Enterprise digital intelligent operation supervision system

Abstract

The invention relates to the field of data analysis, in particular to an enterprise digital intelligent operation supervision system. An enterprise digital intelligent operation supervision system comprises an enterprise operation data acquisition module, an enterprise operation graph structure data construction module and an operation data anomaly detection module. According to the method, the enterprise operation data is constructed into the graph structure data, so that the relation among different main bodies in the operation data can be enhanced, and the operation data is subjected to anomaly detection through the operation data anomaly detection model under the assistance of the relation among different entities, so that the automatic supervision of the enterprise operation is realized; and in the process of carrying out anomaly detection on the operation data through the operation data anomaly detection model, the virtual edges are used for strengthening the relation between the nodes of the same type, and the order nodes are strengthened, so that the relation between different subjects among the operation data is further strengthened, and the accuracy of feature extraction of the enterprise operation data is improved.

Description

Technical Field

[0001] The present invention relates to the field of data analysis, and in particular to an enterprise digital intelligent operation and supervision system. Background Art

[0002] In the context of digital transformation, enterprise operations are facing the challenges of surging data volume and increasing complexity. Traditional manual monitoring methods have been unable to adapt to the current rapidly changing environment and are often unable to detect and respond to anomalies in operational data in real time. This limitation may cause enterprises to miss key opportunities in the decision-making process and even cause serious economic losses. Summary of the invention

[0003] The present invention constructs enterprise operation data as graph structured data, which can strengthen the connection between different subjects in the operation data, and performs anomaly detection on the operation data through an operation data anomaly detection model with the assistance of the connection between different entities, thereby realizing automatic supervision of enterprise operations; and in the process of performing anomaly detection on operation data through an operation data anomaly detection model, virtual edges are used to strengthen the connection between nodes of the same type, and order nodes are also strengthened, thereby further strengthening the connection between different subjects in the operation data and improving the accuracy of feature extraction of enterprise operation data.

[0004] An enterprise digital intelligent operation supervision system, comprising:

[0005] The enterprise operation data acquisition module is used to acquire the enterprise operation data within a preset period. The enterprise operation data includes product purchase data, supply data and order data;

[0006] An enterprise operation graph structure data construction module is used to construct enterprise operation graph structure data based on enterprise operation data. The enterprise operation graph structure data includes customer nodes, product nodes, supplier nodes and order nodes, and obtains customer feature vectors, product feature vectors, supplier feature vectors and order feature vectors corresponding to the customer nodes, product nodes, supplier nodes and order nodes. The edges between the customer nodes, product nodes, supplier nodes and order nodes include purchase relationships, supply relationships and association relationships.

[0007] The operation data anomaly detection module is used to send the enterprise operation graph structure data into the operation data anomaly detection model for processing to output the operation data anomaly label. The enterprise operation personnel supervise the enterprise operation according to the operation data anomaly label.

[0008] The operation data anomaly detection model is established based on the graph convolutional neural network, and specifically includes a virtual edge construction layer, an order node reinforcement layer, a hierarchical feature extraction layer, a global feature extraction layer, a feature fusion layer and anomaly detection layer, wherein the virtual edge construction is used to construct virtual edges between customer nodes, product nodes and supplier nodes in the enterprise operation graph structure data to construct the first enhanced enterprise operation graph structure data; the order node reinforcement layer is used to strengthen the order feature vectors corresponding to the order nodes in the first enhanced enterprise operation graph structure data to construct the second enhanced enterprise operation graph structure data; the hierarchical feature extraction layer is used to perform feature extraction operations on different nodes in the second enhanced enterprise operation graph structure data to construct customer hierarchical feature vectors, product hierarchical feature vectors, supplier hierarchical feature vectors and order hierarchical feature vectors; the global feature extraction layer is used to perform feature extraction operations on the second enhanced enterprise operation graph structure data to construct a global operation data feature vector; the feature fusion layer is used to fuse the customer hierarchical feature vectors, product hierarchical feature vectors, supplier hierarchical feature vectors and order hierarchical feature vectors and then splice them with the global operation data feature vector to construct the operation data feature vector; the anomaly detection layer is used to process the operation data feature vector to output anomaly labels for the operation data.

[0009] Preferably, in the operation data anomaly detection module, virtual edges are constructed between customer nodes, product nodes and supplier nodes in the enterprise operation graph structure data through a virtual edge construction layer to construct the first enhanced enterprise operation graph structure data, which specifically includes the following steps:

[0010] For any two customer nodes, obtain the corresponding two customer feature vectors, perform multiplication operations on the customer feature conversion matrix and the two customer feature vectors respectively, and construct the corresponding first customer feature transformation vector H i and the second customer feature transformation vector H j , the customer similarity score A is calculated by the following formula ij =LeakyReLU(a T [H i / / H j ]), where LeakyReLU() is the activation function, a is the customer similarity evaluation matrix, T is the transposition operation, / / is the concatenation operation, and the customer similarity score A is determined. ij Is it higher than the similarity score threshold? If the customer similarity score is A ij Above the similarity score threshold, a virtual edge is constructed between the two customer nodes with a similarity score of A. ij As the weight of the virtual edge;

[0011] For any two product nodes, obtain the corresponding two product feature vectors, multiply the product feature conversion matrix with the two product feature vectors respectively, and construct the corresponding first product feature transformation vector C i and the second product feature transformation vector C j , and the product feature conversion matrix can automatically adjust the feature representation of the product feature vector, and the product similarity score B is calculated by the following formula ij =LeakyReLU(b T [C i / / C j ]), b is the product similarity evaluation matrix, and the product similarity score B is judged ij Is it higher than the similarity score threshold? If the product similarity score is B ij Above the similarity score threshold, a virtual edge is constructed between the two product nodes and the similarity score B is used. ij As the weight of the virtual edge;

[0012] For any two supplier nodes, obtain the corresponding two supplier feature vectors, multiply the supplier feature conversion matrix with the two supplier feature vectors respectively, and construct the corresponding first supplier feature transformation vector G i and the second supplier feature transformation vector G j , and the supplier feature conversion matrix can automatically adjust the feature representation of the supplier feature vector, and the supplier similarity score U is calculated by the following formula ij =LeakyReLU(u T [G i / / G j ]), where u is the supplier similarity evaluation matrix, and the supplier similarity score U is determined ij Is it higher than the similarity score threshold? If the supplier similarity score is U ij Above the similarity score threshold, a virtual edge is constructed between two supplier nodes and the similarity score U ij As the weight of the virtual edge.

[0013] Preferably, in the operation data anomaly detection module, the order feature vector corresponding to the order node in the first enhanced enterprise operation graph structure data is enhanced through the order node enhancement layer to construct the second enhanced enterprise operation graph structure data, which specifically includes the following steps:

[0014] All order feature vectors are spliced ​​from bottom to top in chronological order from front to back to construct the order feature time series matrix. Then the order feature time series matrix is ​​multiplied by the value weight matrix, key weight matrix and query weight matrix to construct the corresponding order feature value matrix V, order feature key matrix K and order feature query matrix Q. Calculate the attention weight matrix ATT = softmax(QK T / (D) 0.5 ), D is the dimension of the order feature vector, and then the order feature value matrix V is multiplied by the weight matrix ATT to construct an enhanced order feature matrix, and each row in the enhanced order feature matrix is ​​an enhanced order feature vector corresponding to the order feature vector, and the enhanced order feature vector replaces the corresponding order feature vector in the first enhanced enterprise operation graph structure data to construct the second enhanced enterprise operation graph structure data.

[0015] Preferably, in the operation data anomaly detection module, a feature extraction operation is performed on different nodes in the second enhanced enterprise operation graph structure data through a hierarchical feature extraction layer to construct a customer hierarchical feature vector, a product hierarchical feature vector, a supplier hierarchical feature vector and an order hierarchical feature vector, specifically including the following steps:

[0016] The customer feature vectors corresponding to all customer nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct a customer node feature matrix, and the customer node adjacency matrix, the customer node feature matrix and the customer relationship weight matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain a customer hierarchical feature vector, wherein the customer node adjacency matrix stores the normalized weights between customer nodes;

[0017] The product feature vectors corresponding to all product nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct a product node feature matrix, and the product node adjacency matrix, the product node feature matrix and the product relationship weight matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain a product hierarchical feature vector, wherein the product node adjacency matrix stores the normalized weights between product nodes;

[0018] The supplier feature vectors corresponding to all supplier nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct a supplier node feature matrix, and the supplier node adjacency matrix, the supplier node feature matrix and the supplier relationship weight matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain a supplier hierarchical feature vector, wherein the supplier node adjacency matrix stores the normalized weights between supplier nodes;

[0019] The order feature vectors corresponding to all order nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct an order node feature matrix. The order node adjacency matrix, the order node feature matrix and the order relationship weight matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain the order hierarchical feature vector, where the order node adjacency matrix stores the normalized weights between order nodes.

[0020] Preferably, in the operation data anomaly detection module, a feature extraction operation is performed on the second enhanced enterprise operation graph structure data through a global feature extraction layer to construct a global operation data feature vector, which specifically includes the following steps:

[0021] All customer nodes, product nodes, supplier nodes and order nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct an operation data feature matrix. The inverse square root of the correspondence matrix of the operation data feature matrix, the operation data node feature matrix, the inverse square root of the correspondence matrix of the operation data feature matrix, and the operation data relationship matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain a global operation data feature vector.

[0022] Preferably, in the operation data anomaly detection module, the customer stratification feature vector, the product stratification feature vector, the supplier stratification feature vector and the order stratification feature vector are fused through the feature fusion layer and then concatenated with the global operation data feature vector to construct the operation data feature vector, which specifically includes the following steps: concatenating the customer stratification feature vector, the product stratification feature vector, the supplier stratification feature vector and the order stratification feature vector according to the channel, performing a convolution operation, and then concatenating the output content with the global operation data feature vector according to the dimension to construct the operation data feature vector.

[0023] Preferably, training the operation data anomaly detection model in the operation data anomaly detection module specifically includes the following steps:

[0024] Obtain several operation data training samples, which include the constructed enterprise operation graph structure data, and annotate the operation data training samples with operation data anomaly labels, form all annotated operation data training samples into an operation data training set, and train the operation data anomaly detection model with the operation data training set, calculate the loss value, and determine whether the loss value is within a preset range. If the loss value is within the preset range, output the trained operation data anomaly detection model; otherwise, continue to train the operation data anomaly detection model with the operation data training set.

[0025] Preferably, the hyperparameters in the operation data anomaly detection model in the operation data anomaly detection module are simulated and designed by a swarm optimization algorithm.

[0026] The present invention has the following advantages:

[0027] The present invention constructs enterprise operation data as graph structured data, which can strengthen the connection between different subjects in the operation data, and performs anomaly detection on the operation data through an operation data anomaly detection model with the assistance of the connection between different entities, thereby realizing automatic supervision of enterprise operations; and in the process of performing anomaly detection on operation data through an operation data anomaly detection model, virtual edges are used to strengthen the connection between nodes of the same type, and order nodes are also strengthened, thereby further strengthening the connection between different subjects in the operation data and improving the accuracy of feature extraction of enterprise operation data. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a structural diagram of the enterprise digital intelligent operation and supervision system adopted in an embodiment of the present invention.

[0029] Figure 2 A schematic diagram of the structure of the operational data anomaly detection model adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0031] An enterprise digital intelligent operation supervision system, such as Figure 1 As shown, including:

[0032] The enterprise operation data acquisition module is used to obtain the enterprise operation data within a preset period. The preset period is set by the enterprise and is generally one month. The enterprise operation data includes product purchase data, supply data and order data. The purchase data refers to the relevant information of customers purchasing enterprise products, the supply data refers to the product information supplied by upstream suppliers, and the order data refers to the sales situation of products, such as order amount, order status and order time;

[0033] An enterprise operation graph structure data construction module is used to construct enterprise operation graph structure data based on enterprise operation data. The enterprise operation graph structure data includes customer nodes, product nodes, supplier nodes and order nodes, wherein the customer nodes correspond to customers who purchase enterprise products in the supply data, the product nodes correspond to enterprise products and products supplied by suppliers, the supplier nodes correspond to upstream suppliers in the supply data, and the order nodes correspond to order data; and obtain customer feature vectors, product feature vectors, supplier feature vectors and order feature vectors corresponding to customer nodes, product nodes, supplier nodes and order nodes, and the customer feature vectors, product feature vectors, supplier feature vectors and order feature vectors are generally constructed from corresponding statistical features. For example, the customer feature vector is generally composed of the customer's purchase frequency (such as the number of purchases per month), total consumption and return rate (the ratio of products returned by customers to the total number of products purchased by customers), etc. The product feature vector is generally composed of product The supplier feature vector is generally composed of the on-time delivery rate (the ratio of the number of orders delivered on time to the total number of orders received by the supplier), the total number of orders and the return rate (the ratio of the number of orders returned by the enterprise to the total number of orders). The order node is generally composed of the order product code (generally one-hot encoding), order amount and order date; the edges between the customer node, product node, supplier node and order node include purchase relationship, supply relationship and association relationship, among which the purchase relationship is a unidirectional edge from the customer node to the product node, which means that the customer has purchased the corresponding product, and the weight of the corresponding edge of the purchase relationship is the purchase quantity; the supply relationship is a unidirectional edge from the supplier to the product node, which means that the supplier provides the corresponding product, and the weight of the corresponding edge of the supply relationship is the supply quantity; the association relationship is an undirected edge between the customer node, product node and supplier node and the order node, and the weight of the corresponding edge of the association relationship is the default value, generally 1;

[0034] The operation data anomaly detection module is used to feed the enterprise operation graph structure data into the operation data anomaly detection model for processing to output operation data anomaly labels. Operation data anomaly labels include abnormal purchase behavior labels, customer churn anomaly labels, supply chain anomaly labels, and no anomaly labels. Enterprise operation personnel can supervise enterprise operations based on operation data anomaly labels.

[0035] See also Figure 2The operation data anomaly detection model is established based on the graph convolutional neural network, which specifically includes a virtual edge construction layer, an order node reinforcement layer, a hierarchical feature extraction layer, a global feature extraction layer, a feature fusion layer and anomaly detection layer. The virtual edge construction is used to construct virtual edges between customer nodes, product nodes and supplier nodes in the enterprise operation graph structure data to construct the first enhanced enterprise operation graph structure data. The virtual edges can reveal the connections between customers, products and suppliers, making the subsequent feature extraction of enterprise operation data more accurate; the order node reinforcement layer is used to strengthen the order feature vectors corresponding to the order nodes in the first enhanced enterprise operation graph structure data to construct the second enhanced enterprise operation graph structure data. In the process of strengthening the order feature vectors, the order data can be further strengthened according to the dependency relationship between all order information. The layered feature extraction layer is used to perform feature extraction operations on different nodes in the second enhanced enterprise operation graph structure data. It should be noted that the feature extraction operation adopts graph convolution operation to construct customer layered feature vectors, product layered feature vectors, supplier layered feature vectors and order layered feature vectors; the global feature extraction layer is used to perform feature extraction operations on the second enhanced enterprise operation graph structure data to construct a global operation data feature vector; the feature fusion layer is used to perform feature fusion on the customer layered feature vector, product layered feature vector, supplier layered feature vector and order layered feature vector and then splice them with the global operation data feature vector to construct an operation data feature vector; the anomaly detection layer is used to process the operation data feature vector to output anomaly labels for the operation data;

[0036] This application constructs the enterprise operation data as graph structured data, which can strengthen the connection between different subjects in the operation data, and detect anomalies in the operation data through the operation data anomaly detection model with the assistance of the connection between different entities, thereby realizing automatic supervision of the enterprise operation; and in the process of detecting anomalies in the operation data through the operation data anomaly detection model, the connection between nodes of the same type is strengthened by virtual edges, and the order nodes are also strengthened, thereby further strengthening the connection between different subjects in the operation data and improving the accuracy of feature extraction of the enterprise operation data.

[0037] In the operation data anomaly detection module, virtual edges are constructed between customer nodes, product nodes, and supplier nodes in the enterprise operation graph structure data through the virtual edge construction layer to construct the first enhanced enterprise operation graph structure data, which specifically includes the following steps:

[0038] For any two customer nodes, obtain the corresponding two customer feature vectors, perform multiplication operations on the customer feature conversion matrix and the two customer feature vectors respectively, and construct the corresponding first customer feature transformation vector Hi and the second customer feature transformation vector H j It should be noted that the customer feature conversion matrix is ​​built into the operation data anomaly detection module and participates in the training of the operation data anomaly detection module. The customer feature conversion matrix can automatically adjust the feature representation of the customer feature vector and calculate the customer similarity score A through the following formula: ij =LeakyReLU(a T [H i / / H j ]), where LeakyReLU() is the activation function, a is the customer similarity evaluation matrix, which also participates in the training of the operation data anomaly detection module, T is the transposition operation, / / is the concatenation operation, and determines the customer similarity score A ij Is it higher than the similarity score threshold? The similarity score threshold is set manually. If the customer's similarity score is A ij Above the similarity score threshold, a virtual edge is constructed between the two customer nodes with a similarity score of A. ij As the weight of the virtual edge;

[0039] For any two product nodes, obtain the corresponding two product feature vectors, multiply the product feature conversion matrix with the two product feature vectors respectively, and construct the corresponding first product feature transformation vector C i and the second product feature transformation vector C j It should be noted that the product feature conversion matrix is ​​built into the operation data anomaly detection module and participates in the training of the operation data anomaly detection module. The product feature conversion matrix can automatically adjust the feature representation of the product feature vector and calculate the product similarity score B through the following formula: ij =LeakyReLU(b T [C i / / C j ]), b is the product similarity evaluation matrix, which also participates in the training of the operation data anomaly detection module to determine the product similarity score B ij Is it higher than the similarity score threshold? If the product similarity score is B ij Above the similarity score threshold, a virtual edge is constructed between the two product nodes and the similarity score B is used. ij As the weight of the virtual edge;

[0040] For any two supplier nodes, obtain the corresponding two supplier feature vectors, multiply the supplier feature conversion matrix with the two supplier feature vectors respectively, and construct the corresponding first supplier feature transformation vector G i and the second supplier feature transformation vector G jIt should be noted that the supplier feature conversion matrix is ​​built into the operation data anomaly detection module and participates in the training of the operation data anomaly detection module. The supplier feature conversion matrix can automatically adjust the feature representation of the supplier feature vector and calculate the supplier similarity score U by the following formula: ij =LeakyReLU(u T [G i / / G j ]), where u is the supplier similarity evaluation matrix, which also participates in the training of the operation data anomaly detection module to determine the supplier similarity score U ij Is it higher than the similarity score threshold? If the supplier similarity score is U ij Above the similarity score threshold, a virtual edge is constructed between two supplier nodes and the similarity score U ij As the weight of the virtual edge;

[0041] In the operation data anomaly detection module, the order feature vector corresponding to the order node in the first enhanced enterprise operation graph structure data is enhanced through the order node enhancement layer to construct the second enhanced enterprise operation graph structure data, which specifically includes the following steps:

[0042] All order feature vectors are spliced ​​from bottom to top in chronological order from front to back to construct the order feature time series matrix. Then the order feature time series matrix is ​​multiplied by the value weight matrix, key weight matrix and query weight matrix to construct the corresponding order feature value matrix V, order feature key matrix K and order feature query matrix Q. Calculate the attention weight matrix ATT = softmax(QK T / (D) 0.5 ), D is the dimension of the order feature vector, and then the order feature value matrix V is multiplied by the weight matrix ATT to construct an enhanced order feature matrix, and each line in the enhanced order feature matrix is ​​an enhanced order feature vector corresponding to the order feature vector, and the enhanced order feature vector replaces the corresponding order feature vector in the first enhanced enterprise operation graph structure data to construct the second enhanced enterprise operation graph structure data; it should be noted that the value weight matrix, the key weight matrix and the query weight matrix are set based on the self-attention mechanism in the Transformer model, and also participate in the training of the operation data anomaly detection module;

[0043] In the operation data anomaly detection module, the hierarchical feature extraction layer performs feature extraction operations on different nodes in the second enhanced enterprise operation graph structure data to construct customer hierarchical feature vectors, product hierarchical feature vectors, supplier hierarchical feature vectors, and order hierarchical feature vectors, specifically including the following steps:

[0044] The customer feature vectors corresponding to all customer nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct a customer node feature matrix. The customer node adjacency matrix, the customer node feature matrix and the customer relationship weight matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain a customer hierarchical feature vector, wherein the customer node adjacency matrix stores the normalized weights between customer nodes. These weights before normalization are the weights corresponding to the edges between customer nodes. In particular, there is no connection between customer nodes, and the default weight is 0. The customer relationship weight matrix also participates in the training of the operation data anomaly detection module.

[0045] The product feature vectors corresponding to all product nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct a product node feature matrix. The product node adjacency matrix, the product node feature matrix and the product relationship weight matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain a product hierarchical feature vector. The product node adjacency matrix stores the normalized weights between product nodes. These weights before normalization are the weights corresponding to the edges between product nodes. In particular, there is no connection between product nodes, and the default weight is 0. The product relationship weight matrix also participates in the training of the operation data anomaly detection module.

[0046] The supplier feature vectors corresponding to all supplier nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct a supplier node feature matrix. The supplier node adjacency matrix, the supplier node feature matrix and the supplier relationship weight matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain the supplier hierarchical feature vector, where the supplier node adjacency matrix stores the normalized weights between supplier nodes. These weights before normalization are the weights corresponding to the edges between supplier nodes. In particular, there is no connection between supplier nodes, and the default weight is 0. The supplier relationship weight matrix also participates in the training of the operation data anomaly detection module.

[0047] The order feature vectors corresponding to all order nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct an order node feature matrix. The order node adjacency matrix, the order node feature matrix and the order relationship weight matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain the order hierarchical feature vector, where the order node adjacency matrix stores the normalized weights between order nodes. These weights before normalization are the weights corresponding to the edges between order nodes. In particular, there is no connection between order nodes, and the default weight is 0. The order relationship weight matrix also participates in the training of the operation data anomaly detection module;

[0048] In the operation data anomaly detection module, the feature extraction operation is performed on the second enhanced enterprise operation graph structure data through the global feature extraction layer to construct a global operation data feature vector, which specifically includes the following steps:

[0049] All customer nodes, product nodes, supplier nodes and order nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct an operation data feature matrix. The inverse square root of the correspondence matrix of the operation data feature matrix, the operation data node feature matrix, the inverse square root of the correspondence matrix of the operation data feature matrix, and the operation data relationship matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain a global operation data feature vector.

[0050] In the operation data anomaly detection module, the customer stratification feature vector, product stratification feature vector, supplier stratification feature vector and order stratification feature vector are fused through the feature fusion layer and then concatenated with the global operation data feature vector to construct the operation data feature vector, which specifically includes the following steps: concatenating the customer stratification feature vector, product stratification feature vector, supplier stratification feature vector and order stratification feature vector according to the channel, performing a convolution operation, and then concatenating the output content with the global operation data feature vector according to the dimension to construct the operation data feature vector.

[0051] Training the operational data anomaly detection model in the operational data anomaly detection module includes the following steps:

[0052] A number of operation data training samples are obtained. The operation data training samples include constructed enterprise operation graph structure data. It should be noted that these operation data training samples are obtained based on actual enterprise operation data and constructed as graph structure data, and the operation data training samples are labeled with operation data anomaly labels. These operation data anomaly labels are set by experts according to the enterprise operation situation. All labeled operation data training samples are combined into an operation data training set, and the operation data anomaly detection model is trained with the operation data training set to calculate the loss value. The loss value is calculated using a cross entropy loss function to determine whether the loss value is within a preset range. The preset range is set by the developer. If the loss value is within the preset range, the trained operation data anomaly detection model is output; otherwise, the operation data anomaly detection model is continuously trained with the operation data training set; and the hyperparameters in the operation data anomaly detection model are simulated and designed using a swarm optimization algorithm.

[0053] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. An enterprise digital intelligent operation supervision system, characterized in that: include: The enterprise operation data acquisition module is used to acquire the enterprise operation data within a preset period. The enterprise operation data includes product purchase data, supply data and order data; An enterprise operation graph structure data construction module is used to construct enterprise operation graph structure data based on enterprise operation data. The enterprise operation graph structure data includes customer nodes, product nodes, supplier nodes and order nodes, and obtains customer feature vectors, product feature vectors, supplier feature vectors and order feature vectors corresponding to the customer nodes, product nodes, supplier nodes and order nodes. The edges between the customer nodes, product nodes, supplier nodes and order nodes include purchase relationships, supply relationships and association relationships. The operation data anomaly detection module is used to send the enterprise operation graph structure data into the operation data anomaly detection model for processing to output the operation data anomaly label. The enterprise operation personnel supervise the enterprise operation according to the operation data anomaly label. The operation data anomaly detection model is established based on the graph convolutional neural network, which specifically includes a virtual edge construction layer, an order node reinforcement layer, a hierarchical feature extraction layer, a global feature extraction layer, a feature fusion layer and anomaly detection layer, wherein the virtual edge construction is used to construct virtual edges between customer nodes, product nodes and supplier nodes in the enterprise operation graph structure data, and construct the first reinforced enterprise operation graph structure data; The order node strengthening layer is used to strengthen the order feature vector corresponding to the order node in the first strengthened enterprise operation graph structure data to construct the second strengthened enterprise operation graph structure data; The hierarchical feature extraction layer is used to perform feature extraction operations on different nodes in the second enhanced enterprise operation graph structure data to construct a customer hierarchical feature vector, a product hierarchical feature vector, a supplier hierarchical feature vector and an order hierarchical feature vector; The global feature extraction layer is used to perform feature extraction operations on the second enhanced enterprise operation graph structure data to construct a global operation data feature vector; the feature fusion layer is used to fuse the customer layered feature vector, product layered feature vector, supplier layered feature vector and order layered feature vector and then concatenate them with the global operation data feature vector to construct an operation data feature vector; the anomaly detection layer is used to process the operation data feature vector to output anomaly labels for the operation data.

2. According to claim 1, the enterprise digital intelligent operation supervision system is characterized in that: In the operation data anomaly detection module, virtual edges are constructed between customer nodes, product nodes, and supplier nodes in the enterprise operation graph structure data through the virtual edge construction layer to construct the first enhanced enterprise operation graph structure data, which specifically includes the following steps: For any two customer nodes, obtain the corresponding two customer feature vectors, perform multiplication operations on the customer feature conversion matrix and the two customer feature vectors respectively, and construct the corresponding first customer feature transformation vector H i and the second customer feature transformation vector H j , the customer similarity score A is calculated by the following formula ij =LeakyReLU(a T [H i / / H j ]), where LeakyReLU() is the activation function, a is the customer similarity evaluation matrix, T is the transposition operation, / / is the concatenation operation, and the customer similarity score A is determined. ij Is it higher than the similarity score threshold? If the customer similarity score is A ij Above the similarity score threshold, a virtual edge is constructed between the two customer nodes with a similarity score of A. ij As the weight of the virtual edge; For any two product nodes, obtain the corresponding two product feature vectors, multiply the product feature conversion matrix with the two product feature vectors respectively, and construct the corresponding first product feature transformation vector C i and the second product feature transformation vector C j , and the product feature conversion matrix can automatically adjust the feature representation of the product feature vector, and the product similarity score B is calculated by the following formula ij =LeakyReLU(b T [C i / / C j ]), b is the product similarity evaluation matrix, and the product similarity score B is judged ij Is it higher than the similarity score threshold? If the product similarity score is B ij Above the similarity score threshold, a virtual edge is constructed between the two product nodes and the similarity score B is used. ij As the weight of the virtual edge; For any two supplier nodes, obtain the corresponding two supplier feature vectors, multiply the supplier feature conversion matrix with the two supplier feature vectors respectively, and construct the corresponding first supplier feature transformation vector G i and the second supplier feature transformation vector G j , and the supplier feature conversion matrix can automatically adjust the feature representation of the supplier feature vector, and the supplier similarity score U is calculated by the following formula ij =LeakyReLU(u T [G i / / G j ]), where u is the supplier similarity evaluation matrix, and the supplier similarity score U is determined ij Is it higher than the similarity score threshold? If the supplier similarity score is U ij Above the similarity score threshold, a virtual edge is constructed between two supplier nodes and the similarity score U ij As the weight of the virtual edge.

3. According to claim 2, the enterprise digital intelligent operation supervision system is characterized in that: In the operation data anomaly detection module, the order feature vector corresponding to the order node in the first enhanced enterprise operation graph structure data is enhanced through the order node enhancement layer to construct the second enhanced enterprise operation graph structure data, which specifically includes the following steps: All order feature vectors are spliced ​​from bottom to top in chronological order from front to back to construct the order feature time series matrix. Then the order feature time series matrix is ​​multiplied by the value weight matrix, key weight matrix and query weight matrix to construct the corresponding order feature value matrix V, order feature key matrix K and order feature query matrix Q. Calculate the attention weight matrix ATT = softmax(QK T / (D) 0.5 ), D is the dimension of the order feature vector, and then the order feature value matrix V is multiplied by the weight matrix ATT to construct an enhanced order feature matrix, and each row in the enhanced order feature matrix is ​​an enhanced order feature vector corresponding to the order feature vector, and the enhanced order feature vector replaces the corresponding order feature vector in the first enhanced enterprise operation graph structure data to construct the second enhanced enterprise operation graph structure data.

4. According to claim 3, the enterprise digital intelligent operation supervision system is characterized in that: In the operation data anomaly detection module, the hierarchical feature extraction layer performs feature extraction operations on different nodes in the second enhanced enterprise operation graph structure data to construct customer hierarchical feature vectors, product hierarchical feature vectors, supplier hierarchical feature vectors, and order hierarchical feature vectors, specifically including the following steps: The customer feature vectors corresponding to all customer nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct a customer node feature matrix, and the customer node adjacency matrix, the customer node feature matrix and the customer relationship weight matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain a customer hierarchical feature vector, wherein the customer node adjacency matrix stores the normalized weights between customer nodes; The product feature vectors corresponding to all product nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct a product node feature matrix, and the product node adjacency matrix, the product node feature matrix and the product relationship weight matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain a product hierarchical feature vector, wherein the product node adjacency matrix stores the normalized weights between product nodes; The supplier feature vectors corresponding to all supplier nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct a supplier node feature matrix, and the supplier node adjacency matrix, the supplier node feature matrix and the supplier relationship weight matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain a supplier hierarchical feature vector, wherein the supplier node adjacency matrix stores the normalized weights between supplier nodes; The order feature vectors corresponding to all order nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct an order node feature matrix. The order node adjacency matrix, the order node feature matrix and the order relationship weight matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain the order hierarchical feature vector, where the order node adjacency matrix stores the normalized weights between order nodes.

5. According to claim 4, the enterprise digital intelligent operation supervision system is characterized in that: In the operation data anomaly detection module, the feature extraction operation is performed on the second enhanced enterprise operation graph structure data through the global feature extraction layer to construct a global operation data feature vector, which specifically includes the following steps: All customer nodes, product nodes, supplier nodes and order nodes in the second enhanced enterprise operation graph structure data are spliced ​​from top to bottom to construct an operation data feature matrix. The inverse square root of the correspondence matrix of the operation data feature matrix, the operation data node feature matrix, the inverse square root of the correspondence matrix of the operation data feature matrix, and the operation data relationship matrix are multiplied in sequence and then a nonlinear activation operation is performed to obtain a global operation data feature vector.

6. According to claim 5, the enterprise digital intelligent operation supervision system is characterized in that: In the operation data anomaly detection module, the customer stratification feature vector, product stratification feature vector, supplier stratification feature vector and order stratification feature vector are fused through the feature fusion layer and then concatenated with the global operation data feature vector to construct the operation data feature vector, which specifically includes the following steps: concatenating the customer stratification feature vector, product stratification feature vector, supplier stratification feature vector and order stratification feature vector according to the channel, performing a convolution operation, and then concatenating the output content with the global operation data feature vector according to the dimension to construct the operation data feature vector.

7. According to claim 6, the enterprise digital intelligent operation supervision system is characterized in that: Training the operational data anomaly detection model in the operational data anomaly detection module includes the following steps: Obtain several operation data training samples, which include the constructed enterprise operation graph structure data, and annotate the operation data training samples with operation data anomaly labels, form all annotated operation data training samples into an operation data training set, and train the operation data anomaly detection model with the operation data training set, calculate the loss value, and determine whether the loss value is within a preset range. If the loss value is within the preset range, output the trained operation data anomaly detection model; otherwise, continue to train the operation data anomaly detection model with the operation data training set.

8. The enterprise digital intelligent operation supervision system according to claim 7 is characterized in that: The hyperparameters in the operational data anomaly detection model in the operational data anomaly detection module are simulated and designed through a swarm optimization algorithm.

Images