Method and system for constructing enterprise operation system based on dynamic credible causal graph evolution

Abstract

The application discloses an enterprise operation system construction method and system based on dynamic trusted causal diagram evolution, acquires a cloud resource load sequence, constructs a trusted causal diagram through adaptive semantic segmentation, causal structure learning under the constraint of a directed acyclic graph, time sequence causal evolution perception and confidence evaluation, generates time feature embedding and relationship feature embedding in combination with process attention feature extraction and graph attention propagation, and outputs a cloud resource load prediction result through conditional spatio-temporal fusion and cross-section time sequence modeling; the application explicitly mines a directed causal structure among resource indexes and integrates the same into a prediction process as a prior constraint, overcomes the prediction inaccuracy problem caused by learning of false association in the prior art, realizes dynamic tracking and reliability quantization of a causal topology, and significantly improves the accuracy, robustness and explainability of cloud resource load prediction.

Description

Technical Field

[0001] This invention relates to the field of dynamic trusted causal graph technology, and in particular to a method and system for constructing an enterprise operation system based on dynamic trusted causal graph evolution. Background Technology

[0002] In enterprise operations management, the construction of an enterprise operations system is a core prerequisite for ensuring efficient operation and the achievement of strategic goals. This is achieved by analyzing the evolution of multiple core business indicators across various dimensions of business data over time to estimate future business trends and generate operational decision-making plans. However, existing system construction methods either rely on time-series statistical models or recurrent neural networks to model time-series evolution patterns, making it difficult to characterize the complex interactions between cross-business indicators; or they use attention mechanisms or graph neural networks to model the correlations between business indicators. While these methods can capture correlations, they struggle to distinguish between spurious correlations and genuine business causal mechanisms. They cannot explicitly uncover the directed causal structure between business indicators, nor can they dynamically track the evolution of business causal relationships and quantify their reliability. This leads to inaccurate and uninterpretable decisions in scenarios involving changes in the market environment or adjustments to business strategies, failing to meet the needs of enterprise management for decision-making basis and intervention analysis. Summary of the Invention

[0003] In view of this, the purpose of this invention is to propose a method and system for constructing an enterprise operation system based on dynamic trusted causal graph evolution.

[0004] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is as follows: In a first aspect, the present invention provides a method for constructing an enterprise operating system based on dynamic trusted causal graph evolution, comprising: Acquire multi-dimensional business data of enterprise operations; record the evolution of various core business indicators of the enterprise over time. Adaptive semantic segmentation is performed on multi-dimensional business data of enterprise operations to generate a sequence of business semantic segments; Business causal logic is learned based on the sequence of business semantic segments, and a business causal association matrix of each business semantic segment is constructed under the constraint of a directed acyclic graph, resulting in a sequence of business causal graphs. The system perceives the causal evolution of business cause-effect graph sequences, captures the dynamic evolution law of business cause-effect structure, and predicts the future changes in business cause-effect structure. Based on the future changes in the causal structure of business, an evolutionary prior is generated and fed back to the business causal logic learning process to guide the updating of the business causal graph along the historical evolution trend. The confidence of each business causal edge in the business causal graph sequence is evaluated to generate a business causal confidence matrix. The business causal relationship matrix is ​​weighted based on the business causal confidence matrix to generate a reliable business causal graph; Business process attention features are extracted from the business semantic segment sequence to generate business time feature embeddings. Business relationship attention is propagated along the trusted business causal graph to generate business relationship feature embeddings. Conditional spatiotemporal fusion of business time feature embedding and business relationship feature embedding is performed to generate joint business spatiotemporal feature embedding; Cross-segment business time-series modeling is performed on the spatiotemporal feature embedding of joint business to generate a global business situation representation. Based on the global business situation representation, business decision reasoning mapping is performed to generate and output enterprise operation decision schemes.

[0005] In some embodiments, adaptive semantic segmentation is performed on multi-dimensional business data of enterprise operations to generate a sequence of business semantic segments, including: Slide segmented detection windows across multi-dimensional business data of enterprise operations, with each segmented detection window forming an adjacent window pair with the business data in the previous window. For adjacent window pairs, calculate the difference between the average business data within the segmented detection window and the average business data within the previous window to obtain the mean change; The difference between the fluctuation level of business data within the segmented detection window and the fluctuation level of business data within the previous window is calculated to obtain the fluctuation change. The Jensen-Shannon divergence quantification is used to measure the overall difference between the distribution of business data within the segmented detection window and the distribution of business data within the previous window, thus obtaining the change in distribution pattern. After assigning corresponding weight coefficients to mean change, volatility change, and distribution pattern change, the weighted sums are used to generate a comprehensive change score. The comprehensive change score represents the degree of non-stationarity of business data within the segmented detection window. When the overall change score exceeds the judgment threshold, a business mode change occurs at the segment detection window, and the starting boundary of the segment detection window is marked as a candidate segment point. Based on the minimum segment length constraint, adjacent candidate segment points with a spacing less than the minimum segment length are merged, and the candidate segment points are corrected to obtain the segment boundary. The enterprise's multi-dimensional business data is divided according to the segmentation boundaries to generate a sequence of business semantic segments.

[0006] In some embodiments, business causal logic learning is performed based on the business semantic segment sequence, and a business causal association matrix for each business semantic segment is constructed under the constraint of a directed acyclic graph, resulting in a business causal graph sequence, including: A parallel dual-branch business causality perception module is constructed, which consists of a multi-core temporal business causality convolutional branch and a multivariate real-time business causality attention branch. By using a multi-core temporal business causal convolution branch, business causal convolution is performed on each business variable pair in the business semantic segment sequence to obtain the time-delay business causal strength of each business variable pair. Left padding is then performed on the input sequence to ensure that the convolution field of view only covers the current time and historical time, thus preventing the leakage of future information. Through the multivariate real-time business causal attention branch, attention is calculated using the current state of each business variable as the query and key to obtain the real-time business causal strength of each business variable pair. The learnable mask matrix is ​​then applied element by element to the attention score to control the presence and strength of business causal connections between each business variable pair. The business causal intensity of time-delayed business is aggregated with the business causal intensity of real-time business, and the weights corresponding to redundant business causal paths are driven to approach zero by sparse regularization to generate a business causal correlation matrix. Apply a continuous directed acyclic graph constraint to the business causal relationship matrix to generate a business causal relationship matrix that satisfies the directed acyclic graph structure; Perform threshold-based edge pruning on the business causal relationship matrix that satisfies the directed acyclic graph structure to remove weak business causal connections; Inject domain prior knowledge into the business causal relationship matrix after edge clipping, and reset the corresponding association weights of business variables that are known to have no business causal relationship to zero; The above construction and optimization are performed on each business semantic segment in the business semantic segment sequence to obtain the business causal relationship matrix of each business semantic segment, which constitutes the business causal graph sequence.

[0007] In some embodiments, business causal evolution perception is performed on the business causal graph sequence to capture the dynamic evolution law of the business causal structure and predict the future changes in the business causal structure, including: The evolution of the business cause-effect graph sequence on the time axis is explicitly modeled. The business cause-effect association matrix of each business semantic segment in the business cause-effect graph sequence is flattened into a one-dimensional vector along the business variable dimension. The business cause-effect connection strength between each business variable pair is arranged in a fixed order to generate a business cause-effect graph embedding vector sequence. The business cause-effect graph embedding vector sequence is input into the gated loop unit in chronological order. At each time step, the gated loop unit, based on the current business cause-effect graph embedding vector and the hidden state of the previous time step, controls the retention of historical evolution information by updating the gate and controls the forgetting of historical evolution information by resetting the gate. The hidden state is updated by weighted fusion of the candidate hidden state and the hidden state of the previous time step, generating a hidden state sequence. Hidden state sequences characterize the dynamic evolution of the business causal topology over time. The dynamic evolution includes the strengthening trend, weakening trend, and migration trend of the business causal edges over time. By using an evolutionary prediction network to predict and map the hidden state sequence, the direction and magnitude of the change in the business causal correlation matrix of future business semantic segments relative to the current moment are estimated, and the change in the future business causal structure is generated. The change in the future business causal structure quantifies the strengthening, weakening, addition and disappearance trends of future business causal edges, and serves as a priori estimate of the business causal structure at future moments.

[0008] In some embodiments, an evolutionary prior is generated based on future changes in the business causal structure. This evolutionary prior is then fed back into the business causal logic learning process to guide the updating of the business causal graph along historical evolutionary trends. Confidence assessments are performed on each business causal edge in the business causal graph sequence to generate a business causal confidence matrix, including: An evolutionary prior is constructed based on the changes in the causal structure of future business. The evolutionary prior represents the prior constraint that the causal structure of business evolves smoothly and continuously along the historical evolution trend between adjacent business semantic segments. Evolutionary priors are fed back as regularization constraints into the loss function of the business causal logic learning process, so as to constrain the deviation between the actual changes in the business causal structure and the future changes in the business causal structure. By using deviation constraints, the business causal relationship matrix of subsequent business semantic segments is updated in a direction consistent with the historical evolution trend while ensuring the accuracy of decision-making, thereby suppressing abrupt changes in the business causal structure between adjacent business semantic segments. A business causal confidence assessment network is constructed by flattening the business causal association matrix of each business semantic segment in the business causal graph sequence into a one-dimensional vector, which serves as the input to the business causal confidence assessment network. The business causal confidence assessment network independently scores the confidence of each business causal edge corresponding to each pair of business variables in each business semantic segment, and obtains the business causal edge confidence score. The business causal edge confidence score represents the reliability of the existence of a business causal relationship between the corresponding pairs of business variables. The confidence scores of the business causal edges of each business variable pair in each business semantic segment are collected to generate a business causal confidence matrix. The business causal confidence matrix is ​​used to characterize the credibility of each business causal edge.

[0009] In some embodiments, the business causal correlation matrix is ​​weighted based on the business causal confidence matrix to generate a reliable business causal graph, including: The confidence scores of the business causal edges corresponding to each business variable pair in the business causal confidence matrix are multiplied element-wise by the business causal connection strength of the corresponding business variable pair in the business causal association matrix, so that the positions of each business variable pair in the business causal confidence matrix and the business causal association matrix correspond one-to-one. Using the confidence score of the business causal edge as a weight, the business causal connection strength of each business causal edge is scaled so that the effective connection strength of each business causal edge is determined by both the business causal connection strength and the confidence score. The higher the confidence score, the stronger the effective connection strength. Generate a confidence-weighted business causal relationship matrix. The values ​​of the elements in the confidence-weighted business causal relationship matrix represent the effective connection strength of the corresponding business causal edges after confidence scaling. In the confidence-weighted business causal relationship matrix, business causal edges with high confidence scores dominate, while business causal edges with low confidence scores are continuously and adaptively suppressed according to their confidence scores. By continuously and adaptively suppressing false business causal connections with low credibility, and retaining true business causal connections with high credibility, the interference of false associations on the business causal structure is suppressed. The confidence-weighted business causal relationship matrix is ​​used as a trusted business causal graph. The weight of each business causal edge in the trusted business causal graph reflects both the strength of the business causal connection and the degree of trust in the business causal relationship.

[0010] In some embodiments, business process attention features are extracted from the business semantic segment sequence to generate business time feature embeddings, and business relationship attention is propagated along the trusted business causal graph to generate business relationship feature embeddings, including: Construct a Transformer-based business process attention network, which consists of sequentially stacked business process attention layers and feedforward layers; The sequence of business semantic segments is encoded with independent position and cross-channel position. Independent position encoding represents the positional relationship within each business semantic segment, while cross-channel position encoding represents the relative position between different business indicators. The two are then fused to generate a sequence position representation. The business process attention layer maps the sequence position representation into query vector, key vector and value vector, and performs self-attention calculation. The attention calculation adopts orthogonal basis function expansion approximation to capture long-term business time-series dependencies within and across segments and generate business time feature embeddings. The time dimension of the original sequence matrix of each business variable in the trusted business causal graph is mapped and compressed through a fully connected layer to construct the initial node representation; Using a trusted business causal graph as the adjacency matrix, a graph attention mechanism is applied to the initial node representation to aggregate the business causal neighbor information of each business variable and update the node representation layer by layer. The attention coefficient of the graph attention mechanism is obtained by linear transformation of the business causal neighbor node representation pair. The attention coefficient quantifies the propagation strength of the business causal effect along the business causal path. The updated node representations are aggregated using global mean pooling to generate embedded business relationship features.

[0011] In some embodiments, conditional spatiotemporal fusion is performed on the business time feature embedding and the business relationship feature embedding to generate a joint business spatiotemporal feature embedding, including: The business time feature embedding and the business relationship feature embedding are concatenated along the feature dimension to generate an initial joint representation; A gated cross-attention mechanism is used to perform bidirectional conditional information interaction between the embedding of business time features and the embedding of business relationship features. The initial joint representation is transformed into query vector, key vector and value vector through linear transformation. A gate signal is generated by linear transformation of the embedded business time features. The gate signal is used to dynamically modulate the aggregation strength of the key value of the embedded business relationship features. The gate signal determines the strength of each key value information in the embedded business relationship features participating in the aggregation, thereby realizing the conditional control of the aggregation of business relationship features by business time features. The query vector is generated by embedding business relationship features. The query vector guides the focusing direction of business time feature embedding on different frequency domain components, thereby strengthening the frequency domain components related to the business causal structure in the business time feature embedding and realizing the conditional guidance of business relationship features to focus on business time features. The bidirectional cross-attention outputs obtained from bidirectional conditional information interaction are fused after being weighted and modulated by a gated signal, and the single-head computation is extended to multi-head parallel computation through a multi-head mechanism. The output of multi-head parallel computation is processed by feedforward network and residual connection to output joint business spatiotemporal feature embedding.

[0012] In some embodiments, cross-segment business time-series modeling is performed on the spatiotemporal feature embedding of joint business to generate a global business situation representation. Based on the global business situation representation, business decision reasoning mapping is performed to generate and output an enterprise operation decision plan, including: The spatiotemporal features of joint business are embedded and arranged in chronological order of the semantic segments of each business in the original sequence to construct a segmented business feature sequence. The segmented business feature sequence is input into the cross-segment business temporal modeling module, which is composed of a Transformer encoder. The cross-segment business temporal modeling module is used to model the long-term dependency relationship between each business semantic segment and the cross-segment business context evolution mode. Temporal position information is injected into each business semantic segment in the segmented business feature sequence by position encoding in order to preserve the temporal order between each business semantic segment; By utilizing a multi-head self-attention mechanism, attention computation is extended to parallel computation of multiple attention heads to calculate the attention weights between each business semantic segment in the segmented business feature sequence, thereby capturing the long-range dependencies and contextual evolution patterns between each business semantic segment. The output of the multi-head self-attention is aggregated through multiple layers of encoder to output a global business situation representation that characterizes the spatiotemporal features of the entire historical input window. The global business situation representation is input into a multilayer perceptron projection network consisting of fully connected layers, and the global business situation representation is mapped to the decision space through linear transformation. Output enterprise operation decision-making solutions, which include optimization and adjustment strategies and resource allocation suggestions for multiple core business indicators, as the result of the enterprise operation system construction and output.

[0013] In a second aspect, the present invention provides a system for constructing an enterprise operation system based on dynamic trusted causal graph evolution, for executing the method of the first aspect, the system comprising: The adaptive business semantic segmentation module is used to adaptively segment multi-dimensional business data of enterprise operations. It uses a segmentation detection window to detect changes in mean, volatility, and distribution patterns between adjacent windows to determine segmentation boundaries and generate a sequence of business semantic segments. The dual-channel business feature extraction network consists of a business time channel and a business relationship channel. The business time channel includes a business process attention network, which is used to extract business process attention features from the business semantic segment sequence and generate business time feature embeddings. The business relationship channel includes a parallel dual-branch business causality perception module, a directed acyclic graph constraint optimization module, a business temporal causal evolution perception module, a business causal confidence evaluation module, and a business relationship attention propagation module. The parallel dual-branch business causality perception module is used to learn the business causal association matrix of each business semantic segment from the business semantic segment sequence. The directed acyclic graph constraint optimization module is used to apply directed acyclic graph constraints to the business causal association matrix. The business temporal causal evolution perception module is used to capture the dynamic evolution law of the business causal structure, predict the future changes in the business causal structure, and feed the future changes in the business causal structure as evolutionary priors to the business causal logic learning process. The business causal confidence evaluation module is used to generate business causal edge confidence scores for each business causal edge. The business relationship attention propagation module is used to propagate information along the confidence-weighted business causal association matrix and generate business relationship feature embeddings. The conditional business spatiotemporal fusion module is used to perform gated cross-attention bidirectional conditional fusion of business time feature embedding and business relationship feature embedding to generate joint business spatiotemporal feature embedding. The decision-making scheme output module is used to perform cross-segment business time-series modeling based on the spatiotemporal feature embedding of joint business, generate a global business situation representation, and output enterprise operation decision-making schemes based on the global business situation representation through business decision reasoning mapping.

[0014] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: This invention acquires multi-dimensional business data from enterprise operations, performs adaptive semantic segmentation on the data to generate a sequence of business semantic segments, performs business causal logic learning based on these segments, and constructs a business causal association matrix for each segment under a directed acyclic graph constraint to obtain a sequence of business causal graphs. The invention then performs business causal evolution perception on these graphs to capture the dynamic evolution of the business causal structure and predict future changes in that structure. Based on these future changes, an evolutionary prior is generated and fed back into the business causal logic learning process. Simultaneously, confidence assessment is performed on each causal edge in the business causal graph sequence to generate business... A causal confidence matrix is ​​used to generate a credible business causal graph by weighting the business causal correlation matrix based on the business causal confidence matrix. Business process attention features are extracted from the business semantic segment sequence to generate business time feature embeddings. Business relationship attention propagation is performed along the credible business causal graph to generate business relationship feature embeddings. Conditional business spatiotemporal fusion is performed on the business time feature embeddings and the business relationship feature embeddings to generate joint business spatiotemporal feature embeddings. Cross-segment business time series modeling is performed on the joint business spatiotemporal feature embeddings to generate a global business situation representation. Based on the global business situation representation, business decision reasoning mapping is performed to generate and output an enterprise operation decision scheme.

[0015] This invention overcomes the prediction inaccuracies caused by learning spurious associations in scenarios involving workload mode switching or system state changes by explicitly mining the directed causal structure between resource indicators and incorporating it as a priori constraint into the prediction process. Furthermore, it achieves dynamic tracking and reliability quantification of causal topology through temporal causal evolution perception and confidence assessment. The generated credible causal graph can reveal the degree of mutual influence between resource indicators, providing intuitive decision-making basis for cloud operations and maintenance, and significantly improving the accuracy, robustness, and interpretability of cloud resource load prediction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a diagram illustrating the overall structure of the cloud resource prediction model as described in the specific implementation method. Figure 2 This is the relationship channel structure diagram described in the specific implementation method. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 and Figure 2 In a first aspect, this embodiment provides a method for constructing an enterprise operating system based on dynamic trusted causal graph evolution, including: Acquire multi-dimensional business data of enterprise operations; record the evolution of various core business indicators of the enterprise over time. Adaptive semantic segmentation is performed on multi-dimensional business data of enterprise operations to generate a sequence of business semantic segments; Business causal logic is learned based on the sequence of business semantic segments, and a business causal association matrix of each business semantic segment is constructed under the constraint of a directed acyclic graph, resulting in a sequence of business causal graphs. The system perceives the causal evolution of business cause-effect graph sequences, captures the dynamic evolution law of business cause-effect structure, and predicts the future changes in business cause-effect structure. Based on the future changes in the causal structure of business, an evolutionary prior is generated and fed back to the business causal logic learning process to guide the updating of the business causal graph along the historical evolution trend. The confidence of each business causal edge in the business causal graph sequence is evaluated to generate a business causal confidence matrix. The business causal relationship matrix is ​​weighted based on the business causal confidence matrix to generate a reliable business causal graph; Business process attention features are extracted from the business semantic segment sequence to generate business time feature embeddings. Business relationship attention is propagated along the trusted business causal graph to generate business relationship feature embeddings. Conditional spatiotemporal fusion of business time feature embedding and business relationship feature embedding is performed to generate joint business spatiotemporal feature embedding; Cross-segment business time-series modeling is performed on the spatiotemporal feature embedding of joint business to generate a global business situation representation. Based on the global business situation representation, business decision reasoning mapping is performed to generate and output enterprise operation decision schemes.

[0020] In this embodiment, the multi-dimensional business data of enterprise operation is obtained from the continuous collection results of the enterprise operation management system on each core business link. The various core business indicators cover dimensions such as finance, marketing, supply chain, human resources and customer relations. The business data of each dimension are arranged in sequence according to the collection time, which constitutes the time sequence input of the decision model.

[0021] The segmentation boundaries of adaptive semantic segmentation are not preset, but are automatically determined by changes in business data characteristics. This process divides multi-dimensional business data of enterprise operations into a sequence of business semantic segments, making the business patterns within each business semantic segment more stable and providing a relatively stable local observation interval for subsequent learning of business causal logic.

[0022] In the business causal logic learning stage, the causal dependencies between business indicators contained in the observation data are explicitly mined, and the directed acyclic graph constraint is used to eliminate cyclic causality, so that the business causal association matrix corresponding to each business semantic segment maintains a unidirectional causal orientation, thereby forming a business causal graph sequence that depicts the causal topology between business indicators in different time periods.

[0023] Business causal evolution perception models the changes in the business causal graph sequence over time. The captured dynamic evolution patterns reflect the migration trend of the business causal topology over time, and the predicted changes in the future business causal structure provide a basis for the construction of subsequent evolutionary priors.

[0024] The evolutionary priors constructed from the future changes in the causal structure of business are fed back into the business causal logic learning process, ensuring that the learning direction of the business causal graph is consistent with the historical evolution trend. At the same time, the confidence assessment measures the reliability of each business causal edge one by one, forming a business causal confidence matrix.

[0025] The business causal confidence matrix is ​​used as a weight to apply to the business causal association matrix, so that the business causal connections with higher credibility play a greater role in subsequent information dissemination, thereby strengthening the credibility components in the business causal structure and obtaining a credible business causal graph.

[0026] Business process attention feature extraction takes the business semantic segment sequence as input and outputs business time feature embedding. Business relationship attention propagation takes the trusted business causal graph as input and outputs business relationship feature embedding. Business time feature embedding and business relationship feature embedding respectively characterize the dependency relationship in the temporal dimension and the business causal dimension.

[0027] Conditional business spatiotemporal fusion interactively combines the embedding of business time features with the embedding of business relationship features, so that the temporal features and business causal features complement each other, resulting in a joint business spatiotemporal feature embedding.

[0028] The spatiotemporal features of joint business are embedded and integrated into a global business situation representation through cross-segment business time series modeling. The business decision reasoning mapping then maps the global business situation representation to the decision target, and outputs the enterprise operation decision plan.

[0029] This embodiment uses multi-dimensional business data from enterprise operations to form a sequence of business semantic segments through adaptive semantic segmentation, providing a stable local observation window for business causal logic learning. Business causal logic learning, combined with directed acyclic graph constraints, determines a sequence of business causal graphs, explicitly characterizing the business causal topology between business indicators. Business causal evolution perception and confidence assessment yield evolutionary priors and business causal confidence matrices, respectively, tracking the dynamic evolution of the business causal structure in the time dimension and quantifying the reliability of business causal edges in the credibility dimension, thereby weighting to obtain a credible business causal graph. Business process attention feature extraction and business relationship attention propagation output business time feature embeddings and business relationship feature embeddings, respectively. These are integrated into a global business situation representation through conditional business spatiotemporal fusion and cross-segment business time series modeling. Finally, through business decision reasoning mapping, the enterprise operation decision scheme is output, improving the accuracy, robustness, and interpretability of the decision.

[0030] In some embodiments, adaptive semantic segmentation is performed on multi-dimensional business data of enterprise operations to generate a sequence of business semantic segments, including: Slide segmented detection windows across multi-dimensional business data of enterprise operations, with each segmented detection window forming an adjacent window pair with the business data in the previous window. For adjacent window pairs, calculate the difference between the average business data within the segmented detection window and the average business data within the previous window to obtain the mean change; The difference between the fluctuation level of business data within the segmented detection window and the fluctuation level of business data within the previous window is calculated to obtain the fluctuation change. The Jensen-Shannon divergence quantification is used to measure the overall difference between the distribution of business data within the segmented detection window and the distribution of business data within the previous window, thus obtaining the change in distribution pattern. After assigning corresponding weight coefficients to mean change, volatility change, and distribution pattern change, the weighted sums are used to generate a comprehensive change score. The comprehensive change score represents the degree of non-stationarity of business data within the segmented detection window. When the overall change score exceeds the judgment threshold, a business mode change occurs at the segment detection window, and the starting boundary of the segment detection window is marked as a candidate segment point. Based on the minimum segment length constraint, adjacent candidate segment points with a spacing less than the minimum segment length are merged, and the candidate segment points are corrected to obtain the segment boundary. The enterprise's multi-dimensional business data is divided according to the segmentation boundaries to generate a sequence of business semantic segments.

[0031] In this embodiment, the segmented detection window slides along the multi-dimensional business data of enterprise operations segment by segment. The business data carried in the segmented detection window and the business data carried in the immediately preceding window together form an adjacent window pair, which is used to compare the changes in business characteristics between adjacent time periods.

[0032] The average value of business data reflects the overall level of business in a given time period, and changes in the average value are used to depict the shift in the overall level of business between adjacent windows and between windows before and after them.

[0033] The degree of fluctuation in business data reflects the extent to which business data deviates from the mean. Changes in volatility are used to characterize the changes in the magnitude of business fluctuations between adjacent windows and between windows before and after them.

[0034] Jensen-Shannon divergence is used to measure the overall difference between the distribution of business data within a window and before and after. Changes in distribution shape capture changes in distribution structure that are not reflected by changes in mean and volatility.

[0035] Changes in mean, volatility, and distribution pattern reflect changes in different aspects of business data. These changes are weighted and integrated into a comprehensive change score, which serves as a unified measure of the non-stationarity of business data within the segmented detection window.

[0036] The purpose of the judgment threshold is to filter out significant pattern changes. When the comprehensive change score exceeds the judgment threshold, the segment detection window is identified as a business pattern change, and the starting boundary of the segment detection window is then marked as a candidate segmentation point. The candidate segmentation point serves as the initial segmentation position for subsequent correction.

[0037] The minimum segment length constraint is used to ensure that each segment has sufficient length. Adjacent candidate segment points whose spacing does not reach the minimum segment length are merged. The remaining candidate segment points after merging are corrected to form the segment boundary.

[0038] The segmentation boundary divides the enterprise's multi-dimensional business data into a sequence of business semantic segments, which is the output of the adaptive semantic segmentation process.

[0039] This embodiment obtains adjacent window pairs by sliding a segmented detection window along multi-dimensional business data of enterprise operations, transforming the segmentation of the entire sequence into a pairwise comparison of adjacent time periods. Changes in mean, volatility, and distribution pattern are used to characterize changes in business level, volatility amplitude, and distribution structure, respectively, comprehensively capturing the non-stationary differences between adjacent window pairs. A comprehensive change score is obtained through weighted fusion, integrating multi-dimensional changes into a unified non-stationary measure. Significant business model changes are screened by using a judgment threshold, and closely spaced candidate segment points are merged by using a minimum segment length constraint, achieving a balance between segment significance and segment length. Ultimately, a business semantic segment sequence that fits the changes in business characteristics is obtained, providing an internally stable local observation window for subsequent business causal logic learning, thereby improving the stability of decision-making.

[0040] In some embodiments, business causal logic learning is performed based on the business semantic segment sequence, and a business causal association matrix for each business semantic segment is constructed under the constraint of a directed acyclic graph, resulting in a business causal graph sequence, including: A parallel dual-branch business causality perception module is constructed, which consists of a multi-core temporal business causality convolutional branch and a multivariate real-time business causality attention branch. By using a multi-core temporal business causal convolution branch, business causal convolution is performed on each business variable pair in the business semantic segment sequence to obtain the time-delay business causal strength of each business variable pair. Left padding is then performed on the input sequence to ensure that the convolution field of view only covers the current time and historical time, thus preventing the leakage of future information. Through the multivariate real-time business causal attention branch, attention is calculated using the current state of each business variable as the query and key to obtain the real-time business causal strength of each business variable pair. The learnable mask matrix is ​​then applied element by element to the attention score to control the presence and strength of business causal connections between each business variable pair. The business causal intensity of time-delayed business is aggregated with the business causal intensity of real-time business, and the weights corresponding to redundant business causal paths are driven to approach zero by sparse regularization to generate a business causal correlation matrix. Apply a continuous directed acyclic graph constraint to the business causal relationship matrix to generate a business causal relationship matrix that satisfies the directed acyclic graph structure; Perform threshold-based edge pruning on the business causal relationship matrix that satisfies the directed acyclic graph structure to remove weak business causal connections; Inject domain prior knowledge into the business causal relationship matrix after edge clipping, and reset the corresponding association weights of business variables that are known to have no business causal relationship to zero; The above construction and optimization are performed on each business semantic segment in the business semantic segment sequence to obtain the business causal relationship matrix of each business semantic segment, which constitutes the business causal graph sequence.

[0041] In this embodiment, the parallel dual-branch business causality perception module simultaneously captures heterogeneous business causal information from the business semantic segment sequence. The multi-core temporal business causal convolution branch focuses on the time-delayed impact of historical moments on the current moment, and the multivariate real-time business causal attention branch focuses on the instantaneous impact between business variables in the same period.

[0042] The multi-core temporal business causal convolution branch performs causal convolution on business variables with different time delay scales. The resulting time delay business causal strength reflects the degree of influence of historical business variables on current business variables. Left padding is used to ensure temporal causality, so that the cause appears before the effect. Information from future time moments does not participate in the business causal convolution at the current time moment.

[0043] The multivariate real-time business causal attention branch uses queries and keys to measure the correlation between the current states of each business variable. The real-time business causal strength reflects the instantaneous business causal influence between business variables in the same period. The learnable mask matrix is ​​continuously adjusted during training, and the attention score is adjusted one by one according to the position of the elements to determine whether the business causal connection between each pair of business variables is preserved and the strength of the preservation.

[0044] The time-delay business causal strength and the real-time business causal strength are merged to form a business causal association matrix that takes into account both time-delay and real-time business causal information. Sparse regularization makes the weights corresponding to redundant business causal paths converge to zero, retaining only the key business causal connections.

[0045] The continuous directed acyclic graph constraint transforms the directed acyclic structure requirement into an optimizable continuous form, ensuring that the business causal relationship matrix satisfies the directed acyclic graph structure and avoiding the occurrence of cyclic business causal paths.

[0046] Threshold-based edge pruning is used to filter connections with insufficient business causal strength, retaining only connections with significant business causal strength, and discarding weak business causal connections, making the business causal association matrix more concise and reliable.

[0047] Domain prior knowledge is used to correct potential misjudgments in the learning of business causal logic. Business variable pairs that are known to have no business causal relationship will be forcibly excluded, and their corresponding association weights will be set to zero.

[0048] Each business semantic segment undergoes a complete business causal logic learning process to obtain a corresponding business causal association matrix. The business causal association matrices of each business semantic segment are arranged in the order of the business semantic segments and summarized into a business causal graph sequence.

[0049] This embodiment captures both time-delayed and real-time business causality strengths through a parallel dual-branch business causality perception module, taking into account business causal information from both historical and concurrent times. A business causal correlation matrix is ​​obtained through aggregation and sparse regularization, preserving key business causal connections while suppressing redundant business causal paths. A continuous directed acyclic graph constraint ensures the business causal correlation matrix satisfies the directed acyclic graph structure, avoiding cyclic business causal paths. Threshold-based edge pruning and the injection of domain prior knowledge filter weak business causal connections and exclude impossible business causal relationships. The business causal correlation matrices corresponding to each business semantic segment are summarized into a business causal graph sequence, depicting the distribution and changes of business causal topology among business indicators over time, providing a time-evolving business causal structure foundation for business causal evolution perception.

[0050] In some embodiments, business causal evolution perception is performed on the business causal graph sequence to capture the dynamic evolution law of the business causal structure and predict the future changes in the business causal structure, including: The evolution of the business cause-effect graph sequence on the time axis is explicitly modeled. The business cause-effect association matrix of each business semantic segment in the business cause-effect graph sequence is flattened into a one-dimensional vector along the business variable dimension. The business cause-effect connection strength between each business variable pair is arranged in a fixed order to generate a business cause-effect graph embedding vector sequence. The business cause-effect graph embedding vector sequence is input into the gated loop unit in chronological order. At each time step, the gated loop unit, based on the current business cause-effect graph embedding vector and the hidden state of the previous time step, controls the retention of historical evolution information by updating the gate and controls the forgetting of historical evolution information by resetting the gate. The hidden state is updated by weighted fusion of the candidate hidden state and the hidden state of the previous time step, generating a hidden state sequence. Hidden state sequences characterize the dynamic evolution of the business causal topology over time. The dynamic evolution includes the strengthening trend, weakening trend, and migration trend of the business causal edges over time. By using an evolutionary prediction network to predict and map the hidden state sequence, the direction and magnitude of the change in the business causal correlation matrix of future business semantic segments relative to the current moment are estimated, and the change in the future business causal structure is generated. The change in the future business causal structure quantifies the strengthening, weakening, addition and disappearance trends of future business causal edges, and serves as a priori estimate of the business causal structure at future moments.

[0051] In this embodiment, the evolution of the business cause-effect graph sequence on the time axis is explicitly modeled, the business cause-effect association matrix of each business semantic segment is expanded into a one-dimensional vector, and the business cause-effect connection strength of each business variable pair in the expanded vector is stored in a fixed order so that the vectors obtained at different times correspond to the same business variable pair in the same position, thereby forming a business cause-effect graph embedding vector sequence.

[0052] The business cause-effect graph embedding vector sequence is fed into the gate control loop unit time by time. At the current time, the gate control loop unit combines the current business cause-effect graph embedding vector with the hidden state at the previous time. The update gate determines how much historical evolution information is retained, and the reset gate determines how much historical evolution information is forgotten. The candidate hidden state and the hidden state at the previous time are mixed in a weighted manner to obtain the hidden state at the current time. The hidden state sequence is formed by recursion.

[0053] The hidden state sequence carries information about the evolution of the business causal topology over time. The dynamic evolution pattern is reflected in the strengthening, weakening and migration trends of business causal edges over time. Strengthening and weakening reflect the changes in the strength of business causal connections, while migration reflects the transfer of business causal connections between business variable pairs.

[0054] The evolutionary prediction network makes forward predictions starting from the hidden state sequence, and at the same time gives the direction and magnitude of the change of the business causal correlation matrix of the future business semantic segment relative to the current time. The resulting change in the future business causal structure describes the trend of future business causal edges strengthening, weakening, adding and disappearing, and becomes a prior estimate of the business causal structure at the future time.

[0055] This embodiment obtains the business causal graph embedding vector sequence by expanding the business causal correlation matrix, providing comparable input for time series modeling; the hidden state is refreshed time-by-time by a gated recurrent unit, accumulating the evolution information of the business causal topology over time in the hidden state sequence; the retention and forgetting of historical evolution information are determined by update gates and reset gates, respectively, to characterize the strengthening, weakening and migration trends of business causal edges over time; forward prediction is performed by an evolutionary prediction network to obtain an estimate of the changes in the future business causal structure, and the prediction result constitutes the prior estimate of the business causal structure at future time moments.

[0056] In some embodiments, an evolutionary prior is generated based on future changes in the business causal structure. This evolutionary prior is then fed back into the business causal logic learning process to guide the updating of the business causal graph along historical evolutionary trends. Confidence assessments are performed on each business causal edge in the business causal graph sequence to generate a business causal confidence matrix, including: An evolutionary prior is constructed based on the changes in the causal structure of future business. The evolutionary prior represents the prior constraint that the causal structure of business evolves smoothly and continuously along the historical evolution trend between adjacent business semantic segments. Evolutionary priors are fed back as regularization constraints into the loss function of the business causal logic learning process, so as to constrain the deviation between the actual changes in the business causal structure and the future changes in the business causal structure. By using deviation constraints, the business causal relationship matrix of subsequent business semantic segments is updated in a direction consistent with the historical evolution trend while ensuring the accuracy of decision-making, thereby suppressing abrupt changes in the business causal structure between adjacent business semantic segments. A business causal confidence assessment network is constructed by flattening the business causal association matrix of each business semantic segment in the business causal graph sequence into a one-dimensional vector, which serves as the input to the business causal confidence assessment network. The business causal confidence assessment network independently scores the confidence of each business causal edge corresponding to each pair of business variables in each business semantic segment, and obtains the business causal edge confidence score. The business causal edge confidence score represents the reliability of the existence of a business causal relationship between the corresponding pairs of business variables. The confidence scores of the business causal edges of each business variable pair in each business semantic segment are collected to generate a business causal confidence matrix. The business causal confidence matrix is ​​used to characterize the credibility of each business causal edge.

[0057] In this embodiment, the evolutionary prior is derived from the changes in the causal structure of future business operations. The evolutionary prior carries the prior constraint that the causal structure of business operations between adjacent business semantic segments should evolve smoothly along the historical evolution trend. The causal structure of business operations between adjacent time periods continues smoothly along the existing trend, rather than experiencing large fluctuations.

[0058] Evolutionary priors are added as regularization constraints to the loss function of the business causal logic learning process. The deviation between the actual changes in the business causal structure and the future changes in the business causal structure is constrained, and the business causal logic learning process is additionally guided by evolutionary priors.

[0059] Guided by the deviation constraint, the business causal relationship matrix of subsequent business semantic segments maintains decision accuracy and keeps the update direction consistent with the historical evolution trend, thus making it less likely for the business causal structure between adjacent business semantic segments to undergo abrupt changes.

[0060] The business causal confidence evaluation network is used to score each business causal edge. The business causal association matrix of each business semantic segment in the business causal graph sequence is expanded into a one-dimensional vector, which becomes the input of the business causal confidence evaluation network.

[0061] The business causality confidence assessment network independently scores the confidence of each business causal edge corresponding to each pair of business variables. The business causal edges do not interfere with each other. The obtained business causal edge confidence score measures the reliability of the business causal relationship between the corresponding pairs of business variables.

[0062] The confidence scores of the business causal edges of each business variable pair in each business semantic segment are collected and organized into a business causal confidence matrix according to the position of the business variable pair. The business causal confidence matrix is ​​used to characterize the credibility of each business causal edge.

[0063] This embodiment transforms the predicted changes in the business causal structure into prior constraints for subsequent learning by using evolutionary priors formed from future changes in the business causal structure. The evolutionary priors constrain the deviation between the actual changes in the business causal structure and the future changes in the business causal structure in the form of regularization constraints, ensuring that the business causal structure is smoothly updated along the historical evolution trend and suppressing abrupt changes. Each business causal edge is scored independently by a business causal confidence evaluation network, resulting in a business causal edge confidence score that characterizes the reliability of each business causal edge. By aggregating the confidence scores of each business causal edge, a business causal confidence matrix is ​​obtained, providing a quantitative basis for the credibility of each business causal edge.

[0064] In some embodiments, the business causal correlation matrix is ​​weighted based on the business causal confidence matrix to generate a reliable business causal graph, including: The confidence scores of the business causal edges corresponding to each business variable pair in the business causal confidence matrix are multiplied element-wise by the business causal connection strength of the corresponding business variable pair in the business causal association matrix, so that the positions of each business variable pair in the business causal confidence matrix and the business causal association matrix correspond one-to-one. Using the confidence score of the business causal edge as a weight, the business causal connection strength of each business causal edge is scaled so that the effective connection strength of each business causal edge is determined by both the business causal connection strength and the confidence score. The higher the confidence score, the stronger the effective connection strength. Generate a confidence-weighted business causal relationship matrix. The values ​​of the elements in the confidence-weighted business causal relationship matrix represent the effective connection strength of the corresponding business causal edges after confidence scaling. In the confidence-weighted business causal relationship matrix, business causal edges with high confidence scores dominate, while business causal edges with low confidence scores are continuously and adaptively suppressed according to their confidence scores. By continuously and adaptively suppressing false business causal connections with low credibility, and retaining true business causal connections with high credibility, the interference of false associations on the business causal structure is suppressed. The confidence-weighted business causal relationship matrix is ​​used as a trusted business causal graph. The weight of each business causal edge in the trusted business causal graph reflects both the strength of the business causal connection and the degree of trust in the business causal relationship.

[0065] In this embodiment, the business causal confidence matrix and the business causal association matrix maintain the same dimension, the positions of each business variable pair correspond to each other, and the business causal edge confidence score is multiplied by the corresponding business causal connection strength according to the element position, so that the confidence of each business variable pair is directly combined with the business causal strength.

[0066] The confidence score of the business causal edge serves as the weight, and the strength of the business causal connection is scaled according to the weight. The effective connection strength obtained by scaling depends on the business causal connection strength and the confidence score. When the confidence score is higher, the effective connection strength is correspondingly stronger.

[0067] After weighting, a confidence-weighted business causal relationship matrix is ​​formed. The value of each element in the confidence-weighted business causal relationship matrix is ​​the effective connection strength of the corresponding business causal edge after confidence scaling.

[0068] Business causal edges with high confidence scores have greater weight and play a dominant role in the confidence-weighted business causal relationship matrix, while business causal edges with low confidence scores are subject to continuous adaptive suppression, with the degree of suppression increasing as the confidence score decreases.

[0069] Continuous adaptive suppression weakens spurious business causal connections with low credibility while maintaining genuine business causal connections with high credibility, thus suppressing the interference of spurious associations on the business causal structure.

[0070] The confidence-weighted business causal relationship matrix then becomes a credible business causal graph. The weight of each business causal edge in the credible business causal graph reflects both the strength of the business causal connection and the credibility of the business causal relationship.

[0071] This embodiment combines the confidence score of business causal edges with the strength of business causal connections by multiplying the corresponding elements at their positions, directly linking confidence score with business causal strength. By using the confidence score of business causal edges as weights to scale the strength of business causal connections, the effective connection strength is determined jointly by the strength of business causal connections and the confidence score. Through continuous adaptive suppression of spurious business causal connections and maintenance of genuine business causal connections, the interference of spurious associations on the business causal structure is reduced. The resulting confidence-weighted business causal association matrix then becomes a reliable business causal graph, so that the weight of each business causal edge reflects both the strength of business causal connections and the reliability of business causal relationships.

[0072] In some embodiments, business process attention features are extracted from the business semantic segment sequence to generate business time feature embeddings, and business relationship attention is propagated along the trusted business causal graph to generate business relationship feature embeddings, including: Construct a Transformer-based business process attention network, which consists of sequentially stacked business process attention layers and feedforward layers; The sequence of business semantic segments is encoded with independent position and cross-channel position. Independent position encoding represents the positional relationship within each business semantic segment, while cross-channel position encoding represents the relative position between different business indicators. The two are then fused to generate a sequence position representation. The business process attention layer maps the sequence position representation into query vector, key vector and value vector, and performs self-attention calculation. The attention calculation adopts orthogonal basis function expansion approximation to capture long-term business time-series dependencies within and across segments and generate business time feature embeddings. The time dimension of the original sequence matrix of each business variable in the trusted business causal graph is mapped and compressed through a fully connected layer to construct the initial node representation; Using a trusted business causal graph as the adjacency matrix, a graph attention mechanism is applied to the initial node representation to aggregate the business causal neighbor information of each business variable and update the node representation layer by layer. The attention coefficient of the graph attention mechanism is obtained by linear transformation of the business causal neighbor node representation pair. The attention coefficient quantifies the propagation strength of the business causal effect along the business causal path. The updated node representations are aggregated using global mean pooling to generate embedded business relationship features.

[0073] In this embodiment, a Transformer-based business process attention network is constructed, in which the business process attention layer and the feedforward layer are stacked in sequence, alternately extracting temporal features and performing nonlinear transformations, together forming the business process attention network.

[0074] The business semantic segment sequence is encoded by independent position and cross-channel position. Independent position encoding describes the positional relationship within each business semantic segment, while cross-channel position encoding describes the relative position between different business indicators. The sequence position representation is obtained by fusing the two types of encoding.

[0075] The business process attention layer converts the sequence position representation into query vectors, key vectors, and value vectors, and performs self-attention computation. In this embodiment, the attention computation is approximated by orthogonal basis function expansion. Optionally, the orthogonal basis function expansion approximation includes one-dimensional random Fourier feature approximation and Walsh orthogonal basis expansion approximation, which reduces computational overhead. Long-term business time-series dependencies within and across segments are captured, thereby obtaining business time feature embedding.

[0076] Specifically, the one-dimensional stochastic Fourier feature approximation approximates the kernel function in attention calculation by randomly sampling Fourier features, completing the attention operation in the coefficient space. The discrete Walsh orthogonal basis expansion obtains the expansion coefficients by selecting discrete Walsh basis functions to perform orthogonal basis expansion on the business sequence, embedding the self-attention mechanism into the coefficient space of the process neurons, and utilizing the equivalence between integrals and inner products under orthogonal basis conditions to achieve global accumulation. Both of these implementation methods belong to orthogonal basis function expansion approximation, and those skilled in the art can choose one of these two implementation methods to implement based on the requirements for computational accuracy and computational cost.

[0077] The fully connected layer maps and compresses the original sequence matrix of each business variable in the trusted business causal graph in the time dimension, and the compressed result is the initial node representation.

[0078] The trusted business causal graph acts as an adjacency matrix. The graph attention mechanism is applied to the initial node representation. The business causal neighbor information of each business variable is aggregated, and the node representation is updated layer by layer.

[0079] The attention coefficient is calculated by linear transformation of the business causal neighbor node representation pairs. The attention coefficient characterizes the strength of the propagation of business causal effects along the business causal path.

[0080] The updated node representation is aggregated through global mean pooling, and the information of all nodes is summarized into business relationship feature embedding.

[0081] This embodiment obtains a sequence position representation by fusing independent position encoding and cross-channel position encoding, providing position-informed input to the business process attention network; it captures long-range business temporal dependencies within and across segments through self-attention computation of the business process attention layer, obtaining business temporal feature embeddings; it obtains initial node representations through mapping and compression of fully connected layers, providing node inputs for the graph attention mechanism; it aggregates business causal neighbor information using a trusted business causal graph as an adjacency matrix, propagates business causal effects along business causal paths, and updates node representations; and it obtains business relationship feature embeddings through global mean pooling aggregation, which, together with business temporal feature embeddings, characterizes features in both the temporal and business causal dimensions.

[0082] In some embodiments, conditional spatiotemporal fusion is performed on the business time feature embedding and the business relationship feature embedding to generate a joint business spatiotemporal feature embedding, including: The business time feature embedding and the business relationship feature embedding are concatenated along the feature dimension to generate an initial joint representation; A gated cross-attention mechanism is used to perform bidirectional conditional information interaction between the embedding of business time features and the embedding of business relationship features. The initial joint representation is transformed into query vector, key vector and value vector through linear transformation. A gate signal is generated by linear transformation of the embedded business time features. The gate signal is used to dynamically modulate the aggregation strength of the key value of the embedded business relationship features. The gate signal determines the strength of each key value information in the embedded business relationship features participating in the aggregation, thereby realizing the conditional control of the aggregation of business relationship features by business time features. The query vector is generated by embedding business relationship features. The query vector guides the focusing direction of business time feature embedding on different frequency domain components, thereby strengthening the frequency domain components related to the business causal structure in the business time feature embedding and realizing the conditional guidance of business relationship features to focus on business time features. The bidirectional cross-attention outputs obtained from bidirectional conditional information interaction are fused after being weighted and modulated by a gated signal, and the single-head computation is extended to multi-head parallel computation through a multi-head mechanism. The output of multi-head parallel computation is processed by feedforward network and residual connection to output joint business spatiotemporal feature embedding.

[0083] In this embodiment, the business time feature embedding and the business relationship feature embedding are concatenated along the feature dimension. The concatenated representation is the initial joint representation, which retains information from both the time sequence feature and the business causal feature.

[0084] The gated cross-attention mechanism is used to realize bidirectional conditional information interaction between business time feature embedding and business relationship feature embedding. The initial joint representation is transformed by linear transformation to generate query vector, key vector and value vector respectively.

[0085] The business time feature embedding is linearly transformed to obtain a gating signal. The gating signal dynamically modulates the key value aggregation strength of the business relationship feature embedding. The strength of each key value information participating in the aggregation in the business relationship feature embedding is determined by the gating signal. The conditional control of business time features on business relationship feature aggregation is realized.

[0086] The query vector is generated by embedding business relationship features. The query vector provides guidance for the focusing direction of embedding business time features on different frequency domain components. As a result, the frequency domain components related to the business causal structure in the embedding of business time features are strengthened, and the conditional guidance of business relationship features on focusing on business time features is realized.

[0087] The bidirectional cross-attention output obtained from bidirectional conditional information interaction is weighted and modulated by a gated signal. The modulated bidirectional cross-attention output is then fused, and single-head computation is extended to multi-head parallel computation through a multi-head mechanism.

[0088] The output of the multi-head parallel computation is processed sequentially through a feedforward network and a residual connection. The result after processing is the joint business spatiotemporal feature embedding.

[0089] This embodiment obtains an initial joint representation by concatenating business time feature embeddings and business relationship feature embeddings along the feature dimension, providing a fusion basis for bidirectional conditional information interaction; it obtains a gating signal through business time feature embedding and dynamically modulates the key-value aggregation intensity of business relationship feature embeddings, realizing conditional control of business time features on business relationship feature aggregation; it obtains a query vector through business relationship feature embedding, and the query vector acts on the focusing direction of business time feature embeddings in different frequency domain components, realizing conditional guidance of business relationship features on business time feature focusing; it obtains multi-head parallel computation through bidirectional cross-attention output, weighted modulation, and multi-head mechanism expansion; and it outputs the joint business spatiotemporal feature embedding through feedforward network and residual connection processing.

[0090] In some embodiments, cross-segment business time-series modeling is performed on the spatiotemporal feature embedding of joint business to generate a global business situation representation. Based on the global business situation representation, business decision reasoning mapping is performed to generate and output an enterprise operation decision plan, including: The spatiotemporal features of joint business are embedded and arranged in chronological order of the semantic segments of each business in the original sequence to construct a segmented business feature sequence. The segmented business feature sequence is input into the cross-segment business temporal modeling module, which is composed of a Transformer encoder. The cross-segment business temporal modeling module is used to model the long-term dependency relationship between each business semantic segment and the cross-segment business context evolution mode. Temporal position information is injected into each business semantic segment in the segmented business feature sequence by position encoding in order to preserve the temporal order between each business semantic segment; By utilizing a multi-head self-attention mechanism, attention computation is extended to parallel computation of multiple attention heads to calculate the attention weights between each business semantic segment in the segmented business feature sequence, thereby capturing the long-range dependencies and contextual evolution patterns between each business semantic segment. The output of the multi-head self-attention is aggregated through multiple layers of encoder to output a global business situation representation that characterizes the spatiotemporal features of the entire historical input window. The global business situation representation is input into a multilayer perceptron projection network consisting of fully connected layers, and the global business situation representation is mapped to the decision space through linear transformation. Output enterprise operation decision-making solutions, which include optimization and adjustment strategies and resource allocation suggestions for multiple core business indicators, as the result of the enterprise operation system construction and output.

[0091] In this embodiment, the spatiotemporal features of the joint services are embedded in sequence according to the chronological order of each service semantic segment in the original sequence. The resulting feature sequence is the segmented service feature sequence, which maintains the same chronological order as the original sequence.

[0092] The segmented business feature sequences are fed into the cross-segment business temporal modeling module, which is composed of a Transformer encoder. The cross-segment business temporal modeling module is used to mine the long-term dependencies between the semantic segments of each business and the cross-segment business context evolution pattern.

[0093] Location encoding supplements the temporal location information of each business semantic segment in the segmented business feature sequence, thus preserving the temporal order between the business semantic segments.

[0094] The multi-head self-attention mechanism distributes attention computation across multiple attention heads in parallel, and the attention weights between each business semantic segment in the segmented business feature sequence are calculated, thereby capturing the long-range dependencies and context evolution patterns between each business semantic segment.

[0095] The output of the multi-head self-attention is aggregated layer by layer through multi-layer encoding of the encoder, and the final output is a global business situation representation that characterizes the spatiotemporal features of the entire historical input window business.

[0096] The global business situation representation enters a multilayer perceptron projection network composed of fully connected layers and is mapped to the decision space through a linear transformation.

[0097] The enterprise operation decision-making plan is output, which includes optimization and adjustment strategies and resource allocation suggestions for multiple core business indicators, forming the result of the enterprise operation system construction and output.

[0098] This embodiment obtains segmented business feature sequences by arranging the spatiotemporal features of joint business in chronological order, providing ordered input for cross-segment business temporal modeling; it mines long-range dependencies and contextual evolution patterns between semantic segments of each business through a Transformer encoder and a multi-head self-attention mechanism, obtaining a global business situation representation that characterizes the spatiotemporal features of the entire historical input window; through linear transformation of a multilayer perceptron projection network, the global business situation representation is mapped to the decision space, obtaining optimization and adjustment strategies and resource allocation suggestions for multiple core business indicators, and finally forming and outputting an enterprise operation decision plan.

[0099] In a second aspect, the present invention provides a system for constructing an enterprise operation system based on dynamic trusted causal graph evolution, for executing the method of the first aspect, the system comprising: The adaptive business semantic segmentation module is used to adaptively segment multi-dimensional business data of enterprise operations. It uses a segmentation detection window to detect changes in mean, volatility, and distribution patterns between adjacent windows to determine segmentation boundaries and generate a sequence of business semantic segments. The dual-channel business feature extraction network consists of a business time channel and a business relationship channel. The business time channel includes a business process attention network, which is used to extract business process attention features from the business semantic segment sequence and generate business time feature embeddings. The business relationship channel includes a parallel dual-branch business causality perception module, a directed acyclic graph constraint optimization module, a business temporal causal evolution perception module, a business causal confidence evaluation module, and a business relationship attention propagation module. The parallel dual-branch business causality perception module is used to learn the business causal association matrix of each business semantic segment from the business semantic segment sequence. The directed acyclic graph constraint optimization module is used to apply directed acyclic graph constraints to the business causal association matrix. The business temporal causal evolution perception module is used to capture the dynamic evolution law of the business causal structure, predict the future changes in the business causal structure, and feed the future changes in the business causal structure as evolutionary priors to the business causal logic learning process. The business causal confidence evaluation module is used to generate business causal edge confidence scores for each business causal edge. The business relationship attention propagation module is used to propagate information along the confidence-weighted business causal association matrix and generate business relationship feature embeddings. The conditional business spatiotemporal fusion module is used to perform gated cross-attention bidirectional conditional fusion of business time feature embedding and business relationship feature embedding to generate joint business spatiotemporal feature embedding. The decision-making scheme output module is used to perform cross-segment business time-series modeling based on the spatiotemporal feature embedding of joint business, generate a global business situation representation, and output enterprise operation decision-making schemes based on the global business situation representation through business decision reasoning mapping.

[0100] In this embodiment, the adaptive business semantic segmentation module, as the first functional unit of the system, receives multi-dimensional business data of enterprise operations and completes adaptive semantic segmentation. The segmentation detection window slides on the data and detects the mean change, fluctuation change and distribution pattern change between adjacent windows, thereby defining the segmentation boundary and cutting out the business semantic segment sequence.

[0101] The dual-channel business feature extraction network takes over the business semantic segment sequence, with the business time channel and business relationship channel working in parallel. The business time channel uses the business process attention network to extract business process attention features and obtain business time feature embeddings. The business relationship channel is composed of a parallel dual-branch business causality perception module, a directed acyclic graph constraint optimization module, a business temporal causality evolution perception module, a business causality confidence assessment module, and a business relationship attention propagation module cascaded together. The parallel dual-branch business causality perception module learns the business causality association matrix of each business semantic segment using the business semantic segment sequence as input. The directed acyclic graph constraint optimization module applies directed acyclic graph constraints to the business causality association matrix. The business temporal causality evolution perception module tracks the dynamic evolution of the business causality structure and predicts the future changes in the business causality structure. The future changes in the business causality structure are fed back to the business causality logic learning process in the form of evolutionary priors. The business causality confidence assessment module generates business causality edge confidence scores edge by edge. The business relationship attention propagation module propagates information along the confidence-weighted business causality association matrix and obtains business relationship feature embeddings.

[0102] The conditional business spatiotemporal fusion module receives business time feature embeddings and business relationship feature embeddings, and achieves bidirectional conditional fusion with the help of gating cross attention. The fusion result is the joint business spatiotemporal feature embedding.

[0103] The decision-making output module performs cross-segment business time-series modeling based on the spatiotemporal features of joint business to obtain a global business situation representation, and performs business decision reasoning mapping based on the global business situation representation to provide enterprise operation decision-making solutions.

[0104] This embodiment uses an adaptive business semantic segmentation module to segment multi-dimensional business data of enterprise operations into business semantic segment sequences, providing a stable local observation window for the dual-channel business feature extraction network. The business time channel and business relationship channel of the dual-channel business feature extraction network are used to obtain business time feature embeddings and business relationship feature embeddings respectively, taking into account both temporal and causal dimensions. The conditional business spatiotemporal fusion module fuses the business time feature embeddings and business relationship feature embeddings to obtain a joint business spatiotemporal feature embedding. Finally, the decision scheme output module completes cross-segment business temporal modeling and business decision reasoning mapping, outputting the enterprise operation decision scheme.

[0105] Please see Figure 1 and Figure 2 For ease of understanding, the following examples are also provided: This example provides a cloud resource load prediction method based on process attention and dynamic causal graph evolution, including the following steps: S1, the time span is... Resource usage data of a server in a cloud computing cluster ,in Indicates the first Observed at all times Resource load vectors, where resource types include CPU utilization, memory utilization, disk I / O, and network bandwidth utilization. Indicates the first Load values ​​of class resources; S2 will store cloud resource usage data. Divided into training set and test set ; S3, using a length of The data acquisition window for the training set Perform block sampling to obtain The input sample set consists of several sequence segments of equal length. ,in For the data acquisition window Continuous data collected in one session , , Indicates the window number The first collection Observational data at each time step, ; S4. A cloud resource load prediction model, CTG-PAN, based on process attention and dynamic causal graph evolution, is established. This model consists of an adaptive semantic segmentation module, a dual-channel feature extraction network, a conditional spatiotemporal fusion module, and a prediction output module. The dual-channel feature extraction network consists of a temporal channel and a relational channel. The temporal channel is composed of a process attention network (PAN), which is based on the orthogonal basis expansion theory of process neural networks (PNN) and embeds the self-attention mechanism into the coefficient space of process neurons. The relation channel consists of five cascaded parts: a parallel dual-branch causal perception module, a DAG constraint optimization module based on the NOTEARS framework, a temporal causal evolution perception module, a causal confidence evaluation module, and a graph attention propagation module (GAT). It is used to learn the directed causal structure between resources from the data and propagate information along the causal path through GAT. The temporal causal evolution perception module performs temporal modeling of the causal graph sequence through GRU, captures the evolution trend of causal relationships over time, and predicts future changes in causal structure. The causal confidence evaluation module independently generates a confidence score for each directed edge in the causal graph, quantifies the reliability of causal connections, and makes high-confidence causal edges dominate in graph attention propagation. S5, continuously acquires data from the data acquisition window. The input is fed into the adaptive semantic segmentation module, which divides the resource sequence into segments based on the statistical characteristics of the sequence using a segmentation detection window. A set of semantic segments with relatively stable internal relationships , ; S6, the segmented semantic segments The input is fed into the time channel, and then processed through a process attention network (PAN) in three steps: orthogonal basis expansion, dynamic coefficient modulation, and global accumulator, to output a fixed-dimensional temporal feature embedding. ; S7, the segmented semantic segments The input is fed into the relation channel, where a dual-branch causal perception module learns the causal structure. After NOTEARS constraint optimization and post-processing, a sequence of causal adjacency matrices for each semantic segment is obtained. Based on this, a temporal causal evolution perception module models the causal graph sequence to capture its dynamic evolution. A causal confidence evaluation module assigns a confidence score to each causal edge. Finally, the confidence-weighted causal adjacency matrix is ​​input into the graph attention propagation module for spatial dependency modeling, generating feature embeddings representing the relation structure. ; S8, the output of the dual-channel feature extraction network and The input is fed into the conditional spatiotemporal fusion module, where it undergoes interactive deep fusion through a gated attention mechanism to generate a joint spatiotemporal feature embedding. ; S9, semantic segments Corresponding spatiotemporal feature embedding Perform cross-segment temporal modeling according to the original time sequence, capture long-term dependencies between segments, and generate a global spatiotemporal representation. Then, the global spacetime representation The input is a prediction output module composed of a multilayer sensing mechanism, which is mapped to the target space through a linear transformation to output multivariate resource sequence prediction values ​​for multiple future steps. The loss function is constructed through mean squared error, and the model is trained using the Adam optimizer. During the testing phase, the test set is input into the model in the same way to obtain the prediction results.

[0106] Step S1 further includes; S11, collects raw resource usage sequences from the monitoring system of the cloud computing platform, or obtains them from the publicly available cluster tracing dataset cluster-trace-v2018 Alibaba Cloud cluster tracing dataset. .

[0107] Step S2 further includes; S21, resource usage data The training set is divided chronologically in a 7:3 ratio. and test set .

[0108] Step S5 further includes; S51, initialize a string of length... Step size is The segmented detection window, in the input sequence The above calculates its relationship with the previous window (the first window) in turn. The statistical characteristic changes between (a number of windows), including changes in the mean. volatility and changes in distribution patterns ,in, The Jensen-Shannon divergence is used to quantify the overall difference in the data distribution within two windows. S52, the above three changes are weighted and summed using the formula. Calculate the overall change score. This score quantifies the overall non-stationarity within that window. The higher the value, the greater the likelihood of a significant pattern shift occurring at that location; S53, if Exceeding the preset threshold Then, the starting boundary of the current window is marked as a candidate segmentation point, and the minimum segment length constraint (i.e., the merging interval is less than) is applied. Adjacent points, The candidate points are corrected using the minimum segment length threshold to obtain the final segment boundaries.

[0109] Step S6 further includes, S61, Select Discrete Walsh basis functions For each type of resource sequence Perform a discrete Walsh transform to obtain the expansion coefficients: , , ; in, The sampling interval is... For normalized time points. Coefficient matrix. This characterizes the time-frequency energy distribution of the semantic segment. The coefficient matrix is ​​flattened into a one-dimensional vector: ; Where, vector Each dimension The decomposition coefficient of a specific resource on a specific Walsh; S62, the static coefficient vector Mapped to coefficient representation sequence through linear embedding layer : ; Then, Generate query vector through linear transformation Key vector Value vector And calculate the scaled dot product self-attention: ; Output attention Compress back to the original coefficient space dimension through linear mapping: ; in, The coefficient vector after dynamic modulation; S63, based on the Walsh expansion in S61, the time-varying weight function... Expanding on the same set of Walsh bases: ; in, These are the learnable weight expansion coefficients. All... According to Arranged in the same flattened order, they form a weight coefficient vector. The coefficient for each dimension is . By the equivalence theorem of integrals and inner products under the condition of orthogonal basis, for any expansion coefficient vector The time-domain global integral and the coefficient space inner product satisfy: ; Therefore, the modulated coefficients are used. Substitute static coefficient This yields the global cumulative vector: ; in, Each column represents the weight expansion coefficient vector corresponding to different cumulative channels. .

[0110] S64, global cumulative vector The threshold firing characteristics of biological neurons are simulated using the nonlinear activation function Sigmoid of process neurons: ; Then, A deep nonlinear mapping is performed on the input feedforward network. To preserve the underlying information of the original input, the input semantic segments are... Global average pooling is performed along the resource dimension, and after linear projection to align the dimensions, it is added to the output of the feedforward network. Finally, after layer normalization, the temporal feature embedding of this semantic segment is obtained. ; in, Indicates the first In the semantic segment, the first Class resources in Load value at any given time The global scalar mean of this semantic segment is used for residual connection after linear projection.

[0111] Step S7 further includes; S71 constructs a parallel dual-branch causal perception module consisting of a multi-core temporal causal convolution branch and a multivariate instantaneous causal attention branch, which simultaneously captures two kinds of heterogeneous causal information, time-delay causality and instantaneous causality, from the input semantic segment. S72 defines a learnable convolution kernel for multi-kernel temporal causal convolution branches. ( (for the maximum time delay window), for the first... Class resources and the first Each directed variable pair, composed of class resources, is computed through causal convolution. Historical values ​​of variables The lag causality of the current value is determined by left-padding (with zeros padded to a length of 0.5) to ensure temporal causality (i.e., the cause must precede the effect). ),make The convolutional field of view at any given time only covers the current and historical moments, prohibiting the leakage of future information. The causal convolution process is represented as follows: Stack all The calculation results yield the convolution output. The convolution kernel weights The sign and magnitude of the variable reflect the value of the variable. Lag Resources at all times The direction and intensity of causal action are determined, and by applying L1 regularization to the convolution kernel, the weights corresponding to redundant causal paths are driven to approach zero, thereby achieving sparsity selection of the causal structure. S73, for the multivariate immediate causal attention branch, the input sequence segment Mapped through the embedding layer and define the query matrix. Bond matrix By capturing the instantaneous state representation of each resource at the current moment, attention computation is performed. Modeling an instantaneous causal structure, in which The temperature coefficient controls the smoothness of the attention distribution, and the mask matrix can be learned. Element-wise, the attention score is applied to control for the presence and strength of causal connections between variables. Constrained by L1 regularization to encourage sparse, instantaneous causal topologies, output attention weights. Measurement Variable For variables The immediate causal strength; S74, through formula By aggregating the outputs of the two branches, a unified causal adjacency matrix is ​​generated. ,in To balance the hyperparameters, in the first term, Indicates the convolution kernel at the th... Weight matrices on each time-delay channel This indicates that for each pair of variables Take all its time delays independently The maximum absolute value on, i.e. To highlight the strongest causal effect between variables, the second term... Immediate causal attention weights are used to capture the instantaneous utility between contemporaneous variables, and then the continuous DAG constraint of the NOTEARS framework is introduced. This ensures that the learned causal structure is a directed acyclic graph, where This is the Hadamard square operation. As the matrix exponent, the enhanced Lagrange method is used during training to incorporate this constraint into the objective function. ,in The loss is calculated as the mean squared error between the predicted value and the true value output by the model. Used to promote sparsity , and These are the sparse regularization coefficient, the DAG constraint coefficient, and the penalty parameter, respectively. S75, regarding the optimized causal adjacency matrix Post-processing is performed to generate a high-confidence cause-effect graph. The post-processing first employs threshold-based edge clipping, retaining only edges with absolute values ​​greater than a certain threshold. Edge (threshold) Weak connections are eliminated by maximizing the F1 score of the causal structure discovery task on the validation set, followed by the injection of domain prior knowledge. Variables with known causal relationships are forced to have their corresponding weights set to zero. This process is applied to all... The above post-processing is performed on each semantic segment to obtain the causal adjacency matrix of each semantic segment. ,in For the first The causal adjacency matrix of semantic segments; S76, construct a time-series causal evolution perception module to explicitly model the evolution of the causal graph along the time axis, and include all... The causal adjacency matrix corresponding to each semantic segment Arranged in chronological order, forming a cause-effect graph sequence. The gated recurrent unit (GRU) is used to capture the evolution pattern of the causal structure over time, and the causal adjacency matrix of each semantic segment is used. Flattened into a one-dimensional vector As GRU in Time-based input, GRU's hidden state The evolutionary trend of causal topology over time is characterized, and then the causal structure change trend of future semantic segments is predicted through an evolutionary prediction network. The network outputs a predicted value of the change in the causal adjacency matrix, achieving a priori estimation of the causal structure at future time points. This feedback will be used as a regularization constraint in the NOTEARS optimization process of the corresponding semantic segment in subsequent training iterations. Add to the total loss function, where The actual change in the causal graph is used to ensure that the causal graph learning of subsequent semantic segments updates in a direction consistent with the historical evolution trend while maintaining prediction accuracy, thereby guiding the learning of causal structure by evolutionary priors. S77 constructs a causal confidence evaluation module, which assigns a confidence score to each edge in the causal graph. This module executes in parallel with S76, using the causal adjacency matrix of each semantic segment output by S75. As input, evaluate the network using confidence levels. For the first Each pair of variables in each semantic segment Confidence score for generating causal edges ,Right now ,in for The flattened vector, For the first Causal confidence matrix of semantic segments The closer to 1, the better the model is to the "variable". to variable The higher the confidence level of the judgment that "a causal relationship exists," the better the training process will be. The confidence regularization loss is calculated cumulatively for samples from each data acquisition window. ,by The total loss function is added to the weight coefficients, therefore the complete form of the total loss function is: ; S78, the confidence-weighted causal adjacency matrix is ​​input into the graph attention propagation module (GAT) for spatial dependency modeling. First, the confidence matrix output from S77 is used... The original causal adjacency matrix output by S75 Element-wise multiplication yields a confidence-weighted causal adjacency matrix. This allows high-confidence causal edges to dominate subsequent information propagation, while low-confidence causal edges are adaptively suppressed. Then, an initial node representation is constructed for each type of resource, letting... The original sequence matrix is ​​compressed to its temporal dimension using a fully connected layer. The initial embedding matrix of the nodes is obtained. Then, a causal adjacency matrix weighted by confidence level. As an adjacency matrix, a graph attention mechanism is applied to update the node representations. Each layer's node representation is updated by aggregating information from its causal neighbors, and the attention coefficient quantifies the propagation strength of causal effects along causal paths. Finally, global mean pooling aggregates all node representations to form the relational feature embedding for this semantic segment. .

[0112] Step S8 further includes; S81, for the first Each semantic segment is embedded with its temporal features. Relational feature embedding Concatenate along the feature dimensions to form an initial joint representation ; S82 employs a gated cross-attention mechanism to achieve bidirectional information interaction between temporal and relational features. First, the initial joint representation... Generate query matrices through linear transformations respectively , Key matrix , Sum matrix , Then, based on time characteristics Generate gating signal The intensity of key-value aggregation of relation features is dynamically modulated, while the relation features are used as the basis for dynamic modulation. Generate a query vector to guide the focusing direction of temporal features on different frequency domain components. The specific calculation method is: cross-attention output. and The bidirectional outputs are combined into a gated modulation. The single-head computation is then extended to a multi-head parallel version through a multi-head mechanism, and finally the joint spatiotemporal feature embedding of the semantic segment is output through a feedforward network and residual connections. .

[0113] Step S9 further includes; S91, will all Spatiotemporal feature embedding corresponding to each semantic segment Arrange them in chronological order to construct segment-level feature sequences. ; S92, the arranged segment-level feature sequence The input is a cross-segment temporal modeling module composed of a Transformer encoder. It preserves the temporal position information of each semantic segment through positional encoding, and uses multi-head self-attention to calculate the long-range dependencies and contextual evolution patterns between semantic segments. After encoding and aggregation, it outputs a global spatiotemporal representation that represents the entire historical input window. ; S93, representing global spatiotemporal space. The input is a projection network consisting of fully connected layers, which is mapped to the target space through a linear transformation, and the output is a projection of the future. Each time step Predicted value of resource utilization .

[0114] This invention acquires a cloud resource load sequence, performs adaptive semantic segmentation to generate a semantic segment sequence, learns causal structure based on the semantic segment sequence, constructs a causal adjacency matrix for each semantic segment under a directed acyclic graph constraint to obtain a causal graph sequence, performs temporal causal evolution perception on the causal graph sequence to capture the dynamic evolution law of the causal structure and predict future changes in the causal structure, generates evolutionary priors based on future changes in the causal structure and feeds them back into the causal structure learning process, evaluates the confidence of each causal edge in the causal graph sequence to generate a causal confidence matrix, weights the causal adjacency matrix based on the causal confidence matrix to generate a reliable causal graph, extracts process attention features from the semantic segment sequence to generate temporal feature embeddings, performs graph attention propagation along the reliable causal graph to generate relational feature embeddings, performs conditional spatiotemporal fusion of temporal feature embeddings and relational feature embeddings to generate joint spatiotemporal feature embeddings, performs cross-segment temporal modeling on the joint spatiotemporal feature embeddings to generate a global spatiotemporal representation, and performs prediction mapping based on the global spatiotemporal representation to generate and output cloud resource load prediction results.

[0115] This invention overcomes the prediction inaccuracies caused by learning spurious associations in scenarios involving workload mode switching or system state changes by explicitly mining the directed causal structure between resource indicators and incorporating it as a priori constraint into the prediction process. Furthermore, it achieves dynamic tracking and reliability quantification of causal topology through temporal causal evolution perception and confidence assessment. The generated credible causal graph can reveal the degree of mutual influence between resource indicators, providing intuitive decision-making basis for cloud operations and maintenance, and significantly improving the accuracy, robustness, and interpretability of cloud resource load prediction.

[0116] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0117] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0118] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for constructing an enterprise operation system based on dynamic trusted cause-effect graph evolution, characterized in that, include: Acquire multi-dimensional business data of enterprise operations, which records the evolution of various core business indicators of the enterprise over time; Adaptive semantic segmentation is performed on the multi-dimensional business data of the enterprise's operations to generate a sequence of business semantic segments; Based on the business semantic segment sequence, business causal logic learning is performed, and a business causal association matrix of each business semantic segment is constructed under the constraint of a directed acyclic graph to obtain a business causal graph sequence. The business causal graph sequence is subjected to business causal evolution perception to capture the dynamic evolution law of the business causal structure and predict the future changes in the business causal structure. Based on the future changes in the business causal structure, an evolutionary prior is generated and fed back to the business causal logic learning process to guide the business causal graph to be updated along the historical evolution trend. The confidence of each business causal edge in the business causal graph sequence is evaluated to generate a business causal confidence matrix. The business causal relationship matrix is ​​weighted based on the business causal confidence matrix to generate a reliable business causal graph; Business process attention features are extracted from the business semantic segment sequence to generate business time feature embeddings, and business relationship attention is propagated along the trusted business causal graph to generate business relationship feature embeddings. Conditional business spatiotemporal fusion is performed on the business time feature embedding and the business relationship feature embedding to generate a joint business spatiotemporal feature embedding; Cross-segment business time-series modeling is performed on the spatiotemporal feature embedding of the joint business to generate a global business situation representation. Based on the global business situation representation, business decision reasoning mapping is performed to generate and output enterprise operation decision schemes.

2. The method for constructing an enterprise operation system based on dynamic trusted causal graph evolution according to claim 1, characterized in that, Adaptive semantic segmentation is performed on the multi-dimensional business data of the enterprise operations to generate a sequence of business semantic segments, including: Slide a segmented detection window on the multi-dimensional business data of the enterprise operation, and the segmented detection window and the business data in the previous window form an adjacent window pair; For the adjacent window pairs, the difference between the average value of the business data in the segmented detection window and the average value of the business data in the previous window is calculated to obtain the change in average value; The difference between the fluctuation level of business data within the segmented detection window and the fluctuation level of business data within the previous window is calculated to obtain the fluctuation change. The Jensen-Shannon divergence quantification is used to measure the overall difference between the distribution of business data within the segmented detection window and the distribution of business data within the previous window, thus obtaining the change in distribution pattern. The mean change, volatility change, and distribution pattern change are assigned corresponding weight coefficients and then weighted and superimposed to generate a comprehensive change score. The comprehensive change score characterizes the degree of non-stationarity of the business data within the segmented detection window. When the comprehensive change score exceeds the judgment threshold, it is determined that a business mode change has occurred at the segmented detection window, and the starting boundary of the segmented detection window is marked as a candidate segmentation point. Based on the minimum segment length constraint, adjacent candidate segment points with a spacing less than the minimum segment length are merged, and the candidate segment points are corrected to obtain the segment boundary. The enterprise's multi-dimensional business data is divided according to the segmentation boundaries to generate a business semantic segment sequence.

3. The method for constructing an enterprise operation system based on dynamic trusted causal graph evolution according to claim 1, characterized in that, Based on the business semantic segment sequence, business causal logic learning is performed, and a business causal association matrix for each business semantic segment is constructed under the constraint of a directed acyclic graph, resulting in a business causal graph sequence, including: A parallel dual-branch business causal perception module is constructed, which consists of a multi-core temporal business causal convolution branch and a multivariate real-time business causal attention branch. Through the multi-core temporal business causal convolution branch, business causal convolution is performed on each business variable pair in the business semantic segment sequence to obtain the time-delay business causal strength of each business variable pair, and left padding is performed on the input sequence so that the convolution field of view only covers the current time and historical time, and prevents future information leakage. Through the multivariate real-time business causal attention branch, attention is calculated using the current state of each business variable as the query and key to obtain the real-time business causal strength of each business variable pair. The learnable mask matrix is ​​then applied element by element to the attention score to control the presence and strength of business causal connections between each business variable pair. The time-delay business causality strength and the instant business causality strength are aggregated, and the weights corresponding to redundant business causal paths are driven to approach zero by sparse regularization to generate a business causal correlation matrix. Apply a continuous directed acyclic graph constraint to the business causal relationship matrix to generate a business causal relationship matrix that satisfies the directed acyclic graph structure; The business causal association matrix that satisfies the directed acyclic graph structure is subjected to threshold-based edge pruning to remove weak business causal connections. Domain prior knowledge is injected into the business causal relationship matrix after edge trimming, and the association weights of business variables that are known to have no business causal relationship are reset to zero. The above construction and optimization are performed on each business semantic segment in the business semantic segment sequence to obtain the business causal relationship matrix of each business semantic segment, which constitutes the business causal graph sequence.

4. The method for constructing an enterprise operation system based on dynamic trusted causal graph evolution according to claim 1, characterized in that, The business causal graph sequence is subjected to business causal evolution perception to capture the dynamic evolution law of the business causal structure and predict the future changes in the business causal structure, including: The evolution of the business cause-effect graph sequence on the time axis is explicitly modeled. The business cause-effect association matrix of each business semantic segment in the business cause-effect graph sequence is flattened into a one-dimensional vector along the business variable dimension. The business cause-effect connection strength between each business variable pair is arranged in a fixed order to generate a business cause-effect graph embedding vector sequence. The business cause-effect graph embedding vector sequence is input into the gated loop unit in chronological order. At each time step, the gated loop unit updates the degree of retention of historical evolution information by updating the gate and the degree of forgetting of historical evolution information by resetting the gate, based on the current business cause-effect graph embedding vector and the hidden state of the previous time step. The hidden state is updated time step by time by weighted fusion of the candidate hidden state and the hidden state of the previous time step, generating a hidden state sequence. The hidden state sequence characterizes the dynamic evolution of the business causal topology over time, and the dynamic evolution includes the strengthening trend, weakening trend and migration trend of the business causal edge over time. The hidden state sequence is predicted and mapped by an evolutionary prediction network to estimate the direction and magnitude of the change in the business causal correlation matrix of future business semantic segments relative to the current moment, and to generate the change in the future business causal structure. The change in the future business causal structure quantifies the strengthening, weakening, addition and disappearance trends of future business causal edges, and serves as a priori estimate of the business causal structure at future moments.

5. The method for constructing an enterprise operation system based on dynamic trusted causal graph evolution according to claim 1, characterized in that, Based on the future changes in the business causal structure, an evolutionary prior is generated. This evolutionary prior is fed back into the business causal logic learning process to guide the updating of the business causal graph along historical evolutionary trends. Confidence assessments are performed on each business causal edge in the business causal graph sequence to generate a business causal confidence matrix, including: An evolutionary prior is constructed based on the changes in the future business causal structure. The evolutionary prior represents the prior constraint that the business causal structure evolves smoothly and continuously along the historical evolution trend between adjacent business semantic segments. The evolutionary prior is fed back as a regularization constraint term into the loss function of the business causal logic learning process, so as to constrain the deviation between the actual change in the business causal structure and the future change in the business causal structure. The aforementioned deviation constraint ensures that the business causal relationship matrix of subsequent business semantic segments is updated in a direction consistent with historical evolution trends while maintaining decision accuracy, thereby suppressing abrupt changes in the business causal structure between adjacent business semantic segments. A business causal confidence evaluation network is constructed by flattening the business causal association matrix of each business semantic segment in the business causal graph sequence into a one-dimensional vector, which is used as the input of the business causal confidence evaluation network. The business causal confidence evaluation network independently scores the confidence of each business causal edge corresponding to each pair of business variables in each business semantic segment, and obtains the business causal edge confidence score. The business causal edge confidence score represents the reliability of the existence of a business causal relationship between the corresponding pairs of business variables. The confidence scores of the business causal edges of each business variable pair in each business semantic segment are collected to generate a business causal confidence matrix, which is used to characterize the credibility of each business causal edge.

6. The method for constructing an enterprise operation system based on dynamic trusted causal graph evolution according to claim 1, characterized in that, Based on the business causal confidence matrix, the business causal correlation matrix is ​​weighted to generate a reliable business causal graph, including: The confidence score of the business causal edge corresponding to each business variable pair in the business causal confidence matrix is ​​multiplied element by element by the business causal connection strength of the corresponding business variable pair in the business causal association matrix. The business causal confidence matrix and the business causal association matrix correspond one-to-one. Using the confidence score of the business causal edge as a weight, the business causal connection strength of each business causal edge is scaled so that the effective connection strength of each business causal edge is determined by both the business causal connection strength and the confidence score. The higher the confidence score, the stronger the effective connection strength. Generate a confidence-weighted business causal association matrix, wherein the value of each element in the confidence-weighted business causal association matrix represents the effective connection strength of the corresponding business causal edge after confidence scaling; In the confidence-weighted business causal relationship matrix, business causal edges with high confidence scores dominate, while business causal edges with low confidence scores are continuously and adaptively suppressed according to their confidence scores. Through the continuous adaptive suppression, false business causal connections with low credibility are weakened, while true business causal connections with high credibility are retained, thereby suppressing the interference of false associations on the business causal structure. The confidence-weighted business causal relationship matrix is ​​used as a trusted business causal graph. The weight of each business causal edge in the trusted business causal graph reflects both the strength of the business causal connection and the degree of trust in the business causal relationship.

7. The method for constructing an enterprise operation system based on dynamic trusted causal graph evolution according to claim 1, characterized in that, Business process attention features are extracted from the business semantic segment sequence to generate business time feature embeddings. Business relationship attention is propagated along the trusted business causal graph to generate business relationship feature embeddings, including: Construct a Transformer-based business process attention network, which consists of sequentially stacked business process attention layers and feedforward layers; The business semantic segment sequence is subjected to independent position encoding and cross-channel position encoding, wherein independent position encoding represents the positional relationship within each business semantic segment, and cross-channel position encoding represents the relative position between different business indicators. The two are fused to generate a sequence position representation. The business process attention layer maps the sequence position representation into query vector, key vector and value vector, and performs self-attention calculation. The attention calculation adopts orthogonal basis function expansion approximation to capture long-term business time-series dependencies within and across segments, and generates business time feature embeddings. The time dimension of the original sequence matrix of each business variable in the trusted business causal graph is mapped and compressed through a fully connected layer to construct an initial node representation; Using the trusted business causal graph as an adjacency matrix, a graph attention mechanism is applied to the initial node representation to aggregate the business causal neighbor information of each business variable and update the node representation layer by layer; The attention coefficients of the graph attention mechanism are obtained by linear transformation of the business causal neighbor node representation pairs. The attention coefficients quantify the propagation strength of business causal effects along the business causal path. The updated node representations are aggregated using global mean pooling to generate embedded business relationship features.

8. The method for constructing an enterprise operation system based on dynamic trusted causal graph evolution according to claim 1, characterized in that, Conditional spatiotemporal fusion of the business time feature embedding and the business relationship feature embedding is performed to generate a joint business spatiotemporal feature embedding, including: The business time feature embedding and the business relationship feature embedding are concatenated along the feature dimension to generate an initial joint representation; A gated cross-attention mechanism is used to perform bidirectional conditional information interaction between the business time feature embedding and the business relationship feature embedding, and the initial joint representation is transformed into query vector, key vector and value vector through linear transformation respectively; A gate signal is generated by linear transformation of the embedded business time features. The gate signal is used to dynamically modulate the aggregation strength of the key value of the embedded business relationship features. The gate signal determines the strength of each key value information in the embedded business relationship features participating in the aggregation, thereby realizing conditional control of the aggregation of business relationship features by business time features. A query vector is generated from the business relationship feature embedding, and the query vector guides the focusing direction of the business time feature embedding on different frequency domain components, thereby strengthening the frequency domain components related to the business causal structure in the business time feature embedding, and realizing the conditional guidance of the business relationship feature to focus on the business time feature. The bidirectional cross-attention output obtained from the bidirectional conditional information interaction is fused after being weighted and modulated by the gated signal, and the single-head computation is extended to multi-head parallel computation through a multi-head mechanism. The output of the multi-head parallel computation is processed by a feedforward network and a residual connection to output a joint service spatiotemporal feature embedding.

9. The method for constructing an enterprise operation system based on dynamic trusted causal graph evolution according to claim 1, characterized in that, Cross-segment business time-series modeling is performed on the spatiotemporal feature embedding of the joint business to generate a global business situation representation. Based on the global business situation representation, business decision reasoning mapping is performed to generate and output an enterprise operation decision plan, including: The spatiotemporal features of the joint business are embedded and arranged in chronological order of the semantic segments of each business in the original sequence to construct a segmented business feature sequence. The segmented service feature sequence is input into the cross-segment service temporal modeling module composed of a Transformer encoder. The cross-segment service temporal modeling module is used to model the long-term dependency relationship between each service semantic segment and the cross-segment service context evolution mode. Temporal position information is injected into each business semantic segment in the segmented business feature sequence by position encoding in order to preserve the temporal order between each business semantic segment; By utilizing a multi-head self-attention mechanism, attention computation is extended to parallel computation of multiple attention heads to calculate the attention weights between each business semantic segment in the segmented business feature sequence, thereby capturing the long-range dependencies and contextual evolution patterns between each business semantic segment. The output of the multi-head self-attention is aggregated through multiple layers of encoder to output a global business situation representation that characterizes the spatiotemporal features of the entire historical input window business. The global service situation representation is input into a multilayer perceptron projection network consisting of fully connected layers, and the global service situation representation is mapped to the decision space through linear transformation. Output enterprise operation decision-making solutions, which include optimization and adjustment strategies and resource allocation suggestions for multiple core business indicators, as the result of the enterprise operation system construction and output.

10. A system for constructing an enterprise operation system based on dynamic trusted cause-effect graph evolution, characterized in that, The system for performing the method according to any one of claims 1 to 9, the system comprising: The adaptive business semantic segmentation module is used to adaptively segment multi-dimensional business data of enterprise operations. It uses a segmentation detection window to detect changes in mean, volatility, and distribution patterns between adjacent windows to determine segmentation boundaries and generate a sequence of business semantic segments. A dual-channel business feature extraction network consists of a business time channel and a business relationship channel. The business time channel includes a business process attention network for extracting business process attention features from the business semantic segment sequence and generating business time feature embeddings. The business relationship channel includes a parallel dual-branch business causality perception module, a directed acyclic graph constraint optimization module, a business temporal causal evolution perception module, a business causal confidence evaluation module, and a business relationship attention propagation module. The parallel dual-branch business causality perception module learns the business causal association matrix of each business semantic segment from the business semantic segment sequence. The directed acyclic graph constraint optimization module applies directed acyclic graph constraints to the business causal association matrix. The business temporal causal evolution perception module captures the dynamic evolution law of the business causal structure, predicts the future changes in the business causal structure, and feeds back the future changes in the business causal structure as evolutionary priors to the business causal logic learning process. The business causal confidence evaluation module generates business causal edge confidence scores for each business causal edge. The business relationship attention propagation module propagates information along the confidence-weighted business causal association matrix and generates business relationship feature embeddings. The conditional business spatiotemporal fusion module is used to perform gated cross-attention bidirectional conditional fusion of the business time feature embedding and the business relationship feature embedding to generate a joint business spatiotemporal feature embedding. The decision-making scheme output module is used to perform cross-segment business time-series modeling on the spatiotemporal feature embedding of the joint business, generate a global business situation representation, and output an enterprise operation decision-making scheme based on the global business situation representation through business decision reasoning mapping.

Images