A method and system for diagnosing faults of a rolling bearing

Abstract

The present application relates to the technical field of rolling bearing fault diagnosis, in particular to a rolling bearing fault diagnosis method and system, and the specific method is as follows: collecting the vibration signals of each rolling bearing of the same device as original data, marking a one-hot label for all original data, dividing the original data with one-hot labels into a training set, a validation set and a test set according to a certain proportion; performing discrete Fourier transform on the original data with faults in the training set, randomly fusing the amplitude and phase information to obtain composite data, and marking a fusion label by using a proportional fusion method; reassigning the label according to a neighboring smoothing strategy; taking the training set with the re-assigned fusion label and neighboring smoothing label as a new training set, training a rolling bearing fault diagnosis model using the new training set, and performing rolling bearing fault diagnosis. The method can expand the data set and train the model according to the fault characteristics of the rolling bearing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rolling bearing fault diagnosis, in particular to a rolling bearing fault diagnosis method and system. BACKGROUND

[0002] The key position of rolling bearings in rotating machinery directly determines the health status, working performance and safety of mechanical equipment. When the bearing fails, it often causes abnormal vibration, produces a violent impact on the mechanical structure, and further affects other components of the mechanical system, causing a great threat to the safety of mechanical equipment and production efficiency.

[0003] In intelligent bearing fault diagnosis, the collected fault data usually includes vibration signals, temperature signals and acceleration signals, etc. When the bearing has defects and is in working condition, the interaction between components will produce impact. Therefore, when diagnosing bearing faults, the vibration signal of the rolling bearing in operation is usually used. By extracting and classifying the collected rolling bearing vibration signal, the diagnosis and classification of faults are realized.

[0004] However, mechanical equipment often operates in complex working conditions such as high environmental noise and multiple operating conditions. The current intelligent diagnosis technology of mechanical equipment either ignores the existence of complex faults or classifies and identifies faults separately. The bearing fault of rotating machinery is usually complex and interrelated. The occurrence of faults is not completely independent, and there may be different degrees of connection between fault categories. However, in supervised learning, using one-hot encoding as a labeling method is too absolute, which may make the model sensitive to noise and uncertainty during training, resulting in overconfidence and overfitting. SUMMARY

[0005] In view of the above problems, the present application is made, and the purpose of the present application is to provide a rolling bearing fault diagnosis method and system for performing fault diagnosis on rolling bearings.

[0006] The technical scheme of the present application is as follows:

[0007] A rolling bearing fault diagnosis method, specifically comprising:

[0008] S1, collect the vibration signal of each rolling bearing of the same equipment as the original data, label all original data with one-hot label according to the fault condition of each rolling bearing, and divide the original data with one-hot label into a training set, a validation set and a test set according to a certain proportion;

[0009] S2, obtaining corresponding spectrum information by performing discrete Fourier transform on the original data of the faulty rolling bearings in the training set in a set time domain, obtaining amplitude and phase information in the spectrum information, obtaining composite data by randomly fusing the amplitude and phase information, and marking a fusion label of each composite data by using a proportional fusion method;

[0010] Specifically, the composite data in S2 includes first composite data and second composite data, the first composite data is obtained through spectrum information of the faulty rolling bearings, and the second composite data is obtained through a fault frequency of the faulty rolling bearings.

[0011] The specific method for obtaining the first composite data is as follows:

[0012] According to the peak value characteristics of the spectrum, the peak-to-peak value of each faulty rolling bearing is calculated, and all the obtained peak-to-peak values are sorted in descending order.

[0013] According to a preset percentage, all the peak-to-peak values are divided into a high fault segment, a medium fault segment and a low fault segment.

[0014] The amplitude and phase of the original data of the faulty rolling bearings in the high fault segment and the medium fault segment are randomly fused, and the amplitude and phase of the original data of the faulty rolling bearings in the medium fault segment and the low fault segment are randomly fused to generate the first composite data.

[0015] The specific method for obtaining the second composite data is as follows:

[0016] The rolling bearing fault is divided into inner ring fault, outer ring fault or rolling element fault, when one of the faults occurs in the rolling bearing, the vibration signal is taken as a single signal, and when two or more faults occur in the rolling bearing, the vibration signal is taken as a high-order harmonic signal.

[0017] According to the vibration signal of the faulty rolling bearing, the rolling bearing with the high-order harmonic signal is taken as a high-degree rolling bearing, and the rest is taken as a low-degree rolling bearing, the amplitude and phase of the high-degree rolling bearing and the low-degree rolling bearing are randomly fused to generate the second composite data.

[0018] Further, the amplitude and phase information are randomly fused to obtain composite data, and the specific operation is as follows:

[0019]

[0020] Wherein, x a , x b are two original data in the training set with faults, x c is composite data generated by random fusion, represents random fusion, represents element-by-element multiplication, represents x a the amplitude of x represents x b the phase of x, i is an imaginary number.

[0021] The number of composite data obtained by random fusion is denoted as M, the number of original data is denoted as N, and the number relationship between the two is denoted as M=N.

[0022] M=λN,

[0023] wherein λ is a parameter for adjusting the proportion of generated data.

[0024] The fusion label of each composite data is marked by using the proportional fusion method, and the specific operation is as follows.

[0025]

[0026] wherein, is the fusion label of the composite data; β is a parameter between 0 and 1, representing the percentage of the phase feature in the fusion label; x a , x b are two original data with faults in the training set, y a is the label of x a , y b is the label of x b , is the amplitude of y a , is the phase of y b .

[0027] The parameter β of the fusion label is set to 0.2.

[0028] S3, the original data with faults in the training set is re-labeled according to the adjacent smoothing strategy to obtain the adjacent smoothing label of each original data of the faulty rolling bearing.

[0029] Each one-hot label of the original data corresponds to the number of rolling bearings in the device, so that the adjacent smoothing strategy of S3 is specifically as follows.

[0030] The fault probability of the position of the fault rolling bearing in the one-hot label of the fault rolling bearing in the training set is taken as the actual class, and the fault probability of the positions of the rolling bearings on both sides of the fault rolling bearing is taken as the remaining classes, and the sum of the fault probabilities of the actual class and the remaining classes on both sides is 1.

[0031] S4, the original data, the composite data containing the fusion label obtained in S2, and the data labeled with the adjacent smoothing label obtained in S3 are taken as a new training set, and the rolling bearing fault diagnosis model is trained using the new training set to perform rolling bearing fault diagnosis.

[0032] The application also provides a rolling bearing fault diagnosis system, comprising:

[0033] A diversity module; vibration signals of each rolling bearing of the same device are collected as original data, and a one-hot label is marked for all original data according to the fault condition of each rolling bearing, and the original data with the one-hot label is divided into a training set, a validation set and a test set according to a certain proportion;

[0034] A fusion module; in a set time domain, discrete Fourier transform is performed on the original data with faults in the training set to obtain corresponding frequency spectrum information, amplitude and phase information in the frequency spectrum information are obtained, the amplitude and phase information are randomly fused to obtain composite data, and a fusion label of each composite data is marked by using a proportional fusion method;

[0035] A neighboring smoothing module; the one-hot label of the original data with faults in the training set is re-assigned according to a neighboring smoothing strategy to obtain a neighboring smoothing label of the original data of each rolling bearing with faults;

[0036] A machine learning module; the original data, the composite data containing the fusion label obtained in the fusion module and the data marked with the neighboring smoothing label obtained in the neighboring smoothing module are used as a new training set, a rolling bearing fault diagnosis model is trained using the new training set, and rolling bearing fault diagnosis is performed.

[0037] The application has the following beneficial effects:

[0038] The amplitude-phase composite data enhancement technology is adopted to expand the data set, solve the problems of shortage of complex fault samples and difficulty in determining the boundary between faults, ensure more comprehensive training samples to improve accuracy, the label proportional fusion method is used to determine the label of new data, more effectively present data characteristics, increase the robustness of the model to data changes, and improve the adaptability of the model to new data. The neighboring label smoothing technology is adopted to reduce the need for feature normalization, establish the relationship between classes in the training data, enhance the clustering effect of each class sample, in addition, the use of the neighboring label smoothing technology solves the problem of less information entropy in simple classification, makes the label setting more reasonable, and is beneficial to the training of the deep learning model. DETAILED DESCRIPTION

[0039] The technical scheme of the application is as follows:

[0040] A rolling bearing fault diagnosis method, specifically comprising:

[0041] S1, collect the vibration signal of each rolling bearing of the same device as the original data, and mark a one-hot label for all original data according to the fault condition of each rolling bearing, and divide the original data with the one-hot label into a training set, a validation set and a test set according to a certain proportion.

[0042] Taking the mechanical arm fault data as an example, B1, B2, B3, B4, B5 and B6 are rolling bearings in the mechanical arm connected in turn, and B2 and B5 are assumed to be faulty rolling bearings, and B1, B3, B4 and B6 are non-faulty rolling bearings.

[0043] The one-hot label position is (B1, B2, B3, B4, B5, B6), and the one-hot label corresponding to each rolling bearing is B1: (0, 0, 0, 0, 0, 0); B2: (0, 1, 0, 0, 0, 0); B3: (0, 0, 0, 0, 0, 0); B4: (0, 0, 0, 0, 0, 0); B5: (0, 0, 0, 0, 1, 0); B6: (0, 0, 0, 0, 0, 0).

[0044] S2, in a set time domain, discrete Fourier transform is performed on the original data with faults in the training set to obtain corresponding frequency spectrum information, amplitude and phase information in the frequency spectrum information are obtained, the amplitude and phase information are randomly fused to obtain composite data, and a fusion label of each composite data is marked by using a proportional fusion method.

[0045] Specifically, the composite data in S2 includes first composite data and second composite data, the first composite data is obtained through the frequency spectrum information of the faulty rolling bearing, and the second composite data is obtained through the fault frequency of the faulty rolling bearing.

[0046] The specific method for obtaining the first composite data is as follows:

[0047] For the frequency spectrum information of the faulty rolling bearing, the peak-to-peak value of each faulty rolling bearing is calculated according to the peak value characteristics of the frequency spectrum, and all the obtained peak-to-peak values are sorted from large to small.

[0048] According to a predetermined percentage, all the peak-to-peak values are divided into a high fault segment, a medium fault segment and a low fault segment; for example, the first 20% of the data segments are selected as high faults, the middle 40% are selected as medium faults, and the last 40% are selected as low faults.

[0049] The amplitude and phase of the original data of the faulty rolling bearing in the high fault segment and the medium fault segment are randomly fused, and the amplitude and phase of the original data of the faulty rolling bearing in the medium fault segment and the low fault segment are randomly fused to generate the first composite data.

[0050] The specific method for obtaining the second composite data is as follows:

[0051] The rolling bearing fault is divided into inner ring fault, outer ring fault or rolling element fault. When one of the faults occurs in the rolling bearing, the vibration signal is taken as a single signal. When two or more faults occur in the rolling bearing, the vibration signal is taken as a high-order harmonic signal. The frequency is taken as a characterization index to judge whether the rolling bearing has a fault and the fault degree.

[0052] When the inner ring, the outer ring or the rolling element of the rolling bearing has an early fault, a single signal appears in the frequency spectrum of the vibration signal of the rolling bearing. The early fault is one of the inner ring fault, the outer ring fault and the rolling element fault, and the degree is low. The performance is that the fault frequency is single, and the vibration signal of the rolling bearing appears a single signal. With the further deterioration of the fault, the fault of the rolling bearing can be two or more of the inner ring fault, the outer ring fault and the rolling element fault, and the fault degree is high and can also affect other rolling bearings nearby. The high-order harmonic signal of the fault frequency appears in the frequency spectrum of the vibration signal of the rolling bearing.

[0053] Considering that the frequency corresponding to different rotating speeds is also different, the following formula for calculating the critical frequency of the rolling bearing is given:

[0054]

[0055] Wherein, the rotating speed of the rolling bearing is r, the unit is revolution / minute, the number of rolling bearings is n, the diameter of the rolling bearing is d, the pitch diameter of the rolling bearing is D, and the contact angle of the rolling bearing is a.

[0056] The time domain signal of the training set is subjected to Fourier transform to obtain the frequency domain signal, i.e. the frequency spectrum. The inner ring critical frequency, the outer ring critical frequency and the rolling element critical frequency of each rolling bearing are calculated. Whether the frequency spectrum contains the values of the inner ring critical frequency, the outer ring critical frequency and the rolling element critical frequency is observed. If the frequency spectrum of the rolling bearing contains one of the inner ring critical frequency, the outer ring critical frequency and the rolling element critical frequency, it is a single signal. If the frequency spectrum of the rolling bearing contains two or more of the inner ring critical frequency, the outer ring critical frequency and the rolling element critical frequency, it is a high-order harmonic signal.

[0057] According to the vibration signal of the rolling bearing with a fault, the rolling bearing with a high-order harmonic signal is taken as a high-degree rolling bearing, and the rest is taken as a low-degree rolling bearing. The amplitudes and phases of the high-degree rolling bearings and the low-degree rolling bearings are randomly fused to generate second composite data.

[0058] Further, the amplitude and phase information are randomly fused to obtain composite data. The specific operation is as follows:

[0059]

[0060] Wherein, x a , xb are two original data with faults in the training set, x c are composite data generated by random fusion, represents random fusion, represents element-wise multiplication, represents the amplitude of x a , represents the phase of x b , i is an imaginary number.

[0061] In generating the first composite data, x a and x b come from high fault segments and medium fault segments, or from medium fault segments and low fault segments.

[0062] In generating the second composite data, x a and x b come from high degree rolling bearings and low degree rolling bearings.

[0063] The composite data obtained by random fusion contains different types of fault information, which expands the data set and ensures more comprehensive training samples to improve accuracy.

[0064] The number of composite data obtained by random fusion is denoted as M, the number of original data is denoted as N, and their number relationship is denoted as

[0065] M = λN,

[0066] where λ is a parameter for adjusting the proportion of generated data.

[0067] M1 first composite data and M2 second composite data are generated, and M1 + M2 = M, then:

[0068] M1 = μ1N,

[0069] M2 = μ2N,

[0070] where μ1 and μ2 are the ratios of the first composite data and the second composite data to the composite data, respectively, and usually μ2 > μ1.

[0071] Since there are fewer severe fault data in actual working conditions, according to the actual working conditions, M1 new data need to be generated.

[0072] Finally, the first composite data and the second composite data are combined with the original training set to achieve the purpose of expanding the training set. This method makes the newly generated data more consistent with the actual working conditions, that is, the "false" fault signals generated are more close to the real fault signals, not only can expand the complex data samples, but also make the learning more sufficient and comprehensive in subsequent deep learning, so that the fault recognition is more accurate.

[0073] The generated data contains feature information of two different categories of data. In order to improve the adaptability of the model to the composite data, a label proportion fusion method is used to determine the final label, which can more effectively present the data features. The proportion fusion method is used to mark the fusion label of each composite data, and the specific operation is as follows:

[0074]

[0075] wherein, is the fusion label of the composite data; β is a parameter between 0 and 1, representing the percentage of phase features in the fusion label; x a , x b are two original data with faults in the training set, y a is the label of x a , y b is the label of x b , is the amplitude of y a , is the phase of y b .

[0076] Here, the selection of x a , x b is the same as that of x a , x b .

[0077] Through experimental test, the parameter β of the fusion label is preferably set to 0.2.

[0078] Taking β value as 0.2 as an example, the labels y a , y b of x a , x b are (1, 0, 0, 0, 0, 0) and (0, 0, 1, 0, 0, 0) respectively, and the fusion label of the composite data x c is:

[0079] (1-β)×(1, 0, 0, 0, 0, 0)+β(0, 0, 1, 0, 0, 0)(0.8, 0, 0.2, 0, 0, 0)

[0080] The label proportion fusion provides a smoother uncertainty estimation, thereby increasing the classification boundary of the original data, so that the fault classification is no longer too absolute, and the robustness of the model to data changes is increased.

[0081] S3, the original data with faults in the training set is marked with a one-hot label, and the label is re-allocated according to the adjacent smoothing strategy to obtain the adjacent smoothing label of each original data of the rolling bearing with faults;

[0082] Each one-hot label capacity of the original data corresponds to the number of rolling bearings in the device, so the adjacent smoothing strategy of S3 is specific to

[0083] The failure probability of the position of the fault rolling bearing in the one-hot label of the fault rolling bearing in the training set is taken as the actual class, and the failure probability of the position of the rolling bearing on both sides of the fault rolling bearing is taken as the remaining class. The sum of the actual class and the failure probability of the remaining class on both sides is 1.

[0084] The specific calculation method of the adjacent smoothing label is

[0085]

[0086] where k is the number of rolling bearings, is the adjacent smoothing label of the kth rolling bearing, and a is an adjustable parameter. Let y k is the one-hot label of the kth rolling bearing, y k+1 is the one-hot label of the k+1th rolling bearing, y k-1 is the one-hot label of the k-1th rolling bearing, and N is the number of original data.

[0087] y k where if the position of the kth fault rolling bearing is 1, the rest are 0.

[0088] In supervised learning, one-hot encoding is often used to label data, i.e. one-hot encoding will convert the label into a vector, and only the actual class position in the vector is 1, and the rest is 0. However, using one-hot encoding as a labeling method is too absolute, which may make the model sensitive to noise and uncertainty during training, resulting in overconfidence and overfitting. Therefore, a smoothing label strategy is proposed. When one-hot labeling the original data, first sort the different classes according to the characteristics of the fault itself, and through simple human processing, place the most likely fault type in the adjacent position of the label vector of each fault class according to the size of the failure probability.

[0089] Taking the mechanical arm fault data as an example, B1, B2, B3, B4, B5 and B6 are rolling bearings in the mechanical arm connected in turn. The mechanical arm contains multiple rolling bearings connected front and back, so when one of the arm rolling bearings fails, it will affect its adjacent rolling bearings.

[0090] The strength of the fault gradually decreases from the center of the fault to the far end, and the probability of failure also gradually decreases from near to far.

[0091] For example, when B3 fails, B2 and B4 are most likely to be affected and most likely to fail, so the label position is (B1, B2, B3, B4, B5, B6), the corresponding label at the beginning is (0, 0, 1, 0, 0, 0), and the adjacent smoothing label is (0, 0.05, 0.9, 0.05, 0, 0).

[0092] In summary, the data processing module is constructed, including the amplitude-phase composite data enhancement technology and the adjacent label smoothing technology. Now further description is made for the use of the data processing process. The original data set is divided into a training set, a validation set and a test set according to a certain proportion, and the labels of all data are in one-hot form, and the proposed module is only used on the training set. The original signal of the training set is converted to the frequency domain by discrete Fourier transform (DFT), the amplitude and phase of the signal are obtained by spectrum analysis, and new data is generated by random fusion, and then the label is given. At this time, the original data label of the training set is still in one-hot form, and thus the random fusion technology is completed. The original data of the training set is subjected to the adjacent label smoothing strategy, and most of the probability of the label is distributed to the actual class, and the remaining probability is uniformly distributed to the adjacent class. The label is no longer in one-hot form, and thus the data processing is completed.

[0093] S4, the original data, the composite data containing the fusion label obtained in S2, and the data labeled with the adjacent smoothing label obtained in S3 are used as a new training set, and a rolling bearing fault diagnosis model is trained using the new training set to perform rolling bearing fault diagnosis.

[0094] The application also provides a rolling bearing fault diagnosis system, which comprises:

[0095] A diversity module; vibration signals of each rolling bearing of the same device are collected as original data, one-hot labels are marked for all original data according to the fault conditions of each rolling bearing, and the original data with one-hot labels are divided into a training set, a validation set and a test set according to a certain proportion;

[0096] A fusion module; discrete Fourier transform is performed on the original data with faults in the training set in a set time domain to obtain corresponding spectrum information, amplitude and phase information in the spectrum information are obtained, the amplitude and phase information are randomly fused to obtain composite data, and a fusion label of each composite data is marked by using a proportional fusion method;

[0097] An adjacent smoothing module; one-hot labels of the original data with faults in the training set are redistributed according to an adjacent smoothing strategy to obtain adjacent smoothing labels of the original data of each rolling bearing with faults;

[0098] The machine learning module uses the original data, the composite data containing the fusion label obtained in the fusion module, and the data labeled with the adjacent smoothing label obtained in the adjacent smoothing module as a new training set, trains the rolling bearing fault diagnosis model using the new training set, and performs rolling bearing fault diagnosis.

[0099] The rolling bearing fault diagnosis model is trained using a training set containing only original data (referred to as scheme A) and a training set containing original data, composite data containing a fusion label, and data labeled with an adjacent smoothing label (referred to as scheme B), respectively, to obtain the diagnostic accuracy and error of the trained rolling bearing fault diagnosis model.

[0100] The accuracy of the rolling bearing fault diagnosis model trained by scheme A is 97.10%, the accuracy of the rolling bearing fault diagnosis model trained by scheme B is 98.70%, and the accuracy is improved by 1.6%. The error of the rolling bearing fault diagnosis model trained by scheme A is 1.28%, the error of the rolling bearing fault diagnosis model trained by scheme B is 0.60%, and the error is reduced by 0.68%.

[0101] After the data is enhanced by the method of the application, the training set samples are expanded, and the overfitting caused by insufficient samples during training is avoided. The stability of the model is greatly improved by the method of the application, the error of the rolling bearing fault diagnosis model is reduced, and it is proved that the proposed method affects the decision boundary of the model to some extent, thereby improving the generalization ability and robustness of the model. The new training set obtained by combining random fusion and adjacent smoothing labels is inserted into the rolling bearing fault diagnosis model, and the overall performance of the rolling bearing fault diagnosis model is greatly improved, ensuring that the accuracy is increased while the error is reduced, proving the effectiveness and feasibility of the proposed method.

[0102] In order to verify the reliability of the method and system proposed in the application, an anti-noise performance test is also performed. The trained rolling bearing fault diagnosis model is kept unchanged, and Gaussian white noise is continuously added to the test set for verification. The rolling bearing fault diagnosis model trained by the method and system proposed in the application has greatly enhanced anti-noise performance, and the anti-noise ability becomes more significant as the noise increases. This proves that the method and system proposed in the application can enhance the robustness of the rolling bearing fault diagnosis model and enhance the adaptability of the rolling bearing fault diagnosis model in complex environments.

Claims

1. A rolling bearing fault diagnosis method, characterized in that: The method comprises: S1. Collect the vibration signal of each rolling bearing of the same equipment as the raw data. According to the fault condition of each rolling bearing, label all the raw data with a unique hot label. The raw data with the unique hot label are divided into a training set, a validation set, and a test set according to a certain ratio. S2. Within the set time domain, perform discrete Fourier transform on the original data with faults in the training set to obtain the corresponding spectrum information, obtain the amplitude and phase information in the spectrum information, randomly fuse the amplitude and phase information to obtain composite data, and use the proportional fusion method to mark the fusion label of each composite data. The specific operation is as follows; ， in, is the fusion label of composite data; is a parameter between 0 and 1, indicating the percentage of phase features in the fused label; 、 are two original data with faults in the training set, yes Tags, yes Tags, yes The amplitude of yes Phase; S3. The original data with faults in the training set are labeled uniquely, and the labels are reallocated according to the neighboring smoothing strategy. The fault probability of the location of the faulty rolling bearing in the unique label of the faulty rolling bearing in the training set is taken as the actual class, and the fault probabilities of the locations of the rolling bearings on both sides of the faulty rolling bearing are taken as the remaining classes. The sum of the fault probabilities of the actual class and the remaining classes on both sides is 1. Obtaining neighboring smoothed labels of the raw data of each faulty rolling bearing; S4. Use the original data, the composite data with fused labels obtained in S2, and the data with marked adjacent smooth labels obtained in S3 as new training sets, use the new training sets to train the rolling bearing fault diagnosis model, and perform rolling bearing fault diagnosis.

2. A rolling bearing fault diagnosis method according to claim 1, characterized in that: S2 The composite data includes first composite data and second composite data, the first composite data is obtained through frequency spectrum information of the faulty rolling bearing, and the second composite data is obtained through the fault frequency of the faulty rolling bearing.

3. A rolling bearing fault diagnosis method according to claim 2, characterized in that: The specific method for obtaining the first composite data is: Based on the spectrum information of the faulty rolling bearings, calculate the peak-to-peak value of each faulty rolling bearing according to the peak characteristics of the spectrum, and sort all the peak-to-peak values ​​obtained from large to small; Divide all peak-to-peak values ​​into high fault segment, medium fault segment and low fault segment according to preset percentages; The amplitude and phase of the original data of the faulty rolling bearing in the high fault segment and the medium fault segment are randomly fused, and the amplitude and phase of the original data of the faulty rolling bearing in the medium fault segment and the low fault segment are randomly fused to generate the first composite data.

4. A rolling bearing fault diagnosis method according to claim 2, characterized in that: The specific method for obtaining the second composite data is: Rolling bearing faults are divided into inner ring faults, outer ring faults or rolling element faults. When a rolling bearing has one of these faults, the vibration signal is taken as a single signal. When a rolling bearing has two or more of these faults, the vibration signal is taken as a high-order harmonic signal. According to the vibration signal of the faulty rolling bearing, the rolling bearing with high-order harmonic signal is classified as a high-level rolling bearing, and the rest is classified as a low-level rolling bearing. The amplitude and phase of the high-level rolling bearing and the low-level rolling bearing are randomly fused to generate the second composite data.

5. A rolling bearing fault diagnosis method according to claim 1, characterized in that: S2 randomly fuses the amplitude and phase information to obtain composite data, and the specific operation is as follows: ， in, 、 are two original data with faults in the training set, is the composite data generated by random fusion, represents random fusion, represents element-wise multiplication, express The amplitude of express The phase, i Is an imaginary number.

6. A rolling bearing fault diagnosis method according to claim 2, characterized in that: The number of composite data obtained by random fusion is recorded as M, the number of original data is recorded as N, and the quantitative relationship between the two is recorded as; ， Among them, λ is a parameter that customizes the proportion of generated data.

7. A rolling bearing fault diagnosis method according to claim 1, characterized in that: Parameters of the fusion tag Set to 0.

2.

8. A rolling bearing fault diagnosis system, characterized in that: include: Diversity module; The vibration signals of each rolling bearing of the same equipment are collected as raw data. Based on the fault condition of each rolling bearing, all raw data are labeled with a unique hot label. The raw data with the unique hot label are divided into training set, validation set and test set according to a certain ratio. Fusion module; In the set time domain, the original data with faults in the training set are subjected to discrete Fourier transform to obtain the corresponding spectrum information. The amplitude and phase information in the spectrum information are obtained, and the amplitude and phase information are randomly fused to obtain composite data. The proportional fusion method is used to mark the fusion label of each composite data. The specific operation is as follows; ， in, is the fusion label of composite data; is a parameter between 0 and 1, indicating the percentage of phase features in the fused label; 、 are two original data with faults in the training set, yes Tags, yes Tags, yes The amplitude of yes Phase; Neighborhood smoothing module; the original data with faults in the training set are labeled uniquely, and labels are redistributed according to the neighboring smoothing strategy. The failure probability of the location of the faulty rolling bearing in the unique label of the faulty rolling bearing in the training set is used as the actual class, and the failure probability of the locations of the rolling bearings on both sides of the faulty rolling bearing is used as the remaining class. The sum of the failure probabilities of the actual class and the remaining classes on both sides is 1; the neighboring smoothing label of the original data of each faulty rolling bearing is obtained; Machine learning module: uses the original data, the composite data with fusion labels obtained in the fusion module, and the data marked with neighboring smoothing labels obtained in the neighboring smoothing module as new training sets, uses the new training sets to train the rolling bearing fault diagnosis model, and performs rolling bearing fault diagnosis.

Images