Unit internal control management decision-making method and system based on multi-modal data analysis

Abstract

The invention relates to the field of data decision analysis, in particular to a unit internal control management decision method and system based on multi-modal data analysis. The method comprises the following steps: collecting heterogeneous original data streams of a plurality of heterogeneous data sources; performing space-time dimension entropy change differential analysis and multi-dimensional entropy change characteristic mining, performing multi-level semantic analysis on the heterogeneous original data stream, performing layer-by-layer abnormal mode semantic deconstruction, and constructing a precise semantic feature portrait of each abnormal mode; performing multi-target coordination request analysis and dynamic target priority evaluation on the cognitive entropy change reference characteristic spectrum to obtain a coordination request priority; and carrying out maximum decision capacity calculation on the cognitive entropy change reference feature map, and carrying out dynamic optimal decision boundary calculation based on the coordination request priority and the accurate semantic feature portrait so as to generate a dynamic tradeoff constraint range. Through multi-target collaborative decision analysis, the decision target conflict is smoothed, and the decision demand execution efficiency of the user is improved.

Description

Technical Field

[0001] The present invention relates to the field of data decision analysis, and in particular to a unit internal control management decision-making method and system based on multimodal data analysis. Background Art

[0002] With the continuous advancement of informatization and digital governance, organizations are increasingly relying on data for business management, risk control, and decision support. In internal control (IC) management, in particular, data comprehensiveness, accuracy, and intelligent processing capabilities have become key indicators for measuring the efficiency and reliability of internal control systems. At the same time, core management tasks such as financial auditing, risk prevention, and performance evaluation have placed greater demands on the intelligence and sophistication of internal control mechanisms. Traditional internal control management methods primarily rely on regular manual audits, tabular report analysis, and static indicator assessments. These approaches have significant shortcomings in data collection, risk identification, and anomaly warnings. They struggle to respond promptly to complex and ever-changing business scenarios and are unable to effectively integrate heterogeneous data sources to support comprehensive decision-making. Furthermore, faced with massive data volumes and highly fragmented information content, traditional internal control approaches often lack intelligent linkage mechanisms and dynamic analysis capabilities, resulting in lagging risk control and severe information silos, ultimately impacting organizational operational efficiency and management capabilities.

[0003] With the rapid development of emerging technologies such as big data, artificial intelligence, and the Internet of Things, various organizations have gradually accumulated vast amounts of structured data (such as financial data, approval records, and personnel files) and unstructured data (such as meeting minutes, voice recordings, email content, and surveillance videos) in their daily management processes. This data exhibits multimodal characteristics, including diverse sources, complex formats, and rapid dynamic updates, providing new opportunities for in-depth internal control management analysis and decision support. However, current mainstream internal control systems often lack unified data fusion mechanisms, making it impossible to effectively integrate and deeply mine multimodal data. This makes it difficult to identify hidden risks in a timely manner and fails to support higher-dimensional, more forward-looking decision making. Therefore, a more intelligent internal control management decision-making approach is urgently needed. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes a unit internal control management decision-making method and system based on multimodal data analysis to solve at least one of the above technical problems.

[0005] To achieve the above objectives, the present invention provides a unit internal control management decision-making method based on multimodal data analysis, comprising the following steps: Step S1: Collect heterogeneous raw data streams from the unit's internal control management platform; perform temporal and spatial entropy change differential analysis and multi-dimensional entropy change feature mining to generate a cognitive entropy change benchmark feature map; Step S2: Perform multi-level semantic analysis on the heterogeneous original data stream, and perform semantic deconstruction of abnormal patterns layer by layer to construct an accurate semantic feature portrait of each abnormal pattern; Step S3: Obtain the preset decision target instruction, perform multi-target coordination request analysis and dynamic target priority evaluation on the cognitive entropy change benchmark feature map, and obtain the coordination request priority; Step S4: Calculating the maximum decision capacity of the cognitive entropy change benchmark feature map, and calculating the dynamic optimal decision boundary based on the coordination request priority and the precise semantic feature portrait, thereby generating a dynamic trade-off constraint range; Step S5: solving multi-objective dynamic constraint responses based on the dynamic trade-off constraint range, and performing instant decision processing to obtain real-time execution results; Step S6: Identify the decision execution response deviation of the real-time execution effect, perform deep optimization of strategy iteration, and build an internal control management decision-making intelligent body.

[0006] In this specification, a unit internal control management decision-making system based on multimodal data analysis is provided, which is used to execute the unit internal control management decision-making method based on multimodal data analysis as described above, including: The data stream entropy change analysis module is used to collect heterogeneous raw data streams from the unit's internal control management platform; it performs differential analysis of entropy change in the spatiotemporal dimension and multi-dimensional entropy change feature mining to generate a cognitive entropy change benchmark feature map; A semantic parsing module is used to perform multi-level semantic parsing on the heterogeneous raw data streams, and to perform semantic deconstruction of abnormal patterns layer by layer, thereby constructing an accurate semantic feature portrait of each abnormal pattern; The multi-objective coordination module is used to obtain preset decision target instructions, perform multi-objective coordination request analysis and dynamic target priority evaluation on the cognitive entropy change benchmark feature map, and obtain the coordination request priority; A decision boundary module is used to calculate the maximum decision capacity of the cognitive entropy change benchmark feature map and to calculate the dynamic optimal decision boundary based on the coordination request priority and the precise semantic feature portrait, thereby generating a dynamic trade-off constraint range; The constraint solving module is used to solve multi-objective dynamic constraint responses based on the dynamic trade-off constraint range and perform instant decision processing to obtain real-time execution results; The internal control management decision-making module is used to identify decision-making response deviations in real-time execution effects, perform deep optimization of strategy iterations, and build an internal control management decision-making intelligent entity.

[0007] The beneficial effects of the present invention are as follows: By collecting heterogeneous raw data streams from multiple heterogeneous data sources and performing differential analysis of entropy changes in the spatiotemporal dimensions and mining multidimensional entropy change characteristics, high-precision dynamic cognitive modeling of the system's operating status can be achieved. Entropy change, as a sensitive indicator of system complexity and uncertainty, can reveal potential instabilities or trend shifts in data changes, thereby helping the system identify potential anomalies at an early stage. The resulting cognitive entropy change baseline feature map is highly distinguishable and stable, serving as a foundational reference model for subsequent anomaly detection and decision optimization, effectively improving overall perception sensitivity and proactive response. By performing multi-level semantic analysis of heterogeneous raw data streams and layer-by-layer semantic deconstruction of the anomaly patterns contained therein, the unique behavioral characteristics, evolutionary laws, and system impact paths of each anomaly type can be deeply extracted, constructing a highly accurate semantic feature profile. This semantic deconstruction process breaks through traditional rule-based or threshold-based anomaly recognition methods, enabling the system to identify and classify "new, hidden, and gradual" anomalies. Accurate semantic feature profiling not only improves the accuracy of anomaly classification but also provides a rich, structured information foundation for subsequent decision-making trade-offs and prioritization. By receiving preset decision-making target instructions and analyzing and dynamically assessing multi-target coordination requests based on the entropy change feature map, the system effectively resolves conflicts, resource competition, and execution priorities inherent in the multi-target decision-making process. By dynamically generating coordination request priorities, the system can adapt to the urgency, benefit value, and execution cost of targets in different scenarios, achieving a more flexible and intelligent target management mechanism. This process significantly improves the system's scheduling capabilities and execution stability under conditions of limited resources and frequent anomalies. By calculating the maximum decision capacity of the entropy change feature map and combining it with the generated coordination request priorities and anomaly semantic profiles, the boundaries of the decision space that the system can support in its current state can be accurately and dynamically estimated. The resulting dynamic optimal decision boundary not only reflects the system's true perception of current risks and resources, but also significantly enhances the rationality and risk resilience of the policy generation process by dynamically balancing the constraints. This prevents decision-making failures caused by overload or incorrect scheduling, ensuring the security of the system's policies in complex and changing environments. By solving multi-objective dynamic constraint responses within the dynamic trade-off constraints and triggering real-time decision-making processing, it is possible to ensure that the response strategy when multiple objectives exist not only meets the system constraints but also maximizes the target benefits. The core of this step is to achieve "execution and optimization" of decisions, ensuring that the decision-making process has the ability to adapt to environmental conditions, abnormal fluctuations and changes in priorities in real time. Its direct effect is to improve the system's execution efficiency, stability and the accuracy of the implementation of the strategy. It is particularly suitable for complex scenarios with fast dynamic changes and high decision-making frequency, such as energy scheduling and real-time traffic control.By identifying systematic deviations in real-time execution results and embedding them as feedback signals into the strategy optimization process, we can build a continuously evolving, self-learning internal control management decision-making agent. This agent possesses a closed-loop learning capability of "feedback-correction-reoptimization," continuously correcting historical deviations, absorbing execution experience, and enhancing the adaptability and robustness of future strategies. The result is an intelligent decision-making engine that can continuously adapt to the dynamic changes of complex systems and possesses highly adaptive and self-recovering capabilities, fundamentally improving system stability, decision-making efficiency, and sustainable optimization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a schematic flow chart of the steps of a unit internal control management decision-making method based on multimodal data analysis of the present invention; Figure 2 Detailed implementation flow chart of step S1; Figure 3 Detailed implementation flow chart of step S2; Figure 4 Schematic diagram of the detailed implementation steps of step S3. DETAILED DESCRIPTION

[0009] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0010] This application example provides a unit internal control management decision-making method and system based on multimodal data analysis. The execution entities of the unit internal control management decision-making method and system based on multimodal data analysis include but are not limited to: mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc. equipped with the system, which can be regarded as general computing nodes of this application. The data processing platform includes but is not limited to: at least one of an audio and image management system, an information management system, and a cloud data management system.

[0011] See also Figures 1 to 4 The present invention provides a unit internal control management decision-making method based on multimodal data analysis, and the unit internal control management decision-making method based on multimodal data analysis includes the following steps: Step S1: Collect heterogeneous raw data streams from the unit's internal control management platform; perform temporal and spatial entropy change differential analysis and multi-dimensional entropy change feature mining to generate a cognitive entropy change benchmark feature map; Step S2: Perform multi-level semantic analysis on the heterogeneous original data stream, and perform semantic deconstruction of abnormal patterns layer by layer to construct an accurate semantic feature portrait of each abnormal pattern; Step S3: Obtain the preset decision target instruction, perform multi-target coordination request analysis and dynamic target priority evaluation on the cognitive entropy change benchmark feature map, and obtain the coordination request priority; Step S4: Calculating the maximum decision capacity of the cognitive entropy change benchmark feature map, and calculating the dynamic optimal decision boundary based on the coordination request priority and the precise semantic feature portrait, thereby generating a dynamic trade-off constraint range; Step S5: solving multi-objective dynamic constraint responses based on the dynamic trade-off constraint range, and performing instant decision processing to obtain real-time execution results; Step S6: Identify the decision execution response deviation of the real-time execution effect, perform deep optimization of strategy iteration, and build an internal control management decision-making intelligent body.

[0012] In the embodiment of the present invention, see Figure 1 , is a flowchart of a method for internal control management decision-making based on multimodal data analysis of the present invention. In this example, the method includes the following steps: Step S1: Collect heterogeneous raw data streams from the unit's internal control management platform; perform temporal and spatial entropy change differential analysis and multi-dimensional entropy change feature mining to generate a cognitive entropy change benchmark feature map; In this embodiment, after obtaining authorization from relevant departments and users, in the unit's internal control management decision-making scenario, the core of the first step is to systematically collect heterogeneous original data streams from multiple business modules, and conduct in-depth information entropy analysis on these data streams. Combined with the dynamic evolution characteristics of the time and space dimensions, multi-dimensional entropy change features are finally extracted to construct a set of cognitive-level benchmark feature maps, providing a basic reference for subsequent intelligent monitoring, resource scheduling, and decision optimization.

[0013] During the raw data collection phase, data integration across the company's six core business systems is required: the budget management system, revenue management platform, expenditure settlement system, procurement management platform, contract execution system, and material asset management system. The data interfaces for each system are structurally heterogeneous, with some outputting from traditional Oracle databases, others providing Web services or REST APIs, and some systems only accessible through log monitoring or data snapshots. Data extraction and integration tools (such as Apache NiFi or Talend) are used in conjunction with a unified data adapter to connect to data sources, and a field mapping model based on data elements is established. For example, the "budget account number" in the budget system must be compared with the "expenditure account code" in the expenditure system to ensure consistency in subsequent analysis. Data collection is performed at an hourly granularity to ensure temporal continuity and analytical accuracy.

[0014] After the collection is completed, the information entropy density calculation and entropy change differential analysis phase begins. For each type of heterogeneous raw data stream, a data frame (DataFrame) is first constructed with time as the main axis. For example, budget data is aggregated by "month", income and expenditure data are aggregated by "day", contract and procurement data are discretized by "approval node", and material data is formed into an event sequence according to "inventory fluctuation". Then, the entropy value of the state distribution of each type of data in different time windows is calculated using the Shannon information entropy formula H=-∑p(x)logp(x), where p(x) represents the probability distribution of a certain indicator (such as budget execution rate or procurement plan completion rate) under a specific state. To understand the spatiotemporal dynamic evolution of entropy, we further employed the entropy change differential method to calculate the gradient rate of change of information entropy between adjacent time periods or organizational units, thereby constructing a set of data entropy evolution trajectories with time series characteristics. In the experiment, a sliding time window of three days was set, and a first-order difference algorithm was used to process the continuous entropy sequence. We found that the material system exhibited significant entropy change peaks at the beginning and end of the month, reflecting cyclical fluctuations in allocation.

[0015] Next, the data topology and multi-dimensional entropy change feature mining phase begins. By constructing a topological structure of associations between multiple types of business data, a causal path and collaborative relationship network is established between data streams. For example, budget data influences procurement plans, which in turn influence expenditure execution after procurement completion; contract performance is influenced by both budget freezes and material delivery. Node embedding learning based on a graph neural network (GNN) extracts the dynamic coupling characteristics between various data sources and superimposes the entropy change differential features to form a multimodal embedding feature matrix. Finally, dimensionality reduction methods such as t-SNE or UMAP are used to map the high-dimensional features into a two-dimensional space, creating a cognitive entropy change baseline feature map. This map uses a "data channel-time period" coordinate system. Each point corresponds to the entropy change state of a data subsystem within a specific time window. The entropy value is represented by color, and the cluster density reflects the system's complexity and abnormal evolution trends.

[0016] Step S2: Perform multi-level semantic analysis on the heterogeneous original data stream, and perform semantic deconstruction of abnormal patterns layer by layer to construct an accurate semantic feature portrait of each abnormal pattern; In this embodiment, after completing the construction of a cognitive entropy change benchmark map during the unit's multimodal data analysis process, the next key step is to perform multi-level semantic parsing of the heterogeneous raw data streams. This, combined with historical and real-time behavioral trajectories, allows for in-depth analysis of potential anomaly patterns, thereby supporting risk identification and response strategies within internal control management. This step aims to build an anomaly pattern recognition and classification system with semantic interpretation capabilities. By restoring and analyzing the behavioral trajectories and data semantics of each anomaly pattern, a fine-grained semantic feature profile is generated.

[0017] First, multi-level semantic parsing begins with defining the semantic hierarchy of the data. In this scenario, the raw data includes six core business categories: budget, revenue, expenditure, procurement, contracts, and materials management. Each data category contains structured numerical values ​​(such as amounts, percentages, and times) and weakly structured text (such as contract details, procurement requisition reasons, and budget explanations). A unified semantic framework must be established based on the "semantic entity-attribute-behavior" hierarchy. For example, in contract data, "abnormal contract amount," "fluctuation in fulfillment cycle," and "abnormal payment node" are identified as first-level semantic entities. Second-level attributes include "amount deviation rate" and "node time offset." Third-level behaviors can be abstracted into behavioral labels such as "contract execution violates the original budget plan." This stage primarily uses natural language processing tools such as BERT or RoBERTa pre-trained models to extract contextual semantic vectors for weakly structured fields. For structured fields, a label mapping mechanism is used to uniformly abstract them into standard business behavior language (for example, converting "budget execution rate <40%" into the label "severely insufficient execution").

[0018] Next, the system enters the layer-by-layer semantic deconstruction of anomaly patterns. The system must identify potential anomaly patterns in multi-source data, such as "asynchronous timing between contract signing and budget freeze," "purchase amounts exceeding approved limits," and "sudden changes in material withdrawal frequency." To achieve high-accuracy identification, an anomaly detection algorithm based on a graph attention network (GAT) is introduced. This algorithm models the business processes and approval relationships as a heterogeneous graph structure, where nodes represent business entities (such as purchase orders, budgets, invoices, and contracts), and edges represent temporal logic or capital flow relationships between entities. Within this graph model, anomalous subgraph structures are identified by weighted aggregation of node attributes and their neighbor relationships. For example, if the purchase amount in an approval path consistently exceeds 30% of similar contracts and the approval time is significantly below the average, the subgraph will be labeled "High-Risk Quick Procurement."

[0019] During the deconstruction process, each identified anomaly pattern is further reconstructed into the underlying data semantic chain. For example, an anomalous procurement process may involve multiple data anomalies, such as delayed budget releases, inconsistent contract signings, and delayed inventory records. Through causal chain modeling and semantic backtracking, these data entities are chronologically arranged and semantically annotated to construct an "anomaly pattern semantic chain." In an experiment, based on a sample derived from the 2024 financial system of a prefecture-level city, approximately 812 anomaly patterns were identified, each containing an average of 3.2 semantic nodes, spanning time periods ranging from several hours to a week. Cluster analysis of the anomaly semantic chains using similarity metrics (such as cosine similarity or KL divergence) revealed typical anomaly categories, such as "rapid jumps in budget and procurement" and "discontinuous fluctuations in contract amounts."

[0020] Finally, we construct a precise semantic feature profile of the anomaly pattern. Through the above analysis, we extract the following semantic feature vectors for each anomaly type: anomaly triggering conditions (such as amount, time threshold), participating data flow categories, key behavioral nodes, time span, anomaly persistence, number of coordinated anomalies, and historical frequency of occurrence. These features are standardized and stored in a semantic profile template, forming a reusable and comparable knowledge base of anomaly types. This foundation enables organizations to quickly identify similar anomalies, automatically handle them in a tiered manner, and provide early warning for internal control audits.

[0021] Step S3: Obtain the preset decision target instruction, perform multi-target coordination request analysis and dynamic target priority evaluation on the cognitive entropy change benchmark feature map, and obtain the coordination request priority; In this embodiment, during the internal control management process of an enterprise unit, different management modules (such as budget control, procurement approval, contract fulfillment, expenditure monitoring, etc.) often have multiple decision-making objectives running in parallel. These objectives may overlap in time, compete for resources, and even conflict in results. Therefore, after completing the construction of the cognitive entropy change baseline feature map and anomaly semantic profile, the system needs to introduce a multi-objective coordination mechanism to dynamically respond to the different target instructions issued by the internal control management level and coordinate priority assessments of target requests accordingly, thereby ensuring the rationality and efficiency of the entire data analysis and decision-making chain. This step focuses on the core of "multi-objective regulation" and implements real-time dynamic priority assessment based on the data entropy change map.

[0022] The system retrieves pre-set decision-making target directives. These directives originate from the upper-level policy deployment of the internal control management platform and include static targets (such as increasing quarterly budget execution efficiency to over 90%) and dynamic targets (such as optimizing emergency expenditure approval pathways and increasing procurement early warning response priorities). Each target directive is standardized into a multi-dimensional target metadata structure, including target categories, associated data sources, expected outputs, impact cycles, weighting factors, and conflict constraints. For example, a directive might include: "Control the over-budget procurement rate for materials to no more than 5% this quarter (high priority), while ensuring on-time contract fulfillment rates greater than 95% (medium priority)."

[0023] Based on the cognitive entropy change benchmark feature map, the system analyzes the strength of association between each target request and different data entropy change areas in the map. This process is handled through semantic mapping and indicator clustering. Specifically, the system uses the indicator items in the target (such as "budget execution rate", "procurement frequency", and "contract performance deviation rate") as query vectors, and uses the node features in the map (including entropy change trajectories, time gradients, and semantic entities) to calculate vector similarity to evaluate the degree of coupling between the target and the current system state. In the experiment, based on a simulated data set of a provincial government platform, after using the TF-IDF weighted word vector method to model the indicator semantic items, the average matching accuracy with the entropy change nodes reached 91.7%.

[0024] Enter the multi-objective coordination request analysis phase. Because multiple target instructions may be mapped to the same data channel or adjacent entropy change characteristic paths, the system requires conflict identification and coordinated scheduling. During this process, a multi-objective optimization framework (such as the Multi-objective Evolutionary Algorithm (MOEA) algorithm based on the Pareto optimal solution set) is used to model and analyze the coupling, conflict, and resource consumption between multiple targets, generating an inter-target correlation conflict matrix. The matrix elements represent the conflict intensity between two targets in a certain resource domain (such as approval process time, data computation occupancy, signal channel, etc.). By performing spectral clustering on the conflict matrix, high-conflict target pairs that require coordination can be quickly identified.

[0025] The system conducts dynamic goal priority assessments. The assessment method utilizes a combined AHP (Analytic Hierarchy Process) and weighted regression modeling mechanism. First, each goal is scored using the initial weights manually configured by the manager. Then, using system logs and historical data, a secondary weighted adjustment is made to factors such as goal achievement, execution impact, and emergency response level to generate a final coordination request priority sequence. For example, if "Quick Response to Procurement Approval" is found to occur frequently in historical anomaly clusters and has a significant impact on budget and contract execution, this goal will be significantly prioritized in the current cycle.

[0026] The resulting coordination request priority output is fed back into the policy scheduling engine, serving as the foundational input for subsequent process control, resource allocation, and prioritized exception response, providing real-time, goal-oriented optimization strategies for the entire organization's internal control management. This process not only enables structured management of internal control objectives but also improves the response intelligence and resource coordination efficiency of internal control policies. Real-world scenarios have demonstrated that it can increase multi-objective response speed by approximately 36% while significantly reducing the frequency of resource allocation conflicts.

[0027] Step S4: Calculating the maximum decision capacity of the cognitive entropy change benchmark feature map, and calculating the dynamic optimal decision boundary based on the coordination request priority and the precise semantic feature portrait, thereby generating a dynamic trade-off constraint range; In this embodiment, within the unit's internal control management decision-making system based on multimodal data, as multiple business objectives are coordinated and semantic anomaly profiles are constructed, the system needs to quantitatively assess overall decision-making capabilities and, based on this, construct a dynamic trade-off mechanism to achieve boundary control and resource allocation optimization between target instructions. The core purpose of this step is to dynamically model the resource scheduling boundaries of the internal control system during actual operation by calculating the maximum decision-making capacity that the current system can handle and combining the composite characteristics of coordination request priority and anomaly semantics, thereby generating a set of dynamic trade-off constraints with time elasticity, resource matching, and semantic interpretability.

[0028] The first stage involves calculating maximum decision capacity. This step, based on the previously constructed cognitive entropy change benchmark feature map, identifies the processing capacity limits of key data channels, entropy change nodes, and behavioral paths in the system. Specifically, each business entropy change trajectory in the map (e.g., budget fluctuation frequency, expenditure variability, procurement response time) is treated as a "data decision unit." Parameters such as average processing response time, data bandwidth consumption, and maximum load pressure within a historical time window are calculated to form a "unit decision energy index." For example, statistical analysis of historical data from the budget module of a municipal government platform revealed that when the number of budget approval anomalies exceeded 1.8 times the average daily processing capacity, the system's anomaly detection accuracy dropped to 78%, and system response latency increased by 22%. Combined with this data, the system's current "decision pressure limit," or maximum decision capacity, is calculated. This is typically measured by the number of high-priority tasks that can be completed per unit time and their average execution quality.

[0029] The second stage involves the calculation of the dynamic optimal decision boundary. This calculation combines the coordination request priority output from the previous step with precise semantic feature profiles. The core idea is that targets of different priorities occupy different system resources and processing power in different semantic anomaly scenarios. Therefore, a resource "consumption function" must be constructed for each target request based on the semantic profile. For example, the "Procurement Contract Delay Risk" anomaly profile contains three types of behavioral paths and six data entity nodes, and the computing resources required for its processing are more than 20% higher than those for the "Budget Adjustment Deviation" anomaly profile. Therefore, when modeling the multi-target processing boundary, a multivariate function F(x) = w1·R1 + w2·R2 + ... + w is introduced. n ·R n , where R n represents the semantic complexity weighted resource requirement of each abnormal request, w n Represents its priority weight. The system obtains the optimal dynamic scheduling boundary between a set of objectives by solving the optimal boundary point of this function under the maximum decision capacity.

[0030] The third stage is the generation of dynamic trade-off constraint ranges. Based on the above-mentioned dynamic optimal decision boundary calculation results, the system defines the boundary conditions and resource switching thresholds that can be accommodated simultaneously between different objectives, forming a "dynamic trade-off constraint matrix." Each row in the matrix represents the trade-off range of a high-priority objective in its current state (i.e., maximum resource scheduling range, minimum response delay requirement, acceptable target drift, etc.), and each column represents the dependent response relationship of other objectives under its influence. For example, for the priority control objective of "increased risk of delayed revenue collection," the system can allow it to occupy a maximum of 30% of bandwidth resources under the current data resource scheduling model, and limit its delay in the contract fulfillment logic channel to no more than 2 hours; if this boundary is breached, an emergency rescheduling mechanism must be triggered.

[0031] The final output is a dynamic trade-off constraint map covering various target instructions. This map not only reflects the system's resource scheduling boundaries and behavior processing capabilities, but can also be directly embedded in the internal control policy engine for dynamic task allocation, automatic priority adjustment, and intelligent risk avoidance. In practical applications, the introduction of this map reduced the conflict rate of contract approval and procurement scheduling resources for a municipal unit by 41%, and reduced the overall average approval latency by 18%, significantly improving the unit's management flexibility and response efficiency under multi-task and multi-resource conflicts. This step provides a solid boundary foundation for subsequent intelligent policy recommendations and risk response simulations.

[0032] Step S5: solving multi-objective dynamic constraint responses based on the dynamic trade-off constraint range, and performing instant decision processing to obtain real-time execution results; In this embodiment, the system extracts the current schedulable boundary information from the "dynamic trade-off constraint range map" output in the previous stage. Each dynamic trade-off boundary reflects the range of resources that can be scheduled for a specific goal under a specific business state, the allowed response delay, and the priority dependency relationship with other goals. For example, in a certain unit's budget scheduling system, the currently acceptable "budget execution excess float" is ±3.5%, the procurement response delay must not exceed 48 hours, and the concurrent contract approval processing is limited to within 3 paths. The system converts these boundary conditions into constraint function input values ​​for subsequent solver modeling.

[0033] Next, the multi-objective dynamic constraint response solution phase begins. The system constructs a multi-objective optimization mathematical model, treating each business objective as an objective function (such as minimizing approval latency, maximizing contract fulfillment efficiency, and controlling procurement overflow). Resource boundaries, latency constraints, and dependencies are input as inequality constraints. To achieve efficient solution efficiency, the system can employ evolutionary algorithms (such as the multi-objective non-dominated sorting algorithm based on NSGA-II) or convex optimization techniques. For example, NSGA-II uses a population evolution strategy to select, crossover, and mutate candidate solutions, generating a set of non-inferior solutions that satisfy the constraints. The system then adjusts the objective bias based on objective priority and the importance of the current semantic profile. Experiments have shown that in scenarios with intensive internal resource conflicts, the algorithm can achieve over 90% objective coverage and output an executable policy within 100ms.

[0034] The system then converts the scheduling parameters in the optimal solution into immediate decision-making instructions and calls real-time business processes. This includes the following operations: Prioritize the scheduling of tasks that initiate specific paths in the approval system; Elevate the response level to abnormal events (e.g., abnormal purchase order approval). Make tolerance adjustments to budget or contract data to maintain system stability; Dynamically switch data collection priorities to ensure high-value information is processed first.

[0035] When a government agency discovers that the procurement budget exceeds the threshold and the contract fulfillment time offset occurs, the system outputs an immediate processing instruction set of "freeze procurement instructions + schedule contract warning logic + extend the contract tolerance time window to 4 hours" based on the constraint solution, and immediately pushes it to the business middle platform.

[0036] The system collects real-time feedback from each business path after execution, forming a preliminary set of real-time execution effectiveness evaluation indicators. This evaluation set primarily includes key indicators such as response time (such as approval acceleration rate), resource utilization, target achievement deviation, and changes in task conflict frequency. By comparing these indicators with the expected execution boundaries, the effectiveness of immediate decisions can be quantified. For example, in a budget adjustment scenario, real-time feedback showed that after dynamic response solution, the average approval time decreased by 21.4%, and the frequency of abnormal jump path use decreased by 36.2%, demonstrating the significant optimization effect of dynamic solution decision-making.

[0037] Step S6: Identify the decision execution response deviation of the real-time execution effect, perform deep optimization of strategy iteration, and build an internal control management decision-making intelligent body.

[0038] In this embodiment, a real-time monitoring mechanism is established to continuously track the execution of each decision objective. This includes monitoring key indicators such as task completion time, resource consumption, execution order, and task success rate. Real-time monitoring continuously collects data and compares execution results with preset targets to identify deviations. By comparing these results with the preset decision execution results, the system can accurately determine whether there are deviations during execution. If the execution time of a particular objective exceeds a set threshold or resource consumption exceeds expectations, the system automatically flags these deviations. For each decision objective, the system evaluates the degree of execution deviation by comparing the actual results with the target value. Identified deviations are further categorized and analyzed. Possible sources of deviation include uneven resource allocation, incorrect task priority determination, and inappropriate scheduling algorithms. Through in-depth analysis of the causes of deviations, the system can provide targeted solutions for subsequent optimization. Once deviations in decision execution are identified, the system adjusts the current decision model. Adjustments may include re-evaluating objective priorities, reallocating resources, and adjusting the order of task execution. Through a feedback control mechanism, the system transmits deviation information to the decision model and uses this information for dynamic optimization. If a task encounters a performance bottleneck during execution, the system may adjust its priority or alter its resource allocation strategy to ensure efficient overall system operation. To achieve deep optimization, the system can leverage reinforcement learning algorithms to iteratively refine its strategy. Through a reward mechanism, the system continuously learns during execution, gradually improving its decision-making process. Whenever the system identifies a deviation, the reinforcement learning model adjusts based on the type of deviation and optimizes future decision-making strategies. During this process, the decision-making agent simulates decision execution in various scenarios, accumulating experience and gradually improving its decision accuracy. In multi-objective decision-making, conflicting objectives may exist, necessitating an adaptive approach to optimize the decision path for each objective. The system can dynamically adjust the weights of objectives based on factors such as their importance, priority, and resource consumption to ensure optimal resource allocation. This adaptive approach enables the system to adapt to changing environments and objectives and optimize decision outcomes. A framework for internal control management decision-making agents is being developed to enable them to handle and optimize multiple decision-making objectives. This framework will comprise three core components: a perception layer, a decision layer, and an execution layer. The perception layer is responsible for acquiring real-time environmental and execution data, the decision layer is responsible for adjusting decisions based on environmental changes, and the execution layer is responsible for actually executing decisions and providing feedback. The agent possesses self-learning capabilities, extracting experience from each decision execution and gradually improving its decision quality through policy iteration and reinforcement learning. Based on the real-time execution results of each goal and deviation feedback, the agent continuously adjusts resource allocation plans and decision paths, gradually forming an efficient decision-making model.In multi-objective decision-making, intelligent agents must not only solve the problem of a single objective but also consider the relationships between multiple objectives, balancing conflicts between objectives and resource allocation. To this end, intelligent agents must possess a certain level of collaborative decision-making capabilities, able to dynamically coordinate between objectives to ensure the optimality of the overall decision.

[0039] In this embodiment, refer to Figure 2 , is a flowchart of the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: Collect heterogeneous raw data streams from the internal control management platform of the unit; the heterogeneous raw data streams include budget management data, revenue management data, expenditure management data, procurement management data, contract management data, and material management data; Identifying data flow channels of the heterogeneous original data flow, and performing information entropy distribution calculation to generate information entropy density of each data flow channel; Based on the information entropy density, a temporal and spatial dimension entropy change differential analysis is performed to obtain the data entropy change evolution trajectory of each data stream channel; Performing data topology analysis and modeling on the heterogeneous original data streams, and constructing a topology model for each data stream; Multi-dimensional entropy change characteristics are mined based on the topological structure model and the data entropy change evolution trajectory to generate a cognitive entropy change benchmark feature map.

[0040] In this embodiment, after obtaining authorization from relevant departments and users, the first step within the company's internal control management system is to complete the unified collection of heterogeneous data sources across multiple business segments and the structured construction of heterogeneous raw data streams. The core task of this step is to connect multiple data silos and extract data scattered across the budget management system, revenue accounting system, procurement and bidding system, contract archive system, and asset and material management platform in a streaming manner, forming heterogeneous raw data streams with temporal continuity and structural dimension consistency. These heterogeneous raw data streams form the underlying support for subsequent entropy change feature analysis and cognitive modeling.

[0041] First, at the data collection layer, a distributed collection tool based on an ETL (Extract-Transform-Load) strategy (such as Apache NiFi, Talend, or proprietary middleware) is used. Through interface calls (such as RESTful APIs, JDBC connections, or WebService interfaces) or direct database connections, scheduled triggering rules (for example, collection every 10 minutes or upon transaction updates) are set to collect annual and quarterly budget plan details from the budget system; daily unit revenue flow records from the revenue system; each reimbursement or payment record and its approval history from the expenditure system; purchase order, approval, payment, and arrival records from the procurement system; contract terms, amounts, and fulfillment progress from the contract system; and information on inbound and outbound inventory, item transfers, and inventory status from the materials management system. Each data type must include key fields such as timestamps, responsible person codes, and business process IDs for subsequent correlation modeling.

[0042] Secondly, during the data normalization process, various data streams need to be unified and structured. For example, the "amount" field unit is standardized to RMB, and the time field format is standardized to a UTC+8 timestamp (accurate to the second). Field semantics are mapped using a data dictionary to avoid information fragmentation caused by different naming between systems. Collected data is initially divided into structured data (such as tabular financial data), semi-structured data (such as form information extracted from scanned PDF contracts), and log data streams (such as material inbound and outbound logs). Structured data is stored in the DataFrame format, while unstructured data is converted into structure through OCR recognition and text structure analysis.

[0043] In this phase of the experiment, a company's integrated financial platform was selected as a pilot. With a seven-day collection cycle, 34,587 heterogeneous raw data streams were collected, including 7,321 items of budget management data, 4,212 items of revenue data, 10,188 items of expenditure data, 9,876 items of procurement and contract data, and 2,990 items of material data. The total amount of raw data collected reached approximately 1.2GB (after compression), of which nearly 80% was structured form data, with the remainder being semi-structured text.

[0044] Ultimately, through data preprocessing and unified format conversion, a complete data foundation was established for the next steps of data flow channel identification and information entropy analysis, providing multimodal raw data support for the formation of a data-driven internal control model. This step addressed issues such as heterogeneous data formats, dispersed sources, and inconsistent time sequences, creating a controllable data environment for building a unified entropy change modeling system.

[0045] In this embodiment, the specific steps of performing multi-dimensional entropy change feature mining based on the topological structure model and the data entropy change evolution trajectory to generate a cognitive entropy change benchmark feature map for each data stream channel are as follows: Performing cognitive response mapping on the data stream topology structure based on the cognitive entropy change evolution trajectory to obtain the cognitive load hysteresis characteristics of each channel; Performing data flow complexity calculation on the heterogeneous original data flows to obtain complexity parameters of each data flow; Perform instantaneous cognitive entropy mutation detection based on the complexity parameter and the data stream topology to identify the cognitive state volatility of the data stream; The multi-dimensional entropy change characteristics of the data stream are mined for the cognitive load hysteresis characteristics and cognitive state volatility, and a cognitive entropy change benchmark feature map of each data stream channel is generated.

[0046] In this embodiment, after modeling the information entropy evolution trajectory, the temporal variation patterns of the data stream have been quantified as a continuous entropy change curve. The core of this stage is to analyze the response order and response delay of different channels to system state changes, thereby identifying the "cognitive response capability" of each channel. The strength of this capability can be quantified using the "cognitive load hysteresis characteristic." Specifically, this involves determining the time lag after which a channel's information entropy undergoes a significant change when faced with a sudden system change or disturbance. During implementation, the entropy change curves of different data channels are compared, and cross-correlation analysis is used to identify the time delay relationships between channels. The information entropy sequence of each channel is divided into equal-width time windows, and the maximum correlation delay point between each pair of channel entropy value sequences is calculated. If a channel exhibits a persistent time delay compared to its upstream node, meaning its entropy value mutation occurs after the downstream node, it is considered to have a high cognitive hysteresis. In the full network data topology, the hysteresis relationships of all channels can be formed into a hysteresis mapping matrix. By statistically aggregating this matrix, metrics such as the average hysteresis time, hysteresis variance, and minimum and maximum response delay can be calculated for each channel. These metrics collectively define a channel's cognitive load response curve. If a channel's response lag is significantly higher than that of its topological neighbors, it indicates a perception lag in the event chain, high processing pressure, and weak system resilience. In dynamic monitoring, identifying high-lag channels helps proactively identify potential information bottlenecks and propagation delay risks, optimizing flow control strategies and model scheduling. The data structure, change patterns, and predictability of each data channel are quantified to define its complexity parameter. Complexity is an important metric for measuring the structure and uncertainty of a data series, reflecting the processing difficulty and resource consumption imposed by the data on cognitive models. Generally speaking, higher complexity indicates a more irregular and frequently changing data channel, making it more difficult for models to accurately model its state. Time series features of the data stream are extracted, and basic statistics such as volatility (variance), rate of change (mean of first-order differences), and periodicity (peak value of the autocorrelation function) are calculated. These characteristics constitute the foundational time complexity metric. Data structure complexity is assessed by factors such as the number of fields, nesting levels, and data format consistency. Structured data is generally less complex than semi-structured data, but the need for real-time processing of high-frequency data can lead to increased cognitive complexity. Furthermore, methods based on approximate entropy or sample entropy can be introduced to model the predictability of data streams. Such methods evaluate the compressibility and pattern repetitiveness of sequences by comparing the similarities between different subsequences. The more incompressible and irregular a sequence is, the higher its approximate entropy, indicating greater complexity. The complexity indicators of these multiple dimensions are normalized to form a unified complexity parameter vector, which is used to represent the overall complexity of each channel. This parameter is used not only as a model input feature but also in subsequent entropy mutation detection to determine whether the change is "structural" or "anomalous."Determining whether a data stream experiences a sudden change in state at a specific point in time or within a time window—identifying so-called "instantaneous cognitive entropy mutations"—is crucial. Unlike the evolutionary analysis in the previous phase, this phase focuses on dramatic fluctuations in entropy over short periods of time, particularly those that cannot be explained by normal complexity.

[0047] The implementation process consists of three core steps. First, the information entropy of each data channel is monitored over time, and the increase or decrease in entropy is calculated using a sliding time window. If the entropy difference between two consecutive windows exceeds a set threshold and shows a persistent upward (or downward) trend, it is preliminarily identified as a mutation signal. Second, the detected mutation point is compared with the complexity parameters of the channel. If the data itself has high complexity (such as high-frequency noise or strong periodicity), the mutation may be "structured noise" and should not be considered an anomaly. However, if the mutation occurs in a low-complexity channel, it is more likely to be a true abnormal behavior. This joint analysis helps filter out false anomalies and improve detection accuracy. Third, based on the topology of the data stream, it analyzes whether the mutation occurs simultaneously or sequentially in multiple related channels to determine its propagation characteristics. If the mutation affects multiple channels simultaneously or propagates step by step along a specific path, it indicates that the mutation has systemic influence and serves as an important input signal for the cognitive model. Using multiple indicators such as the frequency, intensity, and transmission path length of entropy mutations, a "cognitive state fluctuation index" can be constructed to measure the current system stability and disturbance intensity, providing a basis for decision-making on subsequent anomaly response strategies. The various features extracted above are integrated into a model to form a complete entropy change profile, constructing a "cognitive entropy change baseline feature map" for each data channel. This map serves as structured input for anomaly detection models, scheduling systems, and optimization algorithms, providing a complete and traceable model of data cognitive behavior. Map generation involves four stages: feature aggregation, normalization, structural mapping, and visual representation. Entropy statistical features (such as average entropy, variance, and kurtosis), entropy mutation indicators (frequency and intensity), complexity parameters (structurality and compressibility), topological features (in-degree, out-degree, and betweenness), and cognitive lag indicators (average response delay and variability) for each channel are aggregated into a unified vector space to form a multidimensional feature vector. All features are rescaled using normalization methods to ensure comparability across dimensions. Dimensionality reduction methods (such as principal component analysis or multiscale embedding) are then used to map high-dimensional features into a two- or three-dimensional space for cluster analysis and map visualization. This process classifies different types of data streams into several categories, such as high entropy and high complexity, low entropy and high response, and mutation and diffusion. Each channel corresponds to a complete graph node, whose attributes describe its behavioral characteristics, mutation risk level, propagation role in the network, and response characteristics. This graph is not only useful for offline model analysis but also suitable for rapid screening of key monitoring channels in real-time systems, identifying potentially high-risk nodes, and serving as a crucial reference for optimizing scheduling and response priorities in AI systems.

[0048] In this embodiment, refer to Figure 3 , is a flowchart of the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Performing multi-level semantic analysis on the heterogeneous raw data stream to obtain semantic anomaly features, abnormal time series fluctuations, and abnormal correlation features between data; Adaptive semantic filtering is performed on the semantic abnormality features to obtain filtered optimized abnormal semantic feature data; Perform multi-dimensional semantic space projection on the filtered optimized abnormal semantic feature data to obtain a multi-level semantic space for each data stream; Performing deep semantic analysis on the multi-level semantic space based on abnormal time series fluctuations and abnormal correlation characteristics between data, extracting the deep semantic core and surface semantic representation of the abnormal pattern; The deep semantic core and surface semantic representation are subjected to abnormal behavior semantic mining, and the abnormal pattern semantics are deconstructed layer by layer to construct an accurate semantic feature portrait of each abnormal pattern.

[0049] In this embodiment, within the company's multimodal data-driven internal control decision-making and analysis system, heterogeneous data streams generated by different business systems (such as budgets, revenue, contracts, procurement, and supplies) possess highly diverse structural and semantic characteristics. Therefore, prior to decision modeling, this data must undergo systematic semantic parsing and anomaly feature modeling. The goal of this step is to extract anomaly patterns from the complex raw data through multi-level semantic analysis and adaptive processing, and to construct a "precise semantic feature profile" with deep logical explanatory power, providing an interpretable basis for subsequent anomaly identification and risk response.

[0050] The first stage involves multi-level semantic parsing. Based on natural language processing (NLP) technology, graph neural networks (GNNs), and structured data parsing algorithms, the system performs a hierarchical decomposition of heterogeneous data, including text fields (such as contract terms and procurement instructions), structured table fields (such as budget items and expenditure details), and time series data (such as procurement frequency and budget inflow timing). By constructing a multi-source semantic vector embedding model, both structured and unstructured information are uniformly mapped into a semantic space. During the parsing process, three key anomaly features are extracted: semantic anomalies (such as inconsistencies between budget usage and contract terms); abnormal time series fluctuations (such as expenditures occurring ahead of or behind the budget approval period); and abnormal inter-data associations (such as decreased logical alignment between budget items and procurement plans). In experiments, the Transformer-based embedding model successfully identified over 92% of semantic deviations in an analysis of a unit's 2023 budget-procurement data stream.

[0051] The second stage involves adaptive semantic filtering and enhancement. Because the raw data contains significant noise and fuzzy semantic expressions, the system utilizes a self-attention-based filtering and enhancement model to filter and enhance the initially extracted semantic anomaly features. In practice, a semantic sensitivity factor (SSF) is introduced to measure the impact of semantic variation on internal control rules. The model performs a weighted reconstruction based on the impact of each type of anomaly semantics on the control objective, retaining high-weight anomalies and de-emphasizing low-relevance semantics, resulting in a set of "filtered optimized anomaly semantic feature data."

[0052] The third stage involves multidimensional semantic space projection. Using methods such as t-SNE dimensionality reduction and principal component analysis (PCA), the system maps filtered semantic features into different semantic space dimensions. Each dimension represents the semantic mapping result for a specific business entity (such as budget item, contract type, or procurement process). Taking the "budget procurement semantic space" as an example, dimension one represents "budget intent," dimension two represents "procurement path legitimacy," and dimension three represents "dependency within the approval chain." These multidimensional projections form a visual semantic network, providing a logical structure for further deconstruction of abnormal patterns.

[0053] The fourth stage is deep semantic analysis. Building on the existing multidimensional semantic space and combining the previously identified "abnormal time series fluctuations" and "abnormal inter-data correlation features," a multimodal fusion model (such as Cross-Modal BERT) is used to extract the deep semantic core (i.e., the logical factors that actually trigger the abnormal behavior) and surface semantic representation (i.e., the semantic phenomenon that is externally perceived) of each abnormal pattern. For example, in contract performance data, the identification of "fund use deviating from contract description" reveals that its deep semantic core is "deviation from the implementation of project expenditure classification standards," while its surface semantic representation is "inconsistency between expenditure item names and contractual specifications."

[0054] Finally, the system enters the stage of semantic mining and layer-by-layer semantic deconstruction of abnormal behavior. Based on graph semantic mining and a knowledge rule engine, the system performs semantic path tracing, rule conflict detection, and behavioral pattern clustering based on the combined relationships between deep and surface semantics. By constructing an "abnormal semantic decision tree," each abnormal pattern is deconstructed layer by layer, ultimately generating a "precise semantic feature profile" that includes semantic trigger conditions, data feature ranges, abnormal behavior paths, and possible business consequences.

[0055] In this embodiment, refer to Figure 4 , is a flowchart of the detailed implementation steps of step S3. In this embodiment, the detailed implementation steps of step S3 include: Obtaining a preset decision target instruction; identifying multiple decision targets according to the decision target instruction; calculating real-time weight changes of the decision targets; Based on the multiple decision-making objectives, real-time target conflict strength calculation is performed on the filtered optimized abnormal semantic feature data to obtain the real-time target conflict strength; Performing a conflict situation correlation depth analysis on the real-time weight change and the real-time target conflict intensity to obtain a conflict situation correlation curve; Based on the cognitive entropy change benchmark characteristic map and the conflict situation correlation curve, multi-target coordination request analysis is performed to obtain the coordination request signal for each decision target; A dynamic target priority evaluation is performed on the coordination request signal to obtain the coordination request priority of each decision target.

[0056] In this embodiment, within an organization's internal control system, different decision-making objectives often overlap and conflict in terms of resources and time. For example, between objectives such as budget balance, compliance assurance, efficiency improvement, and procurement execution rate, decision-makers need to dynamically adjust trade-offs in a time-varying environment. Therefore, this step aims to achieve real-time analysis and coordinated optimization of multi-objective conflicts through a systematic conflict modeling and priority assessment approach, ensuring dynamic adaptability and precise execution of internal control actions. The system retrieves pre-set decision-making objective directives from the internal control platform. These directives are typically pre-defined by senior management and include specific objectives (e.g., "budget execution rate for the quarter must not be less than 85%" or "contract performance risk must be controlled within 3%)) and their initial priorities. The system constructs a goal mapping matrix based on rule labels, indicator fields, and policy vectors, abstractly modeling each decision objective in a unified, parameterized form. The system parses the directives and identifies multiple sub-decision objectives with mutually exclusive, competing, or dependent relationships. This process, based on an ontology recognition model and a historical behavior corpus training set, infers a structural relationship graph between objectives through semantic relationships. Taking annual budget and procurement data as an example, structural modeling identified 12 objective combinations with typical conflicting attributes. For example, "timely material supply rate" and "procurement cost savings rate" exhibit a clear inverse coupling relationship. The system incorporates a "real-time weight change calculation mechanism" to adjust the weights of each objective based on dynamic changes in real-time data within the business environment (such as material shortages, contract delays, and approval backlogs). A dynamic assessment model based on a temporal causal Bayesian network (TCBN) is employed to assess the relative impact of objectives on organizational operations in real time. For example, during emergencies, the weight of "emergency procurement completion rate" increases significantly due to its criticality to organizational operations. For scenarios where multiple objectives exist concurrently, the system matches filtered and optimized anomaly semantic feature data with a multi-objective behavioral demand model to calculate the real-time conflict intensity between each objective. A conflict mapping tensor is constructed to measure the degree of conflict overlap in the intersection of different objective resources in the current data scenario, and the conflict intensity values ​​are visualized as a heat map.

[0057] Based on this, the system conducts a deep analysis of the conflict situation correlation between "real-time weight changes" and "goal conflict intensity," constructing a conflict situation association curve. This curve, through time series modeling and multi-goal evolutionary graph analysis, depicts the dynamic evolution of conflict relationships between goals over time, providing a basis for scheduling decisions. For example, it was discovered that some goal conflicts exhibit cyclical fluctuations, corresponding to the end-of-quarter budget sprint or the early-year contract signing period. At the cognitive level, the system jointly models the cognitive entropy change baseline feature graph with the aforementioned conflict situation curve to conduct multi-goal coordination request analysis. This step utilizes a self-attention mechanism to jointly model the semantic drift trends, conflict source dimensions, and risk spillover paths of each goal, and generates corresponding coordination request signals (such as increasing goal flexibility, allowing indicator fluctuation ranges, and invoking external resources for compensation).

[0058] The system comprehensively evaluates the coordination request priority of each target based on target weight, conflict intensity, coordination cost and policy impact. The priority evaluation model is based on fuzzy logic and multi-layer perception network (MLP), dynamically outputs the target coordination priority score, and performs sorting and scheduling in the control center. In this embodiment, step S4 includes the following steps: Calculating the maximum decision capacity of the cognitive entropy change benchmark feature map according to the multiple decision goals to obtain the maximum carrying capacity of each decision goal; Performing multi-time point decision load evaluation based on the precise semantic feature portrait to obtain a multi-time point decision load evaluation curve; Simulating multiple scenario decision-making on the maximum carrying capacity and the multi-time point decision load assessment curve to obtain multiple scenario decision-making simulation data; Dynamic optimal decision boundaries are calculated for multiple scenario decision simulation data based on the coordination request priority, thereby generating a dynamic trade-off constraint range for each decision target.

[0059] In this embodiment, in the multi-objective optimization decision, each decision target has different requirements for system resources, and its impact on the system is also different. Therefore, it is crucial to evaluate the "maximum carrying capacity" of each decision target. This carrying capacity reflects the maximum extent to which a certain target can be effectively executed under the current system state. Specifically, the system evaluates the resource consumption and impact range of each target under specific data characteristics by analyzing the entropy change characteristics of each data channel in the "cognitive entropy change benchmark feature map". A certain target may show higher sensitivity in a high entropy change area, but may cause waste of resources in a low entropy change area. Therefore, by analyzing these characteristics, the system can calculate the maximum carrying capacity of each decision target. In the specific calculation: Input parameters: Each decision-making objective (e.g., anomaly classification, resource scheduling, fault prediction) has a corresponding set of entropy change feature-sensitive dimensions; The entropy structure of different channels in the atlas characterizes its sensory loading capacity and signal permeability; Each channel has known indicators such as volatility, lag, and maximum entropy.

[0060] Based on the information theory model and the system entropy budget model, the maximum decision capacity is defined as: ; Among them, α_j^((i)): the weight of the i-th target on the j-th channel; V_j: channel j volatility; L_j: response hysteresis of channel j; E_j^max: maximum information entropy that can be accommodated by the channel; A linear programming constraint solving model is adopted; a maximum acceptable load is set for each target, and the characteristic spectrum value of each channel in the current state is normalized. The final output is a maximum load capacity C_max^((i)) corresponding to each target.

[0061] Over time, the system's operating state and external environment may change, affecting the performance of each decision objective. Therefore, evaluating changes in decision load at multiple points in time is crucial for dynamic decision optimization. The system analyzes the performance of each objective at different points in time in the "Precise Semantic Feature Profile" and, combined with the characteristics of real-time data streams, calculates the load of each objective at each point in time. These evaluation results are plotted as a "Multi-Time-Point Decision Load Evaluation Curve" to reflect the load variation trends of each objective over different time periods. In experiments, the system typically sets multiple time windows (such as 10 seconds, 30 seconds, and 60 seconds) and calculates the load of each objective within each window. By comparing the evaluation results across different time windows, the system can identify patterns in load variation, providing a basis for subsequent decision optimization. In real-world applications, the system may face a variety of operating scenarios, such as high load, low latency, and large data volumes. To cope with these changing scenarios, the system requires multi-scenario decision simulation. By taking the "maximum load capacity" and the "Multi-Time-Point Decision Load Evaluation Curve" as input, the system can construct multiple scenario models to simulate the performance of each decision objective under different scenarios. These simulation results, known as "scenario-based decision simulation data," are used to evaluate the system's decision-making effectiveness and resource utilization under various conditions. During experiments, the system typically sets up multiple simulation scenarios (such as high-load scenarios, low-latency scenarios, and large data volume scenarios) and conducts decision-making simulations in each scenario. By comparing simulation results across different scenarios, the system can identify the optimal decision-making strategy, providing guidance for practical applications. In multi-objective optimization decisions, objectives may conflict or compete with each other, and effectively balancing these objectives is a key issue. The system analyzes "coordination request priorities" and combines them with "scenario-based decision simulation data" to calculate the optimal decision boundaries for each objective under different scenarios. These boundaries reflect the optimal state that each objective can achieve under specific conditions. By analyzing these boundaries, the system can determine the "dynamic trade-off constraint range" for each decision objective—the optimal range acceptable for each objective under the current environment. The system typically uses multi-objective optimization algorithms, such as genetic algorithms and particle swarm optimization, combined with real-world data for calculations. By comparing the results of different algorithms, the system can select the decision strategy that best suits the current application scenario.

[0062] In this embodiment, step S5 includes the following steps: Solve multi-objective dynamic constraint responses based on dynamic trade-off constraint ranges to generate multiple target decision paths; Conduct risk path deduction analysis on multiple target decision paths to obtain the risk deduction characteristics of each path; Conduct adaptive decision path planning based on the risk deduction characteristics and build a multi-objective collaborative decision execution strategy; Define immediate control instructions based on the multi-objective collaborative decision-making execution strategy and generate multiple decision-making target execution control instructions; Based on the decision target, the control instruction is executed to perform immediate decision processing, and the real-time execution effect of each decision target is collected.

[0063] In this embodiment, the dynamic trade-off constraint range is converted into a set of constraint boundary parameters, corresponding to the load limit, preferred direction, and critical threshold for each objective. Based on these constraints, the system uses combinatorial optimization strategies (such as heuristic search, rule-matching path compression, and weight-adjusted sliding windows) to search for possible optimal or suboptimal paths in the policy space. In experiments, the search depth for path generation is typically set to 5-7 layers, allowing for dynamic shifts in objective priority between nodes. Each path corresponds to a set of policy execution order, impact scope, and expected benefits. Each generated path logically meets current system resource constraints and controllable objective conflicts, while also ensuring execution stability and predictability. The output path set includes not only the execution order for each objective but also dynamic constraint offsets, weight adjustment coefficients, and estimated execution resource usage, facilitating subsequent deduction and scheduling. After path generation, the system evaluates the potential uncertainty, failure probability, and anomaly probability of these paths during execution. This step uses a "risk path deduction" mechanism to simulate and conduct contingency assessments on the execution of each candidate path to identify its risk characteristics. The risk analysis relies on a database of historical anomaly events, the evolutionary trends of entropy patterns, and system sensitivity indicators for each node along the path. In a simulation environment, the system conducts multiple rounds of "time-slice" risk analysis on each path. This involves breaking the path execution into multiple time segments and evaluating the system response (such as latency, mutation, and conflict level) for each segment. In the experimental deployment, each path is simulated for 10 to 20 rounds, generating a set of risk signature data with each round, including but not limited to: risk exposure point (RBP), fluctuation response intensity (FRI), execution time offset (DTA), and resource overrun frequency (ORF). This data is aggregated and modeled to form a "risk prediction signature set" for the path. This signature set is used for subsequent path selection and strategy planning, and can be compared with existing system robustness assessment models to identify high-risk and robust paths. Paths with lower risk are more likely to be used to execute high-priority objectives.

[0064] After understanding the risk characteristics of each path, the system enters the "path selection and combination planning" phase. The goal of this step is to select the optimal or most adaptable path combination based on the deduction results, thereby constructing a decision-making strategy that can coordinate execution across multiple objectives. The system first matches the risk characteristics of each path with the objective requirements, prioritizing paths with low risk exposure, minimal path redundancy, and balanced resource utilization. Next, based on inter-objective synergies (e.g., non-conflicting or synergistic effects), it attempts to combine several paths into "cooperative execution groups." Cooperative path groups can simultaneously meet multiple objectives while sharing some computing resources, improving decision-making efficiency. In experiments, the path planning module typically uses adaptive strategy combination algorithms (such as Pareto frontier heuristic fusion and dynamic swarm evolution strategies) to evaluate the value of path combinations. Evaluation criteria include, but are not limited to, average response time, improved resource utilization, and reduced redundancy and conflict. The output is a set of "multi-objective collaborative decision-making execution strategies." Each strategy defines a set of objectives, a priority ranking, an estimated duration, expected output, and a fallback strategy. This strategy set is directly fed into the control instruction module in the next stage. Abstract execution strategies are converted into control instructions that can be directly invoked within the system, driving real-time adjustments to data flows, model resources, and scheduling behaviors across various processing modules. Each collaborative decision-making strategy is translated into a structured set of control instructions, including but not limited to target identification, resource quotas, execution thresholds, feedback mechanisms, fault-tolerance windows, and fallback path instructions. The system automatically arranges the scheduling order and activation timing of control signals based on the priority assigned to each target's strategy and load forecast. In actual deployment, each control instruction is encoded as a standard operation template (such as a JSON-structured instruction packet) and distributed via a scheduling bus to each compute node, analysis model, response module, or alarm system. Upon receiving the instruction, each node immediately adjusts its operating parameters, such as dynamically switching data filtering accuracy, increasing the rate of model anomaly detection, and shortening alarm delays. This process requires instruction redundancy control to ensure that system load does not surge due to overlapping strategies. Experiments recommend a batch size of no more than 20 instructions per round, with a maximum execution cycle of 30 seconds for each instruction. Interrupt and reset interfaces are also provided to ensure controllability and rollback. After the control instruction is issued, the system enters the "instant decision processing" phase. During this phase, each system module performs tasks such as data processing, model response, and strategy adjustment according to instructions, while simultaneously collecting the execution status and performance indicators of each target in real time. The system employs a data monitoring and feedback mechanism to track the execution process throughout its lifecycle. Key parameters collected include response time, execution success rate, changes in data throughput, anomaly detection accuracy, alert timeliness, and changes in resource utilization. The execution results of each target are continuously recorded in the "Execution Results Record Library" and archived in layers by timestamp for subsequent optimization strategy adjustments and reinforcement learning feedback.In experiments, a sampling frequency of once per second is recommended, with a minimum of 10 data fields recorded for a single target, ensuring that the sampling dimensions cover the four major categories of indicators: model, data, system, and resources. Furthermore, the system has an exception response trigger mechanism. If the execution of a target deviates from the policy expectations by more than a set threshold (e.g., response time deviation greater than 20%), a rollback request or policy reassessment signal will be immediately generated, triggering subsequent adaptive adjustment mechanisms to maintain system stability and agility.

[0065] In this embodiment, step S6 includes the following steps: Identify the decision target benchmark execution effect based on the preset decision target instructions; Identify the decision execution response deviation of each decision target based on the decision target benchmark execution effect, and mark the decision target with the execution deviation; Conduct in-depth research on the causes of deviations in decision-making targets and identify the root causes of decision failures; Based on the fundamental factors of decision failure, the multi-objective collaborative decision execution strategy is weighted and the decision strategy is dynamically reconstructed to obtain the globally optimized decision parameters; Perform deep optimization of global optimization decision parameters through strategy iteration to build an internal control management decision-making intelligent entity.

[0066] In this embodiment, during the multi-objective optimization decision-making process, a baseline performance evaluation for each decision objective must first be defined to facilitate subsequent deviation identification and optimization. The core task of this process is to quantitatively evaluate the initial performance of each decision objective using pre-set decision objective instructions. By setting standardized target performance metrics, such as response time, resource utilization, and execution success rate, the system can accurately identify the expected performance of each objective. In practice, the system extracts the performance criteria for each objective from the decision objective instructions. These criteria are typically based on historical data, business requirements, or expert advice. If the performance criteria for a particular objective is "response time within 30 seconds," the system monitors and records the actual response time for each execution. By comparing these values ​​with the pre-set standard, the system generates a baseline performance for each objective, which is then used to identify deviations. During experiments, these baseline performance results are collected and validated in multiple scenarios to ensure their adaptability and accuracy. Each objective is executed in a standard scenario, and their performance is compared, resulting in an average value that serves as the benchmark. The standard value is typically stable over 20 to 50 experimental cycles to ensure representativeness. After the baseline performance is determined, the system monitors the execution of the decision-making objectives in real time to identify deviations from the baseline performance during actual execution. This process relies primarily on real-time data monitoring and comparative analysis, identifying deviations and flagging decision-making objectives whose performance does not meet expectations.

[0067] The system incorporates a real-time execution monitoring mechanism to dynamically track and collect real-time data on the execution of each decision target. For a real-time response target, the system monitors key metrics such as response time, execution completion rate, and resource usage. Real-time execution performance is compared and analyzed against a preset baseline. Any deviation from a certain threshold is labeled as an "execution deviation." In experiments, a threshold range is typically set (for example, ±10%). Once the target execution deviation exceeds this range, the system immediately records and marks the relevant decision target as a "deviation target." This helps identify issues promptly and initiate further analysis. By continuously tracking these deviation targets, the system accumulates a large amount of execution deviation data, providing data support for subsequent in-depth analysis of the causes of deviations. In-depth analysis of the causes of deviations is a key step in this method, aiming to identify the root causes of execution deviations so that targeted optimization can be implemented. Deviations can arise from multiple factors, such as data anomalies, insufficient resource allocation, and inaccurate algorithmic models. The system uses multi-dimensional analysis techniques to conduct in-depth analysis of deviation targets. The system retraces the target's execution path to analyze whether changes in the external environment (such as sudden fluctuations in data flow or drastic changes in system resources) may have affected the target's execution. Secondly, the system combines historical execution records, log data, and model output to identify potential flaws in the algorithm or model's execution. The system also compares the target's execution pattern with that of other normal targets, identifying potential discrepancies and conducting in-depth analysis.

[0068] In experiments, the system often uses causal inference models or feature importance analysis methods from machine learning to help identify the root causes of deviations. Through feature selection and dimensionality reduction techniques, the system can quickly identify key factors associated with decision failures, such as the impact of changes in certain external parameters on goal execution or issues with data input quality. After identifying the root causes of decision failures, the system enters the "decision strategy reconstruction" phase. The core task of this phase is to adjust the existing multi-objective collaborative decision-making strategy based on the identified failure factors, optimizing its objective weights and decision execution process. The system first quantitatively assesses the key factors influencing decision failures. If abnormal fluctuations in the data flow are the primary factor causing execution deviations, the system adjusts the weights of the corresponding objectives to increase the influence of this factor on the decision. Furthermore, the system can replan the decision strategy based on actual conditions, achieving global optimization by adjusting the order of goal execution and resource allocation priorities. In experiments, methods such as genetic algorithms and particle swarm optimization are often used to dynamically reconstruct decision strategies. The optimal strategy combination is selected by simulating the execution effects of different strategies. After each strategy optimization, the system uses backtesting and forward testing to verify whether the new weighting settings and execution strategies effectively reduce execution deviation and improve target performance. The reconstructed decision-making strategy includes adjusted weights for each objective, a new resource scheduling plan, and execution priorities, ensuring more efficient coordination and optimization across multiple objectives across the entire system. After obtaining the globally optimized decision parameters, the system enters the final policy iteration optimization phase. The goal of this phase is to further optimize the performance of the decision-making agent through continuous policy iteration, enabling it to more efficiently execute multi-objective optimization decisions in future real-time data streams and complex scenarios. The policy iteration process typically consists of two main steps. First, based on the current optimized decision parameters, the system conducts multiple rounds of simulation and testing to evaluate the policy's performance. By collecting various performance indicators during execution, the system further adjusts the policy. Second, the system employs intelligent methods such as reinforcement learning to enable the decision-making agent to autonomously learn and gradually optimize its decision-making process. This iterative optimization based on feedback learning ensures that the system can continuously improve and adapt to new situations in complex and dynamically changing environments. Experiments typically set a set number of iterations (e.g., 10, 20, etc.) and test the robustness and adaptability of the strategy in a dynamically changing data environment. Feedback from each iteration is used to adjust the agent's decision-making behavior, further improving its performance in unknown environments. After repeated policy optimization, the system develops a highly intelligent, adaptively adjusted internal control management decision-making agent capable of continuously performing tasks such as detecting anomalies in complex data streams and optimizing decision-making.

[0069] In this embodiment, a unit internal control management decision-making system based on multimodal data analysis is provided, which is used to execute the unit internal control management decision-making method based on multimodal data analysis as described above, including: The data stream entropy change analysis module is used to collect heterogeneous raw data streams from the unit's internal control management platform; it performs differential analysis of entropy change in the spatiotemporal dimension and multi-dimensional entropy change feature mining to generate a cognitive entropy change benchmark feature map; A semantic parsing module is used to perform multi-level semantic parsing on the heterogeneous raw data streams, and to perform semantic deconstruction of abnormal patterns layer by layer, thereby constructing an accurate semantic feature portrait of each abnormal pattern; The multi-objective coordination module is used to obtain preset decision target instructions, perform multi-objective coordination request analysis and dynamic target priority evaluation on the cognitive entropy change benchmark feature map, and obtain the coordination request priority; A decision boundary module is used to calculate the maximum decision capacity of the cognitive entropy change benchmark feature map and to calculate the dynamic optimal decision boundary based on the coordination request priority and the precise semantic feature portrait, thereby generating a dynamic trade-off constraint range; The constraint solving module is used to solve multi-objective dynamic constraint responses based on the dynamic trade-off constraint range and perform instant decision processing to obtain real-time execution results; The internal control management decision-making module is used to identify decision-making response deviations in real-time execution effects, perform deep optimization of strategy iterations, and build an internal control management decision-making intelligent entity.

[0070] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced within the present invention.

[0071] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A unit internal control management decision-making method based on multimodal data analysis, characterized in that: The following steps are involved: Step S1: Collect heterogeneous raw data streams from the unit's internal control management platform; perform temporal and spatial entropy change differential analysis and multi-dimensional entropy change feature mining to generate a cognitive entropy change benchmark feature map; Step S2: Perform multi-level semantic analysis on the heterogeneous original data stream, and perform semantic deconstruction of abnormal patterns layer by layer to construct an accurate semantic feature portrait of each abnormal pattern; Step S3: Obtain the preset decision target instruction, perform multi-target coordination request analysis and dynamic target priority evaluation on the cognitive entropy change benchmark feature map, and obtain the coordination request priority; Step S4: Calculating the maximum decision capacity of the cognitive entropy change benchmark feature map, and calculating the dynamic optimal decision boundary based on the coordination request priority and the precise semantic feature portrait, thereby generating a dynamic trade-off constraint range; Step S5: solving multi-objective dynamic constraint responses based on the dynamic trade-off constraint range, and performing instant decision processing to obtain real-time execution results; Step S6: Identify the decision execution response deviation of the real-time execution effect, perform deep optimization of strategy iteration, and build an internal control management decision-making intelligent body.

2. The unit internal control management decision-making method based on multimodal data analysis according to claim 1 is characterized in that: The specific steps of step S1 are: Collect heterogeneous raw data streams from the internal control management platform of the unit; the heterogeneous raw data streams include budget management data, revenue management data, expenditure management data, procurement management data, contract management data, and material management data; Identifying data flow channels of the heterogeneous original data flow, and performing information entropy distribution calculation to generate information entropy density of each data flow channel; Based on the information entropy density, a temporal and spatial dimension entropy change differential analysis is performed to obtain the data entropy change evolution trajectory of each data stream channel; Performing data topology analysis and modeling on the heterogeneous original data streams, and constructing a topology model for each data stream; Multi-dimensional entropy change characteristics are mined based on the topological structure model and the data entropy change evolution trajectory to generate a cognitive entropy change benchmark feature map.

3. The unit internal control management decision-making method based on multimodal data analysis according to claim 2 is characterized in that: The specific steps of performing multi-dimensional entropy change feature mining based on the topological structure model and the data entropy change evolution trajectory to generate a cognitive entropy change benchmark feature map for each data stream channel are as follows: Performing cognitive response mapping on the data stream topology structure based on the cognitive entropy change evolution trajectory to obtain the cognitive load hysteresis characteristics of each channel; Performing data flow complexity calculation on the heterogeneous original data flows to obtain complexity parameters of each data flow; Perform instantaneous cognitive entropy mutation detection based on the complexity parameter and the data stream topology to identify the cognitive state volatility of the data stream; The multi-dimensional entropy change characteristics of the data stream are mined for the cognitive load hysteresis characteristics and cognitive state volatility, and a cognitive entropy change benchmark feature map of each data stream channel is generated.

4. The unit internal control management decision-making method based on multimodal data analysis according to claim 1 is characterized in that: The specific steps of step S2 are: Performing multi-level semantic analysis on the heterogeneous raw data stream to obtain semantic anomaly features, abnormal time series fluctuations, and abnormal correlation features between data; Adaptive semantic filtering is performed on the semantic abnormality features to obtain filtered optimized abnormal semantic feature data; Perform multi-dimensional semantic space projection on the filtered optimized abnormal semantic feature data to obtain a multi-level semantic space for each data stream; Performing deep semantic analysis on the multi-level semantic space based on abnormal time series fluctuations and abnormal correlation characteristics between data, extracting the deep semantic core and surface semantic representation of the abnormal pattern; The deep semantic core and surface semantic representation are subjected to abnormal behavior semantic mining, and the abnormal pattern semantics are deconstructed layer by layer to construct an accurate semantic feature portrait of each abnormal pattern.

5. The unit internal control management decision-making method based on multimodal data analysis according to claim 1 is characterized in that: The specific steps of step S3 are: Obtaining a preset decision target instruction; identifying multiple decision targets according to the decision target instruction; calculating real-time weight changes of the decision targets; Based on the multiple decision-making objectives, real-time target conflict strength calculation is performed on the filtered optimized abnormal semantic feature data to obtain the real-time target conflict strength; Performing a conflict situation correlation depth analysis on the real-time weight change and the real-time target conflict intensity to obtain a conflict situation correlation curve; Based on the cognitive entropy change benchmark characteristic map and the conflict situation correlation curve, multi-target coordination request analysis is performed to obtain the coordination request signal for each decision target; A dynamic target priority evaluation is performed on the coordination request signal to obtain the coordination request priority of each decision target.

6. The unit internal control management decision-making method based on multimodal data analysis according to claim 1 is characterized in that: The specific steps of step S4 are: Calculating the maximum decision capacity of the cognitive entropy change benchmark feature map according to the multiple decision goals to obtain the maximum carrying capacity of each decision goal; Performing multi-time point decision load evaluation based on the precise semantic feature portrait to obtain a multi-time point decision load evaluation curve; Simulating multiple scenario decision-making on the maximum carrying capacity and the multi-time point decision load assessment curve to obtain multiple scenario decision-making simulation data; Dynamic optimal decision boundaries are calculated for multiple scenario decision simulation data based on the coordination request priority, thereby generating a dynamic trade-off constraint range for each decision target.

7. The unit internal control management decision-making method based on multimodal data analysis according to claim 1 is characterized in that: The specific steps of step S5 are: Solve multi-objective dynamic constraint responses based on dynamic trade-off constraint ranges to generate multiple target decision paths; Conduct risk path deduction analysis on multiple target decision paths to obtain the risk deduction characteristics of each path; Conduct adaptive decision path planning based on the risk deduction characteristics and build a multi-objective collaborative decision execution strategy; Define immediate control instructions based on the multi-objective collaborative decision-making execution strategy and generate multiple decision-making target execution control instructions; Based on the decision target, the control instruction is executed to perform immediate decision processing, and the real-time execution effect of each decision target is collected.

8. The unit internal control management decision-making method based on multimodal data analysis according to claim 1 is characterized in that: The specific steps of step S6 are: Identify the decision target benchmark execution effect based on the preset decision target instructions; Identify the decision execution response deviation of each decision target based on the decision target benchmark execution effect, and mark the decision target with the execution deviation; Conduct in-depth research on the causes of deviations in decision-making targets and identify the root causes of decision failures; Based on the fundamental factors of decision failure, the multi-objective collaborative decision execution strategy is weighted and the decision strategy is dynamically reconstructed to obtain the globally optimized decision parameters; Perform deep optimization of global optimization decision parameters through strategy iteration to build an internal control management decision-making intelligent entity.

9. A unit internal control management decision-making system based on multimodal data analysis, characterized in that: The method for implementing the unit internal control management decision-making method based on multimodal data analysis as claimed in claim 1 comprises: The data stream entropy change analysis module is used to collect heterogeneous raw data streams from the unit's internal control management platform; it performs differential analysis of entropy change in the spatiotemporal dimension and multi-dimensional entropy change feature mining to generate a cognitive entropy change benchmark feature map; A semantic parsing module is used to perform multi-level semantic parsing on the heterogeneous raw data streams, and to perform semantic deconstruction of abnormal patterns layer by layer, thereby constructing an accurate semantic feature portrait of each abnormal pattern; The multi-objective coordination module is used to obtain preset decision target instructions, perform multi-objective coordination request analysis and dynamic target priority evaluation on the cognitive entropy change benchmark feature map, and obtain the coordination request priority; A decision boundary module is used to calculate the maximum decision capacity of the cognitive entropy change benchmark feature map and to calculate the dynamic optimal decision boundary based on the coordination request priority and the precise semantic feature portrait, thereby generating a dynamic trade-off constraint range; The constraint solving module is used to solve multi-objective dynamic constraint responses based on the dynamic trade-off constraint range and perform instant decision processing to obtain real-time execution results; The internal control management decision-making module is used to identify decision-making response deviations in real-time execution effects, perform deep optimization of strategy iterations, and build an internal control management decision-making intelligent entity.

Images