Rolling bearing RUL prediction method based on fish optimization algorithm global tuning

Abstract

The invention relates to a rolling bearing RUL prediction method based on # imgabs0 # fish optimization algorithm global tuning. Compared with the prior art, the rolling bearing RUL prediction method based on # imgabs0 # fish optimization algorithm global tuning solves the defects that bearing degradation data is scarce and unbalanced, parameter adjustment of a traditional deep learning model is difficult, and prediction model precision is poor. The method comprises the following steps: acquiring rolling bearing data; performing data preprocessing and feature extraction; constructing a data enhancement model; an RUL prediction model is constructed; performing fish global optimization on the data enhancement model and the RUL prediction model; and obtaining an RUL prediction result of the rolling bearing. According to the method, the WGAN-GP is combined for data enhancement, the LSTM is used for time series data analysis, and the ROA is used for global hyper-parameter optimization, so that the problems of rolling bearing data scarcity and prediction model performance limitation are effectively solved, the accuracy and robustness of prediction of the remaining service life are improved, and accurate prediction of the RUL of the rolling bearing is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of rolling bearings, specifically a Rolling bearing RUL prediction method based on global tuning of fish optimization algorithm. Background Art

[0002] As industrial equipment becomes increasingly complex and intelligent, health monitoring and fault prediction of key equipment components have become particularly important. As the core component of rotating equipment, the degradation process of rolling bearings directly affects the overall performance and reliability of the equipment. Therefore, how to effectively predict the remaining useful life (RUL) of bearings and perform maintenance in advance is an important issue to ensure the safe operation of equipment and reduce operating costs.

[0003] In recent years, with the maturity of sensor technology and the widespread application of deep learning algorithms, the research on bearing RUL prediction based on data-driven and deep learning has made significant progress. The data-driven rolling bearing RUL prediction method is a technology that predicts the remaining life based on actual operating data. This method collects relevant parameters such as the vibration signal of the rolling bearing, and uses signal processing algorithms to construct a mapping relationship between data and bearing performance. The health index (HI) that can characterize the performance of the equipment is obtained by this method. After obtaining the HI, a deep learning algorithm is used to predict the HI and estimate the RUL. For example, Catelani et al. obtained relevant degradation indicators through data-driven CM and used recurrent neural networks (RNN) for RUL estimation, Zhou et al. predicted RUL by constructing unsupervised health indicators and using memory-enhanced gated recurrent units (GRUs), and Lu et al. constructed HI based on monotonicity and used physically guided long short-term memory networks (LSTMs) for RUL prediction.

[0004] The above method is based on the assumption that sensor data collection is complete and comprehensive. In practical applications, it is expensive to obtain enough samples, which makes experiments often face problems such as uneven data quality and insufficient samples, which seriously restricts the accuracy and reliability of the model.

[0005] In response to this, some studies have begun to focus on using data enhancement technology to overcome the dilemma of poor data quality and insufficient samples. Grasso et al. proposed a data-driven automated method to achieve data enhancement of rolling bearing vibration signals by optimizing empirical mode decomposition (EMD). Guo et al. adopted a GAN-based data enhancement method to generate rolling bearing data samples through an auxiliary classifier GAN. Yu proposed a reparameterized residual denoising diffusion probability model (ReF-DDPM) to generate high-quality samples by transforming Gaussian noise and original samples through Markov chains. Faced with various challenges, researchers have adopted RUL prediction technology that integrates data-driven, data enhancement and deep learning, and improved the prediction effect of network models.

[0006] Before using deep learning model training, the model's hyperparameters have a decisive influence on prediction accuracy. Setting excellent hyperparameters can improve the model's convergence speed and generalization ability. The traditional manual parameter adjustment method is not only time-consuming and labor-intensive, but also difficult to guarantee the global optimal hyperparameters.

[0007] Therefore, how to design an accurate rolling bearing RUL prediction method has become a technical problem that needs to be solved urgently. Summary of the invention

[0008] The purpose of the present invention is to solve the defects of the existing technology in that the bearing degradation data is scarce and unbalanced, the traditional deep learning model is difficult to adjust the parameters, and the prediction model has poor accuracy. The rolling bearing RUL prediction method based on global tuning of fish optimization algorithm is used to solve the above problems.

[0009] In order to achieve the above object, the technical solution of the present invention is as follows:

[0010] A kind of The rolling bearing RUL prediction method based on global tuning of the fish optimization algorithm includes the following steps:

[0011] Acquisition of rolling bearing data: Acquisition of rolling bearing vibration signals and data enhancement processing;

[0012] Perform data preprocessing and feature extraction: extract degradation information from vibration signals and obtain health indicators HI that reflect the health status of equipment or systems;

[0013] Construction of data enhancement model: WGAN-GP network is used for data enhancement, hyperparameters are globally optimized through ROA optimization algorithm, and a data enhancement model based on ROA-WGAN-GP is constructed;

[0014] Construction of RUL prediction model: Using LSTM network as the prediction model, combined with the data generated by the data enhancement model, and through global tuning of the ROA optimization algorithm, a RUL prediction model based on ROA is constructed;

[0015] Global optimization of data enhancement model and RUL prediction model: Optimize data enhancement model and RUL prediction model using ROA optimization algorithm;

[0016] Obtaining the RUL prediction results of rolling bearings: The vibration signal of the rolling bearing to be predicted is sent to the data enhancement model after preprocessing, and the generated data and the preprocessed data are sent to the RUL prediction model to obtain the RUL prediction results of the rolling bearings.

[0017] The data preprocessing and feature extraction comprises the following steps:

[0018] Obtain the rolling bearing vibration signal after enhancement processing, and use the collected horizontal acceleration to extract time domain features;

[0019] Use pycharm to extract the following 13 features from the .csv dataset file: maximum, mean, variance, standard deviation, root mean square, peak-to-peak value, skewness, kurtosis, waveform factor, peak factor, pulse factor, margin factor, and spectral kurtosis;

[0020] The bearing health index HI based on spectral kurtosis is constructed, and its mathematical formula is as follows:

[0021]

[0022] Where N is the number of samples, x i is the current value at time i, is the sample mean;

[0023] Data feature extraction: The grey correlation degree between each time domain feature and HI is calculated by the grey correlation analysis method, and the features with a value greater than 0.8 are selected as the training data set, which includes 8 time domain features, including the kurt, peak value fz, pulse mc of the horizontal vibration signal and the health indicator variance HI-var, health indicator standard deviation HI_std, health indicator peak HI_fz, health indicator peak-to-peak value HI_ffz, and health indicator margin HI_yd.

[0024] The construction of the data enhancement model includes the following steps:

[0025] The data augmentation model is set as the WGAN-GP network based on ROA global optimization, including the WGAN-GP network and the ROA hyperparameter optimization model. The WGAN-GP network is used to perform data augmentation on the rolling bearing dataset, and the ROA hyperparameter optimization model is used to optimize the hyperparameters of the WGAN-GP network.

[0026] Introducing Wasserstein distance loss and gradient penalty, the total loss function of the WGAN-GP model is defined as follows:

[0027]

[0028] Among them, P r Represents the probability distribution of the real sample, P g represents the probability distribution of samples generated by G, is the mathematical expectation when the generated samples are used as the discriminator input, is the mathematical expectation value when the discriminator input is a real sample,

[0029] is the generator loss G(loss), is the discriminator loss D(loss), is the gradient penalty, satisfy x r ~P r ,x g ~P g ,∈∈0,1.

[0030] The construction of the RUL prediction model includes the following steps:

[0031] The RUL prediction model is set as an LSTM network based on ROA global optimization, and the ROA hyperparameter optimization model is used to optimize the hyperparameters of the LSTM network; the LSTM network is used to process the input data set and output the prediction results. The input data set is the data generated by the data enhancement model to expand the samples;

[0032] The mathematical formula for setting the LSTM network loss function is as follows:

[0033]

[0034] Where n is the number of samples, y i is the true value of the i-th sample, is the predicted value of the ith sample.

[0035] The remora global optimization of the data enhancement model and the RUL prediction model includes the following steps:

[0036] The ROA optimization algorithm is divided into a global exploration phase and a local development phase, which respectively simulate the behavior of remora in the process of searching for and adhering to the host;

[0037] Initialization: In ROA, the calculation formula for initializing the population position is as follows:

[0038] x i =lb+rand(ub-lb)

[0039] In the formula, x i is the position of the individual, lb and ub are the lower and upper bounds of the search space, and rand is a random number between 0 and 1;

[0040] Calculation of fitness function ξ: Fitness is used to determine the optimal solution. The fitness function with a smaller fitness value is considered the optimal solution. The calculation formula is as follows:

[0041]

[0042] Where n is the number of samples,

[0043] L(loss) is the LSTM loss function in the RUL prediction model, G(loss) and D(loss) are the generator loss function and the discriminator loss function in the data augmentation model, respectively. best is the output of the ROA optimization algorithm;

[0044] Setting SFO strategy:

[0045] When the remora attaches to the fast-moving swordfish host, the algorithm uses the optimal individual guidance mechanism and adds a random selection rule;

[0046] Based on the elite idea of ​​the sailfish algorithm, its position update The equation is as follows:

[0047]

[0048] Among them, t is the current iteration number, is the current optimal remora individual, is the current random remora individual, rand is a random number between 0 and 1;

[0049] To set WOA policy:

[0050] When the remora attaches itself to a large whale host, its position is updated by using the spiral search strategy of the whale optimization algorithm. The equation is as follows:

[0051]

[0052] k=rand×(a-1)+1

[0053]

[0054] Where D represents the distance between the optimal position before the update and the current position, k is a random number between -1 and 1, a decreases linearly between [-2,1] during the iteration, and T is the maximum number of iterations;

[0055] Set the global exploration phase:

[0056] In the global exploration phase, the remora continuously makes small movements around the host, and the mathematical formula is as follows:

[0057]

[0058] Among them, X att It was a tentative move by the remora, X pre Indicates the position of the previous generation of remora, randn is a normally distributed random number between 0 and 1;

[0059] To determine whether to change the host, the following formula is used for judgment:

[0060]

[0061] H(i)=round(rand)

[0062] Among them, H(i) determines the host of the remora, and the initial value is 0 or 1. If H(i) is equal to 0, it will absorb whales, and if H(i) is equal to 1, it will absorb swordfish. Round is a rounding function. and f(X att ) are respectively With X att The fitness value of

[0063] Setting the local development stage:

[0064] In the local development stage, the search range is reduced, and the remora uses the resources around the host to further optimize its position. The mathematical expression is as follows:

[0065]

[0066] B=2×V×rand-V

[0067]

[0068] Among them, A is the distance moved by the remora, which is related to the volume of the remora and the host. The remora factor C is used to limit the position of the remora, and its value is 0.1. B and V represent the volume of the host and remora, respectively, to distinguish the host and remora in the search space.

[0069] Remora global optimization: The WGAN-GP model and LSTM network are trained according to the ROA optimization algorithm, and the optimal solution is calculated through the loss function output of the two models. The hyperparameters of the data enhancement model and the RUL prediction model are optimized by using the optimal solution to achieve global optimization based on the ROA optimization algorithm.

[0070] The acquisition of the rolling bearing RUL prediction result comprises the following steps:

[0071] A public data set is used as validation data, and the data set is divided into a training set and a test set. The health index is constructed through the data set as the rolling bearing RUL to be predicted, and the grey correlation analysis method is used for data preprocessing.

[0072] The preprocessed training set is fed into the data enhancement model and RUL prediction model, and the corresponding loss function is output in the WGAN-GP model and LSTM network. The loss function is input into ROA to perform optimization algorithm training.

[0073] Use the trained ROA to optimize the WGAN-GP model, build a data enhancement model through model hyperparameter optimization, generate rolling bearing data, and expand the experimental samples;

[0074] The expanded sample data is sent to the RUL prediction model, and combined with the ROA model optimization, the preprocessed data is predicted, and finally the rolling bearing RUL prediction result is obtained.

[0075] Beneficial Effects

[0076] The present invention is based on Compared with the existing technology, the rolling bearing RUL prediction method with global tuning of the fish optimization algorithm effectively solves the problems of scarce rolling bearing data and limited prediction model performance by combining WGAN-GP for data enhancement, LSTM for time series data analysis, and ROA for global hyperparameter optimization, improves the accuracy and robustness of remaining service life prediction, and realizes accurate prediction of rolling bearing RUL.

[0077] In order to solve the problem of scarce and unbalanced bearing degradation data, the present invention generates high-quality simulation data by using a WGAN-GP network based on ROA to optimize hyperparameters. The remora optimization algorithm simulates the foraging behavior of remora in nature, and has the characteristics of strong global search ability and good adaptability. ROA is used to optimize the hyperparameters of the LSTM model, which improves the learning efficiency and prediction accuracy of the model. The data enhancement (DE) module based on the ROA-WGAN-GP network is integrated into the ROA-LSTM network, which effectively increases the diversity and quantity of training samples and improves the generalization ability of the model.

[0078] This paper innovatively proposes a ROA-DE-LSTM prediction model with ROA global synchronous tuning. The HI that maps the relationship between data and RUL is constructed through prior knowledge, WGAN-GP is used to achieve data enhancement, LSTM is used to carry out RUL prediction, and ROA is used to globally optimize the hyperparameters of the data enhancement model and the RUL prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 is a method sequence diagram of the present invention;

[0080] Figure 2 It is a structural flow chart of the present invention;

[0081] Figure 3 This is the operation table picture of the relevant data set used in this method;

[0082] Figure 4 A visualization diagram of the extracted features of the method for predicting the remaining useful life of a rolling bearing according to the present invention;

[0083] Figure 5 It is a structural schematic diagram of a data enhancement model in the method for predicting the remaining useful life of a rolling bearing involved in the present invention;

[0084] Figure 6 It is a structural schematic diagram of the RUL prediction model in the remaining useful life prediction method of a rolling bearing involved in the present invention;

[0085] Figure 7 It is a schematic diagram of ROA global optimization comparison results of the data enhancement model in the method for predicting the remaining useful life of a rolling bearing involved in the present invention;

[0086] Figure 8 This is the final training set RUL prediction result diagram;

[0087] Fig. 9 This is the final test set RUL prediction result diagram;

[0088] Fig.10The prediction results of the model proposed in this invention and the prediction results of the comparative model are displayed under four indicators. DETAILED DESCRIPTION

[0089] In order to have a further understanding and recognition of the structural features and the effects achieved by the present invention, a preferred embodiment and accompanying drawings are used for detailed description as follows:

[0090] like Figure 1 and Figure 2 As shown, the present invention is based on The rolling bearing RUL prediction method based on global tuning of the fish optimization algorithm includes the following steps:

[0091] The first step is to obtain rolling bearing data: obtain rolling bearing vibration signals and perform data enhancement processing. In the laboratory environment, the data set of the present invention comes from the bearing life test bench of Xi'an Jiaotong University. The instrument structure is as follows Figure 3 , where the data set content comes from the horizontal and vertical vibration signals collected by this test bench. Two data sets (bearing 3_1 and bearing 3_4) under the operating condition of 40Hz / 10kN are selected to verify the effectiveness of the ROA-DE-LSTM proposed in this paper. Bearing 3-1 is used for data enhancement of ROA-WGAN-GP. The enhanced data and bearings 3-1 and 3-4 are used together through ROA-LSTM to verify the prediction effect, and the horizontal vibration signal is used as the detection data.

[0092] The second step is to perform data preprocessing and feature extraction: extract the degradation information in the vibration signal and obtain the health index HI that reflects the health status of the equipment or system.

[0093] (1) Obtaining the rolling bearing vibration signal after enhancement processing, and extracting time domain features using the collected horizontal acceleration;

[0094] The following 13 types of features are extracted from the .csv format data set file using the pycharm tool: maximum, mean, variance, standard deviation, root mean square, peak-to-peak value, skewness, kurtosis, waveform factor, peak factor, pulse factor, margin factor, and spectral kurtosis.

[0095] (2) In RUL prediction, health indicators are usually used to characterize the degradation process of equipment performance over time. These indicators can directly or indirectly reflect the remaining service life of the equipment. To this end, the present invention constructs a bearing health index HI based on spectral kurtosis, and its mathematical formula is as follows:

[0096]

[0097] Where N is the number of samples, x i is the current value at time i, is the sample mean.

[0098] (3) Data feature extraction: The grey correlation between each time domain feature and HI is calculated by grey correlation analysis, and features with a value greater than 0.8 are selected as training data sets, including the kurt, peak value fz, pulse mc of the horizontal vibration signal, and the health indicator variance HI-var, health indicator standard deviation HI_std, health indicator peak value HI_fz, health indicator peak-to-peak value HI_ffz, and health indicator margin HI_yd, a total of 8 time domain features. Grey correlation analysis evaluates the degree of correlation between factors by measuring the similarity of the development trends between them. In time series prediction, strong correlation means that there is a close connection between the data. Changes in one data sequence are likely to affect changes in another data sequence, such as Figure 4 shown.

[0099] The third step is to build a data enhancement model: use the WGAN-GP network for data enhancement, use the ROA optimization algorithm to globally optimize the hyperparameters, and build a data enhancement model based on ROA-WGAN-GP. Its structure is as follows: Figure 5 .

[0100] Using simulation data generated by WGAN-GP structural variants to expand training samples can solve the problems of poor data quality and insufficient samples. However, for the task of bearing data enhancement, the present invention is more concerned with the quality and diversity of generated samples. As a relatively mature and effective model, directly using WGAN-GP for data enhancement is a direct and effective method, which avoids the additional complexity and computational cost that may be introduced by over-optimizing the model. At the same time, combined with the ROA proposed in this paper for hyperparameter tuning, it can provide more effective help for the next model prediction stage.

[0101] (1) The data augmentation model is set as the WGAN-GP network based on ROA global optimization, including the WGAN-GP network and the ROA hyperparameter optimization model. The WGAN-GP network is used to perform data augmentation on the rolling bearing dataset, and the ROA hyperparameter optimization model is used to optimize the hyperparameters of the WGAN-GP network.

[0102] (2) Introducing Wasserstein distance loss and gradient penalty, the total loss function of the WGAN-GP model is defined as follows:

[0103]

[0104] Among them, P r Represents the probability distribution of the real sample, P g represents the probability distribution of samples generated by G, is the mathematical expectation when the generated samples are used as the discriminator input, is the mathematical expectation value when the discriminator input is a real sample,

[0105] is the generator loss G(loss), is the discriminator loss D(loss), is the gradient penalty, satisfy x r ~P r ,x g ~P g ,∈∈0,1.

[0106] The fourth step is to build a RUL prediction model: using the LSTM network as the prediction model, combined with the data generated by the data enhancement model, and globally tuned through the ROA optimization algorithm to build a RUL prediction model. Its structure is as follows: Figure 6 .

[0107] Data-driven prediction technology based on LSTM has been widely used, and its various network structure variants are also complex and varied. Adding various optimization models to the LSTM network has been proven to improve the accuracy of the results to a certain extent. Complex models usually mean lower model generalization performance, and the increase in parameters also makes model training more time-consuming. The deep learning model using the basic LSTM network enriched the data set and improved the model prediction performance after the expanded samples of WGAN-GP. In order to avoid the reduction of model generalization performance and the slowdown of model training, it is of great significance to use intelligent optimization algorithms to optimize model hyperparameters.

[0108] (1) The RUL prediction model is set as an LSTM network based on ROA global optimization, and the ROA hyperparameter optimization model is used to optimize the hyperparameters of the LSTM network; the LSTM network is used to process the input data set and output the prediction results. The input data set is the data generated by the data augmentation model to expand the samples.

[0109] (2) The mathematical formula for setting the LSTM network loss function is as follows:

[0110]

[0111] Where n is the number of samples, y i is the true value of the i-th sample, is the predicted value of the ith sample.

[0112] Step 5: Global optimization of the data enhancement model and RUL prediction model: Use the ROA optimization algorithm to optimize the data enhancement model and RUL prediction model.

[0113] The present invention aims to build an efficient bearing degradation state prediction model by combining ROA global optimization with the data enhancement technology of the GAN network and the time series prediction capability of the LSTM network. This invention can not only provide a new technical path for the predictive maintenance of bearings, but also promote the application and development of intelligent optimization algorithms in the field of deep learning.

[0114] (1) The ROA optimization algorithm is divided into a global exploration phase and a local development phase. These two phases simulate the behaviors of remora in the process of searching for and adhering to hosts.

[0115] A1) Initialization: In ROA, the calculation formula for initializing the population position is as follows:

[0116] x i =lb+rand(ub-lb)

[0117] In the formula, x i is the position of the individual, lb and ub are the lower and upper bounds of the search space, and rand is a random number between 0 and 1;

[0118] A2) Calculation of fitness function ξ: Fitness is used to determine the optimal solution. The fitness function with a smaller fitness value is considered the optimal solution. The calculation formula is as follows:

[0119]

[0120] Where n is the number of samples,

[0121] L(loss) is the LSTM loss function in the RUL prediction model, G(loss) and D(loss) are the generator loss function and the discriminator loss function in the data augmentation model, respectively. best is the output of the ROA optimization algorithm;

[0122] ROA optimizes the hyperparameters of the entire system by comprehensively considering the performance of the RUL prediction model and the data augmentation model. By evaluating the overall performance of the entire system, we can find the parameter configuration that can improve both data quality and prediction accuracy.

[0123] A3) Set up SFO strategy:

[0124] When the remora attaches to the fast-moving swordfish host, the algorithm uses the optimal individual guidance mechanism and adds a random selection rule;

[0125] Based on the elite idea of ​​the sailfish algorithm, its position update The equation is as follows:

[0126]

[0127] Among them, t is the current iteration number, is the current optimal remora individual, is the current random remora individual, rand is a random number between 0 and 1;

[0128] A4) Set up WOA strategy:

[0129] When the remora attaches itself to a large whale host, its position is updated by using the spiral search strategy of the whale optimization algorithm. The equation is as follows:

[0130]

[0131] k=rand×(a-1)+1

[0132]

[0133] Where D represents the distance between the optimal position before the update and the current position, k is a random number between -1 and 1, a decreases linearly between [-2,1] during the iteration, and T is the maximum number of iterations;

[0134] A5) Setting the global exploration phase:

[0135] In the global exploration phase, the remora continuously makes small movements around the host, and the mathematical formula is as follows:

[0136]

[0137] Among them, X att It was a tentative move by the remora, X pre Indicates the position of the previous generation of remora, randn is a normally distributed random number between 0 and 1;

[0138] To determine whether to change the host, the following formula is used for judgment:

[0139]

[0140] H(i)=round(rand)

[0141] Among them, H(i) determines the host of the remora, and the initial value is 0 or 1. If H(i) is equal to 0, it will absorb whales, and if H(i) is equal to 1, it will absorb swordfish. Round is a rounding function. and f(X att ) are respectively With X att The fitness value of

[0142] A6) Setting local development stages:

[0143] In the local development stage, the search range is reduced, and the remora uses the resources around the host to further optimize its position. The mathematical expression is as follows:

[0144]

[0145] B=2×V×rand-V

[0146] #imgpt57#

[0147] Among them, A is the distance moved by the remora, which is related to the volume of the remora and the host. The remora factor C is used to limit the position of the remora, and its value is 0.1. B and V represent the volumes of the host and remora, respectively, which are used to distinguish the host and remora in the search space.

[0148] (2) Global optimization of remora: The WGAN-GP model and LSTM network are trained according to the ROA optimization algorithm. The loss function outputs of the two models are used to calculate and obtain the optimal solution. The hyperparameters of the data augmentation model and the RUL prediction model are optimized using the optimal solution to achieve global optimization based on the ROA optimization algorithm.

[0149] Step 6: Obtaining the RUL prediction results of rolling bearings: The vibration signal of the rolling bearing to be predicted is sent to the data enhancement model after preprocessing, and the generated data and the preprocessed data are sent to the RUL prediction model to obtain the RUL prediction results of the rolling bearings.

[0150] (1) A public dataset is used as validation data, and the dataset is divided into a training set and a test set. The health index is constructed through the dataset as the rolling bearing RUL to be predicted, and the grey correlation analysis method is used for data preprocessing.

[0151] (2) The preprocessed training set is sent to the data enhancement model and RUL prediction model, and the corresponding loss function is output in the WGAN-GP model and LSTM network. The optimization algorithm is trained by inputting the loss function into ROA.

[0152] (3) Use the trained ROA to optimize the WGAN-GP model, build a data enhancement model through model hyperparameter optimization, and generate rolling bearing data to expand the experimental samples.

[0153] (4) The expanded sample data is fed into the RUL prediction model, and combined with the ROA model optimization, the preprocessed data is predicted, and finally the rolling bearing RUL prediction results are obtained. The training set prediction results are as follows: Figure 8 , the test set prediction results are as follows Fig. 9 .

[0154] Data enhancement model result display and comparison: This paper displays the results of optimizing the WGAN-GP model on data enhancement using different intelligent optimization algorithms. MSE is used as the loss function, and the discriminator loss, regularization loss and gradient penalty term of the calculated real data and fake samples are combined into the total discriminator loss Loss D.

[0155] Bearing 3-1 was used as the experimental object, and the basic model, PSO algorithm, HHO algorithm and ROA algorithm were adopted. The whole training process lasted about 3000 cycles. Figure 7 The blue curve represents the ROA optimization result. From the figure, we can see that the ROA optimization effect on the discriminator D is better than other models and has the best convergence effect.

[0156] Results display and comparison of ROA-DE-LSTM prediction model: Due to the long periodicity of health indicators, the sliding window method is used in the LSTM network model to dynamically read data with a step size of 1 and a window size of 50 until it is completely traversed. The health indicator HI is predicted by the constructed network model. In order to verify the effectiveness and performance of the proposed ROA-DE-LSTM and to facilitate comparison with other methods, RMSE, MAE, MAPE, R 2 Do performance indicators, Fig.10 The results of each experimental model under four indicators are shown.

[0157] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A method based on The rolling bearing RUL prediction method based on global tuning of fish optimization algorithm is characterized by: The following steps are involved: 11) Acquisition of rolling bearing data: Acquisition of rolling bearing vibration signals and data enhancement processing; 12) Perform data preprocessing and feature extraction: extract degradation information from the vibration signal and obtain the health index HI that reflects the health status of the equipment or system; 13) Construction of data enhancement model: WGAN-GP network is used for data enhancement, hyperparameters are globally optimized through ROA optimization algorithm, and a data enhancement model based on ROA-WGAN-GP is constructed; 14) Construction of RUL prediction model: Using LSTM network as the prediction model, combined with the data generated by the data enhancement model, and through global tuning of the ROA optimization algorithm, a RUL prediction model based on ROA is constructed; 15) Global optimization of data enhancement model and RUL prediction model: Optimize the data enhancement model and RUL prediction model using ROA optimization algorithm; 16) Obtaining the RUL prediction results of rolling bearings: The vibration signal of the rolling bearing to be predicted is sent to the data enhancement model after preprocessing, and the generated data and the preprocessed data are sent to the RUL prediction model to obtain the RUL prediction results of the rolling bearings.

2. A method according to claim 1 based on The rolling bearing RUL prediction method based on global tuning of fish optimization algorithm is characterized by: The data preprocessing and feature extraction comprises the following steps: 21) Obtaining the rolling bearing vibration signal after enhancement processing, and extracting time domain features using the collected horizontal acceleration; Use pycharm to extract the following 13 features from the .csv dataset file: maximum, mean, variance, standard deviation, root mean square, peak-to-peak value, skewness, kurtosis, waveform factor, peak factor, pulse factor, margin factor, and spectral kurtosis; 22) Construct the bearing health index HI based on spectral kurtosis, and its mathematical formula is as follows: Where N is the number of samples, x i is the current value at time i, is the sample mean; 23) Data feature extraction: The grey correlation degree between each time domain feature and HI is calculated by the grey correlation analysis method, and the features with a value greater than 0.8 are selected as the training data set, which includes 8 time domain features, including the kurt, peak value fz, pulse mc of the horizontal vibration signal and the health indicator variance HI-var, health indicator standard deviation HI_std, health indicator peak HI_fz, health indicator peak-to-peak value HI_ffz, and health indicator margin HI_yd.

3. A method according to claim 1 based on The rolling bearing RUL prediction method based on global tuning of fish optimization algorithm is characterized by: The construction of the data enhancement model includes the following steps: 31) The data augmentation model is set as a WGAN-GP network based on ROA global optimization, including a WGAN-GP network and a ROA hyperparameter optimization model. The WGAN-GP network is used to perform data augmentation on a rolling bearing data set, and the ROA hyperparameter optimization model is used to perform hyperparameter optimization on the WGAN-GP network. 32) Introducing Wasserstein distance loss and gradient penalty, the total loss function of the WGAN-GP model is defined as follows: Among them, P r Represents the probability distribution of the real sample, P g represents the probability distribution of samples generated by G, is the mathematical expectation when the generated samples are used as the discriminator input, is the mathematical expectation value when the discriminator input is a real sample, is the generator loss G(loss), is the discriminator loss D(loss), is the gradient penalty, satisfy x r ~P r ,x g ~P g ,∈∈0,1.

4. A method according to claim 1 based on The rolling bearing RUL prediction method based on global tuning of fish optimization algorithm is characterized by: The construction of the RUL prediction model includes the following steps: 41) The RUL prediction model is set as an LSTM network based on ROA global optimization, and the ROA hyperparameter optimization model is used to optimize the hyperparameters of the LSTM network; the LSTM network is used to process the input data set and output the prediction results, and the input data set is the data generated by the data enhancement model to expand the samples; 42) The mathematical formula for setting the LSTM network loss function is as follows: Where n is the number of samples, y i is the true value of the i-th sample, is the predicted value of the ith sample.

5. A method according to claim 1 based on The rolling bearing RUL prediction method based on global tuning of fish optimization algorithm is characterized by: The remora global optimization of the data enhancement model and the RUL prediction model includes the following steps: 51) The ROA optimization algorithm is divided into a global exploration phase and a local development phase, which respectively simulate the behavior of remora in the process of searching for and adhering to the host; 511) Initialization: In ROA, the calculation formula for initializing the population position is as follows: x i =lb+rand(ub-lb) In the formula, x i is the position of the individual, lb and ub are the lower and upper bounds of the search space, and rand is a random number between 0 and 1; 512) Calculation of fitness function ξ: Fitness is used to determine the optimal solution. The fitness function with a smaller fitness value is considered the optimal solution. The calculation formula is as follows: Where n is the number of samples, L(loss) is the LSTM loss function in the RUL prediction model, G(loss) and D(loss) are the generator loss function and the discriminator loss function in the data augmentation model, respectively. best is the output of the ROA optimization algorithm; 513) Set SFO strategy: When the remora attaches to the fast-moving swordfish host, the algorithm uses the optimal individual guidance mechanism and adds a random selection rule; Based on the elite idea of ​​the sailfish algorithm, its position update The equation is as follows: Among them, t is the current iteration number, is the current optimal remora individual, is the current random remora individual, rand is a random number between 0 and 1; 514) Set WOA policy: When the remora attaches itself to a large whale host, its position is updated by using the spiral search strategy of the whale optimization algorithm. The equation is as follows: k=rand×(a-1)+1 Where D represents the distance between the optimal position before the update and the current position, k is a random number between -1 and 1, a decreases linearly between [-2,1] during the iteration, and T is the maximum number of iterations; 515) Set the global exploration stage: In the global exploration phase, the remora continuously makes small movements around the host, and the mathematical formula is as follows: Among them, X att It was a tentative move by the remora, X pre Indicates the position of the previous generation of remora, randn is a normally distributed random number between 0 and 1; To determine whether to change the host, the following formula is used for judgment: H(i)=round(rand) Among them, H(i) determines the host of the remora, and the initial value is 0 or 1. If H(i) is equal to 0, it will absorb whales, and if H(i) is equal to 1, it will absorb swordfish. Round is a rounding function. and f(X att ) are respectively With X att The fitness value of 516) Setting local development stages: In the local development stage, the search range is reduced, and the remora uses the resources around the host to further optimize its position. The mathematical expression is as follows: B=2×V×rand-V Among them, A is the distance moved by the remora, which is related to the volume of the remora and the host. The remora factor C is used to limit the position of the remora, and its value is 0.

1. B and V represent the volume of the host and remora, respectively, to distinguish the host and remora in the search space. 52) Global optimization of remora: The WGAN-GP model and LSTM network are trained according to the ROA optimization algorithm, and the optimal solution is calculated through the loss function output of the two models; the hyperparameters of the data augmentation model and the RUL prediction model are optimized by using the optimal solution to achieve global optimization based on the ROA optimization algorithm.

6. A method according to claim 1 based on The rolling bearing RUL prediction method based on global tuning of fish optimization algorithm is characterized by: The acquisition of the rolling bearing RUL prediction result comprises the following steps: 61) A public data set is used as validation data, the data set is divided into a training set and a test set, a health index is constructed through the data set as the rolling bearing RUL to be predicted, and the grey correlation analysis method is used for data preprocessing; 62) The preprocessed training set is fed into the data enhancement model and the RUL prediction model, the corresponding loss function is output in the WGAN-GP model and the LSTM network, and the optimization algorithm training is performed by inputting the loss function into the ROA; 63) Use the trained ROA to optimize the WGAN-GP model, build a data enhancement model through model hyperparameter optimization, generate rolling bearing data, and expand the experimental samples; 64) The expanded sample data is sent to the RUL prediction model, combined with the ROA model optimization, the preprocessed data is predicted, and finally the rolling bearing RUL prediction result is obtained.

Images