AI-based dilemma enterprise data information and asset transaction service method

Abstract

The invention discloses an AI-based dilemma enterprise data information and asset transaction service method, and the method comprises the steps: receiving multi-source heterogeneous enterprise data of a dilemma enterprise, and obtaining a structured and desensitized dilemma enterprise panoramic data portrait; portraying the panoramic data of the dilemma enterprise to obtain a multi-dimensional dynamic risk feature vector including an asset state, a potential liability, an operation interruption risk and an industry prosperity association degree; associating the dynamic risk feature vector with real-time market transaction data and an industry merger and purchase case library to obtain a dynamic discount price interval and confidence evaluation for specific dilemma assets; obtaining an optimal potential investor short list and a preliminary conflict-free transaction structure scheme according to the preset preference of the potential investors and the dynamic discount price interval; and obtaining a final traceable dilemma asset transaction record for the transaction which achieves the intention. According to the embodiment of the invention, asset pricing and transaction matching can be dynamically optimized, and the efficiency and security of dilemma asset transaction are improved.

Description

Technical Field

[0001] This invention belongs to the field of AI technology, specifically an AI-based method for providing data information and asset trading services to distressed enterprises. Background Technology

[0002] Currently, distressed companies face challenges in asset transactions, including data fragmentation, insufficient privacy protection, and pricing difficulties. Traditional methods rely on manual assessment, which struggles to integrate multi-source, heterogeneous data (such as financial, operational, and industry data), leading to information asymmetry. Furthermore, cross-institutional data sharing poses privacy risks, while static valuation models fail to dynamically reflect market changes, resulting in asset pricing discrepancies. Existing trading platforms lack intelligent matching and automated execution capabilities, leading to low transaction efficiency and frequent disputes. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-based method for data information and asset trading services for distressed enterprises, in order to address the shortcomings of existing technologies, dynamically optimize asset pricing and transaction matching, and improve the efficiency and security of distressed asset transactions.

[0004] One embodiment of this application provides an AI-based method for providing data information and asset trading services for distressed enterprises, the method comprising:

[0005] It receives multi-source heterogeneous enterprise data of distressed enterprises, performs cross-domain privacy-preserving fusion processing based on a federated learning framework, and obtains a structured and de-identified panoramic data profile of distressed enterprises.

[0006] For the panoramic data profile of the distressed enterprises, a pre-trained deep adversarial generative network is used to extract asset risk features and model value influencing factors, resulting in a multi-dimensional dynamic risk feature vector that includes asset status, potential liabilities, operational interruption risk, and industry prosperity correlation.

[0007] The dynamic risk feature vector is associated with real-time market transaction data and industry M&A case library. Based on reinforcement learning algorithm, dynamic discount rate calculation and optimal asset pricing range deduction are performed to obtain dynamic discount price range and confidence assessment for specific distressed assets.

[0008] Based on the preset preferences of potential investors and the dynamic discount price range, a multi-agent simulation negotiation and conflict resolution algorithm is used to perform intelligent and precise matching and transaction feasibility simulation processing to obtain the optimal shortlist of potential investors and a preliminary conflict-free transaction structure scheme.

[0009] For transactions that have reached an agreement, based on the aforementioned transaction structure, blockchain smart contracts are used to automate the execution of transaction terms, fund custody and transfer, and ownership change registration, resulting in a final traceable record of distressed asset transactions.

[0010] Optionally, the multi-source heterogeneous enterprise data received from distressed enterprises is processed through cross-domain privacy-preserving fusion based on a federated learning framework to obtain a structured and de-identified panoramic data profile of distressed enterprises, including:

[0011] Based on the input multi-source heterogeneous enterprise data, data cleaning and format standardization processes are performed to obtain a standardized data stream;

[0012] Based on the normalized data stream, Gaussian noise is added using the local differential privacy mechanism in the federated learning framework to generate privacy-preserving data fragments;

[0013] Based on privacy-preserving data fragments, heterogeneous data fusion processing is performed using a cross-domain feature alignment algorithm to obtain a fused feature matrix;

[0014] Based on the fused feature matrix, a variational autoencoder is applied for dimensionality reduction and structural reconstruction to generate a structured and desensitized panoramic data profile of distressed enterprises.

[0015] Optionally, the process of creating a comprehensive data profile of the distressed enterprise utilizes a pre-trained deep adversarial generative network to extract asset risk features and model value influencing factors, resulting in a multi-dimensional dynamic risk feature vector that includes asset status, potential liabilities, operational interruption risk, and industry prosperity correlation, including:

[0016] Based on the panoramic data profile of distressed enterprises, the data is input into the generator of a pre-trained Generative Adversarial Network (GAN) for feature generation processing to obtain a preliminary risk feature set.

[0017] Based on the initial risk feature set, the GAN discriminator performs adversarial refinement to obtain a high-confidence risk feature vector.

[0018] Based on the high-confidence risk feature vector, the influencing factors are modeled using a factor analysis model combined with the industry prosperity index to obtain the value influencing factor weight vector.

[0019] Based on the value impact factor weight vector, dynamic feature synthesis and dimensional expansion processing are performed to generate a multi-dimensional dynamic risk feature vector that includes asset status, potential liabilities, operational interruption risk, and industry prosperity correlation.

[0020] Optionally, the step of associating the dynamic risk feature vector with real-time market transaction data and an industry M&A case library, and performing dynamic discount rate calculation and optimal asset pricing range deduction based on reinforcement learning algorithms to obtain a dynamic discount price range and confidence assessment for a specific distressed asset, includes:

[0021] Based on the dynamic risk feature vector and real-time market transaction data, feature enhancement and correlation processing are performed to obtain the spatiotemporal enhanced feature vector;

[0022] Based on the spatiotemporal enhanced feature vectors and the industry M&A case library, case matching is performed using the cosine similarity algorithm to obtain a set of highly similar cases;

[0023] Based on a set of highly similar cases, a reinforcement learning agent is applied to simulate and extrapolate the discount rate, resulting in a preliminary dynamic discount rate strategy.

[0024] Based on the initial dynamic discount rate strategy, the pricing range is optimized using the Monte Carlo tree search algorithm to obtain the pricing range of candidate assets.

[0025] Based on the pricing range of candidate assets, the probability distribution and error range are calculated using a Bayesian confidence model to generate a dynamic discount price range and confidence assessment for specific distressed assets.

[0026] Optionally, based on the preset preferences of potential investors and the dynamic discount price range, the process employs a multi-agent simulation negotiation and conflict resolution algorithm for intelligent and precise matching and transaction feasibility simulation to obtain an optimal shortlist of potential investors and a preliminary conflict-free transaction structure scheme, including:

[0027] Based on the preset preferences of potential investors, preference vectorization and normalization are performed to obtain investor feature vectors;

[0028] Based on investor feature vectors and dynamic discount price ranges, a multi-agent system is created using an agent initialization algorithm to obtain a set of investor agents;

[0029] Based on the set of investor agents, a Nash negotiation game model is applied to conduct transaction simulation negotiation to obtain a preliminary set of transaction proposals.

[0030] Based on the preliminary transaction proposal set, conflict points are identified using conflict detection algorithms, and a conflict report is generated.

[0031] Based on the conflict report, a genetic algorithm was used for multi-objective optimization and resolution to generate an optimal shortlist of potential investors and a preliminary conflict-free transaction structure plan.

[0032] Optionally, for the agreed-upon transaction, based on the transaction structure scheme, blockchain smart contracts are used to automate the execution of transaction terms, fund custody and transfer, and ownership change registration, resulting in a final traceable distressed asset transaction record, including:

[0033] Based on the transaction structure, the terms are digitized and smart contract coding is performed to obtain deployable smart contract code;

[0034] Based on the deployable smart contract code, automatic terms verification is performed through the blockchain network to generate transaction execution event logs;

[0035] Based on the transaction execution event log, a zero-knowledge proof mechanism is applied to process fund custody and automatic transfer, resulting in a secure fund transfer record.

[0036] Based on the secure fund transfer records, the ownership change registration and tamper-proof storage are performed through the hash chain verification protocol to generate the final traceable distressed asset transaction records.

[0037] Another embodiment of this application provides an AI-based data information and asset transaction service system for distressed enterprises, the system comprising:

[0038] The receiving module is used to receive multi-source heterogeneous enterprise data of distressed enterprises, and perform cross-domain privacy-preserving fusion processing based on the federated learning framework to obtain a structured and de-identified panoramic data profile of distressed enterprises.

[0039] The extraction module is used to create a panoramic data profile of the distressed enterprise, and to use a pre-trained deep adversarial generative network to extract asset risk features and model value influencing factors, so as to obtain a multi-dimensional dynamic risk feature vector that includes asset status, potential liabilities, operational interruption risk, and industry prosperity correlation.

[0040] The association module is used to associate the dynamic risk feature vector with real-time market transaction data and industry M&A case library, and to perform dynamic discount rate calculation and optimal asset pricing range deduction based on reinforcement learning algorithm to obtain dynamic discount price range and confidence assessment for specific distressed assets.

[0041] The matching module is used to perform intelligent and accurate matching and transaction feasibility simulation processing based on the preset preferences of potential investors and the dynamic discount price range, using a multi-agent simulation negotiation and conflict resolution algorithm, to obtain an optimal shortlist of potential investors and a preliminary conflict-free transaction structure scheme.

[0042] The execution module is used to automatically execute transaction terms, transfer funds in custody, and register changes in ownership for transactions that have reached an agreement, based on the transaction structure scheme, using blockchain smart contracts, to obtain a final traceable record of distressed asset transactions.

[0043] 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.

[0044] 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.

[0045] Compared with existing technologies, this invention provides an AI-based method for data information and asset transaction services for distressed enterprises. It receives multi-source, heterogeneous enterprise data from distressed enterprises to obtain a structured, anonymized panoramic data profile of the distressed enterprise. From this panoramic data profile, it obtains a multi-dimensional dynamic risk feature vector including asset status, potential liabilities, operational interruption risks, and industry prosperity correlation. This dynamic risk feature vector is then correlated with real-time market transaction data and an industry M&A case database to obtain a dynamic discount price range and confidence level assessment for specific distressed assets. Based on the preset preferences of potential investors and the dynamic discount price range, it obtains an optimal shortlist of potential investors and a preliminary conflict-free transaction structure. For transactions with agreed intentions, it obtains a final traceable distressed asset transaction record, thereby dynamically optimizing asset pricing and transaction matching, and improving the efficiency and security of distressed asset transactions. Attached Figure Description

[0046] Figure 1 A hardware structure block diagram of a computer terminal for an AI-based method of providing data information and asset trading services for distressed enterprises, provided in an embodiment of the present invention;

[0047] Figure 2 A flowchart illustrating an AI-based method for providing data information and asset transaction services to distressed enterprises, as provided in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the structure of an AI-based data information and asset transaction service system for distressed enterprises, provided as an embodiment of the present invention. Detailed Implementation

[0049] 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.

[0050] The present invention first provides an AI-based method for providing data information and asset trading services for distressed enterprises. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0051] 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 an AI-based method of providing data information and asset transaction services for distressed enterprises, as provided in an embodiment of the present invention. Figure 1As 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.

[0052] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause a processor to perform any AI-based method for providing data and asset trading services to distressed enterprises.

[0053] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0054] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When executed by a processor, the computer program enables the processor to perform any AI-based data information and asset transaction service for distressed enterprises.

[0055] 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.

[0056] 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.

[0057] See Figure 2 The present invention provides an AI-based method for providing data information and asset trading services for distressed enterprises, which may include the following steps:

[0058] S201 receives multi-source heterogeneous enterprise data of distressed enterprises, performs cross-domain privacy-preserving fusion processing based on the federated learning framework, and obtains a structured and de-identified panoramic data profile of distressed enterprises.

[0059] Specifically, based on the input multi-source heterogeneous enterprise data, data cleaning and format standardization processing can be performed to obtain a standardized data stream;

[0060] The system receives raw data from distressed companies from multiple sources, including structured data (such as balance sheets, profit and loss statements, and cash flow statements from financial systems), semi-structured data (such as supply chain order logs in XML format and IoT device operation reports in JSON format), and unstructured data (such as supplier payment reminders in PDF format and production line monitoring videos). The data cleaning pipeline is started first:

[0061] Missing value handling: For the missing "Accounts Receivable Turnover" field in the financial statements, Time Series Interpolation (TSI) is used to fill the missing value by weighting data from adjacent quarters (weight factor W=0.7 indicates that recent data has higher weight).

[0062] Outlier correction: Outliers are identified using the Boxplot Rule (BR). For example, if "Management Expenses" suddenly increases by 300% in a certain month, the system combines the invoice information recognized by OCR to confirm whether it is an input error, and calls the Isolation Forest Algorithm (IFA) to automatically mark suspicious data.

[0063] Format standardization: Unstructured text (such as court judgments) is processed by the BERT entity recognition model (Bidirectional Encoder Representations from Transformers Entity Recognition, BERT-ER) to extract key fields (amount involved, judgment result) and convert them into unified key-value pairs (KVP); video data is processed by a 3D Convolutional Neural Network (3D-CNN) to extract equipment downtime frequency indicators, and finally all data is output in Apache Parquet columnar storage format (a high-performance data storage format).

[0064] Addressing the semantic conflict issue of multi-source data:

[0065] The different definitions of "liabilities" in different systems (e.g., system A includes contingent liabilities, while system B does not) are aligned to a unified Enterprise Risk Ontology (ERO) through the Ontology Mapping Engine (OME).

[0066] For data with inconsistent time granularity (e.g., financial data by quarter, production data by minute), a Dynamic Time Warping (DTW) algorithm is used to aggregate high-frequency data into a quarterly average. A Timestamp Alignment Service (TAS) ensures that all data is anchored to the same baseline time axis. This ultimately generates a Normalized Data Stream (NDS) with a Data Lineage Tag (DLT), whose Data Schema (DS) contains 128 standardized fields (such as Asset_Liquidity_Ratio and Pending_Litigation_Count).

[0067] To ensure data quality:

[0068] Deploy a Data Quality Probe (DQP) to calculate the field's Completeness Score (CS, ≥95%) and Consistency Score (CnS, field logical contradiction rate <1%) in real time.

[0069] For geographically dispersed data sources (such as data from overseas subsidiaries), a Sharded Cleaning Strategy (SCS) is employed. After local cleaning is performed on edge nodes (EN), the data is transmitted to the central node (CN) via the gRPC protocol (Google Remote Procedure Call, a high-performance remote call protocol) for federation aggregation (FA). The cleaned NDS is then pushed to the next module in real time using an Apache Kafka message queue (a distributed stream processing platform).

[0070] Based on the normalized data stream, Gaussian noise is added using the local differential privacy mechanism in the federated learning framework to generate privacy-preserving data fragments;

[0071] Normalized data streams (NDS) enter the Federated Learning System (FLS), which employs a star topology (ST): a central server (CS) coordinates multiple data participants (DPs, such as banks and suppliers). Each DP executes locally:

[0072] Data slicing: The NDS is divided into enterprise data fragments (EDF) by entity, such as the financial fragment of a company for the past 3 years, EDF_CompanyX_Financial.

[0073] Noise injection: For sensitive numerical fields in the EDF (such as "net loss"), a Local Differential Privacy (LDP) mechanism is applied. Specifically, a Gaussian Noise Mechanism (GNM) is used, adding noise according to the formula Noise = N(0, σ²), where the noise standard deviation σ is controlled by the privacy budget ε (PBε) (when ε=0.5, σ=Δf / ε, Δf=1 million is the global sensitivity of the field).

[0074] Key steps to enhance privacy:

[0075] Dynamic Noise Scaling (DNS): For high-variance fields (such as "monthly revenue volatility"), the Adaptive Standard Deviation Algorithm (ASDA) is used to dynamically adjust σ (adjustment factor λ=1.2) based on the historical distribution of the field.

[0076] Sparse data protection: For low-frequency category fields (such as "litigation type"), a randomized response mechanism (RRM) is used to retain the true value with a probability p=1 / (1+e^ε) (p≈73% when ε=1). This ultimately generates a noisy privacy-protected data fragment (PPDF) whose privacy level satisfies ε-differential privacy (ε-DP).

[0077] Secure transmission and authentication:

[0078] The PPDF is encrypted with AES-256 (Advanced Encryption Standard 256-bit) and then uploaded to CS via a TLS 1.3 channel (Transport Layer Security version 1.3).

[0079] CS runs a Privacy Leakage Audit (PLA): The Membership Inference Attack Simulator (MIAS) is used to test the PPDF's resistance to attacks (attack accuracy must be <55%). PPDFs that pass the audit are marked as Aggregable State (AS) and stored in a Distributed Hash Table (DHT).

[0080] Based on privacy-preserving data fragments, heterogeneous data fusion processing is performed using a cross-domain feature alignment algorithm to obtain a fused feature matrix;

[0081] The central server (CS) performs cross-domain feature alignment (CFA) on the received multi-party PPDF:

[0082] Feature space mapping: Heterogeneous features are projected onto a unified space using a shared embedding layer (SEL). For example:

[0083] The "number of credit defaults" (numerical) provided by the bank is mapped to dimension D1;

[0084] The "type of litigation" (categorical type) provided by the court is mapped to dimension D2 via word embedding (WE);

[0085] The "payment delay days" (time series) provided by the supplier is mapped to dimension D3 by an LSTM encoder (Long Short-Term Memory Encoder, LSTME).

[0086] Relevance alignment: The difference between feature distributions from different sources (such as the MMD distance between bank data distribution P_bank and court data distribution P_court) is calculated using the Maximum Mean Discrepancy Minimization (MMDM) algorithm, and the embedding parameters are optimized to reduce the MMD value (target threshold MMD < 0.05).

[0087] Key fusion technologies:

[0088] Graph Structure Alignment (GSA): Constructs an Enterprise Association Graph (EAG), where nodes represent enterprise entities and edges represent transaction / litigation relationships. For subgraphs provided by multiple parties, a Graph Convolutional Network (GCN) is used to align node representations (NR) and merge them into a unified graph.

[0089] Time-dimensional fusion: For asynchronously updated data (such as quarterly financial data and monthly operating data), a neural process model (NPM) is applied to learn the latent time function and generate synchronized time-sliced ​​features (TSF).

[0090] Output the Fused Feature Matrix (FFM): The matrix has an N×M dimension (N = number of enterprises, M = fused feature dimension), and each element FFM[i,j] represents the j-th fused feature value of enterprise i.

[0091] Feature Importance Screening (FIS): The Random Forest Algorithm (RFA) is used to calculate the feature importance score (IS), retaining the TOP-K features (K=50, score threshold IS>0.01).

[0092] Finally, the FFM is stored in an OLAP cube (Online Analytical Processing Cube, a multidimensional data analysis storage structure) to support fast feature retrieval.

[0093] Based on the fused feature matrix, a variational autoencoder is applied for dimensionality reduction and structural reconstruction to generate a structured and desensitized panoramic data profile of distressed enterprises.

[0094] The fusion feature matrix (FFM) is input to the variational autoencoder (VAE) for dimensionality reduction.

[0095] Encoding stage: The encoder (ENC) is a 3-layer fully connected network (number of neurons: 256-128-64) that compresses high-dimensional features (e.g., M=100 dimensions) into a low-dimensional vector z (dimension L=16) in the latent space (LS).

[0096] Learn the latent distribution parameters: mean μ and variance σ², and sample z = μ + σ ⊙ ε (ε ∼ N(0,1), ⊙ represents element-wise multiplication) using the reparameterization trick (RT).

[0097] Bottleneck constraint: Apply Kullback-Leibler Divergence Regularization (KLDR) to the latent space to force z to approximate the standard normal distribution (target KL value < 0.1) and enhance generalization.

[0098] Structured reconstruction process:

[0099] Decoding Phase: The decoder (DEC) reconstructs z into an output with the same dimension as the original FFM. A Conditional Generation Architecture (CGA) is employed, injecting industry type tags (e.g., manufacturing tag SIC=35) to guide the reconstruction.

[0100] Feature decoupling: By using a β-VAE variant (β-VAE Variant, βV), setting the decoupling coefficient β=3.5, latent factors (such as...) are separated. Controlling debt repayment ability (Control operational risks).

[0101] Reconstruction loss optimization: The Evidence Lower Bound Loss (ELBO) function is used, which includes reconstruction error (mean squared error MSE) and KL divergence. The optimizer is AdamW (Adam with Weight Decay, a gradient optimization algorithm), with a learning rate lr=0.001.

[0102] Generate a Panoramic Data Portrait (PDP):

[0103] PDP is a structured table (ST) that includes core risk indicators (CRI), such as:

[0104] Asset_Depreciation_Rate (Asset depreciation rate, range 0~1);

[0105] Operation_Interruption_Risk_Score (Operational interruption risk score, ranging from 1 to 10).

[0106] Anonymization guarantee: The profile does not contain original sensitive fields (such as company name, specific amount) and has undergone k-anonymity check (kAC, k≥5) to ensure that no record can be uniquely identified.

[0107] The profiles are stored in a graph database (GDB), with enterprises as nodes and metrics as attributes, supporting queries of complex risk relationships.

[0108] This method first integrates enterprise data from various sources, including financial, operational, and market data, using federated learning techniques, and then processes this data while protecting data privacy. Local differential privacy mechanisms are used to add noise to ensure the security of sensitive information. Simultaneously, feature alignment algorithms are employed to unify data of different formats and standards into a structured format, ultimately forming a comprehensive data profile reflecting the enterprise's situation. This solves the data silo problem in traditional methods, effectively integrating multi-dimensional data while protecting corporate trade secrets and personal privacy. This approach not only meets compliance requirements but also provides a high-quality data foundation for subsequent analysis, making it particularly suitable for handling data from distressed enterprises involving sensitive information.

[0109] S202, for the panoramic data portrait of the distressed enterprise, use a pre-trained deep adversarial generative network to extract asset risk features and model value influencing factors, and obtain a multi-dimensional dynamic risk feature vector including asset status, potential liabilities, operational interruption risk, and industry prosperity correlation.

[0110] Specifically, based on the panoramic data profile of distressed enterprises, the data can be input into the generator of a pre-trained Generative Adversarial Network (GAN) for feature generation processing to obtain a preliminary risk feature set.

[0111] When the system receives a comprehensive data profile of distressed enterprises (including structured, anonymized data on finance, operations, and supply chain) generated through federated learning, it first inputs it into the generator of a pre-trained Generative Adversarial Network (GAN). This generator is a deep neural network (such as a multi-layered convolutional neural network CNN or Transformer architecture) that has been pre-trained on massive corporate bankruptcy and restructuring datasets. The core function of the generator is to simulate the risk characteristic distribution patterns of distressed enterprises. For example, when given a comprehensive profile of a manufacturing enterprise (including parameters such as a debt-to-equity ratio of 65%, accounts receivable turnover days of 120 days, and equipment condition rate of 45%), the generator will capture the complex relationships between parameters through nonlinear transformations of its hidden layers (such as the combination of a high debt ratio and a low equipment condition rate potentially indicating debt repayment risk), and output a preliminary risk feature set (PRFS). PRFS is essentially a vector representation in a high-dimensional feature space, containing implicit risk indicators not explicitly present in the original data, such as derived features like "probability of cash flow disruption" and "collateralization ratio deviation." The generator's network structure typically includes encoder and decoder modules: the encoder compresses the input profile into a low-dimensional latent vector, and the decoder reconstructs this vector into risk features. During training, feature fidelity is ensured by minimizing the reconstruction loss (e.g., mean squared error, MSE).

[0112] The key to the pre-training process lies in the adversarial training mechanism. The generator (denoted as G) needs to be trained alternately with the discriminator (D). D is designed as a binary classifier (such as a fully connected neural network) with the goal of distinguishing the PRFS output by the generator from the expert-annotated risk features of real-world distressed companies (denoted as Real_Risk). In the early stages of training, the PRFS generated by G may be of low quality (e.g., it may not accurately reflect the risk of operational disruption), and D can easily identify the difference between it and Real_Risk. The system adjusts the network weights and biases of G through the backpropagation algorithm, for example, by using the Adam optimizer (Adaptive Moment Estimation) to update the parameters with a learning rate (LR) of 0.001, forcing G to generate features that are closer to Real_Risk. At the same time, D also updates its parameters to improve its discriminative ability. This game continues iteratively (e.g., 10,000 epochs) until the PRFS generated by G can fool D with a high probability (i.e., D's identification accuracy is close to 50%). At this point, G is considered to have grasped the essential pattern of distressed firm risk. After pre-training, G is frozen, and its parameters are no longer updated; it is only used as a feature extractor.

[0113] During the deployment phase, after the panoramic data profile is input into the frozen generator, its output layer generates a fixed-dimensional PRFS. For example, if the input is a 1,000-dimensional enterprise profile vector, the generator outputs a 256-dimensional PRFS, where each dimension corresponds to an abstract risk attribute. Suppose that in a retail enterprise's PRFS, the 48th dimension has a feature value of 0.87 (close to 1 indicating high risk), which is interpreted as "liquidity risk caused by unsold inventory"; the 129th dimension has a feature value of 0.12 (close to 0 indicating low risk), which is mapped to "stability of the core business district." Although these features are not directly present in the input data, they are effectively derived through the generator's learning of historical patterns. The PRFS also needs to undergo standardization (such as Z-score normalization) to ensure that all feature values ​​are of similar magnitude (e.g., a mean of 0 and a standard deviation of 1), avoiding bias in subsequent calculations due to scale differences. The final output PRFS will serve as the input for the next stage of adversarial refinement.

[0114] Based on the initial risk feature set, the GAN discriminator performs adversarial refinement to obtain a high-confidence risk feature vector.

[0115] The preliminary risk feature set (PRFS) generated in the previous step is input into the discriminator (D) of the same GAN. The discriminator, pre-trained, already possesses the ability to distinguish between real risk features (Real_Risk) and generated features (PRFS). During the refinement phase, the discriminator no longer participates in parameter updates but operates as a "feature quality evaluator." Its last layer (usually a sigmoid activation function) outputs a confidence score (CS), ranging from 0 to 1: a score close to 1 indicates that the PRFS is close to real historical risk features, while a score close to 0 indicates a significant difference from the real pattern. For example, if a PRFS scores 0.92 after being input into D, it indicates high confidence; if another PRFS scores only 0.31, it suggests potential noise or distortion. The system sets a confidence threshold (CT, e.g., 0.7). When CS ≥ CT, the PRFS is directly retained; if CS < CT, a feature correction mechanism is triggered.

[0116] The core of adversarial refinement is the iterative correction of low-confidence features. When the CS of the PRFS is less than the CT, the system initiates a gradient feedback loop: the discriminator calculates the gradient of the PRFS relative to its output, which indicates which dimensions in the PRFS need to be adjusted in the Real_Risk direction to improve the CS. For example, the gradient shows that the 48th dimension needs to be increased by 0.15 and the 129th dimension needs to be decreased by 0.08. The generator (at this point, only its decoder part is enabled) receives this gradient signal and fine-tunes the PRFS through one forward-backward pass. The fine-tuning magnitude is controlled by the learning rate decay factor (LRD, e.g., 0.1) to avoid over-correction. The corrected PRFS is then fed back into the discriminator to evaluate the CS until the CS is greater than or equal to the CT or the maximum number of iterations (MI, e.g., 5 times) is reached. The final output is a high-confidence risk feature vector (HCRFV) that has passed confidence checks in all dimensions.

[0117] The refinement process also involves feature importance screening. The activation map of the discriminator's intermediate layer reflects the weight of different feature dimensions on the final confidence score. The system uses Class Activation Mapping (CAM) technology to identify the key dimensions (such as features with scalar values ​​exceeding 0.8) that contribute the most to the discrimination result in the HCRFV. For example, in the HCRFV of an energy company, the CAM weight of the "equipment maintenance delay index" dimension is 0.91, while the weight of "administrative cost ratio" is only 0.05. The former is then marked as a key risk factor. The system finally outputs an HCRFV that has undergone confidence verification and importance weighting. Its number of dimensions may be the same as PRFS (e.g., 256 dimensions), but the confidence of each dimension is significantly improved, providing reliable input for subsequent value modeling.

[0118] Based on the high-confidence risk feature vector, the influencing factors are modeled using a factor analysis model combined with the industry prosperity index to obtain the value influencing factor weight vector.

[0119] The High Confidence Risk Feature Vector (HCRFV) is input into the Factor Analysis Model (FAM). The goal of FAM is to extract a few common factors from the numerous dimensions of the HCRFV that can explain most of the risk variation. The model uses Principal Component Analysis (PCA): first, the covariance matrix of the HCRFV is calculated, and then its eigenvalues ​​and eigenvectors are solved using Singular Value Decomposition (SVD). The principal components (PCs) corresponding to the first k eigenvectors (e.g., k=5) are selected as common factors, and their cumulative variance explained (CVE) must exceed a preset threshold (e.g., 85%). For example, a certain HCRAV contains 256 dimensions, and 5 PCs are extracted by PCA. Their meanings can be interpreted as: "solvency factor" (explaining 40% of variance), "operational resilience factor" (explaining 25% of variance), "asset quality factor" (explaining 15% of variance), "industry correlation factor" (explaining 10% of variance), and "legal risk factor" (explaining 5% of variance).

[0120] Public factors need to be dynamically integrated with the Industry Prosperity Index (IPI). IPI is an externally imported real-time indicator (e.g., from an industry association database) that reflects the overall trend of the target company's industry (e.g., 0-100, >80 indicating high prosperity). The system establishes a Multiple Linear Regression (MLR) equation for each public factor, with IPI as the dependent variable and factor scores as independent variables. Regression coefficients characterize the sensitivity of factors to industry prosperity. For example, the coefficient for the "Industry-Related Factor" is 0.75 (strong positive correlation), and the coefficient for the "Legal Risk Factor" is -0.6 (negative correlation). Simultaneously, the factor loading matrix (FLM) is calculated between the factor scores and the original dimensions of the HCRFV. The loading values ​​(ranging from -1 to 1) represent the degree of contribution of the original features to the public factors (e.g., the loading of "Patent Quantity" on the "Industry-Related Factor" is 0.82).

[0121] The Value Impact Factor Weight Vector (VIFWV) is generated in two steps:

[0122] Factor importance weighting: The weight (Weight_F) of each public factor is obtained by normalizing the product of its variance explained rate (e.g., 40% for the "solvency factor") and the absolute value of the industry sensitivity coefficient (e.g., 0.75).

[0123] Original feature weight backtracking: Weight_F is reassigned back to the original dimensions of HCRFV based on FLM. For example, the final weight of "Patent Quantity" = Weight_F (Industry Relevance Factor) × Loading value (0.82).

[0124] Ultimately, VIFWV is a vector with the same dimensions as HCRFV (e.g., 256 dimensions), and its element values ​​(e.g., 0.05, 0.12, etc.) represent the relative impact of the corresponding risk characteristics on the company's asset value. The sum of all elements is 1.

[0125] Based on the value impact factor weight vector, dynamic feature synthesis and dimensional expansion processing are performed to generate a multi-dimensional dynamic risk feature vector that includes asset status, potential liabilities, operational interruption risk, and industry prosperity correlation.

[0126] The Value Impact Factor Weight Vector (VIFWV) is used to guide feature synthesis. The system first performs a weighted fusion of the High Confidence Risk Feature Vector (HCRFV): each dimension of the HCRFV is multiplied by its corresponding weight in the VIFWV, and then the weighted features are summed to generate the basic components of the four core dimensions.

[0127] Asset Status (AS): A weighted sum of features such as the newness rate of integrated equipment and the validity of patents.

[0128] Potential Liability (PL): A weighted sum of features including implicit guarantees and pending litigation.

[0129] Operation Disruption Risk (ODR): A weighted sum of characteristics including supply chain dependence and key personnel turnover rate.

[0130] Industry Relevance (IR): a weighted sum of characteristics including market share fluctuations and industry technology substitution rates.

[0131] Each component is initially set to a scalar value (e.g., AS = 0.65, range 0-1).

[0132] Dynamic injection is achieved by expanding dimensions through real-time data:

[0133] Time series analysis: Access the company's AS / PL / ODR / IR sub-values ​​for the past 12 months, and calculate the trend slope (TS) using an exponential smoothing algorithm (ES, smoothing coefficient α=0.3). For example, a TS=0.05 for AS indicates that the asset status is slowly improving.

[0134] Industry fluctuation coupling: The correlation strength between the IR sub-item and the real-time industry prosperity index (IPI) is calculated through cointegration test (e.g., cointegration coefficient β=0.8), generating the "industry sensitivity" sub-dimension.

[0135] Emergency Correction: If the external system detects a sudden policy change (such as a new environmental protection regulation), the Event Impact Model (EIM) is activated, and an adjustment factor (AF, such as -0.1) is output to dynamically deduct relevant sub-items.

[0136] The final Dynamic Risk Feature Vector (DRFV) is a 12-dimensional vector:

[0137] The first four dimensions: the current values ​​(0-1) of AS, PL, ODR, and IR.

[0138] 4-dimensional: The trend slope TS of each component (-1 to 1).

[0139] The last four dimensions are dynamically expanded dimensions, including "Industry Sensitivity" (cointegration coefficient of IR and IPI), "Policy Vulnerability" (EIM output value), "Supply Chain Resilience" (ODR sub-feature calculation), and "Asset Realization Premium" (AS sub-feature calculation).

[0140] All dimensions are normalized to the range [0,1] using Min-Max Scaling. For example, the DRFV of a manufacturing company is represented as: [0.72, 0.85, 0.63, 0.58; 0.05, -0.12, 0.08, 0.03; 0.82, 0.15, 0.45, 0.67], which fully includes the four core dimensions and their dynamic correlation characteristics claimed in the claims.

[0141] By employing generative adversarial networks (GANs) to conduct in-depth analysis of a company's comprehensive data, and through adversarial training between the generator and discriminator, the core risk factors affecting company value are accurately identified. The network focuses on key dimensions such as asset quality, hidden debt, going concern capability, and industry environment, quantifying these factors into computable feature vectors. This approach overcomes the limitations of traditional risk assessment methods, enabling the discovery of nonlinear relationships and potential risks hidden within complex data. This dynamic feature extraction method can more accurately assess the true value of a company, providing a scientific basis for subsequent pricing and transactions.

[0142] S203, the dynamic risk feature vector is associated with real-time market transaction data and industry M&A case library, and dynamic discount rate calculation and optimal asset pricing range deduction are performed based on reinforcement learning algorithm to obtain dynamic discount price range and confidence assessment for specific distressed assets.

[0143] Specifically, feature enhancement and correlation processing can be performed based on dynamic risk feature vectors and real-time market transaction data to obtain spatiotemporally enhanced feature vectors;

[0144] Spatiotemporal alignment of dynamic risk feature vectors

[0145] The Dynamic Risk Feature Vector (DRFV) contains data on asset status, potential liabilities, and other dimensions, but lacks real-time market context. The system first accesses the Real-time Market Transaction Data Stream (RMTDS), including high-frequency information such as stock price anomalies (e.g., a sector's daily price fluctuation exceeding 5%), commodity futures prices (e.g., the price of the main copper futures contract), and bond default events (e.g., a corporate bond's yield surging by 200 basis points in a single day). A Sliding Time Window Matching Algorithm (STWMA) aligns the DRFV timestamps with the RMTDS: for example, using a 15-minute window (WindowSize, WS=15min), all market events within the window are aggregated into a Market Sentiment Index (MSI, ranging from 0-100). At the same time, the geospatial encoder (GSE) is used to convert the company's registered location and the location of its main assets into latitude and longitude grid coordinates (such as the GCJ-02 coordinate system) and spatially correlate them with regional policy events (such as sudden environmental protection production restrictions in a certain area).

[0146] Multimodal feature fusion enhancement

[0147] The aligned data is input into the Cross-modal Feature Fusion Layer (CMFFL). This layer contains two processing units:

[0148] The Temporal Feature Enhancement Unit (TFEU) uses a Long Short-Term Memory (LSTM) network (a recurrent neural network that captures long-term dependencies) to analyze the correlation between operational disruption risk indicators (such as negative cash flow for six consecutive months) and the MSI in the DRFV (Disruption Risk Value). For example, when the MSI is below 30 (during a market panic) and the company's cash flow risk value is above 0.8, the LSTM outputs an amplification factor (AF=1.5) for this risk.

[0149] Spatial Feature Association Unit (SFAU): Based on geographic grid coordinates, this unit calculates the Spatial Impact Decay Weight (SIDW) of a company's assets and regional events. For example, if a chemical company is 50 kilometers from the boundary of an environmentally restricted production zone, the SIDW calculated using the Gaussian Decay Model is 0.7, indicating that the impact of this event on the company is 70% of that in the core restricted production zone.

[0150] Generation of spatiotemporal augmentation vectors

[0151] The fused features undergo feature concatenation and standardization: the original DRFV (e.g., 128 dimensions), the temporal enhancement features output by TFEU (e.g., 8 dimensions), and the spatial correlation features output by SFAU (e.g., 4 dimensions) are concatenated into a 140-dimensional vector. Then, Z-score normalization (making the mean of each dimension 0 and the variance 1) is used to eliminate dimensional differences. The final result is a Spatio-temporal Enhanced Feature Vector (STEFV), which has the same dimensions as the original DRFV but includes market spatio-temporal correlation information. For example, the 37th dimension in the vector represents the "equity valuation depreciation rate under regional policy shocks."

[0152] Based on the spatiotemporal enhanced feature vectors and the industry M&A case library, case matching is performed using the cosine similarity algorithm to obtain a set of highly similar cases;

[0153] Vectorized representation of industry M&A case library

[0154] The Industry M&A Case Database (IMCD) stores historical transaction records. Each case includes structured fields such as target asset attributes (e.g., equipment condition rate of 60%) and transaction environment parameters (e.g., the industry PMI index at the time was 45.2). A pre-trained Feature Embedding Model (FEM) transforms each case into a feature vector.

[0155] Numerical fields (such as debt ratio) are directly normalized.

[0156] Categorical fields (such as industry classifications) are mapped to dense vectors using word embedding techniques (such as Word2Vec).

[0157] Text descriptions (such as transaction announcements) extract semantic features using the BERT model.

[0158] Each case is ultimately represented as a 300-dimensional Case Feature Vector (CFV), and the cases in the library are indexed by industry-time partition (e.g., "Manufacturing - 2020 to 2023").

[0159] Cosine similarity calculation and case selection

[0160] Match the STEFV of the asset to be priced with the CFV of the same industry segment in IMCD:

[0161] The Cosine Similarity Algorithm (CSA) calculates the cosine of the angle between two vectors, with the result ranging from -1 to 1. The larger the value, the more similar the vectors are.

[0162] Set a dynamic similarity threshold (DST): automatically adjust the weight based on the timeliness of the case, for example, a threshold of 0.85 for cases within 3 years and a threshold of 0.92 for cases older than 5 years.

[0163] Approximate Nearest Neighbor (ANN) search is used to accelerate matching. For example, the Hierarchical Navigable Small World (HNSW) algorithm can return the Top 50 candidate cases in milliseconds.

[0164] Generation of high similarity case sets

[0165] Perform multidimensional consistency checks on the initially matched cases:

[0166] Consistency of asset attributes: Compare the types of assets (such as patents / factory buildings / land), and exclude those with significant differences.

[0167] Consistency of transaction terms: Check whether the payment method (cash / equity swap), transaction period (e.g., settlement within 90 days) are comparable.

[0168] Environmental disturbance filtering: Excluding cases from special periods.

[0169] Finally, cases with a similarity score > 0.9 and that pass the verification are retained to form a High-Similarity Case Set (HSCS), such as 12 manufacturing equipment transaction cases, each labeled with a similarity score and key differences.

[0170] Based on a set of highly similar cases, a reinforcement learning agent is applied to simulate and extrapolate the discount rate, resulting in a preliminary dynamic discount rate strategy.

[0171] Reinforcement learning environment and agent design

[0172] The core elements of building a Virtual Trading Simulation Environment (VTSE) include:

[0173] State Space (SS): contains a compressed representation of STEFV (e.g., 20-dimensional), the average discount rate of cases in HSCS (e.g., benchmark value of 15%), and real-time buyer inquiry data (e.g., a fund's latest offer is 65% of the appraised value).

[0174] Action Space (AS): Discount rate adjustment step size, defined as a discrete action set {-5%, -1%, +1%, +5%}.

[0175] Reward Function (RF): R = α × probability of successful transaction + β × premium income - γ × penalty for unsold item

[0176] The weight parameters are set according to the asset type. For example, for equipment assets, α=0.6 (emphasizing transaction speed), β=0.3 (focusing on returns), and γ=0.1 (controlling the risk of failed auctions).

[0177] Q-learning-based policy deduction

[0178] The Reinforcement Learning Agent (RLA) employs a Double Deep Q-Network (DDQN) architecture:

[0179] The main network takes the current state S_t as input (e.g., asset risk level B+, market liquidity tight) and outputs the Q value (action value estimate) for each action.

[0180] The target network periodically synchronizes the parameters of the main network to calculate the target Q-value and reduce overestimation.

[0181] The Experience Replay mechanism stores state transition records (S_t, A_t, R_t, S_{t+1}), and random sampling training breaks the correlation of data.

[0182] The agent simulates 1000 trading scenarios in VTSE: starting with an initial discount rate of 10%, it selects price adjustment actions according to the ε-Greedy Policy (ε=0.2, i.e., 20% probability of random exploration), and updates the Q network parameters at each step.

[0183] Dynamic discount rate strategy generation

[0184] After training, the agent outputs the Optimal Action Sequence (OAS) for the current state:

[0185] Short-term strategy: Initial offer discount rate of 18% (3% higher than benchmark to attract buyers).

[0186] Mid-term strategy: If there is no substantial inquiry within 30 days, the action limit will be reduced from -5% to 13%.

[0187] Exit Mechanism: If there is still no intention to sell after two consecutive price adjustments, the auction failure protection will be triggered (the γ term in the reward function will take effect).

[0188] Finally, a preliminary dynamic discount strategy (PDDS) is generated, which includes a phased discount rate adjustment curve and the corresponding expected transaction probability matrix.

[0189] Based on the initial dynamic discount rate strategy, the pricing range is optimized using the Monte Carlo tree search algorithm to obtain the pricing range of candidate assets.

[0190] Game tree construction for Monte Carlo tree search

[0191] The Pricing Decision Tree (PDT) unfolds using the PDDS as the initial strategy:

[0192] Root Node: Current asset status (e.g., appraised value of 10 million yuan, initial discount rate of 18%).

[0193] Branch Actions: Expanded into three types of decisions:

[0194] Aggressive strategy: Reduce the price to 70% of the appraised value in one go (30% discount rate);

[0195] Conservative strategy: Maintain PDDS' tiered price reduction (18%→13%→10%).

[0196] Compromise strategy: 85% of the initial offer (15% discount).

[0197] State Nodes: Record buyer responses (e.g., "rejected and countered with 75%", "expressed interest but requested due diligence").

[0198] Four-stage iterative optimization of tree search

[0199] Monte Carlo Tree Search (MCTS) employs a four-stage iterative optimization strategy:

[0200] Selection: Starting from the root node, select child nodes based on the Upper Confidence Bound (UCB) algorithm, prioritizing nodes with high UCB (such as aggressive strategy nodes selected due to their potential high returns).

[0201] Expansion: When the number of visits to a node reaches a threshold (e.g., N_i>50), expand to include new child nodes (e.g., add a "Buyer requests installment payment" branch).

[0202] Simulation: Run a random simulation on the new node until the transaction ends (successful / failed). The simulation uses a Buyer Behavior Model (BM) to predict the response probability (e.g., 70% probability of accepting 85% of the bids).

[0203] Backpropagation: Update the path node revenue value Q_i in reverse using the simulation results (e.g., the actual transaction price of 8.2 million).

[0204] Extraction of candidate pricing range

[0205] After 10,000 iterations, the distribution of transaction prices at all terminating nodes is statistically analyzed:

[0206] Extract nodes with a transaction probability > 60%, corresponding to a price range of [7.8 million, 8.5 million].

[0207] Exclude extreme values ​​(such as abnormal transactions that are 50% below the appraised value).

[0208] Divide the price range into buckets (Bucket Size=500,000), select the bucket with the densest distribution [8 million, 8.5 million] as the candidate asset pricing range (CAPR), and mark the transaction probability of each range (e.g., 8 million corresponds to 75%, 8.5 million corresponds to 68%).

[0209] Based on the pricing range of candidate assets, the probability distribution and error range are calculated using a Bayesian confidence model to generate a dynamic discount price range and confidence assessment for specific distressed assets.

[0210] Bayesian posterior probability modeling

[0211] The Bayesian Confidence Model (BCM) treats CAPR as observed data to infer the true reasonable price range:

[0212] Prior Distribution: Based on the variance of transaction prices of similar cases in HSCS. For example, if the standard deviation of transaction prices of 12 cases is σ = 1.2 million, then the prior distribution is a normal distribution N(mu_0, sigma_0^2), where mu_0 = CAPR median of 8.25 million and sigma_0 = 1.2 million.

[0213] Likelihood Function: Assuming that the observed prices follow a normal distribution centered at the true price θ, the standard deviation is determined by the number of simulations (the more MCTS iterations, the smaller the standard deviation).

[0214] Posterior distribution calculation: According to Bayes' theorem: P(θ∣CAPR)∝P(CAPR∣θ)×P(θ).

[0215] The posterior distribution remains a normal distribution, and the posterior mean mu_p and variance sigma_p^2 are obtained by weighted average of the prior and observed data.

[0216] Dynamic price range generation

[0217] Generate a dynamic discount price range (DDR) based on the posterior distribution:

[0218] Core interval: The 95% highest density interval (HDI) of the posterior distribution is taken, i.e., [mu_p - 1.96sigma_p, mu_p + 1.96sigma_p]. For example, if mu_p = 8.3 million and sigma_p = 350,000, then DDR = [7.628 million, 8.972 million].

[0219] Range dynamism: For each new market transaction data entry, MCTS and Bayesian updates are re-executed. For example, when a similar asset fails to sell at auction, sigma_p expands to 400,000 and DDR widens to [7.5 million, 9.1 million].

[0220] Confidence assessment and output

[0221] The final output consists of two parts:

[0222] Dynamic discount price range: DDR=[lower limit, upper limit], indicating the discount rate of the benchmark valuation price (e.g., 7.628 million corresponds to a discount rate of 23.7%).

[0223] Confidence Assessment Report (CAR):

[0224] Probability Guarantee: It is stated that "there is a 95% probability that this range contains a reasonable transaction price" (based on HDI).

[0225] Error range: Maximum Expected Error (MEE) = 2.58sigma_p (99% confidence level), for example, 903,000.

[0226] Sensitivity warning: List the impact of key risk factors (e.g., for every 1 point decrease in the industry prosperity index, the lower limit of the range decreases by 50,000).

[0227] This method combines internal corporate risk characteristics with the external market environment, using reinforcement learning algorithms to simulate pricing strategies under different market scenarios. The algorithm references transaction data from similar historical cases while considering current market supply and demand, dynamically adjusting pricing model parameters to ultimately provide a reasonable price range with probabilistic assessment. This achieves intelligent and dynamic pricing of distressed assets, avoiding the subjectivity and lag of traditional pricing methods. Confidence assessment provides clear risk warnings to all parties involved in the transaction, helping to improve transaction success rates and pricing rationality.

[0228] S204. Based on the preset preferences of potential investors and the dynamic discount price range, a multi-agent simulation negotiation and conflict resolution algorithm is used to perform intelligent and precise matching and transaction feasibility simulation processing to obtain the optimal shortlist of potential investors and a preliminary conflict-free transaction structure scheme.

[0229] Specifically, based on the preset preferences of potential investors, preference vectorization and normalization can be performed to obtain investor feature vectors;

[0230] Data Acquisition and Structure

[0231] The system first connects to the Investor Database (IDB), which stores each investor's preset preferences (PP), including investment size threshold (IST, e.g., RMB 5 million to 20 million), industry preference (IT, e.g., semiconductors or new energy), risk tolerance level (RTL, categorized into 1-5 levels), asset type preference (ATP, e.g., factory equipment or intellectual property), and payment cycle requirement (PCR, e.g., lump sum payment or 3-year installment plan). This unstructured text or option data is parsed using a Natural Language Processing (NLP) module: for example, "preferring the new energy industry" is converted into an industry encoding vector (IEV), and one-hot encoding (OHE) is used to convert discrete categories (e.g., "factory" or "patent") into binary vectors (BV). Continuous parameters (such as the minimum value of IST, Min_IST = 5 million, and the maximum value, Max_IST = 20 million) are directly included in the initial vector.

[0232] Vectorization and Feature Fusion

[0233] The parsed data is input into the Preference Vectorization Engine (PVE). This engine employs a Feature Weighted Fusion (FWF) algorithm: a weight (WT) is assigned to each preference dimension, with the weight values ​​derived from historical transaction data (e.g., WT=0.3 for RTL, WT=0.2 for ATP). The fusion process is achieved through weighted concatenation (WC): if an investor's IT vector is [0,1,0] (corresponding to new energy) and their ATP vector is [1,0] (factory), then the weighted fusion result is [0×0.4, 1×0.4, 0×0.4, 1×0.2, 0×0.2] (assuming industry weight WT_IT=0.4, asset type weight WT_ATP=0.2). Finally, a Raw Investor Feature Vector (RIFV) is generated, whose dimension is equal to the sum of all preference dimensions.

[0234] Normalization and standardization processing

[0235] To avoid model bias caused by differences in units, the RIFV needs to be normalized (NP). The Min-Max Scaling (MMS) algorithm is used: for each numerical feature in the RIFV (e.g., Min_IST = 5 million, Max_IST = 20 million), it is scaled to the [0,1] interval according to a formula. For example, for an investor with IST = 10 million, the scaled value is (1000-500) / (2000-500) = 0.333. Discrete features (such as binary vectors generated by OHE) do not need to be scaled. The normalized vector is then subjected to Principal Component Analysis (PCA) to reduce dimensionality, remove redundant features, and output the final Investor Feature Vector (IFV), which can be efficiently processed by the algorithm. The dimensionality is typically compressed to 10-15 dimensions.

[0236] Based on investor feature vectors and dynamic discount price ranges, a multi-agent system is created using an agent initialization algorithm to obtain a set of investor agents;

[0237] Intelligent agent attribute definition

[0238] The system initializes the investor agent (IA) based on IFV. Each IA contains three core attributes:

[0239] Static Attributes (SA): Directly mapped from IFV, such as Risk Level RTL=4 (maximum 5) and Industry Preference IT="New Energy".

[0240] Dynamic Strategy (DS): Defines negotiation behavior rules, such as Bidding Strategy (BS) as "Initial bid = lower limit of dynamic discount price range × 0.9".

[0241] Utility Function (UF): Quantifies transaction satisfaction. For example, UF = WT_Price × (discounted price / expected price) + WT_Time × (1 - payment period / longest period), where the weight parameters WT_Price = 0.6 and WT_Time = 0.4 are generated by training with historical data.

[0242] Construction of multi-agent systems

[0243] The Agent Initialization Algorithm (AIA) iterates through the IFVs of all potential investors, generating an independent IA for each investor. A Multi-Agent System (MAS) is then constructed, employing a Coalition Topology (CT) architecture: IAs can form temporary alliances (e.g., investors in similar industries jointly negotiating). The system injects external environmental parameters into the MAS, including the Dynamic Discount Price Range (DDPR, e.g., range [12 million, 15 million] for device A) and its Confidence Assessment (CA, e.g., range confidence probability Prob=85%). Each IA initializes its bid range (BR) based on the DDPR; for example, a conservative IA's BR=[12 million, 13 million], and an aggressive IA's BR=[14 million, 15 million].

[0244] Loading behavior rule library

[0245] To simulate real investor behavior, each IA loads a pre-defined Behavior Rule Library (BRL). These rules are represented using Production Rules (PRs), for example:

[0246] If the competitor's bid is higher than its own UF threshold (Threshold_UF=0.7);

[0247] THEN Adjust the price step size (Step_Size=5%).

[0248] The rule parameters are optimized through machine learning: the system trains a decision tree model (DTM) based on historical negotiation data to generate the optimal rule set. The final output is an investor agent set (IAS) containing N IAs, where N is the number of potential investors.

[0249] Based on the set of investor agents, a Nash negotiation game model is applied to conduct transaction simulation negotiation to obtain a preliminary set of transaction proposals.

[0250] Game Theory Framework Construction

[0251] The Nash Bargaining Game Model (NBGM) abstracts transaction negotiations into a multi-round game. Each round consists of three phases:

[0252] Proposal Phase (PP): Each IA submits a Transaction Proposal (TP), which includes the price (P), payment cycle (PC), and additional clauses (AC, such as employee resettlement commitments).

[0253] Utility Calculation Phase (UCP): Calculate the utility value (UV) for each proposal based on the utility factor (UF). For example, for a proposal with a proposal P = $13.5 million (within the DDPR) and a PC = 2 years, the UV = 0.6 × (13.5 million / 14 million) + 0.4 × (1 - 2 / 3) = 0.793.

[0254] Response Phase (RP): The IA adjusts its own strategy based on the UV of the opponent's proposal.

[0255] Multi-round negotiation simulation

[0256] The system initiates the Iterative Bargaining Protocol (IBP), setting the maximum number of rounds (Max_Round=10). The process for each round is as follows:

[0257] Random pairing (RP): Randomly pairing IAs in IAS into pairs (Pairing Group, PG).

[0258] Parallel Game (PG): Each pair of IAs negotiates independently based on NBGM. For example, IA_X (buyer) and IA_Y (seller) negotiate the price of device A:

[0259] In the first round: IA_X bid P_X1=12.5 million, and IA_Y asked for P_Y1=14.5 million.

[0260] In the second round: IA_X increases its bid to P_X2 = 13 million based on UF (Step_Size = 500,000).

[0261] Utility Comparison (UC): If the UV difference between the two parties' proposals ΔUV < the tolerance threshold Tol_ΔUV = 0.1, then a tentative agreement (TA) is reached.

[0262] Proposal integration and screening

[0263] After each round, all TAs are collected to form a Candidate Proposal Pool (CPP). Pareto Optimization (PO) is used to filter non-dominated solutions (NDS): for example, proposals TP1 (P=13.5 million, PC=2 years) and TP2 (P=14 million, PC=1 year) are not mutually dominant (because TP1 has a better price but slower payment). Finally, a Preliminary Transaction Proposal Set (PTPS) is output, containing all NDS proposals and their associated IA numbers.

[0264] Based on the preliminary transaction proposal set, conflict points are identified using conflict detection algorithms, and a conflict report is generated.

[0265] Conflict Dimension Definition

[0266] The Conflict Detection Algorithm (CDA) scans the PTPS from three dimensions:

[0267] Resource Conflict (RC): The same asset is bid on by multiple proposals. For example, proposals TP3 (IA_X to purchase equipment A) and TP4 (IA_Y to purchase equipment A) both target the same asset.

[0268] Clause Conflict (CC): The clauses attached to the proposals are mutually exclusive. For example, TP5 requires retaining employees, while TP6 requires a 50% reduction in staff.

[0269] Temporal Conflict (TC): A discrepancy between the payment period and the asset delivery time. For example, TP7 stipulates a two-year installment payment plan, but the asset must be transferred within one year.

[0270] Conflict identification based on rule engine

[0271] The system deploys a rule engine (RE) with a built-in conflict detection rule set (CDRS). Example rule:

[0272] If the proposed assets have the same ID;

[0273] AND the buyer IDs are different;

[0274] THEN is marked as a resource conflict (RC).

[0275] Each proposal's textual clause (such as "employee retention rate ≥ 80%) is converted into a logical expression (LE) by the Semantic Parsing Module (SPM). For example, "retain employees" is parsed as LE1: [employee layoff rate ≤ 20%], and "layoff 50%" is parsed as LE2: [employee layoff rate = 50%]. The engine identifies conflicts between LE1 and LE2 through Logical Contradiction Check (LCC).

[0276] Conflict Quantification and Report Generation

[0277] Detected conflicts are assigned a Conflict Intensity Value (CIV). For example, a resource conflict has a CIV_RC of 1.0 (highest level), while a clause conflict has a CIV_CC of 0.7 (dynamically adjusted based on clause importance). All conflicts are sorted by CIV to generate a Structured Conflict Report (SCR), which includes the Conflict Type (CT), the associated Proposal ID (PID), the Conflict Description (CD), and the CIV value. For example:

[0278] CT: RC | PID: TP3, TP4 | CD: Equipment A resold | CIV: 1.0;

[0279] CT: CC | PID: TP5,TP6 | CD: Conflicting employee resettlement terms | CIV: 0.7.

[0280] Based on the conflict report, a genetic algorithm was used for multi-objective optimization and resolution to generate an optimal shortlist of potential investors and a preliminary conflict-free transaction structure plan.

[0281] Genetic Algorithm Initialization

[0282] The genetic algorithm (GA) uses chromosome encoding (CE) designed as a transaction structure scheme (TSS). Each chromosome consists of a gene segment (GS):

[0283] GS1: Asset Allocation (AA), such as equipment A→IA_X.

[0284] GS2: Payment Cycle (PC), such as a 2-year installment plan.

[0285] GS3: Clause Index (CI), which points to entries in the Clause Library.

[0286] The initial population (IP) consists of 50 individuals, generated by randomly selecting a combination of proposals from PTPS.

[0287] Second section: Multi-objective fitness function optimization

[0288] The fitness function (FF) is defined as: FF = α × total transaction value + β × investor satisfaction - γ × conflict intensity.

[0289] For example, the weighting parameters are α=0.5 (economic benefit weight), β=0.3 (fairness weight), and γ=0.2 (conflict penalty weight). Calculation process:

[0290] Total Transaction Value (TTV): The sum of the selling prices of all assets in the scheme.

[0291] Investor Satisfaction (IS): Average Utility Value ∑(UV) / N.

[0292] Conflict Intensity (CI): The sum of all CIVs in the SCR.

[0293] For example, if a certain scheme has TTV=50 million, IS=0.82, CI=0.7, then FF=0.5×5000+0.3×0.82-0.2×0.7=2500.246.

[0294] Evolutionary operations and scheme generation

[0295] The algorithm performs the following evolutionary operations:

[0296] Selection (SEL): Retain the top 20% of individuals with the highest FF (elite retention strategy).

[0297] Crossover (CO): Randomly select two individuals to exchange GS. For example, exchange GS1 of individual X (device A→IA_X) with GS1 of individual Y (device A→IA_Y).

[0298] Mutation (MUT): Randomly modifies the value in GS. For example, changing PC from 2 years to 1 year.

[0299] After 100 generations of evolution, the individual with the highest FF (Failure Factor) is selected as the Conflict-Free Transaction Structure Scheme (CFTSS). An Optimal Investor Shortlist (OIS) is generated based on the investors involved in this scheme, typically limited to 3-5 core investors. The final scheme includes complete asset allocation, payment terms, and legal annexes, ensuring all conflicts are resolved (e.g., device A is exclusively allocated to IA_X, and employee placement terms are uniformly set at "retention rate ≥ 85%").

[0300] By constructing multiple intelligent agents to simulate the decision-making behavior of different investors, the system automatically analyzes the interests and risk preferences of each party. Game theory algorithms are used to identify potential transaction conflicts, and optimization algorithms are employed to find a balanced solution acceptable to all parties. Ultimately, the most suitable investors are selected, forming a feasible transaction framework. This significantly improves the efficiency and success rate of transaction matching, reducing the risk of negotiation breakdowns due to information asymmetry in traditional transactions. The pre-transaction function helps all parties identify and resolve potential problems in advance, reducing transaction uncertainty.

[0301] S205, for transactions that have reached an agreement, based on the aforementioned transaction structure, blockchain smart contracts are used to automate the execution of transaction terms, fund custody and transfer, and ownership change registration, thereby obtaining a final traceable distressed asset transaction record.

[0302] Specifically, based on the transaction structure, the terms can be digitized and smart contract encoded to obtain deployable smart contract code;

[0303] Once the two parties reach an agreement, the system receives a Transaction Structure Scheme (TSS) generated through multi-agent negotiation. This scheme includes complete legal and commercial terms, such as an asset delivery list, a payment schedule (PS), an earn-out clause (EC), and a liability for breach (LFB). The core of Clause Digitization (CD) is to transform clauses described in natural language into structured data that can be parsed by machines. First, the LegalClause Parsing Engine (LCPE) uses Named Entity Recognition (NER) technology to extract key elements. For example, in the clause "The buyer shall pay the initial payment (IP) of 60% of the total price within T+3 business days after the Closing Date (CD)," the system identifies the entity type (payer, payee, percentage of amount, time constraint) and logical relationships (condition trigger, time sequence). Next, the system calls the Parameter Mapping Table (PMT) to convert the entity into smart contract variables: for example, the settlement date is mapped to the global timestamp variable t_closing, the initial payment ratio is mapped to the floating-point variable ratio_IP=0.6, and the payment period is mapped to the time increment variable delta_T=3*86400 seconds (the number of seconds corresponding to 3 working days).

[0304] After extracting the clause elements, the Smart Contract Coder (SCC) generates executable logic based on a pre-built Clause-Code Template Library (CCTL). For example:

[0305] Installment payment terms are mapped to a time-based State Machine Function (SMF): when the blockchain timestamp block.timestamp satisfies t_closing + delta_T, it automatically checks whether the buyer address buyer_addr has transferred funds total_price * ratio_IP to the escrow contract.

[0306] The betting terms are encoded as an Oracle Trigger Condition (OTC): when the off-chain audit report (signed by the authorized node oracle_node) shows that the company's revenue has reached the revenue_threshold, the final payment_final is automatically released to the seller.

[0307] The default clause is transformed into Auto-arbitration Logic (AL): if the buyer fails to pay on time, the system automatically transfers the security deposit to the seller and triggers the asset repurchase option after the grace period ends. The coding process employs a Modular Encapsulation Strategy (MES), breaking down payment, settlement, and arbitration functions into independent function libraries (such as PaymentModule.sol and DeliveryModule.sol), which are then combined into a complete contract through interface calls.

[0308] To ensure code security and compliance, the system performs triple verification:

[0309] Static Analysis Check (SAC): Uses formal verification tools (such as MythX) to detect vulnerabilities such as reentrancy attacks (RA) and integer overflows (IO).

[0310] Legal Logic Verification (LVV): The generated code is decompiled into a natural language description and compared with the original TSS terms for consistency (e.g., verifying whether ratio_IP=0.6 accurately reflects the 60% down payment requirement).

[0311] Sandbox Simulation Test (SST): Simulates scenarios such as fund transfers and overdue settlements in a private blockchain environment to ensure that function triggering conditions (e.g., block.timestamp > t_closing + delta_T) and behaviors (e.g., automatic penalty calculation penalty = security_deposit * 0.1) meet expectations. Upon successful verification, standardized Deployable Smart Contract Code (DSCC) is output, typically bytecode (BC) written in Solidity or Vyper, along with an ABI interface description file.

[0312] Based on the deployable smart contract code, automatic terms verification is performed through the blockchain network to generate transaction execution event logs;

[0313] Deployable smart contract code (DSCC) is submitted to the target blockchain network (BN), which employs a permissioned blockchain architecture (such as Hyperledger Fabric) or a public blockchain (such as Ethereum Enterprise). The deployment process begins with the Contract Deployment Node (CDN) initiating a transaction request, attaching the DSCC bytecode and initialization parameters (such as seller and buyer addresses `seller_addr` / `buyer_addr`, and total price `total_price`). Consensus nodes (CNs) in the network then execute the verification process.

[0314] Code Compliance Check (CCC): Verifies whether the bytecode hash (BH) of the DSCC is consistent with the off-chain audit report to prevent tampering.

[0315] Permission Signature Verification (PSV): Requires both the buyer and seller to sign the deployment transaction using their private keys, generating a dual-signed message dual_sig=sign(seller_sk) + sign(buyer_sk) to ensure authorization by both parties.

[0316] After consensus is reached, the contract is assigned a unique address, contract_addr, and written into the genesis block (GB). At the same time, the collateral of both parties (such as buyer_deposit and seller_collateral) is frozen.

[0317] After the contract takes effect, Automatic Clause Verification (ACV) runs on-chain in real time:

[0318] Condition Listeners (CL) continuously monitor on-chain and off-chain events. For example, they listen to the block time (block.number, corresponding to physical time) and trigger the delivery function deliverAssets() when the delivery date condition (block.timestamp >= t_closing) is met.

[0319] External Data Access (EDA): Obtaining off-chain data through a decentralized oracle (DO). For example, when verifying a payment, the oracle node oracle_node uploads the SWIFT message hash swift_hash provided by the bank's clearing system to the blockchain. The contract then compares the hash values ​​and executes confirmPayment() after confirming a match.

[0320] Complex Logic Verification (CLV): In the performance-based agreement, when the company's annual financial report is due, the oracle cluster calls the signed audit report PDF hash audit_report_hash, the contract parses the key indicator revenue_actual and compares it with the threshold revenue_threshold to determine whether to release the final payment.

[0321] All verification operations generate a Transaction Execution Event Log (TEEL), which has the following characteristics:

[0322] Structured Storage (SS): Each log entry contains an event type (e.g., PaymentVerified), a triggering block number (block_num), a related address (related_addr), and a data snapshot (e.g., payment amount = 5 million).

[0323] Immutability (IM): After being signed by consensus nodes, logs are stored in a Merkle Tree (MT) structure. Any modification will result in a change to the root hash (root_hash).

[0324] Real-time Push (RTP): This involves pushing event notifications (such as Event: DeliveryConditionMet) to buyers and sellers via the WebSocket protocol. Logs serve as the legal input source for subsequent fund transfers and ownership changes.

[0325] Based on the transaction execution event log, a zero-knowledge proof mechanism is applied to process fund custody and automatic transfer, resulting in a secure fund transfer record.

[0326] Fund escrow (FE) is managed by a multi-signature on-chain escrow contract (OEC). The transfer process is initiated when the event log verification terms are met (e.g., TEEL.event_type == "PaymentDue"). To protect privacy, a zero-knowledge proof mechanism (ZKPM) is used.

[0327] Asset Concealment (AC): The actual amount paid by the buyer, actual_amount, is encrypted as a commitment, commitment = H(actual_amount || nonce) (where H is a hash function and nonce is a random number), proving only that actual_amount >= required_amount to the seller without revealing the specific value.

[0328] Relation Proof (RP): Use zk-SNARK (zero-knowledge concise non-interactive knowledge proof) to generate proof π, verifying that the equation actual_amount - required_amount >= 0 holds true, and that the commitment is consistent with the on-chain commitment.

[0329] Automatic transfer processing (ATP) is executed in stages:

[0330] Fund Locking (FL): The buyer converts fiat currency into stablecoin USD_token and transfers it to the escrow contract, generating a crypto balance commitment balance_commit.

[0331] Proof Verification (PV): The verification function verifyTransferProof() of the managed contract receives the proof π and public parameters (such as required_amount=3 million), and calls the pre-compiled zk-SNARK verification contract (such as Groth16Verifier).

[0332] Conditional Transfer (CT): After successful verification, execute:

[0333] Main transfer: Transfer required_amount from the escrow contract to the seller address seller_addr.

[0334] Overpayment handling: If actual_amount > required_amount, the difference is returned to the buyer (no new proof needs to be generated).

[0335] Collateral Release: Simultaneously release the asset tokens pledged by the seller.

[0336] After the transfer is completed, a Secure Fund Transfer Record (SFTR) is generated, which includes:

[0337] Privacy-preserving Voucher (PPV): Encrypted digest of fund flow tx_digest = H(sender || receiver || commitment).

[0338] Audit Trail Tag (ATT): Associates the block number ref_block and transaction hash tx_hash of the original event log.

[0339] Legal Effect Signature (LES): This signature, jointly signed by the escrow address (escrow_addr) and both parties' private keys (joint_sig), ensures the record can be used as legal evidence. This record fully meets the "verifiable but invisible" requirement, preventing the leakage of sensitive financial information.

[0340] Based on the secure fund transfer records, the ownership change registration and tamper-proof storage are performed through the hash chain verification protocol to generate the final traceable distressed asset transaction records.

[0341] Ownership Change Registration (OCR) is implemented based on the Hash Chain Verification Protocol (HCVP):

[0342] Asset Tokenization (AT): Distressed assets (such as plant and equipment) are represented as non-fungible tokens (NFTs) containing metadata hash (corresponding to an asset valuation report).

[0343] Change Request Construction (CRC): The seller initiates a change request, which includes:

[0344] Source asset NFT identifier asset_id;

[0345] Buyer receives address new_owner_addr;

[0346] The associated secure fund transfer record hash SFTR_hash;

[0347] Seller's signature: seller_sig;

[0348] This request generates a transaction tx_reg to be registered.

[0349] The core operations of the Hash Chain Verification Protocol (HCVP):

[0350] Chained Linking (CL): Concatenate tx_reg with SFTR_hash and the previous ownership record hash prev_owner_hash to calculate the new hash new_hash = H(prev_owner_hash || SFTR_hash || H(tx_reg)).

[0351] Multi-node Consensus (MNC): Registration requests are broadcast to the Asset Registry Node (ARN), and node verification is performed.

[0352] a) The seller is indeed the current owner of the NFT (check ownerOf(asset_id) == seller_addr);

[0353] b) SFTR_hash exists on the chain and its state is "completed";

[0354] c) The seller's signature (seller_sig) is valid.

[0355] Time Anchoring (TA): Obtain the timestamp block_time and block height block_height by writing new_hash to the block header (BH).

[0356] Immutable Storage Processing (ISP) generates a Final Traceable Distressed Asset Transaction Record (FTDATR):

[0357] Record Structuring (RS):

[0358] Integrate the following fields:

[0359] Asset change record (including the addresses of the old and new owners and the change timestamp);

[0360] Fund transfer proof (a cryptographic index pointing to SFTR);

[0361] The transaction lifecycle hash chain is defined as follows: hash_chain = [prev_owner_hash, new_hash, ...].

[0362] Distributed Storage (DS): Records are split into shards shard_1 to shard_n, and after redundant encoding, they are stored in the IPFS (InterPlanetary File System) cluster. The master hash ftdatr_root is written to the blockchain.

[0363] Cross-chain Notarization (CCN): The ftdatr_root is synchronized to the judicial evidence storage chain to obtain a notary receipt. At this point, the record has legal validity, is fully traceable, and cannot be tampered with, forming a closed-loop transaction evidence chain.

[0364] The negotiated transaction plan is encoded into a smart contract, automatically executing key steps such as payment and asset transfer. Blockchain technology ensures the transparency and immutability of the transaction process; all operations generate verifiable transaction records, achieving end-to-end digital management and solving the problems of cumbersome processes, inefficiency, and lack of trust in traditional asset transactions. The application of blockchain technology ensures transaction security and traceability, making it particularly suitable for handling complex distressed asset transactions, reducing dispute risks and legal costs.

[0365] As can be seen, by receiving multi-source, heterogeneous enterprise data from distressed companies, a structured and anonymized panoramic data profile of distressed companies can be obtained. From this panoramic data profile, a dynamic risk feature vector with multiple dimensions, including asset status, potential liabilities, operational interruption risks, and industry prosperity correlation, can be obtained. By linking the dynamic risk feature vector with real-time market transaction data and industry M&A case database, a dynamic discount price range and confidence assessment can be obtained for specific distressed assets. Based on the preset preferences of potential investors and the dynamic discount price range, an optimal shortlist of potential investors and a preliminary conflict-free transaction structure plan can be obtained. For transactions that reach an agreement, a final traceable distressed asset transaction record can be obtained, thereby dynamically optimizing asset pricing and transaction matching, and improving the efficiency and security of distressed asset transactions.

[0366] Another embodiment of the present invention provides an AI-based data information and asset transaction service system for distressed enterprises, see [link to relevant documentation]. Figure 3 The system may include:

[0367] The receiving module 301 is used to receive multi-source heterogeneous enterprise data of distressed enterprises, and perform cross-domain privacy-preserving fusion processing based on the federated learning framework to obtain a structured and desensitized panoramic data profile of distressed enterprises.

[0368] Extraction module 302 is used to create a panoramic data profile of the distressed enterprise, and to use a pre-trained deep adversarial generative network to extract asset risk features and model value influencing factors, so as to obtain a multi-dimensional dynamic risk feature vector that includes asset status, potential liabilities, operational interruption risk, and industry prosperity correlation.

[0369] The association module 303 is used to associate the dynamic risk feature vector with real-time market transaction data and industry M&A case library, and perform dynamic discount rate calculation and optimal asset pricing range deduction based on reinforcement learning algorithm to obtain dynamic discount price range and confidence assessment for specific distressed assets.

[0370] The matching module 304 is used to perform intelligent and accurate matching and transaction feasibility simulation processing based on the preset preferences of potential investors and the dynamic discount price range, using a multi-agent simulation negotiation and conflict resolution algorithm, to obtain an optimal shortlist of potential investors and a preliminary conflict-free transaction structure scheme.

[0371] The execution module 305 is used to automatically execute transaction terms, transfer funds in custody, and register changes in ownership for transactions that have reached an agreement, based on the transaction structure scheme, using blockchain smart contracts, to obtain a final traceable distressed asset transaction record.

[0372] 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.

[0373] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps:

[0374] S201 receives multi-source heterogeneous enterprise data of distressed enterprises, performs cross-domain privacy-preserving fusion processing based on the federated learning framework, and obtains a structured and de-identified panoramic data profile of distressed enterprises.

[0375] S202, for the panoramic data portrait of the distressed enterprise, use a pre-trained deep adversarial generative network to extract asset risk features and model value influencing factors, and obtain a multi-dimensional dynamic risk feature vector including asset status, potential liabilities, operational interruption risk, and industry prosperity correlation.

[0376] S203, the dynamic risk feature vector is associated with real-time market transaction data and industry M&A case library, and dynamic discount rate calculation and optimal asset pricing range deduction are performed based on reinforcement learning algorithm to obtain dynamic discount price range and confidence assessment for specific distressed assets.

[0377] S204. Based on the preset preferences of potential investors and the dynamic discount price range, a multi-agent simulation negotiation and conflict resolution algorithm is used to perform intelligent and precise matching and transaction feasibility simulation processing to obtain the optimal shortlist of potential investors and a preliminary conflict-free transaction structure scheme.

[0378] S205, for transactions that have reached an agreement, based on the aforementioned transaction structure, blockchain smart contracts are used to automate the execution of transaction terms, fund custody and transfer, and ownership change registration, thereby obtaining a final traceable distressed asset transaction record.

[0379] 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.

[0380] 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.

[0381] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0382] S201 receives multi-source heterogeneous enterprise data of distressed enterprises, performs cross-domain privacy-preserving fusion processing based on the federated learning framework, and obtains a structured and de-identified panoramic data profile of distressed enterprises.

[0383] S202, for the panoramic data portrait of the distressed enterprise, use a pre-trained deep adversarial generative network to extract asset risk features and model value influencing factors, and obtain a multi-dimensional dynamic risk feature vector including asset status, potential liabilities, operational interruption risk, and industry prosperity correlation.

[0384] S203, the dynamic risk feature vector is associated with real-time market transaction data and industry M&A case library, and dynamic discount rate calculation and optimal asset pricing range deduction are performed based on reinforcement learning algorithm to obtain dynamic discount price range and confidence assessment for specific distressed assets.

[0385] S204. Based on the preset preferences of potential investors and the dynamic discount price range, a multi-agent simulation negotiation and conflict resolution algorithm is used to perform intelligent and precise matching and transaction feasibility simulation processing to obtain the optimal shortlist of potential investors and a preliminary conflict-free transaction structure scheme.

[0386] S205, for transactions that have reached an agreement, based on the aforementioned transaction structure, blockchain smart contracts are used to automate the execution of transaction terms, fund custody and transfer, and ownership change registration, thereby obtaining a final traceable distressed asset transaction record.

[0387] 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 method for providing data information and asset trading services for distressed enterprises based on AI, characterized in that, The method includes: It receives multi-source heterogeneous enterprise data of distressed enterprises, performs cross-domain privacy-preserving fusion processing based on a federated learning framework, and obtains a structured and de-identified panoramic data profile of distressed enterprises. For the panoramic data profile of the distressed enterprises, a pre-trained deep adversarial generative network is used to extract asset risk features and model value influencing factors, resulting in a multi-dimensional dynamic risk feature vector that includes asset status, potential liabilities, operational interruption risk, and industry prosperity correlation. The dynamic risk feature vector is associated with real-time market transaction data and industry M&A case library. Based on reinforcement learning algorithm, dynamic discount rate calculation and optimal asset pricing range deduction are performed to obtain dynamic discount price range and confidence assessment for specific distressed assets. Based on the preset preferences of potential investors and the dynamic discount price range, a multi-agent simulation negotiation and conflict resolution algorithm is used to perform intelligent and precise matching and transaction feasibility simulation processing to obtain the optimal shortlist of potential investors and a preliminary conflict-free transaction structure scheme. For transactions that have reached an agreement, based on the aforementioned transaction structure, blockchain smart contracts are used to automate the execution of transaction terms, fund custody and transfer, and ownership change registration, resulting in a final traceable record of distressed asset transactions.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous enterprise data received from distressed enterprises is processed using a federated learning framework for cross-domain privacy-preserving fusion, resulting in a structured, anonymized panoramic data profile of distressed enterprises, including: Based on the input multi-source heterogeneous enterprise data, data cleaning and format standardization processes are performed to obtain a standardized data stream; Based on the normalized data stream, Gaussian noise is added using the local differential privacy mechanism in the federated learning framework to generate privacy-preserving data fragments; Based on privacy-preserving data fragments, heterogeneous data fusion processing is performed using a cross-domain feature alignment algorithm to obtain a fused feature matrix; Based on the fused feature matrix, a variational autoencoder is applied for dimensionality reduction and structural reconstruction to generate a structured and desensitized panoramic data profile of distressed enterprises.

3. The method according to claim 2, characterized in that, The aforementioned panoramic data profile of the distressed enterprise utilizes a pre-trained deep adversarial generative network for asset risk feature extraction and value influencing factor modeling, resulting in a multi-dimensional dynamic risk feature vector encompassing asset status, potential liabilities, operational disruption risk, and industry prosperity correlation, including: Based on the panoramic data profile of distressed enterprises, the data is input into the generator of a pre-trained Generative Adversarial Network (GAN) for feature generation processing to obtain a preliminary risk feature set. Based on the initial risk feature set, the GAN discriminator performs adversarial refinement to obtain a high-confidence risk feature vector. Based on the high-confidence risk feature vector, the influencing factors are modeled using a factor analysis model combined with the industry prosperity index to obtain the value influencing factor weight vector. Based on the value impact factor weight vector, dynamic feature synthesis and dimensional expansion processing are performed to generate a multi-dimensional dynamic risk feature vector that includes asset status, potential liabilities, operational interruption risk, and industry prosperity correlation.

4. The method according to claim 3, characterized in that, The process involves associating the dynamic risk feature vector with real-time market transaction data and an industry M&A case database, and using reinforcement learning algorithms to calculate the dynamic discount rate and extrapolate the optimal asset pricing range to obtain a dynamic discount price range and confidence assessment for a specific distressed asset. This includes: Based on the dynamic risk feature vector and real-time market transaction data, feature enhancement and correlation processing are performed to obtain the spatiotemporal enhanced feature vector; Based on the spatiotemporal enhanced feature vectors and the industry M&A case library, case matching is performed using the cosine similarity algorithm to obtain a set of highly similar cases; Based on a set of highly similar cases, a reinforcement learning agent is applied to simulate and extrapolate the discount rate, resulting in a preliminary dynamic discount rate strategy. Based on the initial dynamic discount rate strategy, the pricing range is optimized using the Monte Carlo tree search algorithm to obtain the pricing range of candidate assets. Based on the pricing range of candidate assets, the probability distribution and error range are calculated using a Bayesian confidence model to generate a dynamic discount price range and confidence assessment for specific distressed assets.

5. The method according to claim 4, characterized in that, The process involves using a multi-agent simulation negotiation and conflict resolution algorithm to perform intelligent and precise matching and transaction feasibility simulation based on the preset preferences of potential investors and the dynamic discount price range, resulting in an optimal shortlist of potential investors and a preliminary conflict-free transaction structure plan, including: Based on the preset preferences of potential investors, preference vectorization and normalization are performed to obtain investor feature vectors; Based on investor feature vectors and dynamic discount price ranges, a multi-agent system is created using an agent initialization algorithm to obtain a set of investor agents; Based on the set of investor agents, a Nash negotiation game model is applied to conduct transaction simulation negotiation to obtain a preliminary set of transaction proposals. Based on the preliminary transaction proposal set, conflict points are identified using conflict detection algorithms, and a conflict report is generated. Based on the conflict report, a genetic algorithm was used for multi-objective optimization and resolution to generate an optimal shortlist of potential investors and a preliminary conflict-free transaction structure plan.

6. The method according to claim 5, characterized in that, For transactions where an agreement has been reached, based on the aforementioned transaction structure, blockchain smart contracts are used to automate the execution of transaction terms, fund custody and transfer, and ownership change registration, resulting in a final traceable distressed asset transaction record, including: Based on the transaction structure, the terms are digitized and smart contract coding is performed to obtain deployable smart contract code; Based on the deployable smart contract code, automatic terms verification is performed through the blockchain network to generate transaction execution event logs; Based on the transaction execution event log, a zero-knowledge proof mechanism is applied to process fund custody and automatic transfer, resulting in a secure fund transfer record. Based on the secure fund transfer records, the ownership change registration and tamper-proof storage are performed through the hash chain verification protocol to generate the final traceable distressed asset transaction records.

7. An AI-based data information and asset trading service system for distressed enterprises, characterized in that, The system includes: The receiving module is used to receive multi-source heterogeneous enterprise data of distressed enterprises, and perform cross-domain privacy-preserving fusion processing based on the federated learning framework to obtain a structured and de-identified panoramic data profile of distressed enterprises. The extraction module is used to create a panoramic data profile of the distressed enterprise, and to use a pre-trained deep adversarial generative network to extract asset risk features and model value influencing factors, so as to obtain a multi-dimensional dynamic risk feature vector that includes asset status, potential liabilities, operational interruption risk, and industry prosperity correlation. The association module is used to associate the dynamic risk feature vector with real-time market transaction data and industry M&A case library, and to perform dynamic discount rate calculation and optimal asset pricing range deduction based on reinforcement learning algorithm to obtain dynamic discount price range and confidence assessment for specific distressed assets. The matching module is used to perform intelligent and accurate matching and transaction feasibility simulation processing based on the preset preferences of potential investors and the dynamic discount price range, using a multi-agent simulation negotiation and conflict resolution algorithm, to obtain an optimal shortlist of potential investors and a preliminary conflict-free transaction structure scheme. The execution module is used to automatically execute transaction terms, transfer funds in custody, and register changes in ownership for transactions that have reached an agreement, based on the transaction structure scheme, using blockchain smart contracts, to obtain a final traceable record of distressed asset transactions.

8. The system according to claim 7, characterized in that, The receiving module is specifically used for: Based on the input multi-source heterogeneous enterprise data, data cleaning and format standardization processes are performed to obtain a standardized data stream; Based on the normalized data stream, Gaussian noise is added using the local differential privacy mechanism in the federated learning framework to generate privacy-preserving data fragments; Based on privacy-preserving data fragments, heterogeneous data fusion processing is performed using a cross-domain feature alignment algorithm to obtain a fused feature matrix; Based on the fused feature matrix, a variational autoencoder is applied for dimensionality reduction and structural reconstruction to generate a structured and desensitized panoramic data profile of distressed enterprises.

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