Cigarette retail license full-life-cycle intelligent supervision method

Abstract

The invention discloses a cigarette retail license full-life-cycle intelligent supervision method and a cigarette retail license full-life-cycle intelligent supervision system. The method comprises the following steps: collecting full life cycle data of cigarette retail licenses and carrying out standardized cleaning; constructing a semantic feature vector by utilizing a text embedding model; constructing a teacher-student distillation framework based on a large language model, and training a risk prediction model by adopting a multi-task strategy; calculating a dynamic threshold value by using context features, and carrying out hierarchical risk identification on an access stage, a use stage and an exit stage; and triggering a corresponding supervision action according to the comprehensive risk level. According to the invention, by introducing thinking chain reasoning and an adaptive threshold mechanism, the problems of low coverage rate, high false alarm rate and large manual dependence degree of traditional supervision are effectively solved, full-process intelligent physical examination and differentiated closed-loop disposal are realized, and the administrative permission supervision efficiency is remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of information processing and intelligent supervision technology, and in particular to an intelligent supervision method for the entire lifecycle of cigarette retail licenses. Background Technology

[0002] The tobacco retail license is a core legal document for the state to implement access control over the tobacco retail market, maintain market order, and safeguard national fiscal revenue and the legitimate rights and interests of consumers. With the deepening of the "streamlining administration, delegating power, and improving services" reform and the surge in the number of market entities, the traditional administrative licensing and regulatory model is facing unprecedented challenges. Under the existing technical and management framework, there are several significant deficiencies that severely restrict the improvement of regulatory effectiveness and bring potential legal and integrity risks.

[0003] First, the regulatory coverage is narrow, with numerous blind spots. In the traditional regulatory system, provincial and prefectural tobacco monopoly bureaus mainly adopt a "post-event random inspection" model for handling administrative licenses at the county level. Limited by human resources and technical means, the inspection rate of this vertical supervision is usually less than 10%. This means that more than 90% of administrative licensing cases are essentially "offline," lacking effective review by higher-level authorities.

[0004] Secondly, the system is highly reliant on manual labor, resulting in low efficiency. The existing "two comparisons" (comparing the number of publicly disclosed resources with the actual number of applications processed, and comparing the layout of publicized resources with the actual processing results) and document verification work primarily rely on regulatory personnel manually reviewing paper files and logging into multiple business systems for comparison. According to actual working hours statistics, completing a comprehensive and standardized verification of cigarette retail license management data takes an average of 547.57 minutes (approximately 9 hours). Such high time costs make high-frequency, comprehensive supervision virtually impossible under the current model. Regulatory personnel are wasting a significant amount of energy on tedious data handling and comparison rather than on substantive risk assessment and handling.

[0005] Third, risk identification is lagging and lacks dynamic perception capabilities. Traditional supervision is mostly based on fixed rule engines (such as simple expiration reminders), lacking the ability to identify complex risk patterns. For example, the judgment of "abnormal operation" often relies solely on abnormal sales or reports and complaints after the fact. For highly concealed and logically complex risk behaviors such as "market inspectors staying at merchants for too short a time but reporting normal inspections" (potentially indicating false inspections) or "frequent changes of operators within a short period of time without changing the address" (potentially indicating illegal transfers), there is a lack of real-time correlation analysis and dynamic perception methods based on multi-dimensional data. Fixed threshold judgments (such as "must be greater than 5 minutes") often generate a large number of false alarms or omissions when facing different geographical environments (such as inaccessible mountainous areas and compact urban business districts).

[0006] Fourth, legal and integrity risks coexist. Due to the lack of a unified, intelligent verification tool, law enforcement personnel in different regions objectively differ in their understanding and application of laws and regulations. This can lead to different approval results for the same application in different areas, triggering administrative review and litigation risks. At the same time, the manual operation process has excessive discretionary power and lacks rigid constraints at the system level.

[0007] In response to the national call for "efficiently accomplishing one task" and the new situation of advancing the modernization of the national governance system and governance capabilities, there is an urgent need to introduce advanced intelligent technologies and innovate regulatory methods. Based on this background, this invention proposes to utilize Large Language Model (LLM) distillation technology and dynamic threshold algorithms to construct an intelligent regulatory system covering the entire lifecycle of licenses, thereby achieving a leap from "manual spot checks" to "intelligent full inspection" and solving the aforementioned pain points. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent supervision method for the entire lifecycle of cigarette retail licenses, so as to solve the technical problems in the prior art.

[0009] To achieve the above objectives,

[0010] On the one hand, the present invention provides a method for intelligent supervision of the entire lifecycle of cigarette retail licenses, the method comprising the following steps:

[0011] Step S1: Collect the full lifecycle data of cigarette retail licenses through a preset data interface layer; the full lifecycle data includes data from the entry stage, the usage stage, and the exit stage.

[0012] Multiple imputation is used to fill in missing values ​​in the collected data, unstructured text data is cleaned, and label coding technology is used to convert discrete categorical variables into numerical codes to construct a standardized feature dataset.

[0013] Step S2: Extract features from the preprocessed data, including time interval features, statistical features, spatial distribution features, and compliance features;

[0014] The pre-trained text embedding model is invoked to vectorize the text features of the company name, business address, and case description, mapping them into feature vectors in a high-dimensional semantic vector space. To quantify the semantic similarity between texts;

[0015] Step S3: Construct a knowledge distillation architecture that includes a teacher model and a student model; select a large-parameter, large-language model with thought chain reasoning ability as the teacher model, and select a lightweight large-language model as the student model;

[0016] A fine-tuning dataset of instructions is constructed based on historical regulatory data, and a multi-task learning strategy is used to distill the student model for training; the multi-task learning includes risk level classification task and optimal judgment threshold regression task.

[0017] Define a multi-task loss function The loss function is composed of risk prediction cross-entropy loss. Threshold prediction mean square error loss and distillation loss based on KL divergence Weighted composition;

[0018] Using the trained student model, targeting the input feature vector set Output the risk probability distribution of the current regulated object and the dynamic threshold parameters adapted to the current context.

[0019] Step S4: Based on the feature vector set With dynamic threshold parameters, access verification, usage supervision screening and exit cloud inspection risk identification are performed respectively, and the comprehensive risk level is determined based on the identification results;

[0020] Step S5: Utilize the dynamic threshold function The threshold values ​​in step S4 are corrected in real time, whereby... This is the context feature vector of the current regulatory scenario. These are the tuning coefficients output by the model;

[0021] Based on the final determined comprehensive risk level, corresponding regulatory actions are automatically triggered through the Agent orchestration module. Further, the multi-task loss function in step S3... The mathematical expression is: ,

[0022] in, For hyperparameter weights; Kullback-Leibler divergence was used to calculate the difference between the output probability distribution of the teacher model and the output probability distribution of the student model in order to transfer the reasoning ability of the teacher model.

[0023] During the distillation training process, a course learning strategy is adopted, and training is carried out in order of increasing sample difficulty, and an adversarial training mechanism is introduced.

[0024] Furthermore, the dynamic threshold function in step S5 The calculation logic is as follows:

[0025] For a given variable to be decided and context features Calculate historical data in context The mean of and standard deviation ;

[0026] Combined with the model prediction tuning coefficients Calculate dynamic threshold ;

[0027] If the variable value Then determine the risk label. It is 1 if it is true, otherwise it is 0;

[0028] The objective function for threshold optimization is to maximize... Fraction: ,

[0029] in For precision, To achieve recall, The parameters used to adjust the recall weight.

[0030] Furthermore, in step S4, regulatory screening is used, including an orange risk model targeting information mismatch. The calculation formula is: ,

[0031] in, For the certificate holder, As the legal representative; These are the semantic vectors of the company name and the system matching name after being embedded by the text embedding model, respectively. These are semantic vectors for the business address and the corporate address, respectively. The cosine similarity function; This is the preset similarity threshold.

[0032] Furthermore, in step S4, regulatory screening is used, including a yellow risk model targeting anomalies in the inspection logic. The calculation formula is: ,

[0033] in, The duration of the inspection personnel's stay is calculated based on location data; The results of the inspection to be submitted; The minimum time threshold is calculated based on the dynamic threshold function described in step S5.

[0034] On the other hand, it provides an intelligent monitoring system for the entire lifecycle of cigarette retail licenses, including:

[0035] The data fusion layer is used to connect the business management system and external data platform through API interfaces to perform data extraction, transformation and loading.

[0036] The intelligent computing engine, deployed on a computing server, runs a distilled large language model and data analysis scripts based on a deep learning framework; the engine is responsible for performing feature vectorization, risk model inference, and dynamic threshold calculation.

[0037] The Agent Orchestration Control Center integrates the Model Context Protocol to orchestrate the collaborative work of data retrieval agents, risk analysis agents, and report generation agents.

[0038] The full lifecycle risk monitoring terminal provides a visual interactive interface, displaying a graded risk map and individual household risk profiles in real time;

[0039] The cloud-based inspection and interaction module is used to push self-service verification requests to low-risk users and receive video or image evidence uploaded by users.

[0040] Furthermore, the intelligent computing engine runs in a containerized environment to achieve isolation and rapid deployment of the computing environment; the system backend uses a relational database to store structured license files and risk assessment records, and uses a scientific computing library to perform matrix operations and statistical analysis.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. By introducing a socially guided reverse learning mechanism, the location information of high-level agents is used as a guide for low-level agents to generate reverse points, making the exploration behavior more intelligent and directional, thereby enabling more efficient collaborative search in the complex OPF solution space.

[0043] 2. By introducing fitness-weighted elite centroids as the guiding source and dynamically adjusting the temperature coefficient in the weight calculation, the robustness of the guiding signal is enhanced. In particular, when dealing with problems such as OPF where there are a large number of deceptive local optima, the algorithm's global optimization ability is significantly improved.

[0044] 3. By employing density-based clustering to identify elite subgroups and synthesizing guiding information based on subgroup centers, the algorithm can effectively address multimodal optimization problems (e.g., it can identify multiple high-quality solutions with similar costs but different operating modes in the optimal power flow problem of a power system), fundamentally avoiding the defect that a single centroid may fall in a low-value region.

[0045] 4. By implementing adaptive determination of DBSCAN parameters or adopting the more advanced HDBSCAN algorithm, the algorithm's dependence on manual parameter tuning is reduced, its adaptability to different problems (such as power grids of different sizes and characteristics) is enhanced, and the degree of automation and robustness of the algorithm are improved. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall process of the improved invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the specific embodiments of this invention will be described in detail below with reference to the accompanying drawings. This invention can be executed by a computer (such as a server, personal computer, embedded device, etc.), which includes hardware such as a processor and memory. The method steps can be executed by the processor as computer program instructions stored in the memory. To better illustrate the core idea of ​​this invention, the following embodiments will use the typical complex engineering optimization problem of optimal power flow (OPF) in power systems as a background example.

[0048] Example 1

[0049] This embodiment details the specific process of data preprocessing and feature construction in the intelligent supervision method.

[0050] Step S1: Multi-source heterogeneous data acquisition and standardized cleaning

[0051] The system collects application and public notice data during the access phase, daily operation and regulatory inspection data during the usage phase, and credit evaluation and deregistration data during the exit phase through a preset data interface layer.

[0052] To address the unavoidable missing values ​​in the collected data, this embodiment employs multiple imputation to fill in the missing values. Compared to traditional mean imputation, this method better preserves the data distribution characteristics. For unstructured text data (such as case descriptions), regular expressions are used for cleaning to remove special symbols and irrelevant characters. Label encoding technology is used to convert discrete categorical variables (such as "business type" and "commercial district type") into numerical codes, constructing a standardized feature dataset.

[0053] Step S2: Feature Engineering Construction Based on Semantic Embedding

[0054] Deep feature extraction is performed on the preprocessed data, covering features such as time interval features, historical statistical features, spatial distribution features based on geographical location information, and compliance features based on business logic. To enable the computer to understand the meaning of the text, the system calls a pre-trained text embedding model to vectorize textual features such as company names, business addresses, and case descriptions. In this embodiment, the Nomic-embedded-text model is specifically selected. This model maps text into feature vectors in a high-dimensional semantic vector space. This allows for the quantitative calculation of semantic similarity between texts. This is crucial for solving the problem of "inconsistent address spellings but actually the same location".

[0055] The following details the distillation process of the DeepSeek-R1 model and the mathematical principles of the dynamic threshold algorithm.

[0056] 1. Construct a distillation risk prediction model based on a large language model (Step S3)

[0057] To balance reasoning ability and operational efficiency, this invention constructs a knowledge distillation architecture comprising a teacher model and a student model. A large-parameter, large-language model with thought chain reasoning capabilities is selected as the teacher model (specifically, DeepSeek-R1 70B is used in this embodiment); a lightweight large-language model is selected as the student model (specifically, a quantized Qwen or Llama series model is used in this embodiment, deployed via Ollama). A fine-tuning dataset of instructions is constructed based on historical regulatory data, and a multi-task learning strategy is employed to distill and train the student model. The student model must simultaneously complete two tasks: risk level classification and optimal judgment threshold regression for the current scenario. To effectively train this multi-task model, a loss function is defined. as follows: ,in, These are hyperparameter weights used to balance the impact of various losses. Cross-entropy loss in risk prediction. Mean squared error loss for threshold prediction. Distillation loss, calculated based on Kullback-Leibler (KL) divergence, is used to measure the difference between the output probability distribution of the teacher model and the output probability distribution of the student model, thereby transferring the reasoning ability of the teacher model. During distillation training, a curriculum learning strategy is employed, training is conducted in order of increasing sample difficulty, and an adversarial training mechanism is introduced to add small perturbations to the input samples to enhance the model's robustness.

[0058] 2. Dynamic threshold adjustment algorithm

[0059] Traditional fixed thresholds are prone to false alarms. This invention utilizes a dynamic threshold function. Each judgment threshold is corrected in real time.

[0060] For a given variable to be decided (e.g., inspection duration) and contextual features (e.g., region, time period) Calculate historical data in context The mean of and standard deviation The optimization coefficients are combined with the model predictions. Calculate the dynamic threshold: ,

[0061] If the variable value Then determine the risk label. It is 1 if it is true, otherwise it is 0.

[0062] • Optimization objective: tuning coefficients The learning goal is to maximize Scores, to balance precision and recall: ,

[0063] in This is a parameter used to adjust the recall weight. For example, during a crackdown, it can be adjusted... This can make the model more inclined to recall more potential risks.

[0064] The following illustrates the specific risk calculation logic in the three stages of admission, use, and exit.

[0065] 1. Access Verification Stage

[0066] The system is based on feature vector sets Calculate the overall risk level of a newly issued license:

[0067] Red Risk: Calculating the Application Date With the date of public announcement If the absolute value of the time difference exceeds the inconsistency threshold... This triggers a red risk level.

[0068] Orange Risk: Resource Saturation Based on Level 3 Cells With resource threshold The comparison results.

[0069] Yellow Risk: Based on filing date With decision date Time interval and process time threshold The comparison results.

[0070] Blue Risk: Duplicate Counts and Duplicate Thresholds Based on Company Name in the System The comparison results.

[0071] 2. Use of regulatory screening phase

[0072] This phase focuses on identifying covert violations, specifically including:

[0073] Orange Risk (Information Mismatch): Addressing potential issues such as "mismatch between person and document" or "mismatch between document and address," the orange risk model... The calculation formula is: ,

[0074] in, For the certificate holder, As the legal representative; These are the semantic vectors of the company name and the system matching name after being embedded using Nomic-embed-text, respectively. These are semantic vectors for the business address and the corporate address, respectively. The cosine similarity function; The preset similarity threshold (e.g., 0.85).

[0075] • Yellow Risk (Logical Anomaly): This model addresses fraudulent inspection practices. The calculation formula is: , in, The duration of the inspector's stay at the merchant's premises, calculated based on GPS location data; The results of the inspection to be submitted; This is the minimum time threshold for the current scenario, calculated based on the aforementioned dynamic threshold function.

[0076] 3. Risk Identification for Exiting Cloud-Based Inspection

[0077] Red Risk: Identify seriously dishonest merchants with a credit rating of D.

[0078] Orange Risk: Identify merchants that have stopped supplying goods, have incomplete information, or are associated with individuals involved in the case.

[0079] Yellow Risk: Identify abnormal merchants with a credit rating of A but not included in the normal inspection list.

[0080] 4. Tiered early warning response

[0081] Based on the final determined risk level, the Agent orchestration module automatically triggers regulatory actions: for red risks, it triggers a blocking process; for orange risks, it triggers a key inspection; and for yellow and blue risks, it triggers a routine inspection or cloud-based inspection process.

[0082] Example 2:

[0083] This embodiment constructs an intelligent monitoring system for the entire lifecycle of cigarette retail licenses based on a "cloud-edge-device" collaborative architecture. The system implements five core module functions and employs containerization and database technologies.

[0084] 1. System Logical Architecture

[0085] This system comprises the following five core layers:

[0086] Data Fusion Layer: This layer connects to the franchise management subsystem, cloud collaboration platform (in this embodiment, WPS Enterprise Cloud Collaboration Platform is specifically selected), third-party business data platform (such as Tianyancha), and administrative penalty case system via API interfaces. This layer performs data extraction, transformation, and loading (ETL), unifying the access of multi-source heterogeneous data.

[0087] Intelligent Computing Engine: This is the core brain of the system, deployed on a GPU-accelerated computing server. It runs distilled large language models based on the deep learning framework (PyTorch) and Python data analysis scripts. This engine is responsible for performing feature vectorization, risk model inference, and dynamic threshold calculation.

[0088] Agent Orchestration Control Center: Integrating Model Context Protocol (MCP), it orchestrates the collaborative work of data retrieval agents, risk analysis agents, and report generation agents to automate the workflow of complex tasks. This embodiment preferably uses the Cherry Studio framework as the orchestration hub for intelligent agents.

[0089] Full lifecycle risk monitoring terminal: Provides a visual interactive interface, displaying real-time red, orange, yellow, and blue risk maps and individual risk profiles for regulatory personnel to view. The front-end interface in this embodiment is developed using the PyQt5 framework.

[0090] Cloud-based inspection and verification interaction module: It connects with WeChat mini programs or government service apps to push self-service verification requests to low-risk users and receive video or image evidence uploaded by users, achieving "zero-visit" supervision.

[0091] 2. Hardware and software environment configuration

[0092] To support the above architecture, especially the efficient operation of the intelligent computing engine, this embodiment adopts the following specific configuration:

[0093] Hardware Environment: The system's core computing nodes are configured with high-performance GPU servers. To ensure training stability, the server environment requires strict temperature control, with CPU operating temperature maintained at 100°C. Below C, the GPU temperature is controlled at 120 degrees Celsius. Below C. The storage layer uses an SSD array for high-speed read and write of massive amounts of unstructured documents and structured regulatory records.

[0094] Software Technology Stack and Containerization: Operating System and Containerization: Linux (Ubuntu 22.04 LTS) is used as the base operating system. The system fully adopts Docker containerization technology. Core computing services, database services, and web services are all encapsulated in independent Docker containers to achieve isolation of the computing environment and rapid deployment, ensuring that the system response time is controlled within 20 seconds.

[0095] Database: The system backend uses a MySQL relational database to store structured license files and risk assessment records. Simultaneously, scientific computing libraries such as Pandas and NumPy are used for large-scale matrix operations and statistical analysis.

[0096] Experimental results:

[0097] Table 1: Statistics on System Trial Operation

[0098]

[0099] Table 2: Time Consumption Table for Intelligent Verification of Tobacco Retail License Lifecycle Management Data

[0100]

[0101] Table 3: Statistics on the underreporting rate and false reporting rate of data verification data in four county bureaus from April to June

[0102]

[0103] Table 4: Statistics on the Total Time Spent on Data Verification for the Entire Lifecycle Management of Cigarette Retail Licenses Before and After the Activity

[0104]

[0105] Table 5: Calculation of Work Efficiency

[0106]

[0107] Table 6: Comparison of Fuel Consumption Costs

[0108]

[0109] The team compared the fuel consumption costs of on-site inspections conducted by the Cengong County Bureau from January to March 2025 before the campaign with those after the campaign from July to September 2025. They found that before the campaign, the total fuel consumption for on-site inspections by law enforcement officers in one quarter was approximately 1722 liters. Factors such as traffic conditions could increase the time and fuel consumption. After the campaign, law enforcement officers used the verification system to screen retailers with lower risk levels, allowing for direct "cloud inspection." With a similar number of regulated retailers, the total fuel consumption for on-site inspections in one quarter decreased to 558 liters. Considering the market fuel price of approximately 7.2 yuan / liter, the cost savings for one quarter were 8380.8 yuan, averaging 70.73 yuan per retailer. Based on the 10,898 cigarette retail licenses processed in Qiandongnan Prefecture in 2024, the annual fuel cost savings could reach 770,815.54 yuan, significantly reducing economic costs.

[0110] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent supervision of the entire lifecycle of cigarette retail licenses, characterized in that: The method includes the following steps: Step S1: Collect the full lifecycle data of cigarette retail licenses through a preset data interface layer; the full lifecycle data includes data from the entry stage, the usage stage, and the exit stage. Multiple imputation is used to fill in missing values ​​in the collected data, unstructured text data is cleaned, and label coding technology is used to convert discrete categorical variables into numerical codes to construct a standardized feature dataset. Step S2: Extract features from the preprocessed data, including time interval features, statistical features, spatial distribution features, and compliance features; The pre-trained text embedding model is invoked to vectorize the text features of the company name, business address, and case description, mapping them into feature vectors in a high-dimensional semantic vector space. To quantify the semantic similarity between texts; Step S3: Construct a knowledge distillation architecture that includes a teacher model and a student model; select a large-parameter, large-language model with thought chain reasoning ability as the teacher model, and select a lightweight large-language model as the student model; A fine-tuning dataset of instructions is constructed based on historical regulatory data, and a multi-task learning strategy is used to distill the student model for training; the multi-task learning includes risk level classification task and optimal judgment threshold regression task. Define a multi-task loss function The loss function is composed of risk prediction cross-entropy loss. Threshold prediction mean square error loss and distillation loss based on KL divergence Weighted composition; Using the trained student model, targeting the input feature vector set Output the risk probability distribution of the current regulated object and the dynamic threshold parameters adapted to the current context. Step S4: Based on the feature vector set With dynamic threshold parameters, access verification, usage supervision screening and exit cloud inspection risk identification are performed respectively, and the comprehensive risk level is determined based on the identification results; Step S5: Utilize the dynamic threshold function The threshold values ​​in step S4 are corrected in real time, whereby... This is the context feature vector of the current regulatory scenario. These are the tuning coefficients output by the model; Based on the final determined comprehensive risk level, the corresponding regulatory actions are automatically triggered through the Agent orchestration module.

2. The method according to claim 1, characterized in that, The multi-task loss function in step S3 The mathematical expression is: , in, For hyperparameter weights; Kullback-Leibler divergence was used to calculate the difference between the output probability distribution of the teacher model and the output probability distribution of the student model in order to transfer the reasoning ability of the teacher model. During the distillation training process, a course learning strategy is adopted, and training is carried out in order of increasing sample difficulty, and an adversarial training mechanism is introduced.

3. The method according to claim 1, characterized in that, The dynamic threshold function in step S5 The calculation logic is as follows: For a given variable to be decided and context features Calculate historical data in context The mean of and standard deviation ; Combined with the model prediction tuning coefficients Calculate dynamic threshold ; If the variable value Then determine the risk label. It is 1 if it is true, otherwise it is 0; The objective function for threshold optimization is to maximize... Fraction: , in For precision, To achieve recall, The parameters used to adjust the recall weight.

4. The method according to claim 1, characterized in that, In step S4, regulatory screening is used, including an orange risk model for information mismatch. The calculation formula is: , in, For the certificate holder, As the legal representative; These are the semantic vectors of the company name and the system matching name after being embedded by the text embedding model, respectively. These are semantic vectors for the business address and the corporate address, respectively. The cosine similarity function; This is the preset similarity threshold.

5. The method according to claim 1, characterized in that, In step S4, regulatory screening is used, including a yellow risk model targeting abnormal inspection logic. The calculation formula is: , in, The duration of the inspection personnel's stay is calculated based on location data; The results of the inspection to be submitted; The minimum time threshold is calculated based on the dynamic threshold function described in step S5.

6. A smart monitoring system for the entire lifecycle of cigarette retail licenses, characterized in that: include: The data fusion layer is used to connect the business management system and external data platform through API interfaces to perform data extraction, transformation and loading. The intelligent computing engine, deployed on a computing server, runs a distilled large language model and data analysis scripts based on a deep learning framework; the engine is responsible for performing feature vectorization, risk model inference, and dynamic threshold calculation. The Agent Orchestration Control Center integrates the Model Context Protocol to orchestrate the collaborative work of data retrieval agents, risk analysis agents, and report generation agents. The full lifecycle risk monitoring terminal provides a visual interactive interface, displaying a graded risk map and individual household risk profiles in real time; The cloud-based inspection and interaction module is used to push self-service verification requests to low-risk users and receive video or image evidence uploaded by users.

7. The system according to claim 6, characterized in that, The intelligent computing engine runs in a containerized environment to achieve isolation and rapid deployment of the computing environment; the system backend uses a relational database to store structured license files and risk assessment records, and uses a scientific computing library to perform matrix operations and statistical analysis.

Images