Method for evaluating degradation state of bearing retainer based on digital-analog linkage

Abstract

The invention provides a bearing retainer degradation state evaluation method based on digital-analog linkage, which comprises the following steps of: constructing a real retainer degradation experiment table to acquire a full life cycle signal of bearing retainer degradation, and constructing a physical entity; constructing a characterization mathematical model of the degradation state of the retainer, introducing the characterization mathematical model into the dynamic model, and generating a simulation signal of the degradation process of the retainer; on the basis of an SE attention mechanism, constructing a CycleGAN model, correcting a simulation signal, generating a high-fidelity virtual signal, and enabling the feature distribution of the virtual signal to be further close to a real degraded signal; and constructing an SKformer model, and diagnosing the fault degradation stage of the retainer. According to the method, the SKnet is introduced into the Transform, so that the model automatically adapts to the size of a receptive field from multiple scales, the local information learning ability of the network is improved, and the ability of the SKform to extract the abrupt change feature of the cage fault and the degradation state recognition performance are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing degradation state monitoring, and more particularly, to a method for evaluating the degradation state of a bearing cage based on digital-physical interaction. Background Art

[0002] As an important basic component of rotating machinery, the health status of rolling bearings directly affects the stable operation of equipment. Due to its long-term service under complex conditions, bearings are prone to cracks and eventually fail as the cracks expand. Therefore, accurately evaluating the degradation stage of bearing faults can provide early warnings for equipment and avoid sudden shutdowns of the transmission system. This is of great significance for the health management of rotating machinery. Currently, most studies are based on the degradation process of the outer ring of the bearing. As one of the important components of the bearing, the cage is more likely to fail due to its thinner structure and manufacturing materials compared to the inner ring, outer ring, and rolling elements, and it often shows instantaneous fracture. Therefore, accurately evaluating the current state of the bearing cage is of great significance for extending the service life of mechanical equipment and improving its reliability.

[0003] In recent years, deep learning technology has been widely used in the field of rotating machinery fault diagnosis due to its advantage of being able to adaptively process high-dimensional and complex data. However, most current bearing intelligent diagnosis methods based on deep learning are proposed under the premise of having sufficient and high-quality data samples. In engineering practice, the frequency of equipment failures is low, and collecting a large number of fault samples requires a large amount of cost and time. Therefore, when the available samples are scarce, these intelligent diagnosis methods will face the risk of failure. With the digital transformation and intelligent development in the industrial field, the emergence and rapid rise of digital twin technology provide a possibility to solve this challenge. Summary of the Invention

[0004] In view of the above-mentioned technical problem of insufficient available samples in engineering practice, a method for evaluating the degradation state of a bearing cage based on digital-physical interaction is provided.

[0005] The technical means adopted by the present invention are as follows:

[0006] A method for evaluating the degradation state of a bearing cage based on digital-physical interaction, comprising:

[0007] S1. Build a real cage degradation test bench to collect the full life cycle signals of bearing cage degradation and construct a physical entity;

[0008] S2. Construct a mathematical model for characterizing the degradation state of the cage and introduce the characterization mathematical model into the dynamic model to generate simulation signals of the cage degradation process;

[0009] S3. Based on the SE attention mechanism, construct a CycleGAN model to correct the simulation signal and generate a high-fidelity virtual signal, so that the feature distribution of the virtual signal further approximates the true degraded signal;

[0010] S4. Construct an SKformer model to diagnose the fault degradation stage of the cage.

[0011] Furthermore, step S1 specifically includes:

[0012] S11. Conduct a full-life cycle experiment through a bearing cage fatigue failure test bench;

[0013] S12. Apply an eccentric load of radial load and axial load on the test bench to accelerate the bearing fatigue failure;

[0014] S13. Collect the acceleration signals of the bearing under the full-life cycle and store them in a csv file.

[0015] Furthermore, in step S12, the eccentric load of the applied radial load is 5.1 kN, and the eccentric load of the applied axial load is 1.4 kN.

[0016] Furthermore, step S2 specifically includes:

[0017] S21. When the bearing cage breaks, the clearance of the cage pocket for the ball at the break will become larger, so the ball will deviate from the original position. Therefore, an additional position angle θ f is introduced at the offset ball, and the cage breakage fault is simulated as follows:

[0018]

[0019] where θ 1 is the initial position angle of the No. 1 ball of the bearing, ω r is the inner ring rotation angular velocity, r b and R b respectively represent the inner and outer ring raceway radii of the bearing, and N b is the number of balls in the bearing;

[0020] S22. Based on the Hertz contact theory, considering the bearing housing clearance and the cage breakage fault, the 5-degree-of-freedom nonlinear bearing force at the i-th node of the rotor is expressed as follows:

[0021]

[0022] where α j represents the pressure angle of the j-th ball, and δ j represents the contact deformation between the j-th ball and the raceway, and H(δ j ) is the Heaviside function, and kbj is the Hertz contact stiffness of the j-th ball;

[0023] S23. Using the finite element method and the lumped mass method, establish a system dynamics model and establish the dynamic equation of the rotor-bearing system as follows:

[0024]

[0025] where, M RS , C RS , J RS , K RS are the mass matrix, damping matrix, gyroscopic matrix, and stiffness matrix of the rotating shaft respectively; M SS and J SS are the mass matrix and gyroscopic matrix of the bushing respectively; F b represents the restoring force vector of the rolling bearing; F u represents the rotor unbalance force vector; F r represents the external load vector; G represents the rotor gravity vector;

[0026] S24. Using the Newmark-β method, calculate the dynamic equation formula in step S23 to obtain the numerical solution and extract the vibration response results at the bearing housing;

[0027] S25. Characterize the cage degradation process by changing the additional position angle θ f where θ f ∈[0°, 4°].

[0028] Furthermore, step S3 specifically includes:

[0029] S31. Through the adversarial loss function L GAN , perform adversarial training on the generator and discriminator in the CycleGAN model as follows:

[0030]

[0031] where, L GAN is the adversarial loss, represents the expectation that the actual signal follows the actual physical domain distribution, represents the expectation that the simulation signal follows the simulation domain distribution;

[0032] S32. Introduce a cycle consistency loss in the CycleGAN model to ensure the consistency between the generated signal and the target domain signal. The loss function is as follows:

[0033]

[0034] where, ||·|| 1 represents the L1 norm;

[0035] S33. Optimize and fine-tune the generator G and the discriminator F by minimizing the total loss function as follows:

[0036] L total = L GAN (G, D p ) + L GAN (F, D q ) + λL cyc (G, F)

[0037] where the coefficient λ controls the proportion of the cycle consistency loss in the total loss;

[0038] S34. Introduce the SE attention mechanism into the encoder and decoder, dynamically adjust the channel weights by learning the relationships between features, enabling the model to focus on the features with the greatest contribution, thereby enhancing the feature learning effect of the model and reducing the distortion phenomenon during the signal generation process.

[0039] Furthermore, step S4 specifically includes:

[0040] S41. Perform a simple linear transformation on the input signal and map the signal sequence to a high-dimensional embedding space as follows:

[0041]

[0042] where dim represents the dimension of the time series embedding, and W embedding represents a learnable matrix;

[0043] S42. A randomly initialized trainable embedding is added to the beginning of the embedded token sequence to aggregate the information of each part as follows:

[0044]

[0045] S43. Use a learnable position encoding to more flexibly extract position information, add the position encoding and the token embedding to obtain the input embedding as follows:

[0046]

[0047] where, represents the position embedding, represents the Patch sequence;

[0048] S44. Perform multi-scale feature extraction through SKnet, and combine the two to more comprehensively model the semantic information of the fault signal, thereby improving the diagnostic accuracy and generalization performance of the model;

[0049] S45. For the input embedding, transform the input feature map through a convolutional kernel;

[0050] S46. Implement feature fusion of different branches by element-wise summation of the above feature maps, and then generate statistical features in the channel direction through global average pooling operation to embed global information, and then construct a more compact feature

[0051] to achieve model adaptive selection, as follows:

[0052] U = U + U

[0053]

[0054] d = max(C / r, 16)

[0055] where σ represents the ReLU activation function, represents batch normalization, r represents the reduction rate;

[0056] S47. Through the cross-channel soft attention mechanism, adaptively select information at different spatial scales, and obtain the final feature map by superimposing the attention weights on convolutional kernels of different sizes as follows:

[0057]

[0058] a c + b c = 1

[0059] where, a c and b c respectively represent the soft attention vectors of U c and U c ;

[0060] S48. Through linear transformation, transform a group of feature maps with d model dimensions output by SKnet into keys K, values V, and queries Q with d dimensions respectively, and perform the multi-head self-attention mechanism;

[0061] S49. Input the features into a multi-layer perceptron, obtain the probability values of each fault category through the Softmax function, and obtain the final fault category according to the maximum probability value.

[0062] Further, in step S45, the input feature map is transformed by convolution kernels of sizes 3×1 and 5×1, specifically as follows:

[0063]

[0064] Among them, F * (·) operation consists of grouped convolution, batch normalization, and Relu activation function in sequence.

[0065] Compared with the prior art, the present invention has the following advantages:

[0066] A method for evaluating the degradation state of a bearing cage based on digital-analog linkage provided by the present invention first constructs a high-fidelity virtual model to accurately characterize the degradation state of the cage. Then, CycleGAN is used to correct the simulation signal. The SE attention mechanism is introduced into CycleGAN. By learning the relationship between features, it can dynamically display only the weights of the channels, enabling SECycleGAN to focus on high-contribution features and improve the quality of the generated signal. Finally, SKformer is used to evaluate the degradation stage of the bearing cage. SKnet is introduced into the Transformer, enabling the model to automatically adapt to the size of the receptive field from multiple scales, thereby improving the network's local information learning ability. This method enhances the ability of SKformer to extract the mutation features of cage faults and the degradation state recognition performance.

[0067] Based on the above reasons, the present invention can be widely promoted in the fields such as bearing degradation state monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0069] Figure 1 It is a flowchart of the method of the present invention.

[0070] Figure 2 It is a characterization of the cage fracture fault process provided by the embodiment of the present invention.

[0071] Figure 3 It is the SKnet mechanism provided by the embodiment of the present invention.

[0072] Figure 4 It is the SKformer network structure provided by the embodiment of the present invention

[0073] Figure 5Comparison between the signal generated by the method of the present invention and the measured signal provided by the embodiments of the present invention

[0074] Figure 6 Pearson correlation coefficients between the signals generated by different methods and the original signal provided by the embodiments of the present invention.

[0075] Figure 7 Spectrum diagrams of the signal generated by the method of the present invention and the true signal provided by the embodiments of the present invention.

[0076] Figure 8 Comparison of the evaluation results of the degradation stage between the method of the present invention and other methods provided by the embodiments of the present invention.

[0077] In the figure, (a) 500 samples for each category; (b) 300 samples for each category;

[0078] Figure 9 Confusion matrix results provided by the embodiments of the present invention.

[0079] In the figure, (a) Skformer; (b) DANN; (c) DCAN; (d) WDCNN; Detailed implementation manners

[0080] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0081] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in an order different from those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily need to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0082] As Figure 1 shown, the present invention provides a method for evaluating the degradation state of a bearing cage based on digital-analog linkage, including:

[0083] S1. Set up a real cage degradation test bench to collect the full - life - cycle signals of bearing cage degradation and construct a physical entity;

[0084] S2. Construct a mathematical model for characterizing the degradation state of the cage, and introduce the characterization mathematical model into the dynamic model to generate simulation signals of the cage degradation process;

[0085] S3. Based on the SE attention mechanism, construct a CycleGAN model to correct the simulation signals and generate high - fidelity virtual signals, so that the feature distribution of the virtual signals is further approximated to the real degradation signals;

[0086] S4. Construct an SKformer model to diagnose the fault degradation stage of the cage.

[0087] In specific implementation, as a preferred implementation manner of the present invention, step S1 specifically includes:

[0088] S11. Conduct a full - life - cycle experiment through a bearing cage fatigue failure test bench;

[0089] S12. Apply eccentric forces of radial load and axial load on the test bench to accelerate bearing fatigue failure; in this embodiment, the eccentric force of the applied radial load is 5.1 kN, and the eccentric force of the applied axial load is 1.4 kN.

[0090] S13. Collect the acceleration signals of the bearing under the full - life - cycle and store them in a csv file.

[0091] In specific implementation, as a preferred implementation manner of the present invention, step S2 specifically includes:

[0092] S21. When the bearing cage breaks, the clearance of the cage pocket for the balls at the break will become larger, so the balls will deviate from the original position. As Figure 2 shown, so the method of the present invention introduces an additional position angle θ f at the offset ball to simulate the cage breakage fault, as follows:

[0093]

[0094] Among them, θ 1 is the initial position angle of the No. 1 ball of the bearing, ω r is the inner - ring rotation angular velocity, r b and R b respectively represent the inner and outer - ring raceway radii of the bearing, and N b is the number of bearing balls;

[0095] S22. Based on the Hertz contact theory, considering the bearing housing clearance and the cage fracture fault, the 5-degree-of-freedom nonlinear bearing force at the i-th node of the rotor is expressed as follows:

[0096]

[0097] Among them, α j represents the pressure angle of the j-th ball, and δ j represents the contact deformation between the j-th ball and the raceway. H(δ j ) is the Heaviside function, and k bj is the Hertz contact stiffness of the j-th ball;

[0098] S23. Using the finite element method and the lumped mass method, establish a system dynamics model and establish the dynamic equation of the rotor-bearing system as follows:

[0099]

[0100] Among them, M RS , C RS , J RS , K RS are the mass matrix, damping matrix, gyro matrix, and stiffness matrix of the rotating shaft respectively; M SS and J SS are the mass matrix and gyro matrix of the bearing sleeve respectively; F b represents the restoring force vector of the rolling bearing; F u represents the rotor unbalance force vector; F r represents the external load vector; G represents the rotor gravity vector;

[0101] S24. Using the Newmark-β method, calculate the formula of the dynamic equation in step S23 to obtain the numerical solution, and extract the vibration response result at the bearing housing;

[0102] S25. Characterize the cage degradation process by changing the additional position angle θ f . By comparing the simulation signal and the real experimental signal, it is found that when θ f ∈[0°, 4°], the simulation signal and the real signal match well.

[0103] In specific implementation, as a preferred implementation manner of the present invention, it is characterized in that in step S3, the real signal collected in step S1 is used as the label data, and the simulation signal obtained in step S2 is used as the training data of the SECycleGAN model for correction to make its feature distribution, specifically including:

[0104] S31. Through the adversarial loss function L GAN, perform adversarial training on the generator and discriminator in the CycleGAN model as follows:

[0105]

[0106] Among them, L GAN is the adversarial loss, represents the expectation that the actual signal follows the actual physical domain distribution, represents the expectation that the simulation signal follows the simulation domain distribution;

[0107] S32. Introduce a cyclic consistency loss in the CycleGAN model to ensure the consistency between the generated signal and the target domain signal. The loss function is as follows:

[0108]

[0109] Among them, ||·|| 1 represents the L1 norm;

[0110] S33. Optimize and fine-tune the generator G and the discriminator F by minimizing the total loss function as follows:

[0111] L total = L GAN (G, D p ) + L GAN (F, D q ) + λL cyc (G, F)

[0112] Among them, the coefficient λ controls the proportion of the cyclic consistency loss in the total loss;

[0113] S34. Introduce the SE attention mechanism in the encoder and decoder, dynamically adjust the channel weights by learning the relationship between features, enable the model to focus on the features with the greatest contribution, thereby enhancing the feature learning effect of the model and reducing the distortion phenomenon in the signal generation process.

[0114] Specifically in implementation, as a preferred implementation manner of the present invention, step S4 specifically includes:

[0115] S41. Perform a simple linear transformation on the input signal and map the signal sequence to a high-dimensional embedding space as follows:

[0116]

[0117] Among them, dim represents the dimension of the time series embedding, and W embedding represents a learnable matrix;

[0118] S42. A randomly initialized trainable embedding Is added to the beginning of the embedded token sequence to aggregate the information of each part as follows:

[0119]

[0120] S43. Use learnable positional encoding Extract positional information more flexibly, add positional encoding and token embedding to obtain input embedding as follows:

[0121]

[0122] Among them, Represents positional embedding, Represents Patch sequence;

[0123] S44. Perform multi-scale feature extraction through SKnet, and the combination of the two models the semantic information of the fault signal more comprehensively, thereby improving the diagnostic accuracy and generalization performance of the model;

[0124] S45. For the input embedding, transform the input feature map through a convolutional kernel;

[0125] S46. Implement feature fusion of different branches by element-wise summation of the above feature maps, and then generate statistical features in the channel direction through global average pooling operation To embed global information, and then build a more compact feature through a fully connected layer To achieve model adaptive selection, specifically as follows:

[0126] U = U + U

[0127]

[0128] d = max(C / r, 16)

[0129] Among them, σ represents the ReLU activation function, Represents batch normalization, r represents the reduction rate;

[0130] S47. As Figure 3 Shown, through the cross-channel soft attention mechanism, adaptively select information of different spatial scales, and obtain the final feature map by superimposing the attention weights on convolutional kernels of different sizes As follows:

[0131]

[0132] a c + b c = 1

[0133] Among them, a c and b c respectively represent the soft attention vectors of U c and U c , respectively.

[0134] S48. Through linear transformation, a group of feature maps with d model dimensions output by SKnet are respectively converted into keys K, values V, and queries Q with d dimensions, and the multi-head self-attention mechanism is executed, as Figure 4 shown;

[0135] S49. The features are passed into a multi-layer perceptron, and the probability values of each fault category are obtained through the Softmax function, and the final fault category is obtained according to the maximum probability value.

[0136] Figure 5 shows the comparison between the simulated signals generated by the SECycleGAN model proposed by the present invention and the measured signals. It can be seen that the generated signals maintain almost the same degradation trend as the original signals, but there is an upward shift trend in the healthy stage. In addition, the probability distribution differences between the generated signals and the original signals at 3 time points are randomly analyzed, and the sampling interval is 1 s. It can be seen that there is a deviation trend in the generated signals at 3655 s, but the probability distribution shapes are still similar. The probability distribution curves at 3678 s and 3683 s coincide highly, further illustrating the effectiveness of the generated signals.

[0137] To prove the superiority of the SECycleGAN proposed by the present invention, comparisons are made with the WGAN, DCGAN, and ACGAN models. Here, the Pearson correlation coefficient is used to measure the similarity between the generated signals and the measured signals. The generated signals are evenly divided into 9 segments, Figure 6 showing the Pearson correlation coefficients between each segment of the signals of different models and the original signals. It can be seen that the model proposed by the present invention achieves the optimal diagnostic accuracy in most signal segments, and the average Pearson correlation coefficient is 0.9826, which is better than other methods.

[0138] Figure 7 is the comparison of the time-domain waveforms and envelope spectrograms of the generated signals and the actual signals at 3662 s - 3663 s and 3685 s - 3686 s. It can be seen that the envelope spectrum frequency components of the vibration signals at 3662 s - 3663 s mainly consist of the rotation frequency f r and its multiple frequencies nf r , the rolling bearing variable compliance vibration frequency f vc , and the combined frequency components of f r and the cage frequency f c , i.e., f r ±f cComposition. After the cage breaks down, f r and f vc The related frequency components are gradually overwhelmed by strong noise, and the vibration frequency is mainly the rotation frequency of the cage and its high-order harmonic frequency components nf c dominates, and the frequency amplitude increases significantly, rising from 0.2 m / s 2 to 25 m / s 2 . In addition, through comparison, it is found that the frequency components of the simulated signal corrected by SECycleGAN are highly consistent with the characteristic frequencies of the actual degradation signal, which confirms the potential of SECycleGAN in effectively correcting simulated signals.

[0139] Figure 8 is the evaluation result of the SKformer for the degradation stage of the bearing cage. The models proposed in the present invention have the highest diagnostic accuracies of 98.27% and 97.76% respectively under different numbers of data sets, and the standard deviations are significantly lower than those of other models. By comparing with the Transformer model, it is found that the accuracy of the model with the SK module network increases by about 3%. This is because SKnet can adaptively select the appropriate convolution kernel size according to the input signal, so that the model can better adapt to the characteristics in different scenarios, thereby improving the model's ability to recognize the complex characteristics of fault signals. The diagnostic accuracies of the two DANN and CDAN models with domain transfer modules are both about 90%. The diagnostic accuracy of the WDCNN model is the lowest, and the average diagnostic accuracy on the two data sets is lower than 80%.

[0140] Figure 9 are the confusion matrix results of different models. It can be seen that the method proposed in the present invention achieves 100% recognition accuracy only for the samples in the 1st and 4th stages. There is only a small amount of confusion in the 2nd and 3rd stages, but the diagnostic accuracy still exceeds 95%. DANN and CDAN mainly tend to predict the samples in the 3rd stage as the 2nd stage, and the diagnostic accuracies in the 3rd stage are 51% and 62% respectively. WDCNN shows misclassification to varying degrees in all categories, especially the recognition rate in the fourth stage is only 63%.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for evaluating the degradation state of a bearing cage based on digital-analog linkage, characterized in that: include: S1. Build a real cage degradation test bench to collect the full life cycle signal of bearing cage degradation and construct a physical entity; S2. constructing a mathematical model to characterize the degradation state of the cage, and introducing the mathematical model to the dynamics model to generate a simulation signal of the cage degradation process; S3. Based on the SE attention mechanism, a CycleGAN model is constructed to correct the simulated signal and generate a high-fidelity virtual signal, so that the feature distribution of the virtual signal is closer to the real degraded signal. S4. Construct the SKformer model to diagnose the fault degradation stage of the cage.

2. The method for evaluating the degradation state of a bearing cage based on digital-analog linkage according to claim 1 is characterized in that: Step S1 specifically includes: S11. Conduct full life cycle experiments on the bearing cage fatigue failure test bench; S12, applying radial load and axial load eccentric force on the test bench to accelerate bearing fatigue failure; S13. Collect the acceleration signal of the bearing during its entire life cycle and store it in a csv file.

3. The method for evaluating the degradation state of a bearing cage based on digital-analog linkage according to claim 1 is characterized in that: In step S12, the eccentric load force of the applied radial load is 5.1 kN, and the eccentric load force of the applied axial load is 1.4 kN.

4. The method for evaluating the degradation state of a bearing cage based on digital-analog linkage according to claim 1 is characterized in that: Step S2 specifically includes: S21. When the bearing cage breaks, the clearance between the cage pocket and the ball at the break will increase, so the ball will deviate from the original position. Therefore, an additional position angle θ is introduced at the offset ball. f , simulate the cage fracture failure as follows: Where θ1 is the initial position angle of ball No. 1 in bearing, ω r is the angular velocity of the inner ring, r b and R b Respectively represent the inner and outer ring raceway radius of the bearing, N b is the number of bearing balls; S22. Based on the Hertz contact theory, after considering the bearing seat clearance and cage fracture failure, the 5-DOF nonlinear bearing force at the i-th node of the rotor is expressed as follows: Among them, α j represents the pressure angle of the jth ball, the δth j represents the contact deformation between the jth ball and the raceway, H(δ j ) is the Heaviside function, k bj is the Hertzian contact stiffness of the jth ball; S23. Using the finite element method and the concentrated mass method, a system dynamics model is established, and the dynamics equation of the rotor-bearing system is established as follows: Among them, M RS , C RS , J RS , K RS are the mass matrix, damping matrix, gyro matrix and stiffness matrix of the rotating axis respectively; M SS and J SS are the mass matrix and gyro matrix of the bushing respectively; F b Represents the restoring force vector of the rolling bearing; F u Represents the rotor unbalance force vector; F r represents the external load vector; G represents the rotor gravity vector; S24, using the Newmark-β method, calculating the dynamic equation formula in step S23 to obtain a numerical solution, and extracting the vibration response result at the bearing seat; S25, by changing the additional position angle θ f To characterize the cage degradation process θ f ∈[0°,4°].

5. The method for evaluating the degradation state of a bearing cage based on digital-analog linkage according to claim 1 is characterized in that: Step S3 specifically includes: S31. By adversarial loss function L GAN , the generator and discriminator in the CycleGAN model are trained adversarially as follows: Among them, L GAN To combat losses, represents the expectation that the actual signal follows the actual physical domain distribution, It indicates the expectation that the simulation signal follows the simulation domain distribution; S32. Cycle consistency loss is introduced into the CycleGAN model to ensure the consistency of the generated signal and the target domain signal. The loss function is as follows: Among them, ||·||1 represents the L1 norm; S33. By minimizing the total loss function, the generator G and the discriminator F are optimized and fine-tuned as follows: L total =L GAN (G,D p )+L GAN (F,D q )+λL cyc (G,F) Among them, the coefficient λ controls the proportion of cycle consistency loss in the total loss; S34. The SE attention mechanism is introduced in the encoder and decoder to dynamically adjust the channel weights by learning the relationship between features, so that the model can focus on the features with the greatest contribution, thereby enhancing the feature learning effect of the model and reducing distortion in the signal generation process.

6. The method for evaluating the degradation state of a bearing cage based on digital-analog linkage according to claim 1 is characterized in that: Step S4 specifically includes: S41. Perform a simple linear transformation on the input signal and map the signal sequence to a high-dimensional embedding space as follows: Among them, dim represents the dimension of time series embedding, W embedding represents a learnable matrix; S42. A randomly initialized trainable embedding is added to the beginning of the embedded token sequence, aggregating the information from each part, as follows: S43. Using learnable positional encodings To extract position information more flexibly, add position encoding and tag embedding to obtain input embedding as follows: in, represents position embedding, Indicates Patch sequence; S44, multi-scale feature extraction is performed through SKnet, and the combination of the two more comprehensively models the semantic information of the fault signal, thereby improving the diagnostic accuracy and generalization performance of the model; S45, for input embedding, transform the input feature map through the convolution kernel; S46: The above feature graph is summed up to achieve feature fusion of different branches, and then the statistical features of the channel direction are generated by global average pooling operation. To embed global information, and then build a more compact feature through the fully connected layer To achieve model adaptive selection, as follows: U=U+U d=max(C / r,16) Among them, σ represents the ReLU activation function, represents batch normalization, r represents the reduction rate; S47, through the cross-channel soft attention mechanism, adaptively select information of different spatial scales, and obtain the final feature map by superimposing the attention weights on convolution kernels of different sizes as follows: a c +b c =1 in, a c and b c Respectively represent U c and U c The soft attention vector of S48, through linear transformation, a set of d output by SKnet is transformed model Feature map of dimension Convert them into d-dimensional key K, value V, and query Q respectively, and execute multi-head self-attention mechanism; S49. Pass the features into the multi-layer perceptron, obtain the probability value of each fault category through the Softmax function, and obtain the final fault category according to the maximum probability value.

7. The method for evaluating the degradation state of a bearing cage based on digital-analog linkage according to claim 1 is characterized in that: In step S45, the input feature map is transformed by convolution kernels of sizes 3×1 and 5×1, as follows: Among them, F * The (·) operation consists of grouped convolution, batch normalization, and ReLU activation function in sequence.

Images