Supply chain-oriented intelligent order management method and system

Abstract

The invention relates to the technical field of order management, and discloses a supply chain-oriented intelligent order management method, which comprises the steps of obtaining corresponding multi-modal data through an order demand flow, a production equipment state, logistics sensor dynamic information and an inventory topological graph; analyzing relevance between orders and equipment based on a space-time diagram convolutional network, and generating a capacity allocation scheme; calculating a logistics path planning scheme, predicting a stock stockout risk and generating a replenishment suggestion; if the high-priority order exists, inserting a productivity plan and adjusting an equipment process chain; if resource conflicts occur, dynamically allocating resources; if the path risk value exceeds the threshold value, standby path switching is triggered; adjusting weighting parameters through an adaptive federation algorithm, generating a global strategy and issuing the global strategy to the client; the client dynamically adjusts local configuration and uploads execution effect data in real time; and if abnormity is detected, triggering global strategy regeneration and updating the model through federated learning increment. According to the invention, efficient management of supply chain orders can be realized.

Description

Technical Field

[0001] The present application relates to the technical field of order management, and in particular to a supply chain-oriented intelligent order management method and system. Background Art

[0002] Current supply chain order management systems generally utilize traditional ERP architectures, relying on manual rule configuration and static optimization models for order allocation and resource scheduling. However, these systems lack real-time adaptive closed-loop optimization mechanisms in dynamic environments, making them unable to quickly respond to sudden demand fluctuations and resource conflicts.

[0003] Especially in high-concurrency order scenarios, traditional centralized optimization algorithms have exponentially increasing computational complexity, making it difficult to generate feasible solutions within a limited time, resulting in order response delays and waste of resources.

[0004] From the above, we can see that how to achieve efficient management of supply chain orders still needs to be solved. Summary of the Invention

[0005] In order to achieve efficient management of supply chain orders, this application provides a supply chain-oriented intelligent order management method and system.

[0006] In the first aspect, the present application provides a supply chain-oriented intelligent order management method, which adopts the following technical solutions:

[0007] An intelligent order management method for a supply chain, comprising:

[0008] Extract the corresponding order demand flow through the ERP system. The order demand flow includes order quantity, priority, and delivery time constraints. The corresponding production equipment status is collected through IoT sensors. The production equipment status includes equipment vibration spectrum, temperature curve, and energy consumption fluctuation. The corresponding logistics dynamic information is obtained through logistics sensors. The logistics dynamic information includes transportation path coordinates, vehicle speed, and ambient temperature and humidity. The inventory topology map generated by the warehouse RFID system is obtained. The inventory topology map includes the three-dimensional product location and inventory distribution generated by the warehouse RFID system. The order demand flow, the production equipment status, the logistics dynamic information, and the inventory topology map are aligned through a spatiotemporal synchronization protocol to obtain the corresponding multimodal data.

[0009] Based on the order demand flow and the production equipment status, a pre-set spatiotemporal graph convolutional network is used to analyze the correlation between orders and equipment status to obtain a corresponding capacity allocation plan; based on the logistics dynamic information, a multi-objective optimal path is calculated to obtain a corresponding path planning plan, where the multi-objective optimal path includes cost, timeliness, and carbon emissions; based on the inventory topology map, a Bayesian network is constructed to predict the out-of-stock risk in the next 72 hours and generate corresponding replenishment suggestions;

[0010] Based on the capacity allocation plan, it is determined whether there is a high-priority order to be received. When a high-priority order is received, a new task is inserted into the capacity allocation plan through a lossless order insertion algorithm. Based on the new task, the affected equipment process chain is determined and adjusted; based on the capacity allocation plan, it is determined whether there are multiple orders competing for the same resource. If so, the fuzzy logic arbitrator is called to dynamically allocate resources, generate conflict resolution instructions and execute them; when the risk value corresponding to the path planning plan is greater than the preset risk threshold, the backup path plan is triggered;

[0011] The ERP system, IoT sensors, logistics sensors, and warehouse RFID systems in the supply chain are used as federated learning clients. The warehouse RFID system retrains the ST-GCN model based on equipment status data to extract the corresponding capacity prediction error. The logistics sensors optimize the path model based on the path planning error to predict the probability of carbon emissions exceeding the standard. The warehouse RFID system updates the Bayesian network through the inventory topology map to evaluate the corresponding replenishment response time. An adaptive federated optimization algorithm is used to aggregate the parameters of each client, dynamically weight the weights of nodes related to high-priority orders, and generate a corresponding global strategy. The production equipment of high-priority orders is given higher weights.

[0012] The global policy is sent to each client. After decryption, the client dynamically adjusts the local device configuration, which includes the sensor sampling rate and environmental detection accuracy. After execution, each client obtains the execution effect data corresponding to each client and uploads the execution effect data. If an anomaly is detected, the global policy is regenerated, and the global model is incrementally updated through federated learning.

[0013] Optionally, during the analysis process of the spatiotemporal graph convolutional network, the method also includes: mapping the production equipment status into a spatiotemporal graph structure, wherein the nodes represent the production equipment, and the edge weights are dynamically calculated by the process dependencies between the equipment and the real-time load differences; extracting the short-term fluctuation patterns and long-term trend characteristics of the production equipment status through multi-level spatiotemporal convolution kernels, and generating dynamic capacity allocation weights in combination with order priorities; when it is detected that the health of the equipment has declined, the allocation weights of its associated orders are automatically reduced, and the backup equipment activation protocol is triggered.

[0014] Optionally, in the process of calling the fuzzy logic arbitrator to dynamically allocate resources, the three-dimensional evaluation indicators of resource conflict include order priority, equipment health, and delivery time margin. The method also includes: constructing a fuzzy rule base, when multiple orders compete for the same equipment, giving priority to orders with priority > 0.8 and equipment health > 0.7; if the priority difference of the conflicting orders is < 0.2, then comparing the carbon emission impact factors corresponding to the conflicting orders, and selecting the solution with the lowest carbon emission increase; when generating conflict resolution instructions, synchronously reserving a spare resource window.

[0015] Optionally, during the risk value calculation process of the path planning scheme, the method also includes: constructing a corresponding risk probability model based on real-time traffic flow, weather warning data and historical accident statistics in the logistics dynamic information; obtaining the congestion index and weather disaster warning level corresponding to the path, and when the congestion index of the path is greater than 0.6 or the weather disaster warning level is ≥3, it is marked as a high-risk path; when obtaining the corresponding path planning scheme, the corresponding backup path scheme is synchronously generated, and the backup path scheme pre-generates multiple candidate paths, and dynamically switches the candidate paths according to the real-time risk value, and the switching response time is less than 15 seconds.

[0016] Optionally, in the process of performing anomaly detection and regenerating the global strategy, the method further includes: determining corresponding anomaly criteria, the anomaly criteria including order delivery delay rate > 10% or carbon emissions exceeding the standard > 15%. When an anomaly is detected, the local data of each client is retrieved, and the root device or path of the anomaly is located based on the local data; when incrementally updating the global model, the decision branch of the historical optimal version is retained.

[0017] Optionally, in the process of dynamically adjusting the local device configuration, the method also includes: when the federal policy indicates a high-risk area, the GPS sampling rate of the logistics sensor is increased from 1 Hz to 10 Hz; when the vibration sensor of the production equipment detects a spectrum anomaly, it starts a millisecond-level high-frequency acquisition mode; the scanning interval of the warehouse RFID system is adjusted according to the inventory turnover rate to a scanning interval of ≤2 minutes when the turnover rate is greater than 5 times / day.

[0018] Optionally, the adaptive federated optimization algorithm method also includes: dynamically adjusting the aggregation weight according to the client data quality, and increasing the weight of nodes with data integrity greater than 90% by 30%; encrypting and isolating client parameters associated with high-priority orders, introducing an adversarial verification mechanism, detecting abnormal deviations of parameters uploaded by the client, and automatically downgrading the weight if the deviation is greater than the deviation threshold.

[0019] Secondly, this application provides a supply chain-oriented intelligent order management system, which adopts the following technical solutions:

[0020] An intelligent order management system for supply chain, including:

[0021] A multimodal data acquisition module extracts the corresponding order demand flow through the ERP system. The order demand flow includes order quantity, priority, and delivery time constraints. The corresponding production equipment status is collected through IoT sensors. The production equipment status includes equipment vibration spectrum, temperature curve, and energy consumption fluctuations. The corresponding logistics dynamic information is obtained through logistics sensors. The logistics dynamic information includes transportation path coordinates, vehicle speed, and ambient temperature and humidity. The inventory topology map generated by the warehouse RFID system is obtained. The inventory topology map includes the three-dimensional product location and inventory distribution generated by the warehouse RFID system. The order demand flow, the production equipment status, the logistics dynamic information, and the inventory topology map are aligned through a spatiotemporal synchronization protocol to obtain the corresponding multimodal data.

[0022] The capacity allocation plan acquisition module analyzes the correlation between the order demand flow and the production equipment status through a pre-set spatiotemporal graph convolutional network to obtain the corresponding capacity allocation plan; calculates the multi-objective optimal path based on the logistics dynamic information and obtains the corresponding path planning plan. The multi-objective optimal path includes cost, timeliness, and carbon emissions; and constructs a Bayesian network based on the inventory topology map to predict the out-of-stock risk in the next 72 hours and generate corresponding replenishment suggestions.

[0023] A judgment module is used to determine whether a high-priority order is received based on the capacity allocation plan. When a high-priority order is received, a new task is inserted into the capacity allocation plan through a lossless order insertion algorithm. The affected equipment process chain is determined based on the new task and adjusted. The capacity allocation plan is used to determine whether multiple orders compete for the same resource. If so, a fuzzy logic arbitrator is called to dynamically allocate resources, generate conflict resolution instructions, and execute them. When the risk value corresponding to the path planning plan is greater than a preset risk threshold, an alternative path plan is triggered.

[0024] The global strategy generation module uses the ERP system, IoT sensors, logistics sensors, and warehouse RFID systems in the supply chain as federated learning clients. The warehouse RFID system retrains the ST-GCN model based on equipment status data to extract the corresponding capacity prediction error. The logistics sensors optimize the path model based on the path planning error to predict the probability of carbon emissions exceeding the standard. The warehouse RFID system updates the Bayesian network based on the inventory topology map to evaluate the corresponding replenishment response time. An adaptive federated optimization algorithm is used to aggregate the parameters of each client and dynamically weight the weights of nodes related to high-priority orders to generate the corresponding global strategy. The production equipment of high-priority orders has higher weights.

[0025] The update module sends the global policy to each client. After decryption, the client dynamically adjusts the local device configuration, which includes the sensor sampling rate and environmental detection accuracy. After execution by each client, the execution effect data corresponding to each client is obtained and uploaded. If an anomaly is detected, the global policy is regenerated to incrementally update the global model through federated learning.

[0026] In a third aspect, the present application provides a supply chain-oriented intelligent order management method, which adopts the following technical solutions:

[0027] A supply chain-oriented intelligent order management method includes a processor running a program of any one of the above-mentioned supply chain-oriented intelligent order management methods.

[0028] In a fourth aspect, the present application provides a storage medium, which adopts the following technical solution:

[0029] A storage medium storing a program of any one of the above-mentioned supply chain-oriented intelligent order management methods.

[0030] In summary, this application includes at least one of the following beneficial technical effects:

[0031] First, by integrating multimodal data such as ERP order flow, equipment status, logistics trajectory, and inventory distribution, and based on spatiotemporal graph convolutional networks and Bayesian prediction models, order demand is matched with production, logistics, and warehousing capabilities in real time. This cross-system data alignment and correlation analysis capability can accurately predict capacity gaps, out-of-stock risks, and path risks, thereby planning resource allocation and path optimization in advance, significantly reducing delivery delays and inventory backlogs.

[0032] Secondly, for scenarios such as high-priority order insertion and resource competition, the solution uses a lossless order insertion algorithm and fuzzy logic arbitrator, combined with multi-dimensional indicators such as order priority, equipment health, and carbon emission impact, to achieve dynamic priority allocation of resources and rapid conflict resolution. For example, by reserving backup resource windows and real-time switching of high-risk paths (<15 seconds response), while ensuring the timeliness of critical orders, it avoids waste of production capacity or delivery failures caused by sudden failures or logistics interruptions.

[0033] Finally, by aggregating the data value of each system client through federated learning, data privacy is protected and, through adaptive weight adjustment and anomaly detection mechanisms, the global strategy can dynamically adapt to supply chain fluctuations. For example, strategies such as increasing the weight of devices associated with high-priority orders and adjusting high-frequency sampling rates ensure priority supply of resources for critical paths. The root cause analysis of anomalies and the incremental update capabilities of the model continuously eliminate local decision-making biases and improve the response speed and resource utilization efficiency of the overall supply chain. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The present invention is a flowchart of a supply chain-oriented intelligent order management method according to an exemplary embodiment.

[0035] Figure 2 The figure is a structural block diagram of a supply chain-oriented intelligent order management system according to an exemplary embodiment. DETAILED DESCRIPTION

[0036] Embodiments of the present application are described in detail below, examples of which are illustrated in the accompanying drawings.

[0037] Throughout this specification, reference to the terms "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with the embodiment or example is included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0038] The present application embodiment discloses a supply chain-oriented intelligent order management method, referring to Figure 1 ,include:

[0039] S100 extracts the corresponding order demand flow through the ERP system, collects the corresponding production equipment status through IoT sensors, obtains the corresponding logistics dynamic information through logistics sensors, obtains the inventory topology map generated by the warehouse RFID system, aligns the order demand flow, production equipment status, logistics dynamic information and inventory topology map through the time-space synchronization protocol, and obtains the corresponding multimodal data.

[0040] Among them, for the order demand flow extracted by the ERP system, the order demand flow includes order quantity, priority (such as urgent orders, regular orders), and delivery time constraints (such as deadlines). It can be synchronized through the ERP system API interface or real-time database to extract the order generation time, customer-specified delivery window, SKU (stock keeping unit) involved in the order, etc.; it can provide a dynamic demand benchmark for subsequent production plans to ensure that production capacity allocation matches actual market demand.

[0041] The production equipment status obtained by IoT sensor data includes the equipment vibration spectrum, temperature curve and energy consumption fluctuations; the equipment operating vibration frequency can be monitored by acceleration sensors, and abnormal spectrum may indicate bearing wear or mechanical failure; the temperature curve is a thermocouple or infrared sensor that records the temperature of key equipment components in real time, and high temperature may indicate overload or cooling system failure; energy consumption fluctuations are an electric meter or current sensor that records the power consumption of the equipment, and abnormal fluctuations may reflect a decrease in equipment efficiency or potential failure.

[0042] Sensor data is transmitted to the cloud or edge computing node through an industrial IoT gateway (such as the MQTT protocol), and the sampling frequency is set according to the device type (such as 100Hz for vibration data and 1Hz for temperature data). This allows real-time monitoring of the device health status, providing a basis for production capacity forecasting and fault warning, and avoiding delivery delays caused by equipment downtime.

[0043] The logistics dynamic information obtained by logistics sensors includes transportation route coordinates, vehicle speed and ambient temperature and humidity. The transportation route coordinates can be determined by recording the real-time position of the transport vehicle through GPS or Beidou positioning system; the vehicle speed can be obtained through the on-board OBD device or CAN bus interface to assist in route optimization; the ambient temperature and humidity can be monitored by temperature and humidity sensors to monitor the storage environment of goods during transportation (such as cold chain transportation requires constant temperature).

[0044] By aggregating vehicle data through logistics management platforms (such as TMS systems) and combining historical route data to build dynamic traffic models, it is possible to optimize route planning, avoid congestion or extreme weather, and reduce transportation risks (such as cargo damage caused by temperature control failure).

[0045] The inventory topology map generated by the warehouse RFID system data includes the three-dimensional item location and inventory distribution generated by the warehouse RFID system. The three-dimensional item location is generated by the RFID reader scanning the electronic tags on the warehouse shelves, combined with UWB (ultra-wideband) positioning technology to generate the item's X / Y / Z coordinates. The inventory distribution is calculated by counting the real-time inventory quantity of each SKU and marking out-of-stock or overstock areas. The RFID reader scans periodically (for example, every two minutes) and integrates with the warehouse management system (WMS) to update inventory data. This allows for precise inventory location, supports rapid picking and replenishment decisions, and reduces manual inventory counting errors.

[0046] The specific steps for aligning multimodal data using the spatiotemporal synchronization protocol are as follows: All data sources use UTC timestamps to ensure a consistent time base; heterogeneous data (such as vibration spectrum as time series data and inventory as graph data) is unified into JSON or Parquet format; and, using order generation time as an anchor, production equipment status (such as the temperature of a piece of equipment at the time of order generation), logistics routes (transportation trajectory after order placement), and inventory data (availability of materials required for the order) are mapped to the same time window (e.g., ±5 minutes). Multimodal data is linked to specific production processes or logistics routes through metadata such as equipment ID, order ID, and SKU code.

[0047] Multimodal data can eliminate the time delays and spatial deviations between data sources and provide a consistent data view for subsequent analysis. For example, it can determine whether the raw materials for an order have been received (inventory data) and the production equipment is in an available state (equipment status data), and predict the impact of logistics delays on delivery time (comparison of logistics routes with order deadlines).

[0048] S200 analyzes the correlation between order demand flow and production equipment status through a pre-set spatiotemporal graph convolutional network to obtain the corresponding capacity allocation plan; calculates the multi-objective optimal path based on logistics dynamic information and obtains the corresponding path planning plan. The multi-objective optimal path includes cost, timeliness, and carbon emissions; and constructs a Bayesian network based on the inventory topology map to predict the out-of-stock risk in the next 72 hours and generate corresponding replenishment suggestions.

[0049] Among them, the order and equipment association analysis specifically includes: determining the input data as the order demand flow (quantity, priority, delivery time) and the production equipment status (vibration, temperature, energy consumption), and then using the spatiotemporal graph convolutional network (ST-GCN) to map the production equipment status into a spatiotemporal graph structure, analyzing the correlation between equipment health, process dependencies and order priorities, and thus outputting a dynamic capacity allocation plan to ensure that high-priority orders are matched with healthy equipment first to avoid resource conflicts.

[0050] Through the spatiotemporal graph convolutional network, the equipment health status (such as abnormal vibration and temperature exceeding the limit) is combined with the order priority to adjust the production capacity allocation in real time to avoid production capacity waste caused by equipment failure or inefficiency; high-priority orders (such as urgent orders) are preferentially allocated to equipment with high health and balanced load to ensure the timely delivery of key orders (for example, the on-time delivery rate of high-priority orders is increased to more than 95%); through process dependency analysis, multiple orders are prevented from competing for the same equipment resources, reducing order backlogs or delays caused by resource conflicts.

[0051] Logistics route optimization specifically includes: determining that the input data is logistics dynamic information (route coordinates, speed, temperature and humidity); then calculating the optimal route through a multi-objective optimization algorithm (cost, timeliness, carbon emissions), and simultaneously generating alternative routes to deal with risks; thereby outputting a route planning plan that balances cost and timeliness and reduces the risk of exceeding carbon emissions standards.

[0052] On the premise of meeting delivery time, choose the most cost-effective path (such as choosing a route or vehicle type with higher fuel efficiency) to reduce total logistics costs (for example, transportation costs are reduced by 10% to 15%); reduce carbon emissions by reducing driving distance or choosing low-emission vehicles (such as electric trucks) by optimizing routes (for example, carbon emissions are reduced by more than 20%); monitor risks such as traffic congestion and weather disasters in real time, trigger backup route switching (response time <15 seconds), and avoid delivery delays caused by emergencies (for example, 100% success rate of route switching in extreme weather).

[0053] Inventory forecasting and replenishment recommendations specifically include: determining the input data as an inventory topology map (three-dimensional storage location, inventory quantity); then predicting the risk of out-of-stock in the next 72 hours by building a Bayesian network, combining historical sales and supplier delivery cycles; and outputting replenishment recommendations, including emergency purchase trigger thresholds and dynamic inventory threshold adjustments.

[0054] By predicting SKU-level out-of-stock probabilities 72 hours in advance (e.g., triggering an alert when the probability of a high-value SKU out-of-stock exceeds 30%), order cancellations and customer churn due to stockouts can be avoided. By dynamically adjusting safety stock thresholds (e.g., increasing inventory by 20% when demand fluctuates significantly), inventory backlogs can be reduced (e.g., improving inventory turnover by 25%). Based on the forecast results and supplier delivery cycles, emergency purchase orders can be generated or procurement plans adjusted to ensure timely material delivery (e.g., reducing losses due to stockouts by 40%).

[0055] In an embodiment of the present application, during the analysis of the spatiotemporal graph convolutional network, the method further includes:

[0056] S210, mapping the production equipment status into a spatiotemporal graph structure.

[0057] The nodes represent production equipment. Each production equipment (such as injection molding machine, assembly line) is treated as a node in the graph. The node attributes include real-time status (vibration spectrum, temperature, energy consumption) and health score (based on comprehensive evaluation of sensor data).

[0058] Edge weights are dynamically calculated based on the process dependencies and real-time load differences between devices. Connections between devices are defined based on the production process flow (e.g., injection molding machine → assembly line). Edge weights between devices are dynamically calculated based on two factors: process dependency, where the closer the process, the higher the weight (e.g., weight of adjacent processes = 0.8, weight of non-adjacent processes = 0.3); and load balancing, where the greater the load difference between device A and device B, the lower the weight (e.g., weight difference > 30%, weight reduced by 50%). Edge weights are recalculated every 5 minutes based on the latest sensor data to ensure the graph structure reflects real-time resource status.

[0059] S220 uses multi-level spatiotemporal convolution kernels to extract the short-term fluctuation patterns and long-term trend characteristics of production equipment status, and generates dynamic capacity allocation weights based on order priorities.

[0060] The specific execution steps include: short-term fluctuation pattern extraction, using 1D convolution kernel to analyze equipment sensor data (such as instantaneous anomalies in the vibration spectrum and sudden temperature rise) to identify equipment health risks; long-term trend feature analysis, using time convolution kernel to analyze the long-term changes in the equipment energy consumption curve (such as the average monthly growth trend of energy consumption) to predict equipment aging or performance degradation.

[0061] The dynamic weight generation process specifically includes: health weighting, where equipment with low health (such as abnormal vibration or high temperature) has its weight reduced; load balancing weighting, where equipment with excessive load has its weight reduced; and priority weighting, where high-priority orders have their weight increased (such as a +20% weight for urgent orders). The final capacity allocation weight is generated by weighting health, load, and priority. This generates a matching table for equipment and orders, annotating the estimated completion time and resource conflict markers to output a capacity allocation plan:

[0062] By combining equipment health, load, and order priority, we can maximize the on-time delivery rate of critical orders (such as increasing the on-time delivery rate of high-priority orders to 98%); and reduce the allocation of aging equipment (health < 0.7) to avoid delivery delays caused by failures.

[0063] S230, when it is detected that the health of the device has deteriorated, the allocation weight of its associated orders is automatically reduced, and the backup device activation protocol is triggered.

[0064] The health monitoring and triggering process specifically includes: real-time monitoring of equipment health (such as abnormal vibration, temperature exceeding the limit, and energy consumption fluctuations). When the health level falls below a threshold (such as <0.6), a response is triggered, and the capacity allocation weight of the equipment is automatically reduced. For example, when the health level is <0.6, the weight is reduced by 50%; when the health level is <0.4, almost no new tasks are assigned. Backup equipment is also activated: backup equipment (such as a backup injection molding machine) is started, the preheating time is controlled within 30 minutes, and unfinished orders from the original equipment are reallocated to backup equipment with a health level greater than 0.8. These orders are marked as "high-risk orders" and trigger the manual intervention process.

[0065] For example, when an injection molding machine fails due to high temperature, the system switches to backup equipment within 5 minutes to avoid order delays; through preventive adjustments, the downtime losses caused by sudden equipment failures are reduced (for example, the average repair time is 4 hours).

[0066] Intelligent supply chain management is achieved through three core steps: order and equipment association analysis, logistics path optimization, and inventory forecasting. First, based on the spatiotemporal graph convolutional network, the equipment health status and order priority are dynamically matched to ensure priority delivery of key orders and improve equipment utilization by 15% to 20%. Second, a multi-objective path optimization algorithm is used to balance cost, timeliness, and carbon emissions, reducing transportation costs and environmental impact while ensuring path reliability. Finally, a Bayesian network is used to predict out-of-stock risks and generate replenishment strategies to reduce inventory backlogs and out-of-stock losses, ultimately driving the supply chain from passive response to active prediction, achieving coordinated optimization of efficiency, cost control, and sustainability.

[0067] S300, based on the capacity allocation plan, determines whether there is a high-priority order to be received. When a high-priority order is received, a new task is inserted into the capacity allocation plan through a lossless insertion algorithm, and the affected equipment process chain is determined based on the new task and adjusted; based on the capacity allocation plan, determines whether there are multiple orders competing for the same resource. If so, calls the fuzzy logic arbitrator to dynamically allocate resources, generates conflict resolution instructions and executes them; when the risk value corresponding to the path planning plan is greater than the preset risk threshold, triggers the backup path plan.

[0068] The specific execution process for high-priority order insertion includes the following: Input data is the real-time high-priority order received and the current capacity allocation plan; then, using a lossless order insertion algorithm, the new task is inserted into the capacity plan, and its impact on the equipment process chain (such as delaying other orders or occupying resources) is analyzed; and then an adjusted capacity allocation plan is output. This ensures that high-priority orders are executed first while minimizing interference with the original plan.

[0069] For resource conflict arbitration, the specific execution process includes: the input data is multiple orders in the capacity allocation plan competing for the same resources (such as equipment, raw materials); then by calling the fuzzy logic arbitrator, resources are dynamically allocated based on three-dimensional evaluation indicators (order priority, equipment health, and delivery time margin); thereby conflict resolution instructions are issued, and the resolution instructions include resource allocation results and backup resource reservation strategies.

[0070] For path risk response, the specific execution process includes: the input data is the risk value of the path planning plan (such as congestion index, weather warning level). When the risk value exceeds the threshold, the backup path plan is triggered and the candidate path is dynamically switched in real time; thus, the updated path planning plan ensures transportation reliability.

[0071] In an embodiment of the present application, during the process of calling the fuzzy logic arbiter to dynamically allocate resources, the method further includes:

[0072] S310, construct a fuzzy rule base. When multiple orders compete for the same equipment, orders with priority > 0.8 and equipment health > 0.7 are preferentially allocated.

[0073] The rule base defines fuzzy rules based on business needs. For example, in Rule 1, if the order priority is greater than 0.8 and the equipment health is greater than 0.7, the order will be assigned first. In Rule 2, if the equipment health is less than 0.5, new tasks will be refused.

[0074] When it is detected that multiple orders compete for the same equipment (for example, Order A and Order B both require injection molding machines), the fuzzy rule library is called for evaluation, and then priority determination is performed, that is, the priority score of each order is calculated (in the range of 0-1). If an order meets the conditions of Rule 1, resources are directly allocated.

[0075] By clarifying the priority rules, order backlogs caused by delays in manual intervention can be avoided, ensuring that key orders (such as high-value customer orders) are executed first; it can also achieve equipment health protection, refuse to assign tasks to low-health equipment, and reduce secondary risks caused by equipment failure.

[0076] S320: If the priority difference of the conflicting orders is less than 0.2, the carbon emission impact factors corresponding to the conflicting orders are compared, and the solution with the lowest carbon emission increment is selected.

[0077] Among them, the specific execution steps include: when the priority difference of the conflicting orders is less than 0.2 (such as order A = 0.85, order B = 0.80), enter the carbon emission comparison link, and then perform carbon emission calculations to analyze the incremental carbon emissions of the two orders in production and transportation (such as order A requires additional high-energy consumption equipment to be started, and order B can reuse existing resources); if the carbon emission increment of order A is 10% higher than that of order B, order B will be allocated first to reduce environmental impact.

[0078] When priorities are similar, choose options with lower carbon emissions to comply with ESG (environmental, social, and governance) requirements and reduce carbon tax or compliance risks; it can also reduce energy waste and lower operating costs (such as electricity bills and carbon quota purchase costs) in the long term.

[0079] S330: When generating a conflict resolution instruction, a spare resource window is synchronously reserved.

[0080] Among them, Du Yu's reserved window definition reserves an additional 10% to 15% of equipment capacity or inventory resources as backup when generating conflict resolution instructions. For example, if a certain equipment is allocated 80% of its capacity for order A, 20% is reserved as backup. If a higher-priority order is subsequently inserted, the backup resources can be called upon to quickly respond and avoid triggering another conflict. It should be noted here that if the reserved resources are not used within 48 hours, they will be released back to the public resource pool.

[0081] Reserved resources provide a buffer for sudden orders or equipment failures, reducing delivery failures due to insufficient resources. It can also avoid idle resources caused by excessive reservation and achieve efficient resource utilization through dynamic monitoring.

[0082] In addition, in the embodiment of the present application, during the risk value calculation process of the path planning solution, the method further includes:

[0083] S340, builds a corresponding risk probability model based on real-time traffic flow, weather warning data and historical accident statistics in logistics dynamic information.

[0084] The specific implementation steps include:

[0085] First, data integration is carried out: real-time traffic data, access to navigation APIs (such as Gaode and Baidu) to obtain real-time congestion index; weather warnings, integrating disaster warning levels such as heavy rain and heavy fog from the Meteorological Bureau API; historical accident data, statistics on the number and type of accidents (such as car accidents and landslides) on a certain road section in the past three months.

[0086] Then, the risk probability calculation is performed:

[0087] The path risk value is calculated by weighting the three types of data, with the congestion index accounting for 40%, the weather warning accounting for 30%, and the historical accidents accounting for 30%. For example, if the congestion index of a path is 0.7 (high) and the rainstorm warning is level 3 (medium), the risk value = 0.7×0.4+0.3×0.3+0.3×0.3=0.46 (high risk).

[0088] For threshold setting, a risk value > 0.6 or a weather warning ≥ Level 3 is marked as a high-risk route. By converting abstract risks into calculable values, a basis for route switching is provided, thus avoiding reliance on human experience and reducing route planning errors.

[0089] S350: Obtain the congestion index and weather disaster warning level corresponding to the path. When the congestion index of the path is greater than 0.6 or the weather disaster warning level is greater than or equal to level 3, the path is marked as a high-risk path.

[0090] The specific implementation steps include:

[0091] First, a real-time risk assessment is conducted, updating the route risk value every five minutes. If a route exceeds a threshold (e.g., a congestion index > 0.6), it is marked as high risk. Next, backup routes are pre-generated. Based on historical data and real-time traffic conditions, three to five candidate routes are calculated in advance (e.g., detouring via highways, switching transportation modes). Next, route switching preparation is performed, presetting the conditions for route switching (e.g., if the original route risk value remains > 0.8 for more than 10 minutes) and testing the switching response time (which must be < 15 seconds).

[0092] Switching routes before congestion or disasters occur can reduce the risk of delays (such as switching to inland routes before a typhoon); and route switching within 15 seconds can avoid long-term vehicle stagnation and ensure delivery timeliness.

[0093] S360: When obtaining the corresponding path planning solution, a corresponding backup path solution is generated synchronously. The backup path solution pre-generates multiple candidate paths and dynamically switches the candidate paths according to the real-time risk value, and the switching response time is less than 15 seconds.

[0094] The specific implementation steps include:

[0095] First, a switch is triggered. When the risk value of the original route exceeds a threshold and an alternative route is available, a switch instruction is issued to the logistics management system. Then, real-time monitoring and deviation correction are implemented. Every 30 seconds, the vehicle's actual path is compared with the planned path, and deviations are automatically corrected (for example, re-planning is required if the GPS offset exceeds 500 meters).

[0096] The last step is post-switching evaluation, which records the actual time consumption, carbon emissions and other data after the path switching for subsequent model optimization.

[0097] For example, when a route is marked as high-risk due to heavy rain, the system switches to an alternative route within 5 minutes to ensure that the goods are delivered on time; and the accuracy of the route model is continuously improved through feedback data after the switch.

[0098] Through three core mechanisms: a lossless order insertion algorithm, fuzzy logic arbitration, and dynamic path switching, the S300 achieves flexible execution and intelligent decision-making within the supply chain. First, high-priority orders can be seamlessly inserted into the capacity plan, ensuring that critical tasks are prioritized and minimally disrupted with the original plan (urgent order delay rate ≤ 5%). Second, a fuzzy rule base based on priority, equipment health, and carbon emissions rapidly resolves resource conflicts and reserves backup resources, improving conflict resolution efficiency by 70%. Finally, through real-time risk assessment and millisecond-level path switching, the delivery failure rate caused by logistics disruptions is reduced by over 60%. Overall, the S300 enables the supply chain to rapidly respond to complex scenarios such as sudden orders, equipment failures, or extreme weather, significantly improving delivery on-time rates and resource utilization efficiency.

[0099] S400 uses the ERP system, IoT sensors, logistics sensors, and warehouse RFID systems in the supply chain as federated learning clients. The warehouse RFID system retrains the ST-GCN model based on equipment status data to extract the corresponding production capacity prediction error. The logistics sensor optimizes the path model based on the path planning error and predicts the probability of carbon emissions exceeding the standard. The warehouse RFID system updates the Bayesian network through the inventory topology map to evaluate the corresponding replenishment response time. It uses an adaptive federated optimization algorithm to aggregate the parameters of each client, dynamically weight the weights of nodes related to high-priority orders, and generate a corresponding global strategy. The production equipment of high-priority orders has a higher weight.

[0100] The specific implementation process includes:

[0101] 1. Federated Learning Client Initialization

[0102] Client registration: ERP, IoT sensors, logistics sensors, and warehouse RFID systems are registered as federated learning nodes through security protocols.

[0103] Data permission division: ERP, sharing order history data (excluding sensitive customer information); IoT sensors, sharing desensitized equipment status data (such as statistical characteristics of vibration spectrum); logistics sensors, sharing path planning errors and carbon emission data; warehouse RFID, sharing inventory distribution and replenishment response time data; data encryption, all data is encrypted locally, and only encrypted model parameters are uploaded instead of original data.

[0104] 2. Warehouse RFID System: ST-GCN Model Training and Capacity Error Analysis

[0105] Local model training: Input: equipment status data (vibration, temperature, energy consumption); goal: optimize the ST-GCN model to predict the equipment's production capacity in the next 24 hours; training method: use local historical data (such as equipment failure records and production capacity fluctuations) to adjust model parameters.

[0106] Error analysis: Calculates the error between the model's predicted capacity and actual capacity (e.g., if an injection molding machine predicts a capacity of 100 pieces per day and actually produces 95 pieces, the error is 5%). Flags devices with high errors (e.g., error > 15%), triggering a re-inspection of the equipment's health.

[0107] By optimizing the model with local data, we can reduce production capacity waste caused by forecast deviations (e.g., reducing errors by 20%). It’s important to note that high-error equipment may indicate potential failures, triggering maintenance in advance (e.g., scheduling maintenance if a piece of equipment has an error >10% for three consecutive days).

[0108] 3. Logistics Sensors: Path Model Optimization and Carbon Emissions Prediction

[0109] Local model training: Input: path planning error data (such as the difference between actual and planned time); goal: optimize the path model and predict the incremental carbon emissions of the new path; training method: combine real-time traffic data with historical accident records to adjust path weights (such as reducing the weight of congested sections by 30%).

[0110] Carbon emission prediction: Analyze the carbon emission factors of different pathways (e.g., the emission coefficient of electric trucks is 0.2kg / km, while that of diesel vehicles is 0.5kg / km); predict the over-threshold pathway (e.g., the carbon emissions of a certain pathway are 15% greater than the government limit).

[0111] By performing route optimization, delays caused by route deviations can be reduced (e.g., route time errors can be reduced from 15% to 5%). Fines due to emissions exceeding limits can also be avoided, supporting ESG goals (e.g., annual carbon emissions reduction by 10%).

[0112] 4. Warehouse RFID System: Bayesian Network Update and Replenishment Response Evaluation

[0113] Bayesian network update: Input: inventory topology map (3D storage location distribution, inventory level changes); goal: update the dependencies between nodes (such as the correlation between the out-of-stock probability of a certain SKU and the supplier's delivery cycle); training method: combine new replenishment data with demand fluctuations to adjust the node probability distribution.

[0114] Replenishment response assessment: Calculates the time required for inventory to return to a safe threshold after a replenishment order is issued (e.g., replenishment response time < 2 hours after an emergency order is triggered). This can reduce out-of-stocks or backlogs caused by forecast bias (e.g., increase inventory turnover by 20%) and reduce out-of-stock losses by quickly responding to demand fluctuations (e.g., reduce out-of-stock rates for high-value SKUs by 40%).

[0115] 5. Adaptive federated optimization and global strategy generation

[0116] Parameter aggregation: Each client uploads encrypted model parameters (such as ST-GCN weights and path model coefficients), and uses an adaptive federated optimization algorithm (such as an improved version of FedAvg) to weight and aggregate the parameters: the weight of high-priority orders is increased, and the parameter weight of key production equipment is increased by 30% (for example, the weight of a customer order = 1.3); data quality is weighted, and the weight of nodes with data integrity > 90% is increased by 20%.

[0117] Anomaly detection and isolation: Detects abnormal nodes (e.g., sensor data deviates from the global mean by >30%) and automatically downgrades or isolates them.

[0118] By prioritizing updates to parameters related to high-priority orders, priority supply of critical path resources can be ensured; by integrating multi-client data, the model's adaptability to supply chain fluctuations can be improved (for example, when a new factory joins, its equipment characteristics can be quickly adapted).

[0119] 6. Strategy Execution and Feedback Loop

[0120] Policy distribution: Encrypt and distribute global policies (such as capacity allocation weights and path model parameters) to each client.

[0121] Local execution: ERP updates order priority rules; logistics system switches to optimized routes; warehouse RFID system adjusts replenishment thresholds.

[0122] Effect feedback: Collect execution data (such as order delivery on-time rate and actual carbon emissions) and upload it to the federated learning center; if an anomaly is detected (such as delivery delay rate >10%), trigger an incremental model update (only repair the abnormal module).

[0123] By continuously iterating the model, local policy deviations can be avoided (such as equipment failure in a factory leading to a deviation in global production capacity forecast); and after an anomaly is triggered, model repair and policy updates can be completed within 48 hours.

[0124] By integrating ERP, IoT sensors, logistics sensors, and warehouse RFID systems as distributed clients through federated learning technology, global supply chain optimization is achieved through local model training and global parameter aggregation while protecting data privacy: various clients (such as warehouse RFID optimized capacity forecasting model, logistics sensor optimized path and carbon emission forecast, RFID updated inventory replenishment model) work together to improve forecast accuracy, while dynamically weighting high-priority order-related resources (such as increasing the weight of key equipment) to ensure that urgent orders are executed first (punctuality rate > 98%); and through real-time feedback and adaptive parameter updates, it can quickly respond to sudden scenarios such as equipment failure and path interruption (such as path switching response < 15 seconds), ultimately enabling the supply chain to have self-evolution capabilities, significantly improving efficiency, reducing risks, and becoming the support for the core competitiveness of the enterprise.

[0125] S500 sends the global policy to each client. After decryption, the client dynamically adjusts the local device configuration, which includes the sensor sampling rate and environmental detection accuracy. After execution, each client obtains the execution effect data corresponding to each client and uploads the execution effect data. If an anomaly is detected, the global policy is regenerated and the global model is incrementally updated through federated learning.

[0126] The specific implementation process includes:

[0127] 1. Global policy encryption and delivery

[0128] First, policy encapsulation occurs: the global policy generated by the S400 (such as device configuration parameters, path weights, and replenishment thresholds) is encrypted and packaged to ensure data security. Next, client distribution occurs: the encrypted policy is distributed to various federated learning clients (ERP, IoT sensors, logistics sensors, and warehouse RFID systems) via secure channels (such as HTTPS). Finally, decryption verification occurs: the client decrypts the policy using its private key and verifies data integrity (e.g., through hashing).

[0129] Based on the above three steps, the policy can be prevented from being intercepted or tampered with, ensuring the confidentiality and reliability of the supply chain policy; and ensuring that each client only receives the configuration parameters related to it (for example, the warehouse RFID system only receives the inventory replenishment policy).

[0130] 2. The client dynamically adjusts the local device configuration

[0131] The first step is device configuration parsing: the client parses the parameters in the global policy, for example:

[0132] Sensor sampling rate adjustment: IoT sensors increase the vibration sampling rate of key equipment (such as high-priority order production equipment) from 1Hz to 10Hz; environmental monitoring accuracy optimization: Logistics sensors adjust the temperature and humidity monitoring accuracy based on the path risk level (for example, the accuracy of high-risk paths is increased to ±0.5°C); replenishment threshold update: The warehouse RFID system adjusts the safety stock threshold of a certain SKU from 100 pieces to 120 pieces.

[0133] Then comes the configuration validation: dynamically adjusting device parameters without manual intervention (such as remotely controlling sensors through an API).

[0134] High-frequency sampling of key equipment can detect faults earlier (such as warning of bearing abnormalities 3 hours in advance). It can also improve monitoring accuracy in high-risk scenarios and reduce decision-making bias caused by data errors.

[0135] 3. Execution effect data collection and upload

[0136] The first step is data collection: after the client executes the strategy, the execution effect data is collected in real time, including: on the equipment side, the adjusted production capacity achievement rate and equipment failure rate; on the logistics side, the actual route time and actual carbon emissions; on the warehousing side, the replenishment response time and inventory turnover rate.

[0137] Then comes data encryption and uploading: the data is encrypted and uploaded to the federated learning center to ensure privacy and security.

[0138] For example, verify whether path optimization reduces the actual time consumption (e.g., the original path takes 3 hours → 2.5 hours after optimization); if the failure rate of a certain device suddenly increases, mark it as an abnormal trigger point.

[0139] 4. Anomaly Detection and Global Strategy Regeneration

[0140] The first step is to set abnormal thresholds: preset abnormal thresholds for key indicators (such as delivery delay rate >10%, carbon emissions exceeding the standard >15%).

[0141] Then comes real-time detection: the federated learning center compares the execution performance data with the expected target and triggers an alarm if the threshold is exceeded. For example, the actual carbon emissions of a certain path are 20% higher than the predicted value and are marked as an anomaly.

[0142] Finally, the strategy is regenerated: root cause analysis is performed to trace the relationship between anomalies and strategies (for example, the path model does not consider sudden weather changes); the S200-S400 modules are called to rerun capacity allocation, path planning, and federated learning optimization to generate a new strategy.

[0143] For example, when a typhoon causes a route interruption, an alternative route plan can be generated within 5 minutes; and through continuous iteration, decision-making failures caused by environmental changes can be reduced.

[0144] 5. Federated Learning Incrementally Updates the Global Model

[0145] The first is incremental data integration: the execution data of abnormal scenarios (such as path data under extreme weather conditions) are added to the training set as incremental samples.

[0146] Then comes local model fine-tuning: each client fine-tunes the local model based on incremental data (such as logistics sensors optimizing the path model for heavy rain scenarios).

[0147] Finally, parameter aggregation and updating: Use a federated learning incremental algorithm (such as FedProx) to aggregate fine-tuned parameters and update the global model. Only the anomaly-related modules are updated (for example, only the path model is optimized and the stable parameters of the capacity model are retained). In other words, the decision branch with the historical optimal version is retained.

[0148] After accumulating extreme weather data, the path planning model's prediction accuracy in rainstorm scenarios increased by 30%; and incremental updates avoided retraining the full model, saving computing resources (such as shortening training time by 50%).

[0149] It should be pointed out here that S500, as a closed-loop execution and feedback optimization module for the supply chain, encrypts and sends global policies to each client (such as ERP, sensors, and RFID systems), dynamically adjusts local device configurations (such as sensor sampling rate and environmental monitoring accuracy), and collects execution effect data in real time, forming a closed loop of "policy execution-effect feedback-abnormal response-model evolution". When anomalies such as delivery delays, excessive carbon emissions, or equipment failures are detected, the global policy is regenerated within 5 minutes, and the model is incrementally updated through federated learning (such as optimizing predictions by combining extreme weather path data), ultimately achieving dynamic responsiveness and adaptive evolution of the supply chain. This can not only quickly resolve anomalies (such as path interruption recovery time <15 minutes), but also improve prediction accuracy through continuous learning (such as a 40% increase in path planning accuracy in extreme scenarios), significantly enhancing the efficiency and risk resistance of the supply chain while ensuring data privacy.

[0150] In an embodiment of the present application, during the process of performing anomaly detection and regenerating a global strategy, the method further includes: determining corresponding anomaly criteria, the anomaly criteria including an order delivery delay rate > 10% or a carbon emission exceedance > 15%. When an anomaly is detected, the local data of each client is retrieved, and the root device or path of the anomaly is located based on the local data.

[0151] Specifically, the following steps are included:

[0152] 1. Abnormality Criteria and Threshold Definition: Thresholds are defined, such as when order delivery delays exceed 10% or when carbon emissions exceed 15% to trigger an abnormality. Dynamic adjustments are implemented to temporarily adjust thresholds based on scenarios (such as promotional periods or extreme weather conditions) (e.g., increasing the delay tolerance to 15% during holidays). This avoids false positives and false negatives, and triggers responses only for critical abnormalities.

[0153] 2. Real-time monitoring and anomaly triggering: Data comparison: Real-time indicators are compared against thresholds every 5 minutes. Any exceeding thresholds are flagged as an anomaly (e.g., a path latency rate of 12%). Classification and marking distinguish between local anomalies (equipment failure) and global anomalies (extreme weather). This allows for rapid response, triggering anomaly handling within 5 minutes and minimizing the impact.

[0154] 3. Client Data Retrieval and Root Cause Location: Data requests retrieve relevant client data (such as faulty device logs and congested road conditions). Root cause analysis uses time alignment and correlation analysis to locate the root cause (e.g., "device X's abnormal vibration caused an order delay"). This allows for precise location, avoiding over-adjustment of global policies and fixing only the root cause.

[0155] 4. Global Strategy Regeneration: For local anomalies, only the affected modules are adjusted (e.g., reducing the weight of faulty equipment). For global anomalies, capacity allocation, route planning, and federated learning optimization are rerun to generate a new strategy. Root cause injection incorporates anomaly scenario data into the training set to optimize the model's predictive ability for similar issues. This enables rapid remediation, generating a new strategy within 5 minutes, such as switching routes within 15 minutes in extreme weather conditions.

[0156] By precisely setting exception criteria (order delay rate >10% or carbon emissions exceeding the standard >15%), real-time monitoring and root cause location, the S500's exception handling process can trigger a response within 5 minutes, and generate repair strategies (such as dynamic path switching) within 15 minutes for local problems (such as equipment failure) or global risks (such as extreme weather). At the same time, by injecting abnormal scenario data to optimize the model (such as a 40% increase in the accuracy of path planning in heavy rain), while ensuring data privacy, it achieves precise repair, rapid response and continuous evolution, significantly reducing supply chain risks (delivery failure rate reduced by 60%), and improving resource utilization efficiency and risk resistance.

[0157] In an embodiment of the present application, during the process of dynamically adjusting the local device configuration, the method also includes: when the federal policy indicates a high-risk area, the GPS sampling rate of the logistics sensor is increased from 1Hz to 10Hz; when the vibration sensor of the production equipment detects a spectrum anomaly, it starts the millisecond-level high-frequency acquisition mode; the scanning interval of the warehouse RFID system is adjusted according to the inventory turnover rate to a scanning interval of ≤2 minutes when the turnover rate is greater than 5 times / day.

[0158] The specific implementation process includes the following steps:

[0159] 1. Triggering high-frequency GPS sampling for logistics sensors in high-risk areas: Triggering conditions: Federal policy marks an area as high-risk (such as a congestion index greater than 0.6 or a weather warning ≥ Level 3). Configuration adjustments increase the GPS sampling rate of logistics sensors from 1Hz (1 time / second) to 10Hz (10 times / second), allowing for real-time, high-frequency collection of vehicle location data. This policy is only implemented for logistics equipment in high-risk areas (such as trucks in transit).

[0160] 2. Vibration sensor spectrum anomalies trigger millisecond-level high-frequency acquisition: Trigger condition: The vibration sensor of the production equipment detects a spectrum anomaly (such as a sudden increase in a specific frequency component, indicating bearing wear or looseness); configuration adjustment: The vibration sensor starts the millisecond-level high-frequency acquisition mode (such as increasing the sampling rate from 100Hz to 10kHz) to capture subtle vibration changes; the execution scope is limited to abnormal equipment (such as injection molding machines and CNC machine tools).

[0161] 3. Dynamic adjustment of RFID scanning intervals driven by inventory turnover: Trigger condition: The warehouse RFID system detects that the inventory turnover rate of a certain SKU is greater than 5 times / day (high frequency in and out of the warehouse). Configuration adjustment shortens the RFID scanning interval from the default 5 minutes to ≤ 2 minutes to track inventory changes in real time. Implementation scope: only for high-turnover SKUs (such as best-selling items or out-of-stock materials).

[0162] By dynamically adjusting local device configurations, the supply chain achieves precise and intelligent resource adaptation at key links: In high-risk areas, the GPS sampling rate of logistics sensors is increased to 10Hz, improving positioning accuracy tenfold, supporting millisecond-level path correction (e.g., to avoid sudden congestion), and reducing the risk of delays. In equipment health monitoring, vibration sensors use 10kHz high-frequency data acquisition to provide early warning of faults (e.g., bearing cracks) 3-5 hours in advance, reducing unplanned equipment downtime by 40%. In inventory management, the RFID scanning interval for high-volume SKUs is shortened to 2 minutes, accelerating out-of-stock responses and reducing the out-of-stock rate of high-value SKUs by 30%, while also avoiding over-scanning of low-volume inventory. These adjustments, through scenario-based resource allocation, improve data accuracy and decision-making speed while reducing computing power waste, forming a closed loop of dynamic response and resource optimization, significantly enhancing the supply chain's risk resistance and operational efficiency.

[0163] In an embodiment of the present application, the adaptive federated optimization algorithm and the method also include: dynamically adjusting the aggregation weight according to the client data quality, and increasing the weight of nodes with data integrity greater than 90% by 30%; encrypting and isolating client parameters associated with high-priority orders, introducing an adversarial verification mechanism, detecting abnormal deviations of parameters uploaded by the client, and automatically downgrading when the deviation is greater than the deviation threshold.

[0164] The specific implementation process includes the following steps:

[0165] 1. Dynamic weight adjustment of data quality: First, data assessment is performed to quantify the client data quality (such as data completeness, timeliness, and consistency). Then, weight adjustment rules are implemented. If the client data completeness is greater than 90%, its weight in the global parameter aggregation is increased by 30% (for example, the original weight is 1.0 → 1.3). The corresponding execution scope is all clients participating in federated learning (such as ERP, IoT sensors, and logistics sensors).

[0166] 2. Encrypted isolation and adversarial verification of high-priority order parameters: The first step is encryption isolation. Client parameters associated with high-priority orders (such as key equipment weights and path model parameters) are independently encrypted and stored, isolated from other common order parameters. Then there is adversarial verification, which specifically includes: deviation detection, comparing the client-uploaded parameters with the global model mean, calculating the deviation (such as the deviation between the client value of parameter X and the global mean is >15%); threshold triggering, if the deviation is greater than the threshold (such as 15%), it is judged as abnormal and the client weight is automatically reduced (such as the weight from 1.3 to 0.7). Finally, the corresponding execution scope, high-priority order-related clients (such as key production equipment and core logistics paths).

[0167] Through the adaptive federated optimization algorithm, the supply chain achieves data quality-driven dynamic weight adjustment and secure isolation of high-priority parameters: the weight of clients with data integrity greater than 90% is increased by 30%, ensuring that high-quality data dominates the global model; high-priority order-related parameters are encrypted and isolated, and parameter deviations are detected in real time through adversarial verification (such as automatic weight reduction when the deviation is greater than 15%) to prevent malicious attacks or abnormal data interference; ultimately, the credibility, security and anti-interference ability of the federated learning model are improved (such as reducing the impact of abnormal nodes by 50%), while ensuring the priority supply of key order resources, accelerating the model convergence efficiency, and providing reliable decision-making support for the global optimization of the supply chain.

[0168] The present application embodiment discloses an intelligent order management system for supply chain, referring to Figure 2 ,include:

[0169] Multimodal data acquisition module 001 extracts the corresponding order demand flow through the ERP system. The order demand flow includes order quantity, priority and delivery time constraints. The corresponding production equipment status is collected through the Internet of Things sensor. The production equipment status includes equipment vibration spectrum, temperature curve and energy consumption fluctuation. The corresponding logistics dynamic information is obtained through logistics sensors. The logistics dynamic information includes transportation path coordinates, vehicle speed and ambient temperature and humidity. The inventory topology map generated by the warehouse RFID system is obtained. The inventory topology map includes the three-dimensional goods location and inventory distribution generated by the warehouse RFID system. The order demand flow, production equipment status, logistics dynamic information and inventory topology map are aligned through the time-space synchronization protocol to obtain the corresponding multimodal data.

[0170] The capacity allocation plan acquisition module 002 analyzes the correlation between orders and equipment status based on the order demand flow and production equipment status through a pre-set spatiotemporal graph convolutional network to obtain the corresponding capacity allocation plan; calculates the multi-objective optimal path based on logistics dynamic information and obtains the corresponding path planning plan. The multi-objective optimal path includes cost, timeliness, and carbon emissions; constructs a Bayesian network based on the inventory topology map to predict the out-of-stock risk in the next 72 hours and generate corresponding replenishment suggestions.

[0171] Judgment module 003 is used to determine whether there are high-priority orders to be received based on the capacity allocation plan. When a high-priority order is received, a new task is inserted into the capacity allocation plan through a lossless insertion algorithm, and the affected equipment process chain is determined based on the new task and adjusted; it is used to determine whether there are multiple orders competing for the same resource based on the capacity allocation plan. If so, the fuzzy logic arbitrator is called to dynamically allocate resources, generate conflict resolution instructions and execute them; when the risk value corresponding to the path planning plan is greater than the preset risk threshold, the backup path plan is triggered.

[0172] The global strategy generation module 004 uses the ERP system, IoT sensors, logistics sensors and warehouse RFID system in the supply chain as federated learning clients. The warehouse RFID system retrains the ST-GCN model based on the equipment status data to extract the corresponding production capacity prediction error. The logistics sensor optimizes the path model according to the path planning error and predicts the probability of carbon emissions exceeding the standard. The warehouse RFID system updates the Bayesian network through the inventory topology map to evaluate the corresponding replenishment response time; an adaptive federated optimization algorithm is used to aggregate the parameters of each client and dynamically weight the weights of the nodes related to high-priority orders to generate the corresponding global strategy. The production equipment of high-priority orders has a higher weight.

[0173] Update module 005 sends the global policy to each client. After decryption, the client dynamically adjusts the local device configuration, which includes the sensor sampling rate and environmental detection accuracy. After execution by each client, the execution effect data corresponding to each client is obtained and uploaded. If an anomaly is detected, the global policy is triggered to be regenerated, which is used to incrementally update the global model through federated learning.

[0174] An embodiment of the present application further discloses a supply chain-oriented intelligent order management system, comprising a processor in which a program of any one of the above-mentioned supply chain-oriented intelligent order management methods is running.

[0175] An embodiment of the present application further discloses a storage medium storing a program of any one of the above-mentioned supply chain-oriented intelligent order management methods.

[0176] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A supply chain-oriented intelligent order management method, characterized in that: include: Extract the corresponding order demand flow through the ERP system. The order demand flow includes order quantity, priority, and delivery time constraints. The corresponding production equipment status is collected through IoT sensors. The production equipment status includes equipment vibration spectrum, temperature curve, and energy consumption fluctuation. The corresponding logistics dynamic information is obtained through logistics sensors. The logistics dynamic information includes transportation path coordinates, vehicle speed, and ambient temperature and humidity. The inventory topology map generated by the warehouse RFID system is obtained. The inventory topology map includes the three-dimensional product location and inventory distribution generated by the warehouse RFID system. The order demand flow, the production equipment status, the logistics dynamic information, and the inventory topology map are aligned through a spatiotemporal synchronization protocol to obtain the corresponding multimodal data. Based on the order demand flow and the production equipment status, a pre-set spatiotemporal graph convolutional network is used to analyze the correlation between the order and the equipment status to obtain a corresponding capacity allocation plan; Calculate the multi-objective optimal path based on the logistics dynamic information and obtain the corresponding path planning solution, where the multi-objective optimal path includes cost, timeliness, and carbon emissions; construct a Bayesian network based on the inventory topology map to predict the out-of-stock risk in the next 72 hours and generate corresponding replenishment suggestions; Based on the capacity allocation plan, it is determined whether there is a high-priority order to be received. When a high-priority order is received, a new task is inserted into the capacity allocation plan using a lossless order insertion algorithm. Based on the new task, the affected equipment process chain is determined and adjusted. Based on the capacity allocation plan, it is determined whether there are multiple orders competing for the same resource. If so, a fuzzy logic arbitrator is called to dynamically allocate resources, generate conflict resolution instructions, and execute them. When the risk value corresponding to the path planning solution is greater than the preset risk threshold, the backup path solution is triggered; The ERP system, IoT sensors, logistics sensors, and warehouse RFID systems in the supply chain are used as federated learning clients. The warehouse RFID system retrains the ST-GCN model based on device status data to extract the corresponding capacity prediction error. The logistics sensors optimize the path model based on the path planning error to predict the probability of carbon emissions exceeding the standard. The warehouse RFID system updates the Bayesian network through the inventory topology map to evaluate the corresponding replenishment response time. Use an adaptive federated optimization algorithm to aggregate client parameters, dynamically weight the nodes related to high-priority orders, and generate a corresponding global strategy. Production equipment for high-priority orders is given a higher weight. The global policy is sent to each client. After decryption, the client dynamically adjusts the local device configuration, which includes the sensor sampling rate and environmental detection accuracy. After execution, each client obtains the execution effect data corresponding to each client and uploads the execution effect data. If an anomaly is detected, the global policy is regenerated, and the global model is incrementally updated through federated learning.

2. The supply chain-oriented intelligent order management method according to claim 1, characterized in that: During the analysis of the spatiotemporal graph convolutional network, the method also includes: The production equipment status is mapped into a spatiotemporal graph structure, where nodes represent production equipment and edge weights are dynamically calculated based on the process dependencies between equipment and the real-time load difference. The short-term fluctuation patterns and long-term trend characteristics of production equipment status are extracted through multi-level spatiotemporal convolution kernels, and dynamic capacity allocation weights are generated based on order priorities. When a device's health is detected to be degraded, the allocation weight of its associated orders is automatically reduced, and the backup device activation agreement is triggered.

3. The supply chain-oriented intelligent order management method according to claim 1, characterized in that: In the process of dynamically allocating resources by calling the fuzzy logic arbitrator, the three-dimensional evaluation indicators of resource conflicts include order priority, equipment health, and delivery time margin. The method also includes: Build a fuzzy rule base. When multiple orders compete for the same equipment, prioritize orders with a priority greater than 0.8 and equipment health greater than 0.

7. If the priority difference of the conflicting orders is less than 0.2, the carbon emission impact factors corresponding to the conflicting orders are compared, and the option with the lowest carbon emission increment is selected; When generating a conflict resolution instruction, a spare resource window is reserved synchronously.

4. The supply chain-oriented intelligent order management method according to claim 1, characterized in that: During the risk value calculation process of the path planning solution, the method further includes: Build corresponding risk probability models based on real-time traffic flow, weather warning data, and historical accident statistics in logistics dynamic information; Obtain the congestion index and weather disaster warning level corresponding to the path. If the congestion index of the path is greater than 0.6 or the weather disaster warning level is ≥ level 3, it is marked as a high-risk path. When obtaining the corresponding path planning solution, the corresponding backup path solution is generated synchronously. The backup path solution pre-generates multiple candidate paths and dynamically switches the candidate paths according to the real-time risk value, and the switching response time is less than 15 seconds.

5. The supply chain-oriented intelligent order management method according to claim 1, characterized in that: In the process of detecting anomalies and regenerating the global strategy, the method further includes: Determine the corresponding abnormality criteria, which include order delivery delay rate > 10% or carbon emissions exceeding the standard > 15%. When an abnormality is detected, retrieve the local data of each client and locate the root device or path of the abnormality based on the local data. When incrementally updating the global model, retain the historical optimal version of the decision branch.

6. The supply chain-oriented intelligent order management method according to claim 1, characterized in that: During the process of dynamically adjusting the local device configuration, the method further includes: When federal policies indicate high-risk areas, the GPS sampling rate of logistics sensors is increased from 1Hz to 10Hz; the vibration sensors of production equipment start millisecond-level high-frequency acquisition mode when they detect spectrum anomalies; the scanning interval of the warehouse RFID system is adjusted according to the inventory turnover rate to ≤2 minutes when the turnover rate is greater than 5 times / day.

7. The supply chain-oriented intelligent order management method according to claim 1, characterized in that: Adaptive federated optimization algorithm, the method also includes: The aggregation weight is dynamically adjusted according to the client data quality, and the weight of nodes with data integrity greater than 90% is increased by 30%; the client parameters associated with high-priority orders are encrypted and isolated, and an adversarial verification mechanism is introduced to detect abnormal deviations in the parameters uploaded by the client. If the deviation is greater than the deviation threshold, the weight will be automatically reduced.

8. An intelligent order management system for supply chain, characterized by: include: A multimodal data acquisition module extracts the corresponding order demand flow through the ERP system. The order demand flow includes order quantity, priority, and delivery time constraints. The corresponding production equipment status is collected through IoT sensors. The production equipment status includes equipment vibration spectrum, temperature curve, and energy consumption fluctuations. The corresponding logistics dynamic information is obtained through logistics sensors. The logistics dynamic information includes transportation path coordinates, vehicle speed, and ambient temperature and humidity. The inventory topology map generated by the warehouse RFID system is obtained. The inventory topology map includes the three-dimensional product location and inventory distribution generated by the warehouse RFID system. The order demand flow, the production equipment status, the logistics dynamic information, and the inventory topology map are aligned through a spatiotemporal synchronization protocol to obtain the corresponding multimodal data. A capacity allocation plan acquisition module, which analyzes the correlation between orders and equipment status based on the order demand flow and the production equipment status through a preset spatiotemporal graph convolutional network to obtain a corresponding capacity allocation plan; Calculate a multi-objective optimal path based on the logistics dynamic information and obtain a corresponding path planning solution, where the multi-objective optimal path includes cost, timeliness, and carbon emissions; Based on the inventory topology map, a Bayesian network is constructed to predict the risk of out-of-stock in the next 72 hours and generate corresponding replenishment recommendations; A judgment module is used to determine whether a high-priority order is received based on the capacity allocation plan. When a high-priority order is received, a new task is inserted into the capacity allocation plan using a lossless order insertion algorithm. The affected equipment process chain is determined based on the new task and adjusted accordingly. The capacity allocation plan is also used to determine whether multiple orders compete for the same resource. If so, a fuzzy logic arbitrator is called to dynamically allocate resources, generate conflict resolution instructions, and execute them. When the risk value corresponding to the path planning solution is greater than the preset risk threshold, the backup path solution is triggered; The global strategy generation module uses the ERP system, IoT sensors, logistics sensors, and warehouse RFID systems in the supply chain as federated learning clients. The warehouse RFID system retrains the ST-GCN model based on device status data to extract the corresponding capacity forecast error. The logistics sensors optimize the path model based on the path planning error and predict the probability of carbon emissions exceeding the standard. The warehouse RFID system updates the Bayesian network based on the inventory topology map to evaluate the corresponding replenishment response time. An adaptive federated optimization algorithm is used to aggregate the parameters of each client and dynamically weight the weights of nodes related to high-priority orders to generate corresponding global strategies. Production equipment for high-priority orders is given a higher weight. The update module distributes the global policy to each client. After decryption, the client dynamically adjusts the local device configuration, including sensor sampling rate and environmental detection accuracy. After execution on each client, the execution effect data corresponding to each client is obtained and uploaded. If an anomaly is detected, the global strategy is regenerated to incrementally update the global model through federated learning.

9. An intelligent order management system for supply chain, characterized by: The method comprises a processor in which a program of the supply chain-oriented intelligent order management method according to any one of claims 1 to 7 is run.

10. A storage medium, characterized in that: A program storing the supply chain-oriented intelligent order management method according to any one of claims 1 to 7 is stored.

Images