An openclaw framework business data intelligent analysis method for enterprise high-level decision

Abstract

The present application relates to the technical field of business data analysis, and particularly discloses an OpenClaw framework business data intelligent analysis method for enterprise high-level decision-making, which acquires enterprise internal micro-operation graph structure data and external macro-environment time series data, aligns to an enterprise strategic target ontology through a cross-modal tensor decomposition network to construct a time series strategic intention tensor, extracts multi-scale business state features based on the tensor, and constructs a Markov game state space in combination with resource constraints; decodes high-level strategic decision-making instructions into tactical constraint boundaries, carries out game deduction by using a multi-agent algorithm with a risk entropy penalty term, and applies counterfactual intervention to key macro-environment nodes to generate a strategic evolution trajectory set; and finally, traces back abnormal intervention nodes in a cause-effect chain to generate a multi-modal strategic decision-making report. The present application can realize quantitative deduction and cause-effect explanation of enterprise strategic decision-making, and improve the scientific nature of high-level decision-making and the execution closed-loop efficiency.

Description

Technical Field

[0001] This invention relates to the field of business data analysis technology, specifically to an OpenClaw framework-based business data intelligent analysis method for senior management decision-making in enterprises. Background Technology

[0002] Current commercial data analysis methods have significant limitations when dealing with strategic decision-making by senior management.

[0003] First, internal micro-operational data and external macro-environmental data are often disconnected, lacking a unified semantic alignment and cross-modal fusion mechanism, making it difficult to accurately quantify strategic intent;

[0004] Secondly, traditional analytical methods rely heavily on static reports or single-variable time-series forecasts, which cannot conduct long-term multi-agent game simulations under complex resource constraints involving multiple departments, making it difficult to comprehensively assess the long-term risks and overall benefits of strategic decisions.

[0005] Finally, existing decision support systems lack causal vulnerability analysis of macroeconomic environmental fluctuations. When external environments undergo sudden changes, they cannot provide counterfactual reasoning paths with causal explanations, resulting in a lack of interpretability and risk resistance in high-level decisions. Furthermore, strategic instructions are difficult to automatically decode into executable underlying tactical indicators, thus failing to form an effective decision-making loop.

[0006] Therefore, there is an urgent need for an OpenClaw framework-based business data intelligence analysis method for high-level corporate decision-making to solve the above problems. Summary of the Invention

[0007] The purpose of this invention is to provide an OpenClaw framework-based business data intelligent analysis method for enterprise senior management decision-making, in order to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] An OpenClaw framework-based business data intelligence analysis method for enterprise senior management decision-making includes the following steps:

[0010] S1: Obtain the internal micro-operational diagram structure data and the external macro-environment time series data. Align the micro-operational diagram structure data and the macro-environment time series data to the preset corporate strategic goal ontology through the cross-modal tensor decomposition network of the OpenClaw framework, and construct a time series strategic intent tensor containing latent variables of strategic intent.

[0011] S2: Based on the temporal strategic intent tensor, multi-scale business state features are extracted through the graph-temporal joint encoder of the OpenClaw framework, and combined with the preset enterprise resource constraints, a Markov game state space including state observation space, tactical action space and long-period delayed reward function is constructed.

[0012] S3: Receives high-level strategic decision-making instructions and decodes them into the constraint boundaries of the tactical action space. In the Markov game state space, it uses a multi-agent proximal policy optimization algorithm with a risk entropy penalty term to perform game deduction. By applying counterfactual intervention to key external macroeconomic nodes, it generates a set of strategic evolution trajectories containing the baseline path and counterfactual perturbation paths, and calculates the multi-objective risk-benefit utility matrix and causal vulnerability index for each path.

[0013] S4: Utilize the causal structure learning module of the OpenClaw framework to trace the causal chain of the strategic evolution trajectory set, identify key intervention nodes that lead to anomalies in the causal vulnerability index, generate a multimodal strategic decision report containing decision path explanations and confidence intervals, and decode high-level strategic decision instructions into an executable set of enterprise-level tactical KPI indicators to complete the decision-making loop.

[0014] As can be seen from the technical solution provided by the present invention above, the OpenClaw framework business data intelligent analysis method for enterprise senior management decision-making provided by the present invention has the following beneficial effects:

[0015] By aligning micro-operational graph structure data and macro-environmental time-series data to the enterprise strategic goal ontology through cross-modal tensor decomposition networks, deep integration and semantic unification of heterogeneous data are achieved, and latent variables of strategic intent are accurately extracted, providing a high-dimensional representation basis for state observation in complex business environments.

[0016] We construct a Markov game state space with resource constraints and use a multi-agent algorithm with risk entropy penalty term to perform game inference. Combined with counterfactual intervention, we generate a set of strategic evolution trajectories, which effectively quantifies the multi-objective risks and returns of long-term decision-making and significantly enhances the risk resistance of corporate strategy in the face of external macroeconomic fluctuations.

[0017] By using the causal structure learning module to trace causal chains and generate multimodal strategic decision-making reports, not only can it provide high-level decision-making with a rigorous causal logic and interpretable basis, but it can also automatically decode strategic instructions into a set of underlying tactical indicators, thus completely opening up the decision-making loop from strategic planning to execution feedback and improving the overall operational efficiency of the enterprise. Attached Figure Description

[0018] Figure 1This is a schematic diagram illustrating the steps of a business data intelligent analysis method based on the OpenClaw framework for senior management decision-making in enterprises, as described in this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] To better understand the above technical solutions, the following will provide a detailed description of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, this embodiment of the invention provides a business data intelligent analysis method based on the OpenClaw framework for enterprise senior management decision-making, including the following steps:

[0022] S1: Obtain the internal micro-operational diagram structure data and the external macro-environment time series data. Align the micro-operational diagram structure data and the macro-environment time series data to the preset corporate strategic goal ontology through the cross-modal tensor decomposition network of the OpenClaw framework, and construct a time series strategic intent tensor containing latent variables of strategic intent.

[0023] S2: Based on the temporal strategic intent tensor, multi-scale business state features are extracted through the graph-temporal joint encoder of the OpenClaw framework, and combined with the preset enterprise resource constraints, a Markov game state space including state observation space, tactical action space and long-period delayed reward function is constructed.

[0024] S3: Receives high-level strategic decision-making instructions and decodes them into the constraint boundaries of the tactical action space. In the Markov game state space, it uses a multi-agent proximal policy optimization algorithm with a risk entropy penalty term to perform game deduction. By applying counterfactual intervention to key external macroeconomic nodes, it generates a set of strategic evolution trajectories containing the baseline path and counterfactual perturbation paths, and calculates the multi-objective risk-benefit utility matrix and causal vulnerability index for each path.

[0025] S4: Utilize the causal structure learning module of the OpenClaw framework to trace the causal chain of the strategic evolution trajectory set, identify key intervention nodes that lead to anomalies in the causal vulnerability index, generate a multimodal strategic decision report containing decision path explanations and confidence intervals, and decode high-level strategic decision instructions into an executable set of enterprise-level tactical KPI indicators to complete the decision-making loop.

[0026] In this embodiment, step S1 uses the cross-modal tensor decomposition network of the OpenClaw framework to perform cross-modal alignment and fusion of the enterprise's internal micro-operational graph structure data and the external macro-environment time-series data, and maps the fusion result to the enterprise's strategic goal ontology space. Finally, a time-series strategic intent tensor containing latent variables of strategic intent is constructed, providing a unified semantic representation foundation for subsequent multi-scale business state feature extraction and Markov game inference. The detailed steps are as follows:

[0027] Step S1-1: Extract three types of basic entity nodes from the Enterprise Resource Planning (ERP) system and Customer Relationship Management (CRM) system: organizational structure nodes, business process nodes, and resource flow nodes. Organizational structure nodes cover enterprise departments, positions, and reporting levels; business process nodes cover core business processes such as procurement, production, sales, and after-sales service; and resource flow nodes cover the flow paths of funds, materials, and information.

[0028] An adjacency matrix is ​​constructed based on the semantic relationships between three types of nodes; these semantic relationships include reporting relationships between organizational structure nodes, collaborative relationships between business process nodes, and resource dependency relationships between resource flow nodes; the total number of nodes is assumed to be... Adjacency matrix The dimension is If node With nodes If any of the above semantic relationships exist, then the corresponding elements of the adjacency matrix... The value is 1 if it is not 0 otherwise; at the same time, the attribute feature vectors corresponding to each node are extracted. The attribute feature vectors include the node type code, business weight coefficient and historical activity index. The attribute feature vectors of all nodes are arranged in rows to form a node attribute matrix. Its dimensions are ,in, As the initial attribute dimension; set the adjacency matrix With node attribute matrix Combining these elements yields the micro-operational diagram structure data. , represented as ;

[0029] Step S1-2: Collect three types of macroeconomic environment time-series data from external data sources: macroeconomic indicator time-series data, industry competition situation time-series data, and policy and regulatory change time-series data. Macroeconomic indicator time-series data includes indicators such as GDP growth rate, consumer price index, and manufacturing purchasing managers' index, collected monthly or quarterly. Industry competition situation time-series data includes indicators such as industry concentration index, market share change rate of leading companies, and number of new entrants, collected monthly. Policy and regulatory change time-series data includes indicators such as tax policy adjustment index, industry entry threshold change rate, and environmental compliance cost index, collected weekly or monthly.

[0030] Missing value imputation and outlier correction were performed sequentially on the three types of time series data. Missing value imputation adopted an interpolation method based on local weighted regression of time series, with the values ​​before and after the missing value being adjusted. Using observations at specific time points as a reference window, missing values ​​are estimated by fitting a local polynomial curve using the weighted least squares method. The default value is 5; outlier correction uses the median absolute deviation method. If the observed value at a certain time point deviates from the median of the sliding window by more than three times the median absolute deviation, it is judged as an outlier and replaced with the median of the sliding window.

[0031] After data cleaning, time-granularity aligned sampling was performed on the three types of time-series data; a unified time granularity was set as follows. The default value is monthly. A forward-fill method is used to expand low-frequency data to a monthly granularity, and a monthly mean aggregation method is used to shrink high-frequency data to a monthly granularity. The aligned macro-environment time-series data is assumed to contain... Each time step contains [number] time steps. The standardized time-series data of the macro environment is represented as a matrix based on the following macroeconomic indicator dimensions. Its dimensions are ;

[0032] Step S1-3: Structure the micro-operation diagram data The input graph convolutional encoder performs multi-order neighbor aggregation; the graph convolutional encoder is composed of... The graph convolutional layers are stacked together, and the convolutional operation of each layer is defined as follows: ,in, To add a self-loop adjacency matrix, It is the identity matrix; for The degree matrix, whose diagonal elements ; For the first Layer node embedding matrix; For the first The learnable weight matrix of the layer; The activation function is non-linear, and the rectified linear unit function is used by default; the encoder's first-layer input... Node attribute matrix ,go through After layer-by-layer aggregation, the microstructure embedding matrix is ​​output. Its dimensions are ,in, The output feature dimension of the graph convolutional encoder;

[0033] Simultaneously, macroeconomic environmental time series data The input temporal convolution encoder performs multi-scale sliding window convolution; the temporal convolution encoder consists of... It consists of several parallel temporal convolutional kernels, each with a different receptive field scale; the first... The convolution operation with convolution kernels is defined as follows: ,in, For the first One-dimensional convolution operation; and These are the weight tensor and bias vector of the convolution kernel, respectively; The activation function is represented by *, which indicates a one-dimensional convolution operation. The outputs of each convolutional kernel are concatenated along the feature dimension and then passed through a fully connected layer for dimensionality reduction mapping to obtain the macroscopic temporal embedding matrix. Its dimensions are ,in, This represents the output feature dimension of the temporal convolutional encoder.

[0034] Step S1-4: Embedding the microstructure into the matrix and macroscopic temporal embedding matrix Concatenate along the feature dimension to construct the initial fusion tensor; the specific operation is as follows:

[0035] Will Replicate and expand at the node level Each time step yields a dimension of... tensor;

[0036] Will Broadcast to in the time dimension With nodes, we obtain a dimension of tensor;

[0037] Concatenate the two along the last feature dimension to obtain the initial fusion tensor. Its dimensions are ,in ;

[0038] For the initial fusion tensor Perform tensor decomposition based on alternating least squares, decomposing it into a core tensor and three factor matrices; the mathematical form of the tensor decomposition is: ,in, The core tensor has dimensions of , , , These are the preset ranks for the temporal mode, the nodal mode, and the feature mode, respectively. This is a time-dimensional factor matrix with dimension 1. ; The node dimension factor matrix is ​​also known as the micro-dimensional factor matrix, with dimension 1. ; This is the feature dimension factor matrix, also known as the macro-dimensional factor matrix, with dimensions of... ; Represents tensors and matrices along the first... Modal multiplication;

[0039] Alternating least squares method updates each factor matrix sequentially with fixed other factor matrices, iterating until the reconstruction error converges. The reconstruction error is defined as the square of the Frobenius norm between the original tensor and the reconstructed tensor, and the convergence threshold is set to a default value. After decomposition, the core tensor is obtained. Micro-dimensional factor matrix Macro-dimensional factor matrix and time dimension factor matrix Together, they form a set of factor matrices;

[0040] Steps S1-5: Load the preset corporate strategic goal ontology. The corporate strategic goal ontology defines the hierarchical semantic structure of corporate strategic goals, including three levels: strategic vision layer, strategic goal layer, and strategic action layer. Each level contains several strategic dimension nodes, and each strategic dimension node is associated with a set of semantic keyword vectors. Extract the semantic keyword vectors of all strategic dimension nodes in the corporate strategic goal ontology and construct the ontology constraint matrix. Its dimensions are ,in, The total number of strategic dimension nodes is represented by the matrix, and each column represents the semantic embedding vector of a strategic dimension node.

[0041] Based on ontology constraint matrix Constructing the ontology constraint mask matrix Its dimensions are The ontology constraint mask matrix is ​​calculated as follows: ,in, for Transpose of; for 3D identity matrix; This is the regularization constant, with a default value of [value to be filled in]. ;

[0042] Using ontology constraint mask matrix Orthogonally project the three factor matrices in the factor matrix set to align their column spaces to the semantic subspace spanned by the corporate strategic objective ontology; the alignment operation is defined as follows: , , ,in, This is the aligned time dimension factor matrix; This is the aligned micro-dimensional factor matrix; The aligned macroscopic dimension factor matrix; projection matrix , , The following orthogonal Prouk problem is obtained by solving: ,in, This is the ontology constraint target matrix for the corresponding mode; It is the Frobenius norm; The values ​​are 1, 2, and 3, corresponding to the three modes of time, micro, and macro, respectively;

[0043] Steps S1-6: Convert the core tensor Aligned time dimension factor matrix Aligned micro-dimensional factor matrix and the aligned macroscopic dimension factor matrix Perform tensor reconstruction to obtain the initial strategic intent tensor. : Among them, the initial strategic intent tensor The dimension is ,in For time steps, For the number of nodes, For feature dimensions;

[0044] Tensor of initial strategic intent Reparameterization sampling of the variational autoencoder is performed in the time dimension; firstly, a variational autoencoder is constructed, whose encoder part is... Mean pooling compression is performed along the time dimension, and then the mean vectors of the latent variables are output through two fully connected layers. Sum of logarithmic variance vector Both dimensions are , The default dimension of the latent variable space is 64; the reparameterization sampling operation is defined as: ,in, The element-wise square root of the variance vector; To obtain from the standard normal distribution A random noise vector sampled in the middle; This represents element-wise multiplication;

[0045] The sampled latent variable vector Injecting the initial strategic intent tensor In the feature channels; the specific operation is: to convert the latent variable vector Mapped to dimension [dimensional] through a fully connected layer modulation vector Then, the modulation vector is compared with the initial strategic intent tensor. Each time-step slice is modulated along the feature dimension using element-wise multiplication to obtain a temporal strategic intent tensor containing latent strategic intent variables. : , ,in, The temporal strategic intent tensor represents the tensor in the first order. A slice at each time step, with dimensions of ; It is a vector of all 1s with dimension . ; The modulation vector has a dimension of ; This is the modulation matrix after broadcasting the modulation vector along the node dimension;

[0046] At this point, step S1 is complete, and the output is a temporal strategic intent tensor containing latent variables of strategic intent. Its dimensions are This tensor simultaneously encodes the graph structure information of the enterprise's micro-operations, the temporal evolution information of the external macro environment, and the semantic alignment information of the enterprise's strategic goal ontology. It also injects the latent variable representation of strategic intent through a variational autoencoder, providing a complete data foundation for the multi-scale business state feature extraction in step S2.

[0047] In this embodiment, step S2, based on the temporal strategic intent tensor generated in step S1, extracts multi-scale business state features that fuse spatial topology and temporal evolution using the graph-temporal joint encoder of the OpenClaw framework. Based on this, and combined with preset enterprise resource constraints, a Markov game state space is constructed, including a state observation space, a tactical action space, and a long-period delayed reward function. This provides a complete decision-making environment model for the multi-agent game deduction in step S3. The detailed steps are as follows:

[0048] Step S2-1: Convert the temporal strategic intent tensor output in step S1 into a multi-step ... Input graph-temporal co-encoder spatial graph attention subnetwork; temporal strategic intent tensor The dimension is ,in For time steps, The total number of nodes. For feature dimension; the spatial graph attention subnetwork is composed of It is composed of stacked dynamic graph attention layers; in each dynamic graph attention layer, the temporal strategic intent tensor is first processed. slices at each time step Perform a linear mapping to generate a query matrix. Key matrix Sum matrix : , , ,in, , , These are learnable weight matrices for queries, keys, and values, each with a dimension of [missing information]. , The feature dimension after mapping;

[0049] Then, the spatial dependency attention weights between nodes are calculated; for each time step... ,node For nodes Attention weights Calculated by scaling the dot product: ,in, For query matrix The row vectors Key matrix The row vectors This is the scaling factor;

[0050] Attention weights Normalization is performed using the Softmax function to obtain normalized attention weights. : ,in, It is a natural exponential function;

[0051] Using normalized attention weights Log-value matrix Perform weighted aggregation to obtain nodes At time step Spatial topological feature vectors : ,in, Value matrix The Row vectors;

[0052] Arrange the spatial topological feature vectors of all nodes at all time steps according to the time step and node dimensions to obtain the node-level spatial topological feature tensor. Its dimensions are ;

[0053] Step S2-2: Convert the temporal strategic intent tensor Simultaneously input to the graph-temporal joint encoder is a multi-scale temporal dilated convolutional subnetwork; this subnetwork consists of... It consists of several parallel causal dilated convolutional branches, each configured with a different dilation rate to capture temporal dependencies at different time scales;

[0054] For the Each branch has an expansion rate of [missing information]. The weights of the causal convolution kernel are set to The core size is set to This branch focuses on the temporal strategic intent tensor. Perform causal dilation convolution operations along the time dimension; for each time step , No. Output of each branch The calculation is as follows: ,in, For the first Each branch convolutional kernel is located at... Weighted slices; For the temporal strategic intent tensor at time step Slices; Indicates the expansion offset;

[0055] Will The outputs of each branch are concatenated along the feature dimension and then subjected to dimensionality reduction mapping through a fully connected layer to obtain the sequence-level temporal evolution feature tensor. Its dimensions are ,in, This represents the output feature dimension of the temporal convolutional sub-network.

[0056] Step S2-3: Construct a cross-modal gated fusion unit, and convert the node-level spatial topology feature tensor obtained in step S2-1 into a cross-modal gated fusion unit. and the sequence-level temporal evolution feature tensor obtained in step S2-2 As input, two types of features are adaptively fused through a gating mechanism;

[0057] First of all and Concatenate along the feature dimension, and compute the spatial feature gating weight matrix through a fully connected layer and a sigmoid activation function. Temporal feature gating weight matrix : ; ;

[0058] in, This indicates a concatenation operation along the feature dimension; and The learnable weight matrix has dimensionality 1. ; and It is the bias vector; It is a sigmoid activation function, and its output range is... ;

[0059] Adaptive weighted fusion of the two types of features is performed using gated weights to obtain the fused feature tensor. : ,in, This represents element-wise multiplication;

[0060] For fusion feature tensor A nonlinear activation function is applied for dimensionality reduction mapping to obtain a multi-scale business state feature tensor. Its dimensions are , Preset dimensions for multi-scale business state characteristics: ,in, The dimension-reduced projection matrix has dimensions of . ; It is the bias vector; The activation function for the rectified linear unit;

[0061] Step S2-4: Convert the multi-scale business state feature tensor Mapped to a predefined high-dimensional state manifold space; firstly, for Perform flattening operations on the time and node dimensions to obtain the flattened vector. Its dimensions are ; through a fully connected layer Mapping to a high-dimensional state manifold space yields a high-dimensional state vector. : ,in, The state mapping weight matrix has dimensions of . ; It is the bias vector; Let be the dimension of the high-dimensional state manifold space;

[0062] Introducing the observation noise covariance matrix For high-dimensional state vectors Gaussian perturbation is applied to simulate the uncertainty of state observations in a real-world business environment; the observation noise covariance matrix is... It is a diagonal matrix with dimension 1. Its diagonal elements Indicates the first Observation noise variance of the state components; perturbed state vector The calculation is as follows: ,in, To obtain from multivariate Gaussian distribution The observation noise vector in the middle sampling;

[0063] The perturbed state vector As a state observation space Instantiated representation; state observation space It includes the enterprise's current operational status subspace, the potential strategic intent subspace, and the external environment perception status subspace, which correspond to the following respectively. Component segments of different index intervals;

[0064] Step S2-5: Obtain preset enterprise resource constraints, including the upper limit of cash flow. Human resource quota and supply chain capacity threshold Based on enterprise resource constraints, a multi-dimensional continuous action boundary is constructed; the dimension of the tactical action vector is set to . Then the lower bound vector of the action and action upper bound vector All dimensions are The boundary of a multidimensional continuous action is determined by the resource constraint value corresponding to each action dimension; the boundary of the multidimensional continuous action is defined as: ,in, For tactical action vectors, inequalities represent element-wise constraints;

[0065] Within the action boundaries, an action masking mechanism is used to eliminate invalid action combinations that violate enterprise resource constraints; action mask vector. Each element takes a value of 0 or 1, where 0 indicates that the corresponding action dimension is unavailable and 1 indicates that it is available; the rule for determining the action mask is: if the action vector If the consumption in any resource dimension exceeds the corresponding constraint threshold, the mask vector is set to all 0s; otherwise, the mask vector is set to all 1s; effective tactical action space. Defined as: ,in, For dimension A vector of all 1s;

[0066] Tactical movement space It includes three types of tactical action subspaces: resource allocation action subspace. Market pricing action subspace and R&D investment action subspace The three types of subspaces together constitute the tactical action space. ;

[0067] Step S2-6: Construct a long-term delayed reward evaluation model based on eligibility trace tracking; the long-term delayed reward evaluation model consists of three components: immediate tactical benefit item. Long-term strategic alignment penalty items and resource consumption cost item ;

[0068] Real-time tactical benefits Defined as an intelligent agent executing tactical actions After state from Transfer to Instantaneous gains obtained at that time: ,in, The pre-defined basic return function;

[0069] Long-term strategic alignment penalty Used to measure the current state and steps. The degree of deviation between the strategic goal dimensions in the pre-defined corporate strategic goal ontology; let the strategic goal vector be... If the dimension is consistent with the state vector, then the long-term strategic alignment penalty term is defined as the state vector. With strategic objective vector Weighted cosine distance between them: ,in, Alignment penalty coefficient; Represents the dot product of vectors; The L2 norm of a vector;

[0070] Resource consumption cost item Defined as executing tactical actions The weighted sum of capital, human resources, and production capacity resources consumed: ,in, , , These are the weighting coefficients for capital, human resources, and production capacity costs, respectively. , , Actions The corresponding capital consumption function, human resource consumption function, and production capacity consumption function;

[0071] State observation space The state transition trajectory input to the long-period delayed reward evaluation model uses the eligibility trajectory decay mechanism to evaluate the immediate tactical benefit item. Perform multi-step credit allocation; set the eligibility trace decay factor as follows: The range of values ​​is ; at time step Qualification trace vector The update rules are as follows: ,in, This is a time discount factor, with a value range of [value range missing]. ; The state value function; For the state value function with respect to parameters The gradient;

[0072] Using time discount factor Penalties for long-term strategic alignment Delay decay calculations are performed to obtain the cumulative penalty value for multi-step discounts; finally, the three parts are merged to obtain the long-term delayed reward function. : ,in, This is the current state; For the current tactical action; The next state after the transition;

[0073] Step S2-7: Develop the state observation space constructed in step S2-4. The tactical action space constructed in steps S2-5 and the long-term deferred reward function constructed in steps S2-6 Perform joint encapsulation and define the state transition probability distribution for multiple agents. Construct a complete Markov game state space : ,in, A collection of intelligent agents, containing Each intelligent agent corresponds to a functional department within the enterprise. For each agent, there is a subspace of tactical actions. For tactical movement space The subset corresponding to the functions of the intelligent agent; This is a set of long-term delayed reward functions for each agent. Each agent shares the reward function structure but can have independent reward weight parameters. Let the state transition probability distribution be defined as follows: , indicating the state After all agents jointly execute the action vector, they transition to the state. The probability of;

[0074] At this point, step S2 is complete, and the complete Markov game state space is output. This state space integrates the graph structure topology of the enterprise's micro-operations, the temporal dynamics of the external macro environment, and the enterprise's resource constraints. It also achieves explicit modeling of the trade-off between short-term gains and long-term strategic goals in tactical decisions through a long-term delayed reward function, providing a standardized decision-making environment interface for the multi-agent game deduction in step S3.

[0075] In this embodiment, step S3 receives strategic decision-making instructions from senior management and decodes them into constraints of the tactical action space using a strategic semantic parsing model. In the Markov game state space constructed in step S2, a multi-agent near-end strategy optimization algorithm incorporating a risk entropy penalty term is used for game deduction. Furthermore, by applying counterfactual interventions to key external macro-environment nodes, a set of strategic evolution trajectories containing the baseline path and counterfactual perturbation paths is generated. Finally, the multi-objective risk-benefit utility matrix and causal vulnerability index for each path are calculated, providing a quantitative analysis basis for causal chain tracing and strategic decision-making report generation in step S4. The detailed steps are as follows:

[0076] Step S3-1: Receive strategic decision-making instructions from senior management. These instructions are described in natural language text and cover the company's strategic direction, priority goals, and constraints for a future period. Process these instructions using a pre-trained strategic semantic parsing model. The strategic semantic parsing model is built on a pre-trained language model architecture and fine-tuned on historical strategic documents and industry analysis reports.

[0077] First, entity identification and relationship extraction are performed on high-level strategic decision-making instructions. In the entity identification stage, entities related to enterprise departments, market regions, product lines, and time ranges are extracted from the instructions. In the relationship extraction stage, semantic relationships between entities are identified, including causal relationships, temporal relationships, and conditional dependencies. Based on the results of entity identification and relationship extraction, an instruction semantic graph is constructed, with entities as nodes and semantic relationships as edges. A graph convolution operation is performed on the instruction semantic graph to extract the semantic embedding vectors of the nodes. All node semantic embedding vectors are then averaged to obtain preliminary semantic vectors.

[0078] The initial semantic vector is then semantically aligned with the enterprise strategic objective ontology preset in step S1. An attention query mechanism is constructed using the semantic keyword vectors of each strategic dimension node in the enterprise strategic objective ontology, and the matching degree between the initial semantic vector and each strategic dimension is calculated through multi-head attention. ,in, For instruction semantics and the first Attention matching weights for each strategic dimension; This is a learnable attention query vector; This is the attention weight matrix; It is the bias vector; For the first part of the corporate strategic goal ontology Semantic keyword vectors for each strategic dimension node; Total number of nodes in the strategic dimension; It is the hyperbolic tangent activation function; It is a natural exponential function;

[0079] The semantic keyword vectors of strategic dimension nodes are weighted and summed using attention matching weights to obtain a strategic semantic context vector. This strategic semantic context vector is then concatenated with the initial semantic vector and mapped through a fully connected layer to a predefined strategic intent dimension space, resulting in a high-dimensional strategic intent vector. Its dimensions are ;

[0080] Step S3-2: Based on the high-dimensional strategic intent vector obtained in step S3-1 The multilayer perceptron decoder maps the data to a multidimensional action boundary space within the tactical action space, generating a set of tactical constraint boundaries. The multilayer perceptron decoder consists of three parallel decoder heads, each used to generate an action amplitude threshold vector. Action duration window vector and action priority sorting vector ;

[0081] Action amplitude threshold vector The dimension is equal to the tactical action space constructed in step S2. Dimensions Each of its components Restriction Maximum rate of change for each tactical action dimension: ,in, The weight matrix for decoding the action amplitude has a dimension of . ; It is the bias vector; The activation function is Softplus, which is used by default to ensure that the output is non-negative.

[0082] Action duration window vector The dimension is also Each of its components Constraints The upper and lower bounds of the execution cycle for each tactical action dimension are denoted as follows: and The result is obtained through the following decoding: ; ;

[0083] in, and Decode the weight matrix for duration; and It is the bias vector; for Activation function;

[0084] Action priority sorting vector The dimension is Each of its components Indicates the first The execution priority weight of each tactical action dimension; a larger value indicates that the action dimension is executed with higher priority during scheduling. Function normalization ensures that the sum of priority weights is 1. ,in, Decode the weight matrix for priority; It is the bias vector; for Normalization function;

[0085] motion amplitude threshold vector Action duration window vector and action priority sorting vector Combining them constitutes the tactical constraint boundary set. ;

[0086] Step S3-3: The Markov game state space constructed in step S2 Initialize the multi-agent policy network and multi-agent value network; define each agent in the multi-agent network as a functional department of the enterprise, including finance, marketing, R&D, and operations agents, and denot the total number of agents as . Multiple agents share the state observation space. However, each agent has its own independent tactical action subspace. tactical action subspace Subject to the tactical constraint boundary set generated in step S3-2 constraint;

[0087] No. Policy network of individual agents It adopts a multilayer perceptron architecture and observes local states. As input, output tactical action subspace The action probability distribution on; the policy network parameters are denoted as The action sampling process is represented as follows: ,in, For the first The vector of tactical actions taken by an agent;

[0088] Multi-agent value network Adopting a centralized architecture, with global state For input, i.e. For state observation space state vector Output the estimated cumulative reward of the global state; the value network parameters are denoted as... The value network output is represented as: ,in, The time discount factor follows the definition in step S2-6; This represents the expectation operation; For time offset index; For the long-term deferred reward function, the definition in step S2-6 is used;

[0089] Step S3-4: Construct a risk entropy penalty term to impose risk regularization constraints on the exploration behavior of the multi-agent policy network during game deduction; risk entropy penalty term Penalty sub-item based on policy distribution entropy and tail risk measurement sub-items Weighted composition: ,in, and These are the policy distribution entropy penalty coefficient and the tail risk metric coefficient, respectively, both with a default value of 0.5;

[0090] Policy distribution entropy penalty sub-item The Shannon entropy value is calculated based on the action probability distribution output by the multi-agent policy network; for the Policy network of individual agents Its state The Shannon entropy of the action probability distribution is defined as: ,in, In the state Select action The probability of; For the first The tactical action subspace of each agent;

[0091] The average of the policy distribution entropy penalty sub-items of all agents is taken: ;

[0092] Tail risk measurement sub-item The lower tail quantile of the cumulative reward distribution predicted by the multi-agent value network is calculated; let the cumulative reward sampling set obtained from multiple forward propagations of the multi-agent value network be . , This specifies the number of Monte Carlo samples; the default value is 100.

[0093] Sort the sample sets in ascending order and take the first... The value is used as the lower tail quantile. ,in, This is the lower tail quantile, with a default value of 0.05. For floor operations; the tail risk metric sub-item is defined as: ,in, To preset the revenue threshold, when When the tail risk metric is active, it generates a positive penalty value; otherwise, the tail risk metric is zero. This is a function to find the maximum value.

[0094] Step S3-5: Set the tactical constraint boundary set The policy network parameter update process of the multi-agent proximal policy optimization algorithm is injected; a centralized training and decentralized execution architecture is adopted, and the four stages of execution state observation, action sampling, reward calculation and network parameter update are iteratively executed to generate multiple sets of candidate policy evolution paths;

[0095] During the centralized training phase, the multi-agent value network Based on global state The cumulative reward of all agents is jointly evaluated; the loss function of the value network is defined as the mean square value of the temporal difference error. ,in, This serves as a buffer for experience replay. For time steps The state; For time steps Joint actions; For time steps Long-term delayed reward value ; For time steps The state;

[0096] During the policy network parameter update phase, a pruning agent objective function is introduced to limit the policy update step size, preventing drastic policy changes from causing training instability; for the ... Policy network of individual agents Its pruning agent objective function is defined as: , in, The probability ratio of the strategies. The parameters are the policy network parameters before the update; `clip` is the clipping function, which... Limited to Within the range; This is the cropping range parameter, with a default value of 0.2. The advantage function estimate is calculated based on the output of the multi-agent value network: ;

[0097] At the same time, the risk entropy penalty item will be added. Regularization constraints are applied to the policy gradient; the overall objective function of the policy network is: ,in, This is the weighting coefficient for the risk entropy penalty term, with a default value of 0.1;

[0098] During the decentralized execution phase, each agent observes its local state. Independently sampled tactical actions are not directly affected by the policies of other agents, but are subject to the tactical constraint boundary set. Constraints; The effective actions of an agent must satisfy: ,in, For tactical constraint boundary set The constrained first A subspace of effective tactical actions for each agent; For action The execution duration; For action Priority weights; This is a function to find the maximum value.

[0099] The iterative execution process consists of four phases: state observation, action sampling, reward calculation, and network parameter update. In each iteration, the current state, joint action, immediate reward, and next state are recorded and stored in the experience replay buffer. ;go through Round of simulation, The default value is 500, which generates multiple sets of candidate strategy evolution paths. Each set of paths contains a complete state transition sequence and action sequence.

[0100] Step S3-6: Select a baseline path from the multiple candidate strategy evolution paths generated in Step S3-5; the selection criteria for the baseline path are the minimum variance of strategy payoffs and the maximum mean of strategy rewards; let the first... The cumulative reward sequence of the candidate strategy evolution path is as follows: , The length of the path; calculate the average policy reward for the path. Sum of strategy return variance : ; ;

[0101] Variance of strategy returns And the average reward of the strategy Normalization was performed separately, and the baseline path comprehensive score was calculated. : ,in, and These are the variance penalty weight and the mean reward weight, respectively, both with a default value of 0.5. The normalized variance of the strategy return; This represents the normalized mean of the policy reward.

[0102] Select The largest candidate strategy evolution path is used as the baseline path, denoted as... ;

[0103] Extracting the baseline path The node sequences corresponding to the time series data of the external macro environment are used as the set of candidate intervention nodes. The external macroscopic environment node sequence is derived from the standardized macroscopic environment time-series data in step S1-2. The nodes are composed of the indicator dimensions, and each node corresponds to a macro-environment indicator.

[0104] Calculate the set of candidate intervention nodes using causal effect estimation methods The marginal causal influence of each node on the joint reward of the multi-agent system; for each node Apply a small perturbation to its value space Observe the rate of change of the joint reward of multiple agents and calculate the marginal causal influence. : ,in, Indicates at node When retrieving the current value, the first The long-term delayed reward value of each agent follows the definition in steps S2-6; The state after the perturbation is applied;

[0105] Screening marginal causal effects Exceeding the preset intervention threshold The nodes, as key intervention nodes, constitute the key intervention node set. : ,in, The default value is 0.3;

[0106] Step S3-7: Analyze the set of key intervention nodes identified in step S3-6. Counterfactual interventions are applied at each key intervention node; counterfactual interventions are implemented through a conditional generative adversarial network; the conditional generative adversarial network consists of a generator. and discriminator composition;

[0107] generator Based on the baseline path The historical macroeconomic environmental data distribution is used as a condition, and a random noise vector is input simultaneously. and context vectors under conditions Generate counterfactual values : ,in, From the standard normal distribution Mid-sampling; Key intervention nodes The context vector obtained by encoding the sequence of historical observations in the baseline path;

[0108] Discriminator It is used to distinguish between true observations and counterfactual generated values, and its output is the probability that the input data belongs to the true distribution: ,in, The input data may be actual observations or counterfactual generated values;

[0109] Conditional Generative Adversarial Networks (GANs) are optimized through adversarial training between the generator and the discriminator. After training, counterfactual values ​​are generated. The distribution of the counterfactual values ​​differs from the distribution of the original observations at the key intervention nodes, but remains within a reasonable range; the range of the counterfactual values ​​is set to the standard deviation of the historical observations at the key intervention nodes. Within the preset multiple range: ,in, This represents the historical average of key intervention nodes; The standard deviation of historical observations; This is the disturbance amplitude coefficient, with a default value of 2;

[0110] Obtain the set of counterfactual perturbation nodes Replace the baseline path with the set of counterfactual perturbation nodes. The original key intervention nodes are kept unchanged, while the states of other environmental nodes and the multi-agent policy network parameters in the baseline path are kept constant. The multi-agent game deduction process in steps S3-5 is re-executed to generate the counterfactual perturbation path. ; the baseline path and counterfactual perturbation path Combining these elements yields a set of strategic evolution trajectories. ;

[0111] Step S3-8: Extract the strategic evolution trajectory set For each path, construct a multi-objective risk-reward utility matrix based on the state transition sequence and cumulative reward sequence. Multi-objective risk-benefit utility matrix It includes four evaluation dimensions: financial return dimension Market share dimension Innovation Investment Dimension and risk exposure dimension ;

[0112] Financial returns It is calculated based on the net present value of the cumulative reward sequence; let the cumulative reward sequence be... , Given the path length, the net present value formula is: ,in, The discount rate is set to 0.05 by default.

[0113] Market share dimension The following is calculated based on the time-series mean of the market position indicator in the state transition sequence: ,in, For state The function extracts market position indicators and outputs scalar values ​​related to market share.

[0114] Innovation Investment Dimension The cumulative intensity of R&D investment actions within the tactical action space is calculated as follows: ,in, For time steps Action vector of the R&D investment action subspace; It is an L1 norm;

[0115] Risk Exposure Dimension It is calculated based on the downside volatility of the cumulative reward sequence; let... The mean of the cumulative reward series, and the downside volatility. The root mean square of the below-mean reward bias is defined as follows: Risk exposure dimension values This is the downward volatility. ;

[0116] Combining the four dimensions yields a multi-objective risk-reward utility matrix. For each path, For one The row vector;

[0117] Step S3-9: Calculate the baseline path and counterfactual perturbation path Utility difference vectors across dimensions of a multi-objective risk-reward utility matrix : ,in, The multi-objective risk-benefit utility matrix for the baseline path; A multi-objective risk-benefit utility matrix for counterfactual perturbation paths;

[0118] Calculate the intervention intensity vector of key intervention nodes The intervention intensity vector is based on each key intervention node. Counterfactual value generated Compared with the original observations The normalized difference is calculated as follows: ,in, Key intervention nodes The historical observation standard deviation is defined using the same method as in step S3-7;

[0119] Utility Difference Vector With intervention intensity vector Perform a dot product operation to obtain the scalar value of the causal effect. : ,in, for ∀ represents the transpose of a vector; · represents the dot product of vectors;

[0120] scalar value of causal effect Divide by the intervention intensity vector Normalize the modulus length to obtain the causal vulnerability index. : ,in, L2 norm; causal vulnerability index It reflects the sensitivity of the strategic evolution path to key nodes in the external macro environment. The larger the value, the more significant the impact of the node on multi-objective utility after being subjected to counterfactual disturbances, that is, the higher the vulnerability of the node in the causal structure.

[0121] At this point, step S3 is complete, and the strategic evolution trajectory set T and the multi-objective risk-benefit utility matrix for each path are output. and causal vulnerability index This provides complete quantitative analysis data for the causal chain tracing and multimodal strategic decision-making report generation in step S4.

[0122] In this embodiment, step S4 receives the strategic evolution trajectory set, multi-objective risk-benefit utility matrix, and causal vulnerability index output from step S3. It then uses a structural causal graph construction and path backtracking algorithm to trace the causal chain of key vulnerability nodes. Based on this, a multimodal generation engine is used to transform the quantitative analysis results into a multimodal strategic decision-making report containing natural language strategic narratives, strategic situation maps, and tactical resource allocation scheme tables, providing interpretable and quantifiable strategic decision-making support for senior management. The detailed steps are as follows:

[0123] Step S4-1: Receive the causal vulnerability index output from step S3. and the corresponding set of key intervention nodes Set up key intervention nodes The nodes in the sequence are sorted in descending order according to their corresponding causal vulnerability indices to obtain the sorted key intervention node sequence. ,in, The number of key intervention nodes. This indicates the ranking of the causal vulnerability index. The nodes are sorted by: ,in, For nodes The corresponding causal vulnerability index;

[0124] Based on the sorted key intervention node sequence Construct a structural cause-effect graph Structural cause-effect graph Given a directed acyclic graph with the following set of nodes: ,in, This is a set of auxiliary nodes, which are derived from the micro-operational graph structure data in step S1. The set of nodes and the set of agents in step S3 Corresponding functional department nodes; structural cause-effect diagram A directed edge represents a causal relationship between nodes, with the direction of the edge pointing from the cause node to the result node;

[0125] The directed edges are constructed using a constrained causal discovery algorithm; first, the conditional independence between node pairs is calculated; for any two nodes... and In a given set of conditions Under the condition of, calculate the partial correlation coefficient : ,in, For condition set A variable in; To remove The partial correlation coefficient after; if the partial correlation coefficient The absolute value is less than the independence threshold. The default value is 0.05, which is used for judgment. and In the given Independent under the condition of;

[0126] Based on the conditional independence test results, the PC algorithm is applied to generate a directed skeleton. Starting from a fully connected graph, the PC algorithm progressively removes edges between pairs of conditionally independent nodes, determining a V-shaped structure with causal directions. Finally, the direction of the remaining edges is determined through direction propagation rules, resulting in a structural causal graph. edge set ;

[0127] Structural cause-effect graph Represented as ,in For a set of nodes, It is a set of directed edges;

[0128] Step S4-2: The structural cause-effect graph constructed in step S4-1 In the middle, the sorted sequence of key intervention nodes Each node in the algorithm is the target node, and the causal chain path backtracking algorithm is executed. The causal chain path backtracking algorithm adopts a bidirectional breadth-first search strategy, starting from the target node and simultaneously tracing the cause node upward along the opposite direction of the directed edge and tracing the result node downward along the positive direction.

[0129] For the target node Define its upstream causal chain set For all elements in the structural cause-effect graph The set of directed paths ending at: ,in, This is the starting node of the path, i.e., the root node; This represents the path length. Indicates from node To the node The directed edges;

[0130] Define its downstream influence chain set For all elements in the structural cause-effect graph The set of directed paths originating from: ,in, The terminal node of the path, i.e., the node that ultimately affects it; This represents the path length.

[0131] Set up upstream causal chains and downstream impact chain set Merge to obtain nodes Complete set of causal chains : ;

[0132] For each causal chain path Calculate the strength of path causality The path causality strength is defined as the product of the causality coefficients of all directed edges on the path.

[0133] ,in, For path A directed edge on; For directed edges Causal effect coefficient; causal effect coefficient Estimated using a linear structural equation model; let the edges be... From node Pointing to node Its causal effect coefficient equal right Standardized regression coefficients: ,in, for and covariance; for The variance; and They are respectively and Standard deviation;

[0134] Based on the strength of path causality All causal chain paths are sorted, and the top ones with significant causal effects are selected. These paths, as the core causal chains, constitute the core causal chain set. ; The default value is 5;

[0135] Step S4-3: Select the core causal chain set from step S4-2. With the set of strategic evolution trajectories generated in step S3 Conduct correlation analysis to complete the attribution mapping from macro-environmental causal nodes to micro-tactical actions;

[0136] For each core causal chain Extracting the nodes from the path nodes that belong to the macro-environment time-series data indicators, i.e., macro-causal nodes, denoted as... And nodes belonging to the micro-operation graph structure data or nodes of the agent's functional departments, i.e., micro-attribution nodes, are denoted as... Establish a macro-causal node To the micro attribution node Attribution-related pairs;

[0137] For each attribution association pair, calculate the attribution contribution. The attribution contribution is calculated based on the magnitude of changes in the state of micro-attribution nodes caused by changes in macro-causal nodes in the counterfactual simulation; a baseline path is set. Micro-level attribution nodes At time step The state value is Counterfactual perturbation path The state value of the corresponding node at the same time step is The attribution contribution is defined as follows: ,in, The set of all micro-attribution nodes; and These are micro-attribution nodes. State values ​​in the counterfactual disturbance path and the baseline path;

[0138] Attribution matrix constructed based on attribution contribution. Its dimensions are ,in A set of macroscopic causal nodes; matrix elements Indicates the first The macroscopic causal node for the first Attribution contribution of each micro-attribution node: ;

[0139] Attribution matrix Row-wise normalization ensures that the sum of the attribution contributions of each macroscopic causal node to all microscopic attribution nodes is 1. ,in, This is the normalized attribution matrix; For the set of micro-attribution nodes The sum of its elements;

[0140] Based on the contribution ranking in the normalized attribution matrix, the nodes with the highest attribution contributions are associated with each macro-causal node. Each micro-level attribution node and its corresponding tactical action dimension constitutes an attribution mapping table from macro-level causal nodes to tactical actions. The default value is 3;

[0141] Step S4-4: Based on the core causal chain set in step S4-2 The attribution mapping table in step S4-3 and the multi-objective risk-benefit utility matrix in step S3. The system utilizes a pre-trained large language model to generate natural language strategic narratives. The large language model employs a domain-adaptive fine-tuning version based on a retrieval-enhanced generation framework, embedding an enterprise strategic knowledge base as an external retrieval source during its inference process.

[0142] Construct a prompt template structure, which includes four semantic segments: strategic background overview, causal vulnerability analysis, tactical action suggestions, and strategic outlook.

[0143] The strategic background overview section uses standardized macro-environmental time-series data from step S1. Recent trends and steps in the baseline path of S3 The state evolution overview serves as the input data source; the prompt template guides the large language model to generate macro-level descriptive text about the company's current strategic environment;

[0144] The causal vulnerability analysis section focuses on the core causal chain set. and causal vulnerability index The input data source is provided; the prompt template guides the large language model to describe the causal transmission logic of each core causal chain in turn, and to explain its corresponding causal vulnerability level; the causal chain description template contains the following structured fields: root cause node name, transmission path description, affected micro-node name, and causal vulnerability level; the rule for classifying the causal vulnerability level is: if The level is high; if The level is medium; if The level is low;

[0145] Tactical action recommendations are presented using attribution mapping tables and multi-objective risk-benefit utility matrices. The input data source is provided; the prompt template guides the large language model to generate specific tactical action suggestion text for each causal chain-related micro-attribution node and tactical action dimension, and uses the financial return dimension and risk exposure dimension data in the multi-objective risk-return utility matrix as the quantitative basis for the suggestion.

[0146] The strategic outlook section uses the set of strategic evolution trajectories generated in step S3. The long-term state prediction of the benchmark path and the deviation trend of the counterfactual perturbation path serve as input data sources; the prompt template guides the large language model to generate a prospective description of the enterprise's strategic prospects and clarifies the potential impact of external environmental uncertainties on the strategic implementation path.

[0147] The natural language text generated by the large language model undergoes post-bias removal processing, including terminology consistency verification and numerical accuracy verification. The terminology consistency verification ensures that all technical terms used in the generated text are consistent with the terms in the corporate strategic objective ontology preset in step S1. The numerical accuracy verification ensures that the quantitative indicators cited in the generated text are completely consistent with the values ​​in the multi-objective risk-benefit utility matrix and the causal vulnerability index.

[0148] Step S4-5: Based on structural cause-effect graph Core causal chain set and causal vulnerability index Generate a strategic situation map; the strategic situation map is presented in the form of a directed node link graph, where nodes represent key nodes in the structural causal graph, directed edges represent causal relationships, and the thickness of the edges corresponds to the strength of the causal effect of the path. The color intensity of a node is directly proportional to its causal fragility index. Proportional;

[0149] The graph layout employs a hierarchical force-directed layout algorithm. First, nodes are divided into several levels according to their topological depth in the structural cause-effect graph, with root cause nodes at the top level and terminal influence nodes at the bottom level. Then, within each level, the force-directed algorithm is used to optimize node coordinates, with the optimization objective function being: ,in, For nodes Two-dimensional coordinate vector; This is the repulsive force coefficient between nodes, with a default value of 1000. This is the attraction coefficient for directed edges, with a default value of 0.01.

[0150] The metadata annotations for the graph nodes include the node name, node type, causal vulnerability level label, and the core utility dimension corresponding to the node in the multi-objective risk-benefit utility matrix; the graph output format is scalable vector graphics.

[0151] Steps S4-6: Based on the attribution mapping table and the tactical constraint boundary set from step S3 Generate a tactical resource allocation plan table; the tactical resource allocation plan table contains three types of core resource allocation items: funding allocation items, human resource allocation items, and R&D resource allocation items; the funding allocation items are based on the tactical constraint boundary concentrated action amplitude threshold vector. The calculation of the cost constraints for each action dimension and the attribution contribution of each tactical action in the attribution mapping table; for the first... The first micro-attribution node associated with the Each tactical action dimension has its recommended capital allocation amount. Defined as: ,in, To be related to micro-attribution nodes The row index of the associated macro-causal nodes in the attribution matrix; The preset upper limit for cash flow in step S2; For the corresponding elements in the normalized attribution matrix;

[0152] Human resource allocation items are sorted by action priority vector And attribution contribution calculation; for the first Each micro-attribution node recommends a certain amount of human resource allocation. Defined as: ,in, The preset human resource quota for step S2;

[0153] The R&D resource allocation items are based on the innovation input dimension in the multi-objective risk-benefit utility matrix. and causal vulnerability index Weighted calculation; for the first Each micro-attribution node recommends a certain amount of R&D resource allocation. Defined as: ,in, The number of key intervention nodes; To be related to micro-attribution nodes Corresponding key intervention nodes;

[0154] The items for fund allocation, human resource allocation, and R&D resource allocation are sorted and integrated according to priority to generate a tactical resource allocation plan table. The plan table is arranged in descending order according to the causal vulnerability index of the micro-attribution nodes, and the resource allocation basis and expected utility improvement direction of each allocation item are marked.

[0155] Step S4-7: Compile the natural language strategic narrative generated in Step S4-4, the strategic situation map generated in Step S4-5, and the tactical resource allocation plan table generated in Step S4-6 to form a multimodal strategic decision-making report. The multimodal strategic decision-making report adopts a hierarchical document structure and includes four chapters: Strategic Situation Overview, Causal Vulnerability Diagnosis, Tactical Action Plan, and Resource Allocation Plan.

[0156] The Strategic Situation Overview chapter embeds a strategic situation map and a strategic background overview text, providing a macro-level visual overview of the company's current strategic environment;

[0157] The chapter on causal vulnerability diagnosis embeds a sub-graph of the topological structure of the core causal chain, a bar chart of the distribution of the causal vulnerability index, and a text section on causal vulnerability analysis, providing vulnerability diagnosis results for key external environmental nodes;

[0158] The tactical action plan chapter embeds tactical action suggestion paragraphs and attribution contribution heatmaps, providing tactical action guidance for specific functional departments;

[0159] The resource allocation plan chapter embeds a tactical resource allocation scheme table and a strategic outlook section, providing quantitative resource allocation recommendations and strategic outlook.

[0160] The multimodal strategic decision-making report is output in a portable document format. The output file contains interactive map layers and table layers, which support senior management personnel to zoom, filter and export. After the report is output, step S4 ends, and the entire processing flow of the OpenClaw framework is completed.

[0161] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A business data intelligence analysis method based on the OpenClaw framework for senior management decision-making in enterprises, characterized by: Includes the following steps: S1: Obtain the internal micro-operational diagram structure data and the external macro-environment time series data. Align the micro-operational diagram structure data and the macro-environment time series data to the preset corporate strategic goal ontology through the cross-modal tensor decomposition network of the OpenClaw framework, and construct a time series strategic intent tensor containing latent variables of strategic intent. S2: Based on the temporal strategic intent tensor, multi-scale business state features are extracted through the graph-temporal joint encoder of the OpenClaw framework, and combined with the preset enterprise resource constraints, a Markov game state space including state observation space, tactical action space and long-period delayed reward function is constructed. S3: Receives high-level strategic decision-making instructions and decodes them into the constraint boundaries of the tactical action space. In the Markov game state space, it uses a multi-agent proximal policy optimization algorithm with a risk entropy penalty term to perform game deduction. By applying counterfactual intervention to key external macroeconomic nodes, it generates a set of strategic evolution trajectories containing the baseline path and counterfactual perturbation paths, and calculates the multi-objective risk-benefit utility matrix and causal vulnerability index for each path. S4: Utilize the causal structure learning module of the OpenClaw framework to trace the causal chain of the strategic evolution trajectory set, identify key intervention nodes that lead to anomalies in the causal vulnerability index, generate a multimodal strategic decision report containing decision path explanations and confidence intervals, and decode high-level strategic decision instructions into an executable set of enterprise-level tactical KPI indicators to complete the decision-making loop.

2. The OpenClaw framework business data intelligent analysis method for enterprise senior management decision-making as described in claim 1, characterized in that: Acquire internal micro-operational diagram structure data and external macro-environment time series data. Align the micro-operational diagram structure data and macro-environment time series data to the preset corporate strategic goal ontology through the cross-modal tensor decomposition network of the OpenClaw framework, and construct a time series strategic intent tensor containing latent variables of strategic intent. Specifically, it includes: Organizational structure nodes, business process nodes, and resource flow nodes are extracted from enterprise resource planning systems and customer relationship management systems. Adjacency matrices are constructed based on the reporting, collaboration, and resource dependency relationships between these nodes. The adjacency matrices and the corresponding node attribute matrices are then combined to obtain micro-operational graph structure data. We collect time-series data on macroeconomic indicators, industry competition, and policy and regulatory changes from external data sources. After imputing missing values ​​and correcting outliers in these data, we perform time-granularity aligned sampling to obtain standardized macroeconomic environment time-series data. Microscopic operational graph structure data is input into a graph convolutional encoder for multi-order neighbor aggregation to obtain a microscopic structure embedding matrix; simultaneously, macroscopic environmental time series data is input into a temporal convolutional encoder for multi-scale sliding window convolution to obtain a macroscopic temporal embedding matrix. The microstructure embedding matrix and the macro temporal embedding matrix are concatenated along the feature dimension to form an initial fusion tensor. Tensor decomposition based on alternating least squares is performed on the initial fusion tensor to obtain the core tensor and a set of factor matrices including micro-dimensional factor matrices, macro-dimensional factor matrices and time-dimensional factor matrices. Based on the strategic goal dimension in the preset corporate strategic goal ontology, an ontology constraint mask matrix is ​​constructed. The ontology constraint mask matrix is ​​then used to perform orthogonal projection on the factor matrix set to obtain the aligned micro-dimensional factor matrix, the aligned macro-dimensional factor matrix, and the aligned time-dimensional factor matrix. The initial strategic intent tensor is obtained by tensor reconstruction of the core tensor, the aligned micro-dimensional factor matrix, the aligned macro-dimensional factor matrix, and the aligned time-dimensional factor matrix. The initial strategic intent tensor is then subjected to reparameterized sampling of the variational autoencoder in the time dimension. The sampled latent variable vector is injected into the feature channel of the initial strategic intent tensor to obtain a temporal strategic intent tensor containing strategic intent latent variables.

3. The OpenClaw framework business data intelligent analysis method for enterprise senior management decision-making as described in claim 1, characterized in that: Based on the temporal strategic intent tensor, multi-scale business state features are extracted using a graph-temporal joint encoder within the OpenClaw framework; specifically including: The temporal strategic intent tensor is input into the spatial graph attention subnetwork of the graph-temporal co-encoder. The spatial dependency attention weights of each node in the temporal strategic intent tensor at different time steps are calculated using the dynamic graph attention mechanism. The node features are then weighted and aggregated according to the spatial dependency attention weights to obtain node-level spatial topological features. The temporal strategic intent tensor is input into the multi-scale temporal dilated convolutional sub-network of the graph-temporal joint encoder. By setting causal convolutional kernels with different dilation rates, the temporal strategic intent tensor is subjected to multi-scale time window sliding extraction to obtain sequence-level temporal evolution features containing short-term fluctuation features and long-term trend features. A cross-modal gating fusion unit is constructed, which takes node-level spatial topology features and sequence-level temporal evolution features as inputs, and calculates spatial feature gating weights and temporal feature gating weights through the cross-modal gating fusion unit; By using spatial feature gating weights and temporal feature gating weights, node-level spatial topological features and sequence-level temporal evolution features are adaptively weighted and concatenated, and then dimensionality reduction mapping is performed through a nonlinear activation function to obtain multi-scale business state features.

4. The OpenClaw framework business data intelligent analysis method for enterprise senior management decision-making as described in claim 3, characterized in that: Based on pre-defined enterprise resource constraints, a Markov game state space is constructed, comprising a state observation space, a tactical action space, and a long-period delayed reward function; specifically including: Multi-scale business state features are mapped to a pre-defined high-dimensional state manifold space, and the observation noise covariance matrix is ​​introduced to perform Gaussian perturbation on the mapped features, thereby constructing a state observation space that includes the current operating status of the enterprise, potential strategic intentions, and perception of the external environment. Obtain preset enterprise resource constraints, including cash flow limits, human resource quotas, and supply chain capacity thresholds; based on these constraints, construct a multi-dimensional continuous action boundary, and within this boundary, use an action masking mechanism to eliminate invalid action combinations that violate the constraints, generating a tactical action space that includes resource allocation actions, market pricing actions, and R&D investment actions. A long-term delayed reward evaluation model based on eligibility trace tracking is constructed. The long-term delayed reward evaluation model includes an immediate tactical benefit item, a long-term strategic alignment penalty item, and a resource consumption cost item. The state transition trajectory in the state observation space is input into the long-term delayed reward evaluation model. The immediate tactical benefit item is assigned credit in multiple steps through the eligibility trace decay mechanism, and the long-term strategic alignment penalty item is calculated for delay decay using the time discount factor. Finally, the resource consumption cost item is fused to obtain the long-term delayed reward function. The state observation space, tactical action space, and long-period delayed reward function are jointly encapsulated, and the state transition probability distribution of multiple agents is set to construct the state space of a Markov game.

5. The OpenClaw framework business data intelligent analysis method for enterprise senior management decision-making as described in claim 1, characterized in that: It receives strategic decision-making instructions from high-level authorities and decodes them into the constraint boundaries of the tactical action space. In the state space of a Markov game, it uses a multi-agent proximal policy optimization algorithm with a risk entropy penalty term to perform game deduction. Specifically, it includes: Upon receiving strategic decision-making instructions from high-level authorities, a pre-trained strategic semantic parsing model is used to identify entities and extract relationships from these instructions, extracting a set of strategic keywords. The set of strategic keywords is then mapped to a pre-defined strategic intent dimension space for semantic embedding, resulting in a high-dimensional strategic intent vector. Based on the high-dimensional strategic intent vector, the high-dimensional strategic intent vector is mapped to the multi-dimensional action boundary space of the tactical action space through the multilayer perceptron decoder, generating a tactical constraint boundary set containing action amplitude threshold vector, action duration window vector, and action priority ranking vector. The action amplitude threshold vector limits the maximum rate of change of each tactical action, the action duration window vector constrains the execution cycle of each tactical action, and the action priority ranking vector specifies the execution time sequence dependency of each tactical action. In the Markov game state space, initialize the multi-agent policy network and multi-agent value network. Set each agent in the multi-agent network to correspond to a functional department of the enterprise. Multiple agents share the state observation space but have independent tactical action subspaces. The tactical action subspaces are constrained by the tactical constraint boundary set. A risk entropy penalty term is constructed, which includes a policy distribution entropy penalty sub-term and a tail risk metric sub-term. The policy distribution entropy penalty sub-term is calculated based on the Shannon entropy value of the action probability distribution output by the multi-agent policy network. The tail risk metric sub-term is calculated based on the lower tail quantile of the cumulative reward distribution predicted by the multi-agent value network. The risk entropy penalty term is activated when the variance of the action probability distribution exceeds a preset risk threshold or the lower tail quantile is lower than a preset reward threshold. The tactical constraint boundary set is injected into the policy network parameter update process of the multi-agent proximal policy optimization algorithm. The policy update step size is limited by pruning the agent objective function. At the same time, the policy gradient is regularized by using the risk entropy penalty term. A centralized training and decentralized execution architecture is adopted. Each agent independently samples tactical actions based on local state observations. The multi-agent value network jointly evaluates the cumulative rewards of the multi-agents based on the global state. The game inference process of state observation-action sampling-reward calculation-network parameter update is iteratively executed to generate multiple sets of candidate strategic evolution paths.

6. The OpenClaw framework business data intelligent analysis method for enterprise senior management decision-making as described in claim 5, characterized in that: By applying counterfactual interventions to key external macroeconomic nodes, a set of strategic evolution trajectories containing the baseline path and counterfactual disturbance paths is generated. A multi-objective risk-benefit utility matrix and a causal vulnerability index are calculated for each path; specifically including: From multiple candidate strategic evolution paths, the candidate strategic evolution path with the smallest strategy return variance and the largest strategy reward mean is selected as the benchmark path, and the node sequence corresponding to the time series data of the external macro environment in the benchmark path is extracted as the candidate intervention node set. The marginal causal influence of each node in the candidate intervention node set on the joint reward of the multi-agent agent is calculated using the causal effect estimation method. The marginal causal influence is obtained by applying a small perturbation to the value space of the candidate intervention node set and observing the rate of change of the joint reward of the multi-agent agent. Nodes with a marginal causal influence exceeding the preset intervention threshold are selected as key intervention nodes. Counterfactual intervention is applied to key intervention nodes. The counterfactual intervention is achieved through a conditional generative adversarial network. The conditional generative adversarial network generates counterfactual generated values ​​that are different from the original observation values ​​of the key intervention nodes, based on the historical macroeconomic environmental data distribution of the baseline path. The range of the counterfactual generated values ​​is set within a preset perturbation interval of the standard deviation multiple of the historical observation values ​​of the key intervention nodes, thus obtaining a set of counterfactual perturbation nodes. Replace the original key intervention nodes in the baseline path with the counterfactual perturbation node set, keep the states of other environmental nodes and the multi-agent policy network parameters in the baseline path unchanged, re-execute the Markov game inference process, generate the counterfactual perturbation path, and combine the baseline path and the counterfactual perturbation path to obtain the strategic evolution trajectory set. Extract the state transition sequence and cumulative reward sequence for each path in the strategic evolution trajectory set, and construct a multi-objective risk-return utility matrix. The dimensions of the multi-objective risk-return utility matrix include financial return dimension, market share dimension, innovation input dimension, and risk exposure dimension. The financial return dimension is calculated based on the net present value of the cumulative reward sequence, the market share dimension is calculated based on the time-series mean of the market position indicator in the state transition sequence, the innovation input dimension is calculated based on the cumulative intensity of R&D investment actions in the tactical action space, and the risk exposure dimension is calculated based on the downside volatility of the cumulative reward sequence. The utility difference vectors of the baseline path and the counterfactual perturbation path on each dimension of the multi-objective risk-benefit utility matrix are calculated. The intervention intensity vector of key intervention nodes is calculated. The intervention intensity vector is calculated based on the normalized difference between the counterfactual generated value and the original observation value. The causal effect scalar value is obtained by performing a dot product operation on the utility difference vector and the intervention intensity vector. The causal effect scalar value is normalized by dividing the causal effect scalar value by the magnitude of the intervention intensity vector to obtain the causal vulnerability index. The causal vulnerability index reflects the sensitivity of the strategic evolution path to key external macroeconomic environmental nodes.

7. The OpenClaw framework business data intelligent analysis method for enterprise senior management decision-making as described in claim 1, characterized in that: Using the causal structure learning module of the OpenClaw framework, causal chain tracing is performed on the strategic evolution trajectory set to identify key intervention nodes leading to anomalies in the causal vulnerability index, generating a multimodal strategic decision-making report that includes decision path explanations and confidence intervals; specifically including: The baseline path and counterfactual perturbation path of the strategic evolution trajectory set are mapped into a temporal causal graph structure. The state nodes in the state observation space, the action nodes in the tactical action space, and the external macro environment nodes are used as the vertices of the temporal causal graph. Causal edge weights are constructed based on the action transition probability of the multi-agent policy network and the reward prediction gradient of the multi-agent value network, forming a temporal causal graph containing three levels: state, action, and environment. A causal structure learning algorithm based on continuous optimization is performed on the temporal causal graph. The causal adjacency matrix is ​​iteratively optimized by minimizing the causal sufficiency constraint and the structural sparsity regularization term to obtain the converged causal skeleton graph. On the causal skeleton graph, the end state of the counterfactual perturbation path where the causal vulnerability index exceeds the preset anomaly threshold is used as the backtracking starting point. The set of causal chains from the external macroscopic environment node to the end state is obtained by searching in reverse along the causal edge direction. The contribution of path causal effect is calculated for each causal chain in the causal chain set. The contribution of path causal effect is obtained by multiplying and summing the conditional mutual information of each node in the chain and the strength of counterfactual perturbation. Causal chains whose contribution of path causal effect exceeds the preset contribution threshold are marked as abnormal causal chains. The external macro-environment node at the top of the abnormal causal chain is extracted as the key intervention node that causes the abnormal causal vulnerability index. A decision path explanation text is constructed based on key intervention nodes and abnormal causal chains. The decision path explanation text includes semantic descriptions of key intervention nodes, natural language expressions of causal transmission paths, and descriptions of expected changes in each dimension of the multi-objective risk-reward utility matrix. At the same time, Monte Carlo dropout inference is performed on the multi-agent policy network. While keeping the tactical constraint boundary set unchanged, the multi-agent policy network is subjected to multiple random forward propagations to obtain multiple sampled estimates of the cumulative reward distribution. The confidence interval of the cumulative reward is calculated based on the quantiles of the sampled estimates. The decision path explanation text, confidence intervals, intervention intensity vectors of key intervention nodes, multi-objective risk-benefit utility matrix, and comparative visualization charts of counterfactual perturbation paths and baseline paths are integrated into a multimodal strategic decision report. The multimodal strategic decision report includes four modal sections: structured data tables, causal chain flowcharts, risk heatmaps, and decision recommendation summaries.

8. The OpenClaw framework business data intelligent analysis method for enterprise senior management decision-making as described in claim 7, characterized in that: Decoding high-level strategic decision-making directives into a set of executable, low-level tactical KPI indicators to complete the decision-making loop; specifically including: Strategic adjustment semantic vectors are extracted from the decision recommendation summary of the multimodal strategic decision report. Attention fusion is performed between the strategic adjustment semantic vectors and the causal chain flowchart in the multimodal strategic decision report to obtain the strategic adjustment intention encoding containing causal constraints. The strategic adjustment intention is encoded and input into a pre-defined strategic-tactical mapping knowledge graph. The strategic-tactical mapping knowledge graph contains enterprise functional department nodes, underlying tactical action nodes, and causal constraint edges. Subgraph retrieval is performed on the strategic-tactical mapping knowledge graph through a graph neural network to match a subset of underlying tactical actions that are semantically similar to the encoded strategic adjustment intention and satisfy the causal constraint edges. For each underlying tactical action in the subset of underlying tactical actions, extract the action amplitude threshold vector, action duration window vector, and action priority sorting vector from the corresponding tactical constraint boundary set. Calculate the target execution value of each underlying tactical action based on the action amplitude threshold vector, set the execution time node of each underlying tactical action based on the action duration window vector, and generate the execution order dependency table of each underlying tactical action based on the action priority sorting vector. Combine the target execution value, execution time node, and execution order dependency table to obtain the tactical KPI index of each underlying tactical action. The tactical KPIs for each underlying tactical action are distributed to the corresponding enterprise functional departments' execution systems. Simultaneously, real-time feedback collection rules are set based on these KPIs, including the update frequency of the state observation space and the feedback latency of reward feedback. When an enterprise functional department's execution system executes an underlying tactical action according to the KPIs, it collects the actual execution status and actual accumulated rewards. The deviation between the actual execution status and the expected status in the state observation space is calculated to obtain an execution deviation vector. When the execution deviation vector exceeds a preset deviation threshold, the strategic semantic analysis model is triggered to re-analyze high-level strategic decision-making instructions, initiating a new round of game theory and report generation, forming a closed-loop decision-making process from high-level strategic decision-making instructions to the execution of underlying tactical KPIs and then to strategic adjustments.

Images