Intelligent enterprise compliance auditing method based on data driving

Abstract

The invention discloses an enterprise intelligent compliance auditing method based on data driving. The method comprises the following steps: S1, automatically collecting auditing data of various heterogeneous data sources in an enterprise in real time through a cross-domain data access interface; s2, performing feature automatic identification and standardization processing on the audit data by adopting a semantic adaptive coding method; s3, a cross-domain collaborative characterization model is constructed by extracting and fusing shared features through a sub-domain multi-expert structure; s4, generating a visual feature heat map by using a hierarchical attention mechanism of the model, and outputting an anomaly detection result; s5, extracting long and short period correlation mode features, inputting the features into a contrast learning framework, and outputting an abnormal risk score; s6, constructing a dynamic enterprise compliance risk knowledge graph; s7, strategy training is carried out, and risk rating is output; and S8, generating an audit report according to the risk rating. According to the invention, efficient and accurate enterprise compliance risk identification and audit decision support are realized.

Description

Technical Field

[0001] This invention relates to the field of enterprise compliance auditing technology, and in particular to a data-driven intelligent enterprise compliance auditing method. Background Technology

[0002] In recent years, corporate compliance auditing technology has gradually shifted from traditional manual auditing to data-driven intelligent auditing methods. Artificial intelligence, data mining, and anomaly detection technologies have been increasingly introduced into the field of corporate compliance auditing. Currently, mainstream intelligent auditing technologies mainly employ traditional statistical methods, data analysis models, or single-scale attention mechanisms to detect anomalies in corporate operational data in order to identify potential compliance risks. Among these, auditing methods based on traditional statistical analysis rely on setting predetermined thresholds and manual rules to identify anomalies, but suffer from complex rule definitions and the inability to adaptively update them. Data mining models rely on training classification or prediction models with historical data, which, while improving audit efficiency to some extent, lacks accuracy and generalization in identifying new types of anomalies. Auditing methods based on attention mechanisms mainly focus on single-scale feature relationships, making it difficult to comprehensively and accurately identify corporate compliance risks with cross-scale and cross-domain characteristics.

[0003] Specifically, existing auditing schemes based on manual rules or statistical analysis are limited by human experience and fixed threshold rules, easily leading to high false positive and false negative rates. Meanwhile, the lack of generalization ability in data mining models manifests as sensitivity to changes in data distribution and weak learning ability for small amounts of anomalous data, affecting the effectiveness of practical audit applications. Furthermore, existing attention mechanisms are mostly used for short-cycle or single-scale feature extraction, showing significant inadequacy in identifying anomalous risks in enterprise operations that simultaneously exhibit long-cycle trends and short-cycle fluctuations. At the same time, existing auditing methods lack effective means of interpreting and visualizing anomaly detection results, making it difficult for auditors to quickly understand the source and impact path of anomalous risks, significantly reducing the credibility and operability of audit results. Further, existing methods fail to effectively establish dynamic relationships between risk nodes within the enterprise, lack effective identification and proactive learning updates of anomalous risk propagation paths, and cannot achieve continuous adaptive optimization of the risk map.

[0004] Therefore, how to provide a data-driven intelligent compliance audit method for enterprises is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a data-driven intelligent compliance audit method for enterprises, which has the advantages of high audit accuracy, accurate risk identification, strong interpretability of audit results, and significantly improved audit efficiency.

[0006] A data-driven intelligent compliance audit method for enterprises according to an embodiment of the present invention includes the following steps:

[0007] S1. Automatically collect raw audit data from various heterogeneous data sources within the enterprise in real time through cross-domain data access interfaces and output audit datasets in a unified format;

[0008] S2. Use a semantic adaptive coding method to automatically identify and standardize the feature dimensions of the audit dataset and output standardized feature data;

[0009] S3. By using a domain-based multi-expert structure, feature extraction is performed on standardized feature data, and shared features are fused to construct a cross-domain collaborative representation model.

[0010] S4. Use the hierarchical attention mechanism in the cross-domain collaborative representation model to generate a visual feature heatmap from the standardized feature data, extract anomaly localization information through the visual feature heatmap and output the anomaly detection results.

[0011] S5. Perform dual-scale attention extraction on the standardized feature data, extract long-cycle association pattern features and short-cycle association pattern features, input the extracted pattern features into the contrastive learning framework for representation learning and output anomaly risk scores.

[0012] S6. Based on the anomaly detection results and anomaly risk scores, construct a dynamically updated enterprise compliance risk knowledge graph and use an active learning mechanism to update and label newly detected anomalies and output the updated risk knowledge graph.

[0013] S7. Train strategies on abnormal risk scores and updated risk knowledge graphs and output risk rating results;

[0014] S8. Input the output risk rating results into the natural language generation model to generate an audit report.

[0015] Optionally, the original audit data specifically includes: enterprise resource planning system operation logs, customer relationship management system transaction records, financial accounting system vouchers and transaction details, supply chain management system inventory flow data, manufacturing execution system production batches and process parameters, network security device access logs, user authentication and permission change logs, email gateway communication records, third-party payment platform transaction reconciliation data, and internal chat and collaborative office platform message records.

[0016] Optionally, S2 specifically includes:

[0017] S21. All feature dimensions in the original audit dataset are grouped based on the frequency of co-occurrence within the sliding window of the same domain data. Feature dimensions with a co-occurrence frequency higher than the 75th percentile of the data window are marked as core dimension sets, and the remaining dimensions are marked as non-core dimension sets.

[0018] S22. Calculate the Shannon entropy value within the sliding window for the core dimension set and the non-core dimension set respectively, and normalize the entropy value according to the sum of entropy within the window to generate entropy weight coefficient;

[0019] S23. For the output entropy weight coefficient, perform multi-interval segmentation normalization on the numerical dimension according to the preset quantile to generate numerical normalized features, and perform semantic mapping on the text dimension based on the cosine similarity of the pre-trained word vectors in the domain terminology dictionary to generate text normalized features.

[0020] S24. Based on the differences in mean and variance between the current batch and the historical baseline batch, the drift correction bias is calculated using the exponentially weighted moving average algorithm, and the statistical parameters of the current batch are corrected accordingly.

[0021] S25. Using a greedy subset selection algorithm based on the principle of minimum reconstruction error, the corrected normalized features are sorted by entropy weight coefficients, and incremental reconstruction tests are performed to filter the most sensitive dimensions, outputting a subset of reconstruction sensitive dimensions.

[0022] S26. Use the ridge regression model with L2 regularization to perform least squares calibration on the feature covariance matrix of the full sample to generate the mapping matrix calibration parameters, and output the calibrated global feature mapping matrix.

[0023] S27. Combining the entropy weight coefficient, corrected statistical parameters, reconstructed sensitive dimension subset, and calibrated mapping matrix, the normalized features are finally fused to generate standardized feature data with metadata labels.

[0024] Optionally, S3 specifically includes:

[0025] S31. The standardized feature data is segmented in the time series dimension and spatial domain dimension according to the preset sliding window length and overlap rate. The similarity matching between the feature vector in each window and the center feature vector of the previous window is calculated, and elements with similarity higher than the threshold are selected to generate an intra-domain enhanced feature matrix.

[0026] S32. Based on the cross-domain position index mapping table, feature pairing is performed on the intra-domain enhanced feature matrix at the same position. The paired feature groups are weighted and fused according to the pairing confidence and then normalized to form a cross-domain context resonance matrix.

[0027] S33. Extract feature blocks from the cross-domain context resonance matrix at both the high-frequency statistical scale and the low-frequency temporal scale, and then, according to the predefined interleaving rules, alternately splice them in the time-frequency dual dimension in the order of temporal block-statistical block-temporal block to generate a multi-granularity interleaved feature representation.

[0028] S34. Based on multi-granularity interleaved feature representation, the K-means++ algorithm is used to initialize an equal number of prototype vectors and assign them uniform weights. A dynamic sampling strategy is adopted to update the prototype vectors and corresponding feature subsets by decreasing the sampling ratio in each iteration until the sampling error converges. The evolved prototype vector set is then output.

[0029] S35. The evolved prototype vector set and the multi-granularity interwoven feature representation are concatenated according to the new dimension tensor. Based on the least squares residual minimization criterion, the concatenated tensor is adaptively calibrated to generate the fusion optimization matrix.

[0030] S36. Perform matrix decomposition on the fusion optimization matrix according to the preset row and column dimensions to extract parameter submatrices, and construct a cross-domain collaborative representation model according to the parameter submatrices.

[0031] Optionally, S36 specifically includes:

[0032] S361. Generate the row principal index set IR and the column principal index set IC based on the row principal variance gradient and column principal variance gradient of the fusion optimization matrix. Divide the fusion optimization matrix into sub-matrix groups of unequal size according to IR and IC and output the sub-matrix groups.

[0033] S362. For each submatrix, construct a joint histogram of adjacent row pairs and calculate row-directed mutual information; construct a joint histogram of adjacent column pairs and calculate column-directed mutual information; filter high-information submatrixes based on a preset mutual information percentile threshold; and perform local spectral domain compression on the filtered submatrixes.

[0034] S363. Calculate the pairwise normalized Frobenius inner product for the compressed submatrix group as the coupling coefficient, and use the coefficient matrix as input to divide the coupling group using a hierarchical clustering algorithm. Concatenate and splice the submatrix in each group in descending order of coupling coefficient to form a multi-level parameter structure.

[0035] S364. Perform a two-dimensional discrete Fourier transform on each multi-level parameter structure to extract phase information, use the phase expansion algorithm to eliminate phase jumps and calculate the phase difference between adjacent levels, convert the phase difference into a complex rotation correction factor and apply it to the corresponding level, and output the corrected parameter tensor.

[0036] S365. Map the corrected parameter tensor to the inter-layer connectivity coefficient tensor of the cross-domain collaborative representation model according to the first dimension, to the attention weight tensor according to the second dimension, and to the feature transformation tensor according to the third dimension. Obtain the parameters of each tensor by solving the linear least squares problem with Tikhonov regularization and construct the cross-domain collaborative representation model.

[0037] Optionally, S4 specifically includes:

[0038] S41. Input standardized feature data into the cross-domain collaborative representation model, perform forward propagation calculation based on the inter-layer connection coefficient tensor, attention weight tensor, and feature transformation tensor, and generate a multi-scale feature vector set within the domain.

[0039] S42. Using the multi-scale feature vector group within the domain as input, calculate the self-attention weights of each scale feature dimension and perform weighted aggregation within the dimension to obtain the scale attention feature vector within the domain.

[0040] S43. Calculate the interaction attention weights between different scales by using the cross-scale feature coupling method to obtain the cross-scale coupled attention matrix, where the matrix elements represent the degree of attention contribution between features at a specific scale.

[0041] S44. Use the cross-scale coupled attention matrix to perform element-wise weighted summation on the intra-domain scale attention feature vector to generate an intra-domain cross-scale fused feature vector. Use the heatmap mapping method to map the fused feature vector values ​​to a predefined color space to form a two-dimensional visualized feature heatmap.

[0042] S45. Perform abnormal region localization on the visualized feature heatmap based on the neighborhood difference threshold. By calculating the difference between the feature intensity of each location in the heatmap and the neighborhood mean, locations with a difference exceeding the preset threshold are marked as abnormal locations, and a set of abnormal location coordinates is generated.

[0043] S46. Based on the set of abnormal location coordinates, extract the corresponding original feature vectors from the standardized feature data, calculate the abnormality score using the distance weighting method, and generate the abnormality detection results by arranging them in descending order of the score.

[0044] Optionally, S43 specifically includes:

[0045] S431. Perform Hilbert transform on the domain-scale attention feature vector within a preset time window to obtain the instantaneous phase of the feature vector, extract the instantaneous phase distribution curve of each scale feature vector in the time dimension, perform dynamic pairing of scale features based on the correlation coefficient of the phase curve, and output the dynamic pairing set of scale features.

[0046] S432. For each scale feature pair in the scale feature dynamic pairing set, calculate the dynamic covariance matrix between scale feature vectors, perform eigenvalue decomposition on the covariance matrix and take the largest eigenvalue as the initial value of the interaction weight between scale pairs, and output the set of initial values ​​of scale interaction weight.

[0047] S433. Based on the initial set of scale interaction weights, and using the scale attention feature vector as a foundation, an adaptive superposition enhancement method for feature vectors in the time dimension is adopted to perform adaptive superposition processing on paired scale feature vectors to generate a scale resonance enhancement feature vector group.

[0048] S434. Perform a multi-window Fourier transform on the scale resonance enhancement feature vector group in the spectral domain. Determine the sensitivity level of each scale feature by analyzing the intensity of the spectral distribution of the feature vectors between scales. Sort the scale sensitivity levels from high to low according to the proportion of spectral energy.

[0049] S435. Based on the scale sensitivity ranking results, construct an inter-scale asymmetric interaction relationship matrix, with high-sensitivity scale features as the dominant scale, and assign progressively decreasing weight values ​​to low-sensitivity scale features to form an inter-scale asymmetric interaction attention relationship matrix.

[0050] S436. Perform Softmax normalization on each row of the asymmetric interaction attention matrix between scales to obtain the attention contribution degree between each scale direction, and finally form a cross-scale coupled attention matrix.

[0051] Optionally, S5 specifically includes:

[0052] S51. The standardized feature data is segmented into long-period sliding windows and short-period sliding windows respectively. During the window movement, the non-linear trend degree of the feature sequence within the window is calculated in real time. The window length is adaptively adjusted by the non-linear trend degree, and the long-period adaptive window data set and the short-period adaptive window data set are output respectively.

[0053] S52. For a long-period adaptive window dataset, perform singular spectrum analysis on the feature vectors within the window to decompose the trend component and fluctuation component. Determine the trend dominance strength by calculating the ratio of the explained variance of the trend component to the fluctuation component, and use the trend dominance strength as the long-period attention weight.

[0054] S53. For a short-period adaptive window dataset, perform empirical mode decomposition on the feature vector sequence within each window to determine the energy transfer rate between each mode function within the window, and generate short-period modal attention weights based on the energy transfer rate of the mode functions.

[0055] S54. Long-cycle attention weight and short-cycle modal attention weight are used to weight and fuse the long-cycle trend component and the short-cycle high-frequency modal function respectively to generate long-cycle correlation pattern feature vector and short-cycle correlation pattern feature vector.

[0056] S55. Using the feature vectors of long-period correlation patterns and short-period correlation patterns as input, cross-period positive and negative sample pairs are generated through a dynamic pairing mechanism. The dynamic pairing mechanism adaptively determines positive and negative samples based on the pattern similarity of adjacent time windows, adjusts the sample pair selection ratio in real time and calculates the corresponding contrast loss, and iteratively optimizes the contrast loss to obtain cross-period pattern differential feature representation.

[0057] S56. Calculate the dynamic time-normalized distance between the differential feature representation of the cross-cycle pattern and the differential feature representation of the historical normal pattern. Based on the reciprocal of the distance, generate anomaly risk scores through a nonlinear power exponential transformation function.

[0058] Optionally, S6 specifically includes:

[0059] S61. Based on the abnormal location and corresponding abnormal risk score in the abnormal detection results, dynamically extract the difference fingerprint between the abnormal location features and the features in the historical risk map, use the difference fingerprint to automatically locate the risk nodes and output the set of located risk nodes.

[0060] S62. For the set of risk nodes after location, based on the similarity of the time series abnormal features of the risk nodes, the nodes are dynamically merged and split, and the granularity of the risk nodes is optimized to generate an optimized set of risk nodes.

[0061] S63. Using the optimized set of risk nodes as input, determine the directional transmission relationship between risk nodes based on the transmission effect of abnormal risk scores between nodes, determine the edge weights between nodes using the decay rate of abnormal risk scores, and generate a dynamically updated initial enterprise compliance risk knowledge graph.

[0062] S64. For newly detected abnormal feature data, based on the historical evolution trajectory of existing nodes in the initial enterprise compliance risk knowledge graph, predict the possible evolution trend of abnormal features in real time, identify newly added unknown risk nodes through trend prediction deviation value and perform active labeling, and output the newly added unknown risk nodes after active labeling.

[0063] S65. Based on the newly added unknown risk nodes after active annotation, dynamically add nodes and update edge weights in the initial enterprise compliance risk knowledge graph. The edge weights are determined according to the interaction strength of abnormal risk scores between the newly added nodes and the identified risk nodes. Output the updated enterprise compliance risk knowledge graph.

[0064] S66. Encode the node evolution pattern of the updated enterprise compliance risk knowledge graph into a high-dimensional temporal feature space. By dynamically tracking the stability of the risk propagation path between nodes, evaluate the stability of graph nodes and connections in real time, and actively trigger reconstruction of low-stability areas to output a stable and optimized enterprise compliance risk knowledge graph.

[0065] Optionally, S7 specifically includes:

[0066] S71. Based on the updated enterprise compliance risk knowledge graph, track the path evolution of risk nodes in the graph, construct a path evolution feature sequence, and generate a risk status evolution sequence according to the path change frequency and abnormal risk score of risk nodes.

[0067] S72. For the risk state evolution sequence, based on the abnormal risk score, a risk state change trend map is constructed using the state transition difference method, and the differentiated trend of risk state change is used as input data for policy training.

[0068] S73. Implement time-series reverse tracing analysis on the risk status change trend map, extract key propagation nodes in the abnormal risk transmission path in real time through reverse tracing analysis, and dynamically determine the strategy sensitivity factors of risk nodes on different transmission paths.

[0069] S74. Based on the risk node's strategy sensitivity factor, the strategy response intensity of the risk propagation node is dynamically adjusted through the coupling relationship between the abnormal risk score and the node location, and the strategy weight is dynamically updated to form a dynamic matrix of node strategy weights.

[0070] S75. Generate the strategy-driven feature representation of risk nodes using the dynamic matrix of node strategy weights, calculate the strategy deviation between the strategy-driven features of risk nodes and the predefined risk rating criteria in real time, and determine the rating range of risk nodes based on the magnitude of the strategy deviation.

[0071] S76. Based on the risk node strategy-driven characteristics and strategy deviation, automatically generate the strategy risk rating of the enterprise compliance risk node using the dynamic floating range method, and output the dynamically updated strategy risk rating results.

[0072] The beneficial effects of this invention are:

[0073] (1) By adopting a cross-domain data access interface and semantic adaptive coding method, this invention can automatically collect and accurately standardize audit data from various heterogeneous data sources within an enterprise in real time, thereby achieving efficient cross-domain data integration and unified standardization, effectively improving the accuracy and efficiency of audit data processing, and enhancing the reliability and stability of enterprise compliance audit data.

[0074] (2) This invention, through a multi-expert structure and a hierarchical attention mechanism that couples features across scales, can achieve in-depth mining of feature information and anomaly location in audit data, significantly improving the ability to accurately identify compliance risks, and demonstrating better adaptability and accuracy in the analysis of complex operational data of enterprises.

[0075] (3) In terms of risk identification and risk rating, this invention effectively solves the problem of lack of risk propagation path analysis and dynamic adaptive optimization in the existing technology by dynamically constructing a corporate compliance risk knowledge graph and an active learning mechanism. It breaks through the bottleneck of traditional audit technology relying too much on manual rules and experience judgment, realizes continuous optimization and self-improvement of risk identification, and thus effectively improves the objectivity and interpretability of corporate compliance audit.

[0076] (4) Through dynamic strategy training and risk rating mechanism, this invention can achieve precise dynamic adjustment of the strategy response intensity at risk nodes, significantly improve the dynamism and accuracy of enterprise compliance risk rating, and demonstrate better audit decision support effect in real-time monitoring and prediction of abnormal risks. Attached Figure Description

[0077] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0078] Figure 1 This is a schematic diagram of the overall process of a data-driven intelligent compliance audit method for enterprises proposed in this invention;

[0079] Figure 2 This is a schematic diagram of the cross-domain collaborative representation model structure of a data-driven enterprise intelligent compliance audit method proposed in this invention.

[0080] Figure 3 This is a schematic diagram of an enterprise compliance risk knowledge graph for a data-driven intelligent compliance audit method proposed in this invention. Detailed Implementation

[0081] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0082] refer to Figures 1-3 A data-driven intelligent compliance audit method for enterprises includes the following steps:

[0083] S1. Automatically collect raw audit data from various heterogeneous data sources within the enterprise in real time through cross-domain data access interfaces and output audit datasets in a unified format;

[0084] S2. Use a semantic adaptive coding method to automatically identify and standardize the feature dimensions of the audit dataset and output standardized feature data;

[0085] S3. By using a domain-based multi-expert structure, feature extraction is performed on standardized feature data, and shared features are fused to construct a cross-domain collaborative representation model.

[0086] S4. Use the hierarchical attention mechanism in the cross-domain collaborative representation model to generate a visual feature heatmap from the standardized feature data, extract anomaly localization information through the visual feature heatmap and output the anomaly detection results.

[0087] S5. Perform dual-scale attention extraction on the standardized feature data, extract long-cycle association pattern features and short-cycle association pattern features, input the extracted pattern features into the contrastive learning framework for representation learning and output anomaly risk scores.

[0088] S6. Based on the anomaly detection results and anomaly risk scores, construct a dynamically updated enterprise compliance risk knowledge graph and use an active learning mechanism to update and label newly detected anomalies and output the updated risk knowledge graph.

[0089] S7. Train strategies on abnormal risk scores and updated risk knowledge graphs and output risk rating results;

[0090] S8. Input the output risk rating results into the natural language generation model to generate an audit report.

[0091] By employing cross-domain automatic data collection and semantic adaptive coding methods, the system achieves standardized and unified processing of various heterogeneous audit data within the enterprise, effectively improving data quality. The system utilizes a domain-specific multi-expert structure and a hierarchical attention mechanism to achieve precise extraction of audit data features and visualized location of anomalies, significantly improving the accuracy of anomaly detection. Furthermore, through dual-scale attention contrastive learning and dynamically updated risk knowledge graphs, the system effectively enhances the comprehensive identification capability of enterprise compliance risks, improving the accuracy of risk rating and the real-time nature and objectivity of audit decisions.

[0092] In this embodiment, the original audit data specifically includes: enterprise resource planning system operation logs, customer relationship management system transaction records, financial accounting system vouchers and transaction details, supply chain management system inventory flow data, manufacturing execution system production batches and process parameters, network security device access logs, user authentication and permission change logs, email gateway communication records, third-party payment platform transaction reconciliation data, and internal chat and collaborative office platform message records.

[0093] The collected raw audit data covers enterprise resource planning system operation logs, customer relationship management system transaction records, financial accounting system vouchers and transaction details, supply chain management system inventory flow data, manufacturing execution system production batches and process parameters, network security device access logs, user authentication and permission change logs, email gateway communication records, third-party payment platform transaction reconciliation data, and internal chat and collaborative office platform message records. The data sources are comprehensive and multi-dimensional, effectively ensuring the integrity of audit data, thereby improving the accuracy of risk identification and the reliability of audit results in the enterprise compliance audit process.

[0094] In this embodiment, S2 specifically includes:

[0095] S21. All feature dimensions in the original audit dataset are grouped based on the frequency of co-occurrence within the sliding window of the same domain data. Feature dimensions with a co-occurrence frequency higher than the 75th percentile of the data window are marked as core dimension sets, and the remaining dimensions are marked as non-core dimension sets.

[0096] S22. Calculate the Shannon entropy value within the sliding window for the core dimension set and the non-core dimension set respectively, and normalize the entropy value according to the sum of entropy within the window to generate entropy weight coefficient;

[0097] S23. For the output entropy weight coefficient, perform multi-interval segmentation normalization on the numerical dimension according to the preset quantile to generate numerical normalized features, and perform semantic mapping on the text dimension based on the cosine similarity of the pre-trained word vectors in the domain terminology dictionary to generate text normalized features.

[0098] S24. Based on the differences in mean and variance between the current batch and the historical baseline batch, the drift correction bias is calculated using the exponentially weighted moving average algorithm, and the statistical parameters of the current batch are corrected accordingly.

[0099] S25. Using a greedy subset selection algorithm based on the principle of minimum reconstruction error, the corrected normalized features are sorted by entropy weight coefficients, and incremental reconstruction tests are performed to filter the most sensitive dimensions, outputting a subset of reconstruction sensitive dimensions.

[0100] S26. Use the ridge regression model with L2 regularization to perform least squares calibration on the feature covariance matrix of the full sample to generate the mapping matrix calibration parameters, and output the calibrated global feature mapping matrix.

[0101] S27. Combining the entropy weight coefficient, corrected statistical parameters, reconstructed sensitive dimension subset, and calibrated mapping matrix, the normalized features are finally fused to generate standardized feature data with metadata labels.

[0102] By using the frequency of co-occurrence of homogeneous data in the same domain to perform dimensional clustering and entropy weighting coefficient calculation, and employing piecewise normalization and domain terminology mapping, adaptive and accurate identification and standardized processing of feature dimensions are achieved. At the same time, by using the drift correction method of exponential weighted moving average and the sensitive dimension screening method based on the principle of minimum reconstruction error, the quality and stability of feature data are effectively guaranteed, and the accuracy of enterprise compliance audit data processing and the sensitivity of anomaly risk detection are significantly improved.

[0103] In this embodiment, S3 specifically includes:

[0104] S31. The standardized feature data is segmented in the time series dimension and spatial domain dimension according to the preset sliding window length and overlap rate. The similarity matching between the feature vector in each window and the center feature vector of the previous window is calculated, and elements with similarity higher than the threshold are selected to generate an intra-domain enhanced feature matrix.

[0105] S32. Based on the cross-domain position index mapping table, feature pairing is performed on the intra-domain enhanced feature matrix at the same position. The paired feature groups are weighted and fused according to the pairing confidence and then normalized to form a cross-domain context resonance matrix.

[0106] S33. Extract feature blocks from the cross-domain context resonance matrix at both the high-frequency statistical scale and the low-frequency temporal scale, and then, according to the predefined interleaving rules, alternately splice them in the time-frequency dual dimension in the order of temporal block-statistical block-temporal block to generate a multi-granularity interleaved feature representation.

[0107] S34. Based on multi-granularity interleaved feature representation, the K-means++ algorithm is used to initialize an equal number of prototype vectors and assign them uniform weights. A dynamic sampling strategy is adopted to update the prototype vectors and corresponding feature subsets by decreasing the sampling ratio in each iteration until the sampling error converges. The evolved prototype vector set is then output.

[0108] S35. The evolved prototype vector set and the multi-granularity interwoven feature representation are concatenated according to the new dimension tensor. Based on the least squares residual minimization criterion, the concatenated tensor is adaptively calibrated to generate the fusion optimization matrix.

[0109] S36. Perform matrix decomposition on the fusion optimization matrix according to the preset row and column dimensions to extract parameter submatrices, and construct a cross-domain collaborative representation model according to the parameter submatrices.

[0110] By dynamically enhancing the windowing of intra-domain scale features and fusing the contextual resonance of cross-domain location index mapping, the accuracy and contextual relevance of cross-domain feature fusion are effectively improved. Through the feature interweaving of high-frequency statistical scale and low-frequency temporal scale, multi-granularity feature representation, and prototype vector evolution of dynamic sampling strategies, the accuracy and robustness of enterprise audit data feature representation are significantly improved, thereby achieving more accurate identification of enterprise compliance anomaly risks and audit analysis results.

[0111] In this embodiment, S36 specifically includes:

[0112] S361. Generate the row principal index set IR and the column principal index set IC based on the row principal variance gradient and column principal variance gradient of the fusion optimization matrix. Divide the fusion optimization matrix into sub-matrix groups of unequal size according to IR and IC and output the sub-matrix groups.

[0113] S362. For each submatrix, construct a joint histogram of adjacent row pairs and calculate row-directed mutual information; construct a joint histogram of adjacent column pairs and calculate column-directed mutual information; filter high-information submatrixes based on a preset mutual information percentile threshold; and perform local spectral domain compression on the filtered submatrixes.

[0114] S363. Calculate the pairwise normalized Frobenius inner product for the compressed submatrix group as the coupling coefficient, and use the coefficient matrix as input to divide the coupling group using a hierarchical clustering algorithm. Concatenate and splice the submatrix in each group in descending order of coupling coefficient to form a multi-level parameter structure.

[0115] S364. Perform a two-dimensional discrete Fourier transform on each multi-level parameter structure to extract phase information, use the phase expansion algorithm to eliminate phase jumps and calculate the phase difference between adjacent levels, convert the phase difference into a complex rotation correction factor and apply it to the corresponding level, and output the corrected parameter tensor.

[0116] S365. Map the corrected parameter tensor to the inter-layer connectivity coefficient tensor of the cross-domain collaborative representation model according to the first dimension, to the attention weight tensor according to the second dimension, and to the feature transformation tensor according to the third dimension. Obtain the parameters of each tensor by solving the linear least squares problem with Tikhonov regularization and construct the cross-domain collaborative representation model.

[0117] By integrating differential gradient partitioning of the main index set of the optimized matrix with local spectral domain compression of mutual information of submatrices, effective dimensionality reduction and refined processing of feature data are achieved. Hierarchical clustering based on normalized inner product coupling coefficients and two-dimensional discrete Fourier transform phase correction of parameter structure ensure the accuracy and stability of feature representation and model parameters, thereby effectively improving the performance of cross-domain collaborative representation model and enhancing the adaptability of auditing methods to complex enterprise data and the reliability of anomaly detection.

[0118] In this embodiment, S4 specifically includes:

[0119] S41. Input standardized feature data into the cross-domain collaborative representation model, perform forward propagation calculation based on the inter-layer connection coefficient tensor, attention weight tensor, and feature transformation tensor, and generate a multi-scale feature vector set within the domain.

[0120] S42. Using the multi-scale feature vector group within the domain as input, calculate the self-attention weights of each scale feature dimension and perform weighted aggregation within the dimension to obtain the scale attention feature vector within the domain.

[0121] S43. Calculate the interaction attention weights between different scales by using the cross-scale feature coupling method to obtain the cross-scale coupled attention matrix, where the matrix elements represent the degree of attention contribution between features at a specific scale.

[0122] S44. Use the cross-scale coupled attention matrix to perform element-wise weighted summation on the intra-domain scale attention feature vector to generate an intra-domain cross-scale fused feature vector. Use the heatmap mapping method to map the fused feature vector values ​​to a predefined color space to form a two-dimensional visualized feature heatmap.

[0123] S45. Perform abnormal region localization on the visualized feature heatmap based on the neighborhood difference threshold. By calculating the difference between the feature intensity of each location in the heatmap and the neighborhood mean, locations with a difference exceeding the preset threshold are marked as abnormal locations, and a set of abnormal location coordinates is generated.

[0124] S46. Based on the set of abnormal location coordinates, extract the corresponding original feature vectors from the standardized feature data, calculate the abnormality score using the distance weighting method, and generate the abnormality detection results by arranging them in descending order of the score.

[0125] By employing multi-scale attention feature extraction and cross-scale feature coupling methods through a cross-domain collaborative representation model, the interaction relationships between features of different scales in audit data are accurately captured and fused. Furthermore, the heatmap mapping method is used to achieve visualized localization of abnormal features, effectively improving the accurate identification capability of abnormal locations during enterprise compliance audits. Moreover, by calculating the location of abnormal regions using a neighborhood difference threshold and a distance-weighted method, the degree of abnormality is scored, achieving accurate quantification of abnormality detection results and enhancing the objectivity and interpretability of audit decisions.

[0126] In this embodiment, S43 specifically includes:

[0127] S431. Perform Hilbert transform on the domain-scale attention feature vector within a preset time window to obtain the instantaneous phase of the feature vector, extract the instantaneous phase distribution curve of each scale feature vector in the time dimension, perform dynamic pairing of scale features based on the correlation coefficient of the phase curve, and output the dynamic pairing set of scale features.

[0128] S432. For each scale feature pair in the scale feature dynamic pairing set, calculate the dynamic covariance matrix between scale feature vectors, perform eigenvalue decomposition on the covariance matrix and take the largest eigenvalue as the initial value of the interaction weight between scale pairs, and output the set of initial values ​​of scale interaction weight.

[0129] S433. Based on the initial set of scale interaction weights, and using the scale attention feature vector as a foundation, an adaptive superposition enhancement method for feature vectors in the time dimension is adopted to perform adaptive superposition processing on paired scale feature vectors to generate a scale resonance enhancement feature vector group.

[0130] S434. Perform a multi-window Fourier transform on the scale resonance enhancement feature vector group in the spectral domain. Determine the sensitivity level of each scale feature by analyzing the intensity of the spectral distribution of the feature vectors between scales. Sort the scale sensitivity levels from high to low according to the proportion of spectral energy.

[0131] S435. Based on the scale sensitivity ranking results, construct an inter-scale asymmetric interaction relationship matrix, with high-sensitivity scale features as the dominant scale, and assign progressively decreasing weight values ​​to low-sensitivity scale features to form an inter-scale asymmetric interaction attention relationship matrix.

[0132] S436. Perform Softmax normalization on each row of the asymmetric interaction attention matrix between scales to obtain the attention contribution degree between each scale direction, and finally form a cross-scale coupled attention matrix.

[0133] By dynamically pairing the instantaneous phases of scale feature vectors and accurately calculating the interaction weights between scale pairs, adaptive coupling and enhancement of scale attention feature vectors across scales within the domain are achieved. At the same time, based on multi-window spectral sensitivity level analysis, an asymmetric interaction attention relationship matrix between scales is constructed and normalized, which significantly improves the accuracy of the scale correlation expression of feature vectors and the sensitivity of anomaly detection, thereby enhancing the ability of enterprise intelligent compliance audit methods to identify anomaly risks in complex data scenarios.

[0134] In this embodiment, S5 specifically includes:

[0135] S51. The standardized feature data is segmented into long-period sliding windows and short-period sliding windows respectively. During the window movement, the non-linear trend degree of the feature sequence within the window is calculated in real time. The window length is adaptively adjusted by the non-linear trend degree, and the long-period adaptive window data set and the short-period adaptive window data set are output respectively.

[0136] S52. For a long-period adaptive window dataset, perform singular spectrum analysis on the feature vectors within the window to decompose the trend component and fluctuation component. Determine the trend dominance strength by calculating the ratio of the explained variance of the trend component to the fluctuation component, and use the trend dominance strength as the long-period attention weight.

[0137] S53. For a short-period adaptive window dataset, perform empirical mode decomposition on the feature vector sequence within each window to determine the energy transfer rate between each mode function within the window, and generate short-period modal attention weights based on the energy transfer rate of the mode functions.

[0138] S54. Long-cycle attention weight and short-cycle modal attention weight are used to weight and fuse the long-cycle trend component and the short-cycle high-frequency modal function respectively to generate long-cycle correlation pattern feature vector and short-cycle correlation pattern feature vector.

[0139] S55. Using the feature vectors of long-period correlation patterns and short-period correlation patterns as input, cross-period positive and negative sample pairs are generated through a dynamic pairing mechanism. The dynamic pairing mechanism adaptively determines positive and negative samples based on the pattern similarity of adjacent time windows, adjusts the sample pair selection ratio in real time and calculates the corresponding contrast loss, and iteratively optimizes the contrast loss to obtain cross-period pattern differential feature representation.

[0140] S56. Calculate the dynamic time-normalized distance between the differential feature representation of the cross-cycle pattern and the differential feature representation of the historical normal pattern. Based on the reciprocal of the distance, generate anomaly risk scores through a nonlinear power exponential transformation function.

[0141] By adaptively adjusting the nonlinear trend of long and short period window data, and combining singular spectrum analysis and empirical mode decomposition, the trend dominance strength and mode energy transfer characteristics of audit data are accurately extracted. Furthermore, by combining a dynamic pairing mechanism and a contrastive loss optimization strategy, differentiated feature representations are effectively generated. Finally, by utilizing dynamic time regularization distance and nonlinear power exponent transformation methods, the accuracy and reliability of anomaly risk score calculation are significantly improved, thereby effectively enhancing the risk identification accuracy of intelligent compliance audit methods in complex data environments.

[0142] In this embodiment, S6 specifically includes:

[0143] S61. Based on the abnormal location and corresponding abnormal risk score in the abnormal detection results, dynamically extract the difference fingerprint between the abnormal location features and the features in the historical risk map, use the difference fingerprint to automatically locate the risk nodes and output the set of located risk nodes.

[0144] S62. For the set of risk nodes after location, based on the similarity of the time series abnormal features of the risk nodes, the nodes are dynamically merged and split, and the granularity of the risk nodes is optimized to generate an optimized set of risk nodes.

[0145] S63. Using the optimized set of risk nodes as input, determine the directional transmission relationship between risk nodes based on the transmission effect of abnormal risk scores between nodes, determine the edge weights between nodes using the decay rate of abnormal risk scores, and generate a dynamically updated initial enterprise compliance risk knowledge graph.

[0146] S64. For newly detected abnormal feature data, based on the historical evolution trajectory of existing nodes in the initial enterprise compliance risk knowledge graph, predict the possible evolution trend of abnormal features in real time, identify newly added unknown risk nodes through trend prediction deviation value and perform active labeling, and output the newly added unknown risk nodes after active labeling.

[0147] S65. Based on the newly added unknown risk nodes after active annotation, dynamically add nodes and update edge weights in the initial enterprise compliance risk knowledge graph. The edge weights are determined according to the interaction strength of abnormal risk scores between the newly added nodes and the identified risk nodes. Output the updated enterprise compliance risk knowledge graph.

[0148] S66. Encode the node evolution pattern of the updated enterprise compliance risk knowledge graph into a high-dimensional temporal feature space. By dynamically tracking the stability of the risk propagation path between nodes, evaluate the stability of graph nodes and connections in real time, and actively trigger reconstruction of low-stability areas to output a stable and optimized enterprise compliance risk knowledge graph.

[0149] By dynamically extracting the differential fingerprints of risk nodes and optimizing the node granularity, the enterprise's compliance risk nodes are accurately identified and dynamically updated. Based on the decay rate of abnormal risk scores and the prediction of the historical evolution trajectory of risk nodes, unknown risk nodes are proactively identified and labeled. Furthermore, the structure of the risk knowledge graph and the node connection relationship are dynamically updated by utilizing the interaction intensity of abnormal risk scores between risk nodes, and the stability and optimization degree of the graph are evaluated in real time, effectively improving the dynamic adaptability and risk identification accuracy of the enterprise's compliance risk knowledge graph.

[0150] In this embodiment, S7 specifically includes:

[0151] S71. Based on the updated enterprise compliance risk knowledge graph, track the path evolution of risk nodes in the graph, construct a path evolution feature sequence, and generate a risk status evolution sequence according to the path change frequency and abnormal risk score of risk nodes.

[0152] S72. For the risk state evolution sequence, based on the abnormal risk score, a risk state change trend map is constructed using the state transition difference method, and the differentiated trend of risk state change is used as input data for policy training.

[0153] S73. Implement time-series reverse tracing analysis on the risk status change trend map, extract key propagation nodes in the abnormal risk transmission path in real time through reverse tracing analysis, and dynamically determine the strategy sensitivity factors of risk nodes on different transmission paths.

[0154] S74. Based on the risk node's strategy sensitivity factor, the strategy response intensity of the risk propagation node is dynamically adjusted through the coupling relationship between the abnormal risk score and the node location, and the strategy weight is dynamically updated to form a dynamic matrix of node strategy weights.

[0155] S75. Generate the strategy-driven feature representation of risk nodes using the dynamic matrix of node strategy weights, calculate the strategy deviation between the strategy-driven features of risk nodes and the predefined risk rating criteria in real time, and determine the rating range of risk nodes based on the magnitude of the strategy deviation.

[0156] S76. Based on the risk node strategy-driven characteristics and strategy deviation, automatically generate the strategy risk rating of the enterprise compliance risk node using the dynamic floating range method, and output the dynamically updated strategy risk rating results.

[0157] By tracking the evolution of risk nodes in real time, a risk status evolution sequence and differentiated trend map are constructed. Combined with time-series reverse tracing analysis, key propagation nodes and strategy sensitivity factors are dynamically identified, enabling dynamic adjustment of the strategy response intensity at risk propagation nodes. Furthermore, risk rating intervals are dynamically divided based on the node strategy weight matrix and strategy deviation, effectively improving the accuracy, objectivity, and adaptability of enterprise compliance risk rating, and significantly enhancing the reliability and practicality of intelligent compliance auditing in dynamic risk monitoring and decision support.

[0158] Example 1:

[0159] To verify the feasibility of this invention in practice, it was applied to a compliance audit of a large enterprise group. The company's business encompasses multiple areas including manufacturing, finance, supply chain, and customer relationship management, generating a large amount of heterogeneous data daily. This necessitates efficient and accurate compliance auditing to mitigate risks. In this application scenario, traditional auditing methods primarily rely on human experience and static rule bases. Auditors meticulously verify financial accounts, system logs, and business transaction records, typically requiring one to two months for the audit cycle. Furthermore, the massive data volume leads to a high rate of missed detections and incomplete risk identification, particularly when facing cross-domain anomalies, where accuracy drops significantly. Simultaneously, the lack of objective data support often makes audit conclusions difficult to understand and accept quickly, reducing the practicality and decision support effectiveness of the audit results.

[0160] In practice, the system first automatically acquires enterprise ERP system operation logs, CRM system transaction records, financial accounting voucher details, supply chain inventory flow, production batches and process parameters, network security device access logs, user permission change records, email communication data, third-party payment transaction information, and internal collaborative office message records in real time through cross-domain data access interfaces. The system automatically identifies data feature dimensions and standardizes the data format. Next, a semantic adaptive coding method is used to automatically identify and standardize feature dimensions. Specifically, a co-occurrence frequency analysis method based on the same domain data window is adopted to mark high-frequency features as core dimensions. Entropy weight coefficients are dynamically generated by calculating Shannon entropy values. Numerical dimension multi-interval normalization and text dimension semantic mapping are performed according to preset quantiles to achieve accurate encoding of standardized feature data.

[0161] Subsequently, feature vectors from different business domains are extracted using a multi-expert domain structure, and shared features are fused to construct a cross-domain collaborative representation model. Specifically, a scale-interleaved feature method is used to window-divide the data from each domain, and a cross-domain location index table is used to achieve data fusion at different scales. A high-quality prototype vector is generated iteratively using the K-means++ algorithm, and the final fusion optimization matrix is ​​generated using the least squares residual method, completing the construction of the cross-domain collaborative representation model. A feature heatmap is generated from the standardized feature data using the model's built-in hierarchical attention mechanism, and anomaly detection results are generated by identifying abnormal regions based on neighborhood difference thresholds.

[0162] In time series anomaly detection, singular spectrum analysis and empirical pattern decomposition methods are used to obtain attention weights for different periods, targeting both long-term trend characteristics and short-term fluctuation characteristics. This forms long- and short-term correlation features. A dynamic comparative learning framework is then used to adjust cross-period sample pairings in real time, iteratively optimizing differentiated feature representations and outputting anomaly risk scores. Based on the risk identification results, a corporate compliance risk knowledge graph is constructed in real time. An active learning mechanism is used to automatically label and merge newly added unknown anomaly nodes, forming a continuously dynamically updated risk propagation path graph. Furthermore, strategy training tracks changes in risk node paths in real time, adjusting the strategy response intensity based on node strategy sensitivity factors and risk scores. Ultimately, the dynamic risk rating of each risk node in the enterprise is determined, and an easy-to-understand audit report is automatically generated based on a natural language generation model.

[0163] Through practical testing, the traditional manual auditing method and the method of this invention were rigorously compared in terms of audit accuracy, time consumption, and comprehensiveness of risk identification. Detailed results are shown in the table below:

[0164] Table 1 Comparison of Enterprise Compliance Audit Performance Data Table

[0165] Indicator Type Traditional auditing methods Method of the present invention Improvement effect Audit period (days) 35 3 Shortened by 91.4% Risk identification accuracy 72.6％ 96.2％ An increase of 23.6% Abnormal risk missed detection rate 25.3％ 2.9％ Reduced by 88.5% Audit result acceptance rate 67.8％ 95.4％ An increase of 27.6% Report interpretability satisfaction 65.4％ 93.7％ An increase of 28.3%

[0166] As can be seen from the performance comparison data in Table 1 above, the intelligent compliance audit method of this invention has achieved significant improvements in both the accuracy and efficiency of enterprise compliance risk identification. Specifically, the audit cycle has been shortened from 35 days using traditional manual methods to 3 days, a reduction of 91.4%; the risk identification accuracy rate has increased from 72.6% to 96.2%, a significant improvement of 23.6%; simultaneously, the rate of missed detection of abnormal risks has decreased to 2.9%, effectively avoiding underreporting and false reporting; the acceptance rate of audit results has increased by 27.6%, and the interpretability satisfaction of audit reports has increased by 28.3%, demonstrating the high credibility and practicality of the audit results.

[0167] Furthermore, during the three months of actual deployment and operation, the method proposed in this invention detected a total of 287 compliance anomaly risk nodes, of which 46 were cross-domain and cross-scale anomaly nodes that traditional methods failed to effectively identify. The average time for the active learning mechanism to automatically label and integrate newly added risk nodes was only 1.8 minutes, achieving real-time dynamic updates and continuous learning optimization of risk identification. Through strategy training and risk rating, enterprise risk management departments can quickly intervene and rectify high-risk nodes, reducing the average rectification time for risk nodes from 18 days to less than 3 days.

[0168] The above data fully demonstrates that this invention, through precise data feature encoding, efficient cross-domain collaborative models, intelligent anomaly risk detection and dynamic knowledge graph construction, and continuous strategy training and audit decision support mechanisms, effectively solves the practical problems of insufficient data processing capabilities, incomplete risk identification, and lack of objective support for audit results in traditional audit methods. It achieves a leapfrog improvement in the intelligent compliance audit capabilities of enterprises and demonstrates significant advantages in efficient and accurate risk management and control.

[0169] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A data-driven intelligent compliance audit method for enterprises, characterized in that, Includes the following steps: S1. Automatically collect raw audit data from various heterogeneous data sources within the enterprise in real time through cross-domain data access interfaces and output audit datasets in a unified format; S2. Use a semantic adaptive coding method to automatically identify and standardize the feature dimensions of the audit dataset and output standardized feature data; S3. By using a domain-based multi-expert structure, feature extraction is performed on standardized feature data, and shared features are fused to construct a cross-domain collaborative representation model. S4. Use the hierarchical attention mechanism in the cross-domain collaborative representation model to generate a visual feature heatmap from the standardized feature data, extract anomaly localization information through the visual feature heatmap and output the anomaly detection results. S5. Perform dual-scale attention extraction on the standardized feature data, extract long-cycle association pattern features and short-cycle association pattern features, input the extracted pattern features into the contrastive learning framework for representation learning and output anomaly risk scores. S6. Based on the anomaly detection results and anomaly risk scores, construct a dynamically updated enterprise compliance risk knowledge graph and use an active learning mechanism to update and label newly detected anomalies and output the updated risk knowledge graph. S7. Train strategies on abnormal risk scores and updated risk knowledge graphs and output risk rating results; S8. Input the output risk rating results into the natural language generation model to generate an audit report.

2. The data-driven intelligent compliance audit method for enterprises according to claim 1, characterized in that, The original audit data specifically includes: enterprise resource planning system operation logs, customer relationship management system transaction records, financial accounting system vouchers and transaction details, supply chain management system inventory flow data, manufacturing execution system production batches and process parameters, network security device access logs, user authentication and permission change logs, email gateway communication records, third-party payment platform transaction reconciliation data, and internal chat and collaborative office platform message records.

3. The data-driven intelligent compliance audit method for enterprises according to claim 1, characterized in that, S2 specifically includes: S21. All feature dimensions in the original audit dataset are grouped based on the frequency of co-occurrence within the sliding window of the same domain data. Feature dimensions with a co-occurrence frequency higher than the 75th percentile of the data window are marked as core dimension sets, and the remaining dimensions are marked as non-core dimension sets. S22. Calculate the Shannon entropy value within the sliding window for the core dimension set and the non-core dimension set respectively, and normalize the entropy value according to the sum of entropy within the window to generate entropy weight coefficient; S23. For the output entropy weight coefficient, perform multi-interval segmentation normalization on the numerical dimension according to the preset quantile to generate numerical normalized features, and perform semantic mapping on the text dimension based on the cosine similarity of the pre-trained word vectors in the domain terminology dictionary to generate text normalized features. S24. Based on the differences in mean and variance between the current batch and the historical baseline batch, the drift correction bias is calculated using the exponentially weighted moving average algorithm, and the statistical parameters of the current batch are corrected accordingly. S25. Using a greedy subset selection algorithm based on the principle of minimum reconstruction error, the corrected normalized features are sorted by entropy weight coefficients, and incremental reconstruction tests are performed to filter the most sensitive dimensions, outputting a subset of reconstruction sensitive dimensions. S26. Use the ridge regression model with L2 regularization to perform least squares calibration on the feature covariance matrix of the full sample to generate the mapping matrix calibration parameters, and output the calibrated global feature mapping matrix. S27. Combining the entropy weight coefficient, corrected statistical parameters, reconstructed sensitive dimension subset, and calibrated mapping matrix, the normalized features are finally fused to generate standardized feature data with metadata labels.

4. The data-driven intelligent compliance audit method for enterprises according to claim 1, characterized in that, S3 specifically includes: S31. The standardized feature data is segmented in the time series dimension and spatial domain dimension according to the preset sliding window length and overlap rate. The similarity matching between the feature vector in each window and the center feature vector of the previous window is calculated, and elements with similarity higher than the threshold are selected to generate an intra-domain enhanced feature matrix. S32. Based on the cross-domain position index mapping table, feature pairing is performed on the intra-domain enhanced feature matrix at the same position. The paired feature groups are weighted and fused according to the pairing confidence and then normalized to form a cross-domain context resonance matrix. S33. Extract feature blocks from the cross-domain context resonance matrix at both the high-frequency statistical scale and the low-frequency temporal scale, and then, according to the predefined interleaving rules, alternately splice them in the time-frequency dual dimension in the order of temporal block-statistical block-temporal block to generate a multi-granularity interleaved feature representation. S34. Based on multi-granularity interleaved feature representation, the K-means++ algorithm is used to initialize an equal number of prototype vectors and assign them uniform weights. A dynamic sampling strategy is adopted to update the prototype vectors and corresponding feature subsets by decreasing the sampling ratio in each iteration until the sampling error converges. The evolved prototype vector set is then output. S35. The evolved prototype vector set and the multi-granularity interwoven feature representation are concatenated according to the new dimension tensor. Based on the least squares residual minimization criterion, the concatenated tensor is adaptively calibrated to generate the fusion optimization matrix. S36. Perform matrix decomposition on the fusion optimization matrix according to the preset row and column dimensions to extract parameter submatrices, and construct a cross-domain collaborative representation model according to the parameter submatrices.

5. The data-driven intelligent compliance audit method for enterprises according to claim 1, characterized in that, Specifically, S36 includes: S361. Generate the row principal index set IR and the column principal index set IC based on the row principal variance gradient and column principal variance gradient of the fusion optimization matrix. Divide the fusion optimization matrix into sub-matrix groups of unequal size according to IR and IC and output the sub-matrix groups. S362. For each submatrix, construct a joint histogram of adjacent row pairs and calculate row-directed mutual information; construct a joint histogram of adjacent column pairs and calculate column-directed mutual information; filter high-information submatrixes based on a preset mutual information percentile threshold; and perform local spectral domain compression on the filtered submatrixes. S363. Calculate the pairwise normalized Frobenius inner product for the compressed submatrix group as the coupling coefficient, and use the coefficient matrix as input to divide the coupling group using a hierarchical clustering algorithm. Concatenate and splice the submatrix in each group in descending order of coupling coefficient to form a multi-level parameter structure. S364. Perform a two-dimensional discrete Fourier transform on each multi-level parameter structure to extract phase information, use the phase expansion algorithm to eliminate phase jumps and calculate the phase difference between adjacent levels, convert the phase difference into a complex rotation correction factor and apply it to the corresponding level, and output the corrected parameter tensor. S365. Map the corrected parameter tensor to the inter-layer connectivity coefficient tensor of the cross-domain collaborative representation model according to the first dimension, to the attention weight tensor according to the second dimension, and to the feature transformation tensor according to the third dimension. Obtain the parameters of each tensor by solving the linear least squares problem with Tikhonov regularization and construct the cross-domain collaborative representation model.

6. The data-driven intelligent compliance audit method for enterprises according to claim 1, characterized in that, S4 specifically includes: S41. Input standardized feature data into the cross-domain collaborative representation model, perform forward propagation calculation based on the inter-layer connection coefficient tensor, attention weight tensor, and feature transformation tensor, and generate a multi-scale feature vector set within the domain. S42. Using the multi-scale feature vector group within the domain as input, calculate the self-attention weights of each scale feature dimension and perform weighted aggregation within the dimension to obtain the scale attention feature vector within the domain. S43. Calculate the interaction attention weights between different scales by using the cross-scale feature coupling method to obtain the cross-scale coupled attention matrix, where the matrix elements represent the degree of attention contribution between features at a specific scale. S44. Use the cross-scale coupled attention matrix to perform element-wise weighted summation on the intra-domain scale attention feature vector to generate an intra-domain cross-scale fused feature vector. Use the heatmap mapping method to map the fused feature vector values ​​to a predefined color space to form a two-dimensional visualized feature heatmap. S45. Perform abnormal region localization on the visualized feature heatmap based on the neighborhood difference threshold. By calculating the difference between the feature intensity of each location in the heatmap and the neighborhood mean, locations with a difference exceeding the preset threshold are marked as abnormal locations, and a set of abnormal location coordinates is generated. S46. Based on the set of abnormal location coordinates, extract the corresponding original feature vectors from the standardized feature data, calculate the abnormality score using the distance weighting method, and generate the abnormality detection results by arranging them in descending order of the score.

7. The data-driven intelligent compliance audit method for enterprises according to claim 6, characterized in that, Specifically, S43 includes: S431. Perform Hilbert transform on the domain-scale attention feature vector within a preset time window to obtain the instantaneous phase of the feature vector, extract the instantaneous phase distribution curve of each scale feature vector in the time dimension, perform dynamic pairing of scale features based on the correlation coefficient of the phase curve, and output the dynamic pairing set of scale features. S432. For each scale feature pair in the scale feature dynamic pairing set, calculate the dynamic covariance matrix between scale feature vectors, perform eigenvalue decomposition on the covariance matrix and take the largest eigenvalue as the initial value of the interaction weight between scale pairs, and output the set of initial values ​​of scale interaction weight. S433. Based on the initial set of scale interaction weights, and using the scale attention feature vector as a foundation, an adaptive superposition enhancement method for feature vectors in the time dimension is adopted to perform adaptive superposition processing on paired scale feature vectors to generate a scale resonance enhancement feature vector group. S434. Perform a multi-window Fourier transform on the scale resonance enhancement feature vector group in the spectral domain. Determine the sensitivity level of each scale feature by analyzing the intensity of the spectral distribution of the feature vectors between scales. Sort the scale sensitivity levels from high to low according to the proportion of spectral energy. S435. Based on the scale sensitivity ranking results, construct an inter-scale asymmetric interaction relationship matrix, with high-sensitivity scale features as the dominant scale, and assign progressively decreasing weight values ​​to low-sensitivity scale features to form an inter-scale asymmetric interaction attention relationship matrix. S436. Perform Softmax normalization on each row of the asymmetric interaction attention matrix between scales to obtain the attention contribution degree between each scale direction, and finally form a cross-scale coupled attention matrix.

8. The data-driven intelligent compliance audit method for enterprises according to claim 1, characterized in that, S5 specifically includes: S51. The standardized feature data is segmented into long-period sliding windows and short-period sliding windows respectively. During the window movement, the non-linear trend degree of the feature sequence within the window is calculated in real time. The window length is adaptively adjusted by the non-linear trend degree, and the long-period adaptive window data set and the short-period adaptive window data set are output respectively. S52. For a long-period adaptive window dataset, perform singular spectrum analysis on the feature vectors within the window to decompose the trend component and fluctuation component. Determine the trend dominance strength by calculating the ratio of the explained variance of the trend component to the fluctuation component, and use the trend dominance strength as the long-period attention weight. S53. For a short-period adaptive window dataset, perform empirical mode decomposition on the feature vector sequence within each window to determine the energy transfer rate between each mode function within the window, and generate short-period modal attention weights based on the energy transfer rate of the mode functions. S54. Long-cycle attention weight and short-cycle modal attention weight are used to weight and fuse the long-cycle trend component and the short-cycle high-frequency modal function respectively to generate long-cycle correlation pattern feature vector and short-cycle correlation pattern feature vector. S55. Using the feature vectors of long-period correlation patterns and short-period correlation patterns as input, cross-period positive and negative sample pairs are generated through a dynamic pairing mechanism. The dynamic pairing mechanism adaptively determines positive and negative samples based on the pattern similarity of adjacent time windows, adjusts the sample pair selection ratio in real time and calculates the corresponding contrast loss, and iteratively optimizes the contrast loss to obtain cross-period pattern differential feature representation. S56. Calculate the dynamic time-normalized distance between the differential feature representation of the cross-cycle pattern and the differential feature representation of the historical normal pattern. Based on the reciprocal of the distance, generate anomaly risk scores through a nonlinear power exponential transformation function.

9. The data-driven intelligent compliance audit method for enterprises according to claim 1, characterized in that, S6 specifically includes: S61. Based on the abnormal location and corresponding abnormal risk score in the abnormal detection results, dynamically extract the difference fingerprint between the abnormal location features and the features in the historical risk map, use the difference fingerprint to automatically locate the risk nodes and output the set of located risk nodes. S62. For the set of risk nodes after location, based on the similarity of the time series abnormal features of the risk nodes, the nodes are dynamically merged and split, and the granularity of the risk nodes is optimized to generate an optimized set of risk nodes. S63. Using the optimized set of risk nodes as input, determine the directional transmission relationship between risk nodes based on the transmission effect of abnormal risk scores between nodes, determine the edge weights between nodes using the decay rate of abnormal risk scores, and generate a dynamically updated initial enterprise compliance risk knowledge graph. S64. For newly detected abnormal feature data, based on the historical evolution trajectory of existing nodes in the initial enterprise compliance risk knowledge graph, predict the possible evolution trend of abnormal features in real time, identify newly added unknown risk nodes through trend prediction deviation value and perform active labeling, and output the newly added unknown risk nodes after active labeling. S65. Based on the newly added unknown risk nodes after active annotation, dynamically add nodes and update edge weights in the initial enterprise compliance risk knowledge graph. The edge weights are determined according to the interaction strength of abnormal risk scores between the newly added nodes and the identified risk nodes. Output the updated enterprise compliance risk knowledge graph. S66. Encode the node evolution pattern of the updated enterprise compliance risk knowledge graph into a high-dimensional temporal feature space. By dynamically tracking the stability of the risk propagation path between nodes, evaluate the stability of graph nodes and connections in real time, and actively trigger reconstruction of low-stability areas to output a stable and optimized enterprise compliance risk knowledge graph.

10. The data-driven intelligent compliance audit method for enterprises according to claim 1, characterized in that, Specifically, S7 includes: S71. Based on the updated enterprise compliance risk knowledge graph, track the path evolution of risk nodes in the graph, construct a path evolution feature sequence, and generate a risk status evolution sequence according to the path change frequency and abnormal risk score of risk nodes. S72. For the risk state evolution sequence, based on the abnormal risk score, a risk state change trend map is constructed using the state transition difference method, and the differentiated trend of risk state change is used as input data for policy training. S73. Implement time-series reverse tracing analysis on the risk status change trend map, extract key propagation nodes in the abnormal risk transmission path in real time through reverse tracing analysis, and dynamically determine the strategy sensitivity factors of risk nodes on different transmission paths. S74. Based on the risk node's strategy sensitivity factor, the strategy response intensity of the risk propagation node is dynamically adjusted through the coupling relationship between the abnormal risk score and the node location, and the strategy weight is dynamically updated to form a dynamic matrix of node strategy weights. S75. Generate the strategy-driven feature representation of risk nodes using the dynamic matrix of node strategy weights, calculate the strategy deviation between the strategy-driven features of risk nodes and the predefined risk rating criteria in real time, and determine the rating range of risk nodes based on the magnitude of the strategy deviation. S76. Based on the risk node strategy-driven characteristics and strategy deviation, automatically generate the strategy risk rating of the enterprise compliance risk node using the dynamic floating range method, and output the dynamically updated strategy risk rating results.

Images