Electric power text knowledge discovery method and equipment based on multi-source data fusion

Abstract

The present invention discloses a method and device for discovering knowledge of electric power text based on multi-source fusion of data, and belongs to the technical field of electric power equipment. The method for discovering knowledge of electric power text based on multi-source fusion of data of the present invention can automatically extract key information required for defect diagnosis from electric power natural language data and electric power equipment status text by constructing a text data extraction model, a rule engine judgment model, and a graph data mining model, and integrates graph neural network technology and rule engine technology to evaluate the operating status of electric power equipment recorded in the electric power text, and can determine the fault location and cause; at the same time, it realizes multi-source fusion of at least two types of text, provides more accurate auxiliary decision-making for electric power equipment fault diagnosis, reduces the quality requirements of text record information for operation and maintenance personnel, and the scheme is scientific and reasonable, the calculation is simple, the diagnosis accuracy is high, and it is practical.

Description

Technical Field

[0001] The present invention relates to a method and device for discovering electric power text knowledge based on multi-source data fusion, and belongs to the technical field of electric power equipment. Background Art

[0002] During the daily inspection and maintenance of power equipment, the power grid operation and maintenance department has accumulated a large number of equipment operation status texts. Due to the complex and changeable operation of power equipment, the current judgment and handling of equipment defects still mostly rely on the knowledge and experience of operation and maintenance personnel. However, for inexperienced operation and maintenance personnel, it is very likely that they will not be able to accurately and quickly judge whether there are defects in the power equipment and the defective components and causes based on the inspection results. The large number of equipment defect cases accumulated by power companies can provide guidance for equipment defect diagnosis. The case records information such as equipment defect phenomena, processing processes, and solutions, which contain rich expert experience and have high reference value in the daily operation and maintenance of equipment. However, due to the large number of cases and low daily utilization, equipment defect cases have not yet played their due role. If the rich knowledge contained in the case can be learned with the help of computer algorithms and automatic diagnosis of power equipment based on status text can be realized, the reliability of equipment operation can be greatly improved and labor costs can be reduced.

[0003] The status text of power equipment is generally recorded in natural language, but there are still certain difficulties in the process of extracting, representing and analyzing natural language information with the help of computers. At present, knowledge discovery equipment based on shallow machine learning methods has achieved certain results in data mining of alarm signal texts and distribution line tripping reporting texts. Most of them use stop word removal and part-of-speech tagging to select text features, and then use algorithms such as support vector machines and random forests to analyze text feature vectors. However, due to the professional characteristics and different recording styles of the writers, there are many common words used in the status text of power texts; secondly, the work in daily inspections is complicated, and there are often multi-layer nested relationships between various information; thirdly, power equipment has the characteristics of complex structure, diverse defect causes, and complex logical relationships. At present, the accuracy of fault identification and diagnosis needs to be further improved.

[0004] After searching, "Construction of Transformer Fault Diagnosis System Based on Data Mining" proposed the shortcomings of existing transformer fault diagnosis methods, constructed a transformer fault tree diagnosis model using fault tree analysis, and realized in-depth mining of transformer fault data based on traditional pattern recognition methods such as cluster analysis and decision tree induction classification. However, this method has the following problems: it can only analyze a specific accident, rather than a process or equipment system, and is local; for complex diagnostic models, the fault tree is large in scale, the calculation is complex, and the diagnostic accuracy is low.

[0005] After searching, the "Power Equipment Defect Record Retrieval Method Based on Knowledge Graph Technology" proposed to use the existing power equipment defect record corpus to construct a power equipment defect knowledge graph; then based on the constructed knowledge graph, the power equipment defect retrieval was performed using graph search. However, this method does not use natural language processing technologies such as syntactic analysis, and cannot obtain richer and more effective semantic features in coreference resolution and relationship extraction, so the accuracy and completeness of the knowledge graph are insufficient, and the fault diagnosis accuracy is low.

[0006] After searching, the article "Research on Text Classification Model of Power Equipment Defects Based on Convolutional Neural Network" introduced the convolutional neural network model in power text processing and constructed a power defect text classification model. However, the text fault severity classification processed by the above method is simple (general, important and urgent), and it is difficult to handle the situation where there are multiple nested fault causes in the text, which is not conducive to the accurate diagnosis. Summary of the invention

[0007] In view of the defects of the prior art, the first purpose of the present invention is to provide a method that can automatically extract the key information required for defect diagnosis from the power natural language data and the power equipment status text by constructing a text data extraction model, a rule engine judgment model, and a graph data mining model, and combine the graph neural network technology and the rule engine technology to evaluate the operating status of the power equipment recorded in the power text, and can determine the fault location and cause; at the same time, it realizes the multi-source fusion of at least two data types, provides more accurate auxiliary decision-making for power equipment fault diagnosis, reduces the quality requirements of text record information for operation and maintenance personnel, and has a scientific and reasonable scheme, simple calculation, high diagnostic accuracy, and a practical power text knowledge discovery method based on multi-source fusion of data.

[0008] The second objective of the present invention is to provide a device that can effectively combine basic natural language processing with information extraction, rule engine judgment, graph data mining and other deep learning technologies by setting a touch-screen display, a processor, a power module, a text data extraction module, a graph data mining module, a rule engine judgment module, and a device memory to achieve multi-dimensional fusion of sampled data, and can automatically extract key information required for defect diagnosis from the power equipment status text, and combine graph neural network technology and rule engine technology to evaluate the operating status of the power equipment recorded in the power text, so as to provide more accurate auxiliary decision-making for operation and maintenance personnel. Power text knowledge discovery equipment based on multi-source data fusion.

[0009] To achieve one of the above purposes, the first technical solution of the present invention is:

[0010] Electric power text knowledge discovery method based on multi-source data fusion,

[0011] The following steps are involved:

[0012] The first step is to obtain power natural language data and monitoring indicator data;

[0013] The second step is to build a text data extraction model based on the power natural language data in the first step;

[0014] The text data extraction model is used to process the electric power natural language data and extract key information to obtain key feature information;

[0015] Build a rule engine judgment model based on historical monitoring indicator data;

[0016] The rule engine judgment model is used to analyze the monitoring indicator data based on the rule engine, compare the monitoring indicator data with the comparison rules, obtain faulty equipment information, and quickly locate the faulty components;

[0017] The third step is to build a graph data mining model to analyze the key feature information in the second step and obtain the fault information of the equipment operation phenomenon text description in the power text, so as to reduce the system memory requirements and improve the computing speed and diagnosis accuracy of the equipment;

[0018] The fourth step is to analyze the electric power natural language data and monitoring indicator data to be judged based on the graph data mining model in the third step and the rule engine judgment model in the second step, realize multi-source fusion of at least two data types, and determine the fault location and cause.

[0019] Power natural language data (power text) records equipment failure information such as equipment failure phenomenon, processing process, and solutions. It contains rich equipment quality information and expert processing experience, and has a high reference value when handling similar equipment failures. If a computer can quickly retrieve failure cases with similar failure conditions through an inspection record or sensor data report, it can draw on the experience of predecessors to speed up the progress of fault handling and discover similar latent anomalies or defects, which is of great significance to improving the intrinsic safety level of equipment.

[0020] Therefore, after continuous exploration and experimentation, the present invention can automatically extract the key information required for defect diagnosis from the power natural language data and the power equipment status text by constructing a text data extraction model, a rule engine judgment model, and a graph data mining model, and combine graph neural network technology and rule engine technology to evaluate the operating status of the power equipment recorded in the power text, and can determine the fault location and cause; at the same time, it realizes multi-source fusion of at least two data types, provides more accurate auxiliary decision-making for power equipment fault diagnosis, and reduces the quality requirements of text record information for operation and maintenance personnel. The solution is scientific and reasonable, the calculation is simple, the diagnosis accuracy is high, and it is feasible.

[0021] As the preferred technical measures:

[0022] In the second step, the text data extraction model extracts and preliminarily judges the key indicator information features describing the equipment operation phenomenon in the power text, and obtains a triple including equipment, equipment status type, and equipment status parameter;

[0023] The specific construction method is as follows:

[0024] S1: Preliminary extraction of key features of power natural language data to obtain preliminary extraction results;

[0025] S2: Standardize and generalize the preliminary extraction results in S1 to obtain feature enhanced data;

[0026] S3: Determine the device status parameters based on the feature enhancement data in S2.

[0027] As the preferred technical measures:

[0028] In S1: preliminary extraction of key features, specifically including the following contents:

[0029] When performing triple recognition and extraction of power natural language data, the semantic associations before and after the characters are taken into account, and the bidirectional semantic information of power natural language data is obtained;

[0030] The method for obtaining the bidirectional semantic information is as follows:

[0031] The bidirectional long short-term memory model is used to transmit the forward and reverse information of the power natural language data, and the hidden state sequence (h1,h2,...,h n )∈R n*m , and then connect a linear layer to convert the dimension of the hidden state sequence from m-dimension to k-dimension, and calculate the probability of the key feature information of the characters in the power natural language data;

[0032] At the same time, the random conditional field is used to obtain the connection relationship between the character labels in the power natural language data to obtain the optimal key feature information extraction;

[0033] Under the given condition X, the random conditional field has condition Y that satisfies the Markov property. Each feature template must slide over each position, and the score of the feature template for the entire sentence is obtained by adding them up. Finally, the weighted summation of multiple feature templates is performed to achieve the preliminary extraction of key features. The calculation formula is as follows:

[0034]

[0035] where λ k is the weight of the edge feature function, μ l is the weight of the point feature function, t kis the characteristic function on the edge, s l is the characteristic function at the node, λ k and μ l is the corresponding weight; Z(x) is the normalization factor, k is the serial number of the k-th state feature, and i is the serial number of the i-th character.

[0036] As the preferred technical measures:

[0037] If the equipment directly uses the preliminary extraction results of key features as the input of the graph data mining model, it will lead to low diagnostic accuracy, and the graph data mining model is difficult to apply to a variety of different power equipment;

[0038] This is because the equipment information recorded in the description of equipment operation phenomena in the power text has the problem of different expressions for the same equipment, parts and phenomena during the recording process;

[0039] In S2, the standardization and generalization processing specifically includes the following contents:

[0040] We have carried out knowledge standardization and vocabulary sorting work on synonym relationships and hyponymy relationships;

[0041] Based on the synonym relationships and hyponymy relationships obtained through sorting, the extracted key feature information is subjected to synonym standardization comparison and replacement, thus achieving the standardization of key feature information;

[0042] For example, the standardized synonym of "tank body" is "oil tank body", and the standardized synonym of "oil stain" is "penetrating oil". The present invention realizes the expansion of specific and individual feature information into generalized feature information by revising the upper and lower relationship of feature information, thereby completing the generalization of feature information.

[0043] As the preferred technical measures:

[0044] S3: Device status parameter judgment, specifically includes the following contents:

[0045] S31, identifying the attributes and values ​​corresponding to the target named entity for any given power natural language data;

[0046] S32, according to the attributes and values ​​in S31, obtain a triple of device, device state type, and device state parameter;

[0047] S33, judging whether the device status parameter is normal or abnormal according to the triplet in S32;

[0048] As the preferred technical measures:

[0049] The rule engine judgment model uses business criteria and national standards to establish a rule engine system to judge the monitoring indicator data; that is, the pure digital information set extracted from each monitoring indicator data is taken as an event, and the rule engine system is used to judge whether the event matches the rule, and this is used as a supplementary diagnosis;

[0050] It specifically includes the following steps:

[0051] Step 1: Input the monitoring indicator data into the rule base of the rule engine pre-stored in the device memory in the form of wireless transmission;

[0052] Step 2: Use the pattern matcher to compare the rules in the rule base with the monitoring indicator data in step 1;

[0053] Step 3: If the monitoring index data successfully matches the rules of the rule base in step 2, the corresponding diagnostic results are output; otherwise, the monitoring index data is re-matched with the rules of the rule base until all the rules are matched to accurately determine the fault location and cause, providing more accurate auxiliary decision-making for power equipment fault diagnosis.

[0054] As the preferred technical measures:

[0055] The graph data mining model matches the key indicator information in the power natural language data with the graph structure to form a power heterogeneous graph to diagnose the operating status of the power equipment;

[0056] A graph structure is a data structure consisting of two parts: vertices and edges. All vertices contained in the graph structure G constitute the vertex set V, and all edges contained in it constitute the edge set E. A complete graph structure is composed of the above vertex set V and edge set E. The graph structure can be mathematically expressed as follows:

[0057] G=(V,E)

[0058] The graph neural network algorithm is a neural network algorithm that directly acts on the graph structure G. The principle of GNN is to sample adjacent node information at each layer and continuously update the node state to obtain a state containing adjacent node information and graph topology, and output the required results in a specific way (such as the similarity between the target node and the adjacent node selected in this application).

[0059] The corresponding relationships include: equipment, components, and indicators correspond to nodes in the graph structure, and the association relationships between equipment and components, and between components and indicator nodes correspond to edges in the graph structure;

[0060] Among them, each node in the graph has different properties due to the different corresponding information content.

[0061] As the preferred technical measures:

[0062] The power heterogeneous graph includes two major types of nodes, one is a newly entered text node, the newly entered text node includes the power natural language data to be diagnosed, and the other is a word node, the word node includes the existing power text, and the power text includes equipment information, phenomenon information, and indicator information;

[0063] The process of power equipment status diagnosis is as follows:

[0064] A graph structure with the newly entered text node as the target node and the word node as the adjacent node, and the edge connecting the target node and the adjacent node represents the similarity between the two.

[0065] As the preferred technical measures:

[0066] The graph data mining model uses a one-hot encoding representation method to encode the triples obtained by sequence labeling and extraction based on Bi-LSTM+CRF multi-channel fusion, and expresses the node information with a low-dimensional vector, which includes the following steps:

[0067] Step 31: In order to enable the diagnostic algorithm to successfully process the graph structure information, a recurrent convolutional neural network method is used to vectorize the nodes in each graph and embed the graph structure; the recurrent convolutional neural network receives the input x at time t t After that, the value of the hidden layer is s t , and its calculation formula is as follows:

[0068] s t =f(Ux t +Ws t-1 )

[0069] in, is the input x t The weight matrix of is the previous value s t-1 As the weight matrix of this input, f is the activation function. If the loop is iterated, the output value o is obtained. t , and its calculation formula is as follows:

[0070] o t =g(Vs t )

[0071] =Vf(Ux t +Ws t-1 )

[0072] =Vf(Ux t +Wf(Ux t-1 +Ws t-2 ));

[0073] =Vf(Uxt +Wf(Ux t-1 +Wf(Ux t-2 +Ws t-3 )))

[0074] =Vf(Ux t +Wf(Ux t-1 +Wf(Ux t-2 +Wf(Ux t-3 +...))))

[0075] Among them, V is the matrix of output weights; W is the weight matrix.

[0076] As can be seen from the above, the output value o of the recurrent neural network t It is affected by all previous input values, so the recurrent neural network can see any number of previous input values;

[0077] Step 32: Encode the target node to be diagnosed with its adjacent nodes to obtain the unique hot encoding of each word, which specifically includes the following contents:

[0078] First, we need to create an all-zero vector representing the total length of the vocabulary |V|.

[0079] Then, the index dimension corresponding to each word in the vocabulary is set to 1, and other elements remain unchanged to obtain the final one-hot vector, which is equivalent to training a virtual network to store the structural information of each node;

[0080] After inputting the adjacent nodes, the vector output by the network is the embedding result of the target node to be diagnosed.

[0081] Step 33: Use an attention-based graph convolutional network to iterate nodes, and use the attention mechanism to perform weighted summation on the features of adjacent nodes. The weight of the feature depends entirely on the node feature, and there is no need to obtain the features of the entire graph structure, which is independent of the graph structure. This greatly improves the generalization ability of the diagnostic question-answering model as the scale of the dynamic graph increases. The input of the graph attention layer is the node feature vector set:

[0082]

[0083] Where N is the number of nodes, and F is the number of node features; the matrix dimension is N×F, representing the features of all nodes, and R represents the features of a certain node, and the matrix dimension is F×1; preferably, the number of nodes is 550+m (where 550 is the collected power text nodes, and m is the text node to be diagnosed);

[0084] The corresponding node features are: orientation, fault phenomenon type, and state, and the value of F is 3;

[0085] Step 34: In order to obtain the conversion between input and output, the input node features are at least linearly transformed to obtain the corresponding output features, so it is necessary to set the weight matrix for all nodes

[0086] The initial state of the target node corresponding to the power natural language data to be diagnosed is set as The initial state of the adjacent nodes is set to The similarity between the target node and neighbor nodes in each layer is calculated through the power heterogeneous graph;

[0087] Step 35: To measure the similarity between the adjacent node j in the power heterogeneous graph and the target node i of the power natural language data to be diagnosed, the graph data mining model uses the constructed power heterogeneous graph to perform self-attention mechanism operations on the target node i to be diagnosed and each of its adjacent nodes in turn. The attention coefficient between two nodes in the lth layer is expressed as:

[0088]

[0089] Where: f(﹒) is the local transformation function, which is used to reflect the calculation result of the similarity between the target node to be diagnosed and its adjacent nodes. This function is determined by the gradient descent strategy of the loss function of the pre-trained model;

[0090] l=1,2,3,…;

[0091] The loss function of the pre-trained model is obtained through the supervised nodes. The calculation formula of the loss function is as follows:

[0092]

[0093] Where p(﹒) is the true probability distribution, and q(﹒) is the calculated predicted probability distribution;

[0094] Step 36: In order to make the attention coefficient easier to calculate and compare, the logistic regression Softmax function is introduced to the adjacent node set N of the target node i i All adjacent nodes j in are regularized. The specific method is as follows:

[0095]

[0096] The regularized attention coefficients between nodes in the lth layer obtained through the above operation can be used to predict the output features of each node in the lth layer, that is, as the input features of each node in l+1:

[0097]

[0098] In the formula, σ is a nonlinear activation function, and the linear rectification function ReLU is used as the activation function;

[0099] Step 37: Perform similarity calculation on the features of the target node to be diagnosed obtained through iterative updating of the graph neural network and the diagnostic features of each operating state obtained through embedding, calculate and output through the logistic regression Softmax function, and the output result is the diagnostic result of the natural language data of the power to be diagnosed.

[0100] To achieve one of the above purposes, the second technical solution of the present invention is:

[0101] Power text knowledge discovery equipment based on multi-source data fusion,

[0102] Apply the above-mentioned power text knowledge discovery method based on multi-source data fusion;

[0103] It includes a touch screen, a processor, a power module, a text data extraction module, a graph data mining module, a rule engine judgment module, and a device memory;

[0104] The text data extraction module, graph data mining module, and rule engine judgment module are all stored in the device memory; the text data extraction module is used to automatically process natural language and extract key information;

[0105] The graph data mining module is used to analyze the text description information of equipment operation phenomena using the graph neural network algorithm; the rule engine judgment module is used to analyze and judge the monitoring indicator data using the rule engine.

[0106] The present invention is equipped with a touch-screen display, a processor, a power module, a text data extraction module, a graph data mining module, a rule engine judgment module, and a device memory, and can effectively combine basic natural language processing with information extraction, rule engine judgment, graph data mining and other deep learning technologies to achieve multivariate fusion of sampled data, and can automatically extract key information required for defect diagnosis from the power equipment status text, and combine graph neural network technology and rule engine technology to evaluate the operating status of the power equipment recorded in the power text, so as to provide more accurate auxiliary decision-making for operation and maintenance personnel, with a simple solution and high diagnostic accuracy.

[0107] Compared with the prior art, the present invention has the following beneficial effects:

[0108] After continuous exploration and experiments, the present invention can automatically extract the key information required for defect diagnosis from the power natural language data and the power equipment status text by constructing a text data extraction model, a rule engine judgment model, and a graph data mining model, and combine the graph neural network technology and the rule engine technology to evaluate the operating status of the power equipment recorded in the power text, and can determine the fault location and cause; at the same time, it realizes multi-source fusion of at least two types of text, provides more accurate auxiliary decision-making for power equipment fault diagnosis, and reduces the quality requirements of text record information for operation and maintenance personnel. The scheme is scientific and reasonable, the calculation is simple, the diagnosis accuracy is high, and it is feasible.

[0109] Furthermore, the present invention provides a touch screen, a processor, a power module, a text data extraction module, a The data mining module, rule engine judgment module, and device memory can effectively combine basic natural language processing and information extraction, rule engine judgment, graph data mining and other deep learning technologies to achieve multi-dimensional fusion of sampled data, and can automatically extract the key information required for defect diagnosis from the power equipment status text. It also combines graph neural network technology and rule engine technology to evaluate the operating status of power equipment recorded in the power text, providing operation and maintenance personnel with more accurate auxiliary decision-making, with a simple solution and high diagnostic accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] Figure 1 This is a workflow diagram of the rule engine of the present invention;

[0111] Figure 2 This is a hardware configuration diagram of the electric power text knowledge discovery device based on multi-data fusion of the present invention. DETAILED DESCRIPTION

[0112] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0113] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.

[0114] A specific embodiment of the electric power text knowledge discovery method based on multi-source data fusion of the present invention:

[0115] The electric power text knowledge discovery method based on multi-source data fusion includes the following steps:

[0116] The first step is to obtain power natural language data and monitoring indicator data;

[0117] The second step is to build a text data extraction model based on the power natural language data in the first step;

[0118] The text data extraction model is used to process the electric power natural language data and extract key information to obtain key feature information;

[0119] Build a rule engine judgment model based on historical monitoring indicator data;

[0120] The rule engine judgment model is used to analyze the monitoring indicator data based on the rule engine, compare the monitoring indicator data with the comparison rules, obtain faulty equipment information, and quickly locate the faulty components;

[0121] The third step is to build a graph data mining model to analyze the key feature information in the second step and obtain the fault information of the equipment operation phenomenon text description in the power text, so as to reduce the system memory requirements and improve the computing speed and diagnosis accuracy of the equipment;

[0122] The fourth step is to analyze the electric power natural language data and monitoring indicator data to be judged based on the graph data mining model in the third step and the rule engine judgment model in the second step, realize multi-source fusion of at least two text types, and determine the fault location and cause.

[0123] A specific embodiment of the text data extraction model of the present invention:

[0124] A large number of power equipment status texts usually record rich equipment operating status information in detail, such as basic equipment information, operating status information, equipment detection and test content and results, etc. However, the text content has problems such as scattered content distribution, many professional slangs, and inconsistent recording levels, making it difficult to directly carry out automatic extraction of power text information.

[0125] The text data extraction model of the present invention first extracts and preliminarily judges the key indicator information features describing the equipment operation phenomenon in the power text, and then obtains the "equipment, equipment status type, equipment status parameter" triple, which is the basis of power text knowledge discovery and is mainly divided into three steps: preliminary extraction of key features, feature enhancement, and equipment status parameter judgment.

[0126] The key features are initially extracted, specifically including the following contents:

[0127] When performing triple recognition and extraction of the power equipment operation status text, it is necessary to take into account certain semantic associations before and after the characters. In order to obtain the bidirectional semantic information of the power maintenance text, the present invention adopts a bidirectional long short term memory model (Bi-LSTM) to transmit the forward and reverse information of the sentence, and obtains the hidden state sequence (h1, h2, ..., h n )∈R n*m , and then connect a linear layer to convert the dimension of the hidden state sequence from m dimension to k dimension, and calculate the probability of the key feature information of the characters in the text. At the same time, the present invention uses a random conditional field (CRF) to obtain the connection relationship of the character labels in the state text to obtain the optimal key feature information extraction. CRF is given under the condition of X, Y satisfies the Markov property, each feature template must slide over each position, add up to get the score of the feature template for the entire sentence, and finally perform weighted summation on multiple feature templates. The calculation process is as follows:

[0128]

[0129] where λ k is the weight of the edge feature function, μ l is the weight of the point feature function.

[0130] Based on the above method, the preliminary extraction of key features can be achieved. For example, when the text information "the box body has obvious oil stains" is input to the device, the key feature of the original word extracted is "device (box body)_device status type (oil stains)".

[0131] The feature enhancements specifically include the following:

[0132] If the equipment directly uses the preliminary extraction results of key features as the input of the graph data mining module, it will lead to low diagnostic accuracy, and the graph data mining module is difficult to apply to a variety of different power equipment. This is because the equipment information recorded in the description of the equipment operation phenomenon in the power text has the problem of different descriptions of the same equipment, parts and phenomena during the recording process. For example: "main transformer" and "main transformer", "dielectric loss" and "dielectric loss", "oil-soluble gas" and "dissolved gas in oil" each refer to the same feature information. This will make it difficult for the power equipment defect diagnosis model in the equipment to achieve accurate diagnosis of the operating status.

[0133] The present invention performs standardization and generalization on input data. The present invention performs knowledge standardization and vocabulary sorting on synonym relationships and hyponymy relationships. Based on the sorted synonym relationships and hyponymy relationships, the present invention performs synonym standardization comparison and replacement on the extracted key feature information, thereby achieving standardization of the key feature information. For example, the standardized synonym of "tank body" is "oil tank body", and the standardized synonym of "oil stain" is "penetrating oil". The present invention achieves the generalization of the feature information by revising the hyponymy relationship of the feature information, thereby expanding the specific and individual feature information into generalized feature information.

[0134] When the text information "The box body has obvious oil stains" is input to the device, the initial extracted feature is "Device (Box Body)_Device Status Type (Oil Stain)". The extracted feature after standardization is "Device (Oil Tank Body)_Device Status Type (Penetrating Oil)", and the extracted feature after generalization is "Device (Oil Tank)_Device Status Type (Penetrating Oil)".

[0135] The device status parameter judgment specifically includes the following contents:

[0136] The present invention can identify the attributes and values ​​corresponding to the target named entity for any given text, and finally obtain the triple of "equipment, equipment status type, equipment status parameter", and add a set of rule engine systems based on business guidelines and national standards to judge the structured monitoring index data. The present invention regards the equipment status type extracted from the description of the equipment operation phenomenon in the power text as an event, and uses the rule engine to judge whether the event matches the rule, and judges whether its equipment status parameters are "normal" or "abnormal". For example, when the text information "the box body has obvious oil stains" is input to the device, the triple extracted and judged after generalization is "equipment (oil tank)_equipment status type (penetrating oil)_equipment status parameter (abnormal)". The graph data mining module will carry out the subsequent diagnosis process based on this triple.

[0137] A specific embodiment of the graph data mining model of the present invention:

[0138] The corresponding relationship of key indicator information in the power equipment status text in the graph structure is: equipment, components, indicators, etc. correspond to nodes in the graph structure, and the relationship between equipment and components, components and indicators, etc. corresponds to edges in the graph structure. The nodes in the graph have different properties due to the different corresponding information content, such as: equipment, equipment status type, equipment status parameters, reasons, etc.; the nature of the edge has different properties due to the different nodes connected, such as: "equipment" and "specific equipment components" are inclusion relationships, "fault phenomenon" and "fault cause" are causal relationships, and so on.

[0139] The present invention uses the collected 550 electric power texts as a corpus to construct a large-scale electric power field heterogeneous graph stored in the memory of the device. This heterogeneous graph contains two major types of nodes, one is the newly entered text node (text to be diagnosed), and the other is the word node (a large number of previous electric power texts, including information such as equipment, phenomena, indicators, etc.). In order to realize the diagnosis of the operating status of electric power equipment based on status text, the present invention uses the newly entered text information as the target node, the previous text information as the graph structure of the adjacent node, and the edge connecting the target node and the adjacent node represents the similarity between the two.

[0140] In order to realize the diagnosis of the operating status of power equipment based on key information, the present invention adopts a one-hot coding representation method to encode the triples obtained by sequence labeling and extraction based on Bi-LSTM+CRF multi-channel fusion, and expresses the node information with a low-dimensional vector.

[0141] A specific embodiment of the graph data mining model based on graph neural network of the present invention:

[0142] The graph data mining model based on graph neural network includes the following steps:

[0143] S1: In order to enable the diagnostic algorithm to successfully process the graph structure information, the present invention uses the one-hot encoding method to realize the vectorization of each node in the graph and realize the graph embedding processing of the graph structure. The calculation process is as follows:

[0144] o t =g(Vs t )

[0145] s t =f(Ux t +Ws t-1 )

[0146] The network receives input x at time t t After that, the value of the hidden layer is s t , the output value is o t , U is the weight matrix of input x, W is the previous value s t-1 As the weight matrix of this input, f is the activation function. If the loop is iterated, we can get:

[0147] o t =g(Vs t )

[0148] =Vf(Ux t +Ws t-1 )

[0149] =Vf(Ux t +Wf(Ux t-1 +Ws t-2))

[0150] =Vf(Ux t +Wf(Ux t-1 +Wf(Ux t-2 +Ws t-3 )))

[0151] =Vf(Ux t +Wf(Ux t-1 +Wf(Ux t-2 +Wf(Ux t-3 +...))))

[0152] As can be seen from the above, the output value o of the recurrent neural network t It is affected by all the previous input values, so the recurrent neural network can see any number of input values ​​in the previous column.

[0153] S2: The present invention encodes the target node to be diagnosed with its adjacent nodes. To obtain the one-hot encoding of each word, it is necessary to first establish an all-zero vector representing the total length of the vocabulary |V|, and then set the index dimension corresponding to each word in the vocabulary to 1, and keep other elements unchanged. The final one-hot vector can be obtained, which is equivalent to training a virtual network to store the structural information of each node. After inputting the adjacent nodes, the vector output by the network is the embedding result of the target node to be diagnosed.

[0154] S3: This invention uses an attention-based graph convolutional network to perform node iteration, and uses the attention mechanism to perform weighted summation of the features of adjacent nodes. The weight of the feature depends entirely on the node feature, and there is no need to obtain the features of the entire graph structure, which is independent of the graph structure. This greatly improves the generalization ability of the diagnostic question-answering model in the process of increasing the size of the dynamic graph. The input of the graph attention layer in this method is the node feature vector set:

[0155]

[0156] Where N is the number of nodes, and F is the number of node features. The matrix dimension is N×F, representing the features of all nodes, while R represents the features of a certain node, and the matrix dimension is F×1. In this application, the number of nodes is 550+m (where 550 is the collected power text node and m is the text node to be diagnosed); the corresponding node features are: orientation, fault phenomenon type, and state, so the value of F is 3.

[0157] S4: In order to obtain the conversion between input and output, the present invention performs at least a linear transformation on the input features to obtain the corresponding output features, so it is necessary to define a weight matrix W∈R for all nodes F′×FIn the present invention, the initial state of the target node corresponding to the text to be diagnosed is set to The initial state of the adjacent nodes is set to Based on the constructed heterogeneous graph in the power field, the similarity between the target node and the neighboring nodes in each layer is calculated.

[0158] S5: To measure the similarity between the adjacent node j and the target node i of the text to be diagnosed in the power field heterogeneous graph, the graph data mining module uses the constructed heterogeneous graph to perform self-attention mechanism operations on the target node i to be diagnosed and each of its adjacent nodes in turn. The attention coefficient between two nodes in the lth layer (l = 1, 2, 3, ...) can be expressed as:

[0159]

[0160] Where: f(﹒) is a local transformation function, which is used to reflect the calculation result of the similarity between the target node to be diagnosed and its adjacent nodes. This function is determined by the gradient descent strategy of the loss function of the pre-trained model. The loss function of the training model is obtained by the supervised nodes. The loss function of the GNN model in the present invention is expressed as:

[0161]

[0162] Where p(﹒) is the true probability distribution, and q(﹒) is the predicted probability distribution obtained by the algorithm.

[0163] S6: In order to make the attention coefficient easier to calculate and compare, the present invention introduces the logistic regression Softmax function to the adjacent node set N of the target node i i All adjacent nodes j in are regularized. The specific method is as follows:

[0164]

[0165] The regularized attention coefficients between nodes in the lth layer obtained through the above operation can be used to predict the output features of each node in the lth layer, that is, as the input features of each node in l+1:

[0166]

[0167] Wherein σ is a nonlinear activation function, and the present invention adopts a linear rectification function (ReLU function) as the activation function.

[0168] S7: Finally, the features of the target node to be diagnosed obtained through GNN iterative update are similarly calculated with the diagnostic features of each operating status obtained through embedding, and the results are calculated and output through the logistic regression Softmax function. At this time, the output result of the device is the diagnostic result of the status text of the power equipment to be diagnosed.

[0169] like Figure 1As shown, a specific embodiment of the rule engine judgment model of the present invention is:

[0170] Based on the diagnosis of the text description information of equipment operation phenomena in the power text by the graph neural network algorithm, the present invention adds a set of rule engine systems based on business criteria and national standards to judge the structured monitoring index data. The pure digital information set extracted from each online monitoring data, namely the monitoring index data, is taken as an event, and the rule engine is used to judge whether the event matches the rule, and this is used as a supplementary diagnosis. The specific process is as follows:

[0171] S1: Initial monitoring data is transmitted to the rule base of the rule engine in the form of wireless transmission;

[0172] S2: Extract the pure digital information set from the initial monitoring data and treat the monitoring indicator data as an event;

[0173] S3: The device uses a pattern matcher to search for possible rules in the rule base and compares the rules in the rule base with the events;

[0174] S4: Determine whether the event matches the rule. If the event matches the rule in the rule base successfully, the device outputs the corresponding diagnosis result; otherwise, the device returns to the rule base to search for possible rules and matches the event with the rule in the rule base until all rules are matched.

[0175] Through the processing of the above-mentioned process, the set graph neural network technology and the rule engine technology, the present invention can quickly and accurately determine the fault location and cause according to the key information in the power text, and provide more accurate auxiliary decision-making for power equipment fault diagnosis.

[0176] like Figure 2 As shown, a specific embodiment of the power text knowledge discovery device based on multi-source data fusion of the present invention is:

[0177] An electric power text knowledge discovery device based on multi-source data fusion includes a touch screen, a processor, a power module, a text data extraction module, a graph data mining module, a rule engine judgment module, and a device memory.

[0178] The text data extraction module, graph data mining module, and rule engine judgment module are all stored in the device memory;

[0179] The touch screen is a 10.2-inch 2160x1620 resolution display;

[0180] The processor is AMD Ryzen 9 5900HX processor;

[0181] Text data extraction module, used to automatically process natural language and extract key information;

[0182] The graph data mining module is used to analyze the text description information of equipment operation phenomena using the graph neural network algorithm; the rule engine judgment module is used to analyze and judge the monitoring indicator data using the rule engine.

[0183] The device memory is a storage device with a memory capacity of 6G and a storage capacity of 1TB.

[0184] Explanation of terms in this application:

[0185] Electricity Text

[0186] Refers to a text that fully records the fault phenomena, processing procedures, and solutions of power equipment at different stages. It contains a wealth of information about hidden quality problems of equipment and expert processing experience, and has a high reference value when dealing with similar equipment failures. It mainly includes: failure cases, disposal plans, specifications, guidelines, standards, scientific research papers, etc.

[0187] Multi-source data fusion

[0188] Power text divides information into different forms according to the source, recording format, and time and space characteristics of the information, such as historical data and real-time data, sensor data and social data, fault phenomenon document description and monitoring indicator data, etc. Considering the flexibility of machine learning models and the complexity of heterogeneous data fusion itself, data layer fusion, feature layer fusion, and decision layer fusion are performed respectively.

[0189] Graph Neural Networks

[0190] Refers to a neural network algorithm that acts directly on the graph structure. Its core is to sample adjacent node information at each layer and continuously update the node status to obtain a state containing adjacent node information and graph topology structure, and output the required results in a specific way.

[0191] Rules Engine

[0192] Refers to a rule-based reasoning engine that matches facts, data with production rules, handles conflicting rules, and finally executes corresponding instructions according to business rule outputs. Its core is to acquire knowledge and then apply the acquired knowledge to specific data.

[0193] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0194] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. Electric power text knowledge discovery method based on multi-source data fusion, It is characterized in that The following steps are involved: The first step is to obtain power natural language data and monitoring indicator data; The second step is to build a text data extraction model based on the power natural language data in the first step; The text data extraction model is used to process the natural language data of electric power and extract key information to obtain key feature information; the text data extraction model extracts and initially judges the key indicator information features of the description of the equipment operation phenomenon in the electric power text to obtain a triple including equipment, equipment status type, and equipment status parameter; its specific construction method is as follows: S1: Preliminary extraction of key features of power natural language data to obtain preliminary extraction results; S2: Standardize and generalize the preliminary extraction results in S1 to obtain feature enhanced data; S3: judging the device status parameters according to the feature enhancement data in S2; Build a rule engine judgment model based on historical monitoring indicator data; The rule engine judgment model is used to analyze the monitoring indicator data based on the rule engine, compare the monitoring indicator data with the comparison rules, obtain faulty equipment information, and quickly locate the faulty components; The rule engine judgment model uses business criteria and national standards to establish a rule engine system to judge the monitoring indicator data; it specifically includes the following steps: Step 1: Input the monitoring indicator data into the rule base of the rule engine pre-stored in the device memory in the form of wireless transmission; Step 2: Use the pattern matcher to compare the rules in the rule base with the monitoring indicator data in step 1; Step 3: If the monitoring indicator data successfully matches the rules of the rule base in step 2, the corresponding diagnosis result is output; otherwise, the monitoring indicator data is re-matched with the rules of the rule base until all the rules are matched; The third step is to build a graph data mining model to analyze the key feature information in the second step and obtain the fault information described in the text of the equipment operation phenomenon in the power text; The graph data mining model matches the key indicator information in the power natural language data with the graph structure to form a power heterogeneous graph to diagnose the operating status of the power equipment; The corresponding relationships include: equipment, components, and indicators correspond to nodes in the graph structure, and the association relationships between equipment and components, and between components and indicator nodes correspond to edges in the graph structure; Among them, each node in the graph has different properties due to the different corresponding information content; The properties of edges vary depending on the nodes they connect to; The fourth step is to analyze the electric power natural language data and monitoring indicator data to be judged based on the graph data mining model in the third step and the rule engine judgment model in the second step, realize multi-source fusion of at least two data types, and determine the fault location and cause.

2. The electric power text knowledge discovery method based on multi-source data fusion according to claim 1 is characterized in that: In S1: preliminary extraction of key features, specifically including the following contents: When performing triple recognition and extraction of power natural language data, the semantic association before and after the characters is taken into account, and the bidirectional semantic information of power natural language data is obtained; The method for obtaining the bidirectional semantic information is as follows: The bidirectional long short-term memory model is used to transmit the forward and reverse information of the power natural language data, and the hidden state sequence (h1,h2,...,h n )∈R n*m , and then connect a linear layer to convert the dimension of the hidden state sequence from m-dimension to k-dimension, and calculate the probability of the key feature information of the characters in the power natural language data; At the same time, the random conditional field is used to obtain the connection relationship between the character labels in the power natural language data to obtain the optimal key feature information extraction; Under the given condition X, the random conditional field has condition Y that satisfies the Markov property. Each feature template must slide over each position, and the score of the feature template for the entire sentence is obtained by adding them up. Finally, the weighted summation of multiple feature templates is performed to achieve the preliminary extraction of key features. The calculation formula is as follows: where λ k is the weight of the edge feature function, μ l is the weight of the point feature function, t k is the characteristic function on the edge, s l is the characteristic function at the node, λ k and μ l is the corresponding weight; Z(x) is the normalization factor, k is the serial number of the k-th state feature, and i is the serial number of the i-th character.

3. The electric power text knowledge discovery method based on multi-source data fusion according to claim 1 is characterized in that: In S2, the standardization and generalization processing specifically includes the following contents: We have carried out knowledge standardization and vocabulary sorting work on synonym relationships and hyponymy relationships; Based on the sorted synonym relationships and hyponymy relationships, the extracted key feature information is subjected to standardized synonym comparison and replacement.

4. The electric power text knowledge discovery method based on multi-source data fusion according to claim 1 is characterized in that: S3: Device status parameter judgment, specifically includes the following contents: S31, identifying the attributes and values ​​corresponding to the target named entity for any given power natural language data; S32, according to the attributes and values ​​in S31, obtain a triple of device, device state type, and device state parameter; S33, judging whether the device status parameter is normal or abnormal according to the triplet in S32.

5. The method for discovering electric power text knowledge based on multi-source data fusion according to claim 4 is characterized in that: The power heterogeneous graph includes two major types of nodes, one is a newly entered text node, the newly entered text node includes the power natural language data to be diagnosed, and the other is a word node, the word node includes the existing power text, and the power text includes equipment information, phenomenon information, and indicator information; The process of power equipment status diagnosis is as follows: A graph structure with the newly entered text node as the target node and the word node as the adjacent node, and the edge connecting the target node and the adjacent node represents the similarity between the two.

6. The method for discovering electric power text knowledge based on multi-source data fusion according to claim 5 is characterized in that: The graph data mining model uses a one-hot encoding representation method to encode the triples obtained by sequence labeling and extraction based on Bi-LSTM+CRF multi-channel fusion, and expresses the node information with a low-dimensional vector, which includes the following steps: Step 31: In order to enable the diagnostic algorithm to successfully process the graph structure information, a recurrent convolutional neural network method is used to vectorize the nodes in each graph and embed the graph structure; the recurrent convolutional neural network receives the input x at time t t After that, the value of the hidden layer is s t , and its calculation formula is as follows: st=f(Uxt+Wst-1) in, is the weight matrix of the input xt, The last value st-1 is used as the weight matrix for this input, and f is the activation function. If the loop is iterated, the output value o is obtained. t , and its calculation formula is as follows: Among them, V is the matrix of output weights; W is the weight matrix; Step 32: Encode the target node to be diagnosed with its adjacent nodes to obtain the unique hot encoding of each word, which specifically includes the following contents: First, we need to create an all-zero vector representing the total length of the vocabulary |V|. Then set the index dimension of each word in the vocabulary to 1, and keep other elements unchanged to get the final one-hot vector; After inputting the adjacent nodes, the vector output by the network is the embedding result of the target node to be diagnosed. Step 33: Use the attention-based graph convolutional network to iterate nodes, and use the attention mechanism to perform weighted summation of the features of adjacent nodes; the input of the graph attention layer is the node feature vector set: Where N is the number of nodes, F is the number of node features; the matrix dimension is N×F, representing the features of all nodes, and R represents the features of a certain node, and the matrix dimension is F×1; The corresponding node features are: orientation, fault phenomenon type, and state, and the value of F is 3; Step 34: Perform at least a linear transformation on the input node features to obtain the corresponding output features, and set the weight matrix W∈R for all nodes F′×F ; The initial state of the target node corresponding to the power natural language data to be diagnosed is set as The initial state of the adjacent nodes is set to The similarity between the target node and neighbor nodes in each layer is calculated through the power heterogeneous graph; Step 35: To measure the similarity between the adjacent node j in the power heterogeneous graph and the target node i of the power natural language data to be diagnosed, the graph data mining model uses the constructed power heterogeneous graph to perform self-attention mechanism operations on the target node i to be diagnosed and each of its adjacent nodes in turn. The attention coefficient between two nodes in the lth layer is expressed as: Where: f(﹒) is the local transformation function, which is used to reflect the calculation result of the similarity between the target node to be diagnosed and its adjacent nodes. This function is determined by the gradient descent strategy of the loss function of the pre-trained model; l＝1,2,3,…； The loss function of the pre-trained model is obtained through the supervised nodes. The calculation formula of the loss function is as follows: Where p(﹒) is the true probability distribution, and q(﹒) is the calculated predicted probability distribution; Step 36: Introduce the logistic regression Softmax function to the adjacent node set N of the target node i i All adjacent nodes j in are regularized. The specific method is as follows: The regularized inter-node attention coefficient of the lth layer obtained by the above operation is used to predict the output features of each node in the lth layer: In the formula, σ is a nonlinear activation function, and the linear rectification function ReLU is used as the activation function; Step 37: Perform similarity calculation on the features of the target node to be diagnosed obtained through iterative updating of the graph neural network and the diagnostic features of each operating state obtained through embedding, and calculate and output them through the logistic regression Softmax function.

7. The power text knowledge discovery device based on multi-source data fusion is characterized by: Applying the electric power text knowledge discovery method based on multi-source data fusion as described in any one of claims 1 to 6; It includes a touch screen, a processor, a power module, a text data extraction module, a graph data mining module, a rule engine judgment module, and a device memory; The text data extraction module, graph data mining module, and rule engine judgment module are all stored in the device memory; Text data extraction module, used to automatically process natural language and extract key information; Graph data mining module, used to analyze the text description information of equipment operation phenomena using graph neural network algorithm; The rule engine judgment module is used to analyze and judge the monitoring indicator data using the rule engine.

Images