Standard digital modeling and verification method and system based on artificial intelligence

Abstract

The invention discloses a standard digital modeling and verification method and system based on artificial intelligence. Initial multi-source heterogeneous data are obtained through a multi-source API interface; performing semantic recognition on the text data by utilizing a BERT model, and establishing a dynamic knowledge graph; processing structural features of the dynamic knowledge graph based on a GNN graph neural network, capturing long-process text dependence by utilizing Transform, inputting the dynamic knowledge graph into a mixed neural network model for identification, and quantizing and predicting confidence by utilizing a Monte Carlo Dropout method; and based on the identification result, selecting an optimal model establishment model to establish a digital model, generating a test case for the digital model by using the trained GAN generative adversarial network, and detecting model blind spots. The association relationship and semantic information between the data are fully mined, so that the constructed digital model can reflect an actual scene more truly, and the accuracy and reliability of the model are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of standard digitization, and particularly to a method and system for standard digitization modeling and verification based on artificial intelligence. Background Art

[0002] Standard digitization modeling and verification play a crucial role in many fields such as industrial manufacturing, smart cities, and healthcare. Traditional digitization modeling methods often rely on a single type of data, such as only using text data or image data, which is difficult to comprehensively reflect the real situation of complex systems, resulting in a large deviation between the model and the actual scenario. Moreover, traditional data processing methods have poor compatibility with multi-source heterogeneous data. When processing different types of data such as text, images, videos, and sensors, there is a lack of effective integration and collaborative processing mechanisms, and the correlation information between data cannot be fully mined and utilized. Summary of the Invention

[0003] The purpose of the present invention is to solve the above problems, and a method and system for standard digitization modeling and verification based on artificial intelligence are designed.

[0004] Furthermore, in the above method for standard digitization modeling and verification based on artificial intelligence, the standard digitization modeling and verification method includes the following steps: Collect multi-source heterogeneous data through multi-source API interfaces, and perform data preprocessing on the multi-source heterogeneous data to obtain initial multi-source heterogeneous data; Perform spatial semantic alignment on the image and video data in the initial multi-source heterogeneous data according to the image recognition algorithm, and use the BERT model to perform semantic recognition on the text data to establish a dynamic knowledge graph; Process the structural features of the dynamic knowledge graph based on the GNN graph neural network, and use Transformer to capture long-range text dependencies to obtain a GNN-Transformer hybrid neural network model; 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; Select the best modeling paradigm based on the recognition result 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.

[0005] Furthermore, in the above method for standard digitization modeling and verification based on artificial intelligence, the performing data preprocessing on the multi-source heterogeneous data to obtain initial multi-source heterogeneous data includes: The multi-source heterogeneous data at least includes national standard data, industry standard data, group standard data, enterprise standard data, text data, image data, video data, and sensor data; Extract the quantitative data from the national standard data, industry standard data, group standard data, and enterprise standard data, and use 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; unify the text length to obtain unified text data; Adjust the images with different resolutions in the image data to a unified size, and use the median filtering algorithm for noise reduction. Perform enhancement processing on the denoised image data to obtain denoised image data; Perform frame splitting on the video data in the multi-source heterogeneous data, and sequentially perform noise reduction and enhancement processing on the frame-split image data to obtain frame-split video data; Perform noise reduction processing on the sensor data through the Kalman filtering algorithm, and fill in the missing data of the sensor to obtain complete sensor data.

[0006] Furthermore, in the above method for standard digital modeling and verification based on artificial intelligence, the spatial semantic alignment of the image and video data in the initial multi-source heterogeneous data according to the image recognition algorithm, and the semantic recognition of the text data using the BERT model to establish a dynamic knowledge graph includes: Use the ResNet network to extract features from the images and video frames in the initial multi-source heterogeneous data to obtain the visual feature vectors of the images; Perform geometric transformation on the images through the STN spatial transformation network, perform spatial alignment on the images at different perspectives and scales, calculate the motion information between adjacent frames of the video data using the optical flow method, and optimize the spatial semantic alignment effect by combining time series features; Establish a mapping relationship between the features of the position, shape, and color of the objects in the images and videos and the semantic concepts during the spatial semantic alignment process.

[0007] Furthermore, in the above method for standard digital modeling and verification based on artificial intelligence, the spatial semantic alignment of the image and video data in the initial multi-source heterogeneous data according to the image recognition algorithm, and the semantic recognition of the text data using the BERT model to establish a dynamic knowledge graph also includes: Construct a dynamic knowledge graph based on the mapping relationship after spatial semantic alignment, combined with the entities and relationships in the text data after semantic recognition; Use entity linking technology to match the entities in the text with the nodes in the knowledge graph; The nodes of the dynamic knowledge graph include objects, people, places, events, categories, and attributes, and the edges represent the relationships between entities.

[0008] Furthermore, in the above-mentioned method for standard digital modeling and verification based on artificial intelligence, the structural features of the dynamic knowledge graph are processed by the GNN graph neural network, and the long-range text dependencies are captured by Transformer, resulting in a GNN-Transformer hybrid neural network model. It further includes: Using the GCN graph convolutional network as the basic model of the GNN to process the structural features of the dynamic knowledge graph, and using the Transformer encoder to capture the long-range dependencies in the text data; Concatenate the node feature matrix output by the GCN and the text feature vector output by the Transformer as the input of the hybrid neural network model, and perform feature transformation and dimension unification through a fully connected layer to obtain a hybrid feature representation.

[0009] Furthermore, in the above-mentioned method for standard digital modeling and verification based on artificial intelligence, when the dynamic knowledge graph is input into the GNN-Transformer hybrid neural network model for recognition, the Monte Carlo Dropout method is used to quantify the prediction confidence to obtain the recognition result. It further includes: Using the Monte Carlo Dropout method to perform multiple forward propagations, and maintaining the randomness of Dropout in each propagation to obtain multiple recognition results; Quantify the prediction confidence by calculating the variance of the multiple recognition results. The smaller the variance, the higher the confidence, and output the recognition result with high confidence.

[0010] Furthermore, in the above-mentioned method for standard digital modeling and verification based on artificial intelligence, 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. It further includes: Construct a generator and a discriminator network. Use the generator to generate test cases close to the real data distribution, and use the discriminator to distinguish between the generated test cases and the real data; The generator uses a deep neural network, with a random noise vector as the 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 the input, and the output is a judgment result; Input the generated test cases into the established digital model, and detect whether there are blind spots in the model by analyzing the performance of the model on these test cases.

[0011] Furthermore, in a standard digital modeling and verification system based on artificial intelligence, the standard digital modeling and verification system includes the following modules: A multi-source data acquisition module, which is used to obtain multi-source heterogeneous data through multi-source API interfaces, perform data preprocessing on the multi-source heterogeneous data, and obtain initial multi-source heterogeneous data; A knowledge graph construction module, which 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 by using a BERT model, and construct a dynamic knowledge graph; An identification model construction module, which is used to process the structural features of the dynamic knowledge graph based on a GNN graph neural network, and capture long-range text dependencies by using a Transformer to obtain a GNN-Transformer hybrid neural network model; An identification result obtaining module, which is used to input the dynamic knowledge graph into the GNN-Transformer hybrid neural network model for identification, and use the Monte Carlo Dropout method to quantify the prediction confidence to obtain an identification result; A modeling and verification module, which is used to select an optimal modeling paradigm based on the identification result to establish a digital model, and use a trained GAN generative adversarial network to generate test cases for the digital model to detect model blind spots.

[0012] Furthermore, in a standard digital modeling and verification system based on artificial intelligence, the knowledge graph construction module includes the following sub-modules: An extraction sub-module, which is used to extract features of images and video frames in the initial multi-source heterogeneous data by using a ResNet network to obtain visual feature vectors of the images; An optimization sub-module, which is used to perform geometric transformation on images through an STN spatial transformation network, perform spatial alignment on images at different perspectives and scales, calculate the motion information between adjacent frames of video data by using an optical flow method, and optimize the spatial semantic alignment effect in combination with time series features; A mapping sub-module, which is used to establish a mapping relationship between the features of the position, shape, and color of objects in images and videos and semantic concepts during the spatial semantic alignment process.

[0013] Furthermore, in a standard digital modeling and verification system based on artificial intelligence, the knowledge graph construction module also includes the following sub-modules: A construction sub-module, which is used to construct a dynamic knowledge graph based on the mapping relationship after spatial semantic alignment and in combination with the entities and relationships in the text data after semantic recognition; A matching sub-module, which is used to match the entities in the text with the nodes in the knowledge graph by using entity linking technology; A setting sub-module, which is used to set that the nodes of the dynamic knowledge graph include objects, people, places, events, categories, and attributes, and the edges represent the relationships between entities.

[0014] Its beneficial effects are as follows: text data, image data, video data, and sensor data are obtained through multi-source API interfaces to obtain multi-source heterogeneous data, and the multi-source heterogeneous data is preprocessed to obtain initial multi-source heterogeneous data; the images and video data in the initial multi-source heterogeneous data are subjected to spatial semantic alignment according to image recognition algorithms, and the BERT model is used to perform semantic recognition on the text data 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 Transformer is used to capture long-range text dependencies 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 Monte Carlo Dropout method is used to quantify the prediction confidence to obtain a 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. The correlation relationships and semantic information between data are fully exploited, enabling the constructed digital model to more realistically reflect the actual scenario and greatly improving the accuracy and reliability of the model. 2. It is possible to clearly judge the prediction accuracy of the model under different conditions, enabling decision-makers to make decisions more scientifically based on the model output and avoiding risks brought by model uncertainty. 3. The model can be evaluated and optimized more efficiently and deeply, ensuring the stability and accuracy of the digital model in various scenarios and effectively improving the quality and usability of the model. 4. By establishing a complete technical chain covering "multi-source data integration-cross-modal semantic alignment-dynamic knowledge graph construction-hybrid neural network modeling-uncertainty quantification-adversarial test verification", the present invention not only solves the problem of digital modeling of cross-domain complex systems, but also provides interpretable and verifiable technical support for intelligent decision-making in key fields such as industry, city, and medical care by quantifying model credibility and exposing model blind spots. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] By reading the following detailed description of the preferred embodiments, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present invention.

[0016] Figure 1 Schematic diagram of the first embodiment of a standard digital modeling and verification method based on artificial intelligence in an embodiment of the present invention; Figure 2 Schematic diagram of the second embodiment of a standard digital modeling and verification method based on artificial intelligence in an embodiment of the present invention; Figure 3Schematic diagram of the first embodiment of a standard digital modeling and verification system based on artificial intelligence in the embodiments of the present invention. Detailed implementation manners

[0017] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, 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 used to limit the present invention.

[0018] Those skilled in the art of the present technology can understand that unless specifically 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 description of the present invention means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or their groups.

[0019] The present invention will be specifically described 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: Step 101: Obtain multi-source heterogeneous data through a multi-source API interface, perform data preprocessing on the multi-source heterogeneous data to obtain initial multi-source heterogeneous data; Specifically, in this embodiment, the multi-source heterogeneous data at least includes national standard data, industry standard data, group standard data, enterprise standard data, text data, image data, video data, and sensor data; Extract the quantitative data in the national standard data, industry standard data, group standard data, and enterprise standard data, and use 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; unify the text length to obtain unified text data; Adjust images with different resolutions in the image data to a unified size, perform noise reduction using a median filtering algorithm, and perform enhancement processing on the noise-reduced image data to obtain noise-reduced image data; Perform frame splitting on the video data in the multi-source heterogeneous data, and perform noise reduction and enhancement processing on the frame-split image data in sequence to obtain frame-split video data; Perform noise reduction processing on the sensor data through a Kalman filtering algorithm, and fill in the missing data of the sensor to obtain complete sensor data.

[0020] Specifically, I. Design of multi-source API interface Build an API interface layer that supports the access of multiple data types, adopting the RESTful API architecture to ensure compatibility with different data sources. For text data, access the web crawler API (API encapsulated by the Scrapy framework), document parsing API (supporting formats such as PDF and Word), social media API (Weibo and WeChat open platform APIs), etc., to obtain web page text, document content, user comments, etc. For image data, access the camera device API (adhering to protocols such as RTSP and HTTP), image database API (access interfaces provided by COCO and ImageNet datasets), remote sensing image API (from satellite remote sensing data platforms), etc. Video data is obtained through the monitoring system API, video platform API, and local video file parsing interface. Sensor data is adapted to the communication protocols of various sensor devices (temperature sensors, pressure sensors, acceleration sensors, etc.), and corresponding API interfaces are developed to support data collection via serial ports, Bluetooth, Wi-Fi, etc.

[0021] II. Preprocessing of Multi-source Heterogeneous Data 1. Preprocessing of text data: First, perform data cleaning to remove duplicate text, special symbols, invalid spaces, etc.; then use natural language processing techniques for 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 the problem of inconsistent text lengths, use padding or truncation methods to unify the text to a fixed length for subsequent processing.

[0022] 2. Preprocessing of image data: Perform noise reduction on the image, and methods such as median filtering and Gaussian filtering can be used; perform size normalization to adjust images of different resolutions to a unified size (224x224); for grayscale images, they can be converted to RGB three-channel images according to needs; 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.

[0023] 3. Preprocessing of video data: Decompose the video into single-frame images and process each frame image using the same preprocessing method as image data; at the same time, considering the time series information of the video, extract key frames (using a key frame extraction algorithm based on motion estimation) to reduce the data volume while retaining the main content of the video.

[0024] 4. Preprocessing of sensor data: For the possible noise in sensor data, use filtering algorithms (Kalman filtering, mean filtering) for denoising; for missing data, use interpolation methods (linear interpolation, polynomial interpolation) for filling; perform normalization processing on sensor data to map the data to the interval [0,1] or [-1,1] to eliminate the influence of different sensor data dimensions.

[0025] Step 102: Perform spatial semantic alignment on the image and video data in the initial multi-source heterogeneous data according to the image recognition algorithm, use the BERT model to perform semantic recognition on the text data, and establish a dynamic knowledge graph; Specifically, in this embodiment, the ResNet network is used to extract features from the images and video frames in the initial multi-source heterogeneous data to obtain the visual feature vectors of the images; The STN spatial transformation network is used to perform geometric transformation on the images, perform spatial alignment on the images at different perspectives and scales, use the optical flow method to calculate the motion information between adjacent frames of the video data, and combine the time series features to optimize the spatial semantic alignment effect; In the process of spatial semantic alignment, establish the mapping relationship between the features of the position, shape, and color of the objects in the images and videos and the semantic concepts.

[0026] Based on the mapping relationship after spatial semantic alignment, combined with the entities and relationships in the text data after semantic recognition, construct a dynamic knowledge graph; Use entity linking technology to match the entities in the text with the nodes in the knowledge graph; The nodes of the dynamic knowledge graph include objects, people, places, events, categories, and attributes, and the edges represent the relationships between entities.

[0027] Specifically, I. Spatial semantic alignment of image and video data Use the image recognition algorithm based on the convolutional neural network (CNN) for spatial semantic alignment. Taking ResNet as the basic network, extract features from the images and video frames in the initial multi-source heterogeneous data to obtain the visual feature vectors of the images. Perform geometric transformation on the images through the spatial transformation network (STN) to achieve spatial alignment of the images at different perspectives and scales. For the video data, use the optical flow method to calculate the motion information between adjacent frames, and combine the time series features to further optimize the spatial semantic alignment effect. In the process of spatial semantic alignment, establish the mapping relationship between the spatial features such as the position, shape, and color of the objects in the images and videos and the semantic concepts, providing a spatial semantic basis for the subsequent construction of the knowledge graph.

[0028] II. Semantic recognition of text data Use the BERT model to perform semantic recognition on text data. First, pre-train the BERT model by training it on large-scale general text corpora (Wikipedia, BooksCorpus) to learn the deep semantic representations of the text. For specific tasks, adopt the fine-tuning method. Input the pre-processed text data into the pre-trained BERT model, and fine-tune the model by adding task-specific output layers (classification layer, sequence annotation layer) to make it adapt to the current semantic recognition task. The BERT model can capture the context dependencies in the text and generate word vector representations containing rich semantic information, providing support for the semantic parsing of text data.

[0029] III. Dynamic Knowledge Graph Construction Based on the concepts such as objects and scenes in the spatially semantically aligned image and video data, as well as entities and relationships in the text data after semantic recognition, construct a dynamic knowledge graph. The nodes of the knowledge graph include entities (objects, people, locations, events, etc.), concepts (categories, attributes, etc.), and the edges represent the relationships between entities (belonging to, containing, associating, etc.). Use a graph database (Neo4j) to store and manage the knowledge graph, and through a real-time data update interface, continuously integrate newly acquired multi-source heterogeneous data into the knowledge graph to achieve the dynamic update of the knowledge graph. During the construction process, adopt entity linking techniques (rule-based entity linking, machine learning-based entity linking) to match the entities in the text with the nodes in the knowledge graph to ensure the consistency and accuracy of the knowledge.

[0030] Step 103: Process the structural features of the dynamic knowledge graph based on the GNN graph neural network, and use Transformer to capture long-range text dependencies to obtain a GNN-Transformer hybrid neural network model; Specifically, in this embodiment, based on the GCN graph convolutional network as the basic model of the GNN, process the structural features of the dynamic knowledge graph, and use the Transformer encoder to capture the long-range dependency relationships in the text data; Concatenate the node feature matrix output by the GCN and the text feature vector output by the Transformer as the input of the hybrid neural network model, and perform feature transformation and dimension unification through a fully connected layer to obtain a hybrid feature representation.

[0031] Specifically, I. Processing Structural Features by the GNN Graph Neural Network The Graph Convolutional Network (GCN) is selected as the basic model of GNN to process the structural features of the dynamic knowledge graph. GCN updates the representation of nodes by aggregating the feature information of node neighbors, thereby capturing the structural relationships between nodes in the knowledge graph. For each node in the knowledge graph, its initial feature vector is composed of the feature vectors obtained from the spatial semantic alignment and semantic recognition processes. During the propagation process of GCN, by defining the adjacency matrix and node feature matrix of the graph, the hidden layer representation of the nodes is calculated. II. Transformer Captures Long-Range Text Dependencies The Transformer encoder is adopted to capture the long-range dependency relationships in the text data. The self-attention mechanism of Transformer enables the model to pay attention to the information of all other words in the text when processing each word, thus effectively capturing the long-distance semantic dependencies. For the sequence of word vectors of the text data, positional encoding is first performed to incorporate the position information of the words 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. III. Hybrid Neural Network Model Architecture The structural features of the knowledge graph processed by GNN and the text features processed by Transformer are fused. The fusion methods can include concatenation, addition, or weighted average, etc. For example, 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. Then, through the fully connected layer, feature transformation and dimension unification are performed to obtain the final hybrid feature representation. The hybrid neural network model can make full use of the structural information of the knowledge graph and the semantic dependency relationships of the text data to improve the recognition and reasoning capabilities of the model.

[0032] 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. Specifically, in this embodiment, the Monte Carlo Dropout method is used for multiple forward propagations, and the randomness of Dropout is maintained in each propagation to obtain multiple recognition results. 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.

[0033] Specifically, I. Input of Dynamic Knowledge Graph and Model Recognition Convert the constructed dynamic knowledge graph into a format suitable for input to the hybrid neural network model. For the nodes and edges in the knowledge graph, extract their feature vectors respectively to form a node feature matrix and an edge feature matrix. Input the node feature matrix into the GNN-Transformer hybrid neural network model. After being processed by GNN and Transformer, and through feature fusion and transformation by the fully connected layer, the output result of the model is obtained, that is, the recognition result of the nodes or relationships in the knowledge graph.

[0034] II. Monte Carlo Dropout method for quantifying prediction confidence During the training process of the hybrid neural network model, introduce a Dropout layer to randomly deactivate a part of the neurons to prevent overfitting. In the prediction stage, adopt the Monte Carlo Dropout method to perform multiple forward propagations (T times). Each propagation maintains the randomness of Dropout to obtain multiple prediction results. Quantify the prediction confidence by calculating the variance or entropy value of these prediction results. Specifically, for each prediction category, calculate the frequency of the appearance of this category in T predictions. The higher the frequency, the higher the prediction confidence. At the same time, calculate the variance of the prediction results. The smaller the variance, the higher the confidence. Through confidence quantification, the reliability of the model prediction can be judged, providing a basis for subsequent decisions.

[0035] Step 105: Select the best modeling paradigm based on the recognition result 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.

[0036] Specifically, in this embodiment, construct a generator and a discriminator network. Use the generator to generate test cases close to the real data distribution, and use the discriminator to distinguish between the generated test cases and the real data; The generator uses a deep neural network with a random noise vector as the 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 the input and the judgment result as the output; Input the generated test cases into the established digital model, and detect whether there are blind spots in the model by analyzing the performance of the model on these test cases.

[0037] Specifically, I. Selection of the best modeling paradigm Based on the model recognition results, comprehensively consider data characteristics, task requirements, and model performance to select the best modeling paradigm. Common modeling paradigms include rule-based models, statistical learning models, and deep learning models, etc. For scenarios with clear rules and logic, a rule-based model can be selected; for scenarios with a large amount of data and complex non-linear relationships, a deep learning model (hybrid neural network model) is preferred. During the selection process, evaluate the performance metrics (accuracy, recall rate, F1 value, etc.) of different modeling paradigms through methods such as cross-validation, and select the modeling paradigm with the optimal performance to establish a digital model.

[0038] II. GAN Generative Adversarial Network Training and Test Case Generation GAN Generative Adversarial Network Training: Construct a generator and a discriminator network. The goal of the generator is to generate test cases close to the real data distribution, and the goal of the discriminator is to accurately distinguish between the generated test cases and the real data. The generator uses a deep neural network (multi-layer perceptron, convolutional neural network), with a random noise vector as the input, and generates test cases through layer-by-layer transformation. The discriminator also uses a deep neural network, with real data or generated test cases as the input, and the output is a judgment result (real or fake). During the training process, the generator and the discriminator perform adversarial training. By minimizing the loss function (cross-entropy loss function), the test cases generated by the generator become closer and closer to the real data, and the discrimination ability of the discriminator becomes stronger and stronger.

[0039] Test Case Generation and Model Blind Spot Detection: Use the trained GAN generative adversarial network to generate a large number of test cases. These test cases cover various possible input situations, especially edge cases and extreme cases. Input the generated test cases into the established digital model and observe the output results of the model. By analyzing the performance of the model on these test cases, detect whether there are blind spots in the model (such as misclassification, missed judgment, etc.). For the detected model blind spots, further analyze the reasons, which may be insufficient data preprocessing, unreasonable model architecture, or insufficient training data, etc., and then improve and optimize the problem to improve the reliability and accuracy of the digital model.

[0040] The beneficial effects are as follows: obtaining text data, image data, video data, and sensor data through multi-source API interfaces to obtain multi-source heterogeneous data, performing data preprocessing on the multi-source heterogeneous data to obtain initial multi-source heterogeneous data; performing spatial semantic alignment on the images and video data in the initial multi-source heterogeneous data according to the image recognition algorithm, using the BERT model to perform semantic recognition on the text data, and establishing a dynamic knowledge graph; processing the structural features of the dynamic knowledge graph based on the GNN graph neural network, and using Transformer to capture long-range text dependencies 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; 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. 1. It fully excavates the correlation relationships and semantic information between data, enabling the constructed digital model to more realistically reflect the actual scenario, and greatly improving the accuracy and reliability of the model. 2. It can clearly judge the prediction accuracy of the model in different situations, enabling decision-makers to make decisions more scientifically based on the model output and avoiding risks brought by model uncertainty. 3. It can evaluate and optimize the model more efficiently and deeply, ensuring the stability and accuracy of the digital model in various scenarios, and effectively improving the quality and usability of the model. 4. By establishing a complete technical chain covering "multi-source data integration-cross-modal semantic alignment-dynamic knowledge graph construction-hybrid neural network modeling-uncertainty quantification-adversarial test verification", the present invention not only solves the problem of digital modeling of cross-domain complex systems, but also provides interpretable and verifiable technical support for intelligent decision-making in key fields such as industry, city, and healthcare by quantifying model credibility and exposing model blind spots.

[0041] Specifically, the application scenario of the present invention in the field of financial risk control is as follows: Applied to the retail credit risk control scenario of commercial banks, for the personal consumption loan application process, a full-process digital risk control system covering pre-loan risk assessment, in-loan dynamic monitoring, and post-loan early warning is constructed, focusing on solving core problems such as customer credit assessment, fraud transaction identification, and associated risk transmission. II. Multi-source Data Acquisition and Preprocessing (I) Data Acquisition Layer Integrate multi-dimensional data sources through standardized API interfaces: Text data: Connect to the central bank credit report API (personal credit record), tax system API (tax payment record), e-commerce platform API (consumption record text), and customer service conversation log (application consultation text), and obtain 200,000 structured / unstructured text data per day on average. Image data: Collect ID card OCR images (uploaded by taking pictures through the mobile client), scanned business license copies (for enterprise owner loan scenarios), and face images (captured by live detection cameras), with an average daily processing of 80,000 image data. Video data: Branch network surveillance videos (for abnormal behavior detection), remote video recordings of face-to-face signing (for compliance review), which are transmitted in real-time through the security system API, with an average daily storage of 500GB of video data. Sensor data: Mobile phone GPS location data (for transaction location verification), device fingerprint data (for terminal uniqueness identification), and transaction terminal sensor data (for detecting abnormal operations of accelerometers / gyroscopes). (2) Data preprocessing Cleaning and desensitization: Use financial-grade data cleaning tools to remove fields with a missing rate exceeding 30%, identify and encrypt sensitive information such as ID card numbers and bank card numbers through regular expressions, and retain the features after masking. Format unification: Uniformly scale image data to 224x224 pixels, perform frame extraction on video data (extract 10 key frames per second); perform word segmentation and stop word filtering on text data, and use a special vocabulary library in the financial field (including professional terms such as "overdue", "bad debt", "joint guarantee", etc.). Anomaly detection: Based on the Isolation Forest algorithm, identify outlier data such as abnormal transaction amounts (single consumption exceeding 10 times the average monthly income) and device location jumps (cross-city positioning within 30 minutes), and mark them as suspicious samples for manual review. III. Semantic alignment and dynamic knowledge graph construction (1) Spatial semantic alignment Image and video processing: Adopt the YOLOv8 object detection algorithm to identify key regions in ID card and bank card images, and extract text information in combination with PaddleOCR, with an accuracy rate of 98.7%; Perform action recognition on remote face-to-face signing videos, use OpenPose to detect gesture trajectories (abnormal actions such as frequent occlusion of the camera and multiple people entering the camera), and judge the compliance of face-to-face signing through a spatio-temporal feature fusion model (3D-CNN + LSTM). Geographical semantic association: Match GPS location data with the administrative division knowledge base to establish an association relationship of "customer - common address - risk area", for example, mark counties with a high overdue rate as risk geographical units. (2) Text semantic recognition Domain-adapted BERT model: Fine-tune BERT-base based on financial domain corpus (500,000 credit reports, 300,000 customer service conversations), and focus on optimizing the semantic representations of professional terms such as "default probability" and "guarantee liability", with the F1 value in the credit text classification task increased to 92%. Entity Relationship Extraction: Use the BiLSTM+CRF model to extract entities (customer names, institutional names, loan products) and relationships ("belong to", "guarantee", "associate") in the text. For example, extract the association relationship of "Zhang San - lending institution - XX Bank" from the credit report. (III) Dynamic Knowledge Graph Construction Using the Neo4j graph database as the carrier, construct a risk control knowledge graph containing 5 types of entities (customers, institutions, accounts, transactions, devices) and 8 types of relationships: Customer entity: Includes attributes such as age, occupation, income, etc.; Transaction entity: Includes attributes such as transaction time, amount, channel, etc.; Relationship types: "customer - owns - account", "account - has - transaction", "device - associates - customer", etc.; Dynamic update: Real-time access to transaction data through the Kafka message queue, and update the graph node status every 10 seconds (the change in account balance triggers the update of node attributes). IV. Construction and Risk Identification of the GNN-Transformer Hybrid Model (I) Model Architecture Design GNN layer: Use the Graph Attention Network (GAT) to process the structural features of the knowledge graph, calculate the neighbor risk index of the customer node (if the overdue rate of associated customers > 30%, the node weight increases), and capture the risk conduction effect in the guarantee network and the fund transfer network. Transformer layer: Convert the transaction flow text (consumption records in the past 12 months) into sequence data, use the multi-head self-attention mechanism to capture long-range dependencies, and identify periodic overdue (3 consecutive repayment dates are non-working days) and abnormal consumption patterns (high-frequency large-amount transactions in the early morning). Fusion mechanism: Through the GatingUnit, fuse the structural feature vector output by the GNN and the sequence feature vector output by the Transformer, and output the customer risk feature vector (dimension 128). (II) Model Training and Inference Data annotation: Based on historical loan data, annotate three types of labels: "normal customers", "customers under attention", and "default customers", and use stratified sampling to solve the problem of class imbalance (the positive and negative sample ratio is 1:4). Training strategy: Use the AdamW optimizer, learning rate 1e-4, adopt early stopping to prevent overfitting (stop if the validation set loss does not decrease for 5 consecutive rounds), and it takes 8 hours to train on the NVIDIA A100 GPU. Risk Identification: Input the knowledge graph sub-graph (including neighbor nodes within 3 hops) and transaction text data of the customer to be evaluated, and the model outputs the default probability (between 0 and 1). Uncertainty quantification is performed through Monte Carlo Dropout: Randomly discard 50% of the neurons during each forward pass and repeat the inference 100 times; Calculate the probability mean as the final prediction value and the standard deviation as the confidence index (a standard deviation > 0.15 triggers manual review). V. Digital Model Verification and Optimization (I) Modeling Paradigm Selection Dynamically select the modeling strategy based on the risk identification results: High-confidence scenarios (confidence > 0.9, default probability > 0.7): Immediately trigger a risk warning and reject the loan application; Medium-confidence scenarios (0.5 ≤ confidence ≤ 0.9): Use the XGBoost model for secondary verification and make a comprehensive judgment in combination with expert rules ("overdue more than 3 times in the past 6 months and income-debt ratio > 70%"); Low-confidence scenarios (confidence < 0.5): Enter the manual review process and supplement the submission of asset certificates and other materials. (II) GAN Generative Adversarial Verification Adversarial sample generation: Train the Wasserstein GAN (WGAN-GP) to generate adversarial samples. The input is normal transaction data, and the output is "critical samples" close to the default boundary: Generator network: A multi-layer fully connected neural network that takes an input noise vector (dimension 64) and outputs a 100-dimensional feature vector (simulating features such as transaction amount and frequency); Discriminator network: A binary classifier that distinguishes between real default samples and generated samples, with the gradient penalty coefficient set to 10. Model blind spot detection: Input the generated 20,000 adversarial samples into the risk control model to detect the following blind spots: New fraud patterns: Such as registering multiple accounts using false identity information for fund collection; Rule bypass risks: Such as avoiding single-transaction limit monitoring by splitting the transaction amount; Model iterative optimization: Add the detected blind spot samples to the training data set and update the GNN-Transformer model quarterly to achieve dynamic adaptation to emerging risks.

[0042] VI. Implementation Effects Risk identification accuracy: The F1 value for identifying default customers has increased, and the false rejection rate has decreased; Processing efficiency: The time taken for automated approval of a single loan has been shortened, and the proportion of manual review has decreased; Risk coverage: 23 related guarantee default events were discovered through the knowledge graph, with early warning 15 days ahead of the traditional rule engine.

[0043] Please refer to Figure 2 , in a standard digital modeling and verification method based on artificial intelligence, the data preprocessing of multi-source heterogeneous data to obtain initial multi-source heterogeneous data includes the following steps: Step 201, the multi-source heterogeneous data at least includes national standard data, industry standard data, group standard data, enterprise standard data, text data, image data, video data and sensor data; Step 202, extract the quantitative data in the national standard data, industry standard data, group standard data and enterprise standard data, and use 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; unify the text length to obtain unified text data; Step 203, adjust the images with different resolutions in the image data to a unified size, and use a median filter algorithm for noise reduction, and perform enhancement processing on the noise-reduced image data to obtain noise-reduced image data; Step 204, perform frame splitting on the video data in the multi-source heterogeneous data, and perform noise reduction and enhancement processing on the frame-split image data in sequence to obtain frame-split video data; Step 205, perform noise reduction processing on the sensor data through a Kalman filter algorithm, and fill in the missing data of the sensor to obtain complete sensor data.

[0044] The above introduces the embodiments of a standard digital modeling and verification method based on artificial intelligence of the present invention. Please refer to Figure 3 , in a standard digital modeling and verification system based on artificial intelligence, the standard digital modeling and verification system includes the following modules: Multi-source data acquisition module, used to obtain text data, image data, video data and sensor data through multi-source API interfaces, obtain multi-source heterogeneous data, and perform data preprocessing on the multi-source heterogeneous data to obtain initial multi-source heterogeneous data; Knowledge graph building module, used to perform spatial semantic alignment on the image and video data in the initial multi-source heterogeneous data according to the image recognition algorithm, and use the BERT model to perform semantic recognition on the text data to build a dynamic knowledge graph; Recognition model building module, used to process the structural features of the dynamic knowledge graph based on the GNN graph neural network, and use Transformer to capture long-range text dependencies to obtain a GNN-Transformer hybrid neural network model; The 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 establish a digital model based on the recognition result by selecting the best modeling paradigm, and use the trained GAN (Generative Adversarial Network) to generate test cases for the digital model to detect model blind spots.

[0045] 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 by the above embodiments. The above embodiments and descriptions in the specification are only preferred examples of the present invention and do not limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by 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 described standard digital modeling and verification method includes the following steps: Obtain multi-source heterogeneous data through multi-source API interfaces, and perform data preprocessing on the multi-source heterogeneous data to obtain initial multi-source heterogeneous data; According to the image recognition algorithm, perform spatial semantic alignment on the image and video data in the initial multi-source heterogeneous data, use the BERT model to perform semantic recognition on the text data, and establish a dynamic knowledge graph; Based on the GNN graph neural network, process the structural features of the dynamic knowledge graph, and use Transformer to capture long-range text dependencies to obtain a GNN-Transformer hybrid neural network model; 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; Based on the recognition result, select the best modeling paradigm 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.

2. The method for standard digital modeling and verification 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 at least includes national standard data, industry standard data, group standard data, enterprise standard data, text data, image data, video data, and sensor data; Extract the quantitative data in the national standard data, industry standard data, group standard data, and enterprise standard data, and use 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; unify the text length to obtain unified text data; Adjust the images with different resolutions in the image data to a unified size, and use the median filtering algorithm for noise reduction, and perform enhancement processing on the denoised image data to obtain denoised image data; Perform frame splitting on the video data in the multi-source heterogeneous data, and perform noise reduction and enhancement processing on the frame-split image data in sequence to obtain frame-split video data; Perform noise reduction processing on the sensor data through the Kalman filter algorithm, and fill in the missing data of the sensor to obtain complete sensor data.

3. The method for standard digital modeling and verification based on artificial intelligence according to claim 1, characterized in that, The performing spatial semantic alignment on the image and video data in the initial multi-source heterogeneous data according to the image recognition algorithm, using the BERT model to perform semantic recognition on the text data, and establishing a dynamic knowledge graph includes: Use the ResNet network to extract features from the images and video frames in the initial multi-source heterogeneous data to obtain the visual feature vectors of the images; Perform geometric transformation on the images through the STN spatial transformation network, perform spatial alignment on the images at different perspectives and scales, use the optical flow method to calculate the motion information between adjacent frames of the video data, and optimize the spatial semantic alignment effect by combining time series features; Establish a mapping relationship between the features of the position, shape, and color of the objects in the images and videos and the semantic concepts during the spatial semantic alignment process.

4. The method for standard digital modeling and verification based on artificial intelligence according to claim 1, characterized in that, Performing spatial semantic alignment on the image and video data in the initial multi-source heterogeneous data according to the image recognition algorithm, performing semantic recognition on the text data using the BERT model, and establishing a dynamic knowledge graph, further including: Constructing a dynamic knowledge graph based on the mapping relationship after spatial semantic alignment, combining the entities and relationships in the text data after semantic recognition; Using entity linking technology to match the entities in the text with the nodes in the knowledge graph; The nodes of the dynamic knowledge graph include objects, people, locations, events, categories, and attributes, and the edges represent the relationships between entities.

5. A method for standard digital modeling and verification based on artificial intelligence according to claim 1, characterized in that, Processing the structural features of the dynamic knowledge graph based on the GNN graph neural network, and using Transformer to capture long-range text dependencies, obtaining a GNN-Transformer hybrid neural network model, further including: Based on the GCN graph convolutional network as the basic model of the GNN, processing the structural features of the dynamic knowledge graph, and using the Transformer encoder to capture the long-range dependency relationships in the text data; Concatenating the node feature matrix output by the GCN and the text feature vector output by the Transformer as the input of the hybrid neural network model, and performing feature transformation and dimension unification through a fully connected layer to obtain a hybrid feature representation.

6. A standard digital modeling and verification method based on artificial intelligence according to claim 1, characterized in that 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, further including: Performing multiple forward propagations using the Monte Carlo Dropout method, maintaining the randomness of Dropout in each propagation to obtain multiple recognition results; Quantifying the prediction confidence by calculating the variance of the multiple recognition results, where the smaller the variance, the higher the confidence, and outputting the recognition result with high confidence.

7. The standard digital modeling and verification method based on artificial intelligence according to claim 1, characterized in that 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, further including: Constructing a generator and a discriminator network, using the generator to generate test cases close to the real data distribution, and using the discriminator to distinguish the generated test cases from the real data; The generator uses a deep neural network, the input is a random noise vector, and the test cases are generated through layer-by-layer transformation. The discriminator uses a deep neural network, the input is the real data or the generated test cases, and the output is a judgment result; Inputting the generated test cases into the established digital model, and detecting whether there are blind spots in the model by analyzing the performance of the model on these test cases.

8. An artificial intelligence-based standard digital modeling and verification system, characterized in that, The standard digital modeling and verification system Includes the following modules: A multi-source data acquisition module for acquiring multi-source heterogeneous data through a multi-source API interface, and performing data preprocessing on the multi-source heterogeneous data to obtain initial multi-source heterogeneous data; A knowledge graph establishment module for performing spatial semantic alignment on the image and video data in the initial multi-source heterogeneous data according to the image recognition algorithm, performing semantic recognition on the text data using the BERT model, and establishing a dynamic knowledge graph; The recognition model establishment module is used to process the structural features of the dynamic knowledge graph based on the GNN (Graph Neural Network), and capture long-range text dependencies by using Transformer, so as to obtain a GNN-Transformer hybrid neural network model; The 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 result 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.

9. An artificial intelligence-based standard digital modeling and verification system according to claim 8, characterized in that, The knowledge graph establishment module includes the following sub-modules: The extraction sub-module is used to extract features of images and video frames in the initial multi-source heterogeneous data by using the ResNet network to obtain visual feature vectors of the images; The optimization sub-module is used to perform geometric transformation on the images through the STN (Spatial Transformer Network), perform spatial alignment on the images at different perspectives and scales, calculate the motion information between adjacent frames of the video data by using the optical flow method, and optimize the spatial semantic alignment effect by combining time series features; The mapping sub-module is used to establish the mapping relationship between the features of the position, shape, and color of the objects in the images and videos and the semantic concepts during the spatial semantic alignment process.

10. An artificial intelligence-based standard digital modeling and verification system according to claim 8, characterized in that, The knowledge graph establishment module also includes the following sub-modules: The construction sub-module is used to construct a dynamic knowledge graph based on the mapping relationship after spatial semantic alignment, combined with the entities and relationships in the text data after semantic recognition; The matching sub-module is used to match the entities in the text with the nodes in the knowledge graph by using entity linking technology; The setting sub-module is used to set that the nodes of the dynamic knowledge graph include objects, people, locations, events, categories, and attributes, and the edges represent the relationships between entities.

Images