Large model auxiliary decision-making system for efficient management of supply chain

Abstract

The invention relates to the technical field of supply chain management, in particular to a large-model auxiliary decision-making system for efficient management of a supply chain, and the system comprises a data collection module, a priority calculation module, a decision-making verification module, an abnormity grading judgment module and a feedback adjustment module. The data acquisition module is used for synchronously acquiring data in a preset acquisition period and generating a target data set; the priority calculation module is used for calculating the priority weight of the allocation instruction; the decision verification module verifies the time constraint and the resource constraint of the allocation instruction; the abnormal grading judgment module judges the source type of the abnormal state according to the grading judgment condition; and the feedback adjustment module adjusts and optimizes the adjustable weight in the priority weight calculation model. According to the method, accurate calculation of the priority of the allocation instruction is realized through input data standardization, state estimation, weight synthesis and dynamic feedback adjustment, the scheduling decision precision is improved, the abnormal risk is reduced, and the utilization efficiency of supply chain resources is optimized.

Description

Technical Field

[0001] This invention relates to the field of supply chain management technology, and in particular to a large-scale model-assisted decision-making system for efficient supply chain management. Background Technology

[0002] With the increasing complexity of globalized trade and supply chains, traditional supply chain management decision-making systems face severe challenges. Existing technologies primarily employ rule-based or shallow machine learning models for prediction and scheduling, but their core shortcomings include:

[0003] The existing systems have a weak ability to capture nonlinear dynamic changes: traditional linear or shallow models struggle to accurately capture the highly nonlinear relationships between multiple variables such as inventory, transportation, and demand. Especially when faced with complex scenarios such as sudden fluctuations in market demand and dynamic changes in the saturation of transportation resources, the existing systems exhibit large prediction variance and severely inadequate prediction accuracy and robustness.

[0004] Lack of real-time, fine-grained verification mechanisms: Existing decision-making systems typically lack embedded, time-consistency verification of dispatch instructions after they are generated. Instruction verification is often based on static hard constraints, failing to provide real-time assessment of dynamic, micro-level indicators such as predicted transport delay margins and dynamic resource saturation, resulting in a high risk of default on dispatch instructions.

[0005] The anomaly diagnosis mechanism lacks depth and accuracy: when the system malfunctions, existing technologies mainly rely on human experience or simple threshold alarms, resulting in low efficiency in anomaly localization.

[0006] Furthermore, existing technologies generally lack proactive consideration for data security, business ethics, and decision-making fairness. In particular, when using large models, there is a risk of solidifying biases in historical data, as well as opaque decision-making processes and difficulty in defining responsibilities.

[0007] Chinese Patent Publication No. CN119204886A discloses a supply chain auxiliary decision-making method and system based on logistics big data. The method involves acquiring user behavior data and user location data; obtaining predicted demand based on the user behavior data; obtaining predicted delivery time based on the user location data; inputting the predicted demand and predicted delivery time into an allocation decision model to obtain a decision scheme; and managing the allocation process of the supply chain based on the decision scheme. It is evident that existing supply chain management technologies exhibit poor adaptability when facing complex environmental fluctuations or performance degradation, making it difficult to accurately attribute the root causes of anomalies, ultimately leading to uncontrollable risks in auxiliary decision-making. Summary of the Invention

[0008] To address this, the present invention provides a large-scale model-assisted decision-making system for efficient supply chain management. This system ensures decision reliability by constructing a dual verification mechanism that constrains default and predicts reliability. It also designs a closed-loop architecture for anomaly classification diagnosis and dynamic model weight correction to overcome the problem of uncontrollable risks in assisted decision-making caused by poor model adaptability in existing technologies.

[0009] To achieve the above objectives, the present invention provides a large-scale model-assisted decision-making system for efficient supply chain management, comprising:

[0010] The data acquisition module is used to synchronously collect inventory data, transportation data, and demand data within a preset acquisition period, and generate the target dataset for the current decision-making period.

[0011] The priority calculation module is used to obtain all the original input tensors in the target dataset, generate a set of allocation instructions to be evaluated based on the original input tensors, and calculate the priority weight of the allocation instruction corresponding to each original input tensor based on the priority weight calculation model. The priority weight calculation model is obtained by combining the inventory status, transportation capacity, time deviation and inventory deviation relative to the preset supply target in the current decision cycle.

[0012] The decision verification module is used to perform constraint verification based on the time and resource constraints of the allocation instruction. When a constraint violation is detected, it calls a nonlinear prediction function to obtain its prediction variance in order to determine whether the prediction result is reliable.

[0013] The anomaly classification judgment module is used to determine the fault type when an unreliable prediction result is obtained, based on the inventory deviation, the rate of change of resource saturation, and the cumulative value of prediction error.

[0014] The feedback adjustment module is used to determine the adjustment step size based on a preset step size calculation formula when the abnormal root cause type is determined to be an internal model fault type, and to perform proportional correction on the adjustable weights in the priority weight calculation model according to the adjustment step size, and output the corrected weight value.

[0015] Furthermore, the priority calculation module includes an input tensor generation unit and a weight calculation unit;

[0016] The input tensor generation unit is used to extract all the original input tensors for calculating the transfer instruction from the target dataset, and to construct the corresponding transfer instruction to be evaluated based on the original input tensors.

[0017] The weight calculation unit, connected to the input tensor generation unit, is used to calculate the weights of each of the original input tensors based on the priority weight calculation model, and generate the priority weights of the corresponding allocation instructions.

[0018] Furthermore, the weight calculation unit includes an input standardization subunit, a state estimation subunit, a constraint deviation measurement subunit, and a weight synthesis subunit;

[0019] The input standardization subunit is used to perform normalization processing on the numerical features in each of the original input tensors according to a preset interval mapping rule to obtain the standardized input tensor.

[0020] The state estimation subunit is used to calculate the inventory status and transportation capacity of the current decision-making cycle based on the standardized input tensor within a preset historical time window using a nonlinear prediction function.

[0021] The constraint deviation measurement subunit is used to calculate the time deviation and inventory deviation of the corresponding allocation instruction based on the preset supply target;

[0022] The weighted synthesis subunit is used to combine the inventory status quantity, the transportation capacity quantity, the time deviation quantity, and the inventory deviation quantity based on preset weight parameters to generate the priority weight of the transfer instruction.

[0023] Furthermore, the decision verification module includes a constraint calculation unit, a constraint default judgment unit, and a reliability assessment unit;

[0024] The constraint calculation unit is used to obtain the time constraints and resource constraints of the allocation instructions corresponding to the priority weights, and to calculate the predicted transportation delay margin and resource saturation.

[0025] A constraint default judgment unit, connected to the constraint calculation unit, is used to determine that the allocation instruction has a constraint default when the predicted transportation delay margin is greater than a preset transportation delay threshold and the resource saturation is greater than a preset saturation threshold.

[0026] The credible assessment unit, connected to the constraint default judgment unit, is used to call the prediction variance of the nonlinear prediction function when it is determined that the allocation instruction has a constraint default, and output the prediction unreliable result when the prediction variance is greater than a preset allowable threshold.

[0027] Furthermore, the credibility assessment unit includes a variance acquisition subunit and a credibility threshold determination subunit;

[0028] The variance acquisition subunit is connected to the constraint violation judgment unit and is used to acquire the prediction variance of the nonlinear prediction function when it is determined that the allocation instruction has a constraint violation.

[0029] The confidence threshold determination subunit is connected to the variance acquisition subunit and is used to determine that the prediction result is unreliable when the prediction variance is greater than a preset allowable threshold.

[0030] Furthermore, the anomaly classification judgment module includes a deviation ratio judgment unit and an external fluctuation judgment unit;

[0031] The deviation ratio judgment unit is used to determine whether the ratio of the inventory deviation to the preset supply target is greater than the preset deviation ratio threshold, so as to determine whether it is in an abnormal deviation ratio state.

[0032] An external fluctuation judgment unit, connected to the deviation ratio judgment unit, is used to obtain the periodic change rate of resource saturation when the deviation ratio is in an abnormal state, and compare the periodic change rate with a preset mutation threshold to determine whether the external fluctuation is in an abnormal state.

[0033] Furthermore, the anomaly classification judgment module also includes a model performance judgment unit;

[0034] The model performance judgment unit, connected to the external fluctuation judgment unit, is used to determine whether the statistical cumulative value of the prediction mean square error of the nonlinear prediction function is greater than a preset error cumulative threshold when there is an abnormal external fluctuation.

[0035] Furthermore, the anomaly classification judgment module also includes a hierarchical attribution diagnosis unit;

[0036] The hierarchical attribution diagnosis unit is connected to the model performance judgment unit to obtain abnormal information of each abnormal state and determine the fault type of each abnormal state based on the abnormal information.

[0037] Furthermore, the feedback adjustment module includes a dynamic step size calculation unit and a weight correction unit;

[0038] The dynamic step size calculation unit is used to adjust the step size based on a preset step size calculation formula when the fault type is an internal model fault type.

[0039] The weight correction unit is connected to the dynamic step size calculation unit and is used to correct the adjustable weights in the priority weight calculation model according to the adjustment step size.

[0040] Furthermore, the weight correction unit includes a scaling factor calculation subunit and a weight adjustment subunit;

[0041] The scaling factor calculation subunit is used to calculate the scaling factor based on the adjustment step size;

[0042] The weight adjustment subunit is connected to the scaling factor calculation subunit, and is used to proportionally adjust the adjustable weights in the priority weight calculation model based on the scaling factor and output the adjusted weights.

[0043] Compared with existing technologies, the advantages of this invention lie in forming a complete closed-loop scheduling mechanism through priority weight calculation, constraint default verification, predictive reliability assessment, and hierarchical anomaly diagnosis. This mechanism improves the accuracy of supply chain allocation, isolates risks, and adaptively optimizes the model, thereby enhancing the system's stability and decision reliability in complex dynamic environments. Specifically, all allocation instructions are assigned dynamic priority weights to construct an importance ranking queue that updates in real time with business status. This allows system resources to automatically and centrally process high-priority, high-reliability allocation instructions. Upon identifying instructions with constraint defaults, the system downgrades or postpones these instructions based on the reliability assessment results. This creates an adaptive decision-making resource scheduling mechanism in scenarios with massive concurrent allocation tasks, improving the system's performance under high load conditions. This invention improves overall processing efficiency and output quality; it incorporates prediction variance into the decision verification process and binds it to time and resource constraints to achieve a quantitative judgment on the executability of allocation instructions; it transforms the uncertainty of model prediction from abstract risk into a judgment on the feasibility of executing specific instructions, thereby proactively isolating high-risk allocation instructions and avoiding the execution of allocation operations that would cause losses or chain anomalies when predictions are unstable or have large deviations. It also addresses anomalies caused by external disturbances such as sudden increases in demand and temporary shortages of transportation resources, as well as anomalies caused by internal model performance degradation, avoiding accidental model adjustments due to short-term external fluctuations, thus improving the targeting of correction actions; simultaneously, it only triggers the feedback adjustment module to correct adjustable weights when an internal model fault is determined, making the model update process more robust and efficient.

[0044] Furthermore, by generating input tensors, standardizing processes, estimating states, and synthesizing weights, the system achieves comprehensive weight calculations for inventory state quantities, transportation capacity quantities, time deviation quantities, and inventory deviation quantities, thereby improving the sorting accuracy of allocation instructions and ensuring the consistency and stability of the system's decisions under multi-dimensional inputs. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the structure of a large-scale model-assisted decision-making system for efficient supply chain management according to an embodiment of the present invention;

[0046] Figure 2 This is a logic diagram of the constraint default judgment unit in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the anomaly classification judgment module according to an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of the feedback adjustment module in an embodiment of the present invention. Detailed Implementation

[0049] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0050] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0051] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0052] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0053] Please see Figure 1 As shown, this is a schematic diagram of the structure of a large-scale model-assisted decision-making system for efficient supply chain management according to an embodiment of the present invention. The present invention provides a large-scale model-assisted decision-making system for efficient supply chain management, comprising:

[0054] The data acquisition module is used to synchronously collect inventory data, transportation data, and demand data within a preset acquisition period, and generate the target dataset for the current decision-making period.

[0055] The priority calculation module is used to obtain all the original input tensors in the target dataset, generate a set of allocation instructions to be evaluated based on the original input tensors, and calculate the priority weight of the allocation instruction corresponding to each original input tensor based on the priority weight calculation model. The priority weight calculation model is obtained by combining the inventory status, transportation capacity, time deviation and inventory deviation relative to the preset supply target in the current decision cycle.

[0056] The decision verification module is used to perform constraint verification based on the time and resource constraints of the allocation instruction. When a constraint violation is detected, it calls a nonlinear prediction function to obtain its prediction variance in order to determine whether the prediction result is reliable.

[0057] The anomaly classification judgment module is used to determine the fault type when an unreliable prediction result is obtained, based on the inventory deviation, the rate of change of resource saturation, and the cumulative value of prediction error.

[0058] The feedback adjustment module is used to determine the adjustment step size based on a preset step size calculation formula when the abnormal root cause type is determined to be an internal model fault type, and to perform proportional correction on the adjustable weights in the priority weight calculation model according to the adjustment step size, and output the corrected weight value.

[0059] In this embodiment, the data acquisition module is used to legally and securely collect inventory data, transportation data, and demand data simultaneously within a preset collection period, in accordance with the requirements of applicable laws and regulations such as the Cybersecurity Law, the Data Security Law, and the Personal Information Protection Law, by taking encryption, anonymization, or de-identification measures.

[0060] During the model training phase, the priority calculation module performs bias audits on historical data and introduces minimum fairness constraints into the model to ensure that the average service latency differences among different customer groups are within a preset ethical threshold, so as to ensure that priority allocation is based not only on efficiency but also on fairness.

[0061] The target dataset represents the collection of data compiled from inventory data, transportation data, and demand data for calculating allocation instructions, providing standardized input required by the priority calculation module;

[0062] The original input tensor represents the transformation of various types of data in the target dataset into a multidimensional array according to a unified format, which is then used to input the priority weight calculation model for computation.

[0063] Based on the original input tensor, a set of transfer instructions to be evaluated is generated. The generation of transfer instructions follows a well-known rule system in supply chain management systems, including inventory balancing rules, demand coverage rules, transportation capacity constraint rules, and time window rules. The above rules belong to the constraint models commonly used in existing supply chain optimization systems, and are implemented as standard business logic in typical WMS, TMS, or APS systems. This invention does not limit itself to such rules, but generates feasible transfer actions from the original input tensor based on existing rule parsing methods.

[0064] Priority weights represent the importance of the execution order of allocation instructions, guiding the system to sort and schedule resource allocation. While optimizing supply chain efficiency, the system design of this invention follows the principle of technology for good. Its transparency and fairness constraints help all participants, including small and medium-sized enterprises, to supervise and question the decision-making process, thereby mitigating the social impact of technological solidification of competitive advantages.

[0065] In this embodiment, after the priority calculation module generates the priority weight of the allocation instruction, interpretable artificial intelligence (XAI) technology, such as the LIME algorithm, is used to calculate the contribution of each original input tensor in the target dataset, such as inventory status, transportation capacity, time deviation, and inventory deviation, to the final priority weight value. This will quantify and demonstrate the driving effect of each input factor on the priority ranking, and transform the decision of the weight calculation model into a transparent and understandable attribution report.

[0066] Simultaneously, continuously monitor the output of the decision verification module and the feedback adjustment module, and record detailed logs including timestamps and allocation instruction IDs; record the specific status when the constraint default is triggered, including the values ​​of predicted transportation delay margin and resource saturation; record the basis for determining the prediction as unreliable, namely the prediction variance of the nonlinear prediction function and its comparison with the preset allowable threshold; record the type of abnormal root cause finally determined by the abnormal classification judgment module, such as internal model failure type or external environmental interference type, as well as the underlying data supporting the judgment, such as the statistical cumulative value of the model mean square error;

[0067] A breach of constraint indicates that the allocation order violates time or resource constraints during execution, which triggers the decision verification module to perform anomaly analysis.

[0068] The prediction variance of a nonlinear prediction function represents a measure of the uncertainty of the model's prediction of inventory status or transportation capacity, and is used to judge the reliability of the allocation instruction prediction.

[0069] The prediction of unreliable results indicates that the allocation command is unreliable when the prediction variance is greater than a preset threshold, which is used to trigger the anomaly classification judgment module to perform root cause analysis.

[0070] The anomaly root cause type indicates the category of anomaly source corresponding to unreliable prediction results, such as internal model failure type or external environmental interference type, which is used to guide feedback adjustment or strategy optimization.

[0071] The grading criteria represent a set of rules used to distinguish the source and severity of abnormalities, in order to determine the abnormal state and perform stratified diagnosis.

[0072] Adjustable weights refer to parameters in the priority weight calculation model that can be adjusted and optimized through feedback, in order to improve the accuracy of allocation command prediction.

[0073] The preset adjustment step size represents the magnitude of each update to the adjustable weights during the feedback adjustment process, which is used to control the speed of weight adjustment and system stability.

[0074] A complete closed-loop scheduling mechanism is formed through priority weight calculation, constraint default verification, predictive reliability assessment, and hierarchical anomaly diagnosis. This mechanism improves the accuracy of supply chain allocation, isolates risks, and adaptively optimizes the model, thereby enhancing the system's stability and decision reliability in complex and dynamic environments. Specifically, all allocation instructions are assigned dynamic priority weights to construct an importance ranking queue that updates in real time with business status. This allows system resources to automatically and centrally process high-priority, high-reliability allocation instructions. Upon identifying instructions with constraint defaults, the system downgrades or postpones these instructions based on the reliability assessment results. This creates an adaptive decision-making resource scheduling mechanism in scenarios with massive concurrent allocation tasks, improving the overall processing efficiency and output of the system under high load conditions. Quality: By incorporating prediction variance into the decision verification process and binding it with time and resource constraints, this invention enables a quantitative judgment of the executability of allocation instructions. It transforms the uncertainty of model prediction from abstract risk into a judgment of the feasibility of executing specific instructions, thereby proactively isolating high-risk allocation instructions and avoiding the execution of allocation operations that could cause losses or chain reactions when predictions are unstable or have large deviations. It also addresses anomalies caused by external disturbances such as sudden increases in demand or temporary shortages of transportation resources, as well as anomalies caused by internal model performance degradation. This prevents accidental model adjustments due to short-term external fluctuations, thus improving the targeted nature of correction actions. Furthermore, the feedback adjustment module is only triggered to correct adjustable weights when an internal model fault is identified, making the model update process more robust and efficient.

[0075] Specifically, the priority calculation module includes an input tensor generation unit and a weight calculation unit;

[0076] The input tensor generation unit is used to extract all the original input tensors for calculating the transfer instruction from the target dataset, and to construct the corresponding transfer instruction to be evaluated based on the original input tensors.

[0077] The weight calculation unit, connected to the input tensor generation unit, is used to calculate the weights of each of the original input tensors based on the priority weight calculation model, and generate the priority weights of the corresponding allocation instructions.

[0078] Specifically, the weight calculation unit includes an input standardization subunit, a state estimation subunit, a constraint deviation measurement subunit, and a weight synthesis subunit;

[0079] The input standardization subunit is used to perform normalization processing on the numerical features in each of the original input tensors according to a preset interval mapping rule to obtain the standardized input tensor.

[0080] The state estimation subunit is used to calculate the inventory status and transportation capacity of the current decision-making cycle based on the standardized input tensor within a preset historical time window using a nonlinear prediction function.

[0081] The constraint deviation measurement subunit is used to calculate the time deviation and inventory deviation of the corresponding allocation instruction based on the preset supply target;

[0082] The weighted synthesis subunit is used to combine the inventory status quantity, the transportation capacity quantity, the time deviation quantity, and the inventory deviation quantity based on preset weight parameters to generate the priority weight of the transfer instruction.

[0083] In this embodiment, the original input tensor represents a multidimensional data array extracted from the target dataset, which serves as the basis for calculating the priority of the allocation instruction;

[0084] The standardized input tensor represents the result of normalizing the numerical features in the original input tensor according to a preset interval mapping rule, in order to eliminate the differences in the units of different indicators and improve the stability of subsequent calculations.

[0085] The normalization process uses the linear mapping formula:

[0086]

[0087] Where x represents a certain type of data in the original input;

[0088] x' represents a standardized type of data;

[0089] and These are the minimum and maximum values ​​for each data category, selected from historical observation data from the most recent 12 consecutive decision-making cycles.

[0090] For example, the minimum value of the inventory status quantity is the lowest value of all inventory status quantities within the 12 periods, and the maximum value is the highest value of the corresponding inventory status quantity.

[0091] Inventory status represents the available inventory level within the current decision-making period, predicted by a nonlinear prediction function. The calculation is performed based on the standardized input tensor within a preset historical time window, using the following formula:

[0092]

[0093] in, The set of standardized input tensors within a preset historical time window;

[0094] This represents the inventory status during the current decision-making cycle, reflecting the availability of inventory resources.

[0095] The transport capacity represents the transport capacity available for allocation within the current decision-making period. It is calculated using a nonlinear prediction function based on the same standardized input tensor set. The calculation formula is as follows:

[0096]

[0097] It indicates the transportation capacity during the current decision-making cycle, reflecting the availability of transportation resources;

[0098] The time deviation represents the difference between the actual execution time of the allocation order and the preset target time, and is used to quantify the risk of allocation delay.

[0099] Inventory deviation represents the difference between actual inventory and the preset supply target, and is used to quantify the degree of supply and demand imbalance.

[0100] Time deviation represents the difference between the actual execution time of the allocation order and the preset target time. It is used to quantify the degree of time deviation during the allocation process, and its calculation formula is as follows:

[0101]

[0102] in, This refers to the actual execution time of the allocation order. The preset target time for the corresponding allocation instruction;

[0103] Inventory deviation represents the difference between actual inventory and the preset supply target, used to quantify the degree of deviation from the inventory level; its calculation formula is:

[0104]

[0105] in, This represents the current actual inventory level.

[0106] For the corresponding preset supply target;

[0107] In this embodiment, the weight parameters w1, w2, w3, and w4 represent the relative importance of inventory status, transportation capacity, time deviation, and inventory deviation in the priority weight calculation, respectively.

[0108] w1 corresponds to the inventory status quantity; w2 corresponds to the transportation capacity quantity; w3 corresponds to the time deviation quantity; w4 corresponds to the inventory deviation quantity.

[0109] Each value ranges from [0,1] and satisfies the condition that the sum of the four weight parameters is 1;

[0110] In this embodiment, the optimal values ​​of the weight parameters are obtained through backtesting and simulation calculations of historical allocation records: w1=0.35, w2=0.25, w3=0.20, w4=0.20;

[0111] The final priority weight W is calculated using the following formula:

[0112]

[0113] Where W is the final priority weight of the allocation instruction, which reflects the execution priority of the allocation instruction relative to other instructions.

[0114] By generating input tensors, standardizing, estimating states, and synthesizing weights, the system can calculate the comprehensive weights of inventory state, transportation capacity, time deviation, and inventory deviation, thereby improving the sorting accuracy of allocation instructions and ensuring the consistency and stability of the system's decisions under multi-dimensional inputs.

[0115] See Figure 2 As shown, it is a logic decision diagram of the constraint breach judgment unit in an embodiment of the present invention;

[0116] Specifically, the decision verification module includes a constraint calculation unit, a constraint default judgment unit, and a reliability assessment unit;

[0117] The constraint calculation unit is used to obtain the time constraints and resource constraints of the allocation instructions corresponding to the priority weights, and to calculate the predicted transportation delay margin and resource saturation.

[0118] A constraint default judgment unit, connected to the constraint calculation unit, is used to determine that the allocation instruction has a constraint default when the predicted transportation delay margin is greater than a preset transportation delay threshold and the resource saturation is greater than a preset saturation threshold.

[0119] The credible assessment unit, connected to the constraint default judgment unit, is used to call the prediction variance of the nonlinear prediction function when it is determined that the allocation instruction has a constraint default, and output the prediction unreliable result when the prediction variance is greater than a preset allowable threshold.

[0120] In this embodiment, the time constraint represents the time limit that the allocation instruction must meet during execution, which consists of the planned execution time and the latest execution time that the system can tolerate, and is used to define the time boundary of the allocation task; the resource constraint represents the ratio between the resource scale required to execute the allocation instruction and the current available resource scale, and is used to define the resource feasibility of the allocation task.

[0121] The predicted transport delay margin represents the remaining amount of time between the predicted execution time of a transfer order and the latest tolerable execution time of the system, used to quantify the time urgency; its calculation formula is as follows:

[0122]

[0123] in, This represents the maximum execution time that the system can tolerate for the same type of task over a number of consecutive decision-making cycles in history. In this embodiment, 8 consecutive decision-making cycles are used. This indicates the predicted execution time of the allocation instruction, obtained based on a nonlinear prediction function.

[0124] Resource saturation represents the proportion of currently available transportation resources occupied by the corresponding allocation order, used to quantify resource scarcity; its calculation formula is:

[0125]

[0126] in This indicates the average resource requirement for this allocation order based on statistics from similar historical tasks; This indicates the amount of transportation resources that can be scheduled within the current decision-making cycle;

[0127] A constraint default indicates that the allocation order exceeds both the time constraint and the resource constraint simultaneously, that is, the predicted transportation delay margin is greater than the preset transportation delay threshold and the resource saturation is greater than the preset saturation threshold.

[0128] Transportation delay threshold The value range is [0, 4] hours, and in this embodiment, the optimal empirical value of 2 hours is used; saturation threshold The value range is [0,1]. In this embodiment, the optimal empirical value of 0.75 is taken.

[0129] The constraint calculation unit obtains the time and resource constraints of the allocation instruction, calculates the transportation delay margin and resource saturation, realizes the real-time judgment of constraint violation, and provides risk warning and scheduling reference for the system.

[0130] Specifically, the credibility assessment unit includes a variance acquisition subunit and a credibility threshold determination subunit;

[0131] The variance acquisition subunit is connected to the constraint violation judgment unit and is used to acquire the prediction variance of the nonlinear prediction function when it is determined that the allocation instruction has a constraint violation.

[0132] The confidence threshold determination subunit is connected to the variance acquisition subunit and is used to determine that the prediction result is unreliable when the prediction variance is greater than a preset allowable threshold.

[0133] In this embodiment, the prediction variance represents the degree of prediction uncertainty obtained by the nonlinear prediction function based on historical observation bias, and is used to evaluate the prediction reliability. The calculation formula is as follows:

[0134]

[0135] in, These are actual observed values. To correspond to the predicted value, in this embodiment, n is the number of samples in the most recent 12 consecutive decision periods;

[0136] An unreliable prediction result indicates the assessment of prediction reliability when the prediction variance exceeds a preset tolerance threshold; the tolerance threshold is... The value range is [0,1], and the empirically optimal value of 0.35 is adopted in this embodiment;

[0137] The prediction variance is compared with a preset tolerance threshold. When the prediction variance is greater than the preset tolerance threshold, it is determined to be an unreliable prediction result; when the prediction variance is less than or equal to the preset tolerance threshold, it is determined to be a reliable prediction result.

[0138] By obtaining the prediction variance of the nonlinear prediction function through the reliable evaluation unit and comparing it with the allowable threshold, the reliability of the allocation instruction prediction can be accurately judged, the unreliable prediction state can be identified, and the risk of abnormal allocation can be effectively reduced.

[0139] Specifically, the anomaly classification judgment module includes a deviation ratio judgment unit and an external fluctuation judgment unit;

[0140] The deviation ratio judgment unit is used to determine whether the ratio of the inventory deviation to the preset supply target is greater than the preset deviation ratio threshold, so as to determine whether it is in an abnormal deviation ratio state.

[0141] An external fluctuation judgment unit, connected to the deviation ratio judgment unit, is used to obtain the periodic change rate of resource saturation when the deviation ratio is in an abnormal state, and compare the periodic change rate with a preset mutation threshold to determine whether the external fluctuation is in an abnormal state.

[0142] In this embodiment, the deviation ratio represents the proportion of inventory deviation to the preset supply target, which is used to measure the relative degree of inventory deviation in the overall supply target; the deviation ratio threshold represents the maximum deviation ratio that the system can tolerate, which is used to define the trigger boundary of the abnormal state of deviation ratio; the deviation ratio threshold ranges from [0,1], and in this embodiment, the optimal empirical value is determined to be 0.30 based on historical allocation deviation distribution and stability analysis.

[0143] After obtaining the inventory deviation amount and the preset supply target, the ratio is calculated by comparison and then compared with the deviation ratio threshold: when the ratio is greater than the deviation ratio threshold, it is determined to be an abnormal deviation ratio state; when the ratio is less than or equal to the deviation ratio threshold, the inventory deviation is judged to be in a normal state.

[0144] The periodic change rate represents the magnitude of change in resource saturation between adjacent decision-making periods, used to reflect whether there are sudden fluctuations in the external resource environment; the mutation threshold represents the upper limit of acceptable resource saturation change in the system, used to determine the degree of abnormality of external fluctuations; the mutation threshold ranges from [0,1], and in this embodiment, the optimal empirical value is determined to be 0.20 by backtesting the change characteristics of historical resource saturation time series.

[0145] The external fluctuation judgment unit compares the periodic change rate with the mutation threshold: when the periodic change rate is greater than the mutation threshold, it is determined to be an abnormal external fluctuation state; when the periodic change rate is less than or equal to the mutation threshold, it is determined that the external fluctuation is within the normal range.

[0146] The deviation ratio judgment unit calculates the proportion of inventory deviation to the preset supply target, identifies abnormal deviation ratio states, provides basic data for abnormal root cause analysis, and supports subsequent scheduling optimization.

[0147] Specifically, the anomaly classification judgment module also includes a model performance judgment unit;

[0148] The model performance judgment unit, connected to the external fluctuation judgment unit, is used to determine whether the statistical cumulative value of the prediction mean square error of the nonlinear prediction function is greater than a preset error cumulative threshold when there is an abnormal external fluctuation.

[0149] In this embodiment, the model performance judgment unit adopts a statistical error accumulation method based on a sliding time window; the system obtains the actual inventory change in each prediction period t. Output value of nonlinear prediction function The prediction mean square error is calculated using the following formula:

[0150]

[0151] in, To predict the mean square error;

[0152] To enhance the ability to identify delayed convergence of anomalies, a sliding accumulation model with a window width of W is used to construct the cumulative error statistics:

[0153]

[0154] Where T is the current time point, i.e. the end time of the prediction period;

[0155] W is the width of the sliding window, and its value is determined by combining the business forecasting period and the fluctuation frequency. In this embodiment, W = 12 is used to cover the typical half-day forecasting period.

[0156] Error accumulation threshold Based on the training set error distribution, the calculation method is as follows:

[0157]

[0158] in This represents the mean of the cumulative error (MSE) during the training phase. The standard deviation of the cumulative error MSE during the training phase is represented by k. The value of k ranges from 2.0 to 3.0. In this embodiment, k = 2.5 is selected in combination with the business volatility to cover 99% of the normal error range and improve the recognition accuracy under external disturbance conditions.

[0159] By analyzing the periodic change rate of resource saturation through the external fluctuation judgment unit and comparing it with the mutation threshold, abnormal external fluctuation states can be identified, enabling timely response and scheduling adjustments to sudden changes in resources and the environment.

[0160] See Figure 3 As shown, it is a structural schematic diagram of the anomaly classification judgment module in an embodiment of the present invention;

[0161] Specifically, the anomaly classification judgment module also includes a hierarchical attribution diagnosis unit;

[0162] The hierarchical attribution diagnosis unit is connected to the model performance judgment unit to obtain abnormal information of each abnormal state and determine the fault type of each abnormal state based on the abnormal information.

[0163] In this embodiment, the hierarchical attribution diagnosis unit is used to obtain the abnormal information of each abnormal state in the abnormal classification judgment module, including inventory abnormality, transportation abnormality and demand abnormality, and to determine the fault type of each abnormal state by analyzing the abnormal information, such as inventory backlog, stockout risk or transportation delay; the judgment result is output to the scheduling optimization module or strategy adjustment module to support subsequent optimization decisions for inventory, transportation and demand scheduling.

[0164] The hierarchical attribution diagnostic unit obtains abnormal information from the abnormal classification judgment module and determines the fault type of each abnormal state, such as inventory backlog, stockout, or transportation delay, providing a basis for strategy optimization and scheduling adjustment, and achieving precise abnormal handling.

[0165] See Figure 4 As shown, it is a structural schematic diagram of the feedback adjustment module in an embodiment of the present invention;

[0166] Specifically, the feedback adjustment module includes a dynamic step size calculation unit and a weight correction unit;

[0167] The dynamic step size calculation unit is used to adjust the step size based on a preset step size calculation formula when the fault type is an internal model fault type.

[0168] The weight correction unit is connected to the dynamic step size calculation unit and is used to correct the adjustable weights in the priority weight calculation model according to the adjustment step size.

[0169] In this embodiment, when the dynamic step size calculation unit detects an internal model fault type, it calculates and adjusts the step size Δs based on the internal model's operating status and historical weight information, as shown in the following formula:

[0170]

[0171] in, The adjustable weights for the current decision-making cycle. The weight corresponding to the previous decision-making cycle;

[0172] The adjustable weight set is ;

[0173] Adjustable weights represent the set of weights that can be adjusted in the priority weight calculation model. For example, to optimize the output of the decision model, the system can increase or decrease the weights corresponding to each tensor to reflect the influence of different input tensors. At least one baseline weight should be retained. Keep it unchanged to ensure the stability of the model during the adjustment process;

[0174] The duration of the detected internal model fault, in seconds;

[0175] These are the control parameter values ​​within the current module. These are reference values ​​for normal operation.

[0176] α, β, γ are preset proportionality coefficients used to precisely adjust the contribution of each parameter to the step size. In this embodiment, α = 0.5, β = 0.2, and γ = 0.3.

[0177] The calculated Δs is expressed as a percentage or in absolute weight units;

[0178] The weight correction unit adjusts the adjustable weights based on Δs, using the following formula:

[0179]

[0180] in The importance coefficient corresponding to the input tensor can be quantified according to the influence of the input tensor on the decision result, with a value range of [0,1]. The corrected weight is output to the priority weight calculation model for the input tensor priority calculation in the next decision cycle.

[0181] By using a dynamic step-size calculation unit to calculate and adjust the step size based on the internal model's fault state and historical weight information, the adjustable weights can be accurately updated, thereby improving the system's response speed to internal faults and the accuracy of allocation decisions.

[0182] Specifically, the weight correction unit includes a scaling factor calculation subunit and a weight adjustment subunit;

[0183] The scaling factor calculation subunit is used to calculate the scaling factor based on the adjustment step size;

[0184] The weight adjustment subunit is connected to the scaling factor calculation subunit, and is used to proportionally adjust the adjustable weights in the priority weight calculation model based on the scaling factor and output the adjusted weights.

[0185] In this embodiment, the scaling factor calculation subunit is used to calculate the scaling factor k based on the dynamic step size Δs and the historical adjustable weight deviation, as shown in the following formula:

[0186]

[0187] Δs is the adjustment step size calculated by the dynamic step size calculation unit, which is used to reflect the current internal model fault correction requirements;

[0188] This is the historical average of the adjustable weights, used to quantify the typical level of the adjustable weights;

[0189] To prevent division by zero of small constants;

[0190] This is the scaling factor, used to precisely control the effect of the step size on scaling.

[0191] The weight adjustment subunit adjusts the adjustable weight set according to the scaling factor k. Make proportional adjustments using the following formula:

[0192]

[0193] By using the scaling factor calculation sub-unit and the weight adjustment sub-unit, the dynamic step size and historical adjustable weight deviation are converted into the scaling factor k. The adjustable weight is then adjusted proportionally to achieve dynamic optimization of priority weights, while maintaining the stability of the baseline weights and ensuring the reliability of allocation decisions.

[0194] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0195] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A large model assisted decision system for efficient management of supply chain, characterized in that, The method comprises the following steps: a data collection module is used to synchronously collect inventory data, transportation data and demand data within a preset collection period, and generate a target data set of a current decision period; a priority calculation module is used to obtain all original input tensors in the target data set, generate a set of to-be-evaluated allocation instructions based on the original input tensors, and calculate a priority weight of an allocation instruction corresponding to each original input tensor based on a priority weight calculation model, wherein the priority weight calculation model is obtained based on at least a combined operation of an inventory state quantity, a transportation capacity quantity and a time deviation quantity and an inventory deviation quantity relative to a preset supply target in the current decision period; a decision verification module is used to perform constraint verification based on a time constraint and a resource constraint of the allocation instruction, and when a constraint violation is detected, a nonlinear prediction function is called to obtain a prediction variance, so as to determine whether the prediction result is reliable; an abnormal grading judgment module is used to, when an unreliable prediction result is obtained, determine a fault type to which the unreliable prediction result belongs based on an inventory deviation quantity, a resource saturation change rate and a prediction error cumulative value; a feedback adjustment module is used to, when the abnormal root type is determined as an internal model fault type, determine an adjustment step based on a preset step calculation formula, perform proportional correction on an adjustable weight in the priority weight calculation model according to the adjustment step, and output a corrected weight value.

2. The large model assisted decision system for efficient management of supply chain according to claim 1, wherein, The priority calculation module comprises an input tensor generation unit and a weight calculation unit; the input tensor generation unit is used to extract all original input tensors for allocation instruction calculation from the target data set, and construct corresponding to-be-evaluated allocation instructions based on the original input tensors; the weight calculation unit is used to perform weight calculation on each original input tensor based on a priority weight calculation model, and generate a priority weight of a corresponding allocation instruction.

3. The large model assisted decision system for efficient management of supply chain according to claim 2, wherein, The weight calculation unit comprises an input standardization subunit, a state estimation subunit, a constraint deviation measurement subunit and a weight synthesis subunit; the input standardization subunit is used to perform normalization processing on a numerical feature in each original input tensor according to a preset interval mapping rule, and obtain a standardized input tensor; the state estimation subunit is used to calculate an inventory state quantity and a transportation capacity quantity of the current decision period within a preset historical time window based on the standardized input tensor through a nonlinear prediction function; the constraint deviation measurement subunit is used to calculate a time deviation quantity and an inventory deviation quantity of a corresponding allocation instruction according to a preset supply target; the weight synthesis subunit is used to combine and operate the inventory state quantity, the transportation capacity quantity, the time deviation quantity and the inventory deviation quantity based on a preset weight parameter, and generate a priority weight of an allocation instruction.

4. The large model assisted decision system for efficient management of supply chain according to claim 1, wherein, The decision verification module comprises a constraint calculation unit, a constraint violation judgment unit and a reliable evaluation unit; the constraint calculation unit is used to obtain a time constraint and a resource constraint of an allocation instruction corresponding to the priority weight, calculate a predicted transportation delay margin and a resource saturation degree; The constraint violation judgment unit is connected with the constraint calculation unit, and is configured to judge that the allocation instruction has a constraint violation when the predicted transportation delay margin is greater than a preset transportation delay threshold and the resource saturation degree is greater than a preset saturation threshold. The credible evaluation unit is connected with the constraint violation judgment unit, and is configured to call a prediction variance of the nonlinear prediction function when it is judged that the allocation instruction has a constraint violation, and output an uncredible prediction result when the prediction variance is greater than a preset allowable threshold.

5. The large model assisted decision system for efficient management of supply chain according to claim 4, wherein, The credible evaluation unit comprises a variance acquisition subunit and a credible threshold judgment subunit. The variance acquisition subunit is connected with the constraint violation judgment unit, and is configured to acquire a prediction variance of the nonlinear prediction function when it is judged that the allocation instruction has a constraint violation. The credible threshold judgment subunit is connected with the variance acquisition subunit, and is configured to judge that the prediction is uncredible when the prediction variance is greater than a preset allowable threshold.

6. The large model assisted decision system for efficient management of supply chain according to claim 1, wherein, The abnormality grading judgment module comprises a deviation proportion judgment unit and an external fluctuation judgment unit. The deviation proportion judgment unit is configured to judge whether a proportion of the inventory deviation amount to the preset supply target is greater than a preset deviation proportion threshold, so as to determine whether a deviation proportion abnormality state exists. The external fluctuation judgment unit is connected with the deviation proportion judgment unit, and is configured to acquire a periodic change rate of the resource saturation degree when the deviation proportion abnormality state exists, and compare the periodic change rate with a preset mutation threshold, so as to judge whether an external fluctuation abnormality state exists.

7. The large model assisted decision system for efficient management of supply chain according to claim 6, wherein, The abnormality grading judgment module further comprises a model performance judgment unit. The model performance judgment unit is connected with the external fluctuation judgment unit, and is configured to judge whether a statistical cumulative value of a prediction mean square error of the nonlinear prediction function is greater than a preset error cumulative threshold when the external fluctuation abnormality exists.

8. The large model assisted decision system for efficient management of supply chain according to claim 7, wherein, The abnormality grading judgment module further comprises a hierarchical attribution diagnosis unit. The hierarchical attribution diagnosis unit is connected with the model performance judgment unit, and is configured to acquire abnormality information of each abnormality state, and judge a fault type to which each abnormality state belongs according to the abnormality information.

9. The big model assisted decision system for efficient management of supply chain according to claim 1, wherein, The feedback adjustment module comprises a dynamic step calculation unit and a weight correction unit. The dynamic step calculation unit is configured to adjust a step based on a preset step calculation formula when the fault type is an internal model fault type. The weight correction unit is connected with the dynamic step calculation unit, and is configured to correct an adjustable weight in the priority weight calculation model according to the adjusted step.

10. The large model assisted decision system for efficient management of supply chain according to claim 9, wherein, The weight correction unit comprises a proportional scaling coefficient calculation subunit and a weight adjustment subunit. The proportional scaling coefficient calculation subunit is configured to calculate a proportional scaling coefficient according to the adjusted step. The weight adjustment subunit is connected with the proportional scaling coefficient calculation subunit, and is configured to correct the adjustable weight in the priority weight calculation model in proportion based on the proportional scaling coefficient, and output a corrected weight.

Images