Cross-working-condition bearing fault diagnosis method based on combination of time-frequency analysis and feature migration

Abstract

The invention belongs to the technical field of health monitoring, and particularly relates to a cross-working-condition bearing fault diagnosis method based on combination of time-frequency analysis and feature migration, and the method comprises the steps: S1, collecting vibration sample data, and dividing the vibration sample data into a training set and a test set; s2, performing wavelet packet decomposition on each piece of sample data in the training set and the test set to obtain a four-layer wavelet packet decomposition time-frequency coefficient of each piece of sample data; s3, performing two-dimensional reconstruction on the time-frequency coefficient of each piece of sample data; s4, constructing a four-layer time-frequency graph by using the two-dimensional reconstructed time-frequency coefficient of each piece of sample data according to a Z-shaped splicing method, and normalizing the four-layer time-frequency graph; s5, an MSCNN network model is constructed; s6, the four-layer time-frequency graph of each sample is sent to an MSCNN network, and corresponding combined depth features are extracted; s7, calculating the total cross entropy loss of all samples of the training set; and S8, judging whether the total cross entropy loss meets the requirement or not, if so, entering the next step, and otherwise, optimizing the MSCNN according to the total cross entropy loss.

Description

Technical Field

[0001] The present invention belongs to the technical field of health monitoring, and in particular relates to a cross-operating-condition bearing fault diagnosis method based on the combination of time-frequency analysis and feature migration. Background Art

[0002] The automatic tilt mechanism is an important part of the helicopter control system and is the "link" that converts the pilot's control into blade pitch change. The automatic tilt mechanism bearing is a key component of the automatic tilt mechanism. In actual use, it has the characteristics of complex external load conditions and harsh working environment. At the same time, its design has almost no structural redundancy, and once it fails, it will have serious consequences. Therefore, carrying out automatic tilt mechanism bearing fault diagnosis and identification is of great significance to improving the safety and reliability of helicopters.

[0003] The traditional bearing fault diagnosis algorithm consists of three parts: signal acquisition, feature construction and fault diagnosis. Since the bearing vibration signal contains a lot of dynamic information, vibration sensors are often placed at key locations to collect vibration signals to complete bearing fault diagnosis. At present, bearing fault diagnosis algorithms are divided into two categories: traditional signal analysis and machine learning and deep learning algorithms. Traditional signal analysis and machine learning have high requirements for prior knowledge and feature screening experience, and cannot meet the development needs of big data in the industrial field. Compared with traditional fault diagnosis algorithms, deep learning algorithms can independently extract the original bearing vibration signal features, reduce the influence of human experience, and are more suitable for processing complex bearing vibration data. However, deep learning algorithms cannot fully extract features of non-stationary and nonlinear vibration signals, which affects the final classification accuracy. Therefore, combining time-frequency analysis processing algorithms with deep learning can not only solve the problem of feature loss, but also make full use of the powerful high-precision fault analysis and diagnosis capabilities of deep learning. Although the combination of time-frequency analysis and deep learning has achieved good results under the same working conditions, in the application scenario of variable working conditions, the data of the training set and the test set do not meet the assumption of independent and identical distribution, which makes it difficult for the deep learning model trained with the training set data to achieve good results on the test set data. Therefore, it is necessary to introduce transfer learning methods to solve this problem. Summary of the invention

[0004] Purpose of the invention: To propose a cross-operating-condition bearing fault diagnosis method based on the combination of time-frequency analysis and feature migration, to perform multi-scale refinement analysis on the original signal (including training set and test set) using wavelet packet transform to construct a wavelet packet decomposition time-frequency graph, and to input the time-frequency graph into a multi-core convolutional neural network structure to complete the construction and testing of the migration model, so as to realize cross-operating-condition bearing fault identification.

[0005] Technical Solution

[0006] A bearing fault diagnosis method based on the combination of time-frequency analysis and feature migration includes the following steps:

[0007] S1: Collect vibration sample data and divide the vibration sample data into a training set and a test set;

[0008] S2: Perform wavelet packet decomposition on each sample data in the training set and the test set to obtain the four-layer wavelet packet decomposition time-frequency coefficients of each sample data;

[0009] S3: Reconstruct the time-frequency coefficients of each sample data in two dimensions;

[0010] S4: construct a four-layer time-frequency diagram using the time-frequency coefficients of each sample data after two-dimensional reconstruction according to the zigzag splicing method and normalize them;

[0011] S5: Build a MSCNN network model;

[0012] S6: Send the four-layer time-frequency graph of each sample into the MSCNN network to extract the corresponding combined deep features;

[0013] S7: Calculate the total cross entropy loss of all samples in the training set;

[0014] S8: Determine whether the total cross entropy loss meets the requirements. If so, proceed to the next step. Otherwise, optimize the MSCNN network according to the total cross entropy loss.

[0015] S9: Calculate the MK-MMD distance loss of all samples in the training set and all samples in the test set in the high-dimensional feature space LMK-MMD ;

[0016] S10: Calculate the overall loss L. If the overall loss meets the requirements, proceed to S12, otherwise proceed to the next step;

[0017] S11 determines whether it is equal to the set number of iterations. If so, enter S12 and optimize the MSCNN network according to the overall loss L. Otherwise, return to S6;

[0018] S12: Use the trained MSCNN network to extract test set sample data and determine the fault type.

[0019] Furthermore, in S1, the sample data ratio of the training set and the test set is freely set and can be determined according to the actual data situation; each sample data of the training set contains a fault label, and the sample data of the test set does not contain a data label.

[0020] Furthermore, in S3, the two-dimensional reconstruction process is as follows: each sample data is evenly divided into N segments, the length of each segment is N, and the N segments of data are used as N rows of the two-dimensional reconstruction matrix to construct a matrix with N rows and N columns. The calculation formula of N is as follows:

[0021]

[0022] Among them, Len represents the length of each sample data, and [] means rounding down.

[0023] Furthermore, in S4, the construction process of the four-layer time-frequency graph is as follows: the four-layer wavelet decomposition package decomposes the time-frequency coefficients into two-dimensional reconstruction and is defined as N1-N16 from low frequency to high frequency, and the four-layer time-frequency graph [N1, N2, N6, N7; N3, N5, N8, N13; N4, N9, N12, N14; N10, N11, N15, N16] is constructed according to the Z-shaped splicing method. The purpose of this process is to make the distribution of time-frequency coefficients more uniform, so as to facilitate MSCNN to extract the deep features of the time-frequency graph.

[0024] Furthermore, in S5, the MSCNN network model includes: convolution preprocessing, including convolution layer, ReLu layer and normalization layer, three parallel convolution processes, including 6 convolution layers, 3 ReLu layers, 6 normalization layers and a global average pooling layer, the processing kernels of the three parallel convolution processes are 3*3, 5*5 and 7*7 respectively, and convolution post-processing, including a fully connected layer and a Softmax layer. Compared with the single-core model, the MSCNN model can more comprehensively extract the deep features in the time-frequency graph and can more accurately complete data migration and fault identification.

[0025] Furthermore, in S7, the total cross entropy loss is calculated as follows:

[0026]

[0027] Among them, N represents the total number of samples, M represents the number of classification categories, and F Sic represents the symbolic function, which is 1 if the classification is successful and 0 if the classification fails. ic It represents the predicted probability that sample i belongs to category c. The purpose of calculating the cross entropy loss is to optimize the fully connected layer of the MSCNN network to ensure that the MSCNN network can effectively recognize the training set data.

[0028] Furthermore, in S8, the fully connected layer of the MSCNN network is optimized according to the total cross entropy loss, and the optimization parameters include the weights and biases of the fully connected layer.

[0029] Furthermore, in S9, the MK-MMD distance loss LMK-MMD The calculation formula is as follows:

[0030]

[0031] Among them, P represents the training set, Q represents the test set, φ represents the regenerated Hilbert space mapping, E represents the mathematical expectation of the data, H k represents the kernel k used to reproduce the Hilbert space.

[0032] Furthermore, in S11, the weights and biases of the multi-core convolutional layers of the MSCNN network are optimized according to the overall loss L.

[0033] In summary, the beneficial effects of the present invention are as follows:

[0034] The present invention proposes a bearing fault diagnosis method based on the combination of time-frequency analysis and feature migration. After WPT processing of the signal, multi-scale time-frequency domain information of the vibration signal can be obtained. The time-frequency diagram is constructed using the coefficients obtained after decomposition, and the time-frequency diagram is input into MSCNN. The cross entropy loss and high-dimensional space MK-MDD loss are used to optimize the MSCNN and the fully connected layer, and the migration of the fault diagnosis model from the training set data to the test set data is completed, so as to realize the identification of the bearing fault type and fault scale, and provide reasonable suggestions for bearing maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of a bearing fault diagnosis method based on the combination of time-frequency analysis and feature migration.

[0036] Figure 2 The original time domain image of the vibration signal collected for the bearing test bench.

[0037] Figure 3 For Figure 2 Schematic diagram of the 16-node wavelet coefficients obtained after the vibration signal is decomposed by 4-layer wavelet packet.

[0038] Figure 4 Schematic diagram of Z-shaped splicing method and Figure 3 Two-dimensional time-frequency diagram after splicing the time-frequency coefficients.

[0039] Figure 5 Schematic diagram of the network structure of MSCNN.

[0040] Figure 6 Schematic diagram of the three-layer wavelet packet decomposition of the signal.

[0041] Figure 7 This is the prediction confusion matrix diagram obtained after fault diagnosis of the test set data. DETAILED DESCRIPTION

[0042] A bearing fault diagnosis method based on the combination of time-frequency analysis and feature migration has the following steps:

[0043] a) The bearing fault implantation experiment was completed on the bearing fatigue loading test machine of Luoyang Bearing Factory. Two acceleration measuring points were installed on the test machine. Each measuring point contained a single-axis acceleration sensor in three directions. The bearing used in the fault diagnosis test bench was an automatic tilt mechanism for a certain type of helicopter. The fault implantation method was electric spark engraving. The fault modes included inner ring fault, outer ring fault and ball fault. The fault size was a groove width of 1.5 mm and a groove depth of 0.4 mm (the groove depth of the ball fault was 0.34 mm).

[0044] b) First, the collected vibration data samples are divided into training set and test set data, where the training set data contains bearing data samples and corresponding fault type labels, and the test set data only contains bearing sample data but does not contain corresponding fault type labels.

[0045] c) Perform wavelet packet decomposition on the sample signals of the training set (training set) and the test set (test set) according to formula (1) to obtain the time-frequency coefficients of the three-layer, four-layer and five-layer wavelet packet decomposition of each sample.

[0046]

[0047] d) Reconstruct the time-frequency coefficients in two dimensions.

[0048] e) The two-dimensionally reconstructed time-frequency coefficients are used to construct a four-layer time-frequency diagram using a zigzag splicing method, and normalized according to formula (2).

[0049]

[0050] f) The four-layer time-frequency graphs of the training set and the test set are sent to MSCNN, and MSCNN is used to extract the combined deep features of the four-layer wavelet packet time-frequency graphs.

[0051] g) Using the Softmax classification result probability of the training set sample data and the fault type label of the training set data, the cross entropy loss is calculated according to formula (3).

[0052]

[0053] h) Calculate the MK-MMD distance between the training set data and the test set data in the high-dimensional feature space according to formula (4), and obtain the MK-MDD loss L of the training set feature Fs and the test set data Ft in the high-dimensional space MK-MMD .

[0054]

[0055] i) Reversely optimize the fully connected layer by calculating the cross entropy loss to obtain the optimized MSCNN.

[0056] j) Iterate steps e)-h) until the total loss L and the number of iterative loss times meet the target requirements, and obtain the trained MSCNN and FC.

[0057] L=L c +λL MK-MMD (P,Q) (5)

[0058] The trained MSCNN is used to extract the time-frequency graph features of the test set samples, and the time-frequency graph features are input into the fully connected layer FC to obtain the corresponding fault type.

[0059] Example

[0060] Taking the vibration data collected under 50% load to the vibration data under 75% load as an example, the process of this patent is explained. First, the data is divided into training set data and test set data. The training set is 10 samples for each type of failure mode, a total of 100 samples, and the test set is 20 samples for each type of failure mode, a total of 200 samples. For the time domain data of the training set and the test set, such as Figure 2 As shown, 4 layers of wavelet packet decomposition are performed to obtain 16 wavelet packet coefficients, such as Figure 3 The 16 wavelet packet coefficients are spliced ​​in a Z-shaped pattern to construct the time-frequency diagrams of the training set and the test set, as shown in Figure 4 As shown. The training set data is input into the MSCNN network, and the network results are as follows Figure 5 As shown in the figure, a multi-core convolutional neural network is used to extract the time-frequency features of the test set samples. The splicing layer splices the time-frequency features extracted by each single-core neural network, and the spliced ​​time-frequency features are input into the fully connected layer and the Softmax layer to complete the fault type identification. The cross entropy loss is calculated to reversely optimize the fully connected layer to obtain the optimized MSCNN. The MK-MMD distance between the training set samples and the test set samples in the high-dimensional feature space is calculated to obtain the MK-MDD loss in the high-dimensional space. The weights and biases of the MSCNN multi-core convolutional layer are optimized through the overall loss to obtain a fault diagnosis model that meets the optimization requirements. The optimization process is shown in Figure 1 The test set data is input into the fault diagnosis model after training and optimization to obtain the fault type of the vibration sample. The prediction confusion matrix corresponding to 200 groups of samples is shown as follows: Figure 7 shown.

Claims

1. A bearing fault diagnosis method based on the combination of time-frequency analysis and feature migration, characterized in that: The following steps are involved: S1: Collect vibration sample data and divide the vibration sample data into a training set and a test set; S2: Perform wavelet packet decomposition on each sample data in the training set and the test set to obtain the four-layer wavelet packet decomposition time-frequency coefficients of each sample data; S3: Reconstruct the time-frequency coefficients of each sample data in two dimensions; S4: construct a four-layer time-frequency diagram using the time-frequency coefficients of each sample data after two-dimensional reconstruction according to the zigzag splicing method and normalize them; S5: Build a MSCNN network model; S6: Send the four-layer time-frequency graph of each sample into the MSCNN network to extract the corresponding combined deep features; S7: Calculate the total cross entropy loss of all samples in the training set; S8: Determine whether the total cross entropy loss meets the requirements. If so, proceed to the next step. Otherwise, optimize the MSCNN network according to the total cross entropy loss. S9: Calculate the MK-MMD distance loss L of all samples in the training set and all samples in the test set in the high-dimensional feature space MK-MMD ; S10: Calculate the overall loss L. If the overall loss meets the requirements, proceed to S12, otherwise proceed to the next step; S11 determines whether it is equal to the set number of iterations. If so, enter S12 and optimize the MSCNN network according to the overall loss L. Otherwise, return to S6; S12: Use the trained MSCNN network to extract test set sample data and determine the fault type.

2. The method according to claim 1, characterized in that: In S1, the sample data ratio of the training set and the test set is: each sample data of the training set contains a fault label, and the sample data of the test set does not contain a data label.

3. The method according to claim 2, characterized in that: In S3, the two-dimensional reconstruction process is as follows: divide each sample data into N segments, each segment has a length of N, and use the N segments as N rows of the two-dimensional reconstruction matrix to construct a matrix with N rows and N columns. The calculation formula of N is as follows: Among them, Len represents the length of each sample data, and [] means rounding down.

4. The method according to claim 3, characterized in that: In S4, the construction process of the four-layer time-frequency diagram is as follows: the four-layer wavelet decomposition package decomposes the time-frequency coefficients into two-dimensional reconstruction and is defined as N1-N16 from low frequency to high frequency. The four-layer time-frequency diagram is constructed according to the Z-shaped splicing method, as follows: [N1, N2, N6, N7; N3, N5, N8, N13; N4, N9, N12, N14; N10, N11, N15, N16].

5. The method according to claim 4, characterized in that: In S5, the MSCNN network model includes: convolution preprocessing, including convolution layer, ReLu layer and normalization layer, three parallel convolution processes, including 6 convolution layers, 3 ReLu layers, 6 normalization layers and a global average pooling layer, the processing kernels of the three parallel convolution processes are 3*3, 5*5 and 7*7 respectively, and convolution postprocessing, including a fully connected layer and a Softmax layer.

6. The method according to claim 5, characterized in that: In S7, the total cross entropy loss is calculated as follows: Among them, N represents the total number of samples, M represents the number of classification categories, and F Sic represents the symbolic function, which is 1 if the classification is successful and 0 if the classification fails. ic represents the predicted probability that sample i belongs to category c.

7. The method according to claim 6, characterized in that: In S8, the fully connected layer is optimized according to the total cross entropy loss, and the optimization parameters include the weights and biases of the fully connected layer.

8. The method according to claim 7, characterized in that: In S9, MK-MMD distance loss LMK-MMD The calculation formula is as follows: Among them, P represents the training set, Q represents the test set, φ represents the regenerated Hilbert space mapping, E represents the mathematical expectation of the data, H k represents the kernel k used to reproduce the Hilbert space.

9. The method according to claim 8, characterized in that: In S11, the weights and biases of the multi-core convolutional layers of the MSCNN network are optimized according to the overall loss L.

Images