A standard digital modeling and verification method and system based on artificial intelligence

Abstract

The present invention discloses a standard digital modeling and verification method and system based on artificial intelligence. The method obtains initial multi-source heterogeneous data through a multi-source API interface; uses the BERT model to perform semantic recognition on text data and establish a dynamic knowledge graph; processes the structural characteristics of the dynamic knowledge graph based on the GNN graph neural network, uses the Transformer to capture long-range text dependencies, inputs the dynamic knowledge graph into a hybrid neural network model for recognition, and uses the Monte Carlo Dropout method to quantify the prediction confidence; selects the best modeling paradigm based on the recognition result to establish a digital model, and uses the trained GAN generative adversarial network to generate test cases for the digital model to detect model blind spots. The method fully explores the correlation relationship and semantic information between the data, so that the constructed digital model can more realistically reflect the actual scene, greatly improving the accuracy and reliability of the model.

Description

Technical Field

[0001] The present invention relates to the field of standard digital technology, and in particular to an artificial intelligence-based standard digital modeling and verification method and system. Background Art

[0002] Standardized digital modeling and verification play a vital role in numerous fields, including industrial manufacturing, smart cities, and healthcare. Traditional digital modeling methods often rely on a single type of data, such as text or image data. This makes it difficult to fully reflect the true state of complex systems, resulting in significant discrepancies between the model and the actual scenario. Furthermore, traditional data processing methods lack compatibility with heterogeneous data from multiple sources. When processing disparate data types such as text, images, videos, and sensors, they lack effective integration and collaborative processing mechanisms, and the correlations between data are not fully explored and utilized. Summary of the Invention

[0003] The purpose of the present invention is to solve the above problems and to design a standard digital modeling and verification method and system based on artificial intelligence.

[0004] The technical solution of the present invention to achieve the above object is that, further, in the above-mentioned standard digital modeling and verification method based on artificial intelligence, the standard digital modeling and verification method includes the following steps:

[0005] Through the multi-source API interface multi-source heterogeneous data, data preprocessing is performed on the multi-source heterogeneous data to obtain initial multi-source heterogeneous data;

[0006] Performing spatial semantic alignment on the image and video data in the initial multi-source heterogeneous data using an image recognition algorithm, performing semantic recognition on the text data using a BERT model, and establishing a dynamic knowledge graph;

[0007] Based on the GNN graph neural network, the structural features of the dynamic knowledge graph are processed, and the long-range text dependency is captured using Transformer to obtain a GNN-Transformer hybrid neural network model;

[0008] Inputting the dynamic knowledge graph into the GNN-Transformer hybrid neural network model for recognition, and using the Monte Carlo Dropout method to quantify the prediction confidence to obtain the recognition result;

[0009] Based on the recognition results, the optimal modeling paradigm is selected to establish a digital model, and the trained GAN generative adversarial network is used to generate test cases for the digital model to detect model blind spots.

[0010] Furthermore, in the above-mentioned standard digital modeling and verification method based on artificial intelligence, the data preprocessing of the multi-source heterogeneous data to obtain initial multi-source heterogeneous data includes:

[0011] The multi-source heterogeneous data includes at least national standard data, industry standard data, group standard data, enterprise standard data, text data, image data, video data and sensor data;

[0012] Extracting quantitative data from the national standard data, industry standard data, group standard data, and enterprise standard data, and using a word segmentation tool to perform word segmentation, part-of-speech tagging, and stop word filtering on the text data in the multi-source heterogeneous data; unifying the text length to obtain unified text data;

[0013] Adjusting images of different resolutions in the image data to a uniform size, performing noise reduction using a median filter algorithm, and performing enhancement processing on the noise-reduced image data to obtain noise-reduced image data;

[0014] Performing frame processing on the video data in the multi-source heterogeneous data, and performing noise reduction and enhancement processing on the framed image data in sequence to obtain framed video data;

[0015] The sensor data is denoised using the Kalman filter algorithm, and the missing data of the sensor is filled in to obtain complete sensor data.

[0016] Furthermore, in the above-mentioned standard digital modeling and verification method based on artificial intelligence, the image and video data in the initial multi-source heterogeneous data are spatially semantically aligned according to the image recognition algorithm, and the text data is semantically recognized using the BERT model to establish a dynamic knowledge graph, including:

[0017] Use the ResNet network to extract features from images and video frames in the initial multi-source heterogeneous data to obtain the visual feature vector of the image;

[0018] The STN spatial transformation network is used to perform geometric transformation on images, spatially align images at different perspectives and scales, and use the optical flow method to calculate the motion information between adjacent frames of video data. The spatial semantic alignment effect is optimized by combining time series features.

[0019] In the process of spatial semantic alignment, a mapping relationship between the position, shape and color features of objects in images and videos and semantic concepts is established.

[0020] Furthermore, in the above-mentioned standard digital modeling and verification method based on artificial intelligence, the image and video data in the initial multi-source heterogeneous data are spatially semantically aligned according to the image recognition algorithm, the text data is semantically recognized using the BERT model, and a dynamic knowledge graph is established, which also includes:

[0021] Based on the mapping relationship after spatial semantic alignment and the entities and relationships in the semantically recognized text data, a dynamic knowledge graph is constructed.

[0022] Use entity linking technology to match entities in text with nodes in the knowledge graph;

[0023] The nodes of a dynamic knowledge graph include objects, people, places, events, categories, and attributes, and the edges represent the relationships between entities.

[0024] Furthermore, in the above-mentioned standard digital modeling and verification method based on artificial intelligence, the structural features of the dynamic knowledge graph are processed based on the GNN graph neural network, and the long-range text dependency is captured by Transformer to obtain the GNN-Transformer hybrid neural network model, which also includes:

[0025] Based on the GCN graph convolutional network as the basic model of GNN, it processes the structural characteristics of dynamic knowledge graphs and uses the Transformer encoder to capture long-range dependencies in text data;

[0026] The node feature matrix output by GCN and the text feature vector output by Transformer are concatenated as the input of the hybrid neural network model. Feature transformation and dimension unification are performed through the fully connected layer to obtain a hybrid feature representation.

[0027] Furthermore, in the above-mentioned standard digital modeling and verification method based on artificial intelligence, the dynamic knowledge graph is input into the GNN-Transformer hybrid neural network model for recognition, and the Monte Carlo Dropout method is used to quantify the prediction confidence to obtain the recognition result, which also includes:

[0028] Use the Monte Carlo Dropout method to perform multiple forward propagations, maintain the randomness of Dropout in each propagation, and obtain multiple recognition results;

[0029] The confidence of the prediction is quantified by calculating the variance of the multiple recognition results, wherein the smaller the variance, the higher the confidence, and the recognition result with high confidence is output.

[0030] Furthermore, in the above-mentioned standard digital modeling and verification method based on artificial intelligence, the method of selecting the best modeling paradigm based on the recognition result to establish a digital model, and using the trained GAN generative adversarial network to generate test cases for the digital model to detect model blind spots also includes:

[0031] Build a generator and discriminator network, use the generator to generate test cases close to the real data distribution, and use the discriminator to distinguish the generated test cases from the real data;

[0032] The generator uses a deep neural network, with random noise vectors as input, and generates test cases through layer-by-layer transformation. The discriminator uses a deep neural network, with real data or generated test cases as input, and the output is the judgment result.

[0033] The generated test cases are input into the established digital model, and the model's performance on these test cases is analyzed to detect whether the model has blind spots.

[0034] Furthermore, in an artificial intelligence-based standard digital modeling and verification system, the standard digital modeling and verification system includes the following modules:

[0035] A multi-source data acquisition module is used to obtain multi-source heterogeneous data through a multi-source API interface, perform data preprocessing on the multi-source heterogeneous data, and obtain initial multi-source heterogeneous data;

[0036] A knowledge graph building module is used to perform spatial semantic alignment on the image and video data in the initial multi-source heterogeneous data according to an image recognition algorithm, perform semantic recognition on the text data using a BERT model, and build a dynamic knowledge graph;

[0037] A recognition model building module is used to process the structural features of the dynamic knowledge graph based on the GNN graph neural network, use the Transformer to capture long-range text dependencies, and obtain a GNN-Transformer hybrid neural network model;

[0038] A recognition result obtaining module is used to input the dynamic knowledge graph into the GNN-Transformer hybrid neural network model for recognition, and use the Monte Carlo Dropout method to quantify the prediction confidence to obtain the recognition result;

[0039] The modeling and verification module is used to select the best modeling paradigm based on the recognition results to establish a digital model, and use the trained GAN generative adversarial network to generate test cases for the digital model to detect model blind spots.

[0040] Furthermore, in an artificial intelligence-based standard digital modeling and verification system, the knowledge graph building module includes the following submodules:

[0041] The extraction submodule is used to extract features from images and video frames in the initial multi-source heterogeneous data using the ResNet network to obtain the visual feature vector of the image;

[0042] The optimization submodule is used to perform geometric transformation on images through the STN spatial transformation network, spatially align images at different perspectives and scales, calculate the motion information between adjacent frames of video data using the optical flow method, and optimize the spatial semantic alignment effect by combining time series features;

[0043] The mapping submodule is used to establish the mapping relationship between the position, shape and color features of objects in images and videos and semantic concepts during the spatial semantic alignment process.

[0044] Furthermore, in a standard digital modeling and verification system based on artificial intelligence, the knowledge graph establishment module also includes the following submodules:

[0045] Construct a submodule to build a dynamic knowledge graph based on the mapping relationship after spatial semantic alignment and the entities and relationships in the text data after semantic recognition;

[0046] The matching submodule is used to match entities in the text with nodes in the knowledge graph using entity linking technology;

[0047] The setting submodule is used to set the nodes of the dynamic knowledge graph, including objects, people, places, events, categories and attributes, and the edges represent the relationships between entities.

[0048] Its beneficial effects are that text data, image data, video data and sensor data are obtained through a multi-source API interface to obtain multi-source heterogeneous data, data preprocessing is performed on the multi-source heterogeneous data to obtain initial multi-source heterogeneous data; spatial semantic alignment is performed on the image and video data in the initial multi-source heterogeneous data according to the image recognition algorithm, and semantic recognition of text data is performed using the BERT model to establish a dynamic knowledge graph; the structural features of the dynamic knowledge graph are processed based on the GNN graph neural network, and the long-range text dependency is captured using the Transformer to obtain a GNN-Transformer hybrid neural network model; the dynamic knowledge graph is input into the GNN-Transformer hybrid neural network model for recognition, and the prediction confidence is quantified using the Monte Carlo Dropout method to obtain the recognition result; based on the recognition result, the best modeling paradigm is selected to establish a digital model, and the trained GAN generative adversarial network is used to generate test cases for the digital model to detect model blind spots. 1. It fully explores the correlation relationship and semantic information between data, so that the constructed digital model can more realistically reflect the actual scene, greatly improving the accuracy and reliability of the model. 2. It can clearly judge the prediction accuracy of the model in different situations, so that decision makers can make decisions more scientifically based on the model output and avoid risks caused by model uncertainty. 3. It can evaluate and optimize the model more efficiently and in-depth, ensuring the stability and accuracy of the digital model in various scenarios, and effectively improving the quality and usability of the model. 4. The present invention not only solves the problem of digital modeling of complex cross-domain systems by establishing a complete technology chain covering "multi-source data integration-cross-modal semantic alignment-dynamic knowledge graph construction-hybrid neural network modeling-uncertainty quantification-adversarial testing and verification", but also provides explainable and verifiable technical support for intelligent decision-making in key areas such as industry, cities, and medical care by quantifying model credibility and exposing model blind spots. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Various other advantages and benefits will become apparent to those skilled in the art by reading the following detailed description of the preferred embodiment.The accompanying drawings are only for the purpose of illustrating the preferred embodiment and are not to be considered as limiting the present invention.

[0050] Figure 1 This is a schematic diagram of a first embodiment of a standard digital modeling and verification method based on artificial intelligence in an embodiment of the present invention;

[0051] Figure 2 Schematic diagram of a second embodiment of a standard digital modeling and verification method based on artificial intelligence in an embodiment of the present invention;

[0052] Figure 3This is a schematic diagram of a first embodiment of an artificial intelligence-based standard digital modeling and verification system in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0054] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a," "an," and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0055] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 As shown, a standard digital modeling and verification method based on artificial intelligence, the standard digital modeling and verification method includes the following steps:

[0056] Step 101: Acquire multi-source heterogeneous data through a multi-source API interface, perform data preprocessing on the multi-source heterogeneous data, and obtain initial multi-source heterogeneous data;

[0057] Specifically, the multi-source heterogeneous data in this embodiment includes at least national standard data, industry standard data, group standard data, enterprise standard data, text data, image data, video data and sensor data;

[0058] Extracting quantitative data from the national standard data, industry standard data, group standard data, and enterprise standard data, and using word segmentation tools to perform word segmentation, part-of-speech tagging, and stop word filtering on text data in multi-source heterogeneous data; unifying the text length to obtain unified text data;

[0059] Adjusting images of different resolutions in the image data to a uniform size, performing noise reduction using a median filter algorithm, and performing enhancement processing on the noise-reduced image data to obtain noise-reduced image data;

[0060] Performing frame processing on the video data in the multi-source heterogeneous data, and performing noise reduction and enhancement processing on the framed image data in sequence to obtain framed video data;

[0061] The sensor data is denoised using the Kalman filter algorithm, and the missing data of the sensor is filled in to obtain complete sensor data.

[0062] Specifically,

[0063] 1. Multi-source API interface design

[0064] An API interface layer was built to support access to multiple data types, using a RESTful API architecture to ensure compatibility with diverse data sources. For text data, access was provided via web crawler APIs (encapsulated using the Scrapy framework), document parsing APIs (supporting formats like PDF and Word), and social media APIs (such as the Weibo and WeChat open platform APIs) to retrieve webpage text, document content, and user comments. For image data, access was provided via camera device APIs (compliant with protocols like RTSP and HTTP), image database APIs (accessible via COCO and ImageNet datasets), and remote sensing imagery APIs (from satellite remote sensing data platforms). Video data was acquired through monitoring system APIs, video platform APIs, and local video file parsing interfaces. Sensor data was adapted to the communication protocols of various sensor devices (such as temperature sensors, pressure sensors, and accelerometers), with corresponding APIs developed to support data collection via serial ports, Bluetooth, and Wi-Fi.

[0065] 2. Multi-source heterogeneous data preprocessing

[0066] 1. Text data preprocessing: First, perform data cleaning to remove duplicate text, special symbols, invalid spaces, etc.; then use natural language processing technology to perform word segmentation (using the Jieba word segmentation tool), part-of-speech tagging, and stop word filtering (based on the Harbin Institute of Technology stop word list); for inconsistent text lengths, use padding or truncation to unify the text to a fixed length for subsequent processing.

[0067] 2. Image data preprocessing: Perform image noise reduction using methods such as median filtering and Gaussian filtering. Perform size normalization to adjust images of different resolutions to a uniform size (224x224). Grayscale images can be converted into RGB three-channel images as needed. At the same time, to enhance the generalization ability of the model, data augmentation operations such as random rotation, flipping, and scaling can be performed.

[0068] 3. Video data preprocessing: Decompose the video into single frames and process each frame using the same preprocessing method as image data. At the same time, consider the time series information of the video and extract key frames (using a key frame extraction algorithm based on motion estimation) to reduce the amount of data while retaining the main content of the video.

[0069] 4. Sensor data preprocessing: Filtering algorithms (Kalman filtering, mean filtering) are used to remove noise that may exist in sensor data. Interpolation methods (linear interpolation, polynomial interpolation) are used to fill in missing data. Sensor data is normalized and mapped to the range [0, 1] or [-1, 1] to eliminate the impact of different sensor data dimensions.

[0070] Step 102: Perform spatial semantic alignment on the image and video data in the initial multi-source heterogeneous data using an image recognition algorithm, perform semantic recognition on the text data using a BERT model, and establish a dynamic knowledge graph.

[0071] Specifically, in this embodiment, the ResNet network is used to extract features from images and video frames in the initial multi-source heterogeneous data to obtain visual feature vectors of the images;

[0072] The STN spatial transformation network is used to perform geometric transformation on images, spatially align images at different perspectives and scales, and use the optical flow method to calculate the motion information between adjacent frames of video data. The spatial semantic alignment effect is optimized by combining time series features.

[0073] In the process of spatial semantic alignment, a mapping relationship between the position, shape and color features of objects in images and videos and semantic concepts is established.

[0074] Based on the mapping relationship after spatial semantic alignment and the entities and relationships in the semantically recognized text data, a dynamic knowledge graph is constructed.

[0075] Use entity linking technology to match entities in text with nodes in the knowledge graph;

[0076] The nodes of a dynamic knowledge graph include objects, people, places, events, categories, and attributes, and the edges represent the relationships between entities.

[0077] Specifically,

[0078] 1. Spatial Semantic Alignment of Image and Video Data

[0079] Spatial semantic alignment is performed using an image recognition algorithm based on a convolutional neural network (CNN). Using ResNet as the underlying network, features are extracted from images and video frames in the initial multi-source heterogeneous data to obtain the image's visual feature vector. A spatial transformer network (STN) performs geometric transformations on the images to achieve spatial alignment of images at different perspectives and scales. For video data, optical flow is used to calculate motion information between adjacent frames, and combined with time series features to further optimize the spatial semantic alignment. During the spatial semantic alignment process, a mapping relationship is established between spatial features such as the position, shape, and color of objects in images and videos and semantic concepts, providing a spatial semantic foundation for the subsequent construction of a knowledge graph.

[0080] 2. Semantic Recognition of Text Data

[0081] The BERT model is used to perform semantic recognition on text data. The BERT model is first pre-trained on a large-scale general text corpus (Wikipedia, BooksCorpus) to learn deep semantic representations of text. For specific tasks, a fine-tuning approach is used: pre-processed text data is fed into the pre-trained BERT model. Task-specific output layers (classification layer, sequence annotation layer) are added to fine-tune the model to adapt it to the current semantic recognition task. The BERT model captures contextual dependencies in text and generates word vector representations rich in semantic information, providing support for semantic parsing of text data.

[0082] 3. Dynamic Knowledge Graph Construction

[0083] A dynamic knowledge graph is constructed based on concepts such as objects and scenes in spatially semantically aligned image and video data, as well as entities and relationships in semantically recognized text data. The knowledge graph's nodes include entities (objects, people, places, events, etc.) and concepts (categories, attributes, etc.), while edges represent relationships between entities (belonging to, containing, associated with, etc.). The knowledge graph is stored and managed using a graph database (Neo4j). Newly acquired multi-source heterogeneous data is continuously integrated into the knowledge graph through a real-time data update interface, enabling dynamic updates. During the construction process, entity linking techniques (rule-based entity linking and machine learning-based entity linking) are used to match entities in the text with nodes in the knowledge graph to ensure knowledge consistency and accuracy.

[0084] Step 103: Process the structural features of the dynamic knowledge graph based on the GNN graph neural network, use the Transformer to capture long-range text dependencies, and obtain a GNN-Transformer hybrid neural network model;

[0085] Specifically, this embodiment uses the GCN graph convolutional network as the basic model of GNN to process the structural characteristics of the dynamic knowledge graph and uses the Transformer encoder to capture the long-range dependencies in text data;

[0086] The node feature matrix output by GCN and the text feature vector output by Transformer are concatenated as the input of the hybrid neural network model. Feature transformation and dimension unification are performed through the fully connected layer to obtain a hybrid feature representation.

[0087] Specifically,

[0088] 1. GNN graph neural network processing structure characteristics

[0089] The graph convolutional network (GCN) is selected as the basic model of GNN to process the structural characteristics of dynamic knowledge graphs. GCN updates the representation of nodes by aggregating the feature information of node neighbors, thereby capturing the structural relationship between nodes in the knowledge graph. For each node in the knowledge graph, its initial feature vector is composed of the feature vectors obtained in the process of spatial semantic alignment and semantic recognition. During the propagation process of GCN, the hidden layer representation of the node is calculated by defining the adjacency matrix of the graph and the node feature matrix.

[0090] 2. Transformer captures long-range text dependencies

[0091] The Transformer encoder is used to capture long-range dependencies in text data. The Transformer's self-attention mechanism allows the model to pay attention to the information of all other words in the text when processing each word, thereby effectively capturing long-range semantic dependencies. For the word vector sequence of text data, position encoding is first performed to incorporate the word position information into the vector representation. Then, through the multi-head self-attention mechanism, the attention scores under different heads are calculated to obtain the multi-head attention output.

[0092] 3. Hybrid Neural Network Model Architecture

[0093] The structural features of the knowledge graph obtained through GNN processing are fused with the text features obtained through Transformer processing. This fusion can be achieved through concatenation, addition, or weighted averaging. For example, the node feature matrix output by GCN and the text feature vector output by Transformer are concatenated and used as the input of the hybrid neural network model. Feature transformation and dimensionality unification are then performed through a fully connected layer to obtain the final hybrid feature representation. The hybrid neural network model can fully utilize the structural information of the knowledge graph and the semantic dependencies of the text data, improving the model's recognition and reasoning capabilities.

[0094] Step 104: Input the dynamic knowledge graph into the GNN-Transformer hybrid neural network model for recognition, and use the Monte Carlo Dropout method to quantify the prediction confidence to obtain the recognition result;

[0095] Specifically, in this embodiment, the Monte Carlo Dropout method is used to perform multiple forward propagations, and the randomness of Dropout is maintained in each propagation to obtain multiple recognition results;

[0096] The confidence of the prediction is quantified by calculating the variance of multiple recognition results. The smaller the variance, the higher the confidence, and the recognition result with high confidence is output.

[0097] Specifically,

[0098] 1. Dynamic Knowledge Graph Input and Model Recognition

[0099] The constructed dynamic knowledge graph is converted into a format suitable for input to the hybrid neural network model. Feature vectors are extracted from the nodes and edges in the knowledge graph to form node and edge feature matrices. The node feature matrices are then input into the GNN-Transformer hybrid neural network model. After processing by the GNN and Transformer, as well as feature fusion and transformations in the fully connected layer, the model output is obtained, i.e., the recognition results of the nodes or relationships in the knowledge graph.

[0100] 2. Monte Carlo Dropout Method to Quantify Prediction Confidence

[0101] During the training of the hybrid neural network model, a dropout layer is introduced to randomly inactivate a portion of neurons to prevent overfitting. During the prediction phase, the Monte Carlo dropout method is used, performing multiple forward propagations (T times), maintaining the randomness of the dropout layer in each propagation, to obtain multiple prediction results. The confidence of the predictions is quantified by calculating the variance or entropy of these prediction results. Specifically, for each predicted category, the frequency of that category appearing in the T predictions is calculated. A higher frequency indicates a higher prediction confidence. The variance of the prediction results is also calculated. A smaller variance indicates a higher confidence level. This confidence quantification can be used to assess the reliability of the model's predictions and provide a basis for subsequent decision-making.

[0102] Step 105: Select the best modeling paradigm based on the recognition results to establish a digital model, and use the trained GAN generative adversarial network to generate test cases for the digital model to detect model blind spots.

[0103] Specifically, in this embodiment, a generator and a discriminator network are constructed, the generator is used to generate test cases close to the real data distribution, and the discriminator is used to distinguish the generated test cases from the real data;

[0104] The generator uses a deep neural network, with random noise vectors as input, and generates test cases through layer-by-layer transformation. The discriminator uses a deep neural network, with real data or generated test cases as input, and the output is the judgment result.

[0105] The generated test cases are input into the established digital model, and the model's performance on these test cases is analyzed to detect whether the model has blind spots.

[0106] Specifically,

[0107] 1. Optimal Modeling Paradigm Selection

[0108] Based on the model identification results, the optimal modeling paradigm is selected by comprehensively considering data characteristics, task requirements, and model performance. Common modeling paradigms include rule-based models, statistical learning models, and deep learning models. For scenarios with clear rules and logic, rule-based models are suitable; for scenarios with large data volumes and complex nonlinear relationships, deep learning models (hybrid neural network models) are preferred. During the selection process, the performance indicators (accuracy, recall, F1 score, etc.) of different modeling paradigms are evaluated through methods such as cross-validation, and the modeling paradigm with the best performance is selected to establish the digital model.

[0109] 2. GAN Generative Adversarial Network Training and Test Case Generation

[0110] GAN generative adversarial network training: Build a generator and a discriminator network. The generator's goal is to generate test cases that approximate the distribution of real data, while the discriminator's goal is to accurately distinguish generated test cases from real data. The generator uses a deep neural network (multi-layer perceptron, convolutional neural network) as input, and generates test cases through layer-by-layer transformation. The discriminator also uses a deep neural network, with input being real data or generated test cases, and outputting a judgment result (real or fake). During training, the generator and discriminator perform adversarial training, minimizing the loss function (cross-entropy loss function) to make the test cases generated by the generator increasingly resemble real data, while the discriminator's discriminative ability becomes increasingly stronger.

[0111] Test case generation and model blind spot detection: Leveraging the trained GAN generative adversarial network, a large number of test cases are generated, covering a wide range of possible input scenarios, particularly edge cases and extreme scenarios. These generated test cases are then fed into the established digital model, and the model output is observed. By analyzing the model's performance on these test cases, the model is tested for blind spots (such as misclassifications and missed detections). For any detected model blind spots, further analysis is conducted, potentially leading to insufficient data preprocessing, an inappropriate model architecture, or insufficient training data. Improvements and optimizations are then implemented to address these issues, improving the reliability and accuracy of the digital model.

[0112] Its beneficial effects are that it obtains text data, image data, video data and sensor data through a multi-source API interface to obtain multi-source heterogeneous data, performs data preprocessing on the multi-source heterogeneous data to obtain initial multi-source heterogeneous data; performs spatial semantic alignment on the image and video data in the initial multi-source heterogeneous data according to the image recognition algorithm, uses the BERT model to perform semantic recognition on the text data, and establishes a dynamic knowledge graph; processes the structural features of the dynamic knowledge graph based on the GNN graph neural network, uses the Transformer to capture long-range text dependencies, and obtains a GNN-Transformer hybrid neural network model; inputs the dynamic knowledge graph into the GNN-Transformer hybrid neural network model for recognition, uses the Monte Carlo Dropout method to quantify the prediction confidence, and obtains the recognition result; selects the best modeling paradigm based on the recognition result to establish a digital model, and uses the trained GAN generative adversarial network to generate test cases for the digital model to detect model blind spots. 1. It fully explores the correlation and semantic information between the data, so that the constructed digital model can more realistically reflect the actual scene, greatly improving the accuracy and reliability of the model. 2. It can clearly judge the prediction accuracy of the model in different situations, so that decision makers can make decisions more scientifically based on the model output and avoid risks caused by model uncertainty. 3. It can evaluate and optimize the model more efficiently and in-depth, ensuring the stability and accuracy of the digital model in various scenarios, and effectively improving the quality and usability of the model. 4. The present invention not only solves the problem of digital modeling of complex cross-domain systems by establishing a complete technology chain covering "multi-source data integration-cross-modal semantic alignment-dynamic knowledge graph construction-hybrid neural network modeling-uncertainty quantification-adversarial testing and verification", but also provides explainable and verifiable technical support for intelligent decision-making in key areas such as industry, cities, and medical care by quantifying model credibility and exposing model blind spots.

[0113] Specifically, the application scenarios of the present invention in the field of financial risk control are as follows:

[0114] Applied to the retail credit risk control scenarios of commercial banks, it builds a full-process digital risk control system covering pre-loan risk assessment, dynamic monitoring during the loan, and post-loan early warning for the personal consumer loan application process, focusing on solving core issues such as customer credit assessment, fraudulent transaction identification, and related risk transmission.

[0115] 2. Multi-source data acquisition and preprocessing

[0116] (1) Data Collection Layer

[0117] Integrate multi-dimensional data sources through standardized API interfaces:

[0118] Text data: Access the central bank's credit report API (personal credit records), tax system API (tax payment records), e-commerce platform API (consumption record text), and customer service conversation logs (application consultation texts), obtaining an average of 200,000 pieces of structured / unstructured text data per day.

[0119] Image data: Collects ID card OCR images (taken and uploaded via mobile client), business license scans (for business owner loans), and facial images (captured with liveness detection cameras), processing an average of 80,000 images per day.

[0120] Video data: Branch surveillance videos (abnormal behavior detection) and remote interview recordings (compliance review) are streamed in real time via the security system API, storing an average of 500GB of video data daily.

[0121] Sensor data: mobile phone GPS positioning data (transaction location verification), device fingerprint data (terminal unique identification), transaction terminal sensor data (accelerometer / gyroscope abnormal operation detection).

[0122] (2) Data preprocessing

[0123] Cleaning and desensitizing: Use financial-grade data cleaning tools to remove fields with a missing rate exceeding 30%. Use regular expressions to identify and encrypt sensitive information such as ID card numbers and bank card numbers, retaining masked features.

[0124] Unified format: Image data is uniformly scaled to 224x224 pixels, and video data is frame-processed (10 key frames are extracted per second). Text data is segmented and stop-word filtered, using a specialized vocabulary in the financial field (including professional terms such as "overdue," "bad debt," and "joint guarantee").

[0125] Anomaly detection: Based on the isolation forest algorithm, outlier data such as abnormal transaction amounts (single transaction exceeding 10 times the average monthly income) and device location jumps (cross-city positioning within 30 minutes) are identified and marked as suspicious samples for manual review.

[0126] 3. Semantic Alignment and Dynamic Knowledge Graph Construction

[0127] 1. Spatial Semantic Alignment

[0128] Image and video processing:

[0129] The YOLOv8 target detection algorithm is used to identify key areas in ID card and bank card images, and combined with PaddleOCR to extract text information, with an accuracy rate of 98.7%;

[0130] Perform action recognition on remote interview videos, use OpenPose to detect gesture trajectories (abnormal actions such as frequent camera blocking and multiple people in frame), and use a spatiotemporal feature fusion model (3D-CNN+LSTM) to determine interview compliance.

[0131] Geographic semantic association: Match GPS location data with the administrative division knowledge base to establish a "customer-frequently used address-risk area" association relationship, for example, marking counties with high overdue rates as risky geographic units.

[0132] (2) Text Semantic Recognition

[0133] Domain-adapted BERT model: BERT-base was fine-tuned based on financial domain data (500,000 credit reports and 300,000 customer service conversations), focusing on optimizing the semantic representation of professional terms such as "default probability" and "guarantee liability." The F1 value for credit text classification tasks was increased to 92%.

[0134] Entity Relationship Extraction: Use the BiLSTM+CRF model to extract entities (customer name, institution name, loan product) and relationships ("belongs to," "guaranteed," "associated") from text. For example, extract the relationship "Zhang San - Lending Institution - XX Bank" from a credit report.

[0135] (3) Dynamic Knowledge Graph Construction

[0136] Using the Neo4j graph database as a carrier, we built a risk control knowledge graph consisting of five types of entities (customers, institutions, accounts, transactions, and equipment) and eight types of relationships:

[0137] Customer entity: contains attributes such as age, occupation, and income;

[0138] Transaction entity: includes transaction time, amount, channel and other attributes;

[0139] Relationship types: "Customer-Owns-Account", "Account-Occurred-Transaction", "Device-Associated-Customer", etc.

[0140] Dynamic Update: Transaction data is accessed in real time through the Kafka message queue, and the graph node status is updated every 10 seconds (account balance changes trigger node attribute updates).

[0141] 4. GNN-Transformer Hybrid Model Construction and Risk Identification

[0142] (1) Model architecture design

[0143] GNN layer: Graph Attention Network (GAT) is used to process the structural features of the knowledge graph, calculate the neighbor risk index of the customer node (the node weight increases if the overdue rate of the associated customer is greater than 30%), and capture the risk transmission effect in the guarantee network and capital flow network.

[0144] Transformer layer: Converts transaction flow text (consumption records for the past 12 months) into sequence data, uses a multi-head self-attention mechanism to capture long-range dependencies, and identifies periodic overdue payments (three consecutive non-working days) and abnormal consumption patterns (high-frequency large-value transactions in the early morning).

[0145] Fusion mechanism: The structural feature vector output by GNN and the sequence feature vector output by Transformer are fused through the gating unit (GatingUnit) to output the customer risk feature vector (dimension 128).

[0146] (2) Model training and inference

[0147] Data labeling: Based on historical loan data, we label three categories: "normal customers," "customers of concern," and "defaulting customers." We use stratified sampling to address the class imbalance problem (the ratio of positive and negative samples is 1:4).

[0148] Training strategy: We used the AdamW optimizer with a learning rate of 1e-4 and early stopping to prevent overfitting (stopping if the validation set loss does not decrease for five consecutive rounds). Training took 8 hours on an NVIDIA A100 GPU.

[0149] Risk identification: The model inputs the knowledge graph subgraph (including neighboring nodes within 3 hops) and transaction text data of the customer to be evaluated. The model outputs the probability of default (between 0 and 1) and uses Monte Carlo Dropout to quantify uncertainty:

[0150] Each forward propagation randomly discards 50% of the neurons and repeats the reasoning 100 times;

[0151] The probability mean is calculated as the final prediction value, and the standard deviation is used as the confidence indicator (a standard deviation > 0.15 triggers manual review).

[0152] 5. Digital Model Verification and Optimization

[0153] 1. Modeling Paradigm Selection

[0154] Dynamically select modeling strategies based on risk identification results:

[0155] High confidence scenario (confidence > 0.9, probability of default > 0.7): Directly trigger risk warning and reject the loan application;

[0156] Medium confidence scenario (0.5 ≤ confidence ≤ 0.9): XGBoost model is used for secondary verification, combined with expert rules ("more than three overdue payments in the past six months and debt-to-income ratio > 70%) for comprehensive judgment;

[0157] Low confidence scenario (confidence < 0.5): Enter the manual review process and submit additional materials such as asset proof.

[0158] (2) GAN Generates Adversarial Verification

[0159] Adversarial sample generation: Train WassersteinGAN (WGAN-GP) to generate adversarial samples. The input is normal transaction data, and the output is a "critical sample" close to the default boundary:

[0160] Generator network: A multi-layer fully connected neural network that takes as input a noise vector (64 dimensions) and outputs a 100-dimensional feature vector (simulating transaction amounts, frequency, and other features).

[0161] Discriminator network: a binary classifier that distinguishes true default samples from generated samples, with a gradient penalty coefficient of 10.

[0162] Model blind spot detection: The 20,000 generated adversarial examples are fed into the risk control model to detect the following blind spots:

[0163] New fraud models: such as using false identity information to register multiple accounts to pool funds; rule bypass risks: such as circumventing single transaction limit monitoring by splitting transaction amounts;

[0164] Model iterative optimization: Add detected blind spot samples to the training dataset and update the GNN-Transformer model every quarter to achieve dynamic adaptation to emerging risks.

[0165] VI. Implementation Effect

[0166] Risk identification accuracy: The F1 value of default customer identification is improved, and the false rejection rate is reduced;

[0167] Processing efficiency: The time required for automated approval of individual loans is shortened, and the proportion of manual review is reduced;

[0168] Risk coverage: 23 related guarantee default events were discovered through the knowledge graph, providing early warning 15 days earlier than traditional rule engines.

[0169] See also Figure 2In a standard digital modeling and verification method based on artificial intelligence, data preprocessing is performed on multi-source heterogeneous data to obtain initial multi-source heterogeneous data, including the following steps:

[0170] Step 201: The multi-source heterogeneous data includes at least national standard data, industry standard data, group standard data, enterprise standard data, text data, image data, video data, and sensor data.

[0171] Step 202: extract quantitative data from national standard data, industry standard data, group standard data, and enterprise standard data, and use word segmentation tools to perform word segmentation, part-of-speech tagging, and stop word filtering on text data in multi-source heterogeneous data; unify the text length to obtain unified text data;

[0172] Step 203: resize images of different resolutions in the image data to a uniform size, perform noise reduction using a median filter algorithm, and perform enhancement processing on the noise-reduced image data to obtain noise-reduced image data;

[0173] Step 204: Frame the video data in the multi-source heterogeneous data, and perform noise reduction and enhancement processing on the framed image data in sequence to obtain framed video data;

[0174] Step 205: Perform noise reduction processing on the sensor data using a Kalman filter algorithm, and fill in the missing data of the sensor to obtain complete sensor data.

[0175] The above is an introduction to an embodiment of a standard digital modeling and verification method based on artificial intelligence of the present invention. Figure 3 In an artificial intelligence-based standard digital modeling and verification system, the standard digital modeling and verification system includes the following modules:

[0176] The multi-source data acquisition module is used to obtain text data, image data, video data and sensor data through the multi-source API interface to obtain multi-source heterogeneous data, perform data preprocessing on the multi-source heterogeneous data, and obtain initial multi-source heterogeneous data;

[0177] The knowledge graph building module is used to perform spatial semantic alignment of image and video data in the initial multi-source heterogeneous data based on the image recognition algorithm, perform semantic recognition on text data using the BERT model, and build a dynamic knowledge graph;

[0178] The recognition model building module is used to process the structural features of dynamic knowledge graphs based on the GNN graph neural network, using the Transformer to capture long-range text dependencies and obtain a GNN-Transformer hybrid neural network model;

[0179] The recognition result acquisition module is used to input the dynamic knowledge graph into the GNN-Transformer hybrid neural network model for recognition, and use the Monte Carlo Dropout method to quantify the prediction confidence to obtain the recognition result;

[0180] The modeling and verification module is used to select the best modeling paradigm based on the recognition results to establish a digital model, and use the trained GAN generative adversarial network to generate test cases for the digital model to detect model blind spots.

[0181] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A standard digital modeling and verification method based on artificial intelligence, characterized in that: The standard digital modeling and verification method includes the following steps: Acquire multi-source heterogeneous data through a multi-source API interface, perform data preprocessing on the multi-source heterogeneous data, and obtain initial multi-source heterogeneous data; Use the ResNet network to extract features from images and video frames in the initial multi-source heterogeneous data to obtain the visual feature vector of the image; The STN spatial transformation network is used to perform geometric transformation on images, spatially align images at different perspectives and scales, and use the optical flow method to calculate the motion information between adjacent frames of video data. The spatial semantic alignment effect is optimized by combining time series features. In the process of spatial semantic alignment, the mapping relationship between the position, shape and color features of objects in images and videos and semantic concepts is established; Based on the mapping relationship after spatial semantic alignment and the entities and relationships in the semantically recognized text data, a dynamic knowledge graph is constructed. Use entity linking technology to match entities in text with nodes in the knowledge graph; The nodes of a dynamic knowledge graph include objects, people, places, events, categories, and attributes, and the edges represent the relationships between entities; Based on the GNN graph neural network, the structural features of the dynamic knowledge graph are processed, and the long-range text dependency is captured using Transformer to obtain a GNN-Transformer hybrid neural network model; Inputting the dynamic knowledge graph into the GNN-Transformer hybrid neural network model for recognition, and using the Monte Carlo Dropout method to quantify the prediction confidence to obtain the recognition result; Based on the recognition results, the optimal modeling paradigm is selected to establish a digital model, and the trained GAN generative adversarial network is used to generate test cases for the digital model to detect model blind spots.

2. The standard digital modeling and verification method based on artificial intelligence according to claim 1, characterized in that: The performing data preprocessing on the multi-source heterogeneous data to obtain initial multi-source heterogeneous data includes: The multi-source heterogeneous data includes text data, image data, video data and sensor data, and a word segmentation tool is used to perform word segmentation, part-of-speech tagging and stop word filtering on the text data in the multi-source heterogeneous data; the text length is unified to obtain unified text data; Adjusting images of different resolutions in the image data to a uniform size, performing noise reduction using a median filter algorithm, and performing enhancement processing on the noise-reduced image data to obtain noise-reduced image data; Performing frame processing on the video data in the multi-source heterogeneous data, and performing noise reduction and enhancement processing on the framed image data in sequence to obtain framed video data; The sensor data is denoised using the Kalman filter algorithm, and the missing data of the sensor is filled in to obtain complete sensor data.

3. The standard digital modeling and verification method based on artificial intelligence according to claim 1, characterized in that: The GNN-based graph neural network processes the structural features of the dynamic knowledge graph, uses Transformer to capture long-range text dependencies, and obtains a GNN-Transformer hybrid neural network model, which also includes: Based on the GCN graph convolutional network as the basic model of GNN, it processes the structural characteristics of dynamic knowledge graphs and uses the Transformer encoder to capture long-range dependencies in text data; The node feature matrix output by GCN and the text feature vector output by Transformer are concatenated as the input of the hybrid neural network model. Feature transformation and dimension unification are performed through the fully connected layer to obtain a hybrid feature representation.

4. The standard digital modeling and verification method based on artificial intelligence according to claim 1, characterized in that: The step of inputting the dynamic knowledge graph into the GNN-Transformer hybrid neural network model for recognition, quantifying the prediction confidence using the Monte Carlo Dropout method, and obtaining the recognition result further includes: Use the Monte Carlo Dropout method to perform multiple forward propagations, maintain the randomness of Dropout in each propagation, and obtain multiple recognition results; The confidence of the prediction is quantified by calculating the variance of the multiple recognition results, wherein the smaller the variance, the higher the confidence, and the recognition result with high confidence is output.

5. The standard digital modeling and verification method based on artificial intelligence according to claim 1, characterized in that: The method of selecting the best modeling paradigm based on the recognition results to establish a digital model, and using the trained GAN generative adversarial network to generate test cases for the digital model to detect model blind spots, further includes: Build a generator and discriminator network, use the generator to generate test cases close to the real data distribution, and use the discriminator to distinguish the generated test cases from the real data; The generator uses a deep neural network, with random noise vectors as input, and generates test cases through layer-by-layer transformation. The discriminator uses a deep neural network, with real data or generated test cases as input, and the output is the judgment result. The generated test cases are input into the established digital model, and the model's performance on these test cases is analyzed to detect whether the model has blind spots.

6. A standard digital modeling and verification system based on artificial intelligence, characterized in that: The standard digital modeling and verification system Includes the following modules: A multi-source data acquisition module is used to obtain multi-source heterogeneous data through a multi-source API interface, perform data preprocessing on the multi-source heterogeneous data, and obtain initial multi-source heterogeneous data; The extraction submodule is used to extract features from images and video frames in the initial multi-source heterogeneous data using the ResNet network to obtain the visual feature vector of the image; The optimization submodule is used to perform geometric transformation on images through the STN spatial transformation network, spatially align images at different perspectives and scales, calculate the motion information between adjacent frames of video data using the optical flow method, and optimize the spatial semantic alignment effect by combining time series features; The mapping submodule is used to establish the mapping relationship between the position, shape and color features of objects in images and videos and semantic concepts during the spatial semantic alignment process; Construct a submodule to build a dynamic knowledge graph based on the mapping relationship after spatial semantic alignment and the entities and relationships in the text data after semantic recognition; The matching submodule is used to match entities in the text with nodes in the knowledge graph using entity linking technology; The setting submodule is used to set the nodes of the dynamic knowledge graph, including objects, people, places, events, categories and attributes, and the edges represent the relationships between entities; A recognition model building module is used to process the structural features of the dynamic knowledge graph based on the GNN graph neural network, use the Transformer to capture long-range text dependencies, and obtain a GNN-Transformer hybrid neural network model; A recognition result obtaining module is used to input the dynamic knowledge graph into the GNN-Transformer hybrid neural network model for recognition, and use the Monte Carlo Dropout method to quantify the prediction confidence to obtain the recognition result; The modeling and verification module is used to select the best modeling paradigm based on the recognition results to establish a digital model, and use the trained GAN generative adversarial network to generate test cases for the digital model to detect model blind spots.

Images