A Method and System for Risk Assessment in Industrial Equipment Supply Chain Finance
By constructing a mapping index and multi-layer semantic tree for multi-source heterogeneous data, generating a rule association graph, and using the risk impact transmission tensor for iterative updates, the problem of industrial equipment supply chain financial risk assessment under multi-source heterogeneous data is solved, achieving a more accurate and comprehensive risk assessment.
Patent Information
- Application Number
- CN202611366151.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-04
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies are insufficient to effectively handle industrial equipment supply chain finance risk assessment under multi-source heterogeneous data, leading to inaccurate risk feature extraction and frequent rule conflicts, which affect the accuracy and comprehensiveness of the assessment results.
By constructing a mapping index for multi-source heterogeneous data, the risk rules are parsed into a multi-layer semantic tree, generating a rule association graph, and iteratively updating using the risk impact transmission tensor to achieve quantitative convergence of the risk profile and generate supply chain finance risk assessment results.
It improves the efficiency of rule loading under multi-source heterogeneous data, reduces the redundancy of rule sets, ensures the consistency of the rule system, and enhances the accuracy and comprehensiveness of risk assessment.
Smart Images

Figure CN123115679A_ABST
Abstract
Description
Technical Field
[0001] This application pertains to the field of industrial equipment supply chain, and in particular relates to a method and system for assessing financial risks in the industrial equipment supply chain. Background Technology
[0002] Supply chain finance for industrial equipment involves multiple stages, including raw material procurement, equipment manufacturing, and sales distribution. The flow of funds, logistics, and information are intricately intertwined, requiring the processing of heterogeneous data from multiple channels and dimensions during risk assessment. As supply chain business scenarios become increasingly complex, the scale of data and the number of rules are growing in tandem. When faced with multi-source heterogeneous data, it is often difficult to map and filter key data features against complex risk rules. Risk rules developed in actual business operations often exhibit logical overlap and conditional overlap. When multiple rules are triggered simultaneously, conventional matching mechanisms struggle to reasonably assess and adjudicate rule conflicts in conjunction with business logic, easily leading to chaotic rule activation states and affecting the accuracy of risk feature extraction. Existing rule processing mechanisms often treat each risk rule as a parallel, independent judgment unit, making it difficult to truly identify the transmission effect of risk factors within the supply chain. Since the initial risk states of various dimensions may interfere with each other during evolution, simple rule result accumulation cannot reflect the transmission and convergence process of risk in the global topology. Therefore, how to select and adjudicate risk rules based on the characteristics of multi-source heterogeneous data, deeply analyze the causal relationships and topological structures between rules, and construct a mechanism that can represent the risk transmission process to obtain supply chain finance risk assessment results are problems that need to be solved in this field. Summary of the Invention
[0003] According to one aspect of the present invention, a method for assessing supply chain finance risks in industrial equipment is proposed, which is used to screen and adjudicate risk rules based on the characteristics of multi-source heterogeneous data, and to deeply analyze the causal relationships and topological structures among the rules, including the following steps: Acquire multi-source heterogeneous data of the industrial equipment supply chain, receive a set of risk rule scripts, extract key features from the multi-source heterogeneous data, construct a mapping index based on the correlation between the key features and the risk rule scripts, and filter candidate rule sets from the set of risk rule scripts according to the mapping index. For each rule in the candidate rule set, the rule is parsed into a multi-layer semantic tree containing causal relationship labels and time window constraints. The semantic fingerprint of the multi-layer semantic tree branches is calculated. When the semantic fingerprints of multiple rules meet the similarity condition, the corresponding rules are aggregated into composite rule nodes. The prior confidence is calculated for each rule and composite rule node. When the rule triggering conditions overlap, the conflict resolution is performed based on the fit between the multi-layer semantic tree activation path and the multi-source heterogeneous data and the prior confidence to obtain the active rule subset. A rule association graph is generated based on the causal relationship labels of each node in the activity rule subset. Based on the multi-layer semantic tree parameters corresponding to each node in the rule association graph and combined with the time window constraints, an initial risk profile vector is obtained by mapping according to a preset unified feature dimension. A risk impact transmission tensor is set according to the topology of the rule association graph and the causal relationship of the edges. The initial risk profile vector is iteratively updated based on the risk impact transmission tensor until the risk profile vectors of all nodes converge. The converged risk profile vectors are weighted and aggregated to obtain the financial risk assessment result of the supply chain.
[0004] According to another aspect of the present invention, an industrial equipment supply chain finance risk assessment system is proposed, comprising the following modules: The module is used to acquire multi-source heterogeneous data of the industrial equipment supply chain, receive a set of risk rule scripts, extract key features from the multi-source heterogeneous data, construct a mapping index based on the correlation between the key features and the risk rule scripts, and filter candidate rule sets from the set of risk rule scripts according to the mapping index. The calculation module is used to parse each rule in the candidate rule set into a multi-layer semantic tree containing causal relationship labels and time window constraints, calculate the semantic fingerprint of the multi-layer semantic tree branches, and aggregate the corresponding rules into composite rule nodes when the semantic fingerprints of multiple rules meet the similarity conditions. It calculates the prior confidence for each rule and composite rule node, and when the rule triggering conditions overlap, it makes a conflict resolution based on the fit between the multi-layer semantic tree activation path and the multi-source heterogeneous data and the prior confidence to obtain a subset of active rules. The generation module is used to generate a rule association graph based on the causal relationship labels of each node in the activity rule subset. Based on the multi-layer semantic tree parameters corresponding to each node in the rule association graph and combined with the time window constraints, an initial risk profile vector is obtained by mapping according to a preset unified feature dimension. A risk impact transmission tensor is set according to the topology and edge causal relationship of the rule association graph. The initial risk profile vector is iteratively updated based on the risk impact transmission tensor until the risk profile vectors of all nodes converge. The converged risk profile vectors are weighted and aggregated to obtain the financial risk assessment result of the supply chain.
[0005] This invention improves the efficiency of rule loading under variable data by constructing a mapping index between key features of multi-source heterogeneous data and risk rule scripts, enabling the screening of candidate rule sets. Rules are parsed into multi-layered semantic trees containing causal relationship markers and time window constraints. Semantic fingerprints are used to aggregate similar rules, reducing the redundancy of the rule set. When rule triggering conditions overlap, conflict resolution is performed by combining the fit between the multi-layered semantic tree activation path and the data, as well as prior confidence, ensuring the consistency of the rule system. Furthermore, a correlation graph is generated based on the causal relationships of rule nodes, and the risk impact transmission tensor is used to iteratively update the risk profile vector, representing the interrelationships and network transmission effects among various risk factors in the supply chain, thus improving the accuracy and comprehensiveness of industrial equipment supply chain financial risk assessment. Attached Figure Description
[0006] Figure 1 A flowchart of the first embodiment; Figure 2 A priority sorting bar chart for conflict rule resolution. Detailed Implementation
[0007] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0008] In the first embodiment, the present invention proposes a method for assessing supply chain financial risks in industrial equipment, such as... Figure 1 It includes the following steps: S1. Acquire multi-source heterogeneous data of the industrial equipment supply chain, receive a set of risk rule scripts, extract key features from the multi-source heterogeneous data, construct a mapping index based on the correlation between the key features and the risk rule scripts, and filter candidate rule sets from the set of risk rule scripts according to the mapping index.
[0009] The system collects structured financial transaction records and order statuses, as well as unstructured purchase contracts and industry news texts, as multi-source heterogeneous data for the industrial equipment supply chain. It receives a set of domain-specific language risk rule scripts pre-written by business experts through a rule management module. Unstructured text is segmented and entity recognized, and combined with statistical analysis of structured data, key features representing enterprise creditworthiness, contract fulfillment capabilities, and market environment are extracted. Prerequisite parameters set in each risk rule script are extracted, and the matching degree between the extracted key features and the prerequisite parameters is calculated. Key features with matching degrees exceeding a set threshold are paired with their corresponding risk rule script identifiers to construct a hash-structured mapping index. All extracted key features are traversed, and matching risk rule script identifiers are searched in the mapping index. Rules corresponding to all found identifiers are aggregated, and duplicates are removed to obtain a candidate rule set.
[0010] In an optional embodiment, extracting key features from the multi-source heterogeneous data includes: The multi-source heterogeneous data is classified into enterprise business registration data, transaction records, logistics and warehousing records, and equipment operating condition data. The registered capital, years of establishment, and number of violations recorded are extracted from the enterprise's business registration data as characteristics of the enterprise's qualification. The transaction credit characteristics are obtained by calculating the amount of income and expenditure, transaction frequency, and number of overdue payments in the transaction flow using statistical analysis algorithms. Normalization is performed on the inbound frequency, transit time difference, and return quantity in the logistics warehousing records to obtain logistics operation characteristics; The temperature parameters in the equipment operating condition data are averaged and sampled at a preset time step, the operating status log is numerically encoded, and the sampled temperature parameters and the encoded operating status log are feature-concatenated to obtain the equipment operating characteristics. The key features are obtained by cross-category alignment verification and splicing and recombination of the enterprise entity qualification features, transaction credit features, logistics operation features and equipment operation features.
[0011] The system parses the file extension and structure identifier of the input data packet, parsing and mapping multi-source heterogeneous data in XML or JSON format to four independent structured relational tables: Enterprise Business Registration Data Table, Transaction Flow Table, Logistics and Warehousing Record Table, and Equipment Operating Condition Data Table. For the Enterprise Business Registration Data Table, it extracts fields such as registered capital (e.g., 50 million yuan), years of establishment (e.g., 5.5 years), and number of violations (e.g., 0 times), constructing a 3-dimensional enterprise entity qualification feature vector. For the Transaction Flow Table, it sets a 6-month sliding time window, calls the SQL aggregation function SUM to calculate total income and expenditure, calls COUNT to calculate the average transaction frequency monthly, and counts the number of overdue payments with an overdue flag of 1, generating a transaction credit feature vector.
[0012] For logistics and warehousing records, extract the average monthly inbound frequency, transit time difference, and return quantity, and apply the Min-Max normalization formula. Scaling up, when and When values are equal, the corresponding normalized result is set to a preset fixed value or the normalization process for that dimension is skipped. For example, the inbound frequency of 20 has an extreme value of 0 to 50, the transit time difference of 48 hours has an extreme value of 0 to 120, and the return of 2 items has an extreme value of 0 to 10. A floating-point normalized logistics operation feature vector is then calculated. For equipment operating condition data, a time step is set. =5 minutes, calculate the arithmetic mean of the temperature sensor sequence within the time series, establish a status dictionary to map the text operation log to numerical values, such as normal = 0, warning = 1, shutdown = 2, and concatenate the sampled mean with the coded status to generate the equipment operation feature vector. Using the enterprise unified social credit code as the alignment key, perform a horizontal join operation on the four sets of sub-feature vectors. For missing fields caused by asynchronous data sources, fill in with 0 or the mean, and generate a one-dimensional key feature vector as the basic input data for subsequent calculations.
[0013] In an optional embodiment, the step of constructing a mapping index based on the association between the key features and the risk rule scripts, and filtering candidate rule sets from the risk rule script set according to the mapping index, includes: The form field identifiers and declaration scope parameters in the risk rule script are segmented and tagged to generate rule recognition strings; Calculate the edit distance between the identifier character of the key feature and the rule recognition string; When the edit distance is less than the preset matching threshold, the identifier character of the key feature is associated with the rule recognition string, and the mapping index is constructed by writing it into the mapping logic entity unit through non-conflicting hash addressing; Based on the mapping index, the risk rule script set is traversed, and the rules that are associated with the hit are extracted as the candidate rule set.
[0014] The abstract syntax tree of the DSL risk rule script is traversed to extract form field identifiers and declaration range parameters. The TF-IDF word segmentation tool is then used to remove stop words and special symbols from these identifiers and parameters. The resulting strings are concatenated and tagged according to a unified format of core words plus parameter thresholds to generate rule recognition strings. Metadata identifier characters corresponding to key feature vectors in the feature processing flow are extracted. A (m+1)×(n+1) two-dimensional state matrix is constructed using the Levenshtein programming algorithm. The minimum number of insertion, deletion, and replacement operations between the key feature identifier characters and each rule recognition string is calculated, yielding the edit distance integer value.
[0015] In the judgment phase, the preset matching threshold is set in the configuration item. =5. When the calculated edit distance is less than 5, it is determined that there is a semantic correspondence between the two in business logic. The string hash value of the key feature is extracted as the key, and the rule ID is used as the value. The Cuckoo Hash Algorithm is called, using two sets of independently distributed hash functions. and Alternately calculate memory slots, and resolve conflicts using alternative addressing paths when collisions occur, achieving non-collision hash addressing and writing the key-value pair to logical entity units such as Redis or mapping tables to complete the mapping index construction. Use valid identifier characters from the key feature set as query conditions, retrieve the corresponding value list through the hash mapping index in O(1) time complexity, extract the DSL script code snippets corresponding to all hit rule IDs, and combine them into a candidate rule set for downstream use.
[0016] S2, for each rule in the candidate rule set, the rule is parsed into a multi-layer semantic tree containing causal relationship labels and time window constraints. The semantic fingerprint of the multi-layer semantic tree branches is calculated. When the semantic fingerprints of multiple rules meet the similarity condition, the corresponding rules are aggregated into composite rule nodes. The prior confidence is calculated for each rule and composite rule node. When the rule triggering conditions overlap, the conflict resolution is performed based on the fitting degree between the activation path of the multi-layer semantic tree and the multi-source heterogeneous data and the prior confidence to obtain the active rule subset.
[0017] For each rule in the candidate rule set, the rule code written in the domain-specific language is transformed into an abstract syntax tree using a lexical analyzer and a syntax analyzer. Each node of the abstract syntax tree is traversed, and the condition judgment instructions in the rule are mapped to the branch nodes of the multi-level semantic tree. The execution actions in the rule are extracted into the leaf nodes of the multi-level semantic tree, and the causal relationship tags representing risk deterioration or risk suppression are bound to the corresponding leaf nodes. The effective time periods involved in the rule are uniformly converted into time window constraints and encapsulated in the root node of the multi-level semantic tree, thus constructing the multi-level semantic tree.
[0018] A pre-defined semantic vectorization algorithm is used to vectorize the branch paths from the root node to the leaf node in the multi-layer semantic tree, generating a fixed-length dense feature vector as a semantic fingerprint. The cosine distance between the semantic fingerprints of different rules is calculated as the difference distribution score. When the difference distribution score is less than a pre-defined difference threshold, it indicates that the rules are semantically similar. Further, it is determined whether the causal relationship labels of the corresponding rules are consistent, whether the types of execution actions are the same or belong to the same risk direction, and whether the risk level parameters meet the pre-defined same-level conditions. When all the above conditions are met, the multi-layer semantic trees corresponding to the similar rules are merged into a composite rule node using a cohesive hierarchical clustering algorithm. The consistent causal relationship label is used as the causal relationship label of the composite rule node, and the risk level parameters of each sub-rule are determined as the risk level parameters of the composite rule node according to the pre-defined level synthesis rule. The pre-defined level synthesis rule includes taking the highest risk level among the sub-rules or determining it according to the pre-defined level priority in the rule script. When there are opposite labels for risk deterioration and risk suppression or opposite directions of execution actions, the corresponding rule is retained as an independent rule and sent to the subsequent conflict adjudication process.
[0019] Historical risk assessment logs are retrieved, and the proportion of successful warnings for real risk events by each rule in historical applications is calculated relative to the total trigger frequency. This proportion is used as the prior confidence level for each rule. For composite rule nodes, a weighted average is calculated based on the prior confidence levels of their constituent sub-rules and the historical trigger counts of each sub-rule to obtain the prior confidence level of the composite rule node. For each rule or composite rule node before entering the active rule subset, its fit with the current multi-source heterogeneous data is calculated. The fit level uniformly represents the degree of factual matching of the rule with the multi-source heterogeneous data within the current evaluation period. Specifically, it is the coverage density quotient obtained by dividing the frequency of successfully matched factual judgments within the current evaluation period by the total number of corresponding global data samples within the current evaluation period. For rules containing only a single activation path, the frequency of factual judgments can be determined by the number of nodes in the current activation path where factual judgments are matched. If multiple rules are found to be in the same time window during execution, and their execution actions, risk directions, or risk control conclusions are logically mutually exclusive, the fit degree is multiplied by the corresponding prior confidence to obtain the comprehensive evaluation weight. The rule or composite rule node with the highest comprehensive evaluation weight is retained, and the remaining conflicting rules are removed to obtain a subset of conflict-free active rules.
[0020] In an optional embodiment, for each rule in the candidate rule set, parsing the rule into a multi-layer semantic tree containing causal relationship tags and time window constraints includes: The lexical semantic deconstruction module is invoked to map the condition judgment instructions in the rules to branch nodes of a multi-layer semantic tree, and to extract the execution actions in the rules to leaf nodes of the multi-layer semantic tree, thus constructing a multi-layer tree topology model. Add a positive feedback forward vector label to the leaf node representing risk deterioration, and add an antagonistic vector label to the leaf node representing risk inhibition to generate the causal relationship label; The valid numbers and units of the time field in the rules are extracted, uniformly converted into time window values with preset precision, and encapsulated in the root node of a multi-layer semantic tree to generate the time window constraint.
[0021] The lexical-semantic deconstruction module based on LL(1) top-down lexical analysis is invoked to perform lexical token scanning and syntax tree transformation on the textual DSL statements in the candidate rule set. Boolean logic operators and conditional judgment instructions are extracted and mapped to control branch nodes of a multi-layer semantic tree, and the final execution action that triggers the risk control response is extracted as the leaf node at the end of the tree structure. Parent and child nodes are linked in the memory heap through object reference pointers, and a multi-layer tree topology model with a depth of N containing judgment paths is instantiated.
[0022] For causal relationship labeling, a hash dictionary containing risk deterioration keywords and a hash dictionary containing risk suppression keywords are preloaded. When the text content of a leaf node in the multi-level semantic tree successfully matches the deterioration action dictionary, a 3-dimensional real tensor is bound to that node as a positive feedback forward vector label; when it matches the suppression action dictionary, an antagonistic vector label is assigned. A built-in time regular expression extractor identifies the numerical constants and units of time fields in the rule text, calls a clock conversion function, and converts the extracted units into integer timestamps. This duration value is then set as a time window variable with a preset precision. This data is then encapsulated as a global attribute parameter and written into the root node data structure of the multi-layer semantic tree, thus completing the structured encapsulation of the rule space and time dimension.
[0023] In an optional embodiment, the step of calculating the semantic fingerprints of the multi-layer semantic tree branches, and aggregating the corresponding rules into composite rule nodes when the semantic fingerprints of multiple rules satisfy a similarity condition, includes: The character tags of each node on the decision path of the multi-level semantic tree are concatenated and combined according to the order of logical execution to generate a standard-length path string; The path string is transformed using a preset semantic vectorization algorithm to generate a fixed-length feature vector as the semantic fingerprint; The difference distribution score is obtained by calculating the feature distance between the semantic fingerprints of two rules; When the difference distribution score is less than the preset difference threshold, and the causal relationship markers of the two rules are consistent, the execution action types are the same or belong to the same risk direction, and the risk level parameters meet the preset same level conditions, it is determined that the two rules meet the composite aggregation conditions, and the two rules are recombined and bound to generate the composite rule node.
[0024] A depth-first search algorithm is used to search downwards from the root node of the multi-layer semantic tree along the execution sequence of business logic decisions, extracting identifiers or business tags from each intermediate node and leaf node along the activation path. Underscores are used as delimiters to concatenate and combine strings. Character streams exceeding a set length are truncated, and shorter streams are padded with whitespace at the end, generating a standard path string with a length of 256 characters. A pre-trained BERT network model is then called. This BERT network model consists of a word embedding input layer, a feature extraction layer composed of multiple bidirectional Transformer encoders with multi-head self-attention mechanisms and feedforward networks stacked in series, and a convergent output layer. The input is the standard path string obtained above, and the output is a high-order semantic representation feature vector after deep feature extraction of the input text. The standardized path string is then input into the BERT network model to be converted into a multi-dimensional word embedding representation. After dimensionality reduction of the hidden layer sequence through an average pooling layer, a dense feature vector with a constant dimension of 768 and a data type of Float32 is output. The vector is then cached in the vector library as the global semantic fingerprint of the current rule decision branch.
[0025] In the similarity verification stage, semantic fingerprint vectors of rule pairs triggered within the current execution cycle are extracted in parallel using a tensor computation interface. and Substitute into the cosine distance formula The angle difference between the two rules in multidimensional Euclidean space is calculated, and the result is a scalar floating-point number in the interval 0 to 2, which is taken as the difference distribution score. The calculated difference distribution score is compared with the preset difference threshold issued by the parameter configuration center. Since the actual score is lower than the preset difference threshold, and the causal relationship markers of the two rules are consistent, the execution action types are the same or belong to the same risk direction, and the risk level parameters meet the preset same-level conditions, the two rules are determined to meet the composite aggregation condition. A new heap block is allocated in the runtime environment to create a new logical aggregation wrapper node class, and the root node pointers of the two rules are simultaneously mounted and registered as child elements under this parent class node, completing the composite recombination and binding of similar rules at the structural level.
[0026] In an optional embodiment, the step of obtaining a subset of active rules by resolving conflicts based on the fit between the multi-layer semantic tree activation path and the multi-source heterogeneous data, as well as the prior confidence, when rule triggering conditions overlap, includes: When multiple rules overlap in time sequence and their corresponding execution actions, risk directions, or risk control conclusions are logically mutually exclusive, the frequency of fact judgment hits for each rule in the global data sample is counted. The coverage density quotient obtained by dividing the frequency of the fact determination by the total number of global data samples is used as the goodness of fit. Extract the number of times each rule successfully warned of risk events within the historically set observation period, and divide the number of successful risk warnings by the total number of times the corresponding rule was triggered in history to obtain the confidence ratio score as the prior confidence level; The comprehensive evaluation weight is obtained by multiplying the goodness of fit by the prior confidence level. The conflicting rules are sorted in descending order according to the comprehensive evaluation weights. The rule with the highest weight is added to the active rule subset, and the execution queue of the remaining conflicting rules is removed.
[0027] The rule trigger result queue is continuously monitored using a read-write lock mechanism. Once two or more risk control rules are detected simultaneously reaching the activation threshold for the same enterprise entity record under the same timestamp slice, and the leaf node action labels of the multi-level semantic tree are marked as mutually exclusive (positive and negative), the current process is interrupted and the conflict adjudication subroutine is entered. The SQL analysis module is used to query the multi-source heterogeneous data sample set that has completed cross-category alignment within the current evaluation period. The hit frequency of each specific rule involved in the conflict is counted in the global data sample of the current evaluation period, traversing multiple levels of the tree and hitting the facts. A floating-point division operation is performed to divide the hit frequency by the total number of corresponding samples in the current evaluation period to calculate the coverage density quotient, which is used to indicate the conditional fit between the current multi-source heterogeneous data and the rule itself.
[0028] The historical review period is set for the past twelve months as defined in the online operation logs. Based on audit feedback records, the number of risk events warned by each conflict rule is independently aggregated and calculated. This result is then divided by the total number of historical triggers within this review period to obtain a confidence ratio score representing the model's precision, which serves as the prior confidence parameter. The goodness of fit is multiplied by the prior confidence to obtain the comprehensive evaluation weight. The comprehensive evaluation weights calculated for all mutually exclusive rules within the current decision group are sorted in descending order using a quicksort algorithm with a time complexity of O(NlogN). Only the rule entity with the highest comprehensive evaluation weight is retained and copied to the active rule subset. The remaining rules with lower weights are marked as inactive and removed from the current action execution queue. Figure 2 As shown.
[0029] S3. Generate a rule association graph based on the causal relationship labels of each node in the activity rule subset. Based on the multi-layer semantic tree parameters corresponding to each node in the rule association graph and combined with the time window constraints, obtain an initial risk profile vector according to the preset unified feature dimension mapping. Set the risk impact transmission tensor according to the topology and edge causal relationship of the rule association graph. Iteratively update the initial risk profile vector based on the risk impact transmission tensor until the risk profile vectors of all nodes converge. Weighted aggregation of the converged risk profile vectors yields the financial risk assessment result of the supply chain.
[0030] Each independent and composite rule node in the activity rule subset is used as a vertex in the graph. Field identifiers, comparison operators, and declaration range parameters of the condition judgment branch nodes are extracted from the multi-layer semantic tree corresponding to each node to generate a precondition structure. The conclusion structure is extracted from the execution action, risk status change result, or risk control conclusion fields corresponding to the leaf nodes. The precondition structure and conclusion structure are then converted into semantic identifiers in a unified format. If the semantic matching degree between the conclusion structure of one node and the precondition structure of another node reaches a preset association threshold, a directed edge is established between the two vertices. A rule association graph describing the risk transmission path is generated. Each node in the rule association graph is traversed, and the risk severity is mapped according to the causal relationship label, rule action type, and risk level parameter of the node. The risk level parameter is determined by the risk level label, declaration range parameter, or rule action type in the risk rule script according to a preset level mapping dictionary. The probability of risk occurrence is obtained based on the prior confidence and goodness of fit of the node. The time window decay factor is calculated based on the time difference between the time window constraint encapsulated in the root node of the node and the current evaluation time. The risk severity, the probability of risk occurrence, and the time window decay factor are mapped to the initial risk profile vector of the node according to a preset unified feature dimension and fixed order.
[0031] A risk impact transmission tensor is constructed based on the adjacency matrix of the rule-based association graph. This tensor acts on the risk profile vector of each node according to the node dimension. The values of its elements are determined by the causal relationship strength weights of the directed edges, which are calculated by the feature overlap between the preconditions and conclusions of the two nodes. A damped diffusion process based on a non-negative transmission strength matrix and a signed causal direction matrix is adopted. In each iteration, the matrix composed of the risk profile vectors of all current nodes is multiplied with the risk impact transmission tensor. The non-negative transmission strength matrix is used to represent the risk transmission ratio between nodes, and the signed causal direction matrix is used to distinguish between risk deterioration and risk inhibition relationships. The risk inhibition relationship serves as a directional inhibition correction of the transmission result. The initial risk profile matrix is fused according to a preset damping coefficient to retain the initial risk state of the node itself. The difference between the risk profile vector matrices generated in two adjacent iterations is calculated. The iteration stops when the difference is less than a set convergence limit. The node risk weight coefficient of each node in the association graph is extracted according to the business risk weight dictionary. The node risk weight coefficient is used to perform a weighted summation on the converged risk profile vector to obtain a global risk vector. The global risk vector is then multiplied by the preset risk classification model weight vector and converted into a risk score value between 0 and 100 through a preset mapping function, which serves as the financial risk assessment result for the industrial equipment supply chain.
[0032] In an optional embodiment, the iterative update of the initial risk profile vector based on the risk impact propagation tensor until the risk profile vectors of all nodes converge includes: The initial risk profile vector of each node in the rule association graph is used as the input queue, and matrix multiplication is performed in combination with the normalized risk impact propagation tensor to obtain the updated risk profile vector at the current time step. The updated risk profile vector is used as the input queue for the next time step, and the matrix multiplication operation is performed iteratively to update it. Between two adjacent iterations, the absolute deviation of the risk profile vector for each node is calculated, and the absolute deviations of all nodes are weighted and summed to obtain the global absolute offset. When the global absolute offset is less than the preset threshold base value, the risk profile vector is determined to have reached a convergence state, the iteration is stopped, and the finally converged risk profile vector is output.
[0033] Based on the complex graph topology environment composed of N rule nodes or composite rule nodes in the rule association graph, the parameters such as risk severity, risk occurrence probability, and time window decay factor assigned to each node during the initialization phase are mapped according to a preset unified feature dimension d, and assembled into an initial risk profile matrix R(0) of dimension N×d, which serves as the initial input for the graph neural network propagation. The network structure of the graph neural network includes a node feature input layer, a graph convolutional information transmission layer composed of multiple mapping and activation units, and an output layer representing the final convergence state. The input is the assembled initial risk profile matrix and the node dimension matrix representation of the risk impact transmission tensor representing the causal connection strength between nodes. The output is a converged risk profile matrix that has been iterated through local neighbor feature convergence and global graph structure propagation until an equilibrium state is reached. The normalization algorithm is called to perform scale alignment on the pre-established adjacency weight graph reflecting the causal connection strength between nodes, constructing a risk impact transmission matrix M of dimension N×N, which serves as the matrix representation of the risk impact transmission tensor in the node dimension. First, a non-negative transmission strength matrix P is constructed based on the causal connection strength between nodes, and P is normalized row by row so that the sum of the elements in each row is 1. Then, a signed causal direction matrix C is constructed based on the causal relationship label, with the risk deterioration relationship corresponding to the positive label and the risk inhibition relationship corresponding to the negative label. P and C are multiplied element by element to obtain the signed influence matrix G, which is only used for risk influence direction correction.
[0034] At the beginning of the t=1th time step of the time-series evolution, the instruction executes matrix operations, according to... Perform intermediate updates, among which The damping coefficient is preset and satisfies 0 < <1, the sum of the absolute values of each row of G is no greater than 1, putting the iteration process in a bounded contraction state; perform dimensional constraint projection processing on the intermediate update results, limiting the probability of risk occurrence and the time window decay factor to the interval between 0 and 1, and limiting the severity of risk to a preset non-negative level interval, to obtain ,in This represents a non-negative truncation or interval projection operator based on the value range of each risk profile dimension. It realizes the weighted fusion calculation of the initial risk of the current node and the risk transmitted by the adjacent nodes, and solves the network-wide updated risk profile matrix R(1) after the current iteration cycle is completed.
[0035] After obtaining the first round of update results, the updated risk profile vector is written back and overwrites the original state memory area, becoming the starting point of the input queue for the next cycle. The main loop is then started to repeatedly trigger matrix multiplication operations for continuous iterative transmission of risk impact within the graph. To monitor stability, during the intervals between each matrix multiplication operation, the N×d-dimensional risk profile matrix is expanded element-wise, element-wise subtraction is performed, and the absolute value is taken to obtain the absolute deviation of each network node across each feature dimension. Based on the degree centrality weights calculated from the graph or preset node weights, this absolute deviation is multiplied and added together to form a global absolute offset floating-point number representing the overall network's degree of change. The calculation unit then compares the global absolute offset value obtained in each round with the threshold base values pre-stored in the threshold configuration file. Inequality comparison is performed; once the global absolute offset is detected to be continuously decaying and less than the threshold base value, it can be determined that the risk influence transmission process based on damped diffusion in the entire causal relationship network has reached equilibrium, the loop is exited and the iteration state is exited, and the risk profile vector data structure at this time is saved, and the data output is used as the converged risk profile vector.
[0036] In an optional embodiment, the weighted aggregation of the converged risk profile vectors to obtain the financial risk assessment result of the supply chain includes: The node risk weight coefficients of each node in the association diagram are extracted according to the rules of the business risk weight dictionary; The risk profile vector of the corresponding node after convergence is obtained by scaling the risk profile vector by multiplying the risk profile vector of the corresponding node under the preset unified feature dimension using the node risk weight coefficient. The scaled risk profile vectors of all nodes are added and merged according to the corresponding dimensions in the preset unified feature dimension to obtain a fused one-dimensional risk parameter array; The one-dimensional risk parameter array is multiplied by the preset risk classification model weight vector to obtain a scalar score with a defined numerical range. The scalar score is used as the result of the financial risk assessment of the supply chain.
[0037] By using string matching and key-value binding, the business risk weight dictionary configuration table stored in a distributed non-relational database is accessed. Using the type encoding of the rule-based graph nodes as the search condition, the corresponding node risk weight coefficients are queried and extracted. The multi-dimensional risk profile vectors of N nodes in the graph, which have been confirmed to be converged in the iterative calculation module, are traversed and read. For each node's vector, using the corresponding coefficients extracted in the previous step, a scalar multiplication operation is performed in the vector space using a single-instruction, multi-data-stream parallel computing instruction. This scales the original risk values of each feature dimension proportionally, generating a series of scaled risk profile vector sets with the same dimensions as the original.
[0038] Entering the ensemble dimensionality reduction stage, an accumulator array buffer region is instantiated, initially containing only zeros. The risk profile vectors of all N nodes, scaled according to their respective dimensional position indices, are sequentially accumulated and aggregated. Through feature recombination and removal of node-level dimensional differences, the result is flattened and merged into a one-dimensional risk parameter array. A machine learning risk classifier model, pre-trained and saved using historical blacklist and whitelist data, is loaded. The fixed-dimensional feature coefficients, i.e., the risk classification model weight vector, are extracted from the model. Calculate the inner product of the aforementioned one-dimensional risk parameter array and the model weight vector. The inner product is calculated as follows: The original risk score value z is obtained; when using the Sigmoid mapping, according to A risk score ranging from 0 to 100 is calculated. When using a transformation function, the score is first obtained according to a preset scaling factor and bias term. Then, the score is truncated to a lower limit of 0 and an upper limit of 100 to obtain a risk score within the 0 to 100 range. This scalar score value is stored in the database and displayed on the front-end dashboard as a financial risk assessment result for evaluating the overall health and stability of the supply chain.
[0039] In the second embodiment, the present invention also proposes an industrial equipment supply chain finance risk assessment system, comprising the following modules: The module is used to acquire multi-source heterogeneous data of the industrial equipment supply chain, receive a set of risk rule scripts, extract key features from the multi-source heterogeneous data, construct a mapping index based on the correlation between the key features and the risk rule scripts, and filter candidate rule sets from the set of risk rule scripts according to the mapping index. The calculation module is used to parse each rule in the candidate rule set into a multi-layer semantic tree containing causal relationship labels and time window constraints, calculate the semantic fingerprint of the multi-layer semantic tree branches, and aggregate the corresponding rules into composite rule nodes when the semantic fingerprints of multiple rules meet the similarity conditions. It calculates the prior confidence for each rule and composite rule node, and when the rule triggering conditions overlap, it makes a conflict resolution based on the fit between the multi-layer semantic tree activation path and the multi-source heterogeneous data and the prior confidence to obtain a subset of active rules. The generation module is used to generate a rule association graph based on the causal relationship labels of each node in the activity rule subset. Based on the multi-layer semantic tree parameters corresponding to each node in the rule association graph and combined with the time window constraints, an initial risk profile vector is obtained by mapping according to a preset unified feature dimension. A risk impact transmission tensor is set according to the topology and edge causal relationship of the rule association graph. The initial risk profile vector is iteratively updated based on the risk impact transmission tensor until the risk profile vectors of all nodes converge. The converged risk profile vectors are weighted and aggregated to obtain the financial risk assessment result of the supply chain.
[0040] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A method for assessing supply chain financial risks in industrial equipment, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data of the industrial equipment supply chain, receive a set of risk rule scripts, extract key features from the multi-source heterogeneous data, construct a mapping index based on the correlation between the key features and the risk rule scripts, and filter candidate rule sets from the set of risk rule scripts according to the mapping index. For each rule in the candidate rule set, the rule is parsed into a multi-layer semantic tree containing causal relationship labels and time window constraints. The semantic fingerprint of the multi-layer semantic tree branches is calculated. When the semantic fingerprints of multiple rules meet the similarity condition, the corresponding rules are aggregated into composite rule nodes. The prior confidence is calculated for each rule and composite rule node. When the rule triggering conditions overlap, the conflict resolution is performed based on the fit between the multi-layer semantic tree activation path and the multi-source heterogeneous data and the prior confidence to obtain the active rule subset. A rule association graph is generated based on the causal relationship labels of each node in the activity rule subset. Based on the multi-layer semantic tree parameters corresponding to each node in the rule association graph and combined with the time window constraints, an initial risk profile vector is obtained by mapping according to a preset unified feature dimension. A risk impact transmission tensor is set according to the topology of the rule association graph and the causal relationship of the edges. The initial risk profile vector is iteratively updated based on the risk impact transmission tensor until the risk profile vectors of all nodes converge. The converged risk profile vectors are weighted and aggregated to obtain the financial risk assessment result of the supply chain.
2. The method according to claim 1, characterized in that, The extraction of key features from the multi-source heterogeneous data includes: The multi-source heterogeneous data is classified into enterprise business registration data, transaction records, logistics and warehousing records, and equipment operating condition data. The registered capital, years of establishment, and number of violations recorded are extracted from the enterprise's business registration data as characteristics of the enterprise's qualification. The transaction credit characteristics are obtained by calculating the amount of income and expenditure, transaction frequency, and number of overdue payments in the transaction flow using statistical analysis algorithms. Normalization is performed on the inbound frequency, transit time difference, and return quantity in the logistics warehousing records to obtain logistics operation characteristics; The temperature parameters in the equipment operating condition data are averaged and sampled at a preset time step, the operating status log is numerically encoded, and the sampled temperature parameters and the encoded operating status log are feature-concatenated to obtain the equipment operating characteristics. The key features are obtained by cross-category alignment verification and splicing and recombination of the enterprise entity qualification features, transaction credit features, logistics operation features and equipment operation features.
3. The method according to claim 1, characterized in that, The step of constructing a mapping index based on the association between the key features and the risk rule scripts, and filtering candidate rule sets from the risk rule script set according to the mapping index, includes: The form field identifiers and declaration scope parameters in the risk rule script are segmented and tagged to generate rule recognition strings; Calculate the edit distance between the identifier character of the key feature and the rule recognition string; When the edit distance is less than the preset matching threshold, the identifier character of the key feature is associated with the rule recognition string, and the mapping index is constructed by writing it into the mapping logic entity unit through non-conflicting hash addressing; Based on the mapping index, the risk rule script set is traversed, and the rules that are associated with the hit are extracted as the candidate rule set.
4. The method according to claim 2, characterized in that, For each rule in the candidate rule set, the rule is parsed into a multi-layer semantic tree containing causal relationship markers and time window constraints, including: The lexical semantic deconstruction module is invoked to map the condition judgment instructions in the rules to branch nodes of a multi-layer semantic tree, and to extract the execution actions in the rules to leaf nodes of the multi-layer semantic tree, thus constructing a multi-layer tree topology model. Add a positive feedback forward vector label to the leaf node representing risk deterioration, and add an antagonistic vector label to the leaf node representing risk inhibition to generate the causal relationship label; The valid numbers and units of the time field in the rules are extracted, uniformly converted into time window values with preset precision, and encapsulated in the root node of a multi-layer semantic tree to generate the time window constraint.
5. The method according to claim 1, characterized in that, The calculation of semantic fingerprints for multi-layer semantic tree branches, where multiple rules' semantic fingerprints satisfy a similarity condition, aggregates the corresponding rules into composite rule nodes, includes: The character tags of each node on the decision path of the multi-level semantic tree are concatenated and combined according to the order of logical execution to generate a standard-length path string; The path string is transformed using a preset semantic vectorization algorithm to generate a fixed-length feature vector as the semantic fingerprint; The difference distribution score is obtained by calculating the feature distance between the semantic fingerprints of two rules; When the difference distribution score is less than the preset difference threshold, and the causal relationship markers of the two rules are consistent, the execution action types are the same or belong to the same risk direction, and the risk level parameters meet the preset same level conditions, it is determined that the two rules meet the composite aggregation conditions, and the two rules are recombined and bound to generate the composite rule node.
6. The method according to claim 1, characterized in that, When rule triggering conditions overlap, a conflict resolution process is performed based on the fit between the multi-layer semantic tree activation path and multi-source heterogeneous data, as well as prior confidence, to obtain a subset of active rules, including: When multiple rules overlap in time sequence and their corresponding execution actions, risk directions, or risk control conclusions are logically mutually exclusive, the frequency of fact judgment hits for each rule in the global data sample is counted. The coverage density quotient obtained by dividing the frequency of the fact determination by the total number of global data samples is used as the goodness of fit. Extract the number of times each rule successfully warned of risk events within the historically set observation period, and divide the number of successful risk warnings by the total number of times the corresponding rule was triggered in history to obtain the confidence ratio score as the prior confidence level; The comprehensive evaluation weight is obtained by multiplying the goodness of fit by the prior confidence level. The conflicting rules are sorted in descending order according to the comprehensive evaluation weights. The rule with the highest weight is added to the active rule subset, and the execution queue of the remaining conflicting rules is removed.
7. The method according to claim 1, characterized in that, The iterative update of the initial risk profile vector based on the risk impact propagation tensor until the risk profile vectors of all nodes converge includes: The initial risk profile vector of each node in the rule association graph is used as the input queue, and matrix multiplication is performed in combination with the normalized risk impact propagation tensor to obtain the updated risk profile vector at the current time step. The updated risk profile vector is used as the input queue for the next time step, and the matrix multiplication operation is performed iteratively to update it. Between two adjacent iterations, the absolute deviation of the risk profile vector for each node is calculated, and the absolute deviations of all nodes are weighted and summed to obtain the global absolute offset. When the global absolute offset is less than the preset threshold base value, the risk profile vector is determined to have reached a convergence state, the iteration is stopped, and the finally converged risk profile vector is output.
8. The method according to claim 1, characterized in that, The weighted aggregation of the converged risk profile vectors yields the financial risk assessment results for the supply chain, including: The node risk weight coefficients of each node in the association diagram are extracted according to the rules of the business risk weight dictionary; The risk profile vector of the corresponding node after convergence is obtained by scaling the risk profile vector by multiplying the risk profile vector of the corresponding node under the preset unified feature dimension using the node risk weight coefficient. The scaled risk profile vectors of all nodes are added and merged according to the corresponding dimensions in the preset unified feature dimension to obtain a fused one-dimensional risk parameter array; The one-dimensional risk parameter array is multiplied by the preset risk classification model weight vector to obtain a scalar score with a defined numerical range. The scalar score is used as the result of the financial risk assessment of the supply chain.
9. A supply chain finance risk assessment system for industrial equipment, characterized in that, Includes the following modules: The module is used to acquire multi-source heterogeneous data of the industrial equipment supply chain, receive a set of risk rule scripts, extract key features from the multi-source heterogeneous data, construct a mapping index based on the correlation between the key features and the risk rule scripts, and filter candidate rule sets from the set of risk rule scripts according to the mapping index. The calculation module is used to parse each rule in the candidate rule set into a multi-layer semantic tree containing causal relationship labels and time window constraints, calculate the semantic fingerprint of the multi-layer semantic tree branches, and aggregate the corresponding rules into composite rule nodes when the semantic fingerprints of multiple rules meet the similarity conditions. It calculates the prior confidence for each rule and composite rule node, and when the rule triggering conditions overlap, it makes a conflict resolution based on the fit between the multi-layer semantic tree activation path and the multi-source heterogeneous data and the prior confidence to obtain a subset of active rules. The generation module is used to generate a rule association graph based on the causal relationship labels of each node in the activity rule subset. Based on the multi-layer semantic tree parameters corresponding to each node in the rule association graph and combined with the time window constraints, an initial risk profile vector is obtained by mapping according to a preset unified feature dimension. A risk impact transmission tensor is set according to the topology and edge causal relationship of the rule association graph. The initial risk profile vector is iteratively updated based on the risk impact transmission tensor until the risk profile vectors of all nodes converge. The converged risk profile vectors are weighted and aggregated to obtain the financial risk assessment result of the supply chain.
10. The system according to claim 9, characterized in that, The extraction of key features from the multi-source heterogeneous data includes: The multi-source heterogeneous data is classified into enterprise business registration data, transaction records, logistics and warehousing records, and equipment operating condition data. The registered capital, years of establishment, and number of violations recorded are extracted from the enterprise's business registration data as characteristics of the enterprise's qualification. The transaction credit characteristics are obtained by calculating the amount of income and expenditure, transaction frequency, and number of overdue payments in the transaction flow using statistical analysis algorithms. Normalization is performed on the inbound frequency, transit time difference, and return quantity in the logistics warehousing records to obtain logistics operation characteristics; The temperature parameters in the equipment operating condition data are averaged and sampled at a preset time step, the operating status log is numerically encoded, and the sampled temperature parameters and the encoded operating status log are feature-concatenated to obtain the equipment operating characteristics. The key features are obtained by cross-category alignment verification and splicing and recombination of the enterprise entity qualification features, transaction credit features, logistics operation features and equipment operation features.
