Project whole-process penetrating management and control method and system based on multi-source data fusion

Abstract

The application discloses a project whole-process penetration control method and system based on multi-source data fusion, and relates to the technical field of engineering project risk control and time sequence logic formal verification. The method comprises the following steps: acquiring a control instruction generated by a multi-agent cooperation system; extracting an attention weight distribution in an inference process to lock a target subgraph in a fusion knowledge graph, and cutting a global compliance rule to obtain a local formal context according to the target subgraph; mapping numerical variables in the local formal context into interval logic operators and constructing a local linear time sequence logic formula; and performing safety verification on the formula, and executing the control instruction when the formula passes the safety verification. The application is used to solve the problems of instruction illusion risk caused by the probabilistic decision of the generative model and the calculation space explosion caused by the global system state verification.

Description

Technical Field

[0001] This invention relates to the field of engineering project risk control and temporal logic formal verification technology, and more specifically, to a method and system for full-process penetration control of projects based on multi-source data fusion. Background Technology

[0002] The comprehensive management and control of modern large-scale engineering projects relies on the processing of massive amounts of multi-source heterogeneous data, encompassing high-frequency IoT sensing data and low-frequency discrete business data. The core of achieving penetrating management and control lies in the accurate fusion of cross-frequency domain data, continuous monitoring of anomalies and risks, and the reliable issuance of security and compliance control instructions.

[0003] Currently, most mainstream project management systems employ conventional graph neural networks for feature concatenation and aggregation in Euclidean space, combined with deep neural networks for abnormal behavior classification and early warning, and attempt to introduce generative large models to assist in outputting management strategies. To ensure the security of instructions before execution, some solutions use theorem provers or model checking tools for conventional formal verification based on the global system state space.

[0004] However, existing technologies have significant limitations: First, directly splicing heterogeneous and multi-frequency data can easily cause breakages in the underlying business causal topology, and conventional Euclidean space is prone to over-smoothing of feature representation when processing deep-level architecture graphs; Second, anomaly detection models generally face severe "catastrophic forgetting" when dealing with conceptual drift in dynamic business environments and triggering incremental learning, leading to distortion in the detection of historical core risks; Third, the probabilistic decision-making of generative large models is prone to the risk of industrial instruction illusion. If the state machine of a global system involving continuous numerical variables is directly subjected to formal security deduction, it will inevitably lead to a computational explosion in the state space, which cannot meet the stringent requirements of decision security and computational efficiency in large-scale industrial scenarios. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of existing technologies, this invention provides a project-wide process penetration and control method based on multi-source data fusion. By extracting the attention weights of multiple agents to lock the target subgraph and pruning global compliance rules, the numerical variables in the local formal context are mapped to interval logic operators to construct and verify local linear temporal logic formulas, thereby solving the problems of instruction illusion caused by probabilistic decision-making in generative models and spatial computation explosion under global state verification.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for end-to-end project control based on multi-source data fusion includes the following steps: obtaining control instructions generated by a multi-agent collaborative system; constructing a local linear temporal logic formula associated with the control instructions; performing security verification on the local linear temporal logic formula; and executing the control instructions when the verification passes. Specifically, constructing the local linear temporal logic formula involves: extracting the attention weight distribution of the multi-agent collaborative system during the inference process; locking the target subgraph in a pre-constructed multi-source heterogeneous data fusion knowledge graph based on the attention weight distribution; pruning global business compliance rules based on the target subgraph to obtain a local formal context; mapping the numerical variables in the local formal context to interval logic operators, and constructing the local linear temporal logic formula based on the interval logic operators.

[0007] In addition, this invention also provides a project end-to-end penetration control system based on multi-source data fusion, comprising the following modules: an instruction generation module, used to acquire a multi-source heterogeneous data fusion knowledge graph and abnormal behavior warning sequences constructed from multi-source heterogeneous data of the project; and to generate control instructions for the warning sequences using a multi-agent collaborative system; a graph pruning module, used to extract the attention weight distribution of the multi-agent collaborative system during the inference process, lock the target subgraph associated with the attention weight distribution in the graph, and prune global business compliance rules based on the target subgraph to obtain a local formal context; a logic mapping module, used to map the numerical variables in the local formal context to interval logic operators to construct a local linear temporal logic formula; and an instruction execution module, used to perform security verification on the local linear temporal logic formula and execute the control instructions when the verification is successful.

[0008] The technical effects and advantages of this invention's project end-to-end penetration control method based on multi-source data fusion are as follows: This invention achieves both intelligent and highly reliable project penetration management by combining control commands generated by a multi-agent collaborative system with temporal logic security verification. In the verification preparation phase, by extracting the attention weight distribution during the multi-agent inference process, the target subgraph is accurately located in a multi-source heterogeneous data fusion knowledge graph. Based on this, global business compliance rules are dynamically pruned to obtain a local formal context, thereby significantly compressing the system state to be verified from both the topology network and rule boundaries, objectively solving the inherent state space explosion problem in traditional global formal verification. Furthermore, by mapping numerical variables in the local formal context to interval logic operators and constructing local linear temporal logic formulas, complex continuous numerical variables are transformed into finite symbolic interval logic for security verification. This solution significantly improves verification computation efficiency and shortens the delay in issuing control commands, while rigorous formal logic constraints eliminate the uncertainty brought about by probabilistic decision-making in generative models, ensuring the high security and controllability of the final executed penetration management commands in the underlying system. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the process of the project full-process penetration control method based on multi-source data fusion provided in an embodiment of the present invention; Figure 2 This is a simple complex persistent homology persistence graph provided in an embodiment of the present invention; Figure 3 This is a simulation diagram of the three-dimensional mapping distribution of project entity features based on the hyperbolic manifold of a Poincaré sphere, provided in an embodiment of the present invention. Figure 4 This is a heatmap of the self-attention weight distribution of a multi-agent system provided in an embodiment of the present invention. Figure 5 A block diagram of a project end-to-end penetration control system based on multi-source data fusion provided in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0011] Example 1, Figure 1 This invention presents a method for comprehensive project management and control based on multi-source data fusion, comprising the following steps: S1, obtain the fused knowledge graph and abnormal behavior early warning sequence constructed from multi-source heterogeneous data of the project.

[0012] In this embodiment, the step of acquiring the fused knowledge graph and abnormal behavior early warning sequence constructed from multi-source heterogeneous data of the project includes: S101, map IoT sensing data and discrete business data from the multi-source heterogeneous data of the project to the topological space to construct a simple complex; In this embodiment, the multi-source heterogeneous data for the project originates from different business systems and on-site sensing terminals involved in the entire project management process. Among them, the IoT sensing data is a high-frequency matrix containing timestamps and continuous sensor values. (1) in, Indicates the number of time series sampling points. This indicates sensor characteristic dimensions, including the operating status of construction equipment, current, voltage, vibration frequency, and environmental parameters; Discrete business data consists of low-frequency structured data from bidding systems, contract management systems, and financial settlement systems, represented as follows: (2) in, Indicates the number of business events. This indicates business attribute dimensions, including approval status, contract amount, supplier identification, and payment node.

[0013] Furthermore, firstly, through feature standardization and embedding encoding, the matrix is... and Each sample in the dataset is mapped to a unified Euclidean space. (For example ), to obtain the point set Based on this, the distance threshold used to construct topological adjacency relationships is set as follows: , is used to represent the adjacency distance threshold used to determine whether a connection is established between nodes during the Vietoris-Rips complex construction process.

[0014] Each sample is mapped to a unified Euclidean space, specifically as follows: For continuous IoT sensing data Construct a continuous feature extractor that includes a linear mapping layer and an activation function. Through formula From feature dimensions Mapped to ,in The weight matrix is ​​a learnable matrix. For bias vectors, for The corresponding embedding vector; For discrete business data The categorical attributes are transformed into dense vectors using one-hot encoding or a lookup table, concatenated with the numerical attributes, and then input into the business feature extraction network. Through formula Project the feature dimensions to .in for The input feature matrix is ​​obtained by concatenating the numerical and categorical attributes. To extract learnable weight matrices from the network for business features, To extract bias vectors from the network for business features, for The corresponding embedding vector.

[0015] The adjacency distance threshold for determining whether a connection is established between nodes is based on the following Euclidean distance metric: (3) When satisfied At the node and Establish a one-dimensional simplex between them; further, when any The pairwise distance between any two nodes is no greater than 1. At that time, construct the corresponding The simplex is dimensional, thus generating a set of simplex complexes. .

[0016] In this embodiment, data originally from different systems and frequencies are uniformly embedded into the same topological space and their high-order correlation structure is constructed through the above method. This solves the problem of the lack of a unified relationship modeling mechanism between multi-source heterogeneous data and lays a structural foundation for subsequent topological feature extraction and cross-domain fusion.

[0017] S102, calculate the continuous homology features of the simple complex under different filtering scales to obtain the topological invariant matrix; Furthermore, in the simple complex Based on this, a filtering scale parameter is introduced: (4) in, The preset maximum filtering scale threshold. This represents the filtering scale parameter used for continuous cohomology analysis, which characterizes the evolution of the topology at different spatial resolutions by gradually increasing the adjacency radius.

[0018] along with As the value gradually increases, the simple complex is filtered, and the generation and disappearance processes of topological features (including connected components and ring structures) are recorded. Define the first... The Betti number is: (5) in, of Indicates the number of connected components. This indicates the number of one-dimensional ring structures.

[0019] Furthermore, each topological feature is represented as its "birth-death" interval. A persistent barcode is constructed. By vectorizing the persistent barcode (including interval length statistics, segment counting, or persistent graph embedding), topological information of different dimensions is transformed into fixed-length feature vectors, which are then concatenated at multiple filtering scales to form a topological invariant matrix. Specifically, the vectorized mapping of the persistent barcode and the topological invariant matrix The calculation process is as follows: Step A: Filter the continuous scale intervals Divided into Define the n equally spaced discrete sampling points, and define the nth Each discrete filtering scale is ,in In this embodiment, it can be set .

[0020] Step B: For each discrete filtering scale Traverse all topological feature regions in the persistent barcode If and only if At that time, determine the topological feature at the current scale. It is currently in a "survival" state.

[0021] For the current "survival" topological characteristics, calculate the following three quantitative statistical indicators for 0-dimensional (connected components) and 1-dimensional (one-dimensional ring structure): (1) Number of survivors (segmented count): i.e., the current Betti number and ; (2) Interval length mean (lifetime mean): Calculates the persistence of all current survival features. The average value; (3) Maximum interval length (maximum lifetime): Get the persistence of all currently surviving features. The maximum value.

[0022] By concatenating the statistical indicators from the 0-dimensional and 1-dimensional dimensions mentioned above, we obtain the results at the current scale. Fixed-length feature vectors Since three indicators are extracted from both 0-dimensional and 1-dimensional features, the topological feature dimensions are... ,Right now .

[0023] Step C: [The above...] The feature vectors calculated at discrete scales according to From 0 to The incremental sampling order is used to concatenate the feature vectors corresponding to each sampling point row by row to form a topological invariant matrix. .

[0024] like Figure 2 As shown in the figure, this embodiment of the invention provides a schematic diagram of the simulation results of a persistent graph of a simple complex under continuous homology evolution. The horizontal axis in the figure represents the topological features at the filtering scale. The scatter plot shows the birth radius on the bottom axis and the death radius on the top axis. The scatter plot deviates from the diagonal. The greater the vertical distance, the longer the lifespan and the stronger the structural stability of the topological feature when fusing multi-source heterogeneous data (such as high-frequency sensor data and low-frequency financial data). By extracting the highly significant zero-dimensional homology classes (representing global business connectivity branches) and one-dimensional homology classes (representing business logic loop structures) far from the diagonal in the figure, random high-frequency temporal noise close to the diagonal region is effectively filtered from the mathematical topology level.

[0025] S103, perform feature alignment between the IoT sensing data and the discrete business data based on the topological invariant matrix; In this embodiment, to address the inconsistency between high-frequency time-series data of the Internet of Things and low-frequency discrete data of business in terms of feature space and time scale, a feature alignment mechanism based on topological constraints is introduced.

[0026] Specifically, the original input matrix and With topological invariant matrices respectively Feature concatenation is performed to obtain the enhanced representation: (6) in, This indicates a feature dimension concatenation operation.

[0027] Furthermore, a cross-attention mechanism is introduced for cross-modal alignment. As a query matrix and as a key matrix, calculate the attention mapping: (7) in, For querying the matrix; The key matrix; It is a value matrix; It is a trainable parameter matrix; This is the feature dimension scaling factor.

[0028] Through the aforementioned cross-attention mechanism, IoT enhanced features are output respectively. With business enhancement features By concatenating the two data points along the sample dimension, semantic alignment between IoT sensing data and business data is achieved, outputting a unified, isomorphic feature representation: (8) in, The dimension of the isomorphic feature space is preset.

[0029] In this embodiment, by introducing topological invariants as global structural constraints and combining cross-attention to achieve cross-modal information interaction, the problem of temporal semantic misalignment and feature distortion caused by direct splicing of different frequency data is effectively solved, thereby achieving accurate alignment of multi-source data in a unified semantic space.

[0030] S104 maps the features of the project entities to be aligned to the Poincaré sphere hyperbolic geometry space; In this embodiment, after completing cross-modal feature alignment, a unified representation of the project entity feature matrix is ​​obtained. To characterize the inherent hierarchical structure in engineering projects (such as "group - subsidiary - project department - construction team"), a Poincaré sphere model is introduced to geometrically reconstruct this feature.

[0031] Specifically, define The hyperbolic space of the Viponcaré sphere is: (9) in, Represents the spatial curvature parameter. Indicates the embedding dimension (set in this embodiment) ), is the feature vector of the current node to be aggregated in hyperbolic space.

[0032] Furthermore, the eigenvectors in Euclidean space Treating it as a vector in the tangent space, we project it onto the hyperbolic space using an exponential mapping function: (10) in, Represents a node Eigenvectors in Euclidean tangent space; This represents an exponential mapping function with the origin as its base point; Represents the Euclidean norm; This represents the curvature parameter.

[0033] Through the above mapping, the features that were originally linearly distributed in Euclidean space are transformed into a structure that expands exponentially in hyperbolic space, thus providing a geometric basis for subsequent hierarchical relationship modeling.

[0034] S105, Adaptively adjust the edge curvature of the hyperbolic geometric space according to the hierarchical depth of the project organizational structure; In this embodiment, to enable the hyperbolic space to accurately characterize the structural differences between entities at different levels in an engineering project, hierarchical information from the project's organizational structure is incorporated into the curvature adjustment process. Specifically, each project entity node is mapped to a hierarchical depth variable within the organizational structure. ,For example: Group Headquarters Node: ; Regional / Provincial Company Nodes: ; Subsidiary node: ; Project milestones: ; Construction team milestones: ; Set the basic curvature constant Curvature attenuation coefficient It is used to control the rate at which curvature changes with the depth of the hierarchy.

[0035] The edge curvature of the hyperbolic geometric space is adaptively adjusted based on the hierarchical depth of the project's organizational structure, using the following formula: (11) in, Indicates the first The dynamic edge curvature of the local subgraph where each entity node is located.

[0036] In this embodiment, the curvature adaptation mechanism described above allows nodes at deeper levels (such as construction teams) to correspond to larger absolute curvature values, thereby achieving greater representation capacity in hyperbolic space; while higher-level nodes maintain lower curvature, which is beneficial for global information aggregation. This method avoids the excessive smoothing problem that occurs in deep tree structures during feature propagation from a geometrical perspective, thereby improving the representation capability under complex organizational structures.

[0037] like Figure 3 As shown in the figure, this embodiment of the invention provides a simulation diagram of the three-dimensional mapping distribution of project entity features based on a Poincaré sphere hyperbolic manifold. The scattered points inside the three-dimensional sphere represent business entity nodes at different levels, and the connecting lines represent the causal flow relationships of the business. From the center of the three-dimensional sphere (representing the group headquarters), The edge curvature approaches the Euclidean flat space) to the edge of the spherical shell (representing the basic construction team and the bottom document). (Exhibited by extremely high negative curvature), the exponential mapping distribution of nodes exhibits obvious hierarchical expansion characteristics. Simulation results show that the curvature adaptive algorithm proposed in this invention can effectively expand the embedding space capacity of edge leaf nodes, thereby achieving high-precision feature representation of deep long-tailed entity nodes and ensuring effective decoupling of entity features at different levels in geometric space.

[0038] S106, The project entity features are aggregated using a curvature-adaptive hyperbolic graph neural network to construct the fused knowledge graph; In this embodiment, after completing hyperbolic space embedding and curvature adaptive adjustment, a hyperbolic graph neural network is constructed to model and aggregate the relationships between project entities. The graph structure is determined by project business relationships, including contractual relationships, approval workflows, fund payment paths, and dependencies between equipment and construction tasks.

[0039] Specifically, a hyperbolic graph convolutional layer with 2 to 3 layers is constructed to propagate node features layer by layer. For the first... Layer nodes The feature representation and its update process are as follows: (12) in, For nodes In the The hyperbolic space feature vector of the layer (i.e., the updated layer); For nodes neighboring nodes In the Hyperbolic space feature vectors of the layer; Represents a node The set of neighboring nodes; Indicates from node To the node The normalized weights (which can be obtained from the attention-based weight calculation module); This represents Möbius scalar multiplication; It represents the Möbius strip.

[0040] The Möbius scalar multiplication is defined as follows: (13) in, For scalar weights, and These represent the exponential and logarithmic mappings with the origin as the base point, respectively. Indicates curvature as Möbius scalar multiplication in hyperbolic space.

[0041] The Möbius method is defined as: (14) in, The curvature parameter of the local subgraph where the node is located (determined by step S105); Represents the Euclidean inner product; Describing the vector norm, Let be the feature vector of the neighboring node in hyperbolic space.

[0042] After propagation through multi-layer hyperbolic graph convolution, a node representation matrix that integrates structural relationships and business semantics is obtained: (15) in, This represents the total number of nodes in the graph; This indicates the output embedding dimension.

[0043] Finally, represented by nodes. As attributes of the knowledge graph nodes, and combined with the business relationships between nodes (such as approval relationships, contract dependencies, and payment flow relationships) as edges, a fused knowledge graph is constructed. This graph can uniformly express the entity relationships and behavioral paths throughout the entire project process, providing structured semantic support for subsequent anomaly detection and control reasoning.

[0044] S107 utilizes a linear four-rate detector to monitor the classification boundary performance index of the anomaly detection model in real time. In this embodiment, a real-time performance monitoring mechanism based on a sliding window is constructed for the continuous business data flow (including contract approval flow, fund payment request flow, and on-site construction event flow) generated throughout the entire project process. The sliding window size is set as follows: That is, each cumulative Each business event forms a statistical batch.

[0045] Furthermore, within each sliding window, an anomaly detection model is used to classify and predict business events, and a confusion matrix is ​​constructed by combining the results of post-event manual review or rule verification. (16) Based on this, the linear four-rate index is calculated, including: Recall rate for minority categories (violations): ; Minority class precision: ; Recall rate for the majority category (normal business): ; Majority class precision: ; in: Indicates the number of violations that were correctly identified; This indicates the number of normal events that were mistakenly judged as violations. This indicates the number of violations that were mistakenly classified as normal. This indicates the number of normal events that were correctly identified.

[0046] Furthermore, confidence intervals based on the binomial distribution are constructed for the above four indicators. At a 95% confidence level, the range of these intervals is... for: (17) in, This represents the observed value of the corresponding indicator.

[0047] The above method enables real-time quantitative monitoring of the stability of the classification boundary of the anomaly detection model.

[0048] S108, when the classification boundary performance index is detected to deviate from the preset statistical threshold and concept drift occurs, online incremental learning is triggered; Furthermore, for the time series of the four rate indicators output in step S107, a concept drift determination mechanism based on statistical significance is constructed. A concept drift determination mechanism is established when any indicator meets the following deviation condition: , Then it is determined that the data distribution in which the model is currently located has changed significantly, where, This indicates the value of the indicator currently displayed in the sliding window; This represents the mean of the indicator corresponding to a historically stable period. This represents the allowable fluctuation range calculated based on the confidence interval.

[0049] When the above conditions are met, the control module generates a concept drift interrupt signal: , It also triggers an online incremental learning process to adaptively update the parameters of the anomaly detection model. In engineering business scenarios, this mechanism can identify changes in distribution, such as new patterns of bid rigging and collusion, abnormal changes in funding paths, or abnormal falsification of construction progress, thereby responding promptly to environmental changes.

[0050] In this embodiment, by introducing a drift detection mechanism based on statistical confidence intervals, the problem that traditional models cannot perceive dynamic changes in business distribution is solved, enabling timely identification and response to new risk patterns.

[0051] S109, during the online incremental learning process, the feature gradient space of historical core risk samples is extracted, and the loss gradient of the new samples is projected into the orthogonal complement space of the feature gradient space to update the parameters of the anomaly detection model and generate the abnormal behavior warning sequence. In this embodiment, the pre-stored The historical gradient matrix is ​​obtained by inputting core historical violation samples into the anomaly detection model and taking its derivative. ,in This represents the total number of model parameters. For a new sample, calculate its loss function. parameter gradient The orthogonal projection matrix is ​​solved using the Gram-Schmidt orthogonalization method. The gradient of the new sample is projected onto the orthogonal complement space of the historical feature gradient space to update the anomaly detection model parameters. The update formula is as follows: (18) in, and These represent the model parameters before and after the update, respectively. For the learning rate, we can choose... , It is an identity matrix.

[0052] After the update is complete, the model generates an abnormal behavior warning sequence: , Each of them Includes timestamp eigenvectors Exception type tags and event source identifier T represents the total number of events.

[0053] In this embodiment, an orthogonal gradient projection update mechanism is used to ensure the stability of the weights for judging historical core risks (such as bid rigging and collusion) while the model learns new anomaly features in a dynamic environment, and orthogonal constraints are utilized to ensure the stability of the weights for judging historical core risks (such as bid rigging and collusion). This method achieves smooth incremental updates of the risk control model and effectively solves the catastrophic forgetting problem in the continuous learning process while ensuring the model's adaptability.

[0054] S2, using a multi-agent cooperative system to generate control instructions for the abnormal behavior warning sequence.

[0055] In this embodiment, the step of generating control instructions for the abnormal behavior warning sequence using a multi-agent cooperative system includes: S201, invokes a pre-deployed large language model to build a multi-agent collaborative system; In this embodiment, the multi-agent collaborative system is built upon a finely tuned large-scale language model with hundreds of billions of parameters. It is divided into several heterogeneous agents according to business functions and engineering management needs, forming a collaborative matrix. The financial agent has access to ERP billing interfaces, fund payment flow queries, and budget comparison analysis. Its System Prompt boundary is "only accessing financial information, identifying and reporting budget anomalies, overpayments, and payment rhythm anomalies." The legal agent has access to the enterprise compliance and internal control matrix knowledge base, accessing contract terms, legal provisions, and internal process control rules through the RAG (Retrieval-Augmented Generation) mechanism. Its System Prompt boundary is "responsible for verifying the compliance of business operations, not accessing financial flow or construction progress data." The engineering agent can call the construction progress system and IoT sensor data interfaces to monitor the completion status of construction tasks, on-site equipment operation status, and resource consumption. Its System Prompt boundary is "providing on-site construction progress and equipment status information, not directly handling financial or legal matters."

[0056] The enterprise compliance and internal control matrix knowledge base stores compliance entries and approval rules for each business process within the enterprise. The legal intelligence agent retrieves this knowledge base and, in conjunction with abnormal behavior warning sequences, identifies potential violations and performs risk analysis. This knowledge base can be accessed through standard databases or document retrieval systems, and search results are output in Boolean value or risk level matrix format.

[0057] The multi-agent system architecture employs a shared message pool and a cyclic inference mechanism to achieve communication and information aggregation. The shared message pool stores intermediate inference results and event evidence from each agent. When each agent receives a new anomalous event input or a message from another agent, it performs self-attention encoding to generate a context vector and writes the result to the message pool. Multiple iterations are performed until convergence, ensuring that each agent fully discusses and fuses evidence regarding the anomalous event.

[0058] S202, the abnormal behavior warning sequence is used as an input event to drive each intelligent agent to perform cross-modal evidence verification and multi-round consensus reasoning to generate control instructions; In this embodiment, the abnormal behavior warning sequence generated in step S109 Structured into a vectorized event matrix containing event timestamps, numerical features, category labels or anomaly type codes, and event source identifiers. The structuring process can be expressed by the following formula: , in This is an event vector, with each row corresponding to a matrix. a line of The dimension of the event feature vector. Representing a mapping relationship means mapping elements in the original input space. After feature extraction, the vector space is transformed into the target vector space. In the middle. Then put all of them. The events are stacked in chronological order: , The vectorized sequence is input into the multi-agent collaborative system to initiate a cross-modal evidence verification process. The engineering agent retrieves data from the construction progress system and IoT sensor data to generate matrices representing construction completion rate, equipment operating status, and resource consumption. The financial intelligence agent queries the ERP billing interface and generates a financial transaction matrix. Calculate budget deviations, overpayment percentages, and payment schedule indicators. The legal intelligence agent retrieves the internal control matrix and contract terms, and outputs a compliance verification matrix. Mark potential points of violation or operational risk.

[0059] In each round of consensus reasoning iteration, the agents exchange information through a shared message pool. In each iteration, each agent outputs its own data. The data is written into the pool and then weighted and fused with data from other agents. The fusion formula is as follows: (19) in, Indicates the first Wheel Intelligent Agent Context feature vectors, Indicates the first Wheel Intelligent Agent Context feature vectors, Indicates the first Wheel Intelligent Agent Context feature vectors, This represents the trust weight between agents, which can be dynamically adjusted based on historical inference accuracy. For self-updating weight matrix, Let be a non-linear activation function, and i and j be the agent indices. Iteration starts from... Start, until the convergence condition is met or the preset number of rounds is exceeded. The convergence condition is defined as: , in Denotes the Frobenius norm. For the preset convergence threshold (e.g.) This is used to quantify that the change in the output of multi-round consensus inference is small enough to indicate that the fused information of each agent on the abnormal event has stabilized.

[0060] In this embodiment, a decoding mechanism based on a combination of multilayer perceptron (MLP) and a large language model is constructed to address the problem of decoding the continuous feature matrix after consensus reasoning convergence into discrete structured text.

[0061] Specifically, when multiple iterations converge (or reach the preset number of iterations) When the context feature vectors of each agent tend to be consistent, the output features of all agents in the last round are aggregated (e.g., by global average pooling) to obtain the global decision vector. .

[0062] First, discrete action and target classification and decoding are performed: an action classifier and target mapping network containing a Softmax activation function are constructed, and the global decision vector is... Project onto the preset instruction tag space: (20) (twenty one) in, This represents the probability distribution vector of control actions. A vector representing the probability distribution of the object being operated on; The projected weight matrix is ​​a learnable matrix. This is the bias term. By taking the above probability distribution... For each operation, the category label with the highest probability is selected as the control action and the target, respectively.

[0063] Secondly, natural language generation and decoding of the operational basis are performed: To generate action reasons that conform to business logic, a prefix-tuning mechanism is adopted to optimize the global decision vector. The data is transformed into soft prompt vectors with the same embedding dimension as the input words in the master language model through a linear mapping layer. These soft prompt vectors are then directly appended to the beginning of the original warning event text prompts and input into the master model of the multi-agent system. Guided by these continuous semantic features, the master model autoregressively generates interpretive natural language text as the basis for its operations.

[0064] Finally, the system extracts the current system time as a timestamp, and structurally assembles the discrete action labels and target labels output by the classifier with the text explanation generated by the large model, outputting the following JSON format control instructions: { "action": "freeze_payment", "target": "project_A_account", "reason": "Construction progress is behind schedule, and payments have exceeded 80%." "timestamp": "Year-Month-DayTH:M:S", "source_agents": ["financial_agent","engineering_agent"] } Where action represents the action to be performed, target represents the object to be operated on, reason represents the basis for the operation, timestamp represents the time when the instruction was generated, and source_agents represents the list of agents participating in the decision-making process.

[0065] In this embodiment, a heterogeneous role multi-agent collaboration mechanism is introduced, which realizes the automated generation of anomaly warnings and control instructions, replacing the traditional threshold alarm mode. This enables engineering risk control to proactively respond to complex business scenarios, while ensuring that each agent processes information and makes inferences within its authority, thus ensuring the legality and controllability of the executed operations.

[0066] S3, extract the attention weight distribution of the multi-agent collaborative system during the inference process, lock the target subgraph associated with the attention weight distribution in the graph, and prune the global business compliance rules based on the target subgraph to obtain the local formal context.

[0067] In this embodiment, the step of extracting the attention weight distribution of the multi-agent collaborative system during the inference process, locking the target subgraph associated with the attention weight distribution in the graph, and pruning global business compliance rules based on the target subgraph to obtain the local formal context includes: S301, Extract the attention weight distribution matrix of the multi-agent cooperative system during the reasoning process of the large language model; Furthermore, the feature parsing module delves into the underlying large language model architecture of the multi-agent system. When the multi-agent system generates the control commands, it extracts the self-attention tensor of the final Transformer layer of the large language model. Let the length of the token sequence of the input context be... The extracted attention weight distribution matrix is ​​then expressed as: , in, This represents the self-attention weight matrix. The elements in the matrix... Representative of the generation of the first When the Token is... The system assigns attention weights to each input token. After extracting the attention weight distribution matrix, the system aggregates the weights of tokens belonging to the same business entity according to a preset business entity index table to obtain the normalized attention score of that business entity, which is then compared with a preset threshold.

[0068] like Figure 4 As shown in the figure, this embodiment of the invention provides a thermal simulation diagram of the self-attention weight distribution of a multi-agent system when generating a "forced refund of excess payment work order" instruction. The horizontal and vertical axes in the figure correspond to the input multimodal fusion context sequence Token, respectively. The color depth in the matrix (from cool dark blue to warm dark red) directly corresponds to the normalized attention score. From the densely clustered grid area of ​​bright red pixels in the figure, it can be seen that when the model performs causal reasoning, it assigns extremely high attention weight scores (all...) to the three feature semantic clusters of "progress delay exceeding 30%", "advance payment ratio exceeding limit", and "supplier C". The system extracts the set of nodes in the heatmap whose response values ​​are higher than a preset threshold, maps them as the basis for model decision-making, and uses them as the input anchor points for the extraction of local subgraphs of the hyperbolic spectrum in step S302.

[0069] S302, using entity nodes in the graph whose attention weight is greater than a preset threshold as core anchor points; In this embodiment, after obtaining the normalized attention score, the system sets an attention sensitivity threshold. (In this embodiment, the preferred setting is...) The mapping engine will award an attention score exceeding this threshold. High-weighted tokens are considered the core basis for risk reasoning in the large language model. Subsequently, through entity linking technology, these high-weighted tokens are mapped to a pre-built fusion knowledge graph to accurately locate the corresponding business entity nodes, thereby constructing a core anchor set: , in, Represents the set of core anchor points. Indicates the first in the atlas Each business entity node This represents the normalized attention score of the token corresponding to that node. For example, "Substation A Project Department" and "Supplier B," which the model highly focuses on, will be directly locked as the starting anchor points for the search.

[0070] S303, extract the neighbor nodes of the core anchor point in the graph according to the preset topological hop count, and extract the logical association edges between the core anchor point and the neighbor nodes to form the target subgraph; Furthermore, in order to obtain the complete business execution chain without introducing globally redundant data, the graph traversal engine uses... Perform local subgraph extraction starting from the current point. Set the topology hop count parameter. (In this embodiment, it is set) Using the breadth-first search (BFS) algorithm, the algorithm extracts data layer by layer from the massive fused knowledge graph that is within a certain distance from the core anchor point. Collect all neighboring nodes within the jump range and simultaneously extract the objectively existing business logic connections between these nodes (e.g., "payment relationship" edges indicating fund allocation, "procurement relationship" edges indicating business acceptance). Through the above graph traversal operation, the project holographic graph is cropped into a target subgraph with a reduced volume and highly focused semantics. .

[0071] S304. Based on the entity type and edge attributes of the target subgraph, the global business compliance rule base is dynamically pruned to filter out irrelevant rules and obtain a local formal context.

[0072] As a preferred implementation method, in the actual operation of large engineering enterprises, the global business compliance rule base typically contains tens of thousands of hard standards covering dimensions such as safety production, financial auditing, and material procurement. The context-based pruning engine actively reads the generated target subgraph. The system includes node labels and edge attributes. If a subgraph only involves nodes of the "finance" and "contract" categories, as well as edges of the "funds payment" category, the system will execute a dynamic filtering mechanism to automatically remove rules and logic that are irrelevant to the current control actions, such as "standards for wearing safety helmets during on-site construction" and "annual inspection requirements for special equipment." Only dozens of core logic rules that strictly match the business attributes of the subgraph will be retained. Finally, the retained compliance rules and the target subgraph structure will be encapsulated together to obtain a local formal context used for security verification.

[0073] In summary, traditional formal verification methods, when faced with complex business systems, are prone to "state space explosion" during deduction due to the need to load all state variables and global business rules, resulting in extremely low verification efficiency or even system crashes. This embodiment innovatively utilizes the attention mechanism generated by a large language model during inference to guide the process, accurately segmenting the local business network actually reached by control commands, and accordingly significantly reducing the dimensionality of the rule base. This mechanism completely eliminates the state space explosion problem inevitably caused by full verification from the source of data input, laying a lightweight contextual environment for efficient and computable verification of subsequent theorem provers.

[0074] S4, map the numerical variables in the local formal context to interval logic operators to construct a local linear temporal logic formula.

[0075] It should be noted that various existing techniques can be used to map numerical variables in the local formal context to interval logic operators and construct local linear temporal logic formulas. These include hard-coded discretization methods based on fixed empirical thresholds, equal-width or equal-frequency binning algorithms based on historical data distribution statistics, and clustering and decision tree interval mapping techniques based on uncertainty partitioning. However, when dealing with continuous numerical variables with infinite state domains in the context of engineering process control, these existing techniques often suffer from a lack of mathematical proof of monotonicity, easily leading to approximation errors at interval edges. This can result in false positives or false negatives in the subsequent formal verification engine.

[0076] In this embodiment, to maximize state space compression while ensuring mathematical rigor, it is preferable to construct locally linear temporal logic formulas using abstract interpretation theory, including: S401, Identify continuous numerical variables in the local formal context; Furthermore, the syntax parsing module performs static program analysis on the local formal context output in step S3. The system utilizes pre-compiled regular expressions or constructs an Abstract Syntax Tree (AST) to perform a depth-first traversal of the pruned core business rules and target subgraph attributes to identify continuous numerical variables within the local context. , In the context of full-process project management, these continuous numerical variables specifically correspond to real-time account balances in ERP systems, floating-point numbers representing the proportion of project material payments, or the specific amount of a single fund disbursement. Since these variables have an infinite state domain in the real environment, directly inputting them into a conventional state machine model for exhaustive verification would inevitably lead to the problem of an uncomputable state space.

[0077] S402, classify the continuous numerical variables into a preset logical safety interval to obtain discretized dimensionality-reduced state variables; replace the continuous numerical variables in the local formal context with the dimensionality-reduced state variables as the interval logic operator; As a preferred implementation, to overcome the verification explosion problem caused by continuous variables, the system introduces the Galois Connection mechanism from abstract interpretation theory. Specifically, the system constructs a monotonic mapping relationship from the concrete state space to the abstract state space, defining a relationship defined by abstract functions. With concrete functions The Galois connections formed strictly satisfy the following associated physical constraints: (twenty two) In the formula, Represents continuous numerical values ​​in a specific state space. Represents a discrete abstract interval in the abstract state space; This represents an abstract function that maps specific numerical values ​​to a finite set of discrete intervals. This represents a concrete function used to ensure order preservation and security constraints between the abstract and concrete domains.

[0078] For example, regarding the continuous variable "single payment amount" The system discretizes it into a reduced-dimensional state variable with three discrete state values: representing the daily safe operation range. (when (in ten thousand yuan), indicating the range that requires attention and review. (when (in ten thousand yuan), and indicates an extremely large, high-risk range. (when (Ten thousand yuan). Through this abstract mapping mechanism, continuous numerical variables in the local formal context are replaced with the reduced-dimensional state variables, which serve as the interval logic operators. This dimensionality reduction operation transforms the highly complex traversal and comparison of specific numerical values ​​into symbolic interval logic computation with finite states, thereby reducing the computing power and time required for verification.

[0079] S403, the control command is parsed into a sequence of atomic propositions; In this embodiment, the instruction conversion engine receives the structured JSON format control instructions generated in step S2 and performs semantic decomposition on them. Specifically, the system is pre-configured with an instruction-proposition mapping dictionary. The instruction conversion engine extracts key-value pair features from the JSON instructions, matches them using the mapping dictionary, and transforms them into a set of Boolean atomic proposition sequences that can be directly processed by the subsequent formal verification engine. , Each of them This represents an indivisible, single system state or business action. For example, based on the mapping dictionary, the business prerequisites in JSON instructions are transformed into atomic propositions. (Representing "triggering a payment action"), this transforms the intervention action corresponding to the control strategy (such as dictionary matching to "action": "freeze_payment") into an atomic proposition. (This represents "execute the freeze operation on the corresponding account"). This step refines macro-level control operations into true / false propositions that can be recognized by machines.

[0080] S404, combining the interval logic operator and the local formal context, the atomic proposition sequence is encapsulated using the temporal logic operator to generate a logical expression that conforms to the satisfiability module theory standard as the local linear temporal logic formula.

[0081] Furthermore, to rigorously express the causal and constraint relationships of project flow actions along the timeline at a mathematical level, the timing logic encapsulation module introduces Linear Temporal Logic (LTL) operators, such as those representing "Always". Operators and the "Next" operator Operators. Combining the dimension-reduced interval logic operators and atomic propositions generated above, the syntax is reconstructed according to the business constraint rules in the local formal context to generate a local linear temporal logic formula. Furthermore, to enable this temporal logic formula to be processed by the underlying Satisfiability Modulo Theories (SMT) solver, bounded model checking (BMC) technology is introduced to test the linear temporal logic formula with time attributes at a preset time step. The state is expanded internally, thus transforming it into a normalized logical expression that conforms to the SMT verification standard and does not contain timing operators.

[0082] An example of the constructed local linear temporal logic formula structure is as follows: (twenty three) in, This represents the generated local linear sequential logic formula; A time-series operator indicating that the rule is "always valid" throughout the entire lifecycle of the project; This represents the atomic proposition that triggers the payment request; The logical AND operator; The interval logic operator generated in step S402 indicates that the value is in a high-risk interval; This is a logical implication symbol, indicating that if the precondition is true, the subsequent action will necessarily occur. For representing the timing operator of "the next state immediately following"; This refers to an atomic proposition that enforces the freezing of an account.

[0083] The example formula represents the following engineering compliance logic: "Throughout the project lifecycle, there is always the following rigid constraint: if a payment request is triggered ( And the disbursement amount is in a high-risk range. ), then the next step ( The system must enforce the fund freeze order. After the LTL formula is converted into a pure proposition and interval arithmetic structure through the above bounded expansion encoding, it serves as the input condition for the system to call the theorem prover (such as the Z3 solver) for rigorous verification in the next step.

[0084] In summary, while large language models, such as multi-agent collaborative systems, generate control policies with excellent text and business understanding capabilities, they often lack absolute execution boundaries due to the inherent randomness of natural language processing. This can easily lead to violations caused by "AI illusions" in high-risk scenarios such as finance and engineering. This step, through dimensionality reduction and rank reduction of abstract interpretation and the formulaic reshaping of temporal logic, successfully and losslessly transforms the lack of rigorous boundaries in natural language AI decision-making into formalized machine language that can be deduced through rigorous mathematical theorems, providing a physical foundation for the safe and compliant execution of the system.

[0085] S5, perform security verification on the local linear sequential logic formula, and execute the control instruction when the verification passes.

[0086] It should be noted that various existing technologies can be used to implement security verification of the aforementioned local linear temporal logic formulas. These include standard model detection techniques based on explicit state or symbolic automata (such as Büchi automaton transformations), Boolean logic matching methods based on static rule configuration engines, or runtime verification and monitoring techniques based on asynchronous comparison of dynamic execution trajectories. However, these existing technologies have significant limitations when facing industrial-grade full-process control scenarios: traditional automaton model detection is prone to "state space explosion" anomalies due to the geometrical growth of state combinations when facing multi-source concurrent business state flows, leading to calculation failures; rule configuration engines lack the ability to deeply deduce temporal dependencies and dynamic state transitions (such as the temporal correlation between pre-approval and post-settlement); and runtime verification mainly relies on in-process or post-process triggering, making it impossible to achieve 100% certain pre-interception before irreversible physical control instructions are actually issued and reach the underlying business system interfaces.

[0087] In this embodiment, to overcome the above-mentioned defects, this application preferably utilizes a formal theorem prover for static symbolic reasoning. Specifically, the formal theorem prover employs a satisfiability modular theory solver to perform a full ergodic solution for satisfiability. The specific steps include: S501, The local linear time-series logic formula is input into a preset theorem prover to perform model detection and satisfiability solving; Furthermore, the security verification module invokes a pre-integrated satisfiability modular theory solver (SMTSolver, preferably the Z3 theorem prover in this embodiment) as the underlying execution component for compliance verification. The verification engine first constructs a set of verification constraints for theorem proof. Specifically, the system extracts the action field and target entity field from the structured JSON instruction output in step S2, and transforms them into mathematical update logic for the target entity's state attributes according to the preset underlying business state machine definition, thereby instantiating the state transition function corresponding to the control instruction. For example, if the extracted action field is "freeze_payment" and the target entity is "project_A_account" (project A account), the system queries a preset state machine dictionary to convert it into a specific logical assignment formula: .in, This represents the Boolean state variable of the target account at the next time step, with 0 indicating a dead state. Through this mapping, natural language text is losslessly transformed into symbolic logic recognizable by the solver.

[0088] Let the local state set of the current business system be . The state transition function corresponding to the control command is The normalized logical expression representing the compliance rigid constraint generated in step S4 is as follows: The validation engine constructs a target negation formula. This assumes that "executing this control instruction under the current state would result in a breach of compliance rules." Subsequently, the objective is negated using a formula. The input to the Z3 theorem prover serves as the verification target, enabling satisfiability verification. Thanks to the dimensionality reduction achieved in the preceding steps using attention-based subgraph slicing and abstract interpretation theory, the system effectively eliminates the massive amount of irrelevant rules, successfully removing the "state space explosion" obstacle faced by traditional formal verification. Therefore, the theorem prover can complete a full traversal of the local state space in a very short time, checking whether the control instructions to be executed will lead to business flow deadlocks (such as looping waits between financial and legal states), illegal fund outflows (such as direct payments bypassing hierarchical approval), or underlying data read / write conflicts between heterogeneous business systems.

[0089] S502, when the theorem prover returns an "Unsatisfiable (UNSAT)" signal, the security verification is deemed successful, and the control instruction is executed through the underlying interface; if it returns "Satisfiable (SAT)", the instruction is intercepted and an error path is reported.

[0090] In practice, the control execution engine performs differentiated actions based on the Boolean result output by the theorem prover: when the theorem prover returns an "Unsatisfiable" signal after derivation, it means that the solver cannot find any rule that violates the compliance rules within the given state space. Upon determining the path, the system verifies that the instruction has passed security verification at both the mathematical logic and enterprise regulatory levels. At this point, the instruction issuance module immediately calls the API write interface of the underlying financial ERP system or electronic bidding system (ECP) to transform the abstract control actions generated based on the graph into physical execution instructions at the business level and issues them directly. Depending on the risk level, "soft isolation" (such as restricting certain suppliers' reporting permissions in the system without affecting historical settlements) or "hard blocking" (such as forcibly returning current high-risk payment orders and completely locking related fund accounts at the gateway level) can be implemented.

[0091] Conversely, if the theorem prover returns a "Satisfiable" signal, it means that the solver has successfully found a boundary condition that violates the compliance rules when the instruction is executed. Upon encountering the logical path, the system firewall will immediately intercept the instruction in the sandbox environment, preventing it from penetrating downstream business systems. Simultaneously, the theorem prover generates a satisfying model based on the SAT result. By analyzing this model, it extracts the specific variable assignment sequences that lead to the violation as counterexamples, forming a detailed error proof path. This path is then fed back to the multi-agent collaborative system hub in step S2, forcibly triggering the agents' error reflection mechanism and prompting each agent to reconstruct the instruction based on this rigid error-correcting evidence.

[0092] In summary, in modern complex engineering project management, directly allowing large language models to generate and issue business operation instructions poses significant systemic risks. The innovation of this approach lies in using a rigorous formal theorem proof mechanism as the deterministic verification boundary for instruction issuance. This effectively overcomes the system-level risks arising from the inherent non-deterministic output (i.e., the "illusion" phenomenon) of large language models directly calling underlying business system interfaces. This mechanism ensures that intelligent control instructions achieve "second-level penetration" from macro-level decision-making to the system's underlying layer, while simultaneously building a zero-fault-tolerance security defense at the deepest level of the algorithm, truly meeting the compliance requirements of large engineering enterprises for the secure and controllable core business.

[0093] Example 2, Figure 5 A project end-to-end transparent management and control system based on multi-source data fusion is presented, including: The instruction generation module is used to acquire a multi-source heterogeneous data fusion knowledge graph and abnormal behavior early warning sequence constructed from multi-source heterogeneous data of the project; and to generate control instructions for the early warning sequence using a multi-agent collaborative system. The graph pruning module is used to extract the attention weight distribution of the multi-agent collaborative system during the inference process, lock the target subgraph associated with the attention weight distribution in the graph, and prune the global business compliance rules based on the target subgraph to obtain the local formal context. The logic mapping module is used to map numerical variables in the local formal context to interval logic operators and construct local linear temporal logic formulas. The instruction execution module is used to perform security verification on the local linear sequential logic formula and execute the control instruction when the verification is successful.

[0094] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0095] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0096] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0097] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0098] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0099] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for comprehensive project management and control based on multi-source data fusion, characterized in that, include: The project acquires a fusion knowledge graph and abnormal behavior warning sequence constructed from multi-source heterogeneous data. The abnormal behavior warning sequence includes timestamps, feature vectors, abnormal type labels, and event source identifiers. The abnormal behavior warning sequence is used as an input event to drive the agents of the multi-agent collaborative system to perform cross-modal evidence verification and multi-round consensus reasoning to generate control instructions, which include control actions and operation objects. Construct a local linear sequential logic formula associated with the control command, perform security verification on the local linear sequential logic formula, and execute the control command when the verification passes. The local linear temporal logic formula is constructed as follows: Extract the attention weight distribution matrix of the multi-agent cooperative system during the reasoning process of the large language model; The entity nodes in the fused knowledge graph whose attention weight is greater than a preset threshold are used as core anchor points; According to the preset topology hop count, extract the neighbor nodes of the core anchor point in the fused knowledge graph, and extract the logical association edges between the core anchor point and the neighbor nodes to form the target subgraph; Based on the entity type and edge attributes of the target subgraph, the global business compliance rule base is dynamically pruned to filter out irrelevant rules and obtain a local formal context. Identify continuous numerical variables in the local formal context; Continuous numerical variables are categorized into preset logical safety intervals to obtain discretized, dimension-reduced state variables; The continuous numerical variables in the local formal context are replaced with the reduced-dimensional state variables as interval logic operators; The control instructions are parsed into a sequence of atomic propositions; By combining the interval logic operator and the local formal context, the atomic proposition sequence is encapsulated using the temporal logic operator to generate a logical expression that conforms to the satisfiability module theory standard as the local linear temporal logic formula.

2. The method according to claim 1, characterized in that, The construction process of the fused knowledge graph includes: Map IoT sensing data and discrete business data from the multi-source heterogeneous data of the project to the topological space to construct a simple complex. The continuous homology features of the simple complex at different filtering scales are calculated to obtain the topological invariant matrix; Based on the topological invariant matrix, the IoT sensing data and the discrete business data are feature-aligned to construct the fused knowledge graph.

3. The method according to claim 2, characterized in that, The construction process of the fused knowledge graph also includes: Map the features of the project entities to be aligned to the Poincaré sphere hyperbolic geometry space; The edge curvature of the hyperbolic geometric space is adaptively adjusted according to the hierarchical depth of the project's organizational structure. The project entity features are aggregated using a curvature-adaptive hyperbolic graph neural network.

4. The method according to claim 1, characterized in that, Before obtaining the control instructions generated by the multi-agent cooperative system, the system also includes generating an abnormal behavior warning sequence, specifically: The classification boundary performance index of the anomaly detection model is monitored in real time using a linear four-rate detector; When the classification boundary performance index deviates from the preset statistical threshold and concept drift occurs, online incremental learning is triggered; During the online incremental learning process, the feature gradient space of historical core risk samples is extracted, and the loss gradient of the new samples is projected into the orthogonal complement space of the feature gradient space to update the parameters of the anomaly detection model.

5. The method according to claim 4, characterized in that, The formula used to project the loss gradient of the new sample into the orthogonal complement space of the feature gradient space to update the parameters of the anomaly detection model is as follows: in, and These represent the anomaly detection model parameters before and after the update, respectively. For learning rate, It is the identity matrix. The gradient matrix corresponding to the feature gradient space. The loss gradient for the new sample, It is the orthogonal complementary projection operator matrix.

6. The method according to claim 3, characterized in that, The formula used to adaptively adjust the edge curvature of the hyperbolic geometry is: in, Indicates the first The dynamic edge curvature of the local subgraph where each entity node is located. Based on the fundamental curvature constant, The curvature attenuation coefficient, Indicates the first The hierarchical depth of each entity node in the project's organizational structure.

7. A system using the project full-process penetration control method based on multi-source data fusion as described in any one of claims 1 to 6, characterized in that, include: The instruction generation module is used to acquire a fusion knowledge graph and anomaly behavior warning sequence constructed from multi-source heterogeneous data of the project. The anomaly behavior warning sequence includes a timestamp, feature vector, anomaly type label, and event source identifier. The anomaly behavior warning sequence is used as an input event to drive the agents of the multi-agent collaborative system to perform cross-modal evidence verification and multi-round consensus reasoning to generate control instructions. The control instructions include control actions and operation objects. The graph trimming module is used to extract the attention weight distribution matrix of the multi-agent collaborative system in the large language model reasoning process; and uses entity nodes in the fused knowledge graph whose attention weight is greater than a preset threshold as core anchor points. According to the preset topology hop count, extract the neighbor nodes of the core anchor point in the fused knowledge graph, and extract the logical association edges between the core anchor point and the neighbor nodes to form the target subgraph; Based on the entity type and edge attributes of the target subgraph, the global business compliance rule base is dynamically pruned to filter out irrelevant rules and obtain a local formal context. The logic mapping module is used to identify continuous numerical variables in the local formal context; classify the continuous numerical variables into a preset logical safety interval to obtain discretized dimensionality-reduced state variables; replace the continuous numerical variables in the local formal context with the dimensionality-reduced state variables as the interval logic operator; parse the control instructions into a sequence of atomic propositions; combine the interval logic operator and the local formal context, encapsulate the sequence of atomic propositions using a temporal logic operator, generate a logical expression that conforms to the satisfiability module theory standard, and construct the local linear temporal logic formula. The instruction execution module is used to perform security verification on the local linear sequential logic formula and execute the control instruction when the verification is successful.

Images