A method and system for detecting problems of cataloging multiple equipment with the same name

Abstract

The present application discloses a method, system, device, medium and product for detecting the problem of multiple equipment cataloguing with the same name, which relates to the field of intelligent supervision. The method includes: using a data subset of a reference equipment and an equipment to be detected as a reference entity and a detection entity, and constructing first and second graph structure data; using a graph attention network to obtain the first and second entity embeddings based on the first and second graph structure data, and then generating an adaptive code; determining the distance between the adaptive code and the first entity embedding to form a residual vector; using a convolutional neural network to convert the residual vector into a signature vector; inputting the signature vector into a high-speed network to generate a result probability value of attribute value matching; normalizing the result probability value of attribute value matching through an activation function, generating a product matching label, and determining whether there is a problem of multiple equipment cataloguing with the same name based on this label. The present application can automatically and intelligently solve the problem of "multiple products with one name" in cataloguing.

Description

Technical Field

[0001] The present application relates to the field of intelligent supervision, and in particular to a method, system, device, medium and product for detecting problems in which multiple equipment are catalogued with the same name. Background Art

[0002] Equipment cataloging is a basic and pre-conditioning link in the construction and management of equipment data resources. Cataloging data is the core key data for realizing equipment data governance and application. Its data quality is directly related to the credibility of equipment data resources and the effectiveness of application. The problem of "multiple products with one name" is relatively common in the construction of cataloging data for large and complex equipment. The main reason is that the system structure and scale of the catalog of such equipment are large. Experts use the benchmark name as the basis and combine the equipment attribute data to assign the variety name to a specific equipment variety. However, due to the complex composition structure of large and complex equipment, the field segmentation precision is high. If the communication and collaboration between cross-domain experts is low, different equipment varieties will use the same variety name, which will lead to the problem of "multiple products with one name".

[0003] At present, the detection methods for the problem of "multiple items with one name" in cataloging are mainly based on manual methods of expert collaborative discussion, which is more practical for situations with simple equipment system structures, but not applicable for large and complex equipment cataloging scenarios. After literature search, there are no patents or papers related to the detection of the problem of "multiple items with one name" in cataloging large and complex equipment. Summary of the invention

[0004] The purpose of this application is to provide a method, system, device, medium and product for detecting the problem of cataloging multiple equipment with the same name, which can automatically and intelligently solve the problem of cataloging "multiple products with one name".

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides a method for detecting problems in which multiple equipment are catalogued with the same name, comprising:

[0007] Using the set variety name of the equipment as a keyword, the equipment product data is searched to obtain a data subset of the reference equipment and a data subset of the equipment to be detected; the data subset of the reference equipment and the data subset of the equipment to be detected include multiple attributes, each attribute includes multiple attribute values; the attributes include product name, variety name, natural attributes and application attributes; the self-weight of the detection entity is the weight between nodes in the second graph structure data;

[0008] Taking the data subset of the reference equipment as a reference entity, taking the data subset of the equipment to be detected as a detection entity, taking the attributes or attribute values ​​contained in the reference entity and the detection entity as nodes, determining the weights between the nodes, and expressing the weights between the nodes as associations between the nodes, to obtain first graph structure data and second graph structure data;

[0009] Using a graph attention network to obtain a first node feature vector and a second node feature vector based on the first graph structure data and the second graph structure data;

[0010] Obtaining a first entity embedding and a second entity embedding based on the first node feature vector and the second node feature vector;

[0011] generating an adaptive encoding based on the first entity embedding and the second entity embedding;

[0012] determining a distance between the adaptive encoding and the first entity embedding to form a residual vector;

[0013] Using a convolutional neural network to convert the residual vector into a signature vector;

[0014] Taking the signature vector as input, generating a result probability value of attribute value matching through a high-speed network;

[0015] Normalizing the result probability value of the attribute value matching through an activation function to generate a product matching label;

[0016] Based on the product matching tags, it is determined whether there is a problem of multiple equipment being cataloged with the same name.

[0017] Optionally, using the attributes or attribute values ​​included in the reference entity and the detection entity as nodes, determining weights between nodes includes:

[0018] When the attributes contained in the reference entity and the detection entity are used as nodes, the formula is adopted Determine the weights between nodes;

[0019] When the attributes contained in the reference entity and the detection entity are used as the first node and the attribute value is used as the second node, the formula is adopted determining a weight between the first node and the second node;

[0020] When the attribute values ​​contained in the reference entity and the detection entity are used as nodes, the formula is adopted Determine the weights between nodes;

[0021] In the formula, represents the weight between nodes, Indicates the reverse file frequency, Indicates the frequency of keyword occurrence. Indicates the weights of different attribute value categories; the keyword is a set representation related to the equipment attribute value; different attribute value categories include at least text type and number type.

[0022] Optionally, generating an adaptive encoding based on the first entity embedding and the second entity embedding includes:

[0023] An adaptive weight strategy is adopted to generate an adaptive code based on the first entity embedding and the second entity embedding.

[0024] Optionally, using an adaptive weight strategy to generate an adaptive encoding based on the first entity embedding and the second entity embedding includes:

[0025] Obtaining a weight vector between the first node feature vector and the second node feature vector based on the first entity embedding and the second entity embedding;

[0026] Obtaining a weight encoding of the reference entity based on the weight vector and the second entity embedding;

[0027] The adaptive coding is obtained based on the weight coding of the reference entity and the first entity embedding by adopting an adaptive weight strategy.

[0028] Optionally, determining a distance between the adaptive encoding and the first entity embedding to form a residual vector comprises:

[0029] Based on the Euclidean distance calculation concept, the distance between the adaptive coding and the first entity embedding is determined to form a residual vector.

[0030] Optionally, determining whether there is a problem of multiple equipment cataloguing with the same name based on the product matching tag includes:

[0031] When the product matching tag is 1, it means that the reference equipment and the equipment to be tested are the same product, and there is a problem of multiple equipment being cataloged with the same name;

[0032] When the product matching tag is 0, it indicates that the reference equipment and the equipment to be detected are different products, and there is no problem of cataloging multiple equipment with the same name.

[0033] In a second aspect, the present application provides a system for detecting problems in which multiple equipment are catalogued with the same name, including:

[0034] A graph construction layer is used to retrieve equipment product data using the set variety name of the equipment as a keyword, obtain a data subset of the reference equipment and a data subset of the equipment to be detected, use the data subset of the reference equipment as a reference entity, use the data subset of the equipment to be detected as a detection entity, use the attributes or attribute values ​​contained in the reference entity and the detection entity as nodes, determine the weights between nodes, and express the weights between nodes as associations between nodes, to obtain first graph structure data and second graph structure data; the data subset of the reference equipment and the data subset of the equipment to be detected include multiple attributes, each attribute includes multiple attribute values; the attributes include product name, variety name, natural attributes and application attributes; the self-weight of the detection entity is the weight between nodes in the second graph structure data;

[0035] A graph attention layer, connected to the graph construction layer, for obtaining a first node feature vector and a second node feature vector based on the first graph structure data and the second graph structure data using a graph attention network, and obtaining a first entity embedding and a second entity embedding based on the first node feature vector and the second node feature vector;

[0036] an attribute value comparison layer, connected to the graph attention layer, for generating an adaptive code based on the first entity embedding and the second entity embedding, determining a distance between the adaptive code and the first entity embedding, and forming a residual vector;

[0037] A feature aggregation layer, connected to the attribute value comparison layer, for converting the residual vector into a signature vector using a convolutional neural network;

[0038] The matching prediction layer is connected to the feature aggregation layer, and is used to take the signature vector as input, generate a result probability value of the attribute value matching through a high-speed network, normalize the result probability value of the attribute value matching through an activation function, generate a product matching label, and determine whether there is a problem of cataloging multiple equipment with the same name based on the product matching label.

[0039] In a third aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned method for detecting problems in cataloging multiple equipment with the same name.

[0040] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned method for detecting problems in cataloging multiple equipment with the same name.

[0041] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the method for detecting problems in cataloging multiple equipment with the same name.

[0042] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0043] The present application provides a method, system, device, medium and product for detecting the problem of multiple equipment catalogs with the same name. It uses a graph attention network and a neural network to perform matching analysis on the equipment catalog attribute data text, and can realize intelligent detection of the "multiple items with one name" problem in the cataloging of large and complex equipment, reduce manual participation, improve the objectivity and accuracy of the problem detection results, and reduce the investment cost of manpower, time and other resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 This is an application environment diagram of a method for detecting a problem of cataloguing multiple equipment with the same name in one embodiment of the present application;

[0046] Figure 2 A flowchart of a method for detecting a problem of cataloging multiple equipment with the same name provided in an embodiment of the present application;

[0047] Figure 3 An implementation architecture diagram of a system for detecting problems in which multiple equipment are catalogued with the same name provided in one embodiment of the present application;

[0048] Figure 4 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0050] In the relevant technical field, in the intelligent classification cataloging system based on Radio Frequency Identification (RFID), the cataloging module is based on the book classification results, combined with the data information collected by RFID to intelligently generate the information catalog and compile the corresponding book classification number; the intelligent media cataloging technology based on the convolutional neural network model uses convolutional neural network (CNN) and other algorithm technologies to use the same set of algorithm models in the feature extraction and intelligent retrieval stages to achieve intelligent labeling of media assets. However, the above method can only be transferred to the generation stage of equipment cataloging alternatives, and cannot be promoted and used in the detection of "multiple products and one name" problems in the cataloging of large and complex equipment.

[0051] Therefore, overall, existing technologies are mostly limited to manual means, highly dependent on expert knowledge and experience, and cannot automatically and intelligently solve the problem of "multiple items under one name" in cataloging, and there are problems such as high cost and low efficiency.

[0052] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0053] The problem detection method of multiple equipment catalogs with the same name provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send the data subset of the reference equipment and the data subset of the equipment to be detected to the server 104. After the server 104 receives the data subset of the reference equipment and the data subset of the equipment to be detected, for the data subset of the reference equipment and the data subset of the equipment to be detected, the server 104 uses the data subset of the reference equipment as a reference entity and the data subset of the equipment to be detected as a detection entity, and uses the attributes or attribute values ​​contained in the reference entity and the detection entity as nodes to determine the weights between the nodes, and the weights between the nodes are expressed as the association relationship between the nodes, and the first graph structure data and the second graph structure data are obtained. A graph attention network is used to obtain a first node feature vector and a second node feature vector based on the first graph structure data and the second graph structure data. The first entity embedding and the second entity embedding are obtained based on the first node feature vector and the second node feature vector. Adaptive coding is generated based on the first entity embedding and the second entity embedding. Determine the distance between the adaptive coding and the first entity embedding to form a residual vector. Use a convolutional neural network to convert the residual vector into a signature vector. Use the signature vector as input to generate a result probability value of attribute value matching through a high-speed network. Normalize the result probability value of attribute value matching through an activation function to generate a product matching label. Determine whether there is a problem of multiple equipment cataloging with the same name based on the product matching label. The server 104 can feed back the obtained detection results to the terminal 102. In addition, in some embodiments, the problem detection method of multiple equipment cataloging with the same name can also be implemented separately by the server 104 or the terminal 102, such as the terminal 102 can directly perform a problem detection of multiple equipment cataloging with the same name for the data subset of the reference equipment and the data subset of the equipment to be detected, or the server 104 can obtain the data subset of the reference equipment and the data subset of the equipment to be detected from the data storage system, and perform a problem detection of multiple equipment cataloging with the same name for the data subset of the reference equipment and the data subset of the equipment to be detected.

[0054] The terminal 102 may be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart vehicle-mounted devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers, or may be a cloud server.

[0055] In an exemplary embodiment, Figure 2As shown, a method for detecting the problem of multiple equipment catalogs with the same name is provided. The method is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or can be executed by a terminal and a server together. In the embodiment of the present application, the method is applied to Figure 1 The server 104 in the example is used as an example to illustrate, including the following steps 200 to 209. Among them:

[0056] Step 200: Using the set variety name of the equipment as a keyword, retrieve the equipment product data to obtain a data subset of the reference equipment and a data subset of the equipment to be detected. The data subset of the reference equipment and the data subset of the equipment to be detected include multiple attributes, each of which includes multiple attribute values. The attributes include product name, variety name, natural attributes, and application attributes. The self-weight of the detection entity is the weight between nodes in the second graph structure data.

[0057] Step 201: Take the data subset of the reference equipment as the reference entity, take the data subset of the equipment to be detected as the detection entity, take the attributes or attribute values ​​contained in the reference entity and the detection entity as nodes, determine the weights between the nodes, and express the weights between the nodes as the association relationship between the nodes, to obtain the first graph structure data and the second graph structure data.

[0058] Step 202: Using a graph attention network to obtain a first node feature vector and a second node feature vector based on the first graph structure data and the second graph structure data.

[0059] Step 203: Obtain a first entity embedding and a second entity embedding based on the first node feature vector and the second node feature vector.

[0060] Step 204: Generate an adaptive code based on the first entity embedding and the second entity embedding.

[0061] Step 205: Determine the distance between the adaptive encoding and the first entity embedding to form a residual vector.

[0062] Step 206: Use a convolutional neural network to convert the residual vector into a signature vector.

[0063] Step 207: Using the signature vector as input, a result probability value of attribute value matching is generated through a high-speed network.

[0064] Step 208: normalize the result probability value of the attribute value matching through an activation function to generate a product matching label.

[0065] Step 209: Determine whether there is a problem of multiple equipment catalogs with the same name based on the product matching tags.

[0066] Implementing the above steps 200 to 209 can realize intelligent detection of the "multiple items with one name" problem in the cataloging of large and complex equipment, reduce manual participation, improve the objectivity and accuracy of problem detection results, and reduce the investment cost of resources such as manpower and time.

[0067] In another exemplary embodiment of the present application, the process of obtaining the data subset in step 200 may be: using the specific variety name of the large and complex equipment As keywords, the equipment product data is searched to obtain the data subset with the same product name, among which equipment product The corresponding data subset records are:

[0068] .

[0069] in, is the product name, is the name of the variety, For natural attributes, For application properties.

[0070] In another exemplary embodiment of the present application, the graph attention neural network uses the attention mechanism to assign different weights to nodes, which makes up for the traditional method's neglect of the different importance of nodes and performs well in the name matching task. Therefore, entities and relationships must first be converted into a graph structure. , which may correspond to the same variety name , natural attributes and apply properties This correspondence can not only extract the underlying features through graph structure data, but also build more direct connections, providing a more direct way to integrate external information for data subsets with the same variety name. Graph structure data can distinguish the importance of different nodes, and build a graph of entities that need to be matched by assigning weights, which can reflect the multi-level relationships in the entities. The degree of connection between different nodes in the graph structure is represented by the association weight.

[0071] In order to achieve the above purpose, a set of entities The implementation process of determining the weights between nodes in step 201 is described by taking the attributes or attribute values ​​contained in the reference entity and the detection entity as input as an example. Each entity contains attributes and attribute values ​​in the graph structure. The nodes in the graph are and nodes The relationship between them is expressed by weights, which are:

[0072] (1) When the attributes contained in the reference entity and the detection entity are used as nodes (i.e., nodes and nodes represents entity attributes), the formula Determine the weights between nodes, give common words a smaller weight, whose size is inversely proportional to the commonness of the word, and is used to filter out stop words and other words that are not helpful for matching. Represents the weight between nodes. It represents the Inverse Document Frequency (IDF). For a specific word, the IDF is obtained by dividing the total number of texts by the total number of texts containing the word, and then taking the logarithm of the result. The larger the IDF, the fewer texts contain the specific word, and the word has good classification ability. Otherwise, it does not have better classification ability.

[0073] (2) When the attributes contained in the reference entity and the detection entity are used as the first node and the attribute value is used as the second node (i.e., node and nodes Represent entity attributes and attribute values ​​respectively), the formula is used The weight between the first node and the second node is determined, and the size of the value is positively correlated with the importance, so as to extract attribute values ​​containing more information in the attributes. Indicates the frequency of keyword occurrence, that is, the frequency of keyword occurrence in the text, usually the number of times a specific word appears in the text divided by the sum of the number of times all words appear in the text. The actual operation method is , a high weight is obtained by the high frequency of a specific word in a text and the low text frequency in the entire document. , to filter common words and retain words with more classification characteristics.

[0074] (3) When the attribute values ​​contained in the reference entity and the detection entity are used as nodes (i.e., nodes and nodes represents attribute values), the formula is used Determine the weights between nodes. Here, attribute values ​​refer to text types and numeric types. and The type is represented as and . Represents the weights of different attribute value categories, expressed as:

[0075] .

[0076] In the formula, and Representation Node and The attribute value of .

[0077] When the node and When all attribute values ​​are text type, the node and The weight between The specific calculation expression is:

[0078] .

[0079] in, Represents the node in the matching window and As the probability of text attributes appearing together, and Represents the nodes and The number of times the two attribute values ​​appear respectively. max represents the maximum value function. The larger the value, the higher the degree of correlation between the two.

[0080] When the node and When the attribute values ​​of are all digital inputs, the Langmuir distance can be used to measure the correlation between the two. At the same time, the out-of-range numbers are discarded to reduce the interference of outliers. The graph structure data obtained at this time is an undirected graph, that is, a symmetric matrix. Based on this, steps 202 and 203 can use the eigenvalues ​​and eigenvectors of the Laplace matrix of the undirected graph to realize the Fourier transform of the graph, and then realize the convolution operation of the graph structure based on the frequency domain convolution, and extract the eigenvector embedding.

[0081] In another exemplary embodiment of the present application, the graph attention network can assign different weights to neighboring nodes. In an undirected graph structure, the weight relationship between nodes and edges can be represented by a symmetric matrix. In order to improve its representation performance, it needs to be converted into a matrix with a higher dimension. The nodes are realized in a similar splicing manner. To Node The connection in the direction and the opposite direction makes it asymmetric, which is more in line with the different weights between different nodes in reality. Based on this, in step 202, a linear function with learning ability can be introduced to calculate the similarity factor between nodes based on the first graph structure data and the second graph structure data to obtain the first node feature vector and the second node feature vector, so as to obtain a higher-dimensional feature representation. , calculate all adjacent nodes connected to it Similarity factor :

[0082] .

[0083] In the formula, Representation Node The initial embedding vector of Representation Node The initial embedding vector of Map the feature vector of the node to increase its dimension. Indicates that the node and nodes The feature vectors after dimension increase are concatenated together. It means mapping the connected high-dimensional features into a real number, and the result is used as the similarity factor. Then, the softmax function is used to normalize the attention weights, and the normalized attention weights are obtained as follows: :

[0084] .

[0085] In the formula, represents the activation function, Indicates nodes, Representation Node The number of adjacent nodes, Representation Node With Node The similarity factor.

[0086] According to all the calculated adjacent nodes The attention weights are weighted summed. Nodes representing the output of the graph attention layer The weight of the adjacent node information is integrated. represents the activity function, then The node feature vector can be expressed as :

[0087] .

[0088] In the formula, Indicates Second weighted sum, , Indicates The normalized result of the attention weights of the weighted summation, Indicates The shared parameters of the weighted sum.

[0089] By repeating the weighted summation process Second, extract more information and concatenate the matrices together. The feature size of each node is given by Convert to , a higher-dimensional matrix enhances the representation capability of graph nodes. In order to make the output more stable, the average is taken in each feature generation process, and the node feature vector is finally obtained. , Represents the node feature vector The vector dimension of Indicates that the vector has 1 row and d columns.

[0090] In another exemplary embodiment of the present application, each attribute value in the entity pair is compared with each other to generate a weight code, and the entity generates a self-code according to its own weight. The distance between the weight code and the self-coder is calculated by the quasi-Euclidean distance idea to form a residual vector. The two entities each calculate the residual vector for the other entity as the input of the feature aggregation layer. The embedding calculation method in this embodiment adopts a shared embedding strategy, which assigns weights to matching attributes in the attribute value embedding, and more clearly reflects the distance between the attribute value pairs in the residual vector. Based on this, the implementation process in the above step 204 can be described as:

[0091] (1) Obtain a weight vector between the first node feature vector and the second node feature vector based on the first entity embedding and the second entity embedding.

[0092] For example, the entity pair and , assuming that the entity For reference entities, entities To detect entities, the corresponding entity embedding is generated through the graph attention layer and After that, the attribute value and The weight relationship between It is expressed as follows:

[0093] .

[0094] In the formula, represents a linear function, , and is a parameter, Represents a function variable, or , and Both represent entity embedded attribute values. Representing Entities The number of nodes, Represents the Hadamard product between two matrices. Representing Entities Each attribute value in Comparison Entities All attribute values The weight vector of .

[0095] (2) Based on the weight vector and the embedding of the second entity, the weight encoding of the reference entity is obtained. Based on the above description, the weight encoding is calculated by the weight vector, and the attribute value Weight encoding Through the weight vector and entity The product of the entity embeddings is calculated as follows: .

[0096] At the same time, the entity Weight encoding By setting the attribute value Weight encoding In series, we have: , For Entity The number of nodes.

[0097] Weighted encoding generates a new encoding vector based on the similarity weights of entity pairs. The similarity between the weighted encoding and the original encoding reflects the degree of match between the two entities.

[0098] (3) Adopting the adaptive weight strategy, the adaptive encoding is obtained based on the weight encoding of the reference entity and the first entity embedding. The adaptive weight strategy is to control the degree of fusion between the original word embedding encoding and the weight encoding. For important attribute values, the degree of fusion is reduced to increase the difference. For unimportant attribute values, the degree of fusion is increased to reduce the difference.

[0099] Among them, the adaptive weight strategy Calculated as: In the formula, Indicates that the weight vector dimension is 1 row and n columns, and A parameter indicating the number of control units, Represents entity embedding vector The transpose of .

[0100] Adaptive Encoding Can integrate its own attribute value information, which is encoded by weight and entity embedding (i.e. the first entity embedding) is fused and generated:

[0101] .

[0102] In the formula, is the identity matrix, . yes The left diagonal matrix of It is the vector representation of the adaptive weight of entity N. Since only the weight corresponding to itself is calculated, the diagonal vector value is the corresponding value of its own weight, and the rest of the positions are 0.

[0103] In another exemplary embodiment of the present application, step 205 may determine the distance between the adaptive coding and the first entity embedding based on a concept similar to Euclidean distance calculation to form a residual vector.

[0104] For example, by computing adaptive coding With original encoding The difference between the two entities generates the residual vector .

[0105] .

[0106] Residual vector Adaptive Coding With entity embedding The Euclidean distance between two codes can be used to determine similarity. The element value of the residual vector is small, indicating that Similar to , otherwise they are not similar.

[0107] In another exemplary embodiment of the present application, a convolutional neural network is used in step 206 to convert the residual vector into a signature vector, which is essentially a feature aggregation process.

[0108] Feature aggregation is to transform the residual matrix formed by the residual vector generated by the previous layer into a signature vector with a certain size through a layer of deep neural network (i.e. convolutional neural network). , the residual matrix represents the entity and entities The deep feature distance of the image is 100%, but its matrix form does not meet the processing requirements of HighwayNet (high-speed network) and needs to be further transformed. By connecting the convolution operation, Max pooling operation, dropout operation and activation function in series to form a confusion function (i.e., deep neural network), a matrix form that meets the processing requirements of HighwayNet is generated in the form of matrix transformation. When performing convolution operation, filters are set. Multi-scale matching features can be extracted through window settings of various sizes. Max pooling operation selects the maximum value of matching features at the corresponding filter scale, corresponding to the feature area with the most information content.

[0109] In another exemplary embodiment of the present application, the generated signature vector As input, the final name matching result is generated through the two-layer activation function of HighwayNet. The activation function is the ReLu function, which saves the calculation amount of the error gradient in the back propagation, and at the same time alleviates the overfitting problem through the sparsity of the zero-output network. The loss function of HighwayNet is set to the cross entropy loss function, and finally the loss function is minimized, which is:

[0110] .

[0111] In the formula, for and The cross entropy loss between is the matching distribution output by the matching prediction layer, is the matching tag, is the outcome probability value.

[0112] In another exemplary embodiment of the present application, the implementation process of step 209 may be:

[0113] When the product matching tag is 1, it means that the reference equipment and the equipment to be tested are the same product, and there is a problem of multiple equipment being cataloged with the same name.

[0114] When the product matching tag is 0, it means that the reference equipment and the equipment to be tested are different products, and there is no problem of multiple equipment being cataloged with the same name.

[0115] In another exemplary embodiment of the present application, in the specific implementation process, the above-mentioned entire implementation process is integrated into a detector model for explanation. The detector model, by constructing graph structure data, first performs name character matching training on product names, natural attributes and application attributes, and then inputs a new entity pair to obtain a probability value through the trained model to determine whether it is the same variety. Based on this, the main implementation process of the problem detection method of the same name of multiple equipment catalogs provided by the present application may include two parts, one part is detector training, and the other part is problem detection. Among them, detector training may include 1) graph construction, 2) graph attention, 3) attribute value embedding sharing, 4) feature aggregation, 5) matching prediction and 6) basic model generalization. 1) The implementation process of graph construction can refer to the description of steps 200 and 201 above. 2) The implementation process of graph attention can refer to the description of steps 202 and 203 above. 3) The implementation process of attribute value embedding sharing can refer to the description of steps 204 and 205 above. 4) The process of feature aggregation can refer to the description of step 206 above. 5) The implementation process of matching prediction can refer to the description of the above step 207. 6) The implementation process of basic model generalization can be described as:

[0116] Generate basic models and parameters through sample data training of basic training data sets , defining this version model as The basic model is stored and can be directly tested on the data to be tested in the next part of the problem detection process. For data with professional tendencies, small sample data can be generated by manually setting labels. Retraining the base model Model and parameters , and then test the rest of the data, so that the final model (i.e., the detector model) has a stronger ability to identify and detect professional-oriented data, and save the model. For data with other professional tendencies, the generated model can be annotated based on the basic model. And the corresponding parameters , so that the model has good generalization to other professional data.

[0117] Based on the above description, another part of the problem detection process can be illustrated as follows:

[0118] Equipment products is the reference entity, The word sequence of the two equipment products to be evaluated as testing entities consists of two parts: ① Product name ②Natural attributes of equipment product records , application attribute data , these data can be obtained through word segmentation.

[0119] Reference Entity and detect entities The word sequence is represented as a graph-type data structure through the graph construction layer, where the detection entity The two-part word sequence is represented as two attributes in the graph structure. The attribute contains the attribute value. Calculate the weights between two entity attributes, attributes and attribute values, and attribute values ​​and attribute values ​​to generate graph structure data.

[0120] For detecting entities , through the graph attention network, different weights can be assigned to nodes and neighboring nodes in the graph, and finally a symmetric matrix is ​​used to represent the graph structure. Product Name For example, by calculating this property Similarity factor between adjacent nodes , use softmax to normalize the weights, then perform weighted summation and take the average to get the vector (Indicates the detection entity middle ), whose node feature vector Contains the weights of all neighboring node information.

[0121] In the attribute value comparison stage, the detection entity is generated through the graph attention layer Entity embedding , The matrix is ​​composed of detection entities of , ,and The feature vectors of their respective attribute values ​​are combined to refer to the entity and detect entities All attribute values ​​in two entities are compared to generate weighted encoding (Indicates the detection entity Middle weight encoding of attribute values), detecting entities The weight encoding is achieved by concatenating Get, record as The weight coding is further optimized by using the adaptive weight strategy, and the adaptive coding is calculated. , From the reference entity The feature map is integrated with its own attribute value information, and the adaptive coding is finally calculated. With entity embedding The difference between The residual vector between them forms the difference matrix .

[0122] By calculating the entity The difference matrix between , using a deep neural network to transform the difference matrix Convert to signature vector , through the two-layer activation function of HighwayNet, the result probability value of attribute value matching is generated , the matching probability of different attribute values ​​of the same product is normalized by the softmax function to generate product matching labels. When the label is 1, it means that the entity pair For the same product, there is a "multiple products with one name" problem. When the tag is 0, it does not point to the same product, and there is no "multiple products with one name" problem.

[0123] In summary, compared with the prior art, this application also has the following advantages:

[0124] 1. This application uses a graph attention network, graph structure and attention mechanism for feature extraction and semantic modeling, reduces additional data alignment work, can process different types of data at the same time, and optimizes the problem of high overhead in processing heterogeneous data.

[0125] 2. This application adopts a multi-head attention mechanism to incorporate the similarity between nodes into the attribute value features, fuses structural information and semantic information, enriches the representation of node feature vectors, and more accurately represents the relationship between nodes.

[0126] 3. In the comparison of attribute values, this application adopts an adaptive weight coding strategy, which makes up for the fact that the traditional coding strategy ignores the weight of the attribute value itself, thereby more accurately reflecting the similarity between the two attribute values, supporting the model to make better decisions, and improving the accuracy of the model.

[0127] Based on the same inventive concept, the embodiment of the present application also provides a system for detecting the problem of multiple equipment cataloging with the same name, which is used to implement the method for detecting the problem of multiple equipment cataloging with the same name mentioned above. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more embodiments of the system for detecting the problem of multiple equipment cataloging with the same name provided below can refer to the limitations of the method for detecting the problem of multiple equipment cataloging with the same name above, and will not be repeated here.

[0128] In an exemplary embodiment, Figure 3 As shown, a system for detecting problems of multiple equipment catalogs with the same name is provided, including:

[0129] The graph construction layer is used to retrieve the equipment product data using the set variety name of the equipment as a keyword, obtain the data subset of the reference equipment and the data subset of the equipment to be detected, use the data subset of the reference equipment as the reference entity, use the data subset of the equipment to be detected as the detection entity, use the attributes or attribute values ​​contained in the reference entity and the detection entity as nodes, determine the weights between nodes, and express the weights between nodes as the association relationship between nodes to obtain the first graph structure data and the second graph structure data. The data subset of the reference equipment and the data subset of the equipment to be detected include multiple attributes, and each attribute includes multiple attribute values. Attributes include product name, variety name, natural attributes, and application attributes.

[0130] The graph attention layer is connected to the graph construction layer, and is used to use the graph attention network to obtain the first node feature vector and the second node feature vector based on the first graph structure data and the second graph structure data, and to obtain the first entity embedding and the second entity embedding based on the first node feature vector and the second node feature vector.

[0131] The attribute value comparison layer is connected to the graph attention layer, and is used to generate an adaptive code based on the first entity embedding and the second entity embedding, determine the distance between the adaptive code and the first entity embedding, and form a residual vector. The self-weight of the detected entity is the weight between the nodes in the second graph structure data.

[0132] The feature aggregation layer, connected to the attribute value comparison layer, is used to convert the residual vector into a signature vector using a convolutional neural network.

[0133] The matching prediction layer is connected to the feature aggregation layer and is used to take the signature vector as input, generate the result probability value of the attribute value matching through the high-speed network, normalize the result probability value of the attribute value matching through the activation function, generate the product matching label, and determine whether there is a problem of multiple equipment cataloguing the same name based on the product matching label.

[0134] Based on the above system structure, a pair of entities As the input of the graph construction layer, it is first converted into a word sequence and constructed into graph data through associated weights, which contains the similarity relationship between entities. After obtaining the weight node graph (i.e., graph structure data), the graph attention layer captures the information of adjacent nodes through the graph attention network to generate node embedding. In the attribute value comparison layer, the attribute values-attribute values ​​in the entity pair are compared one by one, and the weight encoding is calculated through an adaptive strategy. The residual vector is calculated with the autoencoder, which contains the similarity between the attribute values. The feature aggregation layer fuses the comparison features generated by the previous layer through a convolutional neural network to form a feature vector of equal length. Finally, the matching prediction layer calculates the function loss of the feature vector through HighwayNet and the activation function to perform matching prediction and obtain the result probability value of the name matching.

[0135] In an exemplary embodiment, the present application can also be applied to data governance. By constructing deep features of data items through feature engineering, it can be applied in multi-source data duplication detection work in the field of data governance.

[0136] In an exemplary embodiment, the present application can also be extended to the prediction of the molecular structure of medical drugs. In the relevant experiments of molecular structure expansion, the structural changes can be judged by matching the newly generated structure with the original structure.

[0137] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 4As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store a variety of equipment catalog detection data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for detecting problems with the same name in the catalog of multiple equipment is implemented.

[0138] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0139] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0140] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0141] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0142] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0143] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0144] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0145] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for detecting problems with multiple equipment catalogs having the same name, characterized in that: The method for detecting the problem of multiple equipment catalogs having the same name includes: Using the set variety name of the equipment as a keyword, the equipment product data is searched to obtain a data subset of the reference equipment and a data subset of the equipment to be detected; the data subset of the reference equipment and the data subset of the equipment to be detected include multiple attributes, each attribute includes multiple attribute values; the attributes include product name, variety name, natural attributes and application attributes; The data subset of the reference equipment is used as a reference entity, the data subset of the equipment to be detected is used as a detection entity, the attributes or attribute values ​​contained in the reference entity and the detection entity are used as nodes, the weights between the nodes are determined, and the weights between the nodes are expressed as associations between the nodes, thereby obtaining first graph structure data and second graph structure data; the self-weight of the detection entity is the weight between the nodes in the second graph structure data; Using a graph attention network to obtain a first node feature vector and a second node feature vector based on the first graph structure data and the second graph structure data; Obtaining a first entity embedding and a second entity embedding based on the first node feature vector and the second node feature vector; generating an adaptive encoding based on the first entity embedding and the second entity embedding; determining a distance between the adaptive encoding and the first entity embedding to form a residual vector; Using a convolutional neural network to convert the residual vector into a signature vector; Taking the signature vector as input, generating a result probability value of attribute value matching through a high-speed network; Normalizing the result probability value of the attribute value matching through an activation function to generate a product matching label; Determining whether there is a problem of multiple equipment being catalogued with the same name based on the product matching tags; The method of using the attributes or attribute values ​​contained in the reference entity and the detection entity as nodes and determining the weights between the nodes includes: When the attributes contained in the reference entity and the detection entity are used as nodes, the formula is adopted Determine the weights between nodes; When the attributes contained in the reference entity and the detection entity are used as the first node and the attribute value is used as the second node, the formula is adopted determining a weight between the first node and the second node; ; When the attribute values ​​contained in the reference entity and the detection entity are used as nodes, the formula is adopted Determine the weights between nodes, distinguish between attribute values ​​of text type and numeric type, and assign nodes and The type is represented as and ; In the formula, represents the weight between nodes, Indicates the reverse file frequency, Indicates the frequency of occurrence of a keyword; the keyword is a representation related to the set equipment attribute value; Represents the weights of different attribute value categories, , and Representation Node and Attribute value of When the node and nodes When all attribute values ​​are text type, the node and nodes The weight between The calculation expression is: ； in, Represents the node in the matching window and nodes As the probability of text attributes appearing together, and Represents the nodes and nodes The number of times each of the two attribute values ​​appears, max represents the maximum value function; When the node and nodes When the attribute values ​​of are all numeric inputs, the Rankin distance is used to measure the correlation between the two; Generating an adaptive encoding based on the first entity embedding and the second entity embedding includes: Obtaining a weight vector between the first node feature vector and the second node feature vector based on the first entity embedding and the second entity embedding; Obtaining a weight encoding of the reference entity based on the weight vector and the second entity embedding; The adaptive coding is obtained based on the weight coding of the reference entity and the first entity embedding by adopting an adaptive weight strategy; the adaptive weight strategy is to control the degree of fusion of the original word embedding coding and the weight coding.

2. The method for detecting the problem of multiple equipment catalogs having the same name according to claim 1, characterized in that: Determining a distance between the adaptive encoding and the first entity embedding to form a residual vector includes: Based on the Euclidean distance calculation concept, the distance between the adaptive coding and the first entity embedding is determined to form a residual vector.

3. The method for detecting the problem of multiple equipment catalogs having the same name according to claim 1, characterized in that: Determine whether there are multiple equipment catalogs with the same name based on the product matching tags, including: When the product matching tag is 1, it means that the reference equipment and the equipment to be tested are the same product, and there is a problem of multiple equipment being cataloged with the same name; When the product matching tag is 0, it indicates that the reference equipment and the equipment to be detected are different products, and there is no problem of cataloging multiple equipment with the same name.

4. A system for detecting problems with multiple equipment cataloguing the same name, characterized in that: The system for detecting problems of multiple equipment catalogs with the same name executes the method for detecting problems of multiple equipment catalogs with the same name as described in any one of claims 1 to 3; The problem detection system of multiple equipment catalogs with the same name includes: A graph construction layer is used to retrieve equipment product data using the set variety name of the equipment as a keyword, obtain a data subset of the reference equipment and a data subset of the equipment to be detected, use the data subset of the reference equipment as a reference entity, use the data subset of the equipment to be detected as a detection entity, use the attributes or attribute values ​​contained in the reference entity and the detection entity as nodes, determine the weights between nodes, and express the weights between nodes as associations between nodes, to obtain first graph structure data and second graph structure data; the data subset of the reference equipment and the data subset of the equipment to be detected include multiple attributes, each attribute includes multiple attribute values; the attributes include product name, variety name, natural attributes and application attributes; the self-weight of the detection entity is the weight between nodes in the second graph structure data; A graph attention layer, connected to the graph construction layer, for obtaining a first node feature vector and a second node feature vector based on the first graph structure data and the second graph structure data using a graph attention network, and obtaining a first entity embedding and a second entity embedding based on the first node feature vector and the second node feature vector; an attribute value comparison layer, connected to the graph attention layer, for generating an adaptive code based on the first entity embedding and the second entity embedding, determining a distance between the adaptive code and the first entity embedding, and forming a residual vector; A feature aggregation layer, connected to the attribute value comparison layer, for converting the residual vector into a signature vector using a convolutional neural network; The matching prediction layer is connected to the feature aggregation layer, and is used to take the signature vector as input, generate a result probability value of the attribute value matching through a high-speed network, normalize the result probability value of the attribute value matching through an activation function, generate a product matching label, and determine whether there is a problem of cataloging multiple equipment with the same name based on the product matching label.

5. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for detecting problems of cataloging multiple equipment with the same name as described in any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for detecting problems of cataloging multiple equipment with the same name as described in any one of claims 1 to 3 is implemented.

7. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for detecting problems of cataloging multiple equipment with the same name as described in any one of claims 1 to 3 is implemented.

Images