Method and system for identifying semantic risk of enterprise operation data based on neural network

Abstract

The application provides a neural network-based enterprise operation data semantic risk identification method and system, relates to the technical field of financial technology and natural language processing, and the method comprises the following steps: acquiring multi-source heterogeneous operation data flow of a target enterprise in a monitoring period, sequentially performing cross-modal cleaning alignment and semantic unit segmentation on the multi-source heterogeneous operation data flow, and obtaining a standardized operation data sequence; mapping the standardized operation data sequence to a high-dimensional continuous vector space to obtain an initial representation matrix; and configuring an initial attribute weight tensor for each semantic feature node in the initial representation matrix. The application can effectively solve the technical problems of lacking high-order dynamic quantification capability for risk evolution acceleration, leading to lagging risk inflection point identification and network weight update response delay in the prior art.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and natural language processing, and in particular to a method and system for identifying semantic risks in enterprise business data based on neural networks. Background Technology

[0002] In the fields of financial risk monitoring and corporate compliance management, building semantic risk identification models based on multi-source heterogeneous data is a current technological hotspot. Existing semantic risk identification methods typically involve acquiring multi-source data such as corporate financial audit reports, supply chain contracts, and logistics settlement vouchers, extracting risk features through static word embeddings or graph neural networks, and tracking and analyzing the first-order state of risk propagation (such as risk propagation rate).

[0003] However, actual business risks (such as supply chain cash flow disruptions) often manifest as a process of risk signals accumulating and then erupting, with the core characteristic being the second-order abrupt change trend in risk propagation. A common deficiency in existing technologies is the lack of high-order dynamic quantification capabilities for this risk evolution acceleration. For example, existing models mostly focus on static or quasi-dynamic tracking of risk propagation rates, making it difficult to effectively calculate the abrupt change gradient and trend slope of risk signals within a continuous monitoring window. This makes it difficult for models to distinguish between regularly deteriorating risks and sudden risks that are accelerating towards a critical point, resulting in a fundamental lag in identifying inflection points in the risk situation. Furthermore, due to the lack of modeling for the acceleration of risk evolution, the network weight update mechanism of existing methods relies heavily on preset fixed step sizes or first-order error feedback. When risk signals experience a surge in acceleration, the model weights struggle to obtain timely adaptive compensation, causing delays in response to sudden violations or liquidity crises and missing the optimal early warning window.

[0004] The aforementioned issues are particularly prominent in scenarios such as dynamic performance evaluation of the supply chain, monitoring of abnormal logistics settlements, and analysis of internal compliance processes. Summary of the Invention

[0005] This invention provides a method and system for identifying semantic risks in enterprise business data based on neural networks. It can effectively solve the technical problems in the prior art, such as the lack of high-order dynamic quantification capabilities for the acceleration of risk evolution, resulting in lag in risk inflection point identification and delayed response to network weight updates.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for identifying semantic risks in enterprise operational data based on neural networks, the method comprising: Step 1: Obtain multi-source heterogeneous operational data streams of the target enterprise within the monitoring period. Perform cross-modal cleaning and alignment and semantic unit segmentation on the multi-source heterogeneous operational data streams in sequence to obtain standardized operational data sequences. Map the standardized operational data sequences to a high-dimensional continuous vector space to obtain an initial representation matrix. Configure initial attribute weight tensors for each semantic feature node in the initial representation matrix. Step 2: Calculate the semantic interaction potential gradient between semantic feature nodes based on the initial representation matrix and the initial attribute weight tensor; construct a dynamic semantic association topology graph based on the semantic interaction potential gradient; perform feature fusion mapping on node pairs that meet the preset bonding convergence threshold; and perform feature rejection masking on node pairs that do not meet the preset bonding convergence threshold to obtain the semantic topology graph. Step 3: Input the semantic topology graph into the preset temporal state evolution neural network to extract the risk propagation rate tensor of each semantic feature node within the continuous sampling window; Step 4: Perform temporal difference operation on the risk propagation rate tensor to obtain the risk evolution acceleration component. Based on the risk evolution acceleration component, dynamically recalibrate the node connection weights in the semantic topology graph to obtain the recalibrated semantic topology graph. Step 5: Input the recalibrated semantic topology diagram into the preset multilayer feedforward risk identification neural network, and use the nonlinear potential energy convergence method of risk semantic features to obtain the semantic risk identification results of enterprise operating data and the corresponding risk confidence level distribution.

[0007] Secondly, a neural network-based enterprise operational data semantic risk identification system includes: The data acquisition module is used to acquire multi-source heterogeneous operational data streams of the target enterprise within the monitoring period, including financial audit reports, supply chain contract performance records, logistics settlement vouchers, and internal compliance approval and processing minutes. The multi-source heterogeneous operational data streams are sequentially subjected to cross-modal cleaning and alignment and semantic unit segmentation to obtain a standardized operational data sequence. The semantic embedding module is used to map standardized business data sequences to a high-dimensional continuous vector space to obtain an initial representation matrix, and to configure an initial attribute weight tensor for each semantic feature node in the initial representation matrix. The potential energy gradient and topology construction module is used to calculate the semantic interaction potential energy gradient between each semantic feature node based on the initial representation matrix and the initial attribute weight tensor, construct a dynamic semantic association topology graph based on the semantic interaction potential energy gradient, perform feature fusion mapping on node pairs that meet the preset bonding convergence threshold, and perform feature rejection masking on node pairs that do not meet the preset bonding convergence threshold to obtain a semantic topology structure graph. The temporal evolution module is used to input the semantic topology graph into a preset temporal state evolution neural network and extract the risk propagation rate tensor of each semantic feature node within a continuous sampling window; The acceleration and recalibration module is used to perform temporal difference operations on the risk propagation rate tensor to obtain the risk evolution acceleration component, and to perform dynamic recalibration on the node connection weights in the semantic topology graph based on the risk evolution acceleration component to obtain a recalibrated semantic topology graph. The risk identification output module is used to input the recalibrated semantic topology diagram into a preset multi-layer feedforward risk identification neural network. It adopts the nonlinear potential energy convergence method of risk semantic features to obtain the semantic risk identification results of enterprise operating data and the corresponding risk confidence level distribution.

[0008] The above-described solution of the present invention has at least the following beneficial effects: By acquiring multi-source heterogeneous operational data streams and constructing an initial representation matrix and attribute weight tensor, the semantic interaction potential gradient is calculated to construct a dynamic semantic association topology graph. Furthermore, a pre-defined temporal state evolution neural network is used to extract the risk propagation rate tensor. Based on this, temporal difference operations are performed on this tensor to obtain the risk evolution acceleration component and the node connection weights are dynamically recalibrated. Finally, the risk identification result is output through a multi-layer feedforward neural network in a nonlinear potential convergence manner. This overcomes the technical problems of existing technologies, such as lack of high-order dynamic quantification capability for risk evolution acceleration, inability to effectively calculate risk mutation gradient and trend slope, and network weight update response delay, thereby achieving the following technical effects.

[0009] The technical effects include accurately capturing the second-order mutation features of risk signals to distinguish between regular and sudden risks, achieving adaptive and rapid compensation of model weights based on acceleration feedback to reduce early warning delays, and improving the signal-to-noise ratio of key risk paths through semantic potential-driven dynamic topology construction and noise isolation, ultimately effectively improving the accuracy, timeliness, and interpretability of enterprise operational risk identification. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the semantic risk identification method for enterprise operation data based on neural networks provided in an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of a neural network-based semantic risk identification system for enterprise business data, provided in an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, embodiments of the present invention propose a method for identifying semantic risks in enterprise operational data based on neural networks. The method includes the following steps: Step 1: Obtain multi-source heterogeneous operational data streams of the target enterprise within the monitoring period. Perform cross-modal cleaning and alignment and semantic unit segmentation on the multi-source heterogeneous operational data streams in sequence to obtain standardized operational data sequences. Map the standardized operational data sequences to a high-dimensional continuous vector space to obtain an initial representation matrix. Configure initial attribute weight tensors for each semantic feature node in the initial representation matrix. Step 2: Calculate the semantic interaction potential gradient between semantic feature nodes based on the initial representation matrix and the initial attribute weight tensor; construct a dynamic semantic association topology graph based on the semantic interaction potential gradient; perform feature fusion mapping on node pairs that meet the preset bonding convergence threshold; and perform feature rejection masking on node pairs that do not meet the preset bonding convergence threshold to obtain the semantic topology graph. Step 3: Input the semantic topology graph into the preset temporal state evolution neural network to extract the risk propagation rate tensor of each semantic feature node within the continuous sampling window; Step 4: Perform temporal difference operation on the risk propagation rate tensor to obtain the risk evolution acceleration component. Based on the risk evolution acceleration component, dynamically recalibrate the node connection weights in the semantic topology graph to obtain the recalibrated semantic topology graph. Step 5: Input the recalibrated semantic topology diagram into the preset multilayer feedforward risk identification neural network, and use the nonlinear potential energy convergence method of risk semantic features to obtain the semantic risk identification results of enterprise operating data and the corresponding risk confidence level distribution.

[0014] In this embodiment of the invention, semantic feature extraction and dynamic weight configuration of multi-source heterogeneous business data streams are realized. The dynamic semantic association topology graph constructed based on the semantic interaction potential energy gradient can adaptively aggregate strong risk coupling features and isolate business noise. The risk propagation rate tensor is extracted by a preset temporal state evolution neural network, and the risk evolution acceleration component is obtained by combining temporal difference operation, realizing high-order dynamic quantification of the second-order mutation trend of risk. The node connection weights are dynamically recalibrated based on the acceleration component, which effectively improves the model's response speed to sudden risks. Finally, the risk decision boundary is sharpened by the nonlinear potential energy convergence of the preset multi-layer feedforward neural network, and the risk category and confidence level distribution are accurately output, thereby improving the accuracy, timeliness and interpretability of enterprise business risk identification.

[0015] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Obtain multi-source heterogeneous operational data streams from the target enterprise within the monitoring period, encompassing financial audit reports, supply chain contract performance records, logistics settlement vouchers, and internal compliance approval process minutes. Specifically, this includes: obtaining multi-source heterogeneous operational data streams from the target enterprise within the monitoring period, encompassing financial audit reports, supply chain contract performance records, logistics settlement vouchers, and internal compliance approval process minutes; where financial audit report data originates from the financial module of the enterprise resource planning system, including periodically generated files of the balance sheet, profit and loss statement, and cash flow statement, in formats including XML, PDF, and CSV; supply chain contract performance records are exported from the supply chain management platform, covering fields such as purchase orders, delivery confirmations, and default recovery notices; logistics settlement vouchers are provided by the data interface of third-party logistics service providers, recording information such as waybill numbers, settlement amounts, and settlement timestamps; internal compliance approval process minutes are collected from the office automation approval terminal, including approval process nodes, approval opinions, and final compliance judgment results; different data sources differ in data format, character encoding, field naming conventions, and timestamp precision, collectively forming a loosely structured and semantically ambiguous multi-source heterogeneous operational data stream.

[0016] Step 1.2 involves performing cross-modal data cleaning protocol parsing and global time-series alignment on the multi-source heterogeneous operational data stream to obtain an initial standardized data sequence with unified dimensions. Specifically, this includes: first, calling a pre-set cross-modal data cleaning protocol, which is configured with cleaning rule subsets corresponding to four data modalities: financial audit reports, supply chain contract performance records, logistics settlement vouchers, and internal compliance approval flow minutes. The protocol automatically matches the corresponding rule set through the data source identifier. The protocol first traverses each record of each data modality, calculates the non-empty mean of its local neighborhood for numerical fields, and replaces missing positions with this mean. For text fields, it extracts the content of fields with the same name from adjacent records in the same modality, and selects the most frequently occurring words as default placeholders for completion. Furthermore, it performs duplicate record detection on each modality's data based on the business primary key combination. When multiple records with the same primary key are identified, only the version with the latest timestamp is retained, and the remaining entries are deleted.

[0017] Based on this, the field formats of each modal data are standardized using the aforementioned protocol: first, the original format of the date field is identified and uniformly output as an ISO8601 standard date string; for all monetary fields, the monetary amounts are converted to the base currency amount according to a preset currency code and exchange rate table; the character encoding of all text fields is uniformly converted to UTF-8; after the above cleaning is completed, the global time series alignment stage begins. First, a target time point sequence is generated according to the start and end times of the monitoring period and a set uniform sampling step size; the original timestamp of each data record is aligned to the nearest target time point through rounding mapping; for cases where multiple records exist at the same target time point after alignment, the median is aggregated for numeric fields and the latest value is taken for text fields; for cases where there are no corresponding records at some target time points, numeric fields are filled with linear interpolation using the effective record values ​​before and after, and text or categorical fields are filled with forward effective values; through the above processing, an initial standardized data sequence that is completely consistent in both the time and field dimensions is finally obtained, where each row in the sequence corresponds to a target time point and each column corresponds to a standardized field.

[0018] Step 1.3: Call the pre-set enterprise business domain ontology library to perform sliding window semantic boundary detection and dependency syntax parsing on the initial standardized data sequence to obtain semantic boundary and syntax relation parsing information; based on the semantic boundary and syntax relation parsing information, the initial standardized data sequence is discretized into several atomic semantic feature units with independent business semantic orientations, and arranged in order according to the sequential position of the atomic semantic feature units in the initial standardized data sequence to obtain a standardized business data sequence. Specifically, the pre-set enterprise business domain ontology library is constructed using a hierarchical structure, with the top layer defining four major business categories: capital flow domain, contract performance domain, logistics settlement domain, and compliance approval domain. Each major category is further subdivided into core business concepts and their synonyms, hierarchical relationships, and causal relationships; the sliding window slides sequentially along the initial standardized data sequence according to the set window length and step size, and the text fragments in each window are synchronously input to the dependency parser. The parser outputs the dependency relation tree of each word in the fragment, identifying subject-predicate relations, verb-object relations, and modification relations.

[0019] Simultaneously, the word sequence within the window is fuzzily matched with the business concepts in the ontology library. The start and end positions of each business semantic segment are determined using a dependency tree. This positional information, along with the corresponding ontology concept type, is then labeled as semantic boundary and syntactic relation parsing information. Based on this semantic boundary and syntactic relation parsing information, a segmentation operation is performed along the initial standardized data sequence, splitting the continuous sequence into several discrete atomic semantic feature units. Each unit points to a single and complete business semantic. For example, the requirement for the seller to complete the delivery of goods and have it accepted by the buyer before a specified date in a contract performance record is segmented into two independent atomic semantic feature units: delivery obligation and acceptance confirmation. After segmentation, the position index of each atomic semantic feature unit in the original sequence is recorded and arranged in ascending order of position index to obtain a standardized business data sequence.

[0020] In a preferred embodiment of the present invention, step 1 above may further include: Step 1.4: Input each atomic semantic feature unit in the standardized business data sequence into a pre-set multi-layer Transformer architecture context semantic encoder. The multi-head self-attention mechanism inside the encoder calculates the long-range semantic dependency correlation strength between each atomic semantic feature unit, and simultaneously maps each atomic semantic feature unit to a high-dimensional continuous vector space to obtain a high-dimensional embedded feature sequence. Specifically, the pre-set context semantic encoder consists of six identical Transformer layers stacked sequentially. Each layer contains a multi-head self-attention sublayer and a feedforward fully connected sublayer, and residual connections and layer normalization operations are introduced around the two sublayers respectively. Specifically, before entering the encoder, the atomic semantic feature unit is first converted into a fixed-dimensional initial vector representation through an embedding layer. This embedding layer is composed of a pre-trained word embedding matrix and business domain fine-tuning parameters. Based on this, according to the position index of the atomic semantic feature unit in the standardized business data sequence, a position code is superimposed on each initial vector representation. This position code is generated using a phase modulation method based on trigonometric functions. The relative temporal relationship information between units is injected into the model through position modulation signals of different frequencies.

[0021] Furthermore, after the encoded vector sequence enters the first Transformer layer, it is mapped in parallel into three independent representations—a query vector, a key vector, and a value vector—in the multi-head self-attention sub-layer. Each mapping is accomplished through a linear transformation. Within each attention head, the inner product similarity between the query vector and the key vector corresponding to any two units is calculated, and this inner product similarity is divided by a preset scaling factor to stabilize the numerical distribution. The preset scaling factor is pre-set according to the dimension of the key vector, specifically taking the square root of the key vector dimension, to prevent the inner product result from being excessively amplified when the dimension is high, thereby maintaining the smoothness of the attention weight distribution and the stability of gradient propagation.

[0022] The scaled similarity values ​​are then mapped to an attention weight distribution between 0 and 1. This distribution directly measures the long-range semantic dependency of each atomic semantic feature unit on other units. Simultaneously, these attention weights are used to perform a weighted summation of the value vector, calculated using the formula: The output of each unit under this attention head is represented as a weighted sum of all value vectors, as shown in the following formula:

[0023] in Indicates the first The output representation vector of each atomic semantic feature unit under the current attention head, that is, the context semantic vector obtained after the unit is aggregated by attention weighting. Indicates the first The unit for the first Attention weights for each unit, Indicates the first A value vector of units, Indicates all The summation of the summations ...

[0024] Step 1.5: Analyze the sequence topological order among the feature vectors in the high-dimensional embedded feature sequence. Based on the sequence topological order, perform multidimensional tensor dimensionality increase and orthogonal basis space coordinate mapping operations on each feature vector to convert each feature vector into semantic feature nodes with spatial distribution structure, obtaining an initial representation matrix. The semantic feature nodes represent cross-business domain fund flow nodes and compliance status elements. Specifically, this includes: traversing the high-dimensional embedded feature sequence and extracting the original position index of the atomic semantic feature unit corresponding to each feature vector in the standardized business data sequence. These indices naturally preserve the chronological order of business events, thus forming the sequence topological order. Based on this order, locate the direct predecessor and direct successor feature vectors of each feature vector to form directed local topological adjacency pairs, thereby clarifying the natural semantic connection relationship between adjacent business links.

[0025] Furthermore, a multidimensional tensor upsizing and reorganization is performed on each feature vector. Specifically, the original one-dimensional high-dimensional embedding vector is copied and reorganized into a three-dimensional tensor structure. The first dimension maintains the semantic representation dimension of the original vector, the second dimension is used to encode the business attribute bias of the node itself, and the third dimension is used to encode the relative positional offset relationship between the node and its preceding and following neighboring nodes. Specifically, the two newly added dimensions are constructed by mixing the feature vector with its direct predecessor and direct successor vectors and the absolute value of the feature vector's own embedding, so that each three-dimensional tensor simultaneously carries independent semantic information, self-attribute bias information, and local topological association information. On this basis, the orthogonal basis space coordinate mapping stage is entered. A preset orthogonal basis vector group space is retrieved. This space is composed of the main semantic directions extracted from the domain's historical data. The variance between each orthogonal basis is maximized and they are linearly independent of each other, which can express the differentiated business semantics separately along different independent axes.

[0026] Furthermore, the three-dimensional tensor of each semantic feature node is sequentially projected along each axis of the orthogonal basis set. Specifically, the three-dimensional tensor is decomposed in an ordered manner according to the directions of each orthogonal basis vector. The semantic representation dimension information, business attribute bias information, and local topological association information encoded in the tensor are matched with the orthogonal basis vectors in the corresponding directions for similarity. The projection response intensity of the tensor in each orthogonal basis direction is calculated. This response intensity reflects the degree of fit of the comprehensive semantic structure of the semantic feature node in the business semantic direction represented by the corresponding orthogonal basis. The stronger the response, the more the semantic attribute of the node conforms to the business concept feature corresponding to the orthogonal basis. Through the above sequential projection operation, a scalar response value is obtained for each semantic feature node in each orthogonal basis direction. These response values ​​together form the coordinate component vector of the node in the orthogonal basis space, thereby accurately locating the node to a unique coordinate point in the orthogonal basis space.

[0027] Specifically, since the orthogonal basis space has pre-absorbed the semantic separability between core concepts of different business domains, the coordinate component values ​​of nodes with similar semantic features after projection also tend to be close. Therefore, they naturally form a clustered distribution in space. Nodes belonging to different stages of capital flow or compliance status elements will have obvious coordinate differences along specific axes, thus widening the spatial distance. Finally, as all feature vectors complete the dimensionality upgrade and reorganization and orthogonal basis space coordinate mapping in sequence, each atomic semantic feature unit is transformed into a semantic feature node carrying three-dimensional spatial coordinates. Multiple semantic feature nodes, together with their coordinate distribution, constitute the initial representation matrix. The semantic feature nodes in this matrix represent cross-business domain capital flow nodes and compliance status elements. The spatial distance between nodes quantitatively reflects the degree of their semantic correlation, providing a high-fidelity initial feature basis for subsequent calculation of semantic interaction potential gradient.

[0028] Step 1.6: For each semantic feature node in the initial representation matrix, extract its data source credibility identifier, time-series decay coefficient, and prior label of business risk sensitivity in parallel to obtain multi-dimensional prior label information; based on the multi-dimensional prior label information, perform cross-aggregation and weight normalization operations on the multi-dimensional features of each semantic feature node through a preset nonlinear feature aggregation network to obtain normalized aggregation weight coefficients; configure initial attribute weight tensors for each semantic feature node according to the weight coefficients, specifically including: the data source credibility identifier is assigned based on the historical data quality records of the data system from which the node originates; the credibility identifier of the financial system is calculated based on the historical error rate of the audit report; the credibility identifier of the supply chain management platform is assigned based on the historical completeness rate of contract performance records; the credibility identifier of the logistics service provider interface is determined with reference to the historical consistency rate of settlement vouchers; and the credibility identifier of the office approval system is assigned based on the historical compliance rate of the approval process. The value of each credibility identifier is between 0 and 1, with higher values ​​indicating more reliable data.

[0029] Based on this, the time-series decay coefficient is determined by the time interval between the data record generation time and the reference time of the current monitoring period. It is calculated using an exponential decay method, assigning a lower time-series decay coefficient to earlier-generated data to reflect the natural decay of information over time. The calculation formula is as follows:

[0030] In the formula Indicates the first The temporal decay coefficient of each semantic feature node, This represents the preset initial attenuation reference value, which is 1. This represents the preset decay rate constant. Indicates the first Each semantic feature node corresponds to a standardized time interval between the generation time of the data record and the benchmark time of the current monitoring period. Furthermore, the prior label for business risk sensitivity is pre-defined based on the statistical significance of various business indicators in historical default cases, categorizing each business indicator into three levels: high sensitivity, medium sensitivity, and low sensitivity, and encoding them as corresponding prior label values. The values ​​are assigned as follows: high sensitivity is 1, medium sensitivity is 0.5, and low sensitivity is 0.

[0031] Based on this, the three types of label information are aggregated by node to form a multi-dimensional prior label information vector corresponding to each semantic feature node. ,in Indicates the first The data source credibility identifier of each semantic feature node; further, based on the obtained multi-dimensional prior label information vector... By using a pre-defined nonlinear feature aggregation network, multidimensional features of each semantic feature node are cross-aggregated and weighted for normalization, resulting in normalized aggregation weight coefficients. This aggregation network is a three-layer fully connected structure, using the multidimensional prior label information vectors of the nodes. As input, the coupling relationship between the confidence identifier, the temporal decay coefficient, and the prior risk sensitivity label is cross-mined in the hidden layer through non-linear activation operations. The output layer maps the combined result to the interval between 0 and 1 through probability distribution transformation operations to generate normalized aggregate weight coefficients. The calculation formula is:

[0032] in The hidden layer weight matrix is... This is the hidden layer bias vector. This represents the non-linear activation operation of the hidden layer. This is the output layer weight vector. This is the output layer bias scalar. This represents the probability distribution transformation operation of the output layer. Finally, based on the normalized aggregate weight coefficients... An initial attribute weight tensor is configured for each semantic feature node. This tensor serves as the initial importance measure of each node in the subsequent semantic interaction potential energy calculation. Nodes with high weights will play a stronger role in the potential energy gradient calculation.

[0033] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Extract the high-dimensional embedding coordinates of each semantic feature node from the initial representation matrix. Perform tensor inner product operation and modulus normalization on the high-dimensional embedding coordinates and the initial attribute weight tensor corresponding to each semantic feature node to obtain the interaction potential baseline matrix of the semantic association strength between the representation nodes. Specifically, this includes: first traversing all semantic feature nodes in the initial representation matrix, and then performing tensor inner product operation and modulus normalization on the first semantic feature node. For each semantic feature node, its position coordinate components are extracted along the first dimension from its corresponding three-dimensional tensor representation to form the high-dimensional embedded coordinate vector of that node. The dimension of this vector is consistent with the dimension of the high-dimensional continuous vector space; at the same time, the initial attribute weight tensor configured for this node in step 1.6 is retrieved. This tensor preserves the importance measurement structure of nodes after cross-aggregation of multidimensional prior information.

[0034] Based on this, for any two semantic feature nodes and Perform bidirectional tensor inner product operation, that is, first convert the nodes... High-dimensional embedded coordinate vector With nodes Initial attribute weight tensor Element-wise multiplication and summation along each feature dimension yields a scalar interaction value. Then the nodes High-dimensional embedded coordinate vector With nodes Initial attribute weight tensor Performing the same inner product operation yields another scalar cross-value. Furthermore, the arithmetic mean of the two scalar interaction values ​​is taken as the node pair. The original potential energy intensity value After completing the traversal calculations for all node pairs, a symmetric matrix is ​​constructed from all the original potential energy values. This matrix is ​​then normalized globally by taking the square root of the sum of the squares of all elements as the normalization denominator and dividing each element by this denominator until the Frobenius norm of the normalized matrix is ​​in unit value. The resulting symmetric matrix is ​​the interactive potential energy reference matrix. , its first Line 1 Column elements Representation Nodes With nodes The initial semantic association strength between nodes, and the diagonal elements represent the self-association benchmark of nodes.

[0035] Step 2.2: Traverse each semantic feature node in the interaction potential energy benchmark matrix, filter the set of locally associated nodes using a preset cosine similarity threshold as the neighborhood radius, and simultaneously calculate the potential energy intensity difference between the central node and each locally associated node to obtain the node local potential energy difference vector. Specifically, this includes: for the interaction potential energy benchmark matrix... China and Israel Row vector centered on semantic feature nodes The format is: ; Combine it with the remaining row vectors in the matrix Perform cosine similarity calculations sequentially: first calculate the inner product of the two vectors. Then calculate their respective modulus lengths. and Dividing the inner product by the product of the two moduli gives the cosine similarity. When the similarity value exceeds a preset threshold (For example, when the value is 0.7), determine the node. With the central node Highly correlated in potential energy distribution patterns, nodes Selected as the central node Locally related node set Based on this, traverse the set of locally related nodes. Each associated node in From the interaction potential reference matrix Read the central node Potential energy intensity values ​​directly stored with the associated node and with the central node The self-correlated potential energy benchmark value By performing difference calculations, the potential energy intensity deviation value in a single associated direction is obtained. ; will the central node The potential energy intensity deviations between all locally associated nodes are arranged sequentially according to the index order of the associated nodes, forming a one-dimensional vector, which represents the node's value. The corresponding nodal local potential energy difference vector Its dimension is equal to The number of nodes in the central node, each component quantitatively describes the potential energy difference between the central node and its specific neighbors.

[0036] Step 2.3: Perform element-wise projection mapping between the local potential energy difference vector of each node and the feature space offset vector of each locally associated node relative to the central node to obtain the projected potential energy difference field; based on the projected potential energy difference field, perform finite difference approximation calculation along each orthogonal dimension of the high-dimensional feature space to obtain the multidimensional potential energy change rate tensor; based on the multidimensional potential energy change rate tensor, extract the potential energy change rate components of each semantic feature node in the multidimensional space, specifically including: for the central node... and its locally associated node set The first in For each associated node, calculate the feature space offset vector of that associated node relative to the central node. Each component of this vector represents the position difference between the two nodes on each coordinate axis in high-dimensional space; based on this, the local potential energy difference vector of the nodes is... The middle corresponds to the node Deviation component offset vector in feature space Perform element-wise projection mapping, that is, scalar Multiplying this by the entire offset vector as a scaling factor yields the projected potential difference vector carrying directional information. .

[0037] The projection potential difference vectors corresponding to all locally related nodes are superimposed and synthesized position by position in the feature space to form a vector covering the central node. Projected potential energy difference field in local neighborhood This field records the distribution of potential energy variation with spatial location within the region surrounding the central node; further, along the first feature space... In the projected potential energy difference field, along orthogonal dimensional directions. Take the center node A tiny step along the positive direction of that dimension The difference between the potential energy value at the current position and the potential energy value at the current position, divided by the step size. Perform a forward finite difference approximation calculation to obtain the center node. In dimensions scalar of the rate of change of potential energy on ,in For the first A unit direction vector in one dimension; indexing the scalar of the rate of change of potential energy in each orthogonal dimension by dimension. Arranged in ascending order, they form the central node. The corresponding multidimensional potential energy rate of change tensor The first tensor Each component is... Based on this, the central node is extracted according to the component distribution of the tensor in each dimension. The potential energy change rate component sequence in each direction of the orthogonal space serves as the multidimensional potential energy change feature of the central node, which is then used for subsequent gradient synthesis.

[0038] Step 2.4 involves vector synthesis and gradient direction normalization of the potential energy change rate component along a preset feature principal axis to obtain a semantic interaction potential energy gradient vector field pointing to a semantically associated dense region, which serves as the semantic interaction potential energy gradient. Specifically, this includes retrieving a preset set of feature principal axis directions. These principal axis directions are determined by performing principal component analysis on the high-dimensional coordinates of all semantic feature nodes in historical business data, and selecting the first set with the largest variance. A number of mutually orthogonal direction vectors As feature principal axes, these principal axes represent the most information-rich and independent spectral directions in the business semantic space; based on this, for the th For each semantic feature node, its potential energy change rate tensor in each orthogonal dimension is... Treated as a complete multidimensional vector, it is sequentially coupled with each preset principal axis direction. Perform inner product projection to obtain the node at the [missing information - likely a specific location or position]. scalar contribution of potential energy changes along each principal axis direction The potential energy gradient vector is calculated by weighting the contribution values ​​along all principal axes according to their corresponding principal axis vectors. This vector integrates potential energy changes along all principal axes, and its final direction is the semantic clustering region where the potential energy growth is steepest, with a magnitude of [missing information]. It is directly proportional to the intensity of the instantaneous change in potential energy in that direction.

[0039] Furthermore, the potential energy gradient vectors at each node are normalized in the magnitude direction: that is, each gradient vector is retained. The original spatial orientation remains unchanged, and its magnitude is scaled proportionally to the maximum observed value of the gradient magnitudes of all nodes, so that the magnitude of the normalized gradient vector falls within the range of the nodes. Within the interval, this eliminates the incomparability caused by differences in the potential energy background of different nodes in their local areas. After the above vector synthesis and normalization processes, the potential energy gradient vectors of all semantic feature nodes together constitute a continuous gradient vector field distribution defined in the entire feature space. This vector field is the semantic interaction potential energy gradient vector field. It fully characterizes the optimal movement direction and relative driving force intensity of each node at its current position towards the semantically densely related region.

[0040] In a preferred embodiment of the present invention, step 2 above may further include: Step 2.5: Analyze the semantic interaction potential energy gradient vector field to obtain the potential energy gradient space analysis information; based on the potential energy gradient space analysis information, construct directed connection edges between the semantic feature nodes, and assign initial connection weights to each directed connection edge to obtain a dynamic semantic association topology graph, specifically including: traversing the semantic interaction potential energy gradient vector field. Each semantic feature node in Extract its gradient vector The orientation angle parameter and magnitude value in space; this orientation angle is determined by... The modulus value is obtained by converting the angles with each coordinate axis, and is then a normalized scalar. Both together constitute a node. The potential gradient space analytical information; based on this, for any two different semantic feature nodes and Calculate from node High-dimensional embedding coordinates Pointing to node coordinates The unit vector of the line connecting the directions Next, calculate the nodes. normalized gradient direction unit vector Unit vector of the connecting line direction The angle between If the included angle Less than the preset gradient direction alignment threshold (like If the node is determined, then the node is determined. The gradient direction points significantly towards the node. In the semantic region where the two nodes are located, a child node is established between them. Pointing to node The directed connection edge; further, an initial connection weight is assigned to the directed connection edge. The weight value is determined by the node. gradient magnitude Combined with the cosine of the included angle, the calculation formula is:

[0041] in Reflecting nodes The driving force converging towards the surrounding area, The consistency between the gradient direction and the actual connection direction is measured. Multiplying the two results in higher connection weights for node pairs whose gradient directions are more aligned and whose magnitudes are larger. By performing the above operation on all node pairs that meet the conditions, a dynamic semantic association topology graph is obtained, which consists of all semantic feature nodes as vertex sets, directed edge sets, and edge weight sets.

[0042] Step 2.6: Perform a bonding convergence determination operation on each node pair in the dynamic semantic association topology graph to obtain a bonding convergence determination signal; based on the bonding convergence determination signal, divide the node pairs into a set of strongly associated node pairs and a set of weakly associated node pairs. Specifically, this includes: for each pair of node pairs connected by directed edges in the dynamic semantic association topology graph... First, extract the high-dimensional embedding coordinate vectors of each of the two nodes. and Simultaneously, based on the time window and business domain type covered by the current business monitoring cycle, the corresponding benchmark anchor vector is retrieved from the pre-set global semantic benchmark anchor set. The anchor vector represents the reference position of the neutral semantic state in a given monitoring scenario; in the multi-dimensional feature orthogonal projection space, these three points are used as vertices to construct a feature geometric triangle. .

[0043] Based on this, calculate the vectors of the two sides of the triangle: and For vectors and Perform the cross product operation and take the modulus to obtain the area of ​​the parallelogram with these two sides as adjacent sides. The area of ​​the triangle is ; This area index quantitatively reflects the degree of deviation between a node pair and the global semantic benchmark: the smaller the area, the flatter the triangle formed by the two nodes and the benchmark point, meaning that the semantic states of the two nodes tend to converge and couple under the constraint of the benchmark anchor point; further, the area index... With preset bonding convergence threshold Compare; if A positive bonding convergence determination signal is generated, indicating that the node pairs are semantically tightly coupled, and they are classified into the set of strongly associated node pairs. ;like A negative bonding convergence determination signal is generated, indicating that the semantic relationship between node pairs is loose, and the node pairs are classified into the set of weakly associated node pairs. .

[0044] Step 2.7: Perform feature fusion mapping on the set of strongly correlated node pairs to obtain fused semantic representation nodes; perform feature rejection masking on the set of weakly correlated node pairs to obtain a masked isolation state graph structure, specifically including: for the set of strongly correlated node pairs... Each pair of nodes in Extract their high-dimensional embedding coordinate vectors respectively. and The first and last parts are concatenated along the feature channel dimension to form a joint feature vector with a length equal to the sum of the two vector dimensions. Based on this, the joint feature vector is input into a pre-defined cross-modal attention-weighted projection module. This module first maps the joint vector into three intermediate representations: query, key, and value. It then calculates the relative importance weight distribution among the feature channels through self-attention operations. Furthermore, it uses these attention weights to process and reorganize the channels of the joint vector, compressing and mapping the high-dimensional joint vector to a shared hidden space with the same dimension as the original node vectors. Finally, it outputs a fused semantic representation node coordinate vector. .

[0045] This fusion node not only resides in a reasonable region of the hidden space, but also simultaneously encodes the business semantic attributes of the original node for both parties, becoming a unified representative of both; for the set of weakly related node pairs... each directed connection edge in Construct a square dynamic sparse mask matrix with the same number of nodes as the current graph. ,initialization All elements are 1, and the matrix positions corresponding to all weakly related edges are further traversed. ,Will The value is assigned to 0, while the corresponding positions of the other strongly correlated edges remain 1; the mask matrix is ​​then... Element-wise multiplication with the adjacency weight matrix of the current graph structure is performed, so that the connection weights corresponding to weakly associated edges are set to 0, and the feature propagation paths between the nodes at both ends of these edges are cut off and blocked. After the above operation, strongly associated node pairs are compressed and integrated into new fused nodes, and weakly associated edges are removed and isolated, thus obtaining a masked isolated state graph structure composed of the remaining active nodes and the blocked edges.

[0046] Step 2.8 involves performing topological fusion and graph structure reorganization on the fused semantic representation nodes and the masked isolation state graph structure to update node adjacency relationships and edge weight distribution, thereby obtaining the semantic topology graph. Specifically, this includes: inserting all the fused semantic representation nodes as new vertices into the node set of the masked isolation state graph structure sequentially; and simultaneously establishing a pair between each fused node and the original node it replaces. A mapping table is generated between nodes to record the source composition of the fused node. Based on this, all nodes in the original graph are traversed. or node There are other nodes connected by directed edges. For each such edge, if the original direction is or Then, a path will be created in the merged graph from... Pointing to the fusion node Directed edges; if the original direction is or Then create a merge node point to The directed edge.

[0047] The weights of new edges inherit the maximum or average weights of the original edges, ensuring that the fused nodes fully inherit all the connections of the original node pairs in the topology. For weakly related edges that are isolated by masks in the original graph, they are completely removed from the edge set, and the endpoint nodes of each deleted edge are checked: if a node loses all incoming and outgoing edges due to the removal of all weakly related edges, it is marked as an isolated node and temporarily moved from the active node set to the archive list, no longer participating in subsequent information propagation. Furthermore, the symmetry of the updated adjacency matrix is ​​checked. If it is found that the weighted out-degree and in-degree are unbalanced due to the introduction of fused nodes, a local weight reallocation fine-tuning is performed to restore the global connectivity consistency of the graph structure. Finally, after node fusion insertion, edge redirection and inheritance, weakly related edge deletion, isolated node archiving, and weight rebalancing adjustment, the generated graph is the semantic topology graph. This semantic topology graph retains a concise expression of strong risk coupling characteristics with fused semantic representation nodes as the core, while minimizing the pollution of key risk transmission paths by business noise through mask isolation.

[0048] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Perform temporal slicing on the semantic topology graph according to a preset time step to obtain a continuously sampled temporal graph frame sequence. Superimpose temporal position encoding on each semantic feature node in the continuously sampled temporal graph frame sequence to obtain a temporally aligned graph sequence. Specifically, this includes: pre-setting a unified time step according to the business monitoring cycle and risk warning timeliness requirements, setting the step to one transaction day or one financial settlement cycle to ensure that the slicing granularity matches the fluctuation frequency of the business risk signal; taking the start time of the monitoring cycle as the reference starting point, uniformly dividing the entire monitoring cycle into multiple consecutive and non-overlapping time windows according to the step, with each time window corresponding to an independent sampling time; for each time window, traversing all semantic feature nodes in the semantic topology graph, detecting whether the business status attribute of each node is triggered or recorded within the window, and extracting the triggered node and its associated directed connection edge; the nodes in the graph frame snapshot retain their static high-dimensional embedding coordinates in the semantic topology graph as initial semantic representations, and the coordinates themselves do not change due to window movement, forming a graph frame snapshot bound to the time window.

[0049] The snapshot of the graph fully preserves the node attribute structure, directed edge connections, and edge weight distribution values ​​in the semantic topology graph, but only reflects the instantaneous topological state within that window, without including historical evolution information across windows. All the snapshots corresponding to all time windows are arranged sequentially to form a continuously sampled temporal graph frame sequence, with adjacent frames exhibiting a dynamic relationship of gradual evolution in node states and edge weights. Based on this, temporal position encoding is superimposed on each semantic feature node within each frame of the temporal graph frame sequence. This encoding is generated using a periodic modulation signal associated with the time window index; different time window index values ​​correspond to modulation signals of different frequencies and phases. The combination of these features allows for a smooth transition between the encoding patterns of adjacent windows, while ensuring that each window possesses globally unique periodic signal features throughout the entire sequence, thus globally and uniquely marking the absolute position of each node on the time axis. The temporal position encoding is then fused with the original high-dimensional embedded coordinates of the nodes element-wise along the feature dimension, tightly integrating the time index features with semantic representation information to obtain an enhanced node representation carrying temporal positioning information. After the above processing, the nodes in each frame of the graph snapshot are simultaneously injected with spatial semantic features and time index features, thereby forming a temporally aligned graph sequence. This sequence provides a spatiotemporally unified graph structure data input for subsequent temporal evolution analysis.

[0050] Step 3.2: Input the temporal alignment graph sequence into the graph message passing layer of the preset temporal state evolution neural network, perform forward propagation of node states and cross-frame feature aggregation along the time axis to obtain aggregated node state features; calculate the representation space coordinate offset of each semantic feature node between adjacent sampling windows based on the aggregated node state features to obtain the semantic state displacement vector. Specifically, this includes: inputting the temporal alignment graph sequence into the preset temporal state evolution neural network, which is composed of a front graph message passing layer, a middle temporal attention gating module, and a rear cross-frame aggregation buffer, specifically used to capture the dynamic evolution of graph structure node states within continuous time windows and output the temporal displacement representation of node semantic states.

[0051] Specifically, the temporally aligned graph sequence is fed frame by frame into the graph message passing layer of the neural network according to the order of the time window. The graph message passing layer performs information passing operations between nodes in each frame of the graph. For each semantic feature node in each frame of the graph, the graph message passing layer first locates all its neighbor nodes in the direction of all incoming edges, receives the state vector of each neighbor node at the previous time step, and concatenates all the received neighbor state vectors with the current node's own temporally enhanced input vector. The concatenated joint information is compressed and fused through a nonlinear transformation unit to generate the updated state vector of the current node in this time window, completing the forward propagation of the state. After completing the state update of all nodes in the current frame, the network transmits the complete state vector matrix of the frame as the initial state input for the next frame. The next frame performs neighbor information aggregation and state update again based on this, so that the state information flows forward frame by frame along the time axis.

[0052] Meanwhile, the graph message passing layer maintains a cross-frame feature aggregation cache, which performs aggregation operations on the state vector sequences of the same node across multiple consecutive time windows. A time-weighted average is used: windows closer to the current time are assigned higher aggregation weights, while windows further away are assigned lower weights. This extracts the smooth evolution trend of the node's state along the time axis, while preserving the local fluctuations of each window's state relative to the trend. After aggregation, the aggregated node state feature is obtained. The aggregated node state feature integrates the node's latest state information at the current sampling time and its evolutionary context information within the recent time range. Based on this, for each semantic feature node, the aggregated node state feature corresponding to the current sampling window and the aggregated node state feature corresponding to the previous adjacent sampling window are retrieved. Both of these aggregated node state features are vectors in a high-dimensional representation space, with each dimension of the vector corresponding to an independent semantic feature channel.

[0053] Specifically, the process of dimension-by-dimensional difference operation is as follows: along each dimension of the vector, the value of the current sampling window in that dimension is subtracted from the value of the previous adjacent sampling window in the same dimension to obtain a difference with a positive or negative sign. The sign of the difference indicates the direction of increase or decrease of the feature value in that dimension, and the absolute value of the difference indicates the magnitude of change of the feature value in that dimension. The differences in all dimensions are arranged in the original dimensional order and combined into a new vector with the same dimension as the original aggregate node state feature dimension. This new vector is the representation space coordinate difference vector, which is the representation space coordinate offset of the node between these two adjacent windows. The coordinate offsets of all semantic feature nodes between every two adjacent windows are systematically organized according to the node identifier and window order to form a set of semantic state displacement vectors. Each vector in this set quantitatively describes the migration direction and migration magnitude of a node's semantic state within a unit time step, reflecting the temporal movement trajectory of the node in the risk semantic space.

[0054] Step 3.3: Based on the semantic state displacement vector, the temporal attention gating module inside the preset temporal state evolution neural network is invoked to calculate the instantaneous flux of risk signal flow along the graph topology edges between each semantic feature node; the instantaneous flux of risk signal flow is subjected to multi-dimensional feature decoupling and tensor projection along the graph topology edges to obtain the risk propagation rate tensor. Specifically, this includes: inputting the semantic state displacement vector into the temporal attention gating module in its entirety according to the correspondence between nodes and time windows. This module first traverses each directed connection edge in the semantic topology graph and extracts the semantic state displacement vectors of the source node and target node connected by the edge within their respective time windows; for each directed edge, the directional consistency of the source node displacement vector and the target node displacement vector in the representation space is calculated: that is, the angle between the two displacement vectors is evaluated. The smaller the angle, the more similar the semantic state change directions of the two nodes are within the window. At the same time, the relative difference in the magnitude of the two displacement vectors is calculated to measure the matching degree between the source node and the target node in terms of temporal state migration rate.

[0055] Based on directional consistency and rate matching information, the attention gating mechanism inside the module comprehensively judges the effective strength of risk signal transmission from the source node to the target node along the edge within the time window through a nonlinear gating unit. If the direction is consistent and the rate matching is reasonable, the gating unit outputs a higher transmission strength value, and vice versa, thereby generating an instantaneous flux scalar that characterizes the activity level of risk flow on the edge at the current moment. Through the above calculation, the set of instantaneous flux scalars of risk signal flow of all directed edges within each time window is obtained. On this basis, the multi-dimensional feature decoupling and tensor projection stage is entered. This stage pre-sets a risk dimension set composed of multiple independent risk analysis dimensions, including but not limited to the dimensions of capital liquidity, compliance deviation, supply chain stability, and credit default transmission. Each dimension corresponds to a set of decoupling basis vectors extracted in advance from historical risk event data.

[0056] Specifically, for each directed edge, the instantaneous flux scalar within a specific time window is projected sequentially onto the decoupling basis vectors of each risk dimension. This involves decomposing the instantaneous flux scalar along each decoupling direction to obtain the flux components of the edge in each risk dimension within that window. Based on this, the flux components of all directed edges in each time window and each risk dimension are systematically organized according to the graph's topology. The time window index is used as the first dimension, the source node as the second dimension, the target node as the third dimension, and the risk dimension as the fourth dimension, forming a four-dimensional risk propagation rate tensor. This tensor fully records the instantaneous rate of the risk signal propagating along each edge of the graph topology at each sampling moment and its fine-grained distribution information at different risk analysis levels, providing a structured dynamic data basis for risk propagation for subsequent high-order time series analysis.

[0057] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1 involves performing a second-order temporal difference operation on the risk propagation rate tensor within a continuous sampling window sequence to extract the abrupt gradient and trend slope of the risk signal propagation rate, thereby obtaining the risk evolution acceleration component. Specifically, this includes denoting the risk propagation rate tensor as... Its internal elements are It means that, among them The position of the indexed continuous sampling window on the time axis and These are the identifiers for the source semantic feature node and the target semantic feature node, respectively. (Note: the characters in steps 2 and 4...) p and q The two symbols have completely identical meanings, both representing the general index identifier of semantic feature nodes in the semantic topology graph. In this embodiment, p and qThe reference to "is always consistent with the node index". Index risk dimension; for each fixed pair of directed edges and each fixed risk dimension Extract the rate value sequence of the edge along this dimension along the continuous sampling window sequence. .

[0058] First, calculate the first-order difference between adjacent windows to obtain the first-order difference sequence. The first-order difference reflects the net change in the risk propagation rate within a unit time step. Based on this, the difference between adjacent terms is calculated again for the first-order difference sequence to obtain the second-order difference sequence. This second-order difference quantitatively characterizes the acceleration feature of the risk propagation rate over time, and is defined as the risk evolution acceleration component and denoted as... ;when When the value is positive, it indicates that the spread rate of the risk signal along the corresponding propagation path is accelerating, showing a positive surge trend; when the value is negative, it indicates that the risk propagation rate is slowing down, showing a negative decay trend; this acceleration component directly provides information on the direction and intensity of the risk situation's evolution.

[0059] Step 4.2: Input the risk evolution acceleration component into a preset graph weight adaptive adjustment network. Calculate the dynamic adjustment coefficients of each connection edge using the nonlinear activation function within the preset graph weight adaptive adjustment network. Based on these dynamic adjustment coefficients, perform connection strength enhancement mapping on connection edges exhibiting a positive surge trend in acceleration components, and perform connection strength suppression mapping on connection edges exhibiting a negative decay trend in acceleration components, to obtain the node connection weight update matrix. Specifically, this includes: first, processing the risk evolution acceleration component... Dimensional aggregation is performed by edge, that is, for each directed edge The weighted summation of the acceleration components across all risk dimensions yields the overall acceleration intensity of that edge. The formula for weighted summation can be expressed as follows:

[0060] in, This represents the total number of risk dimensions. For the risk dimension, the index number is... For the first The weighting coefficients for each risk dimension are determined based on the sensitivity of each risk dimension in business monitoring and its historical early warning contribution rate, and ensure that the sum of the weights of all dimensions is 1; on this basis, the overall acceleration intensity will be... Input a pre-defined graph weight adaptive adjustment network. This network contains a non-linear activation function layer, using the hyperbolic tangent function as the activation function. This function is used to adjust the weights of the graph. Perform nonlinear mapping and output dynamic adjustment coefficients. The calculation formula is as follows:

[0061] In the formula, This represents the hyperbolic tangent nonlinear activation function. This is a preset positive scaling factor used to adjust the sensitivity of acceleration intensity to the output dynamic adjustment coefficient; The base of the natural logarithm; dynamic adjustment coefficient. The range of values ​​is limited to Between them, the positive and negative signs are determined by the combined acceleration intensity. The sign determines the value; a positive value corresponds to a positive surge trend, and a negative value corresponds to a negative decay trend. Furthermore, the current connection weight of each edge is adaptively adjusted based on a dynamic adjustment coefficient: if... Then perform connection strength enhancement mapping, setting the adjusted new weights to... ,in The initial connection weights assigned in step 2.5; if Then, the connection strength suppression mapping is performed according to the above formula, since... If the value is negative, the new weight will be scaled down proportionally relative to the initial weight. After performing the above adjustment on all directed edges, all updated edge weights are aggregated according to their original graph topology to form the node connection weight update matrix. Its elements are .

[0062] Step 4.3 involves element-wise weighted fusion of the node connection weight update matrix and the current adjacency matrix of the semantic topology graph, synchronously updating the edge weight distribution to filter out redundant connection edges below a preset dynamic sparsity threshold, thus obtaining a recalibrated semantic topology graph. Specifically, this includes retrieving the current adjacency matrix corresponding to the semantic topology graph finally generated in step 2.8. Its elements are the connection weights after fusion and recombination. Update the node connection weight matrix. Adjacency matrix Element-wise weighted fusion is performed, and a preset smoothing coefficient is used in the fusion process. (Values ​​between 0 and 1) Perform a convex combination of corresponding elements of the two matrices to obtain the recalibrated adjacency matrix. Its elements are represented in the following form:

[0063] in, Update the node connection weight matrix The middle corresponds to a directed edge In the time window t The new weights after internal acceleration adjustment For the edges in the current adjacency matrix of the semantic topological graph Weight values; smoothing coefficient The closer a value is to 1, the more trust is placed in the immediate weight updates obtained from the acceleration feedback; the closer a value is to 0, the more reliance is placed on the current weight state of the semantic topology graph to maintain model stability. After fusion, the recalibrated adjacency matrix is ​​traversed. All elements in the given set of directed edges Corresponding weight value With the preset dynamic sparsity threshold Compare; if This indicates that the effective transmission strength of the connecting edge under the current risk situation has decayed to a negligible level, and it is regarded as a redundant connecting edge and removed from the graph structure. If the weight is greater than or equal to the threshold, it is retained, and its edge weight is the new weight after fusion. After the above fusion and pruning operations, the final recalibrated semantic topology graph is obtained. The edge weight distribution of this graph not only reflects the original semantic association strength, but also dynamically strengthens the transmission ability of risk mutation path and suppresses the interference of decay path through acceleration feedback.

[0064] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Input the recalibrated semantic topology graph into the first-layer input interface of a preset multi-layer feedforward risk recognition neural network. Obtain the global risk semantic feature vector through global feature aggregation. Specifically, the preset multi-layer feedforward risk recognition neural network consists of a first-layer input interface, multiple consecutively stacked fully connected hidden layers, and a probability output layer. It is specifically used to perform end-to-end mapping from graph structure to risk category on the input recalibrated semantic topology graph and output risk recognition results with confidence level grading. Specifically, the recalibrated semantic topology graph is used as input data and fed into the network through the first-layer input interface of the neural network. The first-layer input interface first traverses all active semantic feature nodes in the recalibrated semantic topology graph, and then... For each node, extract its current high-dimensional embedding coordinate vector and the principal components of its initial attribute weight tensor, and fuse them to form the semantic feature vector of that node. .

[0065] Furthermore, a global attention convergence mechanism is employed to aggregate information from all nodes in the graph, introducing a trainable global query vector. ,Will With the semantic feature vector of each node Perform inner product interaction matching to obtain nodes. The original association scores of all nodes are then mapped to normalized global contribution weights, denoted as [the weights are then calculated]. The calculation process involves dividing the nodes The ratio of the score index value to the sum of the score index values ​​of all nodes is used to calculate the semantic feature vectors of all nodes. According to their respective global contribution weights We perform a weighted summation to obtain the global risk semantic feature vector. The calculation formula is as follows:

[0066] in Represents the global risk semantic feature vector. To recalibrate the total number of active semantic feature nodes in the semantic topology graph. This represents summing over all active nodes. For the first The global contribution weights corresponding to each node; this vector integrates the risk semantic information of all active nodes in the recalibrated semantic topology graph, filters out the interference of noisy nodes that have been isolated by the mask, and provides a compact and information-dense global input representation for the deep processing of the subsequent hidden layers.

[0067] Step 5.2: Input the global risk semantic feature vector into the continuous hidden layer of the preset multi-layer feedforward risk recognition neural network. The continuous hidden layer performs nonlinear potential energy convergence operation on the risk semantic features to obtain the converged discriminative feature tensor. Specifically, this includes: inputting the output global risk semantic feature vector into the first hidden layer of the multi-layer feedforward risk recognition neural network. This continuous hidden layer is composed of multiple fully connected layers stacked sequentially. Each fully connected layer is followed by a nonlinear activation operation and a residual connection path. In each hidden layer, the input vector is first mapped to a higher-dimensional intermediate representation space through a linear transformation. Then, a feature curvature change is introduced through a nonlinear activation operation, so that different categories of risk semantics are gradually pushed away from the overlapping region along different potential energy gradient directions in this high-dimensional space.

[0068] Meanwhile, the residual connection directly transmits the input of the current layer to the output of that layer and adds it to the nonlinear transformation result, ensuring that the information does not decay as the number of layers increases during forward propagation. After layer-by-layer mapping and nonlinear transformation through multiple hidden layers, the risk semantic features are repeatedly squeezed and separated in the high-dimensional potential field formed by the hidden layers. Semantic nodes with similar risk states are guided to a near-stable position by the potential gradient, while semantic nodes with different risk states are pulled towards their respective potential minimum regions in the feature space. When the output change of the subsequent hidden layer is lower than the preset convergence threshold, it indicates that the feature has reached a stable convergence state. The output of the hidden layer at this time is the converged discriminative feature tensor. This tensor is organized in the form of a multi-dimensional array, which completely preserves the spatial distribution of each semantic feature after convergence, and there are clear separable boundaries between semantic clusters of different risk categories.

[0069] Step 5.3 involves dimensionality compression and feature recombination of the convergent discriminative feature tensor to obtain a high-order discriminative feature sequence. Specifically, this includes: transmitting the output convergent discriminative feature tensor to the dimensionality compression layer; firstly, globally averaging the feature sub-tensors corresponding to each node in the convergent discriminative feature tensor along the feature channel dimension; then, averaging the values ​​of all channels in each sub-tensor to obtain a scalar value as the representative response intensity of that channel; further, using principal component selection, sorting the compressed channels from largest to smallest variance contribution, retaining the principal channel components with the largest variance contribution, and discarding secondary channels with variance contributions below a preset cumulative contribution rate threshold, thus achieving dimensionality reduction and compression of the feature space. The compressed channel components are rearranged and combined according to the original spatial distribution order of semantic nodes, expanding the multidimensional tensor structure into a one-dimensional feature sequence. Each element in this sequence corresponds to a high-discriminative, high-order semantic feature. Based on this, the one-dimensional sequence is reorganized, and the feature components belonging to the same semantic cluster are locally aggregated according to their adjacency relationship, while maintaining the order boundaries between different semantic clusters to preserve the structural information at the category level. After the above dimensional compression and feature reorganization processes, the resulting high-order discriminative feature sequence not only significantly reduces the data size to adapt to the input requirements of the output layer, but also retains the core discriminative information used to distinguish different risk categories to the greatest extent.

[0070] Step 5.4: Transmit the high-order discriminative feature sequence to the output layer of the preset multi-layer feedforward risk identification neural network, calculate the probability distribution of each preset risk category, and obtain the initial probability distribution. Specifically, the output layer of the preset multi-layer feedforward risk identification neural network is a fully connected classification layer with the same number of neurons as the total number of preset risk categories. The preset risk categories include at least normal business status, liquidity risk, supply chain default risk, compliance risk, and financial fraud risk. Inside this output layer, a trainable weight matrix is ​​maintained, with each row of the matrix corresponding to a risk category and storing the discriminative weight parameters of that category. At the same time, a bias vector is maintained, with each component corresponding to the discriminative bias of a category.

[0071] Specifically, the input high-order discriminative feature sequence is expanded into a one-dimensional feature vector. This feature vector is then linearly transformed using a weight matrix, which involves multiplying each component of the feature vector element-wise with the corresponding category weight parameter in the weight matrix and summing the results. This sum is then added to the bias value corresponding to that category to obtain the original score value for each preset risk category. Next, the original scores for each category are subjected to exponential normalization. Specifically, each original score value is calculated using an exponent with the natural constant as its base, yielding an exponential value for each category. The exponential value for a single category is then divided by the sum of the exponential values ​​for all categories to obtain the mapped probability value for that category. The mapped probability values ​​for all categories together constitute an initial probability distribution vector. Each element in this vector has a value between 0 and 1, and the sum of all elements is 1, directly representing the probability estimate that the current enterprise's operating data belongs to the corresponding preset risk category. This initial probability distribution provides a complete probabilistic basis for subsequent peak retrieval and confidence level calibration.

[0072] Step 5.5 involves performing peak retrieval and confidence threshold mapping operations on the initial probability distribution to extract the target risk category identifier corresponding to the maximum probability value. The target risk category identifier is then graded according to a pre-set confidence grading interval to obtain the semantic risk identification result of the enterprise operating data and the corresponding risk confidence level distribution. Specifically, this includes: finding the probability value with the largest value in the obtained initial probability distribution vector through element-by-element comparison, locating the index of this maximum value in the vector, and extracting the pre-set risk category corresponding to this index as the target risk category identifier; simultaneously, using this maximum probability value as the first confidence reference value of the identification result, and further, calling the pre-set confidence grading interval. The classification interval divides the confidence level range into at least three levels. For example, the high confidence interval requires a probability value above 0.85, the medium confidence interval requires a probability value between 0.65 and 0.85, and the low confidence interval requires a probability value below 0.65. The maximum probability value is compared with the threshold of each level interval to determine the confidence level interval it falls into, and the level identifier corresponding to the interval is assigned to the target risk category identifier. Finally, the semantic risk identification result of enterprise operating data containing the target risk category identifier and its corresponding confidence level is output. At the same time, the probability values ​​of other risk categories in the initial probability distribution and their corresponding confidence levels can be selectively output to form a complete risk confidence level distribution.

[0073] like Figure 2 As shown, embodiments of the present invention also provide a semantic risk identification system for enterprise operational data based on neural networks, including: The data acquisition module is used to acquire multi-source heterogeneous operational data streams of the target enterprise within the monitoring period, including financial audit reports, supply chain contract performance records, logistics settlement vouchers, and internal compliance approval and processing minutes. The multi-source heterogeneous operational data streams are sequentially subjected to cross-modal cleaning and alignment and semantic unit segmentation to obtain a standardized operational data sequence. The semantic embedding module is used to map standardized business data sequences to a high-dimensional continuous vector space to obtain an initial representation matrix, and to configure an initial attribute weight tensor for each semantic feature node in the initial representation matrix. The potential energy gradient and topology construction module is used to calculate the semantic interaction potential energy gradient between each semantic feature node based on the initial representation matrix and the initial attribute weight tensor, construct a dynamic semantic association topology graph based on the semantic interaction potential energy gradient, perform feature fusion mapping on node pairs that meet the preset bonding convergence threshold, and perform feature rejection masking on node pairs that do not meet the preset bonding convergence threshold to obtain a semantic topology structure graph. The temporal evolution module is used to input the semantic topology graph into a preset temporal state evolution neural network and extract the risk propagation rate tensor of each semantic feature node within a continuous sampling window; The acceleration and recalibration module is used to perform temporal difference operations on the risk propagation rate tensor to obtain the risk evolution acceleration component, and to perform dynamic recalibration on the node connection weights in the semantic topology graph based on the risk evolution acceleration component to obtain a recalibrated semantic topology graph. The risk identification output module is used to input the recalibrated semantic topology diagram into a preset multi-layer feedforward risk identification neural network. It adopts the nonlinear potential energy convergence method of risk semantic features to obtain the semantic risk identification results of enterprise operating data and the corresponding risk confidence level distribution.

[0074] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0075] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying semantic risks in enterprise operational data based on neural networks, characterized in that, The method includes: The target enterprise acquires multi-source heterogeneous operational data streams within the monitoring period. Cross-modal cleaning and alignment and semantic unit segmentation are performed sequentially on the multi-source heterogeneous operational data streams to obtain standardized operational data sequences. The standardized operational data sequences are mapped to a high-dimensional continuous vector space to obtain an initial representation matrix. Initial attribute weight tensors are configured for each semantic feature node in the initial representation matrix. High-dimensional embedding coordinates of each semantic feature node are extracted from the initial representation matrix. The high-dimensional embedding coordinates are then multiplied by the initial attribute weight tensor corresponding to each semantic feature node, and their magnitudes are normalized to obtain the interaction potential benchmark matrix representing the semantic association strength between nodes. Each semantic feature node in the interaction potential benchmark matrix is ​​traversed, and a set of locally associated nodes is selected using a preset cosine similarity threshold as the neighborhood radius. The set of locally associated nodes is then traversed. Each associated node in From the interaction potential reference matrix Read the central node Potential energy intensity values ​​stored directly between the associated nodes and with the central node The self-correlated potential energy benchmark value The difference is calculated to obtain the potential energy intensity deviation value in a single associated direction; the local potential energy difference vector of the node and the feature space offset vector of each local associated node relative to the central node are projected element-wise to obtain the projected potential energy difference field; based on the projected potential energy difference field, finite difference approximation is performed along each orthogonal dimension of the high-dimensional feature space to obtain the multidimensional potential energy change rate tensor; based on the multidimensional potential energy change rate tensor, the potential energy change rate components of each semantic feature node in the multidimensional space are extracted; the potential energy change rate components are vector synthesized and gradient direction normalized along the preset feature principal axis to obtain the semantic interaction potential energy gradient vector field pointing to the semantically associated dense region, which is used as the semantic interaction potential energy gradient; a dynamic semantic association topology graph is constructed based on the semantic interaction potential energy gradient, and feature fusion mapping is performed on the node pairs that meet the preset bonding convergence threshold, and feature rejection masking is performed on the node pairs that do not meet the preset bonding convergence threshold to obtain the semantic topology structure graph. The semantic topology graph is input into a preset temporal state evolution neural network to extract the risk propagation rate tensor of each semantic feature node within a continuous sampling window. The temporal state evolution neural network is composed of a front-end graph message passing layer, a middle temporal attention gating module, and a rear-end cross-frame aggregation buffer. It is used to capture the dynamic evolution of the graph structure node state within a continuous time window and output the temporal displacement representation of the node semantic state. The risk propagation rate tensor is subjected to temporal difference operation to obtain the risk evolution acceleration component. Based on the risk evolution acceleration component, the node connection weights in the semantic topology graph are dynamically recalibrated to obtain the recalibrated semantic topology graph. The recalibrated semantic topology diagram is input into a preset multi-layer feedforward risk identification neural network. The nonlinear potential energy convergence method of risk semantic features is used to obtain the semantic risk identification results of enterprise operating data and the corresponding risk confidence level distribution.

2. The method for identifying semantic risks in enterprise operational data based on neural networks according to claim 1, characterized in that, Acquire multi-source heterogeneous operational data streams of the target enterprise within the monitoring period, and sequentially perform cross-modal cleaning and alignment and semantic unit segmentation on the multi-source heterogeneous operational data streams to obtain a standardized operational data sequence, including: Acquire multi-source heterogeneous operational data streams of the target enterprise within the monitoring period, including financial audit reports, supply chain contract performance records, logistics settlement vouchers, and internal compliance approval and processing minutes; The multi-source heterogeneous operational data stream is subjected to cross-modal data cleaning protocol parsing and global temporal alignment processing to obtain an initial standardized data sequence with unified dimensions; A pre-built enterprise business domain ontology library is invoked to perform sliding window semantic boundary detection and dependency syntactic relation parsing on the initial standardized data sequence, obtaining semantic boundary and syntactic relation parsing information. Based on the semantic boundary and syntactic relation parsing information, the initial standardized data sequence is discretized and segmented into several atomic semantic feature units with independent business semantic orientations. The atomic semantic feature units are arranged in order according to their sequential positions in the initial standardized data sequence to obtain a standardized business data sequence.

3. The method for identifying semantic risks in enterprise operational data based on neural networks according to claim 2, characterized in that, The standardized operational data sequence is mapped to a high-dimensional continuous vector space to obtain an initial representation matrix. Initial attribute weight tensors are then configured for each semantic feature node in the initial representation matrix, including: Each atomic semantic feature unit in the standardized business data sequence is input into a pre-set multi-layer Transformer architecture context semantic encoder. The multi-head self-attention mechanism inside the encoder calculates the long-range semantic dependency correlation strength between each atomic semantic feature unit, and simultaneously maps each atomic semantic feature unit to a high-dimensional continuous vector space to obtain a high-dimensional embedded feature sequence. The sequence topological order between feature vectors in the high-dimensional embedded feature sequence is analyzed. Based on the sequence topological order, multidimensional tensor dimensionality increase and orthogonal basis space coordinate mapping operations are performed on each feature vector to convert each feature vector into semantic feature nodes with spatial distribution structure, thereby obtaining an initial representation matrix. The semantic feature nodes represent cross-business domain fund flow nodes and compliance status elements. For each semantic feature node in the initial representation matrix, its data source credibility identifier, time-series decay coefficient, and business risk sensitivity prior label are extracted in parallel to obtain multi-dimensional prior label information. Based on the multi-dimensional prior label information, the multi-dimensional features of each semantic feature node are cross-aggregated and weighted normalized through a preset nonlinear feature aggregation network to obtain normalized aggregation weight coefficients. Initial attribute weight tensors are configured for each semantic feature node based on the weight coefficients.

4. The method for identifying semantic risks in enterprise operational data based on neural networks according to claim 3, characterized in that, A dynamic semantic association topology graph is constructed based on the semantic interaction potential gradient. Feature fusion mapping is performed on node pairs that satisfy a preset bonding convergence threshold, and feature rejection masks are applied to node pairs that do not satisfy the preset bonding convergence threshold, resulting in a semantic topology structure graph, including: The semantic interaction potential energy gradient vector field is analyzed to obtain potential energy gradient space analysis information; based on the potential energy gradient space analysis information, directed connection edges are constructed between the semantic feature nodes, and initial connection weights are assigned to each directed connection edge to obtain a dynamic semantic association topology graph. A bonding convergence determination operation is performed on each node pair in the dynamic semantic association topology graph to obtain a bonding convergence determination signal; based on the bonding convergence determination signal, the node pairs are divided into a set of strongly associated node pairs and a set of weakly associated node pairs. For strongly correlated nodes, feature fusion mapping is performed on the set to obtain fused semantic representation nodes; for weakly correlated nodes, feature rejection masking is performed on the set to obtain a masked isolation state graph structure. The semantic topology graph is obtained by performing topological fusion and graph structure reorganization on the fused semantic representation nodes and the masked isolated state graph structure to update the node adjacency relationship and edge weight distribution.

5. The method for identifying semantic risks in enterprise operational data based on neural networks according to claim 4, characterized in that, The semantic topology graph is input into a pre-defined temporal state evolution neural network to extract the risk propagation rate tensor of each semantic feature node within a continuous sampling window, including: The semantic topology graph is processed by temporal slicing according to a preset time step to obtain a continuously sampled temporal graph frame sequence. Temporal position encoding is superimposed on each semantic feature node in the continuously sampled temporal graph frame sequence to obtain a temporal alignment graph sequence. The temporal alignment graph sequence is input into the graph message passing layer of the preset temporal state evolution neural network. Node state forward propagation and cross-frame feature aggregation are performed along the time axis to obtain aggregated node state features. Based on the aggregated node state features, the representation space coordinate offset of each semantic feature node between adjacent sampling windows is calculated to obtain the semantic state displacement vector. Based on the semantic state displacement vector, the temporal attention gating module inside the preset temporal state evolution neural network is invoked to calculate the instantaneous flux of risk signal flow along the graph topology edge between each semantic feature node; the instantaneous flux of risk signal flow is subjected to multidimensional feature decoupling and tensor projection along the graph topology edge to obtain the risk propagation rate tensor.

6. The method for identifying semantic risks in enterprise operational data based on neural networks according to claim 5, characterized in that, Perform temporal difference operations on the risk propagation rate tensor to obtain the risk evolution acceleration component. Based on the risk evolution acceleration component, dynamically recalibrate the node connection weights in the semantic topology graph to obtain a recalibrated semantic topology graph, including: The risk propagation rate tensor is subjected to second-order temporal difference operation within a continuous sampling window sequence to extract the abrupt gradient and trend slope of the risk signal propagation rate, thereby obtaining the risk evolution acceleration component. The risk evolution acceleration component is input into a preset graph weight adaptive adjustment network. The nonlinear activation function inside the preset graph weight adaptive adjustment network calculates the dynamic adjustment coefficient of each connection edge. Based on the dynamic adjustment coefficient, connection strength enhancement mapping is performed on connection edges with a positive acceleration component increase trend, and connection strength suppression mapping is performed on connection edges with a negative acceleration component decrease trend, to obtain the node connection weight update matrix. The node connection weight update matrix is ​​fused with the current adjacency matrix of the semantic topology graph element by element, and the edge weight distribution is updated synchronously to filter out redundant connection edges below the preset dynamic sparsity threshold, so as to obtain the recalibrated semantic topology graph.

7. The method for identifying semantic risks in enterprise operational data based on neural networks according to claim 6, characterized in that, The recalibrated semantic topology diagram is input into a preset multi-layer feedforward risk identification neural network. Using a nonlinear potential energy convergence method for risk semantic features, the semantic risk identification results of enterprise operating data and the corresponding risk confidence level distribution are obtained, including: The recalibrated semantic topology diagram is input into the first layer input interface of the preset multi-layer feedforward risk identification neural network, and the global risk semantic feature vector is obtained through global feature aggregation. The global risk semantic feature vector is input into the continuous hidden layer of the preset multi-layer feedforward risk recognition neural network. The continuous hidden layer performs nonlinear potential energy convergence operation on the risk semantic features to obtain the convergent discriminant feature tensor. The convergent discriminative feature tensor is subjected to dimensionality compression and feature recombination to obtain a high-order discriminative feature sequence. The high-order discriminative feature sequence is transmitted to the output layer of the preset multi-layer feedforward risk identification neural network, and the probability distribution of each preset risk category is calculated to obtain the initial probability distribution. Peak retrieval and confidence threshold mapping operations are performed on the initial probability distribution to extract the target risk category identifier corresponding to the maximum probability value. The target risk category identifier is then graded according to a pre-set confidence grading interval to obtain the semantic risk identification results of enterprise operating data and the corresponding risk confidence level distribution.

8. A neural network-based enterprise operational data semantic risk identification system, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous operational data streams of the target enterprise within the monitoring period, including financial audit reports, supply chain contract performance records, logistics settlement vouchers, and internal compliance approval and processing minutes. The multi-source heterogeneous operational data streams are sequentially subjected to cross-modal cleaning and alignment and semantic unit segmentation to obtain a standardized operational data sequence. The semantic embedding module is used to map standardized business data sequences to a high-dimensional continuous vector space to obtain an initial representation matrix, and to configure an initial attribute weight tensor for each semantic feature node in the initial representation matrix. The potential energy gradient and topology construction module is used to calculate the semantic interaction potential energy gradient between each semantic feature node based on the initial representation matrix and the initial attribute weight tensor, construct a dynamic semantic association topology graph based on the semantic interaction potential energy gradient, perform feature fusion mapping on node pairs that meet the preset bonding convergence threshold, and perform feature rejection masking on node pairs that do not meet the preset bonding convergence threshold to obtain a semantic topology structure graph. The temporal evolution module is used to input the semantic topology graph into a preset temporal state evolution neural network and extract the risk propagation rate tensor of each semantic feature node within a continuous sampling window; The acceleration and recalibration module is used to perform temporal difference operations on the risk propagation rate tensor to obtain the risk evolution acceleration component, and to perform dynamic recalibration on the node connection weights in the semantic topology graph based on the risk evolution acceleration component to obtain a recalibrated semantic topology graph. The risk identification output module is used to input the recalibrated semantic topology diagram into a preset multi-layer feedforward risk identification neural network. It adopts the nonlinear potential energy convergence method of risk semantic features to obtain the semantic risk identification results of enterprise operating data and the corresponding risk confidence level distribution.

Images