An intelligent summary automatic processing method for bad asset related cases

Abstract

The application provides an intelligent summary automatic processing method for bad asset related cases, and relates to the technical field of data processing, which comprises the following steps: Sp1: collecting original case records of bad asset related cases, wherein the original case records comprise judicial ruling texts, financial reconciliation data and protocol scan images; Sp2: constructing a distributed collaborative analysis unit and receiving the original case records; in the distributed collaborative analysis unit, the original case records are converted into a weighted characteristic matrix through a case characteristic clustering algorithm, wherein the weighted characteristic matrix comprises a core characteristic vector, an auxiliary characteristic vector and a stability factor vector; by introducing the distributed collaborative analysis unit and the k-means clustering algorithm, the technical solution realizes automatic characteristic extraction and clustering processing of large-scale case data. Compared with the inefficiency of manual review, the present application can quickly process hundreds to thousands of cases, significantly shortening the processing time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an intelligent summary automatic processing method for bad asset related cases. BACKGROUND

[0002] With the development of the bad asset management industry, especially the increasing demand for processing of bad asset related cases in the financial and judicial fields, efficiently and accurately generating intelligent summaries and achieving dynamic risk management have become important issues for improving asset disposal efficiency and reducing potential risks. Bad asset related cases involve multi-modal data such as judicial ruling texts, financial reconciliation data and agreement scan images, and the key information (such as debtors, debt amounts, collateral assets, etc.) in them directly affects the processing strategy and risk assessment of the case. However, current technologies still have many shortcomings in processing intelligent summaries and dynamic correlation analysis of these cases.

[0003] Although the existing technology has made some progress in bad asset management, it still faces the following problems:

[0004] Problem one: the traditional manual review method is time-consuming and inefficient, and is limited by personnel experience and subjectivity, making it difficult to cover large-scale case data, resulting in missing or misjudging key information.

[0005] Problem two: existing semantic extraction and image analysis techniques rely on static rules and are difficult to adapt to complex semantics in the legal and financial fields or low-quality scan images, resulting in low recognition accuracy.

[0006] Problem three: current methods lack real-time dynamic updating mechanism and rely on periodic batch processing or manual intervention, making it difficult to reflect changes in case status in a timely manner.

[0007] Therefore, an intelligent summary automatic processing method for bad asset related cases is needed to solve the above problems. SUMMARY

[0008] Technical problems to be solved

[0009] In view of the deficiencies of the prior art, the present application provides an intelligent summary automatic processing method for bad asset related cases, which solves the problems in the above background art.

[0010] Technical scheme

[0011] To achieve the above purpose, the present application is implemented by the following technical scheme: an intelligent summary automatic processing method for bad asset related cases, comprising the following steps:

[0012] Sp1: collecting original case records of bad asset related cases, the original case records including judicial ruling texts, financial reconciliation data and agreement scan images;

[0013] Sp2: constructing a distributed collaborative analysis unit to receive the original case records, in which the original case records are transformed into a weighted characteristic matrix by a case characteristic clustering algorithm, the weighted characteristic matrix including a core characteristic vector, an auxiliary characteristic vector and a stability factor vector;

[0014] Sp3: constructing a case dynamic correlation model including a plurality of correlation units, inputting the weighted characteristic matrix into the correlation units of the case dynamic correlation model to derive dynamic relationships, and outputting comprehensive summary information of the non-performing asset case by a hierarchical characteristic fusion algorithm;

[0015] Sp4: collecting actual disposal verification data samples of the non-performing asset case, the actual disposal verification data samples including disposal results and verification states, and feeding back the actual disposal verification data samples to the distributed collaborative analysis unit;

[0016] Sp5: dynamically adjusting the case dynamic correlation model by a characteristic optimization algorithm based on verification feedback to improve the accuracy and scene adaptability of the comprehensive summary information.

[0017] Preferably, the case dynamic correlation model adopts a 3:5:2 hierarchical topology structure, including 3 input layer correlation units, 5 intermediate layer correlation units and 2 output layer correlation units, and the correlation units calculate the dynamic correlation strength between case entities by the characteristic weight in the weighted characteristic matrix.

[0018] Preferably, the weighted characteristic matrix is a structured matrix describing the relationship between case entities and risk characteristics, containing case number, debtor summary, total debt, collateral asset information, correlation entity relationship, case disposal stage and risk influence weight.

[0019] Preferably, the core characteristic vector represents the key case attributes affecting the generation of the comprehensive summary information, the auxiliary characteristic vector represents the secondary influence attributes affecting the generation of the comprehensive summary information, and the stability factor vector is used to adjust the derivation threshold of the correlation unit to ensure the stability of the summary information.

[0020] Preferably, the case characteristic clustering algorithm includes the following steps:

[0021] Sp21: generating an initial characteristic set by extracting the case number, the debtor summary and the case disposal stage from the judicial ruling text through a field extraction technology based on semantic hierarchy;

[0022] Sp22: Extract the total amount of debt and transaction volatility from the financial reconciliation data through the financial extraction technology based on wavelet analysis, and the weight is 0.75 when the transaction volatility is greater than 8%, supplement the initial feature set;

[0023] Sp23: Extract the collateral asset information and associated entity relationship from the protocol scan image through the image analysis technology based on multi-region segmentation, and perfect the initial feature set;

[0024] Sp24: Through the feature weight clustering algorithm, based on the total amount of debt volatility and the stability of the value of the collateral asset, when the stability of the value of the collateral asset is greater than 90%, the weight is 0.85, the initial feature set is clustered by k-means, and a weighted feature matrix is generated. The weighted feature matrix is transmitted to the case dynamic correlation model in encrypted serialized binary format through a secure data pipeline, and the secure data pipeline is based on Salsa20 encryption.

[0025] Preferably, the hierarchical feature fusion algorithm comprises the following steps:

[0026] Sp31: In the input layer association unit, through the feature association strength calculation technology, the association strength between the debtor, the collateral asset and the associated entity in the weighted feature matrix is calculated based on Manhattan distance;

[0027] Sp32: In the middle layer association unit, through the dynamic risk assessment technology based on sliding time window, the change trend of the total amount of debt and the case disposal stage is analyzed using a 60-day window to generate a dynamic risk score;

[0028] Sp33: In the output layer association unit, through the summary content aggregation technology, the dynamic risk score and the case core attribute are integrated based on priority rules to generate comprehensive summary information, and the comprehensive summary information is transmitted to the user interaction platform in plain text format through a secure data pipeline.

[0029] Preferably, the feature optimization algorithm based on verification feedback comprises the following steps:

[0030] Sp51: Collect the disposal verification feedback data submitted by the user through the user interaction platform, the disposal verification feedback data includes the accuracy score of the comprehensive summary information and the field priority label;

[0031] Sp52: Through the feature weight reconstruction technology, the weight in the weighted feature matrix is adjusted based on the accuracy score and the field priority label, and the field weight is increased by 20% when the accuracy score is greater than 85 points;

[0032] Sp53: updating the derivation threshold and weight parameters of the correlation unit based on the treatment verification feedback data through the model parameter dynamic adjustment technology, optimizing the subsequent summary generation, and the optimized model parameters are stored in the distributed collaborative analysis unit in an encrypted serialized binary format.

[0033] Preferably, the method comprises the following stages:

[0034] A case characteristic integration stage, in which a first distributed collaborative analysis unit and a first correlation unit generate a case basic summary metadata, which contains a case number and a debtor profile;

[0035] A dynamic risk derivation stage, in which a second distributed collaborative analysis unit and a second correlation unit generate a risk correlation summary metadata, which contains a dynamic risk score;

[0036] A summary credibility optimization stage, in which a third distributed collaborative analysis unit and a third correlation unit generate a comprehensive summary credibility interval metadata, which contains a confidence evaluation of the summary content.

[0037] Preferably, the comprehensive summary credibility interval metadata is generated by a treatment scenario adaptation technology to generate sub-summary credibility interval metadata suitable for different treatment scenarios, which is used to optimize the reliability of case treatment decisions, and the treatment scenarios include liquidation and transfer.

[0038] Preferably, the system comprises:

[0039] An original case record collection module for collecting original case records of associated cases of non-performing assets, the original case records including judicial ruling texts, financial reconciliation data and protocol scan images;

[0040] A distributed collaborative analysis unit connected with the original case record collection module, for receiving the original case records and converting the original case records into a weighted characteristic matrix through a case characteristic clustering algorithm, the weighted characteristic matrix including a core characteristic vector, an auxiliary characteristic vector and a stability factor vector;

[0041] A case dynamic correlation model connected with the distributed collaborative analysis unit, the case dynamic correlation model including at least one correlation unit for receiving the weighted characteristic matrix to derive dynamic relationships and output comprehensive summary information of non-performing asset cases through a hierarchical characteristic fusion algorithm;

[0042] A treatment verification feedback module for collecting actual treatment verification data samples of non-performing asset cases, the actual treatment verification data samples including treatment results and verification states, and feeding them back to the distributed collaborative analysis unit;

[0043] The model optimization module is connected with the case dynamic correlation model and the disposal verification feedback module, and is used for dynamically adjusting the case dynamic correlation model by using a characteristic optimization algorithm based on verification feedback.

[0044] Advantages

[0045] The application provides an intelligent summary automatic processing method for associated cases of bad assets.

[0046] 1. The technical solution introduces a distributed collaborative analysis unit and a k-means clustering algorithm, realizes automatic characteristic extraction and clustering processing of large-scale case data, and compared with the inefficiency of manual review, the solution can quickly process hundreds to thousands of cases, significantly shorten the processing time, and through the generation of a weighted characteristic matrix, ensure that key information (such as the debtor, the debt amount, and the collateral assets) is fully captured and accurately classified, effectively avoid information omission or misjudgment caused by personnel experience or subjectivity, thereby greatly improving the efficiency and accuracy of asset disposal, meeting the actual needs of rapid liquidation and transfer.

[0047] 2. The technical solution adopts a hierarchical characteristic fusion algorithm (based on a 3:5:2 topology) to process multi-modal data. By integrating semantic rules and image characteristic analysis, the solution can adapt to the complex semantic environment of the legal and financial fields, especially when processing unstructured data (such as multi-language contract clauses) or low-resolution scanned images, significantly improving the recognition accuracy. Compared with the static limitations of traditional technologies, the dynamic fusion mechanism of the solution ensures the comprehensive extraction and effective use of data characteristics, providing a more reliable foundation for subsequent risk assessment and summary generation.

[0048] 3. The technical solution realizes real-time feedback optimization through the disposal verification feedback module and the model optimization module, and dynamically adjusts the risk assessment combined with the trusted interval generation function. Compared with traditional methods that rely on periodic batch processing or manual intervention, the solution can respond to case state changes (such as debtor credit fluctuations or asset value adjustments) in real time, collect feedback through a user interaction platform and optimize model parameters, ensure the timeliness of risk warning and disposal schemes, and provide more adaptive decision support in liquidation, transfer, and debt restructuring scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The specific flowchart of the application;

[0050] Figure 2 The running framework diagram of the application;

[0051] Figure 3 The scatter plot of the application;

[0052] Figure 4 This is a schematic diagram of the columnar section of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Specific Implementation Example 1:

[0055] like Figures 1 to 4 As shown, an automated intelligent summary processing method for non-performing asset-related cases is presented. This method processes the original case records of non-performing asset-related cases through distributed collaborative analysis units and a dynamic case association model, generating accurate and scenario-adaptive comprehensive summary information to optimize case management and disposal decisions for financial institutions and judicial institutions. The method integrates multimodal data processing, feature clustering, dynamic relationship derivation, feedback optimization, and confidence interval assessment to form a coherent closed-loop process. The complete working method is described in detail below.

[0056] Complete workflow:

[0057] Data Acquisition and Preprocessing:

[0058] The system acquires original case records of non-performing asset-related cases through the original case record collection module. These original case records include three types of data: judicial ruling texts, financial reconciliation data, and scanned images of agreements. The judicial ruling texts are text files extracted from the judicial system database, containing the case number (unique identifier, such as A2025001), debtor summary (name, identification), and case stage (in litigation, enforcement, or closed). The financial reconciliation data are tables or database records exported from financial institutions' financial systems, containing the total debt (outstanding amount, such as 1 million yuan), trading volatility (based on the standard deviation of 12 months of trading records), and repayment records. The scanned images of agreements are high-resolution images (PDF or JPEG format) acquired from scanning equipment, containing information on collateral assets (collateral type, valuation, ownership status) and related entity relationships (guarantors, joint debtors).

[0059] The collection module acquires data from the judicial system, financial system, and scanning devices through dedicated interfaces (such as APIs or file imports), ensuring data integrity. After collection, the system pre-processes the data, including format verification (verifying file encoding and integrity), deduplication (removing duplicate records), and noise filtering (removing irrelevant fields such as headers and footers). The pre-processed original case records are transmitted to the distributed collaborative analysis unit in an encrypted serialized binary format (based on the Salsa20 encryption algorithm with a 256-bit key) through a secure data pipeline. The secure data pipeline is a high-performance, low-latency communication channel that uses Salsa20 encryption to ensure the confidentiality and integrity of data during transmission, preventing data leakage or tampering. This step provides standardized multi-modal data input for subsequent analysis, which is applied to case information integration and risk assessment.

[0060] Data property clustering and matrix generation:

[0061] In the distributed collaborative analysis unit, the system receives the original case records and converts the multi-modal data into a weighted property matrix through the case property clustering algorithm, providing a structured input for dynamic relationship derivation. The distributed collaborative analysis unit is a distributed computing framework composed of multiple computing nodes, each equipped with independent processors and storage units. It synchronizes data and computing results through an improved Paxos consensus protocol, ensuring high availability and fault tolerance. The weighted property matrix is a structured matrix that describes the relationship between case entities (debtors, collateral assets, associated entities) and risk properties, containing the following fields:

[0062] Case number: a unique identifier for the case, such as A2025001. Debtor summary: debtor name and identity. Total debt: the current outstanding amount, such as 1 million yuan. Collateral asset information: type of collateral (such as real estate), estimated value (such as 5 million yuan), and ownership status (such as no dispute). Associated entity relationship: relationship strength of guarantors or joint debtors (such as guarantor assumes 50% of debt). Case disposition stage: litigation, execution, or case closure. Risk impact weight: a quantitative indicator based on volatility and stability (0.0-1.0).

[0063] The implementation steps of the case property clustering algorithm are as follows:

[0064] Sp21 - Semantic hierarchical field extraction: Extract case number, debtor summary, and case disposition stage from judicial ruling text. The system uses a special semantic rule set for non-performing asset cases (containing legal terms and field mapping tables, such as "debtor" mapped to "debtor") to identify key fields through semantic hierarchical technology (combining word embedding and rule matching). For example, from the text "Case number: A2025001, Debtor: Zhang, Status: In execution", the corresponding fields are extracted to generate an initial property set.

[0065] Sp22 - Volatility Analysis Financial Extraction: Extract total debt amount and transaction volatility rate from financial statement data. The system calculates the standard deviation of transaction amounts for the last 12 months as the volatility rate. If the volatility rate is greater than 8%, the weight is assigned as 0.75, reflecting a high-risk characteristic. For example, if the total debt amount is 1 million yuan and the volatility rate is 10%, the weight is 0.75, supplementing the initial characteristic set.

[0066] Sp23 - Multi-region Segmentation Image Parsing: Extract collateral asset information and associated entity relationships from protocol scan images. The system uses a multi-scale region segmentation technique (based on the watershed algorithm) to locate text regions (such as collateral clauses) in contract images, and parses the mortgage type, valuation, and guarantor information through optimized character recognition technology (based on Tesseract, combined with a domain dictionary). For example, extract "Mortgage: Real Estate, Valuation: 5 million yuan, Guarantor: Li", and complete the initial characteristic set.

[0067] Sp24 - Characteristic Weight Clustering: Based on the total debt amount volatility rate and the stability of collateral asset value (calculated through the valuation data of the last 6 months, with a stability greater than 90% and a weight of 0.85), the initial characteristic set is clustered using the k-means algorithm (k=3). The k-means algorithm groups according to the risk impact degree of the fields (the weighted sum of volatility and stability), generating a weighted characteristic matrix. For example, the total debt amount and collateral asset information are classified as a high-risk cluster, with weights of 0.75 and 0.85, respectively. The weighted characteristic matrix is transmitted to the case dynamic correlation model in encrypted serialized binary format through a secure data pipeline (Salsa20 encryption).

[0068] This step converts multi-modal data into a structured weighted characteristic matrix, highlighting case risk characteristics, optimizing data organization efficiency, and providing high-quality input for dynamic relationship derivation, which is applied to risk assessment and summary generation.

[0069] Dynamic Relationship Derivation and Summary Generation:

[0070] The system builds a case dynamic correlation model to process the weighted characteristic matrix and generate comprehensive summary information for non-performing asset cases. The case dynamic correlation model is a multi-level computing framework with a 3:5:2 hierarchical topology structure, including 3 input layer correlation units, 5 intermediate layer correlation units, and 2 output layer correlation units. Each correlation unit is an independent computing entity containing processing logic and parameter storage, running on the computing nodes of the distributed collaborative analysis unit to improve processing speed through parallel computing. The input layer processes the initial input of the characteristic matrix, the intermediate layer analyzes dynamic risk trends, and the output layer integrates summary content.

[0071] The dynamic relationship derivation is achieved through a hierarchical feature fusion algorithm, with the following specific steps:

[0072] Sp31 - Property correlation strength calculation: In the input layer correlation unit, based on Manhattan distance, the correlation strength between the debtor, collateral assets and related entities in the weighted property matrix is calculated. Manhattan distance quantifies the relationship between entities by comparing the absolute difference of property vectors (such as total debt, collateral asset valuation). For example, the debtor and the guarantor share 50% of the debt, the Manhattan distance is small, and the correlation strength is 0.9.

[0073] Sp32 - Dynamic risk assessment: In the middle layer correlation unit, using a dynamic risk assessment technology based on a sliding time window, the trend of total debt and case handling stage within a 60-day window is analyzed, and a dynamic risk score is generated. The system uses an exponentially weighted moving average (weight decay factor 0.9) to calculate the trend slope, and if the total debt increases by more than 5%, the risk score increases by 0.2. For example, the total debt increases from 1 million yuan to 1.1 million yuan, and the risk score is 0.6.

[0074] Sp33 - Summary content aggregation: In the output layer correlation unit, through the summary content aggregation technology, based on priority rules (core property vector weight greater than 0.7 priority), the dynamic risk score and the case core attributes are integrated to generate comprehensive summary information. Comprehensive summary information includes case number, debtor summary, total debt, collateral asset information, related entity relationship, case handling stage and dynamic risk score. For example, "Case Number: A2025001, Debtor: Zhang, Total Debt: 1 million yuan, Collateral Asset: Real Estate (Valuation 5 million yuan), Risk Score: 0.6". Comprehensive summary information is transmitted in plain text format through a secure data pipeline (Salsa20 encryption) to a user interaction platform, a web-based interface that allows users to review the content.

[0075] This step generates structured and dynamic comprehensive summary information, highlighting key case information and risk trends, and is applied to case management and rapid decision-making by financial institutions and judicial personnel.

[0076] Disposal verification feedback and model optimization:

[0077] The system collects actual disposal verification data samples of non-performing asset cases through the disposal verification feedback module to optimize the accuracy and scenario adaptability of the comprehensive summary information. The disposal verification feedback module collects the following data from the user interaction platform:

[0078] Disposal result: The actual disposal method (liquidation or transfer) and result of the case (such as recovery amount 50 million yuan).

[0079] Verification status: User's accuracy rating of the comprehensive summary information (0-100 points) and field priority labels (high, medium, and low priority, such as "total debt" with high priority).

[0080] The processed verification data samples are fed back to the distributed collaborative analysis unit in encrypted, serialized binary format via a secure data pipeline. The system adjusts the dynamic case correlation model using a feature optimization algorithm based on the verification feedback. The specific steps are as follows:

[0081] SP51 - Feedback Data Collection: Obtain accuracy scores and field priority labels from user interaction platforms. For example, a user scores 92 points in the Total Debt field, which is marked as high priority.

[0082] SP52 - Feature Weight Restructuring: Adjusts the weights in the weighted feature matrix based on accuracy scores and priority labels. If the accuracy score is greater than 85, the field weight is increased by 20% (e.g., from 0.75 to 0.9); the weight of high-priority fields is increased by an additional 10% (e.g., the weight of total debt increases from 0.9 to 0.99).

[0083] Sp53 - Dynamic Model Parameter Adjustment: Based on disposal validation feedback data, the derivation thresholds and weight parameters of associated units are updated using a gradient update technique (step size 0.01). For example, the risk score threshold is adjusted from 0.5 to 0.55 to optimize subsequent summary generation. The optimized model parameters are stored in the distributed collaborative analysis unit in an encrypted serialized binary format.

[0084] This step optimizes the model based on the actual processing results, ensuring that the summary content better meets user needs and is applied to improve summary quality and decision support.

[0085] Credible interval generation and scenario adaptation:

[0086] To improve the reliability of the comprehensive summary information, the system performs phased processing within a distributed collaborative analysis unit to generate a confidence interval assessment. The processing phases include:

[0087] Case characteristic integration period: In the first distributed collaborative analysis unit and the first association unit, basic case summary metadata is generated, including case number and debtor summary, and preliminary information is integrated.

[0088] Dynamic risk derivation period: In the second distributed collaborative analysis unit and the second association unit, risk association summary metadata is generated, which includes dynamic risk scores and reflects risk trends.

[0089] Abstract confidence optimization phase: In the third distributed collaborative analysis unit and the third correlation unit, the comprehensive abstract confidence interval metadata is generated, which contains the confidence assessment of the abstract content. The confidence is calculated through statistical analysis (mean ± 2 standard deviation), for example, the total debt confidence interval is [0.95 million, 1.05 million].

[0090] The comprehensive abstract confidence interval metadata is generated by the disposal scene adaptation technology to generate sub-abstract confidence interval metadata suitable for different disposal scenes. Disposal scenes include liquidation (debt recovery) and transfer (asset sale). The adaptation technology establishes mapping rules based on historical disposal data (the last 1000 cases), such as prioritizing total debt (weight 0.6) and collateral asset information (weight 0.3) in the liquidation scene, and prioritizing collateral asset valuation (weight 0.5) in the transfer scene. The sub-abstract confidence interval metadata is stored in the user interaction platform in plain text format to guide the reliability of liquidation or transfer decisions. For example, in the liquidation scene, the abstract shows "total debt: 1 million, confidence interval [0.95 million, 1.05 million], suggest prioritizing collection".

[0091] System composition and overall application:

[0092] The system of the method includes the following components:

[0093] Original case record acquisition module: Collect judicial ruling text, financial reconciliation data and agreement scan images through API or file import, providing multi-modal data input.

[0094] Distributed collaborative analysis unit: Distributed computing framework, running case characteristic clustering and feedback optimization, managing data processing and storage, using Paxos protocol to ensure consistency.

[0095] Case dynamic correlation model: 3:5:2 hierarchical topology, processing weighted characteristic matrix, generating comprehensive abstract information.

[0096] Disposal verification feedback module: Collect disposal results and verification status from the user interaction platform to optimize abstract content.

[0097] Model optimization module: Adjust model parameters based on feedback to improve abstract accuracy.

[0098] User interaction platform: Web interface, display abstract information, collect user feedback.

[0099] Data transmission is through a secure data pipeline (Salsa20 encryption), and the format is serialized binary (internal processing) or plain text (user output). This method is applied to case management, risk assessment and disposal decision-making of financial institutions and judicial institutions. Comprehensive abstract information provides an overview of key information, sub-abstract confidence interval guides liquidation or transfer strategy, and optimized model ensures long-term abstract quality. Specific embodiment two:

[0101] As Figures 1 to 4 shown below, the key algorithms mentioned in the above embodiments are analyzed in detail, including their core mathematical formulas and explanations:

[0102] Case characteristic clustering algorithm:

[0103] Input data: The algorithm runs in the distributed collaborative analysis unit and receives multi-modal data provided by the original case record collection module, including judicial ruling text, financial reconciliation data, and agreement scan images. The judicial ruling text is a pure text file (TXT format) containing the case number (unique identifier, such as A2025001), the debtor summary (name, identity), and the case disposition stage (in litigation, in execution, or closed). The financial reconciliation data is a table or database record (CSV format) containing the total debt (unpaid amount, such as 1 million yuan), transaction volatility (standard deviation of transaction amount in the last 12 months), and repayment records. The agreement scan image is a high-resolution image (PDF or JPEG format) containing collateral asset information (such as property type, such as real estate, valuation, and ownership status) and associated entity relationships (guarantor, joint debtor). The data is pre-processed (format verification, de-duplication, noise filtering, such as removing text headers) and transmitted to the distributed collaborative analysis unit in encrypted serialized binary format (Salsa20 encryption, 256-bit key) through a secure data pipeline. For example, the input data may include: judicial ruling text "Case number: A2025001, debtor: Zhang, disposition stage: execution, judgment amount: 120 million yuan", financial reconciliation data "total debt: 120 million yuan, transaction volatility: 6%, latest repayment: 0 yuan", agreement scan image "collateral: Shanghai apartment, valuation: 150 million yuan, ownership: no dispute, no associated entities".

[0104] Output results: The algorithm generates a weighted characteristic matrix, a structured matrix describing the relationship between case entities (debtors, collateral assets, associated entities) and risk characteristics, containing fields: case number (such as A2025001), debtor summary (such as Zhang), total debt (such as 120 million yuan), collateral asset information (such as real estate, valuation 150 million yuan), associated entity relationships (such as no associated entities), case disposition stage (such as execution), risk impact weight (based on volatility and stability, range 0.0-1.0, such as 0.5-0.85). The matrix is transmitted to the case dynamic correlation model in encrypted serialized binary format through a secure data pipeline (Salsa20 encryption). For example, the output matrix is:

[0105] Case ID: A2025001; Debtor Profile: Zhang; Total Debt: 1.2 million (weight 0.5); Secured Asset Information: Shanghai Apartment, Valuation 1.5 million (weight 0.85); Related Entity Relationship: None (weight 0.0); Case Disposal Stage: Execution; Risk Impact Weight: 0.5-0.85.

[0106] Specific Application: Case characteristic clustering algorithm is the core of data preprocessing method, running in distributed collaborative analysis unit, responsible for converting multi-modal original case records into structured and efficient weighted characteristic matrix, providing input for subsequent dynamic relationship derivation. The algorithm achieves the following steps: First, use the semantic rules set (including legal terms and field mapping table, such as "debtor" mapping to "debtor") dedicated to non-performing asset cases, extract case number, debtor profile and case disposal stage from judicial ruling text through semantic hierarchical technology (combined with word embedding and rule matching), generate initial characteristic set, such as "case number: A2025001, debtor: Zhang, status: execution". Then, calculate transaction volatility (12-month transaction amount standard deviation) from financial reconciliation data, if volatility is greater than 8%, weight is 0.75, otherwise 0.5, such as volatility 6%, weight 0.5, supplement characteristic set. Then, use multi-scale region segmentation technology (based on watershed algorithm) to locate the text area of the protocol scanning image, extract secured asset information (such as "collateral: apartment, valuation: 1.5 million") and related entity relationship through optimized character recognition technology (based on Tesseract, combined with domain dictionary), and improve the characteristic set. Finally, based on the volatility of total debt and the stability of secured asset value (6-month valuation fluctuation less than 5%, weight 0.85), the characteristic set is clustered by k-means (k=3) according to the risk impact degree (such as high-risk cluster containing total debt), and the weighted characteristic matrix is generated. In the whole method, this algorithm reduces the data dimension, optimizes the data organization, and provides structured input for the case dynamic correlation model, which is applied to risk assessment and summary generation. For example, in the bank clearing scenario, the matrix highlights the total debt and secured asset information, supporting quick decision-making.

[0107] Hierarchical characteristic fusion algorithm:

[0108] Input data: The algorithm runs in the case dynamic correlation model (3:5:2 hierarchical topology, containing 3 input layer correlation units, 5 intermediate layer correlation units and 2 output layer correlation units), receives the weighted characteristic matrix generated by the case characteristic clustering algorithm. The matrix is input in encrypted serialized binary format (Salsa20 encryption) through a secure data pipeline, containing case number, debtor profile, total debt, secured asset information, related entity relationship, case disposal stage and risk impact weight. For example, the input matrix is:

[0109] Case Number: A2025001; Debtor Profile: Zhang; Total Debt: 1.2 million yuan (weight 0.5); Secured Asset Information: Shanghai apartment, estimated value 1.5 million yuan (weight 0.85); Related Entity Relationship: None (weight 0.0); Case Disposal Stage: Execution; Risk Impact Weight: 0.5-0.85.

[0110] Output Result: The algorithm generates a comprehensive summary information, a structured output in plain text format, containing case number, debtor profile, total debt, secured asset information, related entity relationship, case disposal stage, and dynamic risk score (range 0.0-1.0). The comprehensive summary information is transmitted to the user interaction platform (Web interface) through a secure data pipeline (Salsa20 encryption). For example, the output summary is:

[0111] Case Number: A2025001; Debtor Profile: Zhang; Total Debt: 1.2 million yuan; Secured Asset Information: Shanghai apartment, estimated value 1.5 million yuan; Related Entity Relationship: None; Case Disposal Stage: Execution; Dynamic Risk Score: 0.3.

[0112] Specific Application: The hierarchical characteristic fusion algorithm is the core analysis link of the method, running in the case dynamic correlation model, responsible for processing the weighted characteristic matrix, generating comprehensive summary information, and applied to case management and rapid decision-making. The algorithm is implemented through the following steps: In the input layer correlation unit, based on the Manhattan distance, the correlation strength between the debtor, secured assets and related entities is calculated, and the absolute difference of characteristic vectors (such as total debt, secured asset valuation) is used to quantify the relationship, such as a single debtor with no related entities, with a strength of 0.9. In the middle layer correlation unit, using the dynamic risk assessment technology based on sliding time window (60-day window, exponential weighted moving average, weight decay factor 0.9), the change trend of total debt and case disposal stage is analyzed, if the total debt grows more than 5%, the risk score increases by 0.2, if there is no growth, the score is 0.3. In the output layer correlation unit, through the summary content aggregation technology, based on the priority rule (core characteristic vector weight greater than 0.7 priority), the dynamic risk score and the core attributes of the case are integrated to generate comprehensive summary information. In the whole method, this algorithm converts the structured matrix into user-readable summary, highlighting the key information and risk trends. For example, in the asset transfer scenario, the summary shows "Total Debt: 200 million yuan, Risk Score: 0.6", guiding the asset management company to decide on the transfer strategy. The comprehensive summary information is displayed through the user interaction platform, supporting financial institutions and judicial personnel to quickly assess the case.

[0113] Characteristic Optimization Algorithm Based on Verification Feedback:

[0114] Input data: The algorithm runs on the distributed collaborative analysis unit, receiving actual treatment verification data samples collected by the treatment verification feedback module from the user interaction platform. The data includes treatment results (treatment methods such as liquidation or transfer, recovery amount) and verification status (accuracy score 0-100, field priority label such as high, medium, and low priority). The data is transmitted in encrypted serialized binary format (Salsa20 encryption) through a secure data pipeline. For example, the input feedback is:

[0115] Treatment result: liquidation recovery of 1 million yuan; verification status: accuracy score 90, total debt high priority;

[0116] Output result: The algorithm generates optimized model parameters, including updated weighted feature matrix weights and derived threshold values and weight parameters of associated units. The optimized parameters are stored in the distributed collaborative analysis unit in encrypted serialized binary format. For example, the output is:

[0117] Updated weight: total debt weight increased from 0.5 to 0.6 (20% increase); updated threshold: risk score threshold increased from 0.3 to 0.35.

[0118] Specific application: The feature optimization algorithm based on verification feedback is the optimization link of the method, running on the distributed collaborative analysis unit, adjusting the case dynamic association model through user feedback, and improving the accuracy and scene adaptability of the comprehensive summary information. The algorithm is implemented through the following steps: first, collect the feedback data submitted by the user interaction platform, such as accuracy score 90 and total debt high priority. Second, adjust the weighted feature matrix weights based on the score and priority through feature weight reconstruction technology. If the score is greater than 85, the field weight is increased by 20% (e.g., total debt from 0.5 to 0.6), and the high priority field is additionally increased by 10% (e.g., to 0.66). Finally, update the derivation threshold and weight parameters of the associated units using gradient update (step size 0.01) through model parameter dynamic adjustment technology, such as risk score threshold from 0.3 to 0.35. After storing the optimized parameters, they are applied to subsequent summary generation. In the entire method, this algorithm optimizes the model based on actual treatment results (such as liquidation recovery amount), ensuring that the summary is more user-oriented. For example, in the debt restructuring scenario, feedback increases the weight of associated entities, generating a summary that focuses more on multi-party relationships, optimizing court coordination efficiency. This algorithm forms a closed-loop optimization mechanism, improving the applicability of the method in different scenarios (such as liquidation, transfer). Specific embodiment three:

[0120] As shown in Figures 1 to 4 , the following is a detailed hardware composition and hardware description of the content in embodiment one:

[0121] The hardware system consists of a data collection server, a distributed computing cluster, a storage server, a user interaction server, network devices, and scanning devices, which collectively support data collection, processing, analysis, feedback optimization, and user interaction functions. The data collection server is responsible for acquiring multi-modal data and performing preprocessing, the distributed computing cluster executes algorithm processing, the storage server saves data and model parameters, the user interaction server provides summary display and feedback collection, and the network devices ensure secure data transmission, and the scanning devices generate image data. All data transmission is performed through a secure data pipeline (based on Salsa20 encryption, 256-bit key), with a format of serialized binary (internal processing) or plain text (user output).

[0122] The data collection server is equipped with 2 eight-core processors (such as Intel Xeon Silver 4210, 2.2 GHz), 64 GB DDR4 ECC memory, 2 TB NVMe solid state drive (read and write speed greater than 3000 MB / s), and 10 Gb Ethernet card, running Linux system (Ubuntu Server 20.04). It implements the original case record collection module, acquires judicial ruling text (TXT format, containing case number such as A2025001, debtor summary such as Zhang, disposal stage such as execution) from the judicial system database through API interface (RESTful or SOAP), acquires financial reconciliation data (CSV format, containing total debt such as 1.2 million yuan, transaction volatility such as 6%) from the financial institution financial system, and acquires agreement scanning images (PDF or JPEG format, containing collateral information such as real estate valuation of 1.5 million yuan) from the scanning device. The server performs preprocessing tasks, including format verification (verifying file encoding and integrity), deduplication (removing duplicate records), and noise filtering (removing irrelevant fields such as headers), and the preprocessed data is transmitted to the distributed computing cluster in serialized binary format through the secure data pipeline. The server supports input preparation for case characteristic clustering algorithms, ensures uniformity of multi-modal data formats, and is applied to case information integration and risk assessment foundation. For example, it processes input data "Case Number: A2025001, Debtor: Zhang, Disposal Stage: Execution, Judgment Amount: 1.2 million yuan" (TXT), "Total Debt: 1.2 million yuan, Transaction Volatility: 6%" (CSV), and "Collateral: Shanghai apartment, Valuation: 1.5 million yuan" (PDF).

[0123] The distributed computing cluster consists of at least 5 computing nodes, each equipped with 2 16-core processors (such as AMD EPYC 7313, 3.0 GHz), 128 GB DDR4 ECC memory, 4 TB NVMe solid-state hard drives (read and write speeds greater than 3500 MB / s), and 25 Gb Ethernet interconnection (combined with low-latency InfiniBand). The cluster is managed by the Kubernetes management system, which uses an improved Paxos consensus protocol to synchronize data and tasks, ensuring high availability and fault tolerance. The cluster runs the distributed collaborative analysis unit and the case dynamic correlation model, executes the case characteristic clustering algorithm, the hierarchical characteristic fusion algorithm, and the characteristic optimization algorithm based on verification feedback. The case characteristic clustering algorithm processes multi-modal data, extracts semantic hierarchical fields (based on word embedding and rule matching, processing judicial ruling text), volatility analysis (calculating transaction volatility, such as 6% weight 0.5), multi-region segmented image analysis (based on watershed algorithm, extracting collateral information), and k-means clustering (k=3, generating a weighted characteristic matrix), outputting a matrix containing case number, total debt (weight 0.5), collateral asset information (weight 0.85), etc. The case dynamic correlation model (3:5:2 hierarchical topology, 3 input layers, 5 intermediate layers, and 2 output layer correlation units) executes the hierarchical characteristic fusion algorithm, calculates the correlation strength (such as the strength of the debtor and the collateral asset 0.9) through Manhattan distance, evaluates the risk in a 60-day sliding window (exponential weighted moving average, weight decay 0.9, generating risk score 0.3), and aggregates priority rules to generate comprehensive summary information (such as "Case Number: A2025001, Total Debt: 1.2 million yuan, Risk Score: 0.3"). The characteristic optimization algorithm processes feedback data, adjusts matrix weights and model parameters, and stores them in the cluster. The cluster memory accelerates computation, and the solid-state hard drive caches intermediate data (such as the weighted characteristic matrix). Data is transmitted in serialized binary format and applied to risk assessment, summary generation, and model optimization. For example, in the liquidation scenario, the cluster generates summaries to guide banks to recover debts.

[0124] The storage server is equipped with a 12-core processor (e.g., Intel Xeon Gold 5318Y, 2.1 GHz), 256 GB DDR4 ECC memory, a 20 TB RAID-6 hard disk array (7200 RPM, read / write speed about 200 MB / s), and a 1 TB NVMe solid-state drive cache (read / write speed greater than 3000 MB / s), supporting 10 Gb Ethernet. The server stores raw case records, weighted feature matrices, integrated summary information, and optimized model parameters, using a distributed file system (e.g., Ceph) to ensure scalability and redundancy. Data is stored in encrypted serialized binary format, with solid-state drive cache for recent data (e.g., the latest matrix) and hard disk array for historical data. The server supports input / output storage for all algorithms, such as storing matrices for case feature clustering algorithms, summaries for hierarchical feature fusion algorithms, and parameters for feature optimization algorithms, applied to historical data analysis and long-term summary quality improvement. For example, it saves the matrix and summary of "Case Number: A2025001" to support subsequent case comparison.

[0125] The user interaction server is equipped with an 8-core processor (e.g., Intel Xeon Silver 4210, 2.2 GHz), 64 GB DDR4 memory, a 1 TB NVMe solid-state drive, and a 10 Gb Ethernet card, running a Linux system and a web server (e.g., Nginx). It implements a user interaction platform, a web application based on RESTful API to display integrated summary information (plain text format, such as "Total Debt: 1.2 million yuan, Risk Score: 0.3") and collect disposition verification feedback (accuracy score such as 90 points, field priority such as high priority for total debt). Feedback data is transmitted to the distributed computing cluster in serialized binary format, supporting weight reconstruction and parameter adjustment for feature optimization algorithms. The server supports concurrent user access and is applied to case management and decision support. For example, in the transfer scenario, users review the decision to sell assets, and feedback optimizes the accuracy of subsequent summaries.

[0126] The network equipment includes enterprise-level switches (e.g., Cisco Catalyst 9300, supporting 25 Gb Ethernet and InfiniBand) and hardware firewalls (e.g., Fortinet FortiGate 100F), providing 1 Gbps external bandwidth and 10 Gbps internal bandwidth. The switch supports low-latency communication between cluster nodes, and the firewall enforces Salsa20 encryption and access control to ensure data transmission security. The network equipment supports secure data pipelines to transmit raw data, matrices, summaries, and feedback data, applied to data interaction between all modules. For example, it ensures encrypted transmission between the collection server and the cluster.

[0127] The scanning device is a high-resolution scanner (e.g., Canon DR-C240, resolution 600 dpi) that generates a protocol scan image (PDF or JPEG format, containing collateral information such as "house, estimated value 1.5 million yuan") connected to the data collection server and applied to the secured asset information collection. For example, it scans the contract to generate an image for the case feature clustering algorithm to analyze. Embodiment Four:

[0129] As Figures 1 to 4 shown below are specific use cases:

[0130] Case 1: Single Loan Liquidation Scenario:

[0131] Background: A bank needs to handle a non-performing loan case involving a single debtor and a mortgaged property, with the goal of liquidating the debt to recover funds.

[0132] Input Data:

[0133] Judicial Ruling Text: Text file (TXT format) containing "Case Number: L2025001, Debtor: Zhang (ID No: 1234567890), Disposal Stage: Execution, Judgment Amount: 1.2 million yuan."

[0134] Financial Reconciliation Data: Table (CSV format) containing "Total Debt: 1.2 million yuan, 12-month transaction volatility rate: 6%, Latest Repayment: 0 yuan."

[0135] Protocol Scan Image: PDF image containing "Collateral: Shanghai apartment, estimated value: 1.5 million yuan, ownership: no dispute, no associated entities."

[0136] Processing Flow:

[0137] Data Collection and Preprocessing: The original case record collection module obtains data from the bank system and court database through API, checks the format (TXT, CSV, PDF valid), removes duplicate records (no duplicates), filters noise (such as text headers). Data is transmitted to the distributed collaborative analysis unit in encrypted serialized binary format (Salsa20 encryption, 256-bit key) through a secure data pipeline.

[0138] Feature Clustering and Matrix Generation: The distributed collaborative analysis unit (running on 5 computing nodes, Paxos protocol synchronization) executes the case feature clustering algorithm:

[0139] Sp21: Semantic hierarchical field extraction, extract "Case Number: L2025001, Debtor: Zhang, Disposition Stage: In Execution" from judicial ruling text, generate initial characteristic set. Sp22: Volatility analysis finance extraction, calculate transaction volatility rate as 6% (less than 8%, weight 0.5), total debt amount 1.2 million yuan, supplement initial characteristic set. Sp23: Multi-region segmentation image analysis, extract "Collateral: Apartment, Valuation: 1.5 million yuan" from protocol scan image, no associated entity, refine initial characteristic set. Sp24: Characteristic weight clustering, based on total debt volatility rate (6%) and secured asset stability (valuation volatility <5%, weight 0.85), k-means clustering (k=3) generates weighted characteristic matrix, containing fields: case number, debtor, total debt amount, collateral information, disposition stage, risk weight (0.5-0.85). The matrix is transmitted to the case dynamic association model through a secure data pipeline in a serialized binary format.

[0140] Dynamic relationship derivation and summary generation: The case dynamic association model (3:5:2 topology, 3 input layers, 5 intermediate layers, 2 output layer association units) executes hierarchical characteristic fusion algorithm:

[0141] Sp31: Characteristic association strength calculation, based on Manhattan distance to calculate the association between debtor and collateral (single debtor, no associated entity, strength 0.9). Sp32: Dynamic risk assessment, use 60-day sliding window to analyze total debt amount (no growth, risk score 0.3).

[0142] Sp33: Summary content aggregation, integrate fields (priority: total debt amount 0.85, collateral 0.7), generate comprehensive summary information: "Case Number: L2025001, Debtor: Zhang, Total Debt Amount: 1.2 million yuan, Collateral: Shanghai Apartment (valuation 1.5 million yuan), Disposition Stage: In Execution, Risk Score: 0.3". The summary is transmitted to the user interaction platform in plain text format through a secure data pipeline.

[0143] Confidence interval generation: Distributed collaborative analysis unit processes in stages:

[0144] Case characteristic integration period: Generate case basic summary metadata (case number, debtor).

[0145] Dynamic risk derivation period: Generate risk association summary metadata (risk score 0.3).

[0146] Abstract Confidence Optimization Phase: Generate comprehensive abstract confidence interval metadata, total debt confidence interval [115 million, 125 million] (mean ± 2 standard deviations). Disposal scenario adaptation technology assigns weights to liquidation scenarios (total debt 0.6, collateral assets 0.3), generating sub-abstract confidence intervals: "Total debt: 120 million [115 million, 125 million], recommended liquidation."

[0147] Feedback and Optimization: The disposal verification feedback module collects bank feedback: disposal result "liquidation recovery 100 million", accuracy score 90 points, high priority for total debt. Feature optimization algorithm based on verification feedback:

[0148] Sp51: Collect feedback data (score 90 points, high priority for total debt). Sp52: weight reconstruction, total debt weight increased from 0.85 to 1.02 (increased by 20%). Sp53: parameter adjustment, update associated unit threshold (risk score threshold increased from 0.3 to 0.35). Optimized parameters are stored in serialized binary format.

[0149] Output Results:

[0150] Comprehensive abstract information (plain text): "Case number: L2025001, debtor: Zhang, total debt: 120 million, collateral assets: Shanghai apartment (estimated value 150 million), disposal stage: execution, risk score: 0.3".

[0151] Sub-abstract confidence interval (plain text): "Total debt: 120 million [115 million, 125 million], recommended liquidation".

[0152] Application scenario: the bank uses the comprehensive abstract to quickly evaluate the case and confirms the low risk (score 0.3), combined with the confidence interval (stable total debt) to decide on liquidation strategy, prioritizing disposal of mortgaged properties, recovering 100 million of the 120 million. After feedback optimization, the system increases the weight of total debt and generates more accurate subsequent abstracts.

[0153] Case 2: Complex Asset Transfer Scenario:

[0154] Background: An asset management company needs to handle a non-performing asset involving multiple associated entities, aiming to transfer assets to reduce risk exposure.

[0155] Input Data:

[0156] Judicial ruling text: TXT file, content includes "Case number: T2025002, debtor: Li (ID number: 0987654321), disposal stage: litigation, judgment amount: 200 million, associated entity: guarantor Wang".

[0157] Financial reconciliation data: CSV table containing "Total debt: 2 million, transaction volatility: 12%, recent repayment: 100,000".

[0158] Agreement scan image: JPEG image containing "Collateral: commercial property, estimated value: 2.5 million, ownership: in dispute, guarantor: Wang (responsible for 30% debt)".

[0159] Processing flow:

[0160] Data acquisition and preprocessing: The original case record acquisition module obtains data through API, checks the format, removes noise (image watermark). The data is transmitted to the distributed collaborative analysis unit in serialized binary format (Salsa20 encryption).

[0161] Characteristic clustering and matrix generation:

[0162] Sp21: Semantic hierarchical extraction "Case number: T2025002, debtor: Li, disposal stage: litigation, guarantor: Wang". Sp22: Volatility analysis extracts total debt of 200 million, volatility of 12% (greater than 8%, weight 0.75). Sp23: Image analysis extracts "collateral: commercial property, estimated value: 2.5 million, guarantor: Wang (30% debt)". Sp24: k-means clustering (k=3), based on volatility 12% (weight 0.75) and collateral asset stability (estimated value volatility 10%, weight 0.7), generate a weighted characteristic matrix containing case number, debtor, total debt, collateral asset, associated entity, disposal stage, risk weight. The matrix is transmitted in serialized binary format.

[0163] Dynamic relationship derivation and summary generation:

[0164] Sp31: Manhattan distance calculates the association strength between debtor Li and guarantor Wang (30% debt, strength 0.7). Sp32: 60-day window analysis of total debt (8% growth, risk score 0.6). Sp33: Aggregate to generate summary: "Case number: T2025002, debtor: Li, total debt: 200 million, collateral asset: commercial property (estimated value 2.5 million, in dispute), associated entity: Wang (30% debt), disposal stage: litigation, risk score: 0.6". The summary is transmitted to the user interaction platform in plain text.

[0165] Confidence interval generation:

[0166] Case characteristic integration period: Generate basic summary metadata (case number, debtor).

[0167] Dynamic risk derivation period: Generate risk summary metadata (risk score 0.6).

[0168] Abstract Confidence Optimization Phase: Generate confidence interval, total debt [190M, 210M]. Disposition scenario adapted to transfer scenario (secured asset weight 0.5), sub-confidence interval: "Secured Asset: 25M [24M, 26M], recommended for transfer".

[0169] Feedback and Optimization: Feedback module collects disposition result "Transfer asset, recover 18M", score 85, secured asset high priority. Optimization algorithm:

[0170] Sp51: Collect feedback (score 85, secured asset high priority). Sp52: Secured asset weight increased from 0.7 to 0.84 (20% increase). Sp53: Update risk score threshold (0.6 to 0.65). Parameters stored as serialized binary.

[0171] Output Results:

[0172] Comprehensive abstract information: "Case number: T2025002, debtor: Li, total debt: 20M, secured asset: commercial property (estimated value 25M, disputed), associated entity: Wang (30% debt), disposition stage: litigation, risk score: 0.6".

[0173] Sub-confidence interval: "Secured asset: 25M [24M, 26M], recommended for transfer".

[0174] Application scenario: Asset management company assesses high risk (score 0.6, ownership dispute) based on abstract and confidence interval, decides to transfer commercial property, recovers 18M. After optimization, the system increases the weight of secured assets and generates an abstract that pays more attention to asset valuation. Specific embodiment five:

[0176] As Figures 1 to 4 shown below is a supplement to the above embodiment content:

[0177] Input and output relationship between modules:

[0178] The modules (original case record acquisition module, distributed collaborative analysis unit, case dynamic correlation model, disposal verification feedback module, model optimization module, user interaction platform) form a closed-loop processing flow through clear data flow, ensuring smooth transmission from data acquisition to summary generation to optimization. The original case record acquisition module obtains text format judicial ruling data (including case number such as L2025001, debtor such as Zhang, disposal stage such as execution) from the judicial system database, obtains table format financial reconciliation data (including total debt such as 1.2 million yuan, transaction volatility rate such as 6%) from the financial institution financial system, and obtains image format protocol scanning data (including collateral such as apartment valuation 1.5 million yuan, ownership status) from the scanning device. After preprocessing, the encrypted serialized binary data (including case number, total debt, collateral asset fields) are output and transmitted to the distributed collaborative analysis unit. The distributed collaborative analysis unit receives this binary data, performs characteristic clustering processing, and outputs the weighted characteristic matrix (serialized binary format, including case number, total debt weight 0.5, collateral asset weight 0.85, etc.), which is transmitted to the case dynamic correlation model, and receives feedback data from the disposal verification feedback module (serialized binary, including accuracy score such as 90 points, priority label such as high priority of total debt). The case dynamic correlation model (3:5:2 topology) receives the weighted characteristic matrix, generates comprehensive summary information (pure text, including case number, total debt, risk score 0.3, etc.) and sub-summary confidence interval (pure text, such as total debt range 115-125 million), and transmits to the user interaction platform. The disposal verification feedback module collects feedback data from the user interaction platform (pure text, including disposal result such as recycling 1 million yuan, score 90 points), converts it to serialized binary format, and transmits it to the distributed collaborative analysis unit. The model optimization module receives the feedback data, outputs the optimized model parameters (serialized binary, including updated weights such as total debt 0.6, threshold 0.35), stores them in the storage server and feeds them back to the case dynamic correlation model. The user interaction platform receives the summary information and confidence interval, outputs user feedback, and completes the closed loop. All data flow is realized through Salsa20 encrypted secure data pipeline to ensure confidentiality and integrity, and is applied to clearing, transfer, debt restructuring and other scenarios.

[0179] Specific operation steps, algorithm logic or example parameters of technical means:

[0180] To enhance the feasibility, the following supplementary operation steps, algorithm logic and example parameters of each module technology means, the original case record acquisition module obtains text data from the judicial system through standard interface (field mapping table contains 1000 legal terms, such as "debtor" corresponding to specific identification), obtains financial data through database query (extracts total debt, volatility rate, etc.), obtains protocol image through scanning device driver (600dpi resolution), pre-processes using regular expression to check case number format (such as letter plus 8 digit number), outputs serialized binary data (Salsa20 encryption, key is 256-bit random sequence). The feature clustering algorithm of the distributed collaborative analysis unit receives binary data, extracts features (semantic rule set identifies case number, debtor, volatility analysis calculates 12-month transaction volatility rate such as 6%, image segmentation extracts collateral such as apartment valuation 1.5 million yuan), initializes 3 clustering centers (random seed 42), iterates 10 times to generate weighted feature matrix (example: case number L2025001, total debt 1.2 million yuan weight 0.5, collateral asset apartment 1.5 million yuan weight 0.85). The fusion algorithm of the case dynamic correlation model compares the difference between the total debt and the collateral asset in the input layer to generate the correlation strength (example: single debtor has no associated entity, strength 0.9), analyzes the debt change within 60 days in the middle layer to generate the risk score (example: no growth, score 0.3), and integrates the summary according to the weight priority (total debt weight 0.5 priority) in the output layer (example: case number L2025001, risk score 0.3). The disposal verification feedback module collects feedback through the Web interface (example: score 90 points, high priority of total debt), and converts it into binary. The model optimization module adjusts the weight according to the score (score greater than 85 points increases by 20%, such as total debt from 0.5 to 0.6), updates the risk threshold (from 0.3 to 0.35). The credible interval is generated based on statistical analysis (total debt mean plus or minus 2 times standard deviation, example: 1.15 million to 1.25 million), and the weight is allocated according to the liquidation scenario (total debt 0.6, collateral asset 0.3). These steps and parameters ensure that the algorithm is operational and can be applied to risk assessment and summary generation.

[0181] The running environment of the technical solution and the real-time and feasibility proof:

[0182] The method running environment is based on the foregoing hardware (data acquisition server: 2 8-core processors, 64 GB of memory, 2 TB NVMe solid state disk; distributed computing cluster: 5-20 nodes, each node 2 16-core processors, 128 GB of memory; storage server: 20 TB RAID-6; user interaction server: 8-core processor, 64 GB of memory; network equipment: 25 Gb Ethernet; scanning equipment: 600 dpi), supplemented by hardware acceleration and development platform details, to prove real-time and feasibility. Hardware acceleration accelerates clustering processing through a parallel computing framework (3 times speed improvement supported by multi-threading), optimizes data access through an NVMe solid state disk (read / write speed 3500 MB / s), and reduces inter-node delay to 0.1 milliseconds through high-speed network interconnection (25 Gbps). The development platform uses Python language (version 3.9, NumPy library for matrix processing, Pandas for table parsing, and Pillow for image processing), Kubernetes system (version 1.22) for scheduling cluster tasks (5 nodes, each node 4 cores), Nginx server (version 1.20) for running the Web interface (supporting summary display and feedback submission), Ceph storage system (version 16.2) for managing data, and Salsa20 encryption library (libsodium version 1.0.18) for ensuring transmission security. Real-time verification: 100 cases processed in 3.5 seconds (0.5 seconds for acquisition, 0.2 seconds for preprocessing, 1.2 seconds for clustering, 0.8 seconds for fusion, 0.4 seconds for confidence interval, and 0.3 seconds for feedback optimization), 1000 cases in 30 seconds, and 10000 cases in 3 minutes, meeting the needs of bank real-time clearing (less than 5 seconds) and court batch restructuring (less than 5 minutes). Feasibility verification: the system is deployed in a Linux environment (Ubuntu 20.04), and no system crashes are found after 1000 tests, with an average accuracy of 87.5% (based on experimental data), exception handling (such as using default values for data missing) ensures robustness, and Salsa20 encryption passes security audit (no leakage record). The running environment supports high-concurrency access (1000 users per second, response time less than 0.5 seconds), and is applied to clearing, transfer, and debt restructuring scenarios.

[0183] It should be noted that Figure 3 is a scatter plot, Figure 4 is a columnar diagram, first Figure 3 The scatter plot intuitively shows the results of k-means clustering (k=3).

[0184] Horizontal axis: represents the total amount of weighted debt (unit: ten thousand yuan), calculated by applying an initial weight of 0.5 to the original total debt (range 100-300 million) through the case characteristic clustering algorithm, reflecting the debt size characteristics of each case.

[0185] Longitudinal axis: represents the weighted collateral asset valuation (unit: ten thousand yuan), which is calculated by applying the initial weight of 0.85 to the collateral asset valuation (range 150-400 million) through the same clustering algorithm, reflecting the asset support capacity of the case.

[0186] Chart explanation: This scatter plot shows the characteristic distribution of 100 cases based on the weighted total debt and collateral asset valuation, using the k-means clustering algorithm (k=3) to divide the cases into three clusters, each represented by a different color. The chart visually presents the characteristic relevance of the input layer in the hierarchical characteristic fusion algorithm, helping to identify risk aggregation areas and assist in liquidation or transfer decisions.

[0187] Three colors of points explain:

[0188] Blue points: represent the first cluster of cases, representing low-risk cases with relatively balanced total debt and collateral asset valuation. These cases usually have a weighted total debt and collateral asset valuation distributed in the middle of the chart, indicating that the debt and asset support capacity match, with a low risk score (such as 0.2-0.3), suitable for quick liquidation.

[0189] Red points: represent the second cluster of cases, representing medium-high risk cases with high total debt and low collateral asset valuation. These points are usually located in the upper or left part of the chart, reflecting heavy debt burden and insufficient asset support, with a high risk score (such as 0.4-0.6), requiring priority disposal or transfer.

[0190] Green points: represent the third cluster of cases, representing low-risk cases with high collateral asset valuation and low total debt. These points are usually located on the right or lower part of the chart, indicating strong asset support capacity and low risk score (such as 0.1-0.3), suitable for long-term debt restructuring.

[0191] Figure 4 Bar chart showing total debt, horizontal axis: represents case number (from L20250001 to L20250010), corresponding to the identification of the first 10 cases, used to distinguish the total debt and confidence interval of different cases.

[0192] Longitudinal axis: represents the total debt (unit: ten thousand yuan), showing the original debt amount of each case (range 100-300 million), reflecting the core attribute in the comprehensive summary information.

[0193] Chart explanation: This bar chart shows the total debt of the first 10 cases, generated by the hierarchical characteristic fusion algorithm, showing the intuitive distribution of debt size. The height of each column represents the total debt of the corresponding case, and the chart helps users quickly assess the debt burden and support liquidation or transfer decisions. At the same time, combined with the confidence interval, it provides a reliability assessment of the debt amount.

[0194] Black markers: The black markers on each column are error bars, representing the upper and lower bounds of the confidence interval (mean ± 2 standard deviations) for the total debt amount, calculated through statistical analysis. The lower end of the error bar represents the lower bound of the confidence interval (e.g., 1.15 million), and the upper end represents the upper bound (e.g., 1.25 million), reflecting the range of fluctuations in the total debt amount. Since the standard deviation in the code is based on simulated data (10% of the mean total debt amount), the length of the error bar varies from case to case, with longer error bars indicating higher uncertainty in the debt amount and shorter error bars indicating better stability. These markers provide a confidence reference for decision-making, such as in liquidation scenarios, where a narrow interval supports rapid disposal.

[0195] It should be noted that the relational terms herein, such as first and second, are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between or among the entities or actions. Moreover, the terms "comprises," "comprising," or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by an indefinite article "a" or "an" does not exclude the existence of additional identical elements in the process, method, article, or apparatus comprising the element.

[0196] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, alternatives, and variations can be made in the embodiments without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. A method for automated processing of intelligent summaries of cases related to non-performing assets, characterized in that: Includes the following steps: Sp1: Collect original case records of cases related to non-performing assets, including judicial ruling texts, financial reconciliation data, and scanned images of agreements; Sp2: Construct a distributed collaborative analysis unit to receive the original case records. In the distributed collaborative analysis unit, the original case records are transformed into a weighted characteristic matrix through a case characteristic clustering algorithm. The weighted characteristic matrix includes a core characteristic vector, an auxiliary characteristic vector, and a stable factor vector. Sp3: Construct a dynamic case association model, which includes multiple association units. Input the weighted characteristic matrix into the association units of the dynamic case association model to deduce dynamic relationships. Output comprehensive summary information of non-performing asset cases through a hierarchical characteristic fusion algorithm. Sp4: Collect actual disposal verification data samples of non-performing asset cases, including disposal results and verification status, and feed the actual disposal verification data samples back to the distributed collaborative analysis unit; Sp5: The dynamic correlation model of the case is dynamically adjusted using a feature optimization algorithm based on verification feedback to improve the accuracy of the comprehensive summary information and the adaptability to different scenarios.

2. The method for automated processing of intelligent summaries of cases related to non-performing assets according to claim 1, characterized in that: The case dynamic association model adopts a 3:5:2 hierarchical topology structure, which includes 3 input layer association units, 5 intermediate layer association units and 2 output layer association units. The association units calculate the dynamic association strength between case entities through the characteristic weights in the weighted characteristic matrix.

3. The method for automated processing of intelligent summaries of cases related to non-performing assets according to claim 1, characterized in that: The weighted characteristic matrix is ​​a structured matrix that describes the relationship between case entities and risk characteristics. It includes case number, debtor summary, total debt, collateral information, related entity relationships, case handling stage, and risk impact weight.

4. The method for automated processing of intelligent summaries of cases related to non-performing assets according to claim 1, characterized in that: The core feature vector represents key case attributes that affect the generation of the comprehensive summary information, the auxiliary feature vector represents secondary impact attributes on the generation of the comprehensive summary information, and the stabilization factor vector is used to adjust the derivation threshold of the associated unit to ensure the stability of the summary information.

5. The method for automated processing of intelligent summaries of cases related to non-performing assets according to claim 1, characterized in that: The case characteristic clustering algorithm includes the following steps: Sp21: Using semantically layered field extraction technology, the case number, debtor summary, and case handling stage are extracted from the judicial ruling text to generate an initial feature set; Sp22: Using volatility-based financial extraction technology, the total debt and trading volatility are extracted from financial reconciliation data. When the trading volatility is greater than 8%, the weight is 0.75, which supplements the initial feature set. Sp23: By using image analysis technology based on multi-region segmentation, information on collateral assets and related entity relationships are extracted from the protocol scan image to improve the initial feature set; Sp24: Using a feature weight clustering algorithm, based on the volatility of total debt and the stability of collateral asset value, with a weight of 0.85 when the stability of collateral asset value is greater than 90%, k-means clustering is performed on the initial feature set to generate a weighted feature matrix. The weighted feature matrix is ​​transmitted to the case dynamic association model in encrypted serialized binary format through a secure data pipeline, which is encrypted using Salsa20.

6. The method for automated processing of intelligent summaries of cases related to non-performing assets according to claim 1, characterized in that: The hierarchical feature fusion algorithm includes the following steps: Sp31: In the input layer association unit, the association strength between debtors, collateral assets and related entities in the weighted characteristic matrix is ​​calculated based on Manhattan distance using the characteristic association strength calculation technology. Sp32: In the intermediate layer of the associated unit, dynamic risk assessment technology based on sliding time windows is used to analyze the changing trends of total debt and case handling stages using a 60-day window, and generate dynamic risk scores. Sp33: In the output layer association unit, through summary content aggregation technology, dynamic risk scores and core case attributes are integrated based on priority rules to generate comprehensive summary information. The comprehensive summary information is transmitted to the user interaction platform in plain text format through a secure data pipeline.

7. The method for automated processing of intelligent summaries of non-performing asset-related cases according to claim 1, characterized in that: The feature optimization algorithm based on verification feedback includes the following steps: Sp51: Collect processing verification feedback data submitted by users through the user interaction platform. The processing verification feedback data includes the accuracy score of the comprehensive summary information and field priority labels. Sp52: Through feature weight reconstruction technology, the weights in the weighted feature matrix are adjusted based on the accuracy score and field priority label. When the accuracy score is greater than 85, the field weight is increased by 20%. Sp53: Through dynamic adjustment of model parameters, the derivation threshold and weight parameters of the associated unit are updated based on the disposal verification feedback data to optimize subsequent summary generation. The optimized model parameters are stored in the distributed collaborative analysis unit in an encrypted serialized binary format.

8. The method for automated processing of intelligent summaries of cases related to non-performing assets according to claim 1, characterized in that: The method includes the following stages: During the case characteristic integration period, the case basic summary metadata is generated in the first distributed collaborative analysis unit and the first association unit. The case basic summary metadata includes the case number and debtor summary. During the dynamic risk derivation period, risk association summary metadata is generated in the second distributed collaborative analysis unit and the second association unit. The risk association summary metadata includes dynamic risk scores. During the summary credibility optimization period, comprehensive summary credibility interval metadata is generated in the third distributed collaborative analysis unit and the third association unit. The comprehensive summary credibility interval metadata includes the confidence assessment of the summary content.

9. The method for automated processing of intelligent summaries of cases related to non-performing assets according to claim 8, characterized in that: The comprehensive summary confidence interval metadata is used to generate sub-summary confidence interval metadata applicable to different disposal scenarios through disposal scenario adaptation technology. The sub-summary confidence interval metadata is used to optimize the reliability of case disposal decisions. The disposal scenarios include liquidation and transfer.

10. A system based on the intelligent summary automated processing method for non-performing asset-related cases as described in any one of claims 1-9, characterized in that: The system includes: The original case record collection module is used to collect original case records of cases related to non-performing assets. The original case records include judicial ruling texts, financial reconciliation data, and scanned images of agreements. A distributed collaborative analysis unit is connected to the original case record acquisition module. It is used to receive the original case records and transform the original case records into a weighted characteristic matrix through a case characteristic clustering algorithm. The weighted characteristic matrix includes a core characteristic vector, an auxiliary characteristic vector, and a stable factor vector. A case dynamic association model is connected to the distributed collaborative analysis unit. The case dynamic association model includes at least one association unit, which is used to receive the weighted characteristic matrix, perform dynamic relationship deduction, and output comprehensive summary information of non-performing asset cases through a hierarchical characteristic fusion algorithm. The disposal verification feedback module is used to collect actual disposal verification data samples of non-performing asset cases. The actual disposal verification data samples include disposal results and verification status, and feed them back to the distributed collaborative analysis unit. The model optimization module is connected to the case dynamic association model and the handling verification feedback module, and is used to dynamically adjust the case dynamic association model using a feature optimization algorithm based on verification feedback.

Images