly and retrieves historical cases from the strategy knowledge base. When the maximum similarity does not exceed the case reuse threshold, the scheduling agent generates candidate schemes from the pre-set strategy action library. The strategy action library includes five basic scheduling actions: process rerouting, workstation mutual assistance, task resegmentation, buffer release, and outsourcing switching. Applicable action subsets are selected based on the attributes of the affected processes and arranged and combined to generate a set of candidate schemes. The large-scale apparel industry model sorts the candidate schemes according to order priority. When the affected subgraph only contains the source node itself, the scheduling agent enters a local adjustment mode, that is, it only selects applicable actions from the strategy action library for local scheduling for the workstation where the source node is located, without triggering the generation of combined scheduling actions for multiple processes.

[0038] Step S5, Digital Twin Pre-simulation Verification. Each generated candidate solution is sequentially injected into a virtual production line simulation model containing digital mappings of the production line. Using the current actual state of the production line as the initial condition, discrete event simulation is performed. A comprehensive score is calculated for each candidate solution based on three dimensions: completion time offset, equipment utilization rate, and secondary conflict risk. The solution with the highest comprehensive score is selected as the optimal solution. If the comprehensive scores of all candidate solutions are lower than a preset threshold, the process reverts to step S4 to expand the action combination space and regenerate candidate solutions.

[0039] Step S6: Execution and Case Knowledge Accumulation. The executing agent breaks down the optimal solution into specific execution instructions for each workstation and issues them to the corresponding workstation terminals. During execution, the actual recovery time of each affected process is continuously collected. After the solution is executed, the actual recovery effect score is calculated. This anomaly handling process is constructed as a triplet case containing anomaly feature vectors, a structured description of the scheduling scheme, and the actual recovery effect score, and stored in the policy knowledge base. Simultaneously, the edge weight parameters of the anomaly propagation graph are updated in reverse based on the actual observed propagation intensity.

[0040] The process of multi-source data perception and large model semantic diagnosis specifically includes the following processing sub-steps:

[0041] Step S101: The sensing agent continuously collects multi-source heterogeneous data generated in the garment production line. This multi-source heterogeneous data includes equipment IoT data, quality inspection data, material arrival data, and personnel attendance data. Equipment IoT data includes sewing machine speed sequences and thread breakage signals; quality inspection data includes process pass rates and defect type codes; material arrival data includes the time deviation between the arrival time of materials at each workstation and the planned arrival time; and personnel attendance data includes the attendance status of workstation operators. The sensing agent performs anomaly tolerance judgment when collecting multi-source heterogeneous data. When a sensor detects missing data, the sensing agent extracts the most recent valid value to fill in the missing data; when a single data source is completely interrupted, the sensing agent generates a data source interruption alarm and downgrades to a detection mode that relies solely on the remaining data source.

[0042] Step S102: The sensing agent processes the collected IoT device data using a sliding window statistical method. The sliding window length is set to... In each sliding window Within the system, the perceptual agent calculates the mean shift and variance rate of change of the signal sequence. Simultaneously, the perceptual agent calculates a statistical threshold, which is the sum of the mean and the variance. The standard deviation of times, This is a multiple of the statistical threshold. The perceptual agent compares the data segments within the window with the statistical threshold and filters out data segments that exceed the statistical threshold.

[0043] In step S103, the perceptual agent inputs data segments exceeding the statistical threshold into the lightweight temporal detection model for secondary verification, and the lightweight temporal detection model outputs the anomaly confidence level. Based on the secondary confirmation results, the perceptual agent outputs a list of abnormal events. Each record in the list of abnormal events is represented as a quadruple. , Indicates the exception type encoding. Workstation identifiers indicating where an abnormality has occurred. Indicates the time of occurrence. Indicates the degree of confidence in anomalies.

[0044] Step S104, the perceptual agent determines the confidence threshold. The abnormal event quadruple is processed by traffic splitting. (Regarding confidence level...) Exceeding the confidence threshold For abnormal events, the perceptual agent inputs the corresponding four-tuple into step S201. Regarding the confidence level... Below the confidence threshold But higher For any anomalous event, the perceptual agent marks the corresponding four-tuple as a suspected anomaly and moves it to an observation queue for confirmation by subsequent data. When multiple anomalous events occur simultaneously within the same time window, the perceptual agent categorizes them according to confidence level. The exceptions are processed in descending order.

[0045] Step S201: The diagnostic agent receives the abnormal event quadruple (c,s,t, ...) transmitted by the perceptual agent. ), and extract the workstation identifier where the anomaly occurred. The currently associated manufacturing context information includes the work order number, process type, the position sequence number of the corresponding process in the order's process path, and the buffer work-in-process quantity of upstream and downstream processes. The diagnostic agent extracts the field templates of the structured description, concatenates the abnormal event quadruple information with the manufacturing context information, and generates a structured description text.

[0046] In step S202, the diagnostic agent inputs the concatenated structured description text into the large-scale model of the apparel industry. Combining the temperature parameters inferred by the large-scale model, the apparel industry large-scale model performs semantic diagnosis of anomalies based on knowledge of the apparel production domain. The diagnostic process outputs three results: the first is the probability distribution vector of the anomaly root cause type. , For the predefined number of root cause categories, to The first term represents the probability value of the corresponding root cause category; the second term is a textual description of the most likely root cause hypothesis; the third term is the estimated recovery time interval. , This indicates the lower limit of the estimated recovery time. This represents the upper limit of the estimated recovery time. The diagnostic agent will use the probability distribution vector... The root cause description text and the estimated recovery time range are encapsulated into a diagnostic result structure for output.

[0047] Step S203: The diagnostic agent performs judgment and fault tolerance processing on the diagnostic results output by the large-scale model of the apparel industry. The diagnostic agent extracts the probability distribution vector. The maximum probability value in When the maximum probability value When an anomaly is detected, the diagnostic agent marks it as diagnostically uncertain, triggering a manual review process. When the large-scale model service for the apparel industry is unavailable, the diagnostic agent degrades to a rule-based diagnostic mode based on historical statistics, querying the historical frequency table according to the anomaly type code 'c' and the workstation type, and outputting an approximate diagnostic result.

[0048] The process of constructing anomaly propagation maps and assessing their impact includes the following sub-steps:

[0049] Step S301: After the diagnostic agent outputs the diagnostic result, the scheduling agent extracts the production line process dependency topology and real-time data of buffer margins between processes from the production line process dependency topology database. The production line process dependency topology includes material flow relationships between processes and workstation sharing relationships. The scheduling agent uses all processes of the production line as a set of nodes. The set of directed edges is composed of material flow relationships between processes and workstation sharing relationships. Construct a process dependency graph When the process dependency topology database update is delayed, the scheduling agent uses the most recent valid topology snapshot to construct the process dependency graph and marks it with a "topology version delayed" alarm.

[0050] Step S302: Schedule the agent's process dependency graph Each directed edge in Calculate edge weights Border rights The calculation formula is:

[0051]

[0052] In the formula, Indicate process to process The dependency strength, the range of values ​​for which the dependency strength is limited is . The dependence intensity is determined by a combination of the necessity of material flow and the degree of workstation sharing, according to a preset calculation weight ratio. Indicate process With process The ratio of the current quantity of work-in-process buffer to the standard buffer quantity. This represents the buffer attenuation coefficient, which controls the rate at which the buffer margin suppresses the propagation intensity. The weighted process dependency graph constitutes the anomaly propagation graph, used to characterize the propagation path of anomalies between processes.

[0053] Step S303: The scheduling agent uses the process where the anomaly occurred as the source node and sets the influence probability of the source node. The scheduling agent extracts the upper limit of the estimated recovery time from the diagnostic results. The upper limit of the estimated recovery time is set as the initial value of the source node's influence time window. ,Right now The scheduling agent determines whether the source node has a downstream connection edge. When the source node has no downstream connection edge, the affected subgraph only contains the source node itself, and the scheduling agent's control flow directly enters the local adjustment mode in step S4. When the source node has a downstream connection edge, the scheduling agent performs layer-by-layer propagation along the directed edge.

[0054] Step S304: The scheduling agent propagates layer by layer along the directed edges of the process dependency graph. For processes with multiple predecessor nodes... process The probability of impact is calculated by combining independent events. The calculation formula is:

[0055]

[0056] In the formula, Indicate process The complete set of all affected predecessor nodes Represents the predecessor node The probability of influence, Represents the predecessor node to process The weight of the directed edge.

[0057] At the same time, the scheduling agent calculates the impact time window of each node. Affecting the time window The calculation formula is:

[0058]

[0059] In the formula, Represents the predecessor node The time window of influence, Indicate process to process The standard turnaround time.

[0060] Step S305: The scheduling agent calculates the influence probability of each node and the pruning threshold. A comparison was performed. The scheduling agent's probability of influencing the pruning threshold was lower than the threshold. Nodes with an impact probability below the pruning threshold. Pruning the downstream subtree of the node yields the ripple subgraph. In the ripple subgraph The scheduling agent calculates the set of affected orders and the estimated delay time for each order.

[0061] Step S306, the scheduling agent will finalize the affected subgraph. Output, ripple subgraph This includes the set of affected processes, the probability of impact for each process, the impact time window, the list of affected orders, and the estimated delay duration. The scheduling agent statistically analyzes the spillover subgraph. The number of processes included. When the number of processes included in the affected subgraph exceeds 50% of the total number of processes in the production line, the scheduling agent triggers a full-line warning and notifies the scheduling management layer to intervene.

[0062] The process of generating candidate scheduling scheme combinations and performing digital twin pre-simulation verification specifically includes the following sub-steps:

[0063] Step S401: The scheduling agent receives the spillover subgraph, order priority data including delivery urgency and customer level, and current production line capacity constraint data. The scheduling agent extracts the anomaly type, affected process attributes, and spillover scale features, and encodes the anomaly type, affected process attributes, and spillover scale features to calculate the feature vector of the current anomaly. The scheduling agent will use the feature vector of the current anomaly. Cosine similarity is calculated between the feature vectors of historical cases in the strategy knowledge base and the case reuse threshold. When the scheduling agent retrieves the historical solution corresponding to the highest similarity, it fine-tunes the parameters of the historical solution based on the current production line capacity constraints, and selects the fine-tuned solution as the preferred candidate solution. When the strategy knowledge base query times out, the scheduling agent skips the case retrieval and directly enters the action combination generation process.

[0064] Step S402: When no similar cases meet the reuse conditions, the scheduling agent invokes a pre-set strategy action library. The strategy action library contains five categories of basic scheduling actions: process rerouting actions, used to reassign affected processes to alternative workstations with the same processing capacity; workstation mutual assistance actions, used to temporarily allocate operators from adjacent workstations to support affected workstations; task resegmentation actions, used to split the tasks to be processed on the affected workstations into multiple sub-segments and allocate them to multiple available workstations for parallel processing; buffer release actions, used to release buffered work-in-process in downstream processes in advance to maintain the continuous operation of downstream workstations; and outsourcing switching actions, used to transfer tasks corresponding to processes that have exceeded the recovery time limit to external collaborative factories for execution. Each category of basic scheduling action is configured with applicable preconditions.

[0065] Step S403: The scheduling agent extracts the process type, equipment specificity, workstation substitutability, and personnel skill interoperability attributes of the affected processes in the affected subgraph. Based on these attributes, the scheduling agent selects a subset of actions from five basic scheduling actions that meet the applicable preconditions, and then arranges and combines these subsets according to rules to generate a candidate solution set. If the number of candidate solutions generated in a single combination exceeds the upper limit for the number of candidate solutions... When the scheduling agent sorts the candidates by priority, it truncates any candidates exceeding the maximum number of options. If none of the five basic scheduling actions meet the applicable preconditions, the scheduling agent generates a human intervention request plan and marks it as the highest priority.

[0066] Step S404: The scheduling agent assigns priority scores to the affected orders. The formula for calculating the order priority score is:

[0067]

[0068] In the formula, Indicates order Priority rating, Indicates order The number of days remaining until delivery. This indicates the maximum number of days remaining among all currently affected orders. Indicates order The customer level normalized value, and The preset priority weight coefficients are used, and they satisfy the following conditions: When the maximum remaining days At that time, all affected orders have the same delivery urgency, and the first term of the formula is set as follows: Priority scoring is determined solely by customer level.

[0069] In step S405, the scheduling agent inputs the candidate solution set into the large-scale apparel industry model. The large-scale apparel industry model sorts the candidate solutions according to the order priority scoring results, optimizes the time parameters within the candidate solutions, and outputs candidate rescheduling solutions sorted by priority.

[0070] Step S501: The generated candidate solutions are sequentially injected into the virtual production line simulation model for discrete event simulation. The virtual production line simulation model contains a digital mapping of the physical equipment on the production line, which includes equipment status, station capacity, work-in-process queue, and process routing. The virtual production line simulation model uses the current actual state of the production line as its initial condition. When the virtual production line simulation model fails to synchronize its state, the scheduling agent uses the most recently successfully synchronized state snapshot as its initial condition and adds a state lag marker to the subsequent scoring results for decision-making reference.

[0071] In step S502, the virtual production line simulation model adjusts the process allocation and resource configuration according to the parameters set in the candidate scheme, and executes the simulation until all affected orders are completed or the simulation time limit is reached. When the simulation time exceeds the real-time limit, the virtual production line simulation model switches to accelerated simulation mode, ignoring the detailed process simulation of low-priority orders. After the simulation is completed, simulation data is extracted from three dimensions—completion time offset, equipment utilization rate, and secondary conflict risk—for scoring calculation.

[0072] Step S503: Calculate the completion time offset score. The formula for calculating the completion time offset score is as follows:

[0073]

[0074] In the formula, Indicates the completion time offset score. Indicates the total number of affected orders. Indicates the order after the scheme is deduced The expected completion time Indicates order The planned completion time, This indicates the maximum allowable delay time.

[0075] Step S504: Calculate the equipment utilization rate score. The formula for calculating the equipment utilization rate score is:

[0076]

[0077] In the formula, This indicates the equipment utilization rate score. This represents the set of workstations involved in the ripple subgraph. Indicates the workstation during the scheme simulation. Equipment utilization rate.

[0078] Step S505: Calculate the secondary conflict risk score and the overall score. The formula for calculating the secondary conflict risk score is:

[0079]

[0080] In the formula, This indicates the risk score for secondary conflict. This represents the number of new resource conflict events added during the simulation. This indicates the total number of scheduling decisions involved in the simulation.

[0081] The comprehensive score is obtained by weighted summation of the scores from the three dimensions. The formula for calculating the comprehensive score is as follows:

[0082]

[0083] In the formula, This indicates the overall score. , , Let be the dimension weight coefficients, and satisfy . .

[0084] Step S506: Select the solution with the highest comprehensive score as the optimal solution. Determine whether the comprehensive score of the optimal solution reaches a preset threshold. When there is only one candidate solution, it is still necessary to pass the preset threshold. Judgment. If the overall score of all candidate solutions is lower than the preset threshold... If the number of backoffs reaches the preset backoff limit, a backoff signal is sent to step S403. After receiving the backoff signal, the scheduling agent expands the action combination space and regenerates candidate solutions by relaxing the preconditions for action applicability or increasing the depth of action combinations. Then, the scheduling agent terminates the regeneration process and generates a request for human intervention.

[0085] The process of distributing and accumulating case knowledge includes the following sub-steps:

[0086] Step S601: The execution agent receives the optimal solution output in step S5 and the production line execution terminal interface. The execution agent breaks down the optimal solution into specific execution instructions for each workstation, including process reassignment instructions, personnel allocation instructions, and material rerouting instructions. The execution agent sends the specific execution instructions to the corresponding workstation terminals through the production line execution terminal interface. When a workstation terminal communication interruption is detected, the execution agent writes the specific execution instructions into a retry queue and automatically retransmits the specific execution instructions after communication is restored. When the number of retries exceeds the set number, the execution agent marks the workstation where the communication interruption occurred as unavailable and triggers a new round of scheduling.

[0087] In step S602, after the execution agent issues the specific execution instruction, it continuously collects the actual recovery time and execution status of each affected process. If no workstation completion confirmation is received within the estimated recovery time limit, the execution agent triggers an execution timeout alarm and initiates the abnormal event re-sensing process, returning the control process to step S1.

[0088] Step S603: After the solution is executed, the agent calculates the actual recovery effect score. When calculating the actual recovery effect score, the same three-dimensional scoring function as in step S5 is used, and the actual execution data is used instead of the simulation data to calculate the actual recovery effect score. .

[0089] Step S604: The executing agent constructs the current exception handling process as a triplet case. and triplet cases Store in the strategy knowledge base. In the triplet case, The abnormal feature vector encodes the abnormal type, the scale of impact, and the attributes of the affected processes. A structured description of the final execution scheduling scheme; Score the actual recovery effect The strategy knowledge base periodically implements an elimination mechanism. The executing agent performs continuous... Cases that were not retrieved and whose performance score was lower than the median in the strategy knowledge base were marked as inefficient cases and moved to the archive. This is a period of observation for eliminating inefficient cases.

[0090] Step S605: The executing agent updates the edge weight parameters of the anomaly propagation graph in reverse based on the execution effect. For the edges on this propagation path... The dependency strength is updated using an exponential moving average method. The update formula for dependency strength is:

[0091]

[0092] In the formula, Indicates the new dependency strength after the update. Indicates the old dependency strength before the update. Indicates the learning rate. This represents the actual observed propagation intensity. The actual observed propagation intensity is derived by inversely calculating the ratio of the actual affected time window to the predicted time window.

[0093] Step S606, when the actual recovery effect score is... When the score exceeds the threshold for good results, the agent appropriately enhances the credibility of the relevant edge weights; when the actual recovery effect score is... When the performance falls below the inadequate threshold, the executing agent moderately reduces the corresponding edge weights. The executing agent outputs an execution completion confirmation, a case entry confirmation, and an edge weight update log.

[0094] To verify the effectiveness of the method of this invention, an experiment was conducted on a personalized clothing customization production line. This production line comprises four main sections—cutting, sewing, ironing, and quality inspection—comprising 48 processes and 32 workstations, processing approximately 200 orders per day. The overall technical route of the method of this invention is described below. Figure 1 .

[0095] Experiment 1: Verification of the Accuracy of Anomaly Propagation Map Assessment

[0096] A total of 120 abnormal events, including equipment failures, quality anomalies, and material delays, were collected during a continuous operating cycle. For each anomaly, the impact range was predicted using the anomaly propagation graph method, rule engine method, and manual experience assessment method of this invention. Based on the actual set of affected processes observed on the production line, the precision, recall, and F1 score of each method were calculated. The anomaly propagation graph parameters were set as buffer attenuation coefficients. Pruning threshold The rule engine approach uses fixed upstream and downstream two-level influence rules.

[0097] Experimental results are as follows Figure 2As shown, the method of this invention achieved an average precision of 0.897, an average recall of 0.860, and an average F1 score of 0.878 across the three types of anomalies. The rule engine method achieved an average F1 score of 0.667, while human experience-based judgment achieved an average F1 score of 0.549. The method of this invention improved the F1 score by 0.329 compared to human experience and by 0.211 compared to the rule engine. The evaluation performance for equipment failure anomalies was better than the other two types, while material delay anomalies had a relatively low recall rate due to their longer propagation path.

[0098] Experiment 2: Verification of the Improvement of Scheduling Scheme Quality by Digital Twin Pre-simulation

[0099] For 60 abnormal events, two scheduling schemes were generated using different modes: Mode A represents the complete process of this invention (including digital twin pre-simulation verification and optimization), and Mode B removes pre-simulation verification and directly selects the top-ranked scheme from the large model. The synchronization interval between the simulation model and the actual production line status is 5 minutes. Scoring threshold. Dimension weights , , .

[0100] Experimental results are as follows Figure 3 As shown, the success rate of the method of this invention on the first execution is 88.3%, compared to 65.0% in the control group without pre-exercise, representing an improvement of 23.3 percentage points. Regarding average order delay, the method of this invention has a delay of 4.7 minutes, compared to 15.0 minutes in the control group without pre-exercise. As for the secondary conflict rate, the method of this invention has a rate of 3.3%, compared to 21.7% in the control group without pre-exercise. The pre-exercise verification mechanism effectively filters out schemes that are prone to causing secondary conflicts after execution.

[0101] Experiment 3: Verification of the efficiency improvement of case knowledge accumulation

[0102] Record system performance changes over 6 consecutive cycles. Each cycle handles approximately 20 exception events. Case reuse threshold. Edge weight learning rate .

[0103] Experimental results are as follows Figure 4 As shown. In the first period, no historical cases were available, the average response time was 38.5 seconds, and the proportion of large model calls was 100%. As cases accumulated, the average response time in the sixth period dropped to 21.5 seconds, a decrease of 44.2%; the case retrieval hit rate reached 54.9%; and the proportion of large model calls dropped to 45.8%. In the fourth period, a new type of abnormal event occurred in a concentrated manner, the case hit rate temporarily dropped from 35.0% in the third period to 31.0%, and the response time rebounded from 26.1 seconds to 27.3 seconds. Subsequently, after new cases were added to the database, the indicators resumed their downward trend.

[0104] Experimental conclusions

[0105] The average F1 score for anomaly propagation assessment was 0.878, an improvement of 0.329 compared to manual experience. Digital twin simulation increased the first-time success rate of solutions from 65.0% to 88.3%, and reduced the secondary conflict rate from 21.7% to 3.3%. Case studies reduced the average response time from 38.5 seconds to 21.5 seconds within 6 cycles, and the proportion of large model calls decreased from 100% to 45.8%.

[0106] This invention also provides a self-healing scheduling device for a garment production line based on anomaly propagation graphs. The device includes a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement all the steps of the method described in steps S1 to S6 above.

[0107] This invention also provides a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements all the steps of the method described in steps S1 to S6 above.

[0108] For the construction of the communication network of the underlying multi-agent architecture and the data reading and writing of the physical server, those skilled in the art can use existing mature network communication protocols and computer storage technologies. The specific implementation methods of the above-mentioned supporting technologies are well-known technologies in this field and will not be elaborated here.

[0109] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A self-healing scheduling method for a garment production line based on anomaly propagation graphs, characterized in that, The method operates based on an underlying multi-agent hierarchical decision-making architecture and includes: Step S1, Multi-source anomaly perception: The sensing agent collects and integrates the production data continuously generated by the garment production line in real time, extracts statistical features and verifies them, and outputs a list of abnormal events. Step S2, Abnormal Semantic Diagnosis: The diagnostic agent receives the list of abnormal events, extracts the manufacturing context information of the workstation corresponding to the abnormal event in the list of abnormal events, concatenates the abnormal event information with the manufacturing context information into a structured description text, and inputs it into the pre-deployed large model of the apparel industry for semantic diagnosis, and outputs a diagnostic result containing the estimated recovery time range. Step S3, Anomaly Propagation Graph Construction and Impact Assessment: Extract the production line process dependency topology; construct a process dependency graph using all production line processes as the node set and the material flow relationship and workstation sharing relationship between processes as the directed edge set; extract real-time data of buffer margin between each process to calculate the edge weight of the directed edge, wherein the size of the edge weight is determined by multiplying the dependency strength between processes by the negative exponential buffer decay coefficient calculated based on the ratio of the current buffer work-in-process quantity to the standard buffer quantity; based on the diagnostic results, take the process where the anomaly occurred as the source node, and propagate it layer by layer along the directed edge in the process dependency graph, deduce the impact probability and impact time window of each downstream node, prune the nodes whose impact probability is lower than the pruning threshold and their downstream subtrees, and the process dependency graph constitutes the anomaly propagation graph, obtaining the impact subgraph; Step S4, candidate scheduling scheme combination generation: The scheduling agent obtains the ripple subgraph and order priority data, retrieves historical cases in the strategy knowledge base, and when the case reuse condition is not met, it selects applicable action subsets from the preset strategy action library and arranges them to generate a candidate scheme set. The apparel industry big model sorts the candidate scheme set according to the order priority data. Step S5, Digital Twin Pre-simulation Verification: The sorted candidate solutions are sequentially injected into the virtual production line simulation model containing the production line digital mapping, and discrete event simulation is performed. The comprehensive score of each candidate solution is calculated from three dimensions: completion time offset, equipment utilization rate, and secondary conflict risk. The candidate solution with the highest comprehensive score is selected as the optimal solution. Step S6, Execution and Case Knowledge Accumulation: The executing agent decomposes the optimal solution into specific execution instructions and distributes them to the corresponding workstation terminals; after the solution is executed, the same three-dimensional scoring function as the digital twin pre-simulation verification is used to calculate the actual recovery effect score by replacing the simulation data with the actual execution data; the anomaly handling process is constructed as a triplet case containing anomaly feature vectors, a structured description of the scheduling scheme, and the actual recovery effect score, and the triplet case is stored in the policy knowledge base; the actual observed propagation intensity is derived by back-calculating the ratio of the actual affected time window to the predicted time window, and a new dependency intensity is calculated by using an exponential moving average method, combined with the old dependency intensity before the update, the learning rate, and the actual observed propagation intensity, to update the edge weight parameters of the anomaly propagation graph.

2. The self-healing scheduling method for garment production lines based on anomaly propagation graphs according to claim 1, characterized in that, Step S1 includes: The continuously collected production data includes equipment IoT data, quality inspection data, material arrival data, and personnel attendance data; The collected time-series data is traversed using a preset sliding window. Within each time sliding window, the mean offset and variance change rate of the signal sequence are calculated. The signal sequence within the window is then compared with a statistical threshold calculated based on the mean and standard deviation to filter out data segments that exceed the statistical threshold. Data segments exceeding the statistical threshold are input into a lightweight time-series detection model for secondary verification, anomaly confidence is output, and an anomaly event list is generated. Each record in the anomaly event list is a quadruple containing anomaly type code, occurrence station identifier, occurrence time, and anomaly confidence.

3. The self-healing scheduling method for garment production lines based on anomaly propagation graphs according to claim 1, characterized in that, Step S2 includes: The extracted manufacturing context information includes the work order number, process type, the position number of the corresponding process in the order process path, and the buffer work-in-process quantity of upstream and downstream processes; Extract the field template of the structured description, and concatenate the quadruple containing the abnormal event information with the manufacturing context information according to the field template of the structured description to generate the structured description text; By combining temperature parameters inferred from a large model, the large model of the apparel industry is controlled to perform semantic diagnosis of anomalies based on knowledge of the apparel production field. The diagnosis results also include the probability distribution vector of the anomaly root cause category output by the large model of the apparel industry.

4. The self-healing scheduling method for garment production lines based on anomaly propagation graphs according to claim 1, characterized in that, The process of propagating along the directed edges layer by layer in the process dependency graph in step S3 includes: The upper limit of the estimated recovery time interval is set as the initial value of the source node's influence time window; For affected processes with multiple precursor nodes, the impact probability is calculated by combining the independent events of the impact probability of each precursor node. Simultaneously, the impact time window of each node is calculated. The impact time window of each affected node is determined by the maximum value of the sum of the impact time window of its predecessor node and the standard flow time between processes.

5. The self-healing scheduling method for garment production lines based on anomaly propagation graphs according to claim 1, characterized in that, The process of generating and sorting the candidate solution set in step S4 includes: The pre-set strategy action library includes five basic scheduling actions: process rerouting actions, workstation mutual assistance actions, task resegmentation actions, buffer release actions, and external cooperation switching actions. Based on the affected process type, equipment specialization, workstation substitutability, and personnel skill interoperability attributes, a subset of actions that meet the applicable preconditions is selected from the five categories of basic scheduling actions and arranged and combined to generate the candidate solution set. Calculate the priority score of the affected orders, which is a weighted average of the normalized value of the remaining days until delivery and the normalized value of the customer level; control the large apparel industry model to sort the candidate solution set according to the priority score and optimize the time parameters within the solution.

6. The self-healing scheduling method for garment production lines based on anomaly propagation graphs according to claim 5, characterized in that, The process of performing discrete event simulation and calculating the comprehensive score in step S5 includes: Adjust the process allocation and resource configuration according to the parameters set in the candidate scheme, and extrapolate within the virtual production line simulation model until all affected orders are completed or the simulation time limit is reached; Extract simulation data, calculate completion time offset score, equipment utilization rate score and secondary conflict risk score respectively, and sum the scores of the three dimensions by weight to obtain the comprehensive score; wherein, the secondary conflict risk score is equal to one minus the ratio of the number of resource conflict events added during the simulation to the total number of scheduling decisions involved in the simulation; If the overall score of all candidate solutions is lower than the preset threshold, the action combination space is expanded to regenerate candidate solutions, and a manual intervention request is generated after the number of backtracking attempts reaches the preset upper limit.

7. A self-healing scheduling device for a garment production line based on anomaly propagation graphs, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement all the steps of the garment production line self-healing scheduling method based on anomaly propagation graph as described in any one of claims 1-6.

8. A non-volatile computer-readable storage medium, characterized in that, The non-volatile computer-readable storage medium stores a computer program that, when executed by a processor, implements all the steps of the garment production line self-healing scheduling method based on anomaly propagation graph as described in any one of claims 1-6.

Images