Intelligent bearing fault diagnosis method based on finite element simulation and data enhancement

Abstract

The invention is applicable to the technical field of bearing fault diagnosis, and provides an intelligent bearing fault diagnosis method based on finite element simulation and data enhancement, which comprises the following steps of: establishing geometric models of normal and faulty bearing packs, and calculating theoretical characteristic frequencies of bearings; performing finite element dynamics simulation on the normal and faulty bearing packs; extracting simulation vibration signals of the normal bearing pack and the fault bearing pack; comparing the signal characteristic frequencies of the normal bearing pack and the fault bearing pack with a theoretical calculation frequency; the simulation signals are normalized and then mapped into a two-dimensional grey-scale map, and texture features in the grey-scale map are refined; establishing a generative adversarial network model to expand the grey-scale map sample; establishing a deep reinforcement learning model to screen a high-quality grey-scale image synthesis sample; and establishing a convolutional neural network to realize direct fault diagnosis of the synthetic sample. According to the invention, the problems of few bearing fault samples and poor diagnosis effect when the samples are unbalanced are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing fault diagnosis, and in particular relates to an intelligent bearing fault diagnosis method based on finite element simulation and data enhancement. Background Art

[0002] With the rapid development of big data and artificial intelligence, the manufacturing industry is accelerating its transformation towards digitalization and intelligence. As the "industrial mother machine" of modern manufacturing, CNC machine tools are widely used in aerospace, automobile manufacturing, and electronic products. The spindle bearing group is one of the key components in CNC machine tools. It mainly supports and transmits the rotational motion of the spindle; however, these bearings usually work in a high-load, high-speed environment, and are prone to wear and fatigue. Research has found that spindle bearing failures account for about 44% of the total failures of electric spindles. Its performance greatly affects the accuracy of the spindle. Once the rolling bearing fails, it may cause huge economic losses and casualties. Therefore, timely fault diagnosis of the bearing group is of great significance to ensure the continuous operation of the machine tool and improve production efficiency.

[0003] Traditional bearing fault diagnosis methods mainly rely on expert experience and simple signal processing technology, but these methods often have poor diagnostic effects when dealing with complex and changeable working conditions and large amounts of real-time data. As a cutting-edge technology in the field of artificial intelligence, deep learning has powerful data processing and pattern recognition capabilities. By building a multi-layer neural network, it can automatically extract deep features from data, significantly improving the recognition accuracy in many traditional recognition tasks. However, the fault diagnosis model based on deep learning requires sufficient training samples and balanced data sets to achieve ideal diagnostic accuracy. However, it is difficult to directly install sensors on certain key parts, resulting in the inability to obtain a large number of samples of various types of faults of the equipment under specific working conditions; in addition, each fault sample often cannot achieve the same data distribution due to factors such as variable working conditions, noisy non-stationary signals, and dynamic fault evolution, which further weakens the generalization ability of the intelligent diagnosis model. Therefore, the small sample and data imbalance problems seriously restrict the further development of bearing intelligent fault diagnosis technology. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a bearing intelligent fault diagnosis method based on finite element simulation and data enhancement, aiming to solve the problems raised in the above-mentioned background technology.

[0005] The embodiment of the present invention is implemented as follows: a bearing intelligent fault diagnosis method based on finite element simulation and data enhancement, comprising the following steps:

[0006] S1. Establish geometric models of normal and faulty bearing groups and calculate the theoretical characteristic frequencies of the bearings;

[0007] S2, perform finite element dynamics simulation on normal and faulty bearing groups;

[0008] S3, extracting simulated vibration signals of normal and faulty bearing groups;

[0009] S4, comparing the signal characteristic frequencies of the normal and faulty bearing groups with the theoretically calculated frequencies to verify the accuracy of this simulation;

[0010] S5, normalizing the simulation signal and mapping it into a two-dimensional grayscale image, and using a multi-scale local binary pattern method to refine the texture features in the grayscale image;

[0011] S6. Establish a generative adversarial network model to expand the grayscale image samples;

[0012] S7. Establish a deep reinforcement learning model to screen high-quality grayscale image synthesis samples;

[0013] S8. Establish a convolutional neural network to realize direct fault diagnosis of synthetic samples.

[0014] Preferably, in S1, the step of establishing geometric models of normal and faulty bearing groups comprises ten types of models, including one type of normal bearing group and nine types of faulty bearing groups, wherein the fault type is crack damage, the fault locations are the inner ring, the outer ring and the rolling element, and each location is further divided into three types of crack damage with different widths;

[0015] The step of calculating the theoretical characteristic frequency of the bearing, wherein the rolling bearing signal characteristics include two types: normal and fault, is specifically as follows:

[0016] For a normal bearing, when the outer ring is fixed, the inner ring speed frequency expression is:

[0017] f i =n / 60(1)

[0018] In formula (1), n ​​is the shaft speed;

[0019] Since the rolling element is constrained by the cage, the rotation frequency of the two is the same, and the expression is:

[0020]

[0021] In formula (2), D and d are the diameters of the outer and inner rings of the bearing, respectively; α is the initial contact angle;

[0022] When a bearing fails, it will cause vibration with a large amplitude and a specific fault frequency. The outer ring fault frequency expression is:

[0023]

[0024] In formula (3), D and d are the diameters of the outer and inner rings of the bearing, respectively, α is the initial contact angle, Z is the number of rolling elements, and n is the shaft speed;

[0025] The fault frequency expression of the inner ring is:

[0026]

[0027] In formula (4), D and d are the diameters of the outer and inner rings of the bearing, respectively, α is the initial contact angle, Z is the number of rolling elements, and n is the shaft speed;

[0028] The rolling element failure frequency expression is:

[0029]

[0030] In formula (5), D and d are the diameters of the outer and inner rings of the bearing, respectively, α is the initial contact angle, and n is the shaft speed.

[0031] Preferably, in S2, in the step of performing finite element dynamics simulation on normal and faulty bearing groups, the simulation analysis step module adopts implicit dynamics, the outer ring of the bearing is fixed, and the inner ring is rotated, and ten types of bearing groups are simulated in turn.

[0032] Preferably, in S3, the step of extracting simulated vibration signals of normal and faulty bearing groups is specifically: extracting displacement, velocity and acceleration vibration signals of ten types of bearing groups in the vertical direction, and using the acceleration signal as the input of the generative adversarial network model.

[0033] Preferably, in S5, the step of normalizing the simulation signal and mapping it into a two-dimensional grayscale image, the normalized expression is:

[0034]

[0035] In formula (8), x i is the original vibration signal, is the normalized c-th signal, x min and x max are the minimum and maximum values ​​of the signal segment respectively;

[0036] Then convert the normalized signal into a two-dimensional grayscale image of size 64×64;

[0037] The step of using the multi-scale local binary pattern method to refine the texture features in the grayscale image is specifically: using the multi-scale LBP method to further extract local texture features at different scales in the grayscale image, and its expression is:

[0038]

[0039] In formula (9), s(·) is the sign function, P is the number of neighborhood pixels, R is the neighborhood radius, and I c is the gray value of the center pixel, I i is the grayscale value of the i-th pixel in the neighborhood.

[0040] Preferably, in S6, the step of establishing a generative adversarial network model to expand the grayscale image samples, the generative adversarial network includes a generator G and a discriminator D, the preprocessed multi-scale grayscale image is input, G and D are continuously iteratively optimized, and the final model reaches Nash equilibrium, and the total loss function expression of the network is:

[0041]

[0042] In formula (10), E(·) is the mean function, G is the generator, D is the discriminator, and p noise (z) is the noise probability distribution, z is the noise, p data (x) is the real data distribution, x is the real data sample;

[0043] Among them, G captures the characteristic distribution of multi-scale grayscale image samples and continuously optimizes internal parameters to train and generate samples that are as realistic as possible. The loss function expression of G is:

[0044]

[0045] In formula (11), E(·) is the mean function, G is the generator, D is the discriminator, and p z (z) is the noise probability distribution, z is the noise;

[0046] D is a binary classifier that estimates the probability that the sample synthesized by G comes from the training set. Its loss function expression is:

[0047]

[0048] In formula (12), E(·) is the mean function, G is the generator, D is the discriminator, and p data (x) is the real data distribution, x is the real data sample, p z (z) is the noise probability distribution, and z is the noise.

[0049] Preferably, in S7, the step of establishing a deep reinforcement learning model to screen high-quality grayscale image synthesis samples uses a deep policy gradient algorithm to optimize the policy function through gradient ascent, and the loss function expression of the policy gradient algorithm is:

[0050]

[0051] In formula (13), s is the state, a is tis action, π θ (a|s) is the policy function, θ is the parameter in the policy function, τ is the change trajectory of the policy function, and r(τ) is the reward;

[0052] Specifically: the environment is in state s t The state value under is calculated as V t , the agent chooses to perform an action a through the policy function π t ; Then, the environment scores the action and gives the action value Q t , the reward of the environment for the current action is r t =Q t -V t , optimize the parameters θ in the policy function π according to the reward; finally, the environment accepts the new state s t+1 Repeat the above process until the end.

[0053] Preferably, in S8, the step of establishing a convolutional neural network to implement direct fault diagnosis of synthetic samples includes a feature extractor and a classifier. The feature extractor implements feature extraction and refinement through a series of stacked convolution blocks. The classifier uses global average pooling to further extract global features of the feature map and inputs them into the fully connected layer to complete the final classification. The loss function of the classifier selects the cross entropy function, which is expressed as:

[0054]

[0055] In formula (14), p(x i ) is the true probability distribution, q(x i ) is the predicted probability distribution and n is the number of samples.

[0056] An embodiment of the present invention provides an intelligent bearing fault diagnosis method based on finite element simulation and data enhancement. First, through finite element dynamics simulation, simulated vibration signals of normal and faulty bearing groups are extracted, and the characteristic frequencies in the signals are compared with the theoretical frequencies to verify the accuracy of the simulation; secondly, the various extracted simulation signals are normalized and converted into a two-dimensional grayscale image, and a multi-scale local binary pattern method is used to further extract multi-scale texture features in the grayscale image, and a generative adversarial network model is established to achieve the expansion of grayscale image samples; finally, a deep reinforcement learning model is established to screen high-quality grayscale image synthetic samples, and a two-dimensional convolutional neural network is used to perform "end-to-end" fault diagnosis on the synthetic samples. This method aims to solve the problem of poor diagnostic effect when there are few bearing fault samples and unbalanced samples, and use the diagnostic knowledge obtained in a laboratory environment to guide and help solve engineering task goals. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1A flow chart of a bearing intelligent fault diagnosis method based on finite element simulation and data enhancement provided by an embodiment of the present invention;

[0058] Figure 2 A structural diagram of a finite element model of a faulty bearing provided in an embodiment of the present invention;

[0059] Figure 3 Simulated vibration signals of normal and outer ring fault (0.1778mm) bearing groups provided by the embodiments of the present invention;

[0060] Figure 4 The envelope spectrum of the vibration signal of the bearing group with outer ring fault (0.1778mm) provided in the embodiment of the present invention;

[0061] Figure 5 A diagram of signal preprocessing results provided by an embodiment of the present invention;

[0062] Figure 6 A diagram of the structure of a generative adversarial network model provided by an embodiment of the present invention;

[0063] Figure 7 A structural diagram of a deep reinforcement learning model provided by an embodiment of the present invention;

[0064] Figure 8 A schematic diagram for comparing diagnostic results under different numbers of original samples provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0066] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0067] like Figure 1 As shown, a flowchart of a bearing intelligent fault diagnosis method based on finite element simulation and data enhancement is provided in one embodiment of the present invention, comprising the following steps:

[0068] S1. Establish geometric models of normal and faulty bearing groups, and calculate the theoretical characteristic frequencies of the bearings. Specifically, the bearing model used in the embodiment of the present invention is 7014AC, and the specific parameters are shown in Table 1:

[0069] Table 1

[0070]

[0071]

[0072] For a normal bearing, when the outer ring is fixed, the inner ring speed frequency expression is:

[0073] f i =n / 60(1)

[0074] In formula (1), n ​​is the shaft speed;

[0075] Since the rolling element is constrained by the cage, the rotation frequency of the two is the same, and the expression is:

[0076]

[0077] In formula (2), D and d are the diameters of the outer and inner rings of the bearing, respectively; α is the initial contact angle;

[0078] When a bearing fails, it will cause vibration with a large amplitude and a specific fault frequency. The outer ring fault frequency expression is:

[0079]

[0080] In formula (3), D and d are the diameters of the outer and inner rings of the bearing, respectively, α is the initial contact angle, Z is the number of rolling elements, and n is the shaft speed;

[0081] The fault frequency expression of the inner ring is:

[0082]

[0083] In formula (4), D and d are the diameters of the outer and inner rings of the bearing, respectively, α is the initial contact angle, Z is the number of rolling elements, and n is the shaft speed;

[0084] The rolling element failure frequency expression is:

[0085]

[0086] In formula (5), D and d are the diameters of the outer and inner rings of the bearing, respectively, α is the initial contact angle, and n is the shaft speed;

[0087] When the inner ring speed is 1800r / min, the theoretical frequencies calculated by equations (1)-(3) are shown in Table 2:

[0088] Table 2

[0089] type Inner ring speed Cage rotation frequency Rolling element rotation frequency Outer race fault frequency Theoretical value / Hz 30 11.83 11.83 141.94

[0090] S2. Perform finite element dynamics simulation on normal and faulty bearing groups. The finite element model of the faulty bearing group is as follows: Figure 2 The fault information is shown in Table 3:

[0091] Table 3

[0092]

[0093] The specific process of this step is as follows: In the simulation, the material properties of each component are the same, and the density is 7800kg / m 3 , Young's modulus is 210GPa, Poisson's ratio is 0.3; hexahedral units are used as the dominant meshing unit; implicit dynamics method is used in Step module, simulation time is set to 0.5s, of which 0-0.0001s is used to load to rated speed and load; in Load module, the outer ring of the bearing is fixed, the rotation direction of the inner ring is set to 1800r / min, and a load of 200N is applied to the Z direction of the inner ring surface, and the cage has only the degree of freedom in the rotation direction; in the contact properties between the components, the "surface to surface" contact mode is adopted, when the inner and outer ring raceways contact the balls, the raceways are the main surfaces, the contact type is set to friction, the friction coefficient is 0.12, and the friction formula is set to "penalty"; when the balls contact the cages, the ball surfaces are the main surfaces, and the sliding mode is limited slip;

[0094] S3, extract the simulated vibration signals of the normal and faulty bearing groups, the vibration signals of the normal and outer ring faulty bearing groups are as follows: Figure 3 As shown;

[0095] S4. Compare the signal characteristic frequencies of the normal and faulty bearing groups with the theoretically calculated frequencies to verify the accuracy of this simulation. Figure 4 As shown;

[0096] The specific process of this step is as follows: for a normal bearing group, the change period of the inner ring displacement and velocity simulation signal is 0.035s, and the change frequency is 28.57Hz, which is close to the theoretical rotation frequency of the inner ring of 30Hz, and the error is 4.77%; the rolling element is constrained by the cage during actual operation, and the signal change period of both is 0.084s, and the change frequency is 11.90Hz, which is 3.11% different from the theoretical rotation frequency of 11.53Hz; for the outer ring fault bearing group, firstly, the velocity and acceleration signals of a point in the fault area are Hilbert transformed, and then the obtained envelope signal is Fourier transformed to obtain the envelope spectrum;

[0097] The Hilbert transform expression is:

[0098]

[0099] In formula (6), x(t) is a real-valued signal, * represents the convolution operation, is the Hilbert transform kernel;

[0100] The Fourier transform expression is:

[0101]

[0102] In equation (7), f(t) and F(ω) are time-domain and frequency-domain signals, ω is the angular frequency, and e -iωt is a complex exponential function;

[0103] Depend on Figure 4 It can be seen that the frequency of 138 Hz on the envelope spectrum appears in multiples, and the error with the theoretical outer race fault frequency of 141.94 Hz is only 2.78%. Therefore, the error between the characteristic frequency of the simulation signal and the theoretical frequency is kept within 5%, which proves the reliability of the simulation method in the embodiment of the present invention.

[0104] S5, normalize various acceleration signals and map them into two-dimensional grayscale images, and use the multi-scale local binary pattern (MSLBP) method to refine the texture features in the grayscale image, such as Figure 5 As shown;

[0105] The specific process of this step is: each type of acceleration signal is normalized to map to a uniform scale to improve diagnostic efficiency and avoid overfitting. The expression is:

[0106]

[0107] In formula (8), x i is the original vibration signal, is the normalized signal, x min and x max are the minimum and maximum values ​​of the signal in this segment, respectively. Therefore, all signals are normalized to the interval [0,1]. Then the signal value corresponding to each time point is mapped to grayscale to convert it into a two-dimensional grayscale image of size 64×64. Finally, the multi-scale LBP method is used to further extract local texture features at different scales in the grayscale image to improve the robustness of the model. The expression of multi-scale LBP is:

[0108]

[0109] In formula (9), s(·) is the sign function, P is the number of neighborhood pixels, R is the neighborhood radius, and I c is the gray value of the center pixel, I i is the gray value of the i-th pixel in the neighborhood;

[0110] The scale (P, R) is selected as (2, 1), (4, 1) and (4, 2). Each pixel in the grayscale image is compared with its neighbors at these three scales, and LBP features are extracted and further combined, which greatly enhances the texture description ability;

[0111] S6. Establish a generative adversarial network (GAN) model to expand the grayscale image samples. The structure of the generative adversarial network model is as follows: Figure 6 As shown;

[0112] The specific process of this step is as follows: In the established generative adversarial network, the generator and the discriminator are both two-dimensional convolutional neural networks. The two generate new grayscale image samples through adversarial means, continuously optimizing the internal parameters of the two, reducing the total loss of the network, and making the new samples as similar as possible to the training set samples in distribution. The number of network training times is set to 1000, the number of expanded samples for each category is 500, and the total loss function of the network is:

[0113]

[0114] In formula (10), E(·) is the mean function, G is the generator, D is the discriminator, and p noise (z) is the noise probability distribution, z is the noise, p data (x) is the real data distribution, x is the real data sample;

[0115] The generator G starts from a random noise vector and gradually forms a two-dimensional realistic grayscale image sample through a series of convolution operations. The loss function expression of the generator G is:

[0116]

[0117] In formula (11), E(·) is the mean function, G is the generator, D is the discriminator, and p z (z) is the noise probability distribution, z is the noise;

[0118] The discriminator D performs convolution, activation, and pooling on the synthetic grayscale image from the generator, and outputs the probability that the synthetic sample comes from the training set. Its loss function expression is:

[0119]

[0120] In formula (12), E(·) is the mean function, G is the generator, D is the discriminator, and p data (x) is the real data distribution, x is the real data sample, p z (z) is the noise probability distribution, and z is the noise.

[0121] S7. Establish a deep reinforcement learning (DRL) model to screen high-quality grayscale image synthesis samples, and use the deep policy gradient algorithm to optimize the policy function through gradient ascent. The structure of the deep reinforcement learning model is as follows: Figure 7 As shown in Figure 1, it consists of an agent, state, reward, policy function, and value function. The loss function expression of the policy gradient algorithm is:

[0122]

[0123] In formula (13), s is the state, a is the action, and π θ (a|s) is the policy function, θ is the parameter in the policy function, τ is the change trajectory of the policy function, and r(τ) is the reward;

[0124] The specific process of this step is as follows: after expansion, the entire synthetic sample is divided into 10 groups, each group consists of 50 samples from each class, a total of 500 samples, and each group of samples is regarded as a state; then, the state s at time t t After being input into a feature extractor (agent), it is mapped into action probability P and current state value V respectively through two fully connected layers. t ; In the action probability, 0 means discarding the sample, and 1 means retaining the sample; after executing the action, the retained samples are divided into a training set and a validation set in a ratio of 3:1. The training set is used to train the support vector machine, and the validation set is used to obtain the action value Q t ; Then, r t+1 =Q t -V t is regarded as the agent's reward for this action, and the agent continues to accept the next state s t+1 Perform sample screening; the number of samples retained for each type is 300. Finally, 5000 grayscale image samples are screened, and 3000 high-quality synthetic samples are obtained;

[0125] S8. A two-dimensional convolutional neural network (CNN) model is established to realize direct fault diagnosis of synthetic grayscale image samples. In the embodiment of the present invention, the number of original grayscale image samples is set to 10, 20, 50 and 100 respectively, and then sample expansion and high-quality screening are performed. Finally, the average diagnostic accuracy of the model is obtained as shown in Table 4. The diagnostic visualization results in the four cases are shown in Figure 8 As shown:

[0126] Table 4

[0127] Original sample size 10 20 50 100 Average diagnostic rate / % 89.47 91.03 96.58 99.25

[0128] The specific process of this step is as follows: the established two-dimensional convolutional neural network contains 4 convolution blocks, a Dropout layer, two fully connected layers (FC1 and FC2) and an output layer; the division ratio of the training set, validation set and test set is 3:1:1, the model uses the Adam optimizer, the learning rate is 0.001, the batch size is 64, the dropout is 0.5, and the number of network iterations is 50; the loss function of the Softmax classifier selects the cross entropy function, and the expression is:

[0129]

[0130] In formula (14), p(xi ) is the true probability distribution, q(x i ) is the predicted probability distribution and n is the number of samples.

[0131] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A bearing intelligent fault diagnosis method based on finite element simulation and data enhancement, characterized in that: The following steps are involved: S1. Establish geometric models of normal and faulty bearing groups and calculate the theoretical characteristic frequencies of the bearings; S2, perform finite element dynamics simulation on normal and faulty bearing groups; S3, extracting simulated vibration signals of normal and faulty bearing groups; S4, comparing the signal characteristic frequencies of the normal and faulty bearing groups with the theoretically calculated frequencies to verify the accuracy of this simulation; S5, normalizing the simulation signal and mapping it into a two-dimensional grayscale image, and using a multi-scale local binary pattern method to refine the texture features in the grayscale image; S6. Establish a generative adversarial network model to expand the grayscale image samples; S7. Establish a deep reinforcement learning model to screen high-quality grayscale image synthesis samples; S8. Establish a convolutional neural network to realize direct fault diagnosis of synthetic samples.

2. The bearing intelligent fault diagnosis method based on finite element simulation and data enhancement according to claim 1 is characterized in that: In S1, the step of establishing geometric models of normal and faulty bearing groups, the models are divided into ten categories, including one category of normal bearing groups and nine categories of faulty bearing groups, wherein the fault type is crack damage, the fault locations are the inner ring, the outer ring and the rolling element, and each location is divided into three types of crack damage with different widths; The step of calculating the theoretical characteristic frequency of the bearing, wherein the rolling bearing signal characteristics include two types: normal and fault, is specifically as follows: For a normal bearing, when the outer ring is fixed, the inner ring speed frequency expression is: f i =n / 60(1) In formula (1), n ​​is the shaft speed; Since the rolling element is constrained by the cage, the rotation frequency of the two is the same, and the expression is: In formula (2), D and d are the diameters of the outer and inner rings of the bearing, respectively; α is the initial contact angle; When a bearing fails, it will cause vibration with a large amplitude and a specific fault frequency. The outer ring fault frequency expression is: In formula (3), D and d are the diameters of the outer and inner rings of the bearing, respectively, α is the initial contact angle, Z is the number of rolling elements, and n is the shaft speed; The fault frequency expression of the inner ring is: In formula (4), D and d are the diameters of the outer and inner rings of the bearing, respectively, α is the initial contact angle, Z is the number of rolling elements, and n is the shaft speed; The rolling element failure frequency expression is: In formula (5), D and d are the diameters of the outer and inner rings of the bearing, respectively, α is the initial contact angle, and n is the shaft speed.

3. The bearing intelligent fault diagnosis method based on finite element simulation and data enhancement according to claim 2 is characterized in that: In S2, the steps of performing finite element dynamics simulation on normal and faulty bearing groups, the simulation analysis step module adopts implicit dynamics, the outer ring of the bearing is fixed, and the inner ring is rotated, and ten types of bearing groups are simulated in turn.

4. The bearing intelligent fault diagnosis method based on finite element simulation and data enhancement according to claim 2 is characterized in that: In S3, the step of extracting simulated vibration signals of normal and faulty bearing groups is specifically: extracting displacement, velocity and acceleration vibration signals of ten types of bearing groups in the vertical direction, and using the acceleration signals as input of the generative adversarial network model.

5. The bearing intelligent fault diagnosis method based on finite element simulation and data enhancement according to claim 2 is characterized in that: In S4, the step of comparing the signal characteristic frequencies of the normal and faulty bearing groups with the theoretical calculated frequencies is specifically: analyzing the simulated vibration signals of the components of the normal and faulty bearing groups, and comparing the displacement and velocity signal change frequencies of the normal bearing group with the theoretical rotation frequency; performing envelope spectrum analysis on the velocity and acceleration signals of the outer ring faulty bearing group, and comparing the characteristic frequencies with larger amplitudes on the spectrum with the theoretical fault frequencies.

6. The bearing intelligent fault diagnosis method based on finite element simulation and data enhancement according to claim 1 is characterized in that: In S5, the step of normalizing the simulation signal and mapping it into a two-dimensional grayscale image, the normalized expression is: In formula (8), x i is the original vibration signal, is the normalized c-th signal, x min and x max are the minimum and maximum values ​​of the signal segment respectively; Then convert the normalized signal into a two-dimensional grayscale image of size 64×64; The step of using the multi-scale local binary pattern method to refine the texture features in the grayscale image is specifically: using the multi-scale LBP method to further extract local texture features at different scales in the grayscale image, and its expression is: In formula (9), s(·) is the sign function, P is the number of neighborhood pixels, R is the neighborhood radius, and I c is the gray value of the center pixel, I i is the gray value of the i-th pixel in the neighborhood.

7. The bearing intelligent fault diagnosis method based on finite element simulation and data enhancement according to claim 1 is characterized in that: In S6, the step of establishing a generative adversarial network model to expand the grayscale image samples, the generative adversarial network includes a generator G and a discriminator D, the preprocessed multi-scale grayscale image is input, G and D are continuously iterated and optimized, and the final model reaches Nash equilibrium. The total loss function expression of the network is: In formula (10), E(·) is the mean function, G is the generator, D is the discriminator, and p noise (z) is the noise probability distribution, z is the noise, p data (x) is the real data distribution, x is the real data sample; Among them, G captures the characteristic distribution of multi-scale grayscale image samples and continuously optimizes internal parameters to train and generate samples that are as realistic as possible. The loss function expression of G is: In formula (11), E(·) is the mean function, G is the generator, D is the discriminator, and p z (z) is the noise probability distribution, z is the noise; D is a binary classifier that estimates the probability that the sample synthesized by G comes from the training set. Its loss function expression is: In formula (12), E(·) is the mean function, G is the generator, D is the discriminator, and p data (x) is the real data distribution, x is the real data sample, p z (z) is the noise probability distribution, and z is the noise.

8. The bearing intelligent fault diagnosis method based on finite element simulation and data enhancement according to claim 1 is characterized in that: In S7, the step of establishing a deep reinforcement learning model to screen high-quality grayscale image synthesis samples uses a deep policy gradient algorithm to optimize the policy function through gradient ascent. The loss function expression of the policy gradient algorithm is: In formula (13), s is the state, a is the action, and π θ (a|s) is the policy function, θ is the parameter in the policy function, τ is the change trajectory of the policy function, and r(τ) is the reward; Specifically: the environment is in state s t The state value under is calculated as V t , the agent chooses to perform an action a through the policy function π t ; Then, the environment scores the action and gives the action value Q t , the reward of the environment for the current action is r t =Q t -V t , optimize the parameters θ in the policy function π according to the reward; finally, the environment accepts the new state s t+1 Repeat the above process until the end.

9. The bearing intelligent fault diagnosis method based on finite element simulation and data enhancement according to claim 1 is characterized in that: In S8, the step of establishing a convolutional neural network to realize direct fault diagnosis of synthetic samples is performed. The convolutional neural network includes a feature extractor and a classifier. The feature extractor realizes feature extraction and refinement through a series of stacked convolution blocks. The classifier uses global average pooling to further extract global features of the feature map and inputs them into the fully connected layer to complete the final classification. The loss function of the classifier selects the cross entropy function, which is expressed as: In formula (14), p(x i ) is the true probability distribution, q(x i ) is the predicted probability distribution and n is the number of samples.

Images