Multi-task allocation and decision optimization system of communication engineering auditing agent

Abstract

The invention relates to the technical field of artificial intelligence and communication engineering auditing, in particular to a multi-task allocation and decision optimization system for a communication engineering auditing agent, which comprises a data perception analysis module, a task feature vector generation module and a task feature extraction module, and the data perception analysis module is used for acquiring communication engineering settlement documents and unstructured field data and extracting features to generate task feature vectors; the task entropy evaluation module calculates a task information entropy, and divides an audit task into a low-cognitive load task or a high-cognitive load task; the cognitive resource adaptation module constructs an intelligent agent capability model, dynamically constructs an intelligent agent cooperation topological structure based on a cognitive adaptation degree coefficient and allocates tasks; the decision confidence monitoring module tracks an auditing process in real time, calculates a decision confidence fluctuation rate, and executes automatic auditing termination or triggers manual intervention operation; the efficiency feedback adjustment module calculates a dynamic resource efficiency ratio; according to the method, the problem that a fixed rule process cannot cope with high dynamic performance of unstructured data is solved, and dynamic balance between auditing efficiency and computing power cost is realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and communication engineering auditing technology, specifically to a multi-task allocation and decision optimization system for a communication engineering auditing intelligent agent. Background Technology

[0002] Telecommunications engineering auditing is an important means of ensuring the security of funds for base station construction and line maintenance. It mainly involves verifying documents such as settlement statements, concealed works visa forms, and as-built drawings to confirm the authenticity and compliance of the project quantities. Current auditing systems mostly adopt manual review or electronic processes based on static rules, and have good processing capabilities for standardized documents with uniform formats.

[0003] With the deepening of digital transformation, the proportion of unstructured data in audit objects has increased significantly. However, when processing on-site data containing blurry images, handwritten text, and complex layouts, the existing fixed processes lack the ability to dynamically perceive the cognitive load of the task and cannot adaptively allocate computing resources according to the complexity of the documents. This leads to misjudgments or omissions when facing highly uncertain tasks, while wasting expensive inference resources when processing simple tasks, which seriously restricts the overall efficiency in large-scale audit scenarios.

[0004] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention discloses a multi-task allocation and decision optimization system for a communication engineering auditing intelligent agent. Specifically, the technical solution of this invention is as follows: In the cloud, the cloud communication connection includes a data perception and analysis module, a task entropy value evaluation module, a cognitive resource adaptation module, a decision confidence monitoring module, and an efficiency feedback and adjustment module; The data perception and parsing module is used to acquire settlement documents for communication engineering and unstructured field data, extract features from the unstructured field data, and generate task feature vectors. The task entropy evaluation module is used to calculate the task information entropy value based on the task feature vector, and classify the audit task into low cognitive load task or high cognitive load task according to the task information entropy value. The cognitive resource adaptation module is used to build an agent capability model, calculate the cognitive adaptation coefficient between the agent capability model and the task feature vector, and dynamically build an agent collaborative topology and assign audit tasks based on the cognitive adaptation coefficient and task classification results. The decision confidence monitoring module is used to track the audit process of the intelligent agent collaborative topology in real time, calculate the decision confidence volatility, and execute automated audit termination operations or trigger manual intervention operations based on the decision confidence volatility. The performance feedback adjustment module is used to calculate the dynamic resource performance ratio based on the audit results, and to adjust the judgment threshold of the task entropy evaluation module based on the dynamic resource performance ratio.

[0006] Optionally, the data perception and parsing module performs feature extraction on unstructured field data, including: Acquire image data, visa document text, and engineering drawing data from unstructured field data; The semantic ambiguity of the data is calculated by the reciprocal of the average confidence level of optical character recognition, and the missing rate of key elements is calculated based on the missing proportion of required fields in the document. The Euclidean distance between the current data features and the centroids of the historical baseline patterns is calculated as the deviation feature. Combined with semantic ambiguity, key element missing rate and deviation feature, a task feature vector representing the uncertainty of the current audit task is generated.

[0007] Optionally, the process by which the task entropy evaluation module classifies audit tasks into low-cognitive-load tasks or high-cognitive-load tasks includes: Preset entropy threshold; The task feature vector is normalized to obtain the probability distribution of feature elements, and the Shannon entropy of the probability distribution of the feature elements is calculated as the task information entropy value. The task information entropy value is compared with the entropy value judgment threshold. If the task information entropy value is less than the entropy value judgment threshold, the audit task is marked as a low cognitive load task. If the task information entropy value is greater than or equal to the entropy value judgment threshold, the audit task will be marked as a high cognitive load task.

[0008] Optionally, the process by which the cognitive resource adaptation module dynamically constructs the agent collaborative topology and assigns auditing tasks includes: When an audit task is marked as a low cognitive load task, the agent capability model is traversed, the single agent with the highest cognitive fit coefficient is selected, a single-node decision topology is established, and the audit task is assigned to the single agent. When an audit task is marked as a high cognitive load task, multiple agents with a cognitive fit coefficient greater than a preset collaborative admission threshold are selected, a multi-node collaborative debate topology is established, and the audit task is assigned to the multi-node collaborative debate topology.

[0009] Optionally, the process by which the cognitive resource adaptation module calculates the cognitive adaptation coefficient includes: Construct an agent capability feature vector that includes logical verification capability, cost estimation capability, and image and text recognition capability; Calculate the cosine similarity between the agent's capability feature vector and the task feature vector, and use the cosine similarity as the cognitive fit coefficient, where the cognitive fit coefficient characterizes the degree of matching of the agent in handling the current audit task.

[0010] Optionally, the process by which the decision confidence monitoring module performs automated audit termination operations or triggers manual intervention operations based on the decision confidence volatility includes: The quantitative intermediate audit amount output by the intelligent agent collaborative topology during multiple rounds of audit inference is obtained, the variance of the quantitative intermediate audit amount is calculated, and the decision confidence volatility is generated. A preset volatility safety threshold is set. If the decision confidence volatility is less than the volatility safety threshold, the audit results are determined to be converged, the final audit conclusion is output, and an automated audit termination operation is performed. If the decision confidence volatility is greater than or equal to the volatility safety threshold, the audit results are determined to have a high-risk discrepancy. The current agent collaboration topology is suspended, and an abnormal warning signal is generated to trigger manual intervention.

[0011] Optionally, the process by which the efficiency feedback adjustment module calculates the dynamic resource efficiency ratio includes: Obtain the risk correction amount identified in this audit task, and obtain the computing resource cost and time cost consumed in this audit task; Based on the preset performance calculation logic, the risk correction amount is divided by the weighted sum of the calculated resource cost and time cost to obtain the dynamic resource performance ratio.

[0012] Optionally, the process by which the performance feedback adjustment module corrects the judgment threshold of the task entropy evaluation module includes: A preset performance benchmark value is used to determine whether the dynamic resource efficiency ratio is lower than the performance benchmark value. If the dynamic resource efficiency ratio is lower than the efficiency benchmark, the entropy threshold in the task entropy evaluation module will be increased to reduce the triggering frequency of high cognitive load tasks. If the dynamic resource efficiency ratio is higher than or equal to the efficiency benchmark value, the entropy value judgment threshold in the task entropy value assessment module will be maintained or lowered to increase the audit depth of potential risks.

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention, through the collaborative work of the data perception and analysis module and the task entropy evaluation module, utilizes layout fingerprint extraction technology to quantify the semantic ambiguity, key element missing rate, and layout deviation characteristics of unstructured data, and introduces Shannon entropy from information theory to calculate the task information entropy value; it enables the precise division of communication engineering audit tasks into low cognitive load tasks or high cognitive load tasks, effectively solving the problem that traditional fixed rule processes cannot handle highly dynamic unstructured data such as construction site photos and handwritten documents, and providing a scientific quantitative basis for the subsequent on-demand allocation of computing power.

[0014] 2. This invention dynamically constructs an agent collaborative topology based on the cognitive adaptation coefficient through a cognitive resource adaptation module. For standardized low cognitive load tasks, a single-node decision topology is established to greatly improve system throughput. For complex high cognitive load tasks, a multi-node collaborative debate topology including audit executors, norm challengers, and logic judges is automatically formed. By utilizing role-based prompting engineering and finite state machine iterative control algorithms, through multiple rounds of adversarial and collaborative reasoning, the risk of machine illusion when a single agent processes fuzzy data is effectively suppressed, and the reliability and robustness of audit conclusions in complex scenarios are significantly improved.

[0015] 3. This invention tracks the quantitative intermediate audit amount in real time during the intelligent agent collaboration process through a decision confidence monitoring module, calculates the decision confidence volatility reflecting the stability of the conclusion, and combines it with a dynamic safety threshold based on business tolerance to make judgments. When the volatility is less than the threshold, it performs automatic termination to release resources, and when the volatility continues to oscillate, it promptly triggers manual intervention, thus building a reliable risk safety net to ensure that the system only automatically passes when the conclusion is highly convergent. This effectively solves the common problems of blind confidence and misjudgment in automated audit systems.

[0016] 4. This invention constructs an evaluation system based on both technical and economic dimensions through an efficiency feedback adjustment module. It calculates the dynamic resource efficiency ratio based on the risk correction amount, computational resource cost, and time opportunity cost, and uses this to correct the judgment threshold of the task entropy evaluation module. This endows the system with an adaptive evolutionary ability similar to that of a living organism, enabling it to automatically and dynamically switch between extensive and efficient and intensive and in-depth mining modes according to the actual situation of peak business periods or low-risk periods. This ensures that the system maintains the optimal economic balance between audit revenue and computing power cost throughout its entire life cycle. Attached Figure Description

[0017] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0019] Example 1: Please see Figure 1 A multi-task allocation and decision optimization system for a communication engineering auditing intelligent agent includes: a cloud, which is connected to a data perception and parsing module, a task entropy value evaluation module, a cognitive resource adaptation module, a decision confidence monitoring module, and an efficiency feedback adjustment module. The data perception and parsing module is used to acquire settlement documents for communication engineering and unstructured field data, extract features from the unstructured field data, and generate task feature vectors. The task entropy evaluation module is used to calculate the task information entropy value based on the task feature vector, and classify the audit task into low cognitive load task or high cognitive load task according to the task information entropy value. The cognitive resource adaptation module is used to build an agent capability model, calculate the cognitive adaptation coefficient between the agent capability model and the task feature vector, and dynamically build an agent collaborative topology and assign audit tasks based on the cognitive adaptation coefficient and task classification results. The decision confidence monitoring module is used to track the audit process of the intelligent agent collaborative topology in real time, calculate the decision confidence volatility, and execute automated audit termination operations or trigger manual intervention operations based on the decision confidence volatility. The performance feedback adjustment module is used to calculate the dynamic resource efficiency ratio based on the audit results, and to adjust the judgment threshold of the task entropy evaluation module based on the dynamic resource efficiency ratio. This embodiment details the overall architecture and operational logic of the multi-task allocation and decision optimization system for the intelligent agent of communication engineering audit. The system is deployed based on a cloud architecture and uses a data perception and parsing module as the system's sensory nerve endings to obtain communication engineering settlement documents, such as base station construction settlement statements, as well as unstructured on-site data, including concealed works visa forms, construction site photos, and CAD as-built drawings. The system introduces the concept of entropy from information theory through the task entropy evaluation module. Based on the generated task feature vector, the system calculates the task information entropy value. This indicator aims to quantify the intrinsic uncertainty of audit tasks, thereby classifying tasks into standardized low cognitive load tasks or high cognitive load tasks involving complex changes. Based on this, the cognitive resource adaptation module is responsible for the scheduling and networking of agents. By constructing agent capability models and calculating their cognitive adaptation coefficients with task feature vectors, it dynamically constructs agent collaborative topology structures. For example, it constructs a single-node structure for low-load tasks and a multi-node collaborative structure for high-load tasks. During the audit execution process, the decision confidence monitoring module tracks the audit process of the agent collaborative topology structure in real time, calculates the decision confidence volatility that reflects the stability of the conclusions, and executes an automated audit termination operation when the volatility is less than the safety threshold. In response to continuous volatility fluctuations, manual intervention is triggered; the efficiency feedback adjustment module calculates the dynamic resource efficiency ratio based on the final audit results and corrects the judgment threshold of the task entropy evaluation module in reverse, thereby realizing the adaptive evolution of the system. This embodiment constructs an audit ecosystem with adaptive evolution capabilities. In the context of communication engineering auditing, it solves the problem that fixed rule processes cannot cope with the high dynamism of unstructured data. By transforming static rule matching into dynamic cognitive adaptation, the system can autonomously determine the computing power investment based on the ambiguity of the documents. This ensures a high detection rate for violations in hidden engineering projects while avoiding the waste of expensive inference resources on simple documents, thus achieving a dynamic balance between audit efficiency and computing power costs.

[0020] Example 2: The data perception and analysis module performs feature extraction on unstructured field data, including: acquiring image data, visa document text, and engineering drawing data from the unstructured field data; The semantic ambiguity of the data is calculated by the reciprocal of the average confidence level of optical character recognition, and the missing rate of key elements is calculated based on the missing proportion of required fields in the document. The Euclidean distance of the current data features relative to the centroid of the historical baseline pattern is calculated as the deviation feature. Combining the semantic ambiguity, the missing rate of key elements, and the deviation feature, a task feature vector representing the uncertainty of the current audit task is generated. This embodiment is a further specification of the feature extraction steps of the data perception and analysis module based on embodiment 1; the system performs multimodal data acquisition operations to collect image data from the communication engineering site, such as wall photos of the construction site, photos of concealed optical cables during construction, and visa document text containing handwritten signatures and engineering drawing data in CAD or PDF format; The system performs quality quantification calculations on the above data. It processes the document text based on the optical character recognition engine, extracts the average confidence score and calculates its reciprocal to obtain the semantic ambiguity. At the same time, it detects required fields based on the preset document template and calculates the missing rate of key elements. The system calls the historical baseline pattern library to calculate the Euclidean distance of the current data features relative to the centroid of the historical baseline pattern to obtain the deviation features. In this step, to clarify the mathematical definitions of the centroids of the current data features and historical baseline patterns and to fully disclose the calculation process of deviation features, the system employs layout fingerprint extraction technology: to eliminate photo resolution differences caused by different acquisition devices, the system uniformly scales the input document images to a standardized resolution, for example... The system calculates pixels and performs grayscale conversion. To ensure that grid division is performed at a uniform spatial scale and to eliminate ambiguity in the operation objects, the system strictly uses the scaled and grayscale-processed image data as the operation objects, dividing it into pixels. The grid, for each grid cell , Calculate the text box coverage density within the grid. and pixel grayscale average ; Among them, text box coverage density Defined as the first The ratio of the sum of the areas of all text bounding boxes detected by OCR within a grid to the total area of ​​that grid is used to ensure feature normalization. If overlapping bounding boxes cause the sum of their areas to exceed the grid area, then... The value is truncated to 1.0 to ensure that its range is strictly within the specified range. ; To eliminate the noise interference of blank background on feature vectors, the system adopts a grayscale inversion mapping strategy; the corrected formula for calculating the grid composite layout response value is as follows: Among them, items This ensures that when the grid has a pure white background, i.e. At that time, the contribution of the grayscale component is zero; only when ink is present, i.e. A positive response value is generated only at this time. The weighting coefficients of 0.7 and 0.3 here are the optimal projection vector components that maximize the inter-class variance, obtained by Fisher linear discriminant analysis on a training set containing 1000 standard compliant documents and forged documents. Specifically, the system constructs the objective function. : in, Let be the projection vector to be solved. The inter-class scatter matrix, The within-class scatter matrix is ​​used to solve the generalized eigenvalue problem on the training set containing 1000 samples. The eigenvector corresponding to the largest eigenvalue, after normalization, is... This weight is not fixed; the system supports re-triggering the above calculation to update the weight after accumulating 5000 new samples to ensure the dominance of structural features, i.e., density. The system will calculate... Flattened into a 64-dimensional page structure vector, denoted as... The corresponding historical baseline pattern centroid is the center vector obtained by clustering the layout structure vectors of historical compliant documents using the K-Means algorithm, denoted as... ; To ensure the physical meaning of deviation feature calculations and prevent distance calculation failures due to document type mismatches, such as comparing drawings with invoices, the system pre-builds corresponding features for different business types, such as concealed visas, as-built drawings, and main material lists. A set of baseline centroids Before performing the calculation, first perform the calculation. Using the cosine similarity of each centroid, the current document is matched to the nearest centroid category, and the centroid of that specific category is selected as the centroid. By calculating the difference score between the two, i.e., a dimensionless metric calculated based on the Euclidean distance formula, the degree of abnormality in the current document's layout can be effectively quantified. The system integrates the above three dimensions of indicators to generate a task feature vector representing the uncertainty of the current audit task. The relevant calculation formula is as follows: in, Semantic ambiguity, derived from statistical calculations of OCR recognition results, physically refers to the difficulty of recognizing the document text; : No. The confidence score for each character recognition comes from the OCR engine output, and its value range is [value range missing]. ; The total number of characters identified is derived from OCR statistics; this is for the extreme case where no text is detected. At that time, the system forcibly sets the average confidence term to 0, thus reducing semantic ambiguity. Take the maximum value To accurately reflect the high uncertainty of blank documents; A tiny constant, derived from a preset value, used to prevent the denominator from being zero; The key element missing rate, derived from template matching results, physically represents the completeness of document compliance elements. The number of missing required fields is determined by field validation. The total number of required fields is derived from the document specification definition; Deviation features, derived from the Euclidean distance formula mentioned above, are essentially dimensionless difference fractions, and their physical meaning is the degree of difference between the current task and the standard mode. The first feature vector in the current data The dimensional fusion layout feature values ​​are derived from gridded feature extraction and weighted fusion; : Historical baseline model centroid The central eigenvalues ​​of the dimension are derived from the cluster centers; The task feature vector is derived from a combination of the above indicators, and its physical meaning is the digitized task complexity fingerprint. This embodiment constructs a multi-dimensional feature vector that includes ambiguity, missing rate, and deviation. In particular, it solves the problem of dimensionality reduction mapping from multi-dimensional physical features to one-dimensional vector elements within a grid by using a weighted fusion algorithm. When dealing with dirty data scenarios common in communication engineering, it can transform the degree of illegible handwriting and the degree of document damage, which are difficult to quantify, into mathematical indicators that can be processed by computers. Through a category-based centroid matching mechanism, the system can effectively detect non-standard or potentially forged documents, providing a robust quantitative basis for downstream computing power allocation. Furthermore, regarding the issue of obtaining the historical baseline pattern centroids mentioned in the background technology section of the manual during the initial system deployment phase, i.e., the cold start period, the system executes the following initialization construction process: Before the system goes live for the first time, a seed dataset containing a preset number, such as 50 standard compliant documents, is created. The seed dataset is screened and confirmed by human audit experts. The system runs the K-Means algorithm based on this seed dataset, and the generated cluster centers are used as the initial cluster centers. As subsequent actual audit tasks are carried out, the system will dynamically add the feature vectors of newly generated and confirmed compliant documents to the historical database, and recalculate and update the centroid according to a preset cycle, such as weekly, so as to achieve a smooth transition from cold start to adaptive operation.

[0021] Example 3: The task entropy evaluation module classifies audit tasks into low-cognitive-load or high-cognitive-load tasks as follows: a preset entropy threshold is set; the task feature vector is normalized to obtain the probability distribution of feature elements, and the Shannon entropy of the feature element probability distribution is calculated as the task information entropy value; the task information entropy value is compared with the entropy threshold. If the task information entropy value is less than the entropy threshold, the audit task is marked as a low-cognitive-load task; if the task information entropy value is greater than or equal to the entropy threshold, the audit task is marked as a high-cognitive-load task. This embodiment is a further refinement of the classification logic of the task entropy evaluation module based on embodiment 2; the system presets an entropy value judgment threshold, which is set based on the statistical average entropy value of historical data, and typically takes a value range of [value range missing]. To ensure the reproducibility and objectivity of this statistical setting, historical data is specifically defined as data processed and manually verified within a rolling time window, such as the most recent 30 days. For example The set of feature vectors for audit tasks; The average entropy is calculated using the arithmetic mean after removing outliers; preferred range The selection criteria are based on statistical distributions under the law of large numbers: experimental data show that the entropy distribution of standard compliant documents exhibits a narrow-band Gaussian distribution centered at 0.3, while the entropy distribution of problematic documents exhibits a wide-band distribution centered at 0.9; a value of 0.5 as the lower bound corresponds to the distribution of compliant documents. The upper bound effectively covers 99.7% of simple tasks; a value of 0.8 as the upper bound corresponds to the left tail of the high-risk distribution, used to ensure that tasks with extremely high uncertainty are forcibly intercepted. To measure the degree of disorder within the task feature vector, the system uses the Softmax function to normalize the vector elements; in this step, given the task feature vector... Each dimension, namely semantic ambiguity Missing rate Deviation characteristics There are significant differences in physical dimensions and numerical ranges, for example... As a dimensionless difference fraction, its value may be much larger than that of a ratio that is merely a scale. Direct calculation would cause the probability distribution to tilt extremely towards large numerical features, thus artificially suppressing the entropy value; therefore, the system first performs dynamic range scaling on the feature vector: in, and The first Historical sliding window mean and standard deviation of 3D features; based on scaled feature values Calculate the probability distribution of the feature elements: Based on the principles of information theory, the Shannon entropy of this probability distribution is calculated to obtain the task information entropy value: Based on this, the system executes threshold judgment logic: in response to the task information entropy value being less than the entropy value judgment threshold, the system determines that the current document feature distribution has a significant tendency, that is, a certain feature is dominant, such as only extreme layout deviations with clear text, which is obviously abnormal, and marks it as a low cognitive load task to quickly block or pass it. In response to the task information entropy value being greater than or equal to the entropy value judgment threshold, the system determines that the current document features are evenly distributed, that is, ambiguity, missing rate and deviation coexist, and the state is chaotic, and marks it as a high cognitive load task. This embodiment eliminates the bias effect of feature dimensions on entropy calculation by introducing a statistical standardization preprocessing step, ensuring that the entropy value truly reflects the cognitive complexity of the task. The system lacks sufficient historical data, i.e., the number of samples, during the initial operation phase. The statistical window requirements were not met, for example This makes it impossible to calculate. and In this case, the system has a built-in default set of statistical parameters to support cold starts: Based on prior knowledge of communication engineering documents, for values ​​ranging from... semantic ambiguity and key element missing rate Preset initial mean This represents the expected level of high-quality documents and the initial standard deviation. For deviation characteristics Preset initial mean and initial standard deviation This default parameter set is forced to take effect before the cumulative number of samples reaches the threshold. Once the number of samples meets the statistical requirements, the system automatically switches to the real-time statistical calculation mode based on a sliding window, ensuring the feasibility of the system in the absence of historical accumulation.

[0022] Example 4: The process of dynamically constructing the agent collaboration topology and allocating audit tasks by the cognitive resource adaptation module includes: when an audit task is marked as a low cognitive load task, traversing the agent capability model, selecting the single agent with the highest cognitive adaptation coefficient, establishing a single-node decision topology, and allocating the audit task to the single agent; when an audit task is marked as a high cognitive load task, selecting multiple agents with a cognitive adaptation coefficient greater than a preset collaboration admission threshold, establishing a multi-node collaborative debate topology, and allocating the audit task to the multi-node collaborative debate topology; This embodiment is a further specification of the cognitive resource adaptation module topology construction process based on embodiment 3; the system receives task classification tags and adaptively adjusts the audit organization structure according to the load type; in response to the audit task being marked as a low cognitive load task, the system traverses the agent capability model library, selects the single agent with the highest cognitive adaptation coefficient, such as an OCR agent that is good at processing standard invoices, and establishes a single-node decision topology, that is, the input is directly connected to the agent's output preliminary conclusion. In response to the audit task being marked as a high cognitive load task, the system determines that a single model has a risk of hallucination and then selects multiple agents with a cognitive fit coefficient greater than a preset collaboration admission threshold, usually set to 0.75, to ensure that the selected agents have basic competence. For example, agents for cost estimation, engineering specification verification, and image forensics are selected simultaneously to establish a multi-node collaborative debate topology. In this topology, agents form fully connected or ring-connected relationships and use adversarial and collaborative mechanisms to conduct multiple rounds of reasoning and debate. Finally, the audit task is assigned to this multi-node topology for execution. In this process, to concretize the implementation logic of collaborative debate, the multi-node topology adopts a role-based prompting engineering mechanism. The system instantiates three intelligent agent roles: audit executor, normative challenger, and logic judge. In this step, the system does not randomly assign roles, but executes a greedy matching strategy based on capability vectors: the challenger's role requirement weight is defined as emphasizing logical capability. That is, demand vector The role requirement weight for the executor is defined with a focus on cost control. That is, demand vector The system calculates the capability feature vectors of each selected agent. The dot product with the role requirement vector maps the agent with the highest logical ability component to the norm challenger, the agent with the highest cost estimation ability component to the audit executor, and the agent with the highest comprehensive ability modulus among the remaining agents or the agent with the preset arbitration weight to the logic judge, thereby maximizing the professional efficiency of each node in the topology. To address the potential for infinite loops or deadlocks during the debate process, the system incorporates an iterative control algorithm based on a finite state machine into the multi-node collaborative debate topology. The specific steps are as follows: Initialization: Sets the current reasoning round. Maximum number of rounds allowed Set a logical consistency score threshold ; Execution phase: The audit executor outputs a chain of audit conclusions based on the current context. To ensure the feasibility of subsequent calculations, the system mandates that the output format be a standard JSON array in the prompt, such as [{"id": 1,"amount": 500.00,"basis": "..."},...]. The system parses this JSON array to obtain the number of conclusion entries, denoted as... ; Challenge Phase: Standardizing Challenger Scanning Logical flaws in the output constraint violation set In this step, the system injects a knowledge base called Prompt, which contains mandatory standards such as the "Design Specification for Communication Line Engineering," to force the agent to perform clause-level comparison and verification. The system then writes the discovered conflicting clauses into a structured JSON format, with the key "violation_list." The system parses the list and calculates the number of violations. ;like If empty, mark this challenge as invalid; Refereeing Phase: Logical referee calculates logical consistency score. ,in The pre-defined penalty weights for violations are used to ensure that violations have a significant negative impact on the score; This represents the number of array elements counted by the JSON parser; a ratio-based consistency metric is used here instead of the original linear subtraction formula, specifically to address the normalization trap problem: that is, when the number of entries violates the normalization rule... Significantly more than the number of conclusion items The original formula would force the score to zero, making it impossible to distinguish between different degrees of logical fallacies; the new formula ensures that... and Score under equal boundary conditions It still has smooth numerical discrimination, thus ensuring the effectiveness of the score-based convergence judgment mechanism and determining the termination condition; Termination determination: If or If the debate is converged, then the output is determined. As a final conclusion; otherwise, As a negative feedback prompt, it is injected into the next round of input, making Return to the execution phase; This embodiment realizes on-demand allocation and dynamic orchestration of computing resources. When processing a large number of standardized documents, the single-node mode greatly improves the system throughput. When facing high-risk and complex tasks, an expert consultation group is automatically formed for cross-validation. Through a clearly defined iterative convergence mechanism and structured data interaction protocol, the risk of machine illusion when a single intelligent agent processes fuzzy data is effectively suppressed, and the reliability and robustness of audit conclusions are significantly improved.

[0023] Example 5: The process of calculating the cognitive adaptation coefficient by the cognitive resource adaptation module includes: constructing an agent capability feature vector containing logical verification capability, cost estimation capability, and image and text recognition capability; calculating the cosine similarity between the agent capability feature vector and the task feature vector, and using the cosine similarity as the cognitive adaptation coefficient, where the cognitive adaptation coefficient characterizes the degree of matching of the agent in processing the current audit task; This embodiment is a further refinement of the cognitive fit coefficient calculation logic based on embodiment 4; the system constructs an agent capability feature vector for each agent to be scheduled. Each component corresponds to the normalized image and text recognition capability, logic verification capability, and cost estimation capability, respectively, with values ​​ranging from [value range missing]. ; To address the issue of unclear source of capability vectors, the system performs a lookup initialization operation based on the agent's registry when constructing the vector. This registry pre-stores the performance scores of various basic models on standardized test sets; for example, for OCR agents based on large visual models, values ​​are retrieved from the registry and assigned. For cost estimation agents that focus on numerical calculations, assign values... This ensures that vector values ​​have clear physical meaning and data sources; to ensure the computational effectiveness of the vector space model, the system establishes a semantic mapping between capabilities and tasks: Corresponding task characteristics That is, ambiguity. correspond That is, the missing rate. correspond That is, the degree of deviation; For task feature vectors To address the issue of inconsistent numerical scales and prevent large numerical dimensions from dominating similarity calculations, the system performs a feature magnitude remapping operation before calculating similarity; it uses the Sigmoid function to map unrestricted task feature values ​​to... Intervals, to match the scale of the capability vector: in, In order to target the The system sets the adjustment coefficients for the dimensional features; to adhere to the principle of variable traceability and ensure the effectiveness of the mapping function, the system sets the above coefficients based on the statistical distribution of the historical audit task database: Let the first feature be the adjustment coefficient. The historical mean of the dimensional task features is The historical standard deviation is Then define To align the mapping center to the data distribution center, define By controlling the growth slope of the Sigmoid function, approximately 90% of the historical data is distributed within the linear response range of the Sigmoid function. This avoids the gradient vanishing problem caused by excessively large or small feature values; it also calculates the cosine similarity between the agent's capability feature vector and the mapped task feature vector. The calculated cosine similarity As a cognitive fit coefficient, the closer the value of this coefficient is to 1, the more accurately the high capability dimension of the agent covers the high difficulty dimension of the task. For example, a high ambiguity task is matched with an agent with high recognition capability. This embodiment effectively solves the dimension alignment problem of heterogeneous data by introducing an adaptive feature remapping mechanism based on historical statistical moments, achieving precise alignment between the supply of computing power and task requirements. It also addresses the issue of the lack of historical statistical distribution during the system's cold start phase, which makes it impossible to determine the adjustment coefficient. To address this issue, the system employs a linearized default mapping strategy: when historical data accumulation is insufficient, a default mapping is set... For example, for the midpoint of the feature's range of values, Interval features ,set up For unity gain, for example This allows the Sigmoid function to operate in an approximately linear range; This pre-set approach ensures that, in the absence of statistical priors, the relative magnitudes of task features can be losslessly mapped to similarity calculations, preventing feature distortion caused by improper parameter settings. Once the historical task database has sufficient samples, the system then initiates parameter adaptive iteration logic based on historical mean and standard deviation.

[0024] Example 6: The process by which the decision confidence monitoring module performs automated audit termination operations or triggers manual intervention operations based on the decision confidence volatility includes: obtaining the quantified intermediate audit amount output by the intelligent agent collaborative topology during multiple rounds of audit inference, calculating the variance of the quantified intermediate audit amount, and generating the decision confidence volatility. A preset volatility safety threshold is set. If the decision confidence volatility is less than the volatility safety threshold, the audit results are determined to be converged, the final audit conclusion is output, and an automated audit termination operation is performed. If the volatility of decision confidence is greater than or equal to the volatility safety threshold, the audit results are determined to have a high-risk discrepancy. The current agent collaboration topology is suspended, and an abnormal warning signal is generated to trigger manual intervention. This embodiment is a further specification of the decision confidence monitoring logic based on Embodiment 5; during the multi-round reasoning process in the agent collaborative topology, the system extracts the quantified intermediate audit amount of each round's output in real time; for multi-node collaborative debate topologies, this amount... Defined as in the first At the end of the round of debate, the intelligent agent acting as the audit executor outputs a suggested settlement amount after being revised based on feedback from the logic judge. To ensure the accuracy of numerical extraction and avoid interference from non-numeric characters, the system first uses a regular expression compatible with thousands separators: `Total:\s([0-9,]+\.?[0-9]{0,2})` to capture the amount string. Then, a character cleaning function is executed to remove all commas, parsing it as a floating-point number. This avoids numerical parsing errors caused by format issues; to assess the stability of the conclusions, the system collects audit amount data from the most recent rounds and calculates its statistical variance to generate the decision confidence volatility; the system compares this volatility with a preset volatility safety threshold. It should be noted that, in order to prevent false convergence misjudgments caused by sparse sample points (e.g., only one round of data) leading to a constant variance of 0 in the early stages of auditing, the system only performs the current inference round. To reach the preset minimum statistical window length, for example... Only then is the threshold determination logic activated; in response to the decision confidence volatility being less than the volatility safety threshold, the system determines that the opinions of the agents tend to be consistent, the audit results converge, and then outputs the final conclusion and releases computing resources; in response to the volatility being greater than or equal to the safety threshold, the system determines that there are serious disagreements among the agents and that they cannot reach a consensus on their own, immediately suspends the current topology, generates an anomaly warning signal containing relevant documents and inference logs, and pushes it to the human audit expert's end; the formulas involved are as follows: in, Decision confidence volatility, derived from variance calculation, physically represents the degree of oscillation of audit conclusions across multiple rounds of reasoning; The size of the statistical time window, i.e., the number of rounds, is derived from preset system parameters; : No. The suggested audit amount output by the round-robin inference is derived from the agent's output; :recent The average audit amount for each round is derived from statistical calculations; Considering the vast differences in settlement amounts for telecommunications engineering projects, ranging from hundreds of yuan to millions of yuan, directly using variance might lead to judgment failures. Therefore, this embodiment employs a dynamically set logic for the volatility safety threshold, with the formula as follows: ,in For example, a preset limit on the coefficient of variation. This parameter The value of is based on the business's tolerance for audit accuracy, and its value is strictly equal to the square of the maximum allowable relative standard deviation. For example, if auditing standards require that the estimation discrepancies among agents must not exceed the mean. ,but This derivation process ensures that the threshold setting is directly linked to business KPIs. For example, a non-zero stability factor. Its value is determined by a comprehensive consideration of the square of the smallest unit of currency, the cent, and the background noise of the system's floating-point calculation error; when the system amount is in yuan, this factor is set to... This is to ensure the consistency of physical dimensions in the addition operations in the formula; This value can be in the average amount This means that precise numerical damping is provided when there are zero settlement documents to prevent the threshold from collapsing to zero, so that the volatility judgment mechanism maintains the correctness of physical meaning when dealing with amounts ranging from tiers to hundreds of millions of yuan, and adapts to audit projects of different sizes; this embodiment builds a reliable risk bottom-line defense to ensure that the system automatically passes only when the conclusion is highly convergent, effectively solving the problem of blind confidence that is common in automated systems.

[0025] Example 7: The process of calculating the dynamic resource efficiency ratio by the efficiency feedback adjustment module includes: obtaining the risk correction amount identified in this audit task, obtaining the computing resource cost and time cost consumed in this audit task; and dividing the risk correction amount by the weighted sum of computing resource cost and time cost according to the preset efficiency calculation logic to obtain the dynamic resource efficiency ratio. This embodiment further specifies the dynamic resource efficiency ratio calculation process based on Embodiment 6. After the audit task is completed, the system extracts key performance data from the logs, including the risk correction amount identified in this audit (i.e., the amount to be reduced), and the computing resource costs consumed in this task, such as GPU machine time and time costs. Based on the preset efficiency calculation logic, and considering that confirming document compliance is also a valid output of the audit system, the system introduces the total audit amount as the value base. The revised dynamic resource efficiency ratio calculation formula is as follows: in, The total amount of documents submitted for review in this audit task; The compliance confirmation value coefficient is derived from a preset constant and its physical meaning is the equivalent economic value generated by the system endorsing each unit of compliant funds. This correction ensures that even in [the context of the previous sentence] In a fully compliant scenario, It can still be maintained within a reasonable range, avoiding the failure of threshold adjustment logic due to sample bias; Dynamic resource efficiency ratio, derived from input-output calculation, physically represents the economic operating efficiency of the system; Risk adjustment amounts are derived from audit results statistics; Calculate resource costs, derived from system resource monitoring logs, in monetary units such as yuan; Time cost is derived from the time tracking of the entire task lifecycle, and the unit is time, such as seconds. Resource cost weighting coefficient, usually set to 1.0, is used to adjust the proportion of hardware costs in performance evaluation; The time-cost conversion factor is derived from the business's pricing for timeliness, such as 0.5 yuan / second. To ensure sufficient quantitative basis for this factor, this embodiment defines its calculation formula as follows: It should be noted that, and These are not hard-coded constants, but dynamic parameters stored in the cloud configuration center. Users can configure them in real time through the system management interface based on the regional labor costs and project funding scale of the current audit project. In the default scenario, To replace the average hourly wage required for manual auditing, for example, a senior cost engineer earns 300 yuan per hour. The specific formula for calculating the opportunity cost of funds tied up due to audit delays is as follows: ,in This is to obtain the total amount of the project submitted for review by analyzing the cover of the settlement document. The annualized loan prime rate (LPR) configured for the system, for example 3.45%, will be used in the formula to abstract the time dimension. Transform it into an interest loss dimension with clear economic significance to ensure the uniformity of the physical dimensions of the denominator; The tiny constant used to prevent the denominator from being zero comes from preset system parameters; This embodiment transforms the technical audit results into economic performance indicators, focusing not only on the number of problems identified but also on the costs incurred. This indicator provides a quantifiable objective function for the system's self-optimization, preventing the system from consuming excessive computing resources in pursuit of minor improvements in accuracy, thus ensuring the economic rationality of the audit process.

[0026] Example 8: The process by which the performance feedback adjustment module corrects the judgment threshold of the task entropy evaluation module includes: A preset performance benchmark value is used to determine whether the dynamic resource performance ratio is lower than the performance benchmark value. If the dynamic resource performance ratio is lower than the performance benchmark value, the entropy value judgment threshold in the task entropy value evaluation module is increased to reduce the triggering frequency of high cognitive load tasks. If the dynamic resource efficiency ratio is higher than or equal to the efficiency benchmark value, the entropy value judgment threshold in the task entropy value assessment module will be maintained or lowered to increase the audit depth of potential risks. This embodiment is a further refinement of the feedback adjustment mechanism based on Embodiment 7. The system presets a performance benchmark value, which represents the expected average input-output ratio. The system compares the currently calculated dynamic resource performance ratio with the performance benchmark value. In response to a dynamic resource performance ratio lower than the performance benchmark value, the system determines that the current input is too large and the output is insufficient, performs a desensitization operation, and increases the entropy threshold in the task entropy evaluation module. The formula is as follows: This reduces the triggering frequency of high cognitive load tasks; In response to a dynamic resource efficiency ratio higher than or equal to the efficiency benchmark, the system determines that the current efficiency is good or there is a risk of missed detection, and performs a sensitivity enhancement operation to maintain or lower the entropy value judgment threshold. The formula is as follows: This allows more critical tasks to be included in the high-load audit mode; among them To adjust the step size; To eliminate implementation uncertainties caused by the unclear definition of the adjustment step size, this embodiment explicitly defines its setting logic: adjustment step size and It can be set as a fixed constant, with a preferred numerical range of [value missing]. For example, a value of 0.05 can be used to ensure the stability of the system threshold evolution; alternatively, it can be set as a dynamic variable positively correlated with the performance deviation, with the calculation logic as follows: ,in This is a preset gain adjustment coefficient, for example, 0.1. This represents the current dynamic resource efficiency ratio. This serves as a performance benchmark value, enabling adaptive control that allows for rapid correction when performance deviates significantly and fine-tuning when performance deviates slightly. Regarding the strategy of choosing which approach to use under what circumstances, the system incorporates adaptive switching logic based on system load: the system defaults to using dynamic step-size calculation to achieve optimal convergence performance; when the system's real-time monitoring module detects that the CPU load exceeds... In order to reduce the overhead of floating-point operations, the system will automatically switch to fixed-step mode when the system is in a high-concurrency auditing peak period. To further clarify the physical sources of key parameters to meet implementation requirements, the system defines the acquisition methods for performance benchmark values ​​and adjustment gain coefficients as follows: Performance benchmark value Statistical analysis derived from historical data, specifically the process of selecting data from before the system went live. Months, for example The system uses data from manual audit projects to calculate the average return on investment (ROI), which is the ratio of the amount of loss recovered manually to the cost of manual labor hours. This ROI is then used as the minimum acceptable standard for the system. This ensures that the economic benefits of intelligent agent systems are no less than those of traditional manual methods; Adjusting the gain coefficient Derived from offline simulation training of control systems, the system constructs a virtual simulation environment containing historical task flows, using threshold convergence time and oscillation amplitude as joint optimization objectives, and employs gradient descent method... The optimal solution is searched within the parameter space, or, when training computational power is lacking, it is set based on engineering experience. This value corresponds to a typical underdamped response characteristic in control theory, which can ensure the adjustment sensitivity while avoiding system instability caused by a single outlier. This embodiment endows the system with evolutionary capabilities similar to those of a living organism through closed-loop feedback adjustment based on efficiency ratio. In the long-term operation of communication engineering auditing, the system can automatically switch between two modes: extensive and efficient and intensive and in-depth analysis, depending on the actual situation during peak business periods or low-risk periods. This dynamic balancing mechanism ensures that the system always maintains the optimal balance between audit revenue and computing power costs throughout its entire life cycle.

[0027] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-task allocation and decision optimization system for intelligent agents in communication engineering auditing, characterized in that, Including the cloud, the cloud communication connection includes a data perception and analysis module, a task entropy value evaluation module, a cognitive resource adaptation module, a decision confidence monitoring module, and an efficiency feedback adjustment module; The data perception and parsing module is used to acquire settlement documents for communication engineering and unstructured field data, extract features from the unstructured field data, and generate task feature vectors. The task entropy evaluation module is used to calculate the task information entropy value based on the task feature vector, and classify the audit task into low cognitive load task or high cognitive load task according to the task information entropy value. The cognitive resource adaptation module is used to build an agent capability model, calculate the cognitive adaptation coefficient between the agent capability model and the task feature vector, and dynamically build an agent collaborative topology and assign audit tasks based on the cognitive adaptation coefficient and task classification results. The process of calculating the cognitive adaptation coefficient by the cognitive resource adaptation module includes: constructing an agent capability feature vector containing logical verification capability, cost estimation capability, and image and text recognition capability; calculating the cosine similarity between the agent capability feature vector and the task feature vector, and using the cosine similarity as the cognitive adaptation coefficient, where the cognitive adaptation coefficient characterizes the degree of matching of the agent in processing the current audit task; The decision confidence monitoring module is used to track the audit process of the intelligent agent collaborative topology in real time, calculate the decision confidence volatility, and execute automated audit termination operations or trigger manual intervention operations based on the decision confidence volatility. The performance feedback adjustment module is used to calculate the dynamic resource performance ratio based on the audit results, and to adjust the judgment threshold of the task entropy evaluation module based on the dynamic resource performance ratio.

2. The multi-task allocation and decision optimization system for a communication engineering auditing intelligent agent according to claim 1, characterized in that, The data perception and parsing module performs feature extraction on unstructured field data, including the following processes: Acquire image data, visa document text, and engineering drawing data from unstructured field data; The semantic ambiguity of the data is calculated by the reciprocal of the average confidence level of optical character recognition, and the missing rate of key elements is calculated based on the missing proportion of required fields in the document. The Euclidean distance between the current data features and the centroids of the historical baseline patterns is calculated as the deviation feature. Combined with semantic ambiguity, key element missing rate and deviation feature, a task feature vector representing the uncertainty of the current audit task is generated.

3. The multi-task allocation and decision optimization system for a communication engineering auditing intelligent agent according to claim 2, characterized in that, The process by which the task entropy evaluation module classifies audit tasks into low cognitive load tasks or high cognitive load tasks includes: Preset entropy threshold; The task feature vector is normalized to obtain the probability distribution of feature elements, and the Shannon entropy of the probability distribution of the feature elements is calculated as the task information entropy value. The task information entropy value is compared with the entropy value judgment threshold. If the task information entropy value is less than the entropy value judgment threshold, the audit task is marked as a low cognitive load task. If the task information entropy value is greater than or equal to the entropy value judgment threshold, the audit task will be marked as a high cognitive load task.

4. The multi-task allocation and decision optimization system for a communication engineering auditing intelligent agent according to claim 3, characterized in that, The process by which the cognitive resource adaptation module dynamically constructs the agent collaborative topology and assigns auditing tasks includes: When an audit task is marked as a low cognitive load task, the agent capability model is traversed, the single agent with the highest cognitive fit coefficient is selected, a single-node decision topology is established, and the audit task is assigned to the single agent. When an audit task is marked as a high cognitive load task, multiple agents with a cognitive fit coefficient greater than a preset collaborative admission threshold are selected, a multi-node collaborative debate topology is established, and the audit task is assigned to the multi-node collaborative debate topology.

5. The multi-task allocation and decision optimization system for a communication engineering auditing intelligent agent according to claim 4, characterized in that, The process by which the decision confidence monitoring module performs automated audit termination operations or triggers manual intervention operations based on the decision confidence volatility includes: The quantitative intermediate audit amount output by the intelligent agent collaborative topology during the multi-round audit inference process is obtained, the variance of the quantitative intermediate audit amount is calculated, and the decision confidence volatility is generated. A preset volatility safety threshold is set. If the decision confidence volatility is less than the volatility safety threshold, the audit results are determined to be converged, the final audit conclusion is output, and an automated audit termination operation is performed. If the decision confidence volatility is greater than or equal to the volatility safety threshold, the audit results are determined to have a high-risk discrepancy. The current agent collaboration topology is suspended, and an abnormal warning signal is generated to trigger manual intervention.

6. The multi-task allocation and decision optimization system for a communication engineering auditing intelligent agent according to claim 5, characterized in that, The process by which the performance feedback adjustment module calculates the dynamic resource efficiency ratio includes: Obtain the risk correction amount identified in this audit task, and obtain the computing resource cost and time cost consumed in this audit task; Based on the preset performance calculation logic, the risk correction amount is divided by the weighted sum of the calculated resource cost and time cost to obtain the dynamic resource performance ratio.

7. The multi-task allocation and decision optimization system for a communication engineering auditing intelligent agent according to claim 6, characterized in that, The process by which the performance feedback adjustment module corrects the judgment threshold of the task entropy evaluation module includes: A preset performance benchmark value is used to determine whether the dynamic resource efficiency ratio is lower than the performance benchmark value. If the dynamic resource efficiency ratio is lower than the efficiency benchmark, the entropy threshold in the task entropy evaluation module will be increased to reduce the triggering frequency of high cognitive load tasks. If the dynamic resource efficiency ratio is higher than or equal to the efficiency benchmark value, the entropy value judgment threshold in the task entropy value assessment module will be maintained or lowered to increase the audit depth of potential risks.

Images