Digital economic risk intelligent management evaluation system and method

Abstract

The invention relates to the technical field of digital economic risk management, and discloses a digital economic risk intelligent management evaluation system and method. According to the method, source tracing marking is carried out on an original behavior flow obtained from a heterogeneous data source, and a traceable initial behavior sequence is formed. And processing the sequence in parallel by using a multi-channel sensing mechanism, extracting and crossly fusing the features of different abstract levels, and generating a composite feature spectrum. Performing map embedding mapping based on the risk knowledge map, and matching risk nodes and paths to construct a preliminary risk situation map; injecting a plurality of environmental disturbance factors into the dynamic calculation sand table, simulating the dynamic evolution trajectory of the risk situation, and recording the end point state of each trajectory to form a risk evolution set. And through evaluation and convergence of an end point state, a stable risk state is screened out, and a deterministic conclusion is generated. According to the method, through multi-level fusion and dynamic deduction of the features, the comprehensiveness of risk feature characterization and the dynamic prediction capability of risk assessment are improved.

Description

Technical Field

[0001] This invention relates to the field of digital economy risk management technology, specifically to a digital economy risk intelligent management and assessment system and method. Background Technology

[0002] In the field of digital economy risk management, existing technologies typically rely on behavioral features extracted from single or limited data sources for risk analysis. These methods often employ sequential or single-model processing of behavioral data, focusing on extracting features at a specific level. This approach results in extracted features with limited dimensions, and information from different levels of abstraction is fragmented, making it difficult to form a comprehensive picture of complex and hidden risk behaviors. The one-sidedness of feature representation limits the accuracy of subsequent risk assessments and early warning capabilities, hindering effective responses to new and cross-domain digital economy risks.

[0003] Current risk assessment models are mostly based on static analysis or probabilistic statistics of historical data, and their assessment conclusions are built upon the analysis of past states. Even with the introduction of graph technology, it is mainly used for static relationship display and matching, lacking the ability to dynamically extrapolate how the risk situation may evolve in the future, what external environmental factors will influence it, and to what state it may evolve into. Existing solutions are unable to simulate the dynamic interaction and state transition process of risk nodes and paths under the influence of external disturbances such as policy adjustments, market changes, and technological iterations. This results in risk assessment results lagging behind actual changes and failing to provide a dynamic and evolving basis for forward-looking decision-making.

[0004] There is a need for an assessment method that can integrate multi-dimensional and multi-level behavioral characteristics and can dynamically and predictably simulate risk situations in order to solve the problems of incomplete feature expression and static risk assessment. Summary of the Invention

[0005] The purpose of this invention is to provide a digital economy risk intelligent management and assessment system and method to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a digital economy risk intelligent management and assessment method, the method comprising: Obtain the original behavior stream related to the target object from the heterogeneous data source, trace and mark the original behavior stream, and generate an initial behavior sequence with traceable marks; The initial behavior sequence with traceable tags is processed in parallel using a multi-channel sensing mechanism. Feature expressions at different abstraction levels are extracted in each processing channel, and feature expressions at different abstraction levels are cross-fused to generate a composite feature map. Based on the preset risk knowledge graph, the composite feature graph is embedded and mapped. Risk nodes and risk paths that match the composite feature graph are found in the topology of the risk knowledge graph. A preliminary risk situation map is generated based on the matched risk nodes and risk paths. A dynamic calculation sandbox is established, the preliminary risk situation map is imported into the dynamic calculation sandbox, a variety of preset environmental disturbance factors are injected, the evolution trajectory of the preliminary risk situation map under a variety of preset environmental disturbances is simulated, the final state of each evolution trajectory is recorded, and a risk evolution set is formed. The credibility assessment and state convergence of each endpoint state in the risk evolution set are performed. Stable risk states that meet the consensus conditions are selected from the risk evolution set, and a deterministic risk assessment conclusion for the target object is generated.

[0007] Preferably, the step of obtaining the original behavior stream related to the target object from the heterogeneous data source, tracing and marking the original behavior stream, and generating an initial behavior sequence with traceable markers specifically includes: Multiple data probes are configured, each corresponding to a type of heterogeneous data source. The data probes continuously capture raw signal pulses from the corresponding heterogeneous data source in a streaming access manner. The signal integrity of each raw signal pulse captured by the data probe is verified. After the verification is passed, a source identifier, timestamp sequence and integrity check code are added to each raw signal pulse to form a standardized data unit. Standardized data units from different data probes are interleaved and sorted according to the chronological order of the timestamp sequence. During the interleaving and sorting process, logical correlation is detected based on the semantic content of the standardized data units, and standardized data units with strong logical correlation are pre-bound. Apply time window segmentation to the standardized data unit stream after interleaving sorting and pre-binding, dividing the continuous data unit stream into behavioral segments with time spans, and generating a globally unique segment identifier for each behavioral segment; Based on the fragment identifier, all standardized data units within each behavior fragment are aggregated, and a semantic summary describing the core activities of the behavior fragment is generated. The semantic summary is associated with the corresponding behavior fragment and stored, and finally output as an initial behavior sequence with traceable tags.

[0008] Preferably, the parallel processing of the initial behavior sequence with traceable tags using a multi-channel sensing mechanism, extracting feature representations at different abstract levels within each processing channel, specifically includes: Construct a parallel action parsing channel, a pattern recognition channel, and an intent inference channel, and synchronously input the initial behavior sequence with traceable tags into the parallel action parsing channel, the pattern recognition channel, and the intent inference channel; In the parallel action parsing channel, atomic operation parsing is performed on each standardized data unit in the initial action sequence to identify the specific operation instruction, operation object and operation result represented by the data unit. The atomic operations are linked in time order to form an operation chain feature expression. In the pattern recognition channel, the initial behavior sequence is scanned in a sliding window manner. Within the sliding window, the statistical regularity of the standardized data units in terms of value, frequency, or combination is detected to identify periodic patterns, trend patterns, or abnormal fluctuation patterns, thereby forming a statistical pattern feature expression. In the intent inference channel, a causal reasoning network for the behavior sequence is constructed based on the semantic summary of the initial behavior sequence and the contextual relationship of the behavior fragments. Possible behavioral intentions and goals are inferred in the causal reasoning network, forming an intent network feature representation. The process of cross-fusion of feature representations at different abstract levels to generate a composite feature map involves: establishing a feature cross-attention mechanism, enabling mutual calculation of attention weights among operation chain feature representations, statistical pattern feature representations, and intent network feature representations; and performing weighted fusion and information complementarity on operation chain feature representations, statistical pattern feature representations, and intent network feature representations based on the calculated attention weights, thereby generating a multi-layered feature structure that integrates micro-operations, meso-patterns, and macro-intents, i.e., the composite feature map.

[0009] Preferably, the step of performing graph embedding mapping on the composite feature graph based on a preset risk knowledge graph, and finding risk nodes and risk paths matching the composite feature graph in the topology of the risk knowledge graph, specifically includes: The risk knowledge graph includes risk entity nodes, risk relationship edges, and node attribute vectors. The composite feature graph is vectorized and encoded to obtain the vector representation of the composite feature graph. Calculate the semantic similarity between the vector representation of the composite feature graph and the attribute vector of each risk entity node in the risk knowledge graph, and filter out risk entity nodes whose semantic similarity exceeds a preset threshold as a set of candidate risk nodes; Starting from the set of candidate risk nodes, the risk knowledge graph is explored by walking along the risk relationship edges with a finite step length in the topology of the risk knowledge graph. All other risk entity nodes that can be reached from the set of candidate risk nodes are explored, and the sequence of risk relationship edges traversed by the walk is recorded. The sequence of risk relationship edges constitutes a potential risk path. The matching degree between each potential risk path and the composite feature map is evaluated. The matching degree evaluation is based on the synergy between the attribute vectors of all nodes on the risk path and the vector representation of the composite feature map, as well as the degree of consistency between the risk transmission logic represented by the path and the intention network feature expression. The potential risk paths are sorted according to their matching degree, and the several potential risk paths with the highest matching degree and their associated risk entity nodes are selected as risk nodes and risk paths that match the composite feature map.

[0010] Preferably, the step of generating a preliminary risk situation map based on the matched risk nodes and risk paths specifically includes: Using the matched risk nodes as graph nodes and the matched risk paths as directed edges, construct an initial risk topology subgraph; The feature intensity information corresponding to each matched risk node is extracted from the composite feature map, the feature intensity information is quantized into the initial energy value of the risk node, and assigned to the corresponding graph node; Extract the path type and path weight from the matched risk paths, and assign the path type and path weight to the corresponding directed edges in the initial risk topology subgraph. The path weight reflects the strength or probability of risk propagation along the path. The initial risk topology subgraph, the initial energy values ​​of the graph nodes, the path types and path weights of the directed edges are integrated to form a visualized network graph that includes node attributes, edge attributes and topology structure, namely the preliminary risk situation diagram.

[0011] Preferably, the step of establishing a dynamic simulation sandbox involves importing the preliminary risk situation map into the dynamic simulation sandbox, injecting multiple preset environmental disturbance factors, and simulating the evolution trajectory of the preliminary risk situation map under multiple preset environmental disturbances. Specifically, this includes: The dynamic calculation sandbox is a simulation environment that includes node state update rules, edge propagation rules, and environment interaction rules. The graph nodes, directed edges, and their attributes in the preliminary risk situation map are imported into the dynamic calculation sandbox to initialize the sandbox state. Define multiple sets of environmental disturbance factors, each set of environmental disturbance factors simulating a change in an external economic environment or market conditions, including policy shocks, sudden changes in market liquidity, supply chain fluctuations, or disruptive technological events; Each set of environmental disturbance factors is injected into the initialized dynamic calculation sandbox in sequence. According to the node state update rules, edge transmission rules and environmental interaction rules, the energy values ​​of the graph nodes in the preliminary risk situation map are driven to be transmitted, superimposed, dissipated or enhanced along the directed edges, and the sandbox state after each simulation time step is calculated iteratively. Record the changes in the energy values ​​of all graph nodes and the changes in the activation states of directed edges throughout the entire process, from the injection of environmental disturbance factors to the stabilization of the sand table state or the end of the preset simulation time. The changes constitute an evolution trajectory. The process is repeated until all preset environmental disturbance factor groups have been simulated, thereby obtaining multiple evolution trajectories corresponding to multiple sets of environmental disturbance factors.

[0012] Preferably, the recording of the endpoint state of each evolutionary trajectory to form a risk evolution set specifically includes: For each evolution trajectory, extract its dynamic computational sandbox state at the end of the simulation. The sandbox state includes the final energy value of all graph nodes, the final activation state of all directed edges, and the topological energy distribution characteristics of the entire sandbox. The sand table state, the type of injected environmental disturbance factor, and the unique identifier of the evolution trajectory are packaged together to form a risk evolution sample. All risk evolution samples are collected to form the risk evolution set, where each sample represents a final risk pattern that the initial risk situation map may evolve into under specific environmental disturbances.

[0013] Preferably, the credibility assessment and state convergence of each endpoint state in the risk evolution set specifically includes: Extract the sandbox state of all risk evolution samples from the risk evolution set, perform cross-sample alignment comparison of the final energy values ​​of graph nodes in all sandbox states, and calculate the statistical distribution of the final energy value of each graph node in different samples. For each graph node, the consistency of its final energy value in different samples is checked. If the final energy values ​​of the graph node in more than a preset proportion of samples are all in a similar range, then the graph node state is considered to have stable consensus in multiple environments. The similarity between the topological energy distribution characteristics of different risk evolution samples is examined. Samples with high similarity in topological energy distribution characteristics are classified into several state clusters, and each state cluster represents a common final risk pattern. Feature fusion is performed on the sandbox states of all samples within each state cluster to generate a consensus state template representing the state cluster. The consensus state template includes stable node energy values ​​and typical topology within the state cluster.

[0014] Preferably, the step of selecting stable risk states that meet consensus conditions from the risk evolution set and generating a deterministic risk assessment conclusion for the target object specifically includes: Set consensus condition thresholds, which include a node state stability ratio threshold and a sample number ratio threshold within a state cluster; State clusters that meet the node state stability ratio threshold are selected from the graph nodes. Then, state clusters whose sample count in the state cluster exceeds the ratio threshold of the sample count in the risk evolution set are selected from the state clusters. The state clusters are marked as stable risk states that meet the consensus conditions. Extract the consensus state template corresponding to each marked stable risk state, analyze the core risk type represented by the high-energy graph nodes in the consensus state template, as well as the high-risk energy transmission path, and generate descriptive conclusions on the risk exposure, risk correlation, and risk stability of the target object, i.e., the deterministic risk assessment conclusion.

[0015] Preferably, the present invention also includes a digital economy risk intelligent management and assessment system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the digital economy risk intelligent management and assessment method described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: A multi-channel perception mechanism is employed to process labeled behavioral sequences in parallel. Each independent channel specializes in extracting feature representations at different levels of abstraction, and these features, from bottom to top, are cross-fused to generate a composite feature map. This approach avoids the limitations of a single feature perspective, enabling behavioral representations to simultaneously encompass original behavioral patterns, combinatorial logic, and high-level semantics, constructing feature representations with higher information density and stronger correlation. Through the cross-fusion of features at different levels, the model's ability to identify and represent complex, disguised, or cross-chain risky behaviors is enhanced, providing high-quality, multi-dimensional input for accurately matching risk knowledge graphs.

[0017] By establishing a dynamic simulation sandbox, a preliminary risk situation map generated through static matching is placed within it, and various preset environmental disturbance factors are actively injected to simulate the evolution trajectory of the risk situation under multiple possible future environments. This scheme transforms risk assessment from static snapshot analysis to dynamic process extrapolation, revealing the transmission path, evolution direction, and possible final state of risk under different external condition changes through computational simulation. Recording the final states under various disturbances forms a set, enabling the risk assessment conclusions to reflect the multiple potential development directions and stability of the risk, thus outputting a more forward-looking and resilient risk judgment that considers environmental uncertainties. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent management and assessment method for digital economy risks described in this invention. Figure 2 A flowchart for generating an initial sequence of actions with traceable markers; 2 Figure 3 Flowchart for generating a preliminary risk situation map; 5 Figure 4 The graph shows the consensus test results for node state stability. Figure 5 Heatmap of the stability of risk nodes in each state cluster. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides a digital economy risk intelligent management and assessment method. The method includes: first, acquiring original behavioral flows related to the target object from heterogeneous data sources; tracing and marking these original behavioral flows, attaching traceable source, time, and integrity information to each data unit, thereby generating an initial behavioral sequence with traceable markings. Then, using a multi-channel perception mechanism, the initial behavioral sequence is processed in parallel, extracting feature expressions at different abstract levels such as micro-operations, meso-modes, and macro-intentions through different processing channels, and cross-merging these features to form a composite feature map with complementary information. Next, based on a preset risk knowledge graph, the composite feature map is mapped into its topological space. Through semantic matching and path exploration, the risk nodes and risk paths most associated with the current features are found, and a preliminary risk situation map is generated based on this. Subsequently, a dynamic calculation sandbox is established, the preliminary risk situation map is imported into it, and by injecting various preset environmental disturbance factors, the evolution process of the risk situation under different external shocks is simulated, recording the endpoint state of each evolution trajectory to form a risk evolution set. Finally, credibility assessment and state convergence analysis are performed on all endpoint states in the set to screen out stable risk states that can still form consensus under multiple environmental disturbances, thereby generating a deterministic risk assessment conclusion for the target object.

[0021] In one embodiment of the present invention, see [reference] Figure 2Multiple data probes are configured, each corresponding to a heterogeneous data source of a certain type. The data probes continuously capture raw signal pulses from their respective heterogeneous data sources via streaming access. Signal integrity is verified for each raw signal pulse captured by the data probe. After successful verification, a source identifier, timestamp sequence, and integrity check code are appended to each raw signal pulse to form a standardized data unit. The standardized data units from different data probes are interleaved and sorted according to the chronological order of the timestamp sequence. During the interleaving and sorting process, logical correlation is detected based on the semantic content of the standardized data units, and standardized data units with strong logical correlation are pre-bound. A time window segmentation is applied to the interleaved, sorted, and pre-bound standardized data unit stream, dividing the continuous data unit stream into behavioral segments with time spans. A globally unique segment identifier is generated for each behavioral segment. Based on the segment identifier, all standardized data units within each behavioral segment are aggregated, and a semantic summary describing the core activity of that behavioral segment is generated. The semantic summary is associated with the corresponding behavioral segment and stored, ultimately outputting an initial behavioral sequence with traceable tags.

[0022] A parallel action parsing channel, a pattern recognition channel, and an intent inference channel are constructed. The initial behavior sequence with traceable tags is synchronously input into these channels. In the parallel action parsing channel, atomic operations are parsed for each standardized data unit in the initial behavior sequence to identify the specific operation instruction, operation object, and operation result represented by the data unit. The atomic operations are linked in chronological order to form an operation chain feature expression. In the pattern recognition channel, the initial behavior sequence is scanned using a sliding window approach. Within the sliding window, statistical patterns in the numerical values, frequencies, or combinations of standardized data units are detected to identify periodic patterns, trend patterns, or abnormal fluctuation patterns, forming a statistical pattern feature expression. In the intent inference channel, a causal inference network for the behavior sequence is constructed based on the semantic summary of the initial behavior sequence and the contextual relationships of the behavior fragments. The behavioral intent and target are inferred from the causal inference network, forming an intent network feature expression. A feature cross-attention mechanism is established, which enables the operation chain feature expression, statistical pattern feature expression and intention network feature expression to calculate attention weights among each other. Based on the calculated attention weights, the operation chain feature expression, statistical pattern feature expression and intention network feature expression are weighted, fused and complemented to generate a multi-layer feature structure that integrates micro-operations, meso-patterns and macro-intentions, namely the composite feature map.

[0023] In implementation, multiple data probes are configured and deployed. One probe connects to the payment gateway logs, continuously capturing raw signal pulses of transaction amount and payment tool type via streaming access. Another probe connects to the user rating system, capturing raw signal pulses of rating stars and text content. A third probe connects to internal messaging records, capturing raw signal pulses of communication frequency and keywords. Specifically, a raw signal pulse captured by the payment gateway data probe, containing "Transaction Amount: 10,000 yuan, Status: Success," undergoes signal integrity verification. This verifies whether the data format conforms to preset specifications and whether the hash value matches. After successful verification, a source identifier "PAYMENT_GATEWAY_01," a timestamp sequence "2023-10-28T14:30:25.123Z," and an integrity check code "a1b2c3d4" are appended to this raw signal pulse, forming a standardized data unit. Raw signal pulses from the rating system data probe and the internal messaging data probe undergo the same process, being encapsulated into standardized data units with their respective source identifiers, timestamp sequences, and integrity check codes. In practice, standardized data units from payment gateway data probes, evaluation system data probes, and internal message data probes are interleaved and sorted according to their timestamp sequence. During the interleaving and sorting process, a standardized data unit indicating a successful payment and an internal message requesting an offline transaction, with similar timestamps, are detected. Based on the semantic content, a strong logical relationship is identified between "payment" and "offline transaction request," and these two standardized data units are pre-bound. A 5-minute time window is applied to the stream of standardized data units that has completed interleaving, sorting, and pre-binding to segment the continuous data unit stream into multiple behavioral segments. A globally unique segment identifier, such as "SESSION_20231028_1430_1435," is generated for each behavioral segment. In the specific implementation, based on the fragment identifier "SESSION_20231028_1430_1435", all standardized data units within the 5-minute time window are aggregated, and a semantic summary describing "a large payment occurred and was associated with an offline transaction request within the time window from 14:30 to 14:35" is generated. This semantic summary is associated with and stored with the corresponding behavioral fragment, and the final output is an initial behavioral sequence with traceable tags.

[0024] In some embodiments, the constructed parallel action parsing channel, pattern recognition channel, and intent inference channel are activated simultaneously, and the initial behavior sequence with traceable markers is synchronously input into these three processing channels. In the parallel action parsing channel, an atomic operation is parsed on a standardized data unit in the initial behavior sequence, identifying the operation instruction as "transfer," the operation object as "user B's account," and the operation result as "success." This atomic operation is then linked chronologically with preceding and following atomic operations such as "login" and "check balance," forming an operation chain feature expression of "login -> check balance -> transfer to user B." In the pattern recognition channel, a sliding window containing 10 standardized data units scans the initial behavior sequence. Within the sliding window, a rapidly rising statistical pattern is detected in the "transfer amount" value sequence, identifying a trend pattern. Simultaneously, frequent changes in the "login IP address" within a short period are detected, identifying an abnormal fluctuation pattern. These combined features form a statistical pattern feature expression. In the intent inference channel, based on the semantic summary "large payment and associated with offline transaction request" of the behavior segment "SESSION_20231028_1430_1435", and combined with the semantic summary "frequent login attempts" of the preceding and following behavior segments "SESSION_20231028_1425_1430", a causal inference network of behavior sequence is constructed. In the causal inference network, the behavioral intent of "circumventing platform monitoring to transfer funds" is inferred, forming an intent network feature expression.

[0025] In some embodiments, the established feature cross-attention mechanism begins to operate. Operation chain feature representations, statistical pattern feature representations, and intent network feature representations are treated as input sources for queries, keys, and values, respectively, and their roles can be rotated. Attention weights are obtained by calculating the correlation between feature representations. The calculated attention weights reflect the importance associations between features at different abstraction levels. For example, the "avoiding monitoring" intent inferred from the intent inference channel makes the attention mechanism focus more on the "small transfers to different recipients" pattern in the operation chain and the "abnormal transaction time distribution" pattern in the statistical pattern. The operation chain feature representations, statistical pattern feature representations, and intent network feature representations are weighted, fused, and information-complemented based on the calculated attention weights. A specific fusion formula can be expressed as: in: This represents the fused composite feature map. These represent operation chain feature representation, statistical pattern feature representation, and intent network feature representation, respectively. This represents a linear transformation or projection function used to map different feature representations to a unified dimension. The value is calculated through the feature cross-attention mechanism and corresponds to The normalized attention weight coefficients.

[0026] In one embodiment of the present invention, see [reference] Figure 3 The risk knowledge graph contains risk entity nodes, risk relationship edges, and node attribute vectors. A vectorized representation of the composite feature graph is obtained by vectorizing the composite feature graph. The semantic similarity between the vector representation of the composite feature graph and the attribute vector of each risk entity node in the risk knowledge graph is calculated. Risk entity nodes with semantic similarity exceeding a preset threshold are selected as a candidate risk node set. Starting from the candidate risk node set, a finite-step walk is performed along the risk relationship edges in the topology of the risk knowledge graph to explore all other risk entity nodes reachable from the candidate risk node set. The sequence of risk relationship edges traversed is recorded, forming potential risk paths. The matching degree between each potential risk path and the composite feature graph is evaluated. The matching degree is based on the synergy between the attribute vectors of all nodes on the risk path and the vector representation of the composite feature graph, as well as the degree of consistency between the risk transmission logic represented by the path and the intention network feature expression. The potential risk paths are ranked according to the matching degree, and the several potential risk paths with the highest matching degree and their associated risk entity nodes are selected as the risk nodes and risk paths matching the composite feature graph.

[0027] Using matched risk nodes as graph nodes and matched risk paths as directed edges, an initial risk topology subgraph is constructed. Feature intensity information corresponding to each matched risk node is extracted from the composite feature graph, quantified into initial energy values ​​for the risk nodes, and assigned to the corresponding graph nodes. Path type and path weight are extracted from the matched risk paths and assigned to the corresponding directed edges in the initial risk topology subgraph. The path weight reflects the strength or probability of risk propagation along the path. The initial risk topology subgraph, the initial energy values ​​of the graph nodes, and the path types and path weights of the directed edges are integrated to form a visualized network graph containing node attributes, edge attributes, and topological structure—the preliminary risk situation map.

[0028] In practical implementation, taking a peer-to-peer lending platform's credit risk assessment of borrowers as an example, the pre-set risk knowledge graph contains various risk entity nodes. For example, the risk entity node "abnormal transaction behavior" has attribute vectors describing characteristics such as excessively high transaction frequency and irregular transaction times; the risk entity node "multiple borrowing" has attribute vectors describing characteristics such as having loan application records on multiple platforms simultaneously; and the risk entity node "identity information forgery" has attribute vectors describing characteristics such as altered document images and inconsistent identity information verification. Risk relationship edges connect these risk entity nodes. For example, there is a "potentially accompanying" risk relationship edge between the "abnormal transaction behavior" node and the "suspicious use of funds" node. In practical implementation, the previously generated composite feature graph, which integrates features such as "frequent nighttime transfers," "multi-account linked payments," and "intent to evade monitoring," is vectorized and encoded. The encoding process is completed through a pre-trained deep neural network, resulting in a 128-dimensional vector representation, i.e., the vector representation of the composite feature graph. In the implementation of pre-trained deep neural networks, a model pre-trained on a large-scale risk behavior dataset is used. This model learns complex feature patterns in the data and has the ability to map structured or unstructured composite feature maps to a low-dimensional dense vector space. During the encoding process, the composite feature map serves as the input to the neural network. Through multiple layers of nonlinear transformations, the feature information is progressively abstracted and compressed, ultimately generating a 128-dimensional floating-point vector at the output layer. This vector efficiently captures the semantic content of the composite feature map for subsequent semantic similarity calculations. The pre-training objective of the neural network is typically set to maximize the clustering of similar risk behaviors in the vector space, ensuring that the generated vector representation has good discriminative power and representativeness.

[0029] In practice, the semantic similarity between the vector representation of the composite feature graph and the attribute vector of each risk entity node in the risk knowledge graph is calculated using the cosine similarity method. For the attribute vector of the "abnormal transaction behavior" node, the calculated cosine similarity is 0.85; for the attribute vector of the "multiple borrowing" node, the calculated cosine similarity is 0.45; and for the attribute vector of the "identity forgery" node, the calculated cosine similarity is 0.30. A preset threshold of 0.60 for semantic similarity is set, and the risk entity nodes "abnormal transaction behavior" and "multiple borrowing" with semantic similarity exceeding the preset threshold are selected. These two nodes constitute the candidate risk node set.

[0030] In some embodiments, starting with the candidate risk node sets "abnormal transaction behavior" and "multiple borrowing," a finite step-size of 3 is performed along risk relationship edges of types "potentially accompanying," "causing," and "associating" within the risk knowledge graph's topology. Starting from the "abnormal transaction behavior" node, the exploration can reach the "suspicious use of funds" node and the "money laundering risk" node; starting from the "multiple borrowing" node, the exploration can reach the "high debt risk" node and the "fraud risk" node. The sequence of "potentially accompanying" and "causing" edges traversed from the "abnormal transaction behavior" node to the "money laundering risk" node is recorded; this sequence of risk relationship edges constitutes a potential risk path. Similarly, the sequence of risk relationship edges recorded from the "multiple borrowing" node to the "fraud risk" node constitutes another potential risk path. In specific implementations, the matching degree between each potential risk path and the composite feature graph is evaluated. For the potential risk path from "abnormal transaction behavior" to "money laundering risk," the synergy between the attribute vectors of the three nodes—"abnormal transaction behavior," "suspicious use of funds," and "money laundering risk"—and their representations in the composite feature graph is assessed, and a comprehensive synergy score is calculated. Simultaneously, the risk transmission logic represented by this path—"abnormal transactions accompanied by suspicious use of funds ultimately leading to money laundering risk"—is evaluated, along with the degree of consistency with the intent network feature expression of "evasion of monitoring intent" in the composite feature graph, and a logical consistency score is calculated. The matching degree assessment is completed by a comprehensive function, which can be understood as combining the synergy score and the logical consistency score. A specific matching degree calculation formula can be expressed as: in: Represents the final match degree. Vector representation of the composite feature map. Representing the first potential risk path The attribute vector of each node, This represents the total number of nodes on this potentially risky path. This represents the cosine similarity calculation function. It is a weighting coefficient between 0 and 1 used to balance the two contributions. It is a vectorized representation of the risk transmission logic represented by the path. It is a vector representation of the intent network features. This represents a function that calculates the correlation between the two. Based on the matching degree... Multiple potential risk paths are ranked, assuming the path of "abnormal transaction behavior -> money laundering risk" is... A value of 0.88 indicates the "multiple borrowing -> fraud risk" path. With a value of 0.72, the two potential risk paths with the highest matching degree and their associated risk entity nodes are selected as risk nodes and risk paths that match the composite feature map.

[0031] Optionally, using the matched risk nodes such as "abnormal transaction behavior," "multiple borrowing," "money laundering risk," and "fraud risk" as graph nodes, and the two matched risk paths "abnormal transaction behavior -> money laundering risk" and "multiple borrowing -> fraud risk" as directed edges, an initial risk topology subgraph containing four nodes and two edges is constructed. Feature intensity information corresponding to the "abnormal transaction behavior" node is extracted from the composite feature graph, such as the intensity values ​​of "nighttime transfer frequency" and "multi-account correlation." This feature intensity information is quantified into an initial energy value of 85 using a linear weighted model and assigned to the corresponding "abnormal transaction behavior" graph node in the initial risk topology subgraph. Similarly, the initial energy value is assigned to the "multiple borrowing" graph node (65), the "money laundering risk" graph node (40), and the "fraud risk" graph node (50). From the matched risk path "Abnormal Transaction Behavior -> Money Laundering Risk", the path type is extracted as "Transmission", and a path weight of 0.7 is extracted based on historical statistical data; from the path "Multiple Lending -> Fraud Risk", the path type is extracted as "Enhancement", and a path weight of 0.6 is extracted. These path types and path weights are then assigned to the corresponding directed edges in the initial risk topology subgraph.

[0032] In one embodiment of the present invention, the dynamic simulation sandbox is a simulation environment that includes node state update rules, edge propagation rules, and environmental interaction rules. Graph nodes, directed edges, and their attributes from the initial risk situation map are imported into the dynamic simulation sandbox to initialize its state. Multiple sets of environmental disturbance factors are defined, each simulating a change in an external economic environment or market condition, including policy shocks, sudden changes in market liquidity, supply chain fluctuations, or disruptive technological events. Each set of environmental disturbance factors is sequentially injected into the initialized dynamic simulation sandbox. Based on the node state update rules, edge propagation rules, and environmental interaction rules, the energy values ​​of the graph nodes in the initial risk situation map are driven to propagate, superimpose, dissipate, or enhance along the directed edges, and the sandbox state after each simulation time step is iteratively calculated. The changes in the energy values ​​of all graph nodes and the changes in the activation states of directed edges are recorded from the injection of environmental disturbance factors until the sandbox state reaches stability or the preset simulation duration ends; these changes constitute an evolution trajectory. Repeat this process until all preset environmental disturbance factor groups have been simulated, thereby obtaining multiple evolution trajectories corresponding to multiple sets of environmental disturbance factors.

[0033] In practical implementation, taking the assessment of liquidity risk of a cryptocurrency exchange as an example, the dynamic simulation sandbox is established as a simulated environment containing node state update rules, edge propagation rules, and environmental interaction rules. The node state update rules define how the energy value of a graph node decays over time or increases based on input. The edge propagation rules define the proportion and conditions under which energy is transferred from the source node to the target node along directed edges. The environmental interaction rules define how external disturbances affect the node energy value and edge propagation efficiency. In practical implementation, a preliminary risk situation map is imported into the dynamic simulation sandbox. The preliminary risk situation map includes graph nodes representing "spreading market panic," "flash crashes in mainstream coin prices," and "collateral liquidation pressure," as well as directed edges connecting these nodes. The directed edges have an "aggravation" type and a path weight of 0.8. The import process maps these graph nodes, directed edges, and their initial energy values, path types, and path weights into state variables and relationship matrices within the dynamic simulation sandbox, thereby completing the initialization of the sandbox state.

[0034] In some embodiments, multiple sets of environmental disturbance factors are defined for simulation. The first set of environmental disturbance factors simulates the policy shock of "major central banks around the world suddenly announcing interest rate hikes," and this set of factors includes the parameters of "rising risk-free interest rates" and "declining risk asset preferences." The second set of environmental disturbance factors simulates the market liquidity mutation event of "a major stablecoin decoupling," and this set of factors includes the parameters of "obstruction of stablecoin exchange channels" and "sharp drop in market confidence index." The third set of environmental disturbance factors simulates the technological disruption event of "severe congestion in the underlying blockchain network," and this set of factors includes the parameters of "transaction confirmation delay" and "soaring network fees." In specific implementation, each set of environmental disturbance factors is injected into the initialized dynamic calculation sandbox in sequence, starting with the first set of environmental disturbance factors of "major central banks around the world suddenly announcing interest rate hikes." According to the environmental interaction rules, the parameter of "rising risk-free interest rates" will directly weaken the basic energy value of all graph nodes related to investment, while the parameter of "declining risk asset preferences" will amplify the transmission coefficient of directed edges of the "exacerbation" type. According to the node state update rules, the energy value of the "Market Panic Spreads" graph node in the dynamic simulation sandbox increases by 10 points due to the external shock. According to the edge transmission rules, 20% of its energy value is transmitted along the "Aggravation" edge to the "Mainstream Cryptocurrency Price Flash Crash" graph node, with a transmitted energy value of (75+10)*0.2=17 points. After receiving this energy, the dynamic simulation sandbox updates the energy value of the "Mainstream Cryptocurrency Price Flash Crash" graph node according to the node state update rules, changing it from 60 points to 77 points. This process iterates in each simulation time step, driving the graph node energy values ​​in the initial risk situation map to be transmitted, superimposed, dissipated, or enhanced along the directed edges.

[0035] In some embodiments, the entire process is recorded from the injection of the environmental disturbance factor "sudden announcement of interest rate hikes by major central banks worldwide" to the sandbox state reaching stability or the end of the preset simulation period. The recorded content includes the energy value changes of the graph nodes for "spreading market panic," "flash crash in mainstream cryptocurrency prices," and "collateral liquidation pressure" at each time step, as well as the activation status of the directed edges connecting them. These chronologically recorded changes constitute an evolution trajectory under the disturbance of "policy interest rate hikes." The injection and simulation process is repeated. Then, the second set of environmental disturbance factors, "decoupling of a large stablecoin," is injected into the current sandbox state, simulating and recording the second evolution trajectory. Next, the third set of environmental disturbance factors, "severe congestion in the underlying blockchain network," is injected, simulating and recording the third evolution trajectory, until all preset environmental disturbance factor groups have completed the simulation, thus obtaining multiple evolution trajectories corresponding to multiple sets of environmental disturbance factors.

[0036] Optionally, the specific calculation of the node state update rules in the dynamic computation sandbox can be formalized, with a single rule determining the graph node. At time step Energy value The update formula is: in: Representative graph node At time step Energy value at that time It is the energy decay coefficient. It is all of Let be the set of neighboring nodes in the target node's graph. It is one of the source graph nodes. From graph nodes To graph nodes The baseline propagation weight of the edge, In time step The environmental disturbance modulation factor of that side at that time. It is a source graph node At time step Energy value, Graph nodes Sensitivity coefficient to external disturbances In time step The intensity of the injected environmental disturbance factors, Graph nodes At time step The energy dissipation value is recorded. The final state of each evolution trajectory is recorded to form a risk evolution set. In specific implementation, for the evolution trajectory under the disturbance of "major central banks around the world suddenly announcing interest rate hikes", the dynamic calculation sandbox state when it reaches stability at the 85th simulation time step is extracted. The sandbox state includes the final energy value of the graph node of "market panic spread" (92), the final energy value of the graph node of "flash crash of mainstream coin prices" (88), and the final energy value of the graph node of "collateral liquidation pressure" (70). All directed edges are in the "active" state, and the topological energy distribution characteristics of the entire sandbox show that the risk energy is highly concentrated in the price-related nodes.

[0037] In one embodiment of the present invention, for each evolution trajectory, the dynamic computational sandbox state at the end of the simulation is extracted. The sandbox state includes the final energy values ​​of all graph nodes, the final activation state of all directed edges, and the topological energy distribution characteristics of the entire sandbox. The sandbox state, the type of injected environmental disturbance factor, and the unique identifier of the evolution trajectory are packaged to form a risk evolution sample. All risk evolution samples are collected to form a risk evolution set, where each sample represents a final risk pattern evolved from the initial risk situation map under a specific environmental disturbance.

[0038] The sandbox states of all risk evolution samples are extracted from the risk evolution set. The final energy values ​​of graph nodes in all sandbox states are compared across samples, and the statistical distribution of the final energy value of each graph node in different samples is calculated. For each graph node, the consistency of its final energy value across different samples is examined. If the final energy values ​​of the graph node in more than a preset proportion of samples are all within a similar numerical range, then the graph node state is considered to have stable consensus in multiple environments. The similarity between the topological energy distribution characteristics of different risk evolution samples is examined. Samples with high similarity in topological energy distribution characteristics are classified into several state clusters, each representing a common final risk pattern. Feature fusion is performed on the sandbox states of all samples within each state cluster to generate a consensus state template representing the state cluster. The consensus state template includes stable node energy values ​​and typical topological structures within the state cluster.

[0039] In practical implementation, taking a fintech company's credit risk assessment of a cluster of micro and small enterprises in a certain region as an example, the risk evolution set contains eight risk evolution samples, corresponding to eight final risk patterns formed under different environmental disturbances. For the first evolution trajectory under the environmental disturbance factor of "regional industry policy tightening", its dynamic calculation sandbox state at the end of the simulation is extracted. The sandbox state includes the final energy value of 25 for the "Industry A Prosperity" graph node, the final energy value of 65 for the "Enterprise Cluster Default Rate" graph node, and the final energy value of 80 for the "Regional Credit Tightening" graph node. The final activation state of all directed edges such as "leading to" and "exacerbating" is "activated". The topological energy distribution characteristics of the entire sandbox present a chain structure in which risk energy is transmitted from industry nodes to default nodes and then to regional financial environment nodes. The sandbox state containing these contents, the injected "regional industry policy tightening" environmental disturbance factor type, and the unique identifier "Sim_001" of the evolution trajectory are packaged to form a risk evolution sample. All eight risk evolution samples, including those identified as Sim_001 to Sim_008, are aggregated to form the risk evolution set. Each sample represents a final risk pattern that evolves from the initial risk profile of the micro and small enterprise cluster under specific environmental disturbances such as "sharp increase in labor costs", "violent fluctuations in international exchange rates", and "increased tariffs in major export markets".

[0040] In some embodiments, the sandbox states of all eight risk evolution samples are extracted from the risk evolution set, and the final energy values ​​of the graph nodes in all sandbox states are compared across samples. For the graph node "Enterprise Cluster Default Rate," its final energy values ​​in the eight samples are calculated as follows: 65, 70, 15, 68, 72, 18, 67, 20. The statistical distribution of these values ​​is analyzed, their mean and variance are calculated, and the clustering of values ​​is observed. For the "Enterprise Cluster Default Rate" graph node, the consistency of its final energy value across different samples is examined. With a preset proportion set at 60%, five samples have final energy values ​​within a similar range of 60 to 75. These five samples account for more than 60% of the total, therefore, the state of the "Enterprise Cluster Default Rate" graph node is considered to have stable consensus across multiple environments. Using the same method to examine the nodes in the "Regional Credit Tightening" graph, their final energy values ​​across the eight samples were 80, 85, 30, 82, 88, 35, 83, and 38, respectively. Five of these samples fell within the similar range of 80 to 90, thus satisfying the stability consensus condition. However, the energy values ​​of the nodes in the "Industry A Prosperity" graph were dispersed across the samples, failing to meet the consensus condition. The similarity of the topological energy distribution characteristics among different risk evolution samples was assessed by comparing the main paths of risk energy transmission and the hierarchical structure of node energy. Samples with high similarity in topological energy distribution characteristics were categorized into two main state clusters: State cluster Alpha, containing five samples (Sim_001, Sim_002, Sim_004, Sim_005, and Sim_007), is characterized by a high concentration of risk energy at default and credit tightening nodes; State cluster Beta, containing three samples (Sim_003, Sim_006, and Sim_008), is characterized by relatively dispersed risk energy and a lower overall level.

[0041] In practical implementation, feature fusion is performed on the sandbox states of all samples within each state cluster to generate a consensus state template representing the state cluster. For state cluster Alpha, the energy values ​​of the "enterprise cluster default rate" graph node and the "regional credit tightening" graph node, which have reached a consensus on stability, are fused among its five samples. For the "industry A prosperity" graph node, which has not reached a consensus, its energy value is marked as "uncertain" or represented by an interval in the consensus state template. The consensus state template of state cluster Alpha also contains its typical chain topology, namely the transmission path of "industry pressure -> rising default rate -> credit tightening". The formation process of the consensus state template can be reflected by a specific feature aggregation formula. For the fusion of the energy values ​​of nodes that have reached a consensus, the formula can be used: in: Represents the nodes in the consensus state template. The energy value after fusion This represents the set of all risk evolution samples belonging to the same state cluster. This is one of the samples. It is a sample Chinese map nodes The final energy value, It is to assign samples The weights are optional. The settings can be configured based on the probability or severity of the environmental disturbance factors simulated by the sample. If unweighted, then... The fusion of topologies is determined based on the edge connections and propagation directions shared by most samples within the state cluster. See Table 1 for the energy value distribution of key nodes in the risk evolution samples.

[0042] Table 1: Energy values ​​of terminal state nodes for a subset of samples in the risk evolution ensemble Risk Evolution Sample Identifier Types of environmental disturbance factors Enterprise Cluster Default Rate Node Energy Value Regional credit tightening node energy value Industry A Prosperity Node Energy Value Sim_001 Regional industry policies tightening 65 80 25 Sim_002 Labor costs surge 70 85 20 Sim_003 Technological substitution breakthrough 15 30 75 Sim_004 International exchange rates fluctuate wildly 68 82 22 Sim_005 Raw material supply chain disruption 72 88 18 Sim_006 Environmental standards have been significantly raised. 18 35 70 Sim_007 Tariffs increased in major export markets 67 83 24 Sim_008 Local tax incentives canceled 20 38 68 As is understandable, Table 1 visually illustrates the differences and clustering of energy values ​​of the same graph nodes under different environmental disturbances, providing a data foundation for verifying node state consistency. Through the aforementioned credibility assessment and state convergence process, consensus state templates representing different stable outcomes were extracted from discrete risk evolution samples.

[0043] See Figure 4 In the node state stability analysis of intelligent risk management in the digital economy, the consensus condition satisfaction of key risk nodes is presented intuitively in the form of a stacked bar chart. The chart uses 60% as the consensus threshold (marked by a yellow dashed line), and quantifies the state stability of each node under multiple environmental disturbances by dividing it into green (meeting consensus conditions) and red (not meeting consensus conditions) percentages: For the "enterprise cluster default rate" node, 62.5% of the samples meet the consensus conditions, corresponding to the energy value of most samples in its risk evolution set clustering in the 60-75 range, which meets the preset stability consensus standard; the "regional credit tightening" node also has a 62.5% sample percentage meeting the consensus conditions, consistent with its energy value clustering characteristic in the 80-90 range; only 37.5% of the samples for the "Industry A Prosperity" node meet the consensus conditions, reflecting that its energy value is dispersed under different environmental disturbances, and no stability consensus has been reached under multiple environments. The visualization of sample proportions provides an intuitive statistical basis for the credibility assessment and convergence analysis of node states in the risk evolution set, and is one of the core data supports for extracting consensus state templates from discrete risk evolution samples.

[0044] In one embodiment of the present invention, consensus condition thresholds are set, including a node state stability ratio threshold and a state cluster sample quantity ratio threshold. State clusters whose graph node states satisfy the node state stability ratio threshold are selected. Then, state clusters whose sample quantity ratio exceeds the state cluster sample quantity ratio threshold are further selected. These state clusters are marked as stable risk states that satisfy the consensus conditions. A consensus state template corresponding to each marked stable risk state is extracted. The core risk types represented by high-energy graph nodes in the consensus state template, as well as the high-risk energy transmission paths, are analyzed. Based on this, descriptive conclusions regarding the target object's risk exposure, risk correlation, and risk stability pattern are generated, i.e., deterministic risk assessment conclusions.

[0045] In practical implementation, continuing with the scenario of a fintech company assessing the credit risk of a cluster of micro and small enterprises as an example, after credibility assessment and state convergence, a risk evolution set is formed, comprising four state clusters: Alpha, Beta, Gamma, and Delta, with a total of ten samples. Consensus thresholds are set: a node state stability ratio threshold of 60% and a sample size ratio threshold of 50% within a state cluster. In practice, state clusters whose graph node states meet the node state stability ratio threshold are selected. For state cluster Alpha, it is checked that the final energy value of the graph node "enterprise cluster default rate" in all five samples is between 65 and 72, satisfying the requirement that this node is stable in 100% of the samples within state cluster Alpha, exceeding the 60% node state stability ratio threshold; the graph node "regional credit tightening" is also stable in all five samples. All graph nodes in state cluster Alpha meet the node state stability ratio threshold. For the state cluster Beta, among its three samples, the energy value of the graph node "Industry A Prosperity" fluctuates significantly across the three samples, failing to reach the stability ratio threshold. Therefore, the state cluster Beta does not meet the node state stability condition. The state clusters Gamma and Delta also undergo the same test.

[0046] In some embodiments, from state clusters that meet the node state stability ratio threshold, state clusters whose proportion of samples within the state cluster exceeds the threshold for the proportion of samples within the state cluster. State cluster Alpha has five samples, the total number of samples in the risk evolution set is ten, and the proportion of samples within state cluster Alpha is 50%. The threshold for the proportion of samples within the state cluster is set to 50%, and the proportion of samples in state cluster Alpha is equal to this threshold. Optionally, a formula for determining whether a state cluster meets the sample quantity consensus condition can be expressed as: in: Representing the The set of samples within a state cluster, This represents the number of samples within that state cluster. Represents the set of risk evolutions. Represents the total number of samples in the risk evolution set. This represents a preset threshold for the proportion of samples within a state cluster. It sets the state cluster Alpha, state cluster Gamma, and state cluster Delta... and Substituting the values, the state cluster Alpha satisfies... The state cluster Gamma satisfies The judgment result is false, and the state cluster Delta satisfies The judgment result is false. Therefore, only the state cluster Alpha is marked as a stable risky state that satisfies the consensus conditions.

[0047] In practical implementation, the consensus state template corresponding to the marked stable risk state, i.e., state cluster Alpha, is extracted. The consensus state template contains the stable node energy values ​​and typical topology within state cluster Alpha. Analyzing the core risk types represented by high-energy graph nodes in the consensus state template, the graph node "Regional Credit Tightening" has the highest fusion energy value of 84, and the graph node "Enterprise Cluster Default Rate" has the second highest fusion energy value of 70. The high-risk energy transmission path is manifested as a directed edge sequence from the "Industry Policy Shock" graph node to the "Enterprise Cluster Default Rate" graph node, and further to the "Regional Credit Tightening" graph node. Based on this, descriptive conclusions are generated regarding the target object's risk exposure, risk correlation, and risk stability. It can be understood that the deterministic risk assessment conclusion states: Under most external environmental disturbances, the target micro and small enterprise cluster exposes a risk exposure highly sensitive to regional credit policy tightening. Its individual default risk is significantly correlated with regional systemic financial risk, and the stable risk pattern is manifested as a pattern where industry pressure is transmitted to and solidified through the default chain to the regional financial environment.

[0048] See Figure 5In the node stability test of the intelligent risk management assessment of the digital economy, this heatmap uses different state clusters (Alpha, Beta, Gamma, Delta) as the horizontal dimension and risk nodes (enterprise cluster default rate, regional credit tightening, industry A prosperity, market liquidity) as the vertical dimension. It quantifies the stability percentage (%) of each risk node in its corresponding state cluster using a color gradient (dark green to red). Specifically, under the Alpha state cluster, the stability of "enterprise cluster default rate" and "regional credit tightening" both reach 100%, while "industry A prosperity" and "market liquidity" reach 88% and 85% respectively, exhibiting high overall stability. In the Beta state cluster, the stability of "industry A prosperity" is only 40%, making it a low-stability risk node in this cluster. The stability of nodes in the Gamma state cluster is concentrated in the 80%-95% range, placing it at a slightly above-average level. The stability of nodes in the Delta state cluster is generally between 55% and 75%, indicating relatively weak overall stability.

[0049] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A digital economic risk intelligent management evaluation method, characterized in that, The method comprises the following steps: obtaining original behavior flow related to a target object from a heterogeneous data source, performing provenance labeling on the original behavior flow, and generating an initial behavior sequence with traceable labels; processing the initial behavior sequence with traceable labels in parallel using a multi-channel perception mechanism, extracting feature representations at different abstraction levels in each processing channel, cross-fusing the feature representations at different abstraction levels, and generating a composite feature atlas; performing atlas embedding mapping on the composite feature atlas according to a preset risk knowledge graph, finding a risk node and a risk path matching the composite feature atlas in the topological structure of the risk knowledge graph, and generating a preliminary risk situation map based on the matched risk node and risk path; establishing a dynamic evolution sand table, importing the preliminary risk situation map into the dynamic evolution sand table, injecting a plurality of preset environmental disturbance factors, simulating the evolution trajectory of the preliminary risk situation map under a plurality of preset environmental disturbances, recording the end state of each evolution trajectory, and forming a risk evolution set; performing credibility evaluation and state convergence on each end state in the risk evolution set, screening a stable risk state satisfying a consensus condition from the risk evolution set, and generating a deterministic risk assessment conclusion for the target object.

2. The digital economy risk intelligent management evaluation method according to claim 1, characterized in that, The method comprises the following steps: configuring a plurality of data probes, each data probe corresponding to a type of heterogeneous data source, the data probe continuously capturing original signal pulses from the corresponding heterogeneous data source in a streaming access manner; performing signal integrity verification on the original signal pulses captured by each data probe, and after verification, adding a source identifier, a timestamp sequence, and an integrity verification code to each original signal pulse to form a standardized data unit; interleaving and sorting the standardized data units from different data probes according to the chronological order of the timestamp sequence, and in the interleaving and sorting process, performing logical correlation detection according to the content semantics of the standardized data units, and pre-binding standardized data units with strong logical correlation; applying time window segmentation to the interleaved and pre-bound standardized data unit stream, dividing the continuous data unit stream into behavior segments with a time span, and generating a globally unique segment identifier for each behavior segment; based on the segment identifier, aggregating all standardized data units in each behavior segment and generating a semantic summary describing the core activity of the behavior segment, associating and storing the semantic summary with the corresponding behavior segment, and finally outputting an initial behavior sequence with traceable labels.

3. The digital economy risk intelligent management evaluation method according to claim 2, characterized in that, The method comprises the following steps: constructing a parallel action analysis channel, a pattern recognition channel, and an intent inference channel, and synchronously inputting the initial behavior sequence with traceable labels into the parallel action analysis channel, the pattern recognition channel, and the intent inference channel; In the parallel action parsing channel, each standardized data unit in the initial behavior sequence is parsed for atomic operation, specific operation instructions, operation objects and operation results represented by the data unit are identified, the atomic operations are linked in time sequence to form an operation chain feature expression; In the pattern recognition channel, the initial behavior sequence is scanned in a sliding window manner, statistical rules of the appearance of standardized data units in numerical value, frequency or combination are detected within the sliding window, periodic patterns, trend patterns or abnormal fluctuation patterns are identified to form a statistical pattern feature expression; In the intention inference channel, a causal reasoning network of the behavior sequence is constructed based on the semantic summary of the initial behavior sequence and the context relationship of the behavior fragments, possible behavior intentions and targets are inferred in the causal reasoning network to form an intention network feature expression; The feature expressions of different abstraction levels are cross-fused to generate a composite feature map, specifically: a feature cross-attention mechanism is established to calculate attention weights between the operation chain feature expression, the statistical pattern feature expression and the intention network feature expression, the operation chain feature expression, the statistical pattern feature expression and the intention network feature expression are weighted fused and information complementary according to the calculated attention weights to generate a multi-layer feature structure integrating micro operations, medium patterns and macro intentions, i.e. the composite feature map.

4. The digital economic risk intelligent management evaluation method according to claim 3, characterized in that, The composite feature map is mapped into the risk knowledge graph according to a preset risk knowledge graph, and risk nodes and risk paths matching the composite feature map are found in the topological structure of the risk knowledge graph, specifically including: The risk knowledge graph includes risk entity nodes, risk relationship edges and attribute vectors of the nodes, the composite feature map is vectorized and encoded to obtain a vector representation of the composite feature map; The semantic similarity between the vector representation of the composite feature map and the attribute vector of each risk entity node in the risk knowledge graph is calculated, and the risk entity nodes with a semantic similarity exceeding a preset threshold are selected as a candidate risk node set; Starting from the candidate risk node set, walk along the risk relationship edges in the topological structure of the risk knowledge graph for a limited number of steps to explore all other risk entity nodes that can be reached from the candidate risk node set, and record the risk relationship edge sequence passed through, which constitutes a potential risk path; The matching degree of each potential risk path with the composite feature map is evaluated, the matching degree evaluation is based on the cooperativeness of the attribute vectors of all nodes on the risk path and the vector representation of the composite feature map, and the degree of consistency of the risk transmission logic represented by the path and the intention network feature expression; According to the matching degree, the potential risk paths are sorted, and the potential risk paths with the highest matching degree and their associated risk entity nodes are selected as the risk nodes and risk paths matching the composite feature map.

5. The digital economic risk intelligent management evaluation method according to claim 4, characterized in that, The preliminary risk situation map is generated based on the matched risk nodes and risk paths, specifically including: The matched risk node is taken as a graph node, and the matched risk path is taken as a directed edge to construct an initial risk topology subgraph; Feature intensity information corresponding to each matched risk node is extracted from the composite feature map, the feature intensity information is quantified as an initial energy value of the risk node, and the corresponding graph node is given; The path type and path weight are extracted from the matched risk path, and the path type and path weight are given to the corresponding directed edge in the initial risk topology subgraph, wherein the path weight reflects the strength or probability of risk conduction along the path; The initial risk topology subgraph, the initial energy value of the graph node, the path type of the directed edge, and the path weight are integrated to form a visual network graph containing node attributes, edge attributes, and a topology structure, that is, the preliminary risk situation map.

6. The digital economy risk intelligent management evaluation method according to claim 1, characterized in that, The dynamic calculation sand table is established, the preliminary risk situation map is imported into the dynamic calculation sand table, a plurality of preset environmental disturbance factors are injected, the evolution trajectory of the preliminary risk situation map under a plurality of preset environmental disturbances is simulated, and the specific process includes the following steps: The dynamic calculation sand table is a simulation environment containing node state updating rules, edge conduction rules, and environmental interaction rules, the graph node, the directed edge, and the attributes thereof in the preliminary risk situation map are imported into the dynamic calculation sand table, and the sand table state is initialized; A plurality of groups of environmental disturbance factors are defined, each group of environmental disturbance factors simulates a change of an external economic environment or market condition, and the change includes a policy impact, a market liquidity mutation, an industry chain fluctuation, or a technology disruptive event; Each group of environmental disturbance factors is sequentially injected into the initialized dynamic calculation sand table, the node state updating rules, the edge conduction rules, and the environmental interaction rules are used to drive the energy value of the graph node in the preliminary risk situation map to conduct, superimpose, dissipate, or enhance along the directed edge, and the sand table state after each simulation time step is iteratively calculated; The change process of all graph node energy values and the change process of the activation state of the directed edge from the start of injecting the environmental disturbance factor to the end of the sand table state reaching stability or the preset simulation time length are recorded, and the change processes constitute an evolution trajectory; The process is repeatedly executed until all preset environmental disturbance factor groups are simulated, so that a plurality of evolution trajectories corresponding to the plurality of groups of environmental disturbance factors are obtained.

7. The digital economic risk intelligent management evaluation method according to claim 6, characterized in that, The end state of each evolution trajectory is recorded to form a risk evolution set, and the specific process includes the following steps: For each evolution trajectory, the dynamic calculation sand table state at the end of simulation is extracted, and the sand table state includes the final energy value of all graph nodes, the final activation state of all directed edges, and the topological energy distribution characteristics of the entire sand table; The sand table state, the injected environmental disturbance factor type, and the unique identifier of the evolution trajectory are packaged to form a risk evolution sample; All risk evolution samples are collected to constitute the risk evolution set, wherein each sample represents a final risk pattern to which the preliminary risk situation map may evolve under a specific environmental disturbance.

8. The digital economic risk intelligent management evaluation method according to claim 7, characterized in that, The credibility of each end state in the risk evolution set is evaluated and the state is converged, and the specific process includes the following steps: extracting sandtable states of all risk evolution samples from the risk evolution set, comparing final energy values of graph nodes in all sandtable states across samples, and calculating statistical distribution of final energy values of each graph node in different samples; for each graph node, checking consistency of final energy values of the graph node in different samples, and if final energy values of the graph node in more than a preset proportion of samples are in a similar value range, considering that the graph node state has stability consensus in multiple environments; checking similarity between topological energy distribution characteristics of different risk evolution samples, classifying samples with high topological energy distribution characteristic similarity, and forming several state clusters, each state cluster representing a type of final risk pattern with commonality; performing feature fusion on sandtable states of all samples in each state cluster to generate a consensus state template representing the state cluster, the consensus state template including stable node energy values and typical topological structure in the state cluster.

9. The digital economic risk intelligent management evaluation method according to claim 8, characterized in that, The method further includes the following steps of: setting a consensus condition threshold, the consensus condition threshold including a node state stability proportion threshold and a state cluster sample quantity proportion threshold; selecting state clusters in which the graph node state satisfies the node state stability proportion threshold, and further selecting state clusters in which a proportion of sample quantities in the state cluster to a total sample quantity of the risk evolution set exceeds the state cluster sample quantity proportion threshold, and marking the state clusters as stable risk states satisfying the consensus condition; extracting a consensus state template corresponding to each marked stable risk state, analyzing a core risk type represented by a high-energy-value graph node in the consensus state template and a high-risk energy conduction path, and generating a descriptive conclusion on risk exposure, risk correlation and risk stability of the target object, i.e., the deterministic risk assessment conclusion. 10.A digital economy risk intelligent management evaluation system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor, when executing the computer program, implements the steps of the digital economic risk intelligent management evaluation method according to any one of claims 1 to 9.

Images