An AI-based nourishing dry goods intelligent supply chain management system

Abstract

The application discloses an AI-based nourishing dry goods intelligent supply chain management system, belongs to the technical field of data processing and supply chain management, and comprises a data acquisition unit, which is used for collecting historical sales data, meteorological forecast data, real-time road conditions, cold chain sensor data and market intelligence; an AI prediction unit, which is used for predicting demand and logistics paths based on the data collected by the data acquisition unit to generate a recommended adjustment amount; a game equilibrium unit, which is used for performing multi-agent game equilibrium analysis based on the data collected by the data acquisition unit and a preset utility function to solve an elastic buffer threshold value; and a decision execution unit, which is used for comparing and processing the recommended adjustment amount generated by the AI prediction unit and the elastic buffer threshold value solved by the game equilibrium unit; and the application significantly enhances the cohesion and stability of a supply chain alliance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing and supply chain management, in particular to an AI-based intelligent supply chain management system for tonifying dry goods. BACKGROUND

[0002] In the supply chain management of high-end tonifying dry goods such as bird's nest, sea cucumber, fish maw, donkey-hide gelatin, ginseng, longan, and dendrobium, especially for rare products such as wild dendrobium and ginseng, artificial intelligence is widely used for demand prediction and logistics path planning; at the same time, cold chain technology is used to ensure the quality of goods during transportation; however, there is an inherent technical problem when existing technologies integrate AI prediction and cold chain logistics; when facing external strong disturbances such as extreme weather events, the AI system will frequently adjust the model and path in pursuit of real-time accuracy of prediction; this high-frequency decision-making change, i.e. decision oscillation effect, directly impacts the stability of the cold chain logistics transportation plan, which may lead to repeated scheduling of transportation tools, increased waiting time, and unexpected increase in the number of transshipment; this in turn increases the energy consumption of the cold chain and the loss rate of goods, ultimately reducing the overall efficiency and economic benefits of the supply chain. Existing solutions often focus on improving the performance of a single link and fail to solve the technical contradiction between dynamic optimization and stable operation.

[0003] To solve the above technical problems, the present application provides an AI-based intelligent supply chain management system for tonifying dry goods. SUMMARY

[0004] The purpose of the present application is to provide an AI-based intelligent supply chain management system for tonifying dry goods to solve the problems raised in the background technology.

[0005] The technical solution of the present application is as follows:

[0006] A data acquisition unit for acquiring historical sales data, weather forecast data, real-time traffic, cold chain sensor data, and market intelligence;

[0007] An AI prediction unit for predicting demand and logistics paths based on the data acquired by the data acquisition unit to generate a recommended adjustment amount;

[0008] A game equilibrium unit for performing multi-agent game equilibrium analysis based on the data acquired by the data acquisition unit and a preset utility function to calculate an elastic buffer threshold;

[0009] A decision execution unit for comparing the recommended adjustment amount generated by the AI prediction unit with the elastic buffer threshold calculated by the game equilibrium unit; when the recommended adjustment amount exceeds the elastic buffer threshold, a direct adjustment signal is generated; when the recommended adjustment amount does not exceed the elastic buffer threshold, a delayed decision signal is generated;

[0010] a probability decision unit, configured to perform probability evolution calculation on the plurality of candidate decision schemes in response to the delay decision signal to generate posterior probabilities of the schemes, and compare and analyze the posterior probabilities with a preset collapse threshold to generate a deterministic decision instruction.

[0011] Preferably, the game equilibrium analysis process is as follows: utility functions are constructed for the supplier subject, the logistics subject and the retailer subject in the supply chain; a Nash equilibrium strategy combination that makes the utility of all subjects reach local optimum at the same time is found through iterative calculation; the optimal strategy adjustment amount of each subject in the Nash equilibrium strategy combination is extracted; and the minimum value in the optimal strategy adjustment amount is set as the elastic buffer threshold.

[0012] Preferably, the utility function components are measured in unified currency units.

[0013] The utility function includes an individual revenue function, an operation cost function and a penalty function.

[0014] The individual revenue function is derived from a revenue distribution model preset in the smart contract.

[0015] The operation cost function is derived from the financial and operational databases of the enterprise itself.

[0016] The penalty function takes effect when the current strategy deviates from the commitment strategy published on the blockchain.

[0017] Preferably, the penalty function includes a regulation weight of a default penalty associated with reputation.

[0018] The regulation weight is dynamically associated with the subject reputation score recorded by the consensus layer of the blockchain.

[0019] The updating process of the subject reputation score is as follows: the evaluation of the subject's performance is obtained, and the current reputation score and a preset learning rate are combined to perform iterative updating; the evaluation of the performance is automatically executed by the smart contract according to whether the penalty function is triggered.

[0020] Preferably, the probability evolution calculation process is as follows:

[0021] The probability of the decision scheme at the last time is obtained as the prior probability, and the newly collected evidence is obtained; a likelihood function representing the probability of observing the new evidence under the condition that the decision scheme is established is combined to perform Bayesian inference calculation to obtain the posterior probability.

[0022] The likelihood function is constructed based on statistical analysis of historical data.

[0023] Preferably, the probability decision unit is further configured to normalize the posterior probabilities of all candidate decision schemes, and set a decision scheme as a deterministic decision instruction and issue for execution when the posterior probability of the decision scheme exceeds a preset collapse threshold.

[0024] Preferably, the decision execution unit is further configured to instruct the system to bypass the probability decision unit and directly execute an adjustment scheme corresponding to the recommended adjustment amount in response to a direct adjustment signal.

[0025] Preferably, the system runs in a multi-time scale hierarchical framework including a strategy layer, a tactic layer and an operation layer.

[0026] The strategy layer is configured to set long-term supply and demand targets and a game framework at a monthly time scale.

[0027] The tactic layer is configured to plan main logistics trunk lines and inventory strategies at a weekly time scale.

[0028] The operation layer is configured to execute the equilibrium analysis output by the game equilibrium unit and the probability decision output by the probability decision unit at a daily time scale.

[0029] The present application provides an AI-based replenishment dry goods intelligent supply chain management system, which has the following improvements and advantages compared with the prior art.

[0030] 1. The scheme unifies the conflicting interests of each party into a calculable mathematical framework by constructing an utility function for each subject and solving a Nash equilibrium strategy combination. The elastic buffer threshold calculated by the game equilibrium unit is essentially the greatest common divisor recognized by all participants, which is a manifestation of global interest equilibrium. This converts the supply chain from a zero-sum game or asymmetric information confrontation environment to a win-win collaborative ecosystem, significantly enhancing the cohesion and stability of the supply chain alliance.

[0031] 2. The scheme quantifies the intangible asset of commercial credit and directly affects the economic utility of each party. A subject with high credit enjoys a lower default penalty weight in the game, equivalent to a higher decision fault tolerance rate, which is an economic incentive for long-term integrity. On the contrary, dishonest behavior will directly lead to a disadvantaged position in subsequent games. This design goes beyond traditional static contract constraints and builds a trust system that can self-regulate and promote a virtuous cycle, which has far-reaching value in maintaining long-term and stable supply chain cooperation.

[0032] 3. This scheme is not simply ignored, but the probability decision unit is enabled for careful processing; the unit is based on Bayesian inference, and the decision-making process is changed from a transient, deterministic judgment to a dynamic, evidence-driven belief updating process; the system can continuously collect new data as evidence and continuously correct the posterior probability of each candidate decision scheme; only when the credibility of a certain scheme exceeds the preset collapse threshold, the decision is finally determined; this delayed collapse mechanism effectively avoids the system making hasty and wrong decisions based on early, incomplete and noisy information, greatly improving the decision quality and reliability of the system in complex and variable market environments. BRIEF DESCRIPTION OF DRAWINGS

[0033] The application will be further explained below in conjunction with the accompanying drawings and embodiments:

[0034] Figure 1 is a flowchart of an AI-based nourishing dry goods intelligent supply chain management system of the application. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in conjunction with specific embodiments.

[0036] Embodiment 1:

[0037] Please refer to Figure 1 The application provides an AI-based nourishing dry goods intelligent supply chain management system, which comprises:

[0038] A data acquisition unit is configured to acquire historical sales data, weather forecast data, real-time traffic, cold chain sensor data and market intelligence.

[0039] An AI prediction unit is configured to perform demand and logistics path prediction based on the data acquired by the data acquisition unit, so as to generate a recommended adjustment amount.

[0040] A game equilibrium unit is configured to perform multi-agent game equilibrium analysis based on the data acquired by the data acquisition unit and a preset utility function, so as to solve an elastic buffer threshold.

[0041] A decision execution unit is configured to compare and process the recommended adjustment amount generated by the AI prediction unit and the elastic buffer threshold solved by the game equilibrium unit; when the recommended adjustment amount exceeds the elastic buffer threshold, a direct adjustment signal is generated; when the recommended adjustment amount does not exceed the elastic buffer threshold, a delayed decision signal is generated.

[0042] A probability decision unit is configured to respond to the delayed decision signal, perform probability evolution calculation on a plurality of candidate decision schemes, generate posterior probabilities of the schemes, and compare and analyze the posterior probabilities with a preset collapse threshold, so as to generate a deterministic decision instruction.

[0043] An AI-based tonic dry goods intelligent supply chain management system, the core purpose is to solve the technical contradiction between AI high-frequency prediction adjustment and physical world stable operation caused by external disturbance, through the introduction of elastic buffer and probability decision mechanism, realize the balance of supply chain dynamic adaptability and operation stability; In this embodiment, the system is deployed in a distributed computing environment composed of cloud server and edge computing node;

[0044] The overall technical process of the system starts from data collection, then AI prediction, the prediction result is not directly executed, but filtered through an elastic buffer zone determined by game equilibrium analysis, and according to the comparison result, select direct execution or start a probability decision process, finally output stable and multi-party consensus decision instruction;

[0045] The data acquisition unit aims to provide comprehensive and multi-dimensional data input for subsequent intelligent decision-making; The unit is an integrated data interface platform that collects data from different sources through API, application programming interface; Historical sales data comes from enterprise internal ERP, enterprise resource planning or CRM, customer relationship management system, to reveal the periodicity and trend of demand; Weather forecast data comes from national meteorological service agencies to warn of extreme weather that may affect logistics; Real-time traffic is obtained by calling the API of third-party map services to evaluate the real-time efficiency of transportation paths; Cold chain sensor data is uploaded in real time by Internet of Things devices deployed in transportation vehicles and warehouses to monitor the status of goods; Market intelligence is captured through web crawler technology from industry websites, social media and other channels to capture sudden events and competitor dynamics; All collected data is formatted and stored in a unified time series database;

[0046] The AI prediction unit aims to quantitatively predict future market demand and optimal logistics paths based on collected data. In this embodiment, the unit is composed of two deep learning models. The demand prediction uses an LSTM (Long Short-Term Memory) network model. In this embodiment, the LSTM model specifically uses a stacked LSTM structure with two hidden layers, each with 128 neurons. To further improve prediction accuracy, the model at the top also integrates an attention mechanism that dynamically assigns weights to different time steps in the input sequence. For example, when predicting holiday demand, it automatically gives higher attention to sales data from the same period in previous years. The model is trained in advance using historical sales data, weather data, and market intelligence data from the past three years, so it can capture complex nonlinear relationships and time dependencies and generate daily demand prediction values for a future period. The logistics path prediction uses a graph neural network model that abstracts the logistics network as a graph structure, where nodes represent warehouses or distribution centers and edges represent transportation paths. The weights of the edges are dynamically determined by real-time traffic, transportation costs, and estimated duration. When receiving new data inputs, the AI prediction unit outputs a specific recommended adjustment amount, which represents the quantitative adjustment value needed for the current operation plan to respond to predicted changes.

[0047] The game equilibrium unit aims to calculate a system flexibility range that all key participants can accept to absorb minor disturbances by simulating the interest game of each participant in the supply chain. The core of the unit is an algorithm based on the Nash equilibrium idea, which converts the implicit interest demands of each party into explicit mathematical models. Based on real-time data collected by the data collection unit and the utility functions preset by each party, the unit calculates an elastic buffer threshold. The threshold represents the maximum plan adjustment amplitude that the system can tolerate without disrupting the overall stability of the supply chain collaboration.

[0048] The decision execution unit serves as the decision arbitration hub between AI prediction and final decision instructions. The unit receives the recommended adjustment amount generated by the AI prediction unit and the elastic buffer threshold calculated by the game equilibrium unit, and compares their numerical values. When the recommended adjustment amount exceeds the elastic buffer threshold, it means that a significant event has occurred that requires an urgent response. In this case, the unit generates a direct adjustment signal. When the recommended adjustment amount does not exceed the elastic buffer threshold, it indicates that the disturbance is within the system's tolerable flexibility range. In this case, the unit generates a delayed decision signal to start the subsequent probabilistic decision process.

[0049] The probability decision unit aims to respond to the delay decision signal to process small disturbances within the elastic range in a more prudent manner, avoiding rigid deterministic response to every small fluctuation; the unit is designed as a probability evolution calculation engine based on Bayesian inference; it maintains the probability distribution of multiple candidate decision schemes; when new information is obtained by the data acquisition unit, the unit iteratively updates the posterior probability of each scheme according to the new evidence; only when the posterior probability of a scheme grows high enough, exceeding a preset collapse threshold, the unit generates a deterministic decision instruction and issues it to the supply chain execution system;

[0050] The above-mentioned units work together to build a closed-loop intelligent decision system; instead of directly coupling AI prediction results with logistics execution decisions, an elastic buffer interval and a probability decision mechanism determined by multi-agent game are established between them; the implementation of the scheme can effectively decouple the high-frequency prediction update of AI and the stable logistics execution of the physical world, while maintaining the system's high sensitivity to environmental changes, significantly enhancing the stability of operations, reducing the waste of resources caused by decision oscillation, and ultimately improving the overall efficiency and economic benefits of the supply chain;

[0051] In the prior art, the prediction results of artificial intelligence are usually directly applied to the logistics execution level; this direct coupling mechanism will cause decision oscillation when facing continuous small disturbances in the external environment, which is specifically manifested as frequent changes in transportation plans, waste of scheduling resources, and unexpected increase in operating costs;

[0052] The core difference of the present scheme is that two core units, game equilibrium unit and probability decision unit, are innovatively introduced between the AI prediction unit and the decision execution unit; the effective decoupling of the prediction layer and the execution layer is designed and implemented; instead of making a rigid response to every prediction fine-tuning, the system filters through a dynamic elastic buffer threshold determined by the game of multiple interests; the introduction of this mechanism enables the system to absorb and smooth out most of the high-frequency, low-amplitude disturbances, and only respond quickly to major events that exceed the tolerance limit of each party; this processing method mechanismally suppresses decision oscillation, ensuring the stability of transportation plans and warehouse operations in the physical world, directly bringing benefits such as reduced cold chain energy consumption, reduced product damage rate, and improved overall operational efficiency.

[0053] The game equilibrium analysis process is as follows: utility functions are constructed for the supplier agent, logistics agent and retailer agent in the supply chain; the Nash equilibrium strategy combination that makes the utility of all agents reach local optimum at the same time is found through iterative calculation; the optimal strategy adjustment amount of each agent in the Nash equilibrium strategy combination is extracted; and the minimum value in the optimal strategy adjustment amount is set as the elastic buffer threshold;

[0054] In this embodiment, the game equilibrium analysis process is concretized; the core purpose is to provide a calculable and reproducible method to convert abstract multi-party game into a specific and practical guiding flexible buffer threshold;

[0055] The utility function is constructed for the core participants in the supply chain, namely the supplier subject, the logistics subject and the retailer subject; the utility function refers to a mathematical expression to quantify the net income or satisfaction of a subject under a certain strategy selection;

[0056] The system finds the Nash equilibrium strategy combination that makes the utility of all subjects reach the local optimum through iterative calculation; Nash equilibrium refers to a strategy combination in the game, and the utility will not increase if any participant changes his strategy; in this embodiment, the solving process of Nash equilibrium is executed by an off-chain server; the server deploys gradient descent or simulated annealing algorithm to find the strategy combination that makes the utility function of all subjects reach the local optimum; after calculation, the key result of the finally calculated flexible buffer threshold will be submitted to the smart contract of the blockchain for recording and publicity; the subsequent decision execution unit will read this threshold from the smart contract to ensure its authority, transparency and non-tamperability;

[0057] After calculating the Nash equilibrium strategy combination, the system extracts the optimal strategy adjustment amount of each subject in the Nash equilibrium strategy combination; the optimal strategy adjustment amount refers to the strategy adjustment value that can maximize the utility of each subject under the Nash equilibrium state;

[0058] Based on the optimal strategy adjustment amount of each party, the system adopts the rule of taking the minimum value, and sets the minimum value as the flexible buffer threshold; the flexible buffer threshold is determined by the minimum value; for example, if the supplier can tolerate a delay of 2 hours, the logistics can tolerate a delay of 5 hours, and the retailer can only tolerate a fluctuation of 1 hour, then the final flexible buffer threshold will be set to 1 hour;

[0059] The game equilibrium analysis process defined in this way ensures that the generated flexible buffer threshold is the greatest common divisor acceptable to all key participants; the interests of all parties are unified and balanced through mathematical models, so that the threshold setting is no longer subjective, but based on a transparent, fair and quantifiable game result; this greatly enhances the universality of the decision and the execution feasibility in a multi-party cooperation environment, and guarantees the stability of the supply chain alliance;

[0060] The prior art often focuses on the performance improvement of a single link, such as maximizing the prediction accuracy or minimizing the single transportation cost, but ignores the negative impact of these local optimal solutions on other participants in the supply chain; the present scheme unifies the conflicting interests of all parties into a calculable mathematical framework by constructing an utility function for each subject and solving the Nash equilibrium strategy combination; the elastic buffer threshold calculated by the game equilibrium unit is essentially the greatest common divisor recognized by all participants, which is the embodiment of global interest equilibrium; this will change the supply chain from a zero-sum game or asymmetric information confrontation environment to a win-win collaborative ecosystem, significantly enhancing the cohesion and stability of the supply chain alliance.

[0061] The utility function components are measured in unified currency units;

[0062] The utility function includes an individual income function, an operating cost function, and a penalty function;

[0063] The individual income function is derived from the income distribution model preset in the smart contract;

[0064] The operating cost function is derived from the enterprise's own financial and operational databases;

[0065] The penalty function takes effect when the current strategy deviates from the commitment strategy published on the blockchain;

[0066] The utility function is the cornerstone of the system's game theory; it unifies the economic goals of each participant, including maximizing income, operating constraints, minimizing cost, and contract spirit, reputation punishment, into a monetary dimension; the function is not derived by the system, but is constructed based on the real financial data of each party and the preset distribution rules, and its output is the direct basis for the Nash equilibrium solution by the game equilibrium unit;

[0067] In this embodiment, the composition of the utility function is specified in detail based on the above embodiment; the purpose is to ensure the calculability of the utility function and the unity of the dimensions of each component, so that the influence of different dimensions can be weighed in a unified framework;

[0068] The utility function components are measured in unified currency units; this means that whether it is sales revenue, transportation cost or default penalty, it will be converted into the same currency unit, so that direct algebraic addition and subtraction operations can be performed, ensuring the mathematical validity of the model;

[0069] To clarify its composition, the utility function includes an individual income function, an operating cost function, and a penalty function; these three components jointly determine the net utility of a subject;

[0070] To make the construction process of utility function more clear and operable, the embodiment makes the following specific definition to each component; let any subject in the supply chain The net utility when taking strategy , for example, selecting a specific transportation time, storage temperature, etc. is The calculation formula is:

[0071]

[0072] Among them: : any subject in the supply chain The net utility when taking strategy ; i: represents any subject in the supply chain; : the strategy taken by subject i, for example, selecting a specific transportation time, storage temperature, etc. : operation cost function, representing the cost required by subject i to execute strategy ; i: represents any subject in the supply chain; : the cost required by subject i to execute strategy ; i: represents any subject in the supply chain; : penalty function, used to punish the behavior of deviating from the committed strategy; Individual income function, representing the expected income of subject i when taking strategy ; i: represents any subject in the supply chain;

[0073] Individual income function : the function quantifies the expected income of subject i; the form can be defined as:

[0074]

[0075] In the formula, is the preset, unalterable income distribution ratio of subject i in the smart contract; is the total value of the batch goods in the market; is a function of expected loss rate related to strategy , for example, if strategy contains transportation time, the longer the transportation time, the higher the loss rate; Operation cost function

[0076] : the function represents the cost required by subject i to execute strategy ; the data is derived from enterprise financial database and can be modeled as:

[0077]

[0078] Among them, is the fixed cost, such as equipment depreciation, is the variable cost related to strategy​​​ Directly related variable costs, such as fuel costs, cold chain energy consumption, etc. Variable costs are policy variables, such as transportation distance, refrigeration power functions;

[0079] Penalty function : Function for punishing behavior deviating from commitment; Set subject Commitment strategy published on the blockchain is The penalty function can be a quadratic function related to the degree of deviation:

[0080]

[0081] Where, is the adjustment weight associated with the subject's credit points; is the basic penalty coefficient set for the subject, representing its benchmark cost of default, with the dimension of monetary units, representing the basic penalty amount per unit of deviation; is the current strategy The Euclidean distance between the commitment strategy , to quantify the severity of deviation, ensure that minor deviations are less punished, and major deviations will lead to exponential growth in the cost of punishment; To ensure the effectiveness of the Euclidean distance calculation, each component of the strategy vector and with physical units must be pre-processed for dimensionless treatment; For example, each component can be normalized by dividing by its corresponding standard value or allowed fluctuation range; After this processing, becomes a dimensionless pure number, representing the relative severity of policy deviation;

[0082] Individual profit function, the purpose is to quantify the economic return that the subject can obtain in a specific supply chain activity; The source of the function is the profit distribution model preset in the smart contract; The preset profit distribution model refers to a rule that is determined in advance by the parties of the supply chain and is fixed in the blockchain in the form of code; The individual profit function will calculate the expected income of each subject under the rules according to the real-time value of goods and the estimated loss rate;

[0083] Operating cost function, the purpose is to quantify the cost of the subject for executing its strategy; The data source of the function comes from the enterprise's own financial and operational database;

[0084] A penalty function is configured to impose economic constraints on behaviors deviating from the commitment, so as to maintain the contract spirit of the supply chain; the function is triggered when the current strategy deviates from the commitment strategy published on the blockchain; the commitment strategy refers to the operation plan published and jointly recognized by each subject on the blockchain in advance; when the actual executed strategy of a subject deviates from the commitment, the penalty function is triggered to deduct the corresponding penalty amount from the total utility;

[0085] Through the above fine definition of the utility function, the embodiment constructs a highly structured and quantified game model; the multiple factors such as revenue, cost and reputation in business activities are successfully unified into a single mathematical framework measured by currency; this makes the trade-off of the interests of each party accurate, transparent and automatically calculable, providing a solid mathematical foundation for finding the Nash equilibrium solution that truly reflects the core demands of each party, thereby improving the accuracy and fairness of the game equilibrium analysis.

[0086] The penalty function includes a regulation weight of the default penalty associated with the reputation;

[0087] The regulation weight is dynamically associated with the subject reputation score recorded by the consensus layer of the blockchain;

[0088] The updating process of the subject reputation score is as follows: the evaluation of the subject's performance is obtained, and the current reputation score and the preset learning rate are combined to update iteratively; the evaluation of the performance is automatically executed by the smart contract according to whether the penalty function is triggered;

[0089] The embodiment further optimizes the internal mechanism of the penalty function on the basis of the above embodiment; the core purpose is to introduce a dynamic reputation mechanism that is linked to the historical behavior of the subject, so that the penalty is no longer static, but self-adjusting, thereby more effectively encouraging long-term cooperation and honest performance;

[0090] The penalty function includes a regulation weight of the default penalty associated with the reputation; the regulation weight refers to a coefficient for amplifying or reducing the basic penalty;

[0091] The regulation weight is dynamically associated with the subject reputation score recorded by the consensus layer of the blockchain; the subject reputation score refers to a numerical value stored on the blockchain for quantitatively evaluating the historical performance of a subject; the non-tamperable nature of the consensus layer of the blockchain ensures the credibility of the reputation score;

[0092] The updating process of the subject reputation score is designed as an iterative learning process; the system obtains the evaluation of the subject's performance, and combines the current reputation score and the preset learning rate to update iteratively; the specific mathematical model of the process is as follows:

[0093]

[0094] where, is the subject the new reputation score at the next time point; : refers to the subject i; : represents the next time point; : represents the current time point; : the evaluation of the subject's performance at the current time point t, which is a binary evaluation result, for example, 1 for performance and 0 for default; is the subject the reputation score at the current time point, which comes from the historical record on the blockchain, and the value range is limited to an interval such as ; is the preset learning rate, which is a constant between 0 and 1, and the value is set according to the system's convergence and response speed analysis of historical data, balancing stability and sensitivity;

[0095] learning rate The value of the learning rate can be optimized and determined by backtesting the historical performance data of the supply chain in the past 1-2 years; the goal of backtesting is to find an optimal value, so that under the learning rate, the overall game framework composed of the reputation system and the punishment mechanism can produce the lowest total default cost of the supply chain; the process can be formalized as an optimization problem:

[0096]

[0097] where, : the optimal learning rate λ value; : indicates the operation of finding the parameter value that can make the objective function, the total penalty cost, reach the minimum value; : time point, from 1 to T; is the length of historical data, is the total penalty cost at time point , generated by the reputation model with parameter ; by simulating the historical evolution under different values, the that minimizes the total penalty cost can be selected as the final configuration parameter of the system; : the state of the reputation model with parameter λ at time point t;

[0098] is the evaluation of the performance, which is a binary evaluation result of the subject's behavior;

[0099] The evaluation of the performance of the contract is automatically executed by the smart contract according to whether the penalty function is triggered; the smart contract is deployed on the blockchain, and when it detects that the behavior of the subject deviates from the commitment strategy, causing the penalty function to be triggered, the smart contract automatically assigns the value of 0 to the current performance of the subject corresponding to the subject; otherwise, it is assigned a value of 1; this process is fully automated, avoiding human intervention;

[0100] The embodiment introduces a penalty weight dynamically associated with the reputation points to establish an adaptive incentive and constraint mechanism; for a subject with high reputation, the adjustment weight of the default penalty will be reduced accordingly, which is an incentive for long-term honest behavior; on the contrary, a subject with low reputation will face more severe punishment; this design not only effectively deters opportunistic behavior, but also encourages all participants to focus on long-term cooperation and actively maintain their reputation, thereby greatly enhancing the cohesion and long-term stability of the entire supply chain alliance, forming a virtuous cycle of trust ecology;

[0101] The iterative model of reputation points is an exponential moving average model, which describes how a subject's reputation evolves with its behavior; the future reputation of a subject is a weighted average of its historical reputation and the latest behavior performance; the learning rate determines the memory length of the system; a smaller value means that the system places more emphasis on long-term consistent performance, while a larger value makes the reputation more sensitive to recent behavior;

[0102] By introducing a penalty function dynamically associated with the reputation points of the subject, the present scheme quantifies the intangible asset of business reputation and directly affects the economic utility of each party; a subject with high reputation enjoys a lower default penalty weight in the game, which is equivalent to a higher decision fault tolerance rate, which is an economic incentive for long-term honest behavior; on the contrary, dishonest behavior will directly lead to its disadvantaged position in subsequent games; this design goes beyond traditional static contract constraints and builds a trust system that can self-regulate and promote a virtuous cycle, which has far-reaching value in maintaining long-term and stable supply chain cooperation.

[0103] Embodiment 2

[0104] The probability evolution calculation process is as follows:

[0105] The probability of the decision scheme at the previous time is obtained as the prior probability, and the newly collected evidence is obtained; the likelihood function representing the probability of observing new evidence under the condition that the decision scheme is established is combined to perform Bayesian inference calculation to obtain the posterior probability;

[0106] The likelihood function is constructed based on statistical analysis of historical data;

[0107] In this embodiment, the probability evolution calculation process is described in detail; the purpose is to provide a rigorous mathematical method based on Bayesian inference, which is used to dynamically update the credibility of each decision scheme in the case of incomplete information, so as to realize prudent delay decision;

[0108] The core logic of the process is to update the probability of the decision scheme sequentially; the initial step is that the system obtains the probability of the decision scheme at the last time as the prior probability, and obtains the new evidence collected at the current time; the prior probability refers to the initial judgment of the possibility of the establishment of a certain decision scheme before obtaining new information, and the value is derived from the result of the last calculation period; the new evidence refers to the latest information related to the decision obtained by the data acquisition unit at the current time;

[0109] The system combines the likelihood function representing the probability of observing new evidence under the condition that the decision scheme is established to perform Bayesian inference calculation to obtain the posterior probability; the posterior probability refers to the possibility of the establishment of the decision scheme after considering the new evidence; the calculation follows the Bayesian formula:

[0110]

[0111] Among them, is the posterior probability; is a normalization constant, which ensures that the sum of the posterior probabilities of all candidate decision schemes is equal to 1; is the likelihood function, which measures the probability of observing new evidence if the decision is established; is the prior probability; : a certain candidate decision scheme; : new evidence collected at the current time t; : specific new evidence observed at the current time t; : the evidence set at the last time;

[0112] The likelihood function is constructed based on statistical analysis of historical data;

[0113] In this embodiment, the likelihood function model is realized by constructing a Gaussian mixture model; the system extracts a large number of decision scheme, subsequent evidence data pairs from the historical database, and trains the GMM according to these data; after training, for any given decision scheme and newly observed evidence , the GMM can calculate the probability density of observing the evidence under the condition that the decision scheme is established, that is, the likelihood value ;

[0114] The preset likelihood function model refers to a probability model obtained by mining massive amounts of historical data through machine learning methods; when new evidence emerges, the corresponding likelihood probability value can be retrieved from the model.

[0115] By employing a probabilistic evolutionary computation process based on Bayesian inference, this embodiment transforms the decision-making process from a black-and-white instantaneous judgment into a dynamic, evidence-driven belief update process. It allows the system to maintain a probabilistic superposition state of the decision when information is insufficient, and gradually increases confidence in a certain optimal option as new information is continuously injected. This mechanism greatly improves the robustness and prudence of the decision, effectively avoiding the risk of making hasty or erroneous decisions due to early and insufficient information.

[0116] The probabilistic decision unit is also used to: normalize the posterior probabilities of all candidate decision schemes; and when the posterior probability of any decision scheme exceeds a preset collapse threshold, set the decision scheme as a deterministic decision instruction and issue it for execution.

[0117] In this embodiment, the function of the probabilistic decision unit is further explained; the purpose is to define the explicit triggering conditions for decision collapse, that is, how and when a probabilistic state with multiple possibilities coexisting is transformed into a single, deterministic instruction that must be executed.

[0118] The probabilistic decision unit is also used to: normalize the posterior probabilities of all candidate decision options; normalization is a mathematical operation that aims to ensure that the sum of the posterior probabilities of all candidate decision options is exactly equal to 1; this step is a basic requirement of probability theory and ensures the effectiveness of the probability distribution.

[0119] The core function of the unit lies in judging and executing the collapse process of decisions; when the posterior probability of any decision exceeds a preset collapse threshold, the decision is set as a deterministic decision instruction and executed; the collapse threshold is denoted as... This is a preset value; the technical principle behind setting the preset collapse threshold is based on risk cost assessment, and the value can be quantified by statistical analysis of the risk-reward ratio of historical cases.

[0120] collapse threshold The setting is based on a quantitative comparison of the cost of making a decision and the cost of delaying the decision; for any candidate decision option The system assesses the expected cost of immediately implementing it. Compared to the expected costs of continuing to wait and gather more information ;

[0121] Execution cost This includes the direct operational costs of implementing the plan, as well as the risk costs resulting from a flawed plan; assuming... Given its posterior probability, the expected cost is: ;

[0122] Delay costs This refers to the potential loss of opportunities or accumulation of risks due to failure to take timely action; the value can be derived from historical data statistics.

[0123] The system will continue to iterate and calculate until a solution is found. posterior probability This ensures that the expected execution cost is lower than the delay cost for the first time, i.e., the following condition is met:

[0124]

[0125] in, Confirm the plan The cost of correct and successful execution; Due to the plan The risk and cost of errors; The expected cost of delaying decision-making and continuing to wait for information gathering; by solving the critical probability in the above inequality. The collapse threshold can then be obtained. This approach ensures that decisions are triggered based on clear economic trade-offs, rather than subjective settings.

[0126] When the posterior probability of a certain solution At this point, the system considers the evidence for the plan to be sufficient and the uncertainty to be reduced to an acceptable level, because the plan is collapsed from the probabilistic superposition state and becomes a clear instruction that needs to be executed immediately;

[0127] The implementation provides a clear termination and output mechanism for the probabilistic decision-making process; normalization ensures the logical completeness of probability calculation, while the introduction of the collapse threshold finds an adjustable balance between the prudence and timeliness of decision-making; enabling the system to flexibly adjust the deterministic triggering conditions of its decision-making according to the importance and risk level of the task, avoiding missed opportunities due to hesitation and preventing rash actions due to insufficient evidence, thus significantly improving the intelligence and practicality of the entire decision-making system.

[0128] The decision robustness in an uncertain environment is improved; in the face of slight disturbance in the elastic buffer interval, the scheme is not simply ignored, but the probability decision unit is enabled for careful handling; the unit is based on Bayesian inference, and the decision process is changed from a transient and deterministic judgment to a dynamic and evidence-driven belief updating process; the system can continuously collect new data as evidence and continuously correct the posterior probability of each candidate decision scheme; only when the credibility of a scheme exceeds the preset collapse threshold, the decision is finally determined; this delayed collapse mechanism effectively avoids the system making hasty and wrong decisions based on early and incomplete noisy information, greatly improving the decision quality and reliability of the system in a complex and variable market environment.

[0129] Embodiment 3

[0130] The decision execution unit is also used to: in response to the direct adjustment signal, instruct the system to bypass the probability decision unit and directly execute the adjustment scheme corresponding to the recommended adjustment amount;

[0131] In this embodiment, the function of the decision execution unit is supplemented; the purpose is to provide an emergency response procedure for the system to bypass and directly handle major disturbances that are beyond the range of regular fluctuations;

[0132] The decision execution unit is also used to: in response to the direct adjustment signal, instruct the system to bypass the probability decision unit and directly execute the adjustment scheme corresponding to the recommended adjustment amount; the direct adjustment signal is an internal control signal generated by the decision execution unit itself when the recommended adjustment amount predicted by AI is greater than the elastic buffer threshold calculated by game equilibrium; when the signal is triggered, it indicates that the external environment has changed dramatically, and the impact has exceeded the range that the system can smooth and absorb through the delay and buffer mechanism determined by multi-party game; in this case, the system logic will switch, and the recommended adjustment amount will no longer be sent to the probability decision unit, but will be directly converted into an adjustment scheme and issued to the relevant execution system;

[0133] The dual-mode operation capability ensures that the system maintains operational economy through the buffer mechanism under regular disturbance, and ensures the resilience and timeliness of the supply chain through the emergency response procedure when major shocks are encountered, thereby greatly enhancing the overall survival ability and adaptability of the system in a complex and variable environment.

[0134] Embodiment 4

[0135] Running in a hierarchical framework of multiple time scales, the hierarchical framework includes a strategic layer, a tactical layer and an operational layer;

[0136] The strategic layer is used to set long-term supply and demand targets and game frameworks at a monthly time scale;

[0137] The tactical layer is used for planning main logistics lines and inventory strategies on a weekly time scale;

[0138] The operational layer is used for executing the equilibrium analysis output by the game equilibrium unit and the probability decision output by the probability decision unit on a daily time scale;

[0139] In this embodiment, the entire system is arranged in a hierarchical framework operating on multiple time scales; the purpose is to decouple decision-making problems of different frequencies and different levels, so that the stability of high-level strategies and the flexibility of low-level operations can be considered; the framework includes a strategic layer, a tactical layer and an operational layer;

[0140] The strategic layer is used for setting long-term supply and demand targets and game frameworks on a monthly time scale; the decision frequency of the strategic layer is the lowest, and the macro direction is concerned; at this level, the manager will set the overall supply and demand balance target for the next month according to the annual sales target and market trend; more importantly, the core rules in the game framework, such as the constraint range and guiding principle for setting the basic penalty coefficient of each subject, are also defined and adjusted at this level;

[0141] The correlation model of the penalty weight establishes a direct negative correlation between reputation and economic punishment; the system significance is to convert abstract reputation points into actual influence in the game; the higher the reputation points of a subject, the smaller the cost of deviating from the commitment in the game, and vice versa; the model is a bridge connecting the reputation system and the game system;

[0142] The tactical layer is used for planning main logistics lines and inventory strategies on a weekly time scale; the decision of the tactical layer is connected with the upper and lower levels, and the frequency is moderate; according to the monthly target issued by the strategic layer, combined with the medium-term weather forecast and known information such as promotion activities, the main transportation network path for the next week is planned, and the inventory allocation strategy between the central warehouse and the regional warehouse is determined; in addition, the system parameters related to reputation, such as the learning rate of reputation update, are also configured at this level;

[0143] The operational layer is used for executing the equilibrium analysis output by the game equilibrium unit and the probability decision output by the probability decision unit on a daily time scale; the operational layer is the front line of decision execution, and the frequency is the highest; it completely operates according to the core logic described in the present application: every day or even every hour, the system continuously performs data collection, AI prediction, game equilibrium analysis, threshold comparison and possible probability decision; all daily, dynamic and high-frequency adjustments are completed at this level;

[0144] By constructing this hierarchical decision-making framework, the embodiment successfully decomposes the complex and huge supply chain management problem into three different but interrelated sub-problems; the stability of the strategic layer provides a clear direction and game rules for the whole system; the tactical layer converts the strategic intention into an operational weekly plan; the operation layer efficiently handles the daily high-frequency disturbances through its unique elastic buffer and probabilistic decision-making mechanism, while protecting the tactical and strategic level plans from being disturbed by these small fluctuations; this hierarchical governance structure realizes the balance between macro stability and micro flexibility, greatly improving the management efficiency and the overall robustness of the system;

[0145] The evolutionary model of probabilistic decision-making is the direct application of Bayes' theorem in decision science; it describes the updating rule of the system's belief: the updated belief of a certain decision option, the posterior probability, is proportional to the product of the initial belief, the prior probability, and the support degree of the new evidence for the option, the likelihood function; the implementation of the formula enables the decision-making process to have the ability of logical reasoning and evidence accumulation;

[0146] Through the organic combination of the above mechanisms and models, the present scheme constructs an intelligent supply chain management system that can adapt to major environmental changes and resist daily high-frequency disturbances, achieving a delicate balance between dynamic adaptability and operational stability that existing technologies have not achieved, and having significant technical progress and practical value.

[0147] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. An AI-based tonic dry goods intelligent supply chain management system, characterized by, include: The data acquisition unit is used to collect historical sales data, weather forecast data, real-time traffic conditions, cold chain sensor data, and market intelligence. The AI ​​prediction unit is used to predict demand and logistics routes based on data collected by the data acquisition unit, in order to generate suggested adjustments. The game equilibrium unit is used to perform multi-agent game equilibrium analysis based on data collected by the data acquisition unit and a preset utility function, so as to calculate the elastic buffer threshold. The decision execution unit is used to compare the suggested adjustment amount generated by the AI ​​prediction unit with the elastic buffer threshold calculated by the game equilibrium unit; when the suggested adjustment amount exceeds the elastic buffer threshold, a direct adjustment signal is generated; when the suggested adjustment amount does not exceed the elastic buffer threshold, a delayed decision signal is generated. The probabilistic decision unit is used to respond to delayed decision signals, perform probabilistic evolution calculations on multiple candidate decision schemes to generate the posterior probability of each scheme, and compare and analyze the posterior probability with a preset collapse threshold to generate a deterministic decision instruction. The game equilibrium analysis process is as follows: construct utility functions for suppliers, logistics providers, and retailers in the supply chain; find the Nash equilibrium strategy combination that makes the utility of all entities reach local optimum simultaneously through iterative calculation; extract the optimal strategy adjustment amount for each entity in the Nash equilibrium strategy combination; and set the minimum value among the optimal strategy adjustment amounts as the elastic buffer threshold.

2. The AI-based intelligent supply chain management system for nourishing dried goods according to claim 1, characterized in that, The components of the utility function are measured in a uniform monetary unit; Utility functions include individual benefit functions, operating cost functions, and penalty functions; The individual revenue function is derived from the revenue distribution model pre-set in the smart contract; The operating cost function uses data sourced from the company's own financial and operational databases. The penalty function takes effect when the current policy deviates from the committed policy published on the blockchain.

3. The AI-based intelligent supply chain management system for nourishing dried goods according to claim 2, characterized in that, The penalty function includes a weighted adjustment for default penalties that are associated with reputation; Adjust the weights and dynamically link them to the subject's reputation score recorded by the blockchain consensus layer; The process of updating the subject's reputation score is as follows: obtain the evaluation of the subject's performance behavior, and iterate and update it by combining the current reputation score with the preset learning rate; the evaluation of the performance behavior is automatically executed by the smart contract based on whether the penalty function is triggered.

4. The AI-based intelligent supply chain management system for nourishing dried goods according to claim 1, characterized in that, The probability evolution calculation process is as follows: Obtain the probability of the decision scheme at the previous time step as the prior probability, and obtain the new evidence currently collected; combine the likelihood function representing the probability of observing new evidence under the condition that the decision scheme is valid, and perform Bayesian inference calculation to obtain the posterior probability. The likelihood function is constructed based on statistical analysis of historical data.

5. The AI-based intelligent supply chain management system for nourishing dried goods according to claim 4, characterized in that, The probabilistic decision unit is also used to: normalize the posterior probabilities of all candidate decision schemes; and when the posterior probability of any decision scheme exceeds a preset collapse threshold, set the decision scheme as a deterministic decision instruction and issue it for execution.

6. The AI-based intelligent supply chain management system for nourishing dried goods according to claim 1, characterized in that, The decision execution unit is also used to: in response to a direct adjustment signal, instruct the system to bypass the probabilistic decision unit and directly execute the adjustment scheme corresponding to the suggested adjustment amount.

7. The AI-based intelligent supply chain management system for nourishing dried goods according to claim 1, characterized in that, It operates within a multi-timescale hierarchical framework, which includes a strategic layer, a tactical layer, and an operational layer; The strategic layer is used to set long-term supply and demand goals and game theory frameworks on a monthly timescale. The tactical layer is used to plan major logistics routes and inventory strategies on a weekly timescale. The operation layer is used to perform equilibrium analysis output by the game equilibrium unit and probabilistic decisions output by the probabilistic decision unit on a daily time scale.

Images