A trusted artificial intelligence audit control method and system based on asymmetric double execution paths

Abstract

The application provides a kind of asymmetric double execution path-based trusted artificial intelligence audit control method and system, it is related to artificial intelligence control technical field, the method includes: receiving input data containing equipment state parameters, standardization processing, noise filtering and data fragmentation processing are carried out to input data, obtain original feature vector;According to original feature vector, feature extraction and semantic embedding calculation are carried out, and semantic vector is obtained;According to semantic vector, obtain the dynamic logic atlas maintained by equipment operation data;The semantic distance of semantic vector and the existing node in node set is calculated, and according to the semantic distance and the preset double threshold mechanism, node fusion, relationship establishment and new isolated node determination are carried out, and the updated dynamic logic atlas is obtained.The application improves the credibility and security of system output.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence control technology, and in particular to a trusted artificial intelligence audit control method and system based on asymmetric dual execution paths. Background Technology

[0002] With the rapid development of large-scale generative artificial intelligence technology, generative AI systems, represented by large language models, have been widely applied in key areas such as financial risk control, legal document generation, industrial equipment control, and government decision support. However, existing generative AI systems have inherent defects in terms of reliability, security, and controllability. Existing generative AI systems adopt an end-to-end single-path output mechanism, and the output results depend entirely on the probability distribution within the model. Due to the black-box nature of deep neural networks, the system cannot provide a traceable chain of evidence, making it difficult for users to determine whether the output conclusions are based on true and complete information. This poses a significant security risk in high-risk decision-making scenarios. In multi-round reasoning processes, existing models lack a systematic consistency verification mechanism for the generated content. The generated content may contradict the previous reasoning premises or factual statements, and the model itself is unable to detect and correct this, leading to the continuous accumulation of erroneous information and ultimately producing unreliable output results.

[0003] Current technologies for assessing the quality of generative AI output typically employ post-processing verification or manual review, which are essentially ex-post checks and fail to establish an independent, synchronously running audit trail during the generation process. The lack of structural separation between the auditing and generation processes means that audit results cannot enforce output constraints, the auditing process itself can be bypassed, and it's difficult to generate complete audit logs to meet compliance requirements. Existing auxiliary verification models usually operate as supplementary modules, with their output serving only as suggestions or reference scores, lacking mandatory output control. Even if the verification module detects serious problems, problematic generated results may still be directly output to the user. The verification mechanism cannot provide execution-level mandatory constraints on the generation process. At the security control level, existing systems rely entirely on software-level logical judgments for output control, lacking physical-level mandatory cutoff capabilities. When the system encounters adversarial attacks or the software control logic is tampered with, the software control mechanism may fail, leading to untrusted results being output. Existing systems lack hardware-level mandatory security control capabilities, making it difficult to provide users with ultimate security guarantees at the physical level. Summary of the Invention

[0004] This invention provides a trusted artificial intelligence audit control method and system based on asymmetric dual execution paths, which improves the credibility and security of the system output.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a trusted artificial intelligence audit control method based on asymmetric dual execution paths, the method comprising: Step 1: Receive input data containing device status parameters, perform standardization, noise filtering, and data segmentation on the input data to obtain the original feature vector; based on the original feature vector, perform feature extraction and semantic embedding calculation to obtain the semantic vector; Step 2: Based on the semantic vector, obtain the dynamic logical graph of device operation data maintenance; calculate the semantic distance between the semantic vector and the existing nodes in the node set, and perform node fusion, relationship establishment and new isolated node determination according to the semantic distance and the preset dual threshold mechanism to obtain the updated dynamic logical graph; Step 3: Based on the updated dynamic logic graph, perform semantic expansion and multi-path reasoning generation in the generation execution path to obtain a candidate output set; based on the candidate output set and the updated dynamic logic graph, perform integrity detection, consistency detection and logical conflict identification in parallel in the audit execution path to obtain the audit result; Step 4: Based on the confidence parameter corresponding to the audit result, the path priority control module performs path switching judgment. When the confidence parameter is lower than the preset threshold, the output permission to generate the execution path is closed, and a path control instruction is obtained. Step 5: Based on the path control instructions and audit results, the output control module performs output filtering and output locking to obtain the final audit pass result.

[0006] Secondly, a trusted artificial intelligence audit control system based on asymmetric dual execution paths includes: The acquisition module receives input data containing device status parameters, performs standardization, noise filtering, and data segmentation on the input data to obtain the original feature vector; based on the original feature vector, it performs feature extraction and semantic embedding calculation to obtain the semantic vector. The fusion module is used to obtain the dynamic logical graph of device operation data maintenance based on semantic vectors; calculate the semantic distance between the semantic vectors and existing nodes in the node set; and perform node fusion, relationship establishment and new isolated node determination based on semantic distance and preset dual threshold mechanism to obtain the updated dynamic logical graph. The detection module is used to perform semantic expansion and multi-path reasoning generation in the generation execution path based on the updated dynamic logic graph to obtain a candidate output set; based on the candidate output set and the updated dynamic logic graph, it performs integrity detection, consistency detection and logical conflict identification in parallel in the audit execution path to obtain the audit result. The shutdown module is used to determine the path switching based on the confidence parameter corresponding to the audit result through the path priority control module. When the confidence parameter is lower than a preset threshold, the output permission to generate the execution path is closed, and a path control instruction is obtained. The locking module is used to perform output filtering and output locking through the output control module according to the path control instructions and audit results, so as to obtain the final audit pass result.

[0007] Thirdly, a computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0008] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0009] The above-described solution of the present invention has at least the following beneficial effects: By adopting an asymmetric dual-execution path architecture, the generation and auditing functions are separated into two independent and parallel processing paths at the system structure level. The audit execution path is structurally equivalent to and runs independently of the generation execution path, fundamentally eliminating the possibility of outputting unaudited results and significantly improving the credibility of the system output. The output control module enforces that only audited results are allowed as the final system output, forming a structural-level security control barrier at the architectural level, which is mandatory and unbypassable. The path priority control module implements dynamic path switching based on confidence parameters. When the confidence level is lower than the threshold, the generation path output is automatically shut down, effectively reducing the risk of AI hallucination output. The hardware forced switching module provides physical-level security protection. When triggered by hardware, the generation path output is physically cut off and cannot be recovered by software instructions, giving users ultimate autonomous control over the AI ​​output.

[0010] The dynamic logic graph engine employs a dual-threshold, three-stage update mechanism and a node activity weight exponential decay mechanism, enabling the knowledge graph to continuously evolve and providing a consistent knowledge foundation for both paths. The audit execution path integrates three dimensions: integrity detection, consistency detection, and logic verification based on the SMT solver, forming a progressive audit system. This system offers comprehensive audit coverage, high detection accuracy, supports local deployment, does not rely on third-party closed-source APIs, ensures data privacy, and possesses complete technical sovereignty and independent controllability. Attached Figure Description

[0011] Figure 1This is a flowchart illustrating a trusted artificial intelligence audit control method based on asymmetric dual execution paths, provided by an embodiment of the present invention.

[0012] Figure 2 This is a schematic diagram of a trusted artificial intelligence audit control system based on asymmetric dual execution paths, provided by an embodiment of the present invention.

[0013] Figure 3 This is a schematic diagram of the input data preprocessing and feature generation process. Detailed Implementation

[0014] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0015] like Figure 1 As shown, an embodiment of the present invention proposes a trusted artificial intelligence audit control method based on asymmetric dual execution paths, the method comprising the following steps: Step 1: Receive input data containing device status parameters, perform standardization, noise filtering, and data segmentation on the input data to obtain the original feature vector; based on the original feature vector, perform feature extraction and semantic embedding calculation to obtain the semantic vector; Step 2: Based on the semantic vector, obtain the dynamic logical graph of device operation data maintenance; calculate the semantic distance between the semantic vector and the existing nodes in the node set, and perform node fusion, relationship establishment and new isolated node determination according to the semantic distance and the preset dual threshold mechanism to obtain the updated dynamic logical graph; Step 3: Based on the updated dynamic logic graph, perform semantic expansion and multi-path reasoning generation in the generation execution path to obtain a candidate output set; based on the candidate output set and the updated dynamic logic graph, perform integrity detection, consistency detection and logical conflict identification in parallel in the audit execution path to obtain the audit result; Step 4: Based on the confidence parameter corresponding to the audit result, the path priority control module performs path switching judgment. When the confidence parameter is lower than the preset threshold, the output permission to generate the execution path is closed, and a path control instruction is obtained. Step 5: Based on the path control instructions and audit results, the output control module performs output filtering and output locking to obtain the final audit pass result.

[0016] In this embodiment of the invention, an asymmetric dual-execution path architecture is constructed, in which the generation execution path and the audit execution path run in parallel. This achieves a structurally forced separation of generation and audit at the method level. The audit path runs independently and is not affected by the generation process. Through a dynamic path priority switching mechanism based on confidence parameters, the output permission of the generation path is automatically closed when the confidence level is lower than the threshold, effectively suppressing the risk of AI illusion output. The dynamic logic graph adopts a dual-threshold three-stage update mechanism, which enables the knowledge representation to evolve continuously and provides a consistent knowledge foundation for the generation path and the audit path, avoiding judgment bias caused by asynchronous knowledge. The audit execution path integrates three progressive dimensions: integrity detection, consistency detection, and logical conflict identification, to comprehensively verify candidate outputs and ensure the reliability and traceability of the output results. The output control module performs forced filtering and locking based on the audit results, allowing the audited content to be used as the final output, forming an unbypassable security barrier. It supports local deployment, ensures that data privacy is not leaked, and has complete technical autonomy and controllability. This provides a reliable AI audit control method for high-security scenarios such as financial risk control, industrial control, and government decision-making.

[0017] In a preferred embodiment of the present invention, step 1 above may include: Step 11: Receive input data containing text data, structured data, and control command data. Based on the input data, perform data standardization processing to unify the data format and obtain standardized data. Specifically, this includes: receiving raw data input from external systems or users to obtain input data containing text data, structured data, and control command data. Text data includes at least one of the following: query statements described in natural language, document content, and dialogue records. Structured data includes key-value pairs, tabular data, JSON format data, or XML format data. Control command data includes path selection parameters, confidence threshold settings, and audit mode identifiers. Based on the input data, perform data type identification and classification processing to obtain classified and labeled multimodal data. Data type identification adopts a feature-matching-based classification strategy: scan each item of the input data and extract data format features; mark data conforming to text encoding rules as text type; mark data conforming to structured syntax rules as structured type; and mark data conforming to command protocol format as control command type. Based on the classified and labeled multimodal data, perform data standardization processing to obtain standardized data. Data standardization processing uniformly converts heterogeneous data into a preset standard intermediate representation format. The calculation formula is as follows: ; In the formula, This represents the value of the i-th original data field after numerical conversion; This indicates the standardized value corresponding to this field; This represents the mean of all samples belonging to category c of this field; This represents the standard deviation of all samples belonging to category c of this field; This is a smoothing constant used to prevent the denominator from being zero; its typical value is 10. -18 .

[0018] Step 12: Based on the standardized data, perform noise filtering to remove invalid or outlier information, obtaining clean data; based on the clean data, perform data fragmentation to divide the data into multiple data sub-units, obtaining the original feature vector. Specifically, this includes: performing noise filtering on the standardized data to obtain clean data. The noise filtering adopts a statistical filtering method based on outlier detection, determining whether each standardized value deviates from the overall distribution center beyond a preset tolerance range. If it does, it is filtered out; if it does not, it is retained. The determination formula is: ; In the formula, This represents the noise filtering decision function; This represents the i-th standardized value; This represents the median of all standardized values. The median absolute deviation of all standardized values ​​is expressed as follows: ; The tolerance coefficient is typically 3.0, indicating that data points deviating from the median by more than three times the absolute deviation of the median are considered noise. Based on the clean data, data sharding is performed to obtain multiple data sub-units. The data sharding process uses a fixed-length sliding window strategy to divide the complete sequence of clean data into multiple equal-length sub-units. Each sub-unit contains a fixed number of continuous and non-overlapping data elements. The sharding rules are as follows: ; In the formula, This represents the k-th data sub-unit; This represents the complete dataset consisting of clean data obtained after noise filtering. Let represent the j-th data element in the clean data; L represents the preset number of data elements in each shard; k ranges from 1 to 1. ,in This represents the total number of elements in the clean dataset. Based on the data sub-units, a piecewise vectorization process is performed to obtain the original feature vectors. This piecewise vectorization process uses a pre-defined set of feature extractors to map each data sub-unit into a fixed-dimensional numerical vector. The calculation formula is as follows: ; In the formula, This represents the original feature vector corresponding to the kth data sub-unit, with dimension d; This represents the vectorization mapping function; the vectorization mapping function is used to convert the original candidate output entries from text or discrete form into continuous numerical vector representations, which are then used by the downstream auditing module for semantic calculation and similarity comparison. Its function expression is: In the formula, The input space represents the original candidate output entries, that is, the set of original text or data entries produced by the execution path. Represents a d-dimensional real vector space; This represents the q-th original candidate output entry; This represents the d-dimensional vector representation obtained after mapping; This indicates that the t-th feature extractor operates on the data sub-unit. The output scalar eigenvalues, t, range from 1 to d; the superscript T indicates the transpose of a vector or matrix.

[0019] Step 13: Based on the original feature vector, perform feature extraction and vector encoding to obtain the dimensionality-reduced feature vector; based on the dimensionality-reduced feature vector, perform semantic embedding calculation to obtain the mapped semantic vector. Specifically, this includes: based on the original feature vector, performing principal component feature extraction to obtain the dimensionality-reduced feature vector. The principal component feature extraction uses principal component analysis, which maps the high-dimensional original feature vector to a low-dimensional space through linear projection, reducing the feature dimension while preserving the data variance to the maximum extent. The dimensionality reduction formula is as follows: ; In the formula, Let p represent the dimensionality-reduced feature vector corresponding to the k-th slice, with dimension p, and satisfying the following conditions: ; The projection matrix is ​​formed by arranging the eigenvectors corresponding to the first p largest eigenvalues ​​of the original eigenvector covariance matrix column-wise. The matrix size is [size missing]. ; Let represent the mean vector of all original eigenvectors; the superscript T denotes matrix transpose. The process of solving for the projection matrix is ​​as follows: calculate the covariance matrix of all original eigenvectors, and its formula is: ; In the formula, Let the covariance matrix be of size . K represents the total number of slices. The covariance matrix is ​​decomposed into eigenvalues, and the resulting eigenvalues ​​are sorted from largest to smallest. The eigenvectors corresponding to the p largest eigenvalues ​​are then arranged column-wise to obtain the projection matrix, which is expressed as: In the formula, Let represent the q-th feature vector, corresponding to the q-th largest feature value. Based on the dimensionality-reduced feature vector, semantic embedding calculation is performed to obtain the mapped semantic vector. The semantic embedding calculation adopts a nonlinear mapping method based on a feedforward neural network to project the low-dimensional feature vector onto the semantic space. The calculation formula is as follows: ; In the formula, Let h represent the mapped semantic vector corresponding to the k-th slice, with dimension h. This represents the semantic embedding weight matrix, with a size of . ; This represents the semantic embedding bias vector, with a vector dimension of h; This represents a non-linear activation function, defined as follows: Based on the mapped semantic vector, positional encoding overlay processing is performed to obtain a semantic vector carrying positional information. Positional encoding overlay processing injects the positional information of each segment in the original sequence into the semantic vector using sine and cosine encoding, so that the semantic vector simultaneously contains both content semantics and sequence positional information. Its calculation formula is as follows: ; In the formula, This represents the semantic vector after overlaying position encoding; This represents the position encoding vector corresponding to the k-th position, and its dimension is the same as the semantic vector dimension h. This represents the 2i-th component of the position encoding vector; The position encoding vector is represented by the first... There are several components; i is the component index, ranging from 0 to... .

[0020] In this embodiment of the invention, by separating the generation path and the audit path at the structural level and granting the audit path exclusive output permissions, the direct output of unverified content is prevented from the architectural level, thereby improving the credibility and security of the system output. The confidence-driven dynamic path switching mechanism automatically blocks low-quality output when the generation quality is uncertain, effectively suppressing AI illusions. The three-dimensional progressive audit covers integrity, consistency, and logical conflict verification, achieving full-link quality control. Hardware-level physical disconnection provides an ultimate security barrier independent of software. The entire process supports local deployment, and data sovereignty and privacy are completely autonomous and controllable.

[0021] like Figure 3 As shown in the figure, this diagram illustrates the input data preprocessing and feature generation process. It visually presents the comparison and changes of the original data from the vibration sensor of the industrial pump equipment before and after the noise filtering process, and serves as a visual verification of the core effect of the data preprocessing stage.

[0022] In a preferred embodiment of the present invention, step 2 above may include: Step 21: Obtain the dynamic logic graph maintained based on historical equipment operation data. The dynamic logic graph includes a set of nodes and a set of relationships representing historical equipment entities and their associated logic. Extract existing nodes from the node set. Specifically, this includes: obtaining the dynamic logic graph maintained based on historical equipment operation data according to semantic vectors carrying location information. The dynamic logic graph is a structured representation of equipment entities and their associated logic accumulated during the historical equipment operation process, containing two parts: a set of nodes and a set of relationships. Its formal representation is as follows: ; ; In the formula, G represents the dynamic logic graph; N represents the node set, which contains M existing nodes; This represents the m-th node in the node set, where each node is a vector representation of a device entity or logical concept in the semantic space. Represents a set of relations; and These represent the nodes at both ends of the relationship; Represents a node and nodes Based on the strength of the relationship between the nodes, and according to the dynamic logic graph, all existing nodes in the node set are extracted to obtain the existing node list.

[0023] Step 22: Calculate the semantic distance between the semantic vector and existing nodes to obtain the node semantic distance value; based on the node semantic distance value and the preset dual threshold mechanism, perform a three-stage structure update judgment to obtain the graph update instruction. Specifically, this includes: calculating the semantic distance between each semantic vector and each existing node based on the existing node list and the semantic vector carrying position information, obtaining a set of node semantic distance values. The semantic distance uses cosine distance as a metric, reflecting the directional difference between the semantic vector and existing nodes in the high-dimensional semantic space. The calculation formula is as follows: ; In the formula, This represents the semantic distance between the semantic vector corresponding to the k-th partition and the m-th existing node, with a value ranging from... The smaller the value, the closer the semantics are; This represents the semantic vector carrying location information corresponding to the k-th slice; The vector representation of the m-th existing node; The update decision is made, and a graph update instruction is obtained. A dual-threshold mechanism sets two threshold values: a first threshold and a second threshold. The first threshold must be less than the second threshold. The semantic distance value space is divided into three intervals, corresponding to three different graph structure update operations. The decision formula is as follows: ; In the formula, This represents the update decision function for a three-part structure. This represents the semantic distance value between the semantic vector and the existing nodes; This represents the first threshold, typically 0.15. When the semantic distance is less than this value, the semantic vector is determined to be highly similar to the existing node. The second threshold is typically 0.45. When the semantic distance is greater than this value, the semantic vector is determined to be significantly different from all existing nodes. When the semantic distance is between the first and second thresholds, the semantic vector is determined to be related to but not completely identical to existing nodes. If the semantic distance between the semantic vector and multiple existing nodes is less than the first threshold, the node with the smallest semantic distance is selected as the fusion target. If the semantic distance between the semantic vector and multiple existing nodes falls between the two thresholds, a relationship is established with all nodes that meet the conditions.

[0024] Step 23: According to the graph update instruction, node fusion processing is performed when the node fusion condition is met, logical association establishment processing is performed when the relationship establishment condition is met, and isolated node addition processing is performed when the isolated node condition is met, resulting in an updated dynamic logical graph. Specifically, this includes: according to the graph update instruction, when the update instruction is a node fusion instruction, node fusion processing is performed to obtain the fused and updated graph. Node fusion processing integrates the information of the current semantic vector into the existing target nodes, causing the target node's vector representation to absorb new information and update, while maintaining the node's identity. The fusion formula is: ; In the formula, This represents the vector of the m-th node after fusion processing; This represents the vector of the m-th node before fusion processing; This represents the semantic vector to be fused. This represents the fusion coefficient, with a value range of [value range missing]. A typical value of 0.7 is used to control the mixing ratio of old node information and new semantic information. Based on the graph update instruction, when the update instruction is a relation establishment instruction, logical association establishment processing is performed to obtain the updated graph. The logical association establishment processing creates new relation edges between the sharded nodes corresponding to the semantic vectors and existing nodes that meet the conditions, and assigns an initial relation strength value. The relation strength calculation formula is as follows: ; In the formula, This represents the initial strength value of the newly established relationship between the k-th shard node and the m-th existing node; This represents the semantic distance between the two. This represents the second threshold. The formula ensures that the smaller the semantic distance, the stronger the initial relation, and the relation strength takes on a value within... In between, according to the graph update instructions, when the update instruction is an isolated node addition instruction, the isolated node addition process is executed to obtain the graph with expanded nodes. The isolated node addition process uses the current semantic vector directly as the initial vector representation of the new node, and adds the new node to the node set of the graph. Its initialization formula is: ; In the formula, Indicates the newly added number 1 node vector; The current semantic vector is represented by M; M represents the total number of existing nodes in the graph. After adding a node, the total number of nodes increases by 1. Newly added isolated nodes do not establish relationships with any existing nodes in the graph. As more data is input, relationships are gradually established. Based on the results of the three update processes, the updated dynamic logical graph is obtained by integrating them.

[0025] Step 24: Based on the updated dynamic logical graph, perform an exponentially decaying activity weight update process on the nodes within the graph to obtain the current graph state. Specifically, this includes: performing an activity weight update process on each node in the graph one by one based on the updated dynamic logical graph to obtain the current graph state. The activity weight update process adopts an exponentially decaying weight adjustment strategy to ensure that recently visited or updated nodes maintain a high activity weight, while the activity weight of nodes that have not been visited for a long time gradually decreases. The decay formula is as follows: ; In the formula, This represents the active weight value of the m-th node at the current time t; This indicates the time when the m-th node was last updated. The active weight value; This represents the decay rate constant, typically with a value of 0.05, and its unit is the reciprocal of time. This represents the time interval since the last update. For nodes that have undergone merging updates or relationship establishment, their activity weight is reassigned to its initial maximum value at the update time, and then decreases over time according to the decay formula: In the formula, Indicates the time when the node was updated; This represents the maximum value of the activity weight, typically 0.1. For newly added isolated nodes, the initial activity weight is directly set to the maximum value. After updating the active weights of all nodes, the current vector representation of all nodes and their corresponding active weight values, as well as the relationships between nodes and their strength values, together constitute the current graph state.

[0026] In this embodiment of the invention, the present invention achieves accurate matching and adaptive updating of semantic vectors with existing graph structures through a dual-threshold, three-stage judgment mechanism, avoiding misfusion or redundant node problems caused by a single threshold; adopts an exponential decay active weight update strategy, enabling the graph to dynamically reflect the temporal correlation of equipment operation data, with recent data having a greater influence and historical data gradually decaying; node fusion adopts a weighted average method, which smoothly absorbs new observation data while retaining historical accumulated information, ensuring the stability and continuity of the graph structure; relation strength is calculated based on semantic distance normalization, making relation weights quantifiable and comparable.

[0027] In a preferred embodiment of the present invention, step 3 above may include: Step 31: Based on the hyperbolic space graph representation, calculate the geodesic distance between the target node and associated nodes to obtain the spatial geometric distance matrix. Specifically, this includes mapping the dynamic logic graph to hyperbolic space based on the current graph state, obtaining the hyperbolic space graph representation. The hyperbolic space mapping uses the Poincaré sphere model, projecting the node vectors in Euclidean space onto a hyperbolic space with negative constant curvature to more accurately capture the hierarchical structure and power-law distribution characteristics of the data. The mapping formula is as follows: ; In the formula, Let the vector representation of the m-th node in the hyperbolic space Poincaré sphere model satisfy the following condition: That is, all nodes, after mapping, are located inside the unit ball; This represents the vector representation of the m-th node in the current graph state output in step 24 in Euclidean space. Let represent the Euclidean norm. Based on the hyperbolic space graph representation, for each pair of related nodes in the graph, calculate the hyperbolic geodesic distance between that pair of nodes in the Poincaré sphere model to obtain the set of spatial distances between all related node pairs. The calculation formula is as follows: ; In the formula, Represents a node and nodes Geodesic distance in hyperbolic space; The inverse hyperbolic cosine function is defined as follows: ; Let Euclidean norm represent the difference between the hyperbolic embedding vectors of two nodes. The spatial geometric distance matrix is ​​obtained by summing and arranging the hyperbolic geodesic distances between all pairs of associated nodes, and its expression is: ; In the formula, D represents the spatial geometric distance matrix with dimensions of , where M is the total number of graph nodes; This represents the element in the u-th row and v-th column of the matrix, recording the node. and nodes The hyperbolic geodesic distance between them; The distance between two nodes is set to infinity when there is no direct relationship between them.

[0028] Step 32: Based on the spatial geometric distance matrix, perform semantic association calculation based on the shortest geodesic path in the generated execution path to obtain semantic extension features. Specifically, this includes: determining the target node corresponding to the currently input semantic vector in the generated execution path based on the spatial geometric distance matrix, and using this target node as the starting point for semantic association. The method for determining the target node is: calculating the hyperbolic distance between the current semantic vector and the hyperbolic embedding vectors of each node in the graph, and selecting the node with the smallest distance as the target node, i.e.: ; In the formula, Indicates the target node; The hyperbolic space mapping representation of the current semantic vector is used in the same way as the node mapping formula. Let represent the set of all nodes in hyperbolic space. Based on the target node and the spatial geometric distance matrix, semantic association calculation based on the shortest path along the geodesic is performed in the generated execution path to obtain semantic extension features. The semantic association calculation takes the target node as the source point and runs a shortest path search algorithm on the hyperbolic space graph, expanding outward step by step along the geodesic distance to find neighboring nodes that are semantically associated with the target node. The formula for calculating the shortest path is: ; ; In the formula, This represents the semantic association set consisting of all nodes with the target node as the source and whose shortest path distance does not exceed the threshold L; Indicates the target node to the node The shortest path distance; This represents a connected path from the target node to the target node; Representing a path One edge on; Represents the edges in the spatial geometric distance matrix The corresponding geodesic distance; L is a preset semantic association radius threshold, which controls the range of association expansion. The hyperbolic embedding vectors of all nodes in the semantic association set are weighted and aggregated according to their shortest path distance to the target node to obtain the semantic expansion feature vector. The aggregation formula is as follows: ; ; In the formula, Represents the semantic feature vector; represents the distance decay coefficient, which controls the contribution of distant nodes to the aggregation result, with a typical value of 0.8; Z is the normalization factor, which ensures that the sum of the aggregation weights is 1.

[0029] Step 33: Based on the semantic extension features, perform multi-path reasoning generation processing on the multi-dimensional topological manifold to obtain the multi-path reasoning result. Specifically, this includes: constructing a multi-dimensional topological manifold in the generation execution path based on the semantic extension feature vector; embedding the semantic extension features into a continuous manifold space with a multi-branch topological structure. The construction method of the multi-dimensional topological manifold is as follows: using the semantic extension feature vector as the initial embedding point, simultaneously extend paths along multiple different geodesic directions on the hyperbolic space graph. Each extended path corresponds to a branch dimension on the manifold, forming a multi-dimensional topological manifold structure covering multiple semantic directions, which is represented as follows: ; ; In the formula, Represents a multidimensional topological manifold; Let represent the q-th branch curve on the manifold, corresponding to the q-th inference path; Indicates the total number of reasoning paths; This indicates that Viponkalay is kicking off; The embedding representation of semantically extended feature vectors in hyperbolic space; The vector representing the extension direction of the q-th path in hyperbolic space is determined by the direction from the target node to the q-th semantic association node. This represents an addition operation in hyperbolic space, i.e., translation along a geodesic; t is the parameter extending along the path. Let represent the upper limit of the extension length of the q-th path. Based on the multidimensional topological manifold, semantic reasoning generation processing is performed along each branch curve in the generated execution path to obtain the intermediate reasoning results corresponding to each path. The reasoning formula is as follows: ; In the formula, The vector of intermediate inference results generated by the q-th inference path represents the inference mapping function, which is defined as follows: ; Represents the inference weight matrix; Represents the inference bias vector; This represents the coordinates of the manifold point at parameter t along the q-th branch curve; This represents the path attenuation coefficient, which gradually reduces the impact at the far end of the path; a typical value is 0.3. This represents a continuous integration operation along the path. In actual calculations, a discrete sampling approximation is used, which includes all... The intermediate inference results of each path are collected and merged to obtain a multi-path inference result set, which is represented as follows: ; ; In the formula, This represents the set of results from multi-path reasoning. The inference confidence score for the q-th path is represented by a combination of the activity weights of the nodes on that path and the strength of their relationship with the target node. This represents the node's activity weight value; Represents the target node and path The relationship strength value of the m-th node.

[0030] Step 34: Based on the multi-path inference results, perform candidate output construction processing to obtain a candidate output set. This candidate output set serves only as an intermediate result, does not have final output authority, and is transmitted to the audit execution path. Specifically, this includes: based on the multi-path inference result set, sorting the intermediate inference results of each path according to their confidence scores, and selecting the top few inference results with the highest confidence scores to enter the candidate queue. The selection formula is as follows: ; ; In the formula, This represents the initial text representation vector of the q-th candidate output entry; This indicates the output decoding weight matrix; This represents the output decoding bias vector; The decoding transformation function maps the inference result vector from the semantic space to the output representation space. It aggregates all candidate output entries to obtain a candidate output set. This candidate output set serves only as an intermediate output for generating the execution path and does not have the authority to be output to users or external systems. Its entire content will be transmitted to the audit execution path for subsequent verification. The formal representation of the candidate output set is as follows: ; In the formula, This represents the set of candidate outputs; each element contains a vector representation of the candidate output entry and its corresponding confidence score, which is transmitted to the audit execution path.

[0031] In this embodiment of the invention, calculating the geodesic distance between nodes in hyperbolic space more accurately reflects the hierarchical membership and semantic association strength between device entities than Euclidean space distance, thus improving the accuracy of semantic association. A weighted aggregation method based on the shortest geodesic path is used to generate semantic extension features, allowing associated node information to naturally decay and integrate according to distance, avoiding remote noise interference. A multidimensional topological manifold is constructed and parallel reasoning is performed along multiple branch paths, overcoming the limitations of single-path reasoning and achieving multi-angle, multi-level semantic coverage, enhancing the richness and robustness of the output. The candidate output set and confidence score are paired and transmitted to the audit path, providing a quantitative reference for subsequent audit verification, making the audit process verifiable and improving the system's security and credibility.

[0032] In a preferred embodiment of the present invention, step 3 above may include: Step 35: Based on the candidate output set and the current graph state, perform a negative logic scan in the audit execution path to obtain the logic features to be detected. Specifically, this includes: extracting the semantic vector corresponding to each candidate output item in the audit execution path based on the candidate output set and the current graph state, and tracing back the inference source link of the candidate output in the hyperbolic space graph to obtain the inference source tracing path of each candidate output. The backtracking formula is as follows: ; ; In the formula, This represents the set of nodes and their contributions included in the reasoning path of the q-th candidate output; This represents the hyperbolic embedding vector of the j-th node on the source path, where the subscript q indicates the candidate output and j indicates the order position in the path; This represents the semantic contribution of the node to the candidate output. Using distance decay aggregation weights, and based on the inference tracing path, a negative logic scan is performed on adjacent node pairs along each tracing path in the audit execution path to obtain the logic features to be detected. The negative logic scan employs a reverse traversal strategy, detecting the continuity and self-consistency of the logic chain segment by segment from the end node of the tracing path towards the starting target node, extracting potential logical weaknesses, missing information segments, and conflicting node pairs along the path. The scan output formula is: ; ; In the formula, This represents the set of logical features to be detected corresponding to the q-th candidate output; A function to determine the logical connection state between adjacent nodes; and Represents the hyperbolic embedding vector of a pair of adjacent nodes on the tracing path; This represents the strength of the relationship between two corresponding nodes in the current graph state; The lower limit threshold representing the strength of the relationship, typically set to 0.3; The threshold for determining the weakness of a relationship is 0.1; NORMAL, WEAK, and BROKEN represent logically normal segments, logically weak segments, and logically broken segments, respectively. This indicates that there is no indirect connection between the two nodes. The complete set of detectable logical features is obtained by summarizing all candidate outputs, as follows: ; In the formula, This represents the set of logical features to be detected corresponding to all candidate outputs; This represents the vector of the q-th candidate output entry; This indicates the number of candidate outputs after filtering.

[0033] Step 36: Based on the logical features to be detected, perform integrity checks to verify whether the necessary information structure requirements are met, and obtain integrity check results; based on the integrity check results and the current graph state, perform consistency checks to determine whether there are node or relationship conflicts, and obtain consistency check results. Specifically, based on the set of logical features to be detected, perform integrity checks on each candidate output in the audit execution path to verify whether the candidate output meets the preset necessary information structure requirements, and obtain integrity check results. Integrity checks are determined by checking whether the traceability path of the candidate output covers the predefined set of necessary node categories. The detection formula is: ; ; In the formula, This represents the integrity detection score of the q-th candidate output, with a value range of... A higher score indicates more complete information coverage; This represents the set of node categories actually covered on the candidate output tracing path; This represents a pre-defined set of required information node categories, including at least four types: device entity, fault mode, operating parameter, and maintenance operation. This represents the semantic category to which node h belongs. Based on the integrity detection score, the integrity sub-state of each candidate output is determined, and the determination formula is as follows: ; In the formula, Represents the integrity sub-state determination function; The integrity threshold is represented, typically 0.75; COMPLETE indicates the information integrity state; INCOMPLETE indicates the information missing state. Based on the set of logical features to be detected and the current graph state, consistency detection is performed on each candidate output in the audit execution path to determine whether there are semantic conflicts or relational contradictions in the inference link of the candidate output, thus obtaining the consistency detection result. The consistency detection is achieved by calculating the item-by-item comparison of the vectors of each node on the source tracing path, and its detection formula is: ; ; In the formula, This represents the consistency detection score of the q-th candidate output, with a value ranging from 1 to 2. A higher score indicates better internal consistency; This represents the cosine similarity between the hyperbolic embedding vectors of the i-th node and the j-th node on the tracing path; The threshold for conflict determination is represented, typically -0.3. A semantic conflict is determined when the cosine similarity between two nodes is below this negative threshold. I is an indicator function. The summation term counts the number of conflicting node pairs on the path. Based on the consistency detection score, the consistency sub-state of each candidate output is determined using the following formula: ; In the formula, Represents the consistency substate determination function; The consistency threshold is represented, typically 0.80; CONSISTENT indicates an internally consistent state; CONFLICT indicates a conflicting state. Based on the integrity sub-state and the consistency sub-state, the two detection results are merged to obtain the detection result tuple for each candidate output: In the formula, Let q represent the detection result tuple of the q-th candidate output.

[0034] Step 37: Based on the consistency check results, perform logical verification to identify logical contradictions or factual conflicts, obtaining logical verification results; based on the logical verification results, perform adjudication processing to obtain audit results including pass, conflict, and missing states. Specifically, based on the check result tuples, perform logical verification on each candidate output in the audit execution path to identify deep-seated logical contradictions or factual conflicts that may still exist after integrity and consistency checks, obtaining logical verification results. The logical verification adopts a verification strategy based on global graph constraints to check whether the inference relationships contained in the candidate outputs contradict the established high-confidence relationships in the graph. The verification formula is: ; ; In the formula, This represents the logical verification score of the q-th candidate output, with a value range of... A higher score indicates better consistency with the existing knowledge system; This represents a set of reference relationships, defined by the fact that the active weights of the nodes at both ends of the current graph are not less than [a certain value]. The high-confidence relation consists of... A typical value is 0.6; , Represents the original nodes at both ends of the reference relationship. , It is its corresponding hyperbolic embedding vector; This indicates the strength of the established relationship between two nodes in a reference relationship; This represents the strength of the inference relationship between two corresponding nodes in the candidate output tracing path; Representing the set of graph relationships, the logical verification sub-state of each candidate output is determined based on the logical verification score, and the determination formula is as follows: ; In the formula, This represents the logic verification sub-state determination function; This represents the logic verification threshold, typically 0.70; PASS indicates that the logic verification passed; CONTRADICTION indicates that a logical contradiction exists. The integrity check score is used to determine the threshold. Consistency test score and logical verification score In the audit execution path, a comprehensive adjudication process is performed. The scores of the three tests for each candidate output are weighted and summed to obtain the final audit result. The adjudication process uses a tiered weighted scoring method, with each test having different weights to reflect its varying importance within the audit system. The adjudication formula is as follows: ; ; In the formula, This represents the overall audit score of the q-th candidate output; This represents the integrity weight, with a typical value of 0.25. This represents the consistency weight, typically with a value of 0.35. This represents the logical verification weight, typically 0.40. Based on the comprehensive audit score, each candidate output is categorized into one of three audit result states, using the following classification formula: ; In the formula, This represents a classification function for audit results. This represents the minimum score required to pass a given state, typically 0.80. The minimum score for rejection is typically 0.50; APPROVED indicates approval, and the corresponding candidate output is authorized to be output to the user or external system; REECTED indicates conflict, and the corresponding candidate output is marked as having a logical problem, is intercepted, and is logged in the audit log; SUPPRESSED indicates missing information, and the corresponding candidate output is marked as incomplete, is discarded, and a supplementary query process is triggered. The audit results of all candidate outputs are summarized to obtain the final audit result set, which is represented as follows: ; In the formula, This represents the final audit result set, with each entry containing candidate output items. Audit classification results and overall audit score .

[0035] In this embodiment of the invention, a negative logic scanning mechanism traces backward from the inference endpoint to the starting point, systematically checking for logical weaknesses and breaks, achieving in-depth audit coverage from the surface to the core. The three-dimensional progressive audit system confirms the integrity of the information structure through integrity testing, eliminates internal semantic conflicts through consistency testing, and resolves contradictions with the existing knowledge system through logical verification, forming a progressive quality control closed loop. The comprehensive adjudication adopts a graded weighted scoring method, assigning weights differently according to the detection dimensions, making the audit judgment scientific and reasonable. Through the APPROVED, REECTED, and SUPPRESSED three-state classification mechanism, candidate outputs are refined and processed to ensure the normal output of high-quality content and block low-quality content, achieving dual protection of generation quality and output security from the architectural level.

[0036] In a preferred embodiment of the present invention, step 4 above may include: Step 41: Obtain the confidence parameters corresponding to the audit results to obtain the confidence values ​​to be determined; compare the confidence values ​​to be determined with a preset threshold to obtain the path switching determination result. Specifically, this includes: extracting the comprehensive audit score and audit classification result corresponding to each candidate output item in the audit execution path based on the final audit result set, using these as the data source for the confidence parameters to obtain the set of confidence values ​​to be determined. The method for extracting the confidence parameters is as follows: for each item in the audit result set, combine the comprehensive audit score with the corresponding candidate output confidence score to perform a secondary fusion calculation to obtain the comprehensive confidence value. The secondary fusion calculation uses a weighted summation method to proportionally fuse the audit dimension score and the generation dimension confidence. The fusion formula is: ; In the formula, This represents the overall confidence value of the q-th candidate output, with a value ranging from zero to one. This represents the overall audit score of the q-th candidate output; This represents the confidence score of the q-th candidate output during the generation phase; This represents the fusion coefficient, used to adjust the weighting of audit scores and generated confidence levels. A typical value is 0.6, ensuring that the audit dimension's evaluation dominates the fusion result. Based on the set of confidence values ​​to be determined, the comprehensive confidence value of each candidate output is compared one by one with the preset path switching threshold in the audit execution path to obtain the path switching determination result. The preset path switching threshold includes two levels: a high confidence threshold and a low confidence threshold, dividing the confidence value range into three segments, each corresponding to a different path control strategy. The determination formula is as follows: ; In the formula, This represents the path switching determination function; This represents the overall confidence value of the q-th candidate output; This represents the high confidence threshold, typically 0.75. When the overall confidence level is not lower than this value, the generated execution path will be used to output the data directly. The `` represents the low confidence threshold, typically 0.45. When the overall confidence level falls below this value, output is completely blocked. `GEN` represents the direct output mode of the generation execution path, corresponding to the high confidence interval. `AUDIT` represents the forced output mode of the audit execution path, corresponding to the medium confidence interval. `BLOCK` represents the output blocking mode, corresponding to the low confidence interval. Based on the path switching determination results, the determination results of all candidate outputs are summarized to obtain the path switching determination result set, which is represented as follows: ; In the formula, This represents the set of path switching determination results; The vector representation of the q-th candidate output entry; This indicates the path switching determination result for this entry; This indicates the overall confidence level of the entry; This indicates the total number of candidate outputs after filtering.

[0037] Step 42: Based on the path switching determination result, when the confidence value to be determined is lower than a preset threshold, a blocking instruction to close the output permission of the generation execution path is generated, resulting in an initial path control instruction. Specifically, this includes: based on the path switching determination result set, performing generation path blocking processing on all candidate output entries in the audit execution path that are determined to be in non-generation path direct mode, generating a blocking instruction to close the output permission of the generation execution path, and obtaining the initial path control instruction. The generation logic of the blocking instruction is: traversing all entries in the path switching determination result set, when the determination result of any candidate output is not in generation path direct mode, the blocking mechanism is triggered. The triggering determination formula is: ; In the formula, M represents the masking flag bit. A value of 1 triggers the generation of a masking command, and a value of 0 does not trigger it. The quantifier indicates existence, meaning that the trigger condition is met as long as any candidate output path determination result is a non-generated path direct mode. Indicates the total number of candidate outputs; This indicates the path switching determination result of the q-th candidate output. Based on the trigger state of the masking flag, when the masking flag is set to one, a data structure for the masking instruction is constructed in the audit execution path. This data structure contains four control fields: masking target path identifier, masking effective timestamp, masking scope, and masking release condition. Its structure is as follows: ; In the formula, Z represents the initial path control instruction structure; TG represents the target to be masked as the generated execution path; The timestamp of the current time when the instruction was generated is represented; SA indicates that the scope of the masking effect is all candidate output entries; UA indicates that the masking is automatically lifted after the audit execution path completes all outputs in this round. Based on the masking flag and the initial path control instruction structure, the two are jointly encapsulated to obtain the initial path control instruction, which is represented as follows: In the formula, Indicates the initial path control command; M indicates the mask flag; Z indicates the initial path control command structure; This represents the set of path switching determination results. The three components of the initial path control instruction work together. The mask flag determines whether to perform a path switching, and the structure specifies the specific parameters for the switching.

[0038] Step 43: Based on the initial path control instruction, the output permission switch is performed through the path priority control module. This closes the output channel for generating the execution path and locks the audit execution path as the sole output channel, resulting in the final path control instruction. Specifically, this includes: based on the initial path control instruction, calling the path priority control module to parse the mask flag and control fields in the instruction, performing the closing operation of the output permission for generating the execution path, and obtaining the path permission status update result. The path priority control module internally maintains permission status tables for two output paths. Each path independently has two states: enabled and disabled. The state switching operation is represented as follows: When M=1; When M=1; in the formula, This indicates the output permission status for generating the execution path; a value of one indicates enabled, and zero indicates disabled. This indicates the output permission status of the audit execution path; a value of 1 indicates enabled, and 0 indicates disabled. This indicates a status assignment operation; M represents the mask flag. When the mask flag is triggered, the permission to generate execution paths is immediately disabled, and the permission to audit execution paths is enabled simultaneously.

[0039] Based on the path permission status update result, while the output permission of the generated execution path is closed, the path priority control module elevates the output permission of the audited execution path to the highest priority and locks it as the sole output channel. The locking mechanism ensures that during the blocking period, all candidate outputs are prohibited from being sent out through the generated execution path, and output is only allowed through the audited execution path. The routing constraint formula is as follows: ; ; In the formula, Indicates candidate output The assigned output channel route; A0 indicates a forced audited path with exclusive output; GENT indicates output via a generated path; AUDIT indicates output via an audited path; NONE indicates no output channel; This indicates a universal quantifier, meaning that all candidate outputs are subject to this constraint. and These represent the permission status of the generated path and audit path after the switch, respectively; This indicates the path switching determination result. Based on the path permission status update result and routing constraints, after the permission switch is completed, the path priority control module generates the final path control instruction. This instruction records a complete control snapshot of this switch, including the permission status after the switch, the lock expiration time, and the set of path switching determination results, for subsequent auditing, tracing, and permission recovery determination. It is represented as follows: In the formula, Indicates the final path control command; and These represent the permission status of the generated execution path and the audit execution path after the switch, respectively; This indicates the lock expiration time. After this time, the lock will be automatically released and the default path configuration will be restored. Indicates the timestamp at the moment the instruction was generated; Indicates the duration of the lock, typically the remaining duration of the current session or the current request period; This represents the set of path switching determination results.

[0040] In this embodiment of the invention, a secondary fusion calculation of audit score and generation confidence level comprehensively considers quality information from both audit and generation dimensions, making the confidence level determination more comprehensive. A two-level threshold three-stage path switching mechanism accurately diverts candidate outputs to three processing channels: direct access, mandatory audit, and blocking, avoiding a one-size-fits-all approach to control. The instant triggering of the shielding flag and the hard locking mechanism of the path priority control module ensure that low-quality content cannot be leaked. The final path control instruction records a complete control snapshot and the lock deadline, providing verifiable operation logs for audit traceability and permission restoration, enhancing the transparency and controllability of the system.

[0041] In a preferred embodiment of the present invention, step 5 above may include: Step 51: Obtain the final path control instructions and audit results to obtain the output data to be processed. Based on the output data to be processed, perform output filtering to filter out the candidate output set generated by the execution path, obtaining the filtered output data. Specifically, this includes: extracting the execution path permission status field and the audit execution path permission status field from the final path control instructions; and extracting the audit classification result and comprehensive audit score of each candidate output item from the final audit result set. Merge the path control instructions and the audit result set to obtain the output data to be processed, which is represented as follows: In the formula, This represents the set of output data to be processed. Indicates the final path control instruction; This represents the final set of audit results; This indicates a data merging operation. Based on the set of output data to be processed, the output filtering module parses the generation execution path permission status carried by the final path control instruction. When this permission status is disabled, the output filtering module marks all candidate output entries determined to be in the generation path pass-through mode as objects to be blocked and removes them from the output range, obtaining the filtered output data. The filtering formula is: In the formula, This represents the filtered set of data to be output. Represents the q-th candidate output item; A represents the final audit result set; This indicates the path switching determination result; AUDIT indicates the forced output mode of the audit execution path.

[0042] The filtering criteria limit entries to those whose path determination results indicate an audit execution path pattern; all other entries are blocked. Based on the filtered output data set, compliance checks are performed on each entry within the set. This check uses both the audit classification result and the overall audit score as constraints. The audit classification result must be "passed," and the overall audit score must not be lower than a preset release threshold. Only entries meeting both conditions can proceed to the next processing stage. The verification formula is as follows: ; In the formula, This represents the set of data to be output after verification and compliance. This represents the filtered set of data to be output. This indicates the audit classification result of the q-th entry; APPROVED indicates the pass status. This indicates the overall audit score for this item; This indicates the lower threshold for publication, typically set to 0.80, consistent with the pass / fail score line. Represents the logical AND operation.

[0043] Step 52: Based on the filtered output data, perform output locking processing to set the audit execution path as the only legitimate output source, obtaining the locked output data. Specifically, this includes: constructing an output source locking identifier structure in the output locking module based on the verified compliant set of output data. This structure declares the audit execution path as the only legitimate source of the current output data and marks its exclusivity and locking expiration time. The identifier generation formula is: ; In the formula, L represents the output source lock identifier structure; SOURCEAUDIT indicates that the output source is forcibly specified as the audit execution path; This indicates the lock deadline; EXCLUSIVE indicates that this path is an exclusive and unique output source. Based on the output source lock identifier structure, the audit execution path and the verified and compliant set of data to be output are bound and encapsulated in the output locking module to generate an output data structure with a lock tag, and the actual lock expiration time is calculated. The binding formula is as follows: ; ; In the formula, This indicates the output data after locking; This represents the set of compliant data to be output after verification; L represents the output source lock identifier structure. Indicates the actual expiration time of the lock; Indicates the lockout deadline; This represents the current timestamp when the locking operation was performed; `minimum_lock_protection_duration` indicates the minimum lock duration, typically 30 seconds, used to prevent premature release of the lock due to system time drift; `max` indicates the larger of two time values. Based on the output data after locking, access control constraints are applied to the external output interface in the output locking module. This constraint stipulates that before the actual lock expiration time, the response data source of all external output requests is forcibly directed to the audit execution path, and the output interface that generates the execution path remains shielded and cannot be called. The constraint formula is: ; ; In the formula, This represents the function for determining the source of the output data; The quantifier indicates a universal quantifier, meaning that all output requests within the lock period are subject to this constraint; Request indicates any output request initiated by an external system or user; Response indicates the response data returned for this request; AUDIT indicates the audit execution path; DEFAULT indicates the default path configuration, i.e., the normal dual-path mode that is restored after the lock expires; Indicates the time when the request occurred; This indicates the actual expiration time of the lockout.

[0044] Step 53: Based on the locked output data, extract the audit results containing the "passed" status to obtain the final audit pass results. Specifically, this includes: based on the locked output data, traversing the locked output data set in the result extraction module, filtering out candidate output items with an audit classification result of "passed", extracting each item and its corresponding comprehensive audit score, and obtaining the final audit pass result set. The extraction formula is as follows: ; In the formula, This represents the final set of audit approval results; This represents the q-th candidate output entry; This indicates the overall audit score for this item; This indicates the output data after locking; This indicates the audit classification result for this item; APPROVED indicates the pass status. Based on the final audit pass result set, the total number of items in this set is counted, and the average comprehensive audit score of all pass items is calculated to generate an audit quality statistical summary. This summary is used to quantitatively evaluate the output quality level of this round of audit execution path. The statistical formula is as follows: ; ; In the formula, This represents the average overall audit score for the items that passed the audit; This indicates the total number of items that passed the audit; This indicates the number of elements in the audit pass result set; This represents the overall audit score for the q-th passed item; This represents the summation of the comprehensive audit scores for all entries in the set. Based on the audit pass result set and the audit quality statistical summary, the final output data is assembled into a standardized output structure. This structure includes all candidate output entries that passed the audit, the comprehensive audit score for each entry, the total number of entries, the average score, and the output generation time, for direct consumption by downstream systems or users. Its output structure is represented as follows: ; In the formula, This represents the structure of the final, standardized output. This represents the set of audit approval results; Indicates the total number of entries passed; This represents the average overall audit score; The timestamp represents the moment the final output was generated, used to record the output delivery time.

[0045] In this embodiment of the invention, the output filtering process strictly follows the permission status in the path control instruction to completely eliminate low-confidence candidate outputs generated by the execution path, retaining only entries that have been verified as compliant by the audited execution path, thus preventing insufficiently audited content from being mixed into the final output from the source. The output locking process introduces an exclusive locking identifier and access control constraints, forcing all external requests to obtain content only through the audited execution path during the locking period, eliminating the risk of permission bypass that may be caused by parallel dual paths. The result extraction module uses audit classification results and comprehensive audit scores as dual screening conditions to ensure that the final output entries meet the preset standards in the three dimensions of completeness, consistency, and logical verification, and provides a quantifiable output quality metric for downstream systems with an audit quality statistical summary, realizing multi-layered progressive control over the output content.

[0046] like Figure 2 As shown, embodiments of the present invention also provide a trusted artificial intelligence audit control system based on asymmetric dual execution paths, comprising: The acquisition module receives input data containing device status parameters, performs standardization, noise filtering, and data segmentation on the input data to obtain the original feature vector; based on the original feature vector, it performs feature extraction and semantic embedding calculation to obtain the semantic vector. The fusion module is used to obtain the dynamic logical graph of device operation data maintenance based on semantic vectors; calculate the semantic distance between the semantic vectors and existing nodes in the node set; and perform node fusion, relationship establishment and new isolated node determination based on semantic distance and preset dual threshold mechanism to obtain the updated dynamic logical graph. The detection module is used to perform semantic expansion and multi-path reasoning generation in the generation execution path based on the updated dynamic logic graph to obtain a candidate output set; based on the candidate output set and the updated dynamic logic graph, it performs integrity detection, consistency detection and logical conflict identification in parallel in the audit execution path to obtain the audit result. The shutdown module is used to determine the path switching based on the confidence parameter corresponding to the audit result through the path priority control module. When the confidence parameter is lower than a preset threshold, the output permission to generate the execution path is closed, and a path control instruction is obtained. The locking module is used to perform output filtering and output locking through the output control module according to the path control instructions and audit results, so as to obtain the final audit pass result.

[0047] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0048] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0049] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0050] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A trusted artificial intelligence audit control method based on asymmetric dual execution paths, characterized in that, The method includes: Step 1: Receive input data containing device status parameters, perform standardization, noise filtering, and data segmentation on the input data to obtain the original feature vector; based on the original feature vector, perform feature extraction and semantic embedding calculation to obtain the semantic vector; Step 2: Based on the semantic vector, obtain the dynamic logical graph of device operation data maintenance; calculate the semantic distance between the semantic vector and the existing nodes in the node set, and perform node fusion, relationship establishment and new isolated node determination according to the semantic distance and the preset dual threshold mechanism to obtain the updated dynamic logical graph; Step 3: Based on the updated dynamic logic graph, perform semantic expansion and multi-path reasoning generation in the generation execution path to obtain a candidate output set; based on the candidate output set and the updated dynamic logic graph, perform integrity detection, consistency detection and logical conflict identification in parallel in the audit execution path to obtain the audit result; Step 4: Based on the confidence parameter corresponding to the audit result, the path priority control module performs path switching judgment. When the confidence parameter is lower than the preset threshold, the output permission to generate the execution path is closed, and a path control instruction is obtained. Step 5: Based on the path control instructions and audit results, the output control module performs output filtering and output locking to obtain the final audit pass result.

2. The trusted artificial intelligence audit control method based on asymmetric dual execution paths according to claim 1, characterized in that, Step 1: Receive input data containing device status parameters; perform standardization, noise filtering, and data segmentation on the input data to obtain the original feature vector; based on the original feature vector, perform feature extraction and semantic embedding calculation to obtain the semantic vector, including: Receive input data containing text data, structured data, and control command data; perform data standardization processing based on the input data to unify the data format and obtain standardized data; Based on the standardized data, noise filtering is performed to remove invalid or abnormal information, resulting in clean data; based on the clean data, data sharding is performed to split the data into multiple data sub-units, resulting in the original feature vector. Based on the original feature vector, feature extraction and vector encoding are performed to obtain the dimensionality-reduced feature vector; based on the dimensionality-reduced feature vector, semantic embedding calculation is performed to obtain the mapped semantic vector.

3. The trusted artificial intelligence audit control method based on asymmetric dual execution paths according to claim 2, characterized in that, Step 2: Based on the semantic vector, obtain the dynamic logical graph for maintaining device operation data; calculate the semantic distance between the semantic vector and existing nodes in the node set, and perform node fusion, relationship establishment, and determination of newly added isolated nodes based on the semantic distance and a preset dual threshold mechanism to obtain the updated dynamic logical graph, including: Obtain a dynamic logic graph maintained based on historical equipment operation data. The dynamic logic graph contains a set of nodes and a set of relationships representing historical equipment entities and their associated logic. Extract existing nodes from the node set. Calculate the semantic distance between the semantic vector and the existing nodes to obtain the node semantic distance value; based on the node semantic distance value and the preset dual threshold mechanism, perform a three-stage structure update judgment to obtain the graph update instruction; According to the graph update instructions, when the node fusion condition is met, the node fusion process is executed; when the relationship establishment condition is met, the logical association establishment process is executed; and when the isolated node condition is met, the process of adding an isolated node is executed, thus obtaining the updated dynamic logical graph. Based on the updated dynamic logic graph, the active weights of the nodes in the graph are updated exponentially over time to obtain the current graph state.

4. The trusted artificial intelligence audit control method based on asymmetric dual execution paths according to claim 3, characterized in that, Step 3: Based on the updated dynamic logic graph, perform semantic expansion and multi-path reasoning generation in the generation execution path to obtain a candidate output set, including: Based on the hyperbolic space graph representation, the geodesic distance between the target node and the associated nodes is calculated to obtain the spatial geometric distance matrix; Based on the spatial geometric distance matrix, semantic association calculation based on the shortest geodesic path is performed in the generated execution path to obtain semantic extended features; Based on semantic extension features, multi-path reasoning generation processing on multi-dimensional topological manifolds is performed to obtain multi-path reasoning results; Based on the multi-path reasoning results, candidate output construction processing is performed to obtain a candidate output set. The candidate output set is only an intermediate result and does not have the final output permission. It is then transmitted to the audit execution path.

5. The trusted artificial intelligence audit control method based on asymmetric dual execution paths according to claim 4, characterized in that, Based on the candidate output set and the updated dynamic logic graph, integrity checks, consistency checks, and logic conflict identification are performed in parallel within the audit execution path to obtain audit results, including: Based on the candidate output set and the current graph state, a negative logic scan is performed in the audit execution path to obtain the logic features to be detected; Based on the logical features to be detected, an integrity check is performed to verify whether the necessary information structure requirements are met, and the integrity check result is obtained. Based on the integrity check result and the current graph state, a consistency check is performed to determine whether there are any node or relationship conflicts, and the consistency check result is obtained. Based on the consistency check results, perform logical verification to identify logical contradictions or factual conflicts, and obtain the logical verification results; based on the logical verification results, perform adjudication processing to obtain audit results including pass status, conflict status, and missing status.

6. The trusted artificial intelligence audit control method based on asymmetric dual execution paths according to claim 5, characterized in that, Step 4: Based on the confidence parameter corresponding to the audit results, the path priority control module performs a path switching determination. When the confidence parameter is lower than a preset threshold, the output permission for generating execution paths is closed, and path control instructions are obtained, including: Obtain the confidence level parameter corresponding to the audit result to get the confidence level value to be judged; compare the confidence level value to be judged with the preset threshold to make a judgment and get the path switching judgment result; Based on the path switching determination result, when the confidence value to be determined is lower than the preset threshold, a blocking instruction to close the permission to generate and execute path output is generated, and the initial path control instruction is obtained. Based on the initial path control instructions, the output permission is switched through the path priority control module, the output channel for generating execution paths is closed and the audit execution path is locked as the only output channel, thus obtaining the final path control instructions.

7. The trusted artificial intelligence audit control method based on asymmetric dual execution paths according to claim 6, characterized in that, Step 5: Based on the path control instructions and audit results, perform output filtering and output locking through the output control module to obtain the final audit pass result, including: Obtain the final path control instructions and audit results to get the output data to be processed; based on the output data to be processed, perform output filtering to filter out the candidate output set generated by the execution path, and get the filtered output data to be processed. Based on the filtered data to be output, an output locking process is performed to set the audit execution path as the only legitimate output source, resulting in locked output data. Based on the locked output data, extract the audit results that include the pass status to obtain the final audit pass result.

8. A trusted artificial intelligence audit control system based on asymmetric dual execution paths, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module receives input data containing device status parameters, performs standardization, noise filtering, and data segmentation on the input data to obtain the original feature vector; based on the original feature vector, it performs feature extraction and semantic embedding calculation to obtain the semantic vector. The fusion module is used to obtain the dynamic logical graph of device operation data maintenance based on semantic vectors; calculate the semantic distance between the semantic vectors and existing nodes in the node set; and perform node fusion, relationship establishment and new isolated node determination based on semantic distance and preset dual threshold mechanism to obtain the updated dynamic logical graph. The detection module is used to perform semantic expansion and multi-path reasoning generation in the generation execution path based on the updated dynamic logic graph to obtain a candidate output set; based on the candidate output set and the updated dynamic logic graph, it performs integrity detection, consistency detection and logical conflict identification in parallel in the audit execution path to obtain the audit result. The shutdown module is used to determine the path switching based on the confidence parameter corresponding to the audit result through the path priority control module. When the confidence parameter is lower than a preset threshold, the output permission to generate the execution path is closed, and a path control instruction is obtained. The locking module is used to perform output filtering and output locking through the output control module according to the path control instructions and audit results, so as to obtain the final audit pass result.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images