A Big Data Processing Method for Precise Delivery and Intelligent Consultation of Regional Industrial Policy Information

Abstract

This invention relates to the field of computer data processing technology, specifically a big data processing method for precise delivery and intelligent consultation of regional industrial policy information. The invention acquires multimodal features of policy texts to generate a fused context vector stream, which is then encapsulated into a strategy causal graph through graph neural network deduction. It collects heterogeneous enterprise data to synthesize a heterogeneous hierarchical enterprise state-space model, extracts the graph environment vector, and rewrites the internal state transition matrix of this model. Based on the depth of the query signal, it activates specific observation ports of the model, moves the stimulated model into a sandbox to perform path sampling and convergence pruning calculations, and generates a multimodal evolution trajectory set. Finally, it extracts trajectory costs and transition probabilities, applies multi-objective frontier analysis to calculate the optimal strategy set, and outputs a visualized message. This invention breaks down the barriers between policy semantic parsing, dynamic matching of enterprise states, and forward-looking simulation deduction, providing users with quantitative and dynamic optimal strategy decision support.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, and specifically to a big data processing method for precise delivery and intelligent consultation of regional industrial policy information. Background Technology

[0002] In the current process of regional economic development and enterprise transformation and upgrading, industrial support policies play a crucial role. However, enterprises face several challenges in obtaining and utilizing this policy information. Policy documents are typically released in unstructured text format, with complex content, highly technical language, and are scattered across different official public release platforms. Enterprises, especially small and medium-sized enterprises (SMEs), often lack the dedicated human and material resources to systematically track and fully interpret these policies, and accurately assess their applicability and potential impact on their own businesses. This information asymmetry may cause some enterprises to miss development opportunities and also diminishes the expected effects of policies as they are transmitted to market entities.

[0003] To address these challenges, the industry has explored various solutions. Basic solutions include establishing policy information aggregation platforms to push information to enterprises through keyword searches or simple categorized subscriptions. These systems solve the problem of fragmented policy information but suffer from insufficient matching accuracy. A more advanced technical solution involves introducing natural language processing and knowledge graph technologies. For example, extracting clauses, application conditions, and support methods from policy documents into entities and relationships to construct an industry policy knowledge graph. This approach reveals the internal logical structure of policies, allowing users to perform more complex structured queries and improving the relevance of information retrieval compared to simple text matching. Furthermore, some enterprise management software or intelligent workbenches are attempting to integrate policy information modules, making policy reminders a part of enterprise operation and management.

[0004] While the aforementioned solutions alleviate the difficulty of information acquisition to some extent, they still exhibit limitations in application. Push systems based on keywords or simple classifications, failing to deeply understand the semantic connotations of policy texts and the actual needs of enterprise users, often exhibit high recall but low accuracy, burdening users with information filtering. Solutions based on static knowledge graphs, while capable of depicting the relationships between policy clauses, struggle to accurately match these abstract policy requirements with the diverse and dynamically changing internal operating conditions of enterprises. These systems can typically answer questions like "What conditions are required to apply for a certain subsidy?", but they struggle to address more forward-looking and dynamic strategic questions such as "Given my company's current situation, how should we plan to meet the conditions, and what are the costs and success rates of different planning paths?" Existing technological solutions suffer from gaps in data models and analytical capabilities in three key areas: deep analysis of policy semantics, dynamic coupling of policy impact and enterprise status, and future-oriented simulation and strategy optimization.

[0005] Based on the above problems, this invention aims to solve the problem that existing policy services are fragmented in terms of information analysis, enterprise matching, and forward-looking projection, making it difficult to provide quantitative and dynamic decision support. Summary of the Invention

[0006] In view of this, in order to solve the problems mentioned in the background technology above, a big data processing method for accurate delivery and intelligent consultation of regional industrial policy information is proposed.

[0007] The objective of this invention can be achieved through the following technical solution: This invention provides a big data processing method for accurate push and intelligent consultation of regional industrial policy information, including: S1, acquiring the original unstructured policy text data stream, extracting text layout and semantic multimodal features and performing cross-modal feature splicing to generate a fused context vector stream.

[0008] S2 integrates the context vector stream input graph neural network engine to perform dependency inference, extracts policy entity nodes bound to state variable selector encoded signals and global policy environment vectors, and then encapsulates them into a policy causal graph.

[0009] S3 collects raw heterogeneous data streams from multiple frequency bands of enterprises and performs time-domain classification and prediction processing. It maps and constructs the set of enterprise state variables at the bottom, middle and top levels, and connects the trigger feedback transmission pins to synthesize a heterogeneous hierarchical enterprise state space model.

[0010] S4 extracts the global policy environment vector from the policy causal graph as the optimization intervention parameter, and uses the transformation logic to rewrite the state transition matrix coefficients of cross-domain exchanges within the state space model of heterogeneous hierarchical enterprises in the standby state.

[0011] S5 acquires query interaction signals containing targeted policy document identifier parameters, retrieves the matching strategy causal graph, and decodes the associated specific state variable selector encoding signals, thereby deeply activating specific observation ports and the characteristics of the observed state variables within the heterogeneous hierarchical enterprise state space model.

[0012] S6 moves the stimulated heterogeneous hierarchical enterprise state space model into the parallel joint sandbox cache area, introduces policy causal graph approval nodes and sets termination boundaries, initiates continuous path sampling exploration and convergence pruning calculation, and generates a multimodal evolution trajectory set with loss labels.

[0013] S7 extracts the cost and resource consumption library of the multimodal evolution trajectory set and the single-step transition achievement probability array, applies the core of multi-objective frontier analysis to calculate the optimal policy frontier dataset, and converts and packages it into a multi-objective balancing data message for external transmission.

[0014] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention constructs a hybrid coding network and uses visual and semantic dual channels to process unstructured policy texts in parallel. The visual receiving channel in this method can analyze the text layout, table structure and chapter hierarchy to generate a layout structure vector; the semantic receiving channel simultaneously performs autoregressive coding on the text content to generate a text semantic vector. By fusing these two vectors through a cross-modal attention mechanism, the system can associate the semantic information of the text with its structural context in the document. This mechanism enables the model to distinguish between the main text description, the title, and the key constraints in the table or list, thereby preserving the structured information of the original document in the initial stage of machine understanding, improving the accuracy of identifying key information fragments such as application conditions and restrictions in the policy text, and providing higher quality input for subsequent causal relationship extraction.

[0015] (2) This invention designs a docking mechanism between a strategy causal graph and a heterogeneous hierarchical enterprise state space model. It achieves precise mapping of policy requirements to specific state variables within the enterprise through a state variable selector encoding signal. When constructing the strategy causal graph, the system not only identifies the dependencies between policy clauses but also binds address codes pointing to specific enterprise operating indicators to the connection edges representing causal relationships. When a user initiates a consultation, the system can parse this string of address codes from the graph based on the queried policy and directly activate the corresponding observed variables in the enterprise state space model. This design breaks down the barriers between macro-level policy language and micro-level enterprise operational data, transforming a generalized policy consultation request into a focus on and observation of a specific set of quantifiable performance indicators within the enterprise, thus achieving an automated and precise association from "policy text" to "enterprise state."

[0016] (3) This invention uses the global policy environment vector extracted from policy texts as tuning parameters to modulate the dynamic evolution of the heterogeneous hierarchical enterprise state-space model. This method can quantify the macro signals such as the intensity of support and the orientation of support reflected in the policy texts and transform them into corrections to the internal state transition matrix of the enterprise model. An encouraging policy environment will enhance the state transition coefficients representing benign growth in the model, and vice versa. This mechanism ensures that the simulation is not conducted under static and invariant assumptions, but rather that the external policy environment is endogenized into the model as a variable affecting enterprise development. This allows subsequent simulation calculations to be performed under the physical change laws that conform to policy orientation, thereby improving the realism and reference value of the simulation results.

[0017] (4) This invention employs boundary-constrained simulation calculations and Pareto multi-objective frontier analysis to provide users with a quantified set of alternative strategies rather than a single answer. The system uses the final policy approval condition as the search endpoint and leverages the Monte Carlo path sampling algorithm to explore multiple evolutionary trajectories capable of achieving the objective within a vast space of possible actions, recording the time and resource consumption of each trajectory. Subsequently, through Pareto frontier analysis, it selects the optimal strategy combination that constitutes the best trade-off in multiple dimensions such as cost, time, and success rate. This method transforms complex simulation results into a clear set of optimal action plans with well-defined cost and benefit expectations, enabling users to choose based on their risk preferences and resource endowments, and transforming data-driven analysis conclusions into practically guiding business decisions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. 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.

[0019] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0020] Figure 2 This is a timing diagram showing the Kalman filter smoothing comparison of multi-band heterogeneous data in this invention.

[0021] Figure 3 This is a scatter plot of the Monte Carlo tree search trial evolution trajectory and polarization pruning of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1 This invention provides a big data processing method for precise delivery and intelligent consultation of regional industrial policy information, including: S1, acquiring the original unstructured policy text data stream, extracting text layout and semantic multimodal features and performing cross-modal feature splicing to generate a fused context vector stream.

[0024] In a specific embodiment of the present invention, the process of obtaining the original unstructured policy text data stream, extracting text layout and semantic multimodal features and performing cross-modal feature concatenation to generate a fused context vector stream includes: receiving the original unstructured policy text data stream transmitted from an external network and inputting it into a hybrid coding network.

[0025] Start the parallel and independent visual reception channels and semantic reception channels in the hybrid coding network.

[0026] The visual position and layout features of the page are extracted through the visual reception channel to form spatial structure attributes, thereby generating layout structure vectors. At the same time, the full text content is extracted through the semantic reception channel and autoregressive dimensionality reduction is performed to generate text semantic vectors.

[0027] A cross-modal attention interaction graph is created to establish a cross-domain fusion module, which uses layout structure vectors as query criteria and text semantic vectors as key-value pairs and aggregates them side-by-side in the cross-domain fusion module.

[0028] The cross-domain fusion module performs weight ratio correction and multi-dimensional coupling operation extraction at the dual-end spatial level, outputs a joint feature sequence that cancels out the interference of chapter information interruption, and reassembles it into a fusion context vector stream.

[0029] Specifically, the first implementation step S1 of the method of this invention aims to transform the original policy document, which lacks a unified format, into a machine-readable numerical representation containing rich contextual information. This step is performed by a hybrid coding network deployed on the server side, starting with receiving one or more unstructured policy text data streams. This data stream typically originates from government official website publishing systems or is crawled via web crawler interfaces, and its file format includes portable document format PDF, office document format DOCX, or web page format HTML. After receiving the unstructured policy text data stream, the hybrid coding network initiates two parallel and functionally decoupled sub-processing channels: a visual reception channel and a semantic reception channel. In the visual reception channel, the system first converts the input document page, such as a page of a PDF file, into a pixel matrix, i.e., an image, through a built-in rendering engine. Subsequently, a pre-trained visual transformer model, i.e., Vision-Transformer or a variant thereof, analyzes this pixel matrix. The core task of this model is not image content recognition, but rather parsing the spatial arrangement, font size, and style of elements such as text blocks, titles, tables, lists, and images, to understand the physical structural hierarchy of the document. By vectorizing these layout features, the visual reception channel generates a layout structure vector that represents the macroscopic layout and organizational rules of the document. Simultaneously, in the semantic reception channel, the system calls the text extraction module to parse the plain text content from the unstructured policy text data stream and performs necessary cleaning operations, such as removing headers, footers, and formatting noise. The cleaned text sequence is then fed into a large language model based on an autoregressive or autoencoder architecture, such as the BERT model, for deep semantic encoding. This model captures the contextual dependencies between words, sentences, and paragraphs through its internal multi-layer self-attention mechanism, ultimately compressing the core semantics of the entire text and outputting a high-dimensional text semantic vector. After obtaining the layout structure vector and the text semantic vector respectively, these two heterogeneous feature vectors are simultaneously fed into a cross-modal fusion layer. This layer employs a cross-modal attention mechanism to align features from two different sources and perform weighted integration. For example, the layout structure vector can be used as the query vector Q, and the text semantic vector as the key vector K and value vector V. Attention weights are calculated using the following formula and then weighted summed to achieve deep coupling of features.

[0030] Through this interactive process, the model learns the importance of layout features for understanding specific semantics; for example, text within a table structure is often a key application condition. This fusion process ultimately aggregates and generates a fused context vector stream that contains both the literal meaning of the text and structural logical information. This vector stream effectively overcomes the information fragmentation defects in traditional natural language processing methods caused by cross-page and cross-chapter references in policy texts, providing high-quality input for subsequent graph construction.

[0031] Unstructured policy text data streams consist of raw document data sets containing diverse content such as policy titles, body text, attachments, and tables. Hybrid encoding networks are composite deep learning architectures designed to extract information from both visual and textual dimensions simultaneously. The visual receiving channel is the branch within the hybrid encoding network dedicated to processing document visual layout. Its visual converter model is typically pre-trained on a large-scale document layout dataset such as DocLayNet to accurately identify various document layout elements. Layout structure vectors The visual receiving channel outputs a fixed-dimensional real-valued vector, such as a 768-dimensional vector, where each dimension implicitly represents a combination of one or more layout features. The semantic receiving channel is the branch in the hybrid coding network dedicated to processing text content. Its language model, such as BERT-base, typically outputs a 768-dimensional vector to facilitate subsequent feature alignment and fusion. (Text semantic vector) It is output by the semantic receiving channel and is a high-level summary of the semantic content of the entire policy text. The cross-modal fusion layer is a combination of one or more neural network layers, and its core function is to achieve the alignment and information complementarity of heterogeneous vectors. (The formula contains...) and These are the trainable weight matrix and bias vector of the fusion layer, respectively. Their parameter values ​​are learned through end-to-end training on a large corpus for downstream tasks, such as clause relationship classification. This function represents the vector concatenation operation. This represents a non-linear activation function, typically the Modified Linear Unit (ReLU), which introduces non-linear expressive power into the model. It also incorporates context vector streams. This is the final output generated in this step. It is one or a series of high-dimensional vectors that serve as input data for subsequent steps.

[0032] For example, suppose the system receives a PDF policy document titled "Several Measures to Promote the Development of the Artificial Intelligence Industry" as an unstructured policy text data stream. This document has five pages, with the third page containing a table titled "Table 1: Criteria for Identifying Core Technical Talents." First, the visual receiving channel of the hybrid coding network renders this PDF document into five images. After analysis by the visual converter model, the title, paragraphs, and the table structure on the third page are identified, and a 768-dimensional layout structure vector is generated based on this layout information. Its numerical form may be as follows Simultaneously, the semantic receiving channel extracts all text content from the file and inputs it into the BERT model to generate a 768-dimensional text semantic vector. Its numerical form may be as follows Subsequently, these two vectors are fed into the cross-modal fusion layer. Within this layer, the two vectors are first concatenated into a 1536-dimensional vector. Assuming the fusion layer outputs a 768-dimensional vector, its weight matrix... The dimension is Bias vector The dimension is After matrix multiplication Addition with bias After the calculation, a 768-dimensional intermediate result vector is obtained. For example, the calculated intermediate vector is... Finally, this intermediate vector is processed by the ReLU activation function, where all negative values ​​are set to 0 and positive values ​​remain unchanged, resulting in the final fused context vector, for example... This vector is added as an element to the fusion context vector stream and passed to step S2 for processing. Because it incorporates layout information, the vector's internal values ​​already implicitly contain the prior knowledge that "text content within tables has higher weight."

[0033] S2 integrates the context vector stream input graph neural network engine to perform dependency inference, extracts policy entity nodes bound to state variable selector encoded signals and global policy environment vectors, and then encapsulates them into a policy causal graph.

[0034] In a specific embodiment of the present invention, the fused context vector stream is input into a graph neural network engine for dependency inference, and policy entity nodes bound with state variable selector encoded signals and global policy environment vectors are extracted and then encapsulated into a policy causal graph, including: deploying a graph neural network engine with a feature graph modulation array for the received fused context vector stream.

[0035] Within the graph neural network engine, the underlying text boundary distribution patterns contained in the fused context vector stream are extracted. The category probability distribution is then used to extract multiple policy entity nodes that constitute the prerequisite conditions for the declaration business.

[0036] By calculating the relevant attention scores between adjacent policy entity nodes and determining whether they are greater than a preset connection threshold, logical connection guide edges with constraint direction attributes are established.

[0037] The state variable selector encoding signal representing the causal triggering mechanism of the mandatory judgment is retrieved from the preset guidance rule retrieval library and attached and burned into each corresponding logical connection guide.

[0038] The macro-level affixes representing support strength but without conditional constraints in the cleaned and fused context vector stream are tailored, normalized, assigned feature weights, and output as a global policy environment vector. Then, the distribution map of each policy entity node and edge information are packaged to construct a policy causal graph.

[0039] Specifically, after generating the fused context vector stream in step S1, the system proceeds to step S2 to construct a policy graph containing structured causal relationships. This step is executed by a graph neural network engine specifically designed to parse policy logic. First, each high-dimensional vector in the fused context vector stream is considered an initial node in the graph, with each vector representing the initial feature representation of the corresponding node. Next, the graph neural network engine, preferably an attention-based graph neural network model, begins iterative updates of message passing and node features on these initial nodes. In each iteration, the update of node features depends not only on the features of its neighboring nodes but also on the direct modulation of the layout structure information encoded in the fused context vector stream. Specifically, the strength or weight of message passing between two nodes is dynamically adjusted according to their layout relationship in the original document; for example, the message passing weight between sub-items belonging to the same clause is higher than that between clauses belonging to different chapters. After multiple iterations, the output layer of the graph neural network engine performs two parallel tasks. The first task is node classification, which assigns a category label to each node through a fully connected layer and a Softmax function, i.e., outputting a probability distribution representing the policy attribute to which the node belongs, such as "preconditions for application," "core quantitative indicators," or "post-support methods," thereby formally extracting and filtering multiple policy entity nodes. The second task is dependency discrimination, which analyzes the attention scores between nodes calculated by the last layer of the model, and when a node... For nodes Attention score Exceeding the preset connection threshold At that time, the system establishes a path from point to A directed connection edge, which means that the terms must be satisfied. The terms must be met first. After each connection edge is successfully established, the system immediately binds it with a key technical feature: a state variable selector encoding signal. This signal is generated by consulting a pre-defined semantic-to-encoding mapping table. It converts the causal conditions represented by the connection edge, such as "R&D investment ratio greater than 5%", into a machine-readable address encoding pointing to a specific state variable within the enterprise. Simultaneously, a separate quantification module scans the text data stream input to S1, specifically extracting words or phrases describing policy orientation and support levels that are detached from specific clauses, such as macro-level support intensity data like "strong support," "priority development," and "highest reward." This module standardizes, scores, and weights this data, ultimately projecting it to generate a normalized global policy environment vector. Finally, the system packages all extracted policy entity nodes and the connections with established directions and bound state variable selector encoding signals—these two parts together—to form the first type of data group. The global policy environment vector is then used as a top-level attribute of the graph, forming the final output product: the policy causal graph.

[0040] In one specific embodiment of the present invention, the typical range of the preset connection threshold is set between 0.65 and 0.80, with a recommended baseline value of 0.70. The setting of this threshold is primarily based on a dual balance between the mathematical and statistical characteristics of the algorithm and the topological structure of the graph: mathematically, its value range is between (0,1), and values ​​above 0.65 statistically clearly indicate a significant probability of strong logical dependency between two policy entity nodes; from a business perspective of graph construction, if the threshold is set too low, below 0.5, it will lead to an overly dense policy causal graph, incorrectly connecting a large number of weakly related clauses or even non-causal contextual noise as dependency edges, causing an "oversmoothing" problem; conversely, if the threshold is set too high, above 0.85, it will lead to an overly sparse graph, causing the actual policy declaration's pre- and post-declaration dependency paths to break.

[0041] In this embodiment, the node update process of the graph neural network engine can be described by the following formula, which reflects the modulation effect of layout information on message transmission. For any node in the graph... In its first eigenvectors of the layer The calculation process is as follows: Among them, attention coefficient The calculation introduces a format modulation term. : .

[0042] The graph neural network engine is a deep learning model running on the server side. Its preferred architecture is the Graph Attention Network (GAT), which is naturally suitable for modeling dependency strength because it can assign weights to edges. Policy entity nodes are the basic units of the graph, representing an independent, executable logical segment within the policy text. Connecting edges are directed line segments in the graph that connect policy entity nodes; their direction represents the logical order or dependency relationship. Connection threshold. is a hyperparameter whose value is set based on experience and performance on the validation set, typically between 0.6 and 0.8, to strike a balance between graph sparsity and information completeness. The state variable selector encoded signal is structured data, such as a hexadecimal string like "0x01_A03", where "0x01" might represent the set of mid-level mesoscopic state variables in the subsequent enterprise model, and "A03" is the specific address of the "asset turnover efficiency index" within that set. Its generation relies on a pre-built knowledge base maintained by domain experts. Macroeconomic support intensity data are text fragments indicating the overall tone and strength of policies. The global policy environment vector is a quantified and normalized numerical representation of this type of data, with dimensions consistent with the vector dimensions of the fused context vector stream, for example, 768 dimensions. The first data set is the collection of nodes and edges that constitute the core skeleton of the graph. The policy causal graph is the final output of this step; it is a composite data structure containing the graph's topology, edge attributes, and global environment information. In the formula, Represents a node In the The feature vector of the layer. It is a node The set of neighboring nodes. and It is the first The learnable parameter matrix and attention vector of the layer. This indicates a vector concatenation operation. It is a representation node extracted from the fused context vector stream. and A modulated scalar representing the strength of the layout association between the two elements, for example, if both are in the same table. The value is relatively high.

[0043] For example, suppose step S1 generates three fusion context vectors corresponding to the texts "the enterprise must be a high-tech enterprise", "R&D investment accounts for no less than 5%", and "a subsidy of up to 5 million yuan will be granted". ,in These three vectors are initialized as three initial nodes in the graph neural network engine. The characteristics of the node. Through iterative calculation, the engine's attention mechanism calculates the node's characteristics. right Attention score , right Attention score Assuming a preset connection threshold It is 0.7, because and The system then establishes two directed edges: and Subsequently, the system processes the edges. The condition it represents is "high-tech enterprise". By querying the mapping table, the system binds a state variable selector encoding signal to it, such as "0x02_C01". Similarly, for the edge... The system binds a coded signal representing the "ratio of R&D investment," such as "0x01_A03." Simultaneously, the quantification module identifies phrases like "strongly support" and "up to 5 million" from the original document, assigning them scores of +3 and +8 respectively according to preset rules. After standardization, a global policy environment vector is generated, for example, a 768-dimensional vector, where values ​​in corresponding dimensions are set to higher values. Finally, the node set... Edge set The first type of data group, together with the global policy environment vector, is encapsulated to form a complete policy causal graph object, which is then passed to subsequent steps.

[0044] See Figure 2 S3 collects raw heterogeneous data streams from multiple frequency bands of enterprises and performs time-domain classification and prediction processing. It maps and constructs the enterprise state variable sets at the bottom, middle and top levels, and connects the trigger feedback transmission pins to synthesize a heterogeneous hierarchical enterprise state space model.

[0045] In a specific embodiment of the present invention, the original heterogeneous data streams of multiple frequency bands of enterprises are collected and time-domain classification and prediction processing is performed. The bottom, middle and top enterprise state variable sets are mapped and the trigger feedback transmission pins are connected to synthesize a heterogeneous hierarchical enterprise state space model. This includes: opening and intercepting the original heterogeneous data streams of multiple frequency bands of enterprises captured by the input port and performing time-domain span frequency band interception with sampling periods of seconds or milliseconds, sampling periods of minutes or hours, and low-frequency extremely long axes with update periods of months or grades.

[0046] Intercepting, decomposing, and deriving low-latency underlying sensor physical parameters, mid-stage operational flow record central parameters, and slow-cycle top-level financial control parameters.

[0047] The configuration initiates an independent multidimensional Kalman sequence filtering algorithm and a nonlinear function curve fitting operator to map and process the aforementioned three types of parameters with different time delays. It independently produces bottom-level micro-smoothing energy consumption state layer variables, mid-level meso-level asset outflow efficiency state layer variables, and top-level macro-level R&D reserve state layer variables, which are then combined and listed as sets of enterprise state variables at each level.

[0048] Connect the top-level enterprise state variable set to the middle-level enterprise state variable set by linking the lead wire conduit, introduce a closed-loop threshold activation criterion, and trigger when it exceeds the preset change sensitivity threshold.

[0049] A closed negative feedback error adjustment sub-step is generated. By limiting the noise reduction fluctuation range of the model prediction through residual clipping and error covariance matrix reset, the cross-linking coupling constraint of each layer is realized. In the state prediction, the posterior estimated state of the previous moment is used as the input basis for the prior prediction of the current moment, generating a heterogeneous hierarchical enterprise state space model that characterizes the variation of system behavior.

[0050] Specifically, after constructing the policy causal graph in the preceding steps, this invention then constructs a dynamic model on the enterprise side, which is step S3. The execution entity for this step is a state-space modeling engine deployed on the enterprise's private cloud or an authorized third-party server. First, this engine continuously collects multi-dimensional raw heterogeneous data streams generated by the enterprise's internal operations through a pre-authorized secure data interface. These data streams differ significantly in time scale and data type; therefore, the engine performs time-domain segmentation and classification operations on them. Specifically, the data is divided into three independent frequency bands and sent to their respective processing pipelines. The first pipeline processes high-frequency underlying IoT sensor array data streams, such as data from production line motor current sensors and environmental temperature and humidity sensors, with sampling periods typically in the seconds or milliseconds range. The second pipeline processes mid-frequency operational middleware flow record data streams. This type of data originates from Enterprise Resource Planning (ERP) systems, Customer Relationship Management (CRM) systems, or Manufacturing Execution System (MES), recording business events such as order flow, inventory changes, and work order execution, with a generation frequency typically in the minutes or hours range. The third pipeline processes low-frequency financial audit indicator data streams from senior management. This data primarily originates from quarterly or annual financial statements, publicly available patent authorization announcements, and R&D project approval decisions, with update cycles lasting several months to a year. The engine employs targeted algorithms for processing data across different frequency bands. For time-domain slices in the high-frequency underlying IoT sensor array data stream, a multidimensional Kalman sequence filter algorithm is used to filter out measurement noise and estimate the system's true physical state. The output smoothed state sequence is further projected to generate a set of underlying micro-state variables representing the micro-energy consumption fluctuation index. For the mid-frequency operational middleware flow record data stream, the event frequency or flow rate within a statistical window is used, and an S-shaped nonlinear activation projection function is applied to map it to the [0,1] interval, transforming it into a set of mid-level meso-level state variables representing the meso-level asset flow efficiency index. For low-frequency financial audit indicator data streams from senior management, a weighted summation method is typically used to aggregate multiple key indicators, such as the R&D investment ratio and the number of new patents, into a single top-level macro-level state variable set representing the macro-level technological barrier index. The key technical feature of this step lies in establishing a seamless information feedback loop between these three levels, constructing a nonlinear closed-loop system. Specifically, the engine continuously monitors changes in the top-level macroscopic state variable set. Once this change exceeds a preset sensitivity threshold within a single update cycle, a cross-layer correction mechanism is triggered. This mechanism generates a correction gain, which is used to directly adjust the prediction residual matrix in the state prediction model followed by the middle-level mesoscopic state variable set. Specifically, this involves clipping the Kalman gain and resetting the error covariance matrix to prevent filter divergence, thereby limiting the noise reduction fluctuation range.This top-down revision reflects the immediate impact of macro-strategic decisions or major external changes on the expected operational efficiency of middle management, thereby synthesizing a heterogeneous hierarchical enterprise state-space model that can dynamically reflect the complex causal transmission within the enterprise.

[0051] In this step, the correction mechanism of the top-level macroscopic state variable set to the middle-level mesoscopic state variable set can be described by the following variant of the state-space model. For the middle-level mesoscopic state variables... Its standard prediction model is: The update of its predicted residual covariance matrix is ​​as follows: The nonlinear self-excited closed loop of this invention is embodied in the process noise covariance matrix. Regarding dynamic adjustments: Among them, dynamically scaling diagonal matrix .

[0052] The high-frequency, bottom-layer IoT sensor array data stream consists of raw time-series data from sensors in the physical world. The mid-frequency operational middleware flow record data stream is a discrete event log recording the state of enterprise business processes. The low-frequency senior management financial audit indicator data stream is structured data reflecting the enterprise's long-term strategy and financial health. The bottom-layer micro-state variable set, the middle-layer meso-state variable set, and the top-layer macro-state variable set are the sets of quantified state indicators obtained after processing these three layers of data. Kalman sequence filtering is a recursive estimation algorithm used to estimate the state of a dynamic system from a series of incomplete and noisy measurements. The nonlinear activation projection function usually refers to the Sigmoid function or the Tanh function. (Note: The last sentence about sensitivity threshold is unrelated and appears to be a separate, incomplete thought.) This is a vector used to determine whether changes in the top-level state are significant. Its value is typically set based on the statistical volatility of historical data, for example, twice the standard deviation of the historical change. The prediction residual matrix is ​​specifically defined here as the process noise covariance matrix. It represents the magnitude of the uncertainty in the model's predictions. In the formula, It is the prior state prediction vector made for the current moment based on information from the previous moment, while the right side of the equation participates in the multiplication operation. This is the posterior state estimate vector that has already been updated with observations from the previous time step. The two are connected through the intermediate state transition matrix. Establish a time-recursive relationship. It is the covariance matrix of the prediction error. It is a baseline process noise covariance matrix, representing the inherent uncertainty of the system without significant changes at the top level. Is the top-level state variable in The change over time. It is a positive definite correction gain matrix that determines the magnitude of the correction for mid-level uncertainties caused by top-level changes. Multiplying it by the nonlinear function output and taking the negative exponent transforms it into a diagonal matrix. ,pass The conjugate transformation form, from a mathematical foundation, guarantees the modulated covariance matrix. Dimensionality and strict positive definiteness. It is a nonlinear function, such as a vectorized ReLU function, which only produces a correction effect when the change in the top layer exceeds a threshold, ensuring the stability of the system. The dimensions of this formula are consistent because all matrices and vectors involved in the operation are in a dimensionless state space.

[0053] In a specific embodiment of the present invention, the change sensitivity threshold The typical value range is set between 0.05 and 0.15, with a recommended baseline empirical value of 0.10. This threshold is primarily based on a dual consideration of mathematical statistics and the stability of the hierarchical system dynamics: From a mathematical statistics perspective, this is because regular financial fluctuations have approximately a 95.4% probability of falling within this range, and changes exceeding this threshold can be rigorously defined statistically as 'significant non-random mutations'; from a system dynamics perspective, top-level macroeconomic indicators are essentially low-frequency, slowly varying parameters. If the threshold is set too low, below 0.03, the system is highly likely to misinterpret the company's regular quarterly financial fluctuations as strategic leaps, and frequent corrections will cause severe oscillations in the mediator-view state space model, preventing convergence. Conversely, if the threshold is set too high, above 0.20, it will lead to severely sluggish cross-layer feedback links, missing opportunities to correct the company's model trajectory due to significant policy benefits.

[0054] For example, sham Suppose a manufacturing company is modeled. The system collects its main production line's power consumption as high-frequency data, the hourly output of finished products from its ERP system as mid-frequency data, and R&D investment from its quarterly financial reports as low-frequency data. At time point... The mid-level mesoscopic state variable, representing an index of asset turnover efficiency, has a value of 0.8. Its baseline process noise variance... It is 0.01. The top-level macroeconomic state variable, representing the index of technological barriers, is... It is 0.6. At time point The company announced that one of its core technology patents has been authorized. The low-frequency data processing pipeline captures this information and processes the top-level macroscopic state variables. The value is updated to 0.9, therefore the change is... Assuming a preset sensitivity threshold for change. It is 0.1. Because... The cross-layer correction mechanism is triggered. Let the correction gain be... nonlinear functions Based on the improved formula for dynamically scaling diagonal matrices, the scaling factor is first calculated (in this single-state variable example, the diagonal matrix degenerates into a scalar): Subsequently, the new process noise variance is calculated through conjugate transformation. This operation rigorously guarantees the non-negativity and physical meaning of variance from a mathematical foundation: This indicates that, due to the increased technological barriers at the top level, the uncertainty in predicting mid-level operational efficiency has decreased, making the model more certain about future efficiency performance. This complete model, which includes three layers of state variables and a dynamic correction mechanism, is the generated heterogeneous hierarchical enterprise state space model, to be used in subsequent steps.

[0055] S4 extracts the global policy environment vector from the policy causal graph as the optimization intervention parameter, and uses the transformation logic to rewrite the state transition matrix coefficients of cross-domain exchanges within the state space model of heterogeneous hierarchical enterprises in the standby state.

[0056] In a specific embodiment of the present invention, the global policy environment vector in the strategy causal graph is extracted as the tuning intervention parameter, and the state transition matrix coefficients of cross-domain exchange within the heterogeneous hierarchical enterprise state space model in the standby state are rewritten using transformation logic. This includes: cutting off the node local area network instruction addressing in the strategy causal graph ontology during parsing, and focusing on extracting the global policy environment vector residing in the memory block of the head environment pool.

[0057] Activation introduces a vector matrix transformation logic gate specifically for harmonic magnitude parameters.

[0058] The global policy environment vector mapping with positive and negative macroeconomic adjustment characteristics is reduced in order and transformed into a pre-modulation parameter form. Based on the target matrix dimension of the intervened model level, a corresponding correction operator of the same dimension is dynamically generated. The optimization intervention parameters are forcibly assigned to the initializer of the heterogeneous hierarchical enterprise state space model in the idle waiting deduction sequence triggering stage.

[0059] By detecting and comparing the polarity difference between the two ends of the amplitude that is positively expanded or negatively narrowed by the intervention parameters, the coefficients of the original inherent state transition matrix of the corresponding cross-data interaction transmission surface are covered and rewritten according to the Hadamard product criterion. A stability constraint factor based on the spectral radius is introduced to perform forced contraction mapping, which changes the basic derivation law of the target and forces it to conform to the guidance and undergo a hard correction and shift of weight, while preventing the derivation from diverging.

[0060] Specifically, step S4 aims to inject macro-level policy guidance into the micro-level enterprise model, setting a realistic tone for subsequent simulation calculations. This step is executed by a central simulation scheduler. First, the scheduler retrieves a mature policy causal graph from the output of step S2. However, at this stage, the scheduler deliberately ignores the complex node connections and edge-bound state variable selector encoding signals within the graph, temporarily severing its internal micro-communication logic. The scheduler's sole objective is to access and capture the global policy environment vector stored in the top-level attribute table of the graph. This vector, as an independent tuning intervention parameter, carries quantitative information about whether the current policy environment is encouraging, neutral, or suppressive. After acquiring this vector, the scheduler initiates a pre-defined vector-matrix transformation logic gate. This logic gate is essentially an algorithm module whose function is to map the one-dimensional global policy environment vector to a correction of the high-dimensional state transition matrix. This transformation logic is not a simple linear mapping, but rather it corresponds different segments of the vector to state transition matrices at different levels in the heterogeneous hierarchical enterprise state space model. It also incorporates a dynamic tensor reshaping function to ensure that the dimension of the output correction matrix is ​​strictly identical to the dimension of the state transition matrix of the currently intervened layer. At this point, the heterogeneous hierarchical enterprise state space model is in a standby phase, and its internal parameters have not yet been used for actual deduction. During this initialization phase of the allocator, the scheduler forcibly uses the global policy environment vector carrying the pre-emptive policy warning and adjustment tendency as a tuning parameter, which is then processed through a vector-matrix transformation logic gate. The logic gate generates one or more correction matrices of the same dimension as the state transition matrix based on the amplitude of the tuning parameter (i.e., the size of the elements in the vector) and its polarity (i.e., the positive or negative sign of the element values). Finally, the scheduler applies this correction matrix to the original state transition matrix of the corresponding level within the heterogeneous hierarchical enterprise state space model, rewriting its coefficients element by element. For example, a positive, encouraging tuning parameter might strengthen the diagonal elements representing positive growth in the state transition matrix. In this way, the model's internal dynamics are modulated by the external policy environment, forcing its subsequent calculations to follow the rigid physical change laws defined by the corresponding policy issuance cycle from the very beginning, thereby completing the macroscopic calibration of the simulation environment.

[0061] In this embodiment, the process of specifically rewriting the coefficients of the internal state transition matrix of the heterogeneous hierarchical enterprise state-space model, particularly for the middle-level mesoscopic state variables, can be expressed by the following formula. Let the original middle-level state transition matrix be... The global policy environment vector extracted from the strategy causal graph is: Set the intermediate transition matrix. The final rewritten state transition matrix formula is optimized as follows: Among them, the optimized intervention parameters are the global policy environment vectors separated from the strategy causal graph. Vector-matrix transformation logic gates It is a non-linear mapping function that maps the input 768-dimensional vector. Transform it into a state transition matrix related to the target state, for example The same dimension correction operator matrix. The design of this function makes... A specific dimensional range affects the correction of a specific matrix. In engineering implementation, the initializer allocator refers to the software routine that loads and configures the model parameters before the simulation begins. The coefficients of the state transition matrix constitute the state transition matrix, such as... Each numerical element of the vector describes the degree of influence of one state variable on another state variable at the next moment. The magnitude polarity describes the characteristics of the element values ​​in the global policy environment vector; positive values ​​typically represent incentives or gains, while negative values ​​represent inhibition or attenuation. The absolute value represents the intensity of the influence. In the formula, This represents the Hadamard product, which is the element-wise multiplication of matrices. It is a matrix where all elements are 1, and its purpose is to ensure that the original matrix remains unchanged when the correction is zero. Function The output is a correction matrix subject to amplitude penalty. To fundamentally prevent the discrete dynamics system from experiencing exponential numerical divergence (i.e., state explosion) during subsequent long-step sandbox simulations, this system introduces a stability constraint factor. The system acquires the intermediate state matrix. Then, the spectral radius of the matrix will be calculated in real time. That is, the maximum absolute value of the matrix eigenvalues. The value selection logic is as follows: when hour, ,in A preset safe approximation threshold, such as 0.99; when At that time, the Lyapunov anti-collapse protection is triggered. By forcibly shrinking the mapping, the new state transition matrix of the final output is ensured. The largest eigenvalue is always strictly locked within the unit circle. The dimensions of the entire formula remain consistent because all matrices are dimensionless state transition coefficient matrices.

[0062] For example, continuing from the previous steps, the scheduler receives the policy causal graph generated in step S2 and extracts the global policy environment vector representing "strong support" from it. This vector is a 768-dimensional vector, with significant positive values ​​in a specific dimension related to industry support. Simultaneously, the scheduler retrieves the heterogeneous hierarchical enterprise state-space model generated in step S3, assuming that its layer-1 mesoscopic state variable set contains two variables: asset turnover efficiency and R&D result conversion rate, and its original state transition matrix... for: .

[0063] This indicates an asset turnover efficiency with a self-sustaining rate of 0.95 and a positive impact of 0.05 on the R&D results conversion rate. Now, the scheduler will vectorize the global policy environment. Input vector matrix transformation logic gate Assume the logic gate is designed to extract... The values ​​of dimensions 32 to 35 are used to construct a 2x2 correction matrix. Since the policy is strongly supportive, the values ​​of these dimensions are all positive; for example, logic gates... The output correction matrix is: .

[0064] Next, the scheduler calculates the new state transition matrix according to the formula. First, calculate the intermediate transition matrix. : , , .

[0065] Subsequently, the system calculates the spectral radius of the intermediate transition matrix. This is then performed through eigenvalue decomposition. Assume the system has a preset safe approximation threshold. Since 0.9995 > 0.99, the Lyapunov anti-collapse protection is triggered, and the stability constraint factor is calculated. .

[0066] Finally, a new state transition matrix is ​​obtained through forced contraction mapping. : .

[0067] After rewriting, the new state transition matrix All coefficients showed slight increases, particularly the coefficient representing the self-sustaining conversion rate of R&D results, which rose from 0.80 to 0.8241. This allows the modulated heterogeneous hierarchical enterprise state-space model to naturally exhibit a stronger endogenous growth trend in subsequent simulations, consistent with the logic of the external policy environment of "strong support." Furthermore, it benefits from... With the factors locked, the model will never experience a numerical explosion even after tens of thousands of sandbox iterations. The modulated model is now ready and awaits precise activation in step S5.

[0068] S5 acquires query interaction signals containing targeted policy document identifier parameters, retrieves the matching strategy causal graph, and decodes the associated specific state variable selector encoding signals, thereby deeply activating specific observation ports and the characteristics of the observed state variables within the heterogeneous hierarchical enterprise state space model.

[0069] In a specific embodiment of the present invention, a query interaction signal containing a targeted policy document identifier parameter is obtained, a matching strategy causal graph is retrieved, and the associated specific state variable selector encoding signal is decoded, thereby deeply activating specific observation ports and the characteristics of the observed state variables within the heterogeneous hierarchical enterprise state space model, including: connecting to an external security interface to access and lock a query interaction signal initiated by a user with an index number, i.e., a targeted policy document identifier parameter, as a targeted indication request.

[0070] Based on the extracted and decoded target policy document identifier parameters, the unique matching items stored in the data shelf are compared to extract the target strategy causal graph.

[0071] The graph is traversed sequentially and layer by layer, and the key points of the graph are separated and identified by the weighted force surface attached to the associated edge path, and the constraint-specific state variable selector encoded signal is extracted.

[0072] By introducing a built-in decoding and translation dictionary, the encoded strings obtained from the above separation are converted into addresses and corresponding address register bits for searching, and the location is pointed to the internal feature block within the inter-layer branch node that has been modified in advance during the parameter modification process and is ready to carry out simulation operation.

[0073] The execution triggers a Boolean value and activates the specific observation port for the detection and listening associated with the memory anchor point at that coordinate, plus the random port binding association configuration, to be used as the observed state variable feature of the object to be considered.

[0074] Specifically, after macroscopically modulating the overall simulation environment in step S4, the task of step S5 is to precisely activate specific observation points within the model based on specific external queries. This process is executed by a query parsing and signal mapping module. This module first receives user strategic query interaction signals sent from external virtual customer service front-ends or other human-computer interaction interfaces through a standardized API interface. This signal is not free text, but a structured data packet, which must contain a key field, namely the targeted policy document identification parameter, such as the official document number or unique ID of the policy. Upon receiving this signal, the module activates the retrieval function, using the targeted policy document identification parameter as an index to quickly search through multiple strategy graphs stored in the system, retrieving the strategy causal graph that perfectly matches the policy. After successfully retrieving the graph, the module begins deep parsing. The core task of this parsing process is to decompose the set of directed edges representing application conditions distributed along specific path edges in the strategy causal graph. The module traverses these edges, extracting the key condition dependency weight set and associated specific state variable selector encoding signals bound and burned in step S2. These encoded signals are key address probes. After extracting all relevant specific state variable selector encoding signals, the module performs the final mapping operation. It decodes and maps these encoding signals as address instructions into the internal structure of the heterogeneous hierarchical enterprise state-space model, which has been modulated by macroscopic parameters in step S4. Each encoding signal, such as "0x01_A03," is precisely decoded as an access request for a specific variable at a specific level within the model. This request has a dual effect: first, it activates and illuminates the specific observation port associated with that state variable, meaning that the variable's value will be continuously recorded and output in subsequent simulations. Second, it formally activates the characteristics of the observed state variable itself attached to this observation port, making it a key target for evaluation and intervention in subsequent simulations. Through this series of operations, the system accurately transforms a macroscopic strategic question from the user into the activation and focus on specific microscopic variables within the enterprise dynamic model.

[0075] The user strategic query interaction signal is a standardized data request, typically in JSON format, containing user identity, query time, and, most importantly, the target policy document identifier parameter. The state variable selector encoded signal, such as "0x01_A03," is designed as a hierarchical address. Its first part, such as "0x01," can be mapped to a specific layer of a heterogeneous hierarchical enterprise state space model via a lookup table, for example, to a mid-level mesoscopic set of state variables. Its second part, such as "A03," maps to a specific index or dimension of the state vector within that set, corresponding to a specific economic or physical meaning, such as an "asset turnover efficiency index." Read / write register address decoding is a metaphor in computer architecture, here representing the process of converting a symbolic address encoded signal into a specific memory or data structure access path. A specific observation port can be understood as a "probe" or "breakpoint" set in simulation software to monitor the value of a specific variable in real time during simulation runtime. The observed state variable feature is a specific state variable defined in the heterogeneous hierarchical enterprise state space model that is monitored by the probe, such as the "micro-energy consumption fluctuation index" or the "macro-technical barrier index".

[0076] For example, suppose a company user queries through the front-end interface: "How can our company apply for the Smart Manufacturing Special Support Program with the number G-2023-001?" The system receives a user strategic query interaction signal containing the targeted policy document identifier parameter "G-2023-001". The query parsing and signal mapping module immediately uses "G-2023-001" to retrieve the corresponding strategy causal graph. The module analyzes the graph and finds that its key application path contains two core dependency edges. The first edge represents "needs to pass ISO9001 quality system certification", and the state variable selector encoding signal bound to it is "0x02_F05", representing the "certification qualification completeness index" in the top-level macro state variable set. The second edge represents "the unit product energy consumption of the production line must be lower than the industry benchmark by 15%", and the encoding signal bound to it is "0x00_E12", representing the "unit product normalized energy consumption index" in the bottom-level micro state variable set. At this time, the module performs a decoding mapping operation. It decodes the code "0x02_F05" to locate the "Certification Qualification Completeness Index" state variable in the top-level model of the heterogeneous hierarchical enterprise state-space model after parameter modulation in step S4, and sets an observation point there. Simultaneously, the module decodes the code "0x00_E12" to locate the "Unit Product Normalized Energy Consumption Index" variable in the bottom-level model, and similarly sets an observation point there. Thus, the macro-level query is successfully transformed into the activation of two specific variables within the model. These two activated observation points, along with the characteristics of the observed state variables attached to them, allow subsequent simulations to focus on how to optimize these two key indicators through internal and external behavioral adjustments, thereby providing precise data support for answering users' strategic queries.

[0077] See Figure 3 S6 moves the stimulated heterogeneous hierarchical enterprise state space model into the parallel joint sandbox cache area, introduces policy causal graph approval nodes and sets termination boundaries, initiates continuous path sampling exploration and convergence pruning calculation, and generates a multimodal evolution trajectory set with loss labels.

[0078] In a specific embodiment of the present invention, the stimulated heterogeneous hierarchical enterprise state space model is moved into a parallel joint sandbox cache region, a policy causal graph approval node is introduced and a termination boundary is set, continuous path sampling and convergence pruning calculation is initiated, and a multimodal evolution trajectory set carrying loss labels is generated. This includes: instantiating and generating multiple independent isolated virtualized containers to form a parallel joint sandbox cache region, and moving the stimulated state structure with the illuminated and modified lights with imaging detection features into the cache region intact for protection to prevent mutual interference and tampering.

[0079] Pick up the node in the causal graph of the corresponding strategy that is used to indicate the final exit confirmation condition for the pass being achieved, and set it as the termination boundary of the exploration stop barrier in this forward exploration and search process.

[0080] The path deduction algorithm component based on the Monte Carlo Tree Search (MCTS) architecture is invoked to initiate an irregular leap time difference prediction extension pre-draft work with random attempts and variable fluctuations across discrete multi-order continuous branches in the replica located in the trial operation.

[0081] The preset evaluation probe captures the actual coordinate difference change offset polarization distance at the end of each time step along the specified newly excavated downcut path.

[0082] When the simulation determines that the offset polarization distance has lengthened and expanded, deviating from the leading edge limit line, and the cumulative amount of required compensation exceeds the limit budget cap threshold.

[0083] An immediate soft interrupt stop order is issued to the test kernel to forcibly terminate the test run subroutine, perform forced truncation and clear convergence actions, not allocate computing resources, and immediately backpropagate the minimum penalty reward to the upper-level probability tree node to update the confidence upper limit of the branch, preventing the invalid branch from being sampled repeatedly.

[0084] Only those that arrive and smoothly press against the termination boundary baseline, meeting the safety restrictions, are retained, and their time stamps along the way are recorded, and overdraft amounts are converted into specific expenditure dimensions and attached as markers.

[0085] By merging and assembling these paths that have undergone successful bottoming out and breakthroughs after trimming and removing redundancy, along with all the aforementioned historical input and consumption data packets, a multimodal evolution trajectory set data queue containing consumption feature identifiers is synthesized.

[0086] Specifically, after precisely activating the variables to be tested in step S5, step S6 initiates a large-scale parallel simulation process to explore feasible paths to achieve the goal. This step is executed by a built-in integrated verification and inference exploration computing power package. First, the system dynamically creates one or more mutually isolated parallel joint sandbox buffer regions for this inference task, and completely replicates the heterogeneous hierarchical enterprise state space model that has been activated in step S5, migrating it to the sandbox for subsequent dynamic strike tests, ensuring that the original model is not contaminated. Subsequently, the system retrieves the condition node "Review Passed" representing the final approval of the policy in the policy causal graph and sets it as the termination boundary baseline of the global search. After preparation, the inference exploration computing power package activates the underlying fixed Monte Carlo Tree Search (MCTS) path inference algorithm component, initiating continuous Monte Carlo branch fission prediction calculations for the current stimulated model. Specifically, the algorithm component takes the current enterprise state as the root node and the possible planning strategies as branches to expand the probability tree; when it reaches a leaf node that is not fully expanded, it uses Monte Carlo random sampling to perform a fast simulation until the set time step or termination boundary is reached, thereby exploring different behavioral variables in a huge possibility space.

[0087] During the continuous path sampling and convergence pruning process, the system employs a resource blocking and optimal retention mechanism based on multi-dimensional real-time monitoring and dual hard thresholds. Specifically, the system deploys a pre-defined evaluation probe in the sandbox for each newly derived designated excavation path. At the end of each time step in the simulation, the probe captures the actual coordinates of the current evolutionary node in real time and calculates the change in the actual coordinate difference between it and the target approved node, representing the offset polarization distance. Subsequently, the system performs a rigorous simulation judgment: when it detects that the offset polarization distance of a certain trial path not only fails to decrease but instead lengthens and expands, deviating from the leading edge limit line, and the cumulative required compensation amount on this path continues to rise, exceeding the enterprise's set budget cap threshold, the system determines that the path is out of control. In response to such invalid divergent paths, the system immediately issues an immediate soft interrupt stop order to the test kernel, forcibly terminating the trial subroutine, implementing a forced truncation and cleanup convergence action, ceasing the allocation of subsequent computing resources, and simultaneously triggering the Monte Carlo tree's forced backpropagation mechanism. A preset maximum negative reward value is propagated upwards along the truncated path to the parent node, rapidly reducing the node's selection probability in the Unified Confidence Tree (UCT) formula, thus achieving true convergence pruning. Conversely, after this rigorous filtering mechanism, the system retains only valid paths that successfully reach and smoothly reach the termination boundary baseline. For these successful paths that meet the safety restrictions, the system records their timestamps along the way and attaches the estimated overdraft amount as a specific expenditure dimension. Finally, the system merges and assembles these paths—which have undergone redundancy removal and successful bottoming-out—along with the bound historical input and consumption footprints, into a multimodal evolution trajectory set containing consumption feature stamps, and sends it to the next stage.

[0088] The parallel joint sandbox cache area is a temporary isolated memory space partitioned using virtualization technology. The quantized path sampling algorithm component is a high-level heuristic algorithm combining random sampling and decision tree search. The preset evaluation probe is a state monitoring daemon embedded in the sandbox's underlying layer. The actual coordinate difference change offset polarization distance is a comprehensive indicator quantifying the distance between the current state and the target state space (Euclidean distance). The pressure cap threshold is a financial red line rigidly set based on the company's budget limit. The immediate soft interrupt stop order is an operating system-level control signal used to forcibly terminate runaway simulations. Forced truncation and cleanup convergence actions include killing associated processes and reclaiming memory blocks. Converting overdraft funds into specific expenditure dimensions and attaching tags abstracts and encapsulates consumption in the virtual environment into a consumption tag vector.

[0089] For example, suppose a company is simulating a "subsidy for the first set of major technical equipment" policy. The approval condition for this policy, i.e., the termination boundary baseline, is "achieving the technical rating standard and sales exceeding 10 million yuan." Among the numerous fission branches initiated by the probing computing power package, the system detects that path F, at the end of the 4th time step, not only fails to increase sales, leading to a significant expansion of the actual coordinate difference shift and polarization distance, thus deviating from the leading edge limit line, but also suffers from excessive R&D, causing its accumulated required compensation amount to reach 8 million yuan, exceeding the company's 5 million yuan budget cap. At this point, the system immediately issues an immediate soft interrupt stop order, forcibly terminating the trial run subroutine of path F and reclaiming computing power. In another robust path, G, the model smoothly stays on the termination boundary baseline after 10 time steps, with costs controlled at 3 million yuan and no violation risk, meeting the security restrictions. The system ultimately retains only path G, converting its 10-month duration timestamps and 3 million yuan cost into specific cost dimensions and attaching them to it. Ultimately, successful routes and markers, such as path G, are packaged together to form a multimodal evolution trajectory set data queue.

[0090] S7 extracts the cost and resource consumption library of the multimodal evolution trajectory set and the single-step transition achievement probability array, applies the core of multi-objective frontier analysis to calculate the optimal policy frontier dataset, and converts and packages it into a multi-objective balancing data message for external transmission.

[0091] In a specific embodiment of the present invention, the cost and resource consumption database of the multimodal evolution trajectory set and the single-step transition achievement probability array are extracted. The optimal strategy frontier dataset is calculated by applying the core of multi-objective frontier analysis and converted and packaged into a multi-objective balancing data message for external transmission. This includes: retrieving and separating the cost indicators of various maintenance and promotion requirements attached to and implemented on all collected multimodal evolution trajectory sets to form a cost and resource consumption database under the evaluation constraint items.

[0092] The integrated collection and deduction process simulates the multiplication of odds results generated by jump events to construct a single-step transition probability array for the operational risk vulnerability probability safety belt of the system.

[0093] The core takeover operation of the multi-objective front analysis process performs non-dominated sorting and crowding distance calculation to find multivariate composite solutions that can avoid falling into high failure rate policy intervals, minimize cost-saving expenses, and do not constitute exclusion or domination among themselves. These solutions form the front and are output as the optimal policy front dataset of non-dominated group solutions.

[0094] Obtain the coordinates of each key node in the optimal solution derived from the optimal policy frontier dataset.

[0095] Based on the system timestamp when the user request is initiated and the model calculation limit deadline window, the reverse traceability day ratio is calculated to generate a highlighted node warning countdown prompt character field for each time node on the timeline.

[0096] By using graphical templates, the warning prompts are attached to the main structure to present a progressively evolving route leading to the final application approval confirmation endpoint. The output includes a visual report detailing the warning prompts for exceeding resource consumption thresholds, which is used for loading and shipping packages.

[0097] The multi-objective balancing data message, which compresses and assembles all output screens to form the final result form, is delivered to the main router via a push port and sent to the designated customer. The interactive screen carrier realizes the service terminal feedback closed loop.

[0098] Specifically, after generating a multimodal evolution trajectory set containing multiple feasible paths in step S6, the final implementation step S7 of this invention aims to perform multi-dimensional quantitative evaluation of these paths and generate an intuitive decision support report with early warning functions. This step is executed by the decision report generation module. This module first extracts two core data columns for each evolution trajectory from the input multimodal evolution trajectory set: the first is a cost and resource consumption library, which is attached to the trajectory nodes and describes the cost indicators such as funds and time required to put the advancement requirements into practice; the second is a single-step jump achievement probability array, which integrates and collects the multiplicative odds results generated by each jump event during the simulation process, forming a probabilistic safety belt for evaluating the operational risks and vulnerabilities of the system.

[0099] After data extraction, the module drives the Pareto multi-objective front analysis core to take over the computational process. This core algorithm performs non-dominated sorting and crowding-distance calculations on the cost and probability data. Non-dominated sorting is used to divide all evolutionary trajectories into different front levels, while crowding-distance calculation is used to maintain solution diversity within the same front level. This identifies multivariate composite solutions that avoid falling into high-failure-rate policy zones, minimize cost-saving expenses, and do not mutually exclude or dominate each other. These non-dominated solutions constitute the optimal policy front dataset, representing several optimal alternatives selected by the user.

[0100] After generating the optimal solution using pure numerical values, the system uses a window rendering interface to perform planar reprinting and visualization transformation. For each optimal solution in the optimal strategy frontier dataset, the rendering interface first extracts the coordinates of each key node in the decision transformation, i.e., each key application milestone. Subsequently, the system calculates the reverse look-back day ratio based on the system timestamp when the user request is initiated and the final deadline window limited by the model calculation. Based on this ratio, the system generates highlighted node warning countdown prompts at each time node on the timeline and attaches radar alarm layer animations to highlight the urgency of resource preparation. Next, the system uses graphical templates to attach these warning prompts to the main structure, synthesizing a recursive roadmap chart showing a progressive evolution trend that ultimately leads to the final application approval confirmation endpoint. Thus, the system outputs a detailed visualization report containing warning prompts for exceeding resource consumption thresholds and an interactive visualization interface. Finally, the module compresses and assembles all output screens to form a multi-objective balancing data message for the final result form. This message is then delivered to the main router via a push port and sent to the designated customer inquiry and interaction screen, thus realizing a closed-loop feedback mechanism for the entire intelligent consultation service terminal.

[0101] Among them, the Pareto multi-objective frontier analysis core is a multi-objective optimization algorithm engine used to find the optimal trade-off among multiple conflicting objectives. A non-dominated group solution refers to a solution set where no other solution outperforms it in all dimensions. Key node coordinates for decision transitions are milestones on the derivation path involving significant resource input or state changes. Reverse tracing of days allocation is a calculation of the safety reserve time for each stage, based on the policy deadline and working backward to the current time. The radar alert layer is a dynamic UI element that visually indicates the time / resource urgency of the current node through color intensity or flashing frequency. The interactive timeline chart report frame is a scrolling front-end display interface similar to a horizontal timeline. The multi-objective balancing data message is a composite data package that ultimately encapsulates numerical results and front-end rendering code.

[0102] For example, suppose step S6 outputs a set of multimodal evolution trajectories for two successful paths. Path A's dynamic consumption resource library shows its cost as [12 months, 2.5 million yuan], and its total success probability calculated by the deduced transition achievement probability distribution evaluation array is 0.85. Path B's cost is [8 months, 3.5 million yuan], and its total success probability is 0.95. The Pareto multi-objective frontier analysis core inputs these two solutions. In the space of bi-objective cost minimization and success rate maximization, we observe that it cannot be concluded that one solution is absolutely superior to the other: Path A has a lower cost but also a lower success rate, while Path B has a higher success rate but also a higher cost. Since there is no third solution that is superior to or equal to A and B in both objectives, A and B are both non-dominated solutions. Together, they constitute the optimal policy frontier dataset. Subsequently, the report generation module creates an interactive dynamic report interface window for these two solutions. For path A, the system generates a 12-month roadmap, showing "Months 1-6: Startup certification, budget 500,000 yuan; Months 7-12: Equipment upgrade, budget 2 million yuan". For Path B, a more compact but more budgeted 8-month roadmap is generated. These two windows are packaged into a multi-objective balancing data message and pushed to the decision dashboard of the enterprise user who initiated the query via a data forwarding communication gateway. Users can clearly see the trade-offs between the two optimal strategies on the interface: "Strategy A: Slow and reliable, cost-effective but long-term; Strategy B: Fast and accurate, high success rate but high investment." Based on their own risk appetite and financial situation, users can then make a final strategic decision, completing a full closed loop from data to decision.

[0103] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A big data processing method for precise delivery and intelligent consultation of regional industrial policy information, characterized in that, include: S1: Obtain the original unstructured policy text data stream, extract text layout and semantic multimodal features, and perform cross-modal feature concatenation to generate a fused context vector stream; S2, integrates the context vector stream input graph neural network engine to perform dependency inference, extracts policy entity nodes bound with state variable selector encoded signals and global policy environment vectors, and then encapsulates them into a policy causal graph; S3 collects the original heterogeneous data streams of enterprises across multiple frequency bands and performs time-domain classification and prediction processing. It maps and constructs the set of enterprise state variables at the bottom, middle and top levels and connects the trigger feedback transmission pins to synthesize a heterogeneous hierarchical enterprise state space model. S4. Extract the global policy environment vector from the policy causal graph as the optimization intervention parameter, and use the transformation logic to rewrite the state transition matrix coefficients of cross-domain exchange within the state space model of heterogeneous hierarchical enterprises in the standby state. S5, obtain the query interaction signal containing the target policy document identifier parameter, retrieve the matching strategy causal graph and decode the associated specific state variable selector encoding signal, thereby deeply activating the specific observation port and the characteristics of the observed state variable to be measured within the heterogeneous hierarchical enterprise state space model. S6 moves the stimulated heterogeneous hierarchical enterprise state space model into the parallel joint sandbox cache area, introduces policy causal graph approval nodes and sets termination boundaries, initiates continuous path sampling exploration and convergence pruning calculation, and generates a multimodal evolution trajectory set with loss labels. S7 extracts the cost and resource consumption library of the multimodal evolution trajectory set and the single-step transition achievement probability array, applies the core of multi-objective frontier analysis to calculate the optimal policy frontier dataset, and converts and packages it into a multi-objective balancing data message for external transmission.

2. The big data processing method for precise delivery and intelligent consultation of regional industrial policy information according to claim 1, characterized in that: Step S1 includes: Receive raw, unstructured policy text data streams transmitted from external networks and input them into a hybrid encoded network; Start the parallel and independent visual and semantic receiving channels in the hybrid coding network; The visual position and layout features of the page are extracted through the visual receiving channel to form spatial structure attributes and generate layout structure vectors. At the same time, the full text content is extracted through the semantic receiving channel and autoregressive dimensionality reduction is performed to generate text semantic vectors. Create a cross-modal attention interaction graph and establish a cross-domain fusion module. Use layout structure vectors as query criteria and text semantic vectors as key-value pairs to converge side by side in the cross-domain fusion module. The cross-domain fusion module performs weight ratio correction and multi-dimensional coupling operation extraction at the dual-end spatial level, outputs a joint feature sequence that cancels out the interference of chapter information interruption, and reassembles it into a fusion context vector stream.

3. The big data processing method for precise delivery and intelligent consultation of regional industrial policy information according to claim 1, characterized in that: Step S2 includes: Deploy a graph neural network engine with a feature map modulation array for the received fused context vector stream; Within the graph neural network engine, the underlying text boundary distribution patterns contained in the fused context vector stream are extracted, and the category probability distribution is output by the fully connected classification layer to extract multiple policy entity nodes that constitute the prerequisite conditions for the declaration business. By calculating the relevant attention scores between adjacent policy entity nodes and determining whether they are greater than a preset connection threshold, logical connection guide edges with constraint direction attributes are established. The state variable selector encoding signal representing the mandatory judgment causal triggering mechanism is retrieved from the preset guidance rule retrieval library and attached and burned into each corresponding logical connection guide edge. The macro-level affixes representing support strength but without conditional constraints in the cleaned and fused context vector stream are tailored, normalized, assigned feature weights, and output as a global policy environment vector. Then, the distribution map of each policy entity node and edge information are packaged to construct a policy causal graph.

4. The big data processing method for precise delivery and intelligent consultation of regional industrial policy information according to claim 1, characterized in that: Step S3 includes: Enable and intercept the enterprise's multi-band raw heterogeneous data stream captured by the input port and perform time-domain band interception for high-frequency short axis with sampling period in seconds or milliseconds, mid-frequency long axis with sampling period in minutes or hours, and low-frequency extremely long axis with update period in months or grades. Intercepting, decomposing, and deriving low-latency underlying sensor physical parameters, mid-stage operational flow record central parameters, and slow-cycle top-level financial control parameters; Configure and launch an independent multidimensional Kalman sequence filtering algorithm and a nonlinear function curve fitting operator to map and process the three types of parameters with different time delays mentioned above. Independently generate bottom micro-level smooth energy consumption state layer variables, mid-level meso-level asset outflow efficiency state layer variables, and top macro-level R&D reserve state layer variables, and combine them into sets of enterprise state variables at each level. Connect the top-level enterprise state variable set to the middle-level enterprise state variable set by connecting the lead wire conduit and introduce a closed-loop threshold activation criterion, and trigger it when it exceeds the preset change sensitivity threshold. A closed negative feedback error adjustment sub-step is generated. By limiting the noise reduction fluctuation range of the model prediction through residual clipping and error covariance matrix reset, the cross-linking coupling constraint of each layer is realized. In the state prediction, the posterior estimated state of the previous moment is used as the input basis for the prior prediction of the current moment, generating a heterogeneous hierarchical enterprise state space model that characterizes the variation of system behavior.

5. The big data processing method for precise delivery and intelligent consultation of regional industrial policy information according to claim 1, characterized in that: Step S4 includes: Cut off the local area network instruction addressing of the nodes connected in the policy causal graph ontology during parsing, and focus on extracting the global policy environment vector residing in the memory block of the head environment pool; Activation introduces a vector-matrix transformation logic gate specifically harmonic with amplitude parameters; The global policy environment vector mapping with positive and negative macroeconomic adjustment characteristics is reduced and transformed into a pre-modulation parameter form. Based on the target matrix dimension of the intervened model level, a corresponding correction operator of the same dimension is dynamically generated. The optimization intervention parameters are forcibly assigned to the initializer of the heterogeneous hierarchical enterprise state space model in the idle waiting deduction sequence triggering stage. By detecting and comparing the polarity difference between the positive and negative amplitudes contained in the intervention parameters, the coefficients of the original inherent state transition matrix of the corresponding cross-data interaction transmission surface are overwritten according to the Hadamard product criterion. A stability constraint factor based on the spectral radius is introduced to perform forced contraction mapping, which changes the basic derivation law of the target and forces it to conform to the guidance and undergo a hard correction shift in weight, while preventing the derivation from diverging.

6. The big data processing method for precise delivery and intelligent consultation of regional industrial policy information according to claim 1, characterized in that: Step S5 includes: Connect to the external security interface and lock the query interaction signal initiated by the user, which contains the index number, i.e., the target policy document identifier parameter, as a target indication request; Based on the extracted and decoded target policy document identifier parameters, the unique matching items stored in the document rack are compared to extract the target strategy causal graph; The map is traversed sequentially and layer by layer, and the key points of the map are separated and identified by the weighted force surface attached to the associated edge path, and the constraint-specific state variable selector encoding signal is extracted. The built-in decoding and translation dictionary is introduced to convert the above separated encoded strings into addresses and corresponding address register bits for searching, and locates the internal feature block in the inter-layer branch node that has been modified in advance and is ready to carry out simulation operation after the parameter modification process. The execution triggers a Boolean value and activates the specific observation port for detection and listening associated with the memory anchor point at that coordinate, plus the random port binding association configuration, to be used as the observed state variable feature of the object to be considered.

7. The big data processing method for precise delivery and intelligent consultation of regional industrial policy information according to claim 1, characterized in that: Step S6 includes: Multiple independent isolated virtualization containers are instantiated to form a parallel joint sandbox cache area. The entire set of stimulated state structures with display detection feature lights that have been lit and modified is migrated into the sandbox for protection to prevent mutual interference and tampering. The node in the causal graph of the corresponding strategy is picked up and used to indicate the final exit confirmation condition of the pass is set as the termination boundary of the exploration stop barrier in this forward exploration and search simulation. The path deduction algorithm component based on the Monte Carlo Tree Search (MCTS) architecture is invoked to initiate an irregular leap time difference cross-discrete multi-order continuous branch crack prediction extension pre-draft work with random attempts and variable fluctuations on the replica located in the trial operation; By combining the termination boundary monitoring and tracking of each newly derived branch line, the deviation of the reduction and the expected sacrifice ratio and skewness are screened and cut out the invalid divergent paths that have gone out of control, and the execution is terminated and removed. The remaining lines with the possibility of crossing within the limit are recorded and the cumulative time loss and budget exhaustion and other resource statistics are recorded as parameters marked and archived on the additional backplane. By merging and assembling these paths that have undergone successful bottoming out and breakthroughs after trimming and removing redundancy, along with all the aforementioned historical input and consumption data packets, a multimodal evolution trajectory set data queue containing consumption feature identifiers is synthesized.

8. The big data processing method for precise delivery and intelligent consultation of regional industrial policy information according to claim 7, characterized in that: The process combines termination boundary monitoring and tracking of each newly derived branch's deviation magnitude and expected sacrifice ratio to screen and cut out uncontrollable divergent paths, terminate and remove them, and retain paths with the potential to cross within the limit, recording their cumulative time loss plus budget depletion and other resource statistics as parameter markers for additional backplane archiving, including: The preset evaluation probe captures the actual coordinate difference change offset polarization distance at the end of each time step along the specified newly excavated downcutting path; When the simulation determines that the offset polarization distance has lengthened and expanded, deviating from the leading edge limit line, and the cumulative amount of required compensation exceeds the limit budget cap threshold; An immediate soft interrupt stop order is issued to the test kernel to forcibly terminate the trial run subroutine, perform forced truncation and clear convergence actions, not allocate computing resources, and immediately backpropagate the minimum penalty reward to the upper-level probability tree node to update the confidence upper limit of the branch, preventing the invalid branch from being sampled repeatedly. Only those that arrive and smoothly press against the termination boundary baseline, meeting the safety restrictions, are retained, and their time stamps along the way are recorded, and overdraft amounts are converted into specific expenditure dimensions and attached as markers.

9. The big data processing method for precise delivery and intelligent consultation of regional industrial policy information according to claim 1, characterized in that: Step S7 includes: The cost and resource consumption database under the evaluation and constraint items is established by extracting and separating all collected multimodal evolution trajectory sets and identifying the cost indicators of various maintenance and promotion requirements put into practice. The integrated collection and deduction process simulates the multiplication of odds results generated by jump events to construct a single-step transition probability array for the operational risk vulnerability probability safety belt of the synthetic evaluation system. The core takeover operation process of the multi-objective front analysis drives non-dominated sorting and crowding distance calculation to find multivariate composite solutions that can avoid falling into high failure rate strategy intervals, minimize cost-saving expenses as much as possible, and do not constitute exclusion or domination among each other. This forms the front and outputs the non-dominated group solution optimal strategy front dataset. The user-side interactive visualization interface is constructed by using the window rendering interface to reprint its planar structure, including prompts for the execution sequence, timing nodes, resource preparation quantity, and radar alarm layers. The multi-objective balancing data message, which compresses and assembles all output screens to form the final result form, is delivered to the main router via a push port and sent to the designated customer. The interactive screen carrier realizes the service terminal feedback closed loop.

10. The big data processing method for precise delivery and intelligent consultation of regional industrial policy information according to claim 9, characterized in that: The user-side interactive visual display interface, which utilizes the window rendering interface to reprint its planar structure, includes prompts for the execution sequence, timing nodes, resource preparation quantities, and radar alarm layers. Obtain the coordinates of each key node in the optimal solution derived from the optimal policy frontier dataset; Based on the system timestamp when the user request is initiated and the model calculation deadline window, the reverse traceability days ratio is calculated to generate a highlighted node warning countdown prompt character field for each time node on the time axis; By using graphical templates, the warning prompts are attached to the main structure to present a progressively evolving route leading to the final application approval confirmation endpoint. The output includes a visual report detailing the warning prompts for exceeding resource consumption thresholds, which is used for loading and shipping packages.

Images