A method for multi-level weight distribution of signing a project by fusing a supplier portrait and order quality

Abstract

The application belongs to the field of intelligent management of supply chain, and relates to a signing project multi-level weight distribution method combining supplier portrait and order quality. It comprises: generating an order complexity semantic fingerprint with key dimension annotation; generating a supplier capability summary vector with timestamp; constructing a double-flow network topology containing positive activation and negative inhibition connection with the fingerprint as the query node and the summary vector as the candidate node; intercepting the local subgraph of the hit rule for feature alignment, and outputting the adaptation strength score of each supplier in each key dimension; performing adaptive threshold clipping and elastic weight redistribution according to the adaptation strength distribution; and fusing the elastic weight and the supplier basic capability score to generate a comprehensive ranking score and execute a recommendation decision. The application realizes dynamic alignment and adaptive matching of order complexity and supplier capability, and improves matching accuracy and real-time reasoning efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent supply chain management, and in particular to a multi-level weight allocation method for contract signing and project initiation that integrates supplier profiling and order quality. Background Technology

[0002] In existing supplier matching decision-making methods, purchase orders typically use structured fields (such as product attributes, delivery time, and customization level) as the main body of demand description, while suppliers use static profile indicators (such as historical fulfillment rate, capacity benchmark value, and qualification level) as the evaluation basis. When orders involve complex constraints implicit in unstructured technical documents (such as process notes and drawing instructions), this information is difficult to automatically incorporate into the matching calculation system.

[0003] In the matching calculation stage, the importance weights of each dimension of the order are mostly pre-set fixed values, and are not dynamically adjusted according to the actual demand distribution of the current order and the real-time capability distribution of the supplier group. Some graph-based methods require periodic retraining of all nodes to update edge weights, which is difficult to meet the real-time inference requirements of large-scale order scenarios.

[0004] For highly complex orders involving multiple differentiated constraint dimensions, simple weighted summation or linear embedding methods are insufficient to achieve a focused assessment of the core constraints of the order, and the matching results are easily affected by non-critical dimension indicators. Summary of the Invention

[0005] This application provides a multi-level weighting method for contract signing and project initiation that integrates supplier profiles and order quality. It aims to solve the problem of how to achieve accurate and dynamic matching between the complexity of purchase orders and the status of supplier capabilities based on key dimensions while ensuring real-time response capabilities.

[0006] The multi-level weighting method for contract signing and project initiation that integrates supplier profiles and order quality, provided in this application, specifically includes:

[0007] Obtain the structured features and unstructured semantic information of newly accessed purchase orders, and generate a unique semantic fingerprint of order complexity. The semantic fingerprint of order complexity is used to characterize the position of the order in the multi-dimensional feature space and to label key dimensions.

[0008] Based on the time window associated with the semantic fingerprint of the order complexity, real-time capability data of suppliers is collected and time-series compressed to generate a timestamped supplier capability summary vector.

[0009] Based on the mapping rules inferred from historical matching samples, the order complexity semantic fingerprint is used as the query node and the supplier capability summary vector is used as the candidate node to construct a two-stream network topology containing positive connections and negative suppression connections.

[0010] Based on the supplier nodes and their neighborhoods hit by the fine-grained mapping rules in the dual-stream network topology, the corresponding subgraph structure is extracted as a local subgraph, and local feature alignment is performed through feature propagation to output the adaptation strength score of each supplier in the key dimensions.

[0011] Based on the distribution of the adaptation strength score across dimensions, adaptive threshold pruning is performed to identify advantageous suppliers and generate an adjusted elastic weight distribution sequence.

[0012] Obtain the supplier's basic capability score, and then merge the weight distribution sequence with the supplier's basic capability score to generate a comprehensive ranking score that represents the final matching degree between the order and the supplier.

[0013] Supplier recommendation decisions are made based on the comprehensive ranking score. If the performance risk is detected to exceed a preset threshold, manual review is triggered and the decision log is recorded.

[0014] The multi-level weighting method for contract signing and project initiation that integrates supplier profiling and order quality, provided in this application, has the following beneficial effects:

[0015] By generating semantic fingerprints of order complexity with key dimension annotations, the process difficulties and dependencies implicit in unstructured technical documents are incorporated into the formal expression of order requirements. This allows the core constraints of the order to be automatically focused on rather than treated uniformly during matching calculations. By generating timestamped supplier capability summary vectors, the real-time capacity fluctuations and historical performance stability of suppliers are incorporated into capability representation, expanding the matching basis from static profiles to dynamic state descriptions. Based on this, fine-grained mapping rules summarized from historical matching samples drive the construction of positive activation edges and negative inhibition connections in the dual-stream network topology, achieving refined alignment between key order dimensions and supplier capability indicators, avoiding subjective bias from manually setting edge weights. Simultaneously, adaptive threshold pruning and elastic weight redistribution are performed based on the distribution variance of the adaptation strength scores across dimensions, dynamically adjusting the weight distribution according to the actual capability distribution of the current supplier group, preventing a single advantageous dimension from dominating the evaluation results. Furthermore, the elastic weight distribution sequence is nonlinearly fused with the supplier's basic capability score. While retaining the supplier's static qualification information, the coupling gain interaction between dimensions is introduced, so that the comprehensive ranking score can reflect the synergistic effect when multiple dimensional constraints coexist. The local subgraph truncation and feature propagation mechanism allows graph attention reasoning to be performed only on the subset of suppliers that hit the rules. Combined with the performance risk detection and manual review triggering in the recommendation decision-making process, as well as the continuous feedback of the decision log to the rule base, an overall balance is formed in matching accuracy, computational efficiency and system self-optimization capability. Attached Figure Description

[0016] Figure 1This is the main flowchart of a multi-level weight allocation method for contract signing and project initiation that integrates supplier profiles and order quality.

[0017] Figure 2 It is a flowchart of the matching verification and strong matching relationship determination in the multi-level weight allocation method for signing and initiating projects that integrates supplier profiles and order quality.

[0018] Figure 3 It is a flowchart of the logic of extreme value screening and advantage dimension label generation in the multi-level weight allocation method for signing and initiating projects that integrates supplier profiles and order quality. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0021] like Figure 1 As shown, this application provides a multi-level weighting method for contract signing and project initiation that integrates supplier profiles and order quality, specifically including:

[0022] S1: For newly connected purchase orders, extract four types of structured features: process constraints, delivery rigidity, customization granularity, and collaborative interface type. Combined with the semantic parsing results of unstructured technical documents, generate a unique semantic fingerprint of order complexity. This fingerprint encodes the position coordinates of the order in a multi-dimensional complex space in the form of sparse vectors and marks the key dimensions.

[0023] S2: Based on the time window associated with the semantic fingerprint of order complexity, collect and time-series compressed information on the real-time production line scheduling margin, historical similar order fulfillment deviation rate, engineer skill map coverage, and cross-system integration readiness status indicators of registered suppliers, and generate a timestamped supplier capability summary vector.

[0024] S3: Utilizing a fine-grained mapping rule base automatically summarized from historical high-matching order samples, the system dynamically constructs a two-stream network topology containing activation edges and deweighting edges, using the semantic fingerprint of order complexity as the query node and the supplier capability summary vector as the candidate node.

[0025] S4: Based on the two-stream network topology, only the local subgraph consisting of the supplier nodes and their neighborhoods hit by the fine-grained mapping rules is loaded. The local feature alignment calculation is performed through the two-layer graph attention propagation mechanism, and the adaptation strength score of each supplier on the key dimensions is output.

[0026] S5: Based on the distribution variance of the adaptation strength score across each key dimension, perform a threshold adaptive pruning operation to identify the advantageous dimensions of leading suppliers and generate a flexible weight distribution sequence that releases weights to non-advantageous dimensions.

[0027] S6: Input the elastic weight distribution sequence and the supplier basic capability score into a nonlinear fusion function for weighted aggregation operation to generate a comprehensive ranking score that represents the final matching degree between the order and the supplier.

[0028] S7: Execute supplier recommendation decisions based on comprehensive ranking scores. If the fulfillment risk index of the current order is detected to exceed the preset safety threshold, a manual review mark is triggered and the elastic weight allocation log is recorded for subsequent rule base iterations.

[0029] Step S1: For newly integrated purchase orders, extract four types of structured features: process constraints, delivery rigidity, customization granularity, and collaborative interface type. Combined with the semantic parsing results of unstructured technical documents, generate a unique semantic fingerprint of order complexity. This fingerprint encodes the order's position coordinates in a multi-dimensional complex space as a sparse vector and labels key dimensions. Specifically, this includes:

[0030] S1.1: Perform structured field parsing on the raw data stream of newly accessed purchase orders, and extract four types of core structured feature vectors: process constraint parameters, delivery rigidity threshold, customization granularity level, and collaborative interface type, to form a basic feature dataset of orders as input conditions for subsequent semantic fusion.

[0031] The system receives the raw data stream of newly connected purchase orders, parses the JSON format messages transmitted through the ERP system interface, and locates the root node containing process parameters, delivery terms, customization identifiers, and interface definitions. It extracts process constraint parameter fields, parses tolerance grades, surface roughness Ra values, and heat treatment hardness HRC indices, maps non-numerical descriptions to standard quantization ranges, and generates a process complexity feature vector.

[0032] Read the delivery rigidity threshold field, calculate the time deviation ratio between the promised delivery date and the standard production cycle, and construct a scalar of delivery urgency by combining it with the default penalty coefficient. Analyze the customization granularity level, count the proportion of non-standard parts and the number of special assembly processes in the BOM structure, and generate a customization depth index through discretization coding.

[0033] Identify the type of collaborative interface, detect whether the order includes MES system integration, API data feedback, or third-party logistics integration requirements, and convert them into binary state vectors. Perform Z-Score standardization on the above four types of feature vectors to eliminate differences in units and form a unified order basic feature dataset.

[0034] By parsing structured fields and standardizing mapping, heterogeneous raw order data is transformed into a computable dataset of basic order features, enabling underlying data alignment for subsequent multimodal semantic fusion.

[0035] For example, parsing a precision structural component purchase order, the tolerance grade IT6 is mapped to a value of 0.95, and the surface roughness Ra0.8 is mapped to 0.85, generating a process feature vector [0.95, 0.85]. The delivery cycle deviation ratio is calculated to be 1.2, with a penalty coefficient of 0.8, resulting in a delivery scalar of 0.96. The proportion of non-standard parts is 30%, with 2 special processes and a coding customization index of 0.7. An MES integration requirement is detected, generating an interface vector [1, 0]. After Z-score standardization, a four-dimensional basic feature dataset [0.8, 0.9, 0.7, 1.0] is formed, significantly improving feature comparability.

[0036] S1.2: Based on the unstructured technical documents associated with the order basic feature dataset, a domain pre-trained language model is used to perform deep semantic parsing operations to extract text fragments containing implicit technical difficulties and implicit dependencies, in order to generate unstructured semantic feature sequences to supplement the missing explicit features.

[0037] Obtain the order basic feature dataset generated by S1.1 and locate the associated set of unstructured technical documents, including CAD drawing annotations, process instructions, and historical change logs. Perform text cleaning and segmentation on the unstructured technical documents, removing noisy characters and retaining technical description fragments with semantic integrity. Load a pre-trained language model fine-tuned for the manufacturing domain, input the cleaned text fragments into the model's encoder layer, and extract deep semantic embedding vectors containing contextual dependencies. Utilize a multi-head attention mechanism to capture implicit descriptions of process difficulties in the text, such as "strict tolerance fit" or "risk of deformation during heat treatment," and generate a semantic feature matrix of technical difficulties. Simultaneously identify implicit dependencies in the text, such as "process A must be completed before assembly B can proceed," construct a dependency graph, and convert it into sequential dependency feature vectors. Perform concatenation and fusion on the semantic feature matrix of technical difficulties and dependency feature vectors, mapping them to a unified-dimensional semantic space through a fully connected layer to generate an unstructured semantic feature sequence representing implicit complexity. By using deep semantic parsing, the missing structured features from the previous step are transformed into a high-dimensional unstructured semantic feature sequence, thereby achieving the expected technical effect of complementary enhancement between explicit data and implicit knowledge.

[0038] For example, considering a precision bearing purchase order, the input consists of a 5-page PDF technical specification document. The model extracts two key challenges: "inner ring roundness tolerance ≤ 0.005mm" and "requires vacuum quenching," generating corresponding 32-dimensional semantic vectors. Simultaneously, it identifies the time dependency "requires 24-hour resting time after grinding," generating a 16-dimensional dependency vector. These two vectors are concatenated and mapped through a fully connected layer to a 48-dimensional unstructured semantic feature sequence. This sequence accurately supplements the heat treatment process constraints not reflected in the structured fields, significantly improving the coverage of subsequent complexity assessments.

[0039] S1.3: Perform multimodal feature alignment and sparse coding mapping on the order basic feature dataset and unstructured semantic feature sequence. Use a dimensionality reduction projection algorithm to compress high-dimensional heterogeneous features into a unified multidimensional complex space coordinate system to generate an initial sparse vector representing the global complexity of the order.

[0040] We obtain the basic feature dataset of orders and the sequence of unstructured semantic features as inputs for multimodal fusion. We construct a unified multidimensional complex spatial coordinate system and define four orthogonal axes: process, delivery, customization, and interface.

[0041] The structured feature vectors are normalized to convert discrete levels into continuous numerical ranges. Embedding encoding is performed on the unstructured semantic feature sequences to generate high-dimensional dense semantic vectors. A cross-modal attention mechanism is used to calculate the contribution weights of the semantic vectors to each dimension.

[0042] Weighted projection maps high-dimensional semantic vectors to their corresponding dimensional components in a unified coordinate system. The structured mapping components and semantic projection components are then concatenated to form an initial high-dimensional hybrid feature matrix. A sparse autoencoder is applied to perform nonlinear dimensionality reduction on the hybrid feature matrix. Sparsity constraint parameters are set to force most neurons to have zero activation values ​​in order to extract key features.

[0043] By optimizing the reconstruction error through backpropagation, a low-dimensional sparse representation matrix is ​​learned. L1 regularization compression is then applied to this low-dimensional sparse representation matrix to remove redundant and noisy dimensions. An initial sparse vector representing the global complexity of the order is generated, and this vector has unique coordinates in the multidimensional space.

[0044] By using multimodal feature alignment and sparse coding mapping, heterogeneous data is transformed into a unified sparse vector, enabling accurate quantization and dimensionality reduction of order complexity in multidimensional space.

[0045] For example, the multidimensional complex space is set to 4 dimensions, and the process constraint level 3 in the structured features is mapped to 0.75. Unstructured text is used to generate a 768-dimensional semantic vector via the BERT model. Cross-modal attention calculation yields a semantic weight of 0.6 for the process dimension. After projection, the process dimension component is 0.45. These are concatenated to form a 4x768 mixture matrix. The sparse autoencoder's hidden layer is set to 64 nodes with a sparsity parameter of 0.05. After training, a 64-dimensional initial sparse vector is output, with only 8 non-zero elements. After L1 regularization, 5 key dimension values ​​are retained. The final generated initial sparse vector accurately reflects the high complexity of the order in the process and interface dimensions, significantly reducing the subsequent computational load.

[0046] S1.4: Perform key dimension identification calculation based on the distribution entropy values ​​of the initial sparse vector in each dimension. By setting a dynamic threshold, select the key dimension labels with the largest deviation to mark the core constraints in the semantic fingerprint of order complexity and correct the weight distribution of the initial sparse vector.

[0047] Receive the initial sparse vector generated in S1.3, which represents the distribution of orders in a multidimensional complex space. Calculate the information entropy value of the initial sparse vector in each dimension to quantify the uncertainty of each feature dimension. The Shannon entropy formula is used for calculation:

[0048]

[0049] in, For information entropy, For the first The proportion of normalized feature values ​​in each dimension This represents the total number of dimensions. A high entropy value indicates that there are significant technical constraints or demand fluctuations in that dimension, and it is identified as a potentially critical dimension.

[0050] A dynamic threshold filtering mechanism is set up to determine the threshold boundary based on the standard deviation of the historical order complexity distribution.

[0051] The calculated entropy values ​​of each dimension are compared with dynamic thresholds, and dimensions whose entropy values ​​exceed the thresholds are selected as key dimensions.

[0052] Generate a set of key dimension labels, annotating specific constraints such as "sensitive to high-precision machining cycles". Based on the key dimension labels, perform exponential enhancement processing on the weight coefficients of the corresponding dimensions in the initial sparse vector, reducing the weight ratio of non-key dimensions.

[0053] Through the entropy calculation and dynamic threshold filtering process described above, the initial sparse vector from the previous step is transformed into semantic fingerprint intermediate data labeled with key dimensions and with corrected weight distribution. This enables accurate positioning and feature enhancement of the core constraints of the order, providing high-confidence input for targeted matching of the subsequent dual-stream network topology.

[0054] For example, for a precision component purchase order, the initial sparse vector includes four dimensions: process, delivery, customization, and interface. The calculated entropy values ​​are: process dimension 2.85, delivery dimension 1.20, customization dimension 0.95, and interface dimension 0.60. A dynamic threshold of 1.50 is set. Process dimension entropy values ​​exceeding the threshold are identified as critical dimensions and labeled "nanometer-level precision machining dependency." The weight coefficient for this dimension is multiplied by an amplification factor of 3.0, and the weight coefficients for the other dimensions are multiplied by a compression factor of 0.8. The corrected vector highlights the complexity of the process, significantly improving the responsiveness to suppliers with high-precision machining capabilities in subsequent matching.

[0055] S1.5: Perform unique hash verification and encapsulation on the corrected initial sparse vector labeled with key dimension tags to generate an order complexity semantic fingerprint with a globally unique identifier attribute, so as to output the standard data object for the final two-stream network topology query node matching.

[0056] Step S2: Based on the time window associated with the semantic fingerprint of order complexity, collect and time-series compressed data on the real-time production line scheduling margin, historical similar order fulfillment deviation rate, engineer skill map coverage, and cross-system integration readiness indicators of registered suppliers to generate a timestamped supplier capability summary vector. Specifically, this includes:

[0057] S2.1: Based on the time window associated with the semantic fingerprint of order complexity, obtain the original capability monitoring data stream of registered suppliers. The original capability monitoring data stream includes real-time production line scheduling margin values, historical similar order fulfillment deviation rate records, engineer skill map coverage list, and cross-system integration readiness status flag. Perform preprocessing operations such as missing value imputation and outlier removal on the original capability monitoring data stream to generate a standardized supplier multidimensional capability time series.

[0058] The S2.1 sub-step aims to acquire and preprocess the raw capability monitoring data stream of registered suppliers based on the time window associated with the semantic fingerprint of order complexity. This step is located at the beginning of the S2 main step and is responsible for providing standardized, highly reliable input data for subsequent time-series feature extraction.

[0059] Receive the semantic fingerprint of order complexity generated by S1, parse the timestamp identifier contained therein, and determine the effective time window range for data collection, which is usually set to a complete production cycle back from the current moment or a preset dynamic duration.

[0060] Based on the defined time window, four types of raw capability monitoring data are retrieved in parallel from the supplier management database and the IoT monitoring platform: real-time production line scheduling margin numerical sequence, historical similar order fulfillment deviation rate record list, engineer skill map coverage tag set, and cross-system integration readiness status flag sequence.

[0061] The system performs integrity checks on the acquired raw data stream, identifies missing values ​​caused by sensor failures or network transmission interruptions, fills in continuous numerical missing values ​​using linear interpolation, and handles discrete status flag missing values ​​using mode imputation, ensuring the temporal continuity of the data sequence.

[0062] An anomaly detection model based on the isolated forest algorithm is applied to identify outliers in real-time production line scheduling margin and fulfillment deviation rate, eliminating noisy data exceeding three standard deviations to prevent extreme values ​​from distorting the true assessment of supplier capabilities.

[0063] The cleaned multidimensional data is standardized using Z-Score to eliminate numerical differences between indicators of different dimensions and to map the data of each dimension to a standard normal distribution space with a mean of 0 and a variance of 1.

[0064] It should be noted that the time window is the specific implementation of the "time window based on semantic fingerprint association of order complexity" mentioned above. Its length is dynamically set according to the order type: for periodic regular orders, the time window can be set to 30 days backward from the current time; for urgent orders or highly customized orders, the time window can be shortened to 7 days or a complete production cycle. The time window is used to limit the scope of supplier capability data collection; historical data exceeding this window is not included in the generation of the supplier capability summary vector.

[0065] Through the above data cleaning and standardization methods, the heterogeneous and noisy raw capability monitoring data stream is transformed into a standardized supplier multidimensional capability time series, which significantly improves data quality and lays a solid data foundation for subsequent sliding window statistics and feature compression.

[0066] For example, the time window is set to the past 30 days. Two missing values ​​were found in the scheduling margin sequence of a supplier's production line; these were imputed using the average of adjacent time points. One abnormal peak in the fulfillment deviation rate (deviating from the mean by 4.5 times the standard deviation) was detected and removed. Z-score standardization was applied to the remaining 98% of the valid data, normalizing the scheduling margin mean to 0 and the standard deviation to 1. The final output is a standard time-series matrix of 720 hours in length, significantly improving the data signal-to-noise ratio and ensuring the accuracy of subsequent feature extraction.

[0067] S2.2: Based on the standardized supplier multidimensional capability time series, the sliding time window statistical algorithm is used to extract trend features from the real-time production line scheduling margin value and the historical similar order fulfillment deviation rate record. The scheduling fluctuation coefficient, which characterizes the capacity fluctuation characteristics, and the mean fulfillment deviation, which characterizes the fulfillment stability, are calculated. The scheduling fluctuation coefficient and the mean fulfillment deviation are used as numerical capability feature components.

[0068] It should be noted that this step, along with subsequent steps S2.3 and S2.4, collectively implements the time-series compression operation. The time-series compression refers to: first, using sliding window statistics to transform the time-series data into numerical features such as scheduling fluctuation coefficients and mean performance deviations; second, using set mapping encoding to transform discrete labels into structured features; and finally, using principal component analysis for dimensionality reduction to fuse and compress the aforementioned multi-source features into a low-dimensional, dense capability feature compression vector, thereby achieving compression from high-dimensional time-series data to low-dimensional capability representation.

[0069] For the real-time production line scheduling margin values ​​in the standardized supplier multidimensional capability time series, a sliding time window statistical algorithm is used to extract local trend features. The time window length is set to T, the step size to 1, and the time series data points are iterated through to calculate the mean and standard deviation within each window, thus quantifying the short-term fluctuation characteristics of capacity.

[0070] A sliding time window statistical algorithm is used to extract local trend features, calculate the mean and standard deviation within the window, and use the coefficient of variation as the scheduling fluctuation coefficient to quantify the dispersion of production line load.

[0071] The same sliding window processing is applied to historical records of similar order fulfillment deviation rates. The arithmetic mean of the fulfillment deviation rates within the window is calculated as a core indicator characterizing the supplier's long-term fulfillment stability. The mean fulfillment deviation directly reflects the degree of deviation in the supplier's on-time delivery rate in similar past tasks; a lower mean indicates higher fulfillment reliability.

[0072] The calculated scheduling fluctuation coefficient and the mean performance deviation are vector-concatenated to form a numerical capability characteristic component. This component retains the dynamic changes in the time series, avoiding the limitation of a single static indicator failing to reflect supplier status fluctuations.

[0073] By using sliding window statistics and coefficient of variation calculation, the standardized time-series data from the previous step is transformed into a scheduling fluctuation coefficient that characterizes capacity fluctuation and a mean performance deviation that characterizes performance stability. This enables a quantitative description of the supplier's dynamic performance capability and provides high-precision numerical input for subsequent multi-source feature fusion.

[0074] For example, setting the sliding time window length T to 7 days, and considering the production line scheduling margin time series data of a precision machining supplier, the 7 sample values ​​within the window are [0.8, 0.75, 0.9, 0.6, 0.85, 0.7, 0.8]. The calculated mean is 0.771, and the standard deviation is 0.098. Using the ratio of the standard deviation to the mean as the scheduling fluctuation coefficient, the result is 0.127. Meanwhile, the supplier's average fulfillment deviation rate for similar orders over the past 7 days is 0.05. Combining 0.127 and 0.05 forms a two-dimensional numerical feature vector [0.127, 0.05]. This vector accurately captures the supplier's state of slight fluctuations in capacity but overall stable fulfillment, significantly improving the accuracy of identifying potential delivery risks compared to directly using the current value.

[0075] S2.3: Based on numerical capability feature components, set mapping encoding is performed on the engineer skill map coverage list to convert discrete skill tags into high-dimensional sparse skill vectors. At the same time, one-hot encoding transformation is performed on the cross-system integration readiness flag to generate an integration state vector representing interface compatibility. The high-dimensional sparse skill vector and the integration state vector are used as structured capability feature components.

[0076] The S2.3 sub-step aims to transform discrete engineer skill tags and cross-system integration states into computable structured feature components to support subsequent multidimensional capability fusion.

[0077] A global engineer skills ontology is constructed, which includes a set of standardized skill tags covering fields such as mechanics, electronics, and software, and assigns unique index IDs to form a basis vector of a high-dimensional sparse space.

[0078] Traverse the supplier's engineer skill map coverage list, extract the currently available valid skill tags, and map them to the corresponding index positions in the global skill ontology.

[0079] A multi-hot encoding mechanism is adopted, which sets 1 at the corresponding index position and 0 at the rest to generate a high-dimensional sparse skill vector representing the breadth of skill coverage.

[0080] Define an enumeration space for cross-system integration readiness states, including four discrete state values: "not integrated", "under testing", "verified", and "discarded".

[0081] Perform one-hot encoding on the cross-system integration ready status flag, activate the corresponding dimension in the four-dimensional vector space according to the current state value, and generate an integration state vector representing interface compatibility.

[0082] The high-dimensional sparse skill vector and the integrated state vector are concatenated to form a unified structured capability feature component matrix.

[0083] Through the above encoding process, non-numerical heterogeneous capability indicators are transformed into dense or sparse numerical vectors that can be processed by computers, eliminating semantic ambiguity and realizing the quantitative expression of supplier soft capability indicators.

[0084] For example, the global skill ontology is set to 500 dimensions. A supplier possesses two skills: "PLC Programming" and "Python Development," corresponding to indices 10 and 205. The generated skill vector has 1s at positions 10 and 205, and 0s elsewhere. Its integration status is "verified," and the corresponding one-hot vector is [0,0,1,0]. These two are concatenated to form a 504-dimensional structured feature vector, which is directly input into the PCA dimensionality reduction module, ensuring accurate calculation of skill matching.

[0085] S2.4: Based on numerical capability feature components and structured capability feature components, principal component analysis dimensionality reduction algorithm is used to fuse and compress multi-source heterogeneous features, eliminate redundant correlations between features and extract key principal components, and generate low-dimensional dense supplier capability feature compression vectors to achieve data dimension reduction and information density improvement.

[0086] Obtain the numerical capability feature components generated in S2.2 and the structured capability feature components generated in S2.3, and construct a multi-source heterogeneous feature fusion matrix. Perform Z-Score standardization on the fusion matrix to eliminate calculation bias caused by differences in dimensions, generating a standard feature dataset with zero mean and unit variance. Calculate the covariance matrix of the standard feature dataset to quantify the linear correlation between each capability index.

[0087] Eigenvalue decomposition is performed on the covariance matrix to solve for the eigenvalues ​​and their corresponding eigenvectors. The eigenvalues ​​are then sorted in descending order, and the top k principal components whose cumulative contribution rate exceeds a preset threshold are selected to construct the projection transformation matrix.

[0088] The standard feature dataset is mapped to a low-dimensional principal component space, and matrix multiplication is performed to generate a low-dimensional, dense, compressed vector of supplier capability features. This vector retains key variation information from the original multidimensional data and eliminates redundant correlations.

[0089] By using principal component analysis dimensionality reduction algorithm, the high-dimensional heterogeneous features of the previous step are transformed into low-dimensional dense capability feature compressed vectors, thereby achieving data dimensionality reduction and information density improvement, and providing efficient input for subsequent graph neural network node matching.

[0090] For example, numerical features include a scheduling fluctuation coefficient of 0.15 and a mean performance deviation of 0.02, while structured features include a skill vector [0,1,0...] and an integration status [1,0]. A 4-dimensional feature vector is constructed, and the covariance matrix is ​​calculated after standardization. Eigenvalue decomposition yields eigenvalues ​​[2.1,1.8,0.1,0.05]. The first two principal components are selected, with a cumulative contribution rate of 97.5%. After projection, a 2-dimensional compressed vector [1.2,-0.8] is generated, significantly reducing computational complexity while retaining the core capability representation.

[0091] S2.5: Based on the compressed vector of supplier capability features, append the timestamp of the current data acquisition time, perform vector serialization and encapsulation operations to generate a supplier capability summary vector with a time dimension. The supplier capability summary vector is used as the input of candidate nodes in the two-stream network topology to support the subsequent local subgraph construction and adaptation strength score calculation.

[0092] Step S3: Utilizing a fine-grained mapping rule base automatically derived from historical high-matching order samples, the order complexity semantic fingerprint is used as the query node, and the supplier capability summary vector is used as the candidate node to dynamically construct a two-stream network topology containing activation edges and deweighting edges. Specifically, this includes:

[0093] S3.1: Based on a historical high-matching order sample set, extract the co-occurrence patterns between process constraint features, delivery rigidity features, customization granularity features, and collaborative interface type features, and the supplier's real-time production line scheduling margin, historical similar order fulfillment deviation rate, engineer skill map coverage, and cross-system integration readiness status. Use association rule mining algorithms to generate a fine-grained mapping rule base. This fine-grained mapping rule base stores the conditional logic and confidence parameters used to determine the connection relationship between nodes.

[0094] Obtain a historical high-match order sample set and extract four types of order-side features: process constraints, delivery rigidity, customization granularity, and collaborative interface type. Simultaneously extract four types of supply-side features: supplier real-time production line scheduling margin, historical similar order fulfillment deviation rate, engineer skill map coverage, and cross-system integration readiness status.

[0095] The extracted multidimensional heterogeneous features are preprocessed by discretization and standardization. Continuous numerical features are divided into several intervals, and categorical features are mapped to unique identifiers. A transaction database with a unified format is constructed as the basic input for association rule mining.

[0096] The Apriori algorithm or FP-Growth algorithm is used to mine frequent itemsets in the transaction database. A minimum support threshold is set to filter out feature combinations that co-occur frequently in historical high-matching samples, forming a candidate rule set.

[0097] Calculate the confidence and lift indices for each rule. The confidence reflects the probability that the supplier has the corresponding capability under the condition of meeting the order characteristics, and the lift measures the significance of the co-occurrence pattern relative to random occurrence.

[0098] Candidate rules are pruned based on preset confidence and lift thresholds, eliminating rules with low confidence and weak correlation, and retaining rules with strong correlation.

[0099] The retained strongly correlated rules are generalized and logically restructured to generate fine-grained mapping rule entries containing antecedent conditions, consequent conclusions, and confidence parameters. These fine-grained mapping rule entries are then stored in the rule base index structure to establish a fast retrieval mechanism based on feature keys.

[0100] By using the above processing method, the implicit co-occurrence patterns in historical high-matching samples are transformed into an explicit fine-grained mapping rule base, realizing the construction of automated judgment criteria for the connection relationship between nodes, and significantly improving the accuracy and interpretability of the two-stream network topology construction.

[0101] For example, 5000 highly matched order samples are extracted from the historical database, with a minimum support of 0.05 and a minimum confidence level of 0.8. Mining reveals that when an order's characteristics are "high-precision machining" and "delivery cycle less than 7 days," the co-occurrence confidence of suppliers with "five-axis machine tool idle rate greater than 30%" and "senior technician online status" is 0.92, and the lift is 3.5, satisfying the pruning conditions. These are then encapsulated as mapping entries and stored in the rule base. In subsequent new order matching, if a new order matches the preceding item, the system directly activates the supplier node with the following item's characteristics, avoiding wasted computational resources on low-confidence suppliers and significantly improving matching efficiency.

[0102] S3.2: Obtain the semantic fingerprint of the order complexity generated by the newly accessed purchase order and the supplier capability summary vector generated by the registered supplier. Instantiate the semantic fingerprint of the order complexity into a query node in the two-stream network topology, and instantiate the supplier capability summary vector into a candidate node in the two-stream network topology to form a set of nodes to be connected.

[0103] The system receives the order complexity semantic fingerprint vector generated in step S1 and the supplier capability summary vector set generated in step S2 as input data sources. It performs node instantiation processing on the order complexity semantic fingerprint, mapping it to a query node (Query_Node) in a two-stream network topology, and assigning it a unique node identifier (ID_Q) and feature dimension (D_q).

[0104] Key dimension labels are extracted from the semantic fingerprint and transformed into attribute metadata for query nodes, which are then used for conditional matching and verification of subsequent edge weights. The supplier capability summary vector set is traversed in parallel, mapping each supplier's summary vector to a candidate node Cand_Node_i, and assigning it a unique node identifier ID_Si and feature dimension D_s.

[0105] Numerical features such as real-time production line scheduling margin and average performance deviation of candidate nodes are encapsulated as initial state attributes of the nodes. A potential connection pool is established between the query node and all candidate nodes, and the node set topology is initialized. At this time, no explicit edge connections are established between nodes.

[0106] By using tensor concatenation, the feature matrices of query nodes and candidate nodes are aligned to the same latent space dimension, eliminating dimensional differences between heterogeneous data sources. The instantiated query node and candidate node sets are encapsulated into a standardized graph data structure object, containing the node feature matrix V and the empty adjacency matrix E_init.

[0107] Through the node instantiation and feature alignment processes described above, the discrete vector data from the previous step is transformed into a set of nodes with graph topology attributes, thus achieving standardized data structure preparation before topology inference in a two-stream network.

[0108] For example, the semantic fingerprint of order complexity is a 128-dimensional sparse vector, with the key dimension labeled "high-precision machining". The system instantiates this fingerprint as a Query_Node with ID_Q of ORD_20231025_001. Simultaneously, it obtains supplier capability summary vectors from 500 registered suppliers, all of which are 64-dimensional dense vectors. The system instantiates these 500 vectors as Cand_Node_1 to Cand_Node_500, respectively. Cand_Node_1 corresponds to supplier S_01, with a scheduling margin of 0.85 and a skill coverage vector of [0,1,0...]. The system constructs a node set containing one query node and 500 candidate nodes, with the feature matrix V having a shape of [501,128] (after padding and alignment). This process takes less than 5 milliseconds, ensuring that the subsequent graph attention mechanism can directly read the standardized node embedding representation, significantly improving graph construction efficiency.

[0109] S3.3: Traverse each pair of query nodes and candidate nodes in the node set, call the fine-grained mapping rule base for matching verification. If the key dimensions marked by the query node and the capability status indicators of the candidate node meet the preset high confidence condition logic in the fine-grained mapping rule base, then the node pair is determined to have a strong matching relationship.

[0110] like Figure 2 As shown, Figure 2This is the logic flowchart for S3.3 matching verification and strong matching relationship determination. Specifically:

[0111] Obtain the set of key dimension labels from the semantic fingerprint of order complexity and the multi-dimensional index values ​​from the supplier capability summary vector as the initial input data for matching verification.

[0112] Parse the conditional logic expressions stored in the fine-grained mapping rule base to extract the judgment thresholds and confidence weight parameters involving dimensions such as process constraints and delivery rigidity.

[0113] Iterate through each pair of query nodes and candidate nodes in the node set, and index and align the key dimension identifiers of the order fingerprint with the corresponding fields in the supplier capability vector.

[0114] The rules engine is invoked to perform logical comparison, calculating the difference between order demand feature values ​​and supplier capability indicator values, and generating a preliminary matching score. The matching confidence score is calculated using the following formula:

[0115]

[0116] in, To match the confidence level, Assess actual ability. Preset thresholds for rules. This is the sensitivity coefficient.

[0117] If the calculated matching confidence score is higher than the preset high confidence score threshold, it is determined that there is a strong matching relationship between the query node and the candidate node.

[0118] Mark the node pair as a valid connection candidate that meets the high confidence condition logic, and record the specific rule ID and confidence value that were triggered.

[0119] It should be noted that the fine-grained mapping rules are condition-conclusion pairs automatically summarized from historical high-matching order samples. Each rule includes order-side feature antecedents (such as "process accuracy level ≥ IT6"), supplier capability consequents (such as "five-axis linkage machine tool idle rate ≥ 30%)" and corresponding confidence parameters, which are used to determine whether a positive activation edge or a negative suppression and weight reduction edge is established between the query node and the candidate node.

[0120] Through the chain-like derivation process described above, the node set from the previous step is transformed into a list of node pairs with strong matching relationships, achieving the expected technical effect of accurate selection of active edges in the two-stream network topology.

[0121] For example, for order fingerprints labeled with the key dimension of "sensitive to high-precision machining cycle time," the mapping rules for "real-time production line scheduling margin" are retrieved from the rule base. The threshold T is set to 0.2, and the sensitivity coefficient k is 5. A supplier's standardized scheduling margin score S is 0.92. Substituting this into the formula, the calculated confidence level C is approximately 0.993. Since 0.993 is greater than the high confidence threshold of 0.9, the order is determined to have a strong match with this supplier. If another supplier's score S is 0.75, the calculated C is approximately 0.32, which is lower than the threshold, and therefore not considered a strong match. This process effectively identifies supplier nodes with truly high fulfillment potential, significantly improving the computational efficiency and matching accuracy of the subsequent graph attention mechanism.

[0122] S3.4: For node pairs that are determined to have a strong matching relationship, establish an activation edge with positive weight between the query node and the candidate node, and calculate the initial edge weight value of the activation edge based on the confidence parameter output by the fine-grained mapping rule base to complete the construction of the positive connection path in the two-stream network topology.

[0123] Obtain the query nodes and candidate node pairs that have a strong matching relationship as determined by S3.3, as well as the corresponding confidence parameters from the fine-grained mapping rule base. Parse the confidence values ​​output from the rule base and use them as the basic metric for edge weights, reflecting the semantic fit between order requirements and supplier capabilities in specific key dimensions. Introduce a time decay factor to dynamically correct the original confidence values ​​to reflect the time-sensitive impact of supplier capability status. Calculate the initial edge weight value for activated edges using the following formula:

[0124]

[0125] in, To determine the initial weights of the active edges between query node i and candidate node j, The matching confidence score output by the fine-grained mapping rule base. The time decay coefficient, The time difference between the time when the supplier capability summary vector is generated and the time when the current order is processed.

[0126] The calculated initial weights are normalized to ensure that the sum of the weights of all activation edges originating from the same query node is within a preset range, thus avoiding interference from dimensional differences in the feature aggregation of the subsequent graph attention mechanism. In the adjacency matrix of the two-stream network topology, the calculated weight values ​​are filled into the connection positions of the corresponding node pairs to establish activation edges with positive propagation properties.

[0127] By using the above processing method, the discrete logic of rule matching is transformed into continuously differentiable graph structure weights, realizing the quantitative correlation between order complexity and supplier capabilities, and providing a high signal-to-noise ratio topological foundation for subsequent feature propagation of local subgraphs.

[0128] For example, an order fingerprint is labeled with a bottleneck of "high-precision machining," and supplier A's capability snapshot shows "precision machine tool coverage of 95%." The rule base matches the confidence level. It is 0.92. Data acquisition time difference. Set the attenuation coefficient for 2 hours. The value is 0.01. Substituting this into the formula, we get... This weight value is significantly higher than the static weight that does not consider time decay, accurately reflecting the current high adaptability of the supplier and effectively improving the real-time performance and accuracy of graph network inference.

[0129] S3.5: For node pairs where the key dimensions of the query node annotation conflict with the capability status indicators of the candidate node or do not meet the low confidence condition logic in the fine-grained mapping rule base, establish a weighted edge with negative suppression weight between the query node and the candidate node to generate a complete two-stream network topology containing activation edges and weighted edges.

[0130] For node pairs where the key dimensions of the query node annotation conflict with the candidate node capability status indicators or do not meet the low confidence condition logic in the fine-grained mapping rule base, negative suppression weight calculation is performed.

[0131] The conflict determination logic in the fine-grained mapping rule base is analyzed, and the corresponding feature components in the order's key dimension labels and the supplier's capability summary vector are extracted. The semantic distance or numerical deviation between the two is calculated, and if the deviation exceeds a preset safety tolerance threshold, it is determined to be a capability mismatch.

[0132] The initial negative weight values ​​of the reduced-weight edges are constructed based on the degree of mismatch. An exponential decay function is used to map the deviation values ​​to the weight range, ensuring that the weights approach zero in severe mismatch scenarios. In the adjacency matrix of the two-stream network topology, a connection edge with negative attributes is established between the query node and the candidate node, and the calculated negative weight values ​​are written into it.

[0133] This negative weight will play a suppressive role in the subsequent graph attention propagation stage, reducing the contribution of the supplier node to the final adaptation strength score. Traverse all candidate nodes without activated edges, repeating the conflict detection and weighted edge construction process described above, until the topological connections of all candidate nodes are completed.

[0134] By establishing a complete two-stream network topology that includes both active and deweighted edges, the node set from the previous step is transformed into dynamic graph structure data with positive and negative feedback mechanisms. This enables explicit modeling and quantitative suppression of potential performance risks, thereby improving the robustness of the matching results.

[0135] For example, the key dimension of the semantic fingerprint annotation for order complexity is "high-precision machining," with a tolerance requirement of 0.01mm. A candidate supplier's supplier capability summary vector shows a historical fulfillment deviation rate of 5% for similar orders, and a current production line scheduling margin of less than 10%. The fine-grained mapping rule base stipulates that when the tolerance requirement is less than 0.02mm and the fulfillment deviation rate is greater than 3%, it is judged as a high-risk mismatch. The calculated semantic distance deviation value is 0.8, exceeding the safety tolerance threshold of 0.5. An exponential decay function is then used. Calculate the negative weight, where k is 2 and d is the deviation value of 0.8, resulting in a weight of approximately -0.2. Write -0.2 into the corresponding position in the adjacency matrix. This negative weight significantly reduces the supplier's attention coefficient in the graph attention mechanism, causing its final ranking to shift significantly backward, effectively preventing high-difficulty orders from being assigned to suppliers with insufficient capabilities.

[0136] Step S4: Based on the two-stream network topology, only the local subgraph consisting of the supplier nodes and their neighborhoods matched by the fine-grained mapping rules is loaded. Local feature alignment calculation is performed through a two-layer graph attention propagation mechanism, outputting the adaptation strength score of each supplier on key dimensions. Specifically, this includes:

[0137] S4.1: Parse the active and deweighted edges that are hit in the fine-grained mapping rule base to extract the associated supplier node identifier set and its neighborhood connection relationship, and generate a list of local subgraph node indexes to be loaded.

[0138] Receive the two-stream network topology data containing active and deweighted edges generated in step S3, and parse the source node ID, target node ID, and edge type attribute of each edge. Traverse the set of edges in the topology and filter out all connection records whose source node is the semantic fingerprint of the current order complexity.

[0139] Extract the target node identifier from the connection record and construct an initial set of candidate supplier nodes, which contains all supplier nodes that have a rule mapping relationship with the current order.

[0140] The initial set of candidate supplier nodes is deduplicated to eliminate node redundancy caused by multiple rule hits, and a unique list of supplier node identifiers is generated.

[0141] Based on a list of unique supplier node identifiers, the global graph database is queried to obtain the first-order neighbor node set for each supplier node. The first-order neighbor includes other suppliers or associated resource nodes that have historical collaborative relationships with that supplier. The order query node, candidate supplier node, and their first-order neighbor nodes are merged to construct the complete node set of the local subgraph.

[0142] Assign consecutive internal index numbers to the entire set of local subgraph nodes, and establish a mapping dictionary from global node IDs to local subgraph indices. Use a hash lookup method to convert the mapping dictionary into a compact list of local subgraph node indices, removing isolated nodes that do not match the rules.

[0143] By using the above parsing and index mapping processing method, the sparse topological connection relationship of the previous step is transformed into a compact local subgraph node index list data, thereby realizing on-demand loading of computing resources and improving graph reasoning efficiency.

[0144] For example, the current order hits the rule base with 3 active edges and 1 de-weighted edge, involving 5 candidate supplier nodes. The parsed supplier ID set is {S101, S102, S103, S104, S105}. The neighboring nodes of S101 are {S201, S202}, and the neighboring node of S102 is {S203}. The remaining suppliers have no valid neighbors. After merging, the complete set of nodes in the local subgraph is {Order_Q, S101, S102, S103, S104, S105, S201, S202, S203}, a total of 9 nodes. Local indices 0 to 8 are assigned, generating an index list [0, 1, 2, 3, 4, 5, 6, 7, 8]. This process reduces the original thousands of global nodes to 9 nodes, lowering memory usage by over 99% and significantly improving the response speed of subsequent graph attention calculations.

[0145] S4.2: Based on the local subgraph node index list, extract the corresponding order complexity semantic fingerprint query nodes and supplier capability summary vector candidate nodes from the global dual-stream network topology, and construct a local subgraph data instance containing the initial node feature matrix and adjacency matrix.

[0146] Receive a list of local subgraph node indices, which contains the order complexity semantic fingerprint query node IDs filtered by fine-grained mapping rules and the set of candidate node IDs for associated supplier capability summary vectors.

[0147] Based on the node index list, the corresponding node feature data is retrieved in parallel from the distributed storage medium of the global dual-stream network topology, and the sparse complexity vector of the order query node and the dense capability summary vector of the supplier candidate node are extracted.

[0148] A dimension alignment check is performed on the extracted original feature vectors. If the order feature dimension is inconsistent with the supplier feature dimension, a pre-trained linear projection layer is called to map the low-dimensional features to a unified high-dimensional latent space to ensure the isomorphism of the feature space.

[0149] An initial node feature matrix H is constructed based on the aligned node features, where the row vectors correspond to the feature representations of each node and the column vectors correspond to the feature dimensions, forming a feature tensor of shape N×D, where N is the total number of nodes in the local subgraph and D is the feature dimension.

[0150] Analyze the connectivity between active and reduced-weighted edges in a local subgraph, and construct a sparse adjacency matrix A. If there is an active edge between node i and node j, then A... ij Assign a positive confidence weight; if a de-weighting edge exists, assign a negative suppression weight to A_ij; if there is no connection, then A... ij It is zero.

[0151] Symmetric normalization is performed on the sparse adjacency matrix A, a self-loop mechanism is introduced to preserve the node's own feature information, and the normalized Laplacian matrix is ​​calculated to ensure the numerical stability and gradient propagation efficiency during the graph convolution operation.

[0152] The initial node feature matrix H and the normalized adjacency matrix A are encapsulated into a standardized local subgraph data instance. This instance contains node attribute tensors, edge index tensors, and edge weight tensor tuples, which serve as the direct input objects for the graph attention propagation mechanism.

[0153] Through the above data processing flow, discrete node indexes are transformed into structured graph data instances, realizing the mapping from logical topology to computational graph, and providing a standardized data foundation for feature aggregation in subsequent multi-layer graph attention mechanisms.

[0154] For example, the local subgraph contains one order query node and 50 supplier candidate nodes. The order feature vector has a dimension of 128, and the supplier feature vector has a dimension of 64. A projection layer is called to map the supplier features to a 128-dimensional space. A 128×51 node feature matrix H is constructed. In the adjacency matrix A, the weights of the 30 active edges are set to 0.8-0.95, and the weights of the 5 de-weighted edges are set to -0.5. After symmetric normalization, a local subgraph data instance containing 51 nodes and 35 edges is generated, reducing memory usage by 90% and controlling inference latency to within 5ms.

[0155] S4.3: Utilize the first-layer graph attention propagation mechanism to perform weighted aggregation operations on the features of neighboring nodes in the local subgraph data instance to calculate the cross-dimensional feature alignment coefficient within the first-order neighborhood, generating an intermediate-layer node embedding representation after preliminary semantic fusion.

[0156] Obtain the initial node feature matrix and adjacency matrix in the local subgraph data instance, take the order complexity semantic fingerprint query node as the center node, and define the candidate nodes of the supplier capability summary vector directly connected to it as the set of first-order neighbor nodes.

[0157] For the central node and each first-order neighbor node, their corresponding feature vectors are extracted. The heterogeneous order feature space and supplier capability feature space are mapped to the same-dimensional latent space through a linear transformation matrix, eliminating the distribution offset caused by the difference in dimensions.

[0158] The dot product similarity between the feature vector of the central node and the feature vectors of each neighboring node is calculated as the original attention score, which represents the potential response strength of a particular supplier to the current order demand in a certain dimension.

[0159] The LeakyReLU activation function is introduced to perform nonlinear correction on the original attention score, suppressing the interference of negative matching signals and enhancing the weight of positive matching signals on key dimensions.

[0160] The corrected attention score is processed using the Softmax normalization function to ensure that the sum of the attention weights of all neighboring nodes is 1, thereby generating a standardized attention coefficient distribution.

[0161] Based on standardized attention coefficients, the feature vectors of each first-order neighbor node are weighted and aggregated, while retaining the feature information of the central node itself, and information loss is prevented through residual connection mechanism.

[0162] The aggregated feature vectors are input into a fully connected layer for nonlinear transformation to generate an intermediate layer node embedding representation that includes neighborhood context information. This representation integrates the initial alignment features of order requirements and supplier capabilities.

[0163] It should be noted that this step, along with subsequent steps S4.4 and S4.5, collectively implements local feature alignment operations through feature propagation. Feature propagation refers to the process in the local subgraph where the semantic features of the order node's complexity are passed to the supplier node along activation and deweighting edges, while the supplier node's capability features are simultaneously passed to the order node. Initial feature aggregation within the first-order neighborhood is achieved through a first-layer graph attention mechanism, followed by high-order nonlinear feature interaction through a second-layer multi-head graph attention mechanism. Finally, the adaptation strength scores of each supplier across key dimensions are output through dimension mapping.

[0164] Through the graph attention propagation and weighted aggregation processing described above, the local subgraph topology structure constructed in the previous step is transformed into intermediate layer node embedding representation data containing cross-dimensional semantic relationships. This enables the initial feature alignment and fusion of order complexity and supplier capabilities in the first-order neighborhood, providing a structured foundation for subsequent high-order nonlinear interactions.

[0165] For example, the feature dimension of the central order node in the local subgraph is set to 64, and the number of neighboring supplier nodes is 5. The linear transformation matrix is ​​64x64. The original attention scores of the 5 neighboring nodes are calculated to be 0.2, 0.5, -0.1, 0.3, and 0.4. After LeakyReLU correction, they become 0.2, 0.5, -0.01, 0.3, and 0.4. After Softmax normalization, the weight distribution is 0.18, 0.26, 0.09, 0.21, and 0.26. The intermediate layer embedding vector generated after weighted aggregation maintains a 64-dimensional dimension, effectively integrating the capability features of highly matched suppliers and significantly improving the targeting of feature representation.

[0166] S4.4: Based on the intermediate layer node embedding representation, the second layer graph attention propagation mechanism is executed. The feature contribution of key dimensions is further refined through the multi-head attention weight allocation strategy, and the high-order nonlinear feature interaction and alignment operation are completed to output the deep graph node feature vector.

[0167] The intermediate layer node embeddings generated by the first layer graph attention propagation are received and used as the input feature matrix for the second layer graph attention mechanism.

[0168] Initialize the multi-head attention module, set the number of attention heads to 8, and linearly project the input feature vector to 8 independent subspaces to generate a sequence of query vector, key vector and value vector.

[0169] Within each attention head, the query vector and key vector are scaled and dot-producted, and the attention weights are obtained after Softmax normalization. Then, the value vectors are weighted and aggregated to obtain the original attention scores.

[0170] By introducing a mask for key order dimensions, the original attention score is weighted and corrected to suppress feature responses of non-key dimensions and enhance the feature contribution of key dimensions.

[0171] The corrected attention scores are normalized using Softmax to generate the attention weight distribution of each neighbor node to the center node.

[0172] By using attention weights to perform weighted summation and aggregation of the value vectors, we obtain the context-aware feature representation under each attention head.

[0173] The outputs of eight attention heads are concatenated and feature fusion is performed through a linear transformation layer to restore the original feature dimensions and generate deep graph node feature vectors.

[0174] By employing a multi-head attention weight allocation strategy, the preliminary fusion features from the previous step are transformed into deep feature vectors rich in high-order nonlinear interaction information, thereby achieving precise alignment and feature enhancement of key dimensions.

[0175] For example, we set the number of attention heads to 8 and the feature dimension to 64. Linear projection is performed on the intermediate layer embedding vectors to generate 8 sets of query vectors, key vectors, and value vectors. Within each attention head, the query vector and key vector are scaled and normalized to obtain the attention weight distribution. A key dimension mask is applied, increasing the weight coefficient of the "delivery rigidity" dimension by 1.5 times, while keeping other dimensions at 1.0. After weighted aggregation, the context feature vectors of the 8 heads are obtained, concatenated, and transformed through a linear layer to output a 64-dimensional deep feature vector. This process significantly improves the matching sensitivity of high-complexity orders on key dimensions, greatly enhancing the discriminative power of the fit strength score.

[0176] S4.5: Perform dimension mapping and normalization on the feature vectors of deep graph nodes to extract the matching confidence values ​​of each supplier node on the key dimensions of the order label, and generate a sequence of adaptation strength scores that characterize the degree of matching between the order and the supplier.

[0177] Receive the deep graph node feature vector output by S4.4, which integrates the interaction information of order complexity semantics and supplier capability status.

[0178] Construct a linear projection layer to map high-dimensional deep features to a logical space consistent with the number of key dimensions in the order.

[0179] The mapped logits are normalized using the Softmax activation function, and the probability distribution of each supplier on a specific key dimension is calculated.

[0180] The matching confidence scores of each key dimension are weighted and aggregated to generate a comprehensive fit strength score.

[0181] Through the above dimensional mapping and normalization processing, the features of deep graph nodes are transformed into a quantified sequence of adaptation strength scores, thereby achieving accurate quantitative matching between order requirements and supplier capabilities.

[0182] For example, a subgraph contains 5 candidate suppliers, and the order is labeled with 3 key dimensions: process precision, delivery cycle, and system integration. After linear projection, the feature vectors of the deep graph nodes, for the "process precision" dimension, the projected values ​​of the 5 suppliers are 2.0, 1.5, 0.5, -0.5, and -1.0, respectively. After Softmax normalization, the corresponding matching confidence scores are 0.513, 0.307, 0.113, 0.042, and 0.025, respectively. Similarly, the confidence scores of other suppliers and dimensions are calculated, finally generating a 5×3 fit strength score matrix. This matrix clearly quantifies the degree of advantage of each supplier in different key dimensions, providing accurate data support for subsequent flexible weight adjustment and significantly improving the accuracy of matching high-complexity orders.

[0183] Step S5: Based on the variance of the adaptation strength score distribution across each key dimension, perform a threshold adaptive pruning operation to identify the leading supplier's advantageous dimensions and generate a flexible weight distribution sequence that releases weights to non-advantageous dimensions. Specifically, this includes:

[0184] S5.1: Group and aggregate the fit strength scores of each supplier on the key dimensions to calculate the statistical dispersion of the fit strength score set under each key dimension, thereby obtaining the key dimension distribution variance data that characterizes the degree of difference in supplier capabilities.

[0185] Receive the adaptation strength score sequence output by S4.5, which contains the matching confidence values ​​of each supplier on the key dimensions labeled in the order. Perform a dimension grouping operation on the adaptation strength score sequence, and aggregate the adaptation strength scores of all suppliers belonging to the same key dimension into an independent set of values ​​based on the key dimension labels labeled in the semantic fingerprint of order complexity.

[0186] For each key dimension, the arithmetic mean of the score set is calculated to represent the baseline of the overall fit level of the supplier group under that dimension.

[0187] Based on the arithmetic mean, the variance of each score set is calculated to characterize the dispersion of supplier capabilities along this dimension. The larger the variance, the more significant the differences in supplier capabilities along this dimension.

[0188] The distribution variance of each key dimension is standardized and mapped to eliminate the influence of the difference in the units of the scores of different dimensions on the variance value, and normalized distribution variance data of the key dimensions is generated.

[0189] By using the above statistical dispersion calculation method, the adaptation strength score of the previous step is transformed into the distribution variance data of key dimensions that characterize the degree of difference in supplier capabilities, so as to realize the quantitative assessment of the competitive situation of suppliers under each key dimension and provide statistical basis for the subsequent generation of dynamic thresholds.

[0190] For example, for a highly complex order, two key dimensions are extracted: "process precision" and "delivery cycle". In the "process precision" dimension, 50 suppliers participated in the matching, with a set of fit strength scores {0.85, 0.88, ..., 0.92}, a mean of 0.87, and a variance of 0.002, indicating that supplier capabilities are highly concentrated and the differences are minimal in this dimension. In the "delivery cycle" dimension, the set of fit strength scores for the 50 suppliers is {0.60, 0.95, ..., 0.75}, with a mean of 0.78 and a variance of 0.045, indicating that supplier capabilities are highly dispersed and there is significant capability stratification in this dimension. The system outputs 0.002 and 0.045 as the key dimension distribution variance data for subsequent dynamic threshold mapping calculations, thereby identifying "delivery cycle" as a dimension with significant potential for leading suppliers, while "process precision" is a dimension that needs to be considered in a balanced manner.

[0191] S5.2: Perform dynamic threshold mapping operation based on the key dimension distribution variance data, and use the preset variance-confidence nonlinear transformation function to map high variance features to low fault tolerance intervals, thereby generating dimension-specific clipping threshold parameters for defining suppliers with significant advantages.

[0192] The system receives the set of adaptation strength scores for each key dimension and their corresponding statistical dispersion data, i.e., the key dimension distribution variance data, from the output of S5.1. For each key dimension, its corresponding variance value is extracted as the input variable for dynamic threshold mapping. A variance-confidence nonlinear transformation function is constructed. This function is based on the inverse mapping logic of the Sigmoid curve, aiming to map high-variance features to low-tolerance intervals. The mathematical expression of the transformation function is defined as follows:

[0193]

[0194] in, For the generated dimension-specific clipping threshold parameters, The lower limit of the basic threshold, The upper limit is the threshold, and k is the sensitivity adjustment coefficient. The variance of the fit strength score for the current dimension. This serves as the baseline reference value for variance. A sensitivity adjustment coefficient is set based on the historical statistical distribution of order complexity. To control threshold The slope of the changing response.

[0195] Substitute the variance value σ of the current dimension into the above formula, perform exponential and reciprocal calculations to obtain the normalized confidence mapping factor. Multiply this mapping factor by the threshold interval span and add the lower limit of the basic threshold to calculate the specific clipping threshold for this dimension in the current order scenario.

[0196] Repeat the above mapping operation for all key dimensions to generate a vector sequence containing independent threshold parameters for each dimension.

[0197] By using a variance-confidence nonlinear transformation process, the variance data from the previous step is converted into a dimension-specific pruning threshold parameter, enabling dynamic definition of the expected technical effects of suppliers with significant advantages based on the degree of difference in supplier capabilities.

[0198] For example, for the 'delivery rigidity' dimension, the variance σ of its adaptation strength score is 0.04. T_base is set to 0.6, T_max to 0.9, k to 10, and μ to 0.02. Substituting these values ​​into the formula, the exponential part e^(10×(0.04-0.02)) equals e^0.2, approximately 1.2214. The denominator 1+1.2214 equals 2.2214. The fractional part 1 / 2.2214 equals approximately 0.4502. The threshold increment (0.9-0.6)×0.4502 equals 0.1351. Finally, the pruning threshold T for this dimension is 0.6+0.1351, equal to 0.7351. This threshold is significantly higher than the base threshold, effectively filtering out leading suppliers with a significant advantage in delivery rigidity.

[0199] S5.3: Perform extreme value filtering on the fit strength score set using dimension-specific clipping threshold parameters. By comparing the fit strength score of each supplier with the clipping threshold of the corresponding dimension, the top supplier nodes that exceed the threshold are identified, thereby obtaining the top supplier advantage dimension label set that identifies significant capability advantages.

[0200] like Figure 3 As shown, Figure 3 This is a flowchart illustrating the logic of S5.3 extreme value filtering and advantage dimension label generation. Specifically:

[0201] Receive the dimension-specific clipping threshold parameters generated in S5.2 and the set of adaptation strength scores for each supplier on the key dimensions output in S4.5. Group the set of adaptation strength scores according to the key dimensions and construct a two-dimensional mapping matrix with supplier ID as the index and dimension adaptation score as the element.

[0202] Iterate through each key dimension and extract the fit strength score sequence of all candidate suppliers under that dimension. Call the preset extreme value filtering logic to compare the fit strength score sequence of the current dimension with the corresponding dimension pruning threshold calculated by S5.2 element by element.

[0203] For any supplier node, if its fit strength score in a specific key dimension is strictly greater than the pruning threshold for that dimension, then the supplier is determined to have a significant advantage in that dimension. Record all supplier IDs that meet the above criteria and their corresponding key dimension tags to form a preliminary list of advantageous matching pairs.

[0204] The initial list of advantage matching pairs is deduplicated and validated for completeness, removing isolated nodes or invalid dimension labels caused by data anomalies. The validated advantage matching pairs are then aggregated according to dimension categories to generate a header set of supplier advantage dimension tags, with key dimensions as keys and the set of advantage supplier IDs as values.

[0205] By using extreme value screening and aggregation, continuous fit strength scores are transformed into discrete sets of advantage dimension labels, enabling precise targeting of high-matching suppliers and providing clear target objects for subsequent weight shrinkage.

[0206] For example, key dimensions are set to include delivery rigidity and process precision. S5.2 outputs a trimming threshold of 0.85 for the delivery rigidity dimension and 0.90 for the process precision dimension. S4.5 outputs that Supplier A scores 0.92 in delivery rigidity and 0.88 in process precision; Supplier B scores 0.80 in delivery rigidity and 0.95 in process precision. Comparison reveals that Supplier A's delivery rigidity score of 0.92 is greater than the threshold of 0.85, and is marked as having a delivery rigidity advantage; Supplier B's process precision score of 0.95 is greater than the threshold of 0.90, and is marked as having a process precision advantage. The final generated header supplier advantage dimension label set is {Delivery Rigidity: [Supplier A], Process Precision: [Supplier B]}.

[0207] S5.4: Perform weight shrinkage and redistribution calculations based on the leading supplier's advantageous dimension label set. Compress the initial weight coefficients marked as advantageous dimensions proportionally and distribute the released weight increments to non-advantageous dimensions according to the inverse proportional rule, thereby generating an elastic weight distribution sequence that balances the dominant effect of a single dimension.

[0208] Receive the set of advantageous dimension labels from leading suppliers and the initial weight coefficients for each key dimension, and construct the weight mapping matrix to be adjusted. Iterate through each advantageous dimension identifier in the label set and retrieve its corresponding original weight value in the initial weight distribution.

[0209] The weighted contraction factor of the dominant dimension is calculated. This factor is determined based on a preset elastic compression ratio and is used to quantify the degree of suppression of the dominant effect of a single dimension.

[0210] Perform a multiplicative decay operation on the original weights of each dominant dimension to generate new, shrunken weight values, and record the total amount of weight difference released. Calculate the distribution of the basic weight percentages for each dimension within the set of non-dominant dimensions that were not marked as dominant dimensions.

[0211] Based on the inverse distribution rule, the total released weight difference is distributed according to the inverse proportion of the base weight of the non-dominant dimension. A non-linear redistribution function is constructed to ensure that non-dominant dimensions with high base weights receive less increment, while non-dominant dimensions with low base weights receive more increment, thus balancing the overall evaluation perspective.

[0212] The calculated weight increments for each non-dominant dimension are accumulated and added to their corresponding base weights to form updated non-dominant dimension weight values. Normalization is performed on all dimensions (including shrunken dominant dimensions and shrunken non-dominant dimensions) to ensure that the sum of the final weight sequence is always equal to 1.

[0213] By using weight shrinkage and inverse redistribution, the dominant dimension labels identified in the previous step are transformed into an elastic weight distribution sequence that balances the dominant effect of a single dimension, thereby achieving dynamic equilibrium and precise matching of multidimensional evaluation indicators.

[0214] For example, the initial weight vector is set to [0.4, 0.3, 0.2, 0.1], corresponding to the four dimensions of delivery, process, customization, and interface, respectively. If the delivery dimension is detected as an advantageous dimension, the compression ratio is set to 0.5, and the new weight of the delivery dimension is 0.2, releasing the difference of 0.2. The sum of the basic weights of the remaining three non-advantageous dimensions is 0.6, and their reciprocal proportional factors are 1 / 0.3, 1 / 0.2, and 1 / 0.1, respectively. After normalization calculation, the 0.2 difference is inversely distributed to the process, customization, and interface dimensions, ultimately generating a flexible weight sequence, which significantly reduces the excessive influence of a single advantage on the overall score.

[0215] Step S6: Input the elastic weight distribution sequence and the supplier's basic capability score into a nonlinear fusion function for weighted aggregation to generate a comprehensive ranking score that characterizes the final matching degree between the order and the supplier. Specifically, this includes:

[0216] S6.1: Obtain the elastic weight distribution sequence and supplier basic capability score as initial input data, construct a multi-dimensional feature mapping space based on the preset nonlinear activation function, perform normalization processing on the weight values ​​of each key dimension in the elastic weight distribution sequence, and generate a standardized weight vector to eliminate the impact of dimensional differences on subsequent fusion calculations.

[0217] Obtain the elastic weight distribution sequence output from step S5, and obtain the supplier basic capability score as the initial input for nonlinear fusion calculation. The supplier basic capability score is obtained from the supplier management information database, including but not limited to: static profile indicators such as supplier historical fulfillment rate, industry qualification level, quality system certification status, and production capacity benchmark value. This score is independent of the dynamic supplier capability summary vector generated in step S2, and together they constitute the comprehensive evaluation dimension of the supplier.

[0218] The original weight values ​​of each key dimension in the elastic weight distribution sequence are normalized using the L2 norm to eliminate numerical bias caused by differences in units between dimensions. A multidimensional feature mapping space based on the Softmax function is constructed to map the normalized weight values ​​to the probability distribution interval, ensuring that the sum of the weights of each dimension is constant at 1.

[0219] Exponential operations amplify the discriminative power of high-weight dimensions while suppressing noise interference from low-weight dimensions. The calculated standardized weight vector is encapsulated into an adjustment factor matrix for subsequent dynamic scaling of supplier fundamental capability scores.

[0220] Through the above normalization and mapping process, the elastic weight distribution of the previous step is transformed into a standardized weight vector with probabilistic interpretability, thereby achieving the expected technical effect of eliminating dimensional differences and retaining relative importance.

[0221] For example, for an order containing three key dimensions—process constraints, delivery rigidity, and customization granularity—the original elasticity weights output by S5 are 0.6, 0.3, and 0.1, respectively. After Softmax normalization, the corresponding standardized weights are 0.426, 0.316, and 0.258, respectively. This vector eliminates the influence of the original numerical dimensions and has a total sum of 1, significantly improving the numerical stability and matching accuracy of subsequent nonlinear fusion.

[0222] S6.2: Using standardized weight vectors as adjustment factors, dynamic scaling operations are performed on the corresponding dimension scores in the supplier's basic capability score. By introducing a sigmoid nonlinear transformation mechanism, the weighted response values ​​of each dimension are calculated, generating a dimension adaptation strength matrix containing nonlinearity to reflect the marginal effect of weight changes on capability scores.

[0223] Obtain the standardized weight vector and supplier basic capability score matrix after S6.1 normalization processing, and use them as the input data source for dynamic scaling calculation. For each dimension weight value in the standardized weight vector, construct a non-linear mapping channel based on the Sigmoid function. Utilize the saturation characteristic of this function to smoothly compress extreme weight values, avoiding scoring distortion caused by excessive weight in a single dimension.

[0224] The center point parameter of the Sigmoid function is set to 0.5 and the slope parameter to 4 to ensure high sensitivity in the weighted interval and fast convergence at both ends.

[0225] The standardized weight vector is input into the Sigmoid function to calculate the nonlinear adjustment coefficient for each dimension. This coefficient reflects the diminishing marginal contribution of weight changes to the final score.

[0226] Based on the calculated nonlinear adjustment coefficient, element-wise multiplication is performed on the original scores of the corresponding dimensions in the supplier basic capability score matrix to achieve dynamic scaling of the weights.

[0227] By introducing a Sigmoid nonlinear transformation mechanism, the masking effect of high-weight dimensions on low-weight dimensions is effectively suppressed, enhancing the scoring system's sensitivity to weaker dimensions. A dimension adaptation strength matrix containing nonlinear response characteristics is generated, which accurately characterizes the supplier's actual performance potential in each key dimension under a flexible weight distribution.

[0228] By using Sigmoid nonlinear transformation and dynamic scaling, the standardized weights from the previous step are transformed into a dimension adaptation strength matrix with marginal effect constraints. This achieves a technological leap from linear superposition to nonlinear collaboration in weight allocation, significantly improving the robustness of supplier matching in complex order scenarios.

[0229] For example, suppose an order has a standardized weight of 0.8 in the "delivery rigidity" dimension and a standardized weight of 0.2 in the "customization granularity" dimension. Supplier A scores 90 in "delivery rigidity" and 60 in "customization granularity". After applying a Sigmoid nonlinear mapping to the standardized weights (0.8 for "delivery rigidity" and 0.2 for "customization granularity"), the corresponding adjustment coefficients are 0.982 and 0.269, respectively. Scaling supplier A's score, the "delivery rigidity" fit strength is 88.38, and the "customization granularity" fit strength is 16.14. S6.3: Perform cross-dimensional interactive feature extraction based on the dimensional fit strength matrix, and use the multinomial kernel function method to calculate the coupling gain term between different key dimensions to generate a high-order interactive feature vector to capture the nonlinear collaborative relationship between order complexity and supplier capabilities in multidimensional space.

[0230] The system receives a dimension adaptation strength matrix containing scores for each key dimension after nonlinear scaling. A polynomial kernel function mapping space is constructed to capture the nonlinear coupling effects between dimensions. The kernel function order parameter is set, typically to 2 or 3, to balance computational complexity and feature representation capability. A self-outer product operation is performed on the dimension adaptation strength vectors to generate a second-order interaction feature tensor. Off-diagonal elements are extracted from the tensor, representing the strength of cross-influence between different key dimensions. A regularization coefficient is introduced to suppress the overfitting risk caused by high-dimensional sparse data. The high-order interaction feature vector is calculated using the following formula:

[0231]

[0232] Where H is the high-order interactive feature vector, X is the dimension adaptation strength vector, c is the constant offset term, and d is the polynomial order. Expanding the above formula yields a feature combination containing linear, square, and cross terms. Normalization is performed on the cross terms to eliminate the interference of dimensional differences on the coupling gain calculation. A feature selection algorithm is used to select the subset of interactive features with the highest contribution. The selected subset of interactive features is concatenated to the original dimension adaptation strength matrix to form an enhanced feature representation. Through polynomial kernel function mapping, the linear adaptation strength from the previous step is transformed into a high-order interactive feature vector containing nonlinear collaborative relationships between dimensions, achieving a quantitative capture of the deep coupling effect between order complexity and supplier capabilities. For example, for a high-precision machining order, the dimension adaptation strength vector X is [0.8, 0.6, 0.9], corresponding to the process, delivery, and customization dimensions, respectively. The polynomial order d=2 and the constant c=1 are set. The interactive feature vector H is calculated, where the coupling values ​​of cross terms such as process and delivery are significantly improved. After normalization and feature selection, the top 5 interaction features by contribution were retained. In the final enhanced feature representation, the high-order interaction features effectively reflect the supplier capability shortcomings when 'high process requirements' and 'short delivery cycles' coexist. This enables subsequent comprehensive ranking to more accurately identify suppliers who have strong individual capabilities but poor overall coordination, significantly improving the robustness of matching complex orders.

[0233] S6.4: Input the dimension adaptation strength matrix and the high-order interaction feature vector into the weighted aggregation engine, perform element-wise multiplication and accumulation operations and superimpose the bias correction term to generate a preliminary comprehensive matching score to quantify the overall fulfillment potential of the supplier in the current order scenario.

[0234] The dimension adaptation strength matrix generated by S6.2 and the high-order interaction feature vector generated by S6.3 are obtained as the input data source for the weighted aggregation engine.

[0235] An element-wise multiplication and accumulation operation logic is constructed to perform a dot product operation on the scores of each key dimension in the dimension adaptation strength matrix and the corresponding standardized weight coefficients in the elastic weight distribution sequence.

[0236] A bias correction term is introduced to compensate for the systematic deviation caused by the nonlinear transformation. This bias term is dynamically determined based on the mean of the residual distribution of historical order fulfillment data.

[0237] The preliminary overall matching score is calculated using the following formula:

[0238]

[0239] in, To determine the initial overall matching score, For the first Standardized weighting coefficients for each key dimension. For the first The fit strength scores for each key dimension For bias correction term, For interaction gain coefficients, is the aggregated value of the higher-order interactive feature vector, where n is the total number of key dimensions.

[0240] The calculated preliminary comprehensive matching score is subjected to numerical stability verification to ensure that the score falls within the preset valid range and to prevent sorting distortion caused by extreme values.

[0241] Through the above weighted aggregation and bias correction processing, the adaptation strength and interaction characteristics of multidimensional heterogeneity are transformed into a preliminary comprehensive matching score in scalar form, thereby realizing a quantitative representation of the supplier's overall performance potential in the current order scenario.

[0242] For example, for a high-precision machining order, the number of key dimensions n is set to 3, namely process precision, delivery cycle, and collaboration interface. The standardized weight vector w is [0.4, 0.3, 0.3], and the dimension adaptation strength vector A is [0.85, 0.60, 0.75]. The aggregation value H of the higher-order interaction feature vector is calculated to be 0.12, the interaction gain coefficient γ is set to 0.5, and the bias correction term β is set to -0.02. Substituting into the formula, the result is: S = (0.4 × 0.85 + 0.3 × 0.60 + 0.3 × 0.75) + (-0.02) + (0.5 × 0.12) = (0.34 + 0.18 + 0.225) - 0.02 + 0.06 = 0.785. This score accurately reflects the supplier's comprehensive fulfillment potential under the advantage of process precision, and the bias dominated by a single dimension is corrected by the interaction term, which significantly improves the discriminativeness of the matching results.

[0243] S6.5: Perform monotonicity verification and extreme value smoothing on the preliminary comprehensive matching score, use the soft sorting algorithm to generate the final comprehensive ranking score sequence, and output the ranking result representing the final matching degree between the order and the supplier for the downstream recommendation decision module to call.

[0244] Step S7: Execute supplier recommendation decisions based on the comprehensive ranking score. If the fulfillment risk index of the current order exceeds a preset safety threshold, trigger a manual review and record the elastic weight allocation log for subsequent rule base iterations. Specifically, this includes:

[0245] S7.1: Perform multi-dimensional weighted aggregation calculation on the comprehensive ranking score sequence and the corresponding key dimension adaptation strength scores to generate a performance risk index that represents the probability of potential delivery deviation of the current order under a specific supplier combination.

[0246] The system receives the comprehensive ranking score sequence output by S6 and the adaptation strength scores of each supplier in key dimensions generated by S4, as the initial input data for performance risk calculation.

[0247] The adaptation strength score is decoupled by dimension, and the score subsets corresponding to the key dimensions of the order are extracted to construct risk-sensitive feature vectors for specific constraints.

[0248] Based on the standard deviation of the fit strength scores of each dimension, the risk weight coefficient of each key dimension is calculated using the inverse variance weighting method. Dimensions with larger standard deviations are assigned higher risk weights.

[0249] Using the calculated dimensional risk weight coefficients, a punitive correction is applied to high-scoring suppliers in the comprehensive ranking score sequence to generate risk-adjusted potential delivery deviation probability values.

[0250] By performing regression mapping between the potential delivery deviation probability value and the actual performance deviation data of similar historical orders, the performance risk index under the current order scenario is obtained.

[0251] By using multi-dimensional weighted aggregation and risk mapping processing, the comprehensive matching results of the previous step are transformed into a quantitative performance risk index, which enables accurate quantification and early warning of potential delivery risks for highly complex orders.

[0252] For example, an order is labeled with two key dimensions: "process precision" and "delivery cycle." The standard deviations of the supplier fit strength scores for these two dimensions are 0.045 and 0.002, respectively. Using an inverse variance weighting method, dimensions with larger standard deviations are assigned higher risk weights. Based on this, a punitive correction is applied to the overall ranking score to generate a risk-adjusted probability value for potential delivery deviations. After regression mapping, the fulfillment risk index for the current order is obtained as 0.75.

[0253] S7.2: Compare the generated performance risk index with the preset dynamic safety threshold in the risk control module to determine whether the current order is in a high-risk state and output the risk judgment flag.

[0254] Receive the performance risk index generated by S7.1 and the complexity level identifier associated with the current order as input benchmark data for risk assessment.

[0255] The system reads a preset set of dynamic safety threshold parameters from the configuration database of the risk control module. This set includes the baseline threshold, the complexity sensitivity coefficient, and the historical volatility correction factor.

[0256] Based on the complexity level of the current order, the baseline threshold is dynamically adjusted, with more stringent security thresholds applied to orders with higher complexity.

[0257] The generated performance risk index is compared with the calculated dynamic safety threshold, and logical judgment is performed to determine the risk status.

[0258] If the performance risk index is greater than the dynamic safety threshold, the current order is determined to be in a high-risk state, and a risk judgment flag with a value of 1 is generated.

[0259] If the performance risk index is less than or equal to the dynamic safety threshold, the current order is determined to be in a normal risk state, and a risk judgment flag with a value of 0 is generated.

[0260] The risk assessment flag, along with the corresponding dynamic security threshold value and the original performance risk index, are packaged and encapsulated to form a standardized risk decision data package.

[0261] By using the above dynamic threshold comparison processing method, the risk quantification results of the previous step are transformed into binary risk status identifiers, thereby achieving the expected technical effect of differentiated risk management for orders with different complexities.

[0262] For example, a basic security threshold is set to 0.65, and a complexity adjustment coefficient is set to 0.15. For a highly customized order with a normalized complexity score of 0.8, the dynamically adjusted security threshold is 0.728. If the performance risk index is 0.75, since 0.75 is greater than 0.728, it is judged as a high-risk state, and the risk judgment flag is output as 1. For a standard order with a complexity score of 0.2, the dynamically adjusted security threshold is 0.6695. If the performance risk index is 0.60, since 0.60 is less than 0.6695, it is judged as a normal state, and the risk judgment flag is output as 0. This mechanism significantly improves the sensitivity of risk identification for highly complex orders and avoids the problems of missed or false alarms caused by a uniform threshold.

[0263] S7.3: Based on the high-risk status indicated in the output risk assessment flag, suspend the target supplier recommendation list and activate the manual review flag field to block the automated signing process and transfer to the expert intervention channel.

[0264] Receive the risk assessment flag bit and associated supplier recommendation list data output by S7.2, and parse the high-risk status indicator field in the flag bit.

[0265] Identify the set of target supplier nodes marked as high-risk, and extract their corresponding comprehensive ranking score and key dimension fit strength score. Construct a process suspension instruction object, write the unique identifier of the target supplier node into the suspension queue, and lock the automated signing permission for this batch of orders.

[0266] Activate the manual review flag field to generate a review context data package containing the order complexity semantic fingerprint, elastic weight distribution sequence, and performance risk index. Call the expert intervention channel interface to push the review context data package to the procurement expert terminal interface, blocking the execution thread of the subsequent automated contract signing process.

[0267] By using the above processing method, the risk assessment results of the previous step are transformed into specific process control instructions and manual intervention tasks, so as to achieve accurate interception of high-risk orders and expert bottom-line decision-making, and ensure the security of supply chain performance.

[0268] For example, when the performance risk index of a highly customized precision machining order is calculated to be 0.85, exceeding the preset safety threshold of 0.75, the system parses the risk judgment flag as High_Risk. It immediately locks the top three supplier IDs (SUP_01, SUP_02, SUP_03), removes them from the automatic contract signing candidate pool, and adds them to the pending queue. A review data package is generated, containing the order fingerprint hash value, records of abnormal weight fluctuations in the key dimension 'process precision,' and risk contribution factors. This data package is pushed to the senior procurement engineer's workbench, with a 'Manual Review Recommended' warning highlighted on the interface, while the contract generation button in the ERP system is frozen. This process is completed within 50 milliseconds, effectively avoiding potential default risks caused by model confidence bias and ensuring that only low-risk suppliers confirmed by experts enter the contract signing stage.

[0269] S7.4: The elastic weight distribution sequence, fine-grained mapping rule hit records, and performance risk index generated during this decision-making process are structured and encapsulated to form elastic weight allocation log entries containing complete context information.

[0270] S7.5: Write the generated elastic weight allocation log entries into the historical decision sample database so that the subsequent rule base iteration module can extract them and use them to correct the confidence parameters in the fine-grained mapping rule base, thereby achieving continuous self-optimization of the matching model.

[0271] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0272] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0273] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-level weighting method for contract signing and project initiation that integrates supplier profiling and order quality, characterized in that: include: Obtain the structured features and unstructured semantic information of newly accessed purchase orders, and generate a unique semantic fingerprint of order complexity. The semantic fingerprint of order complexity is used to characterize the position of the order in the multi-dimensional feature space and to label key dimensions. Based on the time window associated with the semantic fingerprint of the order complexity, real-time capability data of suppliers is collected and time-series compressed to generate a timestamped supplier capability summary vector. Based on the mapping rules inferred from historical matching samples, the order complexity semantic fingerprint is used as the query node and the supplier capability summary vector is used as the candidate node to construct a two-stream network topology containing positive connections and negative suppression connections. Based on the supplier nodes and their neighborhoods hit by the fine-grained mapping rules in the dual-stream network topology, the corresponding subgraph structure is extracted as a local subgraph, and local feature alignment is performed through feature propagation to output the adaptation strength score of each supplier in the key dimensions. Based on the distribution of the adaptation strength score across dimensions, adaptive threshold pruning is performed to identify advantageous suppliers and generate an adjusted elastic weight distribution sequence. Obtain the supplier's basic capability score, and then merge the weight distribution sequence with the supplier's basic capability score to generate a comprehensive ranking score that represents the final matching degree between the order and the supplier. Supplier recommendation decisions are made based on the comprehensive ranking score. If the performance risk is detected to exceed a preset threshold, manual review is triggered and the decision log is recorded.

2. The multi-level weight allocation method for contract signing and project initiation that integrates supplier profiling and order quality as described in claim 1, is characterized in that, Based on the time window associated with the semantic fingerprint of order complexity, real-time capability data of suppliers is collected and time-series compressed to generate a timestamped supplier capability summary vector, including: Based on the time window associated with the semantic fingerprint of the order complexity, the original capability monitoring data stream of the registered supplier is obtained, and missing value imputation and outlier removal are performed on the original capability monitoring data stream to generate a standardized supplier multidimensional capability time series sequence. Based on the standardized supplier multidimensional capability time series, trend features are statistically extracted using a sliding time window, and the scheduling fluctuation coefficient and the mean performance deviation are calculated as numerical capability feature components. Based on the numerical capability feature components, a set mapping encoding is performed on the engineer skill map coverage list to generate a high-dimensional sparse skill vector, and a one-hot encoding is performed on the cross-system integration ready state flag to generate an integration state vector. The high-dimensional sparse skill vector and the integration state vector are used as structured capability feature components. Based on the numerical capability feature components and the structured capability feature components, principal component analysis is used for dimensionality reduction and fusion compression to generate a compressed vector of supplier capability features. Based on the compressed vector of the supplier capability features, a timestamp is appended and encapsulated to generate the supplier capability summary vector.

3. The multi-level weight allocation method for contract signing and project initiation that integrates supplier profiling and order quality as described in claim 1, is characterized in that... Based on the distribution of the adaptation strength score across all dimensions, adaptive threshold pruning is performed to identify advantageous suppliers, generating an adjusted elastic weight distribution sequence, including: The adaptation strength scores are grouped and aggregated according to the key dimensions, and the statistical dispersion of the score set under each dimension is calculated to obtain the distribution variance data of the key dimensions. Dynamic threshold mapping is performed based on the key dimension distribution variance data to generate dimension-specific clipping threshold parameters. Using the dimension-specific pruning threshold parameter, extreme value filtering is performed on the adaptation strength score to lock the top supplier nodes that exceed the threshold and obtain the top supplier advantage dimension tag set; Based on the set of advantageous dimension labels of the leading suppliers, weight shrinkage and redistribution are performed. The initial weight coefficients of the advantageous dimensions are compressed, and the released weight increments are inversely distributed to the non-advantageous dimensions to generate the elastic weight distribution sequence.

4. The multi-level weight allocation method for contract signing and project initiation that integrates supplier profiling and order quality as described in claim 1, characterized in that, The annotation of the key dimensions in the semantic fingerprint of order complexity includes: Based on the distribution entropy values ​​of the initial sparse vector corresponding to the semantic fingerprint of the order complexity in each dimension, the dimension label with the largest deviation is selected by dynamic thresholding and used as the key dimension.

5. The multi-level weight allocation method for contract signing and project initiation that integrates supplier profiling and order quality as described in claim 1, characterized in that, The generation of the supplier capability summary vector includes: The numerical capability feature components formed by time-series compression of the supplier's real-time capability data are combined with the structured capability feature components by principal component analysis to reduce dimensionality and fuse them, generating a low-dimensional dense capability feature compressed vector. After adding a timestamp, the vector is encapsulated as the supplier's capability summary vector.

6. The multi-level weight allocation method for contract signing and project initiation that integrates supplier profiling and order quality as described in claim 1, characterized in that, The construction of the positive connection and the negative suppression connection in the two-stream network topology includes: Based on a fine-grained mapping rule base derived from historical matching samples, the query node is matched and verified with the candidate node. If the high confidence condition logic is met, the activation edge of the positive connection is established. If there is a conflict or the low confidence condition logic is not met, the deweighted edge of the negative suppression connection is established.

7. The multi-level weight allocation method for contract signing and project initiation that integrates supplier profiling and order quality as described in claim 1, characterized in that, The generation of the adaptation strength score includes: In the local subgraph, the features of the query node and the supplier node are weighted and aggregated and aligned with multi-head attention through a two-layer graph attention propagation mechanism. After dimensional mapping, the adaptation strength score of each supplier on the key dimension is output.

8. The multi-level weight allocation method for contract signing and project initiation that integrates supplier profiling and order quality as described in claim 1, characterized in that, The generation of the elastic weight distribution sequence includes: Based on the distribution variance of the adaptation strength score on each of the key dimensions, a dimension-specific pruning threshold is generated through dynamic threshold mapping. Advantageous supplier nodes that exceed the pruning threshold are locked, the initial weights of the advantageous dimensions are compressed, and the released weight increments are inversely distributed to the non-advantageous dimensions to generate the elastic weight distribution sequence.

9. The multi-level weight allocation method for contract signing and project initiation that integrates supplier profiling and order quality as described in claim 1, characterized in that, The generation of the comprehensive ranking score includes: The normalized elastic weight distribution sequence is used as an adjustment factor to nonlinearly and dynamically scale the supplier basic capability score. After cross-dimensional interactive feature extraction and weighted aggregation, the comprehensive ranking score is generated.

10. The multi-level weight allocation method for contract signing and project initiation that integrates supplier profiling and order quality as described in claim 1, characterized in that, The performance risk detection in the supplier recommendation decision includes: The comprehensive ranking score and the adaptation strength score are aggregated in a multidimensional weighted manner to generate a performance risk index. If the performance risk index exceeds a preset threshold, manual review is triggered.

Images