Construction of fault type of shield tunneling machine based on convolutional neural network and fault diagnosis method

Abstract

The application discloses a kind of based on convolutional neural network's shield machine fault type construction and fault diagnosis method, this method starts from semi-structured shield machine fault record text, formalizes shield machine fault record text data;Second, utilize regular expression to split fault record text, realize the structured storage of shield machine fault record text corpus;Third, do spectrum clustering to fault record text, analyze the common point of each fault cluster data and formulate suitable label, fourth, propose EDA and back translation serial mixed increase data enhancement method to do data enhancement to labeled fault record text, to increase data quantity, solve fault feature imbalance problem;Finally, realize the CNN shield machine fault text classification model based on word2vec, complete the intelligent classification of shield machine fault record text.The method of the application can be used for the structured storage and intelligent classification of shield machine fault text record text data.

Description

TECHNICAL FIELD

[0001] The present application relates to fault type construction and fault diagnosis, more particularly to a shield machine fault type construction and fault diagnosis method based on a convolutional neural network. BACKGROUND

[0002] Shield tunneling machine, referred to as shield machine. It is a special engineering machinery for soft soil tunneling. With the needs of China's urbanization construction, major cities have successively carried out subway construction projects. Shield method has been widely used in subway tunnel construction due to its small impact on the surrounding environment, fast tunneling speed, high engineering quality, safe operation, environmental protection, wide application range and other advantages. Due to the complexity of shield machine system (such as Figure 1 Shield machine fault data accumulated in engineering operation is generally saved in the form of semi-structured shield machine fault record text, but shield fault record data in the form of text is difficult to effectively store and analyze in corpus, which urgently needs to realize the structured storage and intelligent classification of shield fault text record data through statistical natural language processing technology (NLP) and text data analysis technology.

[0003] Currently, the storage and classification of shield machine fault record text mainly rely on manual operation, which has the following three shortcomings:

[0004] (A) Slow speed, manual operation for storage and classification of shield machine fault record text data is far lower than computer processing speed, which will consume a lot of time.

[0005] (B) Low accuracy, shield machine fault record text data can be classified into multiple categories according to different standards, and it is very difficult to correctly classify each fault text data due to the limitation of the knowledge level of the staff, and long-term repetitive work will reduce the judgment ability of the staff, further increasing the risk of incorrect classification.

[0006] (C) Poor consistency, the differences in cognition and knowledge of different staff will make the storage and classification of shield machine fault record data present diverse results, which is not conducive to subsequent text data analysis work.

[0007] In summary, manual processing of shield machine fault record text data cannot complete efficient structured and unified storage and classification at present. SUMMARY

[0008] In order to solve the problems of low efficiency and poor effect of manual processing of shield machine fault record text data, and the problem that manual processing of shield machine fault data cannot be automatically learned in the corpus, the present application provides a shield machine fault type construction and fault diagnosis method based on a convolutional neural network. The shield machine fault type construction method of the present application first divides the fault text into fixed items according to the content by using a regular expression (RE), and stores the structured fault classes; secondly, the spectral clustering (Spectral Clustering) algorithm is used to cluster the shield machine fault record text data, and different classified fault data clusters are obtained; thirdly, the various fault data clusters generated by clustering are labeled; fourthly, in view of the problem of unbalanced data amount of each fault data cluster, the Easy Data Augmentation (EDA) and back-translation serial hybrid enhancement method is used to enhance the label data of each type to the same amount, and a training data set is obtained; fifthly, the fault training data set is fed into a convolutional neural network (Convolutional Neural Networks, CNN) to train the fault text classification model, and a shield machine fault classification model FSM is obtained; finally, the shield machine fault classification model FSM is used to intelligently classify and store the shield machine fault record text data according to the type.

[0009] The shield machine fault type construction method based on the convolutional neural network of the present application comprises the following steps:

[0010] Step one, formalize the shield machine fault record text;

[0011] Any one shield machine fault record text TEXT i Records the relevant information of the shield machine fault; a fault record text TEXT i Records one fault content condition or multiple fault content conditions; therefore, a fault record text TEXT i Composed of one fault content or multiple fault contents;

[0012] Fault content, denoted as CONTENT; any one fault record text TEXT i The fault content existing in the fault record text is expressed in the form of a fault content set And

[0013] Step two, structured storage of shield machine fault record text based on regular expression;

[0014] The field content contained in the fault record text is denoted as fault-field content

[0015] The​

[0016] The fault-field content As an element of the regular expression construction content, the set fault-regular expression content is denoted as

[0017] Step three, spectral clustering analysis of shield machine fault record text data;

[0018] The spectral clustering based on Ncut is realized by calling sklearn.cluster.SpectralClustering in the class library of scikit-learn;

[0019] The input layer is the shield machine fault record text data set FTS={TEXT1,TEXT2,…,TEXT i ,…,TEXT m};

[0020] The parameters that need to be adjusted for spectral clustering based on Ncut are:

[0021] The dimension of spectral clustering is x;

[0022] The clustering of full connection method is affinity;

[0023] The kernel function parameter is gm;

[0024] In the output result of spectral clustering, the shield machine fault record text data set divided into the same fault cluster is expressed in the form of a set and denoted as {C}, and the division set of the same fault cluster is denoted as CTS x , and CTS x ={{C1},{C2},…,{C x}};

[0025] Step four, mark the shield machine fault;

[0026] The shield machine fault record text data contains multiple labels, and the label set is expressed in the form of a set and denoted as T_LABEL, and T_LABEL={LABEL1,LABEL2,…,LABEL k ,…,LABEL z};

[0027] For each fault cluster data in the clustering result CTS x ={{C1},{C2},…,{C x}}, analyze and summarize the common points of the shield machine fault record data in the same fault cluster, and set the corresponding LABEL for each fault cluster data; then the shield machine fault record text data set with complete labels is denoted as CTS_LABEL x,and

[0028] Step 5: Enhance the shield tunneling machine fault record text data by combining EDA and back-translation;

[0029] Step 501: Record a tagged tunnel boring machine fault text. As input, to avoid altering the text's label information during the enhancement process, the first step is to... Tag removal, divided into TEXT i and LABEL k Two parts;

[0030] Step 502, transfer the text i The data is input into the EDA enhancement module for data enhancement. The resulting enhanced data is then input into the translation enhancement module for secondary enhancement. The fault text data obtained after secondary enhancement is expressed as a set, AUG_TEXTS, where AUG_TEXTS = {TEXT...} i_1 ,TEXT i_2 ,…,TEXT i_19 ,TEXT i_ω}, where ω represents the number of faults in the secondary enhancement;

[0031] Step 503, remove the original label LABEL k The text augmented data of tagged tunnel boring machine fault records, added to each text in AUG_TEXTS, is expressed as a set. and

[0032] any The number of texts contained is denoted as N. x The target number of text data augmentations is denoted as N. goal , any The enhancement coefficient is denoted as η. x ,and The number of enhanced texts is denoted as N. aug And N aug =N x ·η x From the enhanced N aug N texts are randomly selected. goal One data point was discarded;

[0033] any The augmented data obtained after hybrid data augmentation is expressed in the form of a set.

[0034] For CTS_LABEL xCarry out mixed data enhancement, get labeled shield machine fault record text enhanced data set, record as AUG_LABEL x , and

[0035] The method for constructing the shield machine fault type has the advantages that:

[0036] ①The method divides the shield machine fault record text into contents such as item name, problem (fault) description, reason analysis and the like by using the regular expression RE, and facilitates structured storage.

[0037] ②The method adopts a clustering algorithm to cluster the shield machine fault record text data, and the spectral clustering algorithm has better adaptability than the traditional K-means algorithm in processing high-dimensional data, and also has smaller calculation amount.

[0038] ③The method adopts the exploratory data analysis EDA and back-translation mixed data enhancement method to obtain richer data, and improves the accuracy and robustness of the model.

[0039] ④The shield machine fault classification model FSM generated by the method can be used for shield machine fault classification.

[0040] ⑤The method can be extended for fault text retrieval, shield machine fault diagnosis and the like. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a structural block diagram of a shield machine system.

[0042] Figure 2 is a flowchart of the shield machine fault type construction based on the convolutional neural network of the application.

[0043] Figure 3 is a shield machine fault record text data sample screenshot.

[0044] Figure 3A is the structured shield machine fault record text data after the regular expression processing of the application.

[0045] Figure 4A is the spectral clustering DBI effect index of the application.

[0046] Figure 4B is the spectral clustering CH effect index of the application.

[0047] Figure 4C is the spectral clustering SC effect index of the application.

[0048] Figure 5 is the visualization result of the shield machine fault record text data after the spectral clustering of the application.

[0049] Figure 6 is the structural framework of the present application for intelligent classification of fault clusters using convolutional neural networks. DETAILED DESCRIPTION

[0050] The present application will be further described in detail below with reference to the accompanying drawings and examples.

[0051] The shield machine fault record text is denoted as TEXT, and a plurality of fault record texts TEXT are expressed in a set form as a fault record text data set FTS, and FTS = {TEXT1, TEXT2, …, TEXT i . m}.

[0052] TEXT1 represents the first fault record text.

[0053] TEXT2 represents the second fault record text.

[0054] TEXT i represents the i-th fault record text.

[0055] TEXT m represents the last fault record text.

[0056] In order to facilitate the description of the present application, the subscript i represents the identification number of the fault record text, and the subscript m represents the total number of fault record texts. The TEXT i is also referred to as any one fault record text.

[0057] Referring to Figure 3 , Figure 3A , the object processed by the method of the present application is any one shield machine fault record text TEXT i , and the subscript i represents the identification number of the fault record text. Each piece of TEXT i is a fault text from a shield machine production enterprise or during shield machine operation and maintenance.

[0058] Referring to Figure 1 , Figure 2 , a shield machine fault type construction method based on a convolutional neural network of the present application includes the following steps:

[0059] Step 1, formalization of the shield machine fault record text;

[0060] In the present application, referring to Figure 3 , Figure 3A , any one shield machine fault record text TEXT i records the relevant information of the shield machine fault. One fault record text TEXT irecorded one fault content or multiple fault contents. Thus, one fault record text TEXT i consists of one fault content or multiple fault contents.

[0061] In the present invention, the fault content is denoted as CONTENT; any one fault record text TEXT i contains a fault content, which is expressed in the form of a fault content set and

[0062] represents the first fault content belonging to TEXT i .

[0063] represents the second fault content belonging to TEXT i .

[0064] represents the jth fault content belonging to TEXT i . The subscript j represents the identification number of the fault content.

[0065] represents the last fault content belonging to TEXT i . The subscript n represents the total number of fault contents.

[0066] Step two, structured storage of the shield machine fault record text based on regular expressions;

[0067] Referring to the shield fault record text sample data shown in Figure 3 , Figure 3A by observing the shield machine fault record text, it is found that each text contains fixed field content.

[0068] In the present invention, the field content contained in the fault record text is denoted as fault-field content .

[0069] The fault-field content

[0070] project represents the project name, i.e. the project name to which the fault record belongs.

[0071] shield_num represents the shield machine number, i.e. the shield machine number in the fault record.

[0072] problem represents the problem (fault) name, i.e. the problem (fault) name in the fault record.

[0073] description represents the brief description, i.e. the brief description of the fault in the fault record.

[0074] analysis represents cause analysis, i.e. cause analysis of the fault in the fault record.

[0075] solution represents solution process and measures, i.e. solution process and measures of the fault.

[0076] partner represents main participants, i.e. main participants of the fault handling.

[0077] summary represents summary, i.e. summary of the fault record.

[0078] recorder represents recorder, i.e. recorder of the fault record text.

[0079] In the present application, the fault-field content As an element of the regular expression construction content, the set fault-regular expression content is recorded as

[0080] In the present application, the fault content set According to The formulated shield machine fault-regular expression is recorded as C_regex, and Wherein (.*) is the corresponding specific content in the shield machine fault content field.

[0081] The fault record text TEXT i is processed by the formulated regular expression C_regex to obtain a structured fault record text, recorded as TEXT i_正则 , i.e.

[0082] represents the content of the project name in TEXT i .

[0083] represents the content of the shield machine number in TEXT i .

[0084] represents the content of the problem (fault) name in TEXT i .

[0085] represents the content of the brief description in TEXT i .

[0086] represents the content of the cause analysis in TEXT i .

[0087] represents the content of the cause analysis in TEXTi The content of the process and measures in the middle.

[0088] The content of the text i The content of the main participants in the middle.

[0089] The content of the text i The content of the summary in the middle.

[0090] The content of the text i The content of the person recorded in the middle.

[0091] The structured shield machine fault record text data is as shown in Figure 3A .

[0092] In the present application, each fault record text in FTS={TEXT1, TEXT2,…, TEXT i ,…,TEXT m} is processed by steps one and two respectively.

[0093] Step three, spectral clustering analysis of shield machine fault record text data;

[0094] In the present application, spectral clustering based on Ncut is realized by calling sklearn.cluster.SpectralClustering in the class library of scikit-learn. The input layer is the shield machine fault record text data set FTS={TEXT1, TEXT2,…, TEXT i ,…,TEXT m}. The parameters that need to be adjusted based on Ncut spectral clustering are:

[0095] The dimension of spectral clustering is x.

[0096] The clustering of full connection method is affinity.

[0097] The kernel function parameter is gm.

[0098] In the present application, in the output result of spectral clustering, the shield machine fault record text data set divided into the same fault cluster is expressed in the form of set and marked as {C}, and the division set of the same fault cluster is marked as CTS x , and CTS x ={C1}, {C2},…, {C x}.

[0099] {C1} represents the set of shield machine fault record text data clustered into the first fault cluster.

[0100] {C2} represents a set of shield machine fault record text data clustered into a second fault cluster.

[0101] {C x} represents a set of shield machine fault record text data clustered into an xth fault cluster.

[0102] In the present application, for example, spectral clustering adopts a full connection method to establish a fault clustering network, and the affinity is a Gaussian kernel function rbf by default; since the dimension of FTS is unknown, the value range of x is set to 2-50, that is, x = [2, 50]; the value of the kernel function parameter is set to gm = [0.01, 0.1, 0.2, 0.3, 0.4, 0.5]; and the spectral clustering effect evaluation index curve is drawn according to the above parameter settings, as shown in Figure 4A 、 Figure 4B 、 Figure 4C

[0103] In the present application, the spectral clustering effect evaluation index includes:

[0104] Silhouette Coefficient, denoted as SC. SC combines the cohesion and separation of clustering, and the larger the SC value is, the better the clustering effect is.

[0105] Calinski Harabasz Score, denoted as CH. The smaller the intra-class distance is, the larger the inter-class distance is, the larger the CH value is, and the better the clustering effect is.

[0106] Davies-Bouldin Index, denoted as DBI. The smaller the DBI is, the smaller the intra-class distance is, and the larger the inter-class distance is.

[0107] According to Figure 4A 、 Figure 4B 、 Figure 4C It can be known that when x = 8 and gm = 0.5, the clustering effect is the best, and the clustering result is shown in Figure 5 Each circle point represents a shield machine fault record text, and different colors represent different fault categories.

[0108] In the present application, the input data of spectral clustering is FTS = {TEXT1, TEXT2, …, TEXT i , …, TEXT m}. The output cluster division is CTS8, and CTS8 = {{C1}, {C2}, {C3}, {C4}, {C5}, {C6}, {C7}, {C8}}.

[0109] {C1} represents a set of shield machine fault record text data clustered into a first fault class.

[0110] ​{C2} represents a set of shield machine fault record text data clustered into a second fault class.

[0111] {C3} represents a set of shield machine fault record text data clustered into a third fault class.

[0112] {C4} represents a set of shield machine fault record text data clustered into a fourth fault class.

[0113] {C5} represents a set of shield machine fault record text data clustered into a fifth fault class.

[0114] {C6} represents a set of shield machine fault record text data clustered into a sixth fault class.

[0115] {C7} represents a set of shield machine fault record text data clustered into a seventh fault class.

[0116] {C8} represents a set of shield machine fault record text data clustered into an eighth fault class.

[0117] Step four, marking shield machine fault;

[0118] In the present application, the label is denoted as LABEL.

[0119] The shield machine fault record text data contains a plurality of labels, and the set of labels is expressed as T_LABEL, and T_LABEL = {LABEL1, LABEL2, …, LABEL k ,…,LABEL z}.

[0120] LABEL1 represents a first type of shield machine fault label. For example, the label symbol of the cutterhead system fault is LABEL1.

[0121] LABEL2 represents a second type of shield machine fault label. For example, the label symbol of the main drive system fault is LABEL2.

[0122] LABEL k represents the kth type of shield machine fault label.

[0123] LABEL z represents the last type of shield machine fault label.

[0124] For the convenience of explaining the present application, the subscript k represents the identification number of the shield machine fault label, and the subscript z represents the total number of shield machine fault labels. The LABEL k is also referred to as any one type of shield machine fault label in the set of shield machine fault labels.

[0125] To the clustering result CTS x{{C1}, {C2}, …, {C x}}. x , and

[0126] represents that the label LABEL1 is assigned to each fault text in {C1}.

[0127] represents that the label LABEL2 is assigned to each fault text in {C2}.

[0128] represents that the label LABEL x is assigned to each fault text in {C x}.

[0129] In the present application, the shield machine fault record text data is labeled in the following manner: the corresponding label is input after each TEXT i , and a shield machine fault record text data with fault labels is obtained at an interval of a fault window width t, denoted as k , and

[0130] In the present application, each shield machine fault record text TEXT i in each fault cluster data in the clustering result CTS8={{C1}, {C2}, {C3}, {C4}, {C5}, {C6}, {C7}, {C8}} is labeled with the label LABEL k that conforms to the common points of the cluster data, and a labeled shield machine fault record text data set CTS_LABEL8 is obtained, and

[0131] represents that the cutterhead system fault label LABEL1 is assigned to each fault text in {C1}.

[0132] represents that the main drive system fault label LABEL2 is assigned to each fault text in {C2}.

[0133] represents that the propulsion and articulation system fault label LABEL3 is assigned to each fault text in {C3}.

[0134] ​The label LABEL4 representing the shield tail sealing system failure is assigned to each failure text in {C4}.

[0135] The label LABEL5 representing the assembly machine system failure is assigned to each failure text in {C5}.

[0136] The label LABEL6 representing the slag modification and slag discharge system failure is assigned to each failure text in {C6}.

[0137] The label LABEL7 representing the synchronous grouting system failure is assigned to each failure text in {C7}.

[0138] The label LABEL8 representing the measurement guiding system failure is assigned to each failure text in {C8}.

[0139] Step five, mixed enhancement of shield machine failure record text data based on EDA and back translation;

[0140] In the present application, when the number of shield machine failure record text data is small, and the CTS_LABEL x The data amount of each cluster in the present application is unbalanced, in order to improve the robustness of the classification model and prevent overfitting, the EDA and back translation serial mixed method is used to perform data enhancement on the labeled shield machine failure record data.

[0141] The EDA in the present application is a word-level data enhancement, which generates new data by changing some words in the text, including four sub-methods, namely synonym replacement (SR), random deletion (RD), random swap (RS) and random insertion (RI).

[0142] In the present application, one data is processed by the four sub-methods in EDA respectively to obtain four data, so as to realize data enhancement.

[0143] Although EDA can retain most of the text information, it has the disadvantage that the similarity between new and old data is too high, so the data enhanced by EDA is enhanced again by back translation method, so as to obtain data with higher richness.

[0144] Back-translation is a data augmentation method that first translates text from one source language into another target language, and then translates it back into the source language. Due to differences in grammar and expression habits between different languages, back-translation can ensure that the semantics of the text remain unchanged while retaining sufficient differences. In this invention, the Baidu translation engine is used to achieve mutual translation between Chinese and foreign languages ​​(such as English, French, Japanese, Russian, Korean, etc.).

[0145] In this invention, a single tunnel boring machine fault record text data can be enhanced by a combination of EDA and back-translation to obtain 20 enhanced data entries.

[0146] In this invention, the process for enhancing the text data of tunnel boring machine fault records is as follows: Figure 2 As shown, the specific steps are as follows:

[0147] Step 501: Record a tagged tunnel boring machine fault text. As input, to avoid altering the text's label information during the enhancement process, the first step is to... Tag removal, divided into TEXT i and LABEL k Two parts;

[0148] Step 502, transfer the text i The data is input into the EDA enhancement module for data enhancement. The resulting enhanced data is then input into the translation enhancement module for secondary enhancement. The fault text data obtained after secondary enhancement is expressed as a set, AUG_TEXTS, where AUG_TEXTS = {TEXT...} i_1 ,TEXT i_2 ,…,TEXT i_19 ,TEXT i_ω}, where ω represents the number of faults in the secondary enhancement. For example, the 20 fault text data obtained after secondary enhancement are represented as a set, AUG_TEXTS, and AUG_TEXTS = {TEXT...} i_1 ,TEXT i_2 ,…,TEXT i_19 ,TEXT i_20}

[0149] TEXT i_1 This indicates the first fault text in the secondary enhancement.

[0150] TEXT i_2 This indicates the second fault text in the secondary enhancement.

[0151] TEXT i_ω This indicates the last fault text in the secondary enhancement.

[0152] Step 503, remove the original label LABEL kEach text added to AUG TEXTS, the resulting labeled shield machine fault record text enhancement data is expressed in the form of a set And For example, the labeled shield machine fault record text enhancement data set

[0153] Is The first piece of data obtained after enhancement.

[0154] Is The second piece of data obtained after enhancement.

[0155] Is The last piece of data obtained after enhancement.

[0156] In the present application, the number of texts contained in any N x . The target number of text data enhancement is N goal , the enhancement coefficient of any η x , and The number of texts after enhancement is N aug , and N aug =N x ·η x . From the N aug texts after enhancement, N goal data is randomly retained, and the rest is discarded.

[0157] In the present application, any The enhanced data obtained after mixed data enhancement is expressed in the form of a set

[0158] In the present application, the CTS_LABEL x Mixed data enhancement is performed to obtain a shield machine fault record text enhancement data set with complete annotations, denoted as AUG_LABEL x , and

[0159] Is a set composed of shield machine fault record text enhancement data with a label of LABEL1.

[0160] Is a set composed of shield machine fault record text enhancement data with a label of LABEL2.

[0161] Is a set composed of shield machine fault record text enhancement data with a label of LABEL xa set of shield machine fault record text augmentation data sets composed of shield machine fault record text augmentation data.

[0162] For example, CTS_LABEL8 is mixed data augmented to obtain a shield machine fault record text augmentation data set with complete annotations, denoted as AUG_LABEL8, and

[0163] Fault type diagnosis by the CNN-based shield machine fault model

[0164] In the present application, the CNN-based shield machine fault model is denoted as FSM model.

[0165] In the present application, the overall architecture of the CNN-based shield machine fault text intelligent classification is as shown in Figure 6 The overall architecture of the shield machine fault model FSM is divided into a data processing layer, a model optimization layer and an intelligent classification layer.

[0166] In the present application, the uppermost layer of the framework is the data processing layer. The data processing layer realizes fault text data word segmentation by using the Jieba word segmentation tool, and converts the fault text vector recognizable and calculable by a computer by using word2vec, thereby realizing fault text data structured processing.

[0167] The middle layer of the framework is the model optimization layer. In the present application, the model optimization layer uses the convolutional neural network CNN to classify faults according to the fault sample data obtained by the data processing layer, and optimizes according to the CNN parameter characteristics. The optimized parameters mainly include the size of the convolution kernel, the number of convolution kernels, the selection of the activation function, the selection of the dropout parameter, and the strength of the regularization.

[0168] The bottom layer of the framework is the intelligent classification layer. In the present application, the intelligent classification layer mainly uses the intelligent classification model obtained by the model optimization layer to automatically classify the shield machine fault record text to be classified.

[0169] Step one of the diagnosis, input the shield machine fault record text, and perform formalized shield machine fault record text;

[0170] Input the fault record text TEXT new , and extract the fault content in the TEXT new , denoted as

[0171] Step two of the diagnosis, structured storage of the shield machine fault record text based on regular expressions;

[0172] According to and in combination with the shield machine fault-regular expression C_regex, the structured fault record text TEXT new_正则 is obtained.

[0173] Step three of the diagnosis, spectral clustering analysis of shield machine fault record text data;

[0174] According to the division set of fault clusters CTS x Spectral clustering processing is performed on the TEXT new_正则 to obtain the fault type

[0175] To illustrate in detail, the shield machine fault text content shown in Table 1 is listed.

[0176] Table 1 Amount of each label text data in CTS_LABEL8 and AUG_LABEL8

[0177]

[0178] In the present application, AUG_LABEL8 is divided into a training set and a test set in a ratio of 7:3.

[0179] In the present application, the convolution kernel size of the convolutional neural network is 3, 4, and 5; the number of convolution kernels is 128; the activation function is relu, and the dropout parameter is 0.5; and the regularization strength is 0.0.

[0180] The test results show that the improved CNN shield machine fault record text data classification model based on word2vec can accurately classify shield machine fault record texts: the precision is 99.9%, the F1 value is 99.8%, and the recall rate is 99.7%. The EDA and back-translation serial mixed enhancement method proposed in the present application solves the problem of small amount of shield machine fault record text data and imbalance of various types of data, and improves the robustness and generalization ability of the classification model.

[0181] In addition, the shield machine fault record text data structured expression method proposed in the present application can store semi-structured texts according to fault content items, which is convenient for fault storage and query.

Claims

1. A method for constructing a model of a fault type of a tunneling machine based on a convolutional neural network, characterized by The steps include: Step one, formalize the shield machine fault record text; Any one of the shield machine fault record text Records the relevant information of the shield machine fault; a failure record text records a failure content situation or a plurality of failure content situations; thus, a failure record text consists of a failure content or a plurality of failure contents; fault content, denoted as ; any one of the fault record texts The fault content existing in the fault record texts is expressed in a set form as a fault content set , and ; indicates the 1st fault content belonging to the 1st fault content; indicates the 2nd fault content belonging to the 2nd fault content; represents a first of the clause fault content; subscript represents an identification number of the fault content; denotes the last fault content belonging to ; the subscript denotes the total number of fault contents; Step two, regular expression-based structured storage of shield machine fault record text; The The field content is noted as fault-field content ; The ; represents the name of the project to which the fault record belongs; represents the shield machine number, i.e. the shield machine number in the fault record; represents a problem fault name, i.e. a problem fault name in the fault record; a brief description, i.e. a brief description of the fault in the fault record; cause analysis, which is an analysis of the cause of the failure in the failure record; representing a resolution process and measures, i.e. a resolution process and measures to the fault; represents the main actors, i.e. the main actors of the fault handling; represents a summary, i.e. a summary of the fault record; represents a recorder, i.e., a person who records the failure record text; The fault-field content is set to As an element of the regular expression build content, the fault-regular expression content set is noted as ; Step three, spectral clustering analysis of shield machine fault record text data; Spectral clustering based on the library of calls is implemented.​​ Input layer is a shield machine fault record text data set ; subscript represents the identification number of the fault record text, subscript represents the total number of fault record texts; Based on The parameters that need to be adjusted for spectral clustering are: The dimensionality of the spectral clustering is ; The clustering of the full connection method is ; The kernel function parameters are ; In the output result of the spectral clustering, the fault record text data sets of the shield machine divided into the same fault cluster are expressed in the form of a set and marked as , the division set of the same fault cluster is marked as , and ; a set of shield machine fault record text data clustered as a first fault cluster; a set of text data of shield machine failure records clustered as a second failure cluster; a shield machine fault record text data set of a fault cluster a shield machine fault record text data set of a fault cluster Step four, mark shield machine faults; The fault record text data of the shield machine contains multiple labels, and a set form is adopted to express a label set as , and ; represents a first type of tunneling machine failure label; represents a second type of tunneling machine failure label; representing a first tunneling machine failure label; represents the last type of tunneling machine failure label; The analysis is performed on each fault cluster data in the clustering results , and the common points contained in the same fault cluster shield machine fault record data are summarized, and the corresponding is set for each fault cluster data; then, the labeled shield machine fault record text data set is denoted as , and ; representative label impart each failure text; representative label impart each failure text; representative label impart each failure text; Step five, mixed enhancement of shield machine fault record text data based on EDA and back translation; Step 501: Record a tagged tunnel boring machine fault text. As input, to avoid altering the text's label information during the enhancement process, the first step is to... Label removal is divided into and Two parts; Step 502, inputting to the EDA enhancement module for data enhancement, obtaining once enhanced data input to the back-translation enhancement module for secondary enhancement, and obtaining the fault text data after secondary enhancement in the form of a set as , and , , representing the number of secondary enhanced fault entries. 1st failure text indicating secondary enhancement; a second failure text indicating a secondary enhancement; represents the last failure text of the secondary enhancement; Step 503, add the original label to each text in , and the obtained labeled shield machine fault record text enhancement data is expressed in the form of a set as , and ; is the first data obtained after enhancement; is the second data obtained after the enhancement; is the last piece of data obtained after enhancement; arbitrary The number of texts contained in the middle is denoted as ; the target number of text data augmentation is denoted as , the augmentation coefficient of arbitrary is denoted as , and ; the number of augmented texts is denoted as , and ; from the augmented texts, random data are retained, and the rest are discarded; arbitrary The enhanced data obtained after mixed data augmentation is expressed in the form of a set as ; right Hybrid data augmentation was performed to obtain a well-annotated text augmentation dataset of tunnel boring machine fault records, denoted as . ,and ; is a collection of text enhanced data sets composed of fault records for a tunneling machine labeled as is a collection of text enhanced data sets composed of fault records for a tunneling machine labeled is a collection of text enhanced data sets of shield machine failure records tagged as ​ The fault training data set is fed into a convolutional neural network to train the fault text classification model, and a shield machine fault classification model is obtained.

2. The method of claim 1, wherein the method is based on a convolutional neural network. The method for tagging the text data of tunnel boring machine fault records is as follows: In each... Then enter the matching and with fault window width At intervals of 1, a textual record of a tunnel boring machine fault with a fault label is obtained, denoted as 1. ,and .

3. A method for fault diagnosis using the model constructed by the method for constructing a model of a fault type of a tunneling machine based on a convolutional neural network according to claim 1, characterized in that The following steps are included: Diagnosis step one, input shield machine fault record text and formalize the shield machine fault record text; input failure record text and extract the failure content in the , denoted as ; Diagnosis step two, regular expression-based structured storage of shield machine fault record text; The The field content is noted as fault-field content ; The ; The fault-field content is set to As an element of the regular expression build content, the fault-regular expression content set is noted as ; According to and in combination with a fault-regular expression of a tunneling machine structured fault record text is obtained ; Diagnosis step three, spectral clustering analysis of shield machine fault record text data; According to the division set of fault clusters To Perform spectral clustering processing to obtain fault types .

4. The method of claim 3, wherein The following steps are included: the uppermost layer is the data processing layer; the data processing layer realizes fault text data segmentation through the Jieba segmentation tool, and converts it into a fault text vector that can be recognized and calculated by a computer through word2vec, realizing structured processing of fault text data.

Images