Gear fault signal identification method based on multi-scale wavelet packet energy entropy-ELM
Through the multi-scale wavelet packet energy entropy-ELM method, the multi-scale energy entropy feature vector is extracted and the ELM neural network is used to solve the problem of incomplete signal feature extraction in gear fault diagnosis, improve diagnostic accuracy and efficiency, and reduce calculation time .
Patent Information
- Application Number
- CN202410757732.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-01-25
- Filing Date
- 2024-06-13
- Publication Date
- 2024-12-03
AI Technical Summary
In the prior art In gear fault diagnosis, the original vibration signal processing cannot fully detect and extract local characteristics of the signal, resulting in low diagnostic accuracy, and the support vector machine is long training time on large-scale data sets and is sensitive to noise and overlapping samples.
The multi-scale wavelet packet energy entropy-ELM method is used to perform multi-scale division and wavelet packet decomposition on time series data, and multi-scale energy entropy feature vectors are extracted, and fault identification is used using ELM neural network.
It improves the accuracy and efficiency of gear fault diagnosis, simplifies signal processing complexity, reduces calculation time, and improves the processing capability of non-stationary signals.
Smart Images

Figure CN119066478A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of gear fault diagnosis, and in particular relates to a gear fault signal recognition method based on multi-scale wavelet packet energy entropy (ELM). Background Art
[0002] As a key part of mechanical equipment, gearboxes play an irreplaceable role in modern industry and are often used in many large-scale machines. For example, gearboxes are used in aerospace, automobiles, and automated mechanical equipment. Since large machines have a large workload during operation and often work in complex environments, gearbox tooth breakage failures may occur, causing a lot of losses. Therefore, gearbox fault diagnosis is particularly important.
[0003] At present, there are many types of gearbox faults, and the signals generated by the faults are relatively complex, making gearbox fault diagnosis more difficult.
[0004] Processing the original vibration signal can improve the accuracy of gear fault diagnosis. However, the original vibration signal only extracts and detects the local features of the signal from a single scale, and cannot fully detect and extract the local features in the signal.
[0005] Energy entropy has the phenomenon of incomplete signal feature extraction when extracting signal features. For example, energy entropy may be sensitive to the amplitude change of the signal when extracting signal features, but is not sensitive to the spectral characteristics, timing characteristics and nonlinear characteristics of the signal. Therefore, when using energy entropy to diagnose gear faults, it is difficult to accurately extract fault features and the diagnostic accuracy is low.
[0006] At the same time, there are many types of gearbox faults. At present, deep learning algorithms are widely used. Among them, the use of learning algorithms such as support vector machine (SVM) can effectively classify gearbox faults. Support vector machine (SVM) can improve the diagnosis accuracy by using high-dimensional feature space, maximizing intervals, focusing on support vectors, processing nonlinear relationships and parameter tuning. Support vector machine (SVM) can process high-dimensional feature space; has good generalization ability; can process nonlinear relationships; and works well for small sample data sets. However, SVM still has many shortcomings, such as long training time for large-scale data sets; sensitive to noise and overlapping samples; and needs to be expanded for multi-category classification problems. Therefore, it is particularly important to choose a suitable learning algorithm. Summary of the invention
[0007] The purpose of the present invention is to provide a gear fault signal recognition method based on multi-scale wavelet packet energy entropy - ELM, so as to solve the problems that the prior art proposed in the above background technology cannot comprehensively detect and extract local features in the signal when processing the original vibration signal, the fault features are difficult to accurately extract and the diagnostic accuracy is low when using energy entropy to diagnose gear faults, and the support vector machine has the problems of long training time for large-scale data sets and sensitivity to noise and overlapping samples.
[0008] To achieve the above object, the present invention adopts the following technical solutions:
[0009] The gear fault signal recognition method based on multi-scale wavelet packet energy entropy (ELM) includes the following steps:
[0010] S1: Collect the time series data of the acceleration of the driving end of the sun gear under normal and different degrees of broken tooth damage fault state, perform multi-scale division processing on the time series data, and decompose the time series data into multiple sub-signals of different scales;
[0011] S2: Decompose the sub-signals after multi-scale division using the wavelet function dmey, and reconstruct the sub-signals of each scale;
[0012] S3: extracting the energy entropy of wavelet packets of different scales under different working conditions, obtaining feature vectors composed of energy entropies at different scales, and classifying and marking each group of feature vectors under different working conditions;
[0013] S4: The ELM neural network is trained using each group of feature vectors and their classification labels under different working conditions, and different fault states of the gear are identified through the trained ELM neural network.
[0014] Preferably, the time series data is processed by multi-scale division in S1, specifically as follows:
[0015] Assume that the signal sequence to be analyzed is X = {x1, x2, ..., x n}, with a length of N, under different scale factors, establish a new coarse-grained sequence:
[0016]
[0017] Where τ is a positive integer, called the scaling factor; for non-zero τ, the original sequence X generates a length The coarse-grained sequence Y of .
[0018] Preferably, S2 specifically includes the following steps:
[0019] S201: Assume that the given orthogonal wavelet function is x(t) and the scaling function is y(t), and let the relationship between the orthogonal wavelet function and the scaling function be:
[0020]
[0021] Among them, z0(t)=y(t) and z1(t)=x(t) correspond to the two-scale equation and wavelet equation respectively;
[0022]
[0023] Among them, {l a} a=z is the low-pass filter function of y(t), {h a} a=z is the high-pass filter function of x(t), where: h(a) = (-1) a (1-a)l;
[0024] The above formula is changed into iterative recursive form:
[0025]
[0026] Iteration function {Z n (t)}(n∈Z + ) is the orthogonal wavelet packet corresponding to z0(t)=y(t);
[0027] S202: Perform j-layer wavelet packet decomposition on the newly established coarse-grained sequence, and obtain 2 j Sub-signals in different frequency bands
[0028] The recursive form of the decomposition algorithm is:
[0029]
[0030] S203: Perform wavelet packet reconstruction on each frequency band of the decomposed sub-signal. The reconstruction coefficient is represented by S j,k (k=0,1,2,…,2 j -1);
[0031] The reconstruction algorithm is:
[0032]
[0033] Preferably, the wavelet packet energy entropy of different scales under different working conditions is extracted in S3, specifically as follows:
[0034] Calculate the energy value E of each frequency band in S203 j,k :
[0035]
[0036] In the formula, A i (t) is the maximum value of the signal at this node, ti-1 and t i are the start time and end time of the node signal respectively;
[0037] Calculate the total energy value E:
[0038]
[0039] Calculate the probability p of each child node energy relative to the total energy j,k :
[0040]
[0041] Calculate the energy entropy value H of the kth node in the jth layer after signal decomposition j,k :
[0042] H j,k =-p j,k (i)lg|p j,k (i)
[0043] Calculate the energy entropy value of the kth wavelet packet in the jth layer after signal decomposition:
[0044]
[0045] Preferably, in S1, the time series data is decomposed into three sub-signals of different scales to obtain three new coarse-grained sequences, wherein the scale factors of the three new coarse-grained sequences are equal to 1, 2, and 3, respectively.
[0046] Preferably, the characteristic vector composed of energy entropy at different scales is obtained in S3, which is as follows:
[0047] The energy entropy values of 8 nodes at scales of 1, 2, and 3 are obtained respectively, and a total of 24 multi-scale energy entropy values are obtained as the eigenvalues of the sample. The obtained eigenvalues are used as the eigenvectors, and the dimension of the eigenvectors is 3×8.
[0048] Preferably, in said S4, each group of feature vectors under different working conditions are classified and marked, specifically, the classification attributes of each group of feature vectors under different working conditions are marked as normal, 30% broken teeth, 60% broken teeth, and 90% broken teeth, respectively.
[0049] Preferably, in S4, each group of feature vectors and their classification labels under different working conditions are used to train the ELM neural network, specifically as follows:
[0050] S401: The feature vector represents the feature value of the sample as the input value, and its classification attribute is used as the expected output of the corresponding sample; the sample is divided into a training set and a test set;
[0051] S402: The number of hidden layer nodes of the ELM neural network is set to 30, and the sigmoid function is selected as the activation function;
[0052] S403: Using the training set to train the ELM neural network, using the feature vector and the classification attribute as the input value and expected output of the ELM neural network for training; using the test set to test the trained ELM neural network.
[0053] Compared with the prior art, the present invention has the following beneficial effects:
[0054] The gear fault diagnosis method of the present invention aims at the problem that the original vibration signal of the gear box only extracts and detects the local features of the signal from a single scale, and it is not easy to detect and extract the local features in the signal. In addition, the original vibration signal has a high dimension and the calculation time is long. The present invention proposes a method for multi-scale division of the original signal, decomposing the original vibration signal into three sub-signals of different scales, which is easier to comprehensively detect and extract the local features in the signal, and it can realize the compression and dimensionality reduction of the original signal, simplify the complexity of signal processing, and improve the calculation efficiency. Combined with the advantage of higher frequency resolution of the wavelet packet algorithm, the wavelet packet algorithm is used to decompose the signal after multi-scale division, which can better process non-stationary signals and reconstruct the decomposed signals. The present invention proposes the concept of multi-scale energy entropy, which can more completely extract the fault features of the original signal. Finally, the extreme learning machine (ELM) model is used for fault diagnosis, which improves the ability of gear fault diagnosis compared with the support vector machine (SVM). BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 It is a flow chart of the gear fault signal identification method based on multi-scale wavelet packet energy entropy-ELM in the present invention;
[0056] Figure 2 A comparison diagram of confusion matrix diagrams of samples subjected to ELM neural network training and SVM model training in Example 1 of the present invention;
[0057] Figure 3 This is a comparison chart of the results of ELM neural network training and SVM model training for the samples in Example 1 of the present invention. DETAILED DESCRIPTION
[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0059] Embodiment 1:
[0060] like Figure 1 As shown, the gear fault signal recognition method based on multi-scale wavelet packet energy entropy-ELM includes the following steps:
[0061] S1: Obtain in advance the time series data of the acceleration of the driving end of the sun gear under normal and different degrees of broken tooth damage fault conditions, perform multi-scale division processing on the time series data, and decompose the time series data into three sub-signals of different scales.
[0062] S101: Assume that the signal sequence to be analyzed is X = {x1, x2, ..., x n}, with a length of N, under different scale factors, a new coarse-grained sequence is established
[0063]
[0064] Where τ is a positive integer, called the scaling factor. Obviously, when τ = 1, the coarse-grained sequence Y(1) is the original sequence. For non-zero τ, the original sequence X generates a (Indicates not greater than The coarse-grained
[0065] Sequence Y.
[0066] S2: By selecting the wavelet function dmey for decomposition, the wavelet packet algorithm decomposes the sub-signals after multi-scale division, and reconstructs the sub-signals of each scale.
[0067] S201: Assume that the given orthogonal wavelet function is x(t) and the scaling function is y(t), and let the relationship between the orthogonal wavelet function and the scaling function be:
[0068]
[0069] Among them, z0(t)=y(t) and z1(t)=x(t) correspond to the two-scale equation and the wavelet equation respectively.
[0070]
[0071] Among them, {l a} a=z is the low-pass filter function of y(t), {h a} a=z is the high-pass filter function of x(t), where: h(a) = (-1) a (1-a)l.
[0072] The above formula is changed into iterative recursive form:
[0073]
[0074] Iteration function {Z n (t)}(n∈Z + ) is the orthogonal wavelet packet corresponding to z0(t)=y(t).
[0075] S202: Create a new coarse-grained sequence in S101 Perform j-layer wavelet packet decomposition, and get 2 j Sub-signals in different frequency bands
[0076] The recursive form of the decomposition algorithm is:
[0077]
[0078] S203: Perform wavelet packet reconstruction on each frequency band of the decomposed sub-signal. The reconstruction coefficient is represented by S j,k (k=0,1,2,…,2 j -1);
[0079] The reconstruction algorithm is:
[0080]
[0081] S3: Extract and calculate the wavelet packet energy entropy of different scale factors under different working conditions, obtain the feature vector composed of the energy entropy under different scale factors, and classify and mark each group of feature vectors under different working conditions.
[0082] S301: Calculate the energy value E of each frequency band in S203 j,k .
[0083]
[0084] In the formula, A i (t) is the maximum value of the signal at this node, t i-1 and t i are the start time and end time of the node signal respectively.
[0085] Calculate the total energy value E:
[0086]
[0087] Calculate the probability p of each child node energy relative to the total energy j,k :
[0088]
[0089] Calculate the energy entropy value H of the kth node in the jth layer after signal decomposition j,k :
[0090] H j,k =-p j,k (i)lg|p j,k (i)|
[0091] Calculate the energy entropy value of the kth wavelet packet in the jth layer after signal decomposition:
[0092]
[0093] Since the original vibration signal is processed by multi-scale division, the original vibration signal is decomposed into three sub-signals of different scales, so the energy entropy values of 8 nodes are obtained when the scale factor is equal to 1, the scale factor is equal to 2, and the scale factor is equal to 3. A total of 24 multi-scale energy entropy values are used as the eigenvalues of the corresponding samples. The obtained eigenvalues are used as the eigenvectors, and their dimension is 3×8.
[0094] S302: Mark the classification attributes of the sample feature values of different working conditions as normal, 30% broken teeth, 60% broken teeth, and 90% broken teeth, respectively, and use the classification attributes as the expected output of the corresponding samples.
[0095] S4: The feature vectors of each group under different working conditions obtained in S3 are used to train the neural network using the extreme learning machine (ELM) model, so that the ELM neural network can identify different fault states of the gear.
[0096] S401: Randomly select 50% of the preprocessed samples as the training set, and the remaining 50% of the samples as the test set. Perform ELM neural network training, set the number of hidden layer nodes to 30, and select the sigmoid function as the activation function.
[0097] S402: Train the ELM neural network with the input value and the expected output to obtain an ELM neural network capable of identifying normal, 30% broken tooth, 60% broken tooth, and 90% broken tooth fault states.
[0098] S403: input the training samples into the network for training, test the trained rolling bearing fault classification network, input the test samples into the network to obtain an output result and compare it with the expected output, and verify the recognition effect of the ELM neural network.
[0099] The present invention first divides the data to be detected into multiple scales, then performs wavelet packet decomposition on the signals under different scale factors and calculates their energy entropy, which is input into the trained fault classification network as the input value to obtain the output result, i.e., the fault classification label, so as to judge whether the gear has a broken tooth fault and the degree of the broken tooth fault. The present invention achieves a good fault detection and classification effect.
[0100] Experimental verification:
[0101] The extreme learning machine (ELM) model of the present invention and the support vector machine (SVM) model of the prior art are respectively selected to train the neural network.
[0102] The training uses the same experimental process and data: start, collect gearbox vibration signal data, divide the original vibration signal into multiple scales, use the wavelet packet algorithm to decompose and reconstruct the scaled signal, extract and calculate the energy entropy of different scale factors under different working conditions, input the eigenvalues into the trained ELM and SVM models, judge the degree of gear tooth breakage based on the output results, and end.
[0103] Experimental results: The fault diagnosis accuracy of ELM and SVM are as follows: Figure 2 As shown, Figure 2 (a) is a confusion matrix diagram of multi-scale division of data of each state of the sun gear and wavelet packet decomposition and energy entropy extraction as eigenvalues for ELM neural network training. Figure 2 (b) is a confusion matrix diagram of the data of each state of the sun gear divided into multiple scales and subjected to wavelet packet decomposition and energy entropy extraction as feature values for SVM model training. The fault diagnosis accuracy of ELM and SVM is 100% and 99.6875% respectively; the classification results of ELM and SVM are shown in Figure 3 As shown, Figure 3 (a) is the result of multi-scale division of the data of each state of the sun gear and wavelet packet decomposition and energy entropy extraction as the eigenvalue for ELM neural network training. Figure 3 (b) is the result of multi-scale division of the data of each state of the sun gear and wavelet packet decomposition and energy entropy extraction as the eigenvalue for SVM model training. It also shows that the ELM fault diagnosis accuracy is 100% and there are wrong predictions in SVM fault diagnosis.
[0104] The above description is only used to help understand the method of the present invention and its core essence, but the protection scope of the present invention is not limited thereto. For those skilled in the art in the art, equivalent replacement or change according to the technical solution and inventive concept of the present invention within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. Gear fault signal recognition method based on multi-scale wavelet packet energy entropy-ELM, It is characterized in that The steps include: S1: Collect the time series data of the acceleration of the driving end of the sun gear under normal and different degrees of broken tooth damage fault state, perform multi-scale division processing on the time series data, and decompose the time series data into multiple sub-signals of different scales; S2: Decompose the sub-signals after multi-scale division using the wavelet function dmey, and reconstruct the sub-signals of each scale; S3: extracting the energy entropy of wavelet packets of different scales under different working conditions, obtaining feature vectors composed of energy entropies at different scales, and classifying and marking each group of feature vectors under different working conditions; S4: The ELM neural network is trained using each group of feature vectors and their classification labels under different working conditions, and different fault states of the gear are identified through the trained ELM neural network.
2. The gear fault signal recognition method based on multi-scale wavelet packet energy entropy-ELM according to claim 1, It is characterized in that In S1, the time series data is processed by multi-scale division, as follows: Assume that the signal sequence to be analyzed is X = {x 1 ,x 2 ,…,x n }, with a length of N, under different scale factors, establish a new coarse-grained sequence: Where τ is a positive integer, called the scaling factor; for non-zero τ, the original sequence X generates a length The coarse-grained sequence Y of .
3. The gear fault signal recognition method based on multi-scale wavelet packet energy entropy-ELM according to claim 2, It is characterized in that The S2 specifically includes the following steps: S201: Assume that the given orthogonal wavelet function is x(t) and the scaling function is y(t), and let the relationship between the orthogonal wavelet function and the scaling function be: Among them, z 0 (t) = y(t) and z 1 (t) = x(t) corresponds to the two-scale equation and the wavelet equation respectively; Among them, {l a } a=z is the low-pass filter function of y(t), {h a } a=z is the high-pass filter function of x(t), where: h(a) = (-1) a (1-a)l; The above formula is changed into iterative recursive form: Iteration function {Z n (t)}(n∈Z + ) is z 0 (t) = orthogonal wavelet packet corresponding to y(t); S202: Perform j-layer wavelet packet decomposition on the newly established coarse-grained sequence, and obtain 2 j sub-signals in different frequency bands; S203: Perform wavelet packet reconstruction on each frequency band of the decomposed sub-signal. The reconstruction coefficient is represented by S j,k (k=0,1,2,…,2 j -1).
4. The gear fault signal recognition method based on multi-scale wavelet packet energy entropy-ELM according to claim 3, It is characterized in that In S3, the wavelet packet energy entropy of different scales under different working conditions is extracted as follows: Calculate the energy value E of each frequency band in S203 j,k : In the formula, A i (t) is the maximum value of the signal at this node, t i-1 and t i are the start time and end time of the node signal respectively; Calculate the total energy value E: Calculate the probability p of each child node energy relative to the total energy j,k : Calculate the energy entropy value H of the kth node in the jth layer after signal decomposition j,k : H j,k =-p j,k (i)lgp j,k (i) Calculate the energy entropy value of the kth wavelet packet in the jth layer after signal decomposition:
5. The gear fault signal recognition method based on multi-scale wavelet packet energy entropy-ELM according to claim 4, It is characterized in that In S1, the time series data is decomposed into three sub-signals of different scales to obtain three new coarse-grained sequences, wherein the scale factors of the three new coarse-grained sequences are equal to 1, 2, and 3, respectively.
6. The gear fault signal recognition method based on multi-scale wavelet packet energy entropy (ELM) according to claim 5, It is characterized in that The characteristic vector composed of energy entropy at different scales is obtained in S3, which is as follows: The energy entropy values of 8 nodes at scales of 1, 2, and 3 are obtained respectively, and a total of 24 multi-scale energy entropy values are obtained as the eigenvalues of the sample. The obtained eigenvalues are used as the eigenvectors, and the dimension of the eigenvectors is 3×8.
7. The gear fault signal recognition method based on multi-scale wavelet packet energy entropy (ELM) according to claim 4 or 6, It is characterized in that In the S4, each group of feature vectors under different working conditions are classified and marked, specifically, the classification attributes of each group of feature vectors under different working conditions are marked as normal, 30% broken teeth, 60% broken teeth, and 90% broken teeth, respectively.
8. The gear fault signal recognition method based on multi-scale wavelet packet energy entropy (ELM) according to claim 7, It is characterized in that In S4, the ELM neural network is trained using each group of feature vectors and their classification labels under different working conditions, as follows: S401: The feature vector represents the feature value of the sample as the input value, and its classification attribute is used as the expected output of the corresponding sample; the sample is divided into a training set and a test set; S402: The number of hidden layer nodes of the ELM neural network is set to 30, and the sigmoid function is selected as the activation function; S403: Using the training set to train the ELM neural network, using the feature vector and the classification attribute as the input value and expected output of the ELM neural network for training; using the test set to test the trained ELM neural network.
Citation Information
Patent Citations
Bearing fault classification method and device based on wavelet packet decomposition weight fuzzy entropy and ELM
CN113790890A
PMSM Demagnetization Fault Diagnosis Method Based on Fuzzy Intelligent Learning of Torque Signals
US20220084732A1
