Rolling bearing composite fault prediction feature selection method based on high-order suppression sorting algorithm

Abstract

The invention provides a full-life-cycle composite fault prediction feature selection method based on a high-order suppression sorting algorithm. The method comprises the following steps: 1) constructing a rolling bearing full-life-cycle composite fault feature data set FD; 2) dividing the whole life cycle of the rolling bearing into three stages of health, degeneration and failure by using the feature data set FD; 3) constructing a causal feature weighted high-order network according to the feature data set FD; 4) constructing high-order PageRank value sets HPR [alpha], HPR [beta] and HPR [gamma] of the three networks G [alpha], G [beta] and G [gamma]; 5) constructing a feature subset Fri for features in the three state data sets FD alpha, FD beta and FD gamma of the rolling bearing according to a high-order PageRank value; 6) performing fault prediction and feature selection; according to the invention, aiming at the physical characteristics of bearing faults, multiple index characteristics of signal data are extracted to construct a causal characteristic weighted high-order network, and an improved centrality index PageRank is used to select fault prediction key characteristics. According to the method, more accurate fault prediction can be realized by using fewer features, and the method has relatively high application value and popularization value.

Description

Technical Field

[0001] The invention relates to the field of bearing composite fault prediction, and in particular to a rolling bearing composite fault prediction feature selection method based on a high-order suppression sorting algorithm, and belongs to the field of mechanical fault diagnosis. Background Art

[0002] As a key component of rotating machinery, rolling bearings are composed of multiple parts and are widely used in modern mechanical equipment. Effectively predicting their health status is crucial to ensuring the stable operation of the equipment. However, due to the complex and changeable environment in which the equipment is located in actual working conditions, rolling bearing failures are often not a single failure, but multiple failures exist at the same time, which not only increases the difficulty of fault prediction, but also poses a serious threat to the reliability and stability of the equipment. In fact, with the development of data element flow technology, in order to improve the effect of fault prediction, engineering generally collects and extracts features from vibration signals from multiple angles to ensure that multi-domain features can reflect fault information under various states of the signal. Selecting multi-dimensional features of composite fault signals is a key link in fault prediction.

[0003] In order to solve the problem of feature selection, the existing technology attempts to use the high-order structure of the network based on complex network theory. In patent CN202311428334.4, Lv Linyuan et al. used a fully connected single-layer perceptron to map the embedding matrix Y of the central layer cell cavity to a one-dimensional vector S h , based on the influence score S of the h-order cell h And calculate the influence ranking matrix R h , determine the important high-order structures in the data; Li Cong et al. in patent CN202211282958.5 established a high-order correlation matrix Sn-1,n between n-1-order groups and n-order groups, and were able to calculate various high-order indicators such as PageRank, thereby screening out the most important group structure, that is, the sequence of key groups; Nan Lu et al. in patent CN202310453531.5 set up a virtual circulation to redirect the active power transmission path between the generator node and the load node from the load node to the generator node, and set the weight of the virtual circulation according to the specific transaction volume to improve the PageRank algorithm, which is used to identify vulnerable lines in the power grid. These ideas of using high-order structures provide new ideas for feature selection. However, these existing methods focus on the identification of important features, have redundant high-dimensional feature indicators, high computational complexity and may reduce recognition accuracy, and are difficult to select effective bearing fault features. Summary of the invention

[0004] The present invention proposes a full life cycle composite fault prediction feature selection method based on a high-order suppression sorting algorithm. Aiming at the physical characteristics of bearing faults, vibration data from multiple angles is collected, and various index features of signal data are extracted to form network structure characteristics. Redundant and irrelevant index features are identified and suppressed according to high-order structural attributes, and key features with large contribution to bearing faults are retained, so as to achieve more accurate fault prediction effect with fewer features. The method comprises: 1) constructing a full life cycle composite fault feature data set FD of rolling bearings; 2) using the feature data set FD, dividing the full life cycle of rolling bearings into three stages: health, degradation, and failure; 3) constructing a causal feature weighted high-order network according to the feature data set FD; 4) constructing three networks G α , G β , G γ The higher-order PageRank value set HPR α , HPR α and HPR γ ; 5) Rolling bearing three state data sets FD α , FD β , FD γ According to the high-order PageRank value, the feature subset Fr is constructed. i ; 6) performing fault prediction and feature selection; Specifically, the method of the present invention comprises the following steps:

[0005] A. Construct the composite fault feature dataset FD of rolling bearings throughout their life cycle. The specific steps are as follows:

[0006] A1. Obtain the composite fault signal data of the rolling bearing throughout its life cycle to form a signal data set SD. The specific steps are as follows:

[0007] A1.1 Select c data sampled in the entire life cycle from the horizontal vibration signal data of the rolling bearing with compound faults;

[0008] A1.2 uses the sliding window method to construct the sample set. The window size is L, the step length is M, and a total of Z groups of samples are obtained in the entire life cycle to form the signal data set SD. The calculation formula is as follows:

[0009]

[0010] A2. Define n features consisting of time domain, frequency domain and entropy value, and calculate the value of n features for each group of samples in the signal data set SD. The specific steps are as follows:

[0011] A2.1 defines h1 dimensional and dimensionless time domain features, and calculates the time domain features of each group of samples. The specific contents are as follows:

[0012] A2.1.1 Dimensional characteristics are peak value, peak-to-peak value, mean amplitude, RMS value, RMS amplitude, variance, and standard deviation;

[0013] A2.1.2 The dimensionless characteristics are peak factor, impulse factor, margin factor, shape factor, kurtosis factor, and skewness;

[0014] A2.2 Perform Fourier transform on each group of samples and calculate h2 frequency domain features of each group of samples, including centroid frequency, mean square frequency, root mean square frequency, frequency variance and frequency standard deviation;

[0015] A2.3 Calculate the power spectrum entropy and singular spectrum entropy of each group of samples, and the number of features is recorded as h3;

[0016] A3. Use the Z*n values ​​of the Z samples to construct the feature data set FD. The specific steps are as follows:

[0017] A3.1 The feature data set FD has n vectors, each of which contains Z values;

[0018] A3.2 The i-th vector is represented by fea i Indicates, recorded as Each element is the value of the i-th feature of the k-th sample;

[0019] B. Using the feature data set FD, the entire life cycle of the rolling bearing is divided into three stages: health, degradation, and failure. The specific steps are as follows:

[0020] B1. Use the values ​​of Z*h1 time domain features to find the inflection point T1 and identify the health status αα of the rolling bearing and its sample number. The specific steps are as follows:

[0021] B1.1 According to the amplitude of the Z*h1 time domain characteristic values ​​obtained in step A2.1 over the entire life cycle, find the first inflection point T1. The specific steps are as follows:

[0022] B1.1.1 For the samples before this inflection point, the amplitude of the dimensional characteristics has almost no fluctuation, while the dimensionless characteristics have repeated small fluctuations;

[0023] B1.1.2 For samples after T1, the amplitude of dimensional features increased significantly and fluctuated repeatedly, and the amplitude of dimensionless features increased or decreased sharply and fluctuated repeatedly;

[0024] B1.2 The number of samples between 0 and T1 is recorded as z1. During the corresponding time period, the rolling bearing is in a healthy state, recorded as α state;

[0025] B2. Use the values ​​of Z*h1 time domain features to find the inflection point T2 and identify the degradation state β of the rolling bearing and its sample number. The specific steps are as follows:

[0026] B2.1 According to the amplitude of the Z*h1 time domain characteristic values ​​obtained in step A2.1 over the entire life cycle, find the second inflection point T2. The specific steps are as follows:

[0027] B2.1.1 Compared with the samples before T1, the amplitude of the dimensional characteristics of the samples between T1 and T2 increased significantly and fluctuated repeatedly, and the dimensionless characteristics increased or decreased sharply and fluctuated repeatedly;

[0028] B2.1.2 For samples after T2, the amplitude of the dimensional feature increases sharply without fluctuation, and the amplitude of the dimensionless feature increases or decreases sharply without fluctuation;

[0029] B2.2 The number of samples from T1 to T2 is recorded as z2. During the corresponding time period, the rolling bearing is in a degraded state, recorded as β state;

[0030] B3. The number of samples from T2 to the end of the life cycle is recorded as z3. During the corresponding time period, the rolling bearing is in failure state Y;

[0031] c. Construct a causal feature weighted high-order network based on the feature data set FD. The specific steps are as follows:

[0032] c1. According to the inflection points T1 and T2 obtained in step B, the feature data set FD is divided into three data sets. The specific steps are as follows:

[0033] C1.1 There are n vectors in the feature data set FD. Each vector is divided into three parts, denoted as

[0034] C1.2 Constructing the health feature dataset FD α , denoted as

[0035] C1.3 Constructing the degradation feature data set FD β , denoted as

[0036] C1.4 Constructing the failure feature data set FD γ , denoted as

[0037] C2. Using three feature datasets FD α , FD β , FD γ Construct causal feature weighted high-order networks for the three states of health, degradation, and failure respectively, and record them as G α , G β , G γ ;

[0038] The three causal feature high-order networks are uniformly denoted as G * = {V * , E * , W *}, *∈{α,β,γ};

[0039] Among them, V * For network G * The node set of Corresponding network G * A node in

[0040] E * For network G * The edge set of If two features and There is a causal relationship between them, then in the network G * Nodes in and There is an edge between

[0041] The specific steps are as follows:

[0042] C2.1 Building Network G α , G β , G γ The specific steps are as follows:

[0043] C2.1.1 Based on G α For example, traverse FD α For each node in the set, get FD α The edge set The specific steps are as follows:

[0044] C2.1.1.1 Using the incremental correlation Markov blanket algorithm, calculate The Markov blanket set

[0045] C2.1.1.2 If the node of There are nodes in Then at the node and nodes Add edges between Otherwise, no edge is added;

[0046] C2.1.2 FD β , FD γ Repeat step C2.1.1 to obtain FD β and FD γThe edge set and

[0047] C2.2 Calculate the absolute value of the Person correlation coefficient between the features as G α , G β , G γ The edge weights between corresponding nodes in the network are recorded as

[0048] C2.3 Identify network G * The higher-order structure in G α Take this as an example, the specific steps are as follows:

[0049] C2.3.1 vs. G α The edge set E α and the node set V α , find the node neighbors, i∈[1,n];

[0050] C2.3.2 Finding Nodes Whether there are common neighbors between them and their neighbors. If so, the three points form a higher-order structure, namely a 2-simplex.

[0051] C2.3.3 Repeat steps C2.3.1 and C2.3.2 until all nodes are traversed and all 2-simplex structures are identified;

[0052] C2.3.4G α The set of all 2-simplices in G α All 1-simplices in

[0053] C2.4 vs. G β , G γ The edge set E β , E γ and the node set V β , V γ Repeat step C2.3 to obtain the causal feature weighted high-order network G β and G γ All 2-simplex and 1-simplex structures in ;

[0054] D. Construct three networks G α , G β , G γ The high-order PageRank value set HPRα, HPR β and HPR γ , the specific steps are as follows:

[0055] D1.Gα For example, generate G α The adjacency matrix of Among them, if With point If there is an edge between Otherwise, 0;

[0056] D2. According to the results obtained in step C2.3, statistics G α The distribution of the 2-simplex of the nodes in the figure divides the neighbors of the nodes into two categories. The specific steps are as follows:

[0057] D2.1 If the node With its neighbors belong to the same 2-simplex, then in Defined as network G α middle The set of high-order neighbors of ;

[0058] D2.2 If the node With its neighbors Only 1-simplex is formed, then in For network G α middle The set of non-high-order neighbors of ;

[0059] D3. Optional parameter θ (0≤θ<1), calculate G α The high-order PageRank value of each node in the , and the high-order PageRank value set HPR α , the specific formula is as follows:

[0060]

[0061] where d is the damping coefficient, For Node The out-degree of n is the network G * Total number of nodes, is the adjacency matrix A * The elements in For Node The high-order PageRank value at the tth iteration;

[0062] The balance function f(θ)(0≤θ<1) is as follows:

[0063]

[0064] in for The set of high-order neighbors of for The set of non-high-order neighbors of is the proportionality coefficient;

[0065] D4. vs. G β and G γ Repeat steps D1, D2 and D3 to obtain the network's high-order PageRank value set HPR β and HPR γ ;

[0066] E. The characteristic data set FD of the three states of the rolling bearing α , FD β , FD γ According to the high-order PageRank value, the feature subset Fr is constructed. i , the following i, j, l∈[1, n], the specific steps are as follows:

[0067] E1. According to HPR α , HPR β and HPR γ The values ​​in are sorted in ascending order from small to large, and the sorting result is recorded as F α , F β , F γ ;

[0068] E2. Based on F α 、F β 、F γ The absolute ranking result of feature i is calculated by the following formula:

[0069]

[0070] in For feature i in F α , F β , F γ The serial number in

[0071] E3. i Arrange in ascending order from smallest to largest;

[0072] E4. Sequentially change r i The features ranked in the first i in the feature subset Fr i , that is, the first feature is the first feature subset Fr 1 , the first two features are the second feature subset Fr 2 , the first three features are the third feature subset Fr 3 ; Similarly, if there are n features in total, there are n feature subsets;

[0073] F. Perform fault prediction and feature selection. The specific steps are as follows:

[0074] F1. Use deep neural network as prediction model and select feature subset Fr i , the corresponding FD α , F.D. β , F.D. γ 70% of the parts are selected as training sets, 20% are selected as validation sets, and 10% are selected as prediction sets. i The corresponding feature data set is input into the deep neural network model, and the model is trained, verified and predicted. 1 Take this as an example, the specific steps are as follows:

[0075] F1.1 will be Fr 1 The part divided into the training set is input into the deep neural network model for training, and the part divided into the validation set is used for validation:

[0076] F1.2 Use the trained deep neural network model and test set data to predict the composite fault state of rolling bearings and calculate relevant prediction indicators; select the average correlation coefficient, root mean square error, and absolute average error as composite fault prediction performance evaluation indicators;

[0077] F1.3 for the remaining n-1 feature subsets Fr i Repeat steps F1.1 and F1.2 to obtain n groups of prediction performance evaluation indicators; the closer the average correlation coefficient is to 1, the better; the closer the root mean square error and absolute mean error are to 0, the better;

[0078] F1.4 takes the feature subset with the best evaluation index as the optimal feature subset under this parameter θ;

[0079] F2. Change the parameter θ (0≤θ<1) (θ takes Q different values), repeat steps D, E and F1, and obtain Q optimal feature subsets under different parameters θ;

[0080] F3. Compare the number of features in the Q optimal feature subsets with the performance evaluation index, and take the parameter θ when the performance evaluation index is optimal as the optimal parameter. The feature subset at this time is the final optimal feature subset.

[0081] The present invention is aimed at the field of bearing fault prediction, combined with the current cross-domain flow technology of data elements, taking into account the redundancy of high-order structures in the network and the uniqueness of low-order structures of composite fault information, and using the improved centrality index PageRank as the basis for judging the importance of features, suppressing the importance of high-order structural features, and improving the importance weights of low-order features, so as to achieve efficient feature selection and obtain prediction results through DNN. The prediction results are evaluated according to the performance evaluation index. Compared with the traditional feature selection method and the centrality index method that only relies on the low-order structure of the network, more accurate fault prediction is achieved with fewer features, which has higher application value and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The accompanying drawings are used to provide further understanding of the technical solution of the present invention and constitute a part of the specification. Together with the implementation of the present method, they are used to explain the technical solution of the present invention and do not constitute a limitation of the technical solution of the present invention.

[0083] The accompanying drawings are as follows:

[0084] Figure 1 A flow chart of the method of the present invention; DETAILED DESCRIPTION

[0085] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0086] The method flow chart is as follows Figure 1 As shown, the method of the present invention includes 1) constructing a composite fault feature data set FD of the whole life cycle of the rolling bearing; 2) using the feature data set FD, dividing the whole life cycle of the rolling bearing into three stages: health, degradation, and failure; 3) constructing a causal feature weighted high-order network based on the feature data set FD; 4) constructing three networks G α , G β , G γ The higher-order PageRank value set HPR α , HPR α and HPR γ ; 5) Rolling bearing three state data sets FD α , FD β , FD γ According to the high-order PageRank value, the feature subset Fr is constructed. i ; 6) Fault prediction and feature selection, the specific implementation steps are as follows:

[0087] 1. Construct a composite fault feature dataset FD for the entire life cycle of rolling bearings. The specific steps are as follows:

[0088] 1.1. Obtain the composite fault signal data of the rolling bearing throughout its life cycle to form a signal data set SD. The specific steps are as follows:

[0089] 1.1.1 From the horizontal vibration signal data of the rolling bearing with compound faults, a total of 1,664,768 valid data samples were selected within 52 minutes of the entire life cycle;

[0090] 1.1.2 The sliding window method is used to obtain samples. The window size is L = 1024, the step size is M = 256, and a total of Z = 6500 groups of samples are obtained to form the signal data set SD. The formula is as follows:

[0091]

[0092] 1.2. Define 20 features consisting of time domain, frequency domain and entropy value, and calculate the values ​​of 20 features for each group of samples in the signal data set SD. Some samples are as follows:

[0093] -0.6089 -0.3597 0.5511 0.4081 -0.5916 -1.2837 -0.4994

[0094] Some feature calculation formulas, where x z 1024 data representing the Zth group of samples:

[0095] Feature Name Calculation formula Peak-to-Peak <![CDATA[max(x z )-min(x z )]]> Peak <![CDATA[max(|x z |)]]>

[0096] Some characteristic calculation results:

[0097] Feature Name Calculation results Peak-to-Peak 6.1563 Peak 3.0938

[0098] 1.3. Using 130,000 values ​​of 6,500 samples, construct the feature data set FD. Some data set examples are as follows; the specific steps are as follows:

[0099] 1.3.1 Feature data set FD has 20 vectors, each containing Z = 6500 values;

[0100] 1.3.2 The i-th vector uses fea i Indicates, recorded as Each element is the value of the i-th feature of the k-th sample; some examples of data sets are as follows:

[0101] <![CDATA[fea 1 ]]> <![CDATA[fea 2 ]]> <![CDATA[fea 3 ]]> <![CDATA[fea 4 ]]> <![CDATA[fea 5 ]]> <![CDATA[fea 6 ]]> <![CDATA[fea 7 ]]> <![CDATA[fea 8 ]]> <![CDATA[fea 9 ]]> <![CDATA[fea 10 ]]> 1.005 6.156 3.093 0.813 1.010 1.005 0.699 0.0489 3.0774 1.2357 0.978 6.156 3.093 0.786 0.957 0.978 0.670 0.0599 3.161 1.2449 0.988 5.998 3.062 0.794 0.977 0.988 0.675 0.034 3.0979 1.2439 1.014 6.055 3.119 0.808 1.029 1.014 0.685 -0.1067 3.0744 1.2552

[0102] 2. Using the feature data set FD, the entire life cycle of the rolling bearing is divided into three stages: health, degradation, and failure. The specific steps are as follows:

[0103] 2.1. Use the values ​​of 13 time domain features to find the time inflection point T1=34, identify the health state αα of the rolling bearing and its sample number, and the number of samples from 0 to T1 is recorded as z1=4250;

[0104] 2.2. Using the values ​​of 13 time domain features, find the time inflection point T2=47, identify the health state β of the rolling bearing and its sample number, and the number of samples from T1 to T2 is recorded as z2=1625;

[0105] 2.3. The number of samples from T2 to the end of the life cycle is recorded as z3 = 625. During the corresponding time period, the rolling bearing is in failure state γ;

[0106] 3. Construct a causal feature weighted high-order network based on the feature data set FD. The specific steps are as follows:

[0107] 3.1. According to the inflection points T1 and T2 obtained in step B, the feature data set FD is divided into three data sets. The specific steps are as follows:

[0108] 3.1.1 Each of the 20 vectors in the feature data set FD is divided into three parts, denoted as

[0109] 3.1.2 Constructing the health feature dataset FD α , denoted as

[0110] 3.1.3 Constructing the degradation feature data set FD β , denoted as

[0111] 3.1.4 Constructing the failure feature data set FD γ , denoted as

[0112] 3.2 Using three feature datasets FD α , FD β , FD γ Construct causal feature weighted high-order networks for the three states of health, degradation, and failure respectively, and record them as G α , G β , G γ ; The specific steps are as follows;

[0113] 3.2.1 Building Network G α , G β , G γ The specific steps are as follows:

[0114] 3.2.1.1 Based on G α For example, traverse FD α For each node in the set, get FD α The edge set The specific steps are as follows:

[0115] 3.2.1.1.1 Using the incremental correlation Markov blanket algorithm, calculate The Markov blanket set

[0116] 3.2.1.1.2 Based on Node For example, if the node of There are nodes in Then at the node and nodes Add edges between Otherwise, no edge is added;

[0117] 3.2.1.1.3 FD β , FD γ Repeat steps 3.2.1.1.1 and 3.2.1.1.2 to obtain FD β and FD γ The edge set and

[0118] 3.2.2 Calculate the absolute value of the Person correlation coefficient between the features as G α , G β , G γ The edge weights between nodes in the network are denoted by

[0119] 3.2.3 Identify the high-order structure in the network, using G α middle Take this as an example, the specific steps are as follows:

[0120] 3.2.3.1 For G α The edge set E α and the node set V α , find the node Neighbors of, i∈[1,20];

[0121] 3.2.3.2 Finding Nodes With its neighbors Are there common neighbors between them? If so, for example, for and common neighbors, then these three points are a high-order structure That is, 2-simplex;

[0122] 3.2.3.3 Repeat steps 3.2.3.1 and 3.2.3.2 until all nodes are traversed and all 2-simplex structures are identified;

[0123] 3.2.3.4Gα The set of all 2-simplices in G α All 1-simplices in The specific results are as follows:

[0124]

[0125] 3.2.4 FD β , FD γ The edge set E β , E γ and the node set V β , V γ Repeat steps 3.2.3 to obtain the causal feature weighted high-order network G β and G γ All 2-simplex and 1-simplex structures in ;

[0126] 4. Build three networks G α , G β , G γ The higher-order PageRank value set HPR α , HPR α and HPR γ ; The specific implementation steps are as follows:

[0127] 4.1. G α For example, generate G α The adjacency matrix of Among them, if With point If there is an edge between Otherwise, 0;

[0128] 4.2. Calculate G based on the results obtained in step 3.2.3 α The distribution of 2-simplex of nodes in the network divides the neighbors of nodes into two categories. The distribution of high-order structure and non-high-order structure of nodes in the three networks is as follows:

[0129]

[0130]

[0131] 4.3. Select parameter θ = 0.1, calculate the high-order PageRank value of each node in Gα, and obtain the high-order PageRank value set HPRα. Some results are as follows:

[0132]

[0133] 4.4. For G β and G γRepeat steps 4.1, 4.2 and 4.3 to obtain the network's high-order PageRank value set HPR β and HPR γ ;

[0134] 5. The three state data sets of rolling bearings are FD α , FD β , FD γ According to the high-order PageRank value, the feature subset Fr is constructed. i , the following i, j, l∈[1, 20], the specific steps are as follows:

[0135] 5.1. According to HPR α , HPR β and HPR γ The values ​​in are sorted in ascending order from small to large, and the sorting result is recorded as F α , F β , F γ , some of the sorting results are as follows:

[0136] 1 2 3 4 5 <![CDATA[F α ]]> <![CDATA[fea 2 ]]> <![CDATA[fea 8 ]]> <![CDATA[fea 10 ]]> <![CDATA[fea 11 ]]> <![CDATA[fea 16 ]]> <![CDATA[F β ]]> <![CDATA[fea 20 ]]> <![CDATA[fea 6 ]]> <![CDATA[fea 4 ]]> <![CDATA[fea 18 ]]> <![CDATA[fea 16 ]]> <![CDATA[F γ ]]> <![CDATA[fea 7 ]]> <![CDATA[fea 12 ]]> <![CDATA[fea 5 ]]> <![CDATA[fea 6 ]]> <![CDATA[fea 4 ]]>

[0137] 5.2. According to F α 、F β 、F γ Sorting, calculating the feature fea i The absolute sorting results are as follows:

[0138] <![CDATA[fea 1 ]]> <![CDATA[fea 2 ]]> <![CDATA[fea 3 ]]> <![CDATA[fea 4 ]]> <![CDATA[fea 5 ]]> <![CDATA[fea 6 ]]> <![CDATA[fea 7 ]]> <![CDATA[fea 8 ]]> <![CDATA[fea 9 ]]> <![CDATA[fea 10 ]]> 2 4 3 7 15 16 14 1 10 11 <![CDATA[fea 11 ]]> <![CDATA[fea 12 ]]> <![CDATA[fea 13 ]]> <![CDATA[fea 14 ]]> <![CDATA[fea 15 ]]> <![CDATA[fea 16 ]]> <![CDATA[fea 17 ]]> <![CDATA[fea 18 ]]> <![CDATA[fea 19 ]]> <![CDATA[fea 20 ]]> 8 6 9 17 18 5 12 13 20 19

[0139] 5.3. Arrange ri in ascending order from small to large, the result is as follows;

[0140] <![CDATA[fea 8 ]]> <![CDATA[fea 1 ]]> <![CDATA[fea 3 ]]> <![CDATA[fea 2 ]]> <![CDATA[fea 16 ]]> <![CDATA[fea 12 ]]> <![CDATA[fea 4 ]]> <![CDATA[fea 11 ]]> <![CDATA[fea 13 ]]> <![CDATA[fea 9 ]]> 1 2 3 4 5 6 7 8 9 10 <![CDATA[fea 10 ]]> <![CDATA[fea 17 ]]> <![CDATA[fea 18 ]]> <![CDATA[fea 7 ]]> <![CDATA[fea 5 ]]> <![CDATA[fea 6 ]]> <![CDATA[fea 14 ]]> <![CDATA[fea 15 ]]> <![CDATA[fea 20 ]]> <![CDATA[fea 19 ]]> 11 12 13 14 15 16 17 18 19 20

[0141] 5.4. Sequentially arrange the features in the first i of ri as feature subset Fr i , that is, the first feature is the first feature subset Fr 1 , the first two features are the second feature subset Fr 2 , the first three features are the third feature subset, and so on, there are 20 feature subsets for a total of 20 features; the first 5 feature subsets are as follows:

[0142] <![CDATA[Fr 1 ]]> <![CDATA[fea 8 ]]> <![CDATA[Fr 2 ]]> <![CDATA[fea 8 ]]> <![CDATA[fea 1 ]]> <![CDATA[Fr 3 ]]> <![CDATA[fea 8 ]]> <![CDATA[fea 1 ]]> <![CDATA[fea 3 ]]> <![CDATA[Fr 4 ]]> <![CDATA[fea 8 ]]> <![CDATA[fea 1 ]]> <![CDATA[fea 3 ]]> <![CDATA[fea 2 ]]> <![CDATA[Fr 5 ]]> <![CDATA[fea 8 ]]> <![CDATA[fea 1 ]]> <![CDATA[fea 3 ]]> <![CDATA[fea 2 ]]> <![CDATA[fea 16 ]]>

[0143] 6. Perform fault prediction and feature selection. The specific implementation steps are as follows:

[0144] 6.1. A deep neural network with 4 hidden layers is used as the prediction model. The learning rate is set to 0.0001, the batch size is 64, and the feature subset Fri is selected. The corresponding FD α , F.D. β , F.D. γ 70% of the parts are selected as training sets, 20% are selected as validation sets, and 10% are selected as prediction sets. i The corresponding data set is input into the deep neural network model, and the model is trained, verified and predicted. 1 Take this as an example, the specific steps are as follows:

[0145] 6.1.1 Fr 1 Belongs to FD α The part in The part divided into the training set is input into the deep neural network model for training and Divide into parts of the validation set for validation;

[0146] 6.1.2 Use the trained deep neural network model and the test set data to predict the composite fault state of the rolling bearing and calculate the relevant prediction indicators, that is, select the average correlation coefficient, root mean square error, and absolute average error as composite fault prediction performance evaluation indicators. The results are as follows:

[0147]

[0148]

[0149] 6.1.3 For the remaining 19 feature subsets Fr i Repeat steps 6.1.1 and 6.1.2 to obtain 20 sets of prediction indicators. Some of the results are as follows:

[0150]

[0151] 6.1.4 The feature subset with the best evaluation index is taken as the optimal feature subset under this parameter θ:

[0152] <![CDATA[Fr 10 ]]> Average correlation coefficient 0.8589 Root mean square error 7.1742 Absolute mean error 0.6348

[0153] 6.2. Take the value of parameter θ as 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9 and repeat steps 4, 5 and 6.1 to obtain 10 optimal feature subsets under different parameters θ. The values ​​of some optimal feature subsets and their performance evaluation indicators are as follows:

[0154] θ=0.1 θ=0.2 θ=0.3 θ=0.4 θ=0.5 Average correlation coefficient 0.8589 0.9637 0.984 0.9722 0.9961 Root mean square error 7.1742 1.8454 0.8159 1.4115 0.1964 Absolute mean error 0.6348 0.2107 0.1638 0.2874 0.1071 Optimal feature subset <![CDATA[Fr 10 ]]> <![CDATA[Fr 7 ]]> <![CDATA[Fr 6 ]]> <![CDATA[Fr 14 ]]> <![CDATA[Fr 11 ]]>

[0155] 6.3. Compare the number of features and performance evaluation index in the 10 optimal feature subsets, and take the parameter θ when the performance evaluation index is optimal as the optimal parameter. The feature subset at this time is the final optimal feature subset. The results are as follows:

[0156] <![CDATA[θ=0.5,Fr 11 ]]> Average correlation coefficient 0.9961 Root mean square error 0.1964 Absolute mean error 0.1071

[0157] Compared with the prior art, the present invention has the following advantages: it takes into account the redundancy of high-order structures and uniqueness of low-order structures of composite fault information in the network, improves the centrality index PageRank so that it can suppress the importance of high-order structural features; obtains prediction results through DNN, and evaluates the prediction results through performance evaluation indicators, which can achieve more accurate fault prediction with fewer features, and has high application value and promotion value.

[0158] Finally, it should be noted that the purpose of publishing the implementation is to help further understand the present invention, but those skilled in the art can understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection claimed by the present invention shall be subject to the scope defined in the claims.

Claims

1. A rolling bearing composite fault prediction feature selection method based on a high-order suppression sorting algorithm, the steps comprising: A. Construct the composite fault feature dataset FD of rolling bearings throughout their life cycle. The specific steps are as follows: A1. Obtain the composite fault signal data of the rolling bearing throughout its life cycle to form a signal data set SD. The specific steps are as follows: A1.1 Select c data sampled in the entire life cycle from the horizontal vibration signal data of the rolling bearing with compound faults; A1.2 uses the sliding window method to construct the sample set. The window size is L, the step length is M, and a total of Z groups of samples are obtained in the entire life cycle to form the signal data set SD. The calculation formula is as follows: A2. Define n features consisting of time domain, frequency domain and entropy value, and calculate the value of n features for each group of samples in the signal data set SD; A3. Use the Z*n values ​​of the Z samples to construct the feature data set FD. The specific steps are as follows: A3.1 The feature data set FD has n vectors, each of which contains Z values; A3.2 The i-th vector is represented by fea i Indicates that i∈[1,n],k∈[1,Z], each element is the value of the i-th feature of the k-th sample; B. Using the feature data set FD, the entire life cycle of the rolling bearing is divided into three stages: health, degradation, and failure. The specific steps are as follows: B1. Use the values ​​of Z*h1 time domain features to find the inflection point T1, and identify the health state α of the rolling bearing and its sample number; B2. Use the values ​​of Z*h1 time domain features to find the inflection point T2, and identify the degradation state β of the rolling bearing and its sample number; B3. The number of samples from T2 to the end of the life cycle is recorded as z3, and during the corresponding time period, the rolling bearing is in failure state γ; C. Construct a causal feature weighted high-order network based on the feature data set FD. The specific steps are as follows: C1. Divide the feature data set FD into three data sets according to the inflection points T1 and T2 obtained in step B; C2. Using three feature datasets FD α , FD β , FD γ Construct causal feature weighted high-order networks for the three states of health, degradation, and failure respectively, and record them as G α , G β , G γ ; The three causal feature high-order networks are uniformly denoted as G * = {V * , E * , W * }, *∈{α,β,γ}; Among them, V * For network G * The node set of Corresponding network G * A node in E * For network G * The edge set of If two features and There is a causal relationship between them, then in the network G * Nodes in and There is an edge between The specific steps are as follows: C2.1 Building Network G α , G β , G γ The edge set of ; C2.2 Calculate the absolute value of the Person correlation coefficient between the features as G α , G β , G γ The edge weights between corresponding nodes in the network are recorded as i, j ∈ [1, n]; C2.3 Identify network G * The higher-order structure in C2.4 vs. G β , G γ The edge set E β , E γ and the node set V β , V γ Repeat step C2.3 to obtain the causal feature weighted high-order network G β and G γ All 2-simplex and 1-simplex structures in ; D. Construct three networks G α , G β , G γ The higher-order PageRank value set HPR α , HPR β and HPR γ , the specific steps are as follows: D1.G α For example, generate G α The adjacency matrix of Among them, if With point If there is an edge between Otherwise, 0; D2. According to the results obtained in step C2.3, statistics G α The distribution of the 2-simplex of the nodes in the figure divides the neighbors of the nodes into two categories. The specific steps are as follows: D2.1 If the node With its neighbors belong to the same 2-simplex, then in Defined as network G α middle The set of high-order neighbors of ; D2.2 If the node With its neighbors Only 1-simplex is formed, then in For network G α middle The set of non-high-order neighbors of ; D3. Optional parameter θ (0≤θ<1), calculate G α The high-order PageRank value of each node in the , and the high-order PageRank value set HPR α , the specific formula is as follows: Where d is the damping coefficient, For Node The out-degree of n is the network G * Total number of nodes, is the adjacency matrix A * The elements in For Node The high-order PageRank value at the tth iteration; The balance function f(θ)(0≤θ<1) is as follows: in for The set of high-order neighbors of for The set of non-high-order neighbors of is the proportionality coefficient; D4. vs. G β and G γ Repeat steps D1, D2 and D3 to obtain the network's high-order PageRank value set HPR β and HPR γ ; E. The characteristic data set FD of the three states of the rolling bearing α , FD β , FD γ According to the high-order PageRank value, the feature subset Fr is constructed. i , the following i, j, l∈[1, n], the specific steps are as follows: E1. According to HPR α , HPR β and HPR γ The values ​​in are sorted in ascending order from small to large, and the sorting result is recorded as F α , F β , F γ ; E2. Based on F α 、F β 、F γ The absolute ranking result of feature i is calculated by the following formula: in For feature i in F α , F β , F γ The serial number in E3. i Arrange in ascending order from smallest to largest; E4. Sequentially change r i The features ranked in the first i in the feature subset Fr i , that is, the first feature is the first feature subset Fr1, the first two features are the second feature subset Fr2, and the first three features are the third feature subset Fr3; and so on, if there are n features in total, there are n feature subsets; F. Perform fault prediction and feature selection. The specific steps are as follows: F1. Use deep neural network as prediction model and select feature subset Fr i , the corresponding FD α , F.D. β , F.D. γ 70% of the parts are selected as training sets, 20% are selected as validation sets, and 10% are selected as prediction sets. i The corresponding feature data set is input into the deep neural network model, and the model is trained, verified and predicted. Taking Fr1 as an example, the specific steps are as follows: F1.1 inputs the part of Fr1 divided into the training set into the deep neural network model for training, and uses the part divided into the validation set for validation; F1.2 Use the trained deep neural network model and test set data to predict the composite fault state of rolling bearings and calculate relevant prediction indicators; select the average correlation coefficient, root mean square error, and absolute average error as composite fault prediction performance evaluation indicators; F1.3 for the remaining n-1 feature subsets Fr i Repeat steps F1.1 and F1.2 to obtain n groups of prediction performance evaluation indicators; the closer the average correlation coefficient is to 1, the better; the closer the root mean square error and absolute mean error are to 0, the better; F1.4 takes the feature subset with the best evaluation index as the optimal feature subset under this parameter θ; F2. Change parameter θ (0≤θ<1) (θ takes Q different values), repeat steps D, E and F1, and obtain Q optimal feature subsets under different parameters θ; F3. Compare the number of features in the Q optimal feature subsets with the performance evaluation index, and take the parameter θ when the performance evaluation index is optimal as the optimal parameter. The feature subset at this time is the final optimal feature subset.

2. A rolling bearing composite fault prediction feature selection method based on a high-order suppression sorting algorithm as claimed in claim 1, defining n features consisting of time domain, frequency domain and entropy value, and calculating the value of n features for each group of samples in the signal data set SD, the specific steps are as follows: A2.1 defines h1 dimensional and dimensionless time domain features, and calculates the time domain features of each group of samples. The specific contents are as follows: A2.1.1 Dimensional characteristics are peak value, peak-to-peak value, mean amplitude, RMS value, RMS amplitude, variance, and standard deviation; A2.1.2 The dimensionless characteristics are peak factor, impulse factor, margin factor, shape factor, kurtosis factor, and skewness; A2.2 Perform Fourier transform on each group of samples and calculate h2 frequency domain features of each group of samples, including centroid frequency, mean square frequency, root mean square frequency, frequency variance and frequency standard deviation; A2.3 Calculate the power spectrum entropy and singular spectrum entropy of each group of samples, and the number of features is recorded as h3.

3. A rolling bearing composite fault prediction feature selection method based on a high-order suppression sorting algorithm as claimed in claim 1, using the values ​​of Z*h1 time domain features to find the inflection point T1 to identify the health status of the rolling bearing α And the number of samples, the specific steps are as follows: B1.1 According to the amplitude of the Z*h1 time domain characteristic values ​​obtained in step A2.1 over the entire life cycle, find the first inflection point T1. The specific steps are as follows: B1.1.1 For samples before this inflection point, the amplitude of the dimensional feature has almost no fluctuation, while the dimensionless feature has repeated small fluctuations; B1.1.2 For samples after T1, the amplitude of the dimensional feature increases significantly and fluctuates repeatedly, while the dimensionless feature increases or decreases sharply and fluctuates repeatedly; B1.2 The number of samples from 0 to T1 is recorded as z1. During the corresponding time period, the rolling bearing is in a healthy state, recorded as α state.

4. A rolling bearing composite fault prediction feature selection method based on a high-order suppression sorting algorithm as claimed in claim 1, using the values ​​of Z*h1 time domain features to find the inflection point T2 and identify the degradation state of the rolling bearing β And the number of samples, the specific steps are as follows: B2.1 According to the amplitude of the Z*h1 time domain characteristic values ​​obtained in step A2.1 over the entire life cycle, find the second inflection point T2. The specific steps are as follows: B2.1.1 Compared with the samples before T1, the amplitude of the dimensional characteristics of the samples between T1 and T2 increased significantly and fluctuated repeatedly, and the dimensionless characteristics increased or decreased sharply and fluctuated repeatedly; B2.1.2 For samples after T2, the amplitude of the dimensional feature increases sharply without fluctuation, and the amplitude of the dimensionless feature increases or decreases sharply without fluctuation; B2.2 The number of samples from T1 to T2 is recorded as z2. During the corresponding time period, the rolling bearing is in a degraded state, recorded as β state.

5. The method for selecting rolling bearing composite fault prediction features based on a high-order suppression sorting algorithm according to claim 1, wherein the feature data set FD is divided into three data sets according to the inflection points T1 and T2 obtained in step B, and the specific steps are as follows: C1.1 There are n vectors in the feature data set FD. Each vector is divided into three parts, denoted as i∈[1,n]; C1.2 Constructing the health feature dataset FD α , denoted as C1.3 Constructing the degradation feature data set FD β , denoted as C1.4 Constructing the failure feature data set FD γ , denoted as 6. A rolling bearing composite fault prediction feature selection method based on a high-order suppression sorting algorithm as claimed in claim 1, constructing a network G α , G β , G γ The specific steps are as follows: C2.1.1 Based on G α For example, traverse FD α For each node in the set, get FD α The edge set The specific steps are as follows: C2.1.1.1 Using the incremental correlation Markov blanket algorithm, calculate The Markov blanket set i∈[1,n]; C2.1.1.2 If the node of There are nodes in Then at the node and nodes Add edges between Otherwise, no edge is added; C2.1.2 FD β , FD γ Repeat step C2.1.1 to obtain FD β and FD γ The edge set and 7. The method for selecting rolling bearing composite fault prediction features based on a high-order suppression sorting algorithm as claimed in claim 1, wherein the network G * The higher-order structure in G α Take this as an example, the specific steps are as follows: C2.3.1 vs. G α The edge set E α and the node set V α , find the node neighbors, i∈[1,n]; C2.3.2 Finding Nodes Whether there are common neighbors between them and their neighbors. If so, the three points form a higher-order structure, namely a 2-simplex. C2.3.3 Repeat steps C2.3.1 and C2.3.2 until all nodes are traversed and all 2-simplex structures are identified; C2.3.4G α The set of all 2-simplices in G α All 1-simplices in

Images