CN112307218A

CN112307218A - Intelligent power plant typical equipment fault diagnosis knowledge base construction method based on knowledge graph

Info

Publication number: CN112307218A
Application number: CN202011131638.0A
Authority: CN
Inventors: 赵健程; 朱文欣; 高诗宁; 赵春晖
Original assignee: Zhejiang University ZJU
Current assignee: Zhejiang University ZJU
Priority date: 2020-10-21
Filing date: 2020-10-21
Publication date: 2021-02-02
Anticipated expiration: 2040-10-21
Also published as: CN112307218B

Abstract

The invention discloses a fault diagnosis knowledge base construction method for typical equipment of an intelligent power plant based on a knowledge graph. The method is directly oriented to the field of intelligent power plant typical equipment fault diagnosis, and a fault diagnosis knowledge map is designed and constructed by combining multi-mode fault diagnosis data from factories and the Internet with expert knowledge and stored in a knowledge base, so that the automation level of fault diagnosis is effectively improved. The tower-shaped knowledge map framework in the form of 'double-layer-three-element' is redesigned, and the retrieval and application are facilitated while the ideographic capability is strong. The invention establishes the description vector of the text in the knowledge graph without supervision by using the bidirectional GRU model, contains the semantic information of the text, can be used for optimizing the fault diagnosis knowledge graph, improves the reasoning and calculating efficiency, and has important significance for the application of the fault diagnosis knowledge graph on the ground.

Description

Intelligent power plant typical equipment fault diagnosis knowledge base construction method based on knowledge graph

Technical Field

The invention belongs to the field of operation fault diagnosis of power plant production equipment, and relates to a design and construction method of a fault diagnosis knowledge graph, an application scheme of a knowledge base and an updating strategy of the knowledge base.

Background

The intelligent power plant is provided under the background of information and industrial deep fusion, and aims to improve the intelligent level of the power industry and realize the technical improvement of unmanned inspection, automatic fault diagnosis and treatment, big data analysis, intelligent control and the like. The fault diagnosis and treatment are crucial to maintaining the stable power generation process and ensuring the production safety. The key to fault diagnosis of the intelligent power plant key equipment is to construct a fault diagnosis knowledge base to realize automatic and reliable fault diagnosis. By collecting and analyzing fault cases and utilizing the leading-edge technologies such as knowledge graph and the like, natural language is effectively processed, a fault diagnosis knowledge base is constructed, and a foundation can be provided for realizing fault diagnosis of typical equipment of an intelligent power plant.

The knowledge graph technology is applied to a search engine of the Google company in 2012 first, so that the quality of the search result is greatly improved. In recent years, the knowledge graph and various industries are deeply integrated, and the aims of solving the problem of industrial pain points and reducing the labor cost are fulfilled. For example, in the medical field where clinical data is relatively limited, a corresponding medical knowledge map is also established for diagnosing common pediatric diseases and partial critical illness, and the clinical average accuracy of the AI diagnosis model on the pediatric diseases reaches 90%, and the performance of the AI diagnosis model can be comparable to that of low-annual-quality main doctors. By constructing the fault diagnosis knowledge base, automatic, reliable and efficient fault diagnosis in the daily operation process of the intelligent power plant can be realized, so that a highly intelligent decision is formed, and specific and effective guidance suggestions are actively provided for the power plant operators. In the next place, most accumulated fault diagnosis knowledge is unstructured data, is difficult to directly apply, and needs to be combed into a knowledge graph form with all parts organically connected, so that the method has important significance on storage, retrieval, reasoning and application of knowledge.

In recent years, the language processing technology is developed vigorously, and the performance of supervised tasks such as emotion analysis and text summarization is increasing. The end-to-end machine translation model of seq2seq and the like also improves the current machine translation effect remarkably. However, for feature extraction of industrial fault diagnosis texts, labels are often absent or obtained quite expensive. In addition, the existing fault diagnosis knowledge graph method usually takes a certain device or a specific fault term as a node, and several specific types such as parameters, representations and reasons are taken as edges to construct a triple, and the method is limited by the ideographic capability of the triple, and is limited in the capability of describing complex reasons and complex solutions. Therefore, the invention adopts the idea of self-supervision learning and combines an encoder-decoder model to obtain the feature vectors of the texts in the nodes in the knowledge graph, thereby providing support for the task needing efficient calculation and reasoning later.

Disclosure of Invention

The invention aims to apply machine learning and knowledge graph technology to the field of fault diagnosis of typical equipment of an intelligent power plant, construct an intelligent power plant fault diagnosis knowledge base by designing a knowledge graph framework aiming at the field of fault diagnosis, and provide support for knowledge storage, retrieval, reasoning, application and updating related to the fault diagnosis of the intelligent power plant.

The purpose of the invention is realized by the following technical scheme: a method for constructing a fault diagnosis knowledge base of typical equipment of an intelligent power plant based on a knowledge graph comprises the following steps:

step 1) raw data is collected. Sources of data include the internet and cooperating power plants. The data quality of the cooperative power plant is high but the quantity is small; the data on the internet is large but the quality is poor. A sufficiently abundant data source can on the one hand increase the size of the knowledge base and on the other hand also provide sufficient training samples for later training.

And 2) performing targeted preprocessing on the multi-modal data, and converting non-text data into text data.

And 3) processing the text data to construct a knowledge graph of double-layer and three-element. "bilayer" refers to a device layer, a failure layer. The method is convenient for the floor application of fault diagnosis and is also beneficial to using equipment to retrieve faults. The device layer is constructed on the basis of keywords extracted by an expert, a domain term dictionary and a TF-IDF algorithm. The failure layer includes the "three elements" of failure diagnosis: fault description, fault diagnosis and opinion processing. Thereby obtaining a fault diagnosis knowledge map;

step 4) further processing the fault description text, the fault diagnosis text and the processing opinion text, including sentence segmentation, word segmentation and BPE processing, and constructing a training set for extracting text characteristics of the bidirectional GRU network;

and 5) constructing and training an encoder-decoder model based on the bidirectional GRU network and the attention mechanism, and obtaining the feature vector of the label-free text from the state output by the encoder. Freezing the trained network parameters and storing the obtained feature vectors;

and 6) aligning the text description generated by the fault site diagram and the process data with the fault description of the original text data by applying the obtained feature vectors and combining key terms provided by the equipment layer and the field dictionary. The redundancy of the existing knowledge base can be eliminated, or reasoning and retrieval are carried out based on the fault diagnosis knowledge map;

after the processes of adding fault diagnosis knowledge, adopting 2), 3) and 4) are processed, if no key terms which are not included in the dictionary appear, the parameters stored in 5) are used to obtain an encoding result, a decoder is used to obtain a decoding result by combining a column search algorithm, the decoding result is compared with the atomic sentence, if the decoding result is consistent with the atomic sentence, the new knowledge is merged into the original knowledge map through inspection, and the update of the knowledge map is realized.

Further, the step 2) is specifically as follows:

aiming at the frequent faults which are enough for the fault site map corresponding to the fault description, generating the fault description by adopting an image text generation technology based on GAN (generic object model); and for the accidental faults with few samples, manually generating a fault description text. For the production data which are collected by the sensor and contain the fault time interval, under the condition of knowing the normal range of the data, the threshold range of 3-Sigma of the normal data is determined, and the abnormal point of the time sequence data is detected, and the data which exceed the range are attributed to the abnormal data. And converting the high early warning or the low early warning of the abnormal data variable and the related variable thereof into a text-form fault description.

Further, the step 3) is specifically as follows:

firstly, fault description, fault diagnosis and treatment opinions are extracted from text data. For the equipment layer, the equipment layer is divided into two sub-layers, the name of the whole typical equipment in the power plant is a top layer node, the specific part of the typical equipment is a bottom layer node, and the name of the whole typical equipment and the corresponding specific part are connected in a 'containing' relationship. Besides the keywords provided by the domain dictionary and the expert, the keywords extracted by the TF-IDF algorithm are introduced, and stop words are removed from the result, as follows:

TF-IDF＝TF*IDF#(20)

wherein TF is the word frequency and IDF is the inverse document frequency. A corpus refers to a collection of collected failure description documents. And taking the word with the TF-IDF value ranked higher as the candidate node of the equipment layer.

The fault layer takes a fault description text, a fault diagnosis text and a processing opinion text as nodes, and the corresponding fault description text and the corresponding fault diagnosis text are connected by taking 'diagnosis' as a relation; and connecting the corresponding fault description text and the processing opinion text by taking 'processing' as a relation to form a fault description layer. And then connecting the fault description text nodes in the fault layer with the specific part nodes involved therein, thereby forming a tower-shaped fault diagnosis knowledge graph framework of two layers and three elements.

Further, the step 4) is specifically as follows: firstly, the text is divided into clauses, and the text is divided into sub-sentences from comma, period and colon. For each clause, word segmentation is performed using a jieba (jieba) segmentation tool. And then carrying out BPE processing to obtain the clauses and the dictionary after the BPE processing, and taking each processed clause as a single training sample.

Further, the step 5) specifically comprises: a Bi-directional circulating Neural Network (Bi-directional circulating Neural Network) using GRUs (gated circulating units) is constructed using an encoder-decoder framework and an attention mechanism. And self-monitoring and acquiring the feature vector of each clause by taking each clause as a source sentence and a target sentence simultaneously. Wherein the GRU specifically includes:

firstly, each word in each clause sample used at the input end of an encoder is converted into a one-dimensional vector by adopting one-hot coding, the length of each vector is the same as the size of a dictionary obtained by BPE processing, wherein only the position corresponding to the word is 1, and the rest positions are 0. And then using an embedding layer to perform dimension reduction mapping, wherein the size of a mapping matrix is K × V, K is the dimension of a set word vector, and V is the size of a dictionary. Multiplying the mapping matrix with a matrix formed by the clause unique hot coding to reduce the dimension of the word vector, and obtaining a word vector group x ═ { x ═¹,x²,…,x^t,…,x^TAnd T represents the corresponding word number in the clause. In the model training process, the expected output sentences corresponding to each sentence sample are processed by the same method to obtain u ═ { u ═ u { (u) }¹,u²,…,u^t,…,u^T}。

Passing through the last moment state h^t-1And input x of the current node^tTo obtain reset gate r and update gate z:

r＝sigmoid(w^r*[h^t-1,x^t]) (21)

z＝sigmoid(w^z*[h^t-1,x^t]) (22)

wherein x ═ { x ═ x¹,x²,…,x^t,…,x^TFor each oneThe word vector group obtained by the clause sample through the above-mentioned one-hot coding and mapping process, t being the current time, represents the position of the currently input word in the clause. The sigmoid function maps values to the range 0-1. w is a^r、w^zAre all parameters that need to be learned.

Then obtaining the state and output of the current time

Where w is the parameter to be learned. h is^tIs the output of the current time step.

For the encoder part, the bidirectional recurrent neural network processes the input sequence in the forward and backward directions, respectively, in the time dimension, and concatenates the outputs at each time step as the final feature vector output.

Wherein x^tIs a word vector obtained by dimensionality reduction after single hot coding,

is a non-linear activation function.

For the decoder part, an attention mechanism is applied, at each instant, based on the context vector c of the t-th word calculated by equation (29)^tEyes of peopleT-th word vector u of target sequence^tAnd hidden state z at time t^tCalculating the next hidden layer state z^t ⁺¹：

Wherein the weight a_ijAnd the attention of the target word i to the source word j is shown, and align is an alignment model used for measuring the matching degree of the target word i to the source word j.

Will z^t+1Obtaining the probability distribution p of the t +1 th word of the target sequence through softmax normalization^t+1The cost of t +1 is obtained by using a cross entropy function, and the total loss function is obtained by averaging all the moments:

p^t+1＝softmax(w_sz^t+1+b) (32)

where avg is the averaging function and cross _ entropy is the cross entropy function. w is a_sAnd b is a parameter to be learned. Freezing the trained network parameters and storing the obtained feature vectors;

further, the step 6) is specifically as follows: using the encoder obtained by training in the step 5) to encode the text description generated by the fault field diagram and the process data to obtain a feature vector, and calculating the cosine similarity with the feature vector of the existing fault description text as follows:

a, B respectively represent the feature vectors of newly calculated and existing fault description texts, and the dimensions of the two vectors are the same. n represents the dimensions of the a and B vectors.

And aligning the group with the highest similarity. Similarity is calculated pairwise through feature vectors of existing fault description texts, fault description nodes with high similarity can be combined, and redundancy is eliminated, so that an intelligent power plant typical equipment fault diagnosis knowledge base is obtained for subsequent application.

Further, the method also comprises a GUI interface step for constructing the industrial floor application, and the GUI function comprises coal mill fault diagnosis, inquiry history, recent overhaul condition and the like.

Further, the fault diagnosis of the coal mill can be carried out through fault description or sensor data, and a fault diagnosis result, a maintenance suggestion and a maintenance fault map are returned, specifically: and after the characteristic vector is obtained by using an encoder, carrying out similarity comparison in the knowledge graph, and returning a fault diagnosis text, a processing opinion text and a fault graph corresponding to the fault description text with the highest similarity.

Further, for the newly added fault diagnosis knowledge, the text of the newly added fault diagnosis knowledge is encoded to obtain a characteristic vector, and then the characteristic vector is obtained, and the column search algorithm is adopted for decoding. During decoding, p is continuously passed^t+1Sample u^t+1. For the column search algorithm, a breadth-first strategy is used for establishing a search tree, nodes are sorted according to the sum of log probabilities of generated words as heuristic cost in each layer of the tree, and then only a predetermined number of nodes are left until a sentence end mark is obtained or the maximum generated length is exceeded.

The knowledge base in the knowledge map form is designed and constructed for the application scene of the fault diagnosis of the typical equipment of the power plant, and unstructured multi-modal data are unified into the structured knowledge map form. And the bidirectional GRU network is used for coding the text description in the nodes in the knowledge graph spectrum, so that a foundation is provided for high-performance reasoning and application tasks based on the knowledge graph. The method has important significance for improving the fault diagnosis of the typical equipment of the intelligent power plant.

Drawings

FIG. 1: the invention is a schematic flow chart;

FIG. 2: a multi-modal data schema;

FIG. 3: an example graph of abnormal differential pressure of a grinding bowl of a coal mill D; wherein a and b are schematic short-period diagrams, wherein b is provided with an early warning mark, and c is a schematic long-period diagram;

FIG. 4: the method comprises the following steps of (1) short-period air-powder mixture pressure (a), primary air pressure (b), a coal mill current diagram (c), long-period air-powder mixture pressure (d), primary air pressure (e) and a coal mill current diagram (f) of a coal mill;

FIG. 5: the 'double-layer-three-element' knowledge map architecture picture;

FIG. 6: encoder-decoder architecture diagram based on bi-directional RNN;

FIG. 7: GRU unit structure diagram;

FIG. 8: a fault cause tree diagram;

FIG. 9: the GUI functional schematic diagram is shown, wherein a is a main interface, b is coal mill fault diagnosis, and c is coal mill fault query history; d is the recent overhaul condition of the coal mill.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings and specific examples.

The invention collects the power generation process fault diagnosis description cases provided by the Internet and a power plant in Zhejiang province, and the power generation process fault diagnosis description cases comprise multi-mode data such as text, images, sensor data in numerical form and the like. By organizing unstructured data into a knowledge base in the form of a structured knowledge map according to specific structures, the automatic fault diagnosis capability is improved.

The intelligent power plant typical equipment fault diagnosis knowledge base construction method based on the knowledge graph comprises the following steps:

step 1) raw data is collected. Sources of data include the internet and cooperating power plants. The fault cases provided by the cooperative power plant are high in quality, are files in a word or pdf form, are in a semi-structured form, contain fault description, fault diagnosis, treatment opinions, matched data, field diagrams and the like, and can be directly extracted. The captured fault descriptions on the internet are used to further augment the knowledge base. A sufficiently abundant data source can on the one hand increase the size of the knowledge base and on the other hand also provide sufficient training samples for later training.

Step 2) performing targeted preprocessing on the multi-modal data, as shown in fig. 2, converting non-text data into text data, and training an image text generation technology based on GAN to generate fault description for data in a picture form if the number of corresponding picture samples under the same fault type is large; and for the accidental faults with few samples, manually generating a fault description text.

In the process of constructing the knowledge graph, aiming at production data which are collected by a sensor and contain fault periods, under the condition that the normal range of each data variable is known, the threshold range of 3-Sigma of each normal data variable is determined, the abnormal point of time series data is detected, and the data variables which exceed the range are attributed to the abnormal data variable. And (4) calculating correlation coefficients of a certain abnormal data variable and other variables pairwise, and considering the variables with the highest correlation coefficients as the correlation variables of the abnormal data variable. And converting the high early warning or the low early warning of the abnormal data variable and the related variable thereof into a text-form fault description.

The 3-Sigma principle is defined as follows: assuming that a group of detection data only contains random errors, calculating and processing the original data to obtain a standard deviation, then determining an interval according to a certain probability, and considering that the error exceeding the interval belongs to an abnormal value. The use of 3-Sigma is premised on the data being normally distributed, and after this condition is met, 99.73% of the data is normal within the 3-Sigma range (μ -3 σ, μ +3 σ), where σ represents the standard deviation, μ represents the mean, and x ═ μ is the axis of symmetry of the graph.

For example, for a fault with an example of "abnormal differential pressure of grinding bowl of coal pulverizer D", as shown in fig. 3, the correlation coefficients of the fault measuring point and other measuring points are calculated using the data during normal periods, and the most relevant are three variables of "selected air-powder mixture pressure at outlet of coal pulverizer D", "primary air pressure of coal pulverizer D" and "# 6D coal pulverizer current", as shown in the following table.

TABLE 1 correlation coefficient table with points of failure

As shown in fig. 4, a graph of a differential pressure high warning curve of a grinding bowl of a coal mill D, a graph of a pressure of a wind-powder mixture after selection of an outlet of the coal mill D, a graph of a primary wind pressure of the coal mill D, and a graph of a current of a #6D coal mill and corresponding warning curves are shown in fig. 4. Using the 3-Sigma principle, for the test, the first 30000min is taken as the normal time, and the data in the time period does not fluctuate greatly. Then μ -2.519 and σ -0.816, so the upper and lower limits of the pulverizer D grinding bowl differential pressure threshold are 4.966 and 0.0720, respectively, and the current data is found to exceed the upper limit, as indicated by the dotted region identified in the figure. The other three variables were also examined and found not to be overrun.

Then for that fault, the fault description generated by the data is then: the pressure difference of a grinding bowl of the coal mill D is early-warned, the pressure of the air-powder mixture selected by the outlet of the coal mill D is normal, the primary air pressure of the coal mill D is normal, and the current of the coal mill is normal.

And 3) processing the text data, and constructing a knowledge graph of 'double-layer-three-element' as shown in the attached figure 5. As shown in FIG. 5, "bilayer" refers to a device layer, a failure layer. The failure layer includes the "three elements" of failure diagnosis: fault description, fault diagnosis and opinion processing. Thereby obtaining a fault diagnosis knowledge map.

For texts containing fault diagnosis descriptions collected by internet channels, carrying out primary positioning based on related words of cause-and-effect relationship, and manually extracting fault descriptions, fault diagnosis and treating opinions in a positioning area; the fault diagnosis text provided by the factory has a certain structure, and the parts of fault description, fault diagnosis and opinion processing are extracted.

According to the triple theory, two layers are respectively expressed in the form of G ═ E, R, S, where E represents a node in the knowledge graph, R represents a relation of the knowledge graph, and S represents a triple in the knowledge graph.

The equipment layer is divided into two sub-layers, the name of the whole typical equipment in the power plant is a top layer node, the specific part of the typical equipment is a bottom layer node, and the name of the whole typical equipment is connected with the corresponding specific part in a 'containing' mode. For example, a typical equipment "coal pulverizer" is connected in a "contained" relationship with its particular component "grinding bowl". In addition to being sourced from domain dictionaries and expert supplies, to prevent some infrequent terms from being missed, the TF-IDF algorithm was introduced to extract keywords and remove stop words in the results as follows:

TF-IDF＝TF*IDF (37)

wherein TF is the word frequency and IDF is the inverse document frequency. A corpus refers to a collection of collected failure description documents. And taking the word with the TF-IDF ranking at the top as the candidate node of the equipment layer.

The fault layer takes a fault description text, a fault diagnosis text and a processing opinion text as nodes, and the corresponding fault description text and the corresponding fault diagnosis text are connected by taking 'diagnosis' as a relation; and connecting the corresponding fault description text and the processing opinion text by taking 'processing' as a relation to form a fault description layer. And then connecting the fault description text nodes in the fault layer with the specific part nodes involved therein, thereby forming a tower-shaped fault diagnosis knowledge graph framework of two layers and three elements. The Neo4j database was used to store the knowledge map.

And 4) further processing the fault description text, the fault diagnosis text and the processing opinion text to construct a training set for extracting text features of the bidirectional GRU network. Firstly, the text is divided into sentences, and the text is divided from commas, periods and colons to form clauses. For each clause, word segmentation is performed using a jieba (jieba) segmentation tool. And then carrying out BPE processing to obtain the clauses and the dictionary after the BPE processing, and taking each processed clause as a single training sample.

And step 5) constructing a bidirectional circulation Neural Network (Bi-direction Current Neural Network) using GRUs (gated circulation units), and adopting an encoder-decoder framework and an attention mechanism as shown in the attached figures 6 and 7. And acquiring the feature vector of each clause in a self-supervision mode by taking each clause as a source sentence and a target sentence simultaneously. Wherein the GRU specifically includes:

firstly, each word in each clause sample used at the input end of an encoder is converted into a one-dimensional vector by adopting one-hot coding, the length of each vector is the same as the size of a dictionary obtained by BPE processing, wherein only the position corresponding to the word is 1, and the rest positions are 0. And then using an embedding layer to perform dimension reduction mapping, wherein the size of a mapping matrix is K x V, K is the dimension of an artificially set word vector, and V is the size of a dictionary. Multiplying the mapping matrix with a matrix formed by the clause one-hot encoding to reduce the dimension of the word vector, and obtaining a vector x ═ { x ═ x¹,x²,…,x^t,…,x^TAnd T represents the corresponding word number in the clause. In the model training process, the expected output sentences corresponding to each sentence sample are processed by the same method to obtain u ═ { u ═ u { (u) }¹,u²,…,u^t,…,u^T}。

By the node state h of the last moment^t-1And input x of the current node^tTo obtain reset gate r and update gate z:

r＝sigmoid(w^r*[h^t-1,x^t]) (38)

z＝sigmoid(w^z*[h^t-1,x^t]) (39)

wherein x ═ { x ═ x¹,x²,…,x^t,…,x^TAnd f, obtaining a vector group by each clause sample through the one-hot coding and mapping process, wherein t is the current time and represents the position of the currently input word in the clause. Sigma is sigmoid function, and the value is mapped to the range of 0-1. w is a^r,w^zAre learnable parameters.

Then obtaining the state and output of the current time

Where w is a learnable parameter. h is^tIs the output of the current time.

For the encoder portion, the bi-directional recurrent neural network processes the input sequence in the forward and backward directions, respectively, in the time dimension and concatenates the outputs at each time step for output.

is a nonlinear activation function, i.e., the operation of equations (38) - (41) above.

For the decoder part, an attention mechanism is applied, each time according toContext vector c of t words^tThe t-th word vector u of the target sequence^tAnd hidden state z at time t^tCalculating the next hidden layer state z^t+1：

p^t+1＝softmax(w_sz^t+1+b) (49)

wherein avg is the averaging function, cross_entropyIs a cross entropy function. w is a_sAnd b is a parameter to be learned. And training the network, freezing the trained network parameters, and storing the obtained feature vectors. The training set containing 43756 samples was constructed in the experiment, and the loss is finally reduced to 0.25, which shows that the feature vector extracted by the encoder is effective, because the decoder can accurately use the feature vectorThe eigenvectors reconstruct the original input.

And 6) for fault description nodes in the knowledge graph, using the feature vectors stored before, using cosine similarity for comparison, combining the fault description nodes with high similarity and basically consistent key terms, and eliminating redundancy of the existing knowledge base.

In the process of reasoning and retrieval, if the input is the fault description in the text form, the input text is encoded by using the encoder trained in the step 5), and the encoding not only contains the information of the key words, but also contains the Chinese language sequence information, the most similar text description can be retrieved in the knowledge map by using cosine similarity, so that the fault diagnosis, the processing opinion and the fault site map can be obtained. And the obtained fault diagnosis is searched as fault description again until the high similarity result cannot be searched, and as shown in figure 8, the results obtained by multiple times of searching are output in a tree form to realize multi-layer deep cause tracking. The related key equipment components are returned in the form of labels through the associated equipment layers, so that the carding summary is convenient for a user.

If a piece of data is input, the threshold generated by processing the data in the normal period by the 3-Sigma principle in the step 2) is used for detection, and therefore the data is converted into a fault description text.

Preferably, a GUI interface for industrial floor applications can also be designed. As shown in FIG. 9, the functions include coal mill fault diagnosis, historical queries, and recent service conditions. Fault diagnosis can be performed through fault description or sensor data, and a fault diagnosis result, a maintenance suggestion and a maintenance fault map are returned.

As another preferable scheme, after the process of steps 2), 3) and 4) for newly adding fault diagnosis knowledge is processed, if no key term which is not included in the dictionary appears, the parameters stored in step 5) are used to obtain an encoding result, a decoder is used to combine with a column search algorithm to obtain a decoding result, and in the decoding process, the p is continuously passed through^t+1Sample u^t+1Establishing a search tree by using a breadth-first strategy, and pairing nodes according to the heuristic cost of the sum of log probabilities of generated words in each layer of the treeThe points are sorted and then only a predetermined number of nodes are left until a sentence end marker is obtained or the maximum generated length is exceeded. And comparing the decoding result with the atomic sentence, and if the decoding result is consistent with the atomic sentence, checking to incorporate the new knowledge into the original knowledge map so as to update the knowledge map.

Claims

1. The method for constructing the intelligent power plant typical equipment fault diagnosis knowledge base based on the knowledge graph is characterized by comprising the following specific steps of:

1) raw data was collected. The method comprises the steps of providing multi-mode data such as a text containing power plant equipment fault diagnosis knowledge, fault data with a fault diagnosis label, a fault site map and the like;

2) performing targeted preprocessing on the multi-modal data, and converting non-text data into text data;

3) and processing the text data to construct a knowledge graph of 'double-layer-three-element'. "bilayer" refers to a device layer, a failure layer. The device layer is constructed on the basis of keywords extracted by an expert, a domain term dictionary and a TF-IDF algorithm. The failure layer includes the "three elements" of failure diagnosis: fault description, fault diagnosis and opinion processing. Thereby obtaining a fault diagnosis knowledge map;

4) further processing the fault description text, the fault diagnosis text and the processing opinion text to construct a training set for extracting text features of the bidirectional GRU network;

5) and constructing and training an encoder-decoder model based on the bidirectional GRU network and the attention mechanism, and obtaining the feature vector of the label-free text from the state output by the encoder. Freezing the trained network parameters and storing the obtained feature vectors;

6) and aligning the text description generated by the fault site diagram and the process data with the fault description of the original text data by using the obtained feature vector and combining key terms provided by the equipment layer and the domain dictionary to obtain the typical equipment fault diagnosis knowledge base of the intelligent power plant.

After processing the processes of adding fault diagnosis knowledge, adopting 2), 3) and 4), if no key terms which are not included in the dictionary do not appear, using a network constructed and trained in 5) to obtain a coding result, using a decoder to combine with a column search algorithm to obtain a decoding result, comparing the decoding result with an atomic sentence, and if the decoding result is consistent with the atomic sentence, checking to incorporate new knowledge into the original knowledge map to update the knowledge map.

2. The intelligent power plant typical equipment fault diagnosis knowledge base construction method according to claim 1, characterized in that: the step 2) is specifically as follows:

3. The intelligent power plant typical equipment fault diagnosis knowledge base construction method according to claim 1, characterized in that: the step 3) is specifically as follows:

TF-IDF＝TF*IDF (3)

wherein TF is the word frequency and IDF is the inverse document frequency. A corpus refers to a collection of collected failure description documents.

4. The intelligent power plant typical equipment fault diagnosis knowledge base construction method according to claim 1, characterized in that: the step 4) is specifically as follows: firstly, the text is divided into clauses, and the text is divided into sub-sentences from comma, period and colon. For each clause, word segmentation is performed using a jieba (jieba) segmentation tool. And then carrying out BPE processing to obtain the clauses and the dictionary after the BPE processing, and taking each processed clause as a single training sample.

5. The intelligent power plant typical equipment fault diagnosis knowledge base construction method according to claim 1, characterized in that: the step 5) is specifically as follows: a Bi-directional circulating Neural Network (Bi-directional circulating Neural Network) using GRUs (gated circulating units) is constructed using an encoder-decoder framework and an attention mechanism. And self-monitoring and acquiring the feature vector of each clause by taking each clause as a source sentence and a target sentence simultaneously. Wherein the GRU specifically includes:

each word in each clause sample for the input to the encoder is first converted to a one-dimensional vector using one-hot encoding,the length of each vector is the same as the size of the dictionary obtained by BPE processing, wherein only the corresponding position of the word is 1, and the rest positions are 0. And then using an embedding layer to perform dimension reduction mapping, wherein the size of a mapping matrix is K × V, K is the dimension of a set word vector, and V is the size of a dictionary. Multiplying the mapping matrix with a matrix formed by the clause unique hot coding to reduce the dimension of the word vector, and obtaining a word vector group x ═ { x ═¹,x²,…,x^t,…,x^TAnd T represents the corresponding word number in the clause. In the model training process, the expected output sentences corresponding to each sentence sample are processed by the same method to obtain u ═ { u ═ u { (u) }¹,u²,…,u^t,…,u^T}。

r＝sigmoid(w^r*[h^t-1,x^t]) (4)

z＝sigmoid(w^z*[h^t-1,x^t]) (5)

wherein x ═ { x ═ x¹,x²,…,x^t,…,x^TAnd f, obtaining a word vector group by each clause sample through the one-hot coding and mapping process, wherein t is the current time and represents the position of the currently input word in the clause. The sigmoid function maps values to the range 0-1. w is a^r、w^zAre all parameters that need to be learned.

Then obtaining the state and output of the current time

Where w is the parameter to be learned.

is a non-linear activation function.

For the decoder part, an attention mechanism is applied, at each instant, based on the context vector c of the t-th word calculated by equation (12)^tThe t-th word vector u of the target sequence^tAnd hidden state z at time t^tCalculating the next hidden layer state z^t+1：

p^t+1＝softmax(w_sz^t+1+b) (15)

where avg is the averaging function and cross _ entropy is the cross entropy function. w is a_sAnd b is a parameter to be learned. And freezing the trained network parameters and storing the obtained feature vectors.

6. The intelligent power plant typical equipment fault diagnosis knowledge base construction method according to claim 1, characterized in that: the step 6) is specifically as follows: using the encoder obtained by training in the step 5) to encode the text description generated by the fault field diagram and the process data to obtain a feature vector, and calculating the cosine similarity with the feature vector of the existing fault description text as follows:

7. The intelligent power plant typical equipment fault diagnosis knowledge base construction method according to claim 1, characterized in that: the method also comprises a GUI interface step of constructing the industrial floor application, wherein the GUI function comprises coal mill fault diagnosis, history inquiry, recent overhaul condition and the like.

8. The intelligent power plant typical equipment fault diagnosis knowledge base construction method according to claim 7, characterized in that: the fault diagnosis of the coal mill can be carried out through fault description or sensor data, and a fault diagnosis result, a maintenance suggestion and a maintenance fault map are returned, specifically comprising the following steps: and after the characteristic vector is obtained by using an encoder, carrying out similarity comparison in the knowledge graph, and returning a fault diagnosis text, a processing opinion text and a fault graph corresponding to the fault description text with the highest similarity.

9. The intelligent power plant typical equipment fault diagnosis knowledge base construction method according to claim 1, characterized in that: and for the newly added fault diagnosis knowledge, coding the text of the newly added fault diagnosis knowledge to obtain a characteristic vector, then obtaining the characteristic vector, and decoding by adopting a column search algorithm. During decoding, p is continuously passed^t+1Sample u^t+1. For the column search algorithm, a breadth-first strategy is used for establishing a search tree, nodes are sorted according to the sum of log probabilities of generated words as heuristic cost in each layer of the tree, and then only a predetermined number of nodes are left until a sentence end mark is obtained or the maximum generated length is exceeded.