CN115408190A - Fault diagnosis method and device - Google Patents

Abstract

The invention provides a fault diagnosis method and device, and relates to the technical field of fault analysis. The method comprises the following steps: constructing a fault diagnosis network model; the fault diagnosis network model comprises a feature extraction network, a feature interaction network and a feature classification network; according to the work order sample data, the feature extraction network and the historical fault database, determining first feature data used for representing text semantics and second feature data used for representing a text fault theme; determining a trained fault diagnosis model according to the first characteristic data, the second characteristic data, the characteristic interaction network and the characteristic classification network; and performing similarity calculation according to the fault diagnosis model and the historical fault database, and determining corresponding work order processing information in the historical fault database. The scheme of the invention can complete the diagnosis of different cases, improve the accuracy of automatic fault diagnosis, and enhance the interpretability of model feature learning and the diagnosis adaptability to different fault subjects.

Description

Fault diagnosis method and device

technical field

The invention relates to the technical field of fault analysis, in particular to a fault diagnosis method and device.

Background technique

Massive fault records stored in manufacturing industries such as aerospace, automotive, and process industries have inspired the application of data mining and text mining in historical data-driven fault diagnosis techniques. The failure record includes the mechanism of product failure, the parts involved and the phenomenon of failure, which can help to carry out product analysis and guide workers to complete the repair of the failure. Industrial fault records are divided into structured data (such as model numbers of parts, operating signals, and common voltage and current observations and other information that can be directly diagnosed using a computer), and unstructured information (usually embedded in text form ). Structured numerical data can be directly used by computers, but information retrieval and diagnosis of unstructured fault records usually rely on the experience of professional technicians, which is time-consuming and inefficient. Mining and analyzing such text records can help maintenance workers to complete the judgment of the type of failure, analyze the cause of the failure and retrieve the corresponding maintenance plan. The failure of complex products often involves different components, and each component will produce different failure phenomena under different failure reasons, corresponding to different failure subject categories and solutions. In addition, due to the existence of common text description problems such as professional words, polysemous words, and modal particles, it may cause deviations in the translation and feature recognition of the text by the computer. The differences in the habits of different manufacturers and maintenance personnel to record text also affect the extraction effect of fault features. The feature extraction results are used as the input of the classifier and similar case retrieval to directly determine the fault diagnosis performance of the model. Therefore, for the case records of different products, there is an urgent need for a general method that can accurately extract fault features and carry out intelligent fault diagnosis.

Contents of the invention

The purpose of the present invention is to provide a fault diagnosis method and device to solve the problem of insufficient efficiency of the fault text feature extraction technology in the existing technical solution, and the features are not detailed, resulting in low fault diagnosis automation and insufficient precision.

In order to achieve the above purpose, an embodiment of the present invention provides a fault diagnosis method, including:

Build a fault diagnosis network model; the fault diagnosis network model includes a feature extraction network, a feature interaction network and a feature classification network;

According to the work order sample data, the feature extraction network and the historical fault database, determine first feature data for characterizing text semantics and second feature data for characterizing text fault topics;

Determine a trained fault diagnosis model according to the first feature data, the second feature data, the feature interaction network and the feature classification network;

Perform similarity calculation based on the fault diagnosis model and the historical fault database to determine corresponding work order processing information in the historical fault database.

Optionally, determining a trained fault diagnosis model according to the first feature data, the second feature data, the feature interaction network, and the feature classification network includes:

Performing vector product interaction on the first feature data and the second feature data to determine an interacted weight coefficient matrix;

determining a processed first weight coefficient according to the normalization function of the feature interaction network and the weight coefficient matrix;

By weighting and summing the first weight coefficient and the first feature data, the third feature data is determined; the third feature data is used to characterize the feature data of the fusion text semantics and theme;

A trained fault diagnosis model is determined according to the third feature data and the target loss function of the feature classification network.

Optionally, the target loss function of the feature classification network is determined as follows:

determining a first loss function of the second feature data passing through the feature classification network, a second loss function of the third feature data passing through the feature classification network, and a third loss function of model parameter regularization loss;

The target loss function is determined according to the weighted sum of the first loss function, the second loss function, and the third loss function.

Optionally, the first loss function is determined according to a preset first cross-entropy function and the number of training samples;

The second loss function is determined according to a preset second cross-entropy function and the number of training samples;

Wherein, both the first cross-entropy function and the second cross-entropy function include: the class label of the sample and the predicted value of the class of the sample.

Optionally, performing vector product interaction on the first feature data and the second feature data to determine an interacted weight coefficient matrix, including:

Performing vector product interaction on the first feature data and the second feature data to determine a semantic feature interaction matrix;

The semantic feature interaction matrix is processed by a layer of convolution network to determine the weight coefficient matrix.

Optionally, according to the work order sample data, the feature extraction network and the historical fault database, determining the first feature data for characterizing text semantics and the second feature data for characterizing text fault topics include:

Determine the first feature data of the work order sample data through the first preset algorithm of the feature extraction network; the first feature data includes semantic length and embedding vector dimension;

Through the second preset algorithm of the feature extraction network, according to the work order sample data and the historical fault database, determine the second feature data of the work order sample data; the second feature data includes the number of topics, the topic of failure phenomena, Failure cause topic, failure action topic.

Optionally, determining a trained fault diagnosis model according to the third feature data and the target loss function of the feature classification network includes:

determining a target fault classification loss value according to the third feature data and the target loss function;

If the target fault classification loss value is lower than the threshold, optimize the feature extraction network, the feature interaction network and the feature classification network according to a preset function until the target fault classification loss value is greater than or equal to the When the threshold value is determined, the trained fault diagnosis model is determined.

Optionally, the above method also includes:

Obtain N historical fault work orders and perform preprocessing to determine the fault characteristic data; N is a positive integer;

Based on the fault feature data, construct the historical fault database; the fault database includes the correspondence between N historical fault work orders and N third feature data; the third feature data is stored in the history in vector form in the fault database;

Wherein, the fault characteristic data includes: one or more of fault mode, fault cause, fault impact, fault detection method, design improvement measures and use compensation measures;

The preprocessing includes one or more of noise information elimination, repeated data deletion, and sensitive word filtering.

In order to achieve the above purpose, an embodiment of the present invention also provides a fault diagnosis device, including:

A building block for constructing a fault diagnosis network model; the fault diagnosis network model includes a feature extraction network, a feature interaction network and a feature classification network;

The first determination module is used to determine the first feature data used to represent the text semantics and the second feature data used to represent the text fault theme according to the work order sample data, the feature extraction network and the historical fault database;

A second determining module, configured to determine a trained fault diagnosis model according to the first feature data, the second feature data, the feature interaction network, and the feature classification network;

The third determination module is configured to perform similarity calculation based on the fault diagnosis model and the historical fault database, and determine corresponding work order processing information in the historical fault database.

In order to achieve the above purpose, an embodiment of the present invention also provides a readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the fault diagnosis method described in any one of the above items is implemented. step.

The beneficial effects of above-mentioned technical scheme of the present invention are as follows:

The above technical solution extracts the first feature data used to represent the semantics of the text and the second feature data used to represent the text fault theme, and combines the first feature data and the second feature data through the feature interaction network, and then classifies the features The network determines the trained fault diagnosis model, calculates the similarity between the fault diagnosis model and the historical fault database, determines the corresponding work order processing information in the historical fault database, improves the accuracy of automatic fault diagnosis, and enhances the model Interpretability of feature learning and diagnostic adaptability to different fault topics.

Description of drawings

Fig. 1 is one of the flowcharts of the fault diagnosis method provided by the embodiment of the present invention;

Fig. 2 is the structural diagram of the fault diagnosis network model that the embodiment of the present invention provides;

FIG. 3 is a structural diagram of a feature interaction network provided by an embodiment of the present invention;

Fig. 4 is the second flow chart of the fault diagnosis method provided by the embodiment of the present invention;

Fig. 5 is a structural diagram of a fault diagnosis device provided by an embodiment of the present invention.

Detailed ways

In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will describe in detail with reference to the drawings and specific embodiments.

It should be understood that reference throughout this specification to "one embodiment" or "an embodiment" means that a particular feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the present invention. Thus, appearances of "in one embodiment" or "in an embodiment" in various places throughout the specification are not necessarily referring to the same embodiment. Furthermore, the particular features, structures or characteristics may be combined in any suitable manner in one or more embodiments.

In various embodiments of the present invention, it should be understood that the sequence numbers of the following processes do not mean the order of execution, and the order of execution of each process should be determined by its functions and internal logic, rather than implementing the present invention. The implementation of the examples constitutes no limitation.

Additionally, the terms "system" and "network" are often used interchangeably herein.

In the embodiments provided in this application, it should be understood that "B corresponding to A" means that B is associated with A, and B can be determined according to A. However, it should also be understood that determining B according to A does not mean determining B only according to A, and B may also be determined according to A and/or other information.

As shown in Figure 1, an embodiment of the present invention provides a fault diagnosis method, including:

Step 101, constructing a fault diagnosis network model; the fault diagnosis network model includes a feature extraction network, a feature interaction network and a feature classification network.

The feature extraction and interaction fault diagnosis network model based on text mining constructed in the present invention is called topic-semantic interaction feature model (Topic-context Interaction Model, TCIM). As shown in Figure 2, according to the specific functions of each part structure, the TCIM model can be divided into three parts: feature extraction network (Gg), feature interaction network (Gf) and feature classification network (Gd). Here, the network structure in Figure 2 It can be used in any of the following steps. The arrows in Figure 2 represent the flow process of fault raw data during model training.

Step 102, according to the work order sample data, the feature extraction network and the historical fault database, determine the first feature data used to represent the semantics of the text and the second feature data used to represent the text fault topic. Among them, the work order sample data is text data of failure cases, such as record data in text form such as maintenance logs and manuals.

In this embodiment, the first feature data can represent the global information of the text mined from the perspective of words or words, and can also be represented as local semantic features determined by the attention mechanism; the second feature data is the feature data of the text fault topic , represents the key information in the text, and if the subject word matches the candidate keyword, it means that the candidate content fully represents the gist of the text. The topic keywords are constructed using the Latent Dirichlet Allocation (LDA) topic model. The basis of topic feature scoring is whether the candidate keywords appear in the topic feature words. If they appear, the weight is doubled, otherwise the weight remains unchanged.

Step 103. Determine a trained fault diagnosis model according to the first feature data, the second feature data, the feature interaction network, and the feature classification network;

In this embodiment, the first feature data and the second feature data are interacted through the feature interaction network, and then the trained fault diagnosis model is further determined through the feature classification network.

Step 104: Perform similarity calculation based on the fault diagnosis model and the historical fault database to determine corresponding work order processing information in the historical fault database.

In the embodiment of the present invention, after the fault diagnosis model is determined, input the work order data to be processed into the fault diagnosis model, the fault diagnosis model can classify the work order data to be processed, determine the classified data or determine the The key information of the work order data is calculated according to the similarity between the classified data or the key information of the work order data to be processed and the historical fault database, so as to find similar work orders and return the processing opinion of the work order.

It should be noted that the process of similarity is: through the feature interaction network, the first feature data and the second feature data are generated into the third feature data containing the first two feature information, and the third feature data set stored in the historical fault database Carry out similarity calculations to find similar fault cases to help solve current problems, and the fault diagnosis model is used to help complete fault classification and intelligent diagnosis.

Optionally, the calculation method of similarity calculation adopts the determination method of cosine similarity, and the degree of correlation between the two texts is judged by the cosine measure. If the cosine similarity is larger, it means that the angle between the two variables is smaller, and the similarity between the two texts is greater. High, if the cosine similarity is smaller, it means that the greater the angle between the two variables, the lower the similarity between the two texts. Finally, after comparing with the historical fault database, determine the work order processing information with the highest similarity in the historical fault database.

The solution of the invention can complete the diagnosis and retrieval tasks for different cases, improves the accuracy of automatic fault diagnosis, and enhances the interpretability of model feature learning and the adaptability of diagnosis to different fault topics.

Optionally, the above step 102 includes:

Step 104: Determine the first feature data of the work order sample data through the first preset algorithm of the feature extraction network; the first feature data includes semantic length and embedding vector dimension;

Step 105, through the second preset algorithm of the feature extraction network, according to the work order sample data and the historical failure database, determine the second characteristic data of the work order sample data; the second characteristic data includes the number of topics, failure Symptom topic, failure cause topic, failure action topic.

It should be noted that the number of topics and the prior distribution of topics are parameters determined in advance (the research field is expected to be divided into several topics), which can be determined manually or by machine algorithms. Here, the subjects of the second characteristic data refer to all the fault subjects contained in the historical fault database, for example, for the automotive field, it includes: engine cylinder intake and exhaust faults, mechanical transmission faults, power battery fault types, etc. After the topic feature module of the feature extraction network is learned, all historical data are divided into several categories corresponding to the number of topics, and there will be corresponding words under each topic (that is, article-topic and topic-word probability distribution). The weighted average of the embedding vectors corresponding to the first few high-frequency words under each topic can obtain the final second feature information, which represents the main fault phenomenon and cause under each topic.

In the embodiment of the present invention, the feature extraction network (Gg) is used to extract effective fault features from the original fault data text (work order sample data), and is divided into a semantic feature extraction module (as shown in Figure 2) and a theme feature extraction module (as shown in picture 2). The semantic feature extraction module uses the convolutional neural network to extract the semantic information of the text based on the learned word vector. After normalization, the linear rectification function (ReLU function) and the pooling layer can obtain the required text semantic feature vector, that is, the first feature data. The topic feature extraction module is based on the latent Dirichlet (LDA) topic model widely used in text mining, and performs topic mining on the records in the case database to obtain the distribution of fault topics that can be observed artificially and the high-frequency words under each topic. The high-frequency word vectors under each topic are weighted and summed to obtain a topic feature vector under each fault topic, that is, the first feature data.

In an optional embodiment, the extraction of semantic features and topic category features is carried out through a feature extraction network (Gg), wherein the semantic features (first feature data) pass through the Word2vec (word to vector) model and convolution for each case text The semantic feature embedding vector D={d ₁ ,d ₂ ,…,d _e }∈R ^s×e output after the layer, s is the length of the document, and e is the dimension of the embedding vector set. The topic category feature (second feature data) is the topic feature vector T={ t ₁ ,t ₂ ,…,t _k }∈R ^k×e , k is the number of topics, and e is the embedding vector dimension. Understandably, R ^k×e represents a vector dimension space, that is, there are k components, and each component is an e-dimensional vector.

Optionally, the above step 103 includes:

Step 104, performing vector product interaction on the first characteristic data and the second characteristic data, and determining the weight coefficient matrix after interaction;

Step 105. Determine the processed first weight coefficient according to the normalization function of the feature interaction network and the weight coefficient matrix;

Step 106. Determine third feature data by weighting and summing the first weight coefficient and the first feature data; the third feature data is used to characterize the feature data of the fusion text semantics and theme;

Step 107: Determine a trained fault diagnosis model according to the third feature data and the target loss function of the feature classification network.

In this embodiment, in the feature interaction network (Gf), the vector product is performed on the extracted topic features and semantic features to obtain the topic-semantic relationship, that is, the vector product interaction between the first feature data and the second feature data , to determine the weight coefficient matrix after the interaction; after the weight coefficient matrix passes through the normalization function (softmax activation function), it is the corresponding topic weight value of the text segment, which can be weighted to the semantic features to get the third content of the text under the guidance of the topic category. Feature data, the third feature data incorporates the topic and semantic information of the case. For the third feature data, it is in the form of a vector that can be easily stored by the computer, and the similarity between the current case and the historical fault cases in the case library can be obtained through the calculation of the cosine vector, and the case retrieval is completed to provide the solution of the current case.

Wherein, in step 107, the feature classification network (Gd) is composed of a multi-layer fully connected neural network, and its input is the theme category feature vector extracted by the theme feature extraction module and the third feature data generated through the interaction network, through the feature classification network ( The target loss function of Gd), the output is the probability value corresponding to all fault categories, which is used to predict the fault classification according to the case characteristics and complete the intelligent fault diagnosis.

Optionally, the target loss function of the feature classification network is determined as follows:

Step 108, determining the first loss function of the second feature data passing through the feature classification network, the second loss function of the third feature data passing through the feature classification network, and the third loss function of the model parameter regularization loss;

Step 109: Determine the target loss function according to the weighted summation of the first loss function, the second loss function, and the third loss function.

In this embodiment, the target loss function is mainly composed of the loss function of the classifier. The total loss function of the fault classifier, that is, the target loss function Lc, is obtained by the weighted sum of the prediction loss of topic features, the prediction loss of interaction features and the regularization loss of model parameters, and is defined as:

L _c ＝L ₁ +PL ₂ +λR(w), Formula 1;

In Formula 1, L1 is the prediction loss obtained from the interaction feature through the classification module, that is, the second loss function of the third feature data through the feature classification network; L2 is the prediction loss obtained from the topic feature through the classification module, that is, the second The feature data passes through the first loss function of the feature classification network; P and λ are the weight values corresponding to the losses of the two, between 0 and 1 (including 0 and 1), R(W)=‖W‖ ² is The quadratic square regularization value of all parameters to be learned in the model.

Specifically, the first loss function is determined according to a preset first cross-entropy function and the number of training samples;

The second loss function is determined according to a preset second cross-entropy function and the number of training samples;

Wherein, both the first cross-entropy function and the second cross-entropy function include: the class label of the sample and the predicted value of the class of the sample.

In this embodiment, the failure prediction loss values of the second loss function L1 and the first loss function L2 are measured by the cross entropy between the prediction and the real label, and the prediction loss formulas of L1 and L2 are as follows:

为交叉熵函数，y₁为当前故障文本的故障类型，y_h＝f(H)为交互特征H经过分类器得到的故障类型预测值，y_t＝f(T)表示为主题特征向量T经过分类器得到的故障类型预测值。Among them, n is the number of training samples;

is the cross-entropy function, y ₁ is the fault type of the current fault text, y _h = f(H) is the fault type prediction value obtained by the interaction feature H through the classifier, y _t = f(T) is expressed as the topic feature vector T through The predicted value of the fault type obtained by the classifier.

The cross entropy function is defined as follows:

为模型对第i个样本分类的预测值。where y ⁽ⁱ⁾ is the category label of the i-th sample, that is, the true value of the sample;

is the predicted value of the model for the i-th sample classification.

In summary, the target loss function Lc is expressed as follows:

Optionally, as shown in FIG. 3, the above step 104 includes:

Step 110, performing vector product interaction on the first feature data and the second feature data to determine a semantic feature interaction matrix.

In this embodiment, the first feature data extracted from the work order sample data and the second feature data are subjected to vector product interaction, that is, the first feature data and the second feature data are input to the feature interaction network, and the feature interaction network The specific network structure is shown in Figure 3. First, carry out the vector product to complete the interactive calculation, and get the category semantic feature interaction matrix G=DT ^T ={g ₁ ,g ₂ ,…,g _s }∈R ^s×k , the semantic feature interaction matrix G enables the model to provide More feature information and the interpretability of the model are preserved, and the re-expression of the text is helped by introducing topic prior knowledge.

Step 111 , subjecting the semantic feature interaction matrix to one layer of convolutional network processing to determine the weight coefficient matrix.

In order to obtain the topic-word pair association, a layer of convolutional network is further used to process the semantic feature interaction matrix G to capture the non-linear relationship of adjacent words.

For each G vector whose center is d, from dn to d+n, the convolution kernel size is n, and the relu activation function is used to obtain the convolutional vector U={u ₁ ,u ₂ ,…,u _s }∈R ^s×k , for the dth convolution vector u _d , the formula 5 of the convolution process is as follows:

u _d = Relu(w _d G _dn:d+n +b _d ), Formula 5;

Among them, w _d and b _d are the parameters to be learned by the convolutional network model, and finally get the final weight attention vector V={v ₁ ,v ₂ ,…,v _S }∈ through maxpooling u _d R ^{s ×1} . V is a vector of length s that stores the attention score between each word and topic in the current text, and the score can be converted into a topic-word weight coefficient vector β with the Softmax function, namely: β=SoftMax(v).

Among them, the d-th topic-word weight coefficient is calculated as:

After the weight coefficient vector β is calculated, the new interaction feature H={h ₁ ,h ₂ ,…,h _S }∈R ^S×E can be generated by corresponding weighting with each word component of the original text semantic feature, which is the final The generated new feature with the interactive relationship between topic and word features, that is, the third feature data, which contains the text information and category information of each case, is calculated as:

Further, as shown in FIG. 2, step 107 of the embodiment of the present invention includes:

Step 112. Determine a target fault classification loss value according to the third characteristic data and the target loss function;

Step 113: If the target fault classification loss value is lower than the threshold, optimize the feature extraction network, the feature interaction network, and the feature classification network according to a preset function until the target fault classification loss value is greater than or When it is equal to the threshold, the trained fault diagnosis model is determined.

In the embodiment of the present invention, the semantic feature extraction network, feature interaction network and feature classification network can also be optimized in determining the trained fault diagnosis model. Fix the trained topic feature extraction module, use the original data to extract semantic features, complete feature interaction and classification prediction, and obtain the target fault classification loss value after a round of training; if the target fault classification loss value is lower than the threshold value, then according to The preset function optimizes the feature extraction network, the feature interaction network, and the feature classification network, that is, for the current loss reverse gradient propagation, repeated optimization, repeating step 112, until the target fault classification loss value is greater than or equal to When the threshold is reached, the model loss convergence ends, and a trained fault diagnosis model is determined.

It should be noted that the optimization objective of the model is to minimize the overall classification loss of the fault classifier. On the one hand, the model needs to minimize the L ₂ loss to constrain the extracted topic features to fit the classification category as much as possible after weighting. On the other hand, after adding semantic features, the model needs to minimize the L ₁ loss to constrain the generated features to represent The complete information of the case enables the feature classification network to distinguish the type of fault, so that the generated fine-grained features have the ability of fault diagnosis. During the training process, the model achieves the above optimization goals by calculating the target fault classification loss value in the feature classification network, and backpropagating to iteratively optimize the parameters of the feature extraction model and the feature interaction model in turn.

It should also be noted that before step 103 above, model initialization is required first. Initialize the weight parameters of the feature extraction network (Gg), feature interaction network (Gf) and feature classification network (Gd); secondly, test and determine the parameters of the topic feature extraction module in the feature extraction network (Gg); use the original data to extract Topic features, using high-frequency words and topic visualization distribution methods to conduct experiments to determine the appropriate topic model parameters, make full use of the experience of relevant technical personnel, so that the topic feature distribution is as close as possible to the real fault feature distribution.

Optionally, the above method also includes:

Step 114, obtaining N historical fault work orders, and performing preprocessing to determine fault characteristic data; N is a positive integer;

Step 115. Based on the fault feature data, construct the historical fault database; the fault database includes the correspondence between N historical fault work orders and N third feature data; the third feature data is stored in vector form in In the historical failure database;

Wherein, the fault characteristic data includes: one or more of fault mode, fault cause, fault impact, fault detection method, design improvement measures and use compensation measures;

The preprocessing includes one or more of noise information elimination, repeated data deletion, and sensitive word filtering.

It should be noted that the constructed historical fault database stores historical fault cases (historical fault work orders) and their corresponding third features (generated during training, including the first and second feature data). The three-feature data is stored in the form of vectors, so compared with traditional text storage and similar text retrieval, this method is easier to store and more convenient to complete the similarity calculation between cases.

In this embodiment, N historical fault work orders are obtained and pre-processed, wherein the determined fault characteristic data includes manual diagnosis results and system diagnosis results, and the system diagnosis results are processed based on pre-stored fault diagnosis system fault characteristic data As obtained, N historical fault work orders include M manual diagnosis results, where M is a positive integer less than or equal to N, and N is a positive integer; wherein, the fault characteristic data is to process the historical fault The fault detailed data of the work order is obtained, and the fault detailed data includes data within a preset time period associated with the occurrence time of the fault in the historical fault work order. The fault characteristic data includes product information (product name, product model, function, material, environmental load, performance parameters) and fault information (failure mode, fault cause, fault impact, fault detection method, design improvement measures and use compensation measures) Wait)

In an optional embodiment: establish a historical failure database, including three database tables of product family/product platform, product detailed information, and function description, and three database tables including failure mode, detailed information, and failure mechanism. Through the form of product family and product tree, organize products with the same internal interface that achieve similar functions, and then add different personality modules on the product platform to form product instances. All failure knowledge belongs to a certain product instance or platform. The database includes two database tables, the relational layer table and the application layer table. The known object relationship is stored in the relationship layer table, and the data relationship between product functions and faults is stored in the application layer table, forming a historical fault database.

In another specific embodiment, as shown in Figure 4, the present invention also provides an overall flow chart, including:

Step 1: Extract historical fault case data (historical fault work orders) from the database, perform text preprocessing, including identifying professional words and removing stop words, and obtain a fault case corpus (historical fault database).

Step 2: According to the feature extraction network, the text semantic features of each fault case and the overall fault topic category features of the case library are extracted in two stages, and the appropriate network layer structure parameters are selected to generate features through the feature interaction network, and through the feature classification network The medium classifier minimizes the overall loss function to train the model.

Step 3: For a new fault case, preprocess the case test sample and directly input the trained model to obtain the feature representation of the current case.

Step 4: The characteristic features of the case can be classified, predicted and diagnosed, and the similarity calculation with the case features in the historical fault database can be performed to find similar cases and complete the diagnosis and solution retrieval of new fault cases.

Step 5: Add the case analysis and inspection to the historical fault database for continuous updating of the database.

In summary, the scheme of the present invention obtains fine-grained case features considering the fault topic-semantic relationship through two-stage extraction and interaction of semantic and topic features, completes diagnosis and retrieval tasks for different cases, and improves automation The accuracy of fault diagnosis enhances the interpretability of model feature learning and the adaptability of diagnosis to different fault topics.

As shown in Figure 5, an embodiment of the present invention also provides a fault diagnosis device, including:

A construction module 501, configured to construct a fault diagnosis network model; the fault diagnosis network model includes a feature extraction network, a feature interaction network and a feature classification network;

The first determining module 502 is configured to determine first feature data for representing text semantics and second feature data for representing text fault topics according to work order sample data, the feature extraction network and historical fault database;

The second determination module 503 is configured to determine a trained fault diagnosis model according to the first feature data, the second feature data, the feature interaction network and the feature classification network;

The third determination module 504 is configured to calculate the similarity between the fault diagnosis model and the historical fault database, and determine the corresponding work order processing information in the historical fault database.

Optionally, the second determining module 503 includes:

The first determining submodule is used to perform vector product interaction on the first characteristic data and the second characteristic data, and determine the weight coefficient matrix after interaction;

The second determination submodule is used to determine the processed first weight coefficient according to the normalization function of the feature interaction network and the weight coefficient matrix;

The third determining submodule is used to determine the third feature data by weighting and summing the first weight coefficient and the first feature data; the third feature data is used to characterize the feature data of the fusion text semantics and theme ;

The fourth determining submodule is used to determine a trained fault diagnosis model according to the third feature data and the target loss function of the feature classification network.

Optionally, in the construction module 501, the target loss function of the feature classification network is determined in the following manner:

A first determining unit, configured to determine a first loss function of the second feature data passing through the feature classification network, a second loss function of the third feature data passing through the feature classification network, and a third loss function of model parameter regularization loss ;

A second determining unit, configured to determine the target loss function according to the weighted summation of the first loss function, the second loss function, and the third loss function.

Specifically, the first determining unit is specifically configured to determine the first loss function according to a preset first cross-entropy function and the number of training samples;

The second determination unit is specifically configured to determine the second loss function according to a preset second cross-entropy function and the number of training samples;

Wherein, both the first cross-entropy function and the second cross-entropy function include: the class label of the sample and the predicted value of the class of the sample.

Optionally, the first determining submodule includes:

A third determining unit, configured to perform vector product interaction on the first feature data and the second feature data to determine a semantic feature interaction matrix;

The fourth determination unit is configured to process the semantic feature interaction matrix with a layer of convolutional network to determine the weight coefficient matrix.

Optionally, the first determining module 502 includes:

The fifth determination unit is configured to determine the first feature data of the work order sample data through the first preset algorithm of the feature extraction network; the first feature data includes a semantic length and an embedding vector dimension;

The sixth determination unit is configured to determine the second feature data of the work order sample data according to the work order sample data and the historical fault database through the second preset algorithm of the feature extraction network; the second feature data includes The number of topics, the topics of fault symptoms, the topics of fault causes, and the topics of fault measures.

Optionally, the fourth determining submodule includes:

A seventh determining unit, configured to determine a target fault classification loss value according to the third feature data and the target loss function;

An eighth determination unit, configured to optimize the feature extraction network, the feature interaction network, and the feature classification network according to a preset function until the target fault classification is performed if the target fault classification loss value is lower than a threshold When the loss value is greater than or equal to the threshold, a trained fault diagnosis model is determined.

In an embodiment of the present invention, the above-mentioned fault diagnosis device further includes:

The obtaining module is used to obtain N historical fault work orders, and perform preprocessing to determine fault characteristic data; N is a positive integer;

The second building module is configured to construct the historical fault database based on the fault feature data; the fault database includes correspondence between N historical fault work orders and N third feature data; the third feature data is Stored in the historical fault database in vector form;

Wherein, the fault characteristic data includes: one or more of fault mode, fault cause, fault impact, fault detection method, design improvement measures and use compensation measures;

The preprocessing includes one or more of noise information elimination, repeated data deletion, and sensitive word filtering.

Wherein, the implementation embodiments of the above-mentioned fault diagnosis method are all applicable to the embodiment of the fault diagnosis device, and can also achieve the same technical effect, so in order to avoid repetition, details are not repeated here.

A readable storage medium according to an embodiment of the present invention, on which programs or instructions are stored, and when the programs or instructions are executed by a processor, the steps in the above-mentioned fault diagnosis method can be realized, and the same technical effect can be achieved, To avoid repetition, details are not repeated here.

Wherein, the processor is the processor in the fault diagnosis method described in the above embodiments. The readable storage medium includes a computer readable storage medium, such as computer read-only memory (Read-Only Memory, ROM for short), random access memory (Random Access Memory, RAM for short), magnetic disk or optical disk, and the like.

In the embodiments of the present invention, the modules may be implemented by software so as to be executed by various types of processors. An identified module of executable code may, by way of example, comprise one or more physical or logical blocks of computer instructions which may, for example, be structured as an object, procedure, or function. Notwithstanding, the executable code of an identified module need not be physically located together, but may include distinct instructions stored in different bits which, when logically combined, constitute the module and implement the specified Purpose.

Indeed, a module of executable code may be a single instruction, or many instructions, and may even be distributed over several different code segments, among different programs and across multiple memory devices. Likewise, operational data may be identified within modules, and may be implemented in any suitable form and organized within any suitable type of data structure. The operational data may be collected as a single data set, or may be distributed in different locations (including on different storage devices), and may exist, at least in part, only as electronic signals on a system or network.

When the module can be realized by software, considering the level of the existing hardware technology, the module that can be realized by software, regardless of the cost, those skilled in the art can build the corresponding hardware circuit to realize the corresponding function. The hardware circuit includes conventional very large scale integration (VLSI) circuits or gate arrays as well as existing semiconductors such as logic chips, transistors, or other discrete components. A module may also be implemented in programmable hardware devices such as field programmable gate arrays, programmable array logic, programmable logic devices, and the like.

The exemplary embodiments described above are described with reference to these drawings. Many different forms and embodiments are possible without departing from the spirit and teachings of the present invention. Therefore, the present invention should not be construed as the exemplary embodiments set forth herein. limits. Rather, these exemplary embodiments are provided so that this disclosure will be thorough and complete, and will convey the scope of the invention to those skilled in the art. In the drawings, component sizes and relative sizes may be exaggerated for clarity. The terminology used herein is for the purpose of describing certain exemplary embodiments only and is not intended to be limiting. As used herein, the singular forms "a", "an" and "the" are intended to include these plural forms unless the context clearly dictates otherwise. It will be further understood that the terms "comprises" and/or "comprises", when used in this specification, indicate the presence of stated features, integers, steps, operations, components and/or components, but do not exclude one or more other The presence or addition of features, integers, steps, operations, components, components and/or groups thereof. Unless otherwise indicated, when stated a range of values includes the upper and lower limits of that range and any subranges therebetween.

The above description is a preferred embodiment of the present invention, it should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, some improvements and modifications can also be made, and these improvements and modifications can also be made. It should be regarded as the protection scope of the present invention.

Claims (10)

1. A fault diagnosis method, characterized in that, comprising: Build a fault diagnosis network model; the fault diagnosis network model includes a feature extraction network, a feature interaction network and a feature classification network; According to the work order sample data, the feature extraction network and the historical fault database, determine first feature data for characterizing text semantics and second feature data for characterizing text fault topics; Determine a trained fault diagnosis model according to the first feature data, the second feature data, the feature interaction network and the feature classification network; Perform similarity calculation based on the fault diagnosis model and the historical fault database to determine corresponding work order processing information in the historical fault database. 2. The method according to claim 1, characterized in that, according to the first feature data, the second feature data, the feature interaction network and the feature classification network, determine the trained fault diagnosis model, include: Performing vector product interaction on the first feature data and the second feature data to determine an interacted weight coefficient matrix; determining a processed first weight coefficient according to the normalization function of the feature interaction network and the weight coefficient matrix; By weighting and summing the first weight coefficient and the first feature data, the third feature data is determined; the third feature data is used to characterize the feature data of the fusion text semantics and theme; A trained fault diagnosis model is determined according to the third feature data and the target loss function of the feature classification network. 3. The method according to claim 2, wherein the target loss function of the feature classification network is determined as follows: determining a first loss function of the second feature data passing through the feature classification network, a second loss function of the third feature data passing through the feature classification network, and a third loss function of model parameter regularization loss; The target loss function is determined according to the weighted sum of the first loss function, the second loss function, and the third loss function. 4. The method according to claim 3, wherein the first loss function is determined according to a preset first cross-entropy function and the number of training samples; The second loss function is determined according to a preset second cross-entropy function and the number of training samples; Wherein, both the first cross-entropy function and the second cross-entropy function include: the class label of the sample and the predicted value of the class of the sample. 5. The method according to claim 2, wherein the vector product interaction is carried out with the first characteristic data and the second characteristic data, and the weight coefficient matrix after the interaction is determined, comprising: Performing vector product interaction on the first feature data and the second feature data to determine a semantic feature interaction matrix; The semantic feature interaction matrix is processed by a layer of convolution network to determine the weight coefficient matrix. 6. The method according to claim 1, characterized in that, according to the work order sample data, the feature extraction network and the historical fault database, determine the first feature data for characterizing text semantics and the first feature data for characterizing text fault topics Second characteristic data, including: Determine the first feature data of the work order sample data through the first preset algorithm of the feature extraction network; the first feature data includes semantic length and embedding vector dimension; Through the second preset algorithm of the feature extraction network, according to the work order sample data and the historical fault database, determine the second feature data of the work order sample data; the second feature data includes the number of topics, the topic of failure phenomena, Failure cause topic, failure action topic. 7. The method according to claim 2, wherein, according to the third feature data and the target loss function of the feature classification network, determining a trained fault diagnosis model includes: determining a target fault classification loss value according to the third characteristic data and the target loss function; If the target fault classification loss value is lower than the threshold, optimize the feature extraction network, the feature interaction network and the feature classification network according to a preset function until the target fault classification loss value is greater than or equal to the When the threshold value is determined, the trained fault diagnosis model is determined. 8. The method according to claim 1, further comprising: Obtain N historical fault work orders and perform preprocessing to determine fault characteristic data; N is a positive integer; Based on the fault feature data, construct the historical fault database; the fault database includes the correspondence between N historical fault work orders and N third feature data; the third feature data is stored in the history in vector form in the fault database; Wherein, the fault characteristic data includes: one or more of fault mode, fault cause, fault impact, fault detection method, design improvement measures and use compensation measures; The preprocessing includes one or more of noise information elimination, repeated data deletion, and sensitive word filtering. 9. A fault diagnosis device, characterized in that, comprising: A building block for constructing a fault diagnosis network model; the fault diagnosis network model includes a feature extraction network, a feature interaction network and a feature classification network; The first determination module is used to determine the first feature data used to represent the text semantics and the second feature data used to represent the text fault theme according to the work order sample data, the feature extraction network and the historical fault database; A second determining module, configured to determine a trained fault diagnosis model according to the first feature data, the second feature data, the feature interaction network, and the feature classification network; The third determination module is configured to calculate the similarity between the fault diagnosis model and the historical fault database, and determine the corresponding work order processing information in the historical fault database. 10. A readable storage medium, on which a program or instruction is stored, characterized in that, when the program or instruction is executed by a processor, the steps in the fault diagnosis method according to any one of claims 1-8 are realized .

Images