Vertical model-based lease risk data processing method and system

Abstract

The invention relates to the technical field of data processing, and discloses a lease risk data processing method and system based on a vertical model, and the method comprises the steps: integrating the related information of a target lease project into structural data; performing supervision fine tuning on the initial language model to obtain a risk perception vertical model; initializing different parameter subsets of the risk perception vertical model to obtain a performance capability evaluation sub-model, a lease item value evaluation sub-model and an external environment evaluation sub-model, and fusing the models into a risk evaluation network; inputting the input features of the structured data into a risk assessment network to obtain a preliminary performance risk score, a preliminary asset risk score and a preliminary environment risk score, and performing weighted fusion to obtain a comprehensive risk score and a risk disposal suggestion; updating the supervision fine tuning data set to obtain an optimized risk perception vertical model; according to the invention, the efficiency of lease risk data processing based on the vertical model can be improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for processing rental risk data based on a vertical model. Background Technology

[0002] In the current process of processing leasing risk data, the low efficiency of integrating multi-source heterogeneous data is a prominent problem. Information on lessees, leased property status, performance history, and external market environment related to leasing transactions are scattered across different systems. These data formats differ significantly and suffer from redundancy and missing data. Traditional data processing methods are cumbersome in the cleaning, standardization, and correlation matching stages, making it difficult to quickly generate unified and standardized structured data. This results in an excessively long preparation period for risk assessment data, impacting overall processing efficiency.

[0003] Traditional risk assessment models lack deep adaptability to leasing scenarios, making it difficult to meet the needs of accurate risk assessment. Existing models mostly adopt a generalized architecture, failing to specifically model the core assessment dimensions of leasing risk, and thus unable to achieve refined analysis of different types of risks such as performance capability, leased asset value, and external environment. Furthermore, the model parameters are not rationally split and shared, and the weight allocation lacks a scientific quantitative basis, resulting in insufficient accuracy of the initial risk score fusion results. Consequently, reliable risk management recommendations cannot be generated, making it difficult to effectively support risk control decisions in leasing businesses. Therefore, improving the efficiency of leasing risk data processing has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a leasing risk data processing method and system based on a vertical model to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a leasing risk data processing method based on a vertical model, comprising: S1. Integrate the lessee information, leased property status information, performance history data, and external market environment data of the target leasing project into structured data of the target leasing project; S2. Construct a supervised fine-tuning dataset based on historical case data, and use the supervised fine-tuning dataset to perform supervised fine-tuning on the initial language model to obtain a risk perception vertical model for the target leasing project; S3. Initialize different parameter subsets of the risk perception vertical model to obtain the performance capability assessment sub-model, leased asset value assessment sub-model and external environment assessment sub-model of the target leasing project, and merge the performance capability assessment sub-model, leased asset value assessment model and external environment assessment model into a risk assessment network; S4. Input the input features of the structured data into the risk assessment network to obtain the preliminary performance risk score, preliminary asset risk score, and preliminary environmental risk score of the target lease project; S5. Based on the weight allocation vector corresponding to the target leasing project, the preliminary performance risk score, the preliminary asset risk score, and the preliminary environmental risk score are weighted and integrated to obtain a comprehensive risk score for the target leasing project, and risk disposal recommendations for the target leasing project are generated based on the comprehensive risk score. S6. Based on the comprehensive risk score, the weight allocation vector, and the risk handling suggestions, update the supervised fine-tuning dataset to obtain the optimized risk perception vertical model.

[0006] In a preferred embodiment, the integration of lessee information, leased asset status information, performance history data, and external market environment data of the target leasing project into structured data for the target leasing project includes: The system extracts lessee information from the business database, the status information of the leased property from the IoT monitoring platform, the performance history data from the contract management system, and the external market environment data from the market data interface. The lessee information, the leased property status information, the performance history data, and the external market environment data are cleaned and standardized. By associating and matching the standardized data using the rental project number, the structured data of the target rental project is obtained.

[0007] In a preferred embodiment, the step of constructing a supervised fine-tuning dataset based on historical case data, and using the supervised fine-tuning dataset to perform supervised fine-tuning on the initial language model to obtain a risk perception vertical model for the target leasing project, includes: From the historical case data, leasing project cases with clear risk labels were selected; Convert the aforementioned rental project case into question-and-answer pair format data; The question-and-answer pair formatted data is combined into a supervisory fine-tuning dataset for the target leasing project; The model parameters of the initial language model were adjusted using the supervised fine-tuning dataset and adapted to the rental risk assessment task. Save the adjusted model parameters to obtain the risk perception vertical model for the target leasing project.

[0008] In a preferred embodiment, initializing different subsets of parameters of the risk perception vertical model to obtain the target leasing project's performance capability assessment sub-model, leased asset value assessment sub-model, and external environment assessment sub-model includes: Based on the dimensions of rental risk assessment, the parameters of the risk perception vertical model are divided into three parameter subsets; The first parameter subset is initialized and fixed into a network structure dedicated to processing the lessee information and the performance history data, thus obtaining the performance capability assessment sub-model of the target leasing project. The second parameter subset is initialized and fixed into a network structure dedicated to processing the state information of the leased property, thus obtaining the leased property value assessment sub-model of the target leasing project; The third parameter subset is initialized and fixed into a network structure specifically for processing the external market environment data, thus obtaining the external environment assessment sub-model of the target leasing project; The shared parameter layer in the risk perception vertical model is retained, so that the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model share the underlying semantic understanding capability.

[0009] In a preferred embodiment, the fusion of the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model into a risk assessment network includes: Define a network convergence architecture; The outputs of the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model are respectively connected to different inputs of the feature fusion layer in the network fusion architecture. After the output of the feature fusion layer, a fully connected processing layer is added to the network fusion architecture; Through the feature fusion layer and the fully connected processing layer, the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model are logically integrated into a single network entity to obtain the risk assessment network of the target leasing project.

[0010] In a preferred embodiment, the step of simultaneously inputting the input features of the structured data into the risk assessment network to obtain a preliminary performance risk score, a preliminary asset risk score, and a preliminary environmental risk score for the target lease project includes: Extract input features corresponding to the target rental project from the structured data; In the performance capability assessment sub-model, the subset of features related to the lessee's credit and behavior in the input features is forward propagated to obtain the preliminary performance risk score of the target lease project; In the leased asset valuation sub-model, the subset of features related to the attributes and status of the leased asset in the input features is forward-propagated to obtain the preliminary asset risk score of the target leased project; in the external environment assessment sub-model, the subset of features related to macroeconomic and market fluctuations in the input features is forward-propagated to obtain the preliminary environmental risk score of the target leased project.

[0011] In a preferred embodiment, the weight allocation vector is calculated using the following formula: ; In the formula, Assign vectors to the weights. This is the preliminary performance risk score. This is the preliminary asset risk score. This is the preliminary environmental risk score. The standard deviation of the historical performance risk score series. The standard deviation of the historical asset risk score series. The standard deviation of the historical environmental risk score series. To assess the significance of overall risks, The sum of the significance of the combined risk across three dimensions. This is a weighted allocation vector corresponding to the initial performance risk score. This is the weight allocation vector corresponding to the initial asset risk score. This is the weight allocation vector corresponding to the preliminary environmental risk score.

[0012] In a preferred embodiment, generating risk management recommendations for the target leasing project based on the comprehensive risk score includes: Based on a preset comprehensive risk score threshold range, the comprehensive risk score of the target leasing project is mapped to the corresponding risk level; Based on the risk level obtained from the mapping, a basic risk management strategy corresponding to the risk level is matched from a preset risk management strategy library; Based on the basic risk management strategy, and combined with the specific risk score composition of the target lease project, extract the portion of any score in the preliminary performance risk score, the preliminary asset risk score, and the preliminary environmental risk score that exceeds its corresponding single threshold, and generate a risk enhancement prompt for the target lease project. By combining the basic risk management strategy with the risk enhancement prompts, risk management recommendations for the target leasing project are obtained.

[0013] In a preferred embodiment, updating the supervised fine-tuning dataset based on the comprehensive risk score, the weight allocation vector, and the risk management recommendations to obtain an optimized risk perception vertical model includes: The structured data of the target leasing project, the final generated comprehensive risk score, the weight allocation vector, and the risk management recommendations are integrated into a new historical case data entry. The validity of new historical case data entries is verified according to the preset verification logic; New historical case data entries will be added to the historical case database of the target leasing project upon verification. Based on the updated historical case database, the supervised fine-tuning dataset was reconstructed; The risk-aware vertical model is incrementally trained using the reconstructed supervised fine-tuning dataset to obtain an optimized risk-aware vertical model.

[0014] To address the aforementioned problems, the present invention also provides a rental risk data processing system based on a vertical model, the system comprising: The data integration and structuring module is used to integrate the lessee information, leased property status information, performance history data and external market environment data of the target leasing project into the structured data of the target leasing project; The vertical model training module is used to construct a supervised fine-tuning dataset based on historical case data, and to use the supervised fine-tuning dataset to perform supervised fine-tuning on the initial language model to obtain a risk perception vertical model for the target leasing project. The risk assessment network construction module is used to initialize different parameter subsets of the risk perception vertical model to obtain the performance capability assessment sub-model, leased asset value assessment sub-model and external environment assessment sub-model of the target leasing project, and to integrate the performance capability assessment sub-model, leased asset value assessment model and external environment assessment model into a risk assessment network. The multi-dimensional risk parallel assessment module is used to simultaneously input the input features of the structured data into the risk assessment network to obtain the preliminary performance risk score, preliminary asset risk score, and preliminary environmental risk score of the target leasing project. The comprehensive decision-making and output module is used to perform weighted fusion of the preliminary performance risk score, the preliminary asset risk score and the preliminary environmental risk score according to the weight allocation vector corresponding to the target leasing project, to obtain the comprehensive risk score of the target leasing project, and to generate risk disposal suggestions for the target leasing project based on the comprehensive risk score. An adaptive optimization module is used to update the supervised fine-tuning dataset based on the comprehensive risk score, the weight allocation vector, and the risk handling suggestions to obtain an optimized risk perception vertical model.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves efficient cleaning, standardization, and correlation matching of leasing business-related data through a standardized multi-source data integration process, forming unified structured data and significantly improving data utilization efficiency. Simultaneously, based on a supervised fine-tuning-built risk perception vertical model, through parameter subset splitting and specialized initialization, it forms assessment sub-models adapted to different risk dimensions, while retaining a shared parameter layer to ensure underlying semantic understanding capabilities. This makes risk analysis in each dimension more targeted, significantly improving the accuracy of leasing risk assessment and overall processing efficiency.

[0016] 2. This invention employs a scientifically quantified weight allocation formula to calculate the weight vector, achieving a reasonable weighted fusion of preliminary risk scores from different dimensions. This makes the comprehensive risk score more closely reflect the risk distribution characteristics of actual business scenarios. Risk management suggestions generated based on the comprehensive risk score and individual risk exceedance scenarios are both universal and targeted, providing direct and effective decision support for risk control. Furthermore, by updating the supervised fine-tuning dataset with new case data and performing incremental training, the risk perception vertical model is continuously optimized, constantly improving its adaptability and evaluation performance, ensuring its reliability and effectiveness in long-term use. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a leasing risk data processing method based on a vertical model, provided in an embodiment of the present invention; Figure 2 A functional block diagram of a leasing risk data processing system based on a vertical model, provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a rental risk data processing method based on a vertical model. The execution entity of this vertical model-based rental risk data processing method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application embodiment: a server, a terminal, etc. In other words, the vertical model-based rental risk data processing method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a leasing risk data processing method based on a vertical model, according to an embodiment of the present invention. In this embodiment, the leasing risk data processing method based on a vertical model includes: S1. Integrate the lessee information, leased property status information, performance history data, and external market environment data of the target leasing project into structured data of the target leasing project; In this embodiment of the invention, the integration of lessee information, leased property status information, performance history data, and external market environment data of the target leasing project into structured data of the target leasing project includes: The system extracts lessee information from the business database, the status information of the leased property from the IoT monitoring platform, the performance history data from the contract management system, and the external market environment data from the market data interface. The lessee information, the leased property status information, the performance history data, and the external market environment data are cleaned and standardized. By associating and matching the standardized data using the rental project number, the structured data of the target rental project is obtained.

[0021] When retrieving lessee information from the business database, the database stores all lessee-related data for all leasing projects in the form of structured data tables. These tables contain fixed fields such as lessee name, ID number, contact information, qualification level, historical leasing cooperation records, and credit rating. Each field is uniquely linked to a specific leasing project using a lease project number. The retrieval operation requires first identifying the lease project number of the target leasing project, then using the database's search function to input that lease project number as the search condition to locate the corresponding data row. Subsequently, the lessee-related fields in the data table are read one by one to accurately extract the specific information corresponding to each field. Record the information to ensure that the extracted lessee information completely matches the target leased project. When retrieving the status information of the leased property from the IoT monitoring platform, the platform will collect various operational and status data of the leased property in real time. This data is stored in categories according to the leased property number, and the stored content includes the real-time location of the leased property, operating parameters, wear and tear assessment results, historical maintenance records, current usage status, fault alarm information, etc. When retrieving the data, you need to enter the leased property number corresponding to the target leased project. Through the platform's data retrieval function, all status data of the leased property from the start of the lease to the day of retrieval are filtered out, and various status indicators of each time point are collected one by one in chronological order. To ensure the completeness and timeliness of leased asset status information; when retrieving historical performance data from the contract management system, which stores complete performance-related records for each lease project, including contract number, lease project number, performance milestone agreements, rent payment records, performance milestone completion status, acceptance results, and breach of contract records, the system uses the lease project number of the target lease project as the search keyword to filter out all performance records corresponding to that project. Each record is then read and organized to ensure no performance-related information is missed; when retrieving external market environment data from the market data interface, the market data... According to the interface, the authoritative industry data platform stores various market dynamic data related to the leasing business, including the average market rental price of similar leased items, the supply and demand ratio in different regions, industry policy changes, macroeconomic indicators, market competition among similar leasing projects, and the impact of raw material price fluctuations on the pricing of leased items. When extracting data, the interface needs to specify key information such as the type of leased item, the leasing area, and the leasing period of the target leasing project. The interface then filters out highly relevant market environment data based on this information, and then extracts and records these data one by one by category to ensure that the extracted external market environment data can accurately match the business scenario of the target leasing project.

[0022] When cleaning tenant information, first check for missing fields. If key fields such as ID number or contact information are found to be empty, supplementary information should be provided by checking the corresponding paper filing documents for the target rental project or contacting the tenant. If supplementary information cannot be provided, clearly mark it as "no relevant record". Next, check for duplicate data. If multiple duplicate records correspond to the same tenant, compare the field information of each record and retain the one with the most complete and up-to-date fields. Finally, verify the accuracy of the data, such as verifying whether the ID number format conforms to national standards, whether the contact information can be reached normally, and whether the qualification level is consistent with the filing information. Correct any errors. When standardizing tenant information, remove all spaces and special characters from the ID number and uniformly retain the ID number number format that conforms to national standards. Contact information is uniformly converted to a standard mobile phone number format. Qualification levels are labeled according to industry-standard terminology. Credit ratings are uniformly labeled using the industry-standard four-level credit rating system, which is divided into four levels: Excellent, Good, Satisfactory, and Needs Improvement. Excellent represents that the lessee has an outstanding credit record with no adverse credit history, and their willingness and ability to fulfill obligations are at the top level in the industry. Good represents that the lessee has a relatively good credit record with no major adverse credit history, and minor deviations in performance details can be corrected in a timely manner. Satisfactory represents that the lessee's credit record meets the standard, with a few minor adverse credit history records, but core performance can be guaranteed. Needs Improvement represents that the lessee's credit record does not meet the industry standard, with multiple adverse credit history records, and significant deficiencies in their ability and willingness to fulfill obligations. This ensures that the format of lessee information is uniform and standardized.

[0023] When cleaning the status information of the leased equipment, first check for abnormal data. For example, if the operating parameters exceed the normal range set by the equipment at the factory, or the location information does not match the agreed area of ​​use, it is necessary to verify the data by referring to the historical operating data and equipment calibration records of the leased equipment. If the data abnormality is caused by equipment failure, contact the maintenance personnel to confirm the actual status and correct the data. If the data transmission error is the cause, remove the abnormal data point, supplement the valid data of the adjacent time point, remove duplicate data, and retain only one record of the same status data collected at the same time point. Finally, fill in the missing data. If status data is missing for a certain period of time, combine the operating pattern of the leased equipment with the status of the preceding and following time points. Trend analysis, inference of reasonable data, and supplementation; when standardizing the status information of leased assets, operating parameters are retained to the appropriate decimal places according to industry data recording standards, location information is converted to latitude and longitude coordinates, and the degree of damage is marked according to the industry-recognized three-level damage classification standard. The three-level damage classification standard is slight damage, moderate damage, and severe damage. Slight damage means that the leased asset has no obvious damage to its appearance, no degradation of core performance indicators, and can be used normally without additional maintenance. Moderate damage means that the leased asset has localized appearance damage and a slight degradation of core performance indicators, and can be used normally after routine maintenance. Severe damage means that the leased asset has large-area damage. If the leased item suffers cosmetic damage or core component failure, resulting in a significant decline in core performance indicators and rendering it unusable, requiring major repairs or replacement of core components, its usage status is standardized according to the industry's four common usage status categories: Idle and Available for Rental, Normal Use, Maintenance and Repair, and Scrapped and Discontinued. Idle and Available for Rental refers to the leased item being awaited allocation after completing the previous rental process, with all performance indicators normal and ready for immediate rental. Normal Use refers to the leased item being used by the lessee within the rental period as agreed in the contract and in good operating condition. Maintenance and Repair refers to the leased item being temporarily out of use due to malfunctions, scheduled maintenance, or other reasons, and is in the repair or testing phase. Scrapped and Discontinued refers to the leased item reaching its service life or suffering damage. If the damage is beyond repair and the item has lost its usability, it should be scrapped. Fault alarm information is categorized and labeled according to the three fault levels in the industry standard: Level 1, Level 2, and Level 3. Level 1 faults are minor faults that only affect some non-core functions of the leased item and do not affect the overall normal use, requiring no emergency handling. Level 2 faults are moderate faults that affect some core functions of the leased item and restrict its overall use, requiring handling within a specified time. Level 3 faults are severe faults where the core functions of the leased item are completely disabled, making the entire item unusable and requiring immediate shutdown and emergency repairs. The format of the leased item status information must be consistent and comparable.

[0024] When cleaning historical performance data, verify whether the amounts in rent payment records are consistent with the contract and whether the payment times are accurate. This is done by comparing with bank transfer vouchers and tenant payment receipts, correcting data with incorrect amounts or times, removing duplicate performance node acceptance records, and retaining the final acceptance results. If the completion status of a performance node is missing, supplement it by reviewing the project manager's work log and acceptance report. When standardizing historical performance data, rent payment amounts are uniformly converted to RMB, and payment times are uniformly formatted as standardized dates and times. Performance compliance is marked according to the industry-standard three levels: fully compliant, basically compliant, and non-compliant. Full compliance means all performance nodes of the lease project are completed strictly according to the time and standards stipulated in the contract without any deviation. Basic compliance means the core performance nodes of the lease project are completed according to the contract. Minor deviations from contractual performance milestones that do not affect the overall progress of the leasing project are considered minor. Failure to meet standards refers to the failure to complete core performance milestones as stipulated in the contract, significantly impacting the overall progress of the leasing project and potentially leading to disputes. Breach of contract records are categorized into three levels of severity recognized in the industry: minor, moderate, and serious. A minor breach indicates minor deviations in performance that do not violate core contract terms and cause no economic loss to the lessor; this can be quickly corrected through negotiation. A moderate breach indicates the lessee violates some non-core contract terms, causing some economic loss to the lessor, requiring the lessee to bear corresponding compensation liability as stipulated in the contract. A serious breach indicates the lessee violates core contract terms, causing significant economic loss to the lessor and preventing the leasing project from progressing normally; in this case, the lessor can terminate the lease and pursue the breaching party's liability according to the contract. This ensures that the format of historical performance data is consistent and logically clear.

[0025] When cleaning external market environment data, abnormal fluctuations in the price data of similar leased assets are investigated. By comparing the average price of the industry during the same period and the overall trend of the regional market, it is determined whether the fluctuations are reasonable. If the data collection is incorrect, the value is removed and replaced with the average price of the adjacent time period. Duplicate policy documents and macroeconomic data are removed, and the latest published content is retained. If the supply and demand ratio data of a certain region is missing, the industry quarterly report and market research data are consulted, and a reasonable supply and demand ratio is estimated and supplemented based on the leasing business transaction volume of the region during the same period. When standardizing external market environment data, the market prices of similar leased assets are uniformly converted to RMB yuan per month. The supply and demand ratio is presented in the form of "supply: demand". Industry policies are organized according to a fixed structure of "policy name - release time - core content". Macroeconomic indicators retain the corresponding decimal places according to industry data recording standards. Market demand changes are labeled with the expression format of "growth", "stable", and "decline" combined with trend descriptions to ensure that the external market environment data is formatted in a standardized manner and is easy to correlate later.

[0026] The lessee information, leased property status information, performance history data, and external market environment data, after being cleaned and standardized, are organized into separate data sets. Each data set contains a unique identifier: the lease project number.

[0027] During the association and matching process, the lease project number of the target lease project is first extracted and used as the core matching basis. First, all data in the lessee information set whose lease project number matches the target number are filtered out. Then, all data corresponding to the same lease project number are filtered out from the leased asset status information set. Manual verification confirms that the lease project numbers in these two types of data are completely consistent, ensuring that the lessee information and leased asset status information corresponding to the same lease project number belong to the same target lease project. Next, the preliminarily associated lessee information and leased asset status information are matched with the performance history data set. Again, based on the lease project number, all performance records corresponding to the same number in the performance history data set are filtered out. The lease project number in each performance record is checked against the numbers in the first two types of data to ensure that each performance record accurately corresponds to the lessee and leased asset of the target lease project. Then, the three types of associated data are... The data is matched with external market environment data sets. Based on information such as the type of leased property and the leased area of ​​the target leasing project, and combined with the lease project number, highly relevant market data is selected from the external market environment data set. The consistency of the lease project number is checked again to ensure that the market environment data and the target leasing project are accurately matched. In the entire association matching process, the lease project number of each data record is reviewed one by one to avoid incorrect, missing, or mismatched numbers. After the matching is completed, according to the preset structured data format, all standardized fields of the lessee information, the full data of the leased property status information, the complete record of the performance history data, and the corresponding external market environment data are systematically integrated to form a structured data of the target leasing project that covers all relevant information and is logically coherent. All information in this structured data is closely linked through the lease project number to ensure data consistency and usability.

[0028] The beneficial effects include ensuring that the lessee information, leased asset status information, performance history data, and external market environment data extracted from the business database, IoT monitoring platform, contract management system, and market data interface accurately correspond to the target leasing project without omissions or mismatches. The system cleanses and processes the data to remove anomalies and duplicates, fill in missing information, and correct errors, ensuring the accuracy and completeness of all data. Standardization then unifies the format of various data types, making data from different sources comparable and compatible. Finally, the lease project number is used as the core identifier for verification and matching, systematically integrating all types of data into a logically coherent and comprehensive structured data set for the target leasing project. All information in this structured data is closely linked, consistent, and reliable, providing sufficient and accurate data support for subsequent management, risk assessment, and decision-making for the leasing project, effectively avoiding business processing errors caused by chaotic, inaccurate, or incomplete data.

[0029] S2. Construct a supervised fine-tuning dataset based on historical case data, and use the supervised fine-tuning dataset to perform supervised fine-tuning on the initial language model to obtain a risk perception vertical model for the target leasing project; In this embodiment of the invention, the step of constructing a supervised fine-tuning dataset based on historical case data, and using the supervised fine-tuning dataset to perform supervised fine-tuning on the initial language model to obtain a risk perception vertical model for the target leasing project includes: From the historical case data, leasing project cases with clear risk labels were selected; Convert the aforementioned rental project case into question-and-answer pair format data; The question-and-answer pair formatted data is combined into a supervisory fine-tuning dataset for the target leasing project; The model parameters of the initial language model were adjusted using the supervised fine-tuning dataset and adapted to the rental risk assessment task. Save the adjusted model parameters to obtain the risk perception vertical model for the target leasing project.

[0030] When selecting rental project cases with clear risk labels from historical case data, the historical case data stores complete information on all past rental projects, including basic project information, lessee information, leased property status records, performance process data, risk occurrence and corresponding handling results, etc. Risk labels are clear identifiers pre-marked based on the actual risk situation of each rental project. Label types cover no risk, low risk, medium risk, high risk, or specific risk types such as rent arrears risk, leased property damage risk, performance breach risk, etc. During the screening process, the risk label marking of each historical rental project case needs to be reviewed one by one to confirm that the label has a clear and specific definition and that the label corresponds completely to the actual risk situation recorded in the case. There should be no cases with ambiguous labels, no labels, or labels that do not match the actual risks. All rental project cases that meet the above conditions are screened out to ensure that each selected rental project case has a clear and accurate risk label.

[0031] When converting lease project cases into question-and-answer pair data, for each selected lease project case, the core information related to risks in the case is first comprehensively reviewed, including the basic information of the lease project, the lessee's qualifications, the usage status of the leased property, performance records, the specific scenario in which the risk occurred, the type of risk, the cause of the risk, the impact of the risk, and the risk handling measures and results. Then, based on this core information, corresponding questions are constructed. The questions must directly point to the key content related to the risks. Subsequently, accurate information that completely corresponds to each question is extracted from the case as the answer, ensuring that the answer fully covers the content involved in the question and is consistent with the actual situation of the case. Each lease project case is converted into multiple question-and-answer pairs, comprehensively covering all kinds of risk-related information of the case, forming question-and-answer pair format data.

[0032] When combining question-and-answer pair formatted data into a supervised fine-tuning dataset for the target leasing project, first, all question-and-answer pair formatted data converted from all leasing project cases are collected. These data are checked one by one to remove duplicate question-and-answer pairs. Then, the completeness of each question-and-answer pair is checked to ensure that each question-and-answer pair contains a clear question and a corresponding complete answer, and that there are no missing questions or incomplete answers. Subsequently, all checked question-and-answer pairs are organized according to a unified format. Each question-and-answer pair is presented with a fixed structure of "question + answer". All the organized question-and-answer pairs are arranged in any order and integrated together to form a complete dataset. This dataset is the supervised fine-tuning dataset for the target leasing project. Each question-and-answer pair in the dataset can provide a clear supervisory signal for the fine-tuning of the initial language model.

[0033] When adjusting the model parameters of the initial language model using the supervised fine-tuning dataset to adapt it for the rental risk assessment task, the initial language model possesses general natural language understanding and information processing capabilities. First, question-answer pairs from the supervised fine-tuning dataset are input into the initial language model sequentially. After receiving each question-answer pair, the model performs semantic understanding of the question, extracting key information and core intent. Then, it combines the corresponding answer to learn the relationship between the question and the answer, as well as the professional expressions and logical rules in the rental risk domain. During the learning process, the model compares its output based on the current parameters with the standard answers in the dataset, identifying the differences between the two. Based on these differences, it automatically adjusts the neuron connections within the model, optimizing its feature extraction capabilities and semantic understanding accuracy for rental risk-related information. This allows the model to gradually master the professional knowledge and judgment logic in the rental risk assessment domain, enabling it to accurately respond to various questions related to rental project risks. All question-answer pairs in the supervised fine-tuning dataset are continuously input until the model's responses to rental risk-related questions are highly consistent with the standard answers, completing the model parameter adjustment and achieving adaptation to the rental risk assessment task.

[0034] Once the adjusted model parameters are saved, and the risk perception vertical model for the target leasing project is obtained, the data input and parameter adjustment process is stopped after the model completes parameter adjustment and adapts to the leasing risk assessment task. All current internal parameters of the model are then fully read. These parameters include key configuration data for various functions such as semantic understanding, feature extraction, and logical reasoning, directly affecting the model's ability to handle leasing risk-related issues and its output results. The reading process must ensure that all parameters are complete and error-free. The read parameters are then organized according to the model's preset storage format, making the parameter data orderly and recognizable by the model. A safe and stable storage path is selected, and the organized parameter data is written to the specified storage file, completing the parameter saving. When the model is needed, the parameter data in the storage file can be called to restore the model to its adjusted state. At this point, the model is a risk perception vertical model specifically designed for the target leasing project, capable of accurately handling various tasks related to leasing project risk assessment.

[0035] The beneficial effects are as follows: by selecting rental project cases with clear risk labels, the data used for model training is guaranteed to have a clear risk orientation and accuracy. The rental project cases are converted into question-answer pair format data, so that the risk-related information in the cases is presented in a form that the model can learn. This ensures that the data is highly adapted to the model training requirements. The combined supervised fine-tuning dataset comprehensively covers various types of rental risk-related information, providing sufficient and effective supervision signals for the initial language model. By adjusting the parameters of the initial language model through supervised fine-tuning, the model can deeply learn the professional knowledge, logical rules and expression habits in the field of rental risk assessment, accurately adapt to the specific tasks of rental risk assessment, and save the adjusted parameters to form a dedicated risk perception vertical model. This model can specifically handle risk assessment-related tasks of target rental projects, significantly improving the professionalism, accuracy and reliability of risk assessment, and providing strong model support for the risk identification, judgment and prevention of target rental projects.

[0036] S3. Initialize different parameter subsets of the risk perception vertical model to obtain the performance capability assessment sub-model, leased asset value assessment sub-model and external environment assessment sub-model of the target leasing project, and merge the performance capability assessment sub-model, leased asset value assessment model and external environment assessment model into a risk assessment network; In this embodiment of the invention, the initialization of different parameter subsets of the risk perception vertical model to obtain the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model of the target leasing project includes: Based on the dimensions of rental risk assessment, the parameters of the risk perception vertical model are divided into three parameter subsets; The first parameter subset is initialized and fixed into a network structure dedicated to processing the lessee information and the performance history data, thus obtaining the performance capability assessment sub-model of the target leasing project. The second parameter subset is initialized and fixed into a network structure dedicated to processing the state information of the leased property, thus obtaining the leased property value assessment sub-model of the target leasing project; The third parameter subset is initialized and fixed into a network structure specifically for processing the external market environment data, thus obtaining the external environment assessment sub-model of the target leasing project; The shared parameter layer in the risk perception vertical model is retained, so that the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model share the underlying semantic understanding capability.

[0037] The process of integrating the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model into a risk assessment network includes: Define a network convergence architecture; The outputs of the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model are respectively connected to different inputs of the feature fusion layer in the network fusion architecture. After the output of the feature fusion layer, a fully connected processing layer is added to the network fusion architecture; Through the feature fusion layer and the fully connected processing layer, the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model are logically integrated into a single network entity to obtain the risk assessment network of the target leasing project.

[0038] Based on the core dimensions of lease risk assessment, the classification is clearly defined into three fixed dimensions: performance capability assessment, leased asset value assessment, and external environment assessment. Each dimension corresponds to processing specific types of risk-related data. The parameters of the risk perception vertical model include various model components adapted to different data processing needs. During the classification process, the functions corresponding to each model parameter are sorted out one by one. Parameters specifically used to analyze lessee information and performance history data to support performance capability judgment are classified into the first parameter subset. Parameters specifically used to analyze leased asset status information to achieve leased asset value assessment are classified into the second parameter subset. Parameters specifically used to process external market environment data to complete external environment assessment are classified into the third parameter subset. This ensures that the function of each parameter subset is fully matched with the corresponding risk assessment dimension, and that the three parameter subsets cover all non-shared parameters of the risk perception vertical model without overlap or omission.

[0039] When initializing the first parameter subset, the feature types of the lessee information and performance history data that the subset needs to adapt to are first determined, including text information and structured data. During the initialization process, a dedicated network structure is configured for the first parameter subset. The input layer of this network structure is set to an input interface that adapts to the above two types of data formats. The hidden layer contains a dedicated processing unit for extracting lessee qualification characteristics, performance behavior patterns, and credit status correlation characteristics. The output layer corresponds to the output format of the performance capability assessment results. The initialization operation assigns initial functional states to each layer of the network structure according to the preset basic configuration. After the initialization is completed, the hierarchical settings, functional allocations, and initial configurations of the network structure are locked, and subsequent adjustments are not allowed. This ensures that the network structure can only specifically process lessee information and performance history data, ultimately forming a performance capability assessment sub-model for the target leasing project.

[0040] When initializing the second parameter subset, based on the core features of the leased asset status information, including leased asset operating status data, wear and tear records, maintenance history information, and usage scenario-related data, a network structure specifically designed for processing this type of data is constructed for the second parameter subset. The input layer of this network structure is adapted to the real-time and multi-dimensional characteristics of the leased asset status data, and can receive different types of status indicator data. The hidden layer is set up with processing modules for extracting core features such as the performance degradation trend, value loss degree, remaining service life assessment, and fault risk correlation of the leased asset. The output layer corresponds to the presentation format of the leased asset value assessment results. The initialization process allocates initial functional parameters to each module of the network structure according to preset standards, determines the processing logic and data flow path of each layer, and fixes all configurations of the network structure after initialization to ensure that it can only specifically process the leased asset status information, thereby obtaining the leased asset value assessment sub-model for the target leased project.

[0041] When initializing the third parameter subset, a dedicated network structure is designed for it, taking into account the types of external market environment data, including market supply and demand data for similar leased assets, information on changes in industry policies, data on the impact of macroeconomic factors, and market price fluctuations. The input layer of this network structure supports the format adaptation of multi-source market data and can receive market environment data from different channels. The hidden layer contains dedicated processing units for analyzing characteristics such as market supply and demand trends, the degree of policy impact, price fluctuation patterns, and industry competition. The output layer corresponds to the assessment result format of the degree of impact of the external environment on the risk of the leasing project. The initialization operation configures the initial functional state of each layer of the network structure according to preset basic rules, clarifies the data processing flow and feature extraction logic, and locks the configuration of the network structure after initialization, so that it can only process external market environment data, thereby obtaining the external environment assessment sub-model of the target leasing project.

[0042] The shared parameter layer in the risk perception vertical model includes a basic semantic understanding unit, a data feature normalization processing unit, and a general logical reasoning module. These units and modules can perform basic analysis and general feature extraction on various types of lease-related data, without relying on specific data types or specific evaluation dimensions. When retaining this shared parameter layer, it is set as the common foundation layer for the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model. When processing their respective input data, the three sub-models must first pass the data into the shared parameter layer. The semantic understanding unit parses the textual meaning in the data, the data feature normalization processing unit unifies the data format and scale, and the general logical reasoning module extracts the basic correlation features of the data. Then, the data processed by the shared parameter layer is passed into their respective dedicated network structures for further precise processing, ensuring that the three sub-models have consistent underlying semantic understanding capabilities and avoiding evaluation result deviations caused by differences in basic processing capabilities.

[0043] When defining the network convergence architecture, the core function is to integrate the outputs of the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model to form a unified risk assessment output. The architecture should include three key modules: an input interface module, a feature fusion layer module, and an output adaptation module. The input interface module should have three pre-defined independent input terminals, each corresponding to the output format of one of the three sub-models. The feature fusion layer module is responsible for receiving and integrating the feature information from the three input terminals. The output adaptation module is used to receive the subsequently added fully connected processing layer and output the final risk assessment result. During the definition process, the functional boundaries and data flow paths of each module must be clearly defined to ensure that the input interface accurately matches the output type of each sub-model, the feature fusion layer has the ability to receive multi-source features, and the output adaptation module provides a stable access foundation for the subsequent fully connected processing layer. This ensures that the logical chain of the entire architecture is complete and can directly adapt to subsequent sub-model connections and layer addition operations.

[0044] When connecting the outputs of the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model to different inputs of the feature fusion layer in the network fusion architecture, the output content type of each sub-model must first be clearly defined. The performance capability assessment sub-model outputs feature information related to the lessee's performance capability; the leased asset value assessment sub-model outputs feature information related to the value of the leased asset; and the external environment assessment sub-model outputs feature information related to the impact of the external market environment on the leasing project. For each sub-model's output, a dedicated signal transmission channel is configured. The channel's transmission format must fully match the sub-model's output format and the receiving format of the corresponding input of the feature fusion layer. During connection, the output of the performance capability assessment sub-model is connected to the first input of the feature fusion layer via a dedicated channel; the output of the leased asset value assessment sub-model is connected to the second input of the feature fusion layer via a dedicated channel; and the output of the external environment assessment sub-model is connected to the third input of the feature fusion layer via a dedicated channel. After connection, the transmission stability of each channel is verified through test signals to ensure that the output features of each sub-model are transmitted completely and without loss to the corresponding input of the feature fusion layer, without signal interference or data loss.

[0045] When adding a fully connected processing layer to the network convergence architecture after the output of the feature fusion layer, first determine the access location of the fully connected processing layer, placing it adjacent to the output of the feature fusion layer to ensure that the integrated features output by the feature fusion layer can be directly transmitted to the fully connected processing layer. The core function of the fully connected processing layer is to further correlate and integrate the multi-dimensional fused features output by the feature fusion layer, forming a unified feature vector that directly reflects the overall risk status of the target leasing project. During the addition process, first configure a feature receiving interface for the fully connected processing layer. The format of this interface should be fully compatible with the output format of the feature fusion layer. Then, connect the feature receiving interface of the fully connected processing layer to the output of the feature fusion layer through a fixed transmission link. After connection, lock the transmission parameters of the link to ensure that the format of the fused features does not change during transmission. At the same time, clarify the output format of the fully connected processing layer so that it can adapt to the output requirements of the final risk assessment results, preparing for the subsequent formation of a single network entity.

[0046] When the performance capability assessment sub-model, leased asset value assessment sub-model, and external environment assessment sub-model are logically integrated into a single network entity through the feature fusion layer and the fully connected processing layer, the workflow of the feature fusion layer is initiated first. The feature fusion layer receives the performance capability features, leased asset value features, and external environment features transmitted from the three sub-models respectively, and then sequentially concatenates the three types of features according to a fixed integration logic, establishing the correlation between different features to form a fused feature encompassing multi-dimensional risk factors of the leasing project. After the fused feature is transmitted to the fully connected processing layer, the fully connected processing layer deeply integrates the fused feature, eliminating dimensional differences between features of different sub-models, and extracting a core feature vector that comprehensively reflects the overall risk level of the target leasing project. In this process, the connection relationships between the three sub-models and the feature fusion layer, the connection relationship between the feature fusion layer and the fully connected processing layer, and the internal processing logic of each layer are locked to ensure that the three sub-models, the feature fusion layer, and the fully connected processing layer form a collaborative whole, that data can flow smoothly along the preset path, and that the functions of each part are closely connected, ultimately forming a single network entity that can receive lessee information, leased property status information, and external market environment data, and output the risk assessment results of the target leasing project. This network entity is the risk assessment network for the target leasing project.

[0047] The beneficial effects include the precise division of parameter subsets of the risk perception vertical model according to the core dimensions of lease risk assessment, ensuring that the functions of each subset are highly adapted to the corresponding assessment tasks. By constructing a dedicated network structure for each parameter subset and completing the initialization and fixed configuration, the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model can respectively process lessee information and performance history data, leased asset status information, and external market environment data. The specialized network structure design ensures the professionalism and accuracy of various data processing, avoiding interference when processing different types of data. At the same time, the shared parameter layer of the risk perception vertical model is retained, allowing the three sub-models to share consistent underlying semantic understanding, general feature extraction, and logical reasoning capabilities, ensuring the uniformity of basic data processing standards, reducing the deviation of assessment results, and ultimately enabling the three sub-models to not only have their own data processing expertise but also form a synergistic effect, comprehensively covering the key dimensions of lease risk assessment, and improving the comprehensiveness, accuracy, and efficiency of risk assessment for target lease projects.

[0048] By defining an adaptive network fusion architecture, clarifying the functional boundaries and data flow paths of each module, a stable foundation is provided for the integration of sub-models. The outputs of the three sub-models are precisely connected to the different inputs of the feature fusion layer, ensuring that the feature information output by each sub-model is transmitted completely, losslessly, and without interference. A fully connected processing layer is added to achieve deep correlation and integration of multi-dimensional fusion features. Finally, the three sub-models and the fusion and processing layer logic are integrated into a single risk assessment network, enabling the functions of each sub-model to work together and comprehensively cover risk assessment dimensions such as performance capability, leased asset value, and external environment. This achieves a comprehensive consideration of the risk factors of the target leasing project and improves the comprehensiveness, accuracy, and reliability of risk assessment.

[0049] S4. Input the input features of the structured data into the risk assessment network to obtain the preliminary performance risk score, preliminary asset risk score, and preliminary environmental risk score of the target lease project; In this embodiment of the invention, the step of simultaneously inputting the input features of the structured data into the risk assessment network to obtain the preliminary performance risk score, preliminary asset risk score, and preliminary environmental risk score of the target lease project includes: Extract input features corresponding to the target rental project from the structured data; In the performance capability assessment sub-model, the subset of features related to the lessee's credit and behavior in the input features is forward propagated to obtain the preliminary performance risk score of the target lease project; In the leased asset valuation sub-model, the subset of features related to the attributes and status of the leased asset in the input features is forward propagated to obtain the preliminary asset risk score of the target leased project; In the external environment assessment sub-model, the subset of input features related to macroeconomic and market fluctuations is forward-propagated to obtain the preliminary environmental risk score of the target leasing project.

[0050] When extracting input features corresponding to the target leasing project from structured data, the structured data has integrated lessee information, leased asset status information, performance history data, and external market environment data. The extraction process first clarifies the screening criteria for input features as information directly related to the risk assessment of the leasing project. Then, it systematically sorts out various types of information in the structured data, selecting core content such as lessee credit information, lessee behavior records, leased asset attribute information, leased asset status data, performance history details, macroeconomic data, and market fluctuation information. This information is transformed into feature forms that can be recognized by the risk assessment network. At the same time, redundant information in the structured data that is not related to risk assessment is removed to ensure that the extracted input features fully cover the three major categories of lessee credit and behavior, leased asset attributes and status, and macroeconomic and market fluctuations. Moreover, each feature accurately corresponds to the target leasing project without confusion or mismatch.

[0051] In the performance capability assessment sub-model, when forward propagating the subset of input features related to the lessee's credit and behavior, it is first clarified that this subset of features includes the lessee's qualification level, credit rating, historical leasing cooperation records, rent payment behavior, default records, etc. The forward propagation process first feeds this subset of features into the input layer of the performance capability assessment sub-model. The input layer performs format validation and adaptation processing on the features to ensure that the features meet the processing requirements of the sub-model. Subsequently, the features enter the shared parameter layer of the sub-model. Through the shared semantic understanding unit, the textual information and structured data meaning in the features are parsed to extract the core features of the lessee's credit level and behavioral patterns. Then, the features are fed into the dedicated network structure of the sub-model. The dedicated network structure analyzes the core features layer by layer according to the preset processing logic to judge the lessee's willingness and ability to perform. Finally, it outputs a value reflecting the performance risk level of the target leasing project. This value is the preliminary performance risk score of the target leasing project.

[0052] In the leased asset valuation sub-model, when forward propagating the subset of features related to the attributes and status of the leased asset from the input features, it is first determined that the subset of features covers information such as the type, specifications, initial value, operating status, degree of wear and tear, and maintenance history of the leased asset. After forward propagation is initiated, the subset of features is first input into the input layer of the leased asset valuation sub-model. The input layer completes the format conversion and adaptation of the features to ensure that the features can be successfully passed into the subsequent processing stage. Then, the features enter the shared parameter layer, where the semantic understanding unit parses the attribute description and status record of the leased asset, extracts the core attributes of the leased asset and the key features of the current status, and then passes these features into the dedicated processing module of the sub-model. The dedicated processing module analyzes the attribute matching degree, status stability, and wear and tear trend of the leased asset to determine the potential asset risks of the leased asset, and finally outputs the corresponding risk value, which is the preliminary asset risk score of the target leased project.

[0053] In the external environment assessment sub-model, when forward propagating the subset of input features related to macroeconomic and market fluctuations, this subset includes macroeconomic data, industry policy changes, supply and demand changes in similar leased assets, and market price fluctuation records. The forward propagation process first imports the feature subset into the input layer of the external environment assessment sub-model. The input layer processes various market-related data in a unified format to eliminate format differences between data from different sources. Subsequently, the features enter the shared parameter layer, where the semantic understanding unit parses textual features such as policy texts and market analysis reports to extract core information on macroeconomic trends and market fluctuations. This information is then passed into the sub-model's dedicated network structure. The dedicated network structure analyzes the potential impact of macroeconomic and market fluctuations on the target leasing project, assesses the asset value fluctuation risk of the leasing project due to changes in the market environment, and finally outputs the corresponding risk value, which is the preliminary environmental risk score of the target leasing project.

[0054] The beneficial effects include the accurate extraction of input features from structured data, covering lessee credit and behavior, leased property attributes and status, and macroeconomic and market fluctuations. Redundant information is eliminated to ensure the relevance and completeness of the features. Through the forward propagation process of the performance capability assessment sub-model, different categories of feature subsets are analyzed layer by layer and specifically. With the help of the basic processing capabilities of the shared parameter layer and the accurate judgment logic of the dedicated network structure, preliminary scores reflecting performance risk, asset risk, and environmental risk are output respectively. Each score corresponds to a clear risk dimension, comprehensively capturing various potential risk factors of the target leasing project. This provides accurate and reliable risk basis for subsequent comprehensive risk assessment, ensuring that the risk assessment is comprehensive and the analysis is in-depth.

[0055] S5. Based on the weight allocation vector corresponding to the target leasing project, the preliminary performance risk score, the preliminary asset risk score, and the preliminary environmental risk score are weighted and integrated to obtain a comprehensive risk score for the target leasing project, and risk disposal recommendations for the target leasing project are generated based on the comprehensive risk score. In this embodiment of the invention, the formula for calculating the weight allocation vector is as follows: ; In the formula, Assign vectors to the weights. This is the preliminary performance risk score. This is the preliminary asset risk score. This is the preliminary environmental risk score. The standard deviation of the historical performance risk score series. The standard deviation of the historical asset risk score series. The standard deviation of the historical environmental risk score series. To assess the significance of overall risks, The sum of the significance of the combined risk across three dimensions. This is a weighted allocation vector corresponding to the initial performance risk score. This is the weight allocation vector corresponding to the initial asset risk score. This is the weight allocation vector corresponding to the preliminary environmental risk score.

[0056] The step of generating risk management recommendations for the target leasing project based on the comprehensive risk score includes: Based on a preset comprehensive risk score threshold range, the comprehensive risk score of the target leasing project is mapped to the corresponding risk level; Based on the risk level obtained from the mapping, a basic risk management strategy corresponding to the risk level is matched from a preset risk management strategy library; Based on the basic risk management strategy, and combined with the specific risk score composition of the target lease project, extract the portion of any score in the preliminary performance risk score, the preliminary asset risk score, and the preliminary environmental risk score that exceeds its corresponding single threshold, and generate a risk enhancement prompt for the target lease project. By combining the basic risk management strategy with the risk enhancement prompts, risk management recommendations for the target leasing project are obtained.

[0057] The preliminary performance risk score is obtained by collecting, organizing, and quantifying the performance-related indicators of the assessed entity. The preliminary asset risk score is obtained by collecting, organizing, and quantifying the asset-related indicators of the assessed entity. The preliminary environmental risk score is obtained by collecting, organizing, and quantifying the environmental-related indicators of the assessed entity. The standard deviation of the historical performance risk score series is calculated as follows: First, multiple past performance risk scores of the assessed entity are collected to form a historical performance risk score series. Then, the mean of this series is calculated by adding all performance risk scores in the series, dividing the sum by the number of scores, and obtaining the mean. Next, the mean is subtracted from each performance risk score in the series, and the square of each difference is calculated. All the squared results are added together, divided by the number of scores minus one, and finally, the square root of the result is taken. The resulting standard deviation is the historical performance risk score series.

[0058] The standard deviation of the historical asset risk score series is calculated as follows: First, collect multiple historical asset risk scores of the assessed object to form a historical asset risk score series. Then, calculate the mean of this series by adding up all asset risk scores in the series, dividing the sum by the number of scores to obtain the mean. Next, subtract the mean from each asset risk score in the series, square each difference, add up all the squared results, divide the sum by the number of scores minus one, and finally take the square root of the result. The result is the standard deviation of the historical asset risk score series.

[0059] The standard deviation of the historical environmental risk score series is calculated as follows: First, collect multiple past environmental risk scores of the assessed object to form a historical environmental risk score series. Then, calculate the mean of the series by summing all environmental risk scores in the series, dividing the sum by the number of scores to obtain the mean. Next, subtract the mean from each environmental risk score in the series, square each difference, sum all the squared results, divide the sum by the number of scores minus one, and finally take the square root of the result. The result is the standard deviation of the historical environmental risk score series.

[0060] The weight allocation vector corresponding to the initial performance risk score is obtained by multiplying the initial performance risk score with the standard deviation of the historical performance risk score series to get the comprehensive risk significance of the corresponding dimension. Then, the comprehensive risk significance corresponding to the initial performance risk score, the comprehensive risk significance corresponding to the initial asset risk score, and the comprehensive risk significance corresponding to the initial environmental risk score are added together to get the sum of the comprehensive risk significance of the three dimensions. The result of dividing the comprehensive risk significance corresponding to the initial performance risk score by the sum of the comprehensive risk significance of the three dimensions is the weight allocation vector corresponding to the initial performance risk score.

[0061] The weight allocation vector corresponding to the initial asset risk score is obtained by multiplying the initial asset risk score with the standard deviation of the historical asset risk score series to obtain the comprehensive risk significance of the corresponding dimension. Then, the comprehensive risk significance corresponding to the initial performance risk score, the initial asset risk score, and the initial environmental risk score are added together to obtain the sum of the comprehensive risk significance of the three dimensions. The comprehensive risk significance corresponding to the initial asset risk score is divided by the sum of the comprehensive risk significance of the three dimensions, and the result is the weight allocation vector corresponding to the initial asset risk score.

[0062] The weight allocation vector corresponding to the preliminary environmental risk score is obtained by multiplying the preliminary environmental risk score with the standard deviation of the historical environmental risk score series to obtain the comprehensive risk significance of the corresponding dimension. Then, the comprehensive risk significance corresponding to the preliminary performance risk score, the comprehensive risk significance corresponding to the preliminary asset risk score, and the comprehensive risk significance corresponding to the preliminary environmental risk score are added together to obtain the sum of the comprehensive risk significance of the three dimensions. The comprehensive risk significance corresponding to the preliminary environmental risk score is divided by the sum of the comprehensive risk significance of the three dimensions, and the result is the weight allocation vector corresponding to the preliminary environmental risk score.

[0063] The calculation process of the weight allocation vector comprehensively considers the magnitude of each preliminary risk score and the dispersion of the corresponding historical risk score sequence. It reflects the prominence of each dimension of risk through comprehensive risk significance, and then obtains the weight allocation vector corresponding to each dimension through normalization. This ensures that the weight allocation result can accurately match the actual significance of each risk dimension, providing a scientific and reasonable weight basis for the weighted fusion of risk scores of each dimension in the subsequent comprehensive risk assessment.

[0064] The weight allocation vector corresponding to the target leasing project is pre-set based on the specific characteristics of the project, the key points of industry risk management and historical project experience. Each weight value in the vector corresponds to the importance ratio of the preliminary performance risk score, the preliminary asset risk score and the preliminary environmental risk score, and the sum of each weight value meets the requirement of fully covering the risk assessment dimensions, ensuring that the impact of different risk dimensions can be accurately reflected through the weights.

[0065] In the weighted fusion process, the specific weights corresponding to each preliminary risk score in the weight allocation vector are first clarified. Then, the preliminary performance risk score is linked and integrated with the corresponding weight to fully consider the impact weight of the lessee's credit and performance behavior on the overall project risk. At the same time, the preliminary asset risk score is linked with the corresponding weight to highlight the impact ratio of risks related to the attributes and status of the leased property. Then, the preliminary environmental risk score is linked with the corresponding weight to take into account the impact of macroeconomic and market fluctuation risks.

[0066] By integrating the risk scores of the three dimensions according to their weights, a comprehensive risk score is obtained that can fully reflect the overall risk level of the target leasing project. This score not only covers the impact of each individual risk, but also distinguishes the importance of different risk dimensions through weights, ensuring that the assessment results are consistent with the actual risk situation of the project.

[0067] When mapping the comprehensive risk score of the target leasing project to the corresponding risk level based on the preset comprehensive risk score threshold range, the preset comprehensive risk score threshold range is a fixed range set in advance according to the risk control standards of the leasing industry and historical project risk data. Each range corresponds to a unique risk level, and the risk level covers fixed categories such as low risk, medium risk, and high risk.

[0068] First, retrieve the preset comprehensive risk score threshold range standard, clarify the score range of each range and the corresponding risk level name, then compare the comprehensive risk score of the target leasing project with the range of each threshold range one by one to determine which threshold range the comprehensive risk score falls into, and directly use the risk level corresponding to that range as the mapping result to complete the mapping from comprehensive risk score to risk level, ensuring that the mapping result is completely consistent with the setting rules of the threshold range and without deviation.

[0069] Based on the risk level obtained from the mapping, when matching the basic risk management strategy corresponding to the risk level from the preset risk management strategy library, the preset risk management strategy library stores basic risk management strategies classified by risk level. Each risk level corresponds to a unique set of basic strategies, and the strategy content includes core elements such as risk monitoring frequency, response framework, responsible entity, and handling process.

[0070] First, clarify the specific name of the risk level obtained from the mapping. Then, using the search function of the strategy library, locate the basic risk management strategy under the corresponding category in the strategy library by using the risk level as the search keyword. Retrieve the complete content of the strategy and perform a completeness check on the retrieved strategy to ensure that the strategy content covers all the response requirements of the common risks under the risk level and that no key information is missing, thus completing the matching of the basic risk management strategy.

[0071] Based on the basic risk management strategy, and combined with the specific risk score composition of the target lease project, extract the portion of any score in the preliminary performance risk score, preliminary asset risk score, and preliminary environmental risk score that exceeds its corresponding individual threshold. When generating risk enhancement prompts for the target lease project, first clarify the individual threshold corresponding to each preliminary risk score. The individual threshold is a fixed value set in advance according to the industry standard of the corresponding risk dimension. Then, compare the preliminary performance risk score with the performance risk individual threshold, the preliminary asset risk score with the asset risk individual threshold, and the preliminary environmental risk score with the environmental risk individual threshold one by one, and screen out the preliminary risk scores and corresponding risk dimensions that exceed the corresponding individual thresholds.

[0072] For each preliminary risk score exceeding the threshold, analyze the corresponding risk factors. For example, a preliminary performance risk score exceeding the threshold corresponds to the risk of insufficient performance capability of the lessee; a preliminary asset risk score exceeding the threshold corresponds to the risk of depreciation or malfunction of the leased property; and a preliminary environmental risk score exceeding the threshold corresponds to the risk of market fluctuations or policy changes. Based on the risk factors obtained from the analysis, write targeted risk enhancement tips. The tips should clearly state the risk points that need to be focused on, the additional monitoring measures that need to be implemented, and the initial response direction, so as to ensure that the risk enhancement tips accurately target the risk dimensions exceeding the threshold.

[0073] When combining basic risk management strategies with risk enhancement tips to obtain risk management recommendations for the target leasing project, first, sort out the complete framework and core content of the basic risk management strategy, and determine the supplementary position of the risk enhancement tips. Usually, the risk enhancement tips are added after the response measures section of the corresponding risk dimension in the basic strategy as specific reinforcement content.

[0074] The specific text of the risk enhancement prompts is integrated into the basic risk management strategy to ensure that the combined content is logically coherent and smoothly expressed. This retains the integrity and universality of the basic risk management strategy while highlighting the specific risk response requirements of the target leasing project through the risk enhancement prompts. After the combination is completed, the entire content is read through and verified to ensure that there is no duplication or conflict and that the core response measures are clear and specific. The final complete content is the risk management recommendation for the target leasing project.

[0075] The beneficial effects include: achieving a precise mapping between comprehensive risk scores and risk levels through preset comprehensive risk scoring threshold ranges, ensuring that risk level determination conforms to industry standards and historical data patterns; matching corresponding basic risk management strategies from the strategy library based on the mapped risk levels, providing a standardized and universal framework for risk response; combining specific risk scores to form risk dimensions that exceed individual thresholds, generating targeted risk enhancement prompts, highlighting key risk points and reinforcement measures, and organically combining basic risk management strategies with risk enhancement prompts. This ensures that risk management recommendations are comprehensive and standardized, while also taking into account the specific risks and operability of specific risks, effectively covering the overall and key risks of the target leasing project, providing clear, specific, and feasible action guidelines for risk management, and improving the accuracy and effectiveness of risk management.

[0076] S6. Based on the comprehensive risk score, the weight allocation vector, and the risk handling suggestions, update the supervised fine-tuning dataset to obtain the optimized risk perception vertical model.

[0077] In this embodiment of the invention, updating the supervised fine-tuning dataset based on the comprehensive risk score, the weight allocation vector, and the risk management recommendations to obtain an optimized risk perception vertical model includes: The structured data of the target leasing project, the final generated comprehensive risk score, the weight allocation vector, and the risk management recommendations are integrated into a new historical case data entry. The validity of new historical case data entries is verified according to the preset verification logic; New historical case data entries will be added to the historical case database of the target leasing project upon verification. Based on the updated historical case database, the supervised fine-tuning dataset was reconstructed; The risk-aware vertical model is incrementally trained using the reconstructed supervised fine-tuning dataset to obtain an optimized risk-aware vertical model.

[0078] The structured data of the target leasing project, the final comprehensive risk score, the weight allocation vector, and the risk disposal recommendations are integrated into a new historical case data entry. The structured data includes complete standardized content such as lessee information, leased property status information, performance history data, and external market environment data. The comprehensive risk score is a fixed value that reflects the overall risk level of the project. The weight allocation vector is the importance ratio configuration corresponding to the three preliminary risk scores. The risk disposal recommendations are a complete action guide that combines basic strategies and risk enhancement prompts.

[0079] The integration process first assigns a unique case identifier to each entry, which is associated with the lease project number of the target lease project. Then, all fields of the structured data, the specific values ​​of the comprehensive risk score, the complete configuration information of the weight allocation vector, and the full text of the risk disposal recommendations are entered in a fixed format to ensure that the information in each part is complete and without omission, and that a clear connection is established between each piece of information through the case identifier, forming a new historical case data entry with a complete structure and comprehensive information.

[0080] When validating the validity of new historical case data entries according to the preset verification logic, the preset verification logic includes three core steps: data integrity verification, data consistency verification, and data authenticity verification. Data integrity verification requires checking each new entry to ensure it contains the four essential parts: structured data, comprehensive risk score, weight allocation vector, and risk management recommendations. It also checks for missing information or empty key fields in each part. Data consistency verification verifies whether the lease project information in the structured data matches the project indicated by the comprehensive risk score and risk management recommendations, whether the weight allocation vector configuration matches the project's risk dimensions, and whether the risk management recommendations correspond to the comprehensive risk score and the preliminary risk score exceeding the threshold. Data authenticity verification confirms the authenticity and reliability of the structured data and various scores by comparing them with the original business records, monitoring data, and contract documents of the target lease project, and that the risk management recommendations conform to the actual risk scenario. New historical case data entries that pass all three verification steps are considered valid entries. If any step fails, the entry must be corrected and re-verified until all verification requirements are met.

[0081] New verified historical case data entries are added to the historical case database of the target leasing project. The historical case database is stored in a structure that is classified by case identifier, and each entry is associated with a specific leasing project category according to the case identifier.

[0082] First, identify the case identifier corresponding to the new verified entry. Then, use the database entry function to find the storage path corresponding to the case identifier. If there is no category corresponding to the case identifier in the database, create a new category folder. If there is already a corresponding category, open the folder directly and enter the complete information of the new entry according to the database's preset storage format. After the entry is completed, perform storage verification on the data to confirm that the entry information has been completely written into the database and can be quickly retrieved by the case identifier. At the same time, record the timestamp of the entry addition to ensure that the addition time of all entries in the database can be traced back, thus completing the addition operation of the new entry in the historical case database.

[0083] When rebuilding the supervised fine-tuning dataset based on the updated historical case database, the updated historical case database includes all existing valid historical case data entries as well as newly added and validated entries. First, all case entries with clear risk labels are selected from the database, ensuring that each selected entry has a clear risk level identifier. Then, according to the previously set question-and-answer pair transformation rules, each case entry is processed. For the core information in the entry, such as structured data, comprehensive risk score, weight allocation vector, and risk management recommendations, question-and-answer pairs covering dimensions such as risk type, risk cause, assessment results, and management measures are constructed. Each case entry is converted into multiple question-and-answer pairs, comprehensively covering the risk-related information in the case.

[0084] Collect all transformed question-answer pairs, remove duplicate content, check whether the questions and answers of each question-answer pair correspond accurately and whether the information is complete, organize them in a unified format, and integrate all valid question-answer pairs into a complete dataset, which is the reconstructed supervised fine-tuning dataset.

[0085] Using the reconstructed supervised fine-tuning dataset, the risk perception vertical model is incrementally trained to obtain the optimized risk perception vertical model. The reconstructed supervised fine-tuning dataset is then sequentially input into the original risk perception vertical model. After receiving each question-answer pair, the model uses the existing parameter base to perform semantic understanding and feature extraction on the question and outputs the corresponding prediction result.

[0086] The model's predictions are compared with the standard answers in the question-and-answer pairs to identify the differences. For the differences, the model adjusts the internal neuron connections to optimize the ability to identify and process risk features in new cases, while retaining the original parameters for effective risk assessment logic to avoid performance degradation of the original model due to incremental training.

[0087] All question-answer pairs in the dataset are processed one by one, and the model parameters are continuously adjusted until the model's prediction results for new question-answer pairs are highly consistent with the standard answers, and the processing accuracy for existing question-answer pairs is not reduced. After training, all parameters of the adjusted model are fully read, organized and saved according to the preset format, and the parameter configuration of the original model is replaced to obtain the optimized risk perception vertical model, which has a more comprehensive risk scenario coverage and higher evaluation accuracy.

[0088] The beneficial effects are as follows: by integrating all relevant information of the target leasing project to form new historical case data entries, the completeness and close correlation of the newly added case information are ensured. The new entries are fully verified by the preset verification logic to ensure their authenticity, completeness and consistency, providing a reliable data foundation for database updates. The inclusion of effective new entries into the historical case database enriches the case reserves and expands the coverage of risk scenarios. Based on the updated database, the dataset is reconstructed and fine-tuned under supervision, so that the dataset contains more comprehensive risk assessment and disposal information, providing more sufficient supervision signals for model training. The parameters of the risk perception vertical model are optimized through incremental training, which improves the ability to identify and handle new risk scenarios while retaining the original effective assessment logic. The final optimized risk perception vertical model has higher assessment accuracy and wider scenario adaptability, continuously improving the reliability and effectiveness of risk assessment for the target leasing project.

[0089] like Figure 2 The diagram shown is a functional block diagram of a leasing risk data processing system based on a vertical model, provided in an embodiment of the present invention.

[0090] The leasing risk data processing system 100 based on a vertical model described in this invention can be installed in an electronic device. Depending on the functions implemented, the leasing risk data processing system 100 based on a vertical model may include a data integration and structuring module 101, a vertical model training module 102, a risk assessment network construction module 103, a multi-dimensional risk parallel assessment module 104, a comprehensive decision-making and output module 105, and an adaptive optimization module 106. The modules described in this invention can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.

[0091] In this embodiment, the functions of each module / unit are as follows: The data integration and structuring module 101 is used to integrate the lessee information, leased property status information, performance history data and external market environment data of the target leasing project into the structured data of the target leasing project. The vertical model training module 102 is used to construct a supervised fine-tuning dataset based on historical case data, and to use the supervised fine-tuning dataset to perform supervised fine-tuning on the initial language model to obtain a risk perception vertical model for the target leasing project. The risk assessment network construction module 103 is used to initialize different parameter subsets of the risk perception vertical model to obtain the performance capability assessment sub-model, the leased asset value assessment sub-model and the external environment assessment sub-model of the target leasing project, and to integrate the performance capability assessment sub-model, the leased asset value assessment model and the external environment assessment model into a risk assessment network. The multi-dimensional risk parallel assessment module 104 is used to simultaneously input the input features of the structured data into the risk assessment network to obtain the preliminary performance risk score, preliminary asset risk score and preliminary environmental risk score of the target leasing project; The comprehensive decision-making and output module 105 is used to perform weighted fusion of the preliminary performance risk score, the preliminary asset risk score and the preliminary environmental risk score according to the weight allocation vector corresponding to the target leasing project, to obtain a comprehensive risk score for the target leasing project, and to generate risk disposal suggestions for the target leasing project based on the comprehensive risk score. The adaptive optimization module 106 is used to update the supervised fine-tuning dataset based on the comprehensive risk score, the weight allocation vector, and the risk handling suggestions to obtain an optimized risk perception vertical model.

[0092] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0093] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0094] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0095] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0096] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0097] Finally, 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 method for processing rental risk data based on a vertical model, characterized in that, The method includes: S1. Integrate the lessee information, leased property status information, performance history data, and external market environment data of the target leasing project into structured data of the target leasing project; S2. Construct a supervised fine-tuning dataset based on historical case data, and use the supervised fine-tuning dataset to perform supervised fine-tuning on the initial language model to obtain a risk perception vertical model for the target leasing project; S3. Initialize different parameter subsets of the risk perception vertical model to obtain the performance capability assessment sub-model, leased asset value assessment sub-model and external environment assessment sub-model of the target leasing project, and merge the performance capability assessment sub-model, leased asset value assessment model and external environment assessment model into a risk assessment network; S4. Input the input features of the structured data into the risk assessment network to obtain the preliminary performance risk score, preliminary asset risk score, and preliminary environmental risk score of the target lease project; S5. Based on the weight allocation vector corresponding to the target leasing project, the preliminary performance risk score, the preliminary asset risk score, and the preliminary environmental risk score are weighted and integrated to obtain a comprehensive risk score for the target leasing project, and risk disposal recommendations for the target leasing project are generated based on the comprehensive risk score. S6. Based on the comprehensive risk score, the weight allocation vector, and the risk handling suggestions, update the supervised fine-tuning dataset to obtain the optimized risk perception vertical model.

2. The leasing risk data processing method based on a vertical model as described in claim 1, characterized in that, The process of integrating lessee information, leased asset status information, performance history data, and external market environment data of the target leasing project into structured data for the target leasing project includes: The system extracts lessee information from the business database, the status information of the leased property from the IoT monitoring platform, the performance history data from the contract management system, and the external market environment data from the market data interface. The lessee information, the leased property status information, the performance history data, and the external market environment data are cleaned and standardized. By associating and matching the standardized data using the rental project number, the structured data of the target rental project is obtained.

3. The leasing risk data processing method based on a vertical model as described in claim 1, characterized in that, The process involves constructing a supervised fine-tuning dataset based on historical case data, and using this dataset to perform supervised fine-tuning of the initial language model to obtain a risk perception vertical model for the target leasing project, including: From the historical case data, leasing project cases with clear risk labels were selected; Convert the aforementioned rental project case into question-and-answer pair format data; The question-and-answer pair formatted data is combined into a supervisory fine-tuning dataset for the target leasing project; The model parameters of the initial language model were adjusted using the supervised fine-tuning dataset and adapted to the rental risk assessment task. Save the adjusted model parameters to obtain the risk perception vertical model for the target leasing project.

4. The leasing risk data processing method based on a vertical model as described in claim 1, characterized in that, The process of initializing different subsets of parameters of the risk perception vertical model yields a sub-model for assessing the performance capability of the target leasing project, a sub-model for assessing the value of the leased asset, and a sub-model for assessing the external environment, including: Based on the dimensions of rental risk assessment, the parameters of the risk perception vertical model are divided into three parameter subsets; The first parameter subset is initialized and fixed into a network structure dedicated to processing the lessee information and the performance history data, thus obtaining the performance capability assessment sub-model of the target leasing project. The second parameter subset is initialized and fixed into a network structure dedicated to processing the state information of the leased property, thus obtaining the leased property value assessment sub-model of the target leasing project; The third parameter subset is initialized and fixed into a network structure specifically for processing the external market environment data, thus obtaining the external environment assessment sub-model of the target leasing project; The shared parameter layer in the risk perception vertical model is retained, so that the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model share the underlying semantic understanding capability.

5. The leasing risk data processing method based on a vertical model as described in claim 4, characterized in that, The process of integrating the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model into a risk assessment network includes: Define a network convergence architecture; The outputs of the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model are respectively connected to different inputs of the feature fusion layer in the network fusion architecture. After the output of the feature fusion layer, a fully connected processing layer is added to the network fusion architecture; Through the feature fusion layer and the fully connected processing layer, the performance capability assessment sub-model, the leased asset value assessment sub-model, and the external environment assessment sub-model are logically integrated into a single network entity to obtain the risk assessment network of the target leasing project.

6. The leasing risk data processing method based on a vertical model as described in claim 1, characterized in that, The input features of the structured data are simultaneously input into the risk assessment network to obtain the preliminary performance risk score, preliminary asset risk score, and preliminary environmental risk score of the target lease project, including: Extract input features corresponding to the target rental project from the structured data; In the performance capability assessment sub-model, the subset of features related to the lessee's credit and behavior in the input features is forward propagated to obtain the preliminary performance risk score of the target lease project; In the leased asset valuation sub-model, the subset of features related to the attributes and status of the leased asset in the input features is forward propagated to obtain the preliminary asset risk score of the target leased project; In the external environment assessment sub-model, the subset of input features related to macroeconomic and market fluctuations is forward-propagated to obtain the preliminary environmental risk score of the target leasing project.

7. The leasing risk data processing method based on a vertical model as described in claim 1, characterized in that, The formula for calculating the weight allocation vector is as follows: ； In the formula, Assign vectors to the weights. This is the preliminary performance risk score. This is the preliminary asset risk score. This is the preliminary environmental risk score. The standard deviation of the historical performance risk score series. The standard deviation of the historical asset risk score series. The standard deviation of the historical environmental risk score series. To assess the significance of overall risks, The sum of the significance of the combined risk across three dimensions. This is a weighted allocation vector corresponding to the initial performance risk score. This is the weight allocation vector corresponding to the initial asset risk score. This is the weight allocation vector corresponding to the preliminary environmental risk score.

8. The leasing risk data processing method based on a vertical model as described in claim 7, characterized in that, The step of generating risk management recommendations for the target leasing project based on the comprehensive risk score includes: Based on a preset comprehensive risk score threshold range, the comprehensive risk score of the target leasing project is mapped to the corresponding risk level; Based on the risk level obtained from the mapping, a basic risk management strategy corresponding to the risk level is matched from a preset risk management strategy library; Based on the basic risk management strategy, and combined with the specific risk score composition of the target lease project, extract the portion of any score in the preliminary performance risk score, the preliminary asset risk score, and the preliminary environmental risk score that exceeds its corresponding single threshold, and generate a risk enhancement prompt for the target lease project. By combining the basic risk management strategy with the risk enhancement prompts, risk management recommendations for the target leasing project are obtained.

9. The leasing risk data processing method based on a vertical model as described in claim 1, characterized in that, The process of updating the supervised fine-tuning dataset based on the comprehensive risk score, the weight allocation vector, and the risk management recommendations to obtain an optimized risk perception vertical model includes: The structured data of the target leasing project, the final generated comprehensive risk score, the weight allocation vector, and the risk management recommendations are integrated into a new historical case data entry. The validity of new historical case data entries is verified according to the preset verification logic; New historical case data entries will be added to the historical case database of the target leasing project upon verification. Based on the updated historical case database, the supervised fine-tuning dataset was reconstructed; The risk-aware vertical model is incrementally trained using the reconstructed supervised fine-tuning dataset to obtain an optimized risk-aware vertical model.

10. A leasing risk data processing system based on a vertical model, characterized in that, The system for implementing the rental risk data processing method based on a vertical model as described in claim 1 includes: The data integration and structuring module is used to integrate the lessee information, leased property status information, performance history data and external market environment data of the target leasing project into the structured data of the target leasing project; The vertical model training module is used to construct a supervised fine-tuning dataset based on historical case data, and to use the supervised fine-tuning dataset to perform supervised fine-tuning on the initial language model to obtain a risk perception vertical model for the target leasing project. The risk assessment network construction module is used to initialize different parameter subsets of the risk perception vertical model to obtain the performance capability assessment sub-model, leased asset value assessment sub-model and external environment assessment sub-model of the target leasing project, and to integrate the performance capability assessment sub-model, leased asset value assessment model and external environment assessment model into a risk assessment network. The multi-dimensional risk parallel assessment module is used to simultaneously input the input features of the structured data into the risk assessment network to obtain the preliminary performance risk score, preliminary asset risk score, and preliminary environmental risk score of the target leasing project. The comprehensive decision-making and output module is used to perform weighted fusion of the preliminary performance risk score, the preliminary asset risk score and the preliminary environmental risk score according to the weight allocation vector corresponding to the target leasing project, to obtain the comprehensive risk score of the target leasing project, and to generate risk disposal suggestions for the target leasing project based on the comprehensive risk score. An adaptive optimization module is used to update the supervised fine-tuning dataset based on the comprehensive risk score, the weight allocation vector, and the risk handling suggestions to obtain an optimized risk perception vertical model.

Images