Supply chain collaborative material state management system based on digital twinning

Abstract

The present application relates to the technical field of digital twin supply chain management, and discloses a supply chain collaborative material state management system based on digital twin, which comprises a physical layer module, a fusion layer module, a mapping layer module and a decision layer module.The physical layer module is used for collecting material state uploading and fusing the layers; the fusion layer module is used for mapping and conflict resolution of heterogeneous data and sending the mapping layer after conflict resolution; the mapping layer module is used for constructing digital twin to detect abnormalities, sending abnormalities and baselines to the decision layer; the decision layer module is used for calculating the demand posterior output fluctuation mutation probability after multi-temperature copy nested sampling, feeding back the wolf fitness, and ranking the wolves according to the fitness to divide them into the front half and the back half, using different strategies, mapping the front 10% scheme parameters to the nested sampling active point, exchanging the temperature collaborative copy and the wolf group every 10 times, and outputting the highest fitness scheme after iteration termination; and the physical layer module is used for receiving scheme execution parameter issuing and automatic equipment.The present application overcomes the robustness deficiency of traditional deterministic decision in a dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of digital twin supply chain management technology, and in particular to a supply chain collaborative material status management system based on digital twins. Background Technology

[0002] Supply chain management is undergoing a profound transformation from informatization to intelligentization. Digital twin technology, by constructing a virtual mirror of a physical system, can accurately reflect the system's behavior, performance, and status, and has already demonstrated its application value in areas such as warehouse management and logistics scheduling. Currently, research is attempting to apply digital twin technology to supply chain management, using real-time data to drive the synchronization of the virtual model with the physical system, achieving visualized monitoring of the supply chain. Some solutions also introduce IoT sensor networks to track the location, status, and environmental conditions of goods, or utilize digital twins for sandbox simulations of emergencies.

[0003] However, existing supply chain digital twin systems still have the following shortcomings: First, the integration of multi-source heterogeneous data is difficult. The supply chain involves multiple independent business entities, and the data formats of their respective ERP, WMS, TMS and other information systems vary, their update frequencies differ, and they lack a unified semantic standard. Existing solutions mostly rely on manual adaptation or fixed rule mapping, which is difficult to cope with the dynamic changes in data sources, resulting in an incomplete data foundation for the digital twin model.

[0004] Secondly, there is a lack of proactive correction capabilities for abnormal situations. Most existing systems remain at the level of "visual monitoring," requiring manual judgment of handling plans after detecting abnormalities in materials. There is a lack of technical means to automatically simulate and evaluate correction plans in a digital twin environment, resulting in low response efficiency.

[0005] Third, supply chain decision-making lacks the quantification of uncertainty. Existing digital twin systems mostly output deterministic results for material demand forecasting and risk assessment, failing to quantify the uncertainty of forecasts and making it difficult to introduce probabilistic risk constraints in the decision-making process, resulting in insufficient robustness of corrective measures in dynamic environments.

[0006] To address the aforementioned issues, this invention proposes a supply chain collaborative material status management system based on digital twins. Summary of the Invention

[0007] One objective of this invention is to propose a supply chain collaborative material status management system based on digital twins. This invention fully integrates replica exchange nested sampling, coyote optimization algorithm, dynamic ontology semantic mapping, sliding window conflict resolution, and temperature coordination mechanism. It describes in detail the entire process of supply chain material status management from multi-source heterogeneous data fusion, material anomaly detection, demand probability prediction, multi-objective search for correction schemes to automated execution and distribution. It has the advantages of high prediction accuracy, strong scheme robustness, fast resource scheduling, and adaptability to dynamic supply chain environments.

[0008] The supply chain collaborative material status management system based on digital twin according to an embodiment of the present invention includes: The physical layer module collects material status data and uploads it to the fusion layer module. The fusion layer module performs semantic mapping and conflict resolution on heterogeneous data and sends it to the mapping layer module. The mapping layer module builds a digital twin model, detects material anomalies, and sends anomaly and baseline data to the decision layer module. The decision layer module runs the first and second processes. The first process uses nested sampling with multiple temperature replicas. The replicas generate sample exchange likelihood values ​​using MCMC, calculate the posterior output of demand fluctuation and mutation probability, and feed it back to the fitness of the second process. Fitness is weighted by time, cost, and risk. The second process maintains the population using the coyote algorithm. The coyote corresponds to candidate solutions including allocation, path, and safety stock. The strategy is selected based on fitness ranking. The first half uses the first strategy based on the alpha wolf and cultural trends, and the second half uses the second strategy based on difference vectors. The parameters of the top 10% of coyote solutions are mapped to the initial live points of the first process, replacing the live points with the lowest fitness. The decision-making layer module is also used to coordinate the first and second processes according to temperature. The high-temperature replica group maps the exploration group using the second strategy, while the low-temperature replica group maps the development group using the first strategy. Every 10 swaps, the top 5% of coyotes are exchanged. The decision-making layer module is also used to terminate the second process when it reaches the upper limit or the change rate is low, and output the candidate solution with the highest fitness. The physical layer module is also used to receive the execution parameters of the candidate solution and send them to the automated execution device.

[0009] Optionally, the collection of material status data includes the following steps: IoT sensor nodes are deployed at various nodes in the supply chain to collect material identification, spatial coordinates, environmental parameters, and operational status at a preset sampling frequency. The collected multi-source data is timestamped and formatted, and the formatted data is uploaded to the fusion layer module.

[0010] Optionally, the semantic mapping and conflict resolution of heterogeneous data performed by the fusion layer module specifically includes: Construct a dynamic ontology library to define semantic models of core entities across the entire supply chain; receive heterogeneous data streams from multiple enterprises, perform field-level semantic mapping on each data field based on the dynamic ontology library, identify the semantic meaning of each field and map it to a unified semantic model; construct an industry knowledge graph based on the mapping results to establish semantic relationships between entities; inject the data after field semantic mapping into a distributed message queue for parallel processing, create dynamic cache partitions according to material object tags and embed timestamp indexes; use a sliding window mechanism to scan the real-time data stream, i.e., the timestamped data within the partition, perform multi-strategy resolution on conflicting data detected within the window, and send the resolved data to the mapping layer module after resolution.

[0011] Optionally, the construction of the digital twin model by the mapping layer module specifically includes: Based on the decomposed data, a 3D geometric model of the physical entity of the supply chain is constructed. The 3D geometric model includes the 3D structure of the warehouse, the layout of the storage locations, the transportation routes, and the 3D model of the materials. The material status data is mapped to the corresponding spatial location in the 3D geometric model in real time, forming a digital twin that is synchronized between the virtual and physical worlds. The real-time synchronization engine receives standardized data streams and drives the continuous updating of the status parameters of the digital twin. It detects whether the material status is abnormal. When the material inventory is lower than the safety threshold or stays in an unexpected location for too long, it is determined to be an abnormal state. The abnormal state data and the normal state baseline data are synchronously sent to the decision-making layer module.

[0012] Optionally, the first process performs the following operations: Initialize several temperature replicas, each temperature replica corresponding to an inverse temperature parameter. Each temperature replica independently initializes a set of live points, and each live point in the set contains a set of parameter vectors for the demand forecasting model. Each temperature replica internally runs a Markov chain Monte Carlo sampler, using the current live point set as the initial state, and generates new samples in the likelihood-constrained prior distribution. The new samples are a set of candidate parameters for the demand forecasting model. Calculate the likelihood value of each new sample, compare this likelihood value with the lowest likelihood value in the current live point set, and if the new sample's likelihood value is higher than the lowest likelihood value, replace the old sample corresponding to the lowest likelihood value with the new sample. Perform a replica exchange operation between different temperature replicas every fixed number of iterations. The exchange operation involves randomly selecting two temperature replicas, calculating the exponential function value of the difference between the current sample likelihood values ​​of the two replicas, and using this exponential function value as the exchange probability to determine whether to exchange all samples of the two replicas. Each temperature replica re-runs the Markov chain Monte Carlo sampler on the swapped sample set, continuing to generate new samples and update the live point set until the live point set of each temperature replica converges. After each temperature replica converges, the posterior predicted distribution of demand is calculated based on the live point set of each replica and the corresponding inverse temperature parameter, outputting the demand fluctuation range and the probability of demand mutation. The demand fluctuation range and the probability of demand mutation are used as risk constraint factors and passed to the fitness function of the second process.

[0013] Optionally, the second process performs the following operations: Initialize several coyote packs, each pack containing a fixed number of coyote individuals. Each coyote individual's position vector corresponds to a candidate correction scheme, which includes inventory allocation parameters, transportation route replanning parameters, and safety stock adjustment parameters; Obtain the risk constraint factor passed from the first process, and define the fitness function as the weighted sum of the correction time, correction cost, and risk constraint factor of the candidate correction scheme; Calculate the fitness value of each coyote individual, sort the coyote individuals in each pack from high to low fitness value, mark the coyote individual with the highest fitness value as the pack alpha, and calculate the arithmetic mean of all coyote position vectors in the pack as the cultural trend vector; Divide the coyote individuals in each pack into a first half and a second half according to their fitness values. The first half of the coyote individuals adopts the first growth strategy, and the second half of the coyote individuals adopts the second growth strategy; The first growth strategy is to generate a first candidate position vector guided by the pack alpha's position and cultural trend vector, and retain the candidate position vector with the higher fitness value than the original position vector to the next generation; The second growth strategy is to use the difference vector between any two different coyotes within the group as the perturbation direction to generate a second candidate position vector. The candidate position vector with the higher fitness value among the original position vector and the second candidate position vector is retained for the next generation. After each iteration, the top 10% of coyotes with the highest fitness values ​​are selected from all groups, and the position vectors of the top 10% of coyotes with the highest fitness values ​​are extracted and mapped to the initial values ​​of the demand prediction model parameter vector. The mapped parameter vector is inserted as a new live point into the initial live point set of the first process nested sampling, and the live points with the lowest fitness values ​​in the live point set are eliminated.

[0014] Optionally, the specific method of temperature-based coordination between the first and second processes is as follows: Each temperature replica in the first process is divided into a high-temperature replica group and a low-temperature replica group based on the inverse temperature parameter value; each coyote group in the second process is divided into an exploratory group and a development group based on the average fitness value within the group, with the average fitness value of the exploratory group being lower than that of the development group; The high-temperature replica group is mapped to the exploratory group, and the low-temperature replica group is mapped to the development group. The mapping relationship is as follows: the current sample set of each temperature replica in the high-temperature replica group serves as the probability constraint boundary for the position vector of each coyote individual in the exploratory group; the current sample set of each temperature replica in the low-temperature replica group serves as the probability constraint boundary for the position vector of each coyote individual in the development group; During the growth process, the coyote individuals in the exploratory group adopt the second strategy, and the generated candidate correction scheme parameter vector is fed back to the Markov chain Monte Carlo sampler in the high-temperature replica group, serving as the initial position for this sampler to generate new samples in the likelihood-constrained prior. In the developmental group, individual coyotes adopt the first strategy during their growth process. The parameter vector of the generated candidate correction scheme is fed back to the Markov chain Monte Carlo sampler in the low-temperature replica group, which serves as the initial position for this sampler to generate new samples in the likelihood-constrained prior.

[0015] Optionally, the method for exchanging coyotes between the high-temperature replica group and the low-temperature replica group is as follows: After every 10 iterations, select the top 5% of coyotes with the highest fitness values ​​from the exploratory group and simultaneously select the top 5% of coyotes with the highest fitness values ​​from the development group; copy the position vectors of the top 5% of coyotes selected from the exploratory group and replace the position vectors of the bottom 5% of coyotes with the highest fitness values ​​in the development group, and copy the position vectors of the top 5% of coyotes selected from the development group and replace the position vectors of the bottom 5% of coyotes with the highest fitness values ​​in the exploratory group; the exchanged coyotes inherit the current cultural trend vector of the target group as the initial value of their own cultural trend vector; after the exchange, the group to which the exchanged coyotes belong recalculates the position and cultural trend vector of the alpha wolf, and uses the updated alpha wolf position and cultural trend vector to guide the growth of coyotes in the next iteration cycle.

[0016] Optionally, the decision-making module outputs the candidate solution with the highest fitness, specifically including: After each iteration of the second process, the highest fitness value in the current iteration is recorded and calculated with the highest fitness value of the previous iteration to obtain the fitness change rate. When the current iteration of the second process reaches the preset maximum number of iterations, the decision layer module terminates the second process. When the average fitness change rate of 5 consecutive iterations is lower than a preset threshold, the decision layer module terminates the second process. After the second process terminates, the decision layer module traverses all coyotes in all coyote populations and selects the coyote with the highest fitness value. The values ​​of inventory allocation parameters, transportation route replanning parameters, and safety stock adjustment parameters are extracted from the location vector of this coyote and combined into a complete set of execution parameters for the candidate correction scheme. The decision layer module sends this complete set of execution parameters to the physical layer module.

[0017] Optionally, the execution parameters for the candidate reception scheme of the physical layer module specifically include: The physical layer module receives inventory transfer parameters, transportation route replanning parameters, and safety stock adjustment parameters; it verifies whether the source and destination warehouse identifiers in the inventory transfer parameters exist in the registered warehouse list, whether the path nodes in the route replanning parameters are connected, and whether the safety stock adjustment parameters are within the preset range; after successful verification, it converts the inventory transfer parameters into automated guided vehicle (AGV) task instructions, including start point coordinates, end point coordinates, and material load; it converts the route replanning parameters into transportation vehicle navigation instructions, including the sequence of nodes passed and the estimated arrival time; and it converts the safety stock adjustment parameters into warehouse management system inventory threshold update instructions, including material codes and updated safety stock values; it then sends the task instructions and navigation instructions to the automated execution equipment via 5G, and sends the inventory threshold update instructions to the warehouse management system via industrial Ethernet; and it receives execution result receipts from the automated execution equipment and the warehouse management system.

[0018] The beneficial effects of this invention are: First, this invention automatically establishes unified semantic relationships between heterogeneous data from multiple enterprises in the supply chain through a dynamic ontology library and field-level semantic mapping. Combined with sliding window multi-strategy conflict resolution, it eliminates the cost and error of manual adaptation, providing a complete and consistent data foundation for digital twin models.

[0019] Second, this invention introduces nested sampling with replica exchange into supply chain demand forecasting. By parallel sampling of multiple temperature replicas and sample exchange between replicas, it can accurately calculate the demand fluctuation range and mutation probability in complex multimodal posterior distributions, thereby quantifying the uncertainty of forecasting and providing probabilistic risk constraint boundaries for decision-making.

[0020] Third, this invention utilizes a dual-strategy adaptive growth mechanism in the coyote optimization algorithm, enabling coyotes with higher fitness to adopt a local refinement strategy guided by the alpha wolf and cultural trends, while coyotes with lower fitness adopt a global exploratory strategy guided by inter-individual difference vectors. This balances search accuracy and search breadth in the modified scheme space.

[0021] Fourth, this invention, through the bidirectional feedback and temperature coordination mechanism between the first and second processes, enables the probabilistic inference process to provide risk constraint factors to the population evolution process, and the population evolution process to feed back high-fitness scheme parameters to the probabilistic inference process as initial live points. The high-temperature replica group and the low-temperature replica group achieve cross-temperature level information migration through coyote exchange, deeply coupling demand uncertainty prediction with the search for correction schemes, so that the output candidate correction schemes have stronger robustness and adaptability in the dynamic supply chain environment. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the modules of the supply chain collaborative material status management system based on digital twin proposed in this invention; Figure 2 This is a schematic diagram of the copy exchange nested sampling of the supply chain collaborative material status management system based on digital twins proposed in this invention.

[0023] Figure 3 This is a schematic diagram illustrating the dual-strategy growth of the Coyote Algorithm in the supply chain collaborative material status management system based on digital twins proposed in this invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0025] refer to Figures 1-3 A digital twin-based supply chain collaborative material status management system includes: The physical layer module collects material status data and uploads it to the fusion layer module. The fusion layer module performs semantic mapping and conflict resolution on heterogeneous data and sends it to the mapping layer module. The mapping layer module builds a digital twin model, detects material anomalies, and sends anomaly and baseline data to the decision layer module. The decision layer module runs the first and second processes. The first process uses nested sampling with multiple temperature replicas. The replicas generate sample exchange likelihood values ​​using MCMC, calculate the posterior output of demand fluctuation and mutation probability, and feed it back to the fitness of the second process. Fitness is weighted by time, cost, and risk. The second process maintains the population using the coyote algorithm. The coyote corresponds to candidate solutions including allocation, path, and safety stock. The strategy is selected based on fitness ranking. The first half uses the first strategy based on the alpha wolf and cultural trends, and the second half uses the second strategy based on difference vectors. The parameters of the top 10% of coyote solutions are mapped to the initial live points of the first process, replacing the live points with the lowest fitness. The decision-making layer module is also used to coordinate the first and second processes according to temperature. The high-temperature replica group maps the exploration group using the second strategy, while the low-temperature replica group maps the development group using the first strategy. Every 10 swaps, the top 5% of coyotes are exchanged. The decision-making layer module is also used to terminate the second process when it reaches the upper limit or the change rate is low, and output the candidate solution with the highest fitness. The physical layer module is also used to receive the execution parameters of the candidate solution and send them to the automated execution device.

[0026] In this embodiment, collecting material status data includes the following steps: IoT sensor nodes are deployed at various nodes in the supply chain to collect material identification, spatial coordinates, environmental parameters, and operational status at a preset sampling frequency. The collected multi-source data is timestamped and formatted, and the formatted data is uploaded to the fusion layer module.

[0027] Specifically, the identification, spatial coordinates, environmental parameters, and operational status include: The identification features include material code, batch number, serial number, category name, specifications, production date, and supplier code; spatial coordinates include latitude and longitude coordinates, warehouse area number, shelf number, shelf number, and storage location coordinates; environmental parameters include ambient temperature, ambient humidity, vibration amplitude, light intensity, and air pressure; and operational status includes pending warehousing, in transit, already shelved, picking, already shipped, and being returned.

[0028] In this embodiment, the fusion layer module performs semantic mapping and conflict resolution on heterogeneous data, specifically including: Construct a dynamic ontology library to define semantic models of core entities across the entire supply chain; receive heterogeneous data streams from multiple enterprises, perform field-level semantic mapping on each data field based on the dynamic ontology library, identify the semantic meaning of each field and map it to a unified semantic model; construct an industry knowledge graph based on the mapping results to establish semantic relationships between entities; inject the data after field semantic mapping into a distributed message queue for parallel processing, create dynamic cache partitions according to material object tags and embed timestamp indexes; use a sliding window mechanism to scan the real-time data stream, i.e., the timestamped data within the partition, perform multi-strategy resolution on conflicting data detected within the window, and send the resolved data to the mapping layer module after resolution.

[0029] Specifically, constructing a dynamic ontology library and defining semantic models for core entities across the entire supply chain includes: Identify the core entity types, including material, organization, facility, order, transportation, personnel, and document entities.

[0030] Define an attribute set for each type of entity. The attributes of material entities include code, name, specifications, unit of measurement, classification and supplier code; the attributes of organization entities include name, type, credit code and contact information; the attributes of facility entities include name, type, geographical location and storage capacity; the attributes of order entities include number, type, order time, delivery deadline and status; and the attributes of transportation entities include tool number, mode of transportation, route information and capacity status. Establish semantic relationships between entities, including supply, storage, transportation, affiliation, and dependency relationships; The ontology library is organized using a layered architecture, including a top-level ontology layer and a domain ontology layer. The top-level ontology defines a common framework, while the domain ontology defines specific concepts in subdomains. A standard supply chain dictionary is pre-entered, specifying the standard names and data types of all entities, attributes, and relationships. A separate field mapping rule table is configured for each type of data source, defining the correspondence between external fields and the standard dictionary. Conflict resolution rules are configured to clarify the adjudication criteria when multiple data sources describe the same material status inconsistently. Rule priorities are set, specifying the execution order when multiple rules are triggered simultaneously. A rule version management mechanism is established to retain historical versions and automatically apply them to subsequent incoming data after new rules take effect. Establish an update mechanism so that the ontology library automatically expands or corrects the corresponding definitions when a new data source is connected or a field changes.

[0031] Specifically, performing field-level semantic mapping for each data field based on the dynamic ontology library includes: Identify the data source type of each field in the input data stream. Data source types include ERP systems, WMS systems, TMS systems, and IoT acquisition devices. For each field, extract the field name, field value format, field value range, and relationships between fields to form the semantic feature vector of that field; The semantic feature vector is matched with the set of class attributes defined in the dynamic ontology library to calculate the semantic similarity score between the field and each defined attribute. When the highest semantic similarity score exceeds a preset threshold, the field is directly mapped to the corresponding class attribute, and the mapping relationship and similarity score are recorded. When the highest semantic similarity score is lower than the preset threshold, the field is determined to be a new field not defined in the dynamic ontology library. Based on the semantic feature vector of the field, a new class attribute definition is created in the dynamic ontology library and a mapping relationship is established. Perform format conversion and unit unification conversion on the mapped field values. The field values ​​include numerical units, date formats, and encoding rules. Output the fields whose field names and values ​​have been converted as standardized fields for use in subsequent steps; When there is no perfectly matching mapping rule in the rule table, the preset word similarity algorithm is called to calculate the similarity between the field name and each standard name in the standard dictionary, and the field is mapped to the standard name with the highest similarity and a mapping record is generated for review.

[0032] Specifically, injecting the data after field semantic mapping into a distributed message queue for parallel processing includes: After the field semantic mapping is completed, the standardized data output is continuously written to the distributed message queue system in chronological order.

[0033] Standardized data is distributed to different message queue partitions based on the material object label, and data for the same material object enters the same partition.

[0034] Data for each partition is appended sequentially according to the time order in which it was injected into the message queue.

[0035] A parallel consumption model is adopted, in which multiple consumer instances simultaneously pull data from different queue partitions for subsequent processing.

[0036] For data on the same material object, a strategy of sequential consumption within a partition and parallel consumption between partitions is adopted to ensure the temporal integrity of the same material data.

[0037] When the amount of data accumulated in a certain partition exceeds a preset threshold, the number of consumer instances in that partition is dynamically increased to speed up processing.

[0038] When a partition is found to be continuously idle, release some consumer instance resources for that partition.

[0039] After each parallel processing node completes its processing, the results are aggregated into a unified output channel.

[0040] Specifically, the multi-strategy resolution of conflict data detected within the window includes: Identify conflicting data records within the sliding window. Conflict types include the same material having two or more different status values ​​at the same time, the same material having inconsistent inventory quantities in different data sources, and the same material having mismatched location information in different data sources.

[0041] Extract information from three dimensions for each data record in the window: timestamp, data source identifier, and field value.

[0042] Invoke the corresponding resolution strategy according to the conflict type: When conflicting data have different timestamps, a timestamp priority strategy is adopted, retaining the data record with the latest timestamp and discarding the data record with the earlier timestamp.

[0043] When conflicting data have the same timestamp or the timestamp difference is within a preset error range, obtain the historical accuracy score of each data source and retain the data record corresponding to the data source with the highest historical accuracy score.

[0044] When neither the timestamp priority strategy nor the confidence priority strategy can determine the conflict, a preset business rule set is invoked to adjudicate the conflicting data. The business rule set includes the valid value range of each field and the constraints between fields.

[0045] The data records that are retained after resolution are marked as valid data, the data records that are discarded or overwritten are marked as conflict-resolved data, and the processing results are written to the log.

[0046] The valid data retained after the digestion is completed will be output to a unified data channel.

[0047] In this embodiment, the construction of the digital twin model by the mapping layer module specifically includes: Based on the decomposed data, a 3D geometric model of the physical entity of the supply chain is constructed. The 3D geometric model includes the 3D structure of the warehouse, the layout of the storage locations, the transportation routes, and the 3D model of the materials. The material status data is mapped to the corresponding spatial location in the 3D geometric model in real time, forming a digital twin that is synchronized between the virtual and physical worlds. The real-time synchronization engine receives standardized data streams and drives the continuous updating of the status parameters of the digital twin. It detects whether the material status is abnormal. When the material inventory is lower than the safety threshold or stays in an unexpected location for too long, it is determined to be an abnormal state. The abnormal state data and the normal state baseline data are synchronously sent to the decision-making layer module.

[0048] The 3D geometric model includes the warehouse's 3D structure, storage location layout, transportation routes, and material 3D models, specifically: The warehouse's 3D structure is constructed based on the facility entity attribute set and spatial coordinate data in the standardized data. It draws the warehouse's outer wall outline, roof structure, ground elevation, column distribution, door and window positions, and the division of each functional area, which includes receiving area, inspection area, storage area, picking area, and shipping area. The warehouse location layout is constructed based on the warehouse location coordinates and shelf layout data in the standardized data. It generates the three-dimensional geometry of each shelf, the spatial coordinate range of each warehouse location, the number of layers and columns of warehouse locations, the aisle width between warehouse locations, and the mapping relationship between warehouse locations and corresponding material codes. Based on the warehouse's three-dimensional structure and cargo location layout, the transportation route is constructed using route information from standardized data, generating the centerline path of the warehouse's internal passageways, the connecting paths between the warehouse entrances and exits and each cargo location, the coordinate positions of each path node, and the connectivity between paths. The material 3D model is constructed based on the warehouse 3D structure and storage location layout, according to the material attribute set in the standardized data, generating a 3D geometric model of each material, material outline dimensions, typical placement posture of the material, and the binding relationship between the material and its storage location.

[0049] Specifically, the continuous updating of the state parameters of the digital twin includes: The real-time synchronization engine receives standardized data sent by the fusion layer. The standardized data includes material identification, spatial coordinates, environmental parameters, and operational status. The current state value carried in the standardized data is compared item by item with the existing state value of the corresponding material in the digital twin, and the change in state is calculated. When the change in status exceeds the preset update trigger threshold, it is determined that the status of the material has changed effectively, and the update process of the corresponding twin object of the material is initiated. Locate the three-dimensional geometric model of the material and its spatial location node in the digital twin.

[0050] Write the new status parameters into the attribute fields of the twin object corresponding to the material. The attribute fields include spatial coordinates, temperature and humidity values, operation status, and associated storage location. For materials that have been moved, update the spatial coordinates of their 3D geometric model in the digital twin, erase the model display at the original location, and re-render at the new location; For materials whose operational status has changed, update the color indicators or status labels displayed on their 3D geometric models; For materials whose environmental parameters change, update the environmental data curve display in their attribute panel; After the update is completed, the updated material identifier, update time, and update content will be written to the data update log; When the deviation between the latest status parameters of a certain material in the digital twin and the actual status parameters of the corresponding material in the physical system exceeds a preset deviation threshold, a status deviation alarm signal is sent to the decision-making layer module.

[0051] In this embodiment, the first process performs the following operations: Initialize several temperature replicas, each corresponding to an inverse temperature parameter. Each temperature replica independently initializes a set of live points, where each live point contains a parameter vector for a demand forecasting model. Each temperature replica internally runs a Markov chain Monte Carlo sampler, using the current live point set as the initial state, and generates new samples in a likelihood-constrained prior distribution. These new samples are a set of candidate parameters for the demand forecasting model. Calculate the likelihood value of each new sample and compare it with the lowest likelihood value in the current live point set. If the new sample's likelihood value is higher than the lowest likelihood value, replace the old sample corresponding to the lowest likelihood value with the new sample. Perform a replica exchange operation between different temperature replicas every fixed number of iterations. The exchange operation involves randomly selecting two temperature replicas, calculating the exponential function value of the difference between the current sample likelihood values ​​of the two replicas, and using this exponential function value as the exchange probability to determine whether to exchange all samples from the two replicas. Each temperature replica re-runs the Markov chain Monte Carlo sampler on the swapped sample set, continuing to generate new samples and update the live point set until the live point set of each temperature replica converges. After each temperature replica converges, the posterior predicted distribution of demand is calculated based on the live point set of each replica and the corresponding inverse temperature parameter, outputting the demand fluctuation range and the probability of demand mutation. The demand fluctuation range and the probability of demand mutation are used as risk constraint factors and passed to the fitness function of the second process.

[0052] Specifically, the reverse temperature parameter is: The inverse temperature parameter is the core parameter that controls the "aggressiveness" of sampling each temperature replica in the nested sampling of replica exchange. It is commonly represented by the symbol β, and its value range is 0 < β ≤ 1. Specifically, when the β value is small (e.g., β=0.2), the likelihood function of the replica is "softened" or "flattened"—that is, the probability difference between the high likelihood region and the low likelihood region is reduced, the sampler is more likely to accept low likelihood samples, the replica has a larger range of activity in the parameter space and a stronger ability to traverse low probability regions, which is suitable for global exploration and crossing energy barriers. When the β value is large (e.g., β=0.9), the likelihood function of the replica remains "steep" or "sharp"—that is, the sampler strictly focuses on the high likelihood region, the samples are concentrated near the peak of the posterior distribution, which is suitable for local fine sampling and high-precision inference. In this invention, the first process configures multiple temperature replicas, each of which is assigned a different inverse temperature parameter value (uniformly distributed from 0.2 to 0.9). This allows low-β replicas to be responsible for searching multiple high-probability regions in the parameter space over a large area, while high-β replicas are responsible for fine sampling within these regions. Replicas with different β values ​​exchange samples and likelihood values ​​every fixed number of iterations, enabling high-β replicas to utilize cross-regional information discovered by low-β replicas. Ultimately, all replicas share the global exploration results and sample centrally in high-probability regions.

[0053] Specifically, each temperature replica independently initializes a set of live points, including: The live point set is a discrete sample set used to approximate the posterior distribution in the replica swap nested sampling algorithm. Each live point corresponds to a parameter vector of a demand prediction model. The specific implementation method for each temperature replica to independently initialize a set of live points is as follows: Each temperature replica is assigned a fixed number of live points, N, which ranges from 50 to 500. N sets of parameter vectors are randomly sampled from the prior distribution of the demand forecasting model to serve as the N initial live points for this temperature replica; Substitute the N sets of parameter vectors into the demand forecasting model, and calculate the likelihood value for each set of parameters based on historical demand data from the supply chain. Arrange the N live points in ascending order of their likelihood values ​​to form the initial set of live points for this temperature replica, where the live point with the smallest likelihood value is marked as the current lowest likelihood live point. In subsequent iterations, after the Markov chain Monte Carlo sampler generates a new sample, it compares the likelihood value of the new sample with the lowest likelihood value in the current live point set. If the likelihood value of the new sample is higher than the lowest likelihood value, the new sample replaces the old sample corresponding to the lowest likelihood value, and the live point set is reordered and the lowest likelihood live point is updated. After multiple iterations, the set of live points gradually migrates from the prior distribution region to the high likelihood region, and finally converges to the high probability density region of the posterior distribution of demand, which is used to calculate the posterior prediction distribution of demand.

[0054] Specifically, the convergence of the set of live points for each temperature replica is as follows: Convergence of the live point set for each temperature replica refers to the state where the live point set of that replica no longer changes significantly. The criterion is: after several consecutive iterations, the change in the lowest likelihood value in the live point set gradually decreases, and when the change in the lowest likelihood value is lower than a preset threshold, the live point set of that temperature replica is considered to have converged.

[0055] Specifically, calculating the posterior predicted distribution of demand based on the liveness set of each replica and the corresponding inverse temperature parameter includes: Demand forecasting models consist of three components: a trend term, a seasonal term, and an external regression term. The trend term is calculated as follows: a trend line that changes monotonically over time is fitted based on historical demand data. The trend coefficient in the current active point is substituted into the trend line equation and multiplied by the distance between the predicted time point and the baseline point to obtain the baseline demand trend value at that time point. The calculation method for the seasonal component is as follows: historical demand data is decomposed according to the cycle length, and seasonal fluctuation components of different frequencies are extracted by Fourier series expansion. The seasonal smoothing coefficient in the current active point is substituted into the seasonal fluctuation function to calculate the phase position of the time point to be predicted in a complete seasonal cycle, and the incremental value generated by the seasonal fluctuation at that time point is output. The external regression term is calculated as follows: Select external variables that affect demand, including supplier capacity index, logistics timeliness index and promotion activity intensity, multiply each regression coefficient in the current active point by the actual observed value of the corresponding external variable at the current moment, sum all the products, and output the adjustment of demand by external factors. The final predicted value is synthesized by adding the values ​​of the trend term, seasonal term, and external regression term, plus the random error term. The random error term follows a normal distribution with a mean of zero, and the variance is estimated by the model fitting residuals. After adding them together, the final demand prediction value at that time point is output. Each live point contains a complete set of model parameters. Substituting these parameters into the above calculations and summing them, a demand prediction value is output. Substituting multiple live points in a set into the calculations results in multiple prediction values. These multiple prediction values ​​constitute a probability distribution sample of the demand at that point in time, which is used for subsequent calculations of confidence intervals and mutation probabilities.

[0056] Specifically, the fitness function passed to the second process, which uses the demand fluctuation range and the probability of demand mutation as risk constraint factors, includes: The first process packages the upper and lower bounds of the demand fluctuation range and the probability of demand mutation at each forecast time point into a risk constraint factor data structure, and sends it to the second process through inter-process communication. The second process parses the data structure and extracts the demand fluctuation range and the probability value of demand mutation within the time range involved in the candidate solutions. The fitness function uses the width of the fluctuation range as a measure of inventory shortage risk and the probability of sudden demand change as a measure of sudden stockout risk. It comprehensively calculates the overall risk penalty value and accumulates all risk penalty values ​​in scenarios with multiple materials or multiple warehouses. The final fitness value is obtained by weighted summation of three factors: correction time, correction cost, and total risk penalty value. The weighting coefficient of the total risk penalty value is preset to a fixed value or dynamically adjusted according to the current supply chain environment.

[0057] In this embodiment, the second process performs the following operations: Initialize several coyote packs, each pack containing a fixed number of coyote individuals. Each coyote individual's position vector corresponds to a candidate correction scheme, which includes inventory allocation parameters, transportation route replanning parameters, and safety stock adjustment parameters; Obtain the risk constraint factor passed from the first process, and define the fitness function as the weighted sum of the correction time, correction cost, and risk constraint factor of the candidate correction scheme; Calculate the fitness value of each coyote individual, sort the coyote individuals in each pack from high to low fitness value, mark the coyote individual with the highest fitness value as the pack alpha, and calculate the arithmetic mean of all coyote position vectors in the pack as the cultural trend vector; Divide the coyote individuals in each pack into a first half and a second half according to their fitness values. The first half of the coyote individuals adopts the first growth strategy, and the second half of the coyote individuals adopts the second growth strategy; The first growth strategy is to generate a first candidate position vector guided by the pack alpha's position and the cultural trend vector, and retain the candidate position vector with the higher fitness value than the original position vector to the next generation; The second growth strategy is to use the difference vector between any two different coyotes within the group as the perturbation direction to generate a second candidate position vector. The candidate position vector with the higher fitness value among the original position vector and the second candidate position vector is retained for the next generation. After each iteration, the top 10% of coyotes with the highest fitness values ​​are selected from all groups, and the position vectors of the top 10% of coyotes with the highest fitness values ​​are extracted and mapped to the initial values ​​of the demand prediction model parameter vector. The mapped parameter vector is inserted as a new live point into the initial live point set of the first process nested sampling, and the live points with the lowest fitness values ​​in the live point set are eliminated.

[0058] The candidate correction schemes, which include inventory transfer parameters, transportation route replanning parameters, and safety stock adjustment parameters, specifically include: The inventory transfer parameters include the warehouse identifier of the warehouse from which the transfer is made, the warehouse identifier of the warehouse from which the transfer is made, the code of the transferred material, the quantity of the transfer, the priority of the transfer, and the expected completion time. The parameters for replanning the transportation route include the current transportation task number, material code, material quantity, origin node identifier, destination node identifier, alternative route node sequence, and estimated transportation time. The safety stock adjustment parameters include the material code, the warehouse identifier, the original safety stock value, the updated safety stock value, and the effective time of the adjustment. In the inventory transfer parameters, the transfer quantity is calculated based on the difference between the upper limit of the demand fluctuation range output by the first process and the current available inventory, and the transfer priority is divided into levels according to the demand mutation probability value corresponding to the material. In the parameters for replanning transportation routes, the sequence of candidate route nodes includes at least two passing nodes and the estimated travel time between adjacent nodes. The total travel time between the starting node and the ending node is the sum of the estimated travel times between each adjacent node. In the safety stock adjustment parameters, the updated safety stock value is calculated based on the product of the upper limit of the demand fluctuation range and the normal supply cycle, and the adjustment takes effect at the moment when the candidate correction plan begins to be implemented.

[0059] The second growth strategy, which uses the difference vector between any two different coyotes within the group as the perturbation direction, specifically involves: In each iteration, for the coyote individual that needs to grow, two different coyote individuals are randomly selected from the coyote group to which the coyote belongs. Neither of the two coyote individuals includes the current coyote individual itself. Subtract the position vectors of the two selected coyotes dimension by dimension. Subtract the corresponding dimension parameter value of the second selected individual from the parameter value of each dimension of the first selected individual to obtain the difference in each dimension. Combine the differences in all dimensions to form a difference vector. The direction of the difference vector is from the second selected individual to the first selected individual, and its length is equal to the distance between the two individuals in the parameter space; Add the current coyote individual's position vector to the difference vector by a preset scaling factor to obtain a new candidate position vector; The preset proportional coefficient ranges from 0.1 to 0.9. It is set to a larger value in the early stage of iteration to increase the perturbation amplitude, and a smaller value in the later stage of iteration to reduce the perturbation amplitude. After the difference vector is added to the current coyote's position vector as a random perturbation direction, the resulting candidate position vector inherits the differential information of other individuals in the group, enabling the current coyote to move to the parameter space region where other individuals in the group are located. Compared to the first growth strategy (converging towards the alpha wolf and cultural trends), the second growth strategy uses the difference vector between two randomly selected individuals to provide unbiased random perturbations, so that the second half of the coyotes will not all gather towards the alpha wolf, but will remain dispersed in the parameter space.

[0060] In this embodiment, the specific method of temperature-based coordination between the first and second processes is as follows: Each temperature replica in the first process is divided into a high-temperature replica group and a low-temperature replica group based on the inverse temperature parameter value. Each coyote group in the second process is divided into an exploratory group and a development group based on the average fitness value within the group. The average fitness value of the exploratory group is lower than that of the development group. The high-temperature replica group is mapped to the exploratory group, and the low-temperature replica group is mapped to the development group. The mapping relationship is as follows: the current sample set of each temperature replica in the high-temperature replica group serves as the probability constraint boundary for the position vector of each coyote individual in the exploratory group, and the current sample set of each temperature replica in the low-temperature replica group serves as the probability constraint boundary for the position vector of each coyote individual in the development group. During the growth process, the coyote individuals in the exploratory group adopt the second strategy, and the generated candidate correction scheme parameter vector is fed back to the Markov chain Monte Carlo sampler in the high-temperature replica group, serving as the initial position for this sampler to generate new samples in the likelihood-constrained prior. In the developmental group, individual coyotes adopt the first strategy during their growth process. The parameter vector of the generated candidate correction scheme is fed back to the Markov chain Monte Carlo sampler in the low-temperature replica group, serving as the initial position for this sampler to generate new samples in the likelihood-constrained prior.

[0061] The distinction between high-temperature replicas and low-temperature replicas in the first process, based on the inverse temperature parameter value, is as follows: Preset an inverse temperature threshold , The value ranges from 0.4 to 0.6, and the inverse temperature parameter value in all temperature replicas is less than... The replicas are divided into a high-temperature replica group, and the inverse temperature parameter value in all temperature replicas is greater than or equal to The replicas are divided into low-temperature replica groups.

[0062] Specifically, the second strategy employed by individual coyotes in exploratory groups during their development is as follows: Two different coyotes are randomly selected from the coyote population. The position vector of the first selected coyote is subtracted from the position vector of the second selected coyote to obtain a difference vector. This difference vector is then multiplied by a preset perturbation coefficient and added to the current coyote's position vector to generate a candidate position vector.

[0063] After the candidate position vector is generated, the system checks whether each dimension parameter in the vector exceeds the probability constraint boundary defined by the high-temperature replica group sample set. If the value of a certain dimension parameter is less than the lower search limit for that dimension, then that dimension parameter value is set as the lower search limit value; if the value of a certain dimension parameter is greater than the upper search limit for that dimension, then that dimension parameter value is set as the upper search limit value. After pruning, the valid candidate position vector is obtained.

[0064] The system calculates the fitness value of the candidate correction scheme corresponding to the valid candidate position vector and compares the candidate fitness value with the fitness value of the current coyote's original position vector. If the candidate fitness value is higher than the original fitness value, the current coyote's position vector is replaced by the candidate position vector; if the candidate fitness value is not higher than the original fitness value, the current coyote's position vector remains unchanged.

[0065] All coyotes in the exploratory group uniformly perform growth operations in the manner described above, using the second strategy in each iteration without switching to the first strategy. The random selection operation in the second strategy ensures that different coyotes receive different sources of difference vectors. Different coyotes within the same group may move in different directions in each iteration, thus maintaining the dispersion of their positional distribution within the spatial range defined by the probability constraint boundary, and enabling continuous exploration of the parameter space surrounding high-probability regions.

[0066] Specifically, the first strategy employed by individual coyotes in the developmental group during their growth process is as follows: Determine the position of the alpha wolf in the coyote pack, which is the position vector of the coyote with the highest fitness value in the pack, and the cultural trend vector, which is the arithmetic mean of each dimension of the position vectors of all coyotes in the pack. The alpha wolf position vector and the cultural trend vector are weighted and superimposed to obtain the guiding direction vector. Each dimension of the guiding direction vector points to the favorable direction for fitness improvement in that dimension. Move the current coyote's position vector along the guiding direction vector by a preset step size, where the step size is a random number between 0.1 and 0.9. After the movement, the candidate position vector is obtained. After the candidate position vector is generated, the system checks whether the parameters of each dimension in the candidate position vector exceed the probability constraint boundary of the low-temperature replica group sample set, and performs pruning on the dimensions that exceed the boundary. The system calculates the fitness value corresponding to the candidate position vector. If the candidate fitness value is higher than the original fitness value, the current coyote position vector is replaced with the candidate position vector; if the candidate fitness value is not higher than the original fitness value, the current coyote position vector is kept unchanged. In the developmental group, all coyotes uniformly perform the above operations, using the first strategy in each iteration to cause the coyotes in the group to gather towards the alpha wolf's position and cultural trend, gradually converging to a high fitness area.

[0067] In this embodiment, the coyote individual exchange method between the high-temperature replica group and the low-temperature replica group is as follows: After every 10 iterations, the top 5% of coyote individuals in terms of fitness value are selected from the exploratory group, and the top 5% of coyote individuals in terms of fitness value are selected from the development group. The position vectors of the top 5% of coyote individuals selected from the exploratory group are copied and used to replace the position vectors of the bottom 5% of coyote individuals in the development group. The position vectors of the top 5% of coyote individuals selected from the development group are copied and used to replace the position vectors of the bottom 5% of coyote individuals in the exploratory group. The exchanged coyote individuals inherit the current cultural trend vector of the target group as the initial value of their own cultural trend vector. After the exchange is completed, the group to which the exchanged coyote individual belongs recalculates the position and cultural trend vector of the alpha wolf, and uses the updated alpha wolf position and cultural trend vector to guide the coyote growth in the next iteration cycle.

[0068] In this embodiment, the decision-making module outputs the candidate solution with the highest fitness, specifically including: After each iteration of the second process, the highest fitness value in the current iteration is recorded, and the difference between this value and the highest fitness value in the previous iteration is calculated to obtain the fitness change rate. When the current iteration of the second process reaches the preset maximum number of iterations, the decision layer module terminates the second process. When the average fitness change rate of five consecutive iterations is lower than a preset threshold, the decision layer module terminates the second process. After the second process terminates, the decision layer module iterates through all coyotes in all coyote populations and selects the coyote with the highest fitness value. The values ​​of inventory allocation parameters, transportation route replanning parameters, and safety stock adjustment parameters are extracted from the location vector of this coyote and combined to form a complete set of execution parameters for the candidate correction scheme. The decision layer module sends this complete set of execution parameters to the physical layer module.

[0069] In this embodiment, the specific execution parameters for the physical layer module's receiving candidate scheme include: The physical layer module receives inventory transfer parameters, transportation route replanning parameters, and safety stock adjustment parameters; it verifies whether the source and destination warehouse identifiers in the inventory transfer parameters exist in the registered warehouse list, whether the path nodes in the route replanning parameters are connected, and whether the safety stock adjustment parameters are within the preset range; after successful verification, it converts the inventory transfer parameters into automated guided vehicle (AGV) task instructions, including start point coordinates, end point coordinates, and material load; it converts the route replanning parameters into transportation vehicle navigation instructions, including the sequence of nodes passed and the estimated arrival time; and it converts the safety stock adjustment parameters into warehouse management system inventory threshold update instructions, including material codes and updated safety stock values; it then sends the task instructions and navigation instructions to the automated execution equipment via 5G, and sends the inventory threshold update instructions to the warehouse management system via industrial Ethernet; and it receives execution result receipts from the automated execution equipment and the warehouse management system.

[0070] Specifically, receiving execution result feedback from automated execution equipment and warehouse management systems includes: The physical layer module receives task execution feedback from automated execution devices via the 5G communication network and inventory threshold update feedback from the warehouse management system via the industrial Ethernet. The execution result receipt includes the instruction execution status code and a completion timestamp. The instruction execution status code is divided into two types: execution success and execution failure. When the execution status code is successful, the physical layer module marks the execution result of the instruction as completed and writes the completion timestamp to the instruction execution log; When the execution status code is execution failure, the physical layer module records the failure reason code and marks the execution result of the instruction as an execution exception status; The physical layer module retryes instructions that are in an abnormal execution state, with a maximum of 3 retries. If the number of retries exceeds 3 and the execution status code is still "execution failed," the physical layer module sends an instruction execution failure alarm message to the decision layer module. The alarm message includes the instruction content, the failure reason code, and the record of the number of retries.

[0071] Example 1: To verify the feasibility and effectiveness of this invention in a real supply chain environment, we used the raw material supply network of a large equipment manufacturing enterprise as the application object. This enterprise has more than 200 direct suppliers, four regional central warehouses and six production plants, generating more than 500,000 data entries of material flow every day. For a long time, this enterprise has faced three typical dilemmas: the ERP systems of various suppliers are incompatible with the internal WMS and TMS systems in terms of material coding and units of measurement, resulting in an average deviation rate of 8.7% between inventory ledgers and actual deliveries; production plans are frequently adjusted, but demand forecasts cannot quantify sudden fluctuations, and safety stock settings are either too high or too low; when transportation is delayed or materials are abnormal, dispatchers rely on experience to make manual adjustments, which takes an average of more than four hours and is often not the most cost-effective solution.

[0072] To address the aforementioned scenario, we have fully deployed the system described in this invention. The physical layer module collects material identification, location, temperature and humidity, and operational status every 5 minutes using UHF RFID readers and sensors, and uploads this data to the fusion layer. The fusion layer performs field-level semantic mapping on data from different supplier ERP systems, logistics provider TMS systems, and internal WMS systems based on a dynamic ontology library. It uniformly maps "Item_Code," "Material_ID," and "Prod_SKU" to "material code," and unifies "Qty," "Amount," and "Num" as standard quantity fields. Subsequently, the data is cached in batches by material tag using a distributed message queue. A sliding window scans the data in 2-second increments and resolves timestamp inconsistencies (such as conflicts between "in transit" and "signed for"), outputting standardized data. The mapping layer uses this standardized data to construct a high-fidelity digital twin covering the warehouse's three-dimensional structure, cargo location distribution, and transportation network. A real-time synchronization engine refreshes the status every 10 seconds and continuously monitors inventory levels and transit time. If a batch of materials is detected as overdue or its inventory falls below a safety threshold, the abnormal data and the baseline data for the most recent 30 days are immediately sent to the decision layer.

[0073] The decision-making level simultaneously initiates two processes. The first process configures eight temperature replicas (inverse temperature parameters ranging from 0.2 to 0.9), each replica initializing 200 live points. Each live point corresponds to a demand forecasting model parameter vector containing seasonality, trend, and regression terms. Within each replica, the MCMC sampler generates candidate parameters in the likelihood-constrained prior. Every 100 iterations, two replicas are randomly selected to calculate the exchange probability. After approximately 5000 iterations, the process converges and outputs the posterior predicted distribution of daily demand for the next two weeks (e.g., demand of 820 units on day 3, 90% confidence interval [750, 910]) and the probability of a sudden demand surge (a 34% probability of a sudden increase in demand for a certain type of steel due to increased downstream production). The second process initializes six coyote packs (30 coyotes per pack). Each coyote's location vector consists of the allocation amount, transportation route number, and safety stock multiple. After receiving the risk constraint factors from the first process, the fitness function is defined as a weighted sum of correction time, transportation costs, and risk penalties. Within the coyote group, individuals are sorted by fitness. The first half of the coyotes uses a first strategy guided by the alpha wolf's position and cultural trends for local refinement, while the second half uses a second strategy involving randomized differential vector perturbations between two companions for global exploration. Every 10 iterations, the position vectors of the top 10% of the coyotes with the best fitness are mapped to initial live points in the demand forecasting model and injected into the first process, while an equal number of live points with the lowest fitness are eliminated. Temperature coordination maps the high-temperature replica group to the exploratory coyote group (using the second strategy to expand the search), and the low-temperature replica group to the development group (using the first strategy to refine the solution). Every 10 rounds, the top 5% of the coyotes with the best fitness in the high-temperature and low-temperature groups are swapped. The process terminates after 200 iterations and if the fitness change rate is below 0.3% for 5 consecutive rounds, outputting the highest fitness solution—an emergency transfer of 280 connectors from the central warehouse to the second factory area, changing the transportation route from the regular highway to an alternative national highway to avoid congestion, and increasing the safety stock of this material from 80 to 130 units. After the physical layer module verifies the warehouse validity and path connectivity, it converts the transfer instruction into an AGV handling task (starting point A3 shelf, ending point B5 assembly line, load capacity 350 kg), path replanning into a navigation instruction (passing through 3 intermediate nodes, estimated arrival time 2 hours), and safety stock changes into WMS threshold updates. All instructions are sent to the automated guided vehicles and warehouse management system via the 5G private network and a success receipt is received. The entire process time is reduced from an average of 4.2 hours to 18 minutes.

[0074] Table 1: Comparison of Key Performance Indicators for Supply Chain Material Management under Different Management Methods

[0075] The comparative data in Table 1 are taken from the actual operation statistics of the manufacturing enterprise for three consecutive months. It can be seen that the system of this invention outperforms traditional manual scheduling and conventional digital management systems in all seven evaluation indicators. Inventory turnover rate increased from 4.2 times / year to 8.7 times / year, meaning that with the same inventory level, it can support more production output, and the capital occupation cycle is significantly shortened. The on-time order delivery rate reached 97.6%, an increase of 8.5 percentage points compared to the conventional system. This is mainly due to the prediction accuracy rate jumping from 81.2% to 94.2%, making material preparation and production scheduling more closely aligned with actual needs. Anomaly response time was further reduced from 2.1 hours in the conventional system to 0.3 hours, meaning that an effective countermeasure can be formed and implemented in an average of 18 minutes. This is because the present invention integrates solution generation and simulation deduction into the same closed loop, eliminating manual discussion and repeated trial and error. Monthly comprehensive operating costs decreased from RMB 2.123 million to RMB 1.784 million, saving nearly RMB 340,000 per month. This included a reduction of approximately RMB 120,000 in inventory holding costs, approximately RMB 150,000 in emergency allocation and overtime transportation expenses, and approximately RMB 70,000 in data governance labor costs. Inventory backlog rate decreased from 22.7% to 11.5%, while the number of stockouts and production stoppages decreased from 4.1 times per month to 1.2 times per month. This demonstrates that the invention reduced redundant inventory without sacrificing supply security, and the introduction of risk constraint factors achieved a better balance between cost and robustness in decision-making. Data consistency maintenance time was reduced from 8.5 hours per week to 1 hour, virtually eliminating the workload of manually verifying data differences across multiple systems. These data fully validate the practical effectiveness of the invention in solving core issues such as heterogeneous data fusion, uncertainty quantification, and automatic solution optimization. Especially in the high-frequency fluctuations of the manufacturing supply chain environment, the system's response speed and decision quality demonstrate significant advantages.

[0076] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A supply chain collaborative material status management system based on digital twins, characterized in that, include: The physical layer module is used to collect material status data and upload it to the fusion layer module; The fusion layer module is used to perform semantic mapping and conflict resolution on heterogeneous data and send it to the mapping layer module; The mapping layer module is used to build a digital twin model, detect material anomalies, and send anomaly and baseline data to the decision layer module. The decision-making layer module is used to run the first and second processes. The first process uses nested sampling with multiple temperature replicas. The replicas generate sample exchange likelihood values ​​through MCMC, calculate the posterior demand output volatility and mutation probability, and feed it back to the fitness of the second process. Fitness is weighted by time, cost, and risk. The second process maintains the population using the coyote algorithm. Coyotes correspond to candidate solutions including allocation, path, and safety stock. The strategy is selected based on fitness ranking. The first half uses the first strategy based on the alpha wolf and cultural trends, and the second half uses the second strategy based on difference vectors. The parameters of the top 10% of coyote solutions are mapped to the initial live points of the first process, and the live points with the lowest fitness are replaced by an equal number of live points. The decision-making layer module is also used to coordinate the first and second processes according to temperature. The high-temperature replica group is mapped to the exploration group using the second strategy, and the low-temperature replica group is mapped to the development group using the first strategy. Every 10 iterations, the top 5% of coyotes are exchanged. The decision-making layer module is also used to terminate the second process when the iteration reaches the upper limit or the rate of change is low, and output the candidate solution with the highest fitness. The physical layer module is also used to receive the execution parameters of the candidate solutions and distribute them to the automated execution equipment.

2. The supply chain collaborative material status management system based on digital twins according to claim 1, characterized in that, The collection of material status data includes the following steps: IoT sensor nodes are deployed at various nodes in the supply chain to collect material identification, spatial coordinates, environmental parameters, and operational status at a preset sampling frequency. The collected multi-source data is timestamped and formatted, and the formatted data is uploaded to the fusion layer module.

3. The supply chain collaborative material status management system based on digital twins according to claim 1, characterized in that, The fusion layer module performs semantic mapping and conflict resolution on heterogeneous data, specifically including: Construct a dynamic ontology library to define semantic models of core entities across the entire supply chain; receive heterogeneous data streams from multiple enterprises, perform field-level semantic mapping on each data field based on the dynamic ontology library, identify the semantic meaning of each field and map it to a unified semantic model; construct an industry knowledge graph based on the mapping results to establish semantic relationships between entities; inject the data after field semantic mapping into a distributed message queue for parallel processing, create dynamic cache partitions according to material object tags and embed timestamp indexes; use a sliding window mechanism to scan the real-time data stream, i.e., the timestamped data within the partition, perform multi-strategy resolution on conflicting data detected within the window, and send the resolved data to the mapping layer module after resolution.

4. The supply chain collaborative material status management system based on digital twins according to claim 1, characterized in that, The construction of the digital twin model by the mapping layer module specifically includes: Based on the decomposed data, a 3D geometric model of the physical entity of the supply chain is constructed. The 3D geometric model includes the 3D structure of the warehouse, the layout of the storage locations, the transportation routes, and the 3D model of the materials. The material status data is mapped to the corresponding spatial location in the 3D geometric model in real time, forming a digital twin that is synchronized between the virtual and physical worlds. The real-time synchronization engine receives standardized data streams and drives the continuous updating of the status parameters of the digital twin. It detects whether the material status is abnormal. When the material inventory is lower than the safety threshold or stays in an unexpected location for too long, it is determined to be an abnormal state. The abnormal state data and the normal state baseline data are synchronously sent to the decision-making layer module.

5. The supply chain collaborative material status management system based on digital twins according to claim 1, characterized in that, The first process performs the following operations: Initializes several temperature replicas, each temperature replica corresponding to an inverse temperature parameter, and each temperature replica independently initializes a set of live points, each live point in the set of live points containing a set of parameter vectors for the demand forecasting model; Each temperature replica internally runs a Markov chain Monte Carlo sampler, using the current set of live points as the initial state, and generates new samples in the likelihood-constrained prior distribution, the new samples being a set of candidate parameters for the demand forecasting model; Calculate the likelihood value of each new sample and compare it with the lowest likelihood value in the current live point set. If the new sample's likelihood value is higher than the lowest likelihood value, replace the old sample corresponding to the lowest likelihood value with the new sample. Perform a replica swap operation between different temperature replicas every fixed number of iterations. The swap operation involves randomly selecting two temperature replicas, calculating the exponential function value of the difference in the likelihood values ​​of the current samples in the two replicas, and using this exponential function value as the swap probability to determine whether to swap all samples between the two replicas. For each temperature replica, re-run the Markov chain Monte Carlo sampler on the swapped sample set to continue generating new samples and updating the live point set until the live point set of each temperature replica converges. After convergence of all temperature replicas, calculate the posterior predicted distribution of demand based on the live point set of each replica and the corresponding inverse temperature parameter, and output the demand fluctuation range and the probability of demand mutation. Use the demand fluctuation range and the probability of demand mutation as risk constraint factors and pass them to the fitness function of the second process.

6. The supply chain collaborative material status management system based on digital twins according to claim 1, characterized in that, The second process performs the following operations: Initialize several coyote packs, each pack containing a fixed number of coyote individuals, and the position vector of each coyote individual corresponds to a candidate correction scheme, which includes inventory transfer parameters, transportation route replanning parameters, and safety stock adjustment parameters; Obtain the risk constraint factor passed from the first process, and define the fitness function as the weighted sum of the correction time, correction cost, and risk constraint factor of the candidate correction scheme; calculate the fitness value of each coyote individual, sort the coyotes in each group from high to low fitness value, and mark the coyote individual with the highest fitness value as the group alpha; calculate the arithmetic mean of the position vectors of all coyotes in the group as the cultural trend vector; divide the coyotes in each group into the first half and the second half according to the fitness value, the first half of the coyotes adopt the first growth strategy, and the second half of the coyotes adopt the second growth strategy; the first growth strategy is to generate the first candidate position vector with the position of the group alpha and the cultural trend vector as the guiding direction, and retain the one with the higher fitness value between the first candidate position vector and the original position vector to the next generation; The second growth strategy is to use the difference vector between any two different coyotes within the group as the perturbation direction to generate a second candidate position vector. The candidate position vector with the higher fitness value among the original position vector and the second candidate position vector is retained for the next generation. After each iteration, the top 10% of coyotes with the highest fitness values ​​are selected from all groups, and the position vectors of the top 10% of coyotes with the highest fitness values ​​are extracted and mapped to the initial values ​​of the demand prediction model parameter vector. The mapped parameter vector is inserted as a new live point into the initial live point set of the first process nested sampling, and the live points with the lowest fitness values ​​in the live point set are eliminated.

7. The supply chain collaborative material status management system based on digital twins according to claim 1, characterized in that, The specific method of temperature-based coordination between the first and second processes is as follows: The temperature replicas of the first process are divided into high-temperature replica groups and low-temperature replica groups based on their inverse temperature parameter values. The coyote groups of the second process are divided into exploratory groups and development groups based on their average fitness values ​​within the group. The average fitness value of the exploratory groups is lower than that of the development groups. The high-temperature replica groups are mapped to the exploratory groups, and the low-temperature replica groups are mapped to the development groups. The mapping relationship is as follows: the current sample set of each temperature replica in the high-temperature replica group serves as the probability constraint boundary for the position vector of each coyote individual in the exploratory group, and the current sample set of each temperature replica in the low-temperature replica group serves as the probability constraint boundary for the position vector of each coyote individual in the development group. During the growth process, the coyote individuals in the exploratory group adopt the second strategy, and the generated candidate correction scheme parameter vector is fed back to the Markov chain Monte Carlo sampler in the high-temperature replica group, serving as the initial position for this sampler to generate new samples in the likelihood-constrained prior. In the developmental group, individual coyotes adopt the first strategy during their growth process. The parameter vector of the generated candidate correction scheme is fed back to the Markov chain Monte Carlo sampler in the low-temperature replica group, which serves as the initial position for this sampler to generate new samples in the likelihood-constrained prior.

8. The supply chain collaborative material status management system based on digital twins according to claim 1, characterized in that, The method for exchanging coyotes between the high-temperature replica group and the low-temperature replica group is as follows: After every 10 iterations, select the top 5% of coyotes with the highest fitness values ​​from both the exploratory and development groups; copy the position vectors of the top 5% of coyotes selected from the exploratory group and replace the position vectors of the bottom 5% of coyotes with the highest fitness values ​​from the development group; and copy the position vectors of the top 5% of coyotes selected from the development group and replace the position vectors of the bottom 5% of coyotes with the highest fitness values ​​from the exploratory group; the exchanged coyotes inherit the current cultural trend vector of the target group as the initial value of their own cultural trend vector. After the exchange is completed, the group to which the exchanged coyote belongs recalculates the alpha wolf position and cultural trend vector, and uses the updated alpha wolf position and cultural trend vector to guide the coyote growth in the next iteration cycle.

9. The supply chain collaborative material status management system based on digital twins according to claim 1, characterized in that, The decision-making layer module outputs the candidate solution with the highest fitness, specifically including: After each iteration of the second process, the highest fitness value in the current iteration is recorded and calculated with the highest fitness value of the previous iteration to obtain the fitness change rate. When the current iteration of the second process reaches the preset maximum number of iterations, the decision layer module terminates the second process. When the average fitness change rate of several consecutive iterations is lower than a preset threshold, the decision layer module terminates the second process. After the second process terminates, the decision layer module traverses all coyotes in all coyote populations and selects the coyote with the highest fitness value. The values ​​of inventory allocation parameters, transportation route replanning parameters, and safety stock adjustment parameters are extracted from the location vector of this coyote and combined into a complete set of execution parameters for the candidate correction scheme. The decision layer module sends this complete set of execution parameters to the physical layer module.

10. The supply chain collaborative material status management system based on digital twins according to claim 1, characterized in that, The specific execution parameters for the candidate reception scheme of the physical layer module include: The physical layer module receives inventory transfer parameters, transportation route replanning parameters, and safety stock adjustment parameters; it verifies whether the source and destination warehouse identifiers in the inventory transfer parameters exist in the registered warehouse list, whether the path nodes in the route replanning parameters are connected, and whether the safety stock adjustment parameters are within the preset range; after successful verification, it converts the inventory transfer parameters into automated guided vehicle (AGV) task instructions, including start point coordinates, end point coordinates, and material load; it converts the route replanning parameters into transportation vehicle navigation instructions, including the sequence of nodes passed and the estimated arrival time; and it converts the safety stock adjustment parameters into warehouse management system inventory threshold update instructions, including material codes and updated safety stock values; it then sends the task instructions and navigation instructions to the automated execution equipment via 5G, and sends the inventory threshold update instructions to the warehouse management system via industrial Ethernet; and it receives execution result receipts from the automated execution equipment and the warehouse management system.

Images