A bearing slip diagnosis method for flexible rotor system based on deflection power threshold

Abstract

The present invention discloses a flexible rotor system bearing slip diagnosis method based on deflection power threshold. First, the deflection, bearing vibration acceleration and cage slip rate data at the maximum point of the flexible rotor modal deformation are collected simultaneously, and a data set is constructed; secondly, a deflection curve is drawn according to the rotor deflection data and the deflection average power point is calculated, and whether the bearing movement is in the resonance zone is determined according to the deflection curve and the deflection average power point; then, the vibration acceleration data in the constructed data set is filtered and denoised, and more training samples are generated based on the feature enhancement generative adversarial network; finally, training and test data sets are constructed when the bearing is in the resonance and non-resonance slip zones, respectively, and bearing resonant slip and non-resonance slip migration diagnosis models are constructed based on the constructed training data set, test data set and deep meta-transfer learning algorithm. The present invention can realize accurate measurement of bearing slip rate based on rotor deflection and bearing vibration acceleration data.

Description

Technical Field

[0001] The invention relates to the technical field of bearing slip diagnosis, and mainly to a bearing slip diagnosis method for a flexible rotor system based on a deflection power threshold. Background Art

[0002] As an important component of high-speed rotating machinery, rolling bearings are subject to increasingly harsh working environments. Among the common bearing failures, bearing slip failures account for about 34% of the total number of failures. In the field of high-end bearings, how to accurately measure the bearing cage slip rate under the conditions of limited sensor installation space and the influence of lubricating oil pollution is a very challenging problem. Regarding the related research on bearing cage slip rate measurement, the main strategies currently adopted include bearing slip rate measurement methods based on optical sensors, bearing slip rate measurement based on acoustic wave signals, and bearing slip rate measurement methods based on electromagnetic effect signals. The relevant methods are difficult to adapt to the harsh working environment of high-end bearings; the bearing slip rate measurement method based on bearing material structure modification and the bearing slip rate measurement method based on embedded chips require improvements to the bearing structure, which is not suitable for the non-destructive monitoring needs of high-end bearings.

[0003] Research has shown that the bearing slip rate is closely related to the characteristics of the rotor system, and measuring the bearing slip rate based on the rotor deflection and bearing vibration acceleration signal is a more practical method. The main reasons include: the measurement of rotor deflection signals and bearing vibration acceleration signals has been widely used in actual systems, the relevant signals are easy to obtain and contain more comprehensive and accurate information; and the relevant signal measurement is not easily affected by the operating environment and has high stability. Therefore, the flexible rotor system bearing slip diagnosis method based on the deflection power threshold provided by the present invention is suitable for non-destructive and accurate measurement of bearing slip rate in harsh working environments, and has high practical value. The idea of ​​measuring the relevant bearing slip rate is proposed for the first time in this field and is highly innovative. Summary of the invention

[0004] Purpose of the invention: Based on the problems existing in the above-mentioned background technology, the present invention provides a bearing slip diagnosis method for a flexible rotor system based on a deflection power threshold, which can accurately measure the bearing slip rate based on the flexible rotor deflection and bearing vibration acceleration data.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] A flexible rotor system bearing slip diagnosis method based on deflection power threshold comprises the following steps:

[0007] Step S1: Perform a modal test on the flexible rotor system to obtain the modal vibration shape of the flexible rotor system and the position of the maximum point of the modal deformation of the flexible rotor, and collect the deflection data R of the flexible rotor at the maximum point position.d , while collecting bearing vibration acceleration data A c and cage slip rate data S d , further construct the deflection data R including the flexible rotor d , bearing vibration acceleration data A c and cage slip rate data S d Data set T1;

[0008] Step S2: Deflection data R of the flexible rotor obtained in step S1 d Draw the deflection curve and calculate the deflection average power point {(t1,A p1 ),(t2,A p1 )}, according to the deflection curve and the deflection average power point {(t1,A p1 ),(t2,A p1 )} Determine whether the bearing motion is in the resonant slip zone; when the flexible rotor deflection is greater than or equal to parameter A p1 When the flexible rotor deflection is less than parameter A, the bearing motion is in the resonant slip zone. p1 When , the bearing motion is in the non-resonance slip zone;

[0009] Step S3: the bearing dynamic acceleration data A in the data set T1 constructed in step S1 c Perform filtering and denoising, and obtain training samples T2 based on the feature enhancement generative adversarial network, wherein the training samples T2 include bearing vibration acceleration data generated by the feature enhancement generative adversarial network, forming a data set F = {T1, T2};

[0010] Step S4: construct a training data set F when the bearing is in the resonant slip zone r1 That is, 60% of the data in the resonant slip zone in the data set F constructed in step S3 and the test data set F are randomly selected. r2 That is, the remaining data in the resonant slip zone in the data set F constructed in step S3, and based on the constructed training data set F r1 , test data set F r2 Using deep meta-transfer learning algorithm to build a bearing resonance slip migration diagnosis model M tr ; Construct training data set F when the bearing is in the non-resonance slip zone n1 That is, 60% of the data in the non-resonance slip zone in the data set F constructed in step S3 and the test data set F are randomly selected. n2 That is, the remaining data in the non-resonance slip zone in the data set F constructed in step S3, and based on the constructed training data set F n1 , test data set F n2 Using deep meta-transfer learning algorithm to build a bearing resonance slip migration diagnosis model M tn .

[0011] Preferably, the implementation process of step S2 is:

[0012] Step S2.1: The flexible rotor deflection data R obtained in step S1 d Draw the deflection curve and calculate the deflection average power point {(t1,A p1 ),(t2,A p1 )};

[0013] Flexible rotor deflection average power E R The calculation is shown in formula (1):

[0014]

[0015] Among them, R d (p) is the measured flexible rotor deflection data; N u is the flexible rotor deflection data R d The length of , p is a parameter;

[0016] Step S2.2: According to the deflection curve and the flexible rotor deflection average power point {(t1, A p1 ),(t2,A p1 )} Distinguish whether the bearing movement is in the resonance slip zone; when the flexible rotor deflection is greater than or equal to A p1 When the flexible rotor deflection is less than A, the bearing motion is in the resonant slip zone; p1 When , the bearing motion is in the non-resonance slip zone.

[0017] Preferably, the implementation process of step S3 is:

[0018] Step S3.1: low-pass filter the vibration acceleration data A in the data set T1 constructed in step S1. c Perform filtering and denoising;

[0019] Step S3.2, obtaining a training sample T2 based on a feature-enhanced generative adversarial network, wherein the training sample T2 includes bearing vibration acceleration data generated by the feature-enhanced generative adversarial network, forming a data set F = {T1, T2};

[0020] The feature-enhanced generative adversarial network includes a discriminator and an improved generator; wherein the loss function L(D) of the discriminator satisfies formula (2):

[0021]

[0022] Where x is the actual sample, i.e. the vibration acceleration data A in the data set T1. c ; y is the generated sample, i.e. the bearing vibration acceleration data A generated by the improved generatorcg ; r is the sample input to the discriminator, including the vibration acceleration data A of the actual sample c And the generated bearing vibration acceleration data A cg ;P da , P g and P r are the distributions of actual samples, generated samples, and discriminator input samples respectively; ||▽ r D(r)||2 is the binary norm of the discriminator function gradient of the input sample; λ is the weight of the gradient penalty term; D(x), D(y) and D(r) are the likelihoods that the input samples belong to the true signal distribution output by the discriminator respectively; and The variables x, y, and r belong to P da , P g and P r Expected value;

[0023] The improved generator loss function L(G) satisfies:

[0024]

[0025] Among them, z is the random noise used to generate samples, which obeys Gaussian distribution; P z is the distribution of random noise; is the random noise z belonging to P z The expected value of; D(G(z)) is the probability of being judged as true by the discriminator; μ is the weight; E n are the parameters of the encoder; G(z) is the sample generated by random noise z; E n (G(z)) is the intermediate layer feature of the generated sample output by the discriminator.

[0026] Preferably, the implementation process of step S4 is:

[0027] Step S4.1: Obtain training data set F when the bearing is in the resonant slip zone r1 And the test dataset F r2 , and based on the obtained training data set F r1 , test data set F r2 Using deep meta-transfer learning algorithm to build a bearing resonance slip migration diagnosis model M tr :

[0028] Bearing resonance slip migration diagnosis model M tr It includes a deep neural network model and a meta-transfer learning model; the design of the deep neural network model is shown in formula (4):

[0029]

[0030] Among them, Θ is the feature extractor; θ is the classifier; α is the learning rate; L Te is the cross entropy loss function; ▽ is the parameter change rate; the deep neural network model is trained by minimizing formula (5):

[0031]

[0032] in, is the true label of the i-th source domain sample; I(·) is the indicator function; is the possibility that the i-th source domain dataset is classified into the j-th label dataset; the parameters obtained after training the deep neural network model are used as the initial parameters of the meta-transfer learning model, and the parameters of the meta-transfer learning model are fine-tuned through scaling and translation operations. The optimization of the classifier of the deep neural network model is shown in formula (6):

[0033]

[0034] Among them, γ is the learning rate; Φ S1 and Φ S2 They are the scaling and translation operation parameters respectively; θ is the gradient of the parameter with respect to θ; at the same time, the subdomain adaptation module is introduced into the deep meta-transfer learning algorithm, and the loss function of the subdomain adaptation module satisfies:

[0035]

[0036] where J(·) is the cross entropy loss function for classification of the deep meta-transfer learning model; ξ is the trade-off parameter; N s is the number of labeled samples in the source domain; f(·) is the transfer learning neural network function; is a source domain label sample; is the source domain sample label; φ and ψ are the source and target domain sample distributions respectively; It is the subdomain loss function of the l-th layer network and satisfies:

[0037]

[0038] Where C is the number of subdomain categories; N t is the number of unlabeled samples in the target domain, i and j are parameters; and They are the activation factors of the l-th layer transfer learning neural network; and are the weights of samples belonging to category c; κ(·) is the multi-kernel local maximum mean difference and satisfies:

[0039]

[0040] Among them, βu is the weight of different kernel functions; m is the number of kernel functions, u is the parameter; k u is the u-th kernel function; k is the multi-kernel function;

[0041] Step S4.2: Construct a training data set F when the bearing is in the non-resonant slip zone n1 And the test dataset F n2 , and based on the constructed training data set F n1 , test data set F n2 Using deep meta-transfer learning algorithm to build a bearing resonance slip migration diagnosis model M tn .

[0042] Beneficial effects:

[0043] (1) Aiming at the problem of difficulty in measuring the slip rate of the bearing cage, or more precisely, the problem of difficulty in measuring the cage rotation speed under the influence of lubricating oil pollution and the limited sensor installation environment, the present invention can accurately measure the slip rate of the bearing cage based on the easily measured rotor deflection signal and vibration acceleration signal, and has strong practical value;

[0044] (2) The present invention uses the average power point of the rotor deflection data to accurately determine whether the bearing motion is in the resonant slip zone or the non-resonant slip zone;

[0045] (3) The feature-enhanced generative adversarial network proposed in the present invention can generate data with greater diversity and accuracy;

[0046] (4) The present invention constructs a slippage migration diagnosis model based on a deep meta-transfer learning algorithm, which can effectively prevent model overfitting and accelerate model convergence. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flow chart of a flexible rotor system bearing slip diagnosis method based on a deflection power threshold provided by the present invention;

[0048] Figure 2 The modal vibration shape of the flexible rotor system provided by the present invention;

[0049] Figure 3 is the bearing vibration acceleration signal in the embodiment of the present invention;

[0050] Figure 4 is a bearing resonance slip region discrimination curve based on deflection data in an embodiment of the present invention;

[0051] Figure 5 The actual local vibration acceleration data of the bearing in the embodiment of the present invention;

[0052] Figure 6is the filtered local vibration acceleration data of the bearing in the embodiment of the present invention;

[0053] Figure 7 (a) is a time-frequency diagram of actual data in an embodiment of the present invention;

[0054] Figure 7 (b) is a time-frequency diagram of data generated in an embodiment of the present invention;

[0055] Figure 8 A bearing slip diagnosis data set under variable working conditions in an embodiment of the present invention;

[0056] Fig. 9 The bearing resonance slip migration diagnosis model in the embodiment of the present invention;

[0057] Fig.10 This is the diagnosis result of bearing slippage of the flexible rotor system in the embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention will be further described below in conjunction with the accompanying drawings.

[0059] See also Figure 1-Figure 10 This embodiment provides a flexible rotor system bearing slip diagnosis method based on deflection power threshold. The process of the method is as follows: Figure 1 As shown, this embodiment uses a certain type of flexible rotor system as an example to verify the algorithm performance. The strategy specifically includes the following steps:

[0060] Step S1: Perform a modal test on the flexible rotor system to obtain the modal vibration shape and the position of the maximum point of the modal deformation of the flexible rotor, and collect the deflection data R at the maximum point of the modal deformation of the flexible rotor. d , bearing vibration acceleration data A c and cage slip rate data S d , and construct data set T1;

[0061] Specifically, in this embodiment, step S1 includes:

[0062] Step S1.1, perform a modal test on the flexible rotor system to obtain the modal vibration shape and the position of the maximum point of modal deformation of the flexible rotor; Figure 2 As shown, it is the modal vibration shape of the flexible rotor system, and the maximum point of the modal deformation of the flexible rotor is located at position 5;

[0063] Step S1.2: Simultaneously collect the deflection data R at the maximum point of the flexible rotor modal deformation d , bearing vibration acceleration data A c and cage slip rate data S d , and construct the data set T1; Figure 3 The following is the collected bearing vibration acceleration data Ac ;

[0064] Step S2: respectively calculate the rotor deflection data R obtained in step S1. d Draw the deflection curve and calculate the deflection average power point {(t1,A p1 ),(t2,A p1 )}, according to the deflection curve and the deflection average power point {(t1,A p1 ),(t2,A p1 )} Determine whether the bearing motion is in the resonance zone; when the flexible rotor deflection is greater than or equal to A p1 When the flexible rotor deflection is less than A, the bearing motion is in the resonant slip zone; p1 When , the bearing motion is in the non-resonance slip zone;

[0065] Specifically, in this embodiment, step S2 includes:

[0066] Step S2.1: The rotor deflection data R obtained in step S1.2 d Draw the deflection curve and calculate the deflection average power point {(t1,A p1 ),(t2,A p1 )};

[0067] The deflection average power is calculated as follows:

[0068]

[0069] Among them, R d (i) is the measured deflection data; N u is the deflection data R d Length;

[0070] Step S2.2: Deflection average power point {(t1, A p1 ),(t2,A p1 )} distinguish whether the bearing movement is in the resonance zone; such as Figure 4 As shown, when the bearing movement time is in the area between t1 and t2, the bearing is in the resonant slip zone; when the bearing is in the remaining area, the bearing is in the non-resonant slip zone.

[0071] Step S3: the vibration acceleration data A in the data set T1 constructed in step S1 c Perform filtering and denoising, and generate more vibration acceleration data training samples T2 based on feature enhancement generative adversarial network to form a data set F = {T1, T2};

[0072] Specifically, in this embodiment, step S3 includes:

[0073] Step S3.1: low-pass filter the vibration acceleration data A in the data set T1 constructed in step S1. c Filter and remove noise; the original local vibration acceleration data of the bearing is as follows Figure 5 As shown in the figure, the filtered bearing local vibration acceleration data is as follows Figure 6 shown.

[0074] Step S3.2, generating more vibration acceleration data training samples T2 based on feature enhancement generative adversarial network to form a data set F = {T1, T2};

[0075] The improved generator loss function and the deep feature enhancement module based on the attention mechanism are used to design a feature-enhanced generative adversarial network; among them, the discriminator loss function satisfies:

[0076]

[0077] Where x is the actual sample, i.e. the vibration acceleration data A in the data set T1. c ; y is the generated sample, i.e. the vibration acceleration data A generated by the improved generator cg ; r is the sample input to the discriminator, including the actual vibration acceleration data A c And the generated vibration acceleration data A cg ;P da , P g and P r are the distributions of actual samples, generated samples, and discriminator input samples respectively; ||▽ r D(r)||2 is the binary norm of the discriminator function gradient of the input sample; λ is the weight of the gradient penalty term; D(x), D(y) and D(r) are the likelihoods that the input samples belong to the true signal distribution output by the discriminator respectively; and The variables x, y, and r belong to P da , P g and P r Expected value;

[0078] The improved generator loss function satisfies:

[0079]

[0080] Among them, z is the random noise used to generate samples, which obeys Gaussian distribution; P z is the distribution of random noise; Is the variable z belonging to P z The expected value of; D(G(z)) is the probability that the data is judged as real by the discriminator; μ is the weight of the loss term; E n are the parameters of the encoder; G(z) is the sample generated by the noise signal z; En (G(z)) is the intermediate layer feature of the generated sample output by the discriminator. The time-frequency diagram of the original data and the generated data is compared as follows Figure 7 (a) and Figure 7 (b) The number of samples generated by the feature-enhanced generative adversarial network is shown in Table 1:

[0081] Table 1 Number of samples generated by feature-enhanced generative adversarial networks

[0082]

[0083] Step S4: construct a training data set F when the bearing is in the resonant slip zone r1 And the test dataset F r2 , and based on the constructed training data set F r1 , test data set F r2 Using deep meta-transfer learning algorithm to build a bearing resonance slip migration diagnosis model M tr ; Construct training data set F when the bearing is in the non-resonance slip zone n1 And the test dataset F n2 , and based on the constructed training data set F n1 , test data set F n2 Using deep meta-transfer learning algorithm to build a bearing resonance slip migration diagnosis model M tn .

[0084] Specifically, in this embodiment, step S4 includes:

[0085] Step S4.1: Construct a training data set F when the bearing is in the resonant slip zone r1 And the test dataset F r2 , and based on the constructed training data set F r1 , test data set F r2 Using deep meta-transfer learning algorithm to build a bearing resonance slip migration diagnosis model M tr ; The collected data set is divided into working conditions according to the flexible rotor speed, and finally the bearing slip diagnosis data set under 17 working conditions is obtained, such as Figure 8 As shown, they are labeled AQ, etc.

[0086] The bearing resonance slip migration diagnosis model includes two parts: deep neural network and meta-transfer learning. Its structure is as follows: Fig. 9 As shown; the deep neural network is designed as follows:

[0087] [Θ; θ]=: [Θ; θ]-α▽L Te (4)

[0088] Among them, Θ is the feature extractor; θ is the classifier; α is the learning rate; L Teis the cross entropy loss function; ▽ is the parameter change rate; the deep neural network model is trained by minimizing formula (5):

[0089]

[0090] in, is the true label of the i-th source domain sample; I(·) is the indicator function; is the probability that the i-th source domain dataset is classified into the j-th label dataset; the parameters obtained after deep neural network training are used as the initial parameters of the meta-transfer learning model, and the parameters of the meta-transfer learning model are fine-tuned through scaling and translation operations, and the classifier is further optimized as follows:

[0091] θ=:θ-γ▽ θ ([Θ;θ],Φ S1 ,Φ S2 ) (6)

[0092] Among them, γ is the learning rate; Φ S1 and Φ S2 They are the scaling and translation operation parameters respectively; θ is the gradient of the parameter with respect to θ; at the same time, the subdomain adaptation module is introduced into the deep meta-transfer learning algorithm, and the subdomain adaptation loss function satisfies:

[0093]

[0094] where J(·) is the cross entropy loss function for classification of the deep meta-transfer learning model; ξ is the trade-off parameter; N s is the number of labeled samples in the source domain; f(·) is the transfer learning neural network function; is a source domain label sample; is the source domain sample label; φ and ψ are the source and target domain sample distributions respectively; is the subdomain loss function of the l-th layer network, which satisfies:

[0095]

[0096] Where C is the number of subdomain categories; N t is the number of unlabeled samples in the target domain; and They are the activation factors of the l-th layer transfer learning neural network; and are the weights of samples belonging to category c; κ(·) is the multi-kernel local maximum mean difference, which satisfies:

[0097]

[0098] Among them, β uis the weight of different kernel functions; m is the number of kernel functions; k u is the uth kernel function; k is a multi-kernel function; the structural parameters of the bearing resonance slip migration diagnosis model are shown in Table 2:

[0099] Table 2 Structural parameters of bearing resonance slip migration diagnostic model

[0100]

[0101] Step S4.2: Construct a training data set F when the bearing is in the non-resonant slip zone n1 And the test dataset F n2 , and based on the constructed training data set F n1 , test data set F n2 Using deep meta-transfer learning algorithm to build a bearing resonance slip migration diagnosis model M tn The diagnosis results of the bearing slip rate of the flexible rotor system obtained based on the bearing slip migration diagnosis model are as follows: Fig.10 As shown, the final diagnostic accuracy reached 96.32%.

[0102] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A flexible rotor system bearing slip diagnosis method based on deflection power threshold, characterized in that: The following steps are involved: Step S1: Perform a modal test on the flexible rotor system to obtain the modal vibration shape of the flexible rotor system and the position of the maximum point of the modal deformation of the flexible rotor, and collect the deflection data R of the flexible rotor at the maximum point position. d , while collecting bearing vibration acceleration data A c and cage slip rate data S d , further construct the deflection data R including the flexible rotor d , bearing vibration acceleration data A c and cage slip rate data S d Data set T1; Step S2: Deflection data R of the flexible rotor obtained in step S1 d Draw the deflection curve and calculate the deflection average power point {(t1,A p1 ),(t2,A p1 )}, according to the deflection curve and the deflection average power point {(t1,A p1 ),(t2,A p1 )} Determine whether the bearing motion is in the resonant slip zone; when the flexible rotor deflection is greater than or equal to parameter A p1 When the flexible rotor deflection is less than parameter A, the bearing motion is in the resonant slip zone; when the flexible rotor deflection is less than parameter A p1 When , the bearing motion is in the non-resonance slip zone; Step S3: the bearing dynamic acceleration data A in the data set T1 constructed in step S1 c Perform filtering and denoising, and obtain training samples T2 based on the feature enhancement generative adversarial network, wherein the training samples T2 include bearing vibration acceleration data generated by the feature enhancement generative adversarial network, forming a data set F = {T1, T2}; Step S4: construct a training data set F when the bearing is in the resonant slip zone r1 That is, 60% of the data in the resonant slip zone in the data set F constructed in step S3 and the test data set F are randomly selected. r2 That is, the remaining data in the resonant slip zone in the data set F constructed in step S3, and based on the constructed training data set F r1 , test data set F r2 Using deep meta-transfer learning algorithm to build a bearing resonance slip migration diagnosis model M tr ; Construct training data set F when the bearing is in the non-resonance slip zone n1 That is, 60% of the data in the non-resonance slip zone in the data set F constructed in step S3 and the test data set F are randomly selected. n2 That is, the remaining data in the non-resonance slip zone in the data set F constructed in step S3, and based on the constructed training data set F n1 , test data set F n2 Using deep meta-transfer learning algorithm to build a bearing resonance slip migration diagnosis model M tn ; The implementation process of step S2 is: Step S2.1: The flexible rotor deflection data R obtained in step S1 d Draw the deflection curve and calculate the deflection average power point {(t1,A p1 ),(t2,A p1 )}; Flexible rotor deflection average power E R The calculation is shown in formula (1): Among them, R d (p) is the measured flexible rotor deflection data; N u is the flexible rotor deflection data R d The length of , p is a parameter; Step S2.2: Determine the average power point {(t1, A p1 ),(t2,A p1 )} Distinguish whether the bearing movement is in the resonance slip zone; when the flexible rotor deflection is greater than or equal to A p1 When the flexible rotor deflection is less than A, the bearing motion is in the resonant slip zone; p1 When , the bearing motion is in the non-resonance slip zone; The implementation process of step S4 is: Step S4.1: Obtain training data set F when the bearing is in the resonant slip zone r1 And the test dataset F r2 , and based on the obtained training data set F r1 , test data set F r2 Using deep meta-transfer learning algorithm to build a bearing resonance slip migration diagnosis model M tr : Bearing resonance slip migration diagnosis model M tr It includes a deep neural network model and a meta-transfer learning model; the design of the deep neural network model is shown in formula (4): Among them, Θ is the feature extractor; θ is the classifier; α is the learning rate; L Te is the cross entropy loss function; is the parameter change rate; the deep neural network model is trained by minimizing formula (5): in, is the true label of the i-th source domain sample; I(·) is the indicator function; is the possibility that the i-th source domain dataset is classified into the j-th label dataset; the parameters obtained after training the deep neural network model are used as the initial parameters of the meta-transfer learning model, and the parameters of the meta-transfer learning model are fine-tuned through scaling and translation operations. The optimization of the classifier of the deep neural network model is shown in formula (6): Among them, γ is the learning rate; Φ S1 and Φ S2 They are the scaling and translation operation parameters; is the gradient of the parameter with respect to θ; at the same time, the subdomain adaptation module is introduced into the deep meta-transfer learning algorithm, and the loss function of the subdomain adaptation module satisfies: where J(·) is the cross entropy loss function for classification of the deep meta-transfer learning model; ξ is the trade-off parameter; N s is the number of labeled samples in the source domain; f(·) is the transfer learning neural network function; is a source domain label sample; is the source domain sample label; φ and ψ are the source and target domain sample distributions respectively; It is the subdomain loss function of the l-th layer network and satisfies: Where C is the number of subdomain categories; N t is the number of unlabeled samples in the target domain, i and j are parameters; and They are the activation factors of the l-th layer transfer learning neural network; and are the weights of samples belonging to category c; κ(·) is the multi-kernel local maximum mean difference and satisfies: Among them, β u is the weight of different kernel functions; m is the number of kernel functions, u is the parameter; k u is the u-th kernel function; k is the multi-kernel function; Step S4.2: Construct a training data set F when the bearing is in the non-resonance slip zone n1 And the test dataset F n2 , and based on the constructed training data set F n1 , test data set F n2 Using deep meta-transfer learning algorithm to build a bearing resonance slip migration diagnosis model M tn .

2. A flexible rotor system bearing slip diagnosis method based on deflection power threshold according to claim 1, characterized in that: The implementation process of step S3 is: Step S3.1: low-pass filter the vibration acceleration data A in the data set T1 constructed in step S1. c Perform filtering and denoising; Step S3.2, obtaining a training sample T2 based on a feature-enhanced generative adversarial network, wherein the training sample T2 includes bearing vibration acceleration data generated by the feature-enhanced generative adversarial network, forming a data set F = {T1, T2}; The feature-enhanced generative adversarial network includes a discriminator and an improved generator; wherein the loss function L(D) of the discriminator satisfies formula (2): Where x is the actual sample, i.e. the vibration acceleration data A in the data set T1. c ; y is the generated sample, i.e. the bearing vibration acceleration data A generated by the improved generator cg ; r is the sample input to the discriminator, including the vibration acceleration data A of the actual sample c And the generated bearing vibration acceleration data A cg ;P da , P g and P r are the distributions of actual samples, generated samples, and discriminator input samples, respectively; is the binary norm of the discriminator function gradient of the input sample; λ is the weight of the gradient penalty term; D(x), D(y) and D(r) are the probabilities that the input samples belong to the true signal distribution output by the discriminator respectively; and The variables x, y, and r belong to P da , P g and P r Expected value; The improved generator loss function L(G) satisfies: Among them, z is the random noise used to generate samples, which obeys Gaussian distribution; P z is the distribution of random noise; is the random noise z belonging to P z The expected value of; D(G(z)) is the probability of being judged as true by the discriminator; μ is the weight; E n are the parameters of the encoder; G(z) is the sample generated by random noise z; E n (G(z)) is the intermediate layer feature of the generated sample output by the discriminator.

Images