Intelligent procurement cooperation method and system

Abstract

The invention relates to an intelligent procurement collaboration method and system, and the method comprises the following steps: S1, building a supplier knowledge graph based on multi-source heterogeneous data fusion through employing a graph embedding algorithm, achieving the associated storage of industrial and commercial information, performance records and quality reports through a Neo4j graph database, calculating the weight of a supplier node through employing a PageRank improved algorithm, and carrying out the calculation of the weight of the supplier node; s2, based on the three-dimensional supplier portrait matrix, using an improved collaborative filtering algorithm to carry out demand matching, analyzing a purchase demand document through an ElasticSearch semantic analysis engine, combining TF-IDF weighted cosine similarity calculation to realize intelligent recommendation, and generating a purchase demand scheme with a weight score. The supplier knowledge graph is constructed through the multi-source heterogeneous data fusion and graph embedding algorithm, the scattered industrial and commercial information, performance records and quality reports are stored in an associated mode, traditional data island limitation is broken through, and effective integration of supplier multi-dimensional features is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of enterprise intelligent procurement and supply chain collaboration, in particular to an intelligent procurement collaboration method and system. BACKGROUND

[0002] The technical field of enterprise intelligent procurement and supply chain collaboration focuses on reconstructing traditional procurement processes through digital technology, deeply integrating big data analysis, artificial intelligence, blockchain and Internet of Things technology, and building a full-link collaboration system covering demand forecasting, intelligent matching of suppliers, automatic negotiation decision-making and real-time tracking of logistics. This field focuses on solving the problem of information asymmetry between multiple subjects in the supply chain, and through the establishment of cross-organizational data sharing mechanisms and intelligent decision-making models, it realizes the standardization of procurement processes, the dynamicization of supplier management and the front-end of risk control. Among them, the intelligent procurement collaboration method and system refers to an intelligent procurement management platform based on a cloud computing architecture, which integrates supplier evaluation models, dynamic pricing engines and electronic contract systems to eliminate data barriers between enterprises and optimize resource allocation efficiency.

[0003] The prior art has the problem of single dimension of supplier evaluation in actual application, and relies on static indicators to cause the problem of not being able to timely reflect the change of supplier credit, which may cause misjudgment of cooperation risk. The cross-organizational data sharing mechanism lacks deep integration capability, and the dispersed supplier data is difficult to form an effective knowledge network, resulting in low demand matching efficiency. The traditional pricing model does not fully consider market dynamic fluctuations and customer differences, and the use of fixed price strategy leads to limited negotiation space, affecting the effect of procurement cost control. SUMMARY

[0004] In order to solve the problem of single dimension of supplier evaluation in actual application of the prior art, and the problem of not being able to timely reflect the change of supplier credit caused by relying on static indicators, which may cause misjudgment of cooperation risk, the present application provides an intelligent procurement collaboration method and system.

[0005] The intelligent procurement collaboration method and system provided by the present application adopts the following technical solution: The intelligent procurement collaboration method comprises the following steps: S1: Based on multi-source heterogeneous data fusion, a graph embedding algorithm is used to construct a supplier knowledge graph, the association storage of industrial and commercial information, performance records and quality reports is realized through a Neo4j graph database, and a PageRank improved algorithm is used to calculate the weight of the supplier node to generate a three-dimensional supplier portrait matrix; S2: Based on the three-dimensional supplier portrait matrix, an improved collaborative filtering algorithm is used for demand matching, a procurement demand document is analyzed through an ElasticSearch semantic analysis engine, intelligent recommendation is realized by combining TF-IDF weighted cosine similarity calculation, and a procurement demand scheme with weighted score is generated; S3: Based on the procurement scheme with weighted scoring, a dynamic pricing game model is used, the Monte Carlo tree search algorithm is used to simulate the bidding strategy, the XGBoost regression model is used to predict the market benchmark price, and the thousand customers thousand prices rule engine is connected to generate a multi-dimensional constrained optimal bidding strategy set. S4: Based on the multi-dimensional constraint optimal pricing strategy set, the smart contract-driven process is initiated. The BiLSTM-CRF contract terms parsing model is adopted, and the compliance verification of the terms is realized through the digital signature timestamp chain. A custom approval routing algorithm is triggered to generate a legally valid electronic contract draft. S5: Based on legally binding electronic contract drafts, construct a digital twin of the supply chain, adopt a multi-agent reinforcement learning framework, adjust procurement strategies in real time through PPO strategy optimization algorithms, and feed them back to the supplier evaluation model to generate adaptive procurement decision suggestions.

[0006] Preferably, generating a three-dimensional supplier profile matrix based on S1 includes the following steps: S101: Based on the integration with the enterprise ERP system, a multi-source data cleaning framework is adopted. The basic information of suppliers is extracted through regular expression matching algorithm, and the isolated forest anomaly detection model is used to filter invalid data and generate standardized supplier structure data. S102: Based on standardized supplier structure data, a graph convolutional neural network is used to construct a supply relationship topology graph through a node feature propagation mechanism, and the TransE embedding algorithm is applied to generate vector representations to generate supplier relationship graph embedding features; S103: Based on the embedded features of the supplier relationship graph, a dynamic weighted PageRank algorithm is designed. The influence of historical cooperation records is adjusted by the time decay factor and normalized by the Sigmoid function to generate a three-dimensional supplier profile matrix.

[0007] Preferably, the generation of a weighted scoring procurement demand scheme based on S2 includes the following steps: S201: Based on the procurement requirement document library, the BiLSTM-Attention model is used to extract technical parameter features through a word segmentation algorithm enhanced by a domain dictionary, and a structured procurement requirement vector is generated. S202: Based on the structured procurement demand vector, a hybrid collaborative filtering algorithm is used to calculate the supplier matching degree through matrix factorization technology, and the profile matrix of S103 is integrated for weighting to generate a primary supplier candidate set. S203: Based on the initial supplier candidate set, a multi-objective optimization model is constructed. The NSGA-II algorithm is used to balance price, delivery time, and quality indicators. The TOPSIS method is used to rank the solutions and generate a procurement requirement solution with weighted scores.

[0008] Preferably, the generation of a multidimensional constrained optimal pricing strategy set based on S3 includes the following steps: S301: Based on historical transaction database, using the extreme gradient boosting tree model, the SHAP value interpretation method is used to identify price influencing factors and generate a market benchmark price prediction curve. S302: Based on the market benchmark price prediction curve, construct an incomplete information game model, use the Monte Carlo tree search algorithm to simulate supplier pricing strategies, and generate a dynamic pricing game tree; S303: Based on the dynamic price game tree, a distributed constraint optimization algorithm is designed. The Lagrange multiplier method is used to solve the multi-objective optimal solution, and the algorithm is connected to the thousand customers thousand prices rule engine for strategy correction to generate a multi-dimensional constraint optimal price strategy set.

[0009] Preferably, generating a legally valid electronic contract draft based on S4 includes the following steps: S401: Based on a contract template library, a bidirectional long short-term memory network-conditional random field model is adopted to parse legal clauses and generate a set of structured contract elements through a domain-adaptive training method. S402: Based on a set of structured contract elements, a smart contract verification chain is constructed. A Merkle tree structure is used to verify the compliance of the terms, and the national cryptographic SM3 algorithm is applied to generate a digital digest and a legal compliance verification report. S403: Based on the legal compliance verification report, trigger the process routing decision tree, dynamically load approval rules through the random forest classification algorithm, and integrate electronic signature services to generate a legally valid electronic contract draft.

[0010] Preferably, generating adaptive procurement decision recommendations based on S5 includes the following steps: S501: Based on IoT sensor data streams, Kalman filtering algorithm is used for real-time data cleaning to build a digital twin of the supply chain and generate a real-time mirror model of the supply chain. S502: Based on the real-time mirror model of the supply chain, a multi-agent deep reinforcement learning framework is designed, and a near-end policy optimization algorithm is used for policy training to generate a dynamic procurement policy matrix. S503: Based on the dynamic procurement strategy matrix, a feedback adjustment mechanism is constructed, and the supplier evaluation model parameters are updated through the gradient backpropagation algorithm to generate adaptive procurement decision suggestions.

[0011] The intelligent procurement collaboration system includes the following modules: a supplier profiling module, which uses graph convolutional neural network algorithm to construct a supply relationship topology based on enterprise ERP system data, calculates node influence through dynamic weight PageRank algorithm, and integrates credit / capability / risk indicators using three-dimensional feature fusion technology to generate a supplier three-dimensional evaluation matrix. The supplier profiling module includes a data cleaning submodule, a graph embedding submodule, and a feature fusion submodule; The intelligent procurement module, based on a supplier three-dimensional evaluation matrix, uses a hybrid collaborative filtering algorithm for demand matching, balances price / delivery time / quality indicators through a multi-objective optimization model, and connects to a thousand customers, thousand prices rule engine to generate dynamic quotations and generate multi-constraint optimal procurement solutions. The intelligent procurement module includes a demand analysis submodule, a strategy optimization submodule, and a dynamic pricing submodule; The contract execution module, based on a multi-constraint optimal procurement scheme, uses a bidirectional long short-term memory network-conditional stochastic model to parse contract terms, verifies the compliance of terms through a smart contract verification chain, and triggers a multi-level approval routing algorithm to generate a legally valid execution contract. The contract execution module includes a terms parsing submodule, a compliance verification submodule, and a process routing submodule.

[0012] Preferably, the data cleaning submodule, based on the original data of the enterprise ERP system, adopts a multi-source data cleaning framework, extracts key fields through regular expression matching algorithms, and uses an isolated forest anomaly detection model to filter noisy data and generate standardized supplier structure data. The graph embedding submodule constructs a supply relationship topology graph based on standardized supplier structure data, uses a graph convolutional neural network algorithm to propagate node features, and applies the TransE embedding algorithm to generate vector representations, thus generating supplier relationship graph embedding features. The feature fusion submodule, based on the embedded features of the supplier relationship graph, designs a dynamic weight PageRank algorithm, adjusts the historical cooperation weights through a time decay factor, and combines it with the Sigmoid function normalization process to generate a three-dimensional supplier evaluation matrix.

[0013] Preferably, the demand parsing submodule, based on the procurement demand document library, adopts a bidirectional long short-term memory network-attention mechanism model, extracts technical parameters through a domain dictionary-enhanced word segmentation algorithm, and generates a structured procurement demand vector; The strategy optimization submodule, based on the structured procurement demand vector, uses a hybrid collaborative filtering algorithm, calculates the matching degree through matrix factorization technology, and integrates the supplier three-dimensional evaluation matrix for weighted sorting to generate a candidate supplier score list. The dynamic pricing submodule constructs a multi-objective optimization model based on the candidate supplier rating list, uses a non-dominated sorting genetic algorithm to balance price / delivery time indicators, and connects to the customer-specific pricing rule engine to generate a multi-constraint optimal procurement plan.

[0014] Preferably, the clause parsing submodule, based on the contract template library, adopts a bidirectional long short-term memory network-conditional random field model and uses a domain-adaptive training method to parse legal clauses and generate a set of structured contract elements. The compliance verification submodule constructs a smart contract verification chain based on a set of structured contract elements, uses a Merkle tree structure for clause comparison, applies the national cryptographic SM3 algorithm to generate digital digests, and generates a legal compliance verification report. The process routing submodule, based on the legal compliance verification report, triggers a multi-level approval routing algorithm, dynamically loads approval rules through a random forest classification model, integrates electronic signature services, and generates legally binding enforcement contracts.

[0015] In summary, this application includes at least one of the following beneficial technical effects: A supplier knowledge graph is constructed through multi-source heterogeneous data fusion and graph embedding algorithms, linking and storing scattered business information, performance records, and quality reports. This breaks through the limitations of traditional data silos and effectively integrates multi-dimensional supplier features. An improved PageRank algorithm introduces a time decay factor and dynamic weight adjustment mechanism, enabling the supplier evaluation matrix to accurately reflect real-time credit status and performance capabilities. Combined with a semantic analysis engine and an improved collaborative filtering algorithm, deep parsing of procurement requirement documents is achieved during the demand matching stage. TF-IDF weighted cosine similarity calculation improves recommendation accuracy, ensuring a high degree of alignment between procurement plans and supplier capabilities. A dynamic pricing game model integrates Monte Carlo methods. The Karlo tree search algorithm and XGBoost regression prediction simultaneously consider market benchmark price fluctuations when simulating pricing strategies. Combined with a personalized pricing rule engine, it generates optimal pricing strategies under multi-dimensional constraints. The smart contract-driven process uses a BiLSTM-CRF model to automate the parsing and compliance verification of contract terms. The digital signature timestamp chain ensures the legal validity of contracts. A custom approval routing algorithm significantly improves contract generation efficiency. The construction of a supply chain digital twin is combined with a multi-agent reinforcement learning framework. The PPO strategy optimization algorithm is used to dynamically adjust procurement strategies, forming a closed-loop feedback mechanism between the evaluation model and decision suggestions. Ultimately, it achieves intelligent and adaptive optimization of the entire procurement process.

[0016] By constructing a supply relationship topology graph using graph convolutional neural network algorithms, discrete data in enterprise ERP systems are deeply correlated, breaking through the limitations of single-dimensional data in traditional supplier evaluation. This achieves three-dimensional dynamic fusion of credit, capability, and risk indicators. The dynamic weighted PageRank algorithm introduces time decay factors and industry-specific parameters, making the calculation of supplier node influence more aligned with actual business scenarios and addressing the shortcomings of static evaluation models in reflecting market changes. The hybrid collaborative filtering algorithm, combined with a multi-objective optimization model, simultaneously optimizes price, delivery time, and quality indicators during the demand matching stage. Pareto front analysis is used to screen the optimal solution set, eliminating suboptimal decision-making problems caused by fragmented indicators in traditional procurement strategies. The personalized pricing rule engine embeds a dynamic pricing mechanism, generating differentiated pricing strategies based on customers' historical procurement data and market fluctuations, overcoming the drawbacks of fixed pricing models' slow market response. The bidirectional long short-term memory network-conditional random field model performs context-aware parsing of contract terms, and combined with a smart contract verification chain, it achieves real-time cross-verification of clause compliance. Through a multi-level approval routing algorithm, it dynamically adapts to enterprise internal control rules, significantly improving contract execution efficiency and legal risk prevention capabilities. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the main steps of this application; Figure 2 This is a detailed schematic diagram of S1 in this application; Figure 3 This is a detailed schematic diagram of S2 in this application; Figure 4 This is a detailed schematic diagram of S3 in this application; Figure 5 This is a detailed schematic diagram of S4 in this application; Figure 6 This is a detailed schematic diagram of S5 in this application; Figure 7 This is a system block diagram of this application. Detailed Implementation

[0018] 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.

[0019] The following is in conjunction with the appendix Figures 1-7 This application will be described in further detail.

[0020] See Figure 1 The intelligent procurement collaboration method includes the following steps: S1: Based on the fusion of multi-source heterogeneous data, a supplier knowledge graph is constructed using a graph embedding algorithm. The Neo4j graph database is used to realize the associated storage of business information, performance records, and quality reports. The PageRank improved algorithm is used to calculate the supplier node weights and generate a three-dimensional supplier profile matrix. S2: Based on a three-dimensional supplier profile matrix, an improved collaborative filtering algorithm is used for demand matching. The ElasticSearch semantic analysis engine is used to parse the procurement demand documents, and TF-IDF weighted cosine similarity calculation is combined to realize intelligent recommendation and generate procurement demand solutions with weighted scores. S3: Based on the procurement scheme with weighted scoring, a dynamic pricing game model is used, the Monte Carlo tree search algorithm is used to simulate the bidding strategy, the XGBoost regression model is used to predict the market benchmark price, and the thousand customers thousand prices rule engine is connected to generate a multi-dimensional constrained optimal bidding strategy set. S4: Based on the multi-dimensional constraint optimal pricing strategy set, the smart contract-driven process is initiated. The BiLSTM-CRF contract terms parsing model is adopted, and the compliance verification of the terms is realized through the digital signature timestamp chain. A custom approval routing algorithm is triggered to generate a legally valid electronic contract draft. S5: Based on legally binding electronic contract drafts, construct a digital twin of the supply chain, adopt a multi-agent reinforcement learning framework, adjust procurement strategies in real time through PPO strategy optimization algorithms, and feed them back to the supplier evaluation model to generate adaptive procurement decision suggestions.

[0021] A supplier knowledge graph is constructed through multi-source heterogeneous data fusion and graph embedding algorithms, linking and storing scattered business information, performance records, and quality reports. This breaks through the limitations of traditional data silos and effectively integrates multi-dimensional supplier features. An improved PageRank algorithm introduces a time decay factor and dynamic weight adjustment mechanism, enabling the supplier evaluation matrix to accurately reflect real-time credit status and performance capabilities. Combined with a semantic analysis engine and an improved collaborative filtering algorithm, deep parsing of procurement requirement documents is achieved during the demand matching stage. TF-IDF weighted cosine similarity calculation improves recommendation accuracy, ensuring a high degree of alignment between procurement plans and supplier capabilities. A dynamic pricing game model integrates Monte Carlo methods. The Karlo tree search algorithm and XGBoost regression prediction simultaneously consider market benchmark price fluctuations when simulating pricing strategies. Combined with a personalized pricing rule engine, it generates optimal pricing strategies under multi-dimensional constraints. The smart contract-driven process uses a BiLSTM-CRF model to automate the parsing and compliance verification of contract terms. The digital signature timestamp chain ensures the legal validity of contracts. A custom approval routing algorithm significantly improves contract generation efficiency. The construction of a supply chain digital twin is combined with a multi-agent reinforcement learning framework. The PPO strategy optimization algorithm is used to dynamically adjust procurement strategies, forming a closed-loop feedback mechanism between the evaluation model and decision suggestions. Ultimately, it achieves intelligent and adaptive optimization of the entire procurement process.

[0022] See Figure 2 The generation of a three-dimensional supplier profile matrix based on S1 includes the following steps: S101: Based on the integration with the enterprise ERP system, a multi-source data cleaning framework is adopted. The basic information of suppliers is extracted through regular expression matching algorithm, and the isolated forest anomaly detection model is used to filter invalid data and generate standardized supplier structure data. Based on the integration with the enterprise ERP system, a data extraction channel was established, and a regular expression pattern library was configured. The supplier name field was matched using the pattern ^[\u4e00-\u9fa5]+company$, and the unified social credit code field was extracted using the \d{18} pattern. The address field was split using a three-level administrative division regular expression. During the data cleaning stage, field integrity verification rules were set. When the missing rate of required fields exceeded 30%, the data discard process was triggered. The anomaly detection stage adopted the isolated forest model, and the anomaly score threshold was set to 0.65. When the ratio of the sample path length to the average path length exceeded the threshold, it was marked as abnormal data. For example, if a supplier's on-time delivery rate field had an anomaly value of 100% for 12 consecutive months, it would be removed after model detection. Finally, a supplier master data table containing unified codes and standard fields was formed.

[0023] S102: Based on standardized supplier structure data, a graph convolutional neural network is used to construct a supply relationship topology graph through a node feature propagation mechanism, and the TransE embedding algorithm is applied to generate vector representations to generate supplier relationship graph embedding features; Based on standardized supplier structure data, a supply relationship graph node attribute table is constructed. The enterprise node feature vector is defined to include registered capital, years of establishment, and industry classification code. The edge relationship weight is calculated using the formula of transaction amount ratio in the past three years: edge weight = annual transaction amount / maximum annual transaction amount × 0.6 + years of cooperation / 10 × 0.4. When applying the graph convolutional neural network, two propagation layers are set. The first layer aggregates the features of directly associated suppliers, and the second layer extends to secondary associated nodes. The TransE embedding training stage is set with an embedding dimension of 128, and the loss function is calculated using the L1 norm. The iteration is stopped when the number of iterations reaches 500. For example, after embedding, the vector value of the cooperation closeness between a certain automotive parts supplier node and the OEM reaches 0.87, which is significantly higher than the industry average.

[0024] S103: Based on the embedded features of the supplier relationship graph, a dynamic weighted PageRank algorithm is designed. The influence of historical cooperation records is adjusted by the time decay factor and normalized by the Sigmoid function to generate a three-dimensional supplier profile matrix.

[0025] The dynamic weighted PageRank algorithm introduces a time decay coefficient λ=0.85. The formula for calculating the weight of historical cooperation records is w_t=λ^(Tt), where T is the current year and t is the year of cooperation. The initial weight of nodes is set using the logarithmic normalized value of registered capital. During the Sigmoid function parameter adjustment stage, the scaling factors for credit dimension α=0.3, capability dimension β=0.5, and risk dimension γ=0.2 are set. During normalization, suppliers with credit scores exceeding the industry average by 15% are non-linearly amplified. For example, an electronic component supplier is calculated to have a credit dimension of 0.82, a capability dimension of 0.75, and a risk dimension of 0.23. The final generated three-dimensional coordinate values ​​are then weighted and fused into the decision matrix.

[0026] See Figure 3 The steps for generating a weighted and scored procurement demand plan based on S2 are as follows: S201: Based on the procurement requirement document library, the BiLSTM-Attention model is used to extract technical parameter features through a word segmentation algorithm enhanced by a domain dictionary, and a structured procurement requirement vector is generated. Based on the procurement requirement document library, a domain-specific dictionary library is established. For the mechanical manufacturing field, professional terms such as bearing models and tolerance grades are added. A multi-threaded word segmentation engine is configured, and special symbol filtering rules are implemented during text preprocessing. Regular expression matching is used on technical parameter description fields to extract key numerical parameters. The feature vector dimension is set to 50 dimensions. A bidirectional long short-term memory network is used to capture contextual semantic relationships. A weight threshold of 0.7 is set in the attention mechanism layer. When the feature importance score exceeds the threshold, the parameter is retained. For example, in a machine tool procurement requirement, "spindle speed ≥ 8000 rpm" is parsed and mapped to feature values ​​in the speed dimension, forming a standardized requirement vector containing material requirements, precision grades, and technical standards.

[0027] S202: Based on the structured procurement demand vector, a hybrid collaborative filtering algorithm is used to calculate the supplier matching degree through matrix factorization technology, and the profile matrix of S103 is integrated for weighting to generate a primary supplier candidate set. Based on the structured procurement demand vector, a user-project scoring matrix is ​​constructed. Pearson correlation coefficient is used to calculate user similarity, and cosine similarity is used to calculate project similarity. The mixed weight ratio is set to 0.6:0.4. The number of potential factors is set to 8 dimensions during matrix decomposition. In the weighting stage of the profile matrix, the credit dimension is set to 0.5, the capability dimension to 0.3, and the risk dimension to 0.2. The comprehensive matching degree is calculated through a weighted formula. For example, a steel supplier with a credit dimension score of 0.85, a capability dimension score of 0.78, and a risk dimension score of 0.15, after weighted calculation, obtains a comprehensive matching value of 0.72, and enters the top 10 of the candidate set.

[0028] S203: Based on the initial supplier candidate set, a multi-objective optimization model is constructed. The NSGA-II algorithm is used to balance price, delivery time, and quality indicators. The TOPSIS method is used to rank the solutions and generate a procurement requirement solution with weighted scores.

[0029] The multi-objective optimization model sets the price weight to 0.4, the delivery time weight to 0.35, and the quality weight to 0.25. The NSGA-II algorithm initializes the population size to 100, with a crossover probability of 0.8 and a mutation probability of 0.1. When performing non-dominated sorting, a crowding calculation threshold of 0.5 is set. When the TOPSIS method determines the ideal solution, the lowest price is taken for the price dimension, the shortest delivery time is taken for the delivery time dimension, and the highest score is taken for the quality dimension. The opposite extreme value is taken for the anti-ideal solution. For example, in a batch of procurement plans, plan A has a price score of 0.85, a delivery time of 0.75, and a quality score of 0.9. After calculation, the proximity is 0.68, which is higher than the 0.62 of plan B. Finally, a list of preferred plans with weighted scores is generated.

[0030] See Figure 4 The generation of a multidimensional constrained optimal pricing strategy set based on S3 includes the following steps: S301: Based on historical transaction database, using the extreme gradient boosting tree model, the SHAP value interpretation method is used to identify price influencing factors and generate a market benchmark price prediction curve. Based on historical transaction databases, a time-series data warehouse was established, and a feature engineering pipeline was configured. Price, transaction volume, and supplier rating fields from the past 36 months of transaction records were selected. The data sliding window was set to a quarterly rolling mode. During the training phase of the extreme gradient boosting tree model, the tree depth was set to 6 layers, the learning rate was 0.05, and the SHAP value was calculated using 500 permutations and combinations. When the feature contribution exceeded the industry benchmark value by 15%, it was marked as a key influencing factor. For example, in the price fluctuation analysis of a certain chemical raw material, the SHAP value of the raw material cost index feature reached 0.78, which was significantly higher than the 0.23 of the transportation cost feature, forming a price prediction feature set containing the weights of the main fluctuation factors.

[0031] S302: Based on the market benchmark price prediction curve, construct an incomplete information game model, use the Monte Carlo tree search algorithm to simulate supplier pricing strategies, and generate a dynamic pricing game tree; Based on the market benchmark price prediction curve, a supplier behavior profile database is constructed. The pricing strategy space is defined as a price fluctuation range of ±8%. The game rounds are set to 5 iterations. The Monte Carlo tree search phase is configured with an exploration factor of 0.5, and the number of simulations is 2000. The node expansion threshold is set to a win rate of over 55%. During backtracking updates, a weighted average method is used to adjust the node value. For example, in a steel procurement game, supplier A obtains an expected return assessment value of 0.68 at the third round pricing strategy node, which is higher than supplier B's 0.52, forming a game tree structure that includes the return assessment values ​​of each strategy path.

[0032] S303: Based on the dynamic price game tree, a distributed constraint optimization algorithm is designed. The Lagrange multiplier method is used to solve the multi-objective optimal solution, and the algorithm is connected to the thousand customers thousand prices rule engine for strategy correction to generate a multi-dimensional constraint optimal price strategy set.

[0033] The distributed constraint optimization algorithm sets a price fluctuation tolerance of ±3%, a delivery time flexibility range of ±2 days, a quality pass rate minimum of 98%, an initial Lagrange multiplier of 0.5, an iteration step size of 0.1, and sets the cost target priority at 0.6, delivery time at 0.3, and quality at 0.1 in the multi-objective weight allocation stage. The thousand-customer-one-price rule engine loads the customer's historical purchase volume tier thresholds, sets the discount coefficient for VIP customers at 0.92, and for ordinary customers at 0.98. For example, after optimization, the price of a certain mechanical equipment procurement plan is 95% of the benchmark price. The final strategy entry is generated after adjusting the customer level coefficient.

[0034] See Figure 5 Generating legally valid electronic contract drafts based on S4 includes the following steps: S401: Based on a contract template library, a bidirectional long short-term memory network-conditional random field model is adopted to parse legal clauses and generate a set of structured contract elements through a domain-adaptive training method. Based on a contract template library, a legal clause annotation system was established, and a domain-adaptive training dataset was configured, including 15 types of document templates such as procurement contracts and service agreements. The bidirectional long short-term memory network was set with a hidden layer dimension of 128, and the conditional random field annotation adopted the BIOES system. In the data preprocessing stage, clause segmentation rules were implemented, and a regular expression matching mode was set for clauses of the "payment method" type to extract the amount value and payment cycle parameters. For example, in a certain equipment procurement contract, the clause "90% of the total contract amount shall be paid within 30 days of delivery" is parsed by the model to generate structured fields such as payment ratio and payment period days, forming a standard element set containing subject information, performance terms, and liability for breach of contract.

[0035] S402: Based on a set of structured contract elements, a smart contract verification chain is constructed. A Merkle tree structure is used to verify the compliance of the terms, and the national cryptographic SM3 algorithm is applied to generate a digital digest and a legal compliance verification report. Based on the set of structured contract elements, a smart contract verification chain node is constructed. A standard clause library containing 200 compliance rules is defined. When constructing the Merkle tree, the hash value of the leaf node is set to be a combination of the clause type code and the content summary. The hash of the parent node adopts a cascading calculation method. A similarity threshold of 0.85 is set in the compliance verification stage. When the clause content matches the standard library below the threshold, an early warning is triggered. The digital digest generation adopts the national cryptographic SM3 algorithm. The full text of the contract is divided into blocks, and the compressed value is calculated iteratively for each 512-byte data block. For example, a technical service agreement is calculated to generate a 64-bit hexadecimal digest value, which is then compared and verified with the blockchain evidence storage system.

[0036] S403: Based on the legal compliance verification report, trigger the process routing decision tree, dynamically load approval rules through the random forest classification algorithm, and integrate electronic signature services to generate a legally valid electronic contract draft.

[0037] The process routing decision tree sets risk level classification standards, using contract amount, partner credit score, and number of abnormal terms as classification features. During random forest model training, the tree depth is set to 8 layers, and the minimum number of samples for node splitting is 50. In the approval rule loading stage, a three-level approval process is defined. When the contract risk score exceeds 0.7, the legal review stage is triggered. In the electronic signature service integration stage, a two-way digital certificate verification mechanism is configured. The timestamp server uses the NTP protocol for synchronization. For example, after a high-risk procurement contract is classified, a four-level approval process of "handler - department head - legal affairs - general manager" is loaded. The time difference between the signatures at each node is controlled to complete the entire process within 300 seconds.

[0038] SeeFigure 6 The process of generating adaptive procurement decision recommendations based on S5 includes the following steps: S501: Based on IoT sensor data streams, Kalman filtering algorithm is used for real-time data cleaning to build a digital twin of the supply chain and generate a real-time mirror model of the supply chain. Based on IoT sensor data streams, a data acquisition frequency configuration table is established, setting the sampling interval to 5 seconds for temperature sensors, 10 seconds for humidity sensors, and 1 second for vibration sensors. In the Kalman filtering stage, the process noise covariance matrix Q is defined as a diagonal matrix, and the observation noise covariance matrix R is dynamically adjusted according to the sensor accuracy. In the data cleaning stage, a residual threshold of 3σ is set. When the deviation between the predicted value and the observed value exceeds three times the standard deviation, data correction is triggered. For example, if the temperature data of a cold chain logistics vehicle shows abnormal fluctuations of ±5℃, a smooth curve is generated after filtering. A digital twin 3D model containing warehouse inventory, transport vehicle location, and production line consumption rate is constructed.

[0039] S502: Based on the real-time mirror model of the supply chain, a multi-agent deep reinforcement learning framework is designed, and a near-end policy optimization algorithm is used for policy training to generate a dynamic procurement policy matrix. Based on the real-time mirror model of the supply chain, an intelligent agent interaction protocol is designed. The action space of the procurement intelligent agent is defined as the order quantity adjustment range of ±20%. The response time of the supplier intelligent agent is discretized into 8 levels. The near-end strategy optimization algorithm sets the hidden layer of the strategy network to 256 neurons, the learning rate of the value function network to 0.0003, and the strategy update interval to 50 time steps. During the training phase, a course learning mechanism is adopted to gradually increase the fluctuation range of market demand. For example, in a certain electronic product procurement scenario, when the raw material price increases by 15%, the intelligent agent automatically reduces the order quantity by 12%, forming a three-dimensional strategy matrix that includes emergency procurement strategy, regular procurement strategy, and strategic reserve strategy.

[0040] S503: Based on the dynamic procurement strategy matrix, a feedback adjustment mechanism is constructed, and the supplier evaluation model parameters are updated through the gradient backpropagation algorithm to generate adaptive procurement decision suggestions.

[0041] The feedback adjustment mechanism establishes a parameter update channel and sets a learning rate decay strategy for the supplier evaluation model. The initial value is 0.01, and the decay coefficient is 0.95 every 24 hours. The momentum optimizer is used in the gradient backpropagation stage with a momentum factor of 0.9. The parameter update threshold is set to the loss function decay rate exceeding 5%. For example, if a supplier delays delivery three times in a row, the delivery reliability weight in its evaluation model is automatically adjusted from 0.6 to 0.4. The order allocation ratio in the procurement strategy is updated synchronously, and a decision suggestion table containing dynamic weight adjustment rules and real-time strategy optimization is generated.

[0042] See Figure 7The intelligent procurement collaboration system includes the following modules: a supplier profiling module, which uses graph convolutional neural network algorithm to construct a supply relationship topology based on enterprise ERP system data, calculates node influence through dynamic weight PageRank algorithm, and integrates credit / capability / risk indicators using three-dimensional feature fusion technology to generate a supplier three-dimensional evaluation matrix. The supplier profiling module includes a data cleaning submodule, a graph embedding submodule, and a feature fusion submodule. The intelligent procurement module, based on a supplier three-dimensional evaluation matrix, uses a hybrid collaborative filtering algorithm for demand matching, balances price / delivery time / quality indicators through a multi-objective optimization model, and connects to a thousand customers, thousand prices rule engine to generate dynamic quotations and generate multi-constraint optimal procurement solutions. The intelligent procurement module includes a demand analysis submodule, a strategy optimization submodule, and a dynamic pricing submodule; The contract execution module, based on a multi-constraint optimal procurement scheme, uses a bidirectional long short-term memory network-conditional random field model to parse contract terms, verifies the compliance of terms through a smart contract verification chain, and triggers a multi-level approval routing algorithm to generate a legally valid execution contract. The contract execution module includes a terms parsing submodule, a compliance verification submodule, and a process routing submodule.

[0043] By constructing a supply relationship topology graph using graph convolutional neural network algorithms, discrete data in enterprise ERP systems are deeply correlated, breaking through the limitations of single-dimensional data in traditional supplier evaluation. This achieves three-dimensional dynamic fusion of credit, capability, and risk indicators. The dynamic weighted PageRank algorithm introduces time decay factors and industry-specific parameters, making the calculation of supplier node influence more aligned with actual business scenarios and addressing the shortcomings of static evaluation models in reflecting market changes. The hybrid collaborative filtering algorithm, combined with a multi-objective optimization model, simultaneously optimizes price, delivery time, and quality indicators during the demand matching stage. Pareto front analysis is used to screen the optimal solution set, eliminating suboptimal decision-making problems caused by fragmented indicators in traditional procurement strategies. The personalized pricing rule engine embeds a dynamic pricing mechanism, generating differentiated pricing strategies based on customers' historical procurement data and market fluctuations, overcoming the drawbacks of fixed pricing models' slow market response. The bidirectional long short-term memory network-conditional random field model performs context-aware parsing of contract terms, and combined with a smart contract verification chain, it achieves real-time cross-verification of clause compliance. Through a multi-level approval routing algorithm, it dynamically adapts to enterprise internal control rules, significantly improving contract execution efficiency and legal risk prevention capabilities.

[0044] The data cleaning submodule, based on the original data of the enterprise ERP system, adopts a multi-source data cleaning framework, extracts key fields through regular expression matching algorithms, and uses an isolated forest anomaly detection model to filter noisy data and generate standardized supplier structure data. Based on the original data from the enterprise ERP system, a multi-source data access channel was established, and a field mapping rule table was configured. Chinese regular expression validation was implemented for the supplier name field, 18-digit numeric validation was performed for the unified social credit code field, and the address field was split into provincial, municipal, and district-level administrative divisions. During the data cleaning stage, field integrity validation rules were set. When the missing rate of required fields exceeded the set threshold, the data compensation process was triggered. The anomaly detection stage adopted the isolated forest model and set a threshold for sample path length difference. When the ratio of a data point to the average path length of trees in the forest exceeded three times the standard deviation, it was marked as an anomaly. For example, if a supplier's on-time delivery rate remained at 100% for 12 consecutive months, it would be transferred to the manual review queue after model detection. Finally, a supplier master data table containing a unified coding system and standard fields was formed.

[0045] The graph embedding submodule constructs a supply relationship topology graph based on standardized supplier structure data, uses a graph convolutional neural network algorithm to propagate node features, and applies the TransE embedding algorithm to generate vector representations, thus generating supplier relationship graph embedding features. Based on standardized supplier structure data, a supply relationship graph node attribute table is constructed. The node feature vector is defined to include the logarithmic transformation value of registered capital, segmented coding of the year of establishment, and industry classification embedding vector. The relationship weight is calculated using the moving average of transaction amount over the past three years. When applying the graph convolutional neural network, two layers of feature aggregation are set. The first layer aggregates the features of direct trading partners, and the second layer extends to secondary related nodes. The boundary distance parameter is set during the TransE embedding training phase. When the triplet scoring function value exceeds the set threshold, the embedding vector is updated. For example, after two feature propagations, the cooperation closeness vector value of a certain auto parts supplier node with the OEM increases to the top 10% level in the industry.

[0046] The feature fusion submodule, based on the embedded features of the supplier relationship graph, designs a dynamic weight PageRank algorithm, adjusts the historical cooperation weights through a time decay factor, and combines it with the Sigmoid function normalization process to generate a three-dimensional supplier evaluation matrix.

[0047] The dynamic weighted PageRank algorithm introduces a time decay coefficient. The weight calculation of historical cooperation records adopts an exponential decay formula. The weight retention coefficient of the cooperation record in the current year is set as the baseline value, and it decays by 15% for each year back. The initial weight setting of nodes is logarithmically standardized by combining registered capital and average annual transaction volume. In the Sigmoid function parameter adjustment stage, the credit dimension scaling factor is set to 1.2 times the industry average, the capability dimension to 1.5 times, and the risk dimension to 0.8 times. During normalization, suppliers with credit scores exceeding the industry average by 20% are non-linearly amplified. For example, an electronic component supplier is calculated to have three-dimensional coordinate values ​​of 0.88 for the credit dimension, 0.79 for the capability dimension, and 0.18 for the risk dimension. These are then weighted and fused to form the decision vector in the evaluation matrix.

[0048] The requirement parsing submodule, based on the procurement requirement document library, adopts a bidirectional long short-term memory network-attention mechanism model, extracts technical parameters through a domain dictionary-enhanced word segmentation algorithm, and generates a structured procurement requirement vector. Based on the procurement requirement document library, a domain-specific terminology dictionary was established. For the mechanical manufacturing field, specialized terms such as bearing models and tolerance grades were added. A multi-threaded word segmentation engine was configured, and special symbol filtering rules were implemented in the preprocessing stage. Regular expression pattern matching was performed on the technical parameter description fields. The feature vector dimension was set to 64 dimensions, the bidirectional long short-term memory network hidden layer was set to 128 neurons, and the attention mechanism layer was configured with a weight threshold of 0.7. When the feature importance score exceeds the threshold, the parameter is retained. For example, in a machine tool procurement requirement, "spindle radial runout ≤ 0.005mm" is parsed and mapped to a feature value of the precision dimension, forming a standardized requirement vector containing material specifications, process requirements, and inspection standards.

[0049] The strategy optimization submodule, based on the structured procurement demand vector, uses a hybrid collaborative filtering algorithm, calculates the matching degree through matrix factorization technology, and integrates the supplier three-dimensional evaluation matrix for weighted sorting to generate a candidate supplier score list. Based on the structured procurement demand vector, a user-project interaction matrix is ​​constructed. User similarity is calculated using improved cosine similarity, and project similarity uses the Jaccard coefficient. The mixed weight ratio is set to 0.55:0.45. During the matrix decomposition stage, the potential factor dimension is set to 10. When weighting the supplier evaluation matrix, the credit dimension has a weight of 0.5, the capability dimension 0.35, and the risk dimension 0.15. The comprehensive matching value is calculated using a weighted formula. For example, a steel supplier has a credit dimension score of 0.88, a capability dimension score of 0.72, and a risk dimension score of 0.18, resulting in a comprehensive score of 0.78. This places the supplier in the top 15% of the candidate set, forming a supplier list sorted in descending order of comprehensive score.

[0050] The dynamic pricing submodule constructs a multi-objective optimization model based on the candidate supplier rating list, uses a non-dominated sorting genetic algorithm to balance price / delivery time indicators, and connects to the customer-specific pricing rule engine to generate a multi-constraint optimal procurement plan.

[0051] The multi-objective optimization model sets the price objective weight to 0.45, the delivery time objective to 0.4, and the quality objective to 0.15. The non-dominated sorting genetic algorithm initializes the population size to 120, with a crossover probability of 0.85 and a mutation probability of 0.15. When performing fast non-dominated sorting, the crowding calculation threshold is set to 0.6. The "one thousand customers, one price" rule engine loads the customer's historical purchase frequency classification, setting the price fluctuation coefficient for VIP customers to 0.9, regular customers to 0.95, and new customers to 1.0. For example, in a certain batch of procurement plans, supplier A's price is 8% lower than the benchmark price and the delivery time is 3 days longer than the standard. After multi-objective optimization, a preferred solution with a comprehensive score of 0.68 is generated.

[0052] The clause parsing submodule, based on the contract template library, adopts a bidirectional long short-term memory network-conditional random field model and uses a domain-adaptive training method to parse legal clauses and generate a set of structured contract elements. Based on a contract template library, a legal clause classification system is established, and 15 types of structured templates for contract documents are configured. The bidirectional long short-term memory network is set with 256 hidden layer nodes, and the conditional random field annotation adopts a four-layer label structure. Clause splitting rules are implemented in the data preprocessing stage. For clauses of the "payment method" type, a regular expression for amount extraction is set to match percentage values ​​and payment cycle parameters. For example, in a certain engineering contract, the clause "30% prepayment shall be paid within 7 working days after the contract is signed" generates structured fields such as prepayment ratio and payment time limit after model parsing, forming a standard element set containing contract subject information, performance conditions, and dispute resolution mechanisms.

[0053] The compliance verification submodule constructs a smart contract verification chain based on a set of structured contract elements, uses a Merkle tree structure for clause comparison, applies the national cryptographic SM3 algorithm to generate digital digests, and generates a legal compliance verification report. Based on the set of structured contract elements, a smart contract verification node network is constructed. A compliance verification rule base containing 200 standard clauses is defined. When constructing the Merkle tree, the hash value of the leaf node is set to be a combination of the clause type code and the key field summary. The hash of the parent node is calculated iteratively. The difference tolerance is set to 15% during the compliance comparison stage. When the key field of the clause is missing or the conflict exceeds the tolerance, an early warning record is generated. The digital summary generation adopts a block processing mechanism. Each block of data is processed three times to generate a compressed value. For example, a lease contract is processed to generate a unique summary identifier, which is then checked for consistency with the record in the evidence storage system.

[0054] The process routing submodule, based on the legal compliance verification report, triggers a multi-level approval routing algorithm, dynamically loads approval rules through a random forest classification model, integrates electronic signature services, and generates legally binding enforcement contracts.

[0055] The process routing decision model sets up a risk level assessment matrix, divides contract amount thresholds into five levels, sets three intervals for partner credit scores, sets four-level early warning standards for the number of abnormal clauses, sets a feature importance screening threshold of 0.6 during random forest model training, uses the Gini coefficient index for node splitting, defines a three-level acceleration channel and a two-level review process for the approval rule configuration module, integrates a digital certificate chain verification mechanism for electronic signature service, and controls the accuracy of timestamp service to the millisecond level. For example, if a cross-border procurement contract exceeds the threshold and has three clause differences, it triggers a four-level approval process of "business department - risk control department - legal department - executive director", and the electronic signature process completes multi-party signing within 120 seconds.

[0056] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An intelligent procurement collaboration method, characterized in that... Includes the following steps: S1: Based on the fusion of multi-source heterogeneous data, a supplier knowledge graph is constructed using a graph embedding algorithm. The Neo4j graph database is used to realize the associated storage of business information, performance records, and quality reports. The PageRank improved algorithm is used to calculate the supplier node weights and generate a three-dimensional supplier profile matrix. S2: Based on a three-dimensional supplier profile matrix, an improved collaborative filtering algorithm is used for demand matching. The ElasticSearch semantic analysis engine is used to parse the procurement demand documents, and TF-IDF weighted cosine similarity calculation is combined to realize intelligent recommendation and generate procurement demand solutions with weighted scores. S3: Based on the procurement scheme with weighted scoring, a dynamic pricing game model is used, the Monte Carlo tree search algorithm is used to simulate the bidding strategy, the XGBoost regression model is used to predict the market benchmark price, and the thousand customers thousand prices rule engine is connected to generate a multi-dimensional constrained optimal bidding strategy set. S4: Based on the multi-dimensional constraint optimal pricing strategy set, the smart contract-driven process is initiated. The BiLSTM-CRF contract terms parsing model is adopted, and the compliance verification of the terms is realized through the digital signature timestamp chain. A custom approval routing algorithm is triggered to generate a legally valid electronic contract draft. S5: Based on legally binding electronic contract drafts, construct a digital twin of the supply chain, adopt a multi-agent reinforcement learning framework, adjust procurement strategies in real time through PPO strategy optimization algorithms, and feed them back to the supplier evaluation model to generate adaptive procurement decision suggestions.

2. The intelligent procurement collaboration method according to claim 1, characterized in that: The steps for generating a three-dimensional supplier profile matrix based on S1 are as follows: S101: Based on the integration with the enterprise ERP system, a multi-source data cleaning framework is adopted. The basic information of suppliers is extracted through regular expression matching algorithm, and the isolated forest anomaly detection model is used to filter invalid data and generate standardized supplier structure data. S102: Based on standardized supplier structure data, a graph convolutional neural network is used to construct a supply relationship topology graph through a node feature propagation mechanism, and the TransE embedding algorithm is applied to generate vector representations to generate supplier relationship graph embedding features; S103: Based on the embedded features of the supplier relationship graph, a dynamic weighted PageRank algorithm is designed. The influence of historical cooperation records is adjusted by the time decay factor and normalized by the Sigmoid function to generate a three-dimensional supplier profile matrix.

3. The intelligent procurement collaboration method according to claim 1, characterized in that: Generating a weighted, scored procurement requirement plan based on S2 involves the following steps: S201: Based on the procurement requirement document library, the BiLSTM-Attention model is used to extract technical parameter features through a word segmentation algorithm enhanced by a domain dictionary, and a structured procurement requirement vector is generated. S202: Based on the structured procurement demand vector, a hybrid collaborative filtering algorithm is used to calculate the supplier matching degree through matrix factorization technology, and the profile matrix of S103 is integrated for weighting to generate a primary supplier candidate set. S203: Based on the initial supplier candidate set, a multi-objective optimization model is constructed. The NSGA-II algorithm is used to balance price, delivery time, and quality indicators. The TOPSIS method is used to rank the solutions and generate a procurement requirement solution with weighted scores.

4. The intelligent procurement collaboration method according to claim 1, characterized in that: Generating a set of multidimensional constrained optimal pricing strategies based on S3 includes the following steps: S301: Based on historical transaction database, using the extreme gradient boosting tree model, the SHAP value interpretation method is used to identify price influencing factors and generate a market benchmark price prediction curve. S302: Based on the market benchmark price prediction curve, construct an incomplete information game model, use the Monte Carlo tree search algorithm to simulate supplier pricing strategies, and generate a dynamic pricing game tree; S303: Based on the dynamic price game tree, a distributed constraint optimization algorithm is designed. The Lagrange multiplier method is used to solve the multi-objective optimal solution, and the algorithm is connected to the thousand customers thousand prices rule engine for strategy correction to generate a multi-dimensional constraint optimal price strategy set.

5. The intelligent procurement collaboration method according to claim 1, characterized in that: Generating a legally valid electronic contract draft based on S4 involves the following steps: S401: Based on a contract template library, a bidirectional long short-term memory network-conditional random field model is adopted to parse legal clauses and generate a set of structured contract elements through a domain-adaptive training method. S402: Based on a set of structured contract elements, a smart contract verification chain is constructed. A Merkle tree structure is used to verify the compliance of the terms, and the national cryptographic SM3 algorithm is applied to generate a digital digest and a legal compliance verification report. S403: Based on the legal compliance verification report, trigger the process routing decision tree, dynamically load approval rules through the random forest classification algorithm, and integrate electronic signature services to generate a legally valid electronic contract draft.

6. The intelligent procurement collaboration method according to claim 1, characterized in that: The process of generating adaptive purchasing decision recommendations based on S5 includes the following steps: S501: Based on IoT sensor data streams, Kalman filtering algorithm is used for real-time data cleaning to build a digital twin of the supply chain and generate a real-time mirror model of the supply chain. S502: Based on the real-time mirror model of the supply chain, a multi-agent deep reinforcement learning framework is designed, and a near-end policy optimization algorithm is used for policy training to generate a dynamic procurement policy matrix. S503: Based on the dynamic procurement strategy matrix, a feedback adjustment mechanism is constructed, and the supplier evaluation model parameters are updated through the gradient backpropagation algorithm to generate adaptive procurement decision suggestions.

7. An intelligent procurement collaboration system, characterized in that... It includes the following modules: Supplier profiling module, which uses graph convolutional neural network algorithm to construct supply relationship topology based on enterprise ERP system data, calculates node influence through dynamic weight PageRank algorithm, and integrates credit / capability / risk indicators using three-dimensional feature fusion technology to generate supplier three-dimensional evaluation matrix; The supplier profiling module includes a data cleaning submodule, a graph embedding submodule, and a feature fusion submodule. The intelligent procurement module, based on a supplier three-dimensional evaluation matrix, uses a hybrid collaborative filtering algorithm for demand matching, balances price / delivery time / quality indicators through a multi-objective optimization model, and connects to a thousand customers, thousand prices rule engine to generate dynamic quotations and generate multi-constraint optimal procurement solutions. The intelligent procurement module includes a demand analysis submodule, a strategy optimization submodule, and a dynamic pricing submodule; The contract execution module, based on a multi-constraint optimal procurement scheme, uses a bidirectional long short-term memory network-conditional random field model to parse contract terms, verifies the compliance of terms through a smart contract verification chain, and triggers a multi-level approval routing algorithm to generate a legally valid execution contract. The contract execution module includes a terms parsing submodule, a compliance verification submodule, and a process routing submodule.

8. The intelligent procurement collaboration system according to claim 7, characterized in that: The data cleaning submodule, based on the original data of the enterprise ERP system, adopts a multi-source data cleaning framework, extracts key fields through regular expression matching algorithms, and uses an isolated forest anomaly detection model to filter noisy data and generate standardized supplier structure data. The graph embedding submodule constructs a supply relationship topology graph based on standardized supplier structure data, uses a graph convolutional neural network algorithm to propagate node features, and applies the TransE embedding algorithm to generate vector representations, thus generating supplier relationship graph embedding features. The feature fusion submodule, based on the embedded features of the supplier relationship graph, designs a dynamic weight PageRank algorithm, adjusts the historical cooperation weights through a time decay factor, and combines it with the Sigmoid function normalization process to generate a three-dimensional supplier evaluation matrix.

9. The intelligent procurement collaboration system according to claim 7, characterized in that: The requirement parsing submodule, based on the procurement requirement document library, adopts a bidirectional long short-term memory network-attention mechanism model, extracts technical parameters through a domain dictionary-enhanced word segmentation algorithm, and generates a structured procurement requirement vector. The strategy optimization submodule, based on the structured procurement demand vector, uses a hybrid collaborative filtering algorithm, calculates the matching degree through matrix factorization technology, and integrates the supplier three-dimensional evaluation matrix for weighted sorting to generate a candidate supplier score list. The dynamic pricing submodule constructs a multi-objective optimization model based on the candidate supplier rating list, uses a non-dominated sorting genetic algorithm to balance price / delivery time indicators, and connects to the customer-specific pricing rule engine to generate a multi-constraint optimal procurement plan.

10. The intelligent procurement collaboration system according to claim 7, characterized in that: The clause parsing submodule, based on the contract template library, adopts a bidirectional long short-term memory network-conditional random field model and uses a domain-adaptive training method to parse legal clauses and generate a set of structured contract elements. The compliance verification submodule constructs a smart contract verification chain based on a set of structured contract elements, uses a Merkle tree structure for clause comparison, applies the national cryptographic SM3 algorithm to generate digital digests, and generates a legal compliance verification report. The process routing submodule, based on the legal compliance verification report, triggers a multi-level approval routing algorithm, dynamically loads approval rules through a random forest classification model, integrates electronic signature services, and generates legally binding enforcement contracts.

Images