Bearing life prediction method based on data feature optimization and deep learning

Abstract

The invention discloses a bearing life prediction method based on data feature optimization and deep learning, and the method comprises the steps: collecting vibration signals of a bearing in different states, carrying out the processing of the signals through MEMD-DisEn, and decomposing the signals into an intrinsic mode function, so as to reveal the features of a bearing fault; expanding fault sample data by using a WGAN-GP strategy so as to improve the processing capability of the model on abnormal samples and enhance the prediction accuracy; a TimeKAN model is applied to analyze the enhanced data, large-scale time series data are processed, a long-term dependency relationship in the data is captured, and the remaining service life of the bearing is accurately predicted; an escape optimization algorithm is improved, and hyper-parameters of the TimeKAN model are optimized through the improved escape optimization algorithm, so that the distinguishing capability and the response speed of the model are improved. The method can effectively improve the prediction accuracy and robustness, and is of great significance to the maintenance and safety production of mechanical equipment.

Description

Technical Field

[0001] The present invention relates to a method for predicting the bearing life, and particularly to a method for predicting the bearing life based on data feature optimization and deep learning. Background Art

[0002] As a key component of mechanical equipment, the performance and life of rolling bearings are crucial for the stable operation of the equipment. Predicting its remaining service life is of great significance for the safe production and maintenance of mechanical equipment. By predicting the remaining life of the bearing, the operation and maintenance of the equipment can be effectively guided, the use safety of the equipment can be ensured, and accidents can be avoided. In addition, accurately predicting the bearing life can reasonably arrange its performance inspection and maintenance replacement work, avoid premature shutdown affecting the production process or late shutdown causing equipment damage, thereby reducing economic losses.

[0003] With the improvement of industrial automation and intelligence levels, higher requirements are put forward for the reliability and maintenance cost-effectiveness of bearings. Life prediction technology can help detect and warn problems existing in the operation of bearings in a timely manner, reduce downtime, and avoid economic losses caused by major accidents. The bearings in operation involve multiple devices and complex physical processes, generating a large amount of data. How to extract useful information from these data and establish an accurate life prediction model is a technical challenge. Machine learning technologies, especially deep learning, data decomposition, and feature extraction technologies, provide the possibility for processing complex data sets and establishing efficient fault diagnosis. In order to improve the accuracy of fault diagnosis, it is necessary to improve and optimize the existing machine learning algorithms to better adapt to specific application scenarios.

[0004] Currently, the solutions to this problem can be mainly divided into three categories, namely physical methods, statistical methods, and data-driven methods. Physical methods mainly collect bearing operation data to construct a mathematical model, and use the mode of combining historical data of mechanical equipment with failure degradation formulas to predict the remaining service life of the equipment. Physical methods require a deep understanding of the physical process, and the construction of the model and the determination of parameters are relatively complex, with high requirements for the quality and quantity of data. Statistical methods calculate the statistical features of the original vibration signal, such as root mean square, kurtosis, peak-to-peak value, etc., and construct some new statistical features to predict the remaining service life of the bearing. However, it may not be able to capture the complex non-linear features of bearing degradation, resulting in limited prediction accuracy. In contrast, data-driven methods mainly use machine learning or deep learning techniques to extract trend features from the original signal to characterize the process of bearing performance degradation, and construct a prediction model, specifically including recursive neural networks, convolutional networks, graph convolutional networks, Transformers, and methods based on transfer learning. These methods cover multiple fields such as machine learning and deep learning, showing great potential in data processing and feature extraction. They can automatically learn and extract useful features from a large amount of data, effectively reduce the dependence on expert experience, have strong applicability, and provide new ideas and technical means for bearing life prediction. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to provide a bearing life prediction method based on data feature optimization and deep learning, which improves the accuracy and prediction efficiency of the bearing service life prediction results.

[0006] Technical Solution: A bearing life prediction method based on data feature optimization and deep learning according to the present invention includes the following steps:

[0007] (1) Use sensors to collect vibration signals of the bearing in healthy and faulty states at different positions.

[0008] (2) Reconstruct the collected vibration signals using the MEMD-DisEn method.

[0009] (3) Use the WGAN-GP data augmentation strategy to augment the bearing vibration data obtained after reconstruction in step (2) and expand the faulty sample data.

[0010] (4) Construct a bearing life prediction model based on TimeKAN, introduce the Tent initialization and Gauss perturbation strategies into the escape optimization algorithm, optimize and improve the initialization stage and population search stage of the escape optimization algorithm, and obtain the improved escape optimization algorithm.

[0011] (5) Optimize the hyperparameters of the TimeKAN model using the improved escape optimization algorithm;

[0012] (6) Train the bearing life prediction model using the bearing vibration data obtained in step (3) and the improved escape optimization algorithm, and perform online fault analysis and life prediction on the real-time collected bearing vibration data using the trained and optimized bearing life prediction model to obtain the bearing remaining life result.

[0013] Preferably, the bearing vibration signal in step (1) is an n-dimensional variable sequence, denoted as where T is the sequence length.

[0014] Preferably, the reconstruction in step (2) is to decompose the n-dimensional bearing vibration signal variable sequence collected in step (1) using the multivariate empirical mode decomposition method into p IMF components and a residual term, and reconstruct the p IMF components obtained by decomposing each sequence using the DisEn entropy.

[0015] Preferably, the data augmentation in step (3) is to train the WGAN-GP model, generate synthetic samples on the original dataset, merge the synthetic samples with the original dataset to create a balanced augmented dataset, and introduce gradient penalty to enforce the K-Lipschitz continuity constraint of the Wasserstein distance to improve the training stability and generate high-quality data augmentation samples.

[0016] Preferably, the goal of the WGAN-GP is to minimize the Wasserstein distance between the generator and the discriminator. During the training process of the WGAN-GP model, the minimization of the generator and the maximization of the discriminator are alternately trained. The goal of the discriminator is to maximize the score for real samples and minimize the score for generated samples; the goal of the generator is to generate samples that can deceive the discriminator, and a gradient penalty term is added to the loss function of the discriminator to satisfy the K-Lipschitz constraint.

[0017] Preferably, step (4) uses the Tent map to generate a chaotic sequence to initialize the population, and the formula is as follows:

[0018]

[0019] where is the chaotic sequence, k is the number of populations, I is the current iteration number, and to maintain the randomness of the algorithm initialization information, the value of u is

[0020] Secondly, the Gauss map is used to generate a more uniformly distributed population.

[0021] Preferably, the hyperparameters in step (5) are the number of network layers Nn, the number of neurons Hu, and the learning rate Lr, and the specific steps are as follows:

[0022] (51) Use the root mean square error index between the life prediction result and the true result as the objective function of the TimeKAN life prediction model;

[0023] (52) Determine that the population type is the learning rate and batch size of the TimeKAN model, which are specifically expressed as follows:

[0024]

[0025] Among them, X represents the population of the non-monopoly search algorithm, and d represents the number of particles in one kind of the population;

[0026] (53) Determine the upper and lower limits of the search space based on the range of the learning rate and batch size, which are specifically expressed as follows:

[0027]

[0028] Preferably, step (6) is to train the bearing life prediction model with the bearing vibration time series data processed in step (3) and the improved escape optimization algorithm, so as to obtain a trained bearing life prediction model. After the various types of sensor data collected are processed in step (3), they are input into the trained bearing life prediction model to predict the life of the bearing vibration data and obtain an accurate remaining life result.

[0029] A computer device includes one or more processors, a memory, and one or more programs, where the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the program is executed by the processor, it implements the steps of the above-mentioned bearing life prediction method based on data feature optimization and deep learning.

[0030] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps of the above-mentioned bearing life prediction method based on data feature optimization and deep learning.

[0031] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: By collecting vibration signals of the bearing at different positions and directions through sensors, the health status and fault characteristics of the bearing can be captured more comprehensively, providing a richer data source for diagnosis; Using the MEMD-DisEn data reconstruction method to process the original bearing signal, the original signal is reconstructed into a series of intrinsic mode functions to further reveal the internal characteristics of bearing faults; Adopting the WGAN-GP data augmentation strategy to expand the reconstructed limited fault sample data to solve the problem of limited abnormal sample quantity in actual working conditions, improving the generalization ability of the model and the accuracy of life prediction; Aiming at the problems existing in the escape optimization algorithm during the optimization process, such as slow convergence speed and easy to fall into local optimum, the Tent initialization and Gauss perturbation strategies are added to the population update process of the escape optimization algorithm to help the algorithm jump out of the local optimum, thus constructing the TGESC algorithm, enhancing the search ability of the original escape optimization algorithm and improving the search efficiency; Aiming at the problems existing in traditional bearing life prediction methods, such as low operating efficiency and large demand for input data, the present invention adopts the TimeKAN model to effectively simplify the complexity of the model and improve the efficiency; And using TGESC to optimize the hyperparameters in the TimeKAN model, constructing the MEMD-DisEn-WGAN-GP-TimeKAN-TGESC model, improving the performance and accuracy of the bearing life prediction model. Description of the Drawings

[0032] Figure 1 This is the structure diagram of the WGAN-GP model of the present invention.

[0033] Figure 2 This is the schematic diagram of the structure of the TimeKAN model of the present invention.

[0034] Figure 3 This is the flow chart of the TGESC algorithm provided by the present invention. Detailed Embodiments

[0035] The technical solution of the present invention will be further described below with reference to the drawings.

[0036] The present invention provides a bearing life prediction method based on data feature optimization and deep learning, including the following steps:

[0037] (1) Using sensors to collect vibration signals of the bearing in healthy and faulty states at different positions;

[0038] (2) Reconstructing the collected vibration signals using the MEMD-DisEn method;

[0039] The reconstruction includes:

[0040] (21) Decompose the collected historical bearing vibration signals using the Multivariate Empirical Mode Decomposition (MEMD) method. The specific decomposition process is as follows:

[0041] (211) Denote the initially collected n-dimensional bearing vibration signal variable sequence as The sequence length is …, and a spherical coordinate system is established in the multi-dimensional space. G direction vectors are drawn from the center of the space to the points on the hypersphere representing the corresponding direction angles in the (n - 1)-dimensional unit sphere The set of direction vectors of, where G represents choosing G uniformly distributed point sets on the (n - 1)-dimensional sphere;

[0042] (212) Calculate along the direction vector on the hypersphere mapping of

[0043] (213) Determine the moments corresponding to when all the mapped signals reach their extreme values, denoted as

[0044] (214) Interpolate each signal extreme point using a multivariate spline interpolation function to obtain G multivariate envelopes

[0045] (215) Calculate the mean a(t) of the multivariate envelopes:

[0046]

[0047] (216) Calculate the intrinsic mode function f(t) = z(t) - a(t), and determine whether f(t) satisfies the judgment conditions for multivariate IMF. If f(t) satisfies the judgment conditions for multivariate IMF, then define f(t) as the first-order IMF component, and use a(t) as the new input signal in step (212). Repeat the iterative process from step (212) to step (216) to obtain a new multivariate IMF component f(t); if f(t) does not satisfy the judgment conditions for multivariate IMF, then use f(t) as the new input signal in step (212), and continue to repeat step (212) to step (216) until the final conditions are met.

[0048] After the above decomposition, each sequence in the initial n-dimensional sequence i z(t) (i = 1, 2..., n) is decomposed into the corresponding p IMF components imf ij (t) = {imf i1 (t), imf i2 (t),..., imf ip(t) and a residual term r i (t), and the number of IMF subsequences generated by the decomposition of each sequence is equal, and the frequency distribution and fluctuation modes are similar. The specific formula is as follows:

[0049]

[0050] Among them, z(t) represents the component of the n-dimensional original bearing vibration sequence after decomposition, n represents the dimension of the original bearing vibration sequence, and p represents the number of generated IMFs.

[0051] (22) Reconstruct the p IMF components imf ij (t) = {imf i1 (t), imf i2 (t),..., imf ip (t)} obtained by decomposing each sequence using DisEn entropy. The specific process is as follows:

[0052] (221) Use the normal cumulative distribution function (NCDF) to map the elements imf ij (t) in the sequence imf is (t) (s = 1, 2,..., p) to the sequence Y = {y1, y2,..., yp}, y s ∈(0, 1); Use a linear algorithm to assign y s as an integer from 1 to w, and obtain a new sequence through where w is the number of categories, represents the discretized sequence.

[0053] (222) Establish the embedding vector Map each to the dispersion pattern where d i+(o-1)l = z o-1 , and the number of dispersion patterns assigned to each is w o , w represents the time delay, o represents the embedding dimension, and l represents the time delay, that is, when constructing the phase space, the time interval between adjacent points.

[0054] (223) Calculate the relative frequency of each potential dispersion pattern. The specific formula is as follows:

[0055]

[0056] Among them, represents the dispersion pattern, represents the relative frequency of the dispersion pattern.

[0057] (224) Calculate the dispersion entropy of the IMF sequence ij of (t), and the formula is as follows:

[0058]

[0059] DisEn(·) represents the dispersion entropy.

[0060] (225) Calculate the dispersion entropy DisEn of each decomposition mode i (i = 1, 2, …, o), and combine to construct the fault feature vector V of the bearing vibration signal = {DisEn1, DisEn2, …, DisEn o}}.

[0061] (3) Use the WGAN-GP data augmentation strategy to augment the bearing vibration data obtained after reconstruction in step (2);

[0062] As Figure 1 shown, by training the WGAN-GP model, generate synthetic samples on the original dataset, merge the synthetic samples with the original dataset, create a balanced augmented dataset, introduce gradient penalty to enforce the K-Lipschitz continuity constraint of the Wasserstein distance, improve the stability of training and generate higher-quality data augmentation samples, and the process is as follows:

[0063] (31) The goal of WGAN-GP is to minimize the Wasserstein distance between the generator (G) and maximize the discriminator (D). Its objective function is expressed as:

[0064]

[0065] where p data is the distribution of real bearing vibration data, p z is the noise distribution, G(z) is the fake sample generated by the generator, D(x) is the score of the discriminator for the sample, and GP is the gradient penalty term.

[0066] (32) WGAN-GP enforces the K-Lipschitz continuity constraint through gradient penalty. The calculation formula of the gradient penalty is:

[0067]

[0068] where is the sample obtained by interpolating p data and p z , γ is the gradient penalty coefficient, is the gradient of the discriminator D for .

[0069] During the training process of the WGAN-GP model, the generator and the discriminator are trained alternately. The goal of the discriminator is to maximize its score for real samples and minimize its score for generated samples. The goal of the generator is to generate samples that can deceive the discriminator. At the same time, a gradient penalty term is added to the loss function of the discriminator to ensure that it satisfies the K-Lipschitz constraint.

[0070] (4) Construct a bearing life prediction model based on TimeKAN, introduce the Tent initialization and Gauss perturbation strategy into the escape optimization algorithm, optimize and improve the initialization stage and population search stage of the escape optimization algorithm, and obtain the improved escape optimization algorithm;

[0071] As Figure 2 shown, it is a schematic diagram of the TimeKAN model structure

[0072] Specifically, the moving average method is used to gradually remove the relatively high-frequency components, generate multi-level sequences, and realize the decomposition of multi-frequency components. The specific formula is expressed as follows:

[0073] X i = AvgPool(Padding(X i-1 ))

[0074] where X i represents the sequence at the i-th layer, i is the index of the frequency band, AvgPool(·) represents the average pooling operation, which is used to remove the relatively high-frequency components in the sequence; Padding(·) represents the padding operation, which is used to add zeros at the edges of the sequence for convolution or Fourier transform.

[0075] Then, each sequence is embedded in a linear layer and mapped to a high-dimensional space, that is:

[0076] X i = Linear(X i )

[0077] Linear(·) represents the linear layer, which is used to map the sequence to a high-dimensional space.

[0078] The sequence is transformed into the frequency domain using the fast Fourier transform (FFT), and upsampling in the frequency domain is performed through zero-padding and the inverse fast Fourier transform (IFFT):

[0079]

[0080] FFT(·) represents the fast Fourier transform, which is used to transform the sequence from the time domain to the frequency domain. IFFT(·) represents the inverse fast Fourier transform, which is used to transform the sequence from the frequency domain back to the time domain.

[0081] Obtain the representation of each frequency band through the difference between sequences:

[0082]

[0083] where Fi represents the representation of the i-th frequency band.

[0084] Use depthwise separable convolution to independently capture the temporal patterns of each channel, and utilize a multi-order Kolmogorov-Arnold network to learn and represent the specific temporal patterns within each frequency band. The specific implementation is as follows:

[0085] F i,1 = ConvD D (F i , group = D)

[0086] F i,2 = KAN(F i , order = b + k - i)

[0087] where ConvD D (·) represents depthwise separable convolution, which is used to independently capture the temporal patterns of each channel; KAN(·) represents the Kolmogorov-Arnold network, which is used to learn and represent the specific temporal patterns within each frequency band; k represents the total number of frequency bands; b represents the lowest order of the Chebyshev polynomial in KAN;

[0088] Add the outputs of the multi-order KANs and depthwise separable convolution to obtain the final representation:

[0089]

[0090] where represents the representation of the i-th frequency band after learning through the M-KAN block.

[0091] Convert the learned representation of each frequency component back to a multi-level sequence:

[0092] X i = IFFT(Padding(FFT(X i+1 )))+ F i

[0093] Finally, use a simple linear layer to map the highest-level sequence to the prediction output:

[0094] X O = Linear(X i )

[0095] where X O represents the final prediction output.

[0096] Specifically, the Escape Optimization Algorithm (ESC) is a meta-heuristic optimization algorithm inspired by crowd evacuation behavior, and its modeling process is as follows:

[0097] Randomly initialize the population, and use the fitness function f to evaluate the fitness f of each individual i = f(x i ). Sort the population according to the fitness, and store the best individuals in the elite pool E, which represents the number of potential safe exits discovered by the population.

[0098] The ESC algorithm classifies the behavior of individuals into three categories: calm, conformist, or panicked, simulating the evolution of crowd behavior during evacuation. Individuals in the calm group behave rationally and move towards the central position C j which represents the collective decision of the population:

[0099]

[0100] where x i,j represents the position of the i-th individual in the j-th dimension; represents the updated position of the i-th individual in the j-th dimension; C j represents the mean of the population in the j-th dimension; m1 is used to control the weight of the influence of different groups on individuals in the calm group when updating their positions; w1 is used to control the weight of the influence of the central position of the population on individuals in the calm group when updating their positions; v c,j represents the movement vector of individuals in the calm group in the j-th dimension, and P(t) represents the panic index.

[0101]

[0102] where R c,j represents the random movement range of individuals in the calm group in the j-th dimension; represents a random perturbation term, increasing the exploration ability of the algorithm; and represent the minimum and maximum random movement ranges of group c in the j-th dimension respectively.

[0103] Conformist individuals imitate the behavior of both the calm group and the panicked group at the same time, and their positions are updated according to the influence of both:

[0104]

[0105] In the formula, x p,j is an individual randomly selected from the panicked group; v h,j represents the movement vector of individuals in the conformist group in the j-th dimension; R h,jrepresents the random movement range of individuals in the j-th dimension of the crowd; m2 is used to control the weight of the influence of different groups on individuals in the panic group when updating their positions; w2 is used to control the weight of the influence of the group center position on individuals in the panic group when updating their positions.

[0106] Panic-driven individuals explore the solution space in a more unstable manner, influenced by potential exits from the elite pool and random directions of other individuals:

[0107]

[0108] where E j is an individual randomly selected from the elite pool; P(t) represents the panic index; v p,j represents the movement vector of individuals in the panic group in the j-th dimension; R p,j represents the random movement range of individuals in the panic group in the j-th dimension.

[0109] As the iteration progresses, beyond T / 2, the algorithm will transition to the exploitation phase, in which all individuals are considered calm.

[0110] Individuals refine their positions by getting closer to the members of the elite pool, simulating the crowd gradually converging towards the determined best exit:

[0111]

[0112] where x i,j represents the position of the i-th individual in the j-th dimension; E j is an individual randomly selected from the elite pool; x rand,j represents the position of an individual selected from random individuals in the j-th dimension, used to simulate the influence of random individuals on individuals in the panic group when updating their positions, increasing the randomness and exploratory nature of their behavior.

[0113] The above formula describes the mathematical modeling process of the escape optimization algorithm, including the initialization of individuals, the reactions of different behaviors during the iteration process, and the position update in the exploitation phase.

[0114] As Figure 3 shown, introducing Tent initialization and Gauss perturbation strategies into ESC realizes the optimization and improvement of the initialization phase and population search phase of the ESC algorithm, and then obtains the improved escape optimization algorithm TGESC. The specific process is as follows:

[0115] This embodiment uses Tent mapping to generate a chaotic sequence to initialize the population, making the initial solutions as evenly distributed as possible in the solution space. Generating a chaotic sequence based on Tent mapping The process is as follows:

[0116]

[0117] where k is the number of populations, I is the current iteration number, and to maintain the randomness of the algorithm initialization information, u takes a value of Combined with the chaotic sequence further generate the initial position sequence of the gray wolf individuals within the search area The process is as follows:

[0118]

[0119] In the formula, are respectively the maximum and minimum values of the above sequence.

[0120] Chaos is a deterministic, random, aperiodic, and non-convergent method found in non-linear dynamic systems.

[0121] In this embodiment, the Gaussian mapping is used to generate a more uniformly distributed population, so that the algorithm has a faster convergence speed. The Gauss mapping is defined as follows:

[0122]

[0123] (5) Use the improved escape optimization algorithm to optimize the hyperparameters of the TimeKAN model;

[0124] The hyperparameters are the number of network layers Nn, the number of neurons Hu, and the learning rate Lr. The specific steps are as follows:

[0125] (51) Take the root mean square error index of the life prediction result and the true result as the objective function of the TimeKAN life prediction model;

[0126]

[0127] where N is the total number of time steps, y t is the true remaining useful life value at time step t, is the predicted remaining useful life value of the model at time step t.

[0128] (52) Determine that the population types are the learning rate lr and the batch size hu of the TimeKAN model, which are specifically expressed as follows:

[0129]

[0130] where X represents the population of the non-monopoly search algorithm (NO), and d represents the number of particles in one population;

[0131] (53) Determine the upper and lower limits of the search space with the ranges of the learning rate lr and the batch size hu, which are specifically expressed as follows:

[0132]

[0133] (6) Use the bearing vibration time series data obtained in step (3) and the improved escape optimization algorithm to train the bearing life prediction model, and obtain the trained MEMD-DisEn-WGANGP-TimeKAN-TGESC bearing life prediction model. After processing the various types of sensor data collected in real time through step (3), input them into the trained bearing life prediction model to predict the bearing vibration data for life, and obtain accurate remaining life results.

Claims

1. A bearing life prediction method based on data feature optimization and deep learning, characterized in that: The following steps are involved: (1) Use sensors to collect vibration signals at different positions of the bearing under healthy and fault conditions; (2) Reconstruct the collected vibration signal using the MEMD-DisEn method; (3) Using the WGAN-GP data enhancement strategy, the bearing vibration data obtained after reconstruction in step (2) is enhanced to expand the fault sample data; (4) A bearing life prediction model based on TimeKAN was constructed, and the Tent initialization and Gauss perturbation strategies were introduced into the escape optimization algorithm. The initialization phase and population search phase of the escape optimization algorithm were optimized and improved to obtain the improved escape optimization algorithm. (5) Use the improved escape optimization algorithm to optimize the hyperparameters of the TimeKAN model; (6) Using the bearing vibration data obtained in step (3) and the improved escape optimization algorithm to train the bearing life prediction model, and using the trained and optimized bearing life prediction model to perform online fault analysis and life prediction on the real-time collected bearing vibration data to obtain the remaining life result of the bearing.

2. A bearing life prediction method based on data feature optimization and deep learning according to claim 1, characterized in that: The bearing vibration signal in step (1) is an n-dimensional variable sequence, denoted as Where T is the sequence length.

3. A bearing life prediction method based on data feature optimization and deep learning according to claim 2, characterized in that: The reconstruction in step (2) is to decompose the n-dimensional bearing vibration signal variable sequence collected in step (1) into p IMF components and a residual term using the multivariate empirical mode decomposition method, and reconstruct the sequence using the DisEn entropy based on the p IMF components obtained from the decomposition of each sequence.

4. The bearing life prediction method based on data feature optimization and deep learning according to claim 1 is characterized in that: The data enhancement described in step (3) is to train the WGAN-GP model, generate synthetic samples on the original data set, merge the synthetic samples with the original data set, create a balanced enhanced data set, introduce gradient penalty to enforce the K-Lipschitz continuity constraint of the Wasserstein distance, improve the stability of training and generate high-quality data enhancement samples.

5. The bearing life prediction method based on data feature optimization and deep learning according to claim 3 is characterized in that: The goal of the WGAN-GP is to minimize the Wasserstein distance between the generator and maximize the discriminator. During the training process of the WGAN-GP model, the minimization generator and the maximization discriminator are trained alternately. The goal of the discriminator is to maximize the score of the real samples and minimize the score of the generated samples; the goal of the generator is to generate samples that can deceive the discriminator. A gradient penalty term is added to the loss function of the discriminator to satisfy the K-Lipschitz constraint.

6. The bearing life prediction method based on data feature optimization and deep learning according to claim 1 is characterized in that: The step (4) generates a chaotic sequence using Tent mapping to initialize the population. The formula is as follows: in, is a chaotic sequence, k is the population number, I is the current iteration number, and to maintain the randomness of the algorithm initialization information, u is taken as Secondly, Gauss mapping is used to generate a more evenly distributed population.

7. The bearing life prediction method based on data feature optimization and deep learning according to claim 1 is characterized in that: The hyperparameters in step (5) are the number of network layers Nn, the number of neurons Hu and the learning rate Lr. The specific steps are as follows: (51) The root mean square error index between the life prediction results and the actual results is used as the objective function of the TimeKAN life prediction model; (52) Determine the learning rate and batch size of the TimeKAN model, which are specifically expressed as follows: Among them, X represents the population of the non-monopolistic search algorithm, and d represents the number of particles of one type in the population; (53) The upper and lower limits of the search space are determined by the range of learning rate and batch size, which are specifically expressed as follows:

8. The bearing life prediction method based on data feature optimization and deep learning according to claim 1 is characterized in that: The step (6) is to train the bearing life prediction model using the bearing vibration time series data processed by the step (3) and the improved escape optimization algorithm, thereby obtaining a trained bearing life prediction model. The various types of sensor data collected are processed by the step (3), and then input into the trained bearing life prediction model to predict the life of the bearing vibration data and obtain the remaining life result.

9. A computer device, characterized in that: It comprises one or more processors, a memory and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of a bearing life prediction method based on data feature optimization and deep learning as described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a bearing life prediction method based on data feature optimization and deep learning as described in any one of claims 1 to 8 are implemented.

Images