Rolling bearing fault diagnosis method based on dispersion pattern multi-scale diversity entropy

Abstract

The application discloses a rolling bearing fault diagnosis method based on dispersion pattern multi-scale diversity entropy, and comprises the following steps: firstly, collecting vibration signals of a rolling bearing under B different faults; then, dividing the collected vibration signals into a plurality of samples, and performing coarse-graining on each sample under different scales; in each sample, calculating dispersion pattern multi-scale diversity entropy of the vibration signal, performing feature extraction, and obtaining entropy features of the sample; extracting entropy features of all samples, constructing a feature set containing different operating states, dividing the feature set into a training set and a test set; finally, using the training set to obtain a fault diagnosis model through optimization; inputting the test set into the fault diagnosis model to obtain a fault diagnosis result, and calculating an accuracy rate. The application solves the problem of information loss in feature extraction by traditional diversity entropy in the prior art.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault diagnosis, and particularly relates to a rolling bearing fault diagnosis method based on dispersion pattern multi-scale diversity entropy. BACKGROUND

[0002] With the increasing complexity of mechanical equipment, more and more factors affect the reliability of equipment operation, and the maintenance difficulty of equipment increases. Thanks to the development of sensing technology, real-time collection of a large amount of data is realized, and fault diagnosis using monitoring data is one of the effective means to ensure the safe and reliable operation of equipment. In mechanical equipment, rolling bearings are core power components, but are affected by factors such as complex operating environment (such as high temperature and humidity, large impact, high load, etc.) and long operating time, which leads to bearing failure and becomes a key to restrict the safe and accurate operation of mechanical equipment. Therefore, monitoring the operating state of rolling bearings and timely detecting faults are of great significance to the safe operation of mechanical equipment.

[0003] When a rolling bearing fails, the complexity of the collected vibration signal (usually in the form of acceleration) changes, and entropy is a way to describe the degree of chaos of a system. Therefore, the complexity of a signal can be measured by entropy. At present, the entropy-based vibration signal feature extraction method is widely studied in the field of fault diagnosis. Diversity entropy is one of them. However, the existing diversity entropy method measures the global characteristics of the sequence by calculating the statistical probability of the cosine similarity of adjacent subsequences. This method measures the complexity of the sequence from a global perspective, but ignores the local angle change trend, resulting in the loss of some information and incomplete features, which affects the accuracy of fault diagnosis. The basic idea of the application to solve this problem is to calculate the cosine similarity dispersion pattern of the vibration signal time sequence as the basis for the entropy algorithm when extracting features from the vibration signal. In this way, the local angle change trend is increased, thereby solving the information loss problem of traditional diversity entropy and improving the accuracy of fault diagnosis. SUMMARY

[0004] The purpose of the application is to provide a rolling bearing fault diagnosis method based on dispersion pattern multi-scale diversity entropy, which solves the problem of information loss in the extraction of features by traditional diversity entropy in the prior art.

[0005] The technical solution adopted by the application is a rolling bearing fault diagnosis method based on dispersion pattern multi-scale diversity entropy, which is implemented according to the following steps:

[0006] Step 1, collect vibration signals of a rolling bearing running under B different faults respectively;

[0007] Step 2, the vibration signal collected in the step 1 is divided into a plurality of samples, and the coarse-grained is carried out on each sample under different scales;

[0008] Step 3, in each sample, the dispersion pattern diversity entropy of the vibration signal is calculated, feature extraction is carried out, and the entropy feature of the sample is obtained;

[0009] Step 4, the entropy features of all samples are extracted, a feature set containing different operating states is constructed, and the feature set is divided into a training set and a test set;

[0010] Step 5, the training set is used to optimize the fault diagnosis model, the test set is input into the fault diagnosis model, the fault diagnosis result is obtained, and the accuracy is calculated.

[0011] The characteristics of the application also lie in,

[0012] Step 2 is implemented according to the following steps:

[0013] Step 2.1, the vibration signal is divided into a plurality of non-overlapping samples S with every L length, and is divided by formula (1) as follows:

[0014] S(i)={v j＝(i-1)·L+1 ,v j＝(i-1)·L+2 ,…,v j＝i·L} (1)

[0015] Wherein, S(i) represents the i-th sample of the vibration signal, v j＝(i-1)·L+1,…,i·L represents the j-th element in the i-th sample;

[0016] Step 2.2, the coarse-grained is carried out on each sample S(i) under different scale factors τ by formula (2);

[0017]

[0018] Wherein, represents the k-th element of the reconstructed sequence of the sample S(i) after coarse-grained under the scale factor τ, v j represents the j-th element in the sample S(i), and L is the sample length, and the reconstructed sequence of the sample under different scale factors τ is obtained by step 2.

[0019] Step 3 is implemented according to the following steps:

[0020] Step 3.1, the vibration signal X after coarse-grained with the length of N is reconstructed into a subsequence matrix Y(m): N N=L-τ+1, and the vibration signal X after coarse-grained with the length of N is reconstructed into a subsequence matrix Y(m):

[0021]

[0022] where m is the embedding dimension, y i (m) = [x i , x i+1 ,..., x i+m-1 ], i = (1, 2,..., N-m+1) represents the i-th element to the i+m-1-th element of the subsequence;

[0023] Step 3.2, calculate the cosine similarity of adjacent subsequence y i (m) and y i+1 (m) in the subsequence matrix Y(m) The cosine similarity vector of the entire sequence matrix is obtained by combining:

[0024] D(m) = [d1,...,d N-m ] = [d(y1(m),y2(m)),...,d(y N-M-1 (m),y N-m (m))] (4)

[0025] where d ranges from [-1, 1], and d represents the similarity of the angles of two subsequence;

[0026] Step 3.3, select c dispersion categories, and assign each element in D(m) to a certain category from 1 to c For , construct the embedding vector where m1 is the embedding dimension, and η is the delay factor;

[0027] Step 3.4, calculate the dispersion pattern Let Then The corresponding dispersion pattern is Since there are c digits, each digit has m1 values, and the corresponding dispersion pattern has ;

[0028] Step 3.5, calculate the probability of each dispersion pattern :

[0029]

[0030] where indicates the number of elements mapped to , i.e. is equal to the number of elements mapped to divided by the number of elements in ;

[0031] ​Step 3.6, calculate the dispersion pattern diversity entropy according to the definition of Shannon entropy:

[0032]

[0033] The dispersion pattern diversity entropy of each coarse-grained sequence is calculated as the entropy feature of the sample.

[0034] Step 4 is implemented according to the following steps:

[0035] Step 4.1, repeat the above step 3 to calculate the dispersion pattern multi-scale diversity entropy MDDE for each sample at different scale factors τ:

[0036]

[0037] MDDE(S, τ, m) represents the dispersion pattern multi-scale diversity entropy of sample S at scale factor τ, m is the embedding dimension, and the size of MDDE of each sample is 1 × τ, DDE represents the dispersion pattern diversity entropy.

[0038] Step 4.2, after calculating the dispersion pattern diversity entropy of all samples at different scale factors τ, a feature set F of the rolling bearing under different operating conditions is formed;

[0039]

[0040] where MDDE(j) represents the dispersion pattern multi-scale diversity entropy calculated for the jth sample, represents the reconstructed sequence of the jth sample at scale factor τ, represents the dispersion pattern diversity entropy calculated for the jth sample at scale factor τ;

[0041] Step 4.3, divide the feature set F into training set and test set in proportion, the training proportion means that the proportion of features selected for training the classification model in the feature set is used as the training set (generally selected as 60%), and the remaining part (40%) is used as the test set. After the training model is successful, the test set is used to evaluate the model.

[0042] Step 5 is implemented according to the following steps:

[0043] Step 5.1, construct B binary classification support vector machines according to the required fault categories, B is the number of fault categories, when constructing the bth (b ∈ [1, … B]) classifier, the training samples of the bth class are taken as a class, and the class number t b = b is positive, and the training samples of all other classes except the bth class are taken as a class, and the class number is -1. The classification output function of the bth classifier can be established by optimization as: ​

[0044]

[0045] wherein f(x) represents the classification of sample x, alpha b represents the weight coefficient of the bth classification surface, K(x, x b ) represents the support vector machine kernel function, u b represents the classification threshold of the bth classifier, t b represents the class number of the bth classifier.

[0046] Step 5.2, input the test set samples into the B binary support vector machines respectively, obtain B output results from the above classification output function formula, compare the maximum fault classification class of the B output results, and the fault class of the test sample is determined.

[0047] The bearing fault diagnosis method of the dispersion pattern improved diversity entropy has the advantages that the signal is reconstructed under different time scales, the similarity between adjacent subsequences is calculated by using cosine similarity, the calculated similarity is mapped to different dispersion patterns by selecting different embedding dimensions m1, and the diversity entropy is calculated by calculating the statistical probability of the dispersion pattern. The calculated entropy value under each scale is used as a characteristic value, and finally the extracted feature set is divided into a training set and a test set according to different proportions and input into a support vector machine to obtain a diagnosis result. The information loss problem in the feature extraction process of the prior art by using the diversity entropy to calculate the cosine similarity statistical probability is solved. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 is the flow chart of the bearing fault diagnosis method of the dispersion pattern improved diversity entropy of the application;

[0049] Fig. 2(a) is an MDE entropy value graph under three fault states of the bearing in the bearing fault diagnosis method of the dispersion pattern improved diversity entropy of the application;

[0050] Fig. 2(b) is an MDDE entropy value graph under three fault states of the bearing in the bearing fault diagnosis method of the dispersion pattern improved diversity entropy of the application;

[0051] Fig. 3(a) is an MDE entropy value scatter plot of the bearing vibration signal in the bearing fault diagnosis method of the dispersion pattern improved diversity entropy of the application;

[0052] Fig. 3(b) is an MDDE entropy value scatter plot of the bearing vibration signal in the bearing fault diagnosis method of the dispersion pattern improved diversity entropy of the application. DETAILED DESCRIPTION

[0053] The application will be described in detail below in combination with the drawings and specific embodiments.

[0054] The rolling bearing fault diagnosis method of dispersion pattern multi-scale diversity entropy provided by the application, a flow chart is shown as Figure 1 The method is implemented according to the following steps:

[0055] Step 1, collect vibration signals of a rolling bearing under B different faults (such as normal, inner ring fault, outer ring fault);

[0056] Step 2, divide the vibration signals collected in step 1 into a plurality of samples, and coarsely granulate each sample at different scales;

[0057] Step 2 is implemented according to the following steps:

[0058] Step 2.1, divide the vibration signal into a plurality of non-overlapping samples S with a length of L, and specifically divide according to formula (1) as follows:

[0059] S(i)={v j＝(i-1)·L+1 ,v j＝(i-1)·L+2 ,...,v j＝i·L} (1)

[0060] Wherein, S(i) represents the i-th sample of the vibration signal, and v j＝(i-1)·L+1,…,i·L represents the j-th element in the i-th sample;

[0061] Step 2.2, coarsely granulate each sample S(i) at different scale factors τ by formula (2):

[0062]

[0063] Wherein, represents the k-th element of the reconstructed sequence of the sample S(i) after coarse granulation at the scale factor τ, v j represents the j-th element in the sample S(i), and L is the sample length. The coarse granulation process is essentially to average the original time series in a window with a length of τ, and then sample with τ as the scale factor. Through step 2, the reconstructed sequence of the sample at different scale factors τ is obtained.

[0064] Step 3, in each sample, calculate the dispersion pattern diversity entropy of the vibration signal, extract features, and obtain the entropy features of the sample;

[0065] Step 3 is implemented according to the following steps:

[0066] Step 3.1, reconstruct the vibration signal X after coarse granulation with a length of N into a subsequence matrix Y(m) as follows: N N=L-τ+1:

[0067]

[0068] where m is the embedding dimension, y i (m) = [x i , x i+1 ,..., x i+m-1 ], i = (1, 2,..., N-m+1) represents the i-th element to the i+m-1-th element of the subsequence;

[0069] Step 3.2, calculate the cosine similarity of adjacent subsequence y i (m) and y i+1 (m) in the subsequence matrix Y(m) Combine to get the cosine similarity vector of the entire sequence matrix:

[0070] D(m) = [d1,...,d N-m ] = [d(y1(m),y2(m)),...,d(y N-M-1 (m),y N-m (m))] (4)

[0071] where the range of d is [-1, 1], and d represents the similarity of the angles of two subsequence;

[0072] Step 3.3, select c dispersion categories, and assign each element in D(m) to a certain category in 1~c For Construct the embedding vector where m1 is the embedding dimension, and η is the delay factor;

[0073] Step 3.4, calculate the dispersion pattern Let Then The corresponding dispersion pattern is This way of constructing the cosine similarity of adjacent subsequence as the dispersion pattern can increase the local feature information, because Each number has m1 values, and the corresponding dispersion pattern has ;

[0074] Step 3.5, calculate the probability of each dispersion pattern :

[0075]

[0076] where points to the number of mapping to , that is equal to mapping to​ the number of elements in the set of elements of the sample S divided by the number of elements in the set of elements of the sample S divided by

[0077] Step 3.6, calculate the dispersion pattern diversity entropy according to the definition of Shannon entropy:

[0078]

[0079] Coarse-grain the sample under different scale factors, and then calculate the dispersion pattern diversity entropy of each coarse-grained sequence by the above step 3 as the entropy feature of the sample.

[0080] Step 4, extract the entropy features of all samples, construct the feature set containing different running states, and divide the feature set into a training set and a test set;

[0081] Step 4 is implemented according to the following steps:

[0082] Step 4.1, repeat the above step 3 to calculate the dispersion pattern multi-scale diversity entropy MDDE for each sample under different scale factors τ as:

[0083]

[0084] MDDE(S, τ, m) represents the dispersion pattern multi-scale diversity entropy of the sample S under the scale factor τ, m is the embedding dimension, and the size of MDDE of each sample is 1 x τ, DDE represents the dispersion pattern diversity entropy.

[0085] Step 4.2, form the feature set F of the rolling bearing under different running states after calculating the dispersion pattern diversity entropy of all samples under different scale factors τ;

[0086]

[0087] wherein MDDE(j) represents the dispersion pattern multi-scale diversity entropy calculated by the jth sample, represents the reconstructed sequence of the jth sample under the scale factor τ, represents the dispersion pattern diversity entropy calculated by the jth sample under the scale factor τ;

[0088] Step 4.3, divide the feature set F into a training set and a test set according to the proportion, the training proportion means that the proportion of the features selected for training the classification model in the feature set is generally selected as 60% to do the training set, and the remaining part (40%) is used as the test set. After the training model is successful, the test set is used to evaluate the model.

[0089] ​Step 5, using the training set, the fault diagnosis model is optimized; the test set is input into the fault diagnosis model to obtain the fault diagnosis result, and the accuracy is calculated.

[0090] Step 5 is implemented according to the following steps:

[0091] Step 5.1, according to the required fault class to be distinguished, B binary classification support vector machines are constructed, B is the number of fault classes, when constructing the b(b∈[1,…B])th classifier, the training samples of the bth class are taken as a class, and the class number t b = b is positive, and the training samples of all the remaining classes except the bth class are taken as a class, and the class number is -1. The classification output function of the bth classifier can be established by optimization:

[0092]

[0093] Wherein, f(x) represents the classification of sample x, α b represents the weight coefficient of the bth classification surface, K(x,x b ) represents the support vector machine kernel function, u b represents the classification threshold of the bth classifier, t b represents the class number of the bth classifier.

[0094] Step 5.2, the test set samples are respectively input into the B binary classification support vector machines, and the B output results are obtained from the above classification output function formula. The fault classification class with the largest output result is compared, that is, the fault class of the test sample is determined.

[0095] Embodiment

[0096] The bearing data set of Paderborn University in Germany is selected. The data set includes normal bearing data and fault bearing data. The fault bearing includes artificial damage bearing data and real damage fault data. The data set includes four working conditions, and the radial force on the bearing and the load torque on the driving system are tested by changing the speed of the driving system. The normal bearing data used in this case is: K001, K002, K003, K004, K005, K006, outer ring bearing fault is KA01, KA03, KA05, KA06, KA07, KA08, KA09, and inner ring fault bearing is selected as: KI01, KI03, KI05, KI07 and KI08. The parameters of MDE and MDDE selected in this case study are: sample length L, embedding dimension m, interval number ε, scale factor τ are selected as 4096, 3, 10, 20, and the parameters of MDDE are: dispersion class c is 6, embedding dimension m1 is 4.

[0097] The three curves in Fig. 2(a) represent the diversity entropy of the bearing data set in normal, outer ring fault and inner ring fault conditions when the scale factor is from 1 to 20. As can be seen from Fig. 2(a), the fluctuation of the entropy value calculated by MDE for the outer ring fault is relatively large, and the entropy value is crossed with those of the other two fault conditions, which is not convenient for classification. The diversity entropy values calculated in the three conditions fluctuate greatly and are not stable enough. Fig. 2(b) shows that the entropy values calculated by MDDE can better separate the normal condition from the outer ring fault, and the calculated entropy values are relatively stable and have good consistency in the normal, inner ring fault and outer ring fault conditions.

[0098] Fig. 3(a) and Fig. 3(b) respectively show the MDE and MDDE clustering scatter plots in normal, outer ring fault and inner ring fault conditions. It can be seen that, due to the use of the dispersion pattern multi-scale diversity entropy algorithm to increase the local feature information, this feature has better clustering effect compared with the traditional diversity entropy.

[0099] Table (1) shows the diagnostic accuracy of the multi-scale diversity entropy and the dispersion pattern multi-scale diversity entropy combined with the support vector machine respectively. In order to avoid the randomness of a single result, the data with a proportion of 0.6, 0.4 and 0.8 are taken as training samples respectively, and independent ten experiments are carried out to take the average value. Compared with the multi-scale diversity entropy, the dispersion pattern multi-scale diversity entropy has higher diagnostic accuracy, and the diagnostic accuracy of the dispersion pattern multi-scale diversity entropy is about 2% higher than that of the multi-scale diversity entropy combined with the support vector machine, which has a significant advantage.

[0100] Table (1) shows the diagnostic accuracy of the multi-scale diversity entropy and the dispersion pattern multi-scale diversity entropy combined with the support vector machine respectively. In order to avoid the randomness of a single result, the data with a proportion of 0.6, 0.4 and 0.8 are taken as training samples respectively, and independent ten experiments are carried out to take the average value. Compared with the multi-scale diversity entropy, the dispersion pattern multi-scale diversity entropy has higher diagnostic accuracy, and the diagnostic accuracy of the dispersion pattern multi-scale diversity entropy is about 2% higher than that of the multi-scale diversity entropy combined with the support vector machine, which has a significant advantage.

[0101]

Claims

1. A method for fault diagnosis of rolling bearings based on dispersion pattern multiscale diversity entropy, characterized in that, The method is implemented according to the following steps: Step 1, collect the vibration signals of the rolling bearing respectively under different faults; Step 2, the vibration signal collected in step 1 is divided into a plurality of samples, and each sample is coarse-grained at different scales; Step 2 is implemented according to the following steps: Step 2.1, divide the vibration signal into non-overlapping samples of length L each , specifically by dividing as follows by equation (1); （1） wherein, represents the first sample of the vibration signal division, represents the first element in the first sample; Step 2.2, for each sample At different scale factors Coarsening by equation (2) at the lower scale; （2） wherein, denotes the sample After coarse-graining at scale factor the first element of the reconstructed sequence, denotes the sample the first element of the reconstructed sequence, is the sample length, the reconstructed sequence of the sample at different scale factors is obtained by step 2; Step 3, in each sample, the dispersion pattern diversity entropy of the vibration signal is calculated, feature extraction is performed, and the entropy feature of the sample is obtained; Step 3 is implemented according to the following steps: Step 3.1: After coarsening, the length is... vibration signal Reconstructed into a subsequence matrix : （3） wherein, is the embedding dimension, represents the first element to the last element of the subsequence. Step 3.2, Compute sub-sequence matrix In the middle, adjacent sub-sequences with cosine similarity , combine to get the cosine similarity vector of the whole sequence matrix: （4） wherein the range is , represents the similarity of the two subsequence angles; Step 3.3, selecting one of the dispersion classes, each of the elements in the set by assigning to one of the classes in the set , and constructing an embedding vector wherein, is the embedding dimension, is a delay factor; Step 3.4, calculating the dispersion pattern Let , , ,..., Then The corresponding dispersion pattern is Since There are bits, each of which has values, the corresponding dispersion pattern has entries; Step 3.5, calculating the probability of each dispersion pattern is:​ （5） wherein, pointing to mapped to the number of elements in is equal to the number of elements mapped to divided by the number of elements in Step 3.6, the dispersion pattern diversity entropy is calculated according to the definition of Shannon entropy: （6） The dispersion pattern diversity entropy of each coarse-grained sequence is calculated as the entropy feature of the sample; Step 4, the entropy features of all samples are extracted, a feature set containing different operating states is constructed, and the feature set is divided into a training set and a test set; Step 5, using the training set, a fault diagnosis model is obtained by optimization; the test set is input into the fault diagnosis model, a fault diagnosis result is obtained, and an accuracy rate is calculated.

2. The rolling bearing fault diagnosis method of dispersed mode multi-scale diversity entropy according to claim 1, characterized in that, Step 4 is implemented according to the following steps: Step 4.1, each sample is repeated in different scale factors The above step 3 is repeated to calculate the dispersion pattern multiscale diversity entropy MDDE as follows: （7） representing samples under a scale factor , the multiscale diversity entropy of the color dispersion pattern, is the embedding dimension, the size of MDDE of each sample is , DDE represents the diversity entropy of the color dispersion pattern; Step 4.2, in all samples at different scale factors The dispersion pattern diversity entropy is calculated below to form a feature set of the rolling bearing under different operating conditions ; （8） in, Indicates the first The multi-scale diversity entropy of the dispersion pattern calculated for each sample. Indicates the first Each sample has a scaling factor of Reconstruction sequence at time, Indicates the first Each sample has a scaling factor of The dispersive mode diversity entropy calculated at that time; Step 4.3, splitting the feature set into training and test sets Splitting into training and test sets in proportion, the training proportion means that the proportion of features selected for training the classification model in the feature set is the training set, and the remaining part is the test set. After the training model is successful, the test set is used to evaluate the model.

3. The rolling bearing fault diagnosis method of dispersed mode multi-scale diversity entropy according to claim 2, characterized in that, Step 5 is implemented according to the following steps: Step 5.1: Construct according to the required fault category. A binary classification support vector machine, For the number of fault categories, in constructing the first... When using the nth classifier, the nth The training samples of a class are treated as one class. Category number If it is a positive number, it will be removed. Training samples from all other categories besides the first category are treated as one category, with the category number being [class number missing]. Optimization can establish the first The classification output function of each classifier is: （9） wherein, represents a sample x classification, represents a weight coefficient of the th classification surface, represents a support vector machine kernel function, represents a classification threshold of the th classifier, represents a class number of the th classifier; Step 5.2, input the test set samples to the two binary support vector machines respectively, and get the output results from the above classification output functions The fault classification category with the largest output result is the fault classification category determined by the test sample.​​

Images