A rolling bearing composite fault feature recognition method based on a causal relationship network

Abstract

The application discloses a rolling bearing composite fault feature recognition method based on a causal relationship network, selects rolling bearing composite fault features, filters out rolling bearing composite fault features, and thus effectively recognizes bearing composite faults. The application uses the causal relationship between composite fault features for composite fault feature selection, constructs a complex network model, uses a complex network node centrality index to quantize the importance of different composite fault features, filters out a smaller number of features, simultaneously realizes a higher composite fault diagnosis accuracy, and has good universality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bearing fault diagnosis, and particularly relates to a rolling bearing composite fault feature selection and identification method based on a causal relationship and a complex network model. BACKGROUND

[0002] Rolling bearings are important components of rotating machinery and are also complex systems composed of multiple parts. The mutual influence of various components in the rolling bearing system makes it prone to composite faults. The causal relationship between the components is contained in the composite fault signal, and the mutual causal action within the system emerges on the scale of bearing signal features, making it difficult to identify the fault type from the bearing composite fault signal.

[0003] Feature selection, as one of the key steps in fault diagnosis, is the basis for improving fault diagnosis accuracy. Therefore, selecting the optimal feature subset from the composite fault features is of great significance to ensuring the stability and reliability of mechanical operation. For the feature selection problem, existing technologies mostly use Pearson correlation coefficient, Fisher score, Laplace score, and Gini index to calculate the distance between bearing fault features and class labels, and select features sensitive to bearing fault class patterns according to the distance. There are also methods that use causal relationships for feature selection to increase the interpretability of the results, such as methods based on Markov blankets and information theory. However, these existing methods only consider the causal effect between bearing fault features and class patterns, ignoring the causal structure information between bearing composite fault features. Existing technologies are difficult to select effective bearing fault features from the causal relationship structure information between these bearing fault features. SUMMARY

[0004] In order to overcome the deficiencies of the existing technologies, the present application proposes a rolling bearing composite fault feature selection and identification method based on a causal relationship network model, which can effectively identify composite fault patterns. This method can select a smaller number of features while achieving higher composite fault diagnosis accuracy, and has good versatility.

[0005] The present application uses a complex network model to express the causal relationship structure between the composite fault features in the rolling bearing system, thereby using the complex network model to construct a bearing composite fault feature selection model. This model can intuitively express various relationships in the bearing system, and can select effective bearing fault features from the causal relationship structure information between these bearing fault features, achieving the effect of effectively identifying bearing composite faults.

[0006] The technical solution of the present application is as follows:

[0007] A rolling bearing composite fault feature recognition method based on a causal relationship network model, by designing a rolling bearing composite fault feature selection method, screening out bearing composite fault features, so as to effectively identify bearing composite faults; comprising:

[0008] 1) Construct a composite fault feature causal relationship network model: this part proposes a network construction method based on Markov blanket discovery algorithm, the proposed method mainly consists of two parts of mining causal relationship structure between features and constructing feature causal relationship network, taking each feature as a target variable, then using Markov blanket discovery algorithm to search for features having causal relationship with the target variable, finally taking features as nodes and causal relationship between features as edges to obtain the composite fault feature causal relationship network; comprising:

[0009] 11) Composite fault feature extraction;

[0010] Select the time domain type, frequency domain type and entropy type of the composite fault signal statistics to form the composite fault feature set; which includes seven composite fault signal dimensional time domain feature parameters; six composite fault signal dimensionless time domain feature parameters; five composite fault signal frequency domain features; two composite fault signal entropy type parameters;

[0011] 12) Obtain the Markov blanket set of each composite fault feature and identify the causal structure between the composite fault features;

[0012] 121) Take each composite fault feature as a target variable and obtain the Markov blanket set of the current target variable;

[0013] 1211) First, initialize the Markov blanket set of the current target variable to an empty set; create a candidate Markov blanket feature set and initialize it to an empty set;

[0014] 1212) Put all composite fault features in the composite fault feature set except the current target variable into the candidate Markov blanket feature set;

[0015] 1213) Calculate the conditional independence of the composite fault features in the candidate Markov blanket feature set and the current target variable, and update the candidate Markov blanket feature set and the Markov blanket set of the current target variable according to the conditional independence; comprising:

[0016] 12131) Calculate the correlation size of each composite fault feature in the candidate Markov blanket feature set and the current target variable, and arrange them in ascending order;

[0017] 12132) Calculate the conditional independence of the current target variable and the first ranked composite fault feature under the condition of Markov blanket; judge whether they are conditionally related or conditionally independent;

[0018] If both are conditionally dependent, the candidate Markov blanket feature set and the Markov blanket set of the current target variable are updated simultaneously: the top-ranked composite fault feature is removed from the candidate Markov blanket feature set and added to the Markov blanket set of the current target variable;

[0019] If both are conditionally independent, the next ranked composite fault feature is judged;

[0020] Until each composite fault feature in the candidate Markov blanket feature set is traversed, the Markov blanket set of the current target variable is obtained;

[0021] 1214) The error nodes in the Markov blanket set of the current target variable are deleted to obtain the final Markov blanket set; including:

[0022] Each feature in the Markov blanket set is sequentially judged for conditional independence with the current target variable under the condition of the remaining features; if conditionally independent, the feature is an error node, and the feature is removed from the Markov blanket set; if conditionally dependent, the next feature is judged until all features are traversed, and the final Markov blanket set of the current target variable is obtained; the composite fault features in the Markov blanket set of the target variable have a causal relationship with the target variable;

[0023] 13) Construct a composite fault feature causal relationship network model according to the causal structure among the composite fault features;

[0024] Each composite fault feature in the composite fault feature set is taken as a network node, and an edge is added between each composite fault feature and the composite fault features in its Markov blanket set, so that the causal relationship structure among the composite fault features is constructed into a composite fault feature causal relationship network;

[0025] 2) Screen the optimal feature subset of composite fault: use the complex network node centrality index to quantify the importance of the composite fault feature, and arrange the features in the feature causal relationship network in ascending order according to the node centrality index. Start from the feature with the lowest centrality index and add it to the feature subset one by one, input it into the deep neural network model that has been constructed to perform composite fault diagnosis, and take the composite fault feature subset corresponding to the peak accuracy as the optimal feature subset under different indexes. Take the optimal feature subset under the index with the highest accuracy as the feature screening result. Including:

[0026] 21) Calculate the complex network node centrality index for quantifying the importance of the composite fault feature; including: the degree centrality, closeness centrality, betweenness centrality and Pagerank centrality of each composite fault feature in the feature causal relationship network;

[0027] 22) According to the node centrality index value of each composite fault feature, an ascending composite fault feature set is obtained in ascending order;

[0028] 23) For the ascending composite fault feature set, the number of composite fault features is increased in sequence, and the corresponding composite fault feature is input into the constructed composite fault diagnosis deep neural network model;

[0029] 24) The composite fault optimal feature subset corresponding to the peak accuracy is the composite fault optimal feature subset, which is the composite fault feature screening result under this index;

[0030] 25) Three kinds of optimal feature subsets are obtained by using different node centrality indexes;

[0031] 26) The accuracy results of different optimal feature subsets are compared, and the optimal feature subset with the highest accuracy is taken as the final result of the application.

[0032] According to the screened composite fault features, the effective identification of the rolling bearing composite fault can be realized.

[0033] Compared with the prior art, the beneficial technical effects of the application are:

[0034] The application provides a bearing composite fault feature selection and identification method based on causal relationship and complex network model, which can screen out features for effectively identifying composite fault patterns. The optimal feature subset screened by the method can achieve higher diagnostic accuracy with fewer features, and has good universality. Compared with the prior art, the technical advantages of the application include:

[0035] (1) The causal relationship is introduced into the field of composite fault feature selection, and the potential mechanism of the internal relationship flow of the system is explained by using the causal relationship;

[0036] (2) The causal relationship structure between the composite fault features is expressed on the complex network model, and the features and their interaction relationships in the rolling bearing system are considered from the overall perspective by the complex network;

[0037] (3) The composite fault feature selection problem is converted into judging the relative importance of the features in the relationship network, and the importance of different composite fault features is quantified by introducing the complex network node centrality index, which realizes the cross-field migration of the fault feature selection basis. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The flowchart of the feature selection method based on causal relationship and complex network provided by the application. DETAILED DESCRIPTION

[0039] The application will be further described by examples in connection with the drawings, but the application is not limited in any way by the scope of the application.

[0040] A bearing composite fault feature selection and identification method based on causality and complex network model is provided herein, a rolling bearing composite fault feature selection method based on causality and complex network model is designed, and bearing composite fault features are screened out, so that the bearing composite fault can be effectively identified; the feature selection method based on causality and complex network of the application realizes selection of features that can accurately identify fault modes from composite fault features containing causality of each component, the complex network expresses the causality structure between the composite fault features, and the node importance is quantified through a node centrality index, and finally the optimal feature subset of the composite fault is obtained. Figure 1 It is the overall flowchart of the application, including the following steps:

[0041] 1) Constructing a composite fault feature causality network: a network construction method based on a Markov blanket discovery algorithm is proposed, mainly including two parts of mining causality structure between composite fault features and constructing a composite fault feature causality network, taking each composite fault feature as a target variable, then using a Markov blanket discovery algorithm to search for features having causality with the target variable, finally taking the composite fault features as nodes of the network, the causality between the composite fault features as edges between the network nodes, and obtaining the composite fault feature causality network;

[0042] 2) Screening to obtain an optimal feature subset of the composite fault: the importance of the composite fault features is quantified by using a complex network node centrality index, the degree centrality, closeness centrality, betweenness centrality and Pagerank centrality of each feature in the feature causality network are ranked in ascending order, the number of features is increased, and a deep neural network (DNN) is input, and when the accuracy reaches the highest, the corresponding feature subset is taken as the feature screening result.

[0043] Specifically, the rolling bearing composite fault feature selection method of the application includes the following steps:

[0044] A. Making a data set

[0045] According to the collected rolling bearing vibration signals, the application pre-processes and makes a data set in a unified format, and this part includes the following steps:

[0046] A1. Mixing the single faults of the three different types of rolling bearings of the inner ring, the outer ring and the rolling body two by two to form three types of composite fault signals of "inner ring + outer ring", "inner ring + rolling body" and "outer ring + rolling body";

[0047] A2. Each type of composite fault signal in A1 is divided into groups of 1024, a total of 2700 groups of fault signal data and form a composite fault data set X={X i}, i∈(1,2700);

[0048] A3. Label each group of fault signal data in A2 composite fault data set with "inner ring + outer ring", "inner ring + rolling body" and "outer ring + rolling body" three types of labels;

[0049] B. Construct a composite fault feature causal relationship network, which includes three steps:

[0050] B1. Composite fault feature extraction

[0051] The present application selects three types of statistics of time domain, frequency domain and entropy value of composite fault signal to constitute a composite fault feature set Fea=[fea1,fea2,...,fea 20}, which involves formulas as shown in Table 1.

[0052] Table 1 Bearing composite fault features and corresponding calculation methods

[0053]

[0054]

[0055]

[0056] In the above formula, N is the length of the signal; K is the total frequency after Fourier transform of the signal; P k is the power spectral density; P x is the singular spectral density.

[0057] The specific implementation is as follows:

[0058] B1.1 Calculate the time domain features of composite fault signal X i ={x n}, n∈(1,102400), including 7 dimension parameters and 6 dimensionless parameters, dimension parameters are fea1~fea7, which are peak value, peak-peak value, average amplitude, root mean square, root amplitude, variance and standard deviation; Dimensionless parameters are fea8~fea 13 , which are peak factor, pulse factor, margin factor, waveform factor, kurtosis factor and skewness;

[0059] B1.2 Fourier transform the signal X i to get Y i ={y j}, j∈(1,K), calculate 5 frequency domain features of composite fault signal, denoted as fea14 ~ fea 18 , respectively, are the center frequency, mean frequency, root mean square frequency, frequency variance and frequency standard deviation;

[0060] B1.3 Calculate the power spectrum entropy and the singular spectrum entropy of the signal X i , denoted as fea 19 20 ;

[0061] B2. Obtain the Markov blanket set of each composite fault feature, and identify the causal structure between the composite fault features;

[0062] B2.1 Take the composite fault feature fea l as the current target variable, and obtain the Markov blanket set of the current target variable;

[0063] First, take the empty set as the initial set of the Markov blanket set of the target variable, denoted as MB(fea1); at the same time, create a candidate Markov blanket feature set and initialize it as an empty set;

[0064] B2.2 Put all composite fault features in the candidate Markov blanket feature set except the target variable fea1, denoted as CMB(fea1);

[0065] B2.3 Judge the conditional independence of the composite fault features in the candidate Markov blanket feature set, which is implemented as follows:

[0066] B2.3.1 Calculate the correlation between each composite fault feature in the candidate Markov blanket feature set and the target variable through the G 2 test method, and rank the composite fault features in descending order according to the correlation;

[0067] B2.3.2 According to the order obtained in B2.3.1, calculate the conditional independence of the target variable fea1 and the first ranked feature Y under the condition of the Markov blanket, and judge whether fea1 and Y are conditionally dependent or conditionally independent according to the conditional independence. If fea1 and Y are conditionally dependent, remove Y from CMB(fea1) and add it to MB(fea1). If fea1 and Y are conditionally independent, continue to judge the next ranked feature until all features are judged, and obtain the Markov blanket set MB(fea1) of fea1;

[0068] ​B2.4 Remove erroneous nodes from the Markov blanket set MB(fea1) to obtain the final Markov blanket set MB(fea1). The specific implementation is as follows: For each feature in the Markov blanket set MB(fea1), conditionally determine its independence from the target variable fea1, given the remaining features. If conditionally independent (i.e., the feature is an erroneous node), remove that feature from MB(fea1). If conditionally dependent (correlated), continue to the next feature until all features have been determined, resulting in the final Markov blanket set MB(fea1).

[0069] B2.5 respectively uses features fea2~fea 20 For the target variable, repeat the above process B2.1 to B2.4 to obtain features fea2 to fea2. 20 Markov blanket collection MB (fea2) ~ MB (fea) 20 );

[0070] B3. Based on the causal relationship structure among composite fault features, construct a causal relationship network model among composite fault features, denoted as A={a ij} 20×20 ;include:

[0071] B3.1 Add nodes v1 to v in the network 20 Nodes v1 to v 20 These represent the composite fault features fea1 to fea1 within the composite fault feature set Fea. 20 ;

[0072] B3.2 Using the Markov blanket set MB(fea1) of fea1, add the edge a connecting node v1 to the network. 1j The specific implementation is as follows:

[0073] B3.2.1 Determine if there is an edge connecting node v2 and node v1: If the feature fea2 corresponding to node v2 is in the Markov blanket set MB(fea1) of feature fea1, then add an edge connecting node v2 and node v1, i.e., a 12 =1; If the feature fea2 corresponding to node v2 is not in MB(fea1), then it is not added, i.e., a 12 =0;

[0074] B3.2.2 Repeat the above process B3.2.1 to obtain node v1 and nodes v3~v1 respectively. 20 The edges connecting them;

[0075] B3.3 Repeat the process in B3.2, adding features fea2 to fea2 to the network. 20 The corresponding nodes v2~v 20 Connecting edges a2j~a 20j ;

[0076] C. Screening the optimal feature subset

[0077] The present application utilizes the composite fault feature causal relationship network A={a ij} 20×20 , four complex network node centrality indexes are adopted to quantify the importance of the composite fault features, i.e. node centrality, including degree centrality, closeness centrality, betweenness centrality and Pagerank centrality, and the composite fault feature set Fea constructed in B1 is screened according to the importance, which includes two steps:

[0078] C1. Calculate the node centrality of the composite fault feature causal relationship network, which is implemented as follows:

[0079] C1.1. Calculate the degree centrality DC of the feature causal relationship network i , the expression of which is formula (1), wherein, k i＝ ∑ i a ij ,a ij i.e. the element in the i-th row and the j-th column of the network adjacency matrix A, N is the number of nodes of the network, and the denominator N-1 is the maximum degree value of the nodes.

[0080]

[0081] C1.2. Calculate the closeness centrality CC of the feature causal relationship network i , the expression of which is formula (2). d ij is the average shortest distance of any node v i to other nodes in the network. If the average shortest distance of a node is smaller, the closeness centrality of the node is greater.

[0082]

[0083] C1.3. Calculate the betweenness centrality BC of the feature causal relationship network i , the expression of which is formula (3). In the formula, g st represents the number of shortest paths from node v s to node v t , represents the number of shortest paths from node v s to node v t that pass through node v i .

[0084]

[0085] C1.4 Calculate the Pagerank centrality PR of the feature-causal relationship network i , whose expression is formula (4), is the out-degree of node v j .

[0086]

[0087] C2. Select important composite fault features;

[0088] C2.1 Sort the 20 composite fault features fea2~fea 20 in ascending order according to the degree centrality index values obtained in C1.1;

[0089] C2.2 According to the sorting results of C2.1, sequentially select the first 1 feature, the first 2 features,..., and the first 20 features as candidate solutions of the optimal feature subset, denoted as FEA1~FEA 20 .

[0090] C2.3 Calculate the fault diagnosis accuracy under each candidate solution, which is implemented as follows:

[0091] C2.3.1 Construct a general four-layer deep neural network model, including: an input layer, a first hidden layer, a second hidden layer, and an output layer; the first hidden layer is set to have 500 neurons, and the second hidden layer is set to have 100 neurons, both of which use the Relu activation function;

[0092] C2.3.2 Select the candidate solution feature set FEA1 of the optimal feature subset;

[0093] C2.3.3 Randomly initialize the weights of the neural network, and randomly divide the candidate solution feature set in C2.3.2 into two parts in a ratio of 8:2, of which 80% is used as the training sample set and 20% is used as the test sample set;

[0094] C2.3.4 Set the number of training iterations, and use the training sample set in C.2.3.3 to iteratively update the weights of the neural network;

[0095] C2.3.5 Input the candidate solution feature set of each sample in the test sample set in C2.3.3 into the model trained in C2.3.4, and output the prediction result of the fault type of each sample;

[0096] C2.3.6 Compare the results output by the model in C2.3.5 with the actual fault types, and define the ratio of the number of samples correctly predicted by the model to the total number of samples in the test set as the accuracy of the diagnosis, denoted as Acc1;

[0097] C.2.3.7 Alternative solutions of selecting optimal feature subsets in turn FEA2~FEA 20 , repeat the steps of C.2.3.3-C2.3.6 to obtain the diagnostic accuracy of each alternative solution, denoted as Acc2~Acc 20 ;

[0098] C2.4 Select the alternative solution corresponding to the maximum accuracy in Acc1~Acc 20 as the optimal feature subset FS d under the node centrality index according to the accuracy value calculated in C2.3.

[0099] C2.5 Repeat the above process C2.1~C2.4 to obtain the optimal feature subsets FS c , FS b , FS p by using the proximity centrality, betweenness centrality and Pagerank centrality in C1.2~C1.4 respectively.

[0100] C2.6 Compare the accuracy results obtained by composite fault diagnosis of the four optimal feature subsets FS d , FS c , FS b and FS p , and select the optimal feature subset with the highest accuracy as the final result of the present application.

[0101] The following examples are based on a certain bearing data set, and the feature selection method based on causal relationship and complex network proposed in the present application is carried out, which specifically includes the following steps:

[0102] A. The present application extracts 9 single fault sample data of different types and different diameters from the bearing data set under the sampling frequency of 12 kHz. The fault types are inner ring fault, outer ring fault and rolling element fault, and the fault diameters are 0.1778mm, 0.3556mm and 0.5334mm respectively.

[0103] A1. Mix the above 9 single faults two by two to obtain 27 different types of composite fault signals;

[0104] A2. Cut off the first 102400 points of each type of composite fault to ensure the consistency of the number of sampling points of each vibration data, and cut the 102400 points of each type of composite fault into 100 segments. The composite fault sample set with a sample number of 2700 is obtained from the 27 types of composite fault data;

[0105] A3. Label the 27 types of composite faults again, including inner ring-outer ring, inner ring-rolling element and outer ring-rolling element;

[0106] B. This invention constructs a Composite Fault Feature Network (CFN) based on the sample set obtained in A:

[0107] B1. The feature fea 1～ fea 20 Defined as nodes v1 to v 20 Calculate dimensional parameters, including peak value, peak-to-peak value, average amplitude, root mean square (RMS), root square amplitude, variance, and standard deviation, denoted as fea1 to fea7; calculate six dimensionless parameters, including peak factor, impulse factor, margin factor, waveform factor, kurtosis factor, and skewness, denoted as fea8 to fea7. 13 Calculate five frequency domain statistics: centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation, denoted as fea. 14 ~fea 18 Calculate two entropy indices, including power spectral entropy and singular spectral entropy, denoted as fea. 19 and fea 20 ;

[0108] B2. Adding edges between features with causal relationships: This invention uses IAMB (Incremental Association Markov Blanket) to obtain nodes v1 to v2. 20 Markov blanket collection MB (fea1) ~ MB (fea1) 20 ), at nodes v1 to v 20 Add connecting edges between features within its Markov blanket set;

[0109] B3. Based on the causal relationship structure among composite fault features, construct a causal relationship network model among composite fault features, denoted as A={a ij} 20×20 When feature fea i and features fea j When a causal relationship exists (i.e., within the other party's Markov blanket set), a ij =a ji =1, otherwise 0;

[0110] C. Selecting the optimal feature subset of compound faults

[0111] C1. Calculate the degree centrality, proximity degree centrality, betweenness centrality, and Pagerank centrality indices of the CFN network;

[0112] C2. Different alternatives are obtained by ascendingly sorting the features of the compound fault according to the size of the centrality index, the optimal feature subset is identified by using the diagnostic accuracy of different alternatives, the diagnostic accuracy of the optimal feature subset under different centrality indexes is compared, and the optimal feature subset selected under the proximity centrality index is taken as the screening result.

[0113] Table 2 Compound fault diagnosis results of the optimal feature subset

[0114]

[0115] The objective evaluation results of the feature selection method based on causality and complex network are shown in Table 2, and it can be seen from Table 2 that the method has obvious technical advantages in compound fault feature screening, and higher fault diagnosis accuracy is achieved with fewer features.

[0116] It should be noted that the purpose of the disclosed embodiments is to help further understand the present application, but those skilled in the art can understand that various replacements and modifications are possible without departing from the scope of the present application and the appended claims. Therefore, the present application should not be limited to the disclosed embodiments, and the scope of the present application is defined by the scope of the claims.

Claims

1. A rolling bearing composite fault feature recognition method based on a causal relationship network, characterized in that, The rolling bearing composite fault feature is selected through the rolling bearing composite fault feature selection, so as to effectively identify the bearing composite fault; comprising: 1) constructing a rolling bearing composite fault feature causal relationship network model: A network based on a Markov blanket discovery algorithm is constructed, including mining the causal relationship structure between the composite fault features and constructing the composite fault feature causal relationship network; each composite fault feature is used as a target variable, and the Markov blanket discovery algorithm is used to search for the features having a causal relationship with the target variable; the composite fault features are used as nodes, and the causal relationship between the features is used as an edge, so that the fault feature causal relationship network is constructed; comprising: 11) composite fault feature extraction; The statistical quantities of the time domain type, the frequency domain type and the entropy value type of the composite fault signal are selected to constitute a composite fault feature set; wherein seven composite fault signal dimensioned time domain feature parameters, six composite fault signal dimensionless time domain feature parameters, five composite fault signal frequency domain features and two composite fault signal entropy value type parameters are included; 12) obtaining a Markov blanket set of each composite fault feature and identifying the causal structure between the composite fault features; 121) taking each composite fault feature as a target variable, obtaining the Markov blanket set of the current target variable; 1211) first, the Markov blanket set of the current target variable is initialized as an empty set; a candidate Markov blanket feature set is created and initialized as an empty set; 1212) all composite fault features in the composite fault feature set except the current target variable are put into the candidate Markov blanket feature set; 1213) the conditional independence of the composite fault features in the candidate Markov blanket feature set and the current target variable is calculated, and the candidate Markov blanket feature set and the Markov blanket set of the current target variable are updated according to the conditional independence; comprising: 12131) the correlation between each composite fault feature in the candidate Markov blanket feature set and the target variable is calculated, and the composite fault features are ranked in descending order according to the correlation; 12132) the conditional independence of the current target variable and the first ranked composite fault feature under the condition of the Markov blanket is calculated; it is judged whether the two are conditionally related or conditionally independent; If the two are conditionally related, the candidate Markov blanket feature set and the Markov blanket set of the current target variable are updated: the first ranked composite fault feature is removed from the candidate Markov blanket feature set and added to the Markov blanket set of the current target variable; If the two are conditionally independent, the next ranked composite fault feature is judged; Until each composite fault feature in the candidate Markov blanket feature set is traversed, the Markov blanket set of the current target variable is obtained; 1214) deleting the error nodes in the Markov blanket set of the current target variable, obtaining the final Markov blanket set; comprising: sequentially judging conditional independence of each feature in the Markov blanket set with the current target variable under the condition of remaining features; if the conditional independence is that the feature is an error node, the feature is removed from the Markov blanket set; if the conditional correlation exists, the next feature is judged until all features are traversed, and a final Markov blanket set of the current target variable is obtained; 13) constructing a composite fault feature causal relationship network model according to the causal structure among the composite fault features; each composite fault feature in the composite fault feature set is taken as a network node, and an edge is added between each composite fault feature and the composite fault features in the Markov blanket set of the composite fault feature, so that the causal relationship structure among the composite fault features is constructed into a composite fault feature causal relationship network; 2) constructing a composite fault diagnosis deep neural network model; calculating a complex network node centrality index; and screening a composite fault optimal feature subset; including: 21) calculating a complex network node centrality index for quantifying the importance of the composite fault features; 22) arranging each composite fault feature in ascending order according to the node centrality index value to obtain an ascending composite fault feature set; 23) increasing the number of composite fault features in sequence for the ascending composite fault feature set, inputting the corresponding composite fault features into the constructed composite fault diagnosis deep neural network model, and calculating the rolling bearing composite fault recognition accuracy; 24) taking the composite fault feature subset corresponding to the peak value of the accuracy as the composite fault optimal feature subset, as the node centrality index corresponding composite fault feature screening result; 25) obtaining the corresponding composite fault optimal feature subset by using different node centrality indexes respectively; 26) comparing the accuracies of different optimal feature subsets, and taking the optimal feature subset with the highest accuracy as the final composite fault optimal feature subset; Through the above steps, the final composite fault optimal feature subset is obtained, and the rolling bearing composite fault can be effectively identified.

2. The rolling bearing complex fault feature identification method based on the cause-effect network according to claim 1, characterized in that, The composite type of the rolling bearing composite fault feature is specifically: A1. Collecting rolling bearing vibration signals and performing pretreatment, mixing single faults of inner ring, outer ring and rolling elements of the rolling bearing two by two to form three types of composite fault signals of "inner ring + outer ring", "inner ring + rolling element" and "outer ring + rolling element"; A2. Each type of composite fault signal in A1 is divided into multiple groups to obtain multiple groups of fault signal data and form a composite fault data set; A3. Labeling the composite fault type label of each group of fault signal data in the composite fault data set in A2, including "inner ring + outer ring", "inner ring + rolling element" and "outer ring + rolling element".

3. The rolling bearing complex fault feature identification method based on the cause-effect network according to claim 2, characterized in that, In A2, each type of composite fault signal in A1 is divided into 2700 groups of fault signal data with 1024 as a group to form a composite fault data set.

4. The rolling bearing complex fault feature identification method based on the cause-effect network according to claim 1, characterized in that, In the composite fault feature set, the dimensioned time domain feature parameters of the composite fault signal include peak value, peak-to-peak value, average amplitude, root mean square, square root amplitude, variance and standard deviation; The dimensionless time domain characteristic parameters of the composite fault signal include: peak factor, pulse factor, margin factor, waveform factor, kurtosis factor and skewness; The frequency domain characteristics of the composite fault signal include: center of gravity frequency, mean square frequency, root mean square frequency, frequency variance and frequency standard deviation; The entropy type parameters of the composite fault signal include: power spectrum entropy and singular spectrum entropy.

5. The rolling bearing complex fault feature identification method based on the cause-effect network according to claim 1, characterized in that, Specifically, by G 2 The test method calculates the correlation between each compound failure feature in the candidate Markov blanket feature set and the target variable.

6. The rolling bearing complex fault feature identification method based on the cause-effect network according to claim 1, characterized in that, In step 13), a causal relationship network between the composite fault characteristics is constructed based on the composite fault characteristic set Fea; including the following specific process: 13.1 Adding nodes v1 ~ v 20 ; nodes v1 ~ v 20 represent composite failure features fea1 ~ fea 20 ; 13.2 Adding edges a to node v1 in the network using the Markov blanket set of feal 1j ; Including: 13.2.1 Determine whether node v2 and node v1 have an edge: if the feature corresponding to node v2 fea2 is in the Markov blanket set of feature fea1, add the edge between node v2 and node v1; if the feature corresponding to node v2 fea2 is not in the Markov blanket set of fea1, do not add the edge; 13.2.2 Repeat 13.2.1 to obtain edges between node v1 and nodes v3-v5, respectively; 20 between node v1 and nodes v3-v5, respectively; 13.3 Repeat the whole process of 13.2 to add features fea2 ~ fea 20 to the causal relationship network between the composite fault features. 20 2j 20j .​​ 7. The rolling bearing complex fault feature identification method based on the cause-effect network according to claim 1, characterized in that, The complex network node centrality index for quantifying the importance of the composite fault characteristics includes: degree centrality, closeness centrality, betweenness centrality and Pagerank centrality of each composite fault characteristic in the feature causal relationship network.

8. The rolling bearing complex fault feature identification method based on the cause-effect network according to claim 1, characterized in that, The method for calculating the rolling bearing composite fault recognition accuracy is achieved by constructing a deep neural network model, which includes: an input layer, a first hidden layer, a second hidden layer and an output layer; the feature set of the test sample candidate scheme is input into the model input layer, and the prediction result of the test sample fault type is output; the rolling bearing composite fault recognition accuracy is the proportion of the number of samples correctly predicted by the model to the total number of samples in the test set.

9. The rolling bearing complex fault feature identification method based on the cause-effect network according to claim 8, characterized in that, The first hidden layer and the second hidden layer of the composite fault diagnosis deep neural network model both adopt the Relu activation function; the output layer adopts the sigmoid activation function; the number of neurons in the model input layer is set to the number of features in the optimal feature subset.

10. The rolling bearing complex fault feature identification method based on a cause-effect network according to claim 1, characterized in that, Specifically, the IAMB (Incremental Association Markov Blanket) is used to test the causal relationship between the bearing composite fault characteristics.

Images