Intelligent and inexpensive assessment and early warning method based on multi-dimensional dynamic risk map

Abstract

The invention provides a multi-dimensional dynamic risk map-based intelligent and inexpensive assessment early warning method, which comprises the following steps of: acquiring multi-modal data from an internal system and an external platform of an enterprise through an API (Application Program Interface), and providing input for subsequent data fusion; fusing the multi-modal data by adopting a multi-modal variational auto-encoder, dynamically adjusting the feature weight through online Bayesian updating, and generating unified potential representation and risk score; constructing a space-time diagram based on the potential representation and the risk score, generating space-time risk features through a space-time diagram neural network, constructing a dynamic knowledge graph, and predicting a risk causal path by adopting a structural causal model; and visualizing the space-time risk characteristics and the risk causal paths through a three-level data cockpit, generating a risk distribution thermodynamic diagram, and pushing early warning information including risk scores, the risk causal paths and the positions of the thermodynamic diagram through a mobile terminal application. The scheme can be used for realizing low-cost early warning and evaluation through multiple dimensions.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence and relates to an intelligent integrity evaluation and early warning method based on a multi-dimensional dynamic risk map. BACKGROUND

[0002] With the deepening of enterprise digital transformation and the increasing of anti-corruption supervision, integrity risk management is facing multiple challenges. Traditional methods rely on static rules or single data sources, which are difficult to capture the dynamics and concealment of corruption behavior and lack the ability to integrate multi-dimensional heterogeneous data. Existing system risk visualization is limited to simple statistical charts, lacking interpretability; static threshold design leads to frequent false positives or false negatives. An intelligent system is urgently needed to integrate multi-modal data, build a dynamic knowledge graph, and provide explainable early warning. SUMMARY

[0003] Therefore, the application provides an intelligent integrity evaluation and early warning method based on a multi-dimensional dynamic risk map, which comprises the following steps: Step S1: Collect multi-modal data from enterprise internal systems and external platforms through API interface to provide input for subsequent data fusion; Step S2: Fuse the multi-modal data using a multi-modal variational autoencoder, dynamically adjust the feature weights through online Bayesian update, generate a unified latent representation and risk score; Step S3: Build a spatio-temporal graph based on the latent representation and risk score, generate spatio-temporal risk features through a spatio-temporal graph neural network, build a dynamic knowledge graph, and predict risk causal paths using a structural causal model; Step S4: Visualize the spatio-temporal risk features and risk causal paths through a three-level data cockpit, generate a risk distribution heat map, and push early warning information including risk score, risk causal path, and heat map location through a mobile application.

[0004] In particular, the multi-modal data fusion in step S2 includes: Extract numerical features from structured data and semantic features from unstructured data to generate multi-dimensional feature representations; Use a multi-modal variational autoencoder to map the above features to a shared latent space, optimize the representation to capture cross-modal risk-related patterns; Dynamically adjust the feature weights based on new data to enhance the adaptability to emerging risk patterns; Generate a risk score based on the latent representation through a classification model to provide input for the spatio-temporal graph and dynamic knowledge graph construction in step S3.

[0005] In particular, the construction of the spatio-temporal graph and the generation of the spatio-temporal risk features in step S3 include: A spatio-temporal graph is constructed using the latent representation and risk score, which contains project nodes and extended features, and reflects spatio-temporal correlation through edges. The spatio-temporal graph is processed using a spatio-temporal graph neural network to extract spatial dependencies and capture temporal dynamics, generating spatio-temporal risk features as input for knowledge graph construction and causal reasoning.

[0006] In particular, the construction of a dynamic knowledge graph in step S3 includes: Entities are extracted from unstructured data using a context embedding method, and global label consistency is optimized; Based on context analysis, the relationship between entities is identified, and a structured relationship set is generated; With the collection of new data, the entity and relationship set is incrementally updated to construct a dynamic knowledge graph to support causal reasoning.

[0007] In particular, the prediction of risk causal paths in step S3 includes: Based on the dynamic knowledge graph and spatio-temporal risk features, a directed acyclic graph is constructed, mapping entities to nodes and relationships to edges; Conditional dependence and intervention effects are estimated in the graph to predict risk causal paths; High-confidence risk causal paths are identified and their causal strength is quantified.

[0008] In particular, step S5 is used to optimize the warning threshold and detect atypical risks, which includes: The spatio-temporal risk features are reduced in dimension to improve computational efficiency; The risk score threshold is optimized based on the trade-off between false positive rate and false negative rate to generate a refined warning threshold; Statistical and semantic anomalies in the risk features are detected to generate an abnormal project list to update the warning information.

[0009] In particular, the optimization of the risk score threshold in step S5 includes: The system state is defined according to the current threshold and the statistical properties of the risk features; The threshold is adjusted based on a cost function that combines the false positives and false negatives, and the optimal adjustment strategy is learned to generate a new threshold for anomaly detection.

[0010] In particular, the anomaly detection in step S5 includes: Statistical outliers are identified based on the reduced risk features to generate a statistical anomaly score; Semantic anomalies are identified through reconstruction error to generate a semantic anomaly score; The statistical and semantic anomaly scores are fused to identify abnormal projects and update the warning information.

[0011] In particular, step S6 is further included for constructing a digital twin sand table simulation model of integrity risk, and the step includes: Based on the multi-modal data, the latent representation and risk score, the spatio-temporal risk feature and the knowledge graph, a digital twin model is constructed to map the enterprise business process and risk feature; Different risk scenarios are simulated to predict the effect of intervention measures and generate risk propagation paths; An interactive heat map and risk causal path diagram are rendered to support variable adjustment and simulation execution, thereby supporting the early warning.

[0012] In particular, the risk scenario simulation in step S6 includes: generating a real risk scenario by using an adversarial simulation technique; predicting the effect of the intervention measures by using causal modeling and quantifying the change in risk score; providing a predictable decision reference for the sand table simulation, thereby enhancing the quality of the early warning information.

[0013] Beneficial effects: Through the technical scheme of the present application, efficient integration and fusion of internal and external multi-modal data of an enterprise can be achieved, structural and semantic features related to integrity risk can be comprehensively captured, and the comprehensiveness and accuracy of risk assessment can be improved.

[0014] Through the technical scheme of the present application, an integrity risk spatio-temporal graph that integrates temporal and spatial features is constructed, which can mine risk propagation paths from the geographical and temporal dependency relationships between multiple projects, and is helpful for early discovery of concentrated and correlated corruption risks.

[0015] Through the technical scheme of the present application, the dynamic knowledge graph can be incrementally updated in terms of entities and relationships according to the latest data, and can be automatically constructed in combination with context relationship recognition capability, thereby effectively improving the accuracy and timeliness of risk event tracing and causal analysis.

[0016] Through the technical scheme of the present application, the cause path of a risk event is modeled in an interpretable manner based on a structural causal model, which can mine risk chain links such as “abnormal payment → collusion → budget overruns”, and provide managers with an intervenable root cause path, thereby improving the pertinence of early warning actions.

[0017] Through the technical scheme of the present application, the early warning threshold of the risk score is dynamically adjusted by using a reinforcement learning strategy, thereby achieving an optimal balance between false positive rate and false negative rate, and significantly improving the response accuracy and adaptive ability of the early warning system in different risk scenarios.

[0018] Through the technical scheme of the present application, a multi-dimensional hybrid anomaly recognition mechanism is formed by combining statistical anomaly detection (such as isolation forest) and semantic anomaly detection (such as VAE reconstruction error), which can discover atypical risk projects that cannot be captured by traditional scoring models.

[0019] Through the technical scheme of the present application, a digital twin sand table simulation environment of integrity risk is constructed, which can visually deduce and test strategies for the integrity risk evolution trend of enterprises in key situations, and provide reliable support for strategic planning, education and training, and emergency drills. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A schematic diagram of an intelligent integrity assessment and early warning method based on a multi-dimensional dynamic risk map is shown in the present application. DETAILED DESCRIPTION

[0021] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0022] The method is shown in Figure 1 The closed-loop risk management process is formed through continuous steps, from data collection to early warning push, and each step is based on the output of the previous step to ensure smooth information flow. Taking the monitoring of 1000 projects (about 5 million records) of a state-owned enterprise as an example, the method is implemented.

[0023] Step 1: Multi-modal data collection Through API docking of enterprise ERP, CRM and external public opinion platform, structured data (such as transaction records, financial indicators), semi-structured data (such as approval logs) and unstructured data (such as employee feedback, public opinion text) are collected. In the embodiment, the system collects about 5 million records, covering multi-modal information of 1000 projects, and uses real-time and batch synchronization mechanism to ensure data freshness and integrity, providing reliable input for subsequent fusion modeling.

[0024] Step 2: Fusion of multi-modal data using multi-modal variational autoencoder, dynamic adjustment of feature weights through online Bayesian update, generation of risk score and potential representation based on XGBoost classifier; Based on the multi-modal data collected in step 1, the system extracts numerical features of structured data through fully connected layers, and generates embedding vectors through BERT and Transformer processing of unstructured text. A multi-modal variational autoencoder (VAE) is used to map the data to a shared latent space to generate a unified feature representation, and the loss function is:

[0025] Wherein: : input multi-modal data; : shared latent representation; : encoder distribution, approximate posterior; : decoder distribution, reconstruct the input; : standard normal prior; : reconstruction loss, ensure Preserve key information; : KL divergence, constraint Distribution approaches normality.

[0026] Information content: Multi-modal fusion: Fusion numerical features (e.g. contract amount), semantic features (e.g. "supplier collusion") and spatio-temporal features (e.g. project location, time variation).

[0027] Probability characteristics: Follow probability distribution, capture data uncertainty, enhance robustness.

[0028] Risk relevance: Encode patterns related to integrity risks, such as the combination of high contract amount and negative public opinion.

[0029] For example, 150 high-risk projects in East China Mean higher than low-risk projects in certain dimensions (e.g. abnormal transaction intensity), reflecting abnormal patterns.

[0030] Shared latent representation Is a low-dimensional vector representation generated by compressing and fusing multi-modal data (including structured, semi-structured, unstructured data) through variational autoencoder (VAE), encoding key features and semantic information related to integrity risks. Through VAE generation, combined with online Bayesian update, dynamically adjust weights, Has probability characteristics and adaptability, supports risk scoring, knowledge graph construction and early warning optimization.

[0031] The following is the generation process and information content: Input data: 5 million records, including structured data (e.g. transaction amount, contract records), semi-structured data (e.g. approval logs) and unstructured data (e.g. employee feedback text).

[0032] Feature extraction: structured data is extracted into a 32-dimensional feature vector through a 3-layer fully connected network. Unstructured data is processed through BERT (768-dimensional embedding) and Transformer, reduced to 32 dimensions.

[0033] Online Bayesian update and parameter meaning To ensure that the latent representation Adapt to new data over time, the system uses online Bayesian update to adjust feature weights, formula:

[0034] The following is the detailed meaning of each parameter in the formula: : In time , No. The weight of a feature or indicator represents its relative importance to risk assessment. For example, the weight of the contract amount. Weight of approval frequency The initial weights are set based on expert experience or historical data.

[0035] In time , No. The updated weights of each feature reflect risk patterns in the latest data. For example, if approval frequency is more correlated with new risk events, It may increase to 0.25.

[0036] In time Given the first Features and dataset Risk events The conditional probability (posterior probability). For example: Risk events, such as "project budget overruns" or "supplier collusion".

[0037] : No. Features such as "contract amount > 10 million" or "approval frequency > 2 times / day" are used to identify key characteristics.

[0038] Deadline The dataset (a subset of 5 million records).

[0039] meaning: Measuring features With risk The strength of the correlation. For example, if high contract amounts are often accompanied by budget overruns, then... Relatively high.

[0040] The sum of weighted conditional probabilities of all features is used for normalization to ensure that the sum of the weights is 1. Normalization prevents the weights from growing indefinitely and maintains model stability. The Bayesian update formula uses conditional probabilities... Quantifying the correlation between characteristics and risk, driving weights Evolution, ensuring Reflects the latest risk patterns.

[0041] Update mechanism: like High, indicating characteristics With risk Strong correlation, weight Increase.

[0042] like Low, weight Decrease.

[0043] For example, new data from East China shows that approval frequency is related to the risk of collusion, with the weight increasing from 0.2 to 0.25, while the weight of contract amount decreases from 0.3 to 0.28.

[0044] Dynamic nature: Weights are updated daily based on newly added data (such as 10 abnormal records), affecting the generation of the VAE encoder. This ensures that potential representations reflect the latest risk patterns.

[0045] Risk scoring: Based on the XGBoost classifier (parameters: tree depth 6, number of trees 200, learning rate 0.1), combined with adversarial training, a 0-1 risk score is generated, and the quantile is used to classify high, medium and low risk.

[0046] For example, in this embodiment: 1000 item ratings are generated, 150 item ratings are >0.7, and the accuracy rate is 90%.

[0047] Step 3: Construct a spatiotemporal graph based on the latent representation and risk score, generate spatiotemporal risk features through the spatiotemporal graph neural network, construct a dynamic knowledge graph, and use a structural causal model to predict risk causal paths; Objective: To construct a spatiotemporal graph using the latent representation and risk score, which includes project nodes and extended features, and reflects spatiotemporal relationships through edges; The spatiotemporal graph is processed using a spatiotemporal graph neural network to extract spatial dependencies and capture temporal dynamics, generating spatiotemporal risk features as input for knowledge graph construction and causal reasoning.

[0048] enter: Potential characteristics: Step 2 .

[0049] Risk rating: 0-1 rating for 1000 items.

[0050] Original data: 5 million records (including a subset of text).

[0051] deal with: Spatiotemporal risk feature generation: Input: Spatiotemporal graph (1000 project nodes, features include risk score and 5-dimensional extended features, such as contract amount), edges based on KNN (…). Generate the adjacency matrix. ,in The time series is a 12-day feature matrix. .

[0052] The Spatiotemporal Graph Neural Network (ST-GNN) consists of the following two parts: Graph Convolutional Networks (GCNs):

[0053] in, , It is an adjacency matrix. It is ReLU.

[0054] Gated Circulation Unit (GRU):

[0055] in, , For GCN output.

[0056] Output: 64-dimensional spatiotemporal features ,matrix .

[0057] Example: The average characteristic of projects in East China is 0.75, which is higher than the national average of 0.6.

[0058] Dynamic knowledge graph construction: Extracting entities from unstructured data using context embedding methods and optimizing global label consistency; Relationships between entities are identified based on context analysis, generating a structured set of relationships; As new data is collected, the entity and relation sets are incrementally updated to build a dynamic knowledge graph to support causal reasoning.

[0059] Objective: To extract entities and relationships from a subset of 5 million records, construct a dynamic knowledge graph that reflects corruption risk patterns and supports causal reasoning.

[0060] enter: Text subset: 500,000 records (employee feedback, approval notes, public opinion).

[0061] Potential characteristics: Step 2 .

[0062] Risk rating: 0-1 rating for 1000 items.

[0063] deal with: Entity Extraction (BERT-CRF): Principle: BERT generates 768-dimensional context embeddings, and CRF optimizes the global label sequence.

[0064] Input: 500,000 text entries, with a maximum sequence length of 512 after word segmentation, and BIO annotations (organization, project, event, etc.).

[0065] Model architecture: BERT: bert-base-chinese, 12-layer Transformer, 768-dimensional hidden state.

[0066] Fully connected layer: Dimensionality reduced to a 10-dimensional label space. , .

[0067] CRF: Transition Matrix Loss function:

[0068] Training: 5000 labeled samples (4000 for training, 1000 for validation), learning rate 2e-5, 3 rounds of fine-tuning.

[0069] Output: Approximately 2000 entities (e.g., “Supplier A”, “Project P023”).

[0070] Example: 1000 organizations and 800 items were extracted from East China, with an F1 score of 0.92.

[0071] Relation Extraction (RE-BERT): Principle: BERT encodes entity pairs and context, and the fully connected layer classifies the relationships.

[0072] Input: 100,000 candidate entity pairs (e.g., “Supplier A - Project P023”), with sequence labels as [E1]Supplier A[ / E1]...[E2]Project P023[ / E2].

[0073] Model architecture: BERT: bert-base-chinese, outputs CLS vectors and entity embeddings.

[0074] Entity representation: splicing .

[0075] Fully connected layer: Mapped to a 5-dimensional relation space. Softmax outputs the probability.

[0076] Loss function: Cross-entropy .

[0077] Training: 2000 labeled samples (1600 for training, 400 for validation), learning rate 3e-5, 2 rounds of fine-tuning.

[0078] Output: Approximately 500 relationships (e.g., “Supplier A - Collusion - Project P023”).

[0079] Example: Extracting 200 "collusion" relationships in East China, with an F1 score of 0.89.

[0080] Dynamically updated: Input: 10,000 new text entries per day, filter out 10 high-risk projects.

[0081] Processing: BERT-CRF extracts 10 new entities, RE-BERT extracts 15 new relations, and inserts them into Neo4j.

[0082] Neo4j storage: 2000 nodes, with attributes such as rating, and 500 edges, with attributes such as confidence.

[0083] Example: "Supplier C - Abnormal Transaction - Project P025" is updated daily, with a query delay of <0.2 seconds.

[0084] Output: Knowledge graph (2000 entities, 500 relations), supporting causal reasoning.

[0085] Example Results: The graph reflects abnormal patterns in East China (such as "Supplier A - Abnormal Payment - Project P023"), reveals 5 risk causal paths, and predicts that intervention can reduce risk by 20%.

[0086] Causal reasoning: Objective: To leverage knowledge graphs and spatiotemporal risk characteristics This allows us to uncover causal pathways of risks, predict the effects of interventions, and enhance the interpretability of risk management.

[0087] enter: Knowledge graph: 2000 entities (e.g., “Supplier A”) and 500 relationships (e.g., “collusion”).

[0088] Spatiotemporal characteristics: 1,000 projects.

[0089] Original data: 5 million records (a subset of 500,000 text records).

[0090] deal with: Cause-effect graph construction: Principle: Map the knowledge graph to a directed acyclic graph (DAG), with entities as nodes and relationships as edges.

[0091] Input: knowledge graph, .

[0092] Processing: KNN ( Add implicit edges and attributes (such as risk scores).

[0093] Example: Construct a DAG containing "Supplier A → Collusion → Project P023 → Budget Overrun".

[0094] Conditional probability estimation: Principle: Bayesian network estimation ,formula:

[0095] Processing: Maximum likelihood estimation calculation ,from estimate .

[0096] Example: , .

[0097] Path mining: Principle: Depth-First Search (DFS) is used to mine the path; computational intensity:

[0098] Processing: Filter high-confidence paths (>0.85) and limit their length to 3.

[0099] Example: Discover 5 paths, such as "abnormal payment → collusion → budget overrun" (intensity 0.792).

[0100] Intervention effect prediction: Principle: Do-Calculus calculation:

[0101] Processing: Simulation ,Compare change.

[0102] Example: Predicting and reducing abnormal payments, It decreased from 0.4 to 0.32 (a 20% reduction).

[0103] Dynamically updated: Input: 10,000 new text entries are added daily, and 10 entities and 15 relationships are updated.

[0104] Processing: Incrementally update the cause-effect graph and CPT, and verify path stability.

[0105] Example: Add "Supplier C - Abnormal Transaction - Project P025" and update the path.

[0106] Output: 5 risk causal paths (e.g., "abnormal payment → collusion → budget overrun"), intervention suggestions, input step 4.

[0107] Example Results: Covering 150 high-risk projects in East China, the predicted risk was reduced by 20%, the query latency was less than 0.2 seconds, and interpretability was enhanced.

[0108] Step 4: Risk Visualization and Early Warning Push Objective: To visualize risk distribution and push out explainable early warnings.

[0109] Input: Spatiotemporal features from step 3 Risk causal path deal with: Visualization: Rendering three-level heatmaps (national, provincial, and municipal) using OpenLayers, based on... Principal component generation risk distribution.

[0110] Warning push notifications: The mobile app "Integrity Index" pushes 10 alerts, including risk scores, risk causal paths (such as "abnormal payments → collusion"), and heat map locations.

[0111] Output: Interactive heatmap, 10 alerts.

[0112] Example Results: Covering high-risk projects in East China, with an alarm click-through rate of 80%, assisting management in decision-making.

[0113] In particular, this embodiment also includes a scheme to increase the detection accuracy in step S5, including the following implementation methods: Objective: To optimize early warning thresholds, detect atypical corruption risks (such as abnormal payments and collusion), and reduce false alarm and false negative rates.

[0114] Input: Spatiotemporal risk characteristics From step 3, we obtain 64-dimensional features of 1000 items, covering a 12-day time series.

[0115] Dynamic knowledge graph: 2000 entities (e.g., “Supplier A”) and 500 relationships (e.g., “collusion”).

[0116] Raw data: 5 million records (financial, approval, public opinion, etc.).

[0117] Initial threshold: 0.75 (based on risk score quantiles).

[0118] deal with: Data preprocessing: right Principal component analysis (PCA) was performed to reduce the dimension to 10, retaining 95% of the variance to ensure computational efficiency.

[0119] Standardize the original data, remove missing values ​​(fill with mean), and normalize to [0, 1].

[0120] Example: After dimensionality reduction of project data in East China, key features (such as approval frequency and contract amount) are retained. The system state is defined based on the statistical characteristics of the current threshold and risk features. Threshold adjustment is performed based on a weighted combination of the cost function for false positives and false negatives. The optimal adjustment strategy is learned, and a new threshold is generated for anomaly detection.

[0121] Deep Q-Network (DQN) Optimization Threshold: Objective: To dynamically adjust risk scoring thresholds to balance false positives (FP) and false negatives (FN).

[0122] enter: Historical FP / FN statistics, initial threshold 0.75.

[0123] DQN model: Status: Current threshold, Statistical characteristics (mean, variance), FP / FN count.

[0124] Action: Threshold adjustment (±0.01, range [0.5, 0.9]).

[0125] Reward function:

[0126] in, , (The penalty for underreporting is higher).

[0127] Network structure: 3-layer fully connected network (64-dimensional input, 128-dimensional hidden layer, 50 discrete actions output), learning rate 0.001, discount factor 0.9.

[0128] Training: 1000 rounds of experience replay, batch size 32, target network updated every 100 rounds.

[0129] Output: Optimized threshold (e.g., 0.68).

[0130] Example: The threshold for projects in East China was optimized from 0.75 to 0.68, the false alarm rate decreased from 8% to 5%, and the false negative rate decreased from 5% to 3%.

[0131] Hybrid anomaly detection: Identifying statistical outliers based on dimensionality-reduced risk features and generating statistical anomaly scores; Semantic anomalies are identified by reconstructing errors, and semantic anomaly scores are generated. By integrating statistical and semantic anomaly scoring, abnormal items are identified, and the warning information is updated.

[0132] Objective: To detect atypical risks by combining statistical and semantic methods.

[0133] Isolated forest detection: Input: 10-dimensional principal components (after PCA dimensionality reduction) ).

[0134] Parameters: Number of trees 100, subsampling size 256.

[0135] Processing: Isolate outliers based on feature distribution and output anomaly scores. .

[0136] Example: Detect 8 statistical anomalies (such as approval frequency > 2 times / day, score > 0.9).

[0137] Variational autoencoder (VAE) detection: Input: Spatiotemporal features .

[0138] Model: 3-layer fully connected network, 32-dimensional latent space, loss function:

[0139] in: Input features; : Reconstruct features; Encoder distribution; : Normal prior; Reconstruction error; KL divergence measures the difference in distribution.

[0140] Output: Anomaly Score .

[0141] Example: Detect 5 semantic anomalies (such as abnormal payment patterns, score > 2.5).

[0142] Mixed score:

[0143] in, The threshold of 0.7 is used to filter out abnormalities.

[0144] Example: Identify 10 atypical risks (such as supplier-related transactions, with a score > 0.7).

[0145] Output: Optimized threshold: 0.68.

[0146] List of abnormal projects: 10 projects (ID, score, description, such as "Project P023: Approval frequency 2 times / day, score 0.85").

[0147] Performance metrics: False alarm rate 5%, false negative rate 3%.

[0148] Update: Feedback the abnormal results and optimization thresholds to step 5, and update the warning information.

[0149] Example Results: The optimized alerts are more accurate, covering 10 abnormal projects in East China, and supporting management decision-making.

[0150] Specifically, this embodiment also includes step S6, which involves performing a corruption risk simulation based on a digital twin, according to existing risk scenarios. The details are as follows: Objective: To build a digital twin model of enterprise operations, create an interactive risk simulation, support "hypothesis-response-feedback" simulations, predict risk scenarios, and optimize intervention strategies.

[0151] enter: Multimodal data: 5 million records from step 1.

[0152] Potential characteristics: Step 2 Risk score.

[0153] Spatiotemporal risk characteristics: Step 3 .

[0154] Knowledge graph: Step 3: 2000 entities, 500 relationships, and 5 risk causal paths.

[0155] deal with: Digital twin model construction: Use Business Process Modeling and Annotation (BPMN) tools to standardize enterprise processes (such as approvals and procurement).

[0156] Map entity relationships (such as “Supplier A - Collusion - Project P023”) and risk characteristics (such as score 0.65) to the virtual environment.

[0157] Example: Constructing virtual operation scenarios for 1,000 projects in East China, including approval processes and supplier relationships.

[0158] Risk scenario simulation: Generative Adversarial Networks (GANs): Generators simulate risky scenarios (such as "the introduction of a new supplier"), and discriminators evaluate the realism. Loss function:

[0159] in, For generator, For discriminator, It is random noise. This is real data.

[0160] Structural causal model (SCM): Predicts the effects of the intervention:

[0161] in, High-risk events (such as budget overruns). For intervention variables (such as reducing abnormal payments). To obfuscate variables (such as project size).

[0162] Example: Simulating "simplified approval process + introduction of new supplier", the predicted risk score increased from 0.65 to 0.75 (an increase of 15%).

[0163] Interactive sand table simulation: Use OpenLayers or Unity to render 3D heatmaps and risk causal path diagrams, supporting user interaction (such as adjusting the approval frequency).

[0164] Provide a console that allows users to set variables (such as "number of suppliers + 10") and run simulations.

[0165] Example: Users in East China adjusted their approval processes, observed changes in risk scores, and identified five new causal paths of risk (such as "new supplier → collusion → budget overrun").

[0166] Feedback and optimization: Analyze the simulation results and generate intervention recommendations (such as "strengthen supplier vetting").

[0167] Use online learning to update the digital twin model and optimize subsequent simulations.

[0168] Example: It is recommended to standardize the approval process, predict and reduce the risk by 20%, and update the model weights.

[0169] Outputs: Interactive sand table view, intervention suggestions, and optimized digital twin model.

[0170] Example Results: Risk in East China was predicted to increase by 15%, employee risk awareness was improved by 30% through training, and optimized strategies reduced risk by 20%.

[0171] It should be understood that the processor in the embodiments of the present invention may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method embodiments can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor described above can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0172] It is understood that the memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (RM), programmable read-only memory (PRM), erasable programmable read-only memory (EPRM), electrically erasable programmable read-only memory (EEPRM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), Synchronous Link Dynamic Random Access Memory (SLDRAM), and Direct Rambus RAM (DRRAM). It should be noted that the memory used in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0173] It should be understood that the above-described memory is exemplary but not restrictive. For example, the memory in the embodiments of this application may also be static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct memory bus RAM (DRRAM), etc. That is to say, the memory in the embodiments of this application is intended to include, but is not limited to, these and any other suitable types of memory.

[0174] This application also provides a computer-readable storage medium for storing computer programs.

[0175] Optionally, the computer-readable storage medium can be applied to the terminal device in the embodiments of this application, and the computer program causes the computer to execute the corresponding processes implemented by the mobile terminal / terminal device in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0176] This application also provides a computer program product, including computer program instructions.

[0177] Optionally, the computer program product can be applied to the terminal device in the embodiments of this application, and the computer program instructions cause the computer to execute the corresponding processes implemented by the mobile terminal / terminal device in the various methods of the embodiments of this application. For the sake of brevity, they will not be described in detail here.

[0178] This application also provides a computer program.

[0179] Optionally, the computer program can be applied to the vehicle autonomous driving device in the embodiments of this application. When the computer program is run on a computer, it causes the computer to execute the corresponding processes implemented by the terminal device in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0180] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0181] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0182] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0183] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A smart integrity assessment and early warning method based on a multi-dimensional dynamic risk map, characterized in that, Includes the following steps: Step S1: Collect multimodal data from internal enterprise systems and external platforms via API interfaces to provide input for subsequent data fusion; Step S2: The multimodal data is fused using a multimodal variational autoencoder, and the feature weights are dynamically adjusted through online Bayesian updates to generate a unified latent representation and risk score; Step S3: Construct a spatiotemporal graph based on the latent representation and risk score, generate spatiotemporal risk features through a spatiotemporal graph neural network, construct a dynamic knowledge graph, and use a structural causal model to predict risk causal paths; Step S4: Visualize the spatiotemporal risk characteristics and risk causal paths through a three-level data cockpit, generate a risk distribution heat map, and push early warning information, including risk scores, risk causal paths, and heat map locations, through a mobile application.

2. The method as described in claim 1, characterized in that, The multimodal data fusion in step S2 includes: Numerical features are extracted from structured data, and semantic features are extracted from unstructured data to generate multidimensional feature representations; The above features are mapped to a shared latent space using a multimodal variational autoencoder, and the representation is optimized to capture risk-related patterns across modalities; The feature weights are dynamically adjusted based on newly added data to enhance the adaptability to emerging risk patterns. Risk scores are generated using a classification model based on latent representations, providing input for the construction of spatiotemporal graphs and dynamic knowledge graphs in step S3.

3. The method as described in claim 2, characterized in that, Step S3, which involves constructing the spatiotemporal graph and generating spatiotemporal risk features, includes: A spatiotemporal graph is constructed using the latent representation and risk score, which includes project nodes and extended features, and reflects spatiotemporal relationships through edges; The spatiotemporal graph is processed using a spatiotemporal graph neural network to extract spatial dependencies and capture temporal dynamics, generating spatiotemporal risk features as input for knowledge graph construction and causal reasoning.

4. The method according to any one of claims 2-3, characterized in that, The construction of the dynamic knowledge graph in step S3 includes: Entities are extracted from unstructured data using context embedding methods, and global label consistency is optimized. Relationships between entities are identified based on context analysis, generating a structured set of relationships; As new data is collected, the entity and relation sets are incrementally updated to build a dynamic knowledge graph to support causal reasoning.

5. The method as described in claim 1, characterized in that, The prediction of the risk causal path in step S3 includes: A directed acyclic graph is constructed based on dynamic knowledge graphs and spatiotemporal risk features, mapping entities to nodes and relationships to edges; Estimate the conditional dependencies and intervention effects in the graph, and predict the causal path of risk. Identify high-confidence risk causal paths and quantify their causal strength.

6. The method as described in claim 1, characterized in that, It also includes step S5, which optimizes the warning threshold and detects atypical risks, the steps of which include: The spatiotemporal risk features are dimensionality reduced to improve computational efficiency; Based on a trade-off between false alarm rate and false negative rate, the risk scoring threshold is optimized to generate a refined early warning threshold; Detect statistical and semantic anomalies in risk characteristics, and generate a list of abnormal items to update the warning information.

7. The method as described in claim 6, characterized in that, Step S5 involves optimizing the risk scoring threshold, including: The system state is defined based on the statistical characteristics of the current threshold and risk features; Threshold adjustment is performed based on a weighted combination of the cost function for false positives and false negatives. The optimal adjustment strategy is learned, and a new threshold is generated for anomaly detection.

8. The method as described in claim 6, characterized in that, The anomaly detection in step S5 includes: Statistical outliers are identified based on the dimensionality-reduced risk features, and statistical anomaly scores are generated. Semantic anomalies are identified by reconstructing errors, and semantic anomaly scores are generated. By integrating statistical and semantic anomaly scoring, abnormal items are identified, and the warning information is updated.

9. The method as described in claim 1, characterized in that, It also includes step S6, which is used to construct a digital twin simulation model of integrity risk, the step of which includes: Based on the multimodal data, the potential representation and risk score, the spatiotemporal risk features and knowledge graph, a digital twin model is constructed to map enterprise business processes and risk features; Simulate different risk scenarios, predict the effects of intervention measures, and generate risk propagation paths; The system renders interactive heatmaps and risk causal path diagrams, supports variable adjustment and simulation execution, and provides support for the aforementioned early warning.

10. The method as described in claim 9, characterized in that, The risk scenario simulation in step S6 includes: generating realistic risk scenarios using adversarial simulation technology; predicting the effects of the intervention measures using causal modeling and quantifying changes in risk scores; providing predictable decision-making references for sand table simulation and enhancing the quality of the early warning information.

Images