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

Abstract

The application belongs to the technical field of health monitoring, and particularly relates to a cross-condition bearing fault diagnosis method based on time-frequency analysis and feature migration, 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 sample data in the training set and the test set, and obtaining four-layer wavelet packet decomposition time-frequency coefficients of each sample data; S3: performing two-dimensional reconstruction on the time-frequency coefficients of each sample data; S4: constructing four-layer time-frequency maps according to a Z-shaped splicing method according to the two-dimensional reconstructed time-frequency coefficients of each sample data, and performing normalization; S5: constructing an MSCNN network model; S6: sending the four-layer time-frequency maps of each sample into the MSCNN network, and extracting corresponding combined deep features; S7: calculating total cross-entropy loss of all samples in the training set; S8: judging whether the total cross-entropy loss meets the requirements, if yes, entering the next step, otherwise, optimizing the MSCNN network according to the total cross-entropy loss.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of health monitoring, and particularly relates to a bearing fault diagnosis method based on time-frequency analysis and feature migration combination across working conditions. BACKGROUND

[0002] The automatic tilt is an important part of the helicopter control system, and is a "link" for converting the pilot's operation into pitch change of the rotor blade. The automatic tilt bearing is a key component of the automatic tilt, and has the characteristics of complex external load conditions and harsh working environment in actual use. Moreover, there is almost no structural redundancy in the design of the automatic tilt bearing, and once the automatic tilt bearing fails, serious consequences will be caused. Therefore, bearing fault diagnosis and identification is of great significance to improve the safety and reliability of the helicopter.

[0003] The traditional bearing fault diagnosis algorithm includes three parts: signal acquisition, feature construction and fault diagnosis. Since the bearing vibration signal contains a large amount of dynamic information, vibration sensors are usually arranged at key positions 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 algorithm. Traditional signal analysis and machine learning have high requirements for prior knowledge and feature selection experience, and cannot meet the development needs of big data in the industrial field. Deep learning algorithm can autonomously extract original bearing vibration signal features, reduce the influence of human experience, and is more suitable for processing complex bearing vibration data. However, the deep learning algorithm cannot fully extract features from non-stationary and nonlinear vibration signals, which affects the final classification accuracy. Therefore, combining time-frequency analysis processing algorithm and deep learning can solve the feature loss problem and fully utilize the powerful high-precision fault analysis and diagnosis capability of deep learning. Although the combination of time-frequency analysis and deep learning achieves good results in the same working condition environment, the data of the training set and the test set do not meet the independent and identically distributed assumption in the variable working condition application scene, so that the deep learning model trained by the training set data is difficult to achieve good results on the test set data. Therefore, it is necessary to introduce a transfer learning method to solve this problem. SUMMARY

[0004] The application aims to provide a bearing fault diagnosis method based on time-frequency analysis and feature migration combination across working conditions, which uses wavelet packet transform to perform multi-scale refinement analysis on the original signal (including the training set and the test set) to construct a wavelet packet decomposition time-frequency diagram, inputs the time-frequency diagram into a multi-kernel convolutional neural network structure to complete the migration model building and testing, and realizes bearing fault identification across working conditions.

[0005] TECHNICAL SCHEME

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

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

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

[0009] S3: two-dimensional reconstruction is performed on the time-frequency coefficients of each sample data;

[0010] S4: the time-frequency coefficients of each sample data after two-dimensional reconstruction are constructed into four-layer time-frequency maps according to a zigzag concatenation method, and are normalized;

[0011] S5: an MSCNN network model is constructed;

[0012] S6: the four-layer time-frequency maps of each sample are input into the MSCNN network, and corresponding combined deep features are extracted;

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

[0014] S8: it is determined whether the total cross-entropy loss meets the requirements, if yes, the next step is entered, otherwise the MSCNN network is optimized according to the total cross-entropy loss;

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

[0016] S10: the total loss L is calculated, if the total loss meets the requirements, S12 is entered, otherwise the next step is entered;

[0017] S11: it is determined whether the iteration number is equal to a set iteration number, if yes, S12 is entered, otherwise the MSCNN network is optimized according to the total loss L, and S6 is returned;

[0018] S12: the test set sample data is proposed by using the trained MSCNN network, and fault type judgment is performed.

[0019] Further, in S1, the sample data proportion of the training set and the test set is freely set, and can be determined according to the actual data; each sample data of the training set includes a fault label, and the sample data of the test set does not include a data label.

[0020] Further, in S3, the two-dimensional reconstruction process is as follows: each sample data is evenly divided into N segments, and the length of each segment is N; N segments of data are taken as N rows of a two-dimensional reconstruction matrix, to construct an N-row-by-N-column matrix, and the calculation formula of N is as follows:

[0021]

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

[0023] Further, in S4, the four-layer time-frequency map construction process is as follows: the four-layer wavelet decomposition package is decomposed, and after two-dimensional reconstruction of the time-frequency coefficients, N1-N16 are defined in turn from low frequency to high frequency, and a four-layer time-frequency map [N1, N2, N6, N7; N3, N5, N8, N13; N4, N9, N12, N14; N10, N11, N15, N16] is constructed by the Z-shaped splicing method. The purpose of this process is to make the time-frequency coefficient distribution more uniform, facilitating the MSCNN to extract the depth features of the time-frequency map.

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

[0025] Further, in S7, the total cross-entropy loss calculation formula is as follows:

[0026]

[0027] wherein, N represents the total number of samples, M represents the number of classification categories, F Sic represents the sign function, 1 if the classification is successful, and 0 if the classification fails, p ic represents the predicted probability of sample i belonging to category c, and the purpose of calculating the cross-entropy loss is to optimize the fully connected layer of the MSCNN network, so that the MSCNN network can realize effective identification of the training set data.

[0028] Further, 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] Further, in S9, the MK-MMD distance loss amount LMK-MMD The calculation formula is as follows:

[0030]

[0031] wherein, P represents the training set, Q represents the test set, φ represents the reproducing Hilbert space mapping, E represents the mathematical expectation of data, H k represents the kernel k used by the reproducing Hilbert space.

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

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

[0034] The bearing fault diagnosis method based on the combination of time-frequency analysis and feature migration is proposed, the multi-scale time-frequency domain information of the vibration signal can be obtained after the WPT processing of the signal, the time-frequency diagram is constructed by using the obtained coefficients after decomposition, the MSCNN is input, the cross-entropy loss and the high-dimensional space MK-MDD loss are used to complete the optimization of the MSCNN and the full connection layer, the migration of the fault diagnosis model from the training set data to the test set data is completed, the bearing fault type and the fault scale are recognized, and reasonable suggestions are provided for bearing maintenance. BRIEF DESCRIPTION OF DRAWINGS

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

[0036] Figure 2 It is the original time domain image of the vibration signal collected by the bearing test bench.

[0037] Figure 3 It is the Figure 2 It is a wavelet coefficient schematic diagram of 16 nodes obtained after 4-layer wavelet packet decomposition of the vibration signal.

[0038] Figure 4 It is a Z-shaped splicing method schematic diagram and the Figure 3 It is a two-dimensional time-frequency diagram after splicing the time-frequency coefficients.

[0039] Figure 5 It is a network structure schematic diagram of the MSCNN.

[0040] Figure 6 It is a 3-layer wavelet packet decomposition schematic diagram of the signal.

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

[0042] The steps of the bearing fault diagnosis method based on the combination of time-frequency analysis and feature migration are as follows:

[0043] a) 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 containing three single-axis acceleration sensors in three directions. The bearing used in the fault diagnosis test bench was a machine part of an automatic tilting device for a certain type of helicopter. The fault implantation method was electric spark carving, and the fault modes included inner ring fault, outer ring fault, and ball fault. The fault size was 1.5mm in groove width and 0.4mm in groove depth (0.34mm for ball fault).

[0044] b) First, the collected vibration data samples were divided into training set and test set data. The training set data included bearing data samples and corresponding fault type labels, and the test set data only included bearing sample data without corresponding fault type labels.

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

[0046]

[0047] d) The time-frequency coefficients were two-dimensionally reconstructed.

[0048] e) The two-dimensionally reconstructed time-frequency coefficients were constructed into four-layer time-frequency graphs according to the zigzag concatenation method, and normalized according to formula (2).

[0049]

[0050] f) The four-layer time-frequency graphs of the training set and the test set were sent to MSCNN, and the four-layer wavelet packet time-frequency graph combination deep features were extracted using MSCNN.

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

[0052]

[0053] h) The MK-MMD distance between the training set data and the test set data in the high-dimensional feature space was calculated according to formula (4), and the high-dimensional space MK-MDD loss amount L MK-MMD .

[0054]

[0055] i) The fully connected layer was optimized by calculating the cross-entropy loss, and the optimized MSCNN was obtained.

[0056] j) iterating steps e)-h) until the total loss L meets the target requirement, obtaining the trained MSCNN and FC.

[0057] L = L c + λL MK-MMD (P, Q) (5)

[0058] Using the trained MSCNN to extract the time-frequency feature of the test set sample, inputting the time-frequency feature into the full connection layer FC to obtain the corresponding fault type.

[0059] Embodiment

[0060] Taking the vibration data collected under 50% load as the vibration data under 75% load as an example, the process of the patent is explained. First, the data is divided into training set data and test set data, the training set is 10 samples for each fault mode, a total of 100 samples, and the test set is 20 samples for each fault mode, a total of 200 samples. The training set and test set time domain data are decomposed by 4-layer wavelet packet as shown in Figure 2 , and 16 wavelet packet coefficients are obtained as shown in Figure 3 . The training set and test set time-frequency diagrams are constructed by Z-shaped splicing according to the 16 wavelet packet coefficients as shown in Figure 4 . The training set data is input into the MSCNN network, and the network result is as shown in Figure 5 . The time-frequency features of the test set sample are extracted by using the multi-core convolutional neural network, the time-frequency features extracted by each single-core neural network are spliced by the splicing layer, the spliced time-frequency features are input into the full connection layer and the Softmax layer to complete the fault type recognition, the cross-entropy loss is calculated to optimize the full connection layer, the optimized MSCNN is obtained, the MK-MMD distance of the training set sample and the test set sample in the high-dimensional feature space is calculated, the high-dimensional space MK-MDD loss is obtained, the MSCNN multi-core convolutional layer weight and bias are optimized by the total loss, and the fault diagnosis model meeting the optimization requirement is obtained, and the optimization process is as shown in Figure 1 . The test set data is input into the trained and optimized fault diagnosis model, and the fault type of the vibration sample is obtained. The prediction confusion matrix diagram of 200 samples is as shown in Figure 7 .

Claims

1. A bearing fault diagnosis method based on time-frequency analysis combined with feature migration, characterized in that: The method comprises the following 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 sample data in the training set and the test set to obtain four-layer wavelet packet decomposition time-frequency coefficients of each sample data; S3: performing two-dimensional reconstruction on the time-frequency coefficients of each sample data as follows: dividing each sample data into N segments, and the length of each segment is N, taking the N segments of data as N rows of a two-dimensional reconstruction matrix, and constructing an N-row-by-N-column matrix, and the calculation formula of N is as follows: Wherein, Len represents the length of each sample data, and [] represents rounding down; S4: constructing a four-layer time-frequency graph according to a Z-shaped splicing method according to the two-dimensional reconstructed time-frequency coefficients of each sample data, and performing normalization; the construction process is as follows: the four-layer wavelet packet decomposition time-frequency coefficients are two-dimensionally reconstructed, and are sequentially defined as N1-N16 from low frequency to high frequency, and a four-layer time-frequency graph is constructed according to a Z-shaped splicing method as follows: [N1, N2, N6, N7; N3, N5, N8, N13; N4, N9, N12, N14; N10, N11, N15, N16]; S5: constructing an MSCNN network model; S6: inputting the four-layer time-frequency graph of each sample into the MSCNN network to extract corresponding combined deep features; S7: calculating the total cross-entropy loss of all samples in the training set; S8: determining whether the total cross-entropy loss meets the requirements, if yes, proceeding to the next step, otherwise, optimizing the MSCNN network according to the total cross-entropy loss; S9: Calculate the MK-MMD distance loss amount L of all samples in the training set and all samples in the test set in the high-dimensional feature space MK-MMD ; S10: calculating the total loss L, if the total loss meets the requirements, proceeding to S12, otherwise, proceeding to the next step; S11: determining whether the iteration number is equal to the set iteration number, if yes, proceeding to S12, otherwise, optimizing the MSCNN network according to the total loss L and returning to S6; S12: using the trained MSCNN network to input the test set sample data to perform fault type judgment.

2. The method of claim 1, wherein: In S1, the sample data ratio of the training set to the test set is 1:1; each sample data of the training set comprises a fault label, and the sample data of the test set does not comprise a data label.

3. The method of claim 2, wherein: In S5, the MSCNN network model comprises: a convolution preprocessing layer comprising a convolution layer, a ReLu layer and a normalization layer, three parallel convolution processing layers comprising six convolution layers, three ReLu layers, six normalization layers and a global average pooling layer, the processing cores of the three parallel convolution processing layers are 3*3, 5*5 and 7*7 respectively, a convolution post-processing layer comprising a full connection layer and a Softmax layer.

4. The method of claim 3, wherein: In S7, the total cross-entropy loss calculation formula is as follows: where N denotes the total number of samples, M denotes the number of classification categories, F Sic denotes the indicator function, which is 1 if the classification is successful and 0 if the classification fails, p ic denotes the predicted probability that sample i belongs to category c.

5. The method of claim 4, wherein: In S8, the full connection layer is optimized according to the total cross-entropy loss, and the optimization parameters comprise the weight and bias of the full connection layer.

6. The method of claim 5, wherein: In S9, the MK-MMD distance loss amount L MK-MMD The calculation formula is as follows: where P denotes the training set and Q the test set, denotes the reproducing Hilbert space mapping, E P denotes the mathematical expectation of the training set, E Q denotes the mathematical expectation of the training set, H k denotes the kernel k used by the reproducing Hilbert space.

7. The method of claim 6, wherein: In S11, the MSCNN network multi-core convolution layer weight and bias are optimized according to the total loss L.

Images