Credit quota recombination charging method and system of SaaS architecture based on AI

Abstract

The invention discloses a creditor's right quota recombination charging method and system based on an AI SaaS architecture, and the method comprises the steps: receiving creditor's right basic data and recombination target constraints inputted by a user through an SaaS platform, and obtaining a debtor behavior feature vector and an associated environment factor set; on the basis of the debtor behavior feature vector and the associated environment factor set, behavior path probability distribution reflecting potential repayment results under different recombination schemes is obtained; obtaining an optimized creditor's right recombination scheme and a corresponding estimated success probability value according to the behavior path probability distribution and the recombination target constraint; based on the optimized creditor's right recombination scheme and the pre-estimated success probability value corresponding to the optimized creditor's right recombination scheme, obtaining charging detail data corresponding to the recombination scheme; and carrying out structured packaging and standardized output processing on the charging detail data to generate a creditor's right recombination charging instruction which can be directly integrated to a SaaS platform charging module. By using the embodiment of the invention, the accuracy and the charging efficiency of the creditor's right recombination service can be improved.

Description

Technical Field

[0001] This invention belongs to the field of AI technology, specifically a debt restructuring billing method and system based on an AI-based SaaS architecture. Background Technology

[0002] Currently, the formulation of traditional debt restructuring plans mainly relies on human experience and static financial models, making it difficult to dynamically assess the debtor's behavioral characteristics and the impact of environmental variables. This results in insufficient success rate and adaptability of the restructuring plans. At the same time, the billing systems of existing SaaS platforms typically adopt fixed rates or simple segmented billing models, failing to combine AI prediction results to achieve dynamic pricing based on risk correlations, thus hindering the intelligent and precise development of debt restructuring services. Summary of the Invention

[0003] The purpose of this invention is to provide a debt restructuring billing method and system based on AI and SaaS architecture, so as to overcome the shortcomings of the existing technology and improve the accuracy and billing efficiency of debt restructuring services.

[0004] One embodiment of this application provides a debt restructuring billing method based on an AI-based SaaS architecture, the method comprising:

[0005] The system receives basic debt data and restructuring target constraints input by users through the SaaS platform. Based on the basic debt data, it extracts the debtor's historical behavioral characteristics and performs environmental variable correlation processing to obtain the debtor's behavioral characteristic vector and the set of associated environmental factors.

[0006] Based on the debtor's behavioral feature vector and the associated environmental factor set, a pre-trained generative behavioral simulation model is used to simulate multi-path future repayment behavior, resulting in a probability distribution of behavioral paths that reflects the potential repayment outcomes under different restructuring schemes.

[0007] Based on the probability distribution of the behavioral path and the constraints of the restructuring target, the success probability of the restructuring plan is calculated and optimized to obtain the optimized debt restructuring plan and its corresponding estimated success probability value.

[0008] Based on the optimized debt restructuring plan and its corresponding estimated success probability, the SaaS billing rule engine is used to perform dynamic billing strategy matching and cost calculation to obtain the billing details data corresponding to the restructuring plan.

[0009] The billing details data are structured, encapsulated, and output in a standardized manner to generate a debt restructuring billing instruction that can be directly integrated into the billing module of a SaaS platform.

[0010] Optionally, the step of receiving the basic debt data and restructuring target constraints input by the user through the SaaS platform, and extracting the debtor's historical behavioral characteristics and performing environmental variable correlation processing based on the basic debt data, yields a debtor behavioral feature vector and a set of associated environmental factors, including:

[0011] Based on the multi-source heterogeneous nature of the basic debt data, data authenticity verification and timestamp alignment are performed using a blockchain consensus mechanism to generate a standardized debt dataset with digital fingerprints.

[0012] The standardized debt dataset is input into a spatiotemporal graph convolutional network to mine historical transaction patterns of debtors and perform cross-entity correlation analysis, extracting behavioral pattern subgraphs of debtors.

[0013] Based on the behavioral pattern subgraph and economic indicators under the current macro environment, environmental sensitivity factor embedding calculation is performed, and environmental weighted feature vector is generated through graph attention mechanism;

[0014] The environmental weighted feature vector is encrypted using federated learning and integrated with privacy-preserving third-party credit data to output a debtor behavior feature vector with dynamic weights and a set of associated environmental factors.

[0015] Optionally, based on the debtor's behavioral feature vector and the associated environmental factor set, a pre-trained generative behavioral simulation model is used to simulate multi-path future repayment behavior to obtain a behavioral path probability distribution reflecting the potential repayment outcomes under different restructuring schemes, including:

[0016] The debtor's behavioral feature vector is input into a pre-trained generative adversarial network, and the debtor's behavioral latent vector is generated through a conditional variational autoencoder.

[0017] Based on the set of associated environmental factors, stochastic differential equations are constructed to perform Monte Carlo environmental disturbance simulation and generate a cluster of dynamic environmental evolution paths.

[0018] The behavioral latent vectors and environmental evolution path clusters are input into the neural control differential equations to perform multi-agent collaborative simulation and output the original behavioral path trajectories.

[0019] By analyzing the topological persistence of the path trajectory, the continuous coherence algorithm is applied to extract key repayment behavior patterns and generate a path topological feature matrix.

[0020] The hidden Markov model is driven by the path topology feature matrix, and the state transition probability is recalibrated by combining the reorganization scheme parameters, outputting the behavioral path probability distribution with confidence intervals.

[0021] Optionally, the step of calculating and optimizing the success probability of the restructuring plan based on the probability distribution of the behavioral path and the restructuring target constraints to obtain the optimized debt restructuring plan and its corresponding estimated success probability value includes:

[0022] Based on the probability distribution of the behavior path, a repayment success surface is constructed, and the restructuring objective constraint is transformed into a Lagrange multiplier. Gradient ascent search under the constraint conditions is performed on the repayment success surface to locate the boundary of feasible solutions.

[0023] Pareto front is constructed based on feasible solution boundary, and solution trade-off analysis is performed using multi-objective Bayesian optimization algorithm to generate Pareto set of candidate solutions.

[0024] Robust stress testing is performed on the Pareto set of candidate solutions, and the stability of the solutions is evaluated through adversarial sample injection. The optimized debt restructuring solution and the estimated success probability value are output.

[0025] Optionally, based on the optimized debt restructuring plan and its corresponding estimated success probability value, the SaaS billing rule engine performs dynamic billing strategy matching and cost calculation to obtain the billing details data corresponding to the restructuring plan, including:

[0026] Input the estimated success probability value into the SaaS billing rule engine, and use the fuzzy logic system to match the pre-designed billing template to generate a basic billing framework;

[0027] Based on the risk exposure characteristics of the optimized debt restructuring plan, the risk value is calculated online to generate a dynamic risk premium coefficient;

[0028] By integrating the entropy features of debtor behavior feature vectors, a service complexity weight factor is generated through a reinforcement learning model;

[0029] Based on the basic billing framework, dynamic risk premium coefficient and service complexity weight factor, a genetic optimization algorithm is used to fine-tune multi-dimensional parameters and generate the optimal combination of billing parameters.

[0030] The system automatically executes the optimal combination of billing parameters through smart contracts, calibrates fees based on real-time market interest rate fluctuations, and outputs structured billing details.

[0031] Optionally, the step of performing structured encapsulation and standardized output processing on the billing details data to generate a debt restructuring billing instruction that can be directly integrated into the billing module of the SaaS platform includes:

[0032] Semantic parsing of billing details data is performed, and a machine-readable billing business entity relationship graph is generated through an ontology reasoning engine;

[0033] A billing instruction template is constructed based on an entity relationship diagram. The logical completeness is ensured through formal verification, and a verified instruction framework is generated.

[0034] The key parameters of the optimized debt restructuring plan will be embedded into the instruction framework through homomorphic encryption to form a tamper-proof billing instruction prototype.

[0035] The instruction prototype is encapsulated using a cross-chain interoperability protocol, with zero-knowledge proof verification credentials added, and the output is a standardized debt restructuring billing instruction that can be directly integrated.

[0036] Another embodiment of this application provides a debt restructuring billing system based on an AI-driven SaaS architecture, the system comprising:

[0037] The receiving module is used to receive the basic debt data and restructuring target constraints input by the user through the SaaS platform, and to extract the debtor's historical behavioral characteristics and perform environmental variable correlation processing based on the basic debt data to obtain the debtor's behavioral feature vector and the set of associated environmental factors.

[0038] The simulation module is used to simulate multi-path future repayment behavior based on the debtor's behavioral feature vector and the set of associated environmental factors, using a pre-trained generative behavioral simulation model to obtain the probability distribution of behavioral paths that reflect the potential repayment results under different restructuring schemes.

[0039] The optimization module is used to calculate and optimize the success probability of the restructuring plan based on the probability distribution of the behavior path and the constraints of the restructuring target, so as to obtain the optimized debt restructuring plan and its corresponding estimated success probability value.

[0040] The matching module is used to perform dynamic billing strategy matching and cost calculation processing through the SaaS billing rule engine based on the optimized debt restructuring plan and its corresponding estimated success probability value, so as to obtain the billing details data corresponding to the restructuring plan.

[0041] The generation module is used to perform structured encapsulation and standardized output processing on the billing details data, and generate debt restructuring billing instructions that can be directly integrated into the billing module of the SaaS platform.

[0042] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0043] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0044] Compared with existing technologies, this invention provides an AI-based SaaS architecture-based debt restructuring billing method. It receives basic debt data and restructuring target constraints input by the user through a SaaS platform, obtaining a debtor behavioral feature vector and a set of related environmental factors. Based on the debtor behavioral feature vector and the set of related environmental factors, it obtains a behavioral path probability distribution reflecting the potential repayment outcomes under different restructuring schemes. According to the behavioral path probability distribution and restructuring target constraints, it obtains an optimized debt restructuring scheme and its corresponding estimated success probability value. Based on the optimized debt restructuring scheme and its corresponding estimated success probability value, it obtains the billing details data corresponding to the restructuring scheme. The billing details data are then structured, encapsulated, and standardized for output, generating a debt restructuring billing instruction that can be directly integrated into the SaaS platform's billing module, thereby improving the accuracy and billing efficiency of debt restructuring services. Attached Figure Description

[0045] Figure 1 A hardware structure block diagram of a computer terminal for a debt amount restructuring and billing method based on AI SaaS architecture provided in an embodiment of the present invention;

[0046] Figure 2 A flowchart illustrating a debt restructuring and billing method based on an AI-enabled SaaS architecture, provided as an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of a debt reorganization and billing system based on an AI-enabled SaaS architecture, provided as an embodiment of the present invention. Detailed Implementation

[0048] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0049] This invention first provides a debt restructuring billing method based on an AI-based SaaS architecture. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0050] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a debt restructuring and billing method based on an AI-driven SaaS architecture, as provided in an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0051] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any AI-based SaaS architecture-based debt restructuring billing method.

[0052] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0053] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When these computer programs are executed by a processor, the processor can execute any AI-based SaaS architecture debt restructuring billing method.

[0054] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0055] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0056] See Figure 2 The present invention provides a debt restructuring billing method based on an AI-based SaaS architecture, which may include the following steps:

[0057] S201, Receive the basic debt data and restructuring target constraints input by the user through the SaaS platform, and extract the debtor's historical behavioral characteristics and perform environmental variable correlation processing based on the basic debt data to obtain the debtor's behavioral characteristic vector and the set of associated environmental factors;

[0058] Specifically, based on the multi-source heterogeneous characteristics of the underlying debt data, data authenticity verification and timestamp alignment can be performed using a blockchain consensus mechanism to generate a standardized debt dataset with digital fingerprints.

[0059] The system receives basic debt claim data input from users of the SaaS platform. This data exhibits significant multi-source heterogeneity, originating from various systems including core banking systems, third-party credit reporting agencies, and enterprise resource management systems. Data formats encompass structured database tables, semi-structured data files, and scanned copies of unstructured documents. To address data trust issues, the system initiates a data authenticity verification process based on a blockchain consensus mechanism. Specifically, the original debt claim data from different sources is divided into several data blocks, each containing data content, a source identifier, and an initial timestamp. These data blocks are broadcast to a permissioned blockchain network composed of financial institution nodes (such as banks and credit reporting companies). The network employs a practical Byzantine fault-tolerant algorithm for consensus verification: each node verifies the authenticity of the data source through digital signatures and cross-checks the logical consistency of key fields (such as the debtor's ID number and debt amount). When more than a preset percentage (e.g., two-thirds) of the nodes pass verification, the data block is deemed valid.

[0060] To address the issue of time inconsistencies in multi-source data (such as discrepancies between bank repayment records and court judgments), the system performs timestamp alignment. First, key event points (such as loan disbursement date, final repayment date, and legal case filing date) are extracted from all basic debt data using event sourcing technology and mapped to a unified timeline. Next, a dynamic time warping algorithm aligns asynchronous events: using the time from an authoritative data source (such as a central bank credit report) as a benchmark, the offset of event timestamps from other data sources from the benchmark time is calculated. For example, if a court judgment's effective date is detected to be 3 days later than the benchmark time, the system automatically calibrates its timestamp to the synchronization position. All data blocks that have passed consensus verification and time calibration are linked in chronological order into an immutable data chain.

[0061] To generate a standardized debt dataset with digital fingerprints, the system performs a standardization transformation on the processed data chain: First, data format standardization converts heterogeneous data into a unified binary format; second, field semantic standardization maps field aliases from different sources to standard field names using a financial ontology library (e.g., unifying "remaining principal" as "outstanding principal"); third, adding digital fingerprints calculates a cryptographic hash value for each standardized data block as a unique identifier, while simultaneously writing the blockchain transaction number and the latest consensus timestamp into the metadata. The final output is a standardized dataset with a unified structure, synchronized time, and traceability for subsequent analysis.

[0062] The standardized debt dataset is input into a spatiotemporal graph convolutional network to mine historical transaction patterns of debtors and perform cross-entity correlation analysis, extracting behavioral pattern subgraphs of debtors.

[0063] A standardized debt dataset is fed into a spatiotemporal graph convolutional network to model debtor behavior. First, a multi-dimensional financial relationship graph is constructed: the debtor is the core node, linked to entity nodes such as their bank account, guarantor, and collateral. Edges between nodes represent transactions, guarantee relationships, and other behaviors. Each node is accompanied by spatiotemporal features: spatial features such as the coordinates of the account's location, and temporal features such as the monthly repayment amount sequence for the most recent 12 months.

[0064] Spatiotemporal graph convolutional networks employ a dual convolution mining model: spatial convolution uses an adjacency matrix to define node connection strength (e.g., using transaction frequency as edge weights) and aggregates neighbor node features through graph convolution operators (e.g., aggregating the average monthly balance of all associated accounts of a debtor); temporal convolution applies a time sliding window (e.g., a 3-month cycle) to the node feature sequence to detect periodic repayment patterns (e.g., the characteristic pattern of large repayments at the end of quarters). The two convolutions are executed alternately to progressively extract high-dimensional spatiotemporal features.

[0065] Cross-entity association analysis is performed based on high-dimensional features: The first layer of analysis identifies implicit associations and calculates the feature similarity between guarantor and collateral nodes. If the similarity exceeds a threshold, virtual edges are added to represent potential risk transmission paths. The second layer of analysis uses graph clustering algorithms to group debtor nodes with frequent repayment delays into high-risk clusters. The final output is a behavioral pattern subgraph, which is a subset of data containing core debtor nodes, strongly related entities, and their interaction features (e.g., extracting the triangular relationship of "debtor-frequently used repayment account-collateralized property" and labeling the account with the "small test repayment on the 25th of each month" behavior tag).

[0066] Based on the behavioral pattern subgraph and economic indicators under the current macro environment, environmental sensitivity factor embedding calculation is performed, and environmental weighted feature vector is generated through graph attention mechanism;

[0067] The system receives real-time access to macroeconomic indicators, including GDP growth rate, industry prosperity index, and monetary policy easing. To quantify the impact of the environment on debtors, environmental sensitivity factors are calculated: industry sensitivity (e.g., volatility of housing price index related to real estate companies), regional sensitivity (e.g., unemployment rate in the registered location of related companies), and scale sensitivity (e.g., credit availability index related to micro and small enterprises).

[0068] The environmental sensitivity factor embedding calculation injects factor values ​​into the behavioral pattern subgraph: node-level embedding adds sensitivity feature dimensions to each debtor node (such as expanding the feature vector to [repayment regularity, debt ratio, housing price fluctuation sensitivity]); edge-level embedding dynamically adjusts the transaction edge weights (for example, during an economic recession, the credibility weight of the guarantee relationship edge is reduced by 30%).

[0069] A graph attention mechanism is used to generate environment-weighted feature vectors: multiple attention heads are set to focus on dimensions such as industry, region, and scale; when calculating the attention coefficient, higher-sensitivity relationships are given greater weight (such as the transaction edge between real estate developers and suppliers); when weighting and fusing neighbor node features, the environmental impact is highlighted. Finally, the environment-weighted feature vector of each debtor is output (for example, the numerical combination [0.82, 1.35, -0.23] represents "strong repayment willingness, high sensitivity to economic cycles, and low regional risk resistance").

[0070] The environmental weighted feature vector is encrypted using federated learning and integrated with privacy-preserving third-party credit data to output a debtor behavior feature vector with dynamic weights and a set of associated environmental factors.

[0071] To protect data privacy, the system implements federated learning encryption on the environment-weighted feature vectors: each financial institution uses the feature vectors locally to train the debtor rating model and generate model parameter gradients; the gradients are encrypted into ciphertext using a homomorphic encryption algorithm to ensure that the original data cannot be parsed in the cloud.

[0072] Integrating third-party credit data in an encrypted environment: receiving and decrypting encrypted data provided by credit reporting agencies through a trusted hardware environment; performing gradient aggregation and data fusion operations within a secure isolation zone (e.g., weighted concatenation of the number of overdue payments in 12 months provided by a third party with the "willingness to repay" dimension of the local feature vector).

[0073] The final feature vector with dynamic weights is generated: weights are assigned based on data freshness and source authority (e.g., 0.9 weight for central bank credit data, 0.6 weight for enterprise self-reported data); core dimensions are retained through feature compression techniques (e.g., [repayment stability score, environmental sensitivity coefficient, associated risk density]); and a set of associated environmental factors is encapsulated (e.g., independently recording "housing price sensitivity factor value and its impact weight"). The final output is a structured feature vector and environmental factor package that can be directly used for behavior prediction.

[0074] This method first receives raw debt data (such as repayment records and debt size) and restructuring objectives (such as extending repayment periods and reducing interest rates) submitted by users through a SaaS platform. Then, it uses data mining techniques to extract key features (such as repayment rates and delinquency frequency) from historical behavior and combines these with macroeconomic indicators (such as interest rate fluctuations and industry prosperity) to construct environmental correlation factors. This forms a structured dataset that comprehensively reflects the debtor's credit status and behavioral tendencies, transforming scattered debt data into quantifiable feature vectors and providing high-quality input for subsequent behavioral simulations. The introduction of environmental factors enhances the model's sensitivity to external risks such as economic cycles, avoiding biases caused by static assessments.

[0075] S202, based on the debtor's behavioral feature vector and the set of associated environmental factors, a pre-trained generative behavioral simulation model is used to simulate multi-path future repayment behavior to obtain the probability distribution of behavioral paths that reflect the potential repayment results under different restructuring schemes.

[0076] Specifically, the debtor's behavioral feature vector can be input into a pre-trained generative adversarial network, and the debtor's behavioral latent vector can be generated through a conditional variational autoencoder.

[0077] Construction and Operation of Pre-trained Generative Adversarial Networks (GANs)

[0078] Generative Adversarial Networks (GENs) consist of two core components: a generator (GEN) and a discriminator (DIS). The generator learns the distribution patterns of debtors' historical behavioral data, while the discriminator distinguishes between real historical data and fake data synthesized by the generator. During the pre-training phase, the system uses a massive historical debt dataset (such as tens of thousands of debtor repayment records) for adversarial training. The specific process is as follows:

[0079] The generator receives a random noise vector (NV, typically 100-dimensional, representing potential behavioral uncertainty) and a debtor behavior feature vector (DBFV) (e.g., containing 30-dimensional features such as repayment frequency, delinquency rate, and debt-to-income ratio).

[0080] The generator maps the input to a simulated sequence of repayment behaviors (such as the distribution of repayment amounts over the next 12 months) using a multi-layer fully connected neural network (FCNN).

[0081] The discriminator simultaneously receives real historical repayment sequences and simulated sequences output by the generator. It extracts temporal features through a convolutional neural network (CNN) and outputs a probability value (DIS_Prob, ranging from 0 to 1), which represents the confidence that the input sequence is real data.

[0082] Through repeated iterations (e.g., 500,000 training cycles), the generator gradually learns the distribution of real data and can synthesize debtor behavior patterns that conform to statistical laws.

[0083] Latent vector generation of Conditional Variational Autoencoder (CVAE)

[0084] The pre-trained GAN generator serves as a foundation, which is then further integrated with a conditional variational autoencoder. The core of CVAE is to compress the DBFV into a low-dimensional behavior latent vector (BLV) while retaining key behavioral features.

[0085] The encoder (ENC) receives the DBFV (e.g., 30-dimensional features) and the associated specific restructuring conditions (e.g., extending the repayment period to 36 months, reducing the interest rate by 2%), and maps them to the mean vector (MV) and variance vector (VV) in the latent space via a neural network.

[0086] Using the reparameterization trick (RT), BLVs (e.g., 10-dimensional vectors) are generated by sampling from a Gaussian distribution. Their mathematical representation is: BLV = MV + VV × ( (This is standard normally distributed noise).

[0087] The decoder (DEC) reconstructs a simulated repayment behavior sequence using the BLV and restructuring scheme conditions as input. The training objective is to minimize the mean squared error (MSE) between the reconstructed sequence and the true sequence. Ultimately, the BLV becomes a low-dimensional abstraction representing the essence of the debtor's behavior, such as a 10-dimensional vector containing implicit factors such as "strength of repayment willingness" and "sensitivity to income fluctuations".

[0088] Financial semantic interpretation and quality control of latent vectors

[0089] The generated BLV must be interpretable to support subsequent decision-making. The system employs the following quality control mechanisms:

[0090] Feature Importance Analysis: The contribution of each original feature in DBFV (such as "number of overdue payments in the last 6 months") to each dimension of BLV is calculated using SHAP (SHapley Additive exPlanations) values ​​to ensure that latent factors are related to business logic (for example, the SHAP value of BLV dimension 3 shows that 70% of its weight comes from the "revenue stability" feature).

[0091] Stability verification: If the generated BLV cosine similarity (CS) is greater than 0.95, the latent vector generation is considered stable, then the DBFV with slight perturbation (5% Gaussian noise) is input to the same debtor.

[0092] Anomaly detection: If the Mahalanobis distance (MD) of BLV exceeds the 99th percentile of the historical training data, a manual review process is triggered to prevent the generation of unrealistic latent vectors.

[0093] Based on the set of associated environmental factors, stochastic differential equations are constructed to perform Monte Carlo environmental disturbance simulation and generate a cluster of dynamic environmental evolution paths.

[0094] Environmental factor modeling and construction of stochastic differential equations

[0095] The Associated Environmental Factor Set (AEFS) includes macroeconomic variables (such as GDP growth rate, unemployment rate, and industry prosperity index) and policy variables (such as benchmark interest rate and credit tightening index). Each factor is modeled as a stochastic process.

[0096] Taking GDP growth rate (GDP_Growth, GDPG) as an example, its dynamic changes are described by the mean regression process, and a stochastic differential equation (SDE) is constructed:

[0097] d(GDPG)=θ(μ-GDPG)dt+σdWt

[0098] Where θ is the regression velocity (set to 0.5), μ is the long-term equilibrium value (e.g., 5%), σ is the volatility (set to 0.8), and dWt is the Wiener Process (WP) increment, representing random shocks.

[0099] The correlation between factors is characterized by a correlation matrix (CM) (e.g., the correlation coefficient between unemployment rate and GDP growth rate is -0.7), and this dependency structure is preserved when generating paths.

[0100] Monte Carlo Simulation (MCS)

[0101] The system performs massively parallel path simulations to predict the future evolution of environmental factors:

[0102] Number of paths: Typically, 10,000 (10K) unique paths are generated, covering 95% of the confidence interval.

[0103] Time discretization: The next 36 months are divided into 216 time steps (TS), each step being 0.5 months long.

[0104] Numerical solution: The Euler-Maruyama Method (EMM) is used to iteratively update the environmental factor values. For example, the GDPG update formula for the i-th step is:

[0105]

[0106] Where Δt=0.5, These are random numbers drawn from a multivariate normal distribution (preserving the correlation defined by CM).

[0107] External shock injection: At a specific point in time (such as the 18th month), an extreme event (such as a financial crisis) is injected with a 5% probability, causing GDPG to drop by 3% instantaneously.

[0108] Generation and compression of dynamic environment evolution path clusters

[0109] The output is a Dynamic Environment Path Cluster (DEPC):

[0110] Data structure: A three-dimensional tensor with dimensions of number of paths × number of time steps × number of environmental factors (e.g., 10,000 × 216 × 8).

[0111] Critical path extraction: The K-means clustering (KMC) algorithm is used to classify 100,000 paths into 10 typical patterns (such as "stable recovery", "mild recession" and "severe fluctuation") based on trajectory similarity. Each pattern is represented by its cluster center (CC) path, which greatly reduces the computational complexity of the subsequent process.

[0112] Path labeling: Label each path with key statistics (such as the 36-month GDPG average and maximum drawdown rate) to facilitate matching debtor behavior characteristics.

[0113] The behavioral latent vectors and environmental evolution path clusters are input into the neural control differential equations to perform multi-agent collaborative simulation and output the original behavioral path trajectories.

[0114] Neural Controlled Differential Equation (NCDE) Model Architecture

[0115] NCDE models behavioral evolution as a continuous dynamic system, overcoming the limitations of discrete time steps:

[0116] The core equation is: the differential form of the debtor's state h(t) is: dh(t) = f(h(t), t; θf) dX(t). Here, X(t) is the driving signal (such as 8-dimensional environmental factors + 3-dimensional scheme parameters) spliced ​​by the environmental path e(t) and the recombination scheme r, and f is a function parameterized by the Neural Network (NN) (θf is the weight).

[0117] Neural network structure:

[0118] f employs a successive version of the Gated Recurrent Unit (GRU), containing 32 hidden units (HUs), capable of memorizing long-term dependencies.

[0119] Initial value setting: The initial state h(0) is directly initialized by the behavioral latent vector BLV (10-dimensional), which carries the debtor's personalized characteristics.

[0120] Multi-agent cooperative simulation process

[0121] The simulation process simulates the interaction between debtors, creditors, and the economic environment:

[0122] Definition of intelligent agent:

[0123] Debtor agent: Determines the monthly repayment amount based on current income (affected by the environment), debt pressure, and willingness to repay (determined by BLV).

[0124] Creditor AI Agent: Dynamically adjusts collection strategies based on repayment performance (e.g., escalating from SMS reminders to legal proceedings).

[0125] Environmental intelligent agent: Pushes economic indicators in real time according to the DEPC path.

[0126] Collaboration mechanism: At the beginning of each month, the environmental intelligent agent updates factors such as GDPG and unemployment rate.

[0127] The debtor's intelligent agent calculates the current repayment ability through NCDE: Repayment ability = BLV_Income Sensitivity × GDPG + BLV_Resilience × Debt ratio.

[0128] The creditor agent selects a response strategy based on the number of overdue days (the strategy table has 10 preset strength levels).

[0129] The debtor adjusts the repayment priority based on the intensity of collection efforts (e.g., prioritizing repayment of the debt in the event of legal proceedings).

[0130] Output and storage of original behavior path trajectories

[0131] Output: Each path contains 36 months of time-series data.

[0132] Monthly repayment amount (RA);

[0133] Accumulated Overdue Days (AOD);

[0134] Collection Level (CL);

[0135] Data scale: If the environmental path cluster is compressed into 10 typical paths, each debtor will simulate 10×10=100 paths (10 environments × 10 random seeds), with a total output of 100×36×3=10,800 data points.

[0136] Storage optimization: Use columnar storage (CS) format (such as Apache Parquet) and partition by path ID to improve the efficiency of subsequent analysis.

[0137] By analyzing the topological persistence of the path trajectory, the continuous coherence algorithm is applied to extract key repayment behavior patterns and generate a path topological feature matrix.

[0138] Topological Persistence Analysis (TPA) Principles

[0139] Viewing repayment behavior as a point cloud in a high-dimensional space, and capturing its shape features through topology:

[0140] Point cloud construction: 36 months of data for a single behavior path, each month is regarded as a 3D point (RA, AOD, CL), forming a trajectory of 36 points.

[0141] Persistent Homology (PH): Calculating point clouds at different scales (distance thresholds). Topological features (such as the number of connected components and the number of holes) under ).

[0142] Persistence Diagram (PD): Records the birth and death thresholds of topological features. For example:

[0143] A connected component in When =0.1, (Birth) appears. It disappears (Death) when its value is 0.5, and its persistence is 0.5-0.1=0.4.

[0144] High persistence (Persist > 0.3) represents a stable behavioral pattern (such as consistently low repayments), while low persistence (Persist < 0.1) is considered noise.

[0145] Extraction of key repayment behavior patterns

[0146] Analyze the product paths (PDs) of all simulation paths to identify common patterns:

[0147] Pattern 1: Early Default (ED)

[0148] Characteristics: High persistent connectivity components appear in the first 6 months (indicating continuous delinquency), accompanied by high-dimensional voids (indicating ineffective collection response).

[0149] PD performance: Components with Birth < 0.2 and Death > 0.6 account for more than 70%.

[0150] Mode 2: Gradual Improvement (GI)

[0151] Characteristics: Over time, connected components split (indicating a decrease in overdue components), with no significant voids.

[0152] PD performance: Birth values ​​are evenly distributed, and the mean Persist value is 0.25.

[0153] Pattern 3: Cyclic Fluctuation (CF)

[0154] Characteristics: Periodic appearance of ring-shaped structures (holes) reflects seasonal repayment difficulties.

[0155] PD performance: Multiple 1D holes with Death values ​​around 0.4.

[0156] Path topology feature matrix generation

[0157] Transform abstract topological features into structured matrices:

[0158] Matrix dimensions: Each row corresponds to a behavioral path, and each column is a topology descriptor.

[0159] Maximum Persistence Component (MPC) of Connected Components;

[0160] Average Persistence of Holes (APH);

[0161] Total Topological Features (TTF);

[0162] The first three principal components (PCs) are extracted from the original PD via PCA.

[0163] The hidden Markov model is driven by the path topology feature matrix, and the state transition probability is recalibrated by combining the reorganization scheme parameters, outputting the behavioral path probability distribution with confidence intervals.

[0164] Initial construction of Hidden Markov Model (HMM)

[0165] Hidden States (HS): Defines five types of debtor's true state:

[0166] S1: Financial Health (Healthy, H);

[0167] S2: Temporary Stress (TS);

[0168] S3: Persistent Difficulty (PD);

[0169] S4: Debt restructuring (R);

[0170] S5: Default (D).

[0171] Observed States (OS): Extracted from the topological feature matrix:

[0172] O1: MPC < 0.3;

[0173] O2: 0.3 ≤ MPC < 0.6;

[0174] O3: MPC ≥ 0.6.

[0175] Initial parameters:

[0176] Transition Matrix (TM): Based on historical data statistics (e.g., probability of S1→S2 is 0.2).

[0177] Emission Matrix (EM): P(O1|S1)=0.8 (low persistence is highly likely to be observed in a healthy state);

[0178] Initial State Distribution (ISD).

[0179] Probabilistic recalibration driven by recombination scheme

[0180] Restructuring plan parameters (such as extended repayment period, interest rate reduction) modification status transition rules:

[0181] Transition probability adjustment:

[0182] If the plan includes "5% interest rate reduction", the probability of S2 (short-term stress) → S1 (healthy) increases by 30%.

[0183] If the plan includes "extending the repayment period to 60 months", the probability of S3 (continued hardship) → S5 (default) decreases by 40%.

[0184] Parameterized adjustment formula: New transition probability TM_new(Si→Sj)=TM_old(Si→Sj)×(1+α) Re).

[0185] Where α is the scheme factor weight (set by business rules), and Re is the strength of restructuring measures (such as the extent of interest rate reduction).

[0186] Bayesian update: Using path data (considered as partial observation sequences) in the topological feature matrix, the HMM parameters are re-estimated using the Baum-Welch algorithm (BWA), making the model more consistent with the current simulation results.

[0187] Behavioral path probability distribution output

[0188] Prediction objective: Calculate the probability curve of being in state S5 (default) over the next 36 months.

[0189] Confidence interval generation:

[0190] Perform 100 Monte Carlo samplings on the recalibrated HMM, generating 1,000 virtual paths each time.

[0191] The mean probability (MP) and standard deviation (SD) of default probability are calculated monthly.

[0192] The 95% confidence interval (CI) is calculated as follows:

[0193]

[0194] Final output: Structured JSON data, example as follows:

[0195] json

[0196] {

[0197] "recovery_probability": {

[0198] "month_12": { "mean": 0.85, "ci_low": 0.82, "ci_high": 0.88},

[0199] "month_24": { "mean": 0.72, "ci_low": 0.69, "ci_high": 0.75},

[0200] "month_36": { "mean": 0.63, "ci_low": 0.60, "ci_high": 0.66}

[0201] },

[0202] "dominant_pattern": "Incremental improvement (pattern 2)"

[0203] }

[0204] Generative AI models (such as GANs or variational autoencoders) simulate possible behavioral paths of debtors under different restructuring schemes, such as whether their willingness to repay increases after interest rate reductions. The model outputs probability distributions for multiple scenarios, such as the probability of timely repayment, partial default, or complete default, breaking through the linear assumptions of traditional statistical models and capturing the nonlinear dynamic characteristics of debtor behavior. Multi-path simulation provides probabilistic decision-making basis for subsequent scheme optimization, significantly improving the robustness of predictions.

[0205] S203, based on the probability distribution of the behavior path and the constraints of the restructuring target, calculate and optimize the success probability of the restructuring plan to obtain the optimized debt restructuring plan and its corresponding estimated success probability value.

[0206] Specifically, a repayment success surface can be constructed based on the probability distribution of the behavior path, the restructuring target constraint can be transformed into a Lagrange multiplier, and a gradient ascent search under the constraint conditions can be performed on the repayment success surface to locate the boundary of feasible solutions.

[0207] The process of constructing the repayment success surface

[0208] The system receives the Behavior Path Probability Distribution (BPPD) generated in the previous step. This distribution, through Monte Carlo simulation, predicts the likelihood of debtors' repayment behavior over the next 12-36 months under different restructuring schemes. The core is extracting three key dimensions from the BPPD: Principal Recovery Rate (PRR, representing the percentage of principal recovered relative to total debt), Tenure Compression Ratio (TCR, reflecting the degree of shortening of the debt cycle after restructuring), and Comprehensive Cost Coefficient (CCC, covering collection / management / capital occupation costs). These three dimensions constitute a three-dimensional coordinate system (X=PRR, Y=TCR, Z=CCC). The discrete probability distribution points are transformed into a continuous surface—the Repayment Success Surface (RSS)—using the Kriging Interpolation algorithm. For example, the coordinates of a point (70%, 1.5, 0.85) represent a 70% principal recovery rate, a 1.5x time-limit reduction efficiency, and an 85% cost control level. Its height value represents the probability of achieving this state (e.g., 0.92). The peak regions on the surface correspond to clusters of high-probability-of-success-probability solutions.

[0209] Constraint transformation and gradient ascent search

[0210] Restructuring Target Constraints (RTCs) include creditor requirements (e.g., PRR ≥ 65%), debtor affordability (e.g., monthly repayments not exceeding 40% of income), and regulatory restrictions (e.g., maximum installment period ≤ 60 months). These constraints are transformed into mathematical boundary conditions using the Lagrange Multiplier Method.

[0211] For example, the minimum recovery rate constraint for creditors is expressed as g1(PRR) = 65% − PRR ≤ 0;

[0212] The debtor's repayment pressure constraint is transformed into g2(TCR) = monthly repayment ratio - 40% ≤ 0.

[0213] The constraint function is multiplied by the Lagrange multiplier λ (physically representing the penalty strength for constraint violation) and then superimposed with the RSS to form the penalized surface (PS). The optimization engine starts from the initial solution point (e.g., the current debt condition) and iteratively moves along the gradient direction of the PS (the steepest path for function value ascent). The gradient is calculated at each step using the Central Difference Method: for example, at the current point (PRR=60%, TCR=1.2), PRR is slightly increased to 60.1%, and the change in PS ΔPS / ΔPRR is observed; similarly, the directional partial derivative of TCR is calculated. The step size is controlled by the Adaptive Learning Rate (ALR) algorithm, increasing the step size in steep regions (e.g., 0.05 units) and decreasing the step size in flat regions (e.g., 0.01 units). The process stops when the improvement amount is less than the threshold ε (Epsilon=1e-5) for 10 consecutive iterations. At this point, the solution is located at the Feasible Solution Boundary (FSB) – the RSS edge region that satisfies all constraints.

[0214] Boundary positioning and stability enhancement

[0215] The Feasible Solution Boundary (FSB) is essentially a complex surface in 3D space formed by constraints cutting the Restricted Segment (RSS). To accurately describe its shape, the Alpha-Shape Algorithm is used to extract the boundary point cloud: setting an α radius parameter (e.g., α=0.3) to connect adjacent points, and removing internal points while retaining the contour point set. Simultaneously, noise robustness processing is injected: a Gaussian perturbation (standard deviation σ=0.5%) is applied to the boundary points, removing points that fail after the perturbation (e.g., PRR falling below 65%), and retaining stable boundary segments. The final output is the Robust Feasible Boundary (RBB), which serves as the search space for subsequent optimization.

[0216] Pareto front is constructed based on feasible solution boundary, and solution trade-off analysis is performed using multi-objective Bayesian optimization algorithm to generate Pareto set of candidate solutions.

[0217] Construction mechanism of Pareto front

[0218] On the reinforced feasible solution boundary (RFB), the system identifies conflicting dimensions of multi-objective optimization: for example, improving the principal recovery rate (PRR) requires shortening the repayment period (increasing the total repayment rate), but this increases the risk of debtor default; reducing costs (CCC) requires extending the repayment period (decreasing the total repayment rate), increasing capital tied up in losses. The Pareto Front (PF) is defined as: no better solution than the PF point exists on the RFB (i.e., there is no point where PRR, TCR, and CCC are all superior to the PF point). The construction process consists of three steps:

[0219] Non-dominated Sorting: Scan all scheme points on the RFB and mark the points that are not surpassed by all other points (e.g., if point A (PRR=75%, TCR=1.8, CCC=0.9) is better than point B (70%, 1.8, 0.92), then B is removed).

[0220] Crowding Distance Calculation: To preserve diversity, calculate the sum of distances between each point and its neighbors in each dimension (e.g., PRR dimension difference + TCR dimension difference).

[0221] Front extraction: Select the K points with the highest crowding among the non-dominated points (K=50) to form the initial front extraction.

[0222] Bayesian optimization trade-off analysis

[0223] The core of Multi-objective Bayesian Optimization (MOBO) is to construct a Gaussian Process Surrogate Model (GPSM) to predict the PRR / TCR / CCC performance in unexplored regions. Specific iterative process:

[0224] Proxy model training: GPSM is trained using 50 points from the initial PF to fit the three-dimensional objective function relationship.

[0225] Maximizing the acquisition function: Expected Hypervolume Improvement (EHVI) is used as the acquisition function. EHVI measures whether a new point can increase the hypervolume (HV) enclosed by the PF and the reference point (e.g., (PRR=0, TCR=0, CCC=1)). The maximum EHVI point (e.g., coordinates (72%, 1.6, 0.88)) is searched on the RFB using the DIRECT algorithm (Dividing Rectangles).

[0226] Solution Evaluation and Update: The success rate of new points is evaluated by calling the Neural Controlled Differential Equation Simulator (NCDES), and the Power Factor (PF) is updated. After repeating 20 iterations, the PF converges to a steady state.

[0227] Generate Pareto candidate set

[0228] The final Product Force (PF) comprises approximately 30 non-dominated solutions, categorized by business requirements:

[0229] Creditor priority type: PRR>78% but TCR>2.0 (e.g., Option C: PRR=79%, TCR=2.1, CCC=0.93);

[0230] Debtor-friendly type: TCR < 1.3 but PRR > 68% (e.g., Option D: PRR = 69%, TCR = 1.2, CCC = 0.87);

[0231] Balanced type: PRR≈73%, TCR≈1.5 (e.g., scheme E: PRR=73.5%, TCR=1.52, CCC=0.89).

[0232] All schemes constitute the Candidate Pareto Set (CPS), and each scheme is accompanied by a baseline success probability (BSP) derived from simulation (e.g., scheme C has a BSP of 81%).

[0233] Robust stress testing is performed on the Pareto set of candidate solutions, and the stability of the solutions is evaluated through adversarial sample injection. The optimized debt restructuring solution and the estimated success probability value are output.

[0234] Stress test scenario design

[0235] Robustness Stress Testing (RST) simulates three types of extreme scenarios:

[0236] Macroeconomic shocks: Injecting a 3 percentage point decrease in GDP growth and a 5% increase in the unemployment rate (by adjusting macroeconomic indicators that are concentrated in related environmental factors).

[0237] Individual debtor risks: simulated income reduction of 30% (modify the income stability parameter in the behavioral feature vector), and sudden increase in medical expenses (add a large consumption tag).

[0238] Market volatility: Benchmark interest rates were raised by 200 basis points, and the value of collateral decreased by 20%.

[0239] 100 adversarial samples are generated for each scenario, which is achieved by modifying the environment and behavior input parameters of the scheme in CPS.

[0240] Stability assessment and probability calibration

[0241] Input the adversarial samples into the Generative Behavior Model (GBM) and rerun the Monte Carlo simulation:

[0242] Success probability decay calculation: Record the decrease in success probability of each scheme under pressure (e.g., the BSP of scheme C changes from 81% to 63%, decay Δ=18%).

[0243] Risk contagion analysis: Detect the correlation of default among multiple debtors using the Dynamic Conditional Correlation (DCC) model (e.g., rising unemployment leads to a synchronous decline in the success rate of the cluster of solutions).

[0244] Stability score: according to formula S stab =BSP-0.5×Δmax (Δmax is the worst-case scenario decline). For example, if scenario E has an unemployment rate shock of Δ=12%, then S stab =76% - 0.5 × 12% = 70%.

[0245] Solution optimization and output

[0246] The final selection is based on a combination of stability scores and business rules.

[0247] Elimination mechanism: Remove S stab Solutions with a success rate of less than 65% or less than 50% in a single scenario (such as Solution C, which was abandoned due to excessive attenuation).

[0248] Parameter fine-tuning: Perform local sensitivity analysis (LSA) on the retained schemes and adjust key parameters (e.g., TCR of scheme D from 1.2 to 1.25, PRR increased to 71%).

[0249] Success Probability Calibration: Based on the stress test results, the BSP is adjusted to generate an estimated success probability (ESP). For example, the ESP of solution E is 0.7 × BSP + 0.3 × S stab = 0.7×76% + 0.3×70% = 74.2%.

[0250] The final output consists of 3-5 optimized restructuring plans (ORPs), each containing:

[0251] Restructuring terms (principal discount percentage / installment period / interest rate floating mechanism);

[0252] ESP value (e.g., in Scheme E: ESP = 74.2%);

[0253] Stress test report (success rate distribution for each scenario).

[0254] By combining the creditor's objectives (such as minimizing the loss rate) with the results of behavioral simulations, constrained optimization algorithms (such as the Lagrange multiplier method) are used to select the restructuring plan with the highest probability of success. For example, extending the repayment period rather than reducing the principal is chosen because it shows a higher probability of performance in the simulation. This realizes the transformation from theoretical prediction to an executable plan, ensuring that the restructuring plan not only meets the creditor's demands but is also compatible with the debtor's behavioral patterns, thereby reducing the risk of default after the plan is implemented.

[0255] S204. Based on the optimized debt restructuring plan and its corresponding estimated success probability value, the SaaS billing rule engine is used to perform dynamic billing strategy matching and cost calculation to obtain the billing details data corresponding to the restructuring plan.

[0256] Specifically, the estimated success probability value can be input into the SaaS billing rule engine, and the pre-designed billing template can be matched through the fuzzy logic system to generate a basic billing framework;

[0257] Once the optimized debt restructuring plan and its estimated success probability value (ESPV, representing the probability that the plan will ultimately successfully recover the debt, ranging from 0% to 100%) are generated, the ESPV is transmitted as a core input parameter to the SaaS Billing Rules Engine (SBRE). SBRE is the intelligent billing decision center built into the SaaS platform, integrating a template matching mechanism based on a fuzzy logic system (FLS). The predefined billing template (PBT) is a set of billing rules pre-defined based on historical business experience, industry standards, and customer types, such as charging based on a percentage of the restructuring amount, tiered rates, and fixed service fees. The core of FLS lies in handling continuous variables with uncertainty, such as ESPV. The system first divides ESPV into multiple fuzzy sets (FS), such as "high risk" (ESPV < 50%), "medium risk" (50% ≤ ESPV < 80%), and "low risk" (ESPV ≥ 80%). For each set, a membership function (MF, such as a triangular or trapezoidal function) is defined, and the degree of membership (DoM) of the current ESPV to each fuzzy set is calculated. For example, if ESPV = 65%, it might simultaneously belong to the "medium risk" set with a DoM of 0.7 and belong to the "high risk" set with a DoM of 0.3.

[0258] The matching process is based on a pre-defined Fuzzy Rule Base (FRB). Each rule is in the form of "IF (ESPV belongs to a certain fuzzy set) AND (other conditions) THEN (select a certain billing template)". For example: "IF ESPV is 'low risk' THEN adopt the 'low base rate + high success reward' template". The FLS Inference Engine (IE) activates all rules related to the membership of the current ESPV and calculates the firing strength (FS) of each rule based on the DoM. The system defuzzifies the results of the firing rules (i.e., the recommended billing templates), often using the centroid method to calculate the best-matching single billing template. For example, if the FS of the three rules are 0.2 (template A), 0.7 (template B), and 0.1 (template C), then template B is selected. This template includes core elements such as a basic rate structure (e.g., an annualized rate of 0.5%-2%), payment cycle (e.g., monthly), and billing trigger points (e.g., when each period's repayment is received), forming the initial Basic Billing Framework (BBF).

[0259] To enhance flexibility, the BBF is not fixed. When outputting the template, FLS also generates an Adaptation Coefficient (AC, ranging from 0.8 to 1.2) to fine-tune the baseline values ​​in the template. AC is determined by the precise value of ESPV and the fuzzy set to which it belongs. For example, if ESPV = 85% (strongly belonging to "low risk"), AC might be 1.1, indicating a 10% increase in the baseline rate to reflect the service premium brought by the high success probability; if ESPV = 55% (weakly belonging to "medium risk"), AC might be 0.9, indicating a need to appropriately lower the rate to balance risk. The final generated BBF is a structured object containing the billing model, baseline parameters, adaptation coefficient, and applicable conditions, laying the foundation for subsequent dynamic adjustments.

[0260] Based on the risk exposure characteristics of the optimized debt restructuring plan, the risk value is calculated online to generate a dynamic risk premium coefficient;

[0261] The risk exposure characteristics (REC) inherent in the optimized debt restructuring plan (ODRP) are a key basis for dynamic billing. REC includes the principal balance (PB), remaining tenor (RT), interest rate type (fixed / floating), and collateral value volatility (CVV). The system uses a Value at Risk (VaR) model to calculate the maximum potential loss of the plan online at a specific confidence level (e.g., 95%) and holding period (e.g., the restructuring period). The calculation employs either historical simulation (HS) or Monte Carlo simulation (MCS). Taking MCS as an example: the system calls upon historical economic data (e.g., interest rates, exchange rates, unemployment rates) and debtor-specific data (e.g., historical repayment volatility) to construct a joint distribution model of risk factors, generating thousands of random paths to simulate future cash flows, and statistically analyzing the quantiles of the loss distribution to obtain the VaR value (e.g., 100,000 yuan).

[0262] To further capture tail risk, the system calculates Conditional Value at Risk (CVaR), which is the average loss when the loss exceeds the VaR threshold (e.g., 120,000 RMB). CVaR reflects extreme risk better than VaR. Based on the CVaR results, the system converts them into a Dynamic Risk Premium Coefficient (DRPC) using a Risk Premium Model (RPM). RPM can be a linear mapping (e.g., DRPC = base coefficient + slope × CVaR) or a non-linear function (e.g., the Sigmoid function). For example, setting a CVaR range of 50,000-200,000 RMB corresponds to a DRPC fluctuating between 1.0 and 2.0: if CVaR = 120,000 RMB, then DRPC = 1.5. This coefficient directly affects the rate benchmark in BBF to compensate creditors for the additional risk they bear.

[0263] Risk calculation is real-time (online) and takes into account the characteristics of the investment plan. For example, if the plan includes a "principal waiver" clause (i.e., a portion of the principal does not need to be repaid), the system will directly include the waived amount (WA) as a certain loss in the risk exposure; if the plan stipulates that "interest rates are linked to the inflation index," the system will enhance the sensitivity analysis to inflation volatility. All risk factors are captured for correlation through a covariance matrix (CM), ensuring that the premium coefficient accurately reflects portfolio risk rather than isolated risk. The final generated DRPC is a floating parameter that can be updated over time, ensuring that billing matches real-time risk.

[0264] By integrating the entropy features of debtor behavior feature vectors, a service complexity weight factor is generated through a reinforcement learning model;

[0265] The Debtor Behavior Feature Vector (DBFV) is high-dimensional data generated in previous steps that characterizes a debtor's repayment habits and stability (e.g., historical delinquency counts, income volatility, frequency of contact information changes). To quantify its disorder or uncertainty, the system calculates the information entropy (IE) of this vector. For example, the Shannon entropy formula (but avoiding the formula) is used to analyze the probability distribution of each feature dimension: if a debtor's feature values ​​are highly concentrated (e.g., stable income), the entropy value is low; if the feature values ​​are dispersed (e.g., irregular repayment dates), the entropy value is high. A high entropy value suggests that managing this debtor requires higher operating costs. This entropy characteristic (EC) is extracted as a scalar (e.g., a 0-1 normalized value).

[0266] The Service Complexity Weight Factor (SCWF) is generated by a Reinforcement Learning Model (RLM). The RLM is built upon a Markov Decision Process (MDP).

[0267] State (S): Includes the current EC value, ODRP term complexity (if cross-border payments are involved), and historical service cost (HSC).

[0268] Action (A): Outputs SCWF (e.g., 0.5-1.5), which directly affects the final rate.

[0269] Reward (R): Designed as a comprehensive indicator: positive rewards come from service fee revenue (Billing Revenue, BR) and customer satisfaction score (CSS); negative rewards come from operational cost (OC) and complaint rate (CR).

[0270] Policy (π): Using Deep Q-Network (DQN) or Proximal Policy Optimization (PPO) algorithms, the strategy learns to select the optimal SCWF to maximize long-term cumulative rewards under a given state through extensive historical trading simulation training.

[0271] When the model runs online, the current state (S) is input into the trained RLM, and the model outputs the optimal action (A), i.e., SCWF. For example, in complex cases with high EC (debtor behavior instability) and high HSC (historical service resource consumption), the RLM may output SCWF=1.4, indicating that a 40% service complexity surcharge should be added to the base fee. The RLM continuously learns from new transactions and regularly updates the Policy Network Parameters (PNP) to ensure that SCWF dynamically adapts to changes in the business environment.

[0272] Based on the basic billing framework, dynamic risk premium coefficient and service complexity weight factor, a genetic optimization algorithm is used to fine-tune multi-dimensional parameters and generate the optimal combination of billing parameters.

[0273] The aforementioned steps generated three core inputs:

[0274] Basic Billing Framework (BBF): Provides the billing structure and initial parameters (such as the base rate R_base=1.2%).

[0275] Dynamic Risk Premium (DRPC): Reflects the risk of the project (e.g., DRPC=1.5).

[0276] Service Complexity Weighting Factor (SCWF): Reflects operational difficulty (e.g., SCWF=1.2).

[0277] The system needs to integrate and optimize these parameters to form the final billing parameters. This translates into a multi-objective optimization problem (MOOP) with objectives including: maximizing the creditor net yield (CNY), minimizing the debtor default probability (DDP), and balancing the platform service cost (PSC). Constraints include: regulatory cap compliance (RC) and debtor affordability threshold (DAT).

[0278] The Genetic Optimization Algorithm (GOA) is used to solve:

[0279] Encoding: The parameters to be optimized (such as the final rate R_final, payment frequency PF, and waiver trigger point (WT)) are encoded as "chromosomes" (CHR). For example, CHR can be represented as a binary string or a vector of real numbers.

[0280] Initial Population: N sets (e.g., 100 sets) of parameter combinations are randomly generated as initial solutions.

[0281] Fitness Function (FF): Evaluates the quality of each set of parameters. FF = w1×CNY - w2×DDP - w3×PSC, where the weights w1, w2, and w3 are set by the business strategy. It also checks whether the RC and DAT constraints are met.

[0282] Selection: Based on FF scores, select the best individuals for the next generation using Roulette Wheel Selection (RWS) or Tournament Selection (TS).

[0283] Crossover: Exchanges parameters of selected individuals (such as single-point crossover) to generate a new solution.

[0284] Mutation: Maintaining population diversity by randomly perturbing certain parameters with low probability (e.g., R_final ± 0.1%).

[0285] The algorithm iteratively performs selection-crossover-mutation operations to gradually approach the Pareto Optimal Set (POS).

[0286] After a preset number of iterations (e.g., 200 generations) or a convergence determination, the algorithm selects an optimal billing parameter combination (OBPC) from the POS based on business preferences (e.g., prioritizing high CNY). For example: R_final = R_base × DRPC × SCWF × adjustment factor = 1.2% × 1.5 × 1.2 × 0.98 ≈ 2.12%, payment frequency PF = quarterly payment, waiver clause WT = triggered by three consecutive on-time payments. OBPC ensures a comprehensive balance of interests and risks among all parties while meeting constraints.

[0287] The system automatically executes the optimal combination of billing parameters through smart contracts, calibrates fees based on real-time market interest rate fluctuations, and outputs structured billing details.

[0288] Once the Optimal Billing Parameter Combination (OBPC) is generated, it is injected into a smart contract (SC) deployed on the blockchain. This SC is pre-written program code that automatically executes the billing logic (only the functionality is described here, not the code itself). Key inputs to the SC include: OBPC, Debtor ID (DID), Creditor Account (CA), and Plan Serial Number (PSN). When contract triggering conditions (such as the repayment due date or the fulfillment of waiver clauses) are met, the SC automatically initiates the billing process.

[0289] To ensure fee fairness, SC connects to a real-time market rate feed (RMRF, such as LIBOR or SHIBOR). SC has an embedded Fee Calibration Module (FCM) that dynamically reads the latest market interest rate at each billing point, based on the calibration rules agreed upon in the OBPC (e.g., "Floating rate = Base rate + Spread × Current SHIBOR"), and recalculates the current period's fees accordingly. For example, if the agreed spread is 1.5% and the current SHIBOR is 3.5%, then the floating rate = 2.0% (fixed portion) + 1.5% × 3.5% = 2.0525%. This process eliminates pricing discrepancies caused by interest rate fluctuations.

[0290] After the smart contract executes, it generates structured billing detail data (SBDD). SBDD contains machine-readable key fields: Billing Cycle (BC), Due Principal (DP), Due Fee (DF), Applied Rate (AR), Calibration Basis (CB), Waiver Status (WS), and Transaction Hash (TH). The data is output in standard formats such as JSON or XML and pushed via API to the SaaS platform's Billing Module (BM), Financial System (FS), and Debtor Portal (DP), achieving a fully automated, auditable, and tamper-proof billing closed loop. At this point, the complete billing detail data for this restructuring scheme has been generated.

[0291] The billing engine automatically matches differentiated rate templates based on the risk level and service complexity of the solution, and generates the final fee by overlaying real-time market parameters. For example, it charges a higher service premium for high-risk solutions, accurately quantifies risk pricing and service value, avoids risk-return mismatch under the traditional fixed rate model, and enhances the sustainability of the business model through dynamic adjustments.

[0292] S205, the billing details data are structured, encapsulated, and output in a standardized manner to generate a debt restructuring billing instruction that can be directly integrated into the billing module of the SaaS platform.

[0293] Specifically, the billing details data can be semantically parsed, and a machine-readable billing business entity relationship diagram can be generated through the ontology reasoning engine;

[0294] The core processing flow of semantic parsing

[0295] Billing Detail Data (BDD) contains multi-dimensional information, such as Base Service Fee (BSF), Risk Premium (RP), and Complexity Weight (CW). The system first uses semantic parsing technology, employing a natural language processing model (such as a financial-domain-adjusted version of BERT) to identify key business entities and their attributes within the BDD. For example, when parsing "Dynamic Risk Premium Coefficient: 1.25," the model identifies "Dynamic Risk Premium Coefficient" as the entity type, "1.25" as the numerical attribute, and associates it with its unit of measurement (such as a multiple). Simultaneously, the system extracts implicit entities (such as the "Credit Rating" entity in "Debtor Credit Rating B") through Named Entity Recognition (NER) to construct an initial Entity-Attribute Pair Set (EAPS).

[0296] Rule-driven mechanism of ontology reasoning engine

[0297] Based on a predefined Billing Ontology Library (BOL), the system inputs EAPS into the Ontology Reasoning Engine (ORE). The BOL is constructed using the OWL (Web Ontology Language) to define the hierarchical relationships and constraint rules of core billing domain concepts (such as ServiceFee and RiskAdjustment). For example:

[0298] Rule 1: RiskPremium is a subclass of AdjustmentFactor and must be associated with a RiskExposure entity.

[0299] Rule 2: The range of ComplexityWeight is [0.5, 2.0].

[0300] ORE automatically completes missing relationships through Description Logic Reasoning (DLR): If a "Service Complexity Weight Factor: 1.8" exists in the BDD but is not associated with a specific service item, the engine associates it with the preset DebtRestructuringService main entity according to BOL rules. The final output is a machine-readable Billing Entity Relationship Graph (MRBERG), stored in RDF triple (Resource Description Framework Triple) format, such as (Debtor 123, Applicable Rate Template, High-Risk Template A).

[0301] Real-time update mechanism of dynamic knowledge graph

[0302] To adapt to the multi-tenant scenario of the SaaS platform, the system introduces an incremental ontology learning (IOL) mechanism. When a new entity (such as "green bond subsidy coefficient") is parsed, ORE automatically triggers the BOL expansion process: first, it compares the similarity of the existing ontology structure using a clustering algorithm (such as DBSCAN), and then, after confirmation by the Human Review Interface (HRI), the new entity is inserted into the BOL hierarchically. At the same time, MRBERG is dynamically updated through a graph database (such as Neo4j) to ensure that entity relationships are always synchronized with business rules.

[0303] A billing instruction template is constructed based on an entity relationship diagram. The logical completeness is ensured through formal verification, and a verified instruction framework is generated.

[0304] Automated generation of billing instruction templates

[0305] Using MRBERG as input, the system extracts the critical path from its topology, which serves as the skeleton of the Billing Instruction Template (BIT). For example, the path identified from the relationship graph is: Debtor → Restructuring Plan → Risk Premium → Basic Service Fee → Total Cost, and mapped to the hierarchical logical structure of the template. The template uses JSON-LD (JSON for Linked Data) format, and each node is bound to a predefined operation function (OF) from the billing rule engine. For example:

[0306] The risk premium node is bound to the function calculateRiskPremium(exposureScore,marketVolatility), where exposureScore (risk exposure score) comes from the behavioral feature vector and marketVolatility (market volatility) comes from the real-time economic indicator API.

[0307] Template variables (TV) are dynamically injected using placeholders (such as ${riskAdjustment}) to ensure compatibility with the input interface of the SaaS billing module.

[0308] Logical completeness guarantee of formal verification

[0309] To avoid instruction logic conflicts, the system employs formal verification (FV) technology. First, the BIT is converted into a finite state machine (FSM) model:

[0310] State: Billing phase (e.g., "Basic Fee Calculation", "Risk Adjustment").

[0311] Transition Condition: Dependency (e.g., "Complexity weight is triggered only when risk premium > 1.0").

[0312] Use model checking tools (such as NuSMV) to perform property verification (PV) on the FSM. The core properties include:

[0313] Deadlock Freedom: Ensures that instructions can continue to execute in any state.

[0314] Boundary Consistency: For example, verifying that "Total Cost = Base Cost × Risk Premium × Complexity Weight" still outputs a reasonable value when the parameter boundary (e.g., premium coefficient = 0).

[0315] If a violation of an attribute is detected (such as an anomaly where the premium coefficient is negative), the system will automatically backtrack to MRBERG to fill in the missing rules.

[0316] Instruction framework generation and fault tolerance enhancement

[0317] The verified BIT is upgraded to a Verified Instruction Framework (VIF). To improve robustness, VIF embeds a Fault-Tolerant Subroutine (FTS):

[0318] If a function's input parameters are missing (e.g., market volatility is not obtained), FTS will call an alternative data source (e.g., historical average volatility).

[0319] If the calculation times out, a degradation strategy is triggered (such as using the cached result and marking the confidence level as 0.8).

[0320] VIF is ultimately encapsulated in XML Schema (Extensible Markup Language Schema), which specifies the data type (e.g., xs:decimal) and validation rules (e.g., minInclusive="0.5") of each field.

[0321] The key parameters of the optimized debt restructuring plan will be embedded into the instruction framework through homomorphic encryption to form a tamper-proof billing instruction prototype.

[0322] Extraction of key parameters and sensitivity classification

[0323] Key parameters (KP) are extracted from the Optimized Debt Restructuring Plan (ODRP) and categorized by sensitivity:

[0324] Level P1 (Highly Sensitive): Debtor ID, Actual Repayment Amount (ARA), Success Probability Value (SPV).

[0325] P2 level (moderately sensitive): Installment Count (IC), Interest Rate Clause (IRC).

[0326] Level P3 (Low Sensitivity): Service Validity Period (SVP).

[0327] The tiered rules are based on GDPR (General Data Protection Regulation) compliance requirements and are dynamically set by the SaaS platform's Data Governance Policy Engine (DGPE).

[0328] Homomorphic encryption parameter embedding mechanism

[0329] Use additive homomorphic encryption (AHE) algorithms (such as the Paillier cryptosystem) to process P1 level parameters:

[0330] Encryption phase: Generate a key pair (KP) for each KP, and encrypt the parameters with the public key (PK). For example, encrypt the SPV value "0.85" into ciphertext C(SPV)=E(PK, 0.85).

[0331] Embedding operation: The ciphertext is directly injected into the corresponding placeholder in the VIF. Due to the characteristics of AHE, when the billing module subsequently executes C (total cost) = C (base cost) × C (risk premium), the calculation can be completed without decryption, ensuring that the original parameters are invisible throughout the process.

[0332] P2 / P3 level parameters employ Attribute-Based Encryption (ABE), dynamically generating decryption strategies based on the data usage scenario (e.g., "for monthly closing reports only").

[0333] Implementation of anti-tampering mechanism and integrity verification

[0334] To create a tamper-proof billing instruction prototype (TPBIP), the system implements dual protection:

[0335] Digital Fingerprint (DF): Calculate the SHA-256 hash value (HashValue, HV) of the entire VIF content and bind it to the encrypted KP for storage.

[0336] Merkle Tree Verification (MTV): Instructions are broken down into data blocks (DB) and a Merkle Tree (MT) is constructed. If any block is tampered with (e.g., by maliciously modifying the rate value), its hash path (HP) immediately becomes invalid.

[0337] TPBIP is ultimately serialized into a binary message (BM), with a version number (VN) and an algorithm ID (AID) appended to the header.

[0338] The instruction prototype is encapsulated using a cross-chain interoperability protocol, with zero-knowledge proof verification credentials added, and the output is a standardized debt restructuring billing instruction that can be directly integrated.

[0339] Layered encapsulation of cross-chain interoperability protocols

[0340] To adapt to multi-blockchain environments (such as SaaS platforms simultaneously connecting to Ethereum and Hyperledger Fabric), the TPBIP is encapsulated using the Cross-Chain Interoperability Protocol (CCIP).

[0341] Adaptation Layer (AL): Converts the binary format of TPBIP into a target chain-compatible format (such as Ethereum's RLP encoding (Recursive Length Prefix, RLP) or Fabric's Protocol Buffers (PB)).

[0342] Routing Layer (RL): Forwards instructions through Inter-Chain Relay (ICR). Routing Table (RT) records the target chain ID (CID) and gateway address (GA).

[0343] When encapsulating, add cross-chain metadata (CCM), which includes source chain hash (SCH), timestamp (TS), and fee deduction account (FDA).

[0344] Generation of zero-knowledge proof verification credentials

[0345] Zero-Knowledge Proof (ZKP) is used to prove the validity of instructions to the billing module without disclosing sensitive information.

[0346] Statement Construction (SC): Generates a Verifiable Claim (VC), such as "The risk premium coefficient calculation in this instruction conforms to the output of the rules engine and does not exceed the threshold of 2.0".

[0347] Proof Generation (PG): This process uses the zk-SNARK (Succinct Non-Interactive Argument of Knowledge) algorithm to generate proof credentials (PC). Key steps include:

[0348] Convert the billing rules into an arithmetic circuit (AC).

[0349] Input encryption parameters and rule constraints to generate a proving key (PK) and a verification key (VK).

[0350] Output PC (approximately 1KB in length), which can independently verify the authenticity of the claim.

[0351] The PC is attached to the end of the instruction package and marked with its corresponding Claim ID (CID).

[0352] The final output and integration readiness of standardized instructions

[0353] Once the instruction is encapsulated, it is converted into a Standardized Debt Restructuring Billing Instruction (SDRBI), which has the following characteristics:

[0354] Structural standardization: Adopting a unified encapsulation format (such as the CCIP message standard defined by the IETF (Internet Engineering Task Force)), including:

[0355] Header: Protocol Version (PV) and Instruction Type (IT).

[0356] Body: Encrypted instruction prototype + cross-chain metadata.

[0357] Trailer: ZKP certificate + digital signature (DS).

[0358] Integration interface compatibility: Provides a RESTful API (Representational State Transition Application Programming Interface) endpoint (APIEndpoint, AE). The SaaS billing module can trigger instruction execution by sending {chain_id,instruction_data} via a POST request (HTTP POST Request, HPR).

[0359] Self-verification capability: After receiving the SDRBI, the billing module first verifies the PC's authenticity using ZKP's VK, and then decrypts the execution content. Upon successful verification, it returns an Integration Success Signal (ISS).

[0360] The cost calculation results are converted into standardized instructions (such as JSON format) containing fields such as billing subject, amount, and effective time, and then transmitted to the SaaS billing system via API encryption. This ensures seamless integration with existing business processes, eliminates errors and delays caused by manual intervention, improves the automation level and audit traceability of the billing process, and meets financial-grade compliance requirements.

[0361] As can be seen, by receiving basic debt data and restructuring target constraints input by users through the SaaS platform, the system obtains debtor behavioral feature vectors and a set of related environmental factors. Based on these, it obtains behavioral path probability distributions reflecting potential repayment outcomes under different restructuring schemes. According to the behavioral path probability distributions and restructuring target constraints, it obtains optimized debt restructuring schemes and their corresponding estimated success probabilities. Based on the optimized debt restructuring schemes and their corresponding estimated success probabilities, it obtains detailed billing data for these schemes. The detailed billing data is then structured, encapsulated, and standardized for output, generating debt restructuring billing instructions that can be directly integrated into the SaaS platform's billing module. This improves the accuracy and efficiency of debt restructuring services and billing.

[0362] Another embodiment of the present invention provides a debt restructuring billing system based on an AI-driven SaaS architecture, see [link to relevant documentation]. Figure 3 The system may include:

[0363] The receiving module 301 is used to receive the basic debt data and restructuring target constraints input by the user through the SaaS platform, and to extract the debtor's historical behavioral characteristics and perform environmental variable correlation processing based on the basic debt data to obtain the debtor's behavioral feature vector and the set of associated environmental factors.

[0364] The simulation module 302 is used to simulate multi-path future repayment behavior based on the debtor's behavioral feature vector and the set of associated environmental factors, using a pre-trained generative behavioral simulation model to obtain a probability distribution of behavioral paths that reflects the potential repayment results under different restructuring schemes.

[0365] The optimization module 303 is used to calculate and optimize the success probability of the restructuring scheme based on the probability distribution of the behavior path and the restructuring target constraint, so as to obtain the optimized debt restructuring scheme and its corresponding estimated success probability value.

[0366] Matching module 304 is used to perform dynamic billing strategy matching and cost calculation processing through the SaaS billing rule engine based on the optimized debt restructuring plan and its corresponding estimated success probability value, so as to obtain the billing details data corresponding to the restructuring plan.

[0367] The generation module 305 is used to perform structured encapsulation and standardized output processing on the billing details data to generate a debt restructuring billing instruction that can be directly integrated into the billing module of the SaaS platform.

[0368] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0369] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps:

[0370] S201, Receive the basic debt data and restructuring target constraints input by the user through the SaaS platform, and extract the debtor's historical behavioral characteristics and perform environmental variable correlation processing based on the basic debt data to obtain the debtor's behavioral characteristic vector and the set of associated environmental factors;

[0371] S202, based on the debtor's behavioral feature vector and the set of associated environmental factors, a pre-trained generative behavioral simulation model is used to simulate multi-path future repayment behavior to obtain the probability distribution of behavioral paths that reflect the potential repayment results under different restructuring schemes.

[0372] S203, based on the probability distribution of the behavior path and the constraints of the restructuring target, calculate and optimize the success probability of the restructuring plan to obtain the optimized debt restructuring plan and its corresponding estimated success probability value.

[0373] S204. Based on the optimized debt restructuring plan and its corresponding estimated success probability value, the SaaS billing rule engine is used to perform dynamic billing strategy matching and cost calculation to obtain the billing details data corresponding to the restructuring plan.

[0374] S205, the billing details data are structured, encapsulated, and output in a standardized manner to generate a debt restructuring billing instruction that can be directly integrated into the billing module of the SaaS platform.

[0375] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0376] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0377] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0378] S201, Receive the basic debt data and restructuring target constraints input by the user through the SaaS platform, and extract the debtor's historical behavioral characteristics and perform environmental variable correlation processing based on the basic debt data to obtain the debtor's behavioral characteristic vector and the set of associated environmental factors;

[0379] S202, based on the debtor's behavioral feature vector and the set of associated environmental factors, a pre-trained generative behavioral simulation model is used to simulate multi-path future repayment behavior to obtain the probability distribution of behavioral paths that reflect the potential repayment results under different restructuring schemes.

[0380] S203, based on the probability distribution of the behavior path and the constraints of the restructuring target, calculate and optimize the success probability of the restructuring plan to obtain the optimized debt restructuring plan and its corresponding estimated success probability value.

[0381] S204. Based on the optimized debt restructuring plan and its corresponding estimated success probability value, the SaaS billing rule engine is used to perform dynamic billing strategy matching and cost calculation to obtain the billing details data corresponding to the restructuring plan.

[0382] S205, the billing details data are structured, encapsulated, and output in a standardized manner to generate a debt restructuring billing instruction that can be directly integrated into the billing module of the SaaS platform.

[0383] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A debt restructuring billing method based on an AI-driven SaaS architecture, characterized in that, The method includes: The system receives basic debt data and restructuring target constraints input by users through the SaaS platform. Based on the basic debt data, it extracts the debtor's historical behavioral characteristics and performs environmental variable correlation processing to obtain the debtor's behavioral characteristic vector and the set of associated environmental factors. Based on the debtor's behavioral feature vector and the associated environmental factor set, a pre-trained generative behavioral simulation model is used to simulate multi-path future repayment behavior, resulting in a probability distribution of behavioral paths that reflects the potential repayment outcomes under different restructuring schemes. Based on the probability distribution of the behavioral path and the constraints of the restructuring target, the success probability of the restructuring plan is calculated and optimized to obtain the optimized debt restructuring plan and its corresponding estimated success probability value. Based on the optimized debt restructuring plan and its corresponding estimated success probability, the SaaS billing rule engine is used to perform dynamic billing strategy matching and cost calculation to obtain the billing details data corresponding to the restructuring plan. The billing details data are structured, encapsulated, and output in a standardized manner to generate a debt restructuring billing instruction that can be directly integrated into the billing module of a SaaS platform.

2. The method according to claim 1, characterized in that, The system receives basic debt data and restructuring target constraints input by users through the SaaS platform. Based on the basic debt data, it extracts historical behavioral characteristics of the debtor and performs environmental variable correlation processing to obtain a debtor behavioral feature vector and a set of associated environmental factors, including: Based on the multi-source heterogeneous nature of the basic debt data, data authenticity verification and timestamp alignment are performed using a blockchain consensus mechanism to generate a standardized debt dataset with digital fingerprints. The standardized debt dataset is input into a spatiotemporal graph convolutional network to mine historical transaction patterns of debtors and perform cross-entity correlation analysis, extracting behavioral pattern subgraphs of debtors. Based on the behavioral pattern subgraph and economic indicators under the current macro environment, environmental sensitivity factor embedding calculation is performed, and environmental weighted feature vector is generated through graph attention mechanism; The environmental weighted feature vector is encrypted using federated learning and integrated with privacy-preserving third-party credit data to output a debtor behavior feature vector with dynamic weights and a set of associated environmental factors.

3. The method according to claim 2, characterized in that, The process involves using a pre-trained generative behavioral simulation model to simulate multi-path future repayment behavior based on the debtor's behavioral feature vector and associated environmental factor set, resulting in a probability distribution of behavioral paths reflecting potential repayment outcomes under different restructuring schemes. This includes: The debtor's behavioral feature vector is input into a pre-trained generative adversarial network, and the debtor's behavioral latent vector is generated through a conditional variational autoencoder. Based on the set of associated environmental factors, stochastic differential equations are constructed to perform Monte Carlo environmental disturbance simulation and generate a cluster of dynamic environmental evolution paths. The behavioral latent vectors and environmental evolution path clusters are input into the neural control differential equations to perform multi-agent collaborative simulation and output the original behavioral path trajectories. By analyzing the topological persistence of the path trajectory, the continuous coherence algorithm is applied to extract key repayment behavior patterns and generate a path topological feature matrix. The hidden Markov model is driven by the path topology feature matrix, and the state transition probability is recalibrated by combining the reorganization scheme parameters, outputting the behavioral path probability distribution with confidence intervals.

4. The method according to claim 3, characterized in that, The step of calculating and optimizing the success probability of the restructuring plan based on the probability distribution of the behavioral path and the constraints of the restructuring target, to obtain the optimized debt restructuring plan and its corresponding estimated success probability value, includes: Based on the probability distribution of the behavior path, a repayment success surface is constructed, and the restructuring objective constraint is transformed into a Lagrange multiplier. Gradient ascent search under the constraint conditions is performed on the repayment success surface to locate the boundary of feasible solutions. Pareto front is constructed based on feasible solution boundary, and solution trade-off analysis is performed using multi-objective Bayesian optimization algorithm to generate Pareto set of candidate solutions. Robust stress testing is performed on the Pareto set of candidate solutions, and the stability of the solutions is evaluated through adversarial sample injection. The optimized debt restructuring solution and the estimated success probability value are output.

5. The method according to claim 4, characterized in that, Based on the optimized debt restructuring plan and its corresponding estimated success probability, the SaaS billing rule engine performs dynamic billing strategy matching and cost calculation to obtain the billing details data corresponding to the restructuring plan, including: Input the estimated success probability value into the SaaS billing rule engine, and use the fuzzy logic system to match the pre-designed billing template to generate a basic billing framework; Based on the risk exposure characteristics of the optimized debt restructuring plan, the risk value is calculated online to generate a dynamic risk premium coefficient; By integrating the entropy features of debtor behavior feature vectors, a service complexity weight factor is generated through a reinforcement learning model; Based on the basic billing framework, dynamic risk premium coefficient and service complexity weight factor, a genetic optimization algorithm is used to fine-tune multi-dimensional parameters and generate the optimal combination of billing parameters. The system automatically executes the optimal combination of billing parameters through smart contracts, calibrates fees based on real-time market interest rate fluctuations, and outputs structured billing details.

6. The method according to claim 5, characterized in that, The process of structurally encapsulating and standardizing the billing details data to generate a debt restructuring billing instruction that can be directly integrated into the billing module of a SaaS platform includes: Semantic parsing of billing details data is performed, and a machine-readable billing business entity relationship graph is generated through an ontology reasoning engine; A billing instruction template is constructed based on an entity relationship diagram. The logical completeness is ensured through formal verification, and a verified instruction framework is generated. The key parameters of the optimized debt restructuring plan will be embedded into the instruction framework through homomorphic encryption to form a tamper-proof billing instruction prototype. The instruction prototype is encapsulated using a cross-chain interoperability protocol, with zero-knowledge proof verification credentials added, and the output is a standardized debt restructuring billing instruction that can be directly integrated.

7. A debt restructuring billing system based on an AI-driven SaaS architecture, characterized in that, The system includes: The receiving module is used to receive the basic debt data and restructuring target constraints input by the user through the SaaS platform, and to extract the debtor's historical behavioral characteristics and perform environmental variable correlation processing based on the basic debt data to obtain the debtor's behavioral feature vector and the set of associated environmental factors. The simulation module is used to simulate multi-path future repayment behavior based on the debtor's behavioral feature vector and the set of associated environmental factors, using a pre-trained generative behavioral simulation model to obtain the probability distribution of behavioral paths that reflect the potential repayment results under different restructuring schemes. The optimization module is used to calculate and optimize the success probability of the restructuring plan based on the probability distribution of the behavior path and the constraints of the restructuring target, so as to obtain the optimized debt restructuring plan and its corresponding estimated success probability value. The matching module is used to perform dynamic billing strategy matching and cost calculation processing through the SaaS billing rule engine based on the optimized debt restructuring plan and its corresponding estimated success probability value, so as to obtain the billing details data corresponding to the restructuring plan. The generation module is used to perform structured encapsulation and standardized output processing on the billing details data, and generate debt restructuring billing instructions that can be directly integrated into the billing module of the SaaS platform.

8. The system according to claim 7, characterized in that, The receiving module is specifically used for: Based on the multi-source heterogeneous nature of the basic debt data, data authenticity verification and timestamp alignment are performed using a blockchain consensus mechanism to generate a standardized debt dataset with digital fingerprints. The standardized debt dataset is input into a spatiotemporal graph convolutional network to mine historical transaction patterns of debtors and perform cross-entity correlation analysis, extracting behavioral pattern subgraphs of debtors. Based on the behavioral pattern subgraph and economic indicators under the current macro environment, environmental sensitivity factor embedding calculation is performed, and environmental weighted feature vector is generated through graph attention mechanism; The environmental weighted feature vector is encrypted using federated learning and integrated with privacy-preserving third-party credit data to output a debtor behavior feature vector with dynamic weights and a set of associated environmental factors.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-6 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-6.

Images