A method, device, equipment and storage medium for generating bearing vibration data
By constructing CVGAN model and autoregression method, using historical vibration data and HI tags to generate full-life vibration data of bearings, the problem of discontinuous vibration data generation in the existing technology is solved, and the clarity and application of data is improved. Equipment health monitoring and prediction provide reliable data support.
Patent Information
- Application Number
- CN202410955659.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-07-17
AI Technical Summary
The prior art is difficult to effectively generate continuous vibration data covering the entire life cycle of the equipment, which affects the effectiveness of the model in practical applications and the application potential of equipment health monitoring and predictive maintenance.
The CVGAN model is constructed, and the vibration data of the bearing is generated by the autoregressive method. The conditional variational autoencoder and the generation adversarial network are used to combine historical vibration data and HI tags to generate high-quality vibration signals.
Improve the clarity and practicality of the vibration signal, enhance the application feasibility and effectiveness of the model, realize the continuous generation of vibration data for the entire life cycle of rolling bearings, and provide richer and more accurate data support for mechanical health monitoring and prediction. .
Smart Images

Figure CN118504623B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data generation, and in particular to a method, device, equipment and storage medium for generating bearing vibration data. Background Art
[0002] Rolling bearings are one of the important components of rotating mechanical equipment. Their performance will gradually degrade during long-term operation, so predictive maintenance is required. Currently, one method of maintaining bearings is to predict the remaining life of rolling bearings. With the development of advanced sensing technology, it has become easy to obtain a large amount of industrial data such as bearing vibration signals. However, in many cases, only limited data can be collected, which makes the network model ineffective. Therefore, data generation methods have become an effective way to solve the small sample problem. In the prior art, although generative models have demonstrated their powerful capabilities in many fields, such as the generation of image and audio data, most of these models are limited to generating data at a specific moment. In the field of vibration data generation, this limitation is particularly obvious. Existing models are difficult to effectively generate continuous vibration data covering the entire life cycle of the equipment, which not only affects the effectiveness of the model in practical applications, but also limits its application potential in equipment health monitoring and predictive maintenance. Summary of the invention
[0003] In order to solve the above technical problems, the present invention provides a method, device, equipment and storage medium for generating bearing vibration data. By constructing a CVGAN model and adopting an autoregressive method to generate bearing vibration data, high-quality generation of vibration data of the rolling bearing throughout its life cycle is achieved, and the clarity and practicality of the generated vibration signal are improved.
[0004] An embodiment of the present invention provides a method for generating bearing vibration data, comprising:
[0005] Collecting raw vibration data of several bearing samples, and preprocessing the raw vibration data to obtain corresponding normalized data, historical data and HI labels, and constructing a triplet data set;
[0006] Combining a conditional variational autoencoder and a generative adversarial network, using the decoder of the conditional variational autoencoder as the generator of the generative adversarial network to construct a generative model; wherein the generative model includes an encoder, a generator, a discriminator, and a classifier;
[0007] Training the generative model according to the triplet data set to obtain a trained generative model;
[0008] Based on the generator in the trained generation model, the data sampled from the normal distribution is used as input, the preset HI label and the historical vibration data are used as conditions, and the vibration data is generated by autoregression; wherein, the historical vibration data of the current round is obtained by splicing the vibration data generated in the previous round with the historical vibration data of the previous round, and the historical vibration data of the first round is generated by the preset initial generator.
[0009] As an improvement of the above solution, the raw vibration data is preprocessed to obtain corresponding normalized data, historical data and HI labels, including:
[0010] Performing low-pass filtering on the raw vibration data;
[0011] Normalizing the filtered original vibration data to obtain corresponding normalized data;
[0012] Calculating the kurtosis of the normalized data, and calculating the initial prediction point using the 3sigma method according to the kurtosis;
[0013] According to the initial prediction point, construct an HI curve corresponding to each normalized data as an HI label;
[0014] The normalized data corresponding to each of the bearing samples is average pooled, and a sliding window is used to construct the historical data corresponding to the normalized data.
[0015] As an improvement of the above solution, when the generative model is trained according to the triple data set, the normalized data, the historical data and the HI label are spliced on the channel, specifically including:
[0016] Before the data is input into the encoder of the generative model, the HI tag is first input into the adapter to generate a first embedding vector of 1*512; the first embedding vector, the normalized data and the historical data are concatenated and input into the encoder;
[0017] Before the data is input into the generator of the generative model, the latent variables output by the encoder are reparameterized to generate a 1*32 first vector sampled from a normal distribution; the HI tag is input into the adapter to generate a 1*32 second embedding vector; the historical data is average pooled and downsampled to generate a 1*32 second vector; the first vector, the second embedding vector and the second vector are concatenated and input into the generator.
[0018] As an improvement of the above solution, the network structure of the generation model includes:
[0019] The encoder includes five convolution blocks, each of which includes a convolution layer, a BN layer, and a LeakyReLU layer; wherein the step sizes of the first four convolution blocks are all 2, and the number of channels is 16, 32, 64, and 128 in order; the step size of the fifth convolution block is 1, and the number of channels is 32; the encoder also includes an output layer: a sigmoid layer;
[0020] The generator includes six transposed convolution blocks, and the number of channels of the transposed convolution blocks is 128, 64, 32, 32, 16, and 2 in sequence; wherein the last transposed convolution block generates a vibration image of two channels and passes it into the discriminator and the classifier;
[0021] The discriminator comprises five convolution blocks, an average pooling layer, a dropout layer and a fully connected layer of the same architecture as the encoder; wherein the fully connected layer of the discriminator outputs the classification of the real image and the generated image;
[0022] The classifier includes five convolution blocks, an average pooling layer, a Dropout layer and a fully connected layer. The first four convolution blocks of the classifier are the same as the first four convolution blocks of the encoder, and the number of channels of the last convolution block is 256; wherein the fully connected layer of the classifier outputs 32 label categories.
[0023] As an improvement of the above solution, when the generation model is trained according to the triple data set, the method further includes:
[0024] The maximum mean difference value and FID value between the generated data of the generation model and the real data are used as evaluation indicators, and the training parameters are optimized and adjusted until the corresponding maximum mean difference value and FID value are less than the preset values. The training is stopped and the parameters of the generation model are fixed to obtain a trained generation model.
[0025] As an improvement of the above solution, the generator based on the trained generative model takes the data sampled from the normal distribution as input, takes the preset HI label and the historical vibration data as conditions, and generates vibration data in an autoregressive manner, specifically including:
[0026] Sampling from a normal distribution to obtain first sampling data;
[0027] Input the preset HI tag into the adapter to obtain the corresponding tag data;
[0028] The first data of the previous round of historical vibration data is discarded, and the vibration data generated in the previous round is added to the end of the historical vibration data to obtain the historical vibration data of the current round; the historical vibration data of the current round is downsampled to obtain the second sampling data; wherein the historical vibration data of the first round is generated by a preset initial generator;
[0029] The first sampling data, the label data and the second sampling data are concatenated and input into a generator in a trained generation model; the generator restores the first sampling data based on the label data and the second sampling data to generate vibration data.
[0030] As an improvement of the above solution, the preset initial generator is an unconditional generative adversarial network model, and the training method of the initial generator includes:
[0031] Using the normalized data of each bearing sample before the corresponding initial prediction point as a training data set to train the initial generator;
[0032] The maximum mean difference value and the FID value are used as evaluation indicators to optimize and adjust the training parameters of the initial generator until the corresponding maximum mean difference value and the FID value are less than the preset values, stop training, and fix the network parameters of the initial generator.
[0033] The embodiment of the present invention further provides a bearing vibration data generating device, comprising:
[0034] A training data module is used to collect the original vibration data of several bearing samples, and pre-process the original vibration data to obtain the corresponding normalized data, historical data and HI labels, and construct a triplet data set;
[0035] A generative model module, which is used to combine a conditional variational autoencoder and a generative adversarial network, use the decoder of the conditional variational autoencoder as the generator of the generative adversarial network, and construct a generative model; wherein the generative model includes an encoder, a generator, a discriminator, and a classifier;
[0036] A model training module, used for training the generation model according to the triple data set to obtain a trained generation model;
[0037] A data generation module is used to generate vibration data in an autoregressive manner based on the generator in the trained generation model, taking the data sampled from the normal distribution as input, taking the preset HI label and the historical vibration data as conditions; wherein the historical vibration data of the current round is obtained by concatenating the vibration data generated in the previous round with the historical vibration data of the previous round, and the historical vibration data of the first round is generated by the preset initial generator.
[0038] An embodiment of the present invention also provides a terminal device, including a processor and a memory, wherein a computer program is stored in the memory, and the computer program is configured to be executed by the processor, and when the processor executes the computer program, the bearing vibration data generation method described in any one of the above items is implemented.
[0039] An embodiment of the present invention further provides a computer-readable storage medium, which stores a computer program. When the computer program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned bearing vibration data generation methods.
[0040] Compared with the prior art, the beneficial effects of a bearing vibration data generation method, device, equipment and storage medium provided by the embodiment of the present invention are: by constructing a CVGAN model, it is possible to use historical vibration data and HI labels as input conditions to directly generate one-dimensional horizontal and vertical vibration signals with physical significance. Compared with the traditional variational autoencoder VAE and the generative adversarial network GAN, CVGAN improves the clarity and practicality of the generated vibration signal, and greatly enhances the feasibility and effectiveness of the model in practical applications; by adopting an autoregressive generation strategy, it is possible to efficiently reuse the historical vibration data generated by the model to provide more accurate guidance for data generation at the current moment, which not only significantly improves the continuity and consistency of data generation, but also simulates the dynamic changes in the actual operation of mechanical equipment to a certain extent, thereby providing richer and more accurate data support for mechanical health monitoring and prediction; by performing a series of refined preprocessing steps on the collected data, the performance and accuracy of the subsequent deep learning model are improved; by using the maximum mean difference value and FID value as evaluation indicators to optimize the training parameters, it is possible to more accurately reflect the similarity between the generated vibration signal and the actual signal when evaluating the results, and provide a reliable quantitative method for evaluating the quality of the generated model. The embodiments of the present invention can effectively use historical vibration data to guide the generation of current vibration data, realize the continuous generation of vibration data of the rolling bearing throughout its life cycle, and provide reliable data for the prediction of the remaining life of the bearing. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 It is a flow chart of a method for generating bearing vibration data provided by an embodiment of the present invention;
[0042] Figure 2 It is a schematic diagram of the structure and operation of a CVGAN model provided by an embodiment of the present invention;
[0043] Figure 3It is a schematic diagram of a process of inputting training data into a CVGAN model provided by an embodiment of the present invention;
[0044] Figure 4 It is a flow chart of a method for generating vibration data by autoregression provided by an embodiment of the present invention;
[0045] Figure 5 It is a RMS analysis result diagram of data generated in the horizontal direction (left figure) and the vertical direction (right figure) in a bearing vibration data generation method provided by an embodiment of the present invention;
[0046] Figure 6 It is a t-SNE analysis result diagram of data generated in the horizontal direction (left figure) and the vertical direction (right figure) in a bearing vibration data generation method provided by an embodiment of the present invention;
[0047] Figure 7 It is a structural schematic diagram of a bearing vibration data generating device provided by an embodiment of the present invention;
[0048] Figure 8 It is a structural diagram of a terminal device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0050] See also Figure 1 , Figure 1 1 is a flow chart of a method for generating bearing vibration data provided by an embodiment of the present invention. The method for generating bearing vibration data comprises:
[0051] S1: Collecting the original vibration data of several bearing samples, and preprocessing the original vibration data to obtain the corresponding normalized data, historical data and HI labels, and constructing a triplet data set;
[0052] S2: Combining a conditional variational autoencoder and a generative adversarial network, using the decoder of the conditional variational autoencoder as the generator of the generative adversarial network to construct a generative model; wherein the generative model includes an encoder, a generator, a discriminator, and a classifier;
[0053] S3: training the generative model according to the triplet data set to obtain a trained generative model;
[0054] S4: Based on the generator in the trained generation model, the data sampled from the normal distribution is taken as input, the preset HI label and the historical vibration data are taken as conditions, and the vibration data is generated by autoregression; wherein, the historical vibration data of the current round is obtained by concatenating the vibration data generated in the previous round with the historical vibration data of the previous round, and the historical vibration data of the first round is generated by the preset initial generator.
[0055] Specifically, in the experiment, the bearing was placed under a heavy load to accelerate its failure process. The present invention collects vibration sensing data through an accelerometer installed on the bearing. Data collection involves two accelerometers installed in the horizontal and vertical directions, respectively. The sampling frequency of these accelerometers is set to 25.6 thousand times per second. They collect data once every 10 seconds, each time lasting 0.1 seconds, so that each data sample contains 2560 data points. The working conditions corresponding to the bearing vibration data collected in this embodiment are a load of 4000N and a speed of 1800r / min. In order to ensure the safety of the test, if the amplitude of the vibration data exceeds 20g, the test is terminated. This moment is the time point of the bearing failure. The collected vibration data includes data from continuous operation to failure.
[0056] After obtaining the raw vibration data of several bearing samples, a series of refined preprocessing steps are performed on the raw vibration data to improve the performance and accuracy of the subsequent deep learning model. First, the raw vibration data is filtered and normalized to adjust all data to the range of 0 to 1. In addition, the kurtosis of the data is calculated and the 3sigma method is applied to determine the initial prediction point for each bearing in order to more accurately capture the key features in the data. Finally, a triplet dataset containing historical data, normalized data, and corresponding category labels is constructed through the sliding window method.
[0057] The structure of the autoencoder AE is divided into an encoder and a decoder. The encoder can compress the input data into a potential representation, and the decoder can reconstruct the compressed potential form into the initial input. The basic principle of the autoencoder is to reduce the difference between the original data and its reconstructed data; the variational autoencoder VAE is an improvement of the autoencoder. Its core lies in its ability to map the input data to a probability distribution in a latent space, rather than mapping it to a fixed encoding vector. This probabilistic encoding and decoding process increases the flexibility of the model; the conditional variational encoder CVAE is an improvement of the variational autoencoder. It combines the idea of the conditional probability model and introduces additional conditional variables, so that the data generation process can be controlled. The generative adversarial network GAN includes two competing networks, namely the generator and the discriminator. The goal of the generator is to generate realistic data, while the task of the discriminator is to distinguish between generated data and real data. The two networks compete with each other during the training process and continuously improve their respective performance.
[0058] The embodiment of the present invention combines the conditional variational autoencoder CVAE and the generative adversarial network GAN to construct a conditional variational generative adversarial network CVGAN, and uses the decoder in the conditional variational autoencoder as the generator in the generative adversarial network. The model adopts the self-supervised learning method in the field of deep learning, which uses historical vibration data and HI labels as conditional inputs and is trained through normalized data of vibration data. After the CVGAN model training is completed, the generator in the CVGAN model is used to generate a series of coherent and realistic vibration data based on historical vibration information and remaining life information in an autoregressive manner, which is a one-dimensional vibration signal in the horizontal and vertical directions. This generation method can not only simulate complex vibration signal patterns, but also predict future vibration signal change trends based on set conditions.
[0059] As one of the optional embodiments, the preprocessing of the original vibration data to obtain the corresponding normalized data, historical data and HI label includes:
[0060] Performing low-pass filtering on the raw vibration data;
[0061] Normalizing the filtered original vibration data to obtain corresponding normalized data;
[0062] Calculating the kurtosis of the normalized data, and calculating the initial prediction point using the 3sigma method according to the kurtosis;
[0063] According to the initial prediction point, construct an HI curve corresponding to each normalized data as an HI label;
[0064] The normalized data corresponding to each of the bearing samples is average pooled, and a sliding window is used to construct the historical data corresponding to the normalized data.
[0065] Specifically, the original vibration data is first filtered using a Butterworth low-pass filter, with a cutoff frequency of 1500 Hz and a filter order of 4 to effectively remove high-frequency noise. After filtering, the data is normalized to ensure that the data values are within a uniform numerical range. The specific formula is:
[0066] ;
[0067] in, To normalize the data, is the original vibration data, and They are the minimum and maximum values in the original vibration data respectively.
[0068] Furthermore, the kurtosis of the normalized data is calculated :
[0069] ;
[0070] in, To normalize the data, yes The mean of , N is the total number of samples.
[0071] Furthermore, the 3sigma method is used to calculate the initial prediction point (FPT) of each bearing, where the basic theory of the 3sigma method is the Chebyshev inequality, which can be expressed as follows:
[0072] ;
[0073] in, is a mean and standard deviation The kurtosis, represents the probability of an event, is a positive number. When set to 3, any data point that falls within 3 The probability of being outside the distance is less than 1 / 9. However, for the actual bearing vibration signal, there are still some abnormal values caused by noise in the normal operation stage, which will affect the accuracy of FPT selection. In this regard, the embodiment of the present invention adopts a strategy for determining FPT, that is, if there are five consecutive data points falling within 3 from time i, In addition to the restrictions, the time i is determined as FPT, and the formula is: ;in, The preferred setting is 5.
[0074] Furthermore, according to the FPT of each bearing, a piecewise function is used to simulate the degradation trend of the bearing to obtain the HI curve corresponding to each bearing. The HI of the bearing remains unchanged at 1 from 0 to the initial prediction time (FPT), and then linearly decays from FPT to obtain the HI label corresponding to each normalized data. In order to reduce the burden of model condition generation, the HI label from 0 to 1 is divided into 32 categories, and the label category is represented as .
[0075] Furthermore, the data of each bearing sample is average pooled to unify the feature dimension of the data to 2*N feature , N feature The preferred setting is 512. Then, a sliding window with a window size of k is used to construct historical data, that is, normalized data of every k length is taken as a sample. Since the vibration data has two channels, horizontal and vertical, the dimension of the historical data v is 2k*N. feature Among them, the first window corresponds to the first input data of the model , and its corresponding historical data is ~ , the corresponding HI tag is Then, starting from the first window, the intercepted window is moved back one time unit along the sampling time axis to obtain the next input data, and the above operation is repeated until the last window. Since the last window lacks a label, no data construction is performed, and a series of input samples are obtained, which includes historical data v, normalized data And a triplet dataset of HI labels y.
[0076] As one of the optional embodiments, when the generative model is trained according to the triple data set, the normalized data, the historical data and the HI label are spliced on the channel, specifically including:
[0077] Before the data is input into the encoder of the generative model, the HI tag is first input into the adapter to generate a first embedding vector of 1*512; the first embedding vector, the normalized data and the historical data are concatenated and input into the encoder;
[0078] Before the data is input into the generator of the generative model, the latent variables output by the encoder are reparameterized to generate a 1*32 first vector sampled from a normal distribution; the HI tag is input into the adapter to generate a 1*32 second embedding vector; the historical data is average pooled and downsampled to generate a 1*32 second vector; the first vector, the second embedding vector and the second vector are concatenated and input into the generator.
[0079] As one of the optional embodiments, the network structure of the generation model includes:
[0080] The encoder includes five convolution blocks, each of which includes a convolution layer, a BN layer, and a LeakyReLU layer; wherein the step sizes of the first four convolution blocks are all 2, and the number of channels is 16, 32, 64, and 128 in order; the step size of the fifth convolution block is 1, and the number of channels is 32; the encoder also includes an output layer: a sigmoid layer;
[0081] The generator includes six transposed convolution blocks, and the number of channels of the transposed convolution blocks is 128, 64, 32, 32, 16, and 2 in sequence; wherein the last transposed convolution block generates a vibration image of two channels and passes it into the discriminator and the classifier;
[0082] The discriminator comprises five convolution blocks, an average pooling layer, a dropout layer and a fully connected layer of the same architecture as the encoder; wherein the fully connected layer of the discriminator outputs the classification of the real image and the generated image;
[0083] The classifier includes five convolution blocks, an average pooling layer, a Dropout layer and a fully connected layer. The first four convolution blocks of the classifier are the same as the first four convolution blocks of the encoder, and the number of channels of the last convolution block is 256; wherein the fully connected layer of the classifier outputs 32 label categories.
[0084] For details, please refer to Figure 2 , Figure 2 This is a schematic diagram of the structure and operation of a CVGAN model provided by an embodiment of the present invention. CVGAN includes an encoder, a generator (decoder), a discriminator and a classifier. When the model is working, the encoder maps the input data to a latent space based on the input conditions of the historical vibration data and the HI label to obtain the corresponding latent vector; then the data is reconstructed through the generator and the corresponding vibration data is output; then the discriminator of the GAN part tries to distinguish between the real data and the generated data, and outputs the corresponding judgment result; the classifier tries to distinguish the different categories of the generated data and corresponds them to 32 HI label categories.
[0085] See also Figure 3 , Figure 3This is a flow chart of the input training data of a CVGAN model provided by an embodiment of the present invention, which specifically shows the network structure of the encoder and decoder. When training the model, the label y in the training data is first input into an adapter Adapter, and a first embedding vector Embedding of dimension 1*512 is generated through a ReLU activation function and a fully connected layer FC-512; then the historical data v and the normalized data and the first embedding vector are spliced and input into the encoder; then the encoder outputs the corresponding latent vector, and after re-parameterizing the latent vector, a vector sampled from the normal distribution is generated, namely the first vector, whose dimension is 1*32; before inputting the decoder, the HI label y is also input into an adapter Adapter, and a second embedding vector Embedding with a dimension of 1*32 is generated through a ReLU activation function and a fully connected layer FC-32; the historical data v is average pooled and downsampled to a size of 32, and then the processed historical data v is spliced with the second embedding vector and the sampled first vector to obtain the input of the generator.
[0086] Furthermore, as shown in Table 1, the specific network structure of the encoder is shown. Its backbone network consists of five convolutional blocks, each of which consists of a convolutional layer, a BN layer, and a LeakyReLU layer. The convolution kernel sizes in the first four convolutional blocks are , the step size is 2, the padding is 1, and the size of the feature map is halved and the channel is doubled after each convolution block; the step size of the last convolution block is 1 and the channel is 32; the network finally passes through a sigmoid layer as the output layer. The decoder's network structure is similar to the encoder, which includes six transposed convolution blocks. The sizes of the convolution kernels in the first five transposed convolution blocks are , with a step size of 2 and padding of 1; the last transposed convolution block generates a vibration image of two channels, which is then passed to the discriminator and classifier. The classifier and discriminator also use an encoder-like architecture in the network, but the last convolution kernel of the classifier has 256 channels; after the convolution block, the feature map size is changed to 1 through the average pooling layer, and then it passes through the Dropout layer and the fully connected layer in sequence; among them, the fully connected layer of the classifier outputs 32 category information, and the fully connected layer of the discriminator outputs 1 classification information whether it is a real image.
[0087] Table 1 Encoder specific network structure table
[0088]
[0089] Specifically, the objective function of CVGAN is as follows:
[0090] ;
[0091] ;
[0092] ;
[0093] ;
[0094] ;
[0095] ;
[0096] in, is the input real vibration image, is the image reconstructed by CVAE, is the reconstruction loss function, is the feature matching loss function, is the KL divergence loss function, is the overall loss function of CVAE, is the classifier loss function, is the discriminator loss function, D(x|y) represents the output of the discriminator, and C(x|y) represents the output of the classifier. represents the intermediate features of the classifier, Represents the intermediate features of the discriminator; is the distribution of real data; is the distribution that generated the data.
[0097] As one of the optional embodiments, when the generation model is trained according to the triple data set, the method further includes:
[0098] The maximum mean difference value and FID value between the generated data of the generation model and the real data are used as evaluation indicators, and the training parameters are optimized and adjusted until the corresponding maximum mean difference value and FID value are less than the preset values. The training is stopped and the parameters of the generation model are fixed to obtain a trained generation model.
[0099] Specifically, in the training of the generative model, the model performance needs to be evaluated to measure the model training effect. In the evaluation process of the model, considering that the mean absolute deviation MAD only provides a rough comparison of the mean difference in the normal stage, in order to avoid the deviation of the evaluation results, the embodiment of the present invention selects the maximum mean difference MMD and Fréchet Inception Distance (FID) as evaluation criteria. These two indicators focus on measuring the difference between the generated data and the real data distribution, so as to more accurately reflect the similarity between the generated vibration signal and the actual signal. According to MMD and FID, the training parameters are carefully adjusted until the model performance reaches the expected optimization goal and the model training is completed.
[0100] It should be noted that during the training of the generative model, the AdamW optimizer is used to minimize the value of the loss function. For the training parameters, the learning rate of the encoder and generator is preferably set to 0.0006, the learning rate of the discriminator and classifier is preferably set to 0.0002, the batch size is preferably set to 1024, the early stop is set to 30, and the network training is set to 100 epochs.
[0101] As one of the optional embodiments, the generator based on the trained generative model uses the data sampled from the normal distribution as input, uses the preset HI label and the historical vibration data as conditions, and generates vibration data in an autoregressive manner, specifically including:
[0102] Sampling from a normal distribution to obtain first sampling data;
[0103] Input the preset HI tag into the adapter to obtain the corresponding tag data;
[0104] The first data of the previous round of historical vibration data is discarded, and the vibration data generated in the previous round is added to the end of the historical vibration data to obtain the historical vibration data of the current round; the historical vibration data of the current round is downsampled to obtain the second sampling data; wherein the historical vibration data of the first round is generated by a preset initial generator;
[0105] The first sampling data, the label data and the second sampling data are concatenated and input into a generator in a trained generation model; the generator restores the first sampling data based on the label data and the second sampling data to generate vibration data.
[0106] For details, please refer to Figure 4 , Figure 4It is a flow chart of a method for generating vibration data by autoregression provided by an embodiment of the present invention. When generating data, we discard the classifier, encoder and discriminator, and only use the trained generator to generate vibration data, which is directly sampled from the normal distribution, and then restored to a vibration signal by the generator, wherein the generator receives HI label category information and historical vibration data as conditions, the HI label category information is preset information, the historical vibration data is obtained from the previously generated vibration data, and the data is generated by autoregression. Due to the lack of initial historical vibration data, the present invention uses an initial generator to generate historical vibration data of a window size, and subsequent historical vibration data is spliced by the currently generated vibration data and the historical vibration data. Specifically, the size of the historical vibration data is the window size, and the first data is discarded each time, and then the vibration data generated in this round is added to the last one for the next round of generation.
[0107] As one of the optional embodiments, the preset initial generator is an unconditional generative adversarial network model, and the training method of the initial generator includes:
[0108] Using the normalized data of each bearing sample before the corresponding initial prediction point as a training data set to train the initial generator;
[0109] The maximum mean difference value and the FID value are used as evaluation indicators to optimize and adjust the training parameters of the initial generator until the corresponding maximum mean difference value and the FID value are less than the preset values, stop training, and fix the network parameters of the initial generator.
[0110] Specifically, the initial generator is a GAN model, and its specific network structure is shown in Table 2 below.
[0111] Table 2 Specific network structure of initial generator
[0112]
[0113] Since the initial generator is an unconditional model, only the vibration data of the vibration data are used. To ensure that all The HI values of all bearings are the same. Since the FPT of all bearings is greater than 8000s, only the first 800 samples of each bearing are taken, that is, the first 8000s of life. During training, the original GAN loss is used for training, and its loss function is:
[0114] ;
[0115] ;
[0116] in, is the input real vibration image, is the corresponding condition information, is a tensor sampled from a normal distribution, is the generator loss function, is the discriminator loss function, D(x|y) represents the output of the discriminator, and G(z|y) represents the output of the generator.
[0117] Furthermore, for the training parameters of the initial generator, the AdamW optimizer was used to minimize the value of the loss function, the learning rate of the generator was set to 0.0003, the learning rate of the discriminator was 0.0006, the batch size was set to 1024, the network was trained for 100 epochs, and the early stop was set to 40. During training, FID and MMD were also used to evaluate the generation results. When the results reached the set threshold range, the training was stopped, the network parameters were fixed, and the model was used to generate the initial historical vibration data.
[0118] Furthermore, the total remaining life of the generated bearing is manually set to 1000 and the FPT is set to 300, and the above autoregressive generation is performed. By utilizing the advantageous characteristics of the autoregressive model, the embodiment of the present invention can effectively utilize historical vibration data to guide the generation of vibration data in the current window, thereby realizing the continuous generation of time series data.
[0119] See also Figure 5 and Figure 6 , Figure 5 It is a RMS analysis result diagram of data generated in the horizontal direction (left figure) and the vertical direction (right figure) in a bearing vibration data generation method provided by an embodiment of the present invention; Figure 6 This is a t-SNE analysis result diagram of the data generated in the horizontal direction (left figure) and the vertical direction (right figure) in a bearing vibration data generation method provided by an embodiment of the present invention. It can be seen that in the data generated by CVGAN, except for the first window, the RMS fluctuation of the remaining data is very small, which is in line with the normal data fluctuation range. The large fluctuation at the beginning is due to the fact that the effect of the initial generator through unconditional generation is not very good, which can be eliminated in practical applications. The differences between different categories in t-SNE are more obvious, and each category can be fully identified in the vertical direction, which shows that the data generation method proposed by the present invention can simulate the distribution of the original data more accurately.
[0120] Furthermore, the data generated by CVGAN in the embodiment of the present invention through autoregression not only enriches the training set of the bearing remaining life prediction model, but also simulates various complex vibration modes that may be encountered under actual operating conditions, thereby providing additional, high-quality training samples for the remaining life prediction model. Furthermore, the present invention provides a model based on convolutional neural network CNN and a hybrid model combining CNN and gated recurrent unit GRU to perform remaining life prediction and quantitatively analyze the data generated by the generated model.
[0121] Specifically, the architecture of the prediction network starts with two one-dimensional convolutional layers, which are responsible for extracting the features of the vibration signal and increasing the number of channels of the data. In order to reduce the size of the feature map, the network then applies a maximum pooling layer and a convolutional layer with a stride, each of which is followed by a normalization layer and a nonlinear activation function ReLU to enhance the model's expressiveness and nonlinear processing capabilities. Subsequently, the network further extracts time series features through three layers of unidirectional GRU units to understand the temporal dependencies in the vibration data. In order to convert the extracted time series features into a fixed dimension, the network uses an average pooling operation to unify the feature dimension to 1. Finally, the prediction results of the remaining life RUL of the bearing are output through two fully connected layers. Among them, after the first fully connected layer, the network introduces a Dropout (p=0.5) layer and a ReLU activation function to improve the generalization ability of the model and avoid overfitting. The prediction model uses the MSE loss function for supervised training. The prediction model also uses the AdamW optimizer, with a learning rate set to 0.0008 and a batch size set to 2048. The network was trained for 150 epochs, and the early stop was set to 20. The prediction model was trained using a two-tuple data set consisting of the historical data v of the bearing and its corresponding HI label y. When the MSE value reached the set threshold, the iterative training was stopped and the model was saved.
[0122] After the prediction model training is completed, three widely recognized indicators are used to measure the model results: root mean square error RMSE, mean absolute error MAE and Score. As shown in Table 3 and Table 4, Table 3 is the statistical indicators of the CNN prediction model, and Table 4 is the statistical indicators of the CNN+GRU prediction model.
[0123] Table 3 CNN prediction model statistical indicators
[0124]
[0125] Table 4 Statistical indicators of CNN+GRU prediction model
[0126]
[0127] It can be found that the CNN prediction model performs very poorly without additional training data, with MAE greater than 0.4; while the training with additional data generated by CVGAN on the bearing samples has greatly improved the results. In the CNN+GRU prediction model, the use of additional data generated by CVGAN can bring a slight improvement in the prediction effect of bearing sample two, but the use of additional data on bearing sample one cannot bring positive benefits. It can be seen that when the CNN+GRU model effect is already very good, only sufficiently high-quality generated data can bring about an improvement in prediction performance. The prediction effect of the CNN model itself is very poor, and any additional training data can improve the performance of the model. When the quality of the additional training data is average, its prediction effect can beat models with much stronger performance. The bearing vibration data generation method provided in an embodiment of the present invention can significantly enhance the accuracy of RUL prediction, especially in scenarios where the model itself is not powerful and the data is very limited.
[0128] By constructing a CVGAN model, the embodiment of the present invention can use historical vibration data and HI labels as input conditions to directly generate one-dimensional horizontal and vertical vibration signals with physical significance. Compared with the traditional variational autoencoder VAE and the generative adversarial network GAN, CVGAN improves the clarity and practicality of the generated vibration signal, and greatly enhances the feasibility and effectiveness of the model in practical applications. By adopting an autoregressive generation strategy, the historical vibration data generated by the model can be efficiently reused to provide more accurate guidance for data generation at the current moment, which not only significantly improves the continuity and consistency of data generation, but also simulates the dynamic changes in the actual operation of mechanical equipment to a certain extent, thereby providing richer and more accurate data support for mechanical health monitoring and prediction. By performing a series of refined preprocessing steps on the collected data, the performance and accuracy of the subsequent deep learning model are improved. By adopting the maximum mean difference value and FID value as evaluation indicators to optimize the training parameters, the similarity between the generated vibration signal and the actual signal can be more accurately reflected in the evaluation results, providing a reliable quantitative method for evaluating the quality of the generated model. The embodiments of the present invention can effectively use historical vibration data to guide the generation of current vibration data, realize the continuous generation of vibration data of the rolling bearing throughout its life cycle, and provide reliable data for the prediction of the remaining life of the bearing.
[0129] Correspondingly, the present invention also provides a bearing vibration data generating device, which can implement all the processes of the bearing vibration data generating method in the above embodiment.
[0130] See also Figure 7 , Figure 71 is a schematic diagram of a bearing vibration data generating device provided by an embodiment of the present invention. The bearing vibration data generating device comprises:
[0131] The training data module 701 is used to collect the original vibration data of several bearing samples, and pre-process the original vibration data to obtain the corresponding normalized data, historical data and HI labels, and construct a triplet data set;
[0132] A generative model module 702 is used to combine the conditional variational autoencoder and the generative adversarial network, use the decoder of the conditional variational autoencoder as the generator of the generative adversarial network, and construct a generative model; wherein the generative model includes an encoder, a generator, a discriminator, and a classifier;
[0133] A model training module 703 is used to train the generation model according to the triple data set to obtain a trained generation model;
[0134] The data generation module 704 is used to generate vibration data by autoregression based on the generator in the trained generation model, taking the data sampled from the normal distribution as input, taking the preset HI label and the historical vibration data as conditions; wherein the historical vibration data of the current round is obtained by concatenating the vibration data generated in the previous round with the historical vibration data of the previous round, and the historical vibration data of the first round is generated by the preset initial generator.
[0135] Preferably, the preprocessing of the original vibration data to obtain corresponding normalized data, historical data and HI labels includes:
[0136] Performing low-pass filtering on the raw vibration data;
[0137] Normalizing the filtered original vibration data to obtain corresponding normalized data;
[0138] Calculating the kurtosis of the normalized data, and calculating the initial prediction point using the 3sigma method according to the kurtosis;
[0139] According to the initial prediction point, construct an HI curve corresponding to each normalized data as an HI label;
[0140] The normalized data corresponding to each of the bearing samples is average pooled, and a sliding window is used to construct the historical data corresponding to the normalized data.
[0141] Preferably, when the generative model is trained according to the triplet data set, the normalized data, the historical data and the HI label are spliced on the channel, specifically including:
[0142] Before the data is input into the encoder of the generative model, the HI tag is first input into the adapter to generate a first embedding vector of 1*512; the first embedding vector, the normalized data and the historical data are concatenated and input into the encoder;
[0143] Before the data is input into the generator of the generative model, the latent variables output by the encoder are reparameterized to generate a 1*32 first vector sampled from a normal distribution; the HI tag is input into the adapter to generate a 1*32 second embedding vector; the historical data is average pooled and downsampled to generate a 1*32 second vector; the first vector, the second embedding vector and the second vector are concatenated and input into the generator.
[0144] Preferably, the network structure of the generative model includes:
[0145] The encoder includes five convolution blocks, each of which includes a convolution layer, a BN layer, and a LeakyReLU layer; wherein the step sizes of the first four convolution blocks are all 2, and the number of channels is 16, 32, 64, and 128 in order; the step size of the fifth convolution block is 1, and the number of channels is 32; the encoder also includes an output layer: a sigmoid layer;
[0146] The generator includes six transposed convolution blocks, and the number of channels of the transposed convolution blocks is 128, 64, 32, 32, 16, and 2 in sequence; wherein the last transposed convolution block generates a vibration image of two channels and passes it into the discriminator and the classifier;
[0147] The discriminator comprises five convolution blocks, an average pooling layer, a dropout layer and a fully connected layer of the same architecture as the encoder; wherein the fully connected layer of the discriminator outputs the classification of the real image and the generated image;
[0148] The classifier includes five convolution blocks, an average pooling layer, a Dropout layer and a fully connected layer. The first four convolution blocks of the classifier are the same as the first four convolution blocks of the encoder, and the number of channels of the last convolution block is 256; wherein the fully connected layer of the classifier outputs 32 label categories.
[0149] Preferably, the device further comprises:
[0150] The maximum mean difference value and FID value between the generated data of the generation model and the real data are used as evaluation indicators, and the training parameters are optimized and adjusted until the corresponding maximum mean difference value and FID value are less than the preset values. The training is stopped and the parameters of the generation model are fixed to obtain a trained generation model.
[0151] Preferably, the generator based on the trained generative model uses the data sampled from the normal distribution as input, takes the preset HI label and the historical vibration data as conditions, and generates vibration data in an autoregressive manner, specifically including:
[0152] Sampling from a normal distribution to obtain first sampling data;
[0153] Input the preset HI tag into the adapter to obtain the corresponding tag data;
[0154] The first data of the previous round of historical vibration data is discarded, and the vibration data generated in the previous round is added to the end of the historical vibration data to obtain the historical vibration data of the current round; the historical vibration data of the current round is downsampled to obtain the second sampling data; wherein the historical vibration data of the first round is generated by a preset initial generator;
[0155] The first sampling data, the label data and the second sampling data are concatenated and input into a generator in a trained generation model; the generator restores the first sampling data based on the label data and the second sampling data to generate vibration data.
[0156] Preferably, the preset initial generator is an unconditional generative adversarial network model, and the device further comprises:
[0157] The initial generator module is used to use the normalized data of each bearing sample before the corresponding initial prediction point as a training data set to train the initial generator; use the maximum mean difference value and FID value as evaluation indicators to optimize and adjust the training parameters of the initial generator until the corresponding maximum mean difference value and FID value are less than the preset value, stop training, and fix the network parameters of the initial generator.
[0158] In specific implementation, the working principle, control process and technical effect achieved by the bearing vibration data generating device provided in the embodiment of the present invention are the same as those of the bearing vibration data generating method in the above embodiment, and will not be repeated here.
[0159] See also Figure 8 , Figure 8 801, a memory 802, and a computer program stored in the memory 802 and configured to be executed by the processor 801. When the processor 801 executes the computer program, the bearing vibration data generation method described in any of the above embodiments is implemented.
[0160] Preferably, the computer program may be divided into one or more modules / units (such as computer program 1, computer program 2, ...), and the one or more modules / units are stored in the memory 802 and executed by the processor 801 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program in the terminal device.
[0161] The processor 801 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor, or the processor 801 may be any conventional processor. The processor 801 is the control center of the terminal device, and various parts of the terminal device are connected using various interfaces and lines.
[0162] The memory 802 mainly includes a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function, etc., and the data storage area can store related data, etc. In addition, the memory 802 can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, and a flash card (Flash Card), etc., or the memory 802 can also be other volatile solid-state storage devices.
[0163] It should be noted that the above terminal device may include, but is not limited to, a processor and a memory. Those skilled in the art will understand that Figure 8 The structural diagram is only an example of the above-mentioned terminal device and does not constitute a limitation of the above-mentioned terminal device. It may include more or less components than shown in the figure, or combine certain components, or different components.
[0164] An embodiment of the present invention further provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the bearing vibration data generation method described in any of the above embodiments.
[0165] The beneficial effects of a bearing vibration data generation method, device, equipment and storage medium provided by an embodiment of the present invention are: by constructing a CVGAN model, it is possible to use historical vibration data and HI labels as input conditions to directly generate one-dimensional horizontal and vertical vibration signals with physical significance. Compared with traditional variational autoencoders VAE and generative adversarial networks GAN, CVGAN improves the clarity and practicality of generated vibration signals, and greatly enhances the feasibility and effectiveness of the model in practical applications; by adopting an autoregressive generation strategy, it is possible to efficiently reuse the historical vibration data generated by the model to provide more accurate guidance for data generation at the current moment, which not only significantly improves the continuity and consistency of data generation, but also simulates the dynamic changes in the actual operation of mechanical equipment to a certain extent, thereby providing richer and more accurate data support for mechanical health monitoring and prediction; by performing a series of refined preprocessing steps on the collected data, the performance and accuracy of the subsequent deep learning model are improved; by using the maximum mean difference value and FID value as evaluation indicators to optimize the training parameters, it is possible to more accurately reflect the similarity between the generated vibration signal and the actual signal when evaluating the results, and provide a reliable quantitative method for evaluating the quality of the generated model. The embodiments of the present invention can effectively use historical vibration data to guide the generation of current vibration data, realize the continuous generation of vibration data of the rolling bearing throughout its life cycle, and provide reliable data for the prediction of the remaining life of the bearing.
[0166] It should be noted that the system embodiment described above is merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the drawings of the system embodiment provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art may understand and implement it without paying any creative effort.
[0167] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A method for generating bearing vibration data, It is characterized in that include: Collect the original vibration data of several bearing samples, and pre-process the original vibration data to obtain the corresponding normalized data, historical data and health index HI label, and construct a triplet data set; Combining a conditional variational autoencoder and a generative adversarial network, using the decoder of the conditional variational autoencoder as the generator of the generative adversarial network to construct a generative model; wherein the generative model includes an encoder, a generator, a discriminator, and a classifier; Training the generative model according to the triplet data set to obtain a trained generative model; Based on the generator in the trained generation model, the data sampled from the normal distribution is used as input, the preset HI label and the historical vibration data are used as conditions, and the vibration data is generated by autoregression; wherein, the historical vibration data of the current round is obtained by splicing the vibration data generated in the previous round with the historical vibration data of the previous round, and the historical vibration data of the first round is generated by the preset initial generator.
2. The bearing vibration data generating method according to claim 1, It is characterized in that The preprocessing of the original vibration data to obtain corresponding normalized data, historical data and HI labels includes: Performing low-pass filtering on the raw vibration data; Normalizing the filtered original vibration data to obtain corresponding normalized data; Calculating the kurtosis of the normalized data, and calculating the initial prediction point using the 3sigma method according to the kurtosis; According to the initial prediction point, construct an HI curve corresponding to each normalized data as an HI label; The normalized data corresponding to each of the bearing samples is average pooled, and a sliding window is used to construct the historical data corresponding to the normalized data.
3. The bearing vibration data generating method according to claim 1, It is characterized in that When the generative model is trained according to the triple data set, the normalized data, the historical data and the HI label are spliced on the channel, specifically including: Before the data is fed into the encoder of the generative model, the HI tag is fed into the adapter to generate 1 512; concatenating the first embedding vector, the normalized data and the historical data, and inputting the concatenated data into the encoder; Before the data is input into the generator of the generative model, the latent variables output by the encoder are reparameterized to generate 1 32 first vector; the HI tag is input into the adapter to generate 1 32; average pooling and downsampling are performed on the historical data to generate 1 32; concatenating the first vector, the second embedding vector and the second vector, and inputting the concatenated vectors into the generator.
4. The bearing vibration data generating method according to claim 1, It is characterized in that The network structure of the generation model includes: The encoder includes five convolution blocks, each of which includes a convolution layer, a BN layer, and a LeakyReLU layer; wherein the step sizes of the first four convolution blocks are all 2, and the number of channels is 16, 32, 64, and 128 in order; the step size of the fifth convolution block is 1, and the number of channels is 32; the encoder also includes an output layer: a sigmoid layer; The generator includes six transposed convolution blocks, and the number of channels of the transposed convolution blocks is 128, 64, 32, 32, 16, and 2 in sequence; wherein the last transposed convolution block generates a vibration image of two channels and passes it into the discriminator and the classifier; The discriminator comprises five convolution blocks, an average pooling layer, a dropout layer and a fully connected layer of the same architecture as the encoder; wherein the fully connected layer of the discriminator outputs the classification of the real image and the generated image; The classifier includes five convolution blocks, an average pooling layer, a Dropout layer and a fully connected layer. The first four convolution blocks of the classifier are the same as the first four convolution blocks of the encoder, and the number of channels of the last convolution block is 256; wherein the fully connected layer of the classifier outputs 32 label categories.
5. The bearing vibration data generating method according to claim 1, It is characterized in that When the generation model is trained according to the triple data set, the method further includes: The maximum mean difference value and FID value between the generated data of the generation model and the real data are used as evaluation indicators, and the training parameters are optimized and adjusted until the corresponding maximum mean difference value and FID value are less than the preset values. The training is stopped and the parameters of the generation model are fixed to obtain a trained generation model.
6. The bearing vibration data generating method according to claim 1, It is characterized in that The generator based on the trained generative model uses the data sampled from the normal distribution as input, uses the preset HI label and the historical vibration data as conditions, and generates vibration data in an autoregressive manner, specifically including: Sampling from a normal distribution to obtain first sampling data; Input the preset HI tag into the adapter to obtain the corresponding tag data; The first data of the previous round of historical vibration data is discarded, and the vibration data generated in the previous round is added to the end of the historical vibration data to obtain the historical vibration data of the current round; the historical vibration data of the current round is downsampled to obtain the second sampling data; wherein the historical vibration data of the first round is generated by a preset initial generator; The first sampling data, the label data and the second sampling data are concatenated and input into a generator in a trained generation model; the generator restores the first sampling data based on the label data and the second sampling data to generate vibration data.
7. The bearing vibration data generating method according to claim 6, It is characterized in that The preset initial generator is an unconditional generative adversarial network model, and the training method of the initial generator includes: Using the normalized data of each bearing sample before the corresponding initial prediction point as a training data set to train the initial generator; The maximum mean difference value and the FID value are used as evaluation indicators to optimize and adjust the training parameters of the initial generator until the corresponding maximum mean difference value and the FID value are less than the preset values, stop training, and fix the network parameters of the initial generator.
8. A bearing vibration data generating device, It is characterized in that include: The training data module is used to collect the original vibration data of several bearing samples, and pre-process the original vibration data to obtain the corresponding normalized data, historical data and health index HI label, and construct a triplet data set; A generative model module, which is used to combine a conditional variational autoencoder and a generative adversarial network, use the decoder of the conditional variational autoencoder as the generator of the generative adversarial network, and construct a generative model; wherein the generative model includes an encoder, a generator, a discriminator, and a classifier; A model training module, used for training the generation model according to the triple data set to obtain a trained generation model; A data generation module is used to generate vibration data in an autoregressive manner based on the generator in the trained generation model, taking the data sampled from the normal distribution as input, taking the preset HI label and the historical vibration data as conditions; wherein the historical vibration data of the current round is obtained by concatenating the vibration data generated in the previous round with the historical vibration data of the previous round, and the historical vibration data of the first round is generated by the preset initial generator.
9. A terminal device, It is characterized in that It comprises a processor and a memory, wherein a computer program is stored in the memory and the computer program is configured to be executed by the processor, and when the processor executes the computer program, the bearing vibration data generating method according to any one of claims 1 to 7 is implemented.
10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a computer program, wherein when the device where the computer-readable storage medium is located executes the computer program, the bearing vibration data generating method according to any one of claims 1 to 7 is implemented.
