Dynamic assessment and intelligent early warning system for cross-border suppliers based on performance of performance

Abstract

The application discloses a kind of based on performance of compliance cross-border supplier dynamic evaluation and intelligent early warning system, belong to supply chain management and artificial intelligence technical field.System includes five layers architecture of perception layer, cognitive layer, game layer, execution layer and evolution layer.Perception layer constructs supplier behavior fingerprint vector by multi-source data;Cognitive layer is based on psychological account model, toughness topological analysis and multi-objective evaluation, quantifies supplier implicit psychological variable, network vulnerability and comprehensive performance;Game layer generates dynamic contract using VCG reverse bidding mechanism, and recommends multi-objective intervention strategy;Execution layer is based on blockchain smart contract and automatically executes performance reward and punishment, and adaptively reconfigures supply chain resource allocation;Evolution layer realizes system self-evolution by meta-learning and multi-agent reinforcement learning.The application solves the problems of single dimension of traditional supplier evaluation, passive early warning, supply chain topology rigidity and the like, and significantly improves the toughness, efficiency and intelligent level of supply chain.

Description

Technical Field

[0001] This invention belongs to the field of supply chain management and artificial intelligence technology, specifically relating to a dynamic evaluation and intelligent early warning system for cross-border suppliers based on performance. Background Technology

[0002] With the deepening of globalization, cross-border e-commerce companies face unprecedented challenges in managing their cross-border suppliers. Suppliers' fulfillment capabilities directly impact a company's operational efficiency, cost control, and customer satisfaction. While existing supply chain management systems (such as Mate and Cloud Chain) have achieved basic functions like order synchronization, product sourcing, and basic pricing, they suffer from fundamental deficiencies in supplier performance evaluation and risk warning.

[0003] 1. Single and static evaluation dimensions: Traditional evaluation systems mainly rely on historical transaction scores or data self-reported by suppliers, failing to capture the supplier's implicit behavioral characteristics, such as risk preferences, stress response patterns, and cooperative habits—dynamic psychological variables. This static evaluation method leads to lagging and distorted results, making it difficult to truly reflect the supplier's willingness to fulfill obligations and potential risks. For example, a supplier who appears to have good performance may exhibit drastically different behaviors when the external environment changes.

[0004] 2. The early warning mechanism is passive and lacks intervention capabilities: Existing systems typically trigger early warnings based on simple thresholds (such as more than 3 delays), and can only send reminders after an early warning, lacking an effective game-theoretic mechanism to mitigate risks. There is a disconnect between early warning and decision-making, failing to form a closed-loop management system. When risk signals appear, enterprises often can only respond passively, lacking proactive intervention measures to reverse the situation.

[0005] 3. Rigid Supply Chain Topology: Supplier knowledge graphs only record static cooperative relationships, making it impossible to dynamically assess the resilience of the supply chain network, let alone automatically reconstruct the supply chain topology to cope with disruptions when risks occur. Once a critical node fails, enterprises often find themselves in a passive position, suffering losses such as order delays and soaring costs.

[0006] 4. Unresolved Conflicts Among Multiple Objectives: In actual procurement decisions, objectives such as cost, timeliness, quality, and risk often conflict with each other. Existing systems are unable to optimize these multiple objectives, leading to suboptimal procurement decisions. For example, pursuing low costs may result in the introduction of high-risk suppliers, or prioritizing timeliness may sacrifice cost control, making it difficult to achieve a balance among multiple objectives.

[0007] 5. Lack of Credibility in Commitments: Verbal commitments from suppliers cannot be verified, and there is a lack of automated reward and punishment mechanisms. Breach of contract often requires manual intervention, which is not only inefficient but also prone to disputes, increasing uncertainty in supply chain management.

[0008] Therefore, there is an urgent need for an intelligent system to accurately assess cross-border suppliers, provide dynamic early warnings, and enable adaptive intervention, thereby improving the overall resilience and operational efficiency of the supply chain. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a dynamic evaluation and intelligent early warning system for cross-border suppliers based on performance, including a five-layer architecture of perception layer, cognition layer, game layer, execution layer and evolution layer, so as to solve the technical problems mentioned in the background art above.

[0010] To achieve the above objectives, this invention proposes a dynamic evaluation and intelligent early warning system for cross-border suppliers based on performance, comprising:

[0011] The perception layer is used to collect behavioral data of suppliers in multi-source business systems and generate supplier behavior fingerprint vectors through feature extraction and dimensionality reduction encoding.

[0012] The cognitive layer, connected to the perception layer, is used to infer the supplier's psychological variables based on the behavioral fingerprint vector using a probabilistic graphical model, analyze the supplier's knowledge graph using a graph neural network to assess supply chain resilience, and use multiple prediction models to conduct multi-objective evaluation of the supplier's performance and generate a comprehensive score.

[0013] The game theory layer, connected to the cognitive layer, is used to conduct supplier bidding and generate dynamic contracts based on the psychological variables, resilience assessment results and comprehensive scores, and to estimate the effect of intervention strategies and generate multi-objective optimization strategies using causal inference models.

[0014] The execution layer, connected to the game layer, is used to automatically execute performance rewards and penalties based on blockchain smart contracts and oracle networks, and to model the supply chain restructuring problem as a combinatorial optimization model to obtain a resource scheduling scheme.

[0015] An evolutionary layer, connected to the execution layer, is used to enable rapid adaptation to new suppliers using a meta-learning framework and to collaboratively optimize the model parameters of each layer of the system using multi-agent reinforcement learning.

[0016] As a further option, the perception layer includes:

[0017] The multi-source data access unit is used to access data from the enterprise business system in real time via API interface, and to clean, align and normalize the data.

[0018] The behavioral feature extraction unit is used to extract time series features, stress response features, social network features, and cost sensitivity features from the processed data;

[0019] The fingerprint encoding and storage unit is used to concatenate the extracted features into an original feature vector, input it into the autoencoder for dimensionality reduction, and output a fixed-dimensional behavioral fingerprint vector for storage.

[0020] As a further option, the module in the cognitive layer that uses a probabilistic graphical model to infer the supplier's psychological variables is specifically used for:

[0021] Risk preference coefficients are obtained by fitting behavioral fingerprint vectors using cumulative prospect theory.

[0022] The cooperative inertia coefficient was obtained using a logistic regression model.

[0023] A dynamic exponential decay model is used to update the trust level, which decays over time and is adjusted by performance behavior.

[0024] Update your psychological balance by accumulating emotional capital;

[0025] The risk preference coefficient, cooperation inertia coefficient, trust level, and psychological balance are used as the hidden states of the hidden Markov model. The behavioral fingerprint vector is used as the observation value to infer the current psychological state in real time and to provide early warning of abnormal changes in psychological variables.

[0026] As a further option, the module in the cognitive layer that uses a graph neural network to analyze the supplier knowledge graph to assess supply chain resilience is specifically used for:

[0027] Use a graph database to store a knowledge graph containing suppliers, logistics providers, warehouse nodes, transaction relationships, and logistics path edges;

[0028] By simulating node removal, we can compute the network connectivity resilience value, redundancy index, and betweenness centrality of nodes.

[0029] The graph neural network is used to encode the network state and output the network's resilience score and the vulnerability probability of the nodes.

[0030] Reinforcement learning models are used to generate reconstruction strategies for vulnerabilities. Actions include replacing suppliers, adding connections, or adjusting order allocation. The reward is the increase in resilience score after reconstruction.

[0031] As a further option, the module in the cognitive layer that uses multiple prediction models to conduct multi-objective evaluation of the supplier's performance and generate a comprehensive score is specifically used for:

[0032] For each supplier, a time-series forecasting model is used to predict the cost of future orders as the cost target, a sequence model is used to predict the probability of order delays as the timeliness target, a machine learning model is used to predict the product quality pass rate as the quality target, and a risk preference coefficient is used as the risk target.

[0033] Each sub-target prediction model shares underlying features and is jointly trained through multi-task learning;

[0034] A comprehensive score is obtained by normalizing each sub-objective using a user preference vector and then summing the results using a weighted summation. The preference vector is then optimized using an online learning algorithm.

[0035] As a further option, the module in the game layer that uses an auction mechanism for supplier bidding is specifically used for:

[0036] When an early warning is triggered or a restructuring is required, a reverse auction is initiated, and a tender document containing the required quantity, delivery date, and quality requirements is issued.

[0037] Collect the quotations and quantities offered by suppliers, solve the combinatorial optimization problem with the goal of maximizing social welfare, and determine the winning bid set;

[0038] For the winning supplier, the payment price is the socially optimal total welfare minus the optimal total welfare after removing the supplier, plus the supplier's bid.

[0039] Multi-agent reinforcement learning is used to simulate supplier pricing strategies and predict the equilibrium pricing distribution.

[0040] As a further option, the module for generating dynamic contracts in the game layer is specifically used for:

[0041] Dynamic terms are generated based on the contract template. The price adjustment terms adopt a function of the benchmark price multiplied by the rate of change of the raw material price index, and the adjustment coefficient is determined by the supplier's risk preference coefficient.

[0042] The performance incentive clauses dynamically adjust the reward and penalty coefficients based on the level of trust.

[0043] The contract content is populated using a template engine and deployed to the blockchain to generate a smart contract.

[0044] As a further option, the module in the game layer that uses a causal inference model to estimate the effect of the intervention strategy and generate a multi-objective optimization strategy is specifically used for:

[0045] Define the policy space, which includes both soft and hard interventions;

[0046] Using causal forests to model historical intervention data, we learn the causal effects of different intervention strategies on the supplier's subsequent performance, and output the estimated value and confidence interval of the change in performance after the intervention.

[0047] A multi-objective optimization algorithm is used to weigh the costs, effects, and risks, generate a Pareto optimal policy set, and select a recommended policy based on user preferences.

[0048] As a further option, the module in the execution layer that automatically executes performance rewards and penalties based on blockchain smart contracts and oracle networks is specifically used for:

[0049] Deploy multiple decentralized oracle nodes to obtain performance data from different data sources, and use a threshold signature mechanism to aggregate and sign the data before submitting it to the smart contract;

[0050] The smart contract receives the performance status reported by the oracle and automatically executes payment, margin deduction and mental balance update according to preset conditions;

[0051] The results of rewards and punishments are transmitted to the mental account chain through a cross-chain bridge, enabling cross-chain data synchronization.

[0052] As a further option, the module in the execution layer that models the supply chain restructuring problem as a combinatorial optimization model to obtain a resource scheduling scheme is specifically used for:

[0053] Receive the refactoring strategy, parse it into specific operation instructions, and orchestrate calls to the APIs of various business systems through the workflow engine;

[0054] The reconstruction problem is modeled as a multi-product flow problem. With the goal of minimizing the total cost, and considering constraints such as supplier capacity, warehouse capacity, and transportation capacity, a column generation algorithm is used to solve the optimal order allocation and logistics scheduling scheme.

[0055] If a step fails during execution, a rollback is automatically triggered and a failure event is sent to the game layer.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] 1. By using the mental accounting model to reveal implicit psychological variables such as suppliers' risk preferences and cooperation inertia, risk warnings can be issued earlier and the accuracy of assessments can be greatly improved.

[0058] 2. The reverse bidding mechanism uses VCG to incentivize suppliers to report their true costs, thereby reducing procurement costs and improving the contract fulfillment rate.

[0059] 3. Resilient topology analysis assesses network vulnerabilities in real time and automatically generates reconstruction plans, reducing the risk of supply chain disruption and increasing the success rate of reconstruction execution.

[0060] 4. Blockchain-based smart contracts automatically execute rewards and penalties, reducing the number of disputes and improving settlement efficiency.

[0061] 5. The meta-learning framework shortens the evaluation and adaptation period for new suppliers, and the system has the ability to quickly start up. Attached Figure Description

[0062] Figure 1 This is a diagram showing the five-layer architecture and module relationships of the system of the present invention.

[0063] Figure 2 A schematic diagram of the process for building supplier behavior fingerprints.

[0064] Figure 3 This is a schematic diagram of the mental accounting model (HMM).

[0065] Figure 4 This is a flowchart of the resilient topology analysis and reconstruction process.

[0066] Figure 5 This is a schematic diagram of multi-objective assessment and conflict resolution.

[0067] Figure 6 This is a schematic diagram of the reverse bidding game and smart contract generation process.

[0068] Figure 7 This is a schematic diagram of the self-evolutionary module meta-learning and reinforcement learning framework. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] The system of this invention adopts a five-layer collaborative architecture: perception layer, cognition layer, game theory layer, execution layer, and evolution layer. Each layer contains multiple modules, which communicate asynchronously with each other through an event bus (such as Apache Kafka), forming a closed-loop intelligent system.

[0071] I. System Overall Architecture and Data Flow

[0072] like Figure 1 As shown, the overall system architecture includes the following core data flows:

[0073] The perception layer outputs the supplier behavior fingerprint vector to the cognition layer;

[0074] The cognitive layer outputs mental account variables, resilience scores, and multi-objective comprehensive scores to the game theory layer.

[0075] The game layer outputs intervention strategies, reverse bidding results, and dynamic contract parameters to the execution layer;

[0076] The execution layer outputs execution results and transaction records to the evolution layer;

[0077] The evolution layer outputs the updated model parameters to each layer module.

[0078] The internal modules of each layer exchange intermediate data through an event bus to ensure that the system has the characteristics of high cohesion and low coupling.

[0079] II. Module Detailed Design and Implementation

[0080] (a) Perception Layer: Supplier Behavior Fingerprint Construction Module

[0081] like Figure 2 As shown, this module is responsible for collecting comprehensive behavioral data from suppliers during cross-border transactions, constructing unique behavioral fingerprint vectors for subsequent psychological modeling and evaluation. It includes the following units:

[0082] 1. Multi-source data access unit

[0083] Real-time access to data from internal enterprise business systems such as ERP (Enterprise Resource Planning), WMS (Warehouse Management System), TMS (Transportation Management System), and CRM (Customer Relationship Management) via standard API interfaces, specifically including:

[0084] Order response time (time from order receipt to confirmation), fulfillment delay (difference between actual delivery / arrival time and promised time), communication records (emails, chat texts), price fluctuations (historical price sequence), and external shock events (such as port blockades or sharp exchange rate fluctuations).

[0085] After data is input, it undergoes cleaning, missing value imputation, time alignment, and normalization to form structured time-series data.

[0086] 2. Behavioral Feature Extraction Unit

[0087] Time rhythm feature extraction: Perform multi-scale entropy analysis on the supplier response time series to extract the regularity features of the series, including the mean entropy value, the slope of change, the peak position, etc., to characterize the supplier's time behavior pattern.

[0088] Stress response feature extraction: When an external shock event occurs, record the supplier's reaction time (time from the occurrence of the event to the first adjustment behavior), the order adjustment magnitude (such as the percentage reduction in order volume), the time required to recover to the normal level, and extract the overshoot (maximum deviation) and the adjustment time (time to enter steady state) to form stress curve features.

[0089] Social network feature extraction: Based on the communication text between suppliers and upstream and downstream partners, semantic vectors are extracted using pre-trained language models (such as BERT); at the same time, a graph network is constructed based on transaction relationships, and graph neural networks (such as GraphSAGE) are used to calculate the embedding representation of each node in the network, reflecting its role in the supply chain (such as core nodes and edge nodes).

[0090] Cost sensitivity feature extraction: By fitting the cost function through the relationship between historical quotations and order volume and raw material prices, the supplier's sensitivity coefficient to order size and external price fluctuations (i.e., the elasticity coefficient of the cost function) is obtained.

[0091] 3. Fingerprint encoding and storage unit

[0092] All the above features are concatenated into an original feature vector (which may have thousands of dimensions), and then input into an autoencoder for dimensionality reduction to obtain a fixed-dimensional behavioral fingerprint vector (e.g., 256 dimensions). The autoencoder is trained to minimize reconstruction error, ensuring that the fingerprint vector retains the main information of the original features. The fingerprint vector is stored in a cache (such as Redis) with a timestamp for subsequent modules to access.

[0093] Output: Supplier behavior fingerprint vector (256-dimensional floating-point array), published to the cognitive layer via message bus.

[0094] (II) Cognitive Level:

[0095] The cognitive layer comprises three core modules that handle mental accounting inference, resilience topology analysis, and multi-objective assessment, respectively.

[0096] 1. Supplier Mental Accounting Model Module

[0097] like Figure 3 As shown, this module infers the supplier's psychological variables based on behavioral fingerprints, including risk preference coefficient, cooperation inertia coefficient, trust level, and psychological balance, and quantifies their implicit motivations.

[0098] (1) Psychological variable definition and modeling unit

[0099] Risk preference coefficient ρ: Its value ranges from [-1, 1], with positive values ​​indicating risk-seeking and negative values ​​indicating risk aversion. This coefficient is obtained by fitting the supplier's historical behavior in the face of uncertainty decisions (such as whether to accept rush orders or maintain inventory) using cumulative prospect theory. Specifically, the risk preference coefficient ρ = (convexity of v'(x)) is calculated by fitting the parameters of the value function v(x) and the probability weight function w(p) based on the decision results, or it can be obtained directly through maximum likelihood estimation.

[0100] The cooperation inertia coefficient γ, ranging from [0, 1], represents the supplier's tendency to maintain the existing cooperative relationship. Using a logistic regression model, input features include cooperation duration, number of historical conflicts, and switching costs (such as the cost of finding a new supplier), and the output is a cooperation inertia score.

[0101] Trust level τ: A dynamic exponential decay model is used to describe the change of trust over time. The initial trust level is set based on historical performance records. Each subsequent performance update the trust level, and the decay rate is dynamically adjusted based on recent performance quality; good performance results in slower decay. The update formula is: τnew = τold * e (-λΔt) + δ * (performance score), where λ is the decay rate and δ is the update step size.

[0102] Psychological balance β: defined as accumulated emotional capital, initially 0, increasing by 1 point for each good performance and decreasing by 2 points for each breach. Points can be used to offset minor mistakes (e.g., a delay of up to 2 hours can be exempted from punishment). Psychological balance affects willingness to cooperate in subsequent games.

[0103] (2) Psychological state inference unit

[0104] Real-time inference is performed using a Hidden Markov Model (HMM). The psychological variables (ρ, γ, τ, β) are combined into a hidden state vector St, and the behavioral fingerprint vector (after dimensionality reduction) is used as the observation Ot. The model parameters λ = (A, B, π) are obtained by training on historical data (labeled psychological states) using the Baum-Welch algorithm. During real-time execution, the Viterbi algorithm is used to calculate the most probable current psychological state St.

[0105] (3) Abnormal psychological early warning unit

[0106] A dynamic threshold (e.g., mean ± 3 standard deviations) is established for each psychological variable, and an alert is triggered when a new value exceeds the threshold. Simultaneously, a cumulative sum (CUSUM) algorithm is used to detect persistent shifts in variables and identify potential trend risks in advance.

[0107] Output: A mental account structure (containing risk preference coefficient ρ, cooperation inertia coefficient γ, trust level τ, and mental balance β) is published to the game layer via a message bus.

[0108] 2. Supply Chain Resilience Topology Analysis and Restructuring Module

[0109] like Figure 4 As shown, this module uses a supplier knowledge graph to dynamically assess the resilience of the supply chain network, identify vulnerabilities, and generate restructuring solutions.

[0110] (1) Knowledge Graph Construction Unit

[0111] Use a graph database (such as Neo4j) to store nodes (suppliers, logistics providers, warehouses) and edges (transaction relationships, logistics routes). Node attributes include type, geographical location, capacity, inventory, etc.; edge attributes include transaction frequency, average latency, reliability score, etc. The graph is updated regularly from the business system.

[0112] (2) Toughness index calculation unit

[0113] Network connectivity resilience: quantified by simulating the change in the maximum connected component of the network after randomly removing nodes. Specifically, the area under the curve is calculated as the network connectivity resilience value, with the proportion of nodes retained (p) as the horizontal axis and the relative size of the maximum connected component (S(p)) as the vertical axis. The closer this value is to 1, the stronger the network resilience.

[0114] Redundancy index R: For each supplier node, calculate the quantity and quality weighted sum of its alternative suppliers: R = Σ (alternative supplier reliability score * distance weight), where the distance weight is determined by geographical distance.

[0115] Betweenness centrality: the degree to which a computation node acts as a bridge in a network, identifying key hub nodes.

[0116] Modularity: Identify the cluster structure of the supply chain and analyze the risks of industrial clusters.

[0117] (3) Resilience assessment model unit

[0118] The current network state is encoded using a graph neural network (such as a GNN). Inputs include node features (capacity, inventory, geographic location embedding, etc.) and an adjacency matrix. Outputs a network resilience score (0-100) and the vulnerability probability of each node. The model is trained using simulated failure scenario data (such as random node and edge removal). The loss function is the sum of the resilience score prediction error and the cross-entropy of node vulnerability classification.

[0119] (4) Reconstruction Strategy Generation Unit

[0120] For each vulnerability, reinforcement learning (Deep Q-Network, DQN) is used to generate a reconstruction strategy. The state consists of the current network embedding and vulnerability information; the action is the reconstruction action type (replace supplier, add connection, adjust order allocation) and the target node; the reward is the improvement in resilience score after reconstruction (calculated through a simulation environment). The agent is trained to select the optimal reconstruction action.

[0121] Outputs include: resilience score, a list of vulnerabilities (node ​​ID and risk value), and a list of refactoring suggestions (each suggestion includes action type, objective, and expected resilience improvement), which are published to the game layer via the message bus.

[0122] 3. Multi-objective dynamic assessment and conflict resolution module

[0123] like Figure 5 As shown, this module dynamically evaluates suppliers under four objectives: cost, timeliness, quality, and risk, resolves conflicts between objectives, and outputs a single comprehensive score to facilitate decision-making.

[0124] (1) Multi-objective evaluation model unit

[0125] For each supplier, construct a multi-objective evaluation vector F = (fc, ft, fq, fr), where:

[0126] Cost target fc: Predict the unit cost of future orders using an LSTM network. Inputs include historical cost series, raw material price index, order quantity, etc.

[0127] Delivery time target ft: Predicts the probability of order delay using a Transformer model. Inputs include historical delay records and logistics route status.

[0128] Quality target fq: Predicts the product quality pass rate using a gradient boosting tree (such as XGBoost). Inputs include historical quality inspection records, production batch information, etc.

[0129] Risk target fr: directly adopts the risk preference coefficient ρ from the mental account.

[0130] Each sub-target prediction model shares underlying features (including behavioral fingerprints, historical performance, etc.) and is jointly trained through a multi-task learning framework, enabling the model to simultaneously optimize the prediction accuracy of multiple targets. The loss function is a weighted sum of the losses of each sub-task (such as mean squared error and cross-entropy).

[0131] (2) Conflict Resolution Unit: To transform the multi-objective evaluation results into a comprehensive score that is easy to compare, a user preference vector ω=(ωc,ωt,ωq,ωr) is introduced, satisfying ∑ω=1. The system uses a linear weighted sum method for conflict resolution, and the specific steps are as follows:

[0132] First, each sub-objective is normalized to eliminate the influence of dimensions. The normalization formula is:

[0133]

[0134] in and These are the minimum and maximum values ​​of the target in historical data, or the ideal boundary preset according to business rules. (Normalized) The value ranges from [0,1], with larger values ​​indicating better performance on that objective (for cost-related indicators such as cost and delay probability, they need to be converted to benefit-related indicators, i.e. Or a similar approach, but for simplicity, each objective can be predefined as benefit-oriented.

[0135] The overall score S is obtained by weighted summation:

[0136]

[0137] The value of S ranges from [0,1], and the larger the value, the better the overall performance of the supplier.

[0138] Users can dynamically adjust their preference vectors through the system interface, and the system calculates the comprehensive score in real time without retraining the model. Simultaneously, the system employs online learning algorithms (such as gradient descent based on the user's historical choices) to implicitly optimize the preference vectors, achieving personalized evaluation.

[0139] Output: Supplier multi-objective comprehensive score S, published to the game layer via message bus.

[0140] (III) Game Theory Layer

[0141] The game theory layer comprises two modules, which handle reverse bidding and dynamic contract generation, and intervention strategy recommendation and effect prediction, respectively.

[0142] 1. Reverse bidding and dynamic contract generation module

[0143] like Figure 6 As shown, when an alert is triggered or reconstruction is required, this module automatically initiates a reverse bidding mechanism to compete for the best bid among the alternative suppliers and generate a dynamic contract.

[0144] (1) Auction Mechanism Design Unit

[0145] A reverse auction is conducted using the VCG (Vickrey-Clarke-Groves) mechanism. The auction process includes:

[0146] Issue tender documents: required quantity, delivery time, and quality requirements.

[0147] Collect supplier quotes: Each supplier submits a quote bi and the quantity qi available.

[0148] Determine the winning bid set: With the goal of maximizing social welfare (buyer's valuation minus total payment), solve the combinatorial optimization problem to obtain the winning bid set W.

[0149] Payment Price: For the winning supplier i, the payment price pi = (socially optimal total welfare - socially optimal total welfare after removing i) + bi. The VCG mechanism incentivizes suppliers to report their true costs, thereby maximizing social welfare.

[0150] (2) Supplier quotation simulation unit

[0151] Multi-agent reinforcement learning (such as MADDPG) is used to simulate possible bidding strategies of suppliers in an offline environment. Each supplier agent learns a bidding strategy based on its own costs, historical bids, market conditions, and other factors. Through multiple rounds of game theory, an equilibrium bid distribution is obtained to predict the actual auction results.

[0152] (3) Dynamic contract generation unit

[0153] Dynamic terms are generated based on the contract template. Price adjustment terms use a formula:

[0154] ,in As the benchmark price, This represents the current raw material price index. As a benchmark index, The adjustment factor is determined by the supplier's risk appetite (risk-seekers are less willing to accept lower risk tolerance). The performance incentive clause dynamically adjusts the reward and penalty coefficients based on the level of trust. The contract content is populated using a template engine and deployed to a blockchain (such as Ethereum or Hyperledger Fabric) to generate a smart contract.

[0155] Output: Reverse bidding results (winning supplier ID, final price), dynamic contract parameters (price formula, reward / penalty coefficients, etc.), and contract address, published to the execution layer via the message bus.

[0156] 2. Multi-objective intervention strategy recommendation and effect prediction module

[0157] After an alert is triggered, this module combines mental accounting, resilience analysis, and multi-objective assessment to recommend specific intervention strategies and predict the intervention effect.

[0158] (1) Strategy space definition unit

[0159] Define soft interventions (sending reminders, adjusting order allocation ratios, providing incentives) and hard interventions (suspending cooperation, switching suppliers, activating alternative logistics) and their combinations. The strategy space is configurable.

[0160] (2) Effect prediction model unit

[0161] Causal forests are used to model historical intervention data to learn the causal effects of different intervention strategies on the supplier's subsequent performance. Input features include the supplier's current state (psychological variables, performance history, network location) and the type of intervention strategy. The output is the improvement in performance rate or cost change over a certain period after the intervention. The model can estimate the expected effect of each strategy and its confidence interval.

[0162] (3) Strategy Recommendation Unit

[0163] A multi-objective optimization algorithm (such as NSGA-II) is used to weigh costs, effectiveness, and risks to generate a Pareto-optimal policy set. Recommended policies are selected from the Pareto set based on user preferences (configurable), and the output list includes predicted effects and confidence intervals.

[0164] Output: A list of recommended intervention strategies (each strategy includes an action description, expected effect, and confidence interval), published to the execution layer via the message bus.

[0165] (iv) Execution layer

[0166] The execution layer comprises two modules, which handle the automatic execution of smart contracts and supply chain restructuring, respectively.

[0167] 1. Trusted Commitment Verification and Smart Contract Automated Execution Module

[0168] This module, based on blockchain and oracles, verifies supplier performance and automatically executes rewards and penalties.

[0169] (1) Oracle Network Unit

[0170] Deploy multiple decentralized oracle nodes (such as Chainlink nodes) to obtain fulfillment data from different data sources (logistics API, quality inspection reports, warehouse monitoring). Employ a threshold signature mechanism where nodes sign the data. When more than 2 / 3 of the nodes collect the same data, aggregate the signatures and submit them to the smart contract, ensuring that the data is tamper-proof and has a trustworthy source.

[0171] (2) Smart contract logic unit

[0172] Smart contracts have the following core logic built in:

[0173] Receive the fulfillment status (whether it is on time, whether the quality is up to standard) reported by the oracle.

[0174] Payments are automatically executed based on preset conditions: if the performance is satisfactory, the payment is released and the supplier's psychological balance is increased (via cross-chain messaging); if the performance is in default, the deposit is deducted and the psychological balance is reduced.

[0175] Trigger a cross-chain event to transmit the reward or punishment result to the chain where the mental account module is located.

[0176] (3) Cross-chain interoperability unit

[0177] Cross-chain data synchronization can be achieved by using cross-chain bridges (such as Polkadot's XCMP) to transfer reward and punishment results from the execution chain to the mental account chain.

[0178] Output: Automatically executed transaction records and mental account update events, stored on the blockchain.

[0179] 2. Supply Chain Adaptive Restructuring and Resource Scheduling Module

[0180] Based on the restructuring recommendations, supply chain adjustments are automatically executed, including order rerouting, inventory reallocation, and logistics switching.

[0181] (1) Reconstruct the execution engine unit

[0182] Receive the refactoring strategy and parse it into specific operation instructions (such as canceling the original order, creating a new order, adjusting inventory reservations, and updating the shipping plan). Orchestrate calls to the APIs of various business systems through a workflow engine (such as Apache Airflow) and monitor the execution status to ensure the atomicity and consistency of operations.

[0183] (2) Resource scheduling optimization unit

[0184] The restructuring problem is modeled as a multi-commodity flow problem, with the objective of minimizing total costs (including transportation costs, inventory costs, and default costs), while considering constraints such as supplier capacity, warehouse capacity, and transportation capacity. A column generation algorithm is used to solve the problem, yielding the optimal order allocation and logistics scheduling scheme.

[0185] (3) Abnormal rollback unit

[0186] If a step fails during execution, a rollback is automatically triggered: the executed operation is undone, the original state is restored, and a failure event is sent to the game layer to trigger a new round of game.

[0187] Output: Refactoring execution result (success / failure), scheduling plan, rollback record, stored in the execution log.

[0188] (V) Evolutionary Layer: Meta-learning and Reinforcement Learning Self-Evolutionary Module

[0189] like Figure 7 As shown, this module learns from historical intervention results and continuously optimizes the models of each module to achieve system self-evolution.

[0190] 1. Meta-learning framework unit

[0191] Model-independent meta-learning (MAML) is used to train the initial model, enabling it to quickly adapt to new suppliers or scenarios with only a small number of gradient updates. For example, after training the mental accounting model with MAML, new suppliers only require a small amount of historical data (such as 5 transactions) to fine-tune the model, achieving a rapid cold start.

[0192] 2. Multi-agent reinforcement learning unit

[0193] Key modules in the perception, cognition, game theory, and execution layers (such as feature extractors, psychological inferators, policy recommenders, and schedulers) are treated as agents. A multi-agent proximal policy optimization (MAPPO) algorithm is used to collaboratively optimize the system's long-term benefits (such as improved overall performance and cost savings). Each agent outputs an action based on local observations, a central value network evaluates the global state, and all agents share rewards (such as weighted comprehensive performance). The overall system performance is continuously improved through training in a simulation environment.

[0194] 3. Experience Replay and Knowledge Distillation Unit

[0195] Historical intervention experiences are stored in an experience pool (Replay Buffer) and periodically sampled for training. Knowledge distillation is used to compress the trained complex multi-agent model into a lightweight model, ensuring online real-time performance while retaining key performance characteristics.

[0196] Output: Continuously evolving model parameters (published periodically to each module), training logs.

[0197] III. Example of Inter-Module Collaborative Workflow

[0198] Taking a complete early warning cycle as an example, the system workflow is as follows:

[0199] 1. The perception layer processes data in real time, discovers that supplier A's order response time has shortened and its price fluctuations have increased, updates the behavior fingerprint vector and publishes it to the message bus.

[0200] 2. The cognitive layer subscribes to behavioral fingerprints, and the mental accounting model infers that the risk preference coefficient is abnormally high, triggering a psychological warning; resilience analysis finds that A is a key node, and its failure may lead to a significant decrease in network resilience; multi-objective evaluation shows that the comprehensive score drops below the threshold, and an evaluation event is released.

[0201] 3. The game layer integrates various events, generates a strategy through the intervention strategy recommendation module (switching 30% of orders to alternative supplier B, while giving A a warning), and initiates reverse bidding to select B as the winning bidder from the alternative pool, generating dynamic contract parameters.

[0202] 4. The execution layer deploys smart contracts to execute order switching. The oracle monitors B's performance status and automatically executes rewards and penalties, while updating A's mental account.

[0203] 5. The evolutionary layer records all data from the entire intervention process for subsequent model updates.

[0204] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic evaluation and intelligent early warning system for cross-border suppliers based on performance, characterized in that, include: The perception layer is used to collect behavioral data of suppliers in multi-source business systems and generate supplier behavior fingerprint vectors through feature extraction and dimensionality reduction encoding. The cognitive layer, connected to the perception layer, is used to infer the supplier's psychological variables based on the behavioral fingerprint vector using a probabilistic graphical model, analyze the supplier's knowledge graph using a graph neural network to assess supply chain resilience, and use multiple prediction models to conduct multi-objective evaluation of the supplier's performance and generate a comprehensive score. The game theory layer, connected to the cognitive layer, is used to conduct supplier bidding and generate dynamic contracts based on the psychological variables, resilience assessment results and comprehensive scores, and to estimate the effect of intervention strategies and generate multi-objective optimization strategies using causal inference models. The execution layer, connected to the game layer, is used to automatically execute performance rewards and penalties based on blockchain smart contracts and oracle networks, and to model the supply chain restructuring problem as a combinatorial optimization model to obtain a resource scheduling scheme. An evolutionary layer, connected to the execution layer, is used to enable rapid adaptation to new suppliers using a meta-learning framework and to collaboratively optimize the model parameters of each layer of the system using multi-agent reinforcement learning.

2. The system according to claim 1, characterized in that, The sensing layer includes: The multi-source data access unit is used to access data from the enterprise business system in real time via API interface, and to clean, align and normalize the data. The behavioral feature extraction unit is used to extract time series features, stress response features, social network features, and cost sensitivity features from the processed data; The fingerprint encoding and storage unit is used to concatenate the extracted features into an original feature vector, input it into the autoencoder for dimensionality reduction, and output a fixed-dimensional behavioral fingerprint vector for storage.

3. The system according to claim 1, characterized in that, The module in the cognitive layer that uses a probabilistic graphical model to infer the supplier's psychological variables is specifically used for: Risk preference coefficients are obtained by fitting behavioral fingerprint vectors using cumulative prospect theory. The cooperative inertia coefficient was obtained using a logistic regression model. A dynamic exponential decay model is used to update the trust level, which decays over time and is adjusted by performance behavior. Update your psychological balance by accumulating emotional capital; The risk preference coefficient, cooperation inertia coefficient, trust level, and psychological balance are used as the hidden states of the hidden Markov model. The behavioral fingerprint vector is used as the observation value to infer the current psychological state in real time and to provide early warning of abnormal changes in psychological variables.

4. The system according to claim 3, characterized in that, The module in the cognitive layer that uses graph neural networks to analyze supplier knowledge graphs to assess supply chain resilience is specifically used for: Use a graph database to store a knowledge graph containing suppliers, logistics providers, warehouse nodes, transaction relationships, and logistics path edges; By simulating node removal, we can compute the network connectivity resilience value, redundancy index, and betweenness centrality of nodes. The graph neural network is used to encode the network state and output the network's resilience score and the vulnerability probability of the nodes. Reinforcement learning models are used to generate reconstruction strategies for vulnerabilities. Actions include replacing suppliers, adding connections, or adjusting order allocation. The reward is the increase in resilience score after reconstruction.

5. The system according to claim 4, characterized in that, The module in the cognitive layer that uses multiple prediction models to conduct multi-objective evaluations of supplier performance and generate a comprehensive score is specifically used for: For each supplier, a time-series forecasting model is used to predict the cost of future orders as the cost target, a sequence model is used to predict the probability of order delays as the timeliness target, a machine learning model is used to predict the product quality pass rate as the quality target, and a risk preference coefficient is used as the risk target. Each sub-target prediction model shares underlying features and is jointly trained through multi-task learning; A user preference vector is introduced to normalize each sub-objective, and then a weighted sum is obtained to obtain a comprehensive score. The preference vector is then optimized through an online learning algorithm.

6. The system according to claim 1, characterized in that, The module in the game layer that uses an auction mechanism for supplier bidding is specifically used for: When an early warning is triggered or a restructuring is required, a reverse auction is initiated, and a tender document containing the required quantity, delivery date, and quality requirements is issued. Collect the quotations and quantities offered by suppliers, solve the combinatorial optimization problem with the goal of maximizing social welfare, and determine the winning bid set; For the winning supplier, the payment price is the socially optimal total welfare minus the optimal total welfare after removing the supplier, plus the supplier's bid. Multi-agent reinforcement learning is used to simulate supplier pricing strategies and predict the equilibrium pricing distribution.

7. The system according to claim 6, characterized in that, The module that generates dynamic contracts in the game layer is specifically used for: Dynamic terms are generated based on the contract template. The price adjustment terms adopt a function of the benchmark price multiplied by the rate of change of the raw material price index, and the adjustment coefficient is determined by the supplier's risk preference coefficient. The performance incentive clauses dynamically adjust the reward and penalty coefficients based on the level of trust. The contract content is populated using a template engine and deployed to the blockchain to generate a smart contract.

8. The system according to claim 7, characterized in that, The module in the game theory layer that uses a causal inference model to estimate the effect of intervention strategies and generate multi-objective optimization strategies is specifically used for: Define the policy space, which includes both soft and hard interventions; Using causal forests to model historical intervention data, we learn the causal effects of different intervention strategies on the supplier's subsequent performance, and output the estimated value and confidence interval of the change in performance after intervention. A multi-objective optimization algorithm is used to weigh the costs, effects, and risks, generate a Pareto optimal policy set, and select a recommended policy based on user preferences.

9. The system according to claim 1, characterized in that, The module in the execution layer that automatically executes performance rewards and penalties based on blockchain smart contracts and oracle networks is specifically used for: Deploy multiple decentralized oracle nodes to obtain performance data from different data sources, and use a threshold signature mechanism to aggregate and sign the data before submitting it to the smart contract; The smart contract receives the performance status reported by the oracle and automatically executes payment, margin deduction and mental balance update according to preset conditions; The results of rewards and punishments are transmitted to the mental account chain through a cross-chain bridge, enabling cross-chain data synchronization.

10. The system according to claim 9, characterized in that, The module in the execution layer that models the supply chain restructuring problem as a combinatorial optimization model to obtain a resource scheduling scheme is specifically used for: Receive the refactoring strategy, parse it into specific operation instructions, and orchestrate calls to the APIs of various business systems through the workflow engine; The reconstruction problem is modeled as a multi-product flow problem. With the goal of minimizing the total cost, and considering constraints such as supplier capacity, warehouse capacity, and transportation capacity, a column generation algorithm is used to solve the optimal order allocation and logistics scheduling scheme. If a step fails during execution, a rollback is automatically triggered and a failure event is sent to the game layer.

Images