An ERP system resource scheduling and collaborative optimization method based on big data analysis

Abstract

The application discloses an ERP system resource scheduling and collaborative optimization method based on big data analysis, relates to the technical field of enterprise resource planning and operation scheduling, and comprises the following steps: S1, generating a business correlation topology; S2, generating a business initial cause-effect diagram; S3, based on an improved NOTEARS model, introducing a foraging network growth mechanism of multiple head velvet bubble fungi, based on error gradient driving dynamic evolution and pruning proliferation of a graph structure, and outputting an evolved adjacency matrix; driving structure optimization solution of the improved NOTEARS model by the evolved adjacency matrix; S4, generating an abnormal local cause-effect subdiagram; S5, generating a root cause positioning result; S6, generating a revised scheduling parameter set; and S7, generating a resource scheduling revision scheme. The application overcomes the limitation that a traditional method ignores deep cause-effect dependence and dynamic evolution rules, and provides an efficient solution for ERP system resource scheduling and collaborative optimization.

Description

Technical Field

[0001] This invention relates to the field of enterprise resource planning and operations scheduling technology, and in particular to a method for resource scheduling and collaborative optimization of ERP systems based on big data analysis. Background Technology

[0002] As the scale and depth of collaboration within ERP systems continue to expand, massive amounts of high-dimensional streaming data and complex entity interaction networks are generated in enterprise data processing chains. Traditional resource scheduling methods face severe challenges in accurately locating and responding to dynamic and ever-changing business anomalies and resource conflicts. Existing root cause analysis methods based on topological association or rule matching, while improving anomaly tracing efficiency by utilizing the statistical characteristics and path connectivity of business entities, primarily rely on shallow temporal deviations in streaming data or pre-defined expert rules for association determination. This method, based solely on shallow statistics and static association, ignores the complex conditional dependency structure and deep causal transmission mechanisms implicit in streaming data, leading to misjudgments of causal direction and the introduction of false associations when constructing business association networks, thus limiting the accuracy of anomaly root cause location. Furthermore, classic resource scheduling methods often fail to fully utilize the impact propagation characteristics and dynamic evolution patterns of root cause entities along causal paths to constrain resource allocation space, resulting in resource allocation redundancy or boundary overflow when handling high-frequency sudden anomalies, reducing the collaborative optimization efficiency of system supply and demand matching.

[0003] Therefore, how to provide a resource scheduling and collaborative optimization method for ERP systems based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a resource scheduling and collaborative optimization method for ERP systems based on big data analysis. Through a dynamic evolution optimization mechanism of causal graphs based on the growth mechanism of *Myxomyces cerevisiae* foraging network and an improved NOTEARS model, the edges of the initial causal graph are mapped to myxomycete protoplasmic channels. The residual error gradient is generated by differentiating the reconstruction error value and mapped to the nutrient flow within the channels. The channel flux is calculated by combining causal attraction pressure and distance decay coefficient. The shrinkage and proliferation thresholds are adaptively determined based on the distribution characteristics of the channel flux. Dynamic pruning and proliferation operations are performed on the candidate adjacency weight matrix, outputting an evolved adjacency matrix. The evolved adjacency matrix drives the structural optimization solution of the improved NOTEARS model. The combined acyclic constraints and reconstruction error iterative solution outputs the target directed graph. Furthermore, it identifies mixed paths and intermediate paths, performs demixing and decoupling processing, and generates a business-quantified causal graph. This mechanism effectively eliminates false associations and directional misjudgments during the construction of business-related networks by establishing a dynamic evolutionary path from "error gradient-driven network growth" to "constrained optimization of causal structure," ensuring that the generated business-quantified causal graph accurately reflects the deep causal transmission mechanism in complex pipeline data. Furthermore, it quantifies resource allocation weights based on the impact propagation characteristics of root cause entities along the causal path and combines this with resource consumption quantile constraints to limit allocation, achieving the technical effect of improving the accuracy of root cause localization while enhancing the precision of resource boundary constraints and dynamic allocation. This invention overcomes the limitations of traditional methods that ignore deep causal dependencies and dynamic evolution laws, providing an efficient solution for resource scheduling and collaborative optimization in ERP systems.

[0005] According to an embodiment of the present invention, a resource scheduling and collaborative optimization method for an ERP system based on big data analysis specifically includes: S1. Collect transaction data from the ERP system to extract business entities and their associated paths, and integrate the attribute characteristics of business entities with their associated paths to generate a business association topology. S2. Calculate the conditional mutual information between business entity pairs in the business association topology, strip the conditional independent association paths according to the conditional independence constraint, and generate the initial causal graph of the business. S3. Input the initial business cause-effect graph and ERP system flow data into the improved NOTERAS model, introduce the foraging network growth mechanism of *Hylocereus multicephalomycetes*, drive the dynamic evolution and pruning proliferation of the graph structure based on the error gradient, and output the evolutionary adjacency matrix; use the evolutionary adjacency matrix to drive the structural optimization solution of the improved NOTERAS model, output the target directed graph, and identify hybrid paths and intermediate paths to perform de-hybridization and intermediate decoupling processing, and generate a business quantitative cause-effect graph; S4. Identify abnormal business entities based on the ERP system's transaction data, and extract related sub-networks from the business quantitative cause-effect graph centered on the abnormal business entities to generate abnormal local cause-effect sub-graphs. S5. Calculate the causal correlation between candidate business entities and abnormal business entities in the abnormal local causal subgraph to screen and determine the root cause business entities and generate root cause location results. S6. Extract the abnormal feature status of the root cause business entity based on the root cause localization results, calculate the resource allocation weight and resource availability boundary according to the abnormal feature status, and aggregate to generate a set of corrected scheduling parameters. S7. Load the revised scheduling parameter set, update resource allocation constraints, and perform supply and demand matching and allocation optimization calculations on business entities and resource pools to generate a resource scheduling revision scheme.

[0006] Optionally, S1 specifically includes: S11. Based on the transaction data of the ERP system, extract the multi-dimensional attribute features of business entities and the time-series interaction sequence between business entities, statistically determine the interaction frequency threshold by statistically analyzing the distribution characteristics of the time-series interaction sequence, and filter out business entity pairs with higher interaction frequency thresholds to generate entity association paths. S12. Calculate the feature similarity of business entities based on the multi-dimensional attribute features of business entities, and extract the path topology connectivity based on the associated paths of business entities. Calculate the joint affinity by the feature similarity of associated business entities and the path topology connectivity. S13. Adaptively extract the clustering truncation threshold based on the distribution characteristics of statistical joint affinity, and perform graph segmentation operation based on joint affinity and clustering truncation threshold to generate business association topology.

[0007] Optionally, S2 specifically includes: S21. Based on the business association topology, calculate the local Markov blanket features of business entities to generate conditional dependency sets, calculate the conditional mutual information between business entity pairs based on the conditional dependency sets, and adaptively determine the independence judgment threshold by statistically analyzing the distribution characteristics of the conditional mutual information. S22. Compare the conditional mutual information with the independence judgment threshold, strip away the association paths between business entity pairs that are below the independence judgment threshold, and generate the initial undirected graph of the business. S23. Update the local Markov blanket features based on the initial undirected graph of the business, perform a direction orientation operation on the initial undirected graph of the business based on the updated local Markov blanket features, and output the initial causal graph of the business.

[0008] Optionally, the improved NOTEARS model includes a structural parameter initialization layer, a slime mold foraging network growth layer, and a causal inference layer: The structural parameter initialization layer is used to map business entities in the initial business causal graph to graph nodes and ERP system transaction data to node observation matrices. Based on the topological connection relationship of the initial business causal graph, for node pairs with connections, initial weights are adaptively generated based on the data dependency features of the corresponding node data in the observation matrix. For node pairs without connections, the weights are reset to zero to generate candidate adjacency weight matrices. Matrix reconstruction operations are performed based on the candidate adjacency weight matrices and the node observation matrices to generate reconstruction error values. The slime mold foraging network growth layer is used to introduce the foraging network growth mechanism of *Vorticella multicephala*, and the specific execution process includes: The edges of the current candidate adjacency weight matrix are mapped to slime mold protoplasmic clusters; the residual error gradient is generated by differentiating the current candidate adjacency weight matrix based on the reconstruction error value, and then mapped to the nutrient flow within the channel. For node pairs with non-zero weights, the causal attraction pressure between node pairs is determined based on the correlation features of the corresponding residual error gradients and the statistical dependency features extracted under the multidimensional joint distribution of the source node observation data and the global residual of the target node. The distance decay coefficient is determined based on the topological connectivity features of the current candidate adjacency weight matrix. The pipeline flux is generated by combining the causal attraction pressure and the distance decay coefficient. The current candidate adjacency weight matrix is ​​dynamically updated based on pipeline flux. The shrinkage threshold and proliferation threshold are adaptively determined according to the distribution characteristics of pipeline flux in the current candidate adjacency weight matrix. Pipelines with flux below the shrinkage threshold are subjected to shrinkage pruning. For node pairs with a current weight of zero, potential connection strength is evaluated based on the data dependency characteristics of the corresponding node observation data. Pipeline proliferation is performed on node pairs with potential connection strength higher than the proliferation threshold and which do not form a directed loop after addition. Finally, the dynamically grown adjacency weight matrix is ​​output. The causal inference layer is used to construct an acyclicity penalty term based on the dynamically grown adjacency weight matrix, jointly reconstruct error values ​​to construct a constrained optimization objective and perform iterative solution to output a directed target graph; extract the directed connection paths in the directed target graph, classify them into hybrid paths and intermediate paths based on the convergence and transmission topology features between nodes, and perform dehybridization processing and intermediate decoupling processing respectively, and comprehensively output a business quantification causal graph.

[0009] Optionally, S4 specifically includes: S41. Extract the time-series observation sequence of business entities based on the ERP system's transaction data, statistically determine the distribution characteristics of the time-series observation sequence to adaptively determine the anomaly judgment threshold, and compare the deviation of the time-series observation sequence with the anomaly judgment threshold to identify and generate abnormal business entities. S42. Taking the abnormal business entity as the central node, extract the multi-hop neighbor nodes and directed edges of the central node based on the business quantification cause-effect graph, and splice them to generate the associated sub-network. S43. Extract local observation data of corresponding nodes in the associated subnetwork of the ERP system's transaction data, and merge the associated subnetwork and local observation data to generate an abnormal local causal subgraph.

[0010] Optionally, S5 specifically includes: S51. Based on the directed edges in the abnormal local causal subgraph, a directed causal path is generated by tracing the directed walk sequence from the candidate business entity to the abnormal business entity along the positive direction of the directed edges. S52. Calculate the path span and edge weight product of the directed causal path to generate the path topology connectivity, and extract the Pearson correlation coefficient and mutual information of candidate business entities and abnormal business entities on the corresponding local observation data as data dependency features. S53, the causal correlation between the weighted summation path topology connectivity and the data dependency feature output candidate business entities; S54. Sort the causal correlation in descending order to generate a correlation sequence, calculate the difference gradient of the causal correlation values ​​of adjacent sorted positions in the correlation sequence, locate the cutoff point corresponding to the maximum difference gradient, and adaptively determine the root cause screening threshold. S55. Compare the causal correlation degree with the root cause screening threshold, select candidate business entities that are higher than the root cause screening threshold and determine them as root cause business entities, and extract the directed causal path and causal correlation degree corresponding to the root cause business entity to generate root cause feature pairs. S56. Aggregate root cause business entities and root cause feature pairs to generate root cause location results.

[0011] Optionally, S6 specifically includes: S61. Based on the root cause feature pairs in the root cause localization results, calculate the deviation and volatility of the root cause business entity observation time series to generate the observation time series deviation. S62. Accumulate edge weights and causal correlation along the directed causal path in the root cause feature pair to generate path impact propagation features, and splice observation time series deviation and path impact propagation features to generate abnormal feature states. S63. The abnormal feature states are weighted and mapped according to the resource demand dimension to generate the initial allocation weight. The resource consumption quantile of the historical resource scheduling records under the equidistant binning of the abnormal feature states is calculated to adaptively determine the resource availability boundary. The initial allocation weight is normalized and adjusted, and the allocation upper limit is constrained based on the resource availability boundary to generate the boundary constraint allocation weight. S64. Assign weights to the root cause business entities and boundary constraints to generate a modified scheduling parameter set.

[0012] Optionally, S7 specifically includes: S71. Load and correct the scheduling parameter set, update resource allocation constraints, generate dynamic constraint boundaries, and collect real-time load data of business entities and capacity data of resource pools to construct a supply and demand state matrix. S72. Based on the dynamic constraint boundary filtering supply and demand state matrix, calculate the supply and demand matching degree between business entities and resource pools, statistically determine the distribution characteristics of supply and demand matching degree to adaptively determine the matching screening threshold, and filter entities that meet the matching screening threshold to generate supply and demand matching relationship pairs with resource pools. S73. Integrate the supply and demand matching relationship with the supply and demand state matrix, and perform allocation optimization calculation under dynamic constraint boundary to output the resource scheduling correction scheme.

[0013] The beneficial effects of this invention are: (1) This invention achieves deep decoupling and dynamic reconstruction of pseudo-association stripping and true causal orientation in business association topology by constructing a multi-headed myxomycete foraging network growth mechanism and an improved NOTEARS model causal graph dynamic evolution optimization process. The residual error gradient is used as the nutrient flow within the myxomycete protoplasm channel to drive network growth. The channel flux is calculated by combining causal attraction pressure and distance decay coefficient. Based on the channel flux distribution characteristics, shrinkage and proliferation thresholds are adaptively determined to perform pruning and proliferation operations, outputting a dynamic evolutionary adjacency matrix. The improved NOTEARS model is combined with the evolutionary adjacency matrix to drive structural optimization, and combined with acyclic constraints and reconstruction error iterative calculation, outputting a target directed graph. This mechanism maps sparse business associations to a continuously dynamically growing implicit network space, accurately filtering out conditional independence noise and static topological deviations in entity interaction data, generating a business-quantified causal graph that accurately reflects the deep causal transmission mechanism of complex flow data.

[0014] (2) This invention establishes a causal topology-driven anomaly tracing and resource quantification allocation system by employing directed causal path tracing and differential gradient adaptive truncation. A directed walk sequence is generated by tracing along the positive direction of the directed edges in the business quantification causal graph. The path topology connectivity is generated by calculating the product of the path span and edge weights. Data dependency features are constructed by combining Pearson correlation coefficients and mutual information, and a weighted sum is used to output the causal correlation degree. The differential gradient of adjacent positions in the causal correlation degree sequence is calculated to locate the truncation point, and the root cause screening threshold is adaptively determined to extract root cause feature pairs. Based on the root cause features, the cumulative edge weights and causal correlation degree generate path impact propagation features. Anomaly feature states are generated by splicing observation time-series deviations. Initial allocation weights are generated by weighted mapping according to resource demand dimensions. The allocation upper limit is constrained by historical resource consumption quantiles, and the boundary constraint allocation weights are output. This system achieves accurate cross-space mapping from causal topology propagation to physical resource allocation through causal path impact quantification, differential gradient adaptive bounding, and physical capacity boundary constraints, ensuring that the generated modified scheduling parameter set has extremely high constraint accuracy and response efficiency for dynamic supply and demand imbalances. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a resource scheduling and collaborative optimization method for ERP systems based on big data analysis proposed in this invention; Figure 2 This is a flowchart illustrating the working principle of the improved NOTEARS model, a resource scheduling and collaborative optimization method for ERP systems based on big data analysis proposed in this invention. Detailed Implementation

[0016] The 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.

[0017] refer to Figure 1 and Figure 2 A resource scheduling and collaborative optimization method for ERP systems based on big data analytics, specifically including: S1. Collect transaction data from the ERP system to extract business entities and their associated paths, and integrate the attribute characteristics of business entities with their associated paths to generate a business association topology. S2. Calculate the conditional mutual information between business entity pairs in the business association topology, strip the conditional independent association paths according to the conditional independence constraint, and generate the initial causal graph of the business. S3. Input the initial business cause-effect graph and ERP system flow data into the improved NOTERAS model, introduce the foraging network growth mechanism of *Hylocereus multicephalomycetes*, drive the dynamic evolution and pruning proliferation of the graph structure based on the error gradient, and output the evolutionary adjacency matrix; use the evolutionary adjacency matrix to drive the structural optimization solution of the improved NOTERAS model, output the target directed graph, and identify hybrid paths and intermediate paths to perform de-hybridization and intermediate decoupling processing, and generate a business quantitative cause-effect graph; S4. Identify abnormal business entities based on the ERP system's transaction data, and extract related sub-networks from the business quantitative cause-effect graph centered on the abnormal business entities to generate abnormal local cause-effect sub-graphs. S5. Calculate the causal correlation between candidate business entities and abnormal business entities in the abnormal local causal subgraph to screen and determine the root cause business entities and generate root cause location results. S6. Extract the abnormal feature status of the root cause business entity based on the root cause localization results, calculate the resource allocation weight and resource availability boundary according to the abnormal feature status, and aggregate to generate a set of corrected scheduling parameters. S7. Load the revised scheduling parameter set, update resource allocation constraints, and perform supply and demand matching and allocation optimization calculations on business entities and resource pools to generate a resource scheduling revision scheme.

[0018] In this embodiment, S1 specifically includes: S11. Read the business entity attribute fields and interaction record fields in the ERP system transaction data, extract the multi-dimensional attribute features of the business entities and the time-series interaction sequence between the business entities, calculate the mean and standard deviation of the interaction frequency in the time-series interaction sequence, add 1.5 times the standard deviation to the mean to calculate the interaction frequency threshold, select business entity pairs with interaction frequency higher than the interaction frequency threshold to construct basic association edges, and record the corresponding interaction frequency as the basic edge weight of the basic association edge. Traverse the basic association edges to extract the edge sequence that is connected end to end and splice it to generate the entity association path. S12. Calculate the feature similarity of business entities based on the multi-dimensional attribute features of business entities. Specifically, calculate the cosine similarity of the multi-dimensional attribute feature vectors of two business entities as the feature similarity value. Extract multi-hop topological connectivity based on entity association paths. Specifically, locate the shortest path between two business entities on the entity association path, calculate the arithmetic mean of all basic edge weights on the shortest path and divide it by the global maximum basic edge weight for normalization. Multiply the normalized arithmetic mean by 0.5 raised to the power of the number of edges on the shortest path to generate the multi-hop topological connectivity value. Calculate the product of the feature similarity value and the multi-hop topological connectivity value and divide it by 2 to generate a cross feature scalar. Input the cross feature scalar into the Sigmoid activation function to calculate the nonlinear interaction coefficient. Multiply the nonlinear interaction coefficient by the arithmetic mean of the feature similarity value and the multi-hop topological connectivity value to generate the joint affinity. S13. Calculate the mean and variance of the joint affinity, subtract twice the variance from the mean to generate the clustering truncation threshold, compare the joint affinity between business entity pairs with the clustering truncation threshold, disconnect the low affinity basic association edges with joint affinity lower than the clustering truncation threshold to perform graph segmentation operation, extract the joint affinity corresponding to the remaining basic association edges in the segmented connected subgraph as edge weights and assign them to the remaining basic association edges to generate the business association topology.

[0019] In this embodiment, S2 specifically includes: S21. Read the business association topology. For each pair of business entities with basic association edges, calculate the direct mutual information value between the multi-dimensional attribute feature vectors of the two business entities as the unconditional dependency. Traverse all business entity pairs and combine the calculated unconditional dependency values ​​into a dependency distribution set. Calculate the mean and standard deviation of the dependency distribution set. Add 1.5 times the standard deviation to the mean to generate a dependency truncation threshold of 0.85. Remove weak association edges with unconditional dependency values ​​lower than the dependency truncation threshold of 0.85. Combine the remaining association edges to generate an undirected graph of the business skeleton.

[0020] S22. Read the undirected graph of the business skeleton. For each pair of business entities with directly related edges, extract the set of all neighboring nodes except for the pair of business entities as the candidate set of condition variables. Increase the subset order of the candidate set of condition variables one by one. Calculate the condition mutual information value of the two business entities under the given subset order and extract the minimum condition mutual information value. Set the independence benchmark threshold to 0.10. If the minimum condition mutual information value is lower than the independence benchmark threshold of 0.10, then determine that the business entity is conditionally independent and disconnect the corresponding related edges. After traversal, generate a simplified undirected graph of the business.

[0021] S23. Read the simplified undirected graph of the business and extract the set of neighbor nodes of each business entity as the Markov blanket feature. For each undirected triple consisting of a central entity and two edge entities in the simplified undirected graph of the business, extract the unconditional dependency between the two edge entities. If the unconditional dependency is lower than the independence benchmark threshold of 0.10, calculate the conditional mutual information value of the two edge entities under the condition of only the central entity variable as the collision activation index. If the collision activation index is higher than the independence benchmark threshold of 0.10, determine that the triple is a collision structure. Perform direction orientation operation on the candidate edge of the two edge entities pointing to the central entity respectively. After traversing all undirected triples to complete the direction orientation, output the initial causal graph of the business.

[0022] In this embodiment, the improved NOTEARS model includes a structural parameter initialization layer, a slime mold foraging network growth layer, and a causal inference layer: The structural parameter initialization layer is used to map business entities in the initial causal graph of the business to graph nodes, and to concatenate the ERP system transaction data column by column to map it to a node observation matrix. For node pairs with connections, the absolute value of the Pearson correlation coefficient of the corresponding two columns of data in the node observation matrix is ​​calculated and multiplied by a scaling factor of 1.0 to generate initial weights. For node pairs without connections and the weights of the node's own diagonal, they are directly set to zero. The candidate adjacency weight matrix is ​​then generated by summing the results. The candidate adjacency weight matrix is ​​right-multiplied by the node observation matrix to generate the reconstructed observation matrix. The Frobenius norm distance between the reconstructed observation matrix and the node observation matrix is ​​calculated to generate the reconstruction error value.

[0023] The slime mold foraging network growth layer is used to introduce the foraging network growth mechanism of *Vorticella multicephala*, and the specific execution process includes: Map the non-zero weighted edges in the current candidate adjacency weight matrix to slime mold protoplasmic clusters; calculate the partial derivatives of the candidate adjacency weight matrix based on the reconstruction error value to generate the residual error gradient and map it to the nutrient flow within the channel; For node pairs with non-zero weights, the mutual information value between the source node's observation data and the target node's global residual is calculated as a statistical dependency feature. The product of this statistical dependency feature and the residual error gradient is divided by the sum of the variance of the source node's observation data and 1.0 to generate the causal attraction pressure. The shortest path hop count between two nodes in the candidate adjacency weight matrix is ​​calculated plus the reciprocal of 1.0 as the distance decay coefficient. The causal attraction pressure is multiplied by the distance decay coefficient to generate the pipeline flux.

[0024] Multiply the current pipeline flux by the update rate of 0.8 and add the current candidate adjacency weight matrix by the retention rate of 0.2 to generate the updated candidate adjacency weight matrix; calculate the mean of the non-zero weights in the updated candidate adjacency weight matrix and subtract 1.0 standard deviation to generate the shrinkage threshold, and calculate the mean of the non-zero weights and add 2.0 standard deviation to generate the proliferation threshold; perform a shrinkage pruning operation to reset the weights to zero for edges with pipeline flux below the shrinkage threshold; for node pairs with a current weight of zero, calculate the mutual information value of the observation data of the two nodes as the potential connection strength, assign an initial weight of 0.01 to node pairs with a potential connection strength higher than the proliferation threshold and which do not form a directed loop after addition, perform a pipeline proliferation operation, and output the dynamically grown adjacency weight matrix.

[0025] The causal inference layer calculates the square of the Frobenius norm of the dynamically grown adjacency weight matrix, multiplied by a penalty coefficient of 0.5 to generate an acyclicity penalty term. The reconstruction error value is then added to the acyclicity penalty term to construct a constrained optimization objective function. An Adam optimizer with a learning rate of 0.01 is used to iteratively solve for minimizing the objective function, outputting a target directed graph. Multiple directed connection paths with an in-degree greater than or equal to 2 are extracted from the target directed graph as promiscuous paths. The residual offsets at common child nodes of the promiscuous paths are calculated and subtracted from the target node observation data to perform depromiscation. A continuous directed path consisting of intermediate nodes with an in-degree of 1 and an out-degree of 1 is extracted from the target directed graph as an intermediate path. The indirect transit weight product from the start node to the end node on the intermediate path is calculated, and this indirect transit weight product is subtracted from the direct connection weights from the start node to the end node to perform intermediate decoupling. The resulting comprehensive output is a business-quantified causal graph.

[0026] The improved NOTEARS model proposed in this step is similar to the traditional NOTEARS model in that it is based on structural equation modeling and continuous optimization theory. That is, by transforming the discrete graph structure search into a continuous adjacency matrix optimization problem, the target directed graph is solved iteratively by minimizing the reconstruction error and the acyclicity penalty term, and the subsequent causal effect identification and decoupling process is performed by extracting the directed path.

[0027] The difference lies in that this invention breaks through the limitations of the traditional NOTEARS model, which relies on static continuous relaxation and fixed L1 regularization, leading to network topology rigidity and easy getting trapped in local optima. It adds a slime mold foraging network growth layer to replace the traditional static gradient descent, maps the residual error gradient to the pipeline nutrient flow, and replaces the traditional fixed threshold pruning with the adaptive shrinkage and proliferation operation of the multi-headed velvet mold foraging mechanism. It calculates the pipeline flux based on causal attraction pressure and distance decay coefficient to dynamically remove invalid edges and discover potential causal edges, rather than a single L1 norm forced sparsity constraint.

[0028] The beneficial effects of this improvement are that, through the adaptive dynamic growth and topology reshaping of the multicephalomycetes foraging network, the physical evolution constraints of the graph structure are rigidly implanted into the matrix iterative optimization process. This breaks through the limitations of traditional methods in complex ERP flow data, which suffer from incomplete removal of pseudo-associations and omission of true causality due to topological solidification. It achieves a precise transformation from static sparse constraints to bio-inspired dynamic growth optimization. This design significantly enhances the dynamic discovery capability of implicit causal pathways in high-dimensional sparse business associations. It can accurately prune redundant paths and proliferate potential causal links under the drive of error gradient. Combined with decontamination and mediator decoupling, it effectively improves the topological accuracy of business quantification causal graph reconstruction and the absolute reliability of anomaly root cause tracing.

[0029] In this embodiment, S4 specifically includes: S41. Extract the time-series observation sequence of business entities based on the ERP system's transaction data, calculate the mean and standard deviation of the time-series observation sequence, add 3.0 times the standard deviation to the mean to generate a dynamic upper bound threshold, and subtract 3.0 times the standard deviation from the mean to generate a dynamic lower bound threshold; calculate the absolute difference between the value of each data point in the time-series observation sequence and the mean as the deviation, and determine the data points with a deviation greater than 3.0 times the standard deviation as abnormal data points, and count the proportion of abnormal data points in each business entity to the total length of the time-series observation sequence of that entity. Set the abnormal proportion judgment threshold to 0.05. When the proportion value is greater than the abnormal proportion judgment threshold of 0.05, the business entity is identified as an abnormal business entity.

[0030] S42. Taking the abnormal business entity as the central node, read the business quantification cause-effect graph, extract the nodes directly connected to the central node through directed edges in the business quantification cause-effect graph as one-hop neighbor nodes, extract the nodes directly connected to the one-hop neighbor nodes through directed edges but not the central node as two-hop neighbor nodes, merge the central node, one-hop neighbor nodes and two-hop neighbor nodes to generate a multi-hop neighbor node set, extract all directed edges connecting the nodes within the multi-hop neighbor node set in the business quantification cause-effect graph, and concatenate the multi-hop neighbor node set with all directed edges to generate an associated subnetwork.

[0031] S43. Based on the ERP system's pipeline data, extract the local time-series observation data corresponding to the multi-hop neighborhood node set in the associated sub-network, and concatenate the local time-series observation data by row alignment to generate a local observation matrix; map the directed edges in the associated sub-network to directed edges in the graph, attach the row vector data of the corresponding nodes in the local observation matrix to the graph nodes to generate node feature vectors, and fuse the topology of the associated sub-network and the node feature vectors to generate an abnormal local causal subgraph.

[0032] In this embodiment, S5 specifically includes: S51. Based on the directed edges in the abnormal local causal subgraph, a directed causal path is generated by tracing the directed walk sequence from the candidate business entity to the abnormal business entity along the positive direction of the directed edges.

[0033] S52. Calculate the number of directed edges contained in the directed causal path as the path span, and take the root of the path span by multiplying the weights of all directed edges on the directed causal path to generate the path topology connectivity; extract the absolute value of the Pearson correlation coefficient and the mutual information value of the corresponding row vector data of candidate business entities and abnormal business entities in the local observation matrix, and multiply the absolute value of the Pearson correlation coefficient by the weight coefficient 0.4 and add the mutual information value by the weight coefficient 0.6 to generate data dependency features.

[0034] S53. Multiply the path topology connectivity by a weight coefficient of 0.5, add the data dependency feature by a weight coefficient of 0.5, and sum to generate the causal correlation degree of the candidate business entities.

[0035] S54. Sort the causal correlation in descending order to generate a correlation sequence. Calculate the difference between the causal correlation values ​​of adjacent sorted positions in the correlation sequence as the differential gradient. Extract the maximum value of the differential gradient to locate the cutoff point corresponding to the maximum differential gradient. Determine the causal correlation value of the current sorted position corresponding to the cutoff point in the correlation sequence as the root cause screening threshold.

[0036] S55. Compare the causal correlation degree of candidate business entities with the root cause screening threshold, and select candidate business entities whose causal correlation degree is greater than or equal to the root cause screening threshold as root cause business entities. Extract the directed causal path and causal correlation degree corresponding to the root cause business entity to generate root cause feature pairs.

[0037] S56. Aggregate root cause business entities and root cause feature pairs to generate root cause location results.

[0038] The causal correlation degree adaptive root cause localization process proposed in this step is similar to the traditional root cause localization process based on graph topology or statistical indicators in that it is based on graph structure path tracing and data dependency measurement theory. That is, it calculates topological connectivity by extracting the walking path from candidate nodes to abnormal nodes in the associated network, and evaluates the dependency strength between nodes by using the statistical correlation characteristics of the observed data. Finally, it ranks and locates the root cause entity by combining topological and data features.

[0039] The difference lies in that this invention breaks away from the limitations of traditional methods that rely on manual experience to set fixed thresholds for root cause screening. It adds a differential gradient truncation step, which weights and fuses the path topology connectivity with Pearson and mutual information data dependency features to generate causal correlation. The root cause screening threshold is adaptively determined by the maximum differential gradient of adjacent positions in the correlation sequence. At the same time, the directed causal path corresponding to the root cause entity is extracted and aggregated with the causal correlation to generate root cause feature pairs, rather than a single node sorting list output.

[0040] The beneficial effects of the improvements are that, through multi-dimensional feature fusion and differential gradient adaptive truncation, the natural discontinuity boundary of data distribution is rigidly embedded into the root cause discrimination standard, breaking the limitations of traditional methods that lead to root cause misjudgment or omission due to fixed thresholds. This achieves a precise transformation from qualitative screening based on manual experience to data-driven adaptive quantitative truncation. This design significantly enhances the ability to accurately capture the multi-level attenuation characteristics of abnormal propagation in complex business networks. It can adaptively peel off non-critical related entities in dynamically fluctuating pipeline data. Combined with the extraction of root cause feature pairs, it effectively improves the robustness of root cause localization and provides a high-confidence causal path basis for subsequent quantitative resource scheduling.

[0041] In this embodiment, S6 specifically includes: S61. Based on the root cause feature pairs in the root cause localization results, extract the time series observation sequence of the root cause business entity, calculate the absolute difference between the value of each data point in the time series observation sequence and the mean as the deviation, calculate the square root of the quotient of the sum of squares of the differences between adjacent data points in the time series observation sequence divided by the total number of data points minus 1.0 to generate the volatility, multiply the deviation by the weight coefficient 0.6 and add the volatility multiplied by the weight coefficient 0.4 to generate the observation time series bias.

[0042] S62. Along the directed causal path in the root cause feature pair, extract the weight values ​​of each directed edge on the path, calculate the cumulative weight of all directed edge weights, extract the causal correlation degree in the root cause feature pair, multiply the cumulative edge weight by the weight coefficient 0.5, add the causal correlation degree multiplied by the weight coefficient 0.5, and sum to generate the path impact propagation feature.

[0043] S63. Multiply the observation time series deviation by the resource demand dimension mapping coefficient of 0.8 to generate the initial allocation weight; extract historical resource scheduling records, set the step size of 0.1 along the path impact propagation feature value interval to perform equidistant binning, locate the target bin interval to which the path impact propagation feature belongs, and calculate the 90.0 percentile of the resource consumption value of the historical resource scheduling record in the target bin interval as the resource availability boundary; compare the initial allocation weight with the resource availability boundary, when the initial allocation weight is greater than the resource availability boundary, truncate the initial allocation weight and replace it with the resource availability boundary, when the initial allocation weight is less than or equal to the resource availability boundary, retain the original value of the initial allocation weight to generate the truncated allocation weight, divide the truncated allocation weight by the sum of all truncated allocation weights to perform normalization adjustment to generate the boundary constraint allocation weight.

[0044] S64. Extract the root cause business entity from the root cause localization result, aggregate the root cause business entity and the boundary constraint allocation weight, construct key-value pair combination with the root cause business entity as key and the boundary constraint allocation weight as value, and generate the corrected scheduling parameter set.

[0045] The boundary constraint scheduling parameter generation process proposed in this step is similar to the traditional abnormal resource allocation parameter generation process in that it is based on the theory of abnormal state assessment and resource demand mapping. That is, the fault intensity is quantified by extracting the observed deviation features of abnormal entities and mapping them as the initial weights for resource allocation to guide the subsequent generation of scheduling strategies.

[0046] The difference lies in that this invention breaks away from the limitations of traditional methods that rely solely on the degree of single-point time-series anomalies for unbounded allocation while ignoring the physical capacity limits of the system. It adds the steps of path impact propagation characteristics and equidistant bin boundary constraints, accumulates edge weights and causal correlation along directed causal paths to generate impact propagation characteristics, and splices and maps them with the observed time-series deviation. At the same time, it adaptively determines the resource availability boundary based on historical resource consumption quantiles, and performs normalization and upper limit truncation on the initial allocation weights to generate boundary constraint allocation weights, rather than a single unbounded linear mapping prediction.

[0047] The beneficial effects of the improvements are that, by accumulating causal path impacts and truncating bin quantile boundaries, this invention forcibly embeds the physical constraints of resource capacity and historical consumption patterns into the parameter generation process. This breaks the limitations of traditional methods that are prone to triggering resource overload and starvation deadlocks under abnormal impacts, and achieves a precise conversion from unbounded black-box allocation to boundary-constrained safe allocation. This design significantly enhances the defense against the amplification effect of cross-node propagation of abnormal states, can accurately truncate resource overflow risks in the historical steady-state distribution space, and, combined with causal path feature pairs, effectively improves the dynamic adaptability of scheduling parameters and the absolute security of resource collaborative allocation.

[0048] In this embodiment, S7 specifically includes: S71. Load the modified scheduling parameter set, extract the boundary constraint allocation weights from the modified scheduling parameter set, multiply the boundary constraint allocation weights by the total capacity value of the resource pool to generate dynamic constraint boundaries; collect real-time load data of business entities and capacity data of resource pools, calculate the ratio of real-time load data to capacity data as the load rate, perform cross expansion of the load rate vector of business entities and the capacity vector of resource pools to construct a supply and demand state matrix with the number of rows being the number of business entities and the number of columns being the number of resource pools.

[0049] S72. Extract the load rate corresponding to each row in the supply and demand status matrix and compare it with the dynamic constraint boundary. Remove the business entity rows whose load rate is greater than the dynamic constraint boundary to generate a filter status matrix. Calculate the product of the difference between 1.0 and the load rate and the capacity data as the supply and demand matching degree. Calculate the mean and standard deviation of the supply and demand matching degree in the filter status matrix. Add 2.0 times the standard deviation to the mean to generate the matching screening threshold. Select business entities with a supply and demand matching degree greater than or equal to the matching screening threshold and combine them with the resource pool to generate supply and demand matching relationship pairs.

[0050] S73. Extract the corresponding business entities and resource pools of the supply and demand matching relationship pairs, extract the corresponding matrix elements in the filtering state matrix to construct the matching supply and demand sub-matrix; based on the matching supply and demand sub-matrix, under the dual upper limit constraints of dynamic constraint boundary and resource pool capacity data, with the goal of minimizing the sum of the differences between real-time load data and allocation amount value, calculate the resource amount value allocated to the corresponding resource pool for each business entity, and output the resource scheduling correction scheme.

[0051] Example 1: To verify the feasibility of this invention in ERP system resource scheduling and collaborative optimization, the method of this invention was applied to the intelligent supply chain and production collaboration ERP system of a large manufacturing group (hereinafter referred to as "Group M"). In traditional ERP system resource scheduling mechanisms, anomaly response strategies based on static business rules or shallow temporal deviations are typically adopted. These methods struggle to remove conditionally independent noise and pseudo-correlation paths from massive amounts of flow data, and cannot accurately quantify the impact propagation effect of anomalies along causal links. This easily leads to inaccurate root cause localization and boundary overflow of resource allocation, causing local resource depletion or large-scale idleness. To solve the above problems, Group M decided to adopt the ERP system resource scheduling and collaborative optimization method based on big data analysis proposed in this invention.

[0052] During implementation, Group M first collects flow data of production work orders, material circulation, and equipment operation through the ERP system. It then extracts multi-dimensional attribute features and temporal interaction sequences of business entities, calculates feature similarity and path topology connectivity to generate joint affinity, and performs graph segmentation to construct a business association topology. Next, it calculates the conditional mutual information between business entity pairs, removes conditionally independent association paths based on conditional independence constraints, generates an initial undirected graph of the business, and performs direction orientation operations based on the updated local Markov blanket features, outputting an initial causal graph of the business, which serves as the baseline input data for subsequent dynamic evolution.

[0053] Group M improves the NOTEARS model by mapping the edges of the initial causal graph of the business to slime mold protoplasmic clusters as pipelines. Based on the reconstruction error value, it differentiates the candidate adjacency weight matrix to generate a residual error gradient, which is then mapped to nutrient flow within the pipeline. By combining the multidimensional joint distribution of source node observation data and target node global residuals, statistical dependency features are extracted to determine the causal attraction pressure. The pipeline flux is calculated using the joint distance decay coefficient. Based on the pipeline flux distribution, shrinkage and proliferation thresholds are adaptively determined, and dynamic shrinkage pruning and pipeline proliferation operations are performed on the network, outputting an evolutionary adjacency matrix. Subsequently, the evolutionary adjacency matrix drives the structural optimization solution of the improved NOTEARS model. Iterative solutions using acyclic constraints and reconstruction errors are combined to output the target directed graph. Furthermore, it identifies hybrid paths and intermediate paths, performs de-hybridization and intermediate decoupling processing, and generates a business-quantified causal graph.

[0054] In the core tracing and scheduling phase, this invention identifies abnormal business entities based on ERP system transaction data. Centering on the abnormal entity, it extracts related sub-networks from the business quantitative causal graph to generate an abnormal local causal subgraph. Within the subgraph, it traces directed walk sequences to generate directed causal paths, calculates path topology connectivity and data dependency features to output causal correlation, and adaptively determines root cause screening thresholds by locating truncation points based on the differential gradient of causal correlation, accurately extracting root cause feature pairs. Furthermore, it accumulates edge weights along the directed causal paths in the root cause feature pairs and generates path impact propagation features based on causal correlation, splices observation time-series deviations to generate abnormal feature states, weights and maps them according to resource demand dimensions, and combines historical resource consumption quantile constraints to allocate upper limits, generating boundary constraint allocation weights and aggregating them into a corrected scheduling parameter set. Finally, it loads the corrected scheduling parameter set to update dynamic constraint boundaries, performs supply-demand matching degree screening and allocation optimization calculations on business entities and resource pools, and outputs a resource scheduling correction scheme.

[0055] During implementation, the technical team of Group M discovered that, compared with traditional scheduling methods based on static rules and shallow associations, the method of this invention significantly improves the penetration of root cause localization and the accuracy of dynamic resource allocation under complex business anomalies. This invention effectively achieves closed-loop collaboration from deep causal tracing to precise matching of physical resources by dynamically evolving the vesicular foraging network to remove noise, adaptively truncating the differential gradient to locate the root cause, and quantifying the allocation boundary based on causal path impact propagation characteristics.

[0056] To further verify the actual performance of the method of the present invention, Group M conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1: Table 1. Performance Comparison of Resource Scheduling and Collaborative Optimization in Group M's ERP System

[0057] As shown in Table 1, the performance of resource scheduling and collaborative optimization in the ERP system was comprehensively improved after applying the method of this invention. The accuracy of anomaly root cause localization increased from 74.5% with traditional methods to 95.2%, and the pseudo-related path stripping rate jumped from 38.0% to 89.5%, significantly enhancing the accuracy of deep causal transmission mechanism mining and providing a reliable basis for subsequent scheduling. The resource allocation boundary overflow rate was sharply reduced from 15.2% to 1.3%, and the anomaly impact propagation quantification error decreased from 22.4% to 2.8%, effectively avoiding the risks of physical capacity overflow and quantification distortion. The scheduling response delay was significantly shortened from 300 seconds to 25 seconds, significantly improving the system's timeliness. In addition, the number of local resource depletion events decreased from 8.5 times / month to 0.6 times / month, the comprehensive resource utilization rate increased from 76.0% to 94.5%, and the cost of manual intervention in scheduling decreased from 2 million yuan / year to 850,000 yuan / year, significantly reducing operation and maintenance expenses. Satisfaction with business collaboration and scheduling has also improved significantly, increasing from 81.0% to 97.5%.

[0058] Through the method of this invention, Group M successfully achieved accurate causal tracing and dynamic collaborative optimization of resources for complex business anomalies in the ERP system. This effectively mitigated the risks of resource allocation boundary breaches and supply-demand mismatches, ensured the high resilience of the supply chain and production chain, significantly improved the intelligence and precision of ERP system resource scheduling, significantly reduced the decision-making burden of scheduling personnel, enhanced the system's robustness in responding to sudden anomalies, and provided strong technical support for intelligent resource collaboration in the industrial internet environment.

[0059] 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 resource scheduling and collaborative optimization method for ERP systems based on big data analytics, characterized in that, Includes the following steps: S1. Collect transaction data from the ERP system, extract business entities and their associated paths, and integrate the attribute characteristics of business entities with their associated paths to generate a business association topology. S2. Calculate the conditional mutual information between business entity pairs in the business association topology, strip the conditional independent association paths according to the conditional independence constraint, and generate the initial causal graph of the business. S3. Input the initial business cause-effect graph and ERP system flow data into the improved NOTERAS model, introduce the foraging network growth mechanism of *Hylocereus multicephalomycetes*, drive the dynamic evolution and pruning proliferation of the graph structure based on the error gradient, and output the evolutionary adjacency matrix; use the evolutionary adjacency matrix to drive the structural optimization solution of the improved NOTERAS model, output the target directed graph, and identify hybrid paths and intermediate paths to perform de-hybridization and intermediate decoupling processing, and generate a business quantitative cause-effect graph; S4. Identify abnormal business entities based on ERP system transaction data, and extract related sub-networks from the business quantitative cause-effect graph with the abnormal business entities as the center to generate abnormal local cause-effect sub-graphs. S5. Calculate the causal correlation between candidate business entities and abnormal business entities in the abnormal local causal subgraph to screen and determine the root cause business entities and generate root cause location results. S6. Extract the abnormal feature status of the root cause business entity based on the root cause localization results, calculate the resource allocation weight and resource availability boundary according to the abnormal feature status, and aggregate to generate a set of corrected scheduling parameters. S7. Load the revised scheduling parameter set, update resource allocation constraints, and perform supply and demand matching and allocation optimization calculations on business entities and resource pools to generate a resource scheduling revision scheme.

2. The resource scheduling and collaborative optimization method for an ERP system based on big data analysis according to claim 1, characterized in that, S1 specifically includes: S11. Based on the transaction data of the ERP system, extract the multi-dimensional attribute features of business entities and the time-series interaction sequence between business entities, statistically determine the interaction frequency threshold by statistically analyzing the distribution characteristics of the time-series interaction sequence, and filter out business entity pairs with higher interaction frequency thresholds to generate entity association paths. S12. Calculate the feature similarity of business entities based on the multi-dimensional attribute features of business entities, and extract the path topology connectivity based on the associated paths of business entities. Calculate the joint affinity by the feature similarity of associated business entities and the path topology connectivity. S13. Adaptively extract the clustering truncation threshold based on the distribution characteristics of statistical joint affinity, and perform graph segmentation operation based on joint affinity and clustering truncation threshold to generate business association topology.

3. The resource scheduling and collaborative optimization method for an ERP system based on big data analysis according to claim 1, characterized in that, S2 specifically includes: S21. Based on the business association topology, calculate the local Markov blanket features of business entities to generate conditional dependency sets, calculate the conditional mutual information between business entity pairs based on the conditional dependency sets, and adaptively determine the independence judgment threshold by statistically analyzing the distribution characteristics of the conditional mutual information. S22. Compare the conditional mutual information with the independence judgment threshold, strip away the association paths between business entity pairs that are below the independence judgment threshold, and generate the initial undirected graph of the business. S23. Update the local Markov blanket features based on the initial undirected graph of the business, perform a direction orientation operation on the initial undirected graph of the business based on the updated local Markov blanket features, and output the initial causal graph of the business.

4. The resource scheduling and collaborative optimization method for an ERP system based on big data analysis according to claim 1, characterized in that, The improved NOTEARS model includes a structural parameter initialization layer, a slime mold foraging network growth layer, and a causal inference layer: The structural parameter initialization layer is used to map business entities in the initial causal graph of the business as graph nodes and ERP system flow data as node observation matrices. Based on the topological connectivity of the initial causal graph of the business, for node pairs with connections, initial weights are adaptively generated based on the data dependency features of the corresponding node data in the observation matrix. For node pairs without connections, the weights are reset to zero to generate a candidate adjacency weight matrix. A matrix reconstruction operation is performed based on the candidate adjacency weight matrix and the node observation matrix to generate a reconstruction error value. The slime mold foraging network growth layer is used to introduce the foraging network growth mechanism of *Vorticella multicephala*, and the specific execution process includes: The edges of the current candidate adjacency weight matrix are mapped to slime mold protoplasmic clusters; the residual error gradient is generated by differentiating the current candidate adjacency weight matrix based on the reconstruction error value, and then mapped to the nutrient flow within the channel. For node pairs with non-zero weights, the causal attraction pressure between node pairs is determined based on the correlation features of the corresponding residual error gradients and the statistical dependency features extracted under the multidimensional joint distribution of the source node observation data and the global residual of the target node. The distance decay coefficient is determined based on the topological connectivity features of the current candidate adjacency weight matrix. The pipeline flux is generated by combining the causal attraction pressure and the distance decay coefficient. The current candidate adjacency weight matrix is ​​dynamically updated based on pipeline flux. The shrinkage threshold and proliferation threshold are adaptively determined according to the distribution characteristics of pipeline flux in the current candidate adjacency weight matrix. Pipelines with flux below the shrinkage threshold are subjected to shrinkage pruning. For node pairs with a current weight of zero, potential connection strength is evaluated based on the data dependency characteristics of the corresponding node observation data. Pipeline proliferation is performed on node pairs with potential connection strength higher than the proliferation threshold and which do not form a directed loop after addition. Finally, the dynamically grown adjacency weight matrix is ​​output. The causal inference layer is used to construct an acyclicity penalty term based on the dynamically grown adjacency weight matrix, jointly reconstruct error values ​​to construct a constrained optimization objective and perform iterative solution to output a directed target graph; extract the directed connection paths in the directed target graph, classify them into hybrid paths and intermediate paths based on the convergence and transmission topology features between nodes, and perform dehybridization processing and intermediate decoupling processing respectively, and comprehensively output a business quantification causal graph.

5. The resource scheduling and collaborative optimization method for an ERP system based on big data analysis according to claim 1, characterized in that, S4 specifically includes: S41. Extract the time-series observation sequence of business entities based on the transaction data of the ERP system, statistically determine the distribution characteristics of the time-series observation sequence to adaptively determine the anomaly judgment threshold, and compare the deviation of the time-series observation sequence with the anomaly judgment threshold to identify and generate abnormal business entities. S42. Taking the abnormal business entity as the central node, extract the multi-hop neighbor nodes and directed edges of the central node based on the business quantification cause-effect graph, and splice them to generate the associated sub-network. S43. Extract local observation data of corresponding nodes in the associated subnetwork of the ERP system's transaction data, and merge the associated subnetwork and local observation data to generate an abnormal local causal subgraph.

6. The resource scheduling and collaborative optimization method for an ERP system based on big data analysis according to claim 1, characterized in that, S5 specifically includes: S51. Based on the directed edges in the abnormal local causal subgraph, a directed causal path is generated by tracing the directed walk sequence from the candidate business entity to the abnormal business entity along the positive direction of the directed edges. S52. Calculate the path span and edge weight product of the directed causal path to generate the path topology connectivity, and extract the Pearson correlation coefficient and mutual information of candidate business entities and abnormal business entities on the corresponding local observation data as data dependency features. S53, the causal correlation between the weighted summation path topology connectivity and the data dependency feature output candidate business entities; S54. Sort the causal correlation in descending order to generate a correlation sequence, calculate the difference gradient of the causal correlation values ​​of adjacent sorted positions in the correlation sequence, locate the cutoff point corresponding to the maximum difference gradient, and adaptively determine the root cause screening threshold. S55. Compare the causal correlation degree with the root cause screening threshold, select candidate business entities that are higher than the root cause screening threshold and determine them as root cause business entities, and extract the directed causal path and causal correlation degree corresponding to the root cause business entity to generate root cause feature pairs. S56. Aggregate root cause business entities and root cause feature pairs to generate root cause location results.

7. The resource scheduling and collaborative optimization method for an ERP system based on big data analysis according to claim 1, characterized in that, S6 specifically includes: S61. Based on the root cause feature pairs in the root cause localization results, calculate the deviation and volatility of the root cause business entity observation time series to generate the observation time series deviation. S62. The cumulative edge weights of the directed causal path in the root cause feature pair and the causal correlation degree generate the path impact propagation feature. The observation time series deviation and the path impact propagation feature are spliced ​​together to generate the abnormal feature state. S63. The abnormal feature states are weighted and mapped according to the resource demand dimension to generate the initial allocation weight. The resource consumption quantile of the historical resource scheduling records under the equidistant binning of the abnormal feature states is calculated to adaptively determine the resource availability boundary. The initial allocation weight is normalized and adjusted, and the allocation upper limit is constrained based on the resource availability boundary to generate the boundary constraint allocation weight. S64. Assign weights to the root cause business entities and boundary constraints to generate a modified scheduling parameter set.

8. The resource scheduling and collaborative optimization method for an ERP system based on big data analysis according to claim 1, characterized in that, Specifically, S7 includes: S71. Load and correct the scheduling parameter set, update resource allocation constraints, generate dynamic constraint boundaries, and collect real-time load data of business entities and capacity data of resource pools to construct a supply and demand state matrix. S72. Based on the dynamic constraint boundary filtering supply and demand state matrix, calculate the supply and demand matching degree between business entities and resource pools, statistically determine the distribution characteristics of supply and demand matching degree to adaptively determine the matching screening threshold, and filter entities that meet the matching screening threshold to generate supply and demand matching relationship pairs with resource pools. S73. Integrate the supply and demand matching relationship with the supply and demand state matrix, and perform allocation optimization calculation under dynamic constraint boundary to output the resource scheduling correction scheme.

Images