A document management risk intelligent analysis method and system based on multi-source data fusion

Abstract

This invention discloses a document management risk intelligent analysis method and system based on multi-source data fusion, relating to the fields of artificial intelligence and data fusion processing. The method includes: acquiring multi-source heterogeneous data in parallel around the clock and performing independent encoding preprocessing to output a multimodal initial feature tensor set; extracting entities from unstructured text as anchor nodes of a knowledge graph and attaching high-frequency structured temporal features as dynamic attributes of the corresponding nodes to construct an initial graph structure; dynamically updating and reconstructing the graph using topology reconstruction based on semantic drift and local multi-hop message passing based on topological distance; guiding a multi-gated expert network based on graph features to perform collaborative computation of risk assessment and opportunity value identification; modeling the strategy generation process as a Markov decision process and using a deep reinforcement learning algorithm with soft actors and commentators to output an executable sequence of operation control instructions. This effectively solves the problems of fragmented environmental data, lagging static association mechanisms, and lack of forward-looking inference, comprehensively improving the intelligence and executability of risk management strategies.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and data fusion processing, and in particular to a document management risk intelligent analysis method and system based on multi-source data fusion. Background Technology

[0002] Improving enterprise operational security and profitability highly depends on the effective management of document and business risks. Currently, risk analysis mainly involves multi-source heterogeneous data such as structured data, unstructured text, and unstructured images. In existing technologies, traditional methods have obvious limitations. For long-term macro texts such as quarterly maintenance reports and high-frequency micro values ​​such as millisecond-level SCADA voltage fluctuations, existing technologies have serious mismatch problems in terms of time span and semantic space. This directly leads to the fragmentation of environmental data and makes it impossible to achieve deep fusion and cross-alignment of cross-modal data.

[0003] To overcome the limitations of single data sources and modality mismatch, some fusion analysis schemes have emerged in recent years. However, these existing traditional methods still face the following difficulties in practical applications: existing graph or correlation analysis techniques mainly adopt static correlation mechanisms, which have obvious lag and are difficult to adapt to the high-frequency dynamic updates and response requirements of massive data in industrial application scenarios; in the risk assessment and decision-making stage, existing technologies lack forward-looking extrapolation capabilities and cannot balance the trade-off between avoiding equipment safety risks and seizing market opportunities; in addition, in traditional multi-task learning networks, when the system processes different tasks simultaneously, risk warning and benefit optimization often experience gradient conflicts with opposite optimization directions, such as safe shutdown instructions and instructions to maximize power generation, and are prone to negative transfer problems, leading to contamination of model features.

[0004] In summary, existing technologies still suffer from pain points when processing multi-source heterogeneous data, such as fragmented environmental data, lagging static correlation mechanisms, lack of forward-looking analysis, and gradient conflicts in traditional networks. There is an urgent need for a document management risk intelligent analysis method and system that can achieve deep integration of multi-source data. Summary of the Invention

[0005] The main objective of this invention is to provide a document management risk intelligent analysis method and system based on multi-source data fusion, which solves the problems of fragmented environmental data, lagging static correlation mechanisms, and lack of forward-looking analysis.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a document management risk intelligent analysis method based on multi-source data fusion, which includes the following steps: S1. Acquire multi-source heterogeneous data in parallel around the clock, perform independent encoding preprocessing, and output multimodal initial feature tensor sets; S2. Cross-modal data fusion to construct an initial graph structure containing a set of nodes and a set of edges; S3. The topology reconstruction based on semantic drift and the local multi-hop message passing based on topological distance are used to dynamically update and reconstruct the graph, and the updated macroscopic graph features and local subgraph feature tensors are output. S4. A customized multi-gated expert network collaborative evaluation model based on graph features performs collaborative computation for risk assessment and opportunity value identification. S5. Model the strategy generation process as a Markov decision process, and use a deep reinforcement learning algorithm with soft actors and critics to output an executable sequence of operation control instructions.

[0007] In the preferred embodiment, in step S1: We acquire multi-source heterogeneous data in parallel around the clock through API interfaces, web crawlers, and local database connections; the multi-source heterogeneous data is divided into structured datasets, unstructured text datasets, and unstructured image datasets. For structured datasets, a sliding time window and standardized mapping are used to extract temporal feature vectors, as shown in the formula: (1); in, For a moment Structured temporal feature vectors; For a moment The original structured numerical matrix collected; This is the mean vector of the data within the historical sampling window; This is the variance vector of the data within the historical sampling window; To prevent extremely small constraint constants with a denominator of zero; It is a learnable linear mapping weight matrix; It is the bias vector; It is a non-linear activation function; Indicates splicing; Output standardized structured features For alignment purposes; For unstructured text datasets, a pre-trained Natural Language Processing (NLP) deep encoder is used to extract semantic feature vectors: (2); in, For the first A high-dimensional semantic feature vector extracted from an unstructured text using a self-attention mechanism; For the first The initial word embedding matrix of the text; This is the position encoding matrix for the corresponding text sequence; This represents the multi-head self-attention computation function; Presentation layer normalization operation; Output text semantic features For use in extracting entities; For unstructured image datasets, spatial feature vectors are extracted using computer vision (CV) convolutional neural networks: (3); in, For the first Spatial feature vectors of an image flattened after convolutional pooling; This is the original input pixel matrix; This is a two-dimensional convolution operation; For batch normalization operations; It is a linear rectification activation function; Output image features Provide additional node attributes for the map.

[0008] In the preferred embodiment, in step S2: Extract the semantic features of the text output in step S1 The key entities in the text are parsed into anchor nodes of an organization-specific environmental element knowledge graph; and the high-frequency structured temporal feature vectors output from step S1 are used to... Mapping and mounting dynamic attributes as corresponding anchor nodes solves the mismatch problem between long-term macro text and high-frequency micro values ​​in terms of time span and semantic space, completes cross-modal data fusion, and constructs an initial graph structure containing a set of nodes and a set of edges. Specifically, this includes entity recognition and anchor binding mechanisms, utilizing conditional random fields or sequence labeling models from... Extract entity set Subsequently, for the initial edge set of the knowledge graph, the system extracts the initial semantic relationships between entities as edges based on the dependency parsing results of unstructured text and a pre-built industry-specific ontology library. It also establishes initial topological connections for node pairs with physical affiliation or logical influence relationships, forming the initial network topology. Finally, it calculates the structured feature flow for each edge. With the extracted entity nodes The semantic alignment confidence scores between them form the mounting mapping, and the formula is: (4); in, Indicates the first The structured feature flow belongs to the first... Probability score of each entity anchor point; and These are the query and key projection matrices for the structured space and the text entity space, respectively; For a specific structured data channel vector; From The portion extracted corresponds to the entity semantic embedding vector; The dimension of the projection space; Based on score For mapping relationships exceeding a preset alignment threshold, structured data is concatenated as attributes into node features to construct an initial knowledge graph. : (5); in, For a moment The knowledge graph node set consists of all anchor entities extracted from unstructured text, i.e. ; For a moment The knowledge graph edge set represents the initial semantic or topological association between each anchor node; For a moment The set of feature tensors of all nodes in the entire graph; For the map Specific nodes At any moment The complete fusion feature representation; This represents a vector concatenation operation; This indicates that the node has been successfully mapped. A set of structured feature stream indexes; This is a subset of image features associated with the entity; if none exists, it is a zero vector. When integrating unstructured image feature sets into node features, the system accurately assigns and mounts the corresponding image feature tensors to the text entity nodes with the highest correlation based on the physical layout position of the image or chart in the original unstructured document and the cross-modal semantic correlation between the chart title and the surrounding paragraphs, thus completing the cross-alignment of visual and textual features; the output is the initial topology of the current time-to-time graph. and fusion node characteristics It will be used for dynamic reconstruction calculations.

[0009] In the preferred embodiment, in step S3: To balance the millisecond-level response requirements for massive data in industrial scenarios with the control of underlying computing power costs, a topology reconstruction based on semantic drift and a local multi-hop message passing based on topological distance are adopted to dynamically update and reconstruct the graph, wherein: For unstructured policy and report data with low update frequency, topology reconstruction based on semantic drift employs a node threshold triggering mechanism based on high-dimensional feature cosine similarity. When the entity feature update caused by the new round of policy text parsed and entered into the database in step S1 meets the following conditions, the global calculation of the macro-structure edge weights is triggered: (6); in, For macroscopic entity nodes semantic drift distance; The characteristics of newly connected nodes after mapping according to step S2; Features of historical baseline nodes; It is an L2 norm; The preset macro-semantic drift trigger threshold; If triggered, recalculate the result. All connected edges Attention weight matrix; For SCADA structured attribute data with high update frequency, local multi-hop messaging based on topological distance is used. When the absolute value or rate of change of dynamic attributes attached to entity nodes exceeds a safety preset limit, a graph neural network (GNN) is used to perform local ripple-like state updates, limiting the impact to adjacent subgraphs with a specific number of hops. (7); in, micro nodes In the The hidden state vector after message passing in a subgraph neural network; during initialization. ; For nodes The set of first-order neighbor nodes; For the first The graph convolution weight matrix of the layer; The degree of the node; The edge attention coefficient; This is the topological distance attenuation factor; Let be the number of hops on the shortest path in the graph structure, when That is, when the maximum number of hops is reached, the transmission path is cut off; After reconstruction, the system [processes the mapping]. Apply pooling operations to output the macroscopic environmental context embedding feature vector of the entire image. and local subgraph embedding feature vectors related to specific business scenarios : (8); in, Embed feature vectors into the output subgraph; A subset of the graph nodes involved in a specific analysis task; To read out the multilayer perceptron; This represents the final hidden state of the node. and As a priori representation, it is input into S4.

[0010] In the preferred embodiment, in step S4: To eliminate gradient conflicts and negative transfer problems, the underlying model is decoupled into a risk-specific expert group, an opportunity-specific expert group, and an environment-sharing expert group. The working mechanism and mathematical model are defined as follows: In expert networks, define A dedicated network of risk experts A dedicated network of experts and opportunities Each environment shares an expert network; each expert network is a nonlinear mapping function containing several fully connected layers. For direct input, define the first... A risk expert output ;No. Opportunity Expert Output ;No. Shared expert output ; In the dynamic routing gating guided by knowledge graph environment representation, the graph features dynamically updated in step S3 are... As a priori guiding signal, it is concatenated with the input data and then dynamically calculated through the corresponding task gating network to assign weights to different expert networks. The graph-guided gating mechanism for the risk assessment task is as follows: (9); in, The weight allocation vector for the risk task gating output; A learnable projection matrix for risk gating; Indicates splicing; Temperature hyperparameters for controlling the degree of gated polarization; The function ensures that the sum of the elements of the weight vector is 1; Changes in the macro-level graph environment are translated into a mathematical implementation that allocates underlying computing resources to specific risk experts; similarly, the output of the graph-guided gating mechanism for opportunity mining tasks is... ; In feature fusion and multi-task collaborative output, the outputs of each expert network are weighted and summed using a gated weight vector, and then fed into a specific task tower to generate the final prediction. The calculation formula is as follows: (10); (11); in, The model outputs a risk index vector containing specific occurrence probabilities and impact levels. A vector of opportunity value quantification indicators output by the model; and These are top-level decision-making multi-layer perception modules for risks and opportunities; The scalar weights assigned to a specific expert in the gating vector; To further isolate conflicts, during the model backpropagation iterative optimization phase, asymmetric gradient constraints are set based on the current true labels: (12); in, The objective loss function for overall network optimization; Binary cross-entropy loss for risk identification; The mean squared error loss for opportunity value prediction; Labeled as authentic history; and These are task weight coefficients, derived from global graph features. Dynamically determined; To square the Frobenius norm, an orthogonal constraint penalty term is introduced to force the risk expert feature matrix. With shared expert feature matrix Spatially orthogonal to prevent feature contamination at the shared layer; After the model is trained, output in real time. and Simultaneously, the SHAP value analysis algorithm is embedded in parallel in this step to calculate the marginal contribution of each input feature to risk warning based on the model output, providing mathematical support for interpreting fault prediction.

[0011] In the preferred embodiment, in step S5: Risk index for receiving collaborative assessment results With opportunity value and map features The strategy generation process is modeled as a Markov decision process, defined as a tuple. Through deep reinforcement learning algorithms involving soft actors and critics, an executable sequence of operation control instructions is output. First, define the intelligent agent. The state space observed at each time step is a composite tensor. ; Ensure the agent perceives the entire environment; Define the agent's action space as a continuous or discrete vector of operation instructions. For example, adjusting the percentage reduction of wind turbine speed, modifying the threshold for photovoltaic inverter switching, and setting the charging and discharging power of energy storage devices; Next, to balance the trade-off between avoiding safety risks and seizing market opportunities when issuing control actions, a reward function dynamically adjusted by the graph is established: (13); in, In the state Next action The scalar reward value obtained immediately; and To input full-spectrum features The nonlinear mapping function is used to output the dynamic weight adjustment factor; The benchmark return coefficient; The benchmark risk penalty coefficient; This is the safety boundary threshold; This indicates a corrected linear calculation; When knowledge graph indicators are in a period of stringent policy compliance. When the output increases dramatically, the intelligent system prioritizes outputting conservative control commands. To avoid penalties for exceeding limits.

[0012] Finally, the SAC algorithm is used to find the optimal randomization strategy. To maximize the expected sum of cumulative dynamic reward and policy entropy: (14); in, For the strategic objective functional; For mathematical expectation calculation; In strategy Trajectory distribution below; Temperature is a parameter used to weigh exploration against utilization. To output the entropy of the action distribution; The algorithm employs a dual-critic network trained in parallel to predict action value, and an actor network to generate actions. The actor network updates its parameters according to the following formula to improve the action output: (15); in, For actor network parameters The gradient operator; For experience replay pool; To utilize reparameterization techniques to extract Gaussian noise Action generation function for mid-sampling; A Q-network with smaller output values ​​is used in the dual-critic network to alleviate the overestimation problem.

[0013] Through multiple iterations of reinforcement learning, the system ultimately bases its current... State, through optimal policy Output specific command actions Combined with the SHAP interpretation value output by S4, and simultaneously... The action deduction logic and execution suggestions are filled into the preset text model or the private large language model interface is called to generate a comprehensive analysis report including a dynamic risk dashboard.

[0014] This invention also provides a document management risk intelligent analysis system based on multi-source data fusion, comprising: The multi-source heterogeneous data acquisition and preprocessing module is equipped with API interfaces and a crawler engine to acquire internal company documents, equipment and environmental market data. It also has a built-in feature mapping algorithm array to clean, standardize and generate multimodal feature tensors of structured text and images. The graph structure alignment and cross-modal fusion module executes graph structure alignment technology based on entity anchors, uses NLP tools to extract unstructured text entities to construct core anchors, and uses a time-series mapping algorithm to mount high-frequency structured data to the corresponding anchors to build the underlying knowledge graph. The hierarchical graph reconstruction strategy management module has a built-in dual graph update engine. It triggers global graph updates based on macroscopic semantic drift cosine similarity and performs local subgraph adaptive pruning and weight adjustment based on microscopic attribute out-of-bounds and GNN multi-hop message passing mechanism. The customized multi-gated expert collaborative evaluation module, namely the CGC collaborative processing hub, includes a parallel distributed risk expert network group, opportunity expert network group, and shared expert network group. It uses graph features to generate gating vectors to dynamically guide the flow and calculation of different data, and adopts an asymmetric joint loss function to control the training. The dynamic reward-based reinforcement learning decision control module deploys an agent decision machine based on the SAC framework. It transforms the output of the upstream evaluation model into the state space of the agent and constructs a dynamic reward and penalty coefficient mechanism using graph features. It outputs operable work orders and control instructions and outputs real-time charts and early warning reports based on the business engine and SHAP values.

[0015] This invention provides a document management risk intelligent analysis method and system based on multi-source data fusion. It acquires heterogeneous data from multiple sources, including structured and unstructured text and images, in parallel around the clock, and performs independent encoding preprocessing to generate a multimodal initial feature tensor set. An entity recognition and anchor binding mechanism is used to construct the initial graph topology, accurately mapping and attaching high-frequency structured temporal features and image features to entity nodes parsed from macroscopic unstructured text. This effectively solves the mismatch problem between long-period macroscopic text and high-frequency microscopic numerical values ​​in terms of time span and semantic space, achieving deep fusion and cross-alignment of cross-modal data. In the graph reconstruction and update stage, it innovatively combines global topology reconstruction based on macroscopic semantic drift with a microscopic local multi-hop message passing mechanism based on topological distance, perfectly balancing the millisecond-level response requirements of massive data in industrial application scenarios with the control of underlying computing power costs. In the model evaluation stage, the underlying network is... The system decouples expert networks into risk-specific, opportunity-specific, and environment-shared expert groups. By introducing spatial orthogonal constraint penalty terms to control training, it eliminates gradient conflicts and negative transfer problems in the model optimization process, prevents feature pollution at the shared level, and provides highly interpretable mathematical support for risk and failure prediction through parallel embedded SHAP value analysis algorithms. The final process models the strategy generation process as a Markov decision process, combining a reward function adjusted by dynamic graph features with the Soft Actor Commentator (SAC) deep reinforcement learning algorithm. When issuing control actions, it accurately balances the game between avoiding security risks and seizing market opportunities, and uses a dual commentator network mechanism to alleviate overestimation problems. It solves the pain points of traditional methods such as fragmented environmental data, lagging static correlation mechanisms, and lack of forward-looking inference. Finally, it generates a comprehensive analysis report including a dynamic risk dashboard, which comprehensively improves the intelligence, forward-looking early warning capabilities, and executability of document and business risk management. Attached Figure Description

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of a document management risk intelligent analysis method based on multi-source data fusion according to the present invention; Figure 2 This is a schematic diagram of a document management risk intelligent analysis system module based on multi-source data fusion according to the present invention; Detailed Implementation Example 1 To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. In specific industrial engineering practice, this embodiment takes the multi-source heterogeneous data collaborative early warning and asset operation decision-making of new energy projects covering wind power, photovoltaic and energy storage systems as the application scenario, and provides an engineering-specific and parameter-specific description of the document management risk intelligent analysis method and system based on multi-source data fusion proposed in this invention. like Figure 1-2 As shown, a document management risk intelligent analysis method and system based on multi-source data fusion is presented, wherein: A document management risk intelligent analysis system based on multi-source data fusion includes: a multi-source heterogeneous data acquisition and preprocessing module, a graph structure alignment and cross-modal fusion module, a hierarchical graph reconstruction strategy management module, a customized multi-gated expert collaborative evaluation module, and a dynamic reward shaping reinforcement learning decision control module.

[0017] In the preferred embodiment, the multi-source heterogeneous data acquisition and preprocessing module is configured with an API interface and a crawler engine to acquire internal company documents, equipment, and environmental market data. It also incorporates a feature mapping algorithm array to clean, standardize, and generate structured, text, and image multimodal feature tensors. At the hardware and deployment level, for the acquisition of structured equipment data, this embodiment selects the Advantech ECU-1251 series edge gateway, which communicates with IEC via OPC UA. The 61850 industrial communication protocol connects to the main control PLC of the field equipment, continuously collecting SCADA time-series data such as generator speed, gearbox oil temperature, and nacelle vibration RMS values ​​at a preset sampling frequency. For unstructured data, the system deploys a data processing server in the cloud to run a distributed web crawler program, which acquires meteorological forecast data such as wind speed, wind direction, and irradiance through a preset meteorological interface according to a set period, and periodically acquires electricity market transaction data and policy documents. For image acquisition, on-site inspection images are acquired through industrial PTZ cameras or drones equipped with thermal imaging modules. In the data processing stream, the feature mapping algorithm array runs on a streaming computing framework, performing outlier removal on the time-series data and performing standardized mapping calculations based on a set sliding time window. At the same time, a pre-trained visual convolutional neural network model is called to extract high-dimensional feature tensors from infrared images, and finally outputs a standardized multimodal initial feature tensor set.

[0018] In the preferred embodiment, the graph structure alignment and cross-modal fusion module executes graph structure alignment technology based on entity anchors. It uses NLP tools to extract unstructured text entities to construct core anchors, and then uses a temporal mapping algorithm to attach high-frequency structured data to the corresponding anchors, building the underlying knowledge graph. In this embodiment, the underlying knowledge graph is stored in a graph database that supports large-scale topological relationships. This module has a built-in natural language processing pipeline, which is fine-tuned by a new energy professional technical corpus. When inputting unstructured text such as equipment maintenance manuals or logs, a named entity recognition algorithm is used to extract physical component entities from the text as core anchor nodes in the graph database. Subsequently, the temporal mapping algorithm module is activated, calculating the cosine alignment confidence between the label information of the structured data channel and the corresponding entity semantic vector, and mapping the structured dynamic numerical stream with a confidence greater than a preset threshold to the attribute fields of the corresponding entity nodes. This process achieves alignment of multi-source heterogeneous data at the physical component dimension at the database level, solving the semantic mismatch problem of cross-modal data.

[0019] In the preferred scheme, the hierarchical graph reconstruction strategy management module incorporates a dual graph update engine. It triggers global graph updates based on macroscopic semantic drift cosine similarity and performs local subgraph adaptive pruning and weight adjustment based on microscopic attribute out-of-bounds behavior in conjunction with the GNN multi-hop message passing mechanism. This module is deployed on an AI server cluster equipped with NVIDIA A100 Tensor Core GPUs and developed using the PyTorch Geometric Graph Neural Network framework. The parallel computation mechanism of the dual graph update engine is as follows: when a newly added macroscopic unstructured document, such as an extreme weather warning file, is parsed by the model, and the change in the cosine similarity between its semantic feature vector and the historical baseline features exceeds a preset semantic drift threshold, the system triggers a global environmental context vector. The system performs global topology reconstruction operations. When a micro-attribute triggers an update condition where, for example, the effective vibration value of a wind turbine gearbox exceeds the set safety envelope for multiple consecutive sampling periods, the system activates the local update engine, driving the GNN to perform feature aggregation and multi-hop message passing on the local subgraph within the graph topology. The system adaptively reduces or filters out edge weights of irrelevant data through an attention mechanism, concentrating computational resources on the abnormal device subgraph nodes and outputting the updated local subgraph feature tensor. This mechanism ensures effective feature extraction while avoiding redundancy and computational overload in global map calculations caused by high-frequency abnormal numerical fluctuations.

[0020] In the preferred scheme, the customized multi-gated expert collaborative evaluation module, namely the CGC collaborative processing hub, includes parallel distributed risk expert network groups, opportunity expert network groups, and shared expert network groups. It utilizes graph features to generate gating vectors to dynamically guide the flow and computation of different data, and employs an asymmetric joint loss function to control training. In terms of system architecture, this hub is deployed within a high-performance inference acceleration framework, with its underlying structure consisting of multiple expert networks based on multilayer perceptrons (MLPs). These include risk expert networks for handling tasks such as mechanical fatigue and electrical short circuits, opportunity expert networks for maximizing power generation and revenue, and shared expert networks for extracting environmental parameter fluctuation patterns. During forward inference, the graph-guided gating network receives dynamically updated global features. With local features The system dynamically generates corresponding weight allocation vectors. When a specific anomaly occurs, the gating network automatically increases the weight of the feature data flowing into the corresponding risk expert network and decreases the weight of the unrelated expert network by outputting the polarized allocation vector. During the offline training and iterative update phase of the model, a spatial orthogonality constraint penalty term with preset weights is introduced. This constraint term is integrated into the asymmetric joint loss function. Through constraint calculation, it forces the risk expert feature extraction matrix and the shared expert feature extraction matrix to remain orthogonal in the spatial vector, preventing data pollution and negative transfer caused by the extraction of minor anomaly risk features by the conventional shared environment features.

[0021] In the preferred embodiment, the dynamic reward shaping reinforcement learning decision control module deploys an agent decision machine based on the SAC framework. It transforms the output of the upstream evaluation model into the agent's state space and constructs a dynamic reward and penalty coefficient mechanism using graph features. This mechanism outputs operable work orders and control commands, and, based on the business engine and SHAP values, outputs real-time charts and early warning reports. This embodiment constructs soft actor and commentator agent models based on a deep reinforcement learning framework. In application, the agent acquires a high-dimensional composite state space including risk index assessment results, opportunity value quantification indicators, and environmental graph features. The dynamic reward function mechanism dynamically adjusts the baseline reward coefficient and safety risk penalty coefficient according to the environmental graph features. For example, when receiving a macroscopic graph vector containing features of severe meteorological disasters, the reward function adaptively increases the weight of the safety risk penalty coefficient, guiding the agent to generate control actions biased towards equipment safety protection through the policy network. The control action corresponds to a power-limiting operation or shutdown risk avoidance command issued to the edge coordination controller. At the same time, the system's built-in business engine quantifies and extracts the weight distribution of key factors affecting the current risk decision based on the SHAP values ​​of each input feature vector output by the model. Finally, the system encapsulates the control command deduction logic, the probability of risk occurrence, and the feature contribution, and generates a comprehensive analysis and early warning report with quantitative mathematical support, which is then sent to the operation and maintenance terminal to improve the interpretability and reliability of the model early warning results and control strategies in actual industrial applications.

[0022] This invention provides a document management risk intelligent analysis method based on multi-source data fusion, the method comprising the following steps: S1. Acquire multi-source heterogeneous data in parallel around the clock, perform independent encoding preprocessing, and output multimodal initial feature tensor sets; S2. Cross-modal data fusion to construct an initial graph structure containing a set of nodes and a set of edges; S3. The topology reconstruction based on semantic drift and the local multi-hop message passing based on topological distance are used to dynamically update and reconstruct the graph, and the updated macroscopic graph features and microscopic subgraph feature tensors are output. S4. A customized multi-gated expert network collaborative evaluation model based on graph features performs collaborative computation for risk assessment and opportunity value identification. S5. Model the strategy generation process as a Markov decision process, and use a deep reinforcement learning algorithm with soft actors and critics to output an executable sequence of operation control instructions.

[0023] In the preferred scheme, in step S1: the system acquires multi-source heterogeneous data in parallel around the clock through API interfaces, web crawlers and local database connections; the multi-source heterogeneous data is divided into structured datasets, unstructured text datasets and unstructured image datasets. In the new energy industrial environment, the structured datasets mainly come from the SCADA systems of wind turbines and photovoltaic inverters, meteorological station sensors and real-time electricity price waveforms in the electricity trading market. For structured datasets, a sliding time window and standardized mapping are used to extract time-series feature vectors. The calculation logic is based on formula (1), with the sliding time window size set to 10 minutes to capture the short-term dynamic response of wind turbine yaw or inverter MPPT tracking. The core variables in the formula include time. Structured temporal feature vectors and time The original structured numerical matrix collected To eliminate numerical bias caused by different sensor dimensions, the mean vector of data within the historical sampling window is used. With variance vector Standardization was implemented; to prevent program crashes caused by zero variance due to wind turbine shutdown or sensor disconnection, a minimal constraint constant was introduced. ; learned linear mapping weight matrix Bias vector and nonlinear activation functions After processing, perform the splicing operation. The final output is the standardized structured features. For alignment purposes; For unstructured text datasets, the sources include regular operation and maintenance logs of wind farms, dispatch instructions issued by the power grid, and the latest carbon emission policy documents; the system uses a pre-trained RoBERTa model, such as one fine-tuned based on industrial corpora, to extract semantic feature vectors, and the calculation process is referenced in formula (2); by introducing a position encoding matrix The text retains the chronological order of maintenance steps or the logical hierarchy of policy provisions. After processing with a multi-head self-attention function and layer normalization, the text semantic features are output. For subsequent extraction of equipment and environmental entities; in addition, for unstructured image datasets such as infrared thermal images of photovoltaic panels or RGB images of cracks on the surface of wind turbine blades taken by drone inspections, spatial feature vectors are extracted using computer vision convolutional neural networks, and the mathematical expression is referenced in formula (3); through cascaded feature extraction of two-dimensional convolution operations, batch normalization operations and linear rectified activation functions, complex background noise is filtered out, and highly robust image features are output. Provides additional node attributes for the knowledge graph.

[0024] In the preferred embodiment, step S2 involves extracting the semantic features of the text output from step S1. The key entities in the text are parsed into anchor nodes of an organization-specific environmental element knowledge graph; and the high-frequency structured temporal feature vectors output from step S1 are used to... Mapping and mounting dynamic attributes as corresponding anchor nodes solves the mismatch problem in time span and semantic space between long-term macro text such as quarterly maintenance reports and high-frequency micro values ​​such as millisecond-level SCADA voltage fluctuations, completes cross-modal data fusion, and constructs an initial graph structure containing a set of nodes and a set of edges. Specifically, this includes entity recognition and anchor binding mechanisms; the system utilizes conditional random fields or sequence labeling models to... Extract entity set For example, it can identify nodes such as "No. 3 wind turbine main bearing", "photovoltaic array B" or "State Grid peak shaving policy"; for the edge set of the initial knowledge graph, the system extracts the initial semantic relationship between entities as the connection edge based on the dependency parsing results of unstructured text and the pre-built new energy industry basic ontology library, and establishes the initial topology connection for node pairs that have physical attribution such as "main bearing" belonging to "No. 3 wind turbine" or logical influence relationship such as "peak shaving policy" affecting "energy storage charging and discharging strategy", thus forming the initial network topology structure; Subsequently, each structured feature flow is calculated using formula (4). With the extracted entity nodes Semantic alignment confidence scores between modules are used to form a mounting mapping; this is to achieve cross-modal attitude metrics. and These are used as query and key projection matrices for the structured space and text entity space, respectively, to represent specific structured data channel vectors. With corresponding entity semantic embedding vector The dimension of the projected space when mapped to the same latent space. The dimension is set to 256 to balance expressive power and computational efficiency; based on the score using formula (5). For mapping relationships exceeding a preset alignment threshold, structured data is concatenated as attributes into node features to construct an initial knowledge graph. ,in Let be the set of nodes consisting of all extracted anchor points. Indicates a relationship. For nodes The complete fusion feature representation, which is composed of text embeddings Structured feature flow set and image feature subsets It is pieced together; When integrating unstructured image feature sets into node features, the system allocates and mounts the corresponding image feature tensors to the text entity nodes with the highest correlation based on the physical layout of the UAV image in the inspection document and the cross-modal semantic correlation between the chart title and the context paragraphs, thus completing the cross-alignment of visual and textual features; the final output is the initial topology of the current time-to-time map. and fusion node characteristics It will be used directly for high-frequency dynamic reconstruction calculations.

[0025] In the preferred scheme, in step S3: the system uses topology reconstruction based on semantic drift and local multi-hop message passing based on topological distance to dynamically update and reconstruct the graph, taking into account the millisecond-level response and computing cost control requirements of large amounts of data in industrial scenarios. For unstructured policy and market analysis report data with low update frequency, the system adopts a node threshold triggering mechanism based on high-dimensional feature cosine similarity to perform topology reconstruction based on semantic drift; when the entity feature update caused by the new round of policy text parsed into the database in step S1 satisfies the condition of formula (6), the global calculation of the macro-structure edge weights is triggered, where For macroscopic entity nodes The semantic drift distance is calculated by determining the features of the newly accessed node. Features of historical baseline nodes The L2 norm between them is realized; in the application of the new energy market, the preset macroscopic semantic drift trigger threshold is used. The value is set to 0.15. The system will recalculate the distance only when the new power grid trading rules cause a substantial change in the meaning of a node, i.e., the distance is greater than 0.15. All connected edges The attention weight matrix is ​​used to save the computational power consumption of meaningless global graph convolution; Local multi-hop message passing based on topological distance is mainly aimed at SCADA structured attribute data with high update frequency. When the absolute value or rate of change of the dynamic attribute mounted on the entity node exceeds the safety preset limit, a local ripple-like state update is performed using a graph neural network. The update logic is shown in formula (7). This mechanism limits the impact to adjacent subgraphs with a specific number of hops, where micro nodes In the The hidden state vector after message passing in the subgraph neural network; through edge attention coefficients. Graph convolution weight matrix and node degree Through comprehensive calculations, abnormal signals will be transmitted to surrounding physical or logically related nodes, such as from the "gearbox" to the "generator"; to further prevent excessive information smoothing and computing power explosion, a topological distance attenuation factor is introduced. Set to 0.8, and when the shortest path hop count on the graph structure Greater than the maximum number of jumps At that time, the transmission path will be forcibly truncated; After dynamic reconstruction, as shown in formula (8), the system reconstructs the spectrum. Applying pooling operations to read out the multilayer perceptron Process the final hidden state of the node Output the macroscopic environmental context embedding feature vector of the entire image. And specific business scenarios such as the embedding feature vectors of local subgraphs related to a single wind farm. This is used as a priori representation input into S4 to guide subsequent steps.

[0026] In the preferred scheme, in step S4: In traditional multi-task learning networks, risk warning and benefit optimization often encounter gradient conflicts, such as the optimization directions of safe shutdown instructions and maximum power generation instructions being opposite. In order to eliminate this gradient conflict and negative transfer problem, the underlying model is creatively decoupled into risk-specific expert groups, opportunity-specific expert groups and environment-sharing expert groups. Architecturally, the system includes A dedicated network of risk experts, such as equipment fatigue prediction experts and power grid curtailment prediction experts. Opportunities such as high-electricity-price-period discharge experts, over-generation reward experts, dedicated expert networks, and An environment-sharing expert network for extracting common device health baseline features; Each expert network is a nonlinear mapping function containing several fully connected layers, with... For direct input, output respectively , and In the dynamic routing gating guided by knowledge graph environment representation, the system dynamically updates the subgraph features in step S3. As a priori guiding signal, it is concatenated with the input data and dynamically calculated through the corresponding task gating network to determine the weight allocation vectors of the input data flow to different expert networks. Specifically, the graph-guided gating mechanism for the risk assessment task is implemented based on formula (9), and the output risk task gating weight allocation vector is... Controlled by learnable projection matrix Temperature hyperparameter , Setting it to 0.5 increases the determinism of routing decisions; the Softmax function ensures that the sum of the elements of the weight vector is 1. This mathematically translates changes in the macro-level graph environment into a priority of computing resources and data flow towards specific risk experts. Similarly, the output of the graph-guided gating mechanism for opportunity mining tasks is... ; In the feature fusion and multi-task collaborative output stage, the outputs of each expert network are weighted and summed using gating weight vectors, and then fed into a specific top-level decision multi-task tower through formulas (10) and (11). and Generate the final prediction; the model's final output is a risk index vector containing the specific probability of occurrence and the degree of impact. and opportunity value quantification index vector ; To further isolate conflicts during the backpropagation iterative optimization phase, the system sets asymmetric gradient constraints based on the current true labels and calculates the overall objective loss function using formula (12). Binary cross-entropy loss including risk identification Mean squared error loss of chance prediction At the same time, it introduces features derived from global graphs. Dynamically determined task weight coefficients and Furthermore, by introducing a penalty term based on the square of the Frobenius norm. The loss surface is forced to ensure that the feature matrices of risk experts and shared experts are spatially orthogonal, fundamentally preventing feature contamination at the shared level; after the model is trained and deployed, the output is provided in real time. and Meanwhile, the SHAP value analysis algorithm is embedded in parallel in this step, which calculates the marginal contribution of each input feature to risk warning based on the model output, providing mathematical support for on-site operation and maintenance engineers to explain the fault prediction logic.

[0027] In the preferred scheme, in step S5: the decision control module receives the risk index of the collaborative evaluation result. With opportunity value and map features The strategy generation process is modeled as a Markov decision process, defined as a tuple. And through the soft actor and critic (SAC) deep reinforcement learning algorithm, it outputs an executable sequence of operation control instructions; First, define the intelligent agent in The state space observed at each moment is a composite tensor This ensures that the agent perceives both the long-term and short-term overall picture of the environment, as well as its risk-opportunity preferences, and defines the agent's action space as a continuous or discrete vector of operation instructions. In the context of new energy, these actions are specifically manifested as a percentage reduction in wind turbine speed, modification of photovoltaic inverter cut-in threshold, and continuous setting of energy storage device charging and discharging power, etc.; in order to accurately balance the trade-off between avoiding equipment safety risks and seizing the high premium opportunity in the electricity market when issuing control actions, the system establishes a reward function represented by the graph through formula (13), in which, In the state Next action The scalar reward value obtained immediately; and To input full-spectrum features A nonlinear mapping function is used to output dynamic weight adjustment factors in real time; combined with the benchmark return coefficient. Benchmark risk penalty coefficient And to ensure safety, it is set and safety boundary thresholds This enables adaptive strategy adjustments; for example, when the knowledge graph indicates that the current period is a stringent frequency response policy compliance assessment period for the power grid, The output will increase dramatically. To avoid violating network protocols and incurring huge fines, the intelligent agent will prioritize outputting conservative control commands. To avoid penalties for exceeding limits; Subsequently, the SAC algorithm is used to find the optimal randomization strategy. The objective functional is to maximize the expected sum of cumulative dynamic reward and policy entropy. The structure is shown in formula (14); the entropy of the strategy distribution is introduced. With temperature parameters To prevent the agent from getting stuck in local optima and to maintain the exploration of unknown control ranges; in the network update stage, the value of actions is estimated by training a dual-commentator network in parallel, while the actor network updates its own parameters according to formula (15). To improve action output; the action generation function is sampled from Gaussian noise using a reparameterization technique, and combined with a Q-network that outputs smaller values ​​in a dual critic network. This alleviates the common value overestimation problem in the traditional DDPG algorithm; Through multiple rounds of reinforcement learning and interactive iteration with the digital twin environment, the system ultimately determines the optimal strategy based on the current composite state. Output specific physical layer commands and actions To improve the clinical and logical fit in industrial settings, the system combines the SHAP interpretation value output by S4 with the quantified risk index. Enter the action deduction logic and execution suggestions into the preset report template or call the privately deployed large language model interface to generate a comprehensive analysis report containing a dynamic risk dashboard.

[0028] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A document management risk intelligent analysis method based on multi-source data fusion, characterized in that, The method includes the following steps: S1. Acquire multi-source heterogeneous data in parallel around the clock, perform independent encoding preprocessing, and output multimodal initial feature tensor sets; S2. Extract entities from unstructured text in multi-source heterogeneous data as anchor nodes of the knowledge graph, and map and mount high-frequency structured time-series features in multi-source heterogeneous data as dynamic attributes of the corresponding anchor nodes to complete cross-modal data fusion and construct an initial graph topology structure containing a set of nodes and a set of edges. S3. Reconstruct the initial graph topology based on the update characteristics of multimodal data. Specifically, global graph reconstruction is triggered by semantic changes in unstructured data, and subgraph reconstruction is triggered by changes in structured attribute data and local multi-hop message passing. Output the reconstructed macro-environment context vector and local subgraph feature tensor. S4. Based on the feature tensor of the local subgraph, a gated vector is generated to dynamically guide the input data flow to the mutually decoupled risk assessment expert network group, opportunity identification expert network group and environment sharing expert network group, and collaboratively calculate and output the risk index vector and the opportunity value quantification index vector. S5. Construct the agent's state space using the risk index vector, opportunity value quantification index vector, and macro-environmental context vector. Dynamically adjust the reward function based on the macro-environmental context vector to balance risk and opportunity. Output the state control parameters and operation instruction sequence for the target execution node through reinforcement learning algorithm.

2. The document management risk intelligent analysis method based on multi-source data fusion according to claim 1, characterized in that, In step S1, multi-source heterogeneous data is acquired in parallel around the clock, and independent encoding preprocessing is performed to output a multimodal initial feature tensor set, specifically including: Multi-source heterogeneous data are categorized into structured datasets, unstructured text datasets, and unstructured image datasets. For structured datasets, a sliding time window is used to extract time series data, and the mean vector and variance vector of the data within the historical sampling window are used to perform standardized mapping processing to output a structured feature stream; For unstructured text datasets, a pre-trained natural language processing deep encoder is used, combined with a positional encoding matrix and a multi-head self-attention computation function to extract high-dimensional semantic feature vectors. For unstructured image datasets, a computer vision convolutional neural network is used to extract spatial feature vectors through two-dimensional convolution operations, batch normalization operations, and linear rectified activation functions. The structured feature stream, high-dimensional semantic feature vector, and spatial feature vector are combined into a multimodal initial feature tensor set.

3. The document management risk intelligent analysis method based on multi-source data fusion according to claim 1, characterized in that, In step S2, mapping and attaching the high-frequency structured time-series features in the multi-source heterogeneous data as dynamic attributes of the corresponding anchor nodes specifically includes: Specific structured data channel vectors and corresponding anchor node semantic embedding vectors are mapped to the same latent space through query and key projection matrices; Calculate the semantic alignment confidence score between the structured feature flow and the anchor node in the latent space; Based on the mapping relationship where the semantic alignment confidence score is greater than the preset alignment threshold, the corresponding structured data is concatenated as attributes into the node features to construct the initial graph topology.

4. The document management risk intelligent analysis method based on multi-source data fusion according to claim 1, characterized in that, In step S3, the global reconstruction of the graph is triggered based on semantic changes in unstructured data, specifically including: A node threshold triggering mechanism based on high-dimensional feature distance is adopted to calculate the distance parameter between the node features parsed from the newly accessed document and the historical baseline node features; When the distance parameter crosses the preset macro-semantic drift trigger threshold, a global network topology operation is triggered to recalculate the attention weight matrix of all edges connected to the corresponding entity node.

5. The document management risk intelligent analysis method based on multi-source data fusion according to claim 1, characterized in that, In step S3, the subgraph reconstruction based on changes in structured attribute data triggering local multi-hop message passing specifically includes: When the absolute value or rate of change of the dynamic attribute attached to the anchor node exceeds the safety preset limit, the graph neural network is used to perform feature aggregation and multi-hop message passing of the local subgraph on the graph topology. During multi-hop message passing, the hidden state vector of the micro-node is updated by combining the node degree, edge attention coefficient and preset topological distance decay factor, and the passing path is truncated when the shortest path hop count on the graph structure is greater than the maximum limit hop count. Pooling operations and readout multilayer perceptron processing are applied to the reconstructed graph to output macroscopic environment context vectors and local subgraph feature tensors.

6. The document management risk intelligent analysis method based on multi-source data fusion according to claim 1, characterized in that, In step S4, a gating vector is generated based on the local subgraph feature tensor to dynamically guide the input data flow to the mutually decoupled risk assessment expert network group, opportunity identification expert network group, and environment sharing expert network group. This collaboratively calculates and outputs the risk index vector and the opportunity value quantification index vector, specifically including: The local subgraph feature tensor is used as a priori guiding signal and concatenated with the input data. Then, a task gating network with temperature hyperparameter is introduced to dynamically generate the weight allocation vector of the input data flow to different expert network groups. The outputs of the risk assessment expert network group, the opportunity identification expert network group, and the environment sharing expert network group are weighted and summed using a weighted allocation vector. The weighted summed features are fed into the risk decision task tower to generate a risk index vector, and into the opportunity decision task tower to generate an opportunity value quantification index vector.

7. The document management risk intelligent analysis method based on multi-source data fusion according to claim 1, characterized in that, In step S4, the risk assessment expert network group and the environment sharing expert network group use an asymmetric joint loss function for iterative optimization training. The joint loss function includes a binary cross-entropy loss for risk identification, a mean squared error loss for opportunity prediction, and a regularization loss based on a spatial orthogonality constraint penalty term. Among them, the spatial orthogonality constraint penalty term is constructed based on the square of the Frobenius norm. It is used to force the risk assessment expert feature extraction matrix and the environmental shared expert feature extraction matrix to remain orthogonal in the spatial vector on the loss surface, so as to prevent the environmental shared features from causing data pollution to the risk feature extraction.

8. The document management risk intelligent analysis method based on multi-source data fusion according to claim 1, characterized in that, In step S5, the reward function is dynamically adjusted based on the macro-environmental context vector to balance risk and opportunity. A reinforcement learning algorithm is used to output state control parameters and operation instruction sequences for the target execution node. Specifically, this includes: A nonlinear mapping function is used to process the macroscopic environmental context vector, and a dynamic weight adjustment factor is output in real time. By combining the preset benchmark return coefficient, safety boundary threshold, and benchmark risk penalty coefficient, the reward function of the reinforcement learning algorithm is adaptively adjusted using a dynamic weight adjustment factor. We employ a soft actor and critic deep reinforcement learning algorithm that includes a dual critic network and an actor network. Based on maximizing the expected value of the cumulative dynamic reward and policy entropy, we search for the optimal random policy. In the network update stage, we use reparameterization techniques to sample and generate actions from Gaussian noise, and output the state control parameters and operation instruction sequence for the target execution node.

9. A document management risk intelligent analysis system based on multi-source data fusion, characterized in that, include: The multi-source heterogeneous data acquisition and preprocessing module is used to acquire multi-source heterogeneous data in parallel around the clock, perform independent encoding preprocessing, and output a multimodal initial feature tensor set; The graph structure alignment and cross-modal fusion module is used to extract entities from unstructured text in multi-source heterogeneous data as anchor nodes of the knowledge graph, and to map and mount high-frequency structured temporal features in multi-source heterogeneous data as dynamic attributes of the corresponding anchor nodes, thereby completing cross-modal data fusion and constructing an initial graph topology structure containing a set of nodes and a set of edges. The hierarchical graph reconstruction strategy management module is used to reconstruct the initial graph topology based on the update characteristics of multimodal data. Specifically, the global reconstruction of the graph is triggered by semantic changes in unstructured data, and the subgraph reconstruction of local multi-hop message passing is triggered by changes in structured attribute data. The reconstructed macro-environment context vector and local subgraph feature tensor are output. A customized multi-gated expert collaborative evaluation module is used to generate gating vectors based on local subgraph feature tensors, dynamically guide the input data flow to mutually decoupled risk assessment expert network groups, opportunity identification expert network groups, and environment sharing expert network groups, and collaboratively calculate and output risk index vectors and opportunity value quantification index vectors. The dynamic reward-forming reinforcement learning decision control module is used to construct the agent's state space by combining the risk index vector, the opportunity value quantification index vector, and the macro-environmental context vector. Based on the macro-environmental context vector, the reward function is dynamically adjusted to balance risk and opportunity. Through reinforcement learning algorithms, the module outputs state control parameters and operation instruction sequences for the target execution node.

10. The document management risk intelligent analysis system based on multi-source data fusion according to claim 9, characterized in that: The hierarchical graph reconstruction strategy management module contains a parallel macro-global reconstruction engine and a micro-local transfer engine. The macro-global reconstruction engine calculates the high-dimensional distance between the features of newly accessed document nodes and the features of historical baselines, and triggers global topology operations when the preset drift threshold is crossed. The micro-local transfer engine responds to the out-of-bounds triggering of node dynamic attributes and performs local feature aggregation and multi-hop message passing based on graph neural networks with a topology distance decay mechanism. The customized multi-gated expert collaborative evaluation module internally deploys risk assessment expert network groups, opportunity identification expert network groups, and environment sharing expert network groups in parallel. It includes a task-gated network that receives local subgraph feature tensors as prior guidance signals to dynamically generate weight allocation vectors for input data flow to different expert network groups. During the model training phase, an asymmetric joint loss function based on spatial orthogonality constraint penalty term is introduced to enforce that the feature extraction matrices of the risk assessment expert network group and the environment sharing expert network group remain orthogonal in space.

Images