A cloud operation knowledge graph construction method based on a lightweight pre-training model

Abstract

The present application relates to the technical field of cloud operation and maintenance knowledge graph, and specifically relates to a cloud operation and maintenance knowledge graph construction method based on a lightweight pre-training model, comprising the following steps: S1: obtaining cloud operation and maintenance reports and cloud operation and maintenance log information, and using a regular matching rule to generate a sample set from the obtained data; S2: performing data cleaning on the sample set, and using a BIO labeling strategy to perform text labeling; S3: using an ALBERT-LSTNet-CRF model to perform entity extraction on the labeled sample set; S4: constructing a triple according to the extraction result and storing it in a graph database, thereby completing the construction of the knowledge graph. The present application effectively solves the data sparsity and cold start problems in the field of cloud operation and maintenance by using an ALBERT-based lightweight pre-training model and a high-quality knowledge graph, and simultaneously improves the accuracy and efficiency of the operation and maintenance of the State Grid cloud platform.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cloud operation knowledge graph, and particularly relates to a cloud operation knowledge graph construction method based on a light pre-training model. BACKGROUND

[0002] With the rapid development of digitalization and cloud computing technology, the State Grid Corporation of China releases the State Grid cloud platform, which gradually becomes the core component of power grid digitalization construction. However, in the aspect of cloud platform operation and maintenance, although relying on various indicators, logs and request tracking information has a certain effect, these traditional methods still face many challenges, including data sparsity problem, lack of deep semantic analysis of data, and cold start problem of the algorithm itself.

[0003] Under this background, knowledge graph technology is attracting more and more attention. It provides a new solution path for information retrieval and data mining by constructing complex semantic relationships between entities. However, in the aspects of text entity recognition and information extraction, although the BERT model shows strong ability, it is limited in the application of professional fields such as cloud operation and maintenance due to its large parameter quantity and long training requirement. In contrast, although the light ALBERT model reduces the parameter quantity, there is no mature application case in the construction of cloud operation knowledge graph.

[0004] At the same time, the explosive growth of Internet information also makes the traditional search engine technology gradually expose its limitations. Knowledge graph technology, especially Google's related research since 2012, has shown great potential in information organization and retrieval.

[0005] In summary, in the face of the increasingly complex cloud platform operation and maintenance environment and the growing information retrieval demand, a new solution is urgently needed. SUMMARY

[0006] Based on the above purpose, the present application provides a cloud operation knowledge graph construction method based on a light pre-training model.

[0007] A cloud operation knowledge graph construction method based on a light pre-training model includes the following steps:

[0008] S1: Obtain cloud operation reports and cloud operation log information, and use regular matching rules to generate a sample set from the obtained data;

[0009] S2: Perform data cleaning on the sample set, and use the BIO labeling strategy for text labeling;

[0010] S3: Use the ALBERT-LSTNet-CRF model to extract entities from the labeled sample set;

[0011] S4: Construct triples based on the extraction results and store them in the graph database, thus completing the construction of the knowledge graph.

[0012] Furthermore, the acquisition of cloud operation and maintenance reports and cloud operation and maintenance log information specifically adopts the Internet crawler technology of the Selenium framework to directionally crawl publicly accessible websites and databases containing the keyword "cloud operation and maintenance", thereby obtaining relevant cloud operation and maintenance reports and cloud operation and maintenance log information.

[0013] Furthermore, the specific steps of generating samples from the acquired data using regular matching rules in S1 are as follows:

[0014] S11: creating a predefined regular expression rule set, the rule set including regular expressions for matching keywords and phrases;

[0015] S12: Scan the acquired cloud operation and maintenance reports and cloud operation and maintenance log information, and perform text matching using the regular expression rule set predefined in S11;

[0016] S13: Filtering out text fragments matching a predefined regular expression from the scanning results, and saving the text fragments as preliminary samples;

[0017] S14: Perform secondary screening and integration on the saved preliminary samples to generate a final sample set that can be used for subsequent model training.

[0018] Furthermore, the S2 specifically includes:

[0019] S21: Remove non-critical information such as special characters, spaces, and punctuation marks from the sample set to obtain pure text data;

[0020] S22: Decompose the clean text data into words or phrases using a word segmentation algorithm, and assign a unique identifier to each word or phrase;

[0021] S23: Apply the BIO tagging strategy to tag each word or phrase using a predefined tag set, where B represents the start of an entity, I represents the inside of an entity, and O represents a non-entity;

[0022] S24: Verify and correct the annotated words or phrases to ensure the accuracy and consistency of the annotations, thereby generating the final annotated sample set for model training.

[0023] Furthermore, the steps of using the ALBERT-LSTNet-CRF model to extract entities from the labeled sample set in S3 are as follows:

[0024] S31: The original input sentence first flows through the ALBERT layer for feature vectorization to obtain the final feature vector T;

[0025] S32: The generated feature vector T is then passed through the LSTNet layer for capturing and integrating context information.

[0026] S33: Sequence labeling is performed through the CRF layer to output the optimal entity label.

[0027] Further, the S31 specifically comprises:

[0028] S311: The original input sentence includes monitoring information abnormal alarm, system log error, public component class fault, distributed service bus normal, cloud platform public component library unable to link centralized cloud resource management system.

[0029] S312: The input layer obtains the collected information sentence of the original input sentence, which is processed into a word sequence w=(w1, w2, …, w n ) as the input of the ALBERT layer;

[0030] S313: The ALBERT layer obtains the subsequence w=(w1, w2, …, w n ), converts the word sequence into a vector sequence X required by the subsequent neural network, and X is as follows:

[0031]

[0032] In the formula, H is the vector dimension; E i t is the word embedding coding of the word sequence; E i p is the position information coding of the word sequence; E i s is the sentence information coding of the word sequence, and the above three are mapped to high dimension addition to obtain the input sequence of the ALBERT layer, and finally the complete features of the multi-layer Transformer are calculated to obtain the final feature vector T.

[0033] Further, the S32 specifically comprises:

[0034] S321: After the convolution layer obtains the feature vector T, it performs dimension increasing feature processing on it, and the input data matrix is processed by the kth convolution kernel to obtain the hidden layer h k :

[0035]

[0036] In the formula is the convolution operation;

[0037] S322: After the input data passes through the convolutional layer, the output matrix enters the recurrent layer and the recurrent skip layer. The recurrent layer and the recurrent skip layer use the RELU function as the activation function of the hidden layer. Specifically, at time t, the hidden state of the recurrent unit in the recurrent layer is:

[0038] r t =σ(x t W xr +h t-1 W hr +b r )

[0039] u t =σ(x t W xu +h t-1 W hu +b u )

[0040] c t =RELU(x t W xc +r t ☉(h t-1 W hc )+b c )

[0041] h t =(1-u t )☉h t-p +u t ☉c t

[0042] In the formula, ⊙ represents element-wise multiplication, σ is the sigmoid function, x t is the data input at time t, r t , u t , c t are the values of the reset gate, the update gate, and the candidate hidden state at time t, respectively, and W and b are the weights and biases of each module, respectively.

[0043] S323: The calculation process of the LSTNet recurrent skip layer is as follows:

[0044] r t =σ(x t W xr +h t-p W hr +b r )

[0045] u t =σ(x t W xu +h t-p W hu +b u )

[0046] c t = RELU(x t W xc + b t ) t-p W hc ) c )

[0047] h t = (1 - u t ) * h t-p + u t * c t

[0048] where p is the number of skip of hidden units, the final output vector sequence h w = [h1 h2 … h n ] is obtained after the recurrent layer and the recurrent skip layer.

[0049] Further, the S33 specifically comprises:

[0050] S331: the vector sequence hw ＝ [h1 h2 … h n ] of information enters into a conditional random field (CRF) layer for optimization, and the CRF formula is specifically as follows:

[0051]

[0052]

[0053] S332: a globally optimal label sequence according to the adjacent relationship is finally obtained in the output layer, and the optimal entity can be output.

[0054] Further, the S4 comprises:

[0055] S41: key entities including fault codes, fault types and cloud platform components are parsed and recognized from the optimal entity labels output from the CRF layer.

[0056] S42: a directed triple is constructed by using a predefined relationship template and combining the key entities recognized in the previous step, and each triple includes entities, entities and the relationship therebetween.

[0057] S43: nodes and edges are created in a graph database to represent the entities and the relationship in the triple, and for each new triple, it is checked whether the corresponding nodes and edges exist in the graph database, and if not, they are added.

[0058] S44: Apply the transaction management mechanism of the graph database to ensure that all newly added triples are persistently stored, thereby completing the construction of the knowledge graph.

[0059] Beneficial effects of the present invention:

[0060] By using a lightweight pre-trained model based on ALBERT, the present invention can achieve high accuracy on a smaller data set while greatly shortening the model training time. This advancement is particularly suitable for complex environments such as the State Grid Cloud Platform, which requires efficient and accurate acquisition and processing of operation and maintenance information.

[0061] By constructing a high-quality knowledge graph, the present invention provides a more comprehensive and semantically rich cloud operation and maintenance solution. This beneficial effect not only improves the accuracy of cloud platform operation and maintenance, but also provides strong information support for the implementation of more advanced operation and maintenance strategies, such as predictive maintenance and resource optimization. Therefore, the present invention has important application value in improving the efficiency and accuracy of cloud platform operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0063] Fig. 1 This is a flow chart of a method for constructing a cloud operation and maintenance knowledge graph according to an embodiment of the present invention;

[0064] Fig. 2 Schematic diagram of the ALBERT-LSTNet-CRF model according to an embodiment of the present invention;

[0065] Fig. 3 Schematic diagram of the Albert-MA-LSTNet model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0067] It should be noted that, unless otherwise defined, technical terms or scientific terms used in the present application shall be understood as having the usual meaning as understood by a person with ordinary skill in the art to which the present application pertains. The terms "first", "second", and the like used in the present application do not represent any order, number, or importance, but are only used to distinguish different components. The terms "include" or "contain" and the like mean that the elements or objects before the terms encompass the elements or objects listed after the terms and their equivalents, and do not exclude other elements or objects. The terms "connected" or "connected" and the like are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms "up", "down", "left", "right", and the like are only used to represent relative positional relationships, and when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0068] As shown in Figs. 1-3 A cloud operation knowledge graph construction method based on a lightweight pre-training model, comprising the following steps:

[0069] S1: Obtain cloud operation report and cloud operation log information, and use regular matching rules to generate sample sets from the obtained data;

[0070] S2: Data cleaning is performed on the sample set, and text labeling is performed using the BIO labeling strategy;

[0071] S3: Use the ALBERT-LSTNet-CRF model to extract entities from the labeled sample set;

[0072] S4: Construct a triple according to the extraction result and store it in a graph database, thereby completing the construction of the knowledge graph.

[0073] The cloud operation report and cloud operation log information are obtained by using the Internet crawler technology of the Selenium framework, and the publicly accessible websites and databases containing the keyword "cloud operation" are targeted to be crawled, thereby obtaining the relevant cloud operation report and cloud operation log information.

[0074] The specific steps for generating samples from the obtained data using regular matching rules in S1 are as follows:

[0075] S11: Create a predefined regular expression rule set, which includes regular expressions for matching keywords and phrases, such as regular expressions for matching "fault code", "fault type", and "cloud platform components";

[0076] S12: Scan the obtained cloud operation report and cloud operation log information, and use the predefined regular expression rule set in S11 to perform text matching;

[0077] S13: screening text fragments matching the predefined regular expression from the scanning results, and saving the text fragments as preliminary samples;

[0078] S14: performing secondary screening and integration on the saved preliminary samples to generate a final sample set that can be used for subsequent model training.

[0079] S2 specifically includes:

[0080] S21: removing special characters, blanks, and non-key information such as punctuation marks in the sample set to obtain pure text data;

[0081] S22: using a word segmentation algorithm to decompose the pure text data into words or phrases, and assigning a unique identifier to each word or phrase;

[0082] S23: applying a BIO (Begin-Inside-Outside) labeling strategy to label each word or phrase using a predefined label set, where B represents the beginning of an entity, I represents the inside of an entity, and O represents a non-entity;

[0083] S24: checking and correcting the labeled words or phrases to ensure the accuracy and consistency of the labeling, thereby generating a final labeled sample set for model training.

[0084] The steps of using the ALBERT-LSTNet-CRF model in S3 to perform entity extraction on the labeled sample set are as follows:

[0085] S31: The original input sentence is first input into the ALBERT layer for feature vectorization to obtain the final feature vector T;

[0086] S32: The generated feature vector T is then passed through the LSTNet layer to capture and integrate context information;

[0087] S33: Perform sequence labeling through the CRF layer to output the optimal entity label.

[0088] S31 specifically includes:

[0089] S311: The original input sentence includes monitoring information abnormal alarm, system log error, public component class fault, distributed service bus normal, and cloud platform public component library unable to link centralized cloud resource management system;

[0090] S312: The input layer obtains the collected information sentences of the original input sentence, which are processed into word sequences w=(w1, w2, …, w n ) as the input of the ALBERT layer;

[0091] S313: The ALBERT layer obtains a subsequence w = (w1, w2, …, w n ), and converts the word sequence into a vector sequence X required by the subsequent neural network. X is as follows:

[0092]

[0093] In the formula, H is the vector dimension; E i t is the word embedding coding of the word sequence; E i p is the position information coding of the word sequence; E i s is the sentence information coding of the word sequence. The above three are mapped to high-dimensional addition to obtain the input sequence of the ALBERT layer, and finally the complete features of the multi-layer Transformer are calculated to obtain the final feature vector T.

[0094] S32 specifically includes:

[0095] S321: After the convolution layer obtains the feature vector T, it performs dimension-increasing feature processing on it. After the input data matrix is processed by the kth convolution kernel, the hidden layer h k is obtained:

[0096]

[0097] In the formula is the convolution operation;

[0098] S322: After the input data is processed by the convolution layer, the output matrix enters the recurrent layer and the recurrent skip layer. The recurrent layer and the recurrent skip layer use the RELU function as the activation function of the hidden layer. Specifically, at time t, the hidden state of the recurrent unit in the recurrent layer is:

[0099] r t = σ (x t W xr + h t-1 W hr + b r )

[0100] u t = σ (x t W xu + h t-1 W hu + b u )

[0101] c t = RELU (x t W xc + r t · (h t-1 Whc )+b c )

[0102] h t =(1-u t )☉h t-p +u t ☉c t

[0103] Where ⊙ represents element-wise multiplication, σ is the sigmoid function, and x t is the data input at time t, r t 、u t 、c t are the reset gate, update gate, and candidate hidden state values ​​at time t, respectively. W and b are the weights and biases of each module, respectively. The loop skip layer will capture the correlation of long time spans, ensuring that the problem of gradient vanishing cannot accurately extract the correlation of sequences over long time spans. By analyzing historical information, the relevance of contextual information in long time series can be extracted.

[0104] S323: The calculation process of the LSTNet loop skip layer is as follows:

[0105] r t =σ(x t W xr +h t-p W hr +b r )

[0106] u t =σ(x t W xu +h t-p W hu +b u )

[0107] c t =RELU(x t W xc +r t ☉(h t-p W hc )+b c )

[0108] h t =(1-u t )☉h t-p +u t ☉c t

[0109] Where p is the number of skips in the hidden unit, which depends on the periodic pattern of the time series data. After the recurrent layer and the recurrent skip layer, the final output vector sequence is h w =[h1 h2 … hn ]。

[0110] S33 specifically comprises:

[0111] S331: the vector sequence h of information w = [h1 h2 … h n ] into the conditional random field CRF layer for optimization, so as to better predict the final output sequence by summarizing the context information between different layers before, and increase the amount of information required for the model to make good predictions, and the CRF formula is as follows:

[0112]

[0113]

[0114] S332: finally obtaining the globally optimal label sequence according to the adjacent relationship in the output layer, that is, outputting the optimal entity.

[0115] S4 comprises:

[0116] S41: from the optimal entity label output from the CRF layer, parse and identify the key entity, which includes the fault code, fault type and cloud platform component;

[0117] S42: using a predefined relationship template, combined with the key entity identified in the previous step, a directed triple is constructed, each triple includes an entity, an entity and their relationship, such as fault code → affects → cloud platform component;

[0118] S43: create nodes and edges in the graph database to represent entities and relationships in triples, for each new triple, check whether the corresponding node and edge exist in the graph database, if not, add;

[0119] S44: apply the transaction management mechanism of the graph database to ensure that all newly added triples are persistently stored, thereby completing the construction of the knowledge graph.

[0120] The ALBERT-MA-LSTNet model is used instead of the ALBERT-LSTNet-CRF model, which specifically comprises:

[0121] The ALBERT-MA-LSTNet model makes some changes based on the ALBERT-LSTNet-CRF model introduced in the previous section, and uses the ALBERT-MA-LSTNet model for entity extraction, and its structure diagram is as follows Fig. 3As shown, the original input sentence is converted into a word sequence through the input layer and enters the ALBERT layer. The role of this layer is to convert the sub-sequence into a vector sequence input into the convolutional neural network, and then the vector sequence obtains the final feature vector T by calculating the complete features of the multi-layer Transformer. Then the feature vector T enters the convolutional layer and the subsequent recurrent layer and recurrent skip layer, which can capture the language features of the text based on the context information, output the vector sequence considering the context information, and input the vector sequence into the self-attention layer to finally obtain the optimal relationship sequence, that is, the relationship between entities can be output. The specific steps are as follows:

[0122] According to the ALBERT-LSTNet-CRF model, the first two steps of ALBERT-MA-LSTNet are the same, which will not be described here.

[0123] Through the previous processing of the input sentence, the vector sequence O=[o1 o2 … o n ] considering the context information is obtained, and then relation extraction (RE) is performed on the vector sequence, which is one of the core tasks of knowledge extraction. The purpose is to identify the relationship between entities in the sentence. Relation extraction is the basis for downstream tasks such as building a cloud platform operation and maintenance knowledge graph and an intelligent operation and maintenance question and answer system.

[0124] Cloud platform operation and maintenance entity relation extraction is also based on the ALBERT pre-training model, and introduces a multi-head attention mechanism (Multi-head Attention) based on LSTNet. Its multi-head calculation and mapping are as follows:

[0125] Q=K=V=H

[0126]

[0127] head i =Attention(Q,K,V)

[0128] MultiHead=Concat(head1,...,head n )W o

[0129]

[0130] h t D =h t ′+W M MultiHead t +b

[0131] In the formula, Q, K, V are respectively query, key, value vectors in the attention function, H is all hidden layer unit vectors of the recurrent layer, Attention is the attention function, d k is the dimension of the key vector, is the output of the i-th attention head, Multihead represents the output after multi-head mapping, and Concat is a concatenation function, h t D is the prediction result at time k, that is, the output through the convolution layer, the recurrent layer, the recurrent skip layer and the self-attention layer.

[0132] The self-attention mechanism helps to give more weight to the key information in the relationship classification, reduces the weight of the interference meaningless information, thereby further improving the accuracy of the model feature extraction, and can perform weighted transformation on the state information sequence predicted by the MA-LSTNet, and automatically gives different weights according to the importance of information.

[0133] Finally, the extracted knowledge entity relationship T is output, and the whole relationship extraction task is completed, and finally if the two are related, the output Bool value is 1, and if the two are unrelated, the output Bool value is 0.

[0134] The present application is intended to cover all such alternatives, modifications, and variations as fall within the broad scope of the appended claims. Accordingly, any omission, modification, equivalent replacement, improvement, etc. made in the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A cloud operation knowledge graph construction method based on a lightweight pre-training model, characterized in that, The method comprises the following steps: S1: obtaining cloud operation and maintenance report and cloud operation and maintenance log information, and generating a sample set from the obtained data using a regular matching rule; S2: data cleaning is performed on the sample set, and text labeling is performed using a BIO labeling strategy; S3: an ALBERT-LSTNet-CRF model is used to perform entity extraction on the labeled sample set; S4: a triple is constructed according to the extraction result and stored in a graph database, thereby completing the construction of a knowledge graph; The specific steps of generating a sample from the obtained data using a regular matching rule in S1 are as follows: S11: a pre-defined regular expression rule set is created, which includes regular expressions for matching keywords and phrases; S12: the obtained cloud operation and maintenance report and cloud operation and maintenance log information are scanned, and text matching is performed using the pre-defined regular expression rule set in S11; S13: text segments matching the pre-defined regular expression are selected from the scanning results, and the text segments are saved as preliminary samples; S14: the saved preliminary samples are subjected to secondary screening and integration to generate a final sample set that can be used for subsequent model training; S2 specifically comprises: S21: special characters, blanks, and non-key information such as punctuation marks in the sample set are removed to obtain pure text data; S22: a word segmentation algorithm is used to decompose the pure text data into words or phrases, and each word or phrase is assigned a unique identifier; S23: a BIO labeling strategy is applied, and a pre-defined label set is used to label each word or phrase, wherein B represents the beginning of an entity, I represents the inside of an entity, and O represents a non-entity; S24: the labeled words or phrases are checked and corrected to ensure labeling accuracy and consistency, thereby generating a final labeled sample set for model training; The steps of using an ALBERT-LSTNet-CRF model to perform entity extraction on the labeled sample set in S3 are as follows: S31: the original input sentence is first input into the ALBERT layer for feature vectorization to obtain the final feature vector T; S32: the generated feature vector T is then input into the LSTNet layer for context information capture and integration; S33: sequence labeling is performed through the CRF layer to output the optimal entity label; S31 specifically comprises: S311: the original input sentence includes monitoring information abnormal alarm, system log error, public component class fault, distributed service bus normal, and cloud platform public component library unable to link centralized cloud resource management system; S312: The input layer obtains the collected information sentences of the original input sentence, which are processed into word sequences w=(w1, w2, …, w n ) as the input of the subsequent ALBERT layer; S313: The ALBERT layer obtains a sub-sequence w = (w1, w2, …, w n ), and converts the word sequence into a vector sequence X required by the subsequent neural network. X is as follows: In the formula, H is the vector dimension; E i t is to encode the word sequence for word embedding; E i p is to encode the word sequence for position information; E i s is to encode the word sequence for sentence information, the above three are mapped to high-dimensional addition to obtain the input sequence of the ALBERT layer, and finally the complete features of the multi-layer Transformer are calculated to obtain the final feature vector T; S32 specifically comprises: S321: After the convolution layer obtains the feature vector T, it performs dimension increasing feature processing on it. After the input data matrix is processed by the kth convolution kernel, the hidden layer h is obtained k is: In the formula is a convolution operation; S322: after the input data passes through the convolution layer, the output matrix enters the recurrent layer and the recurrent skip layer, and the recurrent layer and the recurrent skip layer use the RELU function as the activation function of the hidden layer. Specifically, at time t, the hidden state of the recurrent unit in the recurrent layer is: r t = σ(x t W xr +h t-1 W hr +b r ) u t = σ(x t W xu + h t-1 W hu + b u ) c t = RELU(x t W xc + r t ☉(h t-1 W hc ) + b c ) h t = (1 - u t )☉h t-p + u t ☉c t In the formula, ⊙ represents element-wise multiplication, σ is a sigmoid function, x t is the data input at time t, r t , u t , c t are the values of the reset gate, update gate, and candidate hidden state at time t, respectively, and W and b are the weights and biases of the respective modules. S323: the calculation process of the LSTNet recurrent skip layer is as follows: r t = σ(x t W xr +h t-p W hr +b r ) u t = σ(x t W xu + h t-p W hu + b u ) c t = RELU(x t W xc + r t ☉(h t-p W hc ) + b c ) h t = (1 - u t )☉h t-p + u t ☉c t In the formula, p is the number of skips of hidden units, and the final output vector sequence h is obtained through the recurrent layer and the recurrent skip layer w = [h1, h2…h n ] S33 specifically comprises: S331: The vector sequence h of information w = [h1, h2...h n ] enters a conditional random field CRF layer for optimization, and the CRF formula is as follows: S332: the globally optimal label sequence obtained according to the adjacent relationship is finally obtained in the output layer, and the optimal entity is output; S4 comprises: S41: From the optimal entity label output from the CRF layer, parse and identify the key entity, including fault code, fault type, cloud platform component; S42: Using the predefined relationship template, combined with the key entity identified in the previous step, construct a directed triple, each triple including entity, entity and the relationship between them; S43: Create nodes and edges in the graph database to represent entities and relationships in triples. For each new triple, check whether the corresponding node and edge exist in the graph database. If not, add them; S44: Apply the transaction management mechanism of the graph database to ensure that all newly added triples are persistently stored, thereby completing the construction of the knowledge graph; The cloud platform operation and maintenance entity relationship extraction is also based on the ALBERT pre-training model, and introduces a multi-head attention mechanism based on LSTNet. Its multi-head calculation and mapping are as follows: Q=K=V=H head i = Attention(Q, K, V) MultiHead = Concat(head1,..., head n )W o h t D = h t + w M MultiHead t + b In the formula, Q, K, V are respectively query, key, value vectors in the attention function, H is all hidden layer unit vectors of the recurrent layer, Attention is the attention function, d k is the dimension of the key vector, is the output of the i-th attention head, Multihead represents the output after multi-head mapping, and Concat is a concatenation function, h t D is the prediction result at time k, that is, the output through the convolution layer, the recurrent layer, the recurrent skip layer, and the self-attention layer.

2. The cloud operation knowledge graph construction method based on a lightweight pre-training model according to claim 1, characterized in that, The cloud operation and maintenance report and cloud operation and maintenance log information are obtained by using the Internet crawler technology of the Selenium framework to directionally crawl the publicly accessible websites and databases containing the keyword "cloud operation and maintenance", thereby obtaining the relevant cloud operation and maintenance report and cloud operation and maintenance log information.

Images