Gearbox fault diagnosis method and system based on local feature sparse coding representation

Abstract

The application discloses a kind of gear box fault diagnosis method and system based on local feature sparse coding representation, first, the vibration signal sample of different health states of gear box is obtained to obtain sample set;Then each sample in sample set is fragmented to obtain fragment set with the same overlap rate, after whitening processing, a group of dictionaries is obtained by training sparse coding model;Then the sparse representation of each fragment in each sample is calculated using the obtained dictionary, and the feature of each vibration signal sample is obtained after average pooling processing;Finally, the vibration signal feature extracted is trained Softmax classifier, and the trained Softmax classifier is used to realize the intelligent diagnosis and identification of different health states of gear box;The method directly processes original vibration signal, realizes the automatic extraction of feature and the automatic identification of health state, reduces human intervention, and reduces the dependence on diagnosis knowledge and diagnosis experience.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gear box fault diagnosis, in particular to a gear box fault diagnosis method and system based on local feature sparse coding representation. BACKGROUND

[0002] With the continuous progress and development of science and technology, wind turbine equipment is developing towards high performance, complexity and further large-scale integration. The gear box is an important power transmission device in the wind turbine, which contains a large number of gears, bearings and other key components. Long-term operation in harsh environments, the gear box is prone to bearing wear, gear crack and tooth shedding and other faults, resulting in high maintenance costs and significant economic losses. According to statistics, the maintenance cost caused by gear box failure accounts for 13% of the entire wind turbine operation and maintenance cost. Therefore, it is of great significance to study efficient and reliable gear box fault diagnosis methods.

[0003] Intelligent fault diagnosis relies on machine learning technology to quickly and effectively process a large number of vibration signals and automatically obtain accurate diagnosis results, and is widely studied in the field of gear box fault diagnosis. Intelligent fault diagnosis usually includes three steps: signal acquisition, feature extraction and fault classification. Among them, feature extraction is a key step, because the quality of the extracted features directly determines the fault classification result. Traditional feature extraction methods usually use advanced signal processing techniques such as wavelet packet decomposition and empirical mode decomposition combined with rich diagnostic experience to design features to represent the health status of the gear box. However, this kind of method relies heavily on signal processing techniques and diagnostic expertise, and when faced with complex operating conditions and massive signals, the designed features are difficult to fully characterize the health status of the equipment. SUMMARY

[0004] In view of the problems existing in the prior art, the present application provides a gear box fault diagnosis method and system based on local feature sparse coding representation, which extracts features from original vibration signals to realize efficient and reliable fault diagnosis of the gear box.

[0005] The present application is realized by the following technical solutions:

[0006] A gear box fault diagnosis method based on local feature sparse coding representation, comprising the following processes:

[0007] Step 1, obtaining vibration signal samples under different health states of the gear box to form a sample set;

[0008] Step 2, fragmenting each vibration signal sample in the sample set to obtain a local sample fragment set at the same overlap rate, then whitening the local sample fragment set, and training a sparse coding model using the whitened local sample fragment set to obtain a dictionary;

[0009] Step 3, calculating sparse representation of each segment in each vibration signal sample according to the dictionary, and determining the feature of each vibration signal sample according to the sparse representation;

[0010] Step 4, training a Softmax classifier using the vibration signal sample features, and using the trained Softmax classifier to realize fault diagnosis of the gearbox.

[0011] Preferably, after the vibration signal sample is normalized in step 1, a sample set is constructed according to the normalized vibration signal sample.

[0012] Preferably, the whitening method is as follows:

[0013] A local sample segment set Synthesize a data matrix

[0014] Perform eigenvalue decomposition on the covariance matrix of the data matrix to obtain the eigenvalues, and the decomposition method is as follows:

[0015] cov(S) = EVE T

[0016] Wherein, V is a diagonal matrix, the diagonal elements are the eigenvalues of cov(S), and E is the eigenvector corresponding to the eigenvalues of cov(S);

[0017] According to the eigenvalues, the whitening result corresponding to the local sample segment set is calculated, and the expression is as follows:

[0018]

[0019] Wherein, W white = EV -1 / 2 E T is the whitening transformation matrix.

[0020] Preferably, the damage function of the sparse coding model in step 2 is as follows:

[0021]

[0022] s.t:||d j ||2≤ε,j=1,2,L,J

[0023] In the formula, X is a data matrix composed of whitened segments, D is a dictionary containing J atoms, H is a sparse representation, β is a sparse penalty parameter, and ε is a constraint coefficient.

[0024] Preferably, the calculation method of the sparse representation in step 3 is as follows:

[0025] Optimization problem solving is performed on the whitening processing result of the vibration signal sample segment to obtain the corresponding sparse representation

[0026]

[0027] where D is a dictionary, β is a sparsity penalty parameter, is the l-th segment of the sample x m corresponding to the whitening result.

[0028] Preferably, the sparse representation of each vibration signal sample in step 3 is averaged and pooled to obtain a feature vector of the vibration signal sample, and the feature of each vibration signal sample is determined according to the feature vector.

[0029] Preferably, the minimization cost function of the Softmax classifier in step 4 is as follows:

[0030]

[0031] where 1{g} represents an indicator function, and γ>0 represents a weight decay factor.

[0032] Preferably, the fault diagnosis method of the gearbox in step 4 is as follows:

[0033] The features of the new vibration signal sample are determined by steps 2 and 3, the features are input into the trained Softmax classifier, the probability of the health state of the new vibration signal sample is obtained, and the sample label corresponding to the maximum probability is taken as the health label of the new vibration signal sample x.

[0034] Preferably, the expression of the health state is as follows:

[0035]

[0036] where c is a class, f is a feature vector, is a parameter of the Softmax classifier.

[0037] A gearbox fault diagnosis method system based on local feature sparse coding representation, comprising,

[0038] A sample module for obtaining vibration signal samples of a gearbox in different health states to form a sample set;

[0039] A dictionary module for fragmenting each vibration signal sample in the sample set at the same overlap rate to obtain a local sample segment set, then whitening the local sample segment set, and training a sparse coding model using the whitened local sample segment set to obtain a dictionary;

[0040] A feature extraction module for calculating the sparse representation of each segment in each vibration signal sample according to the dictionary, and determining the feature of each vibration signal sample according to the sparse representation.

[0041] The diagnostic module is used to train a Softmax classifier using vibration signal sample features, and then uses the trained Softmax classifier to diagnose gearbox faults.

[0042] Compared with the prior art, the present invention has the following beneficial technical effects:

[0043] This invention provides a gearbox fault diagnosis method based on sparse coding representation of local features. First, vibration signal samples of the gearbox under different health states are acquired, and each sample is normalized to obtain a sample set. Then, each sample in the sample set is segmented with the same overlap rate to obtain a fragment set. After whitening, a sparse coding model is trained to obtain a dictionary. Next, the sparse representation of each fragment in each sample is calculated using the obtained dictionary, and after average pooling, the features of each vibration signal sample are obtained. Finally, the extracted vibration signal features are used as input to train a Softmax classifier, which is then used to achieve intelligent diagnosis and identification of different health states of the gearbox. This invention utilizes a sparse coding model to automatically extract features directly from gearbox vibration signals, overcoming the dependence of traditional signal processing-based intelligent diagnosis methods on diagnostic experience and expert knowledge, and can efficiently and reliably achieve intelligent diagnosis and identification of gearbox faults. The sample segmentation strategy used in this invention helps to eliminate the time-shift characteristics in vibration signals, extract more discriminative local features, and improve computational efficiency. Sparse coding adaptively learns a dictionary from vibration signals, where each atom is analogous to a one-dimensional Gabor filter with rich and interpretable time-frequency features. Attached Figure Description

[0044] Figure 1 This is a flowchart of the gearbox fault diagnosis method of the present invention;

[0045] Figure 2 This is a flowchart of the feature extraction process for the signal in this invention;

[0046] Figure 3 The time-domain curves of atomic vectors in the dictionary of this invention and their Gabor fitting curves;

[0047] Figure 4 The frequency domain curves and their Gabor fitting curves of the atomic vectors in the dictionary of this invention are shown. Detailed Implementation

[0048] The present invention will now be described in further detail with reference to the accompanying drawings. These descriptions are intended to explain the invention and not to limit it.

[0049] like Figure 1 As shown, a gearbox fault diagnosis method based on sparse coding representation of local features includes the following steps:

[0050] S1, collect vibration signal samples of the gearbox C different health states, and normalize each sample to 0-1, and record the normalized sample set as

[0051] wherein, is the mth sample, consisting of N sampling points, y m ∈{1,2,L,C} is the label of x m ;

[0052] S2, overlap slicing is performed on each sample x m in the sample set to obtain L segments The overlap rate between adjacent segments is η, and the segment length is I (I=N), that is,

[0053] S3, the segments of all samples are combined to form a local sample segment set wherein N s =LM; After whitening, a whitened local sample segment set is obtained wherein

[0054] The method of the above whitening is as follows:

[0055] First, the local sample segment set is synthesized into a data matrix

[0056] Then, the covariance matrix of S is decomposed: cov(S)=EVE T ,

[0057] Wherein, V is a diagonal matrix, the diagonal elements are the eigenvalues of cov(S), and E is the eigenvector corresponding to the eigenvalue of cov(S);

[0058] Finally, the whitened representation of x s is calculated Wherein W white =EV -12 E T is the whitening transformation matrix.

[0059] The whitened local sample segment set is used to train a sparse coding model to obtain a group of dictionaries

[0060] wherein, represents the jth atom in the dictionary, j=1,2,L,J;

[0061] Training a sparse coding model is essentially equivalent to solving the following loss function:

[0062]

[0063] st:||d j ||2≤ε,j=1,2,L,J

[0064] In the formula, A data matrix composed of whitened fragments. It is a dictionary containing J atoms. in yes The sparse representation; β > 0 is the sparse penalty parameter, and ε > 0 is the constraint coefficient.

[0065] S4, according to Figure 2 Using the obtained dictionary D, calculate the m-th sample x. m The lth segment sparse representation Then, average pooling is performed on the L sparse representations to obtain x. m eigenvector f m ;

[0066] Calculate sample x using dictionary D m eigenvector f m The specific steps are as follows:

[0067] First, for x m The lth segment Whitening transformation is performed to obtain

[0068] Then, by solving the following optimization problem, we obtain... Corresponding sparse representation

[0069]

[0070] Finally, for the L sparse representations Perform average pooling to obtain x m eigenvectors

[0071] S5. Utilizing the sample set after feature extraction Training a Softmax classifier, i.e., minimizing the cost function:

[0072]

[0073] In the formula, 1{g} represents an indicator function, 1{true} = 1, and 1{false} = 0. γ > 0 represents a weight attenuation factor. After training the Softmax, an optimal parameter set can be obtained

[0074] S6, a newly collected vibration signal sample is given First, it is fragmented to obtain L segments The overlap rate between adjacent segments is η, and each segment After whitening transformation and local feature extraction, the corresponding sparse representation is obtained Then, L sparse representations are averaged and pooled to obtain the feature vector f of x. Then, f is input into the trained Softmax classifier to obtain the probability that sample x belongs to the cth (c = 1, 2, L, C) health state. The calculation formula is as follows:

[0075]

[0076] Finally, the sample label corresponding to the maximum probability is taken as the health label of sample x.

[0077] The application also provides a gearbox fault diagnosis method based on local feature sparse coding representation, comprising,

[0078] A sample module is configured to acquire vibration signal samples in different health states of the gearbox to form a sample set.

[0079] A dictionary module is configured to fragment each vibration signal sample in the sample set to obtain a local sample segment set at the same overlap rate, then perform whitening processing on the local sample segment set, and train a sparse coding model using the whitened local sample segment set to obtain a dictionary.

[0080] A feature extraction module is configured to calculate the sparse representation of each segment in each vibration signal sample according to the dictionary, and determine the feature of each vibration signal sample according to the sparse representation.

[0081] A diagnosis module is configured to train a Softmax classifier using the vibration signal sample features, and implement the fault diagnosis of the gearbox using the trained Softmax classifier.

[0082] Embodiment 1

[0083] The gearbox fault data set collected by the drivetrain dynamics simulator (DDS) is used to verify the feasibility of the application.

[0084] The specific information of the data set is shown in Table 1, including the health states under two working conditions, four gear states and four bearing states. The two working conditions are respectively 20Hz rotation frequency without load and 30Hz rotation frequency with load. The sampling frequency of the vibration signal is 5120Hz, and each health state under a single working condition includes 300 vibration signal samples, each sample having 2400 data points. 10% of the samples from each health state are selected for model training, and the remaining 90% of the samples are used for model testing.

[0085] Table 1 Description of the gearbox data set

[0086]

[0087] In the experimental setup, the segment length I is selected as 75, the number of atoms J is 150, and the overlap rate η between adjacent segments is 0.97, so that a vibration signal with a length of 2400 can be divided into 6984 segments with a length of 75. The sparse penalty parameter β in sparse coding is selected as 0.2, the constraint coefficient ε is selected as 0.01, and the weight attenuation γ of Softmax is selected as 1E-5. The commonly used F score As an evaluation index, it is defined as follows:

[0088]

[0089] In the formula, TP, FN and FP respectively represent the number of samples predicted as positive class, the number of samples predicted as negative class, and the number of samples predicted as positive class. Using the present application, the F score values of different health states on the test set are shown in Table 2.

[0090] In order to verify the effectiveness of the present application, the results of the present application are compared with the results of the conventional diagnosis method based on signal processing technology. In the conventional method, common features in gearbox fault diagnosis are artificially designed, including 12 commonly used statistical parameters: mean, root mean square, root amplitude, standard deviation, skewness, kurtosis, waveform index, peak value index, pulse index, margin index, skewness index and kurtosis index, 8 empirical mode decomposition energy ratios and 16 wavelet packet energy ratio features. These features are used as the input of the Softmax classifier to realize gearbox fault diagnosis. The diagnosis results of the conventional method are shown in Table 2:

[0091] Table 2 F score values of different health states of the gearbox (%)

[0092]

[0093] From the comparison results in Table 2, it can be seen that the diagnostic accuracy of the method of the present application for 7 of the 9 gear box health states is higher than that of the conventional method, the average diagnostic accuracy of the present application is 97.43%, which is higher than the 93.57% of the conventional method, verifying the effectiveness of the method of the present application compared with the conventional method. In addition, the method of the present application can automatically extract features from the vibration signal, overcoming the deficiency of the conventional method in over-reliance on diagnostic experience and prior knowledge in feature extraction.

[0094] The present application learns a dictionary containing 150 atoms from the gear box vibration signal using a sparse coding model, and draws time domain and frequency domain graphs of any 16 atoms, and further fits the time domain and frequency domain graphs with one-dimensional Gabor filters, respectively, and the results are shown in Figs. 1 and 2. Figure 3 and Figure 4 As can be seen, each atom in the learned dictionary is similar to a one-dimensional Gabor filter, with rich and obvious time-frequency characteristics, so the present application can well reveal the physical mechanism of feature extraction, and has certain interpretability. The present application can directly process the original vibration signal, realize automatic feature extraction and automatic identification of health state, reduce human intervention, and reduce the dependence on diagnostic knowledge and diagnostic experience.

[0095] The above content is only to illustrate the technical idea of the present application, and cannot limit the protection scope of the present application, and any modification made according to the technical idea of the present application on the basis of the technical scheme falls within the protection scope of the claims of the present application.

Claims

1. A gearbox fault diagnosis method based on sparse coding representation of local features, characterized in that, Includes the following processes: Step 1: Obtain vibration signal samples of the gearbox under different health conditions to form a sample set; Step 2: Divide each vibration signal sample in the sample set into local sample fragment sets with the same overlap rate, then whiten the local sample fragment sets, and use the whitened local sample fragment sets to train a sparse coding model to obtain a dictionary. The whitening treatment method is as follows: Local sample fragment set Synthesize a data matrix ; The eigenvalues ​​are obtained by performing eigenvalue decomposition on the covariance matrix of the data matrix. The decomposition method is as follows: in, It is a diagonal matrix with diagonal elements. eigenvalues, for The eigenvectors corresponding to the eigenvalues; The whitening result for a local sample fragment set is calculated based on the feature values, as shown in the following expression: in, Here is the whitening transformation matrix; Step 3: Calculate the sparse representation of each segment in each vibration signal sample based on the dictionary, and determine the features of each vibration signal sample based on the sparse representation; The calculation method for the sparse representation is as follows: The optimization problem of the whitening process of vibration signal sample segments is solved to obtain the corresponding sparse characterization. ; in, For a dictionary, For sparsity penalty parameters, For the sample The first in The whitening result corresponding to each segment; Step 4: Train a Softmax classifier using vibration signal sample features, and use the trained Softmax classifier to diagnose gearbox faults.

2. The gearbox fault diagnosis method based on sparse coding representation of local features according to claim 1, characterized in that, After normalizing the vibration signal samples in step 1, a sample set is constructed based on the normalized vibration signal samples.

3. The gearbox fault diagnosis method based on sparse coding representation of local features according to claim 1, characterized in that, The impairment function of the sparse coding model described in step 2 is as follows: In the formula, A data matrix composed of whitened fragments. For inclusion A dictionary of atoms As a sparse characterization, For sparsity penalty parameters, This is the constraint coefficient.

4. The gearbox fault diagnosis method based on sparse coding representation of local features according to claim 1, characterized in that, In step 3, the sparse representation of each vibration signal sample is processed by average pooling to obtain the feature vector of each vibration signal sample. The features of each vibration signal sample are determined based on the feature vector.

5. The gearbox fault diagnosis method based on sparse coding representation of local features according to claim 1, characterized in that, The cost function for minimizing the Softmax classifier in step 4 is as follows: In the formula, Indicates an indicator function, This represents the weight decay factor.

6. The gearbox fault diagnosis method based on sparse coding representation of local features according to claim 1, characterized in that, The fault diagnosis method for the gearbox described in step 4 is as follows: Using steps 2 and 3, the features of the new vibration signal sample are determined. These features are then input into a trained Softmax classifier to obtain the probability of the new vibration signal sample's health status. The label corresponding to the sample with the highest probability is then taken as the new vibration signal sample. The health label.

7. The gearbox fault diagnosis method based on sparse coding representation of local features according to claim 6, characterized in that, The expression for the health status is as follows: in, As a category, Feature vector These are the parameters of the Softmax classifier.

8. A system for performing the gearbox fault diagnosis method based on local feature sparse coding representation as described in any one of claims 1-7, characterized in that, include, The sample module is used to acquire vibration signal samples of the gearbox under different health conditions to form a sample set; The dictionary module is used to segment each vibration signal sample in the sample set with the same overlap rate to obtain a local sample fragment set, then whiten the local sample fragment set, and use the whitened local sample fragment set to train a sparse coding model to obtain the dictionary. The feature extraction module is used to calculate the sparse representation of each segment in each vibration signal sample according to the dictionary, and to determine the features of each vibration signal sample based on the sparse representation. The diagnostic module is used to train a Softmax classifier using vibration signal sample features, and then uses the trained Softmax classifier to diagnose gearbox faults.

Images