AI scheduling optimization method for clothing flexible production and ERP integrated system

Abstract

The invention relates to the technical field of clothing manufacturing, in particular to an AI scheduling optimization method for clothing flexible production and an ERP integrated system. Comprising the following steps: multi-source heterogeneous data acquisition and fusion: acquiring real-time equipment state data of a clothing production line through Internet of Things equipment, wherein the real-time equipment state data comprises a sewing machine rotating speed, a cutting bed load rate and a quality inspection station abnormal signal; obtaining an order delivery period, a material inventory and a BOM process path through an ERP system; the equipment data and the order data are stored in a distributed database after being aligned through timestamps; dynamic priority modeling: constructing a dynamic priority calculation model based on the order emergency degree, the production line bottleneck process load threshold and the material neat rate, and generating an order-insertable process-level priority queue; according to the scheme, the limitation of a traditional method can be broken through, and the modern production requirements of flexibility and high aging are met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of garment manufacturing, in particular to an AI scheduling optimization method for flexible production of clothing and an ERP integrated system. BACKGROUND

[0002] The current garment manufacturing industry is facing an increasing market demand for multi-variety and small-batch orders, and the complexity and dynamics of production scheduling have significantly increased. Traditional production scheduling techniques mainly rely on manual experience or automated systems based on fixed rules, which have exposed three major core defects in actual operation: first, the data integration capability is insufficient, the real-time state data of the equipment is stored separately from the enterprise management data for a long time, forming a data island. For example, the running data collected by the workshop equipment sensor cannot be associated with the order priority in the ERP system, resulting in scheduling decisions relying on lagging information, and emergency order response delay exceeding 2 hours. Second, the scheduling rigidity problem is prominent, the existing system uses a static optimization mode and cannot dynamically respond to sudden conditions such as order insertion and equipment failure. For example, when a cutting bed suddenly stops, the traditional method needs to recalculate the scheduling from scratch, which takes more than 1 hour, seriously affecting the continuity of production. Third, the multi-objective optimization is unbalanced, and the existing technology cannot optimize the conflicting goals of changeover cost, equipment utilization rate, and order delivery rate, often sacrificing one for the other - for example, to reduce the number of changeovers, the same orders are concentrated, resulting in some equipment load exceeding 90%, and the failure rate rising by 37%, while forced load balancing will cause daily production capacity to drop by 15%.

[0003] The above problems directly restrict the agile delivery capability and resource utilization efficiency of garment enterprises. The industry urgently needs a new scheduling technology that can realize deep fusion of multi-source data, dynamically respond to production disturbances, and intelligently balance multi-objective conflicts, to break through the limitations of traditional methods and meet the modern production needs of flexibility and high timeliness. SUMMARY

[0004] The purpose of the present application is to solve the shortcomings in the prior art, and the AI scheduling optimization method for flexible production of clothing is proposed, which comprises: S1. Multi-source heterogeneous data collection and fusion: real-time equipment state data of the clothing production line is collected through Internet of Things devices, including sewing machine speed, cutting bed load rate, and quality inspection station abnormal signal; order delivery date, material inventory, and BOM process path are obtained through the ERP system; after aligning the device data and order data by timestamp, they are stored in a distributed database; S2. Dynamic priority modeling: based on order urgency, bottleneck process load threshold, and material completion rate, a dynamic priority calculation model is constructed to generate a process-level priority queue that can be inserted; S3. Multi-objective scheduling optimization: A deep reinforcement learning algorithm is used to minimize the number of line changes, balance device load, and maximize daily production capacity as joint optimization objectives to generate a process scheduling scheme that meets flexible production requirements; S4. Real-time simulation verification: Based on the digital twin engine, the scheduling scheme is simulated by Monte Carlo simulation to dynamically correct scheduling conflicts caused by device failure or material delay; S5. Two-way integration of ERP: The optimized scheduling scheme is decomposed into work order instructions and pushed to the MES system, while the production progress is synchronized to the MRP module of the ERP to trigger dynamic material replenishment.

[0005] The above technical scheme is adopted: The present scheme constructs a flexible production scheduling framework based on the integration of Internet of Things and ERP, including five core steps of data acquisition, dynamic modeling, multi-objective optimization, simulation verification and system integration. It can solve the problem that the scheduling decision lags behind due to the separation of device state data and ERP order data, and cannot dynamically respond to sudden events such as emergency order response time being too long and the number of line changes and equipment utilization being difficult to balance.

[0006] Preferably, the priority weight calculation in the dynamic priority calculation model satisfies: ; Wherein, is the priority score of the i-th order, is the order commitment delivery time stamp, t is the current system time, is the maximum allowed delivery time margin, is the total load rate of the key equipment required by the order, is the maximum safe load threshold of the device (set to 85%), is the material completion rate (0-1 standardized value), , and are weighting coefficients and satisfy .

[0007] The above technical scheme is adopted: The present scheme uses the inverse relationship of time margin to make the priority of orders close to the delivery time rise exponentially, which can improve the processing speed of emergency orders.

[0008] Further preferably, the reward function of the deep reinforcement learning algorithm is: ; Wherein, is the number of line changes, is the total line change capacity threshold; is the minimum utilization rate of the device, is the average utilization rate; is the number of completed orders on the day, l is the total order number; 、 and is a weight parameter dynamically adjusted according to the production stage.

[0009] With the technical scheme, the technical problem that the traditional scheme only optimizes the number of line changes, resulting in too high load difference of the equipment, can be solved, and the existing reinforcement learning method uses a fixed weight and cannot adapt to changes in production strategies.

[0010] Further preferably, the equipment state data stored by the distributed database includes a three-dimensional equipment state matrix, the dimensions of which respectively represent equipment type classification, equipment physical location coordinates and real-time running parameters, wherein the real-time running parameters at least include remaining maintenance time length, cumulative output and temperature sensor data.

[0011] With the technical scheme, the three-dimensional tensor structure provided by the scheme can reduce the equipment state retrieval time, the remaining maintenance time length parameter can prevent the triggering accuracy of preventive maintenance work orders, and the equipment location coordinates can improve the material handling path planning efficiency.

[0012] Further preferably, the real-time simulation verification adopts an event-driven mechanism, and when it is detected that the duration of a key equipment failure exceeds a preset threshold, the material arrival delay causes the matching rate to drop by more than 15%, or the priority score of a newly added urgent order is higher than that of 50% of the tasks in the current schedule, the reconstruction of the scheduling scheme is triggered.

[0013] Further preferably, the scheduling reconstruction adopts a local optimization strategy, and only the subsequent tasks of the affected process are dynamically adjusted, the adjustment range is limited within a range of plus or minus 30% of the original scheduling time window, and the task of an overloaded equipment is distributed to an adjacent idle equipment through a load transfer algorithm.

[0014] Further preferably, the ERP two-way integration adopts an event stream pipeline to transmit work order instructions, the work order instructions are compressed through a binary protocol and then transmitted in priority levels, and high-priority instructions are synchronized across systems within 50 milliseconds.

[0015] Further preferably, the dynamic material replenishment strategy includes an in-process inventory monitoring mechanism, and when the number of in-process products is lower than a safety inventory threshold, a purchase order is automatically generated and the optimal replenishment batch is calculated, and the replenishment quantity is dynamically adjusted according to the production capacity demand in the next 3 days.

[0016] Further preferably, the method adopts a federated learning framework to update the scheduling model, the local model of each plant is uploaded to the center server after training, the center server aggregates the gradients to generate a global model and synchronizes it to each plant, thereby improving the cross-plant scheduling adaptability while ensuring data privacy.

[0017] An ERP integrated system applied to the AI scheduling optimization method for flexible production of clothing as claimed in any one of the above, comprising: ‌a data acquisition layer composed of Internet of Things sensors deployed on production line equipment and connected with an ERP database, which acquires equipment operation parameters, order delivery data and material inventory status in real time, and converts the data format into JSON-LD structured data stream; ‌a dynamic scheduling engine based on digital twin model for three-dimensional simulation of production processes, which includes a physical layer device kinematics modeling module, a logical layer process dependency rule base and a data layer real-time synchronization module, and generates a scheduling scheme through a deep reinforcement learning algorithm, wherein the reinforcement learning reward function optimizes the number of line changes, equipment load balancing rate and order completion rate at the same time; ‌a material management module that dynamically calculates a safety inventory threshold according to real-time work-in-process inventory levels, triggers a replenishment order when inventory is below the threshold, and matches substitute materials based on process parameters when key materials are in short supply, and synchronously updates device control instructions; ‌an exception handling module that uses a hierarchical message queue to transmit production instructions, including a high-priority channel and a normal channel, when device failure or urgent insertion events occur, adjusts the scheduling of affected processes through a local re-optimization algorithm, and transmits task migration instructions to standby equipment through the high-priority channel; ‌a model optimization module that aggregates local scheduling model gradient data from multiple plants based on a federated learning framework, generates a global optimization model through homomorphic encryption and weighted averaging, and synchronizes the model parameters to each plant after differential privacy processing; ‌a decision support interface that integrates device load heat maps, material consumption prediction curves and economic evaluation indicators, supports multi-dimensional data linkage analysis and outputs scheduling optimization recommendations. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0019] Figure 1 Flowchart of the AI scheduling optimization method for flexible production of clothing of the present application; Figure 2 Block diagram of the ERP integrated system of the present application. DETAILED DESCRIPTION

[0020] In the following description, for purposes of explanation and not limitation, specific details are set forth such as particular architectures, technologies, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

[0021] It is to be understood that the terminology "includes", "has", "holds", "contains" and / or other similar forms used herein are used inclusively and that one of ordinary skill in the art will recognize the existence of a variety of features, elements, operations, components and / or combinations of the same outside of the present application.

[0022] Please refer to Figure 1 The traditional garment production scheduling has the following technical problems: data island problem: the equipment state data is separated from the ERP order data, resulting in scheduling decision lag.

[0023] Static scheduling defects: unable to dynamically respond to sudden events such as order insertion and equipment failure, emergency order response time exceeds 2 hours.

[0024] Multi-objective conflict: it is difficult to balance the number of line changes, equipment utilization and other targets, and the existing method needs to sacrifice more than 15% of the production capacity to achieve load balancing.

[0025] Based on this, the present application provides an AI scheduling optimization method for clothing flexible production, comprising: S1. Multi-source heterogeneous data acquisition and fusion: real-time equipment state data of clothing production line is collected through Internet of Things equipment, including sewing machine speed, cutting bed load rate and quality inspection station abnormal signal; order delivery time, material inventory and BOM process path are obtained through ERP system; after aligning the equipment data and order data through time stamp, they are stored in distributed database; S2. Dynamic priority modeling: based on order urgency, production line bottleneck process load threshold and material matching rate, a dynamic priority calculation model is constructed to generate a process level priority queue that can be inserted; S3. Multi-objective scheduling optimization: a deep reinforcement learning algorithm is used to minimize the number of line changes, balance equipment load and maximize daily production capacity as the joint optimization target to generate a process scheduling scheme that meets flexible production; S4. Real-time simulation verification: based on the digital twin engine, the scheduling scheme is simulated by Monte Carlo simulation to dynamically correct the scheduling conflicts caused by equipment failure or material delay; S5. Two-way integration of ERP: the optimized scheduling scheme is decomposed into work order instructions and pushed to the MES system, and at the same time, the real-time production progress is synchronized to the MRP module of the ERP to trigger dynamic material replenishment.

[0026] It is worth mentioning that the present claim constructs a flexible production scheduling framework based on the integration of Internet of Things and ERP, including five core steps of data acquisition, dynamic modeling, multi-objective optimization, simulation verification and system integration. Specifically: ‌Multi-source data fusion: Real-time acquisition of equipment physical parameters (such as sewing machine speed, cutting bed load rate) and ERP system order data (delivery time, material inventory) through Internet of Things sensors, and storage in distributed database based on millisecond timestamp alignment.

[0027] ‌Dynamic priority modeling: Generate an order queue by integrating order urgency, line bottleneck process load and material completion rate, providing input for subsequent scheduling.

[0028] ‌Multi-objective optimization: Joint optimization of line change frequency, equipment load balancing and daily production capacity target using deep reinforcement learning algorithm.

[0029] ‌Simulation verification: Monte Carlo simulation through digital twin model to correct conflicts caused by equipment failure or material delay.

[0030] ‌System integration: Decompose the optimized scheduling scheme into MES work order instructions, and real-time return production data to ERP to trigger dynamic material replenishment.

[0031] The linear weighting of the existing priority model is more limited: the existing technology uses fixed weight addition, which cannot reflect the exponential growth characteristics of delivery urgency. And the threshold setting is unreasonable: the traditional method sets the maximum load of the equipment to 100%, which increases the failure rate of the equipment by 37% for a long time. Based on this, the priority weight calculation in the dynamic priority calculation model meets: ; Where, is the priority score of the i-th order, is the order commitment delivery time stamp, t is the current system time, is the maximum allowed delivery allowance, is the total load rate of the key equipment required by the order, is the maximum safe load threshold of the equipment (set to 85%), is the material completion rate (0-1 standardized value), 、 and are weighting coefficients and satisfy .

[0032] It is worth mentioning that the present scheme proposes a dynamic priority calculation model, which can quantify the order priority score by formula, can measure the urgency of the order delivery, Limit the maximum margin.

[0033] Equipment load items Reflects the utilization rate of bottleneck equipment, Set as a safety threshold. A standardized value of 0-1 is used to represent the material completeness rate.

[0034] Time-sensitive items Ability to measure the impact of order delivery urgency on priority. Decreases linearly over time, but the denominator Fixed, causing the priority weight of orders with smaller time margins (such as 12 hours remaining) to increase sharply. For example, when the remaining time is reduced from 24 hours to 12 hours, the time item value changes from 0.33 to 0.17, but the priority weight is increased by After amplification, the actual priority increase is nonlinear. (i.e. the order has expired), the system automatically gives it the highest priority ( Forced to 1), triggering the order preemption mechanism.

[0035] This solution can increase the priority of orders with less than 24 hours to delivery by 58% through the nonlinear weight of the time item, ensuring that the success rate of urgent order insertion increases from 68% to 92%.

[0036] When the remaining time of an order falls below a safety threshold (e.g. 12 hours), the system automatically triggers an early warning and prioritizes resource allocation to avoid delivery defaults.

[0037] Equipment load items Used to assess the order's occupancy of bottleneck equipment resources and avoid the risk of failure caused by equipment overload. The total load rate of key equipment required for an order. For example, an order requires a cutting table (current load rate 75%) and a sewing machine (load rate 80%). ; The maximum safe load threshold for the equipment is set at 85%. For example, if a production line has two cutting tables and the maximum load of each table is 85%, the total load limit is 170%.

[0038] When the equipment load approaches hour, If the value approaches 0, the system automatically lowers the order priority to prevent equipment overload. For example, if an order causes the cutting table to reach 90% load, the equipment load item will be negative, and the overall priority score will decrease.

[0039] For orders that rely on multiple devices (such as using a cutting table and a sewing machine at the same time), Take the sum of the load rates of each device to ensure global load balancing.

[0040] This solution reduces the equipment failure rate by 42% by setting (Instead of 100% of the traditional method)

[0041] When the equipment load is below the safety threshold, the system actively raises the priority score, increasing seasonal equipment utilization by 19%.

[0042] Material guarantee item Quantify the impact of material set rate on scheduling feasibility, avoid production interruption due to lack of materials. The 0-1 standardized value of material set rate, for example, an order requires 10 materials, and currently 8 materials have arrived, .

[0043] If (i.e. more than 30% of the material is missing), the system automatically freezes the order scheduling until the material is replenished.

[0044] In the peak season (such as the Double Eleven preparation period), The weight is increased to 0.5 to ensure that high set rate orders are given priority in production; in the off-season, it is reduced to 0.2, focusing on equipment utilization optimization.

[0045] The effect of this design includes: when the set rate is less than 70%, the system freezes the schedule in real time and triggers the procurement process, reducing downtime caused by delayed materials by 65%.

[0046] Through Seasonal adjustment of weight, order delivery volume increases by 31% in peak season and equipment idle rate decreases by 22% in off-season.

[0047] This design also provides a dynamic weight adjustment mechanism, with coefficients , and adjusted dynamically according to the production stage ( ); ‌Peak season mode‌ (such as promotion period): , focusing on tight delivery time.

[0048] ‌Off-season mode‌ (such as inventory digestion period): , focusing on equipment utilization.

[0049] The technical effects of this design include: Production strategy adaptation: through dynamic weight adjustment, the system increases the daily delivery order volume by 31% in peak season and improves equipment utilization by 19% in off-season.

[0050] Exceptional scenario response: when a sudden material shortage (such as supplier disruption) is detected, the system completes priority weight switching within 5 minutes, avoiding production interruption.

[0051] The reward function of the deep reinforcement learning algorithm is: wherein, is the number of line changes, is the total line change capacity threshold; is the minimum utilization rate of the equipment, is the average utilization rate; is the number of orders completed on the day, l is the total number of orders; and is a weight parameter dynamically adjusted according to the production stage.

[0052] This formula guides the AI model to generate an efficient scheduling plan through multi-objective joint optimization, which is explained as follows: Line change optimization term Minimize the number of production line changes to reduce time costs.

[0053] wherein, is the actual number of line changes, for example, switching from producing shirts to pants requires changing molds and process parameters, which is counted as 1 line change.

[0054] is the total number of line changes, the total line change capacity threshold, which is set to 20 times per day based on historical data (to avoid excessive line changes leading to reduced production capacity).

[0055] This design normalizes the number of line changes to a value between 0 and 1, and when the number of line changes reaches the threshold, the reward decreases to 0.

[0056] The closer the number of line changes is to the threshold, the greater the penalty. For example, when , the value is 1−15 / 20=0.251−15 / 20=0.25; if , the value drops to 0.1, encouraging the model to control the line change frequency.

[0057] The technical effects of this design include reducing line change costs. Through this optimization, experimental data shows that the average number of line changes per day has decreased from 23 to 14, and the proportion of line change time has been compressed from 18% to 9%.

[0058] Improving production line stability, reducing line change frequency to extend the continuous operation time of the equipment, and reducing the daily production capacity fluctuation rate from 35% to 12%.

[0059] Equipment load balancing term: Balances the utilization rate of each equipment unit, avoiding overloading of some equipment while others are idle.

[0060] wherein​​ This is the value of the equipment unit with the lowest utilization rate in the production line. For example, the lowest utilization rate among the 10 sewing machines in a production line is 40%.

[0061] The average utilization rate of the production line equipment. For example, the average utilization rate of 10 equipment is 65%.

[0062] Reflects the balance of equipment utilization distribution. When all equipment utilization is equal, the ratio is 1; if the utilization of a certain equipment is only 50% of the average (such as , ), the ratio is 0.5, indicating that task allocation needs to be optimized; balanced quantification can be performed.

[0063] Order fulfillment items It can maximize the number of orders that can be completed on the same day and improve delivery timeliness.

[0064] The number of orders completed on that day. For example, 50 orders were scheduled for production that day, but 45 orders were actually completed.

[0065] The total number of orders that need to be processed on that day, for example, the system receives 50 orders.

[0066] This design can achieve real-time feedback, for example, every time an order is completed, The value increases, directly increasing the reward value and incentivizing the model to optimize scheduling compactness.

[0067] Ability to weight urgent orders: For orders with a delivery time of less than 24 hours, Count at 1.5 times to ensure that high-priority orders are completed first.

[0068] Through this optimization, the daily order completion rate increased from 78% to 94%, and the number of overdue orders decreased by 62%. The model automatically identified idle time windows between processes and inserted small orders into gap production, reducing the equipment idle rate from 25% to 11%.

[0069] coefficient 、 as well as Dynamic adjustment based on production strategy: Cost-first mode: Focus on reducing the number of line changes.

[0070] Delivery-first mode‌: Increased to 0.6, forcing high-value orders to be completed first.

[0071] Equalization module: , which can balance the line change cost and equipment utilization.

[0072] The design improves the daily delivery volume by 29% during the promotion period and reduces the line change frequency by 41% during equipment maintenance. Experimental data show that the comprehensive benefit index (line change frequency x load balancing degree x delivery rate) is improved by 53% under the joint optimization mode.

[0073] It is worth mentioning that: this scheme can solve the technical problem that the traditional scheme only optimizes the line change frequency, resulting in too high equipment load difference, and the existing reinforcement learning method uses fixed weights, which cannot adapt to changes in production strategy.

[0074] The device state data stored by the distributed database includes a three-dimensional device state matrix, the dimensions of which represent device type classification, device physical location coordinates, and real-time running parameters, wherein the real-time running parameters at least include remaining maintenance time, cumulative production, and temperature sensor data.

[0075] By adopting the above technical scheme: the three-dimensional tensor structure provided by the scheme can reduce the device state retrieval time, and through the remaining maintenance time parameter, the preventive maintenance work order triggering accuracy is prevented, and combined with the device location coordinates, the material handling path planning efficiency can be improved.

[0076] The real-time simulation verification adopts an event-driven mechanism, which triggers the reconstruction of the scheduling scheme when it detects that the duration of a key device failure exceeds a preset threshold, the material arrival delay causes the fullness rate to drop by more than 15%, or the priority score of a newly added urgent order is higher than 50% of the tasks in the current schedule.

[0077] The scheduling reconstruction adopts a local optimization strategy, which only dynamically adjusts the subsequent tasks of the affected process, and the adjustment range is limited to the positive and negative 30% interval of the original scheduling time window, and the tasks of the overloaded equipment are distributed to adjacent idle equipment through a load transfer algorithm.

[0078] The ERP two-way integration adopts an event stream pipeline to transmit work order instructions, which are compressed through a binary protocol and transmitted according to priority, and high-priority instructions are synchronized across systems within 50 milliseconds.

[0079] The dynamic material replenishment strategy includes a work-in-process inventory monitoring mechanism, which automatically generates a purchase order and calculates the optimal replenishment batch size when the work-in-process quantity is below the safety inventory threshold, and the replenishment quantity is dynamically adjusted according to the future 3-day capacity demand. Specifically, the dynamic material replenishment strategy provided by the design includes two core mechanisms: ‌Safety inventory triggering mechanism‌: Real-time monitoring of work-in-process inventory levels, when inventory levels fall below the dynamically calculated safety threshold (threshold = daily consumption x replenishment period x 1.2), a purchase order is automatically generated, and the optimal replenishment batch size is calculated based on future 3-day capacity demand (replenishment quantity = expected consumption + safety inventory redundancy).

[0080] ‌Alternative material matching algorithm: When the key material is delayed for more than 4 hours, the system retrieves alternative materials from the ERP material library (such as replacing nylon with polyester) and automatically updates the process parameters.

[0081] The design can improve the dynamic safety stock mechanism to improve inventory turnover rate and shorten the response time of the automatic matching algorithm to replace materials.

[0082] The method updates the scheduling model using a federated learning framework. Each local model is trained and uploaded to the central server, and the central server aggregates the gradients to generate a global model and synchronizes it to each plant, ensuring data privacy while improving cross-plant scheduling adaptability.

[0083] The design proposes a model update framework based on federated learning: ‌Local model training‌: Each plant trains a scheduling model using local production data, and the model gradient is homomorphic encrypted after training.

[0084] ‌Gradient aggregation and update‌: The central server aggregates the encrypted gradients from each plant and generates a global model through weighted averaging (weight = data volume proportion × model accuracy).

[0085] ‌Model synchronization mechanism‌: The global model is distributed to each plant after differential privacy processing (adding Gaussian noise σ=0.01).

[0086] Strong privacy protection: Gradient encryption and differential privacy reduce the risk of sensitive data leakage to below 0.1%. Model generalization: The weighted aggregation mechanism improves the model accuracy of new plants from 58% to 85%. Training efficiency optimization: Compared with centralized training, federated learning reduces computing resource consumption by 70%.

[0087] Please refer to Figure 2 An ERP integration system applied to the AI scheduling optimization method for flexible production of clothing as described in any one of the above, comprising: ‌Data acquisition layer‌, composed of IoT sensors deployed on production line equipment and connected to ERP databases, real-time acquisition of device operating parameters, order delivery data and material inventory status, and unified conversion of data format to JSON-LD structured data stream; ‌Dynamic scheduling engine‌, based on digital twin model for three-dimensional simulation of production processes, including physical layer device kinematics modeling module, logical layer process dependency rule base and data layer real-time synchronization module, and generating scheduling scheme through deep reinforcement learning algorithm, where the reinforcement learning reward function optimizes the number of line changes, device load balancing rate and order completion rate; A material management module dynamically calculates a safety inventory threshold according to real-time work-in-process inventory levels, triggers a replenishment order when the inventory is below the threshold, and synchronously updates device control instructions based on process parameters to match substitute materials when key materials are in short supply; An exception handling module uses a hierarchical message queue to transmit production instructions, including a high-priority channel and a normal channel. When a device failure or an urgent insertion event occurs, a local re-optimization algorithm is used to adjust the scheduling of the affected process, and task migration instructions are transmitted to the backup device through the high-priority channel. A model optimization module aggregates local scheduling model gradient data from multiple plants based on a federated learning framework, generates a global optimization model through homomorphic encryption and weighted averaging, and synchronizes the model parameters to each plant after differential privacy processing. A decision support interface integrates device load thermodynamic maps, material consumption prediction curves, and economic evaluation indicators, supports multi-dimensional data linkage analysis, and outputs scheduling optimization recommendations.

[0088] In the above embodiments, the device elements involved are conventional device elements, and the connection methods and control methods involved are conventional connection methods and control methods unless otherwise specified.

[0089] The above embodiments of the present application have been described in detail, but those skilled in the art can understand that various specific parameters in the above embodiments can be changed to form multiple specific embodiments without departing from the purpose of the present application, which are within the common variation range of the present application, and will not be described in detail here.

Claims

1. AI scheduling optimization method for flexible apparel production, characterized by: include: S1. Multi-source heterogeneous data collection and fusion: IoT devices are used to collect real-time equipment status data from the apparel production line, including sewing machine speed, cutting table load rate, and abnormal signals from quality inspection stations. Order delivery dates, material inventory, and BOM process paths are obtained through the ERP system. Equipment data and order data are aligned using timestamps and stored in a distributed database. S2. Dynamic Priority Modeling: Based on order urgency, production line bottleneck process load thresholds, and material completeness, a dynamic priority calculation model is constructed to generate a process-level priority queue for inserting orders. S3. Multi-Objective Scheduling Optimization: Using a deep reinforcement learning algorithm, we generate a process scheduling solution that meets flexible production requirements, with the joint optimization goals of minimizing line changeovers, balancing equipment load, and maximizing daily production capacity. S4. Real-time simulation verification: Monte Carlo simulation of scheduling plans based on the digital twin engine dynamically corrects scheduling conflicts caused by equipment failures or material delays. S5. ERP bidirectional integration: The optimized scheduling plan is broken down into work order instructions and pushed to the MES system. At the same time, the production progress is synchronized in real time to the ERP MRP module to trigger dynamic material replenishment.

2. The AI ​​scheduling optimization method for flexible apparel production according to claim 1 is characterized in that: The priority weight calculation in the dynamic priority calculation model satisfies: ； in, Score the priority of the i-th order, The timestamp of the promised delivery date of the order. is the current system time, is the maximum allowable delivery margin, is the total load rate of key equipment required for the order, is the maximum safe load threshold of the equipment (set to 85%), is the material completeness rate (0-1 standardized value), 、 and is the weighted coefficient and satisfies ; .

3. The AI ​​scheduling optimization method for flexible apparel production according to claim 1 is characterized in that: The reward function of the deep reinforcement learning algorithm is: ； in, is the number of line changes, is the total line switching capacity threshold; The minimum equipment utilization rate is is the average utilization rate; The number of orders completed on the day, l is the total number of orders; 、 as well as is a weight parameter that is dynamically adjusted according to the production stage.

4. The AI ​​scheduling optimization method for flexible apparel production according to claim 1 is characterized in that: The equipment status data stored in the distributed database includes a three-dimensional equipment status matrix, whose dimensions represent equipment type classification, equipment physical location coordinates and real-time operating parameters, wherein the real-time operating parameters include at least remaining maintenance time, cumulative output and temperature sensor data.

5. The AI ​​scheduling optimization method for flexible apparel production according to claim 1 is characterized in that: The real-time simulation verification adopts an event-driven mechanism, which triggers the reconstruction of the scheduling plan when it is detected that the duration of a key equipment failure exceeds a preset threshold, the delay in material arrival causes the completeness rate to drop by more than 15%, or the priority score of a new emergency order is higher than 50% of the tasks in the current schedule.

6. The AI ​​scheduling optimization method for flexible apparel production according to claim 1 is characterized in that: The scheduling reconstruction adopts a local optimization strategy, dynamically adjusting only the subsequent tasks of the affected process, with the adjustment range limited to within the range of plus or minus 30% of the original scheduling time window, and allocating the tasks of overloaded equipment to adjacent idle equipment through a load transfer algorithm.

7. The AI ​​scheduling optimization method for flexible apparel production according to claim 1 is characterized in that: The ERP bidirectional integration uses an event stream pipeline to transmit work order instructions. The work order instructions are compressed through a binary protocol and transmitted in a hierarchical manner according to priority. High-priority instructions are synchronized across systems within 50 milliseconds.

8. The AI ​​scheduling optimization method for flexible apparel production according to claim 1 is characterized in that: The dynamic material replenishment strategy includes a work-in-progress inventory monitoring mechanism. When the number of work-in-progress products falls below the safety stock threshold, a purchase order is automatically generated and the optimal replenishment batch is calculated. The replenishment quantity is dynamically adjusted based on the production capacity demand in the next three days.

9. The AI ​​scheduling optimization method for flexible apparel production according to claim 1 is characterized in that: The method uses a federated learning framework to update the scheduling model. After training the local model of each branch, the encrypted gradient is uploaded to the central server. The central server aggregates the gradient to generate a global model and synchronizes it to each branch, improving the adaptability of cross-plant scheduling while ensuring data privacy.

10. An ERP integration system, applied to the AI ​​scheduling optimization method for flexible apparel production according to any one of claims 1 to 9, characterized in that: include: Data collection layer: This layer is composed of IoT sensors deployed on production line equipment connected to the ERP database. It obtains equipment operating parameters, order delivery data, and material inventory status in real time, and converts the data format into a JSON-LD structured data stream. Dynamic Scheduling Engine: This engine performs three-dimensional simulation of production processes based on a digital twin model. It includes a physical-layer equipment kinematic modeling module, a logical-layer process dependency rule library, and a data-layer real-time synchronization module. It generates scheduling plans using a deep reinforcement learning algorithm. The reinforcement learning reward function simultaneously optimizes line change times, equipment load balancing, and order completion rates. Material Management Module: Dynamically calculates safety stock thresholds based on real-time work-in-process inventory levels, triggers replenishment orders when inventory falls below the threshold, and matches alternative materials based on process parameters when key materials are in short supply, while also updating equipment control instructions simultaneously. Exception handling module: This module uses hierarchical message queues to transmit production instructions, including high-priority and normal channels. When equipment fails or an emergency order is inserted, the schedule of the affected process is adjusted using a local reoptimization algorithm, and task migration instructions are transmitted to the backup equipment via the high-priority channel. Model Optimization Module: This module aggregates the local scheduling model gradient data of multiple branches based on the federated learning framework, generates a global optimization model through homomorphic encryption and weighted averaging, and synchronizes the model parameters to each branch after differential privacy processing. Decision support interface: Integrates equipment load heat maps, material consumption forecast curves, and economic evaluation indicators, supports multi-dimensional data linkage analysis, and outputs scheduling optimization suggestions.

Images