Digital archive management system based on multi-modal knowledge graph and block chain
By using multimodal knowledge graphs and blockchain technology, a cross-institutional trusted and shared digital archive management system was built, which solved the problems of authenticity and multimodal data processing in the verification of students' competition results, realized intelligent result storage and personalized evaluation, and improved the system's adaptability and data security.
Patent Information
- Application Number
- CN202610027700.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-09
- Publication Date
- 2026-03-31
AI Technical Summary
Existing technologies for verifying student competition results suffer from several problems, including difficulty in tracing the authenticity of the results, insufficient multimodal data processing capabilities, immature knowledge graph construction, and a lack of adaptability in the evaluation system. These issues result in high verification costs, low efficiency, insufficient credibility, and a lack of personalized evaluation results.
A digital archive management system based on multimodal knowledge graphs and blockchain is adopted, including an adaptive blockchain evidence storage layer, a multimodal knowledge graph intelligent construction layer, and a deep learning-driven evaluation engine, to achieve trusted data sharing across institutions, unified understanding of multimodal data, intelligent evaluation, and personalized services.
It enables automated understanding, reliable storage, and intelligent evaluation of students' competition results, improves data security and credibility, enhances the system's adaptability and personalized service capabilities, and reduces system maintenance costs and update cycles.
Smart Images

Figure CN121765141A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of educational informatization and blockchain technology integration, and in particular to a digital archive management system based on multimodal knowledge graphs and blockchain. Background Technology
[0002] With the deepening of digital transformation in my country's higher education, academic competitions, as an important vehicle for cultivating innovative talents, face significant challenges in their management and evaluation. Existing technologies suffer from the following key issues: I. Outdated Technology for Verifying the Authenticity of Results: Current verification of student competition results mainly relies on traditional paper certificates or simple electronic signature technology, which has fundamental flaws such as ease of forgery and difficulty in tracing the source. Especially in cross-institutional verification scenarios, the lack of a unified and trusted infrastructure leads to high verification costs, low efficiency, and insufficient credibility.
[0003] Second, insufficient multimodal data processing capabilities: Student competition results cover multiple modalities of data, such as text reports, design drawings, demonstration videos, and defense recordings. The existing system is mainly designed for single modality and lacks a unified ability to understand and analyze multimodal data, thus failing to fully explore the value of the data.
[0004] Third, the knowledge graph construction technology is immature: Traditional knowledge graph construction is mainly based on rule-driven or simple machine learning methods, which have low accuracy when dealing with complex semantic relationships in the education field, and lack adaptive learning and dynamic update capabilities, making it difficult to adapt to the rapidly changing education environment.
[0005] Fourth, the lack of intelligent assessment models: the existing assessment system mainly relies on fixed weights and rules formulated manually, which cannot be adaptively adjusted according to the characteristics and development trends of different disciplines. The assessment results lack personalization and foresight, making it difficult to support accurate educational decisions. Summary of the Invention
[0006] The purpose of this invention is to solve the problem of how to construct an intelligent digital archive management system for universities that integrates multimodal knowledge graphs, blockchain technology, and deep learning models, so as to realize the automated understanding, reliable evidence storage, intelligent evaluation, and personalized services of students' competition results.
[0007] To achieve the above objectives, this invention provides a digital archive management system based on multimodal knowledge graphs and blockchain, comprising: an adaptive blockchain evidence storage layer and a multimodal knowledge graph intelligent construction layer; The adaptive blockchain evidence storage layer adopts a consortium blockchain architecture to achieve trusted data sharing across institutions. The intelligent construction layer of the multimodal knowledge graph adopts a pre-trained language model based on a multimodal fusion architecture to achieve unified understanding and representation learning of multimodal data including text, images, audio, and video.
[0008] Preferably, the authoritative nodes of the consortium blockchain architecture are deployed by the education authorities, the participating nodes are deployed by the universities, and the observation nodes are deployed by the competition organizers; the adaptive blockchain evidence storage layer integrates a smart contract understanding model based on the Transformer architecture, a dynamic consensus mechanism that combines Byzantine fault tolerance algorithm and proof-of-stake mechanism, and protects data privacy through a federated learning framework combined with differential privacy technology, and uses the zk-SNARKs protocol to implement zero-knowledge proof verification.
[0009] Preferably, the adaptive blockchain evidence storage layer dynamically optimizes the evidence storage strategy through reinforcement learning algorithms, and adaptively adjusts the evidence storage granularity and verification mechanism according to data importance, access frequency and verification requirements; the dynamic consensus mechanism uses a deep learning model with a recurrent neural network architecture to predict node behavior patterns, and dynamically adjusts the fault tolerance threshold and block production weight through a Bayesian optimization algorithm.
[0010] Preferably, the entity recognition module of the multimodal knowledge graph intelligent construction layer adopts the BERT-CRF architecture, uses a large-scale corpus in the education field for pre-training and fine-tuning through self-supervised learning, and integrates adversarial training and data augmentation techniques; The relationship extraction module adopts an end-to-end learning framework based on graph neural networks, and supports the identification of complex one-to-many and many-to-many relationships through the attention mechanism. The relationship types include participation, acquisition, manifestation, guidance, and cooperation. Multimodal fusion employs a cross-modal attention mechanism to generate a unified multimodal representation vector. The fusion weights adaptively adjust the contribution of each modality according to different task requirements using a meta-learning algorithm. The contrastive loss function for multimodal alignment is: ; in, To compare the loss values, It is an exponential function. For similarity, For the feature vector of the sample, for The corresponding positive sample feature vector, To adjust the hyperparameters of the similarity distribution, For the first k The feature vector of each candidate sample Let be the similarity index of all candidate samples.
[0011] Preferably, the relation extraction module supports differentiated modeling of different types of nodes and edges. Node updates use a message passing mechanism, where each node collects information from its neighbors and updates its own representation using a learnable aggregation function. The graph sampling strategy employs importance sampling, determining the probability of adoption based on the centrality and connection strength of the nodes. A dynamic graph update algorithm based on incremental learning is used, prioritizing the updating of nodes that have a greater impact on the global structure through importance sampling. The knowledge graph update of the intelligent construction layer of the multimodal knowledge graph adopts a continuous learning framework; and an elastic weight consolidation method is used to add regularization constraints to important model parameters; the identification of new entities and relations adopts few-shot learning technology, and adapts to new entity types and relation patterns through a meta-learning framework; the prototype network is used to learn the prototype representation of entities and relations, and new samples are classified by similarity calculation with the prototype.
[0012] Preferably, the system also includes a deep learning-driven evaluation engine and an intelligent education evaluation support layer; The deep learning-driven evaluation engine constructs an evaluation model based on neural networks. The evaluation model adopts a multi-task learning framework with hard parameter sharing to realize intelligent evaluation and personalized recommendations of students' abilities and competition results. The intelligent education assessment support layer is used to provide subject development prediction, education quality assessment and abnormal risk warning, and support education management decisions.
[0013] Preferably, the multi-task learning framework includes a bottom-level shared encoder learning a general feature representation and a top-level task-specific decoder predicting different evaluation dimensions, wherein the shared encoder uses residual connections and layer normalization; The evaluation engine integrates an adaptive weight learning algorithm based on a meta-learning framework. It automatically balances the weights of each evaluation task through a gradient normalization algorithm. Weight updates are based on the relative rate of change of task loss, and the weight update formula is as follows: ; in, The weight at time t+1, Let α be the weight at time t, and α be the learning rate. Let f(x;w) be the gradient of the weight w, and L be the loss function. t ) is the prediction function of the model, y is the true label corresponding to sample x, and the weights are continuously optimized through the gradient descent algorithm.
[0014] Preferably, in the deep learning-driven evaluation engine, the competition results evaluation adopts a sequence-to-sequence model based on an attention mechanism. It is pre-trained on historical winning works through self-supervised learning to analyze the innovation, technical difficulty, and application value of the competition results. The attention mechanism includes self-attention, cross-attention, and hierarchical attention. The attention calculation adopts sparsity technology and reduces computational complexity through Top-K selection and threshold truncation. The deep learning-driven evaluation engine includes a personalized recommendation system that employs a hybrid architecture of graph neural networks and collaborative filtering to model the interaction relationship between users, competitions, and skills, generating personalized recommendations. The recommendation strategy is optimized through reinforcement learning, iterating the strategy with the goal of improving users' long-term engagement and capabilities. It uses a temporal difference learning algorithm and a stable training process through experience replay and target network.
[0015] Preferably, the intelligent education assessment support layer uses a spatiotemporal graph neural network to construct a subject development prediction model, models the time dimension dependency through gated recurrent units, and learns the spatial dimension correlation through graph convolutional networks; the prediction model supports multi-step prediction and probabilistic prediction, and outputs prediction results that include uncertainty quantification; uncertainty estimation adopts a Bayesian neural network framework, and learns the posterior distribution of parameters through variational inference. The intelligent education assessment support layer integrates an anomaly detection system based on variational autoencoders and generative adversarial networks. The risk warning adopts a time series anomaly detection algorithm combined with statistical methods and deep learning methods. The LSTM autoencoder learns the pattern of normal sequences, and combines statistical process control methods to set dynamic thresholds for real-time risk monitoring.
[0016] Preferably, the intelligent education assessment support layer adopts an integrated learning framework, which automatically extracts education quality assessment indicators through factor analysis and principal component analysis to generate a comprehensive assessment report; and enhances the interpretability of the assessment results through attention visualization, SHAP value calculation and counterfactual interpretation techniques.
[0017] This invention employs the aforementioned digital archive management system based on multimodal knowledge graphs and blockchain, and its beneficial effects are as follows: (1) This invention adopts the latest deep learning and large model technology to realize the fundamental transformation from traditional rule-driven to intelligent data-driven; the multimodal fusion capability and knowledge graph construction provide stronger technical support for educational data analysis; (2) Through meta-learning and reinforcement learning mechanisms, the system can automatically adjust parameters and strategies according to different application scenarios and data characteristics, effectively improving the system's adaptability and significantly reducing system maintenance costs and update cycles. (3) The deep integration of blockchain technology and artificial intelligence in this invention realizes the immutable storage and intelligent verification of data, greatly improves data security and credibility, and lays the foundation for building a nationwide education data ecosystem; (4) Based on deep learning prediction and analysis models, it can discover hidden patterns that traditional methods cannot identify, providing more scientific and forward-looking support for educational decisions; intelligent personalized recommendations and automated services significantly improve the user experience. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the overall architecture of a digital archive management system based on multimodal knowledge graph and blockchain according to the present invention; Figure 2 This invention provides a flowchart of the multimodal knowledge graph construction process for a digital archive management system based on multimodal knowledge graphs and blockchain. Figure 3 This is a flowchart of a deep learning evaluation engine for a digital archive management system based on multimodal knowledge graphs and blockchain, according to the present invention. Figure 4 This is a prediction flowchart of an intelligent education assessment support layer for a digital archive management system based on multimodal knowledge graphs and blockchain, according to the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.
[0020] like Figure 1 As shown, an embodiment of the present invention provides a digital archive management system based on multimodal knowledge graph and blockchain, comprising: an adaptive blockchain evidence storage layer, a multimodal knowledge graph intelligent construction layer, a deep learning-driven evaluation engine, and an intelligent education evaluation support layer; 1. Adaptive Blockchain Evidence Storage Layer: Adopting a consortium blockchain architecture to achieve trusted data sharing across institutions, the authoritative node is deployed by the education authority, the participating nodes are deployed by various universities, and the observation node is deployed by the competition organizer. The adaptive blockchain evidence storage layer integrates a smart contract understanding model based on the Transformer architecture, a dynamic consensus mechanism that combines Byzantine fault tolerance algorithm and proof-of-stake mechanism, and protects data privacy through a federated learning framework combined with differential privacy technology, and uses the zk-SNARKs protocol to achieve zero-knowledge proof verification.
[0021] (1) Smart Contract Understanding Model: Based on the Transformer natural language understanding model, it can automatically parse complex evidence storage rules and permission logic. The contract code is semantically analyzed through a pre-trained language model to identify potential logical vulnerabilities and security risks. Contract optimization adopts a reinforcement learning framework, with multiple objectives such as execution efficiency, security, and cost as reward functions, and continuously optimizes the contract execution strategy through a policy gradient algorithm. During the optimization process, the system automatically collects execution logs and performance indicators, builds an experience replay buffer, and supports offline policy learning.
[0022] The evidence storage strategy is dynamically optimized through reinforcement learning algorithms, and the granularity of evidence storage and the verification mechanism are adaptively adjusted according to the importance of data, access frequency and verification requirements. (2) Dynamic consensus mechanism: It integrates the Byzantine fault tolerance algorithm and the proof-of-stake mechanism, and adopts a deep learning model with a recurrent neural network architecture to analyze the network status and predict the node behavior pattern in real time, and dynamically adjust the consensus parameters; it inputs multi-dimensional features such as historical transaction data, network latency, and node performance, and outputs the optimal consensus configuration parameters. By dynamically adjusting the fault tolerance threshold and block weight through a Bayesian optimization algorithm, the transaction processing throughput is maximized while ensuring network security. The system maintains a probabilistic model with parameter configuration and continuously updates the parameter distribution based on historical performance data.
[0023] (3) Privacy protection enhancement technology: The federated learning framework is adopted to protect user privacy. Each participating node only needs to share model parameters and does not need to share the original data. Differential privacy technology is used in the federated learning process to add noise to the model parameters to prevent malicious nodes from inferring privacy information through the parameters.
[0024] Zero-knowledge proof technology is used to verify the authenticity of students' work, proving that it meets specific conditions without disclosing the specific content of the work; the proof generation adopts the zk-SNARKs protocol, which supports zero-knowledge verification of complex logical relationships.
[0025] 2. Multimodal Knowledge Graph Intelligent Construction Layer: Employs a pre-trained language model based on a multimodal fusion architecture (based on the Vision-Language Transformer architecture) to achieve unified understanding and representation learning of multimodal data including text, images, audio, and video; model pre-training tasks include various cross-modal tasks such as image-text matching, video-text alignment, and audio semantic understanding. Figure 2 As shown.
[0026] The model employs a hierarchical attention mechanism. The bottom-level attention learns local feature associations within a modality, while the top-level attention learns global semantic correspondences across modalities. Attention weights are dynamically adjusted through a gating mechanism, adaptively allocating computational resources based on the quality and importance of the input data.
[0027] (1) Entity recognition module: The BERT-CRF architecture is adopted. The model is pre-trained and fine-tuned using a large-scale corpus in the education field through self-supervised learning, which supports the accurate recognition of multiple types of entities such as students, competitions, achievements, skills, and institutions. The model integrates adversarial training and data augmentation techniques to improve robustness to noisy data and data from outside the domain. (2) The relationship extraction module adopts an end-to-end learning framework based on graph neural networks, and supports the identification of complex one-to-many and many-to-many relationships through the attention mechanism. The relationship types include participation, acquisition, manifestation, guidance, and cooperation. The relation extraction module supports differentiated modeling of different types of nodes and edges. Node updates use a message passing mechanism, where each node collects information from its neighbors and updates its own representation through a learnable aggregation function. The graph sampling strategy employs importance sampling, determining the sampling probability based on node centrality and connection strength to improve training efficiency while maintaining the integrity of the graph structure. For large-scale graph data, graph sampling algorithms such as GraphSAINT are used, supporting mini-batch training; a dynamic graph update algorithm based on incremental learning is employed, prioritizing updates to nodes with a greater impact on the global structure through importance sampling. (3) Multimodal fusion adopts a cross-modal attention mechanism to learn the correspondence and complementarity between features of different modalities, and generates a unified multimodal representation vector. The fusion weights are adaptively adjusted by a meta-learning algorithm according to different task requirements. The contrastive loss function for multimodal alignment is: ; in, To compare the loss values, It is an exponential function. For similarity, For the feature vector of the sample, for The corresponding positive sample feature vector, To adjust the hyperparameters of the similarity distribution, For the first k The feature vector of each candidate sample Let be the similarity index of all candidate samples.
[0028] (4) Knowledge graph update and continuous optimization: The knowledge graph adopts a continuous learning framework to avoid forgetting the learned knowledge while adding new knowledge; and adopts an elastic weight consolidation method to add regularization constraints to important model parameters to prevent catastrophic forgetting; the identification of new entities and relations adopts few-shot learning technology and adapts to new entity types and relation patterns through a meta-learning framework; the prototype network is used to learn the prototype representation of entities and relations, and new samples are classified by similarity calculation with the prototype.
[0029] 3. A deep learning-driven evaluation engine is constructed based on a neural network. This evaluation model employs a multi-task learning framework with shared hard parameters to achieve intelligent assessment and personalized recommendations of student abilities and competition results; for example... Figure 3 As shown.
[0030] (1) Multi-task learning framework: including a bottom-level shared encoder that learns general feature representations and a top-level task-specific decoder that predicts different evaluation dimensions. The shared encoder uses residual connections and layer normalization. The evaluation engine integrates an adaptive weight learning algorithm based on a meta-learning framework. It automatically balances the weights of each evaluation task through gradient normalization. Weight updates are based on the relative rate of change of task loss, preventing any single task from dominating the overall training process and ensuring that all tasks are adequately trained. The weight update formula is: ; in, The weight at time t+1, Let α be the weight at time t, and α be the learning rate. Let f(x;w) be the gradient of the weight w, and L be the loss function. t ) is the prediction function of the model, y is the true label corresponding to sample x, and the weights are continuously optimized through the gradient descent algorithm.
[0031] (2) Evaluation of Competition Results: A sequence-to-sequence model based on attention mechanisms was adopted. The model was pre-trained on historical award-winning works through self-supervised learning to analyze the innovativeness, technical difficulty, and application value of the competition results. Attention mechanisms included self-attention, cross-attention, and hierarchical attention. Self-attention was used to learn the dependencies within the input sequence, cross-attention was used to fuse different types of input information, and hierarchical attention was used to process the hierarchical structure of long sequences. The model was pre-trained on a large number of historical award-winning works through self-supervised learning to learn the common characteristics and evaluation criteria of outstanding results.
[0032] Attention computation employs sparsity techniques, reducing computational complexity through Top-K selection and threshold truncation; the sparse attention pattern is dynamically adjusted through a learnable mask matrix, adaptively selecting important attention connections based on the characteristics of the input data.
[0033] (3) Personalized recommendation system: A hybrid architecture of graph neural network and collaborative filtering (deep Q network framework) is adopted to model the interaction relationship between user-competition-skill and generate accurate personalized recommendations. The state space includes multi-dimensional information such as user's historical behavior, current ability level and environmental context. The action space is the recommended competition, resource, partner and other options. The recommendation strategy is optimized by reinforcement learning. The strategy is iterated with the goal of improving the user's long-term participation and ability. The temporal difference learning algorithm is adopted and the network is stabilized through experience playback and target network training.
[0034] The reward function is designed to consider both short-term feedback and long-term goals. Short-term rewards are based on the user's immediate reaction, while long-term rewards are based on the user's ability improvement and goal achievement.
[0035] 4. Intelligent Education Assessment Support Layer: This layer provides predictions of subject development, assessments of educational quality, and early warnings of abnormal risks, supporting educational management decisions. For example... Figure 4 As shown.
[0036] (1) Discipline development prediction model: The discipline development prediction model is constructed using a spatiotemporal graph neural network. The time dimension dependency is modeled through gated recurrent units, and the spatial dimension correlation is learned through graph convolutional networks. The prediction model supports multi-step prediction and probabilistic prediction, and outputs prediction results containing uncertainty quantification. Uncertainty estimation adopts a Bayesian neural network framework, and the posterior distribution of the learning parameters is learned through variational inference. The model models the dynamic evolution of relationships among disciplines, universities, students, and competitions, predicting trends in discipline development and changes in talent demand. It identifies key influencing factors through an attention mechanism, providing interpretable prediction results.
[0037] (2) Anomaly detection and risk warning: An anomaly detection system based on variational autoencoder and generative adversarial network is integrated to identify abnormal patterns through reconstruction error; the encoder compresses the input data into a low-dimensional latent space, and the decoder reconstructs the original data from the latent representation. Samples with large reconstruction errors are identified as anomalies; the system automatically identifies abnormal patterns and potential risks in educational data to provide early warning support for educational management decisions.
[0038] The risk warning system employs a time series anomaly detection algorithm combined with statistical and deep learning methods; the LSTM autoencoder learns the patterns of normal sequences, and combined with statistical process control methods, dynamic thresholds are set for real-time risk monitoring.
[0039] (3) Education quality assessment and interpretability enhancement: The education quality assessment adopts an integrated learning framework, which automatically extracts education quality assessment indicators through factor analysis and principal component analysis to generate a comprehensive assessment report; The interpretability of the evaluation results is achieved through techniques such as attention visualization, feature importance analysis, and counterfactual interpretation. Attention weights are mapped to the original input to identify the data segments most important to the prediction results; SHAP value calculation is used to quantify the contribution of each feature to the prediction results, supporting both global and local interpretations; counterfactual generation uses adversarial example generation techniques to demonstrate the impact of small changes in the input on the prediction results.
[0040] Therefore, this invention employs a digital archive management system based on multimodal knowledge graphs and blockchain. This invention utilizes the latest deep learning and large-scale model technologies to achieve a fundamental shift from traditional rule-driven to intelligent data-driven approaches. Multimodal fusion capabilities and knowledge graph construction provide stronger technical support for educational data analysis. Through meta-learning and reinforcement learning mechanisms, the system can automatically adjust parameters and strategies according to different application scenarios and data characteristics, effectively improving the system's adaptability and significantly reducing system maintenance costs and update cycles. The deep integration of blockchain technology and artificial intelligence enables tamper-proof data storage and intelligent verification, significantly improving data security and credibility, laying the foundation for building a nationwide educational data ecosystem. Deep learning-based prediction and analysis models can discover hidden patterns that traditional methods cannot identify, providing more scientific and forward-looking support for educational decision-making. Intelligent personalized recommendations and automated services significantly enhance the user experience.
[0041] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.
[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A digital archive management system based on multi-modal knowledge graph and blockchain, characterized in that, The system comprises: an adaptive blockchain storage layer and a multi-modal knowledge graph intelligent construction layer. The adaptive blockchain storage layer realizes cross-institutional data trusted sharing by adopting a consortium chain architecture. The multi-modal knowledge graph intelligent construction layer realizes unified understanding and representation learning of multi-modal data including text, image, audio and video by adopting a pre-training language model based on a multi-modal fusion architecture.
2. The digital archive management system based on multi-modal knowledge graph and blockchain according to claim 1, characterized in that, The authoritative nodes of the consortium chain architecture are deployed by the education authorities, the participating nodes are deployed by the universities, and the observation nodes are deployed by the competition organizers; the adaptive blockchain storage layer integrates an intelligent contract understanding model based on a Transformer architecture, a dynamic consensus mechanism that combines a Byzantine fault-tolerant algorithm and a proof-of-stake mechanism, and protects data privacy by combining a federated learning framework with differential privacy technology, and realizes zero-knowledge proof verification by using a zk-SNARKs protocol. 3.The digital archive management system based on multi-modal knowledge graph and blockchain according to claim 2, characterized in that, The adaptive blockchain storage layer dynamically optimizes the storage strategy by reinforcement learning algorithm, and adaptively adjusts the storage granularity and verification mechanism according to the data importance, access frequency and verification demand; the dynamic consensus mechanism uses a deep learning model using a recurrent neural network architecture to predict node behavior patterns, and dynamically adjusts the fault tolerance threshold and block weight by a Bayesian optimization algorithm. 4.The digital archive management system based on multi-modal knowledge graph and blockchain according to claim 1, wherein, The entity recognition module of the multi-modal knowledge graph intelligent construction layer adopts a BERT-CRF architecture, pre-trains and fine-tunes using large-scale corpus in the education field through self-supervised learning, and integrates adversarial training and data enhancement technology; The relationship extraction module adopts an end-to-end learning framework based on a graph neural network, supports one-to-many and many-to-many complex relationship recognition through an attention mechanism, and the relationship types include participation, acquisition, embodiment, guidance and cooperation; Multi-modal fusion adopts a cross-modal attention mechanism to generate a unified multi-modal representation vector, and the fusion weight is adaptively adjusted according to different task requirements by a meta-learning algorithm; the contrast loss function of multi-modal alignment is: ; wherein, is a contrast loss value, is an exponential function, is a similarity, is a feature vector of a sample, is is a corresponding positive sample feature vector, is a hyper parameter to adjust the similarity distribution, is a feature vector of the k th candidate sample, is a sum of similarity exponentials of all candidate samples.
5. The multi-modal knowledge graph and blockchain based digital archive management system as claimed in claim 4, wherein, The relationship extraction module supports differential modeling of different types of nodes and edges, and the node update uses a message passing mechanism, each node collects information of neighbor nodes, and updates its own representation through a learnable aggregation function; the graph sampling strategy adopts importance sampling, and determines the probability according to the centrality and connection strength of the node; a dynamic graph update algorithm based on incremental learning is adopted, and the nodes with greater impact on the global structure are updated preferentially through importance sampling; The knowledge graph update of the multi-modal knowledge graph intelligent construction layer adopts a continuous learning framework; and an elastic weight consolidation method is used to add regularization constraints to important model parameters; the identification of new entities and relationships uses few-shot learning technology, and adapts to new entity types and relationship patterns through a meta-learning framework; the prototype network is used to learn the prototype representation of entities and relationships, and new samples are classified by similarity calculation with the prototype. 6.The digital archive management system based on multi-modal knowledge graph and blockchain according to claim 1, wherein, The system further comprises a deep learning driven evaluation engine and an intelligent education evaluation support layer. The deep learning driven evaluation engine constructs an evaluation model based on a neural network, and the evaluation model adopts a multi-task learning framework with hard parameter sharing to realize intelligent evaluation and personalized recommendation of student ability and competition results. The intelligent education evaluation support layer is used for providing subject development prediction, education quality evaluation and abnormal risk early warning, and supporting education management decision-making.
7. The multi-modal knowledge graph and blockchain based digital archive management system as claimed in claim 6, wherein, The multi-task learning framework includes a bottom-layer shared encoder learning a general feature representation and a top-layer task-specific decoder predicting different evaluation dimensions, and the shared encoder uses residual connection and layer normalization; The evaluation engine integrates an adaptive weight learning algorithm based on a meta-learning framework, automatically balances the weights of each evaluation task through a gradient normalization algorithm, and updates the weights based on the relative change rate of task loss, with the weight update formula being: ; wherein, is the weight at time t + 1, is the weight at time t, and a is the learning rate, is the gradient of the weight w, L is the loss function, f(x; w t ) is the prediction function of the model, y is the true label corresponding to the sample x, and the weight is continuously optimized by the gradient descent algorithm. 8.The digital archive management system based on multi-modal knowledge graph and blockchain according to claim 7, characterized in that, In the deep learning driven evaluation engine, the competition achievement evaluation adopts a sequence-to-sequence model based on an attention mechanism, is pre-trained on historical award-winning works through self-supervised learning, and analyzes the innovation, technical difficulty and application value of the competition achievement; the attention mechanism includes self-attention, cross-attention and hierarchical attention, attention calculation adopts sparsification technology, and Top-K selection and threshold truncation are used to reduce the computational complexity; The deep learning driven evaluation engine includes a personalized recommendation system, adopts a hybrid architecture of graph neural network and collaborative filtering, models the interaction relationship among users, competitions and skills, and generates personalized recommendations; The recommendation strategy is optimized through reinforcement learning, and strategy iteration is performed with the user long-term participation and ability improvement as the target, a time difference learning algorithm is adopted, and the training process is stabilized through experience replay and target network. 9.The digital archive management system based on multi-modal knowledge graph and blockchain of claim 6, wherein, The intelligent education evaluation support layer adopts a spatio-temporal graph neural network to build a subject development prediction model, models the time dimension dependency relationship through a gated recurrent unit, and learns the spatial dimension correlation relationship through a graph convolution network; the prediction model supports multi-step prediction and probability prediction, and outputs prediction results containing uncertainty quantification; the uncertainty estimation adopts a Bayesian neural network framework, and learns the posterior distribution of parameters through variational inference; The intelligent education evaluation support layer integrates an anomaly detection system based on a variational autoencoder and a generative adversarial network, and the risk early warning adopts a time series anomaly detection algorithm combined with statistical methods and deep learning methods; The LSTM autoencoder learns the pattern of normal sequences, and sets a dynamic threshold through a statistical process control method to perform real-time risk monitoring.
10. The multi-modal knowledge graph and blockchain based digital archive management system as claimed in claim 9, wherein, The intelligent education evaluation support layer adopts an ensemble learning framework, automatically extracts education quality evaluation indicators through factor analysis and principal component analysis, and generates a comprehensive evaluation report; the explainability of the evaluation results is enhanced through attention visualization, SHAP value calculation and counterfactual explanation technology.
Citation Information
Patent Citations
Financial talent education evaluation system based on block chain technology
CN112258017A
Block chain-based after-class service system and method
CN118096453A
Multi-modal education evaluation system and method based on super-automation and federal cognitive calculation
CN120258606A
Cross-modal travel knowledge graph construction system based on large language model
CN120542539A
Financial data risk control system and method based on big data
CN120746695A
