Enterprise financial health degree AI diagnosis method

Abstract

The invention relates to the technical field of financial science and technology and data processing, and discloses an enterprise financial health degree AI diagnosis method, which comprises the following steps: acquiring multi-dimensional heterogeneous data including enterprise finance, strategic modes and behavior preferences, and constructing a multi-dimensional dynamic state vector for each entity in a commercial ecosystem; constructing a business entity relationship network and determining an asymmetric interaction relationship between entities; applying a preset disturbance event, and generating a state evolution path of the target enterprise in a future period of time through coupling dynamics simulation; and extracting an evolution path, comprehensively comparing the evolution path with a plurality of preset health degree references, combining the path similarity based on dynamic time warping and the terminal state distance, and performing weighted summation to determine a health degree category closest to the evolution process and result of the target enterprise as a final diagnosis result. The method can simulate the future performance of an enterprise in a specific pressure situation, reveals the internal risk and development potential of the enterprise, and improves the depth and accuracy of financial health diagnosis.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and data processing technology, specifically to an AI-based diagnostic method for corporate financial health. Background Technology

[0002] Accurate diagnosis of a company's financial health is a core element of modern financial risk management, credit approval, and investment decision-making. Currently, the mainstream method for assessing a company's financial condition typically relies on the analysis of its historical financial statements, using a series of financial ratios to construct credit rating models or financial early warning systems.

[0003] However, existing technologies have significant limitations in providing profound and forward-looking diagnostics. On the one hand, these methods often analyze enterprises as isolated entities, or even when considering upstream and downstream relationships, they struggle to accurately quantify and model the complex, multidimensional, and often unequal interactions between entities within a business ecosystem. This results in incomplete underlying models upon which diagnostic analysis relies. On the other hand, traditional assessments are essentially static snapshots based on historical data. While they reflect a company's past operating conditions and current state, they cannot effectively simulate and predict the dynamic response processes and evolutionary paths of enterprises when faced with specific market shocks or supply chain disruptions in the future. Furthermore, even when some models attempt future predictions, their assessment dimensions are relatively singular, typically focusing on the final financial outcome (such as default), while ignoring the stability and resilience of the enterprise's trajectory during the response to shocks. Consequently, they cannot provide a comprehensive and quantitative assessment of the enterprise's true resilience and systemic risk. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an AI-based diagnostic method for corporate financial health. This method solves the problem that existing financial risk assessment methods are typically based on static financial statements, making it difficult to capture the dynamic interactions between entities in the business environment and effectively predict the trajectory of future financial status changes of enterprises under specific disturbances.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-based diagnostic method for enterprise financial health, comprising the following steps:

[0006] S1. Acquire multi-source heterogeneous data, and define the target enterprise and related business entities as multiple business entities. Determine the multi-dimensional dynamic state vector of each business entity. The multi-dimensional dynamic state vector is used to characterize the internal state and behavioral tendencies of the business entity.

[0007] S2. Construct a business entity relationship network that includes all the business entities. In the business entity relationship network, based on the multidimensional dynamic state vector of each business entity, determine the asymmetric interaction relationship between any two business entities.

[0008] S3. In response to a preset disturbance event, perform coupled dynamic simulation in the business entity relationship network, and apply the asymmetric interaction relationship determined in step S2 to iteratively update the multidimensional dynamic state vector of each business entity to obtain a series of multidimensional dynamic state vectors for each business entity.

[0009] S4. Extract a series of multidimensional dynamic state vectors of the target enterprise from a series of multidimensional dynamic state vectors of each of the business entities obtained in step S3.

[0010] S5. Determine a series of multidimensional dynamic state vectors of the target enterprise obtained in step S4 as a future state evolution path of the target enterprise, and determine the financial health of the target enterprise based on the future state evolution path.

[0011] Preferably, the multi-source heterogeneous data includes: financial data, business relationship data, and dynamic behavior data of the target enterprise and the associated business entities.

[0012] Preferably, in step S1, the step of determining the multidimensional dynamic state vector of each business entity specifically involves: determining the internal financial state vector, strategic pattern vector, and behavioral preference vector of the business entity, and combining the internal financial state vector, the strategic pattern vector, and the behavioral preference vector to form the multidimensional dynamic state vector. In a specific embodiment, the multidimensional dynamic state vector of business entity i at time t... Represented as:

[0013]

[0014] in, This represents the multidimensional dynamic state vector of business entity i at time t; This represents the internal financial state vector of business entity i at time t; This represents the strategic pattern vector of business entity i at time t; This represents the behavioral preference vector of business entity i at time t.

[0015] Preferably, in step S1, the step of determining the internal financial state vector, the strategic pattern vector, and the behavioral preference vector specifically involves: processing the financial data to determine the internal financial state vector; and using a natural language processing model and a pattern recognition model to process the dynamic behavioral data to determine the strategic pattern vector and the behavioral preference vector.

[0016] Preferably, in step S2, the asymmetric interaction relationship is further determined by the relationship type and strength between the business entities. The asymmetric interaction influence A of business entity j on business entity i. ji This can be determined by a pre-defined interaction function, an example of which is as follows:

[0017]

[0018] It should be noted that, since the effects are mutual but asymmetrical, A is usually... ji ≠A ij .

[0019] Among them, A ji This represents the interaction effect exerted by business entity j on business entity i; f(·) represents a pre-defined interaction function used to calculate the interaction between entities; The multidimensional dynamic state vector of the receiving business entity i at time t; R represents the multidimensional dynamic state vector of the business entity j that exerts the action at time t; ji This indicates the type of relationship from business entity j to business entity i, such as supplier-customer relationship, competitor relationship, or investment relationship; W ji This indicates the strength of the relationship from business entity j to business entity i.

[0020] Preferably, the relationship type is determined based on the business relationship data, and the relationship strength is quantified based on the transaction amount and transaction frequency in the financial data.

[0021] Preferably, step S2 further includes: monitoring a preset business event; when the preset business event occurs, identifying a business entity directly associated with the preset business event, and then, based on the multidimensional dynamic state vector of the directly associated business entity, reconfirming all the asymmetric interaction relationships involving the business entity.

[0022] Preferably, the preset disturbance event is selected from a group of event types including node disturbance, relationship disturbance, and macroscopic environment disturbance.

[0023] Preferably, in step S3, the step of iteratively updating the multidimensional dynamic state vector of each of the business entities specifically involves: at the initial moment of the dynamic simulation, adjusting the multidimensional dynamic state vectors of the business entities related to the preset disturbance event to set the initial conditions for the dynamic simulation; and at each time step starting from the initial moment, first calculating the net effect of the asymmetric interaction relationship acting on each of the business entities. For business entity i, the net effect it experiences at time t is... The calculation method is as follows:

[0024]

[0025] Among them, A ji This represents the interaction effect exerted by business entity j on business entity i; This indicates that the following terms are summed; N represents the total number of business entities.

[0026] Then, using a preset evolution function, the multidimensional dynamic state vector for the next time step is calculated based on the current multidimensional dynamic state vector of the business entity and the net effect. Its evolution process can be described by the following equation:

[0027]

[0028] in, It represents the net effect of the action on business entity i at time t, and is the vector sum of the effects of all other entities on business entity i in the network; This represents the multidimensional dynamic state vector of business entity i at time t; G(·) represents the multidimensional dynamic state vector of business entity i at the next time step t+Δt; G(·) represents a preset evolution function used to calculate the next state based on the current state and net effect; Δt represents the time step of the dynamic simulation.

[0029] Finally, the multidimensional dynamic state vector of the next time step is set as the current multidimensional dynamic state vector of the business entity in the subsequent time steps.

[0030] Preferably, in step S5, the step of determining the financial health of the target enterprise based on the future state evolution path specifically involves: determining a series of internal financial state vectors from the future state evolution path, and determining the terminal state of the series of internal financial state vectors; comparing the series of internal financial state vectors and the terminal state with a preset health benchmark to determine the financial health of the target enterprise.

[0031] This invention provides an AI-based diagnostic method for corporate financial health. It offers the following advantages:

[0032] 1. This invention determines a multi-dimensional dynamic state vector that characterizes the internal state and behavioral tendencies of a business entity and establishes an asymmetric interaction relationship in the business entity relationship network. This enables quantitative modeling of the multi-dimensional and directional influence between entities in the business ecosystem, thereby improving the completeness of the diagnostic analysis basis.

[0033] 2. This invention iteratively updates the state of each entity by performing coupled dynamic simulation in a business entity relationship network in response to preset disturbance events. It can reveal the evolution path of the target enterprise's future state under specific stress scenarios, realizing the transformation from static evaluation to dynamic prediction.

[0034] 3. This invention compares a series of internal financial state vectors in the future state evolution path of the target enterprise with the terminal state and a health benchmark, thereby simultaneously assessing the stability of the enterprise's state trajectory and its final financial condition in the process of responding to disturbances, and thus obtaining a comprehensive quantitative assessment result of the enterprise's resilience and risk. Attached Figure Description

[0035] Figure 1 This is an overall flowchart of an AI-based diagnostic method for enterprise financial health according to an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of a business entity relationship network and interactions according to an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram illustrating the application of a disturbance event according to an embodiment of the present invention. Detailed Implementation

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] See attached document Figure 1 , Figure 1 This is an overall flowchart of an AI-based diagnostic method for enterprise financial health according to an embodiment of the present invention. The present invention provides an AI-based diagnostic method for enterprise financial health, including steps S1 to S5.

[0040] In step S1, data acquisition and multidimensional dynamic state vector construction are performed. This step first acquires multi-source heterogeneous data for analysis. The multi-source heterogeneous data includes financial data, business relationship data, and dynamic behavior data.

[0041] Financial data includes, but is not limited to, the balance sheets, income statements, and cash flow statements of the target company and its related business entities. The data in these statements may be presented at a quarterly or annual timeframe. Business relationship data includes, but is not limited to, lists of suppliers, major customers, equity investment agreements, memorandums of understanding, and records of guarantees between companies. Dynamic behavioral data includes, but is not limited to, publicly released press releases, management discussion and analysis sections of annual reports, stock exchange announcements, and textual content from third-party industry research reports.

[0042] The data acquisition process includes: automatically acquiring structured financial data from public databases or commercial data providers through application programming interfaces (APIs); and using web crawling technology to selectively crawl unstructured or semi-structured text data published from sources such as official corporate websites and regulatory agency websites, based on preset entity lists and keyword rules.

[0043] After acquiring the raw data, it is preprocessed. Preprocessing operations include data cleaning, timestamp alignment, and data structuring. In the data cleaning step, missing values, outliers, and duplicate records are identified and addressed. One specific implementation method is to fill in missing numerical fields in financial data using the mean, median, or regression imputation methods based on the data of companies of similar size and industry size during the same period.

[0044] In the timestamp alignment step, data from different sources are correlated to a unified timeline. Specifically, all data points are mapped to a uniform time granularity based on when they occurred or were published, such as to a specific calendar day or financial reporting quarter, to ensure the consistency of state vectors in the time dimension in subsequent dynamic simulations.

[0045] For dynamic behavioral data of text type, structured processing is performed. This process begins with sentence segmentation, word segmentation, stop word removal, and part-of-speech tagging. Subsequently, the preprocessed text data is converted into a numerical vector representation. In one embodiment, a term frequency-inverse document frequency (IF-IVF) algorithm can be used to calculate the weights of each word and generate the IF-IVF vector of the text. In another embodiment, a pre-trained language model, such as the BERT model, can be used. A sentence or the entire document is input into the model, and its hidden layer output is extracted as a sentence vector or document vector with contextual semantic information. This numerical vector representation provides input for subsequent determination of strategic pattern vectors and behavioral preference vectors.

[0046] In step S1, after acquiring and preprocessing the multi-source heterogeneous data, the multi-dimensional dynamic state vector is determined. This step constructs a numerical vector for each business entity that can characterize its internal state and behavioral tendencies.

[0047] In one specific embodiment, the multidimensional dynamic state vector is composed of a concatenated combination of an internal financial state vector, a strategic pattern vector, and a behavioral preference vector. For business entity i, its multidimensional dynamic state vector at time t... Represented as:

[0048]

[0049] in, This represents the multidimensional dynamic state vector of business entity i at time t; This represents the internal financial state vector of business entity i at time t, with dimension N. F Representatives are used to construct the internal financial state vector. The total number of financial indicators; This represents the strategic pattern vector of business entity i at time t, with dimension N. M Representatives are used to construct strategic pattern vectors The total number of strategic features; This represents the behavioral preference vector of business entity i at time t, with dimension N. B This represents the vector used to construct behavioral preference vectors. The total number of behavioral characteristics; therefore, It is (N) F +N M +N B A 1 / 2 dimensional vector.

[0050] Internal financial state vector The determination of N is based on preprocessed financial data. This process includes selecting N from the balance sheet, income statement, and cash flow statement. F Several pre-defined financial indicators are used, and their values ​​at time t are calculated. These indicators include the current ratio, quick ratio, debt-to-equity ratio, total asset turnover, and return on equity, etc. The calculated N... F Each index value is normalized (e.g., its values ​​are mapped to intervals using minimum and maximum scaling), and then arranged in a predetermined order to form a vector. Represented as:

[0051]

[0052] Where, r i,k (t) represents the normalized financial indicator value of business entity i at time t, where k = 1, 2, ..., N FN F Representatives are used to construct the internal financial state vector. The total number of financial indicators.

[0053] Strategic Pattern Vector With behavioral preference vector The determination of the values ​​is based on the preprocessed dynamic behavioral data that has been converted into vector representations. This process is accomplished through a pre-trained machine learning model.

[0054] Specifically, for strategic pattern vectors The document vectors of dynamic behavioral data corresponding to time t are input into one or more pre-defined classification or regression models. The output dimension of the model is N. M Each dimension corresponds to a specific strategic pattern, such as expansionary tendency, contractionary tendency, or level of innovation investment. The numerical values ​​output by the model constitute the strategic pattern vector.

[0055] Similarly, for behavioral preference vectors The document vectors are input into another classification or regression model. The model's output dimension is N. B Each dimension corresponds to a specific behavioral preference, such as risk tolerance or willingness to cooperate. The numerical values ​​output by the model constitute the behavioral preference vector.

[0056] See attached document Figure 2 , Figure 2 This is a schematic diagram illustrating the business entity relationship network and interactions according to an embodiment of the present invention. In step S2, based on all N business entities determined in step S1, a business entity relationship network is constructed that can describe the interrelationships between them. Mathematically, this network is constructed as a directed graph, where each business entity constitutes a node in the graph, and the specific business relationships between business entities constitute the directed edges connecting the nodes.

[0057] In one specific embodiment, the business entity relationship network is mediated by a relationship adjacency matrix M. R Represent the matrix M using data structures. R It is an N×N square matrix, where N is the total number of business entities. Each element M in the matrix... R (i,j) is used to characterize the relationship type from business entity j to business entity i.

[0058] The adjacency matrix M of this relationship RThe construction process includes: First, initializing an N×N zero matrix. Then, systematically parsing the business relationship data acquired and preprocessed in step S1. For each specific business relationship record in the data, for example, a record showing that business entity j is the main supplier of business entity i, the element M in the i-th row and j-th column is located in the matrix. R (i,j).

[0059] The element M R The value of (i,j) is set to a predefined, non-zero category identifier to uniquely specify the type of relationship. For example, a value of 1 can represent a supplier-customer relationship, 2 a competitor relationship, 3 an investor-investee relationship, and 4 a guarantee relationship. If there is no direct relationship defined by business relationship data between business entity j and business entity i, pointing from j to i, then element M... R The value of (i,j) remains at the initial value of 0.

[0060] Due to the inherent directionality of business relationships—for example, the fact that business entity j is a supplier of i does not mean that i is also a supplier of j—the relational adjacency matrix M constructed in this embodiment is... R M is an asymmetric matrix, meaning that in the usual case, M R (i,j)≠M R (j,i). This asymmetric structure provides the basic topological structure for determining the asymmetric interaction relationships in the subsequent step S2.

[0061] In step S2, after constructing the topology of the business entity relationship network, the asymmetric interaction relationship between any two business entities in the network is further determined. This interaction relationship is a quantitative description of the influence of one business entity (the actor) on another business entity (the receiver), and the result is a numerical interaction vector.

[0062] For any pair of business entities i and j in the network, the asymmetric interaction from business entity j to business entity i affects A. ji This is determined through a pre-defined interaction function f(·). This function integrates information from multiple dimensions, and its specific form is:

[0063]

[0064] Among them, A ji represents the interaction influence vector exerted by business entity j on business entity i; f(·) represents the preset interaction function; The multidimensional dynamic state vector of the receiving business entity i at time t; R represents the multidimensional dynamic state vector of the business entity j that exerts the action at time t; ji This represents the relation type from j to i, and its value comes from the relation adjacency matrix M. R The category identifier in (i,j); W ji This indicates the strength of the relationship from j to i.

[0065] Relationship strength W ji It is a scalar used to quantify the tightness of a relationship. In one embodiment, this strength is calculated based on transaction amounts and frequencies in financial data. Specifically, the total transaction amount V between business entities j and i within a preset time window is first obtained. ji Total transaction frequency F ji Then, these two values ​​are normalized to obtain the normalized transaction amount. and trading frequency Relationship strength W ji Then it is calculated using the following formula:

[0066]

[0067] Here, α and β are preset weighting coefficients, and α+β=1.

[0068] The internal logic of the interaction function f(·) depends on the type of the input relation R. ji Perform selective calculations. For example: when R ji When the relationship is identified as supplier-customer (j is a supplier of i), the function f(·) calculates the financial health status of j (reflected in...). The impact of the financial state component on the supply chain stability of i. If the current ratio of j decreases, then R ji A negative component will be generated in the corresponding dimension, the magnitude of which is related to the strength of the relationship W. ji Proportional. When R ji When identified as a competitive relationship, the function f(·) calculates the strategic pattern of j (reflected in...). The impact of the strategic model component on the market environment of i. If the expansion tendency index of j increases, then R ji This will generate a negative competitive pressure component in the corresponding dimension.

[0069] By performing the above calculation on all entity pairs with relationships in the network, we finally obtain an N×N asymmetric interaction matrix A(t) that varies with time t, where the element in the i-th row and j-th column of the matrix is ​​the interaction vector R. ji This matrix provides direct input for the dynamic simulation in step S3.

[0070] In another embodiment of step S2, the step further includes a dynamic update mechanism for responding to changes in the business environment. This mechanism ensures the real-time nature and accuracy of the business entity relationship network and its interactions.

[0071] The mechanism first continuously monitors pre-defined business events. This monitoring process is achieved through periodic information extraction and analysis of the dynamic behavioral data sources (e.g., press releases, stock exchange announcements) from step S1. Pre-defined business event types include, but are not limited to: merger and acquisition announcements between companies, major asset restructuring, changes in core suppliers or customers, bankruptcy filings, and the signing or termination of major long-term contracts.

[0072] When a new piece of information matching a preset business event type is detected, the system first performs entity recognition, precisely extracting the unique identifier of the business entity directly associated with the event from the information text. For example, in an announcement about business entity g acquiring business entity h, the system identifies entity g and entity h as the core related entities in this event.

[0073] After identifying the associated business entities, the system triggers a local or global update to the relationship network. Specifically, the system first updates the constructed relationship adjacency matrix M. R Taking the aforementioned acquisition as an example, the system will modify matrix M. R All entries related to entity h are updated, and the relationship types between entity g and other entities are updated according to the new corporate structure following the acquisition. For example, existing supplier relationships of entity h are now transferred or integrated into entity g.

[0074] In the relational adjacency matrix M R Upon completion, the system immediately redetermines all asymmetric interactions involving the modified relation. Specifically, for each relation type R... ji Or relationship strength W ji For entity pairs (i,j) that have changed, the system calls the interaction function f(·), using the updated relation parameters and the latest multidimensional dynamic state vector. and Recalculate the interaction vector A ji The corresponding elements in the asymmetric interaction matrix A(t) are then updated with the new results. Through this mechanism, the present invention achieves dynamic maintenance of business relationship networks.

[0075] See attached document Figure 3 , Figure 3This is a schematic diagram illustrating the application of a disturbance event according to an embodiment of the present invention. The first stage of step S3 is to respond to a preset disturbance event. This response is achieved by making a one-time, instantaneous adjustment to the system state at the initial time t0 of the dynamic simulation, thereby setting initial conditions for subsequent state evolution. The disturbance event is selected from a preset group of event types, which includes node disturbances, relational disturbances, and macroscopic environmental disturbances.

[0076] In one embodiment, the selected perturbation event is a node perturbation. A node perturbation represents a specific event that directly impacts the internal state of a single business entity. This perturbation is applied by modifying the multidimensional dynamic state vector of the perturbed entity at an initial time t0. If business entity g is the perturbed entity, its adjusted initial state vector... Calculated using the following formula:

[0077]

[0078] in, This represents the initial multidimensional dynamic state vector of the business entity g after the perturbation is applied; This represents the original multidimensional dynamic state vector of the business entity g before the perturbation is applied; This represents the node perturbation vector corresponding to the current node perturbation event, and its dimension is... same.

[0079] The non-zero elements of this vector are set on the dimensions that are directly related to the disturbance event. For example, if the event is the failure of critical equipment, a negative value is set on the corresponding vector component that represents operational capability.

[0080] In another embodiment, the selected disturbance event is a relationship disturbance. A relationship disturbance represents a specific event that alters the way two business entities interact. This disturbance is achieved by modifying the corresponding elements in the asymmetric interaction matrix A(t0). For example, if the supply relationship from business entity j to business entity i is interrupted due to force majeure, the initial interaction vector A(t0) will be modified. ji (t0) is directly set to the zero vector, that is If the relationship is partially damaged, the original interaction vector can be multiplied by a decay coefficient between 0 and 1.

[0081] In another embodiment, the selected disturbance event is a macro-environmental disturbance. A macro-environmental disturbance represents a systematic event that simultaneously affects all or most business entities in the network. This disturbance is achieved by applying a common macro-perturbation vector to the initial state vectors of all affected entities. For each affected business entity i, its adjusted initial state vector... Calculated using the following formula:

[0082]

[0083] in, This represents the original multidimensional dynamic state vector of business entity i before the perturbation is applied; This represents the macroscopic disturbance vector corresponding to the current macroscopic environmental disturbance event.

[0084] For example, if the event is a central bank interest rate hike, then the vector has a positive value in the component related to financing costs, and this positive value is uniformly applied to the state vectors of all affected entities.

[0085] In step S3, after applying the initial perturbation, the system performs a coupled dynamic simulation in the business entity relationship network. This simulation starts at an initial time t0 and iterates at preset time steps Δt until a preset simulation termination time T is reached. Within each time step, the system synchronously updates the multidimensional dynamic state vector for each business entity in the network.

[0086] For any business entity i in the network, its state update process at time t includes the following calculations. First, calculate the net impact of the action on the business entity at time t. The net effect is the asymmetric interaction effect vector A exerted by all other business entities j on entity i in the network. ji The vector sum is calculated using the following formula:

[0087]

[0088] Among them, A ji This represents the asymmetric interaction effect vector exerted by business entity j on business entity i at time t, which is determined by step S2; This indicates that the following terms are summed; N represents the total number of business entities.

[0089] In calculating the net effect Then, through a pre-defined evolution function G(·), based on the multi-dimensional dynamic state vector of business entity i at the current time t, Net effect The multidimensional dynamic state vector of the next time step t+Δt is calculated. This evolutionary process can be described by the following formula:

[0090]

[0091] Where G(·) represents the preset evolution function; Δt represents the multidimensional dynamic state vector of business entity i at the current time t; Δt represents the discrete time step of the dynamic simulation. This represents the net impact of business entity i at the current time t.

[0092] In one specific embodiment, the evolution function G(·) is expressed as a set of coupled difference equations. This function represents the multidimensional dynamic state vector. Decomposed into internal financial state vector Strategic Pattern Vector With behavioral preference vector Evolutionary calculations were then performed separately. The specific evolutionary formulas are as follows:

[0093]

[0094] in, These are the internal financial state vectors corresponding to the next time step t+Δt. Strategic Pattern Vector With behavioral preference vector The vector; Δt represents the discrete time step of the dynamic simulation; W F Represented as a dimension N F ×(N F +N M +N B A preset weight matrix is ​​used to linearly transform the net effect vector into the rate of change of each component of the internal financial state vector; W M Represented as a dimension N M ×(N F +N M +N B The preset weight matrix is ​​used to linearly transform the net effect vector into the rate of change of each component of the strategy pattern vector; W B Represented as a dimension N M ×(N F +N M +N B The preset weight matrices are used to linearly transform the net effect vector into the rate of change of each component of the behavioral preference vector. The parameter values ​​of these weight matrices are calibrated using historical data.

[0095] Then, the updated component vectors are recombine to obtain...

[0096] Finally, the multidimensional dynamic state vector for the next time step is calculated. This is set as the current state vector of the business entity in subsequent iterations. The system repeats this process for all N entities in the network, completing a global state update for one time step. This iterative process continues until the simulation time reaches the termination time T, ultimately generating a multi-dimensional dynamic state vector time series from t0 to T for each business entity.

[0097] See attached document Figure 1 Step S4 is the target enterprise data extraction step. This step is performed after the coupled dynamics simulation in step S3, and its purpose is to accurately separate the time series data that is only relevant to the target enterprise of this diagnostic method from the huge dataset containing the state evolution information of all business entities.

[0098] After step S3 is completed, the system generates a set containing the state evolution data of all N business entities from the initial time t0 to the final time T. In a specific embodiment, this data set is stored as a three-dimensional data structure with dimensions N×K×D, where N is the total number of business entities; K is the total number of time steps in the simulation; and D is the dimension of the multidimensional dynamic state vector. Each element in this data structure corresponds to a component of the state vector of a business entity at a specific point in time.

[0099] The core operation of step S4 is to extract data specific to the target enterprise from the three-dimensional data structure based on a preset target enterprise identifier. This target enterprise identifier is the index or ID uniquely assigned to the target enterprise when the target enterprise and associated business entities are defined as multiple business entities in step S1.

[0100] The extraction operation specifically involves selecting the index corresponding to the target company identifier on the first dimension (entity dimension) of the three-dimensional data structure, thereby obtaining a two-dimensional data matrix. This two-dimensional data matrix has a dimension of K×D, where K is the total number of time steps in the simulation; and D is the dimension of the multidimensional dynamic state vector. Each row represents the multidimensional dynamic state vector of the target company at one time step in the simulation process, while each column represents the change of a specific component of the state vector over time.

[0101] This two-dimensional data matrix is ​​a series of multi-dimensional dynamic state vectors for the target company, which completely records the trajectory of the target company's state changes after responding to preset disturbance events. This matrix serves as the unique and independent input data for the subsequent step S5 to determine financial health, ensuring the specificity and accuracy of the final diagnostic results.

[0102] See attached document Figure 1 Step S5 involves analyzing the future state evolution path and determining financial health. This step first analyzes the future state evolution path extracted in step S4 that pertains to the target company. The input for this analysis is a series of multi-dimensional dynamic state vectors recording the target company's state from initial time t0 to final time T.

[0103] The first step in the analysis is to identify a time series containing only internal financial state information from the future state evolution path. Specifically, this involves identifying each multidimensional dynamic state vector within that path. Extract its internal financial state vector components.

[0104]

[0105] in, This represents the internal financial state vector of the target company at time t; This represents the strategic pattern vector of the target company at time t; This represents the target company's behavioral preference vector at time t.

[0106] In one specific embodiment, this extraction operation is a data subset selection operation. Because in the vector construction of step S1, the internal financial state vector... Preset to occupy the top N of the multidimensional dynamic state vector F Since it has multiple dimensions, the specific operation is as follows: for each row of the K×D two-dimensional data matrix output in step S4, extract the first N rows. F The values ​​in the column. The output of this operation is a new K×N array. F A dimensional data matrix, which is a series of internal financial state vectors of the target company.

[0107] The second step of the analysis is to determine the terminal states of a series of internal financial state vectors. These terminal states are defined as the target firm's internal financial state vectors at the end time T of the simulation process, denoted as... In the specific implementation, this operation is to directly select the aforementioned K×N. F The last row of the two-dimensional data matrix.

[0108] Based on the above analysis, this step generates two outputs: a time series representing the trajectory of the internal financial state evolution process. and a terminal state vector representing the final financial outcome. These two outputs will be used together as input data for subsequent comparisons with health benchmarks.

[0109] In step S5, after analyzing the future state evolution path, a comparison is performed with a health benchmark to ultimately determine the financial health of the target company. This process first requires setting a preset health benchmark.

[0110] The health baseline includes M preset baseline categories, where M is an integer greater than or equal to 2. Each baseline category m (where m = 1, 2, ..., M) is defined by a template path. and a template terminal status Composition. These benchmark categories are defined either through cluster analysis of historical business entity data or by domain experts based on financial criteria. Each category corresponds to a specific description of financial health, such as robust, deteriorating, or collapsing.

[0111] The comparison process consists of two parallel parts. The first part calculates the time series of the evolution of the target company's internal financial state. Template path for each baseline category m The path similarity between them. In one specific embodiment, this similarity is calculated using a dynamic time warping algorithm, and the result is a scalar distance value D. path (target,m).

[0112] Part Two: Calculating the Terminal State Vector of the Target Enterprise Template terminal state for each baseline category m The terminal state distance between the two vectors. In one specific embodiment, this distance is obtained by calculating the Euclidean distance between the two vectors, resulting in a scalar distance value D. term (target, m). Its calculation formula is:

[0113]

[0114] Where, r target,k (T) represents the k-th normalized financial indicator value of the target company at terminal time T; This represents the k-th normalized financial indicator value in the template terminal state of the m-th benchmark category; Indicates the content within parentheses The difference is squared to ensure that r target,k (T) and The differences between them are all included in the sum as positive numbers, and larger differences are given higher weights; N F Represents the internal financial state vector The total number of financial indicators; This indicates that the following item, i.e. Accumulate.

[0115] Subsequently, for each baseline category m, a weighted function is used to combine the path similarity and the terminal state distance into a comprehensive distance score D. total (m). The specific form of this function is:

[0116] D total (m)=ω path ·D path (target,m)+ω term ·D term(target,m);

[0117] Where, ω path D represents the preset path similarity weight coefficient; path (target,m) represents the internal financial state vector sequence of the target company. Template path of the m-th benchmark category Path similarity between them; ω term D represents the preset terminal state distance weighting coefficient; term (target,m) represents the terminal state vector of the target enterprise. Template terminal state of the m-th benchmark category Distance between terminal states; ω path and ω term All are non-negative real numbers, and satisfy ω path +ω term =1.

[0118] Finally, the category with the lowest overall distance score among all M benchmark categories is identified. The financial health description corresponding to this benchmark category with the lowest score is determined as the financial health result of the target company in this diagnosis.

[0119] A system embodiment of the present invention provides a computing device having the hardware foundation for performing the aforementioned method steps S1 to S5.

[0120] In one specific embodiment, the computing device includes at least one processor, a memory, a non-volatile storage medium, an input / output interface, and a network interface. All of the above components are communicatively connected via one or more system buses.

[0121] The processor can be a central processing unit (CPU), a graphics processing unit (GPU), or a dedicated artificial intelligence (AI) accelerator chip. The processor's function is to read and execute computer program instructions stored in a non-volatile storage medium and loaded into memory to control the computing device to perform all the method steps disclosed in this invention.

[0122] The memory is a volatile storage device, such as random access memory (RAM). Its function is to provide a high-speed data read and write channel for the processor, and to temporarily store the operating system, application programs, and intermediate data generated during method execution, such as the multi-dimensional dynamic state vector generated for each time step for each business entity in step S3.

[0123] The non-volatile storage medium can be a hard disk drive (HDD), a solid-state drive (SSD), or flash memory. Its function is to store data and programs long-term; specifically, it stores computer program instructions for implementing the method of the present invention, as well as the raw multidimensional heterogeneous data acquired in step S1.

[0124] A network interface, such as an Ethernet adapter or a wireless network interface card, provides the computing device with the ability to connect to an external network (such as the Internet). This interface is specifically used to perform the data acquisition operation in step S1, namely, to obtain financial data, business relationship data, and dynamic behavioral data from online data sources such as public databases, commercial data providers, or corporate websites.

[0125] The input / output interface connects to external input devices (such as a keyboard and mouse) and output devices (such as a monitor). This interface is used to receive instructions from the operator, such as starting a diagnostic task, setting the type of disturbance event, and to present the final target enterprise financial health results determined in step S5 to the operator.

[0126] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor in the aforementioned computing device, it can implement the aforementioned AI diagnostic method for enterprise financial health. The computer program can be logically divided into multiple functional modules.

[0127] The computer program's functional modules include: a data acquisition and vector construction module, a relation network and interaction determination module, a coupled dynamics simulation module, a target data extraction module, and a health analysis and determination module.

[0128] The data acquisition and vector construction module is configured to perform the functions of step S1. This module contains instructions that enable it to acquire financial data, business relationship data, and dynamic behavioral data from multiple pre-defined web data sources via a web interface. The module further includes subroutines for data cleaning, timestamp alignment, and converting unstructured text data into numerical vectors (such as TF-IDF vectors or sentence vectors). The final output of this module is to construct and store a multidimensional dynamic state vector for each business entity at its initial moment.

[0129] The relation network and interaction determination module is configured to perform the function of step S2. This module reads business relation data and constructs and stores a relation adjacency matrix M based on it. R This module also implements a predefined interaction function f(·), used to calculate and generate an asymmetric interaction matrix A(t) by combining the multidimensional dynamic state vectors of business entities, relationship types, and relationship strengths. The module also includes an event monitoring subroutine for dynamically updating the relationship network.

[0130] The Coupled Dynamics Simulation module is configured to perform the function of step S3. This module receives the perturbation event type and parameters set by the user through the input / output interface, and modifies the state vectors of one or more entities or their interactions at the initial time step accordingly. At the core of this module is an iterative computation loop that calculates the net impact for each business entity in the network within each time step Δt. The module applies the evolution function G(·) to update the multidimensional dynamic state vector. It stores the state evolution trajectories of all entities throughout the simulation in memory or a non-volatile storage medium.

[0131] The target data extraction module is configured to perform the function of step S4. This module receives a unique identifier of a target enterprise as input. It performs a data selection operation to accurately extract a series of multi-dimensional dynamic state vectors from t0 to T corresponding to the target enterprise identifier from the complete state evolution dataset generated by the coupled dynamics simulation module, and passes the results to the next module.

[0132] The health analysis and determination module is configured to perform the function of step S5. This module first extracts the internal financial status vector sequence from the received target enterprise data. Then, it loads a preset health baseline (including template paths and template terminal states) from non-volatile storage. This module contains program code implementing a dynamic time warping algorithm and Euclidean distance calculation to calculate path similarity and terminal state distance, and calculates a comprehensive distance score based on preset weights. Finally, the module determines the baseline category with the lowest score and outputs the corresponding financial health description as the final result through the input / output interface.

Claims

1. An AI-based diagnostic method for corporate financial health, characterized in that, Includes the following steps: S1. Acquire multi-source heterogeneous data, and define the target enterprise and related business entities as multiple business entities. Determine the multi-dimensional dynamic state vector of each business entity. The multi-dimensional dynamic state vector is used to characterize the internal state and behavioral tendencies of the business entity. S2. Construct a business entity relationship network that includes all the business entities. In the business entity relationship network, based on the multidimensional dynamic state vector of each business entity, determine the asymmetric interaction relationship between any two business entities. S3. In response to a preset disturbance event, perform coupled dynamic simulation in the business entity relationship network, and apply the asymmetric interaction relationship determined in step S2 to iteratively update the multidimensional dynamic state vector of each business entity to obtain a series of multidimensional dynamic state vectors for each business entity. S4. Extract a series of multidimensional dynamic state vectors of the target enterprise from a series of multidimensional dynamic state vectors of each of the business entities obtained in step S3. S5. Determine a series of multidimensional dynamic state vectors of the target enterprise obtained in step S4 as a future state evolution path of the target enterprise, and determine the financial health of the target enterprise based on the future state evolution path.

2. The AI ​​diagnostic method for enterprise financial health according to claim 1, characterized in that, The multi-source heterogeneous data includes: The target company and its associated business entities' financial data, business relationship data, and dynamic behavior data.

3. The AI ​​diagnostic method for enterprise financial health according to claim 2, characterized in that, In step S1, the specific steps for determining the multidimensional dynamic state vector of each of the business entities are as follows: The internal financial state vector, strategic pattern vector, and behavioral preference vector of each business entity are determined, and the internal financial state vector, strategic pattern vector, and behavioral preference vector are combined to form the multidimensional dynamic state vector.

4. The AI ​​diagnostic method for enterprise financial health according to claim 3, characterized in that, In step S1, the specific steps for determining the internal financial state vector, strategic pattern vector, and behavioral preference vector are as follows: Process the financial data to determine the internal financial state vector; The dynamic behavior data is processed using natural language processing models and pattern recognition models to determine the strategic pattern vector and the behavior preference vector.

5. The AI ​​diagnostic method for enterprise financial health according to claim 2, characterized in that, In step S2, the asymmetric interaction relationship is also determined by the relationship type and strength between the business entities.

6. The AI ​​diagnostic method for enterprise financial health according to claim 5, characterized in that, The relationship type is determined based on the business relationship data, and the relationship strength is quantified based on the transaction amount and transaction frequency in the financial data.

7. The AI ​​diagnostic method for enterprise financial health according to claim 1, characterized in that, Step S2 also includes: Monitor pre-set business events; When the preset business event occurs, the business entities directly associated with the preset business event are identified. Then, based on the multidimensional dynamic state vector of the directly associated business entities, the asymmetric interaction relationships involving all of the business entities are reconfirmed.

8. The AI ​​diagnostic method for enterprise financial health according to claim 1, characterized in that, The preset disturbance events are selected from a group of event types including node disturbances, relationship disturbances, and macro-environmental disturbances.

9. The AI ​​diagnostic method for enterprise financial health according to claim 1, characterized in that, In step S3, the step of iteratively updating the multidimensional dynamic state vector of each of the business entities is as follows: At the initial moment of the dynamic simulation, the initial conditions of the dynamic simulation are set by adjusting the multidimensional dynamic state vector of the business entities related to the preset disturbance event in one go. At each time step starting from the initial moment, the net effect of the asymmetric interaction relationship acting on each of the business entities is calculated; Then, using a preset evolution function, the multidimensional dynamic state vector for the next time step is calculated based on the current multidimensional dynamic state vector of the business entity and the net effect. The multidimensional dynamic state vector of the next time step is set as the current multidimensional dynamic state vector of the business entity in the subsequent time steps, and the iterative update is completed.

10. The AI ​​diagnostic method for enterprise financial health according to claim 3, characterized in that, In step S5, the step of determining the financial health of the target enterprise based on the future state evolution path is as follows: From the future state evolution path, determine a series of internal financial state vectors, and determine the terminal state of the series of internal financial state vectors; The series of internal financial status vectors and the terminal status are compared with a preset health benchmark to determine the financial health of the target enterprise.

Images