Bearing fault detection method based on EMD-CNN-Transformer model

Abstract

The invention relates to a bearing fault detection method based on an EMD-CNN-Transfromer model. The model is trained through the following steps that S1, vibration acceleration data, collected by a three-axis sensor, of a bearing in three directions of different working states (x, y and z) are obtained; s2, performing resampling and normalization on the data in each direction, and then performing three-dimensional data splicing; s3, EMD decomposition and splicing are carried out on the data in the step S2, then labels are added, and a data set is made; during EMD decomposition, ineffective components are removed; s4, loading a data set, and performing feature extraction through CNN; and S5, inputting the data processed in the step S4 into a Transform network for encoding and decoding, and finally classifying and outputting the data. Then, the fault type of the bearing is determined based on the trained model. According to the method, the EMD and the CNN are firstly utilized to perform two times of feature extraction, interference of invalid components is removed during EMD decomposition, and finally, the Transform is utilized to perform classification output, so that the final classification result has higher accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault detection, and in particular to a bearing fault detection method based on an AI model of EMD-CNN-Transformer. Technical Background

[0002] In a bearing, the "Inner Race", "Ball" and "Outer Race" are three important components, which together constitute the basic structure of the bearing.

[0003] Inner Race: The inner ring is an annular component of the bearing, usually located inside the ball or roller, and matched with the shaft. The inner ring rotates on the shaft by contacting the rolling element, bearing and transmitting axial and radial loads.

[0004] Ball: Balls are rolling elements in bearings. They are usually small spherical objects that can roll between the inner and outer rings. Balls play a role in supporting and transferring loads in bearings, reducing friction and sliding.

[0005] Outer Race: The outer race is another annular component of the bearing, located outside the balls or rollers. The outer race cooperates with the inner race and balls to provide external support and fixation of the bearing.

[0006] These three parts work together to form the basic structure of the bearing to support and transfer the load on the shaft, reduce friction, and enable the bearing to operate smoothly. Their materials and designs vary depending on the application and bearing type to meet specific load and working environment requirements.

[0007] Bearing failures include:

[0008] 1) Wear: The relative movement between the raceway and the ball body will cause mechanical wear. The intrusion of dust, debris, metal powder, and hard particles will aggravate the wear. Poor lubrication and improper speed will also cause wear on the contact surface. When the wear reaches a certain extent, the clearance between the bearing components will increase, resulting in a decrease in the operating accuracy of the machine.

[0009] 2) Fatigue spalling: It is mainly caused by the alternating load force on the inner and outer raceways of the rolling bearing. The raceway surface elastically deforms and forms fine cracks at the point of maximum stress. With continuous load operation, the cracks develop to the contact surface and form spalling pits, resulting in large-scale spalling. Fatigue spalling will accelerate the impact load, vibration and noise intensity of the bearing during rolling, seriously affecting the service life of the bearing.

[0010] 3) Plastic deformation: Due to excessive impact load, thermal deformation or hard object intrusion, the bearing is subjected to excessive extrusion stress, forming cracks or indentations on the raceway surface, which further aggravates the impact load and causes the inner and outer ring surfaces of the bearing to peel off.

[0011] 4) Corrosion: including rust and electrical corrosion. Factors such as poor lubrication and environmental humidity may cause corrosion on the bearing surface.

[0012] In addition, bearing damage may also be caused by the following reasons:

[0013] 5) Overload: Especially under high-speed operation or frequent repeated loads, it is easy to cause bearing damage.

[0014] 6) Solid particle intrusion: Contaminants entering the bearing will aggravate wear.

[0015] 7) Use in situations with strong vibration: Vibration will accelerate bearing wear and damage.

[0016] 8) Improper use of lubricating oil: Too much or too little grease may affect the performance of the bearing.

[0017] 9) Improper installation: leads to uneven force, further affecting the normal operation of the bearing.

[0018] The above are the common types and causes of bearing failures, which are crucial for the maintenance and repair of bearings. The main significance of bearing fault detection is to detect the corresponding degree of failure and give early warning to avoid the degree of damage affecting the qualified rate of the product. Therefore, bearing fault detection aims to improve the qualified rate of products and reduce damage to equipment, which is of great significance to the operation and maintenance of equipment.

[0019] At present, there are many methods for detecting bearing faults, including statistical methods, such as root mean square (RMS), kurtosis, skewness, principal component analysis (PCA), adaptive filtering algorithm; methods based on fault diagnosis expert system; methods based on time-frequency domain fusion, such as instantaneous frequency analysis, instantaneous amplitude analysis, etc.; machine learning methods based on models, such as support vector machine (SVM), decision tree, random forest, naive Bayes, etc.; methods based on deep learning, such as convolutional neural network (CNN), long short-term memory network (LSTM), attention mechanism, etc.; vibration analysis methods, such as Fourier transform and its variants, short-time Fourier transform, fast Fourier transform, etc. Fast Fourier transform, wavelet transform and its variants; methods based on spectrum analysis, energy spectrum analysis, order analysis, etc.; methods based on time domain feature extraction, such as peak index, square root amplitude, pulse counting, etc.; and methods based on the fusion of the above, such as short-time Fourier transform STFT, which decomposes the signal into a superposition of a series of continuous sine waves of different frequencies. However, since the Fourier transform transforms the entire signal from the time domain to the frequency domain to obtain an overall spectrum, it loses the time information and cannot determine when a specific signal occurs based on the results of the Fourier transform. Therefore, the Fourier transform lacks time-frequency analysis capabilities and multi-resolution analysis capabilities, and it is difficult to analyze non-stationary signals.

[0020] In order to make a more objective comparison of several existing methods with high accuracy, ten classifications are adopted, including one normal sample and nine fault samples, and the bearing data set of Western Reserve University is uniformly adopted. The prediction accuracy is detailed in Table 1. The present invention aims to provide a new bearing fault detection method and further improve the fault diagnosis accuracy.

[0021] Table 1

[0022] Summary of the invention

[0023] The purpose of the present invention is to provide a new bearing fault detection method and further improve the fault diagnosis accuracy.

[0024] The technical solution for realizing the invention purpose of the present invention is as follows: a bearing fault detection method based on EMD-CNN-Transfromer model (i.e., empirical mode decomposition deep convolution transformer multi-attention model), characterized in that the model is trained in the following manner:

[0025] S1. Obtain the vibration acceleration data of the bearing in three directions (x, y, z) under different working conditions collected by the triaxial sensor;

[0026] S2, resampling and normalizing the data in each direction, and then performing three-dimensional data splicing;

[0027] S3, perform EMD decomposition on the data in step S2 and concatenate them, then add labels and make a data set; pay attention to removing invalid components during EMD decomposition;

[0028] S4, load the data set and perform feature extraction through CNN;

[0029] S5: Input the data processed by S4 into the Transformer network for encoding and decoding, and finally output the classified data;

[0030] Then, the fault type of the bearing is determined based on the trained model.

[0031] In step S3, the effective component and the invalid component are determined by comparing the correlation coefficients of the IMF components decomposed by the EMD with the original signal.

[0032] The steps of EMD decomposition in step S3 are as follows:

[0033] a. The maximum and minimum points of the original signal x(t), these extreme points are fitted by the curve interpolation method to obtain the upper envelope of the signal x max (t) and the lower envelope x min (t);

[0034] b. Calculate the average value of the uplink and downlink routes:

[0035]

[0036] c. Subtract the original signal x(t) from the average envelope g(t) to obtain the remaining signal d1(t);

[0037] d. Replace the original signal x(t) with the remaining signal d1(t) and perform the processing from step a to step c until SD is less than the screening threshold value, and the threshold value range is 0.2-0.3, so as to obtain the final suitable first-order modal component c1(t), that is, the first IMF;

[0038] The SD is calculated as follows:

[0039]

[0040] Where: T represents the number of sample points, d k (t) is the residual signal of the kth iteration;

[0041] e. Calculate the difference between the original signal x(t) and c1(t) to obtain the first-order residual r1(t). Use r1(t) to replace the original signal x(t) and perform steps a to e. Repeat n times to obtain the nth-order modal function c n(t) and the final standard residual r n (t);

[0042] The expression of the original signal x(t) decomposed by EMD is:

[0043]

[0044] Then, the IMF components c decomposed from EMD are k (t) is compared with the original signal x(t) by correlation coefficient and the invalid components are removed.

[0045] The correlation coefficient calculation formula is as follows:

[0046]

[0047] When the IMF component c k (t) is the effective amount.

[0048] Preferably, the overlap rate of resampling in step S2 is 50%.

[0049] Step S4 extracts the signal features by performing convolution and pooling operations twice using one-dimensional CNN.

[0050] Step S5 first encodes through TransformerEncoderLayer(), then uses LayerNorm() for normalization, then performs average pooling AdaptiveAvgPool1d(), and finally makes a fully connected classification FC output.

[0051] The encoding layer of TransformerEncoderLayer() has two layers. The first layer is multi-head attention. After passing through the 8-head attention module, the residual network connection is performed, and the result is input to the next forward propagation layer, and then the residual network connection method is used to output the encoding layer result of this layer.

[0052] Step S5 includes 4 coding layers.

[0053] Beneficial effects:

[0054] The present invention first uses EMD and CNN to perform feature extraction twice, and removes invalid components during EMD decomposition, thereby avoiding interference caused by invalid components, and then uses Transformer for classification output, so that the final classification result has higher accuracy than the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1A data processing and training flow chart of a bearing fault detection method based on an EMD-CNN-Transfromer model according to a preferred embodiment of the present invention;

[0056] Figure 2 This is the EMD decomposition diagram;

[0057] Figure 3 The results of 10 classifications after sampling the three-axis sensor are shown in the figure. DETAILED DESCRIPTION

[0058] The scheme of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0059] The present invention provides a bearing fault detection method based on the EMD-CNN-Transformer model, wherein the model is trained in the following manner: Figure 1 As shown:

[0060] S1. Obtain the vibration acceleration data of the bearing in three directions (x, y, z) under different working conditions collected by the triaxial sensor

[0061] This embodiment uses a three-axis vibration temperature intelligent sensor VTall-T163E-A, with a sampling frequency of 26.7k and a sampling number of 26700*300 in 5 minutes, which can capture sufficient instantaneous acceleration data for distinguishing fault characteristics. This time, 9 types of faults and 1 type of normal operation data are collected, and the collected data types are shown in Table 2.

[0062] Table 2 Data categories

[0063]

[0064] S2: Resample and normalize the data in each direction, and then perform three-dimensional data splicing

[0065] The data collected in the three directions are resampled respectively, and for each direction, the overlap rate of the resampling is set to 50%, such as the first sample series starts from 1 to 1024, and the second sample series starts from 513 to 1536. The advantages of resampling and setting a higher overlap rate in this embodiment are as follows: first, repeated correlation can be removed; second, data enhancement can be achieved, and more sample series can be generated.

[0066] Then the data is normalized and 3D data splicing is performed. 3D data splicing is to splice the data of the three directions at the same time point in the sample series together to form a 3D sample. For the above sample series, the data length of a 3D sample is 1024*3 (1024 data for each acceleration in the x, y, and z directions). 1000 3D samples are collected for each category in Table 2, such as normal, B007, etc.

[0067] S3: Perform EMD decomposition and concatenation on the data in step S2, then add labels and create a data set

[0068] The EMD decomposition principle is as follows:

[0069] The basic idea of ​​empirical mode decomposition is to transform a wave with irregular frequency into multiple waves of single frequency + residual waves. The original waveform = ∑IMFs + residual waves.

[0070] The decomposition steps are as follows:

[0071] a. The maximum and minimum points of the original signal x(t), these extreme points are fitted by the curve interpolation method to obtain the upper envelope of the signal x max (t) and the lower envelope x min (t).

[0072] b. Calculate the average value of the uplink and downlink routes:

[0073]

[0074] c. Subtract the original signal x(t) from the average envelope g(t) to obtain the remaining signal d1(t).

[0075] Generally speaking, for a stationary signal, it is the first modal function (IMF) of the original signal x(t). However, for a non-stationary signal, the signal is not monotonically increasing in a certain area, but has inflection points. If these inflection points that reflect the specific characteristics of the original signal x(t) are not selected, the first-order modal function obtained is not accurate, that is, the d1(t) usually obtained does not meet the two conditions of the IMF, so here we choose to continue screening.

[0076] Two conditions for the modal function IMF:

[0077] 1) The number of local extreme points and zero-crossing points of the function must be equal or differ by at most one in the entire time range.

[0078] 2) At any point in time, the envelope of the local maximum (upper envelope) and the envelope of the local minimum (lower envelope) must average zero.

[0079] d. Use the remaining signal d1(t) to replace the original signal x(t) and perform the processing from step a to step c until SD is less than the screening threshold (usually 0.2-0.3), thus obtaining the final suitable first-order modal component c1(t), i.e. the first IMF. The SD is calculated as follows:

[0080]

[0081] Where: T represents the number of sample points, d k (t) is the residual signal of the kth iteration.

[0082] e. Calculate the difference between the original signal x(t) and c1(t) to obtain the first-order residual r1(t). Use r1(t) to replace the original signal x(t) and perform steps a to e. Repeat n times to obtain the nth-order modal function c n (t) and the final standard residual r n (t).

[0083] The expression of the original signal x(t) after EMD decomposition is:

[0084]

[0085] Because the components decomposed by EMD are uncertain, and the decomposed modal signals are from high-frequency signals to low-frequency signals, there are invalid components in the components, so the present invention first removes the invalid components by the correlation coefficient method. After the invalid components are removed, the interference caused by them can be avoided, and the subsequent calculation speed can also be improved.

[0086] The correlation coefficient method is to use each IMF component (i.e., c k (t)) and the original signal x(t) sequence to find the correlation coefficient, the expression is:

[0087]

[0088] By comparing the correlation coefficient with the original signal, it is found that IMF1-IMF7 are effective components Therefore, each sample is decomposed into 7 components, such as Figure 2 As shown, a data sample has 1024*3*7 data values, where 1024 represents the 1024 accelerations collected, 3 represents the data in the x, y, and z directions, and 7 represents the first 7 valid components selected after EMD decomposition.

[0089] Then, the 7 effective components of EMD decomposition are spliced ​​to form a new sample, that is, the 7 components are connected together in the order of the first and last sequences. At this time, the sample data volume is 1024*3*7=21504. Data labels are added and divided into training set, validation set and test set in the ratio of 70%, 20% and 10%.

[0090] Label: In machine learning, a label refers to a target value or expected output associated with a sample. It represents the result that the model is expected to learn and predict from the input data. Labels play a very important role in supervised learning because they provide the model with learning goals and standards. For example, in this embodiment, each sample (1024*3*7) corresponds to a label, and different samples can correspond to the same or different categories of labels in Table 2.

[0091] Dataset: refers to an important data set used to "feed" the model during the model training process so that it can learn and understand data characteristics, distribution laws and patterns. These data sets have a crucial impact on the performance, accuracy and generalization ability of the model. The data set can be divided into a training set (Training Set): the data used to train the model, and the parameters of the fitting curve are determined by the data in the training set; a validation set (Validation Set): an optional set used for model selection, final optimization and determination of the model, to assist in the construction of the model, and to train hyperparameters; a test set (Test Set): to test the accuracy of the trained model.

[0092] S4. Load the data set and extract features through CNN

[0093] In the data set, a single sample of 21504 data is taken as a sample, and the parameters are set as input = 336 (because 21504 = 64*336, which is the number of channels 336), output = 1024, kernel_size = 3, as data input to the first one-dimensional CNN for convolution pooling operation to extract signal features. The parameters of the second one-dimensional convolution are set as: input = 1024, output = 1024, kernel_size = 3, so the output result of the second convolution is also 1024.

[0094] S5: Input the data processed by S4 into the Transformer network for encoding and decoding, and finally output the classification

[0095] Transformer network, the parameters are set as follows:

[0096] input=1024

[0097] num_heads=8 #8 attention heads

[0098] hidden_dim = 1024 #Number of features in the hidden layer

[0099] dropout=0.5

[0100] batch_first=True

[0101] num_layers = 4 # 4 coding layers

[0102] batch_size=32 #batch size is 32 samples

[0103] learn_rate=0.0001#Learning rate is 0.0001

[0104] epoch=50

[0105] After encoding by TransformerEncoderLayer(), normalization LayerNorm() is used, and then average pooling AdaptiveAvgPool1d() is performed, and finally a fully connected classification FC is performed to output 10 categories.

[0106] For the specific structure of TransformerEncoderLayer(), see Figure 1 There are two encoding layers. The first layer is multi-head attention. After the 8-head attention module, the residual network connection is performed to input the result to the next forward propagation layer, and then the residual network connection is used to output the first encoding layer result. After 4 encoding layers, 10 classification results are output through normalization and full connection.

[0107] The training results are as follows Figure 3 As shown, this embodiment classifies 10 data categories, and the accuracy rate reaches 99.76% after 50 epochs (all sample data are input into the model training once as 1 epoch).

[0108] After the model training is completed, the fault type of the bearing is subsequently determined based on the trained model.

[0109] Ten classifications are adopted, including one normal sample and nine fault samples, and the bearing data set of Western Reserve University is uniformly adopted. The comparison of the detection accuracy of the method of the present invention and various vibration detection methods is shown in Table 3. As can be seen from the table, the method of the present invention, using the same data set, has an accuracy of 99.01%, which is better than other algorithms. The accuracy of the method of the present invention for data collected by three-dimensional channels can reach 99.76%.

[0110] Table 3

[0111]

Claims

1. A bearing fault detection method based on EMD-CNN-Transfromer model, characterized in that: The model is trained in the following way: S1. Obtain the vibration acceleration data of the bearing in three directions (x, y, z) under different working conditions collected by the triaxial sensor; S2, resampling and normalizing the data in each direction, and then performing three-dimensional data splicing; S3, perform EMD decomposition on the data in step S2 and concatenate them, then add labels and make a data set; pay attention to removing invalid components during EMD decomposition; S4, load the data set and perform feature extraction through CNN; S5: Input the data processed by S4 into the Transformer network for encoding and decoding, and finally output the classified data; Then, the fault type of the bearing is determined based on the trained model.

2. The bearing fault detection method according to claim 1, characterized in that: In step S3, the effective component and the invalid component are determined by comparing the correlation coefficients of the IMF components decomposed by the EMD with the original signal.

3. The bearing fault detection method according to claim 2, characterized in that: The steps of EMD decomposition in step S3 are as follows: a. The maximum and minimum points of the original signal x(t) are fitted by the curve interpolation method to obtain the upper envelope of the signal x max (t) and the lower envelope x min (t); b. Calculate the average value of the uplink and downlink routes: c. Subtract the original signal x(t) from the average envelope g(t) to obtain the remaining signal d1(t); d. Replace the original signal x(t) with the remaining signal d1(t) and perform the processing from step a to step c until SD is less than the screening threshold value, and the threshold value range is 0.2-0.3, so as to obtain the final suitable first-order modal component c1(t), that is, the first IMF; The SD is calculated as follows: Where: T represents the number of sample points, d k (t) is the residual signal of the kth iteration; e. Calculate the difference between the original signal x(t) and c1(t) to obtain the first-order residual r1(t). Use r1(t) to replace the original signal x(t) and perform steps a to e. Repeat n times to obtain the nth-order modal function c n (t) and the final standard residual r n (t); The expression of the original signal x(t) decomposed by EMD is: Then, the IMF components c decomposed by EMD are k (t) is compared with the original signal x(t) by correlation coefficient and the invalid components are removed.

4. The bearing fault detection method according to claim 3, characterized in that: The correlation coefficient calculation formula is as follows: ρ xck >0, determine the IMF component c k (t) is the effective amount.

5. The bearing fault detection method according to claim 1, characterized in that: The overlap rate of the resampling in step S2 is 50%.

6. The bearing fault detection method according to claim 1, characterized in that: Step S4 extracts the signal features by performing convolution and pooling operations twice using one-dimensional CNN.

7. The bearing fault detection method according to claim 1, characterized in that: Step S5 first encodes through TransformerEncoderLayer(), then normalizes using LayerNorm(), then performs average pooling AdaptiveAvgPool1d(), and finally makes a fully connected classification FC output.

8. The bearing fault detection method according to claim 7, characterized in that: The encoding layer of TransformerEncoderLayer() has two layers. The first layer is multi-head attention. After passing through the 8-head attention module, the residual network connection is performed, and the result is input to the next forward propagation layer, and then the residual network connection method is used to output the encoding layer result of this layer.

9. The bearing fault detection method according to claim 8, characterized in that: Step S5 includes 4 coding layers.

Images