An AI large model-based enterprise project digital management and control method, device and medium

Abstract

The application discloses an enterprise project digital management and control method and device based on an AI large model, and relates to the technical field of artificial intelligence, which comprises the following steps: receiving a feasibility study report and a project data table, extracting a text feature vector and a numerical feature matrix through a lightweight large model, and injecting Laplace noise to generate an encrypted feature vector set; inputting the feasibility study report into a knowledge privacy protection review intelligent agent for content analysis, performing risk checking and historical case matching through real-time query of a dynamic knowledge graph, and generating a structured review result; performing logical checking on the structured review result, generating interpretable evidence, adding the interpretable evidence to the structured review result, and obtaining a review report with an evidence chain; and introducing a lightweight large model in the feature extraction stage, combining a Laplace noise injection and an encryption mechanism, and significantly reducing the reverse inference risk caused by gradient leakage.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to an enterprise project digital management and control method based on an AI large model, equipment and a medium. BACKGROUND

[0002] With the deepening of enterprise digital transformation, the intelligentization and data-driven decision-making of project management have become the key link to improve the efficiency of organizational operation. Project management and control systems based on artificial intelligence (AI) technology have gradually evolved from traditional rule engines and statistical analysis models to deep learning and large model-driven directions. In particular, in key links such as feasibility study review, resource allocation optimization and risk early warning, the industry has begun to explore the use of pre-trained language models (Pre-trained Language Models, PLMs) to perform semantic analysis on unstructured text (such as feasibility study reports), and to realize multi-modal information fusion by combining structured project data tables. By converging data from various business domains within an enterprise to a central data lake, end-to-end neural networks are used to complete project evaluation tasks. Knowledge graph technology is widely introduced to enhance the reasoning ability of the system, and through the static construction of ontology relationships, the correlation analysis between project elements is realized to assist decision support.

[0003] The existing technology still faces significant challenges in balancing cross-organizational data collaboration and privacy protection. Especially in complex project scenarios involving multiple business departments or cross-enterprise collaboration, since project data is usually distributed in different trust domains, direct centralized processing not only violates data compliance requirements, but also may cause the risk of sensitive business information leakage. Although the current mainstream federated learning framework can achieve a distributed training mechanism of "data not moving, model moving", the gradient aggregation process in the high-dimensional feature space may still expose the statistical characteristics of the original data, posing a risk of reverse inference attacks. Existing methods generally lack effective modeling of dynamic knowledge evolution processes, making it difficult to integrate the implicit relationships between new project information and historical cases in real time, resulting in insufficient contextual adaptability of review conclusions. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an enterprise project digital management and control method based on an AI large model to solve the problem of high gradient leakage risk and lack of dynamic knowledge fusion capability leading to low credibility of review results.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides an AI large model-based enterprise project digital management and control method, which comprises the following steps: receiving a feasibility study report and a project data table, extracting a text feature vector and a numerical feature matrix through a lightweight large model, injecting Laplace noise to generate an encrypted feature vector set; aggregating the encrypted feature vector set to generate a global fusion feature vector, extracting project entities using an AI large model and associating them with historical project data tables to construct a dynamic knowledge graph; using the AI large model as a backbone encoder, reconstructing a multi-task output layer and integrating the dynamic knowledge graph to construct an evaluation model base architecture; using the dynamic knowledge graph, adding Gaussian noise to the aggregated gradient in the gradient descent stage, and performing multi-task training on the evaluation model base architecture to obtain a knowledge privacy protection evaluation agent; inputting the feasibility study report into the knowledge privacy protection evaluation agent for content analysis, performing risk verification and historical case matching through real-time query of the dynamic knowledge graph to generate a structured evaluation result; performing logical verification on the structured evaluation result to generate interpretable evidence, and adding the interpretable evidence to the structured evaluation result to obtain an evaluation report with an evidence chain.

[0008] As a preferred scheme of the AI large model-based enterprise project digital management and control method, the method comprises the following steps:

[0009] Receiving the feasibility study report and the project data table, and performing SHA-256 integrity verification and format compliance review to obtain a verified feasibility study report and a verified project data table;

[0010] Calling the lightweight large model DeepSeek-Coder to segment and analyze the verified feasibility study report, and generating a text feature vector through a Transformer encoder; simultaneously extracting core fields in the verified project data table to construct a numerical matrix, and obtaining a numerical feature vector after PCA dimension reduction;

[0011] Injecting Laplace noise into the text feature vector and the numerical feature vector to generate a noisy feature vector, and performing secondary encryption on sensitive fields through an SM4-CBC algorithm and adding a timestamp to output an encrypted feature vector set.

[0012] As a preferred scheme of the AI large model-based enterprise project digital management and control method, the method comprises the following steps:

[0013] Aggregating the encrypted feature vector set through a secure aggregation protocol to generate a global fusion feature vector;

[0014] The global fusion feature vector is input into the AI ​​large model, which automatically identifies and extracts project entities and attributes through semantic understanding capabilities;

[0015] The project entity's attributes are matched with records in the historical project data table. By calculating the similarity between attributes, the historical project node with the highest similarity is selected.

[0016] The project entity is treated as a new node, and a semantic relationship is established between it and the historical project node with the highest similarity. The association confidence is recorded to obtain the project node association results.

[0017] Based on the project node association results, new project entity nodes are created and timestamp attribute information is added to establish relationship edges between project entities. According to the association confidence, the weight values ​​of the relationships between project nodes and project entities are calculated and updated. By continuously integrating new project nodes and timestamp attributes, the relationship network between project entities is continuously expanded to form a dynamic knowledge graph.

[0018] As a preferred embodiment of the AI-based large-scale model-based digital management and control method for enterprise projects described in this invention, the specific steps for using the AI ​​large-scale model as the main encoder, reconstructing the multi-task output layer and integrating the dynamic knowledge graph to construct the review model base architecture are as follows.

[0019] The large AI model is encapsulated into the backbone encoder using Docker containers, and GPU resources are allocated using Kubernetes to generate a containerized encoder entity.

[0020] Based on the containerized encoder entity, a three-task parallel processing layer is constructed. Neo4jConnector is embedded in the input layer. The entity relationships of the dynamic knowledge graph are queried in real time through the Bolt protocol, and the fusion model prototype is obtained through graph neural network transformation.

[0021] The prototype of the fusion model is encapsulated as a gRPC microservice, a standardized prediction interface is defined, a two-layer elastic scaling strategy is configured, and a review model base architecture is formed.

[0022] As a preferred embodiment of the AI-based large-scale model-based digital management and control method for enterprise projects described in this invention, the specific steps for obtaining the knowledge privacy protection review intelligent agent are as follows:

[0023] Based on dynamic knowledge graphs, entity relationship triples are extracted using Cypher query statements, and then converted into feature tensors using graph neural networks to form a knowledge-enhanced training dataset.

[0024] The global parameters of the review model base architecture are distributed to all clients participating in the training. Each client uses local data and knowledge-enhanced training dataset to train and backpropagate the review model base architecture to obtain local gradients.

[0025] The local gradient is decrypted and the average gradient is calculated. Gaussian noise is added to the average gradient, and the review model base architecture is updated to obtain the knowledge privacy-preserving review agent.

[0026] As a preferred embodiment of the AI-based large-scale model-based digital management and control method for enterprise projects described in this invention, the steps of decrypting the local gradient and calculating the average gradient, adding Gaussian noise to the average gradient, updating the review model's base architecture, and obtaining the knowledge privacy-preserving review agent are as follows.

[0027] Receive all local gradients uploaded by participating clients in encrypted mode, perform joint decryption using a secure aggregation protocol to obtain plaintext local gradients, and obtain the global average gradient by weighted averaging.

[0028] The Gaussian noise intensity is calculated based on the privacy budget, a corresponding noise vector is generated, and it is injected into the global average gradient to obtain the noisy gradient.

[0029] Based on the noisy gradient, the global parameters of the review model's base architecture are updated using the stochastic gradient descent algorithm, and multiple iterations are performed to solidify and save the global parameters of the review model's base architecture, generating a knowledge privacy-protected review agent.

[0030] As a preferred embodiment of the AI-based large-scale model-based digital management and control method for enterprise projects described in this invention, the specific steps for generating structured review results are as follows:

[0031] The feasibility study report was converted into UTF-8 standardized text, and sensitive fields were de-identified to form an encrypted text feature vector;

[0032] The encrypted text feature vector is input into the knowledge privacy protection review agent, decrypted using a key to obtain the plaintext item feature representation, and then mathematically transformed by the backbone encoder to generate a deep semantic representation.

[0033] The deep semantic representation is converted into parameters of the Cypher query statement, and the query is sent to the Neo4j graph database through the efficient Bolt protocol, and the returned subgraph information is obtained in real time.

[0034] The returned graph subgraph information is transformed through a graph neural network to form a knowledge context vector. The vector is then independently computed using a multi-task parallel processing layer to generate structured review results.

[0035] As a preferred embodiment of the AI-based large-scale model-based digital management and control method for enterprise projects described in this invention, the specific steps for obtaining a review report with a chain of evidence are as follows:

[0036] The structured review results are subjected to deterministic logical verification, while probabilistic deep analysis is performed through neural networks to obtain multi-dimensional verification conclusions.

[0037] Relevant historical case node attributes are extracted from the dynamic knowledge graph and concatenated with the rule clause text to generate interpretable evidence. Combined with the structured review results, a review report with a chain of evidence is formed.

[0038] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements any step of the enterprise project digital management and control method based on AI big model as described in the first aspect of the present invention.

[0039] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the enterprise project digital management and control method based on AI big model as described in the first aspect of the present invention.

[0040] The beneficial effects of this invention are as follows: by introducing a lightweight large model combined with Laplacian noise injection and encryption mechanism in the feature extraction stage, distributed preprocessing and privacy enhancement of feasibility study report text and structured data are achieved. It not only enhances the learning depth of implicit relationships between project entities by leveraging the semantic association capabilities of knowledge graphs, but also significantly reduces the risk of reverse inference caused by gradient leakage through the synergistic effect of dual-layer encryption channels and secure aggregation protocols. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Fig. 1 This is a flowchart of a digital management and control method for enterprise projects based on AI big data models.

[0043] Fig. 2 A flowchart for generating encrypted feature vectors.

[0044] Fig. 3 A flowchart for constructing a dynamic knowledge graph.

[0045] Fig. 4A flowchart for generating the review model's base architecture. Detailed Implementation

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0047] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0048] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0049] Reference Figs. 1-4 This is one embodiment of the present invention, which provides a method for digital management and control of enterprise projects based on an AI large model, including the following steps:

[0050] S1. Receive the feasibility study report and project data table, and extract text feature vectors and numerical feature matrices through a lightweight large model, and inject Laplace noise to generate an encrypted feature vector set.

[0051] S1.1. Receive the feasibility study report and project data table, and perform SHA-256 integrity verification and format compliance review to obtain the verified feasibility study report and project data table.

[0052] Specifically, upon receiving the feasibility study report and structured data table, the SHA-256 hash algorithm is used to perform a one-way hash on the entire content of the feasibility study report, generating a fixed-length hash value to ensure the integrity of the source of the feasibility study report. At the same time, a format compliance review is performed on the structured data table, checking the field naming conventions, data type consistency, completeness of required fields, and whether the table structure meets the template requirements. For example, the field "Project Number" must be a string type with a length of 12 digits, numeric fields must not contain non-numeric characters, and missing values ​​must be represented by a standard null identifier. Through item-by-item verification, it is ensured that the structured data table meets the input conditions for subsequent processing. After integrity verification and format compliance review, the verified feasibility study report and structured data table are output.

[0053] S1.2. Call the lightweight large model DeepSeek-Coder to parse the verified feasibility study report in segments, and generate text feature vectors through the Transformer encoder. Simultaneously extract the core fields from the verified project data table to construct a numerical matrix. After dimensionality reduction by PCA, obtain the numerical feature vectors.

[0054] Specifically, the validated feasibility study report and structured data tables are accessed, and the lightweight large model DeepSeek-Coder is used to segment and parse the feasibility study report. The Transformer encoder inside the DeepSeek-Coder model processes the segmented text and generates text feature vectors. At the same time, core fields, such as total investment amount and expected rate of return, are extracted from the validated structured data tables and organized into a numerical matrix. Principal component analysis (PCA) is then applied to the numerical matrix to reduce its dimensionality and obtain numerical feature vectors.

[0055] It should be noted that the training of the lightweight large model DeepSeek-Coder adopts a two-stage paradigm of pre-training and supervised fine-tuning. In the pre-training stage, a large-scale code and natural language corpus is used. The corpus includes open-source code from GitHub, technical documents, and Stack Overflow question and answer text. The training objective is to perform autoregressive language modeling based on the Transformer architecture. By calculating the prediction loss of the next word and using the AdamW optimizer for gradient descent, the lightweight large model DeepSeek-Coder is trained to master programming language syntax, project-related terminology, and basic logical reasoning ability. In the supervised fine-tuning stage, a specially constructed project analysis and review instruction dataset is used, which includes paired samples of manually annotated feasibility study report fragments and corresponding review comments. For example, if the input is "This project plans to use quantum computing technology", the expected output is "The technology is highly cutting-edge, but the team's technical reserves and supply chain risks need to be evaluated."

[0056] During training, the parameters of the bottom Transformer layer of the lightweight large model DeepSeek-Coder are frozen, and only the top layer and task-specific output headers are fine-tuned. The cross-entropy loss function and a small learning rate, such as 0.00001, are used to adapt the lightweight large model DeepSeek-Coder to downstream tasks such as feasibility study report parsing and structured feature extraction, resulting in the trained lightweight large model DeepSeek-Coder.

[0057] S1.3. Inject Laplacian noise into the text feature vector and numerical feature vector to generate noisy feature vectors, and then encrypt sensitive fields twice using the SM4-CBC algorithm and add timestamps to output the encrypted feature vector set.

[0058] Specifically, Laplace noise that meets differential privacy requirements is injected into text feature vectors and numerical feature vectors. The scale parameter of the Laplace noise is determined according to the privacy budget. For example, when the value is 1.0, the scale parameter is 1.0. Privacy protection is achieved by adding random noise sampled from the Laplace distribution to each feature vector element, generating a noisy feature vector.

[0059] The SM4-CBC algorithm is used to perform secondary encryption on the sensitive fields in the noisy feature vector. The SM4-CBC algorithm uses a 256-bit key and a random initialization vector. The encryption process includes 16 rounds of nonlinear transformation and appends the current UTC timestamp to the encrypted data packet, outputting an encrypted feature vector set.

[0060] It should be noted that sensitive fields refer to dimensions in the feature vector that can directly or indirectly identify a specific enterprise or individual or disclose core business secrets, such as enterprise name, project leader's name, precise investment amount, rare technology combination features, technology formula parameters, or financial model details.

[0061] Differential privacy requires that the presence or absence of any single data point in the output of a data query or analysis does not significantly affect the result through a rigorous mathematical mechanism (usually achieved by adding noise), thus preventing attackers from inferring information about a specific individual from the output.

[0062] S2. Aggregate the encrypted feature vector set to generate a global fusion feature vector, use an AI large model to extract project entities and associate them with historical project data tables to construct a dynamic knowledge graph.

[0063] S2.1. Aggregate the encrypted feature vector set through a secure aggregation protocol to generate a globally fused feature vector.

[0064] Specifically, through a secure aggregation protocol, the server receives encrypted feature vector sets from all participating clients. Using secret sharing technology, each client generates key fragments and uploads encrypted feature vectors. The server uses a homomorphic encryption algorithm to accumulate the encrypted feature vector sets, generating an encrypted aggregate vector. By aggregating the key fragments from all clients, the server decrypts the aggregate vectors to obtain a globally fused feature vector.

[0065] It should be noted that the secure aggregation protocol is an encryption protocol used in federated learning. Its goal is to achieve secure aggregation computation of data uploaded by multiple clients, provided that the central server cannot decrypt the data of a single client. Through secret sharing technology, each client decomposes the encryption key into fragments and stores them in a distributed manner. This allows the server to use homomorphic encryption algorithms to blindly accumulate the encrypted feature vectors uploaded by the clients (e.g., Paillier homomorphic encryption supports ciphertext addition) to generate an encrypted aggregation vector. By aggregating the key fragments of all clients, the decryption key is recovered, and the encrypted aggregation vector is decrypted to obtain the plaintext global fusion feature vector. This ensures that the server cannot reverse-engineer the original data of any single client.

[0066] S2.2. Input the global fusion feature vector into the AI ​​large model, and automatically identify and extract project entities and attributes through semantic understanding capabilities.

[0067] Specifically, the global fusion feature vector is input into the AI ​​big model. The AI ​​big model uses the deep semantic understanding capability of the Transformer architecture to analyze the global feature information contained in the global fusion feature vector, automatically identify the feature dimensions representing the project entity (such as key attributes such as technology field and investment scale), and extract the specific entity (such as "quantum computing project") and related attribute values ​​(such as "investment amount: 120 million yuan").

[0068] S2.3. Based on the attributes of the project entity, perform association matching with the records in the historical project data table, and filter the historical project node with the highest similarity by calculating the similarity between attributes.

[0069] Specifically, the project entity's attributes are matched with records in the historical project data table. By calculating the cosine similarity between attributes (e.g., comparing the vectorized representations of core attributes such as technical field, investment scale, and team size), the historical project node with the highest similarity to the current project entity's attributes is selected from the historical project data table.

[0070] It should be noted that the expression for calculating the cosine similarity between attributes is:

[0071] ;

[0072] in, Represents the cosine similarity between attributes. This represents the attribute vector of the current project entity. An attribute vector representing a historical project node. This represents the Euclidean norm of the attribute vector of the current project entity. The Euclidean norm of the attribute vector of a historical project node.

[0073] S2.4. Take the project entity as a new node, establish a semantic relationship of derivation and optimization with the historical project node with the highest similarity, and record the association confidence to obtain the project node association result.

[0074] Specifically, the project entity is treated as a new node, and similarity is calculated based on the attributes of the project entity and the attributes of historical project nodes (for example, using the cosine similarity algorithm to compare core attributes such as technical field and investment scale). The derivative and optimized semantic relationship between the project entity and historical project nodes is determined based on the similarity score, and the similarity score is recorded as the association confidence in the relationship attribute to generate the project node association result.

[0075] It should be noted that "derived from" indicates the inheritance and development relationship between the new project and the historical project in terms of technical solutions, design ideas, or core features;

[0076] Optimization is a semantic connection in a dynamic knowledge graph that describes the iterative improvement relationship between projects, meaning that the current project has made significant improvements in performance, efficiency or technology compared to the previous project.

[0077] The project node association results include the complete attribute set of the newly added project node, the identifier of the matched historical project node, the clear semantic relationship type (such as "derived from" or "optimized from"), the numerical association confidence based on similarity calculation, the list of core attributes involved in the matching, the difference markers between the current project and historical projects, and the data source identifier. All elements are organized in a structured format to update the dynamic knowledge graph and support intelligent review and decision-making.

[0078] S2.5. Based on the project node association results, create new project entity nodes and add attribute information with timestamps to establish relationship edges between project entities; calculate and update the weight values ​​of the relationships between project nodes and project entities according to the association confidence, and continuously expand the relationship network between project entities by continuously integrating new project nodes and timestamp attributes to form a dynamic knowledge graph.

[0079] Specifically, based on the project node association results, project entities are added as new nodes to the dynamic knowledge graph, and time-stamped attribute information is added to the new nodes (e.g., recording the node creation time using UTC time format); corresponding relationship edges are created based on the derived and optimized semantic relationships established in the project node association results; based on the association confidence recorded in the project node association results, the association confidence is converted into a weight value using a linear normalization method (e.g., mapping a similarity score of 0.85 to a weight value of 0.85), and updated to the corresponding relationship attributes between project nodes and project entities, forming a dynamic knowledge graph.

[0080] S3. Using the large AI model as the backbone encoder, reconstruct the multi-task output layer and integrate the dynamic knowledge graph to build the review model base architecture.

[0081] S3.1. Encapsulate the large AI model into a backbone encoder using Docker containers, and use Kubernetes to allocate GPU resources to generate a containerized encoder entity.

[0082] Specifically, the AI ​​large model runtime environment is defined through Dockerfile, a Docker image containing the AI ​​large model and its dependent libraries is built, and the AI ​​large model is packaged into a portable container; the Docker image is deployed using Kubernetes, the GPU resource allocation strategy is configured (for example, specifying the NVIDIA GPU driver and resource limit to 4GB of video memory), and the container is scheduled to available GPU nodes through the Kubernetes scheduler to generate running container instances as containerized encoder entities.

[0083] S3.2. Based on the containerized encoder entity, a three-task parallel processing layer is constructed. Neo4jConnector is embedded in the input layer. The entity relationships of the dynamic knowledge graph are queried in real time through the Bolt protocol, and the prototype of the fusion model is obtained through graph neural network transformation.

[0084] Specifically, based on the containerized encoder entity, a three-task parallel processing layer architecture is constructed, encompassing financial feasibility assessment, technical risk identification, and compliance checks. The input layer of this architecture pre-embeds the Neo4jConnector component interface for subsequent runtime queries. The architecture supports querying dynamic knowledge graphs via the Bolt protocol and integrates GraphSAGE functionality to convert query results into dense vector representations. This architecture also supports concatenating and fusing the dense vector representations with the project feature vectors output by the containerized encoder entity to form multimodal feature representations. The fusion model architecture generates multi-task prediction results through its internal three-task parallel processing layer, thus completing the initial construction of the fusion model.

[0085] It should be noted that the training process of the graph neural network GraphSAGE adopts an unsupervised learning approach based on graph structure. Node sequences are generated from the dynamic knowledge graph through random walk sampling, constructing positive sample pairs (adjacent nodes) and negative sample pairs (random non-adjacent nodes). The node embedding representation is iteratively updated using GraphSAGE's neighbor aggregation function (e.g., mean aggregator or LSTM aggregator). The loss function is optimized using negative sampling techniques to maximize the similarity of positive sample pair node embeddings in the vector space (measured by cosine similarity) while minimizing the similarity of negative sample pairs. The graph contrast loss is minimized using the stochastic gradient descent algorithm and the Adam optimizer, and the weight parameters of GraphSAGE are updated to obtain the trained graph neural network GraphSAGE.

[0086] It should also be noted that the financial feasibility rating task in the three-task parallel processing layer outputs the probability distribution of financial risk level through a fully connected layer and a softmax activation function, and uses a regression layer to predict key indicators such as internal rate of return.

[0087] The technology risk identification task adopts a multi-label classification architecture, and uses the sigmoid activation function to output an independent probability value for each technology risk in the preset risk label library;

[0088] The compliance inspection task uses a binary classification network structure, employs a sigmoid activation function to output compliance probabilities, and combines threshold judgments to generate Boolean compliance conclusions. The three tasks are processed in parallel to generate structured review results.

[0089] S3.3. Encapsulate the prototype of the fusion model as a gRPC microservice, define a standardized prediction interface, configure a two-layer elastic scaling strategy, and form the base architecture of the review model.

[0090] Specifically, the prototype of the fusion model is encapsulated as a gRPC microservice, and a standardized prediction interface (including ProjectRequest request message and MultiTaskResponse response message structure) is defined using Protocol Buffers. In the Kubernetes environment, an elastic scaling strategy for the number of instances based on Horizontal Pod Autoscaler (e.g., setting the CPU utilization threshold to 70%) and an elastic scaling strategy for resource quotas based on Vertical Pod Autoscaler (e.g., dynamically adjusting the GPU memory allocation range to 4GB-16GB) are configured to form the base architecture of the review model.

[0091] S4. Using a dynamic knowledge graph, Gaussian noise is added to the aggregated gradient during the gradient descent phase to train the review model base architecture in multiple tasks, thus obtaining a knowledge privacy-preserving review agent.

[0092] S4.1. Based on dynamic knowledge graphs, entity relationship triples are extracted using Cypher query statements, and then converted into feature tensors using graph neural networks to form a knowledge-enhanced training dataset.

[0093] Specifically, based on the dynamic knowledge graph, the Cypher query is used to extract entity relationship triples. The nodes and relationships in the triples are converted into low-dimensional dense feature tensors through the graph neural network GraphSAGE. The feature tensors are then fused with the original project data to form a knowledge-enhanced training dataset.

[0094] S4.2. Distribute the global parameters of the review model base architecture to all clients participating in the training. Each client uses local data and knowledge-enhanced training dataset to train and backpropagate the review model base architecture to obtain local gradients.

[0095] Specifically, the global parameters of the review model's base architecture are distributed to all participating clients via an encrypted channel. Each client uses local data and a knowledge-enhanced training dataset to perform forward propagation calculations on the review model's base architecture. The financial feasibility rating loss, technical risk identification loss, and compliance check loss are calculated separately through a multi-task output layer, and then weighted and summed to obtain the joint loss of the multi-tasks. The partial derivatives of the joint loss of the multi-tasks with respect to all parameters of the review model's base architecture are calculated using a backpropagation algorithm to obtain the complete local gradient.

[0096] It should be noted that local data is stored in the local storage environment of each participating client and is not private project data such as feasibility study reports, structured project data tables, historical review records, and operational data that have not been uploaded to the central server.

[0097] S4.3. Decrypt the local gradient and calculate the average gradient, add Gaussian noise to the average gradient, update the review model base architecture, and obtain the knowledge privacy-preserving review agent.

[0098] S4.3.1. Receive all local gradients uploaded by participating clients via encryption, perform joint decryption using a secure aggregation protocol to obtain plaintext local gradients, and obtain the global average gradient through weighted averaging.

[0099] Specifically, the system receives all local gradients uploaded by participating clients in encrypted form, performs joint decryption using a secure aggregation protocol, collects all client encrypted gradients, decrypts them using a distributed key to obtain plaintext local gradients, and performs a weighted average based on the number of local data points of each client to obtain the global average gradient.

[0100] S4.3.2. Calculate the Gaussian noise intensity based on the privacy budget, generate the corresponding noise vector, and inject it into the global average gradient to obtain the noisy gradient.

[0101] Specifically, based on the privacy budget parameter, the standard deviation of Gaussian noise is calculated; a noise vector with the same dimension as the global average gradient is generated by sampling from the Gaussian distribution; the noise vector is added to the global average gradient to obtain the noisy gradient.

[0102] It should be noted that the privacy budget parameter is the core control indicator for differential privacy, and it is set by the data governance party based on GDPR, the privacy-utility trade-off curve of the business scenario, and industry best practices.

[0103] It should be noted that the expression for calculating the standard deviation of Gaussian noise is:

[0104] ;

[0105] in, Indicates privacy budget, This represents the allowed probability of failure, a low-probability event where privacy might be compromised. Indicates the gradient clipping threshold. The standard deviation of Gaussian noise is represented by... This represents an empirical constant used for tightening tail boundary estimation. This represents the natural constant in the tail decay of the Gaussian distribution.

[0106] It should be noted that the gradient clipping threshold is determined by an empirical initial value (e.g., ...). =1.0), dynamically adjusted based on the actual distribution of the gradient norm during training: if the gradient norm is consistently much lower than 1.0. Then lower the threshold to reduce excessive noise; if the gradient frequently approaches or exceeds... To avoid over-pruning, the threshold is increased. It is determined by grid search (e.g., testing C=0.5, 1.0, 2.0) and combined with the privacy-utility trade-off curve to achieve a balance between differential privacy protection and model performance. An exemplary value range is 0.1 to 10.0.

[0107] S4.3.3. Based on the noisy gradient, the global parameters of the review model base architecture are updated through the stochastic gradient descent algorithm, and multiple iterations are performed to solidify and save the global parameters of the review model base architecture, generating a knowledge privacy-protected review agent.

[0108] Specifically, based on the noisy gradient, the global parameters of the review model's base architecture are updated using the stochastic gradient descent algorithm. The process involves adjusting the parameters of the review model's base architecture by multiplying the noisy gradient with the learning rate. Multiple rounds of iterative training are performed until the required number of iterations is reached. In each iteration, the gradient calculation, noisy addition, and parameter update processes are repeated. The global parameters of the trained review model's base architecture are then serialized and saved as a binary file to generate a knowledge privacy-preserving review agent.

[0109] S5. Input the feasibility study report into the knowledge privacy protection review agent for content parsing, perform risk verification and historical case matching by querying the dynamic knowledge graph in real time, and generate structured review results.

[0110] S5.1. Convert the feasibility study report into UTF-8 standardized text, and desensitize sensitive fields to form an encrypted text feature vector.

[0111] Specifically, the feasibility study report is converted into standardized text in UTF-8 encoding format. Regular expressions are used to match and anonymize sensitive fields (e.g., replacing names with hash values ​​prefixed with "name" and amounts with ranges). Text features are extracted using the Transformer encoder of the lightweight large model DeepSeek-Coder to generate text feature vectors. Simultaneously, PCA dimensionality reduction is performed on the anonymized structured numerical fields to generate numerical feature vectors. The text feature vectors and numerical feature vectors are concatenated and fused, Laplacian noise is injected, and the SM4-CBC algorithm is used to encrypt the fused feature vectors, forming encrypted text feature vectors.

[0112] S5.2. Input the encrypted text feature vector into the knowledge privacy protection review agent, decrypt it using the key to obtain the plaintext item feature representation, and then perform mathematical transformation on the plaintext item feature representation through the backbone encoder to generate a deep semantic representation.

[0113] Specifically, the encrypted text feature vector is input into the knowledge privacy protection review agent, and decrypted using the corresponding key of the SM4-CBC algorithm to obtain the plaintext item feature representation. The plaintext item feature representation is then input into the backbone encoder (the Transformer architecture of the AI ​​large model), and the contextual association weights between feature vectors are calculated through a multi-head self-attention mechanism. After nonlinear transformation via a feedforward neural network and layer normalization, a deep semantic representation is generated.

[0114] S5.3. Convert the deep semantic representation into parameters of the Cypher query statement, send the query to the Neo4j graph database through the efficient Bolt protocol, and obtain the returned subgraph information in real time.

[0115] Specifically, the key dimensions in the deep semantic representation vector are mapped to the parameters of the Cypher query statement to construct a complete Cypher query statement; a Bolt protocol connection with the graph database is established through the Neo4j Connector, and the query statement is sent using a binary transmission format; the Neo4j graph database returns subgraph data in real time after executing the query.

[0116] S5.4. The returned graph subgraph information is transformed through a graph neural network to form a knowledge context vector. The vector is then independently computed using a multi-task parallel processing layer to generate structured review results.

[0117] Specifically, the returned graph subgraph information is input into the graph neural network GraphSAGE. Through neighbor node sampling and feature aggregation operations (e.g., using a mean aggregator to calculate node embeddings), a knowledge context vector representing the semantics of the graph is generated. The knowledge context vector is then concatenated and fused with the deep semantic representation output by the backbone encoder to form an enhanced feature representation.

[0118] The enhanced feature representations are input into a multi-task parallel processing layer. The financial feasibility rating task outputs rating results through a fully connected layer and a softmax activation function, the technical risk identification task outputs risk labels through a sigmoid activation function, and the compliance inspection task outputs compliance status through a binary classifier. Each task is calculated independently to generate structured review results.

[0119] It should be noted that the rating results include classification rating labels, corresponding confidence probabilities, and predicted values ​​of key quantitative indicators. The structured review results are a machine-readable, multi-dimensional JSON object containing financial feasibility ratings (such as grade "A" with a confidence level of 0.85 and a predicted investment payback period of 5.2 years), a list of technical risks (such as the risk label "technology dependence on imports" and associated evidence node IDs), compliance status (such as violation clauses ["ISO 19582 §5.2"]), and metadata such as project ID and timestamps.

[0120] S6. Perform logical verification on the structured review results, generate interpretable evidence, and add the interpretable evidence to the structured review results to obtain a review report with a chain of evidence.

[0121] S6.1. Perform deterministic logical verification on the structured review results, and at the same time perform probabilistic deep analysis through neural networks to obtain multi-dimensional verification conclusions.

[0122] Specifically, when performing deterministic logic verification on the structured review results, the business rule base is invoked to perform item-by-item rule matching, checking whether each output in the review results conforms to the hard rule constraints; through a neural network component (loading an AI large model as the basic architecture, using historical project review data to build a training dataset, and adjusting parameters through supervised fine-tuning techniques to enable the AI ​​large model to learn to identify technical logic contradictions, contextual relationships, and implicit risk patterns, forming a neural network component), the AI ​​large model is used to perform probabilistic deep analysis to evaluate the internal logical consistency, contextual rationality, and potential implicit risks of the review results; after the AI ​​large model completes the evaluation and calculation of the internal logical consistency, contextual rationality, and potential implicit risks of the review results, it outputs a confidence score scalar value representing the overall reliability of the conclusion (e.g., 0.92), and a list of strings representing implicit risk issues not covered by the business rule base as a risk list (e.g., ["budget and technical solution mismatch"]), combined with the rule compliance status output by the deterministic logic verification (e.g., field-level verification result: {"compliance": false, "violation clause": "CLAU-5.2", "field name": "financial risk level", "actual value": "low", "Constraints": "When the debt-to-asset ratio is greater than 0.7, the financial risk level must not be low"}), generating multi-dimensional verification conclusions.

[0123] It should be noted that the business rule base is a collection of hard constraints parsed from industry standards (such as ISO 19582) and corporate internal control policies, which are then digitized into structured logical rules (such as if-then expressions). These rules are stored in a database and provided with API calls for automated compliance verification of review results.

[0124] S6.2. Extract relevant historical case node attributes from the dynamic knowledge graph, and concatenate them with the rule clause text to generate interpretable evidence. Combine this with the structured review results to form a review report with a chain of evidence.

[0125] Specifically, historical case node attributes (such as project ID, technical risk type, and actual impact data) related to the current review result are extracted from the dynamic knowledge graph through Cypher query. The historical case node attributes are then combined with the corresponding rule clause text in the business rule base. An interpretable evidence paragraph with natural language description is generated through an AI big model. The interpretable evidence paragraph is then embedded as a new field into the JSON format of the structured review result to form a review report with an evidence chain.

[0126] This embodiment also provides a computer device applicable to the enterprise project digital management and control method based on AI large model, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the enterprise project digital management and control method based on AI large model proposed in the above embodiment.

[0127] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0128] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for digital management and control of enterprise projects based on an AI-driven large model, as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0129] In summary, this invention achieves distributed preprocessing and privacy enhancement of feasibility study report text and structured data by introducing a lightweight large model combined with Laplace noise injection and encryption mechanisms during the feature extraction stage. It not only enhances the learning depth of implicit relationships between project entities by leveraging the semantic association capabilities of knowledge graphs, but also significantly reduces the risk of reverse inference caused by gradient leakage through the synergistic effect of a two-layer encryption channel and a secure aggregation protocol.

[0130] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. 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 be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A digital management and control method for enterprise projects based on AI large-scale models, characterized in that: include, Receive the feasibility study report and project data table, and extract text feature vectors and numerical feature matrices through a lightweight large model and inject Laplace noise to generate an encrypted feature vector set; The encrypted feature vector set is aggregated to generate a globally fused feature vector. A large AI model is then used to extract project entities and associate them with historical project data tables to construct a dynamic knowledge graph. The specific steps are as follows: By a secure aggregation protocol, encrypted feature vector sets are aggregated to generate a globally fused feature vector. The global fusion feature vector is input into the AI ​​large model, which automatically identifies and extracts project entities and attributes through semantic understanding capabilities; The project entity's attributes are matched with records in the historical project data table. By calculating the similarity between attributes, the historical project node with the highest similarity is selected. The project entity is treated as a new node, and a semantic relationship is established between it and the historical project node with the highest similarity. The association confidence is recorded to obtain the project node association results. Based on the project node association results, create new project entity nodes and add attribute information with timestamps to establish relationship edges between project entities; Based on the association confidence, the weight values ​​of the relationships between project nodes and project entities are calculated and updated. By continuously integrating new project nodes and timestamp attributes, the relationship network between project entities is continuously expanded to form a dynamic knowledge graph. Using the large AI model as the backbone encoder, we reconstructed the multi-task output layer and integrated dynamic knowledge graph to build the review model base architecture. By using dynamic knowledge graphs, Gaussian noise is added to the aggregated gradient during the gradient descent phase, and the review model base architecture is trained in multiple tasks to obtain a knowledge privacy-preserving review agent. The feasibility study report is input into the knowledge privacy protection review intelligent agent for content parsing. Risk verification and historical case matching are performed by querying the dynamic knowledge graph in real time to generate structured review results. The structured review results are logically validated to generate interpretable evidence, which is then added to the structured review results to obtain a review report with a chain of evidence. The specific steps are as follows: The structured review results are subjected to deterministic logical verification, while probabilistic deep analysis is performed through neural networks to obtain multi-dimensional verification conclusions. Relevant historical case node attributes are extracted from the dynamic knowledge graph and concatenated with the rule clause text to generate interpretable evidence. Combined with the structured review results, a review report with a chain of evidence is formed.

2. The enterprise project digital management and control method based on AI large model as described in claim 1, characterized in that: The process of receiving the feasibility study report and project data table, extracting textual feature vectors and numerical feature matrices using a lightweight large model, and injecting Laplacian noise to generate an encrypted feature vector set, is detailed below. Receive the feasibility study report and project data sheet, and conduct SHA-256 integrity verification and format compliance review to obtain the verified feasibility study report and project data sheet; The lightweight large model DeepSeek-Coder is called to parse the verified feasibility study report in segments, and the Transformer encoder is used to generate text feature vectors. The core fields of the verified project data table are extracted simultaneously to construct a numerical matrix. After dimensionality reduction by PCA, the numerical feature vectors are obtained. Laplacian noise is injected into text feature vectors and numerical feature vectors to generate noisy feature vectors. Sensitive fields are then encrypted a second time using the SM4-CBC algorithm and a timestamp is added to output an encrypted feature vector set.

3. The enterprise project digital management and control method based on AI large model as described in claim 1, characterized in that: The specific steps for constructing a review model base architecture by using a large AI model as the backbone encoder, reconstructing the multi-task output layer and integrating a dynamic knowledge graph are as follows. The large AI model is encapsulated into the backbone encoder using Docker containers, and GPU resources are allocated using Kubernetes to generate a containerized encoder entity. Based on containerized encoder entities, a three-task parallel processing layer is constructed. The Neo4j Connector is embedded in the input layer. The entity relationships of the dynamic knowledge graph are queried in real time through the Bolt protocol, and the prototype of the fusion model is obtained through graph neural network transformation. The prototype of the fusion model is encapsulated as a gRPC microservice, and a review model base architecture is formed by defining a standardized prediction interface and configuring a two-layer elastic scaling strategy.

4. The enterprise project digital management and control method based on AI large model as described in claim 1, characterized in that: The specific steps for obtaining the knowledge privacy protection review agent are as follows: Based on dynamic knowledge graphs, entity relationship triples are extracted using Cypher query statements, and then converted into feature tensors using graph neural networks to form a knowledge-enhanced training dataset. The global parameters of the review model base architecture are distributed to all clients participating in the training. Each client uses local data and knowledge-enhanced training dataset to train and backpropagate the review model base architecture to obtain local gradients. The local gradient is decrypted and the average gradient is calculated. Gaussian noise is added to the average gradient, and the review model base architecture is updated to obtain the knowledge privacy-preserving review agent.

5. The enterprise project digital management and control method based on AI large model as described in claim 4, characterized in that: The process of decrypting the local gradient and calculating the average gradient, adding Gaussian noise to the average gradient, and updating the review model's base architecture to obtain the knowledge privacy-preserving review agent is detailed below. Receive all local gradients uploaded by participating clients in encrypted mode, perform joint decryption using a secure aggregation protocol to obtain plaintext local gradients, and obtain the global average gradient by weighted averaging. The Gaussian noise intensity is calculated based on the privacy budget, a corresponding noise vector is generated, and it is injected into the global average gradient to obtain the noisy gradient. Based on the noisy gradient, the global parameters of the review model's base architecture are updated using the stochastic gradient descent algorithm, and multiple iterations are performed to solidify and save the global parameters of the review model's base architecture, generating a knowledge privacy-protected review agent.

6. The enterprise project digital management and control method based on AI large model as described in claim 1, characterized in that: The specific steps for generating structured review results are as follows: The feasibility study report was converted into UTF-8 standardized text, and sensitive fields were de-identified to form an encrypted text feature vector; The encrypted text feature vector is input into the knowledge privacy protection review agent, decrypted using a key to obtain the plaintext item feature representation, and then mathematically transformed by the backbone encoder to generate a deep semantic representation. The deep semantic representation is converted into parameters of the Cypher query statement, and the query is sent to the Neo4j graph database through the efficient Bolt protocol, and the returned subgraph information is obtained in real time. The returned graph subgraph information is transformed through a graph neural network to form a knowledge context vector. The vector is then independently computed using a multi-task parallel processing layer to generate structured review results.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the enterprise project digital management and control method based on AI large model as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the enterprise project digital management and control method based on AI large model as described in any one of claims 1 to 6.

Images