AI-based enterprise special asset diversified value-added service method and system

Abstract

The invention discloses an AI-based enterprise special asset diversified value-added service method and system, and the method comprises the steps: receiving multi-source heterogeneous data of enterprise special assets, and obtaining an enterprise special asset data graph containing asset attributes, market environments and historical disposal records; inputting the enterprise special asset data atlas into a pre-trained AI-based deep value evaluation model, and generating an asset value evaluation report including basic values, potential value-added paths and risk factors; according to the asset value evaluation report, utilizing a multi-target optimization algorithm to generate a plurality of preliminary value-added scheme sets for different value-added targets in parallel; and inputting the initial value-added scheme set into an intelligent matching engine, carrying out scheme feasibility verification and suitability screening, and outputting an optimal customized value-added service scheme package. By utilizing the embodiment of the invention, the accuracy of asset value analysis and the potential mining efficiency can be improved, and dynamic optimization and risk controllability of an asset value-added path are realized.

Description

Technical Field

[0001] This invention belongs to the field of AI technology, specifically an AI-based method and system for diversified value-added services of special enterprise assets. Background Technology

[0002] Currently, the disposal and value enhancement of special corporate assets (such as non-performing assets, idle equipment, and intellectual property) face pain points such as fragmented data, inaccurate value assessment, and a lack of diverse strategies. Traditional methods rely on human experience and struggle to integrate multi-source heterogeneous data (such as financial, legal, and market data), resulting in low assessment efficiency and insufficient exploration of value-added pathways. Furthermore, the rapid changes in market dynamics and industry trends further increase the complexity of decision-making, and existing technologies lack dynamic optimization capabilities, making it difficult to generate personalized solutions that balance risk and return. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-based method and system for diversified value-added services of special corporate assets, in order to address the shortcomings of existing technologies, improve the accuracy of asset value analysis and the efficiency of potential mining, and achieve dynamic optimization and risk control of asset value-added paths.

[0004] One embodiment of this application provides an AI-based method for diversified value-added services of enterprise special assets, the method comprising:

[0005] Receive multi-source heterogeneous data on special assets of enterprises, and clean, associate and structure the multi-source heterogeneous data based on preset data fusion rules to obtain a data map of special assets of enterprises that includes asset attributes, market environment and historical disposal records;

[0006] The enterprise's special asset data map is input into a pre-trained AI-based deep valuation model. Combined with real-time market dynamics and industry trends, multi-dimensional value analysis and potential mining are carried out to generate an asset valuation report that includes basic value, potential value-added paths and risk factors.

[0007] Based on the asset valuation report, a multi-objective optimization algorithm is used to generate multiple preliminary value-added scheme sets for different value-added objectives in parallel. Each scheme set contains multiple specific service strategy combinations.

[0008] The initial set of value-added solutions is input into the intelligent matching engine. Based on the preset enterprise profile and real-time constraints, the feasibility of the solutions is verified and the adaptability is screened, and the optimal customized value-added service solution package is output.

[0009] Optionally, the multi-source heterogeneous data of the received enterprise's special assets is cleaned, correlated, and structured based on preset data fusion rules to obtain a data map of the enterprise's special assets containing asset attributes, market environment, and historical disposal records, including:

[0010] By using knowledge graph ontology mapping technology, multi-source heterogeneous data of enterprise special assets are mapped into a unified topology structure, and an initial data topology graph with weight labels is output.

[0011] Based on the initial data topology, an adaptive sliding window is used to detect outliers in time-series data. Combined with an industry rule base, contradictory fields are automatically corrected to generate a cleaned spatiotemporally aligned data stream.

[0012] The spatiotemporally aligned data stream is input into the multi-head attention association engine to identify implicit association rules between asset attributes and market environment, and output an enhanced data network with cross-modal association factors;

[0013] By leveraging an enhanced data network-driven neural radiation field model, a data map of enterprise-specific assets containing three-dimensional spatiotemporal relationships is generated, in which historical disposal records are encoded as traceable time-dimensional vectors.

[0014] Optionally, the step of inputting the enterprise's special asset data map into a pre-trained AI-based deep valuation model, combined with real-time market dynamics and industry trends, to perform multi-dimensional value analysis and potential mining, generates an asset valuation report that includes basic value, potential value-added paths, and risk factors, including:

[0015] Input the enterprise's special asset data map into the spatiotemporal graph convolutional network, extract the dynamic evolution features of asset attributes and the spatiotemporal coupling features of the market environment, and output a multidimensional feature tensor.

[0016] The real-time industry trend data is compressed into a latent space vector by a variational encoder, and then the latent space vector is fused with a multidimensional feature tensor by tensor product to generate an industry-enhanced feature matrix.

[0017] Using the industry-enhanced feature matrix as input, a deep reinforcement learning-driven Monte Carlo tree search is employed to explore potential value-added paths, and the output path branch tree with probability weights is generated.

[0018] Based on path branching trees, stochastic differential equations are used to model the interaction between market volatility and asset characteristics, generating a dynamic risk factor field and sensitivity parameter set.

[0019] The report generation engine inputs dynamic risk factors into the report, creating an asset valuation report with a sandwich structure that decouples the core layer of basic value, the intermediate layer of value-added path, and the outer shell of risk.

[0020] Optionally, based on the asset valuation report, a multi-objective optimization algorithm is used to generate multiple preliminary value-added solution sets for different value-added objectives in parallel. Each solution set contains multiple specific service strategy combinations, including:

[0021] The basic value, appreciation path, and risk factors in the asset valuation report are analyzed and mapped into a three-dimensional target space coordinate system of economic benefits, disposal cycle, and risk entropy value.

[0022] Simulated annealing algorithm is used in parallel optimization in three-dimensional target space to generate a set of candidate solutions that satisfy the Pareto front.

[0023] The candidate schemes are input into the genetic algorithm, and the service strategy is combined through crossover and mutation operations, and the strategy combination chromosome group is output.

[0024] Apply a corporate compliance constraint filter to the strategy portfolio chromosome group to generate an initial set of value-added solutions with fitness scores.

[0025] Optionally, the initial set of value-added solutions is input into the intelligent matching engine, and based on the preset enterprise profile and real-time constraints, the feasibility of the solutions is verified and the adaptability is screened, outputting the optimal customized value-added service solution package, including:

[0026] By injecting enterprise profile parameters into a pre-set digital twin engine, a dynamic simulation model of the enterprise is constructed, which includes at least financial structure and production capacity flexibility.

[0027] The system uses Apache Flink to process constraint stream data containing at least liquidity information in real time, generating a dynamic set of constraint boundary conditions.

[0028] The initial set of value-added solutions is input into the enterprise's dynamic simulation model. The implementation process of the solutions is simulated in a neural radiation field, and a feasibility heatmap with failure probability is output.

[0029] Based on the feasibility heatmap and dynamic constraint boundary condition set, an adaptive weighted algorithm with fuzzy logic control is used to dynamically assign weights to the three-dimensional objectives of the scheme and generate an optimized scheme pool.

[0030] The optimization solution pool is topologically compressed, redundant strategy combinations are eliminated, and finally a lightweight customized value-added service solution package is output.

[0031] Another embodiment of this application provides an AI-based enterprise special asset diversification and value-added service system, the system comprising:

[0032] The receiving module is used to receive multi-source heterogeneous data of enterprise special assets, and clean, associate and structure the multi-source heterogeneous data based on preset data fusion rules to obtain a data map of enterprise special assets containing asset attributes, market environment and historical disposal records.

[0033] The analysis module is used to input the enterprise's special asset data map into a pre-trained AI-based deep value assessment model, and combine it with real-time market dynamics and industry trends to perform multi-dimensional value analysis and potential mining, generating an asset value assessment report that includes basic value, potential value-added paths and risk factors.

[0034] The generation module is used to generate multiple preliminary value-added scheme sets for different value-added objectives in parallel using a multi-objective optimization algorithm based on the asset valuation report. Each scheme set contains multiple specific service strategy combinations.

[0035] The output module is used to input the preliminary value-added solution set into the intelligent matching engine, and based on the preset enterprise profile and real-time constraints, to perform feasibility verification and adaptability screening of the solutions, and output the optimal customized value-added service solution package.

[0036] 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.

[0037] 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.

[0038] Compared with existing technologies, this invention provides an AI-based method for diversified value-added services of enterprise special assets. It receives multi-source heterogeneous data on enterprise special assets to obtain a data map of these assets, including asset attributes, market environment, and historical disposal records. The data map is then input into a pre-trained AI-based deep valuation model to generate an asset valuation report containing basic value, potential value-added paths, and risk factors. Based on the asset valuation report, a multi-objective optimization algorithm is used to generate multiple preliminary value-added solution sets targeting different value-added objectives. These preliminary value-added solution sets are then input into an intelligent matching engine. Based on a preset enterprise profile and real-time constraints, the engine verifies the feasibility and adaptability of the solutions, outputting the optimal customized value-added service package. This improves the accuracy of asset value analysis and the efficiency of potential discovery, enabling dynamic optimization of asset value-added paths and controllable risks. Attached Figure Description

[0039] Figure 1A hardware structure block diagram of a computer terminal for an AI-based method for diversified value-added services of special enterprise assets, provided in an embodiment of the present invention;

[0040] Figure 2 A flowchart illustrating an AI-based method for diversified value-added services of special enterprise assets, provided as an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of the structure of an AI-based enterprise special asset diversified value-added service system provided in an embodiment of the present invention. Detailed Implementation

[0042] 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.

[0043] The present invention first provides an AI-based method for diversified value-added services of special enterprise assets. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0044] 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 diversifying and adding value to special enterprise assets, provided as an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0045] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any AI-based method for diversifying and adding value to enterprise-specific assets.

[0046] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0047] 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 method for diversifying and adding value to enterprise-specific assets.

[0048] 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.

[0049] 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.

[0050] See Figure 2 The present invention provides an AI-based method for diversified value-added services of special enterprise assets, which may include the following steps:

[0051] S201, Receive multi-source heterogeneous data of enterprise special assets, and clean, associate and structure the multi-source heterogeneous data based on preset data fusion rules to obtain a data map of enterprise special assets containing asset attributes, market environment and historical disposal records;

[0052] Specifically, knowledge graph ontology mapping technology can be used to map multi-source heterogeneous data of a company's special assets into a unified topology, and output an initial data topology graph with weighted labels.

[0053] Multi-source heterogeneous data acquisition and ontology definition

[0054] The system first accesses multi-source heterogeneous data on the enterprise's special assets, including:

[0055] Structured data (such as net asset value statements in financial systems and equipment depreciation records in ERP systems);

[0056] Semi-structured data (such as PDF documents of asset appraisal reports, HTML pages of judicial auction announcements);

[0057] Unstructured data (such as drone aerial videos of asset site inspections and audio logs of equipment operating status).

[0058] For different data types, a Domain Ontology Library (DOL) is pre-built. For example, for industrial equipment assets, the ontology library defines core entities (such as "equipment ID", "geographic location", and "technical parameters") and relationships (such as "own factory - equipment" and "mortgage status - legal risk"). The Ontology Mapping Engine (OME) uses semantic alignment algorithms (such as word vector similarity matching based on Word2Vec) to automatically classify raw data fields to pre-defined ontology nodes. For example, "factory building valuation" in a PDF report is mapped to the "asset value" node in the ontology, and "roof damage" in an aerial video is mapped to the "physical depreciation" node, forming a preliminary semantic association network.

[0059] Unified topology construction and weight labeling

[0060] The mapped data is transformed into a unified topology using a Graph Constructor (GC). The topology employs a Property Graph Model (PGM), where nodes represent asset entities (e.g., "an idle factory") and edges represent relationships (e.g., "adjacent to a highway"). Weight labels are injected using a dual-weight calculation model.

[0061] Data Source Credibility Weight (DSCW): Dynamically assigned based on the authority of the data source (e.g., DSCW=0.9 for government registration systems, DSCW=0.3 for social media rumors).

[0062] Temporal Freshness Weight (TFW): Calculated based on the decay of data update time (e.g., TFW=1.0 for data within the most recent month, and TFW=0.6 for data from 1 year ago).

[0063] For example, the geographical location node of a factory is associated with the edge "2 kilometers from the subway station". Because it comes from the GIS data of the Land and Resources Bureau (DSCW=0.95) and was updated last week (TFW=0.98), the final edge weight composite value is 0.95×0.98=0.931.

[0064] Conflict resolution and topology optimization

[0065] When conflicting data exist for the same asset attribute (e.g., two appraisal reports differ by more than 20% in the residual value of equipment), the ontology-based Conflict Resolver (OCR) is activated. This module automatically selects high-weight data sources based on preset priority rules in the ontology library (e.g., "court ruling > asset appraisal report > company self-inspection") and adds conflict marker edges (CME) to the topology graph to record the original conflicts. The final output Initial Data Topology Graph (IDTG) is stored in JSON-LD format, with each node / edge carrying weight labels such as DSCW and TFW, laying the foundation for subsequent time-series processing.

[0066] Based on the initial data topology, an adaptive sliding window is used to detect outliers in time-series data. Combined with an industry rule base, contradictory fields are automatically corrected to generate a cleaned spatiotemporally aligned data stream.

[0067] Timing Anomaly Detection

[0068] Extract time-series data streams (such as the rental fluctuation curve of a commercial real estate property over the past 3 years) from IDTG and input them into an Adaptive Sliding Window Detector (ASWD). The window size (WS) is dynamically adjusted according to the data frequency.

[0069] High-frequency data (such as minute-level prices of stock market-linked REITs funds) are presented in a small window (WS=30 data points).

[0070] Low-frequency data (such as the annual assessment price of industrial land) are processed using a large window (WS=5 years).

[0071] Within each window, the Local Outlier Factor (LOF) algorithm is used to calculate the anomaly score (AS) for each data point. For example, if a sudden 60% drop in rent in a given month is detected (AS > 2.5), it is marked as an anomaly and triggers industry rule validation.

[0072] Industry rule-driven revision

[0073] Anomalies are input into the Industry Rule Base (IRB). The IRB contains three types of rules:

[0074] Market fluctuation rules (such as "maximum monthly drop threshold for commercial real estate rents = 30%));

[0075] Policy-related rules (such as "the valuation decline rate of high-pollution equipment is ≥15% within 6 months after the release of new environmental protection policies").

[0076] Physical constraints (such as "the annual depreciation rate of equipment shall not exceed the tax law limit of 20%").

[0077] When an abnormal 60% drop in rent is detected, the IRB searches the related event database and finds that the area was under lockdown that month due to the pandemic (event credibility weight = 0.88). It automatically corrects the drop to the industry rule-allowed upper limit of 30% and adds a correction flag (CF) to the data stream.

[0078] Spatiotemporal alignment and data reconstruction

[0079] The revised data stream input spatiotemporal alignment engine (SAE). The engine performs two steps:

[0080] Spatial alignment: unify scattered coordinates (such as "longitude of a factory building" in different GIS standards) to the WGS84 coordinate system;

[0081] Time alignment: Normalizes timestamps from multiple time zones (such as contract dates in UTC+8 and exchange data in UTC-5) into a standard timeline.

[0082] The final output is a cleaned spatiotemporal data stream (CSDS), which is an Avro data sequence with timestamp (TS) and spatial encoding (GeoHash (GH) to ensure the basic consistency of subsequent correlation analysis.

[0083] The spatiotemporally aligned data stream is input into the multi-head attention association engine to identify implicit association rules between asset attributes and market environment, and output an enhanced data network with cross-modal association factors;

[0084] Multimodal feature embedding

[0085] The CSDS data stream is input into the Feature Embedding Layer (FEL), which transforms different types of data into a unified vector space.

[0086] Numerical features (such as asset area and valuation) are mapped to the [0,1] interval through quantile normalization (QN);

[0087] Textual features (such as judicial auction announcements) are used to generate 768-dimensional semantic vectors through a BERT fine-tuning model;

[0088] Spatiotemporal features (such as geographic location) are converted into 64-dimensional dense vectors using GeoHash grid encoding (GE).

[0089] For example, the textual description of the "pollution level" of an industrial site, after being encoded by BERT, can be compared with the numerical feature of "surrounding housing prices" in the same vector space.

[0090] Multi-head attention association mining

[0091] The engine uses embedded vector inputs and employs a multi-head attention association engine (MAAE). The engine comprises eight independent attention heads (AHs), each focusing on a different association dimension.

[0092] AH1 calculates the correlation between the physical attributes of assets and market supply and demand (such as "equipment capacity - raw material futures prices").

[0093] AH2 explores the correlation between legal status and financing costs (such as "collateral status - LPR interest rate fluctuations").

[0094] Each head outputs an Attention Weight Matrix (AWM) to indicate the strength of correlations between cross-modal features (e.g., the AWM value for "remaining equipment lifespan" and "inquiries on secondhand trading platforms" is 0.76). Key implicit rules are extracted using gradient-based saliency analysis (GSA), for example, finding that "industrial park policy support" has a weight of 0.63 on "valuation of environmental protection equipment".

[0095] Cross-modal factor injection and network enhancement

[0096] The outputs of each attention head are concatenated into a Cross-modal Association Factor (CAF), formatted as [factor type:weight] key-value pairs (e.g., "policy sensitivity:0.82"). This factor is injected into the initial data topology.

[0097] Add new virtual nodes (such as "Regional Economic Prosperity Index") to support CAF;

[0098] Add CAF-weighted edges between the original nodes (e.g., update the weight of the edge connecting "a factory" and "logistics enterprise clustering" to 0.75×CAF=0.62).

[0099] The final result is an Enhanced Data Network (EDN), whose edge weights include the combined value of the original DSCW / TFW and the newly added CAF (weight = α × DSCW + β × CAF, α + β = 1), forming a topological structure with quantifiable implicit rules.

[0100] By leveraging an enhanced data network-driven neural radiation field model, a data map of enterprise-specific assets containing three-dimensional spatiotemporal relationships is generated, in which historical disposal records are encoded as traceable time-dimensional vectors.

[0101] Spatiotemporal modeling of neural radiation fields

[0102] The EDN input neural radiance field model (NeRF) consists of a three-layer structure:

[0103] Spatial Encoding Layer: The geographic coordinates of the asset (longitude X, latitude Y, altitude Z) are expanded into a high-dimensional vector (e.g., 64-dimensional) through Positional Encoding (PE).

[0104] Time encoding layer: Converts timestamps into periodic time vectors (PTV) (e.g., quarterly fluctuations are encoded as sin / cosin function pairs);

[0105] Feature fusion layer: Input spatial encoding, temporal encoding and EDN node features into multilayer perceptron (MLP), and output voxel density value (σ) and feature vector (F).

[0106] For example, the state of a warehouse in Q2 of 2023 is modeled as a joint function of spatial voxels [114.12°E, 22.5°N, 15m] and time vectors [sin(π / 2), cos(π / 2)].

[0107] Historical disposal record coding

[0108] Historical disposal records (such as "judicial auction failures in 2020") are encoded using a Temporal Convolutional Network (TCN):

[0109] Input layer: One-hot encoding of the event type (auction / lease / restructuring);

[0110] Convolutional layers: Use dilated convolution to capture long-term dependencies (such as the impact of a failed auction 3 years ago on the current valuation);

[0111] Output layer: Generates a 128-dimensional temporal dimension vector (TDV), whose direction reflects the nature of the disposal (the failed bid vector points to the negative quadrant), and whose magnitude represents the intensity of the impact.

[0112] TDV is dynamically bound to the corresponding asset node. For example, the TDV of a certain land parcel is [-0.34, 0.72,...], which means "strong negative auction event + moderate positive lease event".

[0113] 3D map generation and traceability implementation

[0114] NeRF model rendering generates Enterprise Special Asset DataAtlas (ESADA):

[0115] Spatial dimension: Assets are distributed in a 3D map according to their actual geographical location (e.g., industrial equipment is precisely located in the factory BIM model).

[0116] Time dimension: Dynamically display the changes in asset status through a sliding timeline (TS) (such as the evolution of a land value heatmap from 2019 to 2023).

[0117] Attribute Dimension: Clicking on a node will expand the attribute panel with the TDV traceability chain (e.g., "2020 failed auction → TDV component V12 = -0.34 → resulting in an 8% decrease in the current valuation").

[0118] The map is stored in an interactive WebGL format, which supports querying the asset status at any point in time and space through backward ray tracing (BRT), meeting the needs of full lifecycle tracing.

[0119] This step integrates enterprise-specific asset data from various channels (such as financial statements, market transaction records, and legal documents), employing data fusion technology to address issues like inconsistent data formats and timelines. By cleaning and removing noise from the data, and linking entity relationships across different data sources, a structured data graph is constructed. This transforms scattered information into a network structure with clearly defined relationships, achieving standardization and interoperability of multi-source data and providing a high-quality data foundation for subsequent valuation. The construction of the data graph enables unified analysis and mining of asset attributes, market environment, and historical disposal records, avoiding analytical biases caused by data silos.

[0120] S202, input the enterprise's special asset data map into a pre-trained AI-based deep value assessment model, combine real-time market dynamics and industry trends to conduct multi-dimensional value analysis and potential mining, and generate an asset value assessment report that includes basic value, potential value-added paths and risk factors.

[0121] Specifically, the data map of a company’s special assets can be input into a spatiotemporal graph convolutional network to extract the dynamic evolution features of asset attributes and the spatiotemporal coupling features of the market environment, and output a multidimensional feature tensor.

[0122] The enterprise-specific asset data graph, as input, is essentially a dynamic knowledge network encompassing spatiotemporal relationships. Nodes in this graph represent asset entities (such as idle factory buildings or patent portfolios) and market factors (such as regional GDP growth rates and industry policies), while edges represent the connections between nodes (such as the relationship between factory location and regional logistics costs). The core of the Spatial-Temporal Graph Convolutional Network (STGCN) is to simultaneously capture spatial topological relationships and temporal evolution patterns. First, the spatial convolutional layer aggregates neighbor node information based on the graph's adjacency matrix (ADJ, describing node connection strength) using graph filters. For example, when calculating a node representing industrial land, it aggregates the attributes of adjacent transportation hub nodes and upstream / downstream enterprise nodes in the industrial chain (e.g., land use type ADJ weight 0.8, transportation node weight 0.5). Secondly, the temporal convolutional layer employs dilated causal convolution (DCC) to scan historical disposal record vectors (such as quarterly rent change sequences) along a time sliding window (e.g., the past 12 months) to extract trend features (such as the annualized rent growth rate). Finally, spatial and temporal features are fused in the spatiotemporal coupling module: the lagged impact of market environment changes (such as interest rate increases) on asset attributes (such as mortgage valuation) is modeled as a spatial-temporal attention weight (STAW), outputting a four-dimensional multi-dimensional feature tensor (MDFT), whose dimensions are asset quantity × time step × feature channel × spatial relationship, such as the joint feature vector of a certain equipment asset in Q3 2023, which is "technical depreciation rate - regional demand heat".

[0123] The key to dynamic feature extraction lies in handling non-stationarity. For example, if the valuation of a commercial real estate property is affected by sudden policy changes (such as new policies in free trade zones), STGCN responds through the following mechanism:

[0124] Adaptive Graph Learning (AGL): The adjacency matrix ADJ is updated in real time. When a policy node appears, AGL automatically calculates its association strength with all asset nodes (e.g., the ADJ weight between a new policy node and a bonded warehouse node increases from 0 to 0.9).

[0125] Gated Temporal Unit (GTU): Introduces a Forget Gate (FG) and an Input Gate (IG) into temporal convolution. If the data for a certain month is abnormal (such as zero transactions due to the pandemic), the FG reduces the weight of that time step, and the IG strengthens trend smoothing.

[0126] In the final MDFT output, each asset corresponds to a dynamic feature cube: the bottom layer represents static attributes (such as area and property rights term), the middle layer represents short-term fluctuations (such as quarterly inquiry volume), and the top layer represents long-term trends (such as 5-year appreciation potential). This tensor becomes the input base for subsequent value analysis.

[0127] To verify the effectiveness of the features, the system incorporates a Feature Interpretability Module (FIM). For example, for a certain patent portfolio asset, FIM analyzes the feature channel with the highest weight in the MDFT as the "technology life cycle - market competition density" coupling factor (weighting 35%), which is consistent with industry expert experience. Simultaneously, by visualizing the spatiotemporal region of interest through Grad-weighted Class Activation Mapping (Grad-CAM), it shows a surge in potential demand for this patent in the Yangtze River Delta electronics market in Q2 2024 (STAW=0.92), providing a basis for further potential exploration.

[0128] The real-time industry trend data is compressed into a latent space vector by a variational encoder, and then the latent space vector is fused with a multidimensional feature tensor by tensor product to generate an industry-enhanced feature matrix.

[0129] Real-time industry trend data includes macroeconomic indicators (such as the PMI index), industry reports (such as IC Insights semiconductor forecasts), and social media sentiment (such as the popularity of new energy keywords on Twitter). This data is high-dimensional, sparse, and multimodal (text, time series, statistical charts). The Variational Encoder (VE) achieves efficient compression through an encoder-sampling-decoder structure.

[0130] Encoder: Uses a multi-head Transformer to process text reports (e.g., generating semantic vectors for “chip shortage will continue until Q4 2025”), and a TCN network to process time-series indicators (e.g., the 12-month fluctuation cycle of PMI), outputting raw latent space distribution parameters (mean μ, variance σ).

[0131] Reparameterization Trick (RT): Samples from an N(μ,σ²) distribution to generate a low-dimensional latent space vector (LSV), typically with 128-256 dimensions. For example, the 37th dimension of the vector represents the "strength of support for green energy policies".

[0132] This process ensures that LSVs with similar industry trends (such as photovoltaic and wind power subsidy policies) are close in potential space, thus improving generalization.

[0133] Tensor Product Fusion (TPF) is the core of industry information injection. It involves performing a tensor product operation between the LSV and the MDFT output by STGCN:

[0134] Dimensional alignment: Extend the LSV to a tensor of the same order as the MDFT (e.g., LSV

[256] → extended tensor[1×1×256×1]).

[0135] Outer Product (OP): The extended tensor is outer-productted with each local feature block of the MDFT to generate a fusion block. For example, the outer product of the feature vector [0.7, -0.2] (technology depreciation rate, demand intensity) of an asset with the "carbon tax policy" dimension (value = 1.5) in the LSV produces a new vector [1.05, -0.3], which represents the superimposed impact of the policy on the asset.

[0136] The final output is the Industry-enhanced Feature Matrix (IEFM), whose dimensions are: number of assets × time step × (original feature channels × latent space dimension). Each element in the matrix is ​​the interaction result of asset characteristics and industry factors.

[0137] The fusion effect is optimized through cross-modal contrastive learning (CCL). Positive sample pairs: IEFM and actual value-added results for an asset during periods of favorable industry reports; negative sample pairs: IEFM and results for the same asset during periods of unfavorable industry reports. The model learns to maximize the similarity of positive sample pairs (cosine similarity > 0.85) and minimize the similarity of negative sample pairs (< -0.6). For example, when the environmental policy LSV=2.0, the "elimination risk" channel value in the IEFM for a certain chemical equipment rises to 0.9, matching records of actual low-price disposal, thus validating the fusion effectiveness.

[0138] Using the industry-enhanced feature matrix as input, a deep reinforcement learning-driven Monte Carlo tree search is employed to explore potential value-added paths, and the output path branch tree with probability weights is generated.

[0139] Deep Reinforcement Learning (DRL) frameworks model value-added path exploration as Markov decision processes:

[0140] State (S): Asset-industry joint characteristics of the IEFM matrix.

[0141] Action (A): Value-added strategies (such as renovation leasing, securitization, and technology upgrades).

[0142] Reward (R): Forecasted IRR (Internal Rate of Return), risk reduction, and reduction in disposal cycle.

[0143] Monte Carlo Tree Search (MCTS) is responsible for efficiently finding the optimal solution in the state space.

[0144] Selection: Starting from the root node (initial asset state), select child nodes based on the UCB formula (Upper Confidence Bound) to balance exploration (low-visit nodes) and utilization (high-reward nodes).

[0145] Expansion: When a leaf node is selected, the DRL Policy Network (PN) generates K candidate actions (e.g., K=5) and expands them into new child nodes.

[0146] Simulation: The Value Network (VN) evaluates the future cumulative rewards of new nodes (e.g., the 10-year IRR of a securitization action = 15.2%).

[0147] Backpropagation: Update the path node statistics (number of visits N, average reward Q) in reverse order of the simulation results.

[0148] Example of the path branch tree (PBT) generation process:

[0149] Root node: An old hotel asset (status characteristics: occupancy rate 62%, renovation cost 3M).

[0150] First-level branches: Action A1 (renovate nursing home, probability weight P=0.4), A2 (joint brand operation, P=0.3), A3 (land conversion development, P=0.3).

[0151] Second-level branch: If A1 is selected, the sub-actions include A1a (introducing a medical partner, P=0.6) and A1b (applying for government subsidies, P=0.4).

[0152] The probability weight P of each node is derived by normalizing the number of visits N (P = N_node / ∑N_siblings). The final PBT contains all high-potential paths from the current state to the endgame (such as asset sale, stable operation).

[0153] DRL training employs the Proximal Policy Optimization (PPO) algorithm, with key innovations including:

[0154] Reward Shaping (RS): Introduces domain knowledge constraints, such as a -10% reward if the disposal cycle exceeds 36 months.

[0155] Risk-aware Exploration (RAE): Introducing risk entropy weights into UCB reduces the probability of exploring high volatility paths by 20%.

[0156] Final PBT output example: The probability weight of the path "patent pool licensing → derivative technology development → securitization exit" of a certain patent portfolio is 0.55, and the cumulative reward IRR is 28.7%, making it the main recommended option.

[0157] Based on path branching trees, stochastic differential equations are used to model the interaction between market volatility and asset characteristics, generating a dynamic risk factor field and sensitivity parameter set.

[0158] The core of the stochastic differential equation (SDE) is to characterize uncertainty: dX_t=μ(Xt,t)dt+σ(Xt,t)dW_t, where:

[0159] X_t: Key metrics for asset value or path (such as rental cash flow).

[0160] μ: Drift Term, driven by PBT node reward R (e.g., the expected annual return growth rate of 8% for renovating a nursing home).

[0161] σ: Diffusion Term, influenced by market volatility (fitted by the historical VIX index) and asset-specific risks (such as plant age).

[0162] dW_t: Wiener Process, which simulates random shocks of Brownian motion.

[0163] Generation of the Dynamic Risk Factor Field (DRFF):

[0164] Path discretization: Each path of PBT is sliced ​​by time (e.g., quarter), and each slice is used as the initial state of SDE.

[0165] Monte Carlo Simulation (MCS): For each path, 10,000 random samples are taken, and a different dW_t (following N(0,dt)) is injected each time.

[0166] Risk field construction: Statistical analysis of the distribution of indicators (such as the mean and variance of IRR) for all paths across all time slices. For example, the DRFF layer of a certain logistics asset securitization path in Q3 2024 includes:

[0167] Interest rate risk factor: If the benchmark interest rate rises by 1%, the IRR will decrease by 2.3% (mean).

[0168] Vacancy rate risk factor: When the regional vacancy rate is >15%, the IRR variance increases by 40%.

[0169] The field structure is a three-dimensional grid with time, space, and risk factor dimensions.

[0170] The Sensitivity Parameter Set (SPS) is quantified using Greek letters:

[0171] Delta (Δ): The sensitivity of asset value to underlying market variables (e.g., the Δ of real estate valuation to GDP growth rate = 0.7).

[0172] Vega (ν): The sensitivity of value to volatility (e.g., the sensitivity of patent licensing revenue to the rate of technological iteration, ν = 1.2).

[0173] Theta (Θ): Time decay effect (e.g., the monthly value loss of idle equipment Θ = -0.5%).

[0174] Example of a parameter set: SPS for a certain value-added path = {Δinterest rate = -2.3, Δpolicy = +1.8, νcommodities = 0.6}, which guides subsequent adjustments to the plan.

[0175] The report generation engine inputs dynamic risk factors into the report, creating an asset valuation report with a sandwich structure that decouples the core layer of basic value, the intermediate layer of value-added path, and the outer shell of risk.

[0176] The Report Generation Engine (RGE) employs a sandwich layered architecture:

[0177] Core Value Layer (CVL):

[0178] Input: Risk-free metrics in DRFF (such as asset replacement cost, discounted current net income).

[0179] Generation logic: Deterministic models (such as income-based DCF) calculate the most conservative value.

[0180] Output example: CVL of a certain equipment asset = $1.2 million (based on remaining useful life cash flow).

[0181] Appreciation Path Layer (APL):

[0182] Input: The reward distribution of the Top-K path in PBT (e.g., the IRR distribution of the nursing home renovation [12%, 25%]).

[0183] Generation logic: Bayesian optimization (BO) filters the Pareto optimal path set.

[0184] Output example: Expected added value and probability for the three main promotion paths (brand operation / renovation / securitization).

[0185] Risk Shell Layer (RSL):

[0186] Input: Risk factors and SPS parameters in DRFF.

[0187] Generation logic: Value at Risk (VaR) model (maximum loss of $150,000 at 95% confidence level).

[0188] Output example: The sensitivity matrix shows that "for every 1% increase in interest rates, the value decreases by $23,000".

[0189] Key technologies for layered decoupling:

[0190] Information Distillation Network (IDN): Uses a gating mechanism to separate the information flow across three layers. For example, the CVL layer only allows historically stable data to pass through (filtering nodes with fluctuations >10%).

[0191] Cross-layer Consistency Constraint (CCC): Ensures logical consistency. For example, the path revenue of APL must be higher than the CVL base value, otherwise path correction is triggered (e.g., a certain renovation plan is eliminated because the base rent is too low).

[0192] Example of structured output for the final report:

[0193] Basic Value Core Layer

[0194] Liquidation value: $1 million (based on the assumption of rapid disposal);

[0195] Going concern value: $1.2 million (DCF model, discount rate 8%)

[0196] Value-added path intermediate layer

[0197] Path A: Joint brand operation (probability weight 0.45);

[0198] 3-year median IRR: 18.7% → Increase of $450,000;

[0199] Key move: Bring in an international brand (cost $200,000);

[0200] Path B: Property renovation (probability weight 0.35)

[0201] 5-year median IRR: 22.1% → Added value of $600,000;

[0202] Risk shell

[0203] Key risk factors:

[0204] Interest rate risk (Δ = -2.3): If the interest rate rises by 1%, the value added in Path A will decrease to $400,000;

[0205] Policy risk (ν=1.5): Changes in local subsidies may cause the IRR of path B to fluctuate by ±6%;

[0206] Overall Value at Risk (VaR 95%): Maximum loss $250,000;

[0207] This structure enables the report to combine robustness (CVL), forward-looking perspective (APL), and risk management guidance (RSL).

[0208] This step utilizes AI models to conduct in-depth analysis of the data landscape, not only assessing the current market value of assets but also predicting their potential for appreciation by incorporating industry trends. The model dynamically learns from market changes, identifies the core value drivers and risks of assets, and generates a comprehensive assessment report covering both short-term returns and long-term potential. This provides intelligent analysis that goes beyond traditional valuation methods, helping companies discover hidden asset appreciation opportunities. By quantifying risks and potential, it provides a scientific basis for subsequent plan development, reducing the risk of uninformed decisions.

[0209] S203, Based on the asset valuation report, a multi-objective optimization algorithm is used to generate multiple preliminary value-added scheme sets for different value-added objectives in parallel, wherein each scheme set contains multiple specific service strategy combinations;

[0210] Specifically, the basic value, appreciation path, and risk factors in the asset valuation report can be analyzed and mapped into a three-dimensional target space coordinate system of economic benefits, disposal cycle, and risk entropy value.

[0211] After receiving the asset valuation report generated by the deep valuation model, the system's primary task is to deconstruct its core elements. The report employs a sandwich structure, clearly presenting a layered structure: a core layer of basic value (e.g., equipment residual value assessment, land use right benchmark price), an intermediate layer of value-added paths (e.g., technological upgrade and transformation plans, asset securitization feasibility), and a risk shell layer (e.g., policy change probability, market volatility sensitivity). The analysis engine first identifies key numerical fields in the basic value layer (e.g., "equipment net value: RMB 8.5 million," "land benchmark price: RMB 1200 / square meter"), categorizing them as basic inputs for the Economic Benefit (EB) dimension. Simultaneously, each path in the value-added path layer (e.g., "cooperating with new energy companies to renovate photovoltaic power stations") includes an estimated disposal cycle (DC, typically in months), a parameter directly reflecting the time cost of implementing the plan. The risk outer shell layer requires extracting dynamic field data of risk factors (such as "Regional Policy Risk Index: 0.65" and "Market Demand Volatility: 18%)". This unstructured data will be quantified using the Risk Entropy (RE) model. Risk entropy, drawing on information entropy theory, calculates the negative sum of the logarithms of the products of the probability of occurrence of each risk factor and its degree of influence (the specific calculation process is abstracted), comprehensively reflecting the total uncertainty in the asset appreciation process. For example, an asset involving the disposal of highly polluting equipment may have a high entropy value due to environmental compliance risk factors.

[0212] After element extraction is complete, the system activates the mapping engine. The core function of this engine is to normalize three types of heterogeneous data (numerical EB, temporal DC, and probabilistic RE) into a unified three-dimensional space. Economic benefit (EB), measured in monetary units (e.g., RMB 10,000), is mapped to the X-axis of the three-dimensional coordinate system, and magnitude differences are eliminated through range standardization (e.g., compressing the EB value of all assets to the 0-100 range). Disposal cycle (DC) is mapped to the Y-axis, and its reciprocal (e.g., 1 / DC) is used to convert it into a "disposal efficiency" indicator, ensuring that points with shorter cycles are positioned higher on the Y-axis. Risk entropy (RE) is mapped to the Z-axis; higher entropy values ​​represent greater risk, extending further in the negative Z-axis direction. Thus, each specific asset to be disposed of obtains a unique coordinate point (X, Y, Z) in this three-dimensional objective space coordinate system, for example (EB=75.2, 1 / DC=0.15, RE=-32.8). This coordinate system forms the mathematical basis for subsequent multi-objective optimization. Its design principle is that the ideal solution should pursue high X (economic benefits), high Y (disposal efficiency), and low |Z| (controllable risk).

[0213] To ensure the operational rationality of the coordinate system, a dynamic calibration mechanism is introduced during the mapping process. When there are significant differences in asset types (such as real estate vs. intellectual property), the system calls the preset Industry Weight Template (IWT). For example, for technology-related intangible assets, the IWT will increase the EB weight of technology licensing revenue in the value-added path; for heavy asset equipment, it will focus on the EB calculation of residual value auctions. At the same time, the Real-time Constraint Module (RCM) continuously monitors changes in external conditions (such as a sudden interest rate hike by the central bank) and dynamically adjusts the calculation parameters of RE. All mapping rules and calibration records are stored in the Coordinate Metadata Repository (CMR) to support audit backtracking. The final output three-dimensional coordinate system is not only a data container, but also constructs a mathematical framework for quantitatively evaluating the merits of value-added schemes, laying the foundation for subsequent parallel optimization.

[0214] Simulated annealing algorithm is used in parallel optimization in three-dimensional target space to generate a set of candidate solutions that satisfy the Pareto front.

[0215] Based on the constructed 3D target space, the system initiates the Simulated Annealing Algorithm (SAA) for multi-objective parallel optimization. Inspired by metal heat treatment processes, this algorithm gradually reduces system randomness by controlling the temperature parameter (TP). In the initial stage, a high temperature is set (e.g., initial TP = 1000), allowing the algorithm to accept inferior solutions (i.e., solutions with temporarily deteriorating EB / DC / RE indices) with a higher probability, avoiding getting trapped in local optima. Parallelism is achieved by simultaneously launching multiple annealing threads (AT), each independently searching for a specific incremental objective (e.g., "maximizing EB", "minimizing DC", or "RE ≤ threshold"). The core operation of each thread is neighborhood perturbation (NP): randomly generating a new point (point B) near the current solution's coordinates (e.g., point A), and calculating the merits of the two points using an energy function (EF). The energy function is defined as a weighted form of the three-dimensional Euclidean distance: EF = w1*(EB_A-EB_B) + w2*(1 / DC_A-1 / DC_B) + w3*|RE_A-RE_B|, where the weights w1, w2, and w3 are dynamically allocated according to the current optimization objective.

[0216] The core of the annealing process is the Metropolis Criterion (MC). When the energy EF_B of a new point B is better than that of the current point A (EF_B < EF_A), point B is automatically accepted; when EF_B is worse than EF_A, point B is accepted with probability P = exp(-(EF_B - EF_A) / TP). At high temperatures, the P value is larger, encouraging extensive exploration of the solution space. As the annealing schedule (AS) gradually decreases TP (e.g., TP is multiplied by the cooling coefficient α = 0.95 every 100 iterations), the P value decreases, and the algorithm gradually focuses on high-quality regions. The exploration results of all annealing threads (i.e., the discovered high-quality coordinate points) are merged into the Shared Solution Pool (SSP) in real time. When TP drops to the termination temperature (e.g., TP_final = 0.1), the system extracts the Non-dominated Solution Set (NDSS) from the SSP—meaning no solution can simultaneously outperform it in the EB, DC, and RE dimensions. These solutions constitute the Pareto Frontier (PF), representing the optimal trade-off boundary between economic efficiency, disposal efficiency, and risk control.

[0217] To improve optimization efficiency, the algorithm employs three enhancement strategies: First, Adaptive Neighborhood (AN), which dynamically adjusts the perturbation magnitude based on historical acceptance rates (e.g., expanding the search radius when acceptance rate > 30%); second, Elite Retention (ER), which retains the top K (e.g., K=5) optimal solutions in each generation without participating in the perturbation, preventing the loss of high-quality solutions; and third, Hot Restart (HR), which resets the TP to a higher value and retains the current optimal solution to re-explore when the solution quality has not improved for N consecutive generations (e.g., N=20). The final output CandidateSolution Set (CSS) contains dozens to hundreds of coordinate points on the Pareto front, each implicitly corresponding to an asset disposal strategy direction (e.g., "rapid auction", "deep transformation", "split and reorganization"), providing input for the next strategy combination.

[0218] The candidate schemes are input into the genetic algorithm, and the service strategy is combined through crossover and mutation operations, and the strategy combination chromosome group is output.

[0219] Each coordinate point in the candidate solution set (CSS) only represents the target direction and needs to be transformed into a specific executable service strategy combination. The system models this process as an evolutionary problem using a genetic algorithm (GA). First, chromosome encoding (CE) is performed: each candidate solution is decoded into a chromosome (CHR), whose gene locus (GL) corresponds to an optional atomic service strategy. The strategy library contains five categories: ① Value enhancement (e.g., technological transformation, brand reshaping); ② Circulation acceleration (e.g., split auction, lease buyback); ③ Risk hedging (e.g., insurance purchase, performance-based agreements); ④ Policy utilization (e.g., tax incentive application, industry fund connection); ⑤ Collaborative operation (e.g., supply chain integration, managed operation). Each chromosome consists of M (e.g., M=8) gene loci, and each gene locus selects a specific strategy from its respective strategy class (e.g., gene locus 1 selects "equipment refurbishment" instead of "technology licensing").

[0220] The initialization generates the first generation population (FGP), containing N chromosomes (e.g., N=100). Each chromosome inherits its target tendency from the candidate solution set (CSS) through Roulette Wheel Selection (RWS)—high EB solutions are more likely to include value-enhancing strategies, while low DC solutions tend to favor circulation-accelerating strategies. Then, a crossover operation (CO) is initiated: chromosomes are randomly paired (e.g., CHR_A and CHR_B), and segments are exchanged at random breakpoints (e.g., after the 3rd gene position) to generate offspring chromosomes. For example, crossover of CHR_A=[Technical Improvement, Split, Insurance] and CHR_B=[Brand, Leasing, Gambling] may produce [Technical Improvement, Leasing, Insurance] and [Brand, Split, Gambling]. Next, a mutation operation (MO) is performed: the strategy of a certain gene position is randomly changed with a small probability (e.g., mutation rate PM=0.05) (e.g., mutating "split auction" to "whole transfer"). Crossover promotes strategy recombination and innovation, while mutation introduces diversity and prevents premature convergence.

[0221] Each generation of the population undergoes a fitness evaluation (FE). The fitness function F is defined as a weighted sum of three-dimensional objectives: F = β1*EB_norm + β2*(1 / DC)_norm + β3*(1-RE_norm), where β is the weight and norm represents the normalized value. A strategy conflict detector (SCD) is also introduced. For example, if both "long-term trusteeship" (requiring continuous investment) and "rapid auction" (requiring immediate termination of operation) are detected on the same chromosome, a fitness penalty is applied. After evaluation, Tournament Selection (TS) is used to retain superior individuals: K chromosomes (e.g., K=3) are randomly selected, and the individuals with the highest fitness are retained for the next generation. After G generations (e.g., G=50), the output strategy combination chromosome group (SCCG) is generated, containing hundreds of complete service strategy chains optimized by natural selection.

[0222] Apply a corporate compliance constraint filter to the strategy portfolio chromosome group to generate an initial set of value-added solutions with fitness scores.

[0223] The chromosome population (SCCG) output by the genetic algorithm needs to be validated for legality through a Compliance Constraint Filter (CCF). The filter incorporates a three-layer constraint network: the first layer is a Mandatory Regulation Base (MRB), for example, the "Measures for the Supervision and Management of Transactions of State-owned Assets of Enterprises" stipulates that asset transfers exceeding 5 million yuan must be listed for trading; the second layer is Industry Self-regulation Rules (ISR), such as the requirement that the disposal of chemical equipment must comply with environmental dismantling standards; the third layer is Enterprise Specific Redlines (ESR), such as shareholder agreements prohibiting the licensing of core patents. After each chromosome is parsed into a strategy sequence, it is pattern matched (PM) with the constraint base. For example, if a "direct agreement transfer" strategy is detected and the asset valuation exceeds the threshold, an MRB warning is triggered; if an "on-site equipment modification" strategy is found but lacks an environmental impact assessment document, an ISR interception is triggered.

[0224] The constraint handling employs a Dynamic Relaxation Mechanism (DRM). For minor conflicts (such as missing non-critical approval documents), the system initiates an Auto-repair Program (ARP): generating a commitment letter via the electronic signature interface or inserting a "reissue license" sub-policy. For major conflicts (such as violations of ESR), the chromosome is directly eliminated. Verified chromosomes enter the Fitness Refinement (FR) stage: based on the original fitness F, a Compliance Score (CS) and Implementation Complexity (IC) are added for correction. The Compliance Score (CS) is calculated based on the degree of conflict remediation (1.0 for a perfect match, 0.6~0.9 after remediation); the Implementation Complexity (IC) is calculated using a policy dependency graph (e.g., "cross-border auction" requires coordination with 5 departments including customs / taxation) to determine the topological order depth. The final fitness is adjusted to F_final = F × CS / log(IC+1).

[0225] The calculated chromosome populations are sorted in descending order of F_final, and the Solution Aggregation Engine (SAE) generates a Preliminary Value-added Solution Set (PVSS). Each solution contains three core modules: ① Strategy Chain (SC): Service strategies arranged in execution order, such as "equipment evaluation → package splitting → online auction → payment settlement"; ② 3D Objective Values ​​(3OV): Estimated EB / DC / RE values; ③ Fitness Score (FS): F_final value (range 0~100). The solution set adopts a Diversity Protection Strategy (DPS), which forcibly retains the top 10 FS solutions, the most innovative (largest crossover mutation depth), and the most robust (lowest RE) solutions. The final output includes dozens of preliminary value-added solutions that meet compliance requirements, have clear quantifiable objectives, and include implementation paths, providing raw materials for subsequent intelligent matching.

[0226] This step, based on multi-dimensional indicators (such as returns, cycle, and risk) in the assessment report, uses an optimization algorithm to generate multiple possible value-added paths. The algorithm simultaneously considers objectives such as economic benefits and feasibility, outputting differentiated strategy combinations, such as asset restructuring, leasing operations, or securitization, overcoming the limitations of a single strategy and providing diversified value-added options. Parallel optimization ensures that the solutions cover the needs of different enterprises, laying the foundation for customized services.

[0227] S204. Input the preliminary value-added solution set into the intelligent matching engine. Based on the preset enterprise profile and real-time constraints, verify the feasibility and adaptability of the solutions, and output the optimal customized value-added service solution package.

[0228] Specifically, enterprise profile parameters can be injected into a pre-set digital twin engine to build a dynamic simulation model of the enterprise that includes at least financial structure and production capacity flexibility;

[0229] Enterprise profile parameter injection process

[0230] Enterprise Profile Parameters (EPP) are structured data extracted from enterprise ERP, SCM, and other systems, containing core indicators:

[0231] Financial Structure Parameters (FSPs): such as Asset-Liability Ratio (ALR), Cash Flow Coverage Ratio (CFCR), and Current Ratio (CR).

[0232] Capacity Elasticity Parameters (CEP): such as Line Switch Response Time (LSRT) and Equipment Utilization Fluctuation Threshold (EUFT).

[0233] The EPP is imported into the Digital Twin Engine (DTE) via the Parameter Injection Interface (PII). This interface uses the OAuth 2.0 protocol (Open Authorization Protocol) for secure authentication, ensuring real-time data synchronization. For example, when the injected debt-to-equity ratio (ALR) is 0.45, the engine automatically maps it to the capital and liability sub-model of the financial simulation module.

[0234] Dynamic simulation model construction mechanism

[0235] The Digital Twin Engine (DTE) builds a three-layer simulation architecture based on enterprise profile parameters:

[0236] Financial Structure Layer: The System Dynamics Model (SDM) is used to simulate cash flow. For example, by inputting a Cash Flow Coverage Ratio (CFCR) of 3.2 (target value ≥ 2.5), the model automatically generates a causal loop diagram showing the impact of accounts receivable turnover on cash flow.

[0237] Capacity elasticity layer: The production line is modeled using Discrete Event Simulation (DES). For example, a threshold for equipment utilization fluctuation, EUFT=15%, is set. When market demand changes abruptly, the model calculates the optimal response path for expanding or reducing production capacity.

[0238] Environmental Coupling Layer: Integrates external economic indicator APIs (such as GDP growth rate and industry prosperity index) and uses time series analysis (TSA) to predict the impact of market shocks on enterprise resilience.

[0239] Real-time data-driven and verification

[0240] The model connects to the enterprise's real-time database via the OPC UA protocol (Open Process Control Unified Architecture). For example, it collects production line sensor data every 5 minutes to update the capacity elasticity parameter CEP, ensuring that the simulation is synchronized with the physical world.

[0241] The Historical Backtesting Validation (HBV) method is adopted: the model is fed back with the enterprise's operational data from the past 12 months, requiring a financial forecast error rate of ≤5% and a capacity scheduling accuracy of ≥92%. Once these standards are met, the model is activated and put into operation.

[0242] The system uses Apache Flink to process constraint stream data containing at least liquidity information in real time, generating a dynamic set of constraint boundary conditions.

[0243] Constrained Flow Data Acquisition and Preprocessing

[0244] Data sources for Funding Liquidity Constraints (FLC) include:

[0245] Real-time account balance (RAB) of a bank account.

[0246] Short-term Debt Maturity Calendar (SDMC);

[0247] Credit Line Utilization Rate (CLUR);

[0248] Access heterogeneous data sources through Apache Flink's DataStream API. For example, obtain cross-border fund flow data from the SWIFT system (Society for Worldwide Interbank Financial Telecommunication) and aggregate account changes at a 1-minute granularity using the TumblingWindow function.

[0249] Real-time computing engine operation logic

[0250] The Flink engine executes a three-layer processing pipeline:

[0251] Constraint Extraction Layer: Employs CEP (Complex Event Processing) rules to detect urgent constraints. For example, a "financing warning" event is triggered when CLUR > 85%.

[0252] Boundary Quantization Layer: Dynamic boundaries are calculated using sliding window statistics. For example, based on the past 30 days of RAB data, the safe funding threshold (SFT = average daily balance × 70%) is calculated on a rolling basis.

[0253] Conditional Synthesis Layer: Multiple types of constraints are associated through a Constraint Relation Graph (CRG). For example, the debt of 5 million yuan due next week in SDMC is associated with the current RAB = 8 million yuan to generate the boundary condition "minimum weekly cash holdings = 6 million yuan".

[0254] Dynamic constraint set generation and update

[0255] The output is a Dynamic Constraint Boundary Set (DCBS), which contains two core types of data:

[0256] Hard Boundaries (HB): such as "the quarterly debt-to-equity ratio (ALR) must not exceed 0.6";

[0257] Flexible Boundaries (FB): such as "capacity utilization is allowed to fluctuate within the range of [65%, 90%]";

[0258] DCBS is pushed to downstream modules in real time via a Kafka message queue (a distributed stream processing platform), with an update frequency of up to 200ms / time. When the CLUR is detected to exceed the threshold, the boundary condition is immediately recalculated.

[0259] The initial set of value-added solutions is input into the enterprise's dynamic simulation model. The implementation process of the solutions is simulated in a neural radiation field, and a feasibility heatmap with failure probability is output.

[0260] Interaction mechanism between the scheme and the simulation model

[0261] The Preliminary Value-added Solution Set (PVSS) contains a variety of strategy combinations, such as:

[0262] Option A: Asset securitization (issue size of 100 million, fee rate of 2.5%) + capacity leasing (leasing 30% of idle equipment);

[0263] Option B: Strategic restructuring (introducing strategic investors to hold 15% of the shares) + technology upgrade (investment of 20 million).

[0264] The solution parser (SP) converts text solutions into machine-executable instructions. For example, "issuance scale of 100 million" is mapped to the "new liabilities of 100 million" parameter in the financial module.

[0265] Core technology for simulating neural radiation fields

[0266] Neural radiance fields (NeRF) are here extended to multidimensional decision fields:

[0267] Spatial dimension: mapping the enterprise's physical architecture (such as factory location, office network);

[0268] Time dimension: Load the historical processing record time vector (TDV);

[0269] Logical dimension: Simulation rules for embedded digital twin models (DTE).

[0270] The simulation process employs Monte Carlo Path Tracing (MCPT):

[0271] For each scheme, 1000 implementation paths are randomly generated;

[0272] Each path is infused with market volatility parameters (such as interest rate ±1.5%).

[0273] The dynamic simulation model (DTE) calculates financial indicators and production capacity status on a daily basis.

[0274] Feasibility heatmap generation logic

[0275] Output a heatmap with failure probability (Feasibility Heatmap with Failure Probability, FHF):

[0276] X-axis: Time span (e.g., 0 - 36 months);

[0277] Y-axis: Key performance indicators (e.g., return on equity ROE, capacity utilization CU);

[0278] Chroma value: Failure probability (Failure Probability, FP), dark red indicates FP > 30%.

[0279] For example, when simulating Scenario A, it is found that in the 18th month, due to the increase in interest rates, the FP suddenly rises to 45%, which is shown as a red patch in the heatmap. The heatmap is rendered as a three-dimensional visual matrix through OpenGL (Open Graphics Library).

[0280] Based on the feasibility heatmap and the dynamic constraint boundary condition set, an adaptive weighted algorithm using fuzzy logic control is adopted to dynamically allocate weights to the three-dimensional objectives of the scenario, generating an optimized scenario pool;

[0281] Multi-source data fusion analysis

[0282] The feasibility heatmap (FHF) and the dynamic constraint boundary set (DCBS) are aligned through a constraint-risk coupler (Constraint-Risk Coupler, CRC):

[0283] Extract the high failure probability areas (FP > 25%) in the heatmap;

[0284] Compare the constraint boundary conditions (e.g., mark as "high-risk area" when FP > 25% and cash balance < SFT);

[0285] Generate a risk-constraint correlation matrix (Risk-Constraint Correlation Matrix, RCCM).

[0286] The core operation of fuzzy logic control

[0287] The adaptive weighting algorithm (Adaptive Weighting Algorithm, AWA) consists of three stages:

[0288] Fuzzy input: The three-dimensional objectives of the solution (Economic Benefit, Dur of disposal, and Risk) are converted into linguistic variables. For example, "high / medium / low" economic benefits correspond to triangular membership functions.

[0289] Rule-based reasoning: Includes 200+ pre-defined fuzzy rules (FR). For example:

[0290] If risk entropy is high and liquidity is tight, then the economic weighting is reduced by 30%.

[0291] Adaptive adjustment: Adjusts rule strength in real time based on boundary conditions in DCBS. If CLUR > 90% is detected, the weight of risk control rules is automatically strengthened.

[0292] Dynamic weight allocation and scheme optimization

[0293] Perform weighted TOPSIS (Technique for Order Preference by Similarity to IdealSolution).

[0294] Calculate the dynamic weights of the three-dimensional objectives for each option (e.g., economic: 0.5→0.4, risk: 0.3→0.4).

[0295] Priority of schemes is reordered based on weights.

[0296] Generate an Optimized Solution Pool (OSP):

[0297] Retain 80% of the original PVSS scheme;

[0298] Eliminate all schemes with FP > 40% in the heatmap;

[0299] Add 5% of hybrid strategy options (such as a cross combination of options A and B).

[0300] Each scheme is labeled with a dynamic weight vector (e.g., [Economic weight: 0.45, Cycle weight: 0.25, Risk weight: 0.3]).

[0301] The optimization solution pool is topologically compressed, redundant strategy combinations are eliminated, and finally a lightweight customized value-added service solution package is output.

[0302] Scheme topology modeling and analysis

[0303] The optimization pool (OSP) is constructed as a strategy dependency graph (SDG):

[0304] Node: Individual service strategy (such as "equipment leasing", "patent licensing");

[0305] Edge: The strength of the association between strategies (calculated using the covariance matrix based on historical data).

[0306] The Community Detection Algorithm (CDA) is used to identify policy clusters.

[0307] Discover strongly correlated clusters (such as the combination of "asset securitization + credit enhancement");

[0308] Mark weakly associated isolated nodes (such as "land conversion for development" which require independent evaluation).

[0309] Redundancy Removal and Structural Compression

[0310] Perform triple filtering:

[0311] Redundancy filtering: Rank strategies within the same cluster based on their cost-benefit ratio (CBR), retaining the top 30% of strategies with the highest CBR. For example, eliminate options with a yield of less than 8% in equipment leasing.

[0312] Risk hedging compression: The minimum spanning tree (MST) algorithm is used to simplify risk hedging paths. For example, two similar exchange rate hedging strategies can be merged.

[0313] Constraint conflict resolution: A constraint satisfaction problem solver (CSPS) is used to detect conflicting strategies. For example, if both "expanding production capacity" and "selling equipment" exist simultaneously, an alarm will be triggered.

[0314] Lightweight solution package generation

[0315] Output a customized value-added service package (CVSP):

[0316] Core layer: 3-5 mandatory strategies (such as "patent portfolio licensing"), which shorten the critical path by 50% after topology compression;

[0317] Flexible layer: Dynamically pluggable modules (such as "selecting Class A or Class B financing instruments");

[0318] Risk isolation chamber: Embedded hedging solution (such as automatically activating interest rate swaps when the interest rate rises by more than 1%).

[0319] The final solution package size is reduced by 60% compared to the input, the inter-strategy dependencies are simplified from a mesh structure to a star topology, and it is delivered by encapsulation using JSON-LD (Associative Data Format).

[0320] This step dynamically filters solutions based on the company's actual situation (such as cash flow and compliance requirements). By simulating the implementation effects of solutions using digital twin technology, options that do not meet the constraints are eliminated. The final output is a lightweight solution package highly matched to the company's needs, ensuring the feasibility and adaptability of the solution and avoiding strategies that are theoretically optimal but practically infeasible. Customized output reduces the company's decision-making costs and improves asset disposal efficiency.

[0321] As can be seen, by receiving multi-source heterogeneous data on enterprise special assets, a data map of enterprise special assets containing asset attributes, market environment, and historical disposal records can be obtained. This data map is then input into a pre-trained AI-based deep valuation model to generate an asset valuation report containing basic value, potential value-added paths, and risk factors. Based on the asset valuation report, a multi-objective optimization algorithm is used to generate multiple preliminary value-added solution sets targeting different value-added objectives in parallel. These preliminary value-added solution sets are then input into an intelligent matching engine, which verifies the feasibility and suitability of the solutions based on a pre-set enterprise profile and real-time constraints. The result is the output of an optimized, customized value-added service solution package, thereby improving the accuracy of asset value analysis and the efficiency of potential discovery, and achieving dynamic optimization of asset value-added paths and controllable risks.

[0322] Another embodiment of the present invention provides an AI-based diversified value-added service system for enterprise special assets, see [link to relevant documentation]. Figure 3 The system may include:

[0323] The receiving module 301 is used to receive multi-source heterogeneous data of enterprise special assets, and clean, associate and structure the multi-source heterogeneous data based on preset data fusion rules to obtain a data map of enterprise special assets containing asset attributes, market environment and historical disposal records.

[0324] The parsing module 302 is used to input the enterprise's special asset data map into a pre-trained AI-based deep value assessment model, and combine real-time market dynamics and industry trends to perform multi-dimensional value analysis and potential mining, and generate an asset value assessment report that includes basic value, potential value-added paths and risk factors.

[0325] The generation module 303 is used to generate multiple preliminary value-added scheme sets for different value-added objectives in parallel using a multi-objective optimization algorithm based on the asset valuation report. Each scheme set contains multiple specific service strategy combinations.

[0326] The output module 304 is used to input the preliminary value-added solution set into the intelligent matching engine, and perform feasibility verification and adaptability screening of the solutions based on the preset enterprise profile and real-time constraints, and output the optimal customized value-added service solution package.

[0327] 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.

[0328] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps:

[0329] S201, Receive multi-source heterogeneous data of enterprise special assets, and clean, associate and structure the multi-source heterogeneous data based on preset data fusion rules to obtain a data map of enterprise special assets containing asset attributes, market environment and historical disposal records;

[0330] S202, input the enterprise's special asset data map into a pre-trained AI-based deep value assessment model, combine real-time market dynamics and industry trends to conduct multi-dimensional value analysis and potential mining, and generate an asset value assessment report that includes basic value, potential value-added paths and risk factors.

[0331] S203, Based on the asset valuation report, a multi-objective optimization algorithm is used to generate multiple preliminary value-added scheme sets for different value-added objectives in parallel, wherein each scheme set contains multiple specific service strategy combinations;

[0332] S204. Input the preliminary value-added solution set into the intelligent matching engine. Based on the preset enterprise profile and real-time constraints, verify the feasibility and adaptability of the solutions, and output the optimal customized value-added service solution package.

[0333] 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.

[0334] 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.

[0335] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0336] S201, Receive multi-source heterogeneous data of enterprise special assets, and clean, associate and structure the multi-source heterogeneous data based on preset data fusion rules to obtain a data map of enterprise special assets containing asset attributes, market environment and historical disposal records;

[0337] S202, input the enterprise's special asset data map into a pre-trained AI-based deep value assessment model, combine real-time market dynamics and industry trends to conduct multi-dimensional value analysis and potential mining, and generate an asset value assessment report that includes basic value, potential value-added paths and risk factors.

[0338] S203, Based on the asset valuation report, a multi-objective optimization algorithm is used to generate multiple preliminary value-added scheme sets for different value-added objectives in parallel, wherein each scheme set contains multiple specific service strategy combinations;

[0339] S204. Input the preliminary value-added solution set into the intelligent matching engine. Based on the preset enterprise profile and real-time constraints, verify the feasibility and adaptability of the solutions, and output the optimal customized value-added service solution package.

[0340] 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 diversified value-added services for enterprise special assets based on AI, characterized in that, The method includes: Receive multi-source heterogeneous data on special assets of enterprises, and clean, associate and structure the multi-source heterogeneous data based on preset data fusion rules to obtain a data map of special assets of enterprises that includes asset attributes, market environment and historical disposal records; The enterprise's special asset data map is input into a pre-trained AI-based deep valuation model. Combined with real-time market dynamics and industry trends, multi-dimensional value analysis and potential mining are carried out to generate an asset valuation report that includes basic value, potential value-added paths and risk factors. Based on the asset valuation report, a multi-objective optimization algorithm is used to generate multiple preliminary value-added scheme sets for different value-added objectives in parallel. Each scheme set contains multiple specific service strategy combinations. The initial set of value-added solutions is input into the intelligent matching engine. Based on the preset enterprise profile and real-time constraints, the feasibility of the solutions is verified and the adaptability is screened, and the optimal customized value-added service solution package is output.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous data on the received enterprise's special assets is cleaned, correlated, and structured based on preset data fusion rules to obtain a data map of the enterprise's special assets containing asset attributes, market environment, and historical disposal records, including: By using knowledge graph ontology mapping technology, multi-source heterogeneous data of enterprise special assets are mapped into a unified topology structure, and an initial data topology graph with weight labels is output. Based on the initial data topology, an adaptive sliding window is used to detect outliers in time-series data. Combined with an industry rule base, contradictory fields are automatically corrected to generate a cleaned spatiotemporally aligned data stream. The spatiotemporally aligned data stream is input into the multi-head attention association engine to identify implicit association rules between asset attributes and market environment, and output an enhanced data network with cross-modal association factors; By leveraging an enhanced data network-driven neural radiation field model, a data map of enterprise-specific assets containing three-dimensional spatiotemporal relationships is generated, in which historical disposal records are encoded as traceable time-dimensional vectors.

3. The method according to claim 2, characterized in that, The process involves inputting the enterprise's special asset data map into a pre-trained AI-based deep valuation model, combining real-time market dynamics and industry trends to perform multi-dimensional value analysis and potential mining, and generating an asset valuation report that includes basic value, potential value-added paths, and risk factors, including: Input the enterprise's special asset data map into the spatiotemporal graph convolutional network, extract the dynamic evolution features of asset attributes and the spatiotemporal coupling features of the market environment, and output a multidimensional feature tensor. The real-time industry trend data is compressed into a latent space vector by a variational encoder, and then the latent space vector is fused with a multidimensional feature tensor by tensor product to generate an industry-enhanced feature matrix. Using the industry-enhanced feature matrix as input, a deep reinforcement learning-driven Monte Carlo tree search is employed to explore potential value-added paths, and the output path branch tree with probability weights is generated. Based on path branching trees, stochastic differential equations are used to model the interaction between market volatility and asset characteristics, generating a dynamic risk factor field and sensitivity parameter set. The report generation engine inputs dynamic risk factors into the report, creating an asset valuation report with a sandwich structure that decouples the core layer of basic value, the intermediate layer of value-added path, and the outer shell of risk.

4. The method according to claim 3, characterized in that, Based on the asset valuation report, a multi-objective optimization algorithm is used to generate multiple preliminary value-added solution sets for different value-added objectives in parallel. Each solution set contains multiple specific service strategy combinations, including: The basic value, appreciation path, and risk factors in the asset valuation report are analyzed and mapped into a three-dimensional target space coordinate system of economic benefits, disposal cycle, and risk entropy value. Simulated annealing algorithm is used in parallel optimization in three-dimensional target space to generate a set of candidate solutions that satisfy the Pareto front. The candidate schemes are input into the genetic algorithm, and the service strategies are combined through crossover and mutation operations, and the combined chromosome population of the strategies is output. Apply a corporate compliance constraint filter to the strategy portfolio chromosome group to generate an initial set of value-added solutions with fitness scores.

5. The method according to claim 4, characterized in that, The process involves inputting the initial set of value-added solutions into an intelligent matching engine. Based on a preset enterprise profile and real-time constraints, the engine verifies the feasibility and adaptability of the solutions, and outputs the optimal customized value-added service solution package, including: By injecting enterprise profile parameters into a pre-set digital twin engine, a dynamic simulation model of the enterprise is constructed, which includes at least financial structure and production capacity flexibility. The system uses Apache Flink to process constraint stream data containing at least liquidity information in real time, generating a dynamic set of constraint boundary conditions. The initial set of value-added solutions is input into the enterprise's dynamic simulation model, and the implementation process of the solutions is simulated in the neural radiation field, outputting a feasibility heatmap with failure probability. Based on the feasibility heatmap and dynamic constraint boundary condition set, an adaptive weighted algorithm with fuzzy logic control is used to dynamically assign weights to the three-dimensional objectives of the scheme and generate an optimized scheme pool. The optimization solution pool is topologically compressed, redundant strategy combinations are eliminated, and finally a lightweight customized value-added service solution package is output.

6. An AI-based diversified value-added service system for enterprise special assets, characterized in that, The system includes: The receiving module is used to receive multi-source heterogeneous data of enterprise special assets, and clean, associate and structure the multi-source heterogeneous data based on preset data fusion rules to obtain a data map of enterprise special assets containing asset attributes, market environment and historical disposal records. The analysis module is used to input the enterprise's special asset data map into a pre-trained AI-based deep value assessment model, and combine it with real-time market dynamics and industry trends to perform multi-dimensional value analysis and potential mining, generating an asset value assessment report that includes basic value, potential value-added paths and risk factors. The generation module is used to generate multiple preliminary value-added scheme sets for different value-added objectives in parallel using a multi-objective optimization algorithm based on the asset valuation report. Each scheme set contains multiple specific service strategy combinations. The output module is used to input the preliminary value-added solution set into the intelligent matching engine, and based on the preset enterprise profile and real-time constraints, to perform feasibility verification and adaptability screening of the solutions, and output the optimal customized value-added service solution package.

7. The system according to claim 6, characterized in that, The receiving module is specifically used for: By using knowledge graph ontology mapping technology, multi-source heterogeneous data of enterprise special assets are mapped into a unified topology structure, and an initial data topology graph with weight labels is output. Based on the initial data topology, an adaptive sliding window is used to detect outliers in time-series data. Combined with an industry rule base, contradictory fields are automatically corrected to generate a cleaned spatiotemporally aligned data stream. The spatiotemporally aligned data stream is input into the multi-head attention association engine to identify implicit association rules between asset attributes and market environment, and output an enhanced data network with cross-modal association factors; By leveraging an enhanced data network-driven neural radiation field model, a data map of enterprise-specific assets containing three-dimensional spatiotemporal relationships is generated, in which historical disposal records are encoded as traceable time-dimensional vectors.

8. The system according to claim 7, characterized in that, The parsing module is specifically used for: Input the enterprise's special asset data map into the spatiotemporal graph convolutional network, extract the dynamic evolution features of asset attributes and the spatiotemporal coupling features of the market environment, and output a multidimensional feature tensor. The real-time industry trend data is compressed into a latent space vector by a variational encoder, and then the latent space vector is fused with a multidimensional feature tensor by tensor product to generate an industry-enhanced feature matrix. Using the industry-enhanced feature matrix as input, a deep reinforcement learning-driven Monte Carlo tree search is employed to explore potential value-added paths, and the output path branch tree with probability weights is generated. Based on path branching trees, stochastic differential equations are used to model the interaction between market volatility and asset characteristics, generating a dynamic risk factor field and sensitivity parameter set. The report generation engine inputs dynamic risk factors into the report, creating an asset valuation report with a sandwich structure that decouples the core layer of basic value, the intermediate layer of value-added path, and the outer shell of risk.

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 1-4 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 1-4.

Images