CN113486931B - Rolling bearing enhanced diagnosis method based on PDA-WGANGP - Google Patents

Abstract

The invention discloses a rolling bearing enhancement diagnosis method based on PDA-WGANGP, and relates to the field of mechanical equipment fault data enhancement. Firstly, training a pre-training network by using health data and fault data under laboratory conditions, and taking the structure and parameters of the pre-training network as feature extraction layers of the first several layers of a discriminator and a Dense classifier; secondly, introducing a residual error network, constructing a brand new generator, solving the gradient descent problem of the convolutional neural network, and preventing the generation of overfitting, so that a few high-quality samples are stably generated by using an unbalanced training set; finally, model training is completed, high-quality fault samples are generated, and verification is carried out on the generated samples. The method can accurately generate the high-quality fault samples, improve the running efficiency of the model, solve the problem of unbalanced data and improve the accuracy and stability of the generation of the fault samples of the rolling bearing.

Description

An enhanced diagnosis method for rolling bearings based on PDA-WGANGP

Technical Field

The invention relates to the field of bearing fault diagnosis, and in particular to a rolling bearing enhanced diagnosis method based on PDA-WGANGP.

Background Art

As one of the most important components in mechanical and power systems, rolling bearings often work in harsh environments with heavy loads and high-speed rotation, which greatly increases the probability of failure. Once a failure occurs, it will cause huge economic losses or safety accidents. Therefore, it is necessary to monitor the health status of rolling bearings and conduct fault diagnosis research.

The data-driven fault diagnosis method is an important means to detect the health status of equipment. However, in practical applications, rolling bearings are mostly in normal working state, and the collected fault data is relatively small. Therefore, the problem of data imbalance occurs. The so-called data imbalance means that the amount of collected fault data is much smaller than the normal sample, which leads to model bias in the data-driven fault diagnosis method. Therefore, it is of great significance to improve the prediction performance of data imbalance problem, especially the prediction performance of minority category data. The auxiliary classification generative adversarial network based on gradient penalty and Wasserstein distance (ACWGAN-GP) provides a feasible solution to the data imbalance problem in fault diagnosis. ACWGAN-GP can learn the data distribution characteristics of the original samples, generate new synthetic samples with similar distribution, and add Wasserstein distance to increase the authenticity of the generated samples, and add gradient penalty to prevent model collapse. However, ACWGAN-GP also has certain shortcomings. First, the network structure of the model is trained from scratch, and more iterations are required to generate fault signal samples that are relatively similar to the naked eye. Although training a model from scratch for a specific task will eventually fit a good sample, it takes a greater time cost, which greatly reduces the computational efficiency of the algorithm. Secondly, the model adds a classification layer to the last layer of the discriminator, which has low generalization ability and poor classification effect. Moreover, for the classifier, if only the original samples are used for training, it is easy to cause overfitting problems. Finally, in the generator, as the number of network layers increases, the convolutional neural network will experience the phenomenon of gradient disappearance, which will lead to low precision and overfitting problems of the generator.

In summary, the data enhancement method based on ACWGAN-GP can make up for the problem of unbalanced data in bearing fault data, but it also produces some new problems, such as low quality of model generated samples, low training efficiency, poor generalization ability of classifier, and easy gradient vanishing phenomenon of generator. This will reduce the training efficiency of the model and restrict the accuracy of subsequent model fault diagnosis. Therefore, a rolling bearing enhanced diagnosis method that can quickly and effectively improve the quality and training efficiency of model generated samples, enhance the generalization ability of classifier, avoid the gradient vanishing problem of generator, and achieve more accurate and efficient fault sample generation needs to be studied urgently.

Summary of the invention

In order to solve the above problems, the present invention discloses a rolling bearing enhanced diagnosis method based on PDA-WGANGP, a pre-trained data enhanced generative adversarial network method based on Wasserstein distance and gradient penalty, the pre-trained network is trained using health data and fault data under laboratory conditions, the structure and parameters of the pre-trained network are used as the feature extraction layer of the discriminator and the first few layers of the Dense classifier, the unbalanced training set is used to stably generate high-quality samples of the minority class, and finally, the model training is completed, high-quality fault samples are generated, and the generated samples are verified.

A rolling bearing enhanced diagnosis method based on PDA-WGANGP, characterized in that it includes the following steps:

Step 1: The sensor collects healthy signal and fault signal data, pre-processes the data based on the PDA method, and pre-trains the network model;

Step 2, using the data as input and calling the pre-trained network model, generating samples based on the WGANGP method, and evaluating the quality of the generated samples to output a high-quality fault model;

Step 3: Use the high-quality fault model as an input sample to train a fault diagnosis model, input the data into the trained fault diagnosis model, output the fault diagnosis result, and verify the diagnostic capability of the network model.

Preferably, step 1 specifically includes: step 1.1, performing fast Fourier transform on the healthy signal and fault signal collected by the sensor, so that the health and fault features corresponding to the signal are more significant, which is convenient for the subsequent model to extract features, and the data is normalized to between (-1, 1); step 1.2, dividing the data into a pre-training data set, an unbalanced training set, and a test set. Step 1.3, improving the discriminator and the dense classifier: using the pre-training data set in step 1.2 to pre-train the first few layers of the one-dimensional convolutional network of the discriminator and the dense classifier, and then adding a high-dimensional convolutional network structure, so that the network has a more generalized feature extraction layer.

Preferably, the pre-training method of the improved discriminator and the improved Dense classifier is: in the improved discriminator and the Dense classifier, the pre-training data set is used to pre-train the first few layers of the one-dimensional convolutional network, and then the high-dimensional convolutional network structure is added to make the network have a more generalized feature extraction layer. This can accelerate the arrival of the Nash equilibrium and quickly generate a good quality signal. The pre-training network is modified with reference to the VGG network. Since the fault data and healthy status data under laboratory conditions are very rich, the training data of the pre-training network uses the data set generated by the faulty bearings and the healthy bearings under laboratory conditions.

Preferably, step 2 specifically includes: step 2.1, establishing a WGANGP model and initializing parameters; step 2.2, training the WGANGP model: introducing a residual network to construct an improved generator, inputting random noise and fault category labels into the improved generator to generate labeled synthetic samples; labeling the unbalanced training set with category labels as the original samples input; inputting the synthetic samples and the original samples into the improved discriminator and the Dense classifier to respectively judge the authenticity and classify the samples; step 2.3, if the improved discriminator judges that the data is a raw data, If the improved discriminator determines that the data is an original sample, a "false" discrimination result is given. If the improved discriminator determines that the data is an original sample, a "true" discrimination result is given; the dense classifier finally outputs the fault category of the generated sample and the original sample; step 2.4, when the improved discriminator cannot accurately determine whether a sample is an original sample or a generated sample, the entire model reaches a Nash equilibrium and outputs a generated sample based on the Nash equilibrium; step 2.5, evaluate the quality of the generated samples output by step 2.4, and use the high-quality generated samples as the input data for bearing fault diagnosis under the balanced training set condition in step 3.

批量大小m，Adam超参数α、β₁、β₂，超参数ρ，Preferably, the pseudo code for establishing the WGANGP model in step 2.1 is as follows, where λ=10, n _D =5,

Batch size m, Adam hyperparameters α, β ₁ , β ₂ , hyperparameter ρ,

Preferably, the improved generator in step 2.2 is a new generator constructed after introducing the residual network. Specifically, the residual block A is added to the network structure of the generator, the output of the residual block is G(A), and F(A) is the mapping of the residual block: G(A) = F(A) + A. This solves the problem that the gradient of the convolutional neural network disappears as the number of convolutional layers increases.

As a preference, the generator network uses the Instance Normalization (IN) algorithm instead of the Batch Normalization (BN) algorithm for model optimization, so that the network has better fitting ability. Using IN in the generator can not only accelerate model convergence, but also maintain the independence of each signal instance.

其中，m为需产生样本的数量，k为故障类别数。Preferably, in step 2.2, random noise and fault category labels are input into the improved generator, specifically: random noise vector set Z = (z ₁ , z ₂ , z ₃ ... z _m ), m = (1, 2 ...) and label set Y = (y ¹ , y ² , y ³ ... y ^k ), k = (1, 2 ...) are simultaneously input into the improved generator, so that the generator generates synthetic samples with labels.

Among them, m is the number of samples to be generated, and k is the number of fault categories.

As a preferred method, the Dense classifier adopts an alternating pooling method and an activation function to reduce the influence of noise during training. The original auxiliary classification generative adversarial network is a simple method to provide different category controls during sample generation by adding a classification layer to the discriminator. The classifier is extracted from the discriminator separately, and a Dense-based classifier is proposed to focus on the classification task, so that it has better classification ability, thereby constraining the generation of the generator-controlled category and effectively improving the classification accuracy. However, the classifier was originally designed for real data sets, not for data generated by adversarial networks. Since the generator takes a random noise vector as input and generates synthetic data, the proposed classifier should consider the influence of noise input. Therefore, the alternating pooling method and the activation function of the classifier are used to reduce the influence of noise during training.

As a preferred method, the principles of the improved discriminator and the Dense classifier are as follows:

L_C是改进的判别器、Dense分类器对真实数据的分类、Dense分类器对生成数据的分类、Dense分类器的损失函数。P_r(s)是真实数据的分布，P_z(z)是噪声向量z的先验分布。θ_D、θ_G、θ_C是改进的判别器、生成器、Dense分类器的参数。c_g是从条件噪声中采样的类别，c_r是真实数据的类别。ρ是一个超参数，用来控制改进的生成器对真实数据分类的准确性和对生成数据分类的准确性的重要性。

为真实数据和生成数据的插值，λ是梯度惩罚项的权重以满足1-lipschitz条件。 _LD ,

L _C is the classification of real data by the improved discriminator, the classification of generated data by the Dense classifier, and the loss function of the Dense classifier. _{P r} (s) is the distribution of real data, and P _z (z) is the prior distribution of the noise vector z. _{θ D} , θ _G , θ _C are the parameters of the improved discriminator, generator, and Dense classifier. c _g is the category sampled from the conditional noise, and _cr is the category of the real data. ρ is a hyperparameter used to control the accuracy of the improved generator in classifying real data and the importance of the accuracy of the generated data classification.

is the interpolation of the real data and the generated data, and λ is the weight of the gradient penalty term to satisfy the 1-lipschitz condition.

Preferably, in step 2.5, two evaluation indicators, cosine similarity and Pearson correlation coefficient, are used to evaluate the similarity between the generated samples and the real samples.

Preferably, the fault diagnosis in step 3 includes: step 3.1, expanding the unbalanced training set: inputting the generated samples based on Nash equilibrium outputted in step 2.4 into fault diagnosis models such as convolutional neural networks and support vector machines for model training; step 3.2, inputting the test set divided in step 1 into the trained fault diagnosis model for model diagnosis capability testing, and finally obtaining the enhanced diagnosis results of rolling bearings based on PDA-WGANGP.

Beneficial effects:

(1) The present invention adopts a pre-trained network structure in the discriminator and the classifier, and has a more generalized feature extraction layer, which to some extent alleviates the problem of weak generalization of the model, reduces the amount of calculation, and improves the efficiency of sample generation;

(2) The present invention not only trains the classifier with original data, but also adds generated data to reduce overfitting and ensure the training stability of the generator;

(3) The present invention adopts a new generator network structure, designs a new deep residual network to fit the data distribution, and uses Instance Normalization (IN) instead of Batch Normalization (BN), so as to have better fitting ability;

(4) The present invention separates the classifier from the discriminator and proposes a Dense-based classifier, which focuses on the classification task and has better classification ability, thereby constraining and supervising the generation of the generator control category;

(5) The technical method proposed in the present invention can be applied in the field of production operations related to rolling bearing fault diagnosis, realize the monitoring of the health status of rolling bearings in industrial production, extend the service life of rolling bearings, and ensure the continuous airworthiness of equipment.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of a rolling bearing enhanced diagnosis method based on PDA-WGANGP according to an embodiment of the present invention;

FIG2 is a sample generation scheme diagram based on WGANGP according to an embodiment of the present invention;

FIG3 is a comparison diagram of the diagnostic accuracy of an unbalanced training set and four methods according to an embodiment of the present invention.

DETAILED DESCRIPTION

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.

Aiming at the needs of rolling bearing fault and health data imbalance in mechanical systems, this paper proposes a rolling bearing enhanced diagnosis method based on PDA-WGANGP with the goal of solving the problems of low training efficiency and low quality of generated samples existing in ACWGAN-GP.

In conjunction with Figures 1 to 3, a rolling bearing enhanced diagnosis method based on PDA-WGANGP, step 1, the sensor collects health signal and fault signal data, pre-processes the data based on the PDA method and pre-trains the model. In order to specifically illustrate the method of the present invention, the present invention experimentally verifies the 2016 Paderborn University bearing data set. In this data set, bearing damage is widely distributed and will be divided into three main groups: 6 healthy bearings, 12 artificially damaged bearings and 14 actually damaged bearings. The vibration signal is obtained at a sampling rate of 64khz.

The time domain signal of the data is converted into a frequency domain signal through fast Fourier transform to enhance the signal characteristics and normalize the signal. Fast Fourier transform is performed on the vibration signal to make the corresponding health and fault characteristics of the signal more significant, which is convenient for the subsequent model to extract features and normalize the data to between (-1, 1).

The data is divided into pre-training data set, unbalanced training set and test set. Because the fault data and health status data under laboratory conditions are very rich, the data set generated by the faulty bearings and the healthy bearings under laboratory conditions is used as the pre-training data, and the Dense classifier and discriminator are pre-trained based on the pre-training data set. However, in actual industrial production, most rolling bearings are in normal working condition, so the healthy samples are far more than the fault samples. Based on this condition, it is assumed that in the training set, the number of each type of fault samples in the original sample is n, and the number of samples in the healthy state is 5n, which is 5 times different, which means that the generator needs to generate 4n samples for each type of fault. In the test set, the number of samples of the healthy state and each type of fault condition is 2n. Among them, the length of each sample is 1024. The operating condition N09_M07_F10 represents a rotation speed of 900rpm, a loading torque of 0.7NM, and a radial force of 1000N. The specific pre-training data set is shown in Table 1.

Table 1. Pre-trained network datasets

For the unbalanced data set, 8 categories, including K001, KA04, KA15, KA16, KI04, KI14, KI17, and KI18, are selected as training and test sets, including 1 healthy bearing and 7 real faulty bearings. The operating condition used is N09_M07_F10. K001 is a healthy bearing, and the specific information of the faulty bearing is shown in Table 2.

Table 2. Damage description of faulty bearings

In real situations, due to the influence of noise, there are some differences in the signal measurement of different operation times under the same working conditions and the same degree of damage, which brings a considerable obstacle to the accuracy of fault diagnosis. Therefore, this paper selects two different operation times, divided into time one and time two, to more realistically reflect the actual situation. The training set and test set are shown in Table 3 and Table 4.

Table 3. Imbalanced training dataset

Table 4. Test dataset

Pre-training is performed on the discriminator and the dense classifier so that they can obtain more general feature extractors through learning under a large amount of data. In the improved discriminator and dense classifier, the pre-training data set is used to pre-train the first few layers of the one-dimensional convolutional network, and the structure of the high-dimensional convolutional network is added afterwards, so that the network has a more generalized feature extraction layer. This can accelerate the arrival of Nash equilibrium and quickly generate high-quality generation signals. For the 1st to 9th layers of the discriminator, the pre-trained network structure and parameters are used, and the 1st to 9th layers are set as untrainable, and the remaining layers are added layers, and training is started from scratch. For the 1st to 19th layers of the classifier, the pre-trained network structure and parameters are used, and the 1st to 11th layers are set as untrainable, and the 12th to 19th layers are set as trainable. The remaining layers are added layers and training is started from scratch. Among them, the pre-trained network structure, the dense classifier network structure, and the improved discriminator network structure are shown in Tables 5, 6, and 7. When the network pre-training is completed, the pre-trained model is saved for calling during formal training.

Table 5. Pre-trained network structure

Table 6. Dense classifier network structure

Table 7. Improved discriminator network structure

Step 2: Use the data in step 1 as input and call the pre-trained network model, generate samples based on the WGANGP method, and evaluate the quality of the generated samples to output a high-quality fault model.

First, establish the WGANGP model and initialize the parameters:

As shown in Figure 2, it is a sample generation scheme diagram based on WGANGP. Random noise and fault category labels are input into the improved generator to generate labeled synthetic samples; the category labels are attached to the imbalanced training set as input; the synthetic samples and input are input into the improved discriminator and Dense classifier to respectively judge the authenticity and classify the samples.

批量大小m，Adam超参数α、β₁、β₂，超参数ρ：The pseudo code of the training model of WGANGP is as follows, where λ = 10, n _D = 5,

Batch size m, Adam hyperparameters α, β ₁ , β ₂ , hyperparameter ρ:

其中，m为需产生样本的数量，k为故障类别数。其中，生成器G的网络结构如表8所示。To train the WGANGP model, first, random noise z is combined with seven fault category labels Y = {y ¹ , y ² ...y ⁷ } and input into the generator G so that the generator generates synthetic samples with labels.

Among them, m is the number of samples to be generated, and k is the number of fault categories. The network structure of the generator G is shown in Table 8.

In order to solve the problem of gradient vanishing of convolutional neural network as the number of convolutional layers increases, the residual block A is added to the network structure to improve the generator. The output of the residual block is G(A), and F(A) is the mapping of the residual block:

G(A)＝F(A)+A

In addition, Instance Normalization (IN) is used instead of Batch Normalization (BN) to make the network have better fitting ability. Using IN in the generator can not only accelerate the convergence of the model, but also keep each signal instance independent. Therefore, IN is used in the generator.

Then, the generator will continue to generate new samples. The continuously generated new samples and the original 100 fault samples of each category are randomly input into the Dense classifier C and the improved discriminator D in the form of batches for sample category classification and authenticity judgment respectively.

ACWGAN-GP is a simple method to provide different category controls in the sample generation process by adding a classification layer to the discriminator. The classifier is extracted from the discriminator and a Dense-based classifier is proposed to focus on the classification task, which has better classification capabilities and thus plays a role in constraining and supervising the generation of the generator control category.

However, the classifier was originally designed for real datasets, not for data generated by adversarial networks. Since the generator takes a random noise vector as input and generates synthetic data, the proposed classifier should consider the impact of the noise input. Therefore, an alternating pooling method and the activation function of the classifier are used to reduce the impact of noise during training. The principle of the improved discriminator and dense classifier is:

L_C是改进的判别器、Dense分类器对真实数据的分类、Dense分类器对生成数据的分类、Dense分类器的损失函数。P_r(s)是真实数据的分布，P_z(z)是噪声向量z的先验分布。θ_D、θ_G、θ_C是改进的判别器、生成器、Dense分类器的参数。c_g是从条件噪声中采样的类别，c_r是真实数据的类别。ρ是一个超参数，用来控制改进的生成器对真实数据分类的准确性和对生成数据分类的准确性的重要性。

为真实数据和生成数据的插值，λ是梯度惩罚项的权重以满足1-lipschitz条件。 _LD ,

L _C is the classification of real data by the improved discriminator, the classification of generated data by the Dense classifier, and the loss function of the Dense classifier. _{P r} (s) is the distribution of real data, and P _z (z) is the prior distribution of the noise vector z. _{θ D} , θ _G , θ _C are the parameters of the improved discriminator, generator, and Dense classifier. c _g is the category sampled from the conditional noise, and _cr is the category of the real data. ρ is a hyperparameter used to control the accuracy of the improved generator in classifying real data and the importance of the accuracy of the generated data classification.

is the interpolation of the real data and the generated data, and λ is the weight of the gradient penalty term to satisfy the 1-lipschitz condition.

Finally, when the discriminator cannot determine whether the sample comes from the generator or the original training sample, the entire model reaches a Nash equilibrium. At this time, the sample generated by the generator is almost consistent with the original sample and can be used as a training sample for subsequent fault diagnosis to expand the unbalanced training data set.

As shown in Table 3, the number of samples in the healthy state under the unbalanced training set is 500. Therefore, when the model reaches the Nash equilibrium, the 100 fault samples in each of the seven fault states are expanded to 500 to prepare for the experimental verification of fault diagnosis.

Table 8. Improved generator network structure

In order to verify the quality of generated samples, this paper introduces cosine similarity and Pearson correlation coefficient to evaluate the similarity between generated samples and real samples, and verifies the superiority of the WGANGP method.

Cosine similarity: measures the correlation between two sample vectors. The given similarity ranges from -1 to 1: -1 means that the two vectors point in opposite directions, 1 means that they point in exactly the same direction, 0 usually means that they are independent, and the values in between indicate intermediate similarities or dissimilarities. The cosine similarity of sample vectors x and y is:

Pearson correlation coefficient: measures the correlation between two random sample variables. The larger the absolute value, the greater the correlation. The Pearson correlation coefficient of random sample variables X and Y is:

Where Cov(X,Y) is the covariance of X and Y, _σx is the variance of X, and _σY is the variance of Y.

The three methods of synthetic minority oversampling (SMOTE), adaptive synthetic oversampling (ADASYN), and ACWGAN-GP will be used to complete the comparative analysis. The cosine similarity comparison of the four methods is shown in Table 9, and the Pearson correlation coefficient comparison is shown in Table 10.

Table 9. Cosine similarity comparison of four methods

Table 10. Comparison of Pearson correlation coefficients of four methods

Through the calculation of cosine similarity and Pearson correlation coefficient in Table 9 and Table 10, it can be seen that the fault signal generated by WGANGP has a high degree of authenticity, so it will play a positive role in fault diagnosis after data balancing. The cosine similarity and Pearson correlation coefficient of SMOTE and ADASYN are higher than those of WGANGP and AC-WGANGP. It can be seen that AC-WGANGP does not fit the data distribution of the original fault signal well, so the fault signal generated by its generator is of poor quality and cannot effectively help the fault diagnosis in real situations. Traditional SMOTE and ADASYN generate new signals by linear interpolation, so even under small sample conditions, the generated signals have a high degree of similarity, which has a certain effect on data enhancement.

Step 3, expand the unbalanced training set through the high-quality fault model output by step 2, expand the number of n minority class fault samples to the same number of healthy state samples 5n, that is, generate 4n samples for each type of fault, input the expanded training set into fault diagnosis models such as convolutional neural networks and support vector machines for model training, input the test set into the trained fault diagnosis model, output the fault diagnosis results, verify the diagnostic ability of the network model, and obtain the final enhanced diagnosis results of rolling bearings based on PDA-WGANGP.

The rolling bearing enhanced diagnosis method based on PDA-WGANGP proposed in this invention is adopted. In order to compare and verify the effectiveness and superiority of the proposed method, the three methods of SMOTE, ADASYN and ACWGAN-GP will be used to complete the comparative analysis. This paper will use these four methods to supplement the minority fault signals and obtain the data enhanced training data set. Finally, the unbalanced training data set and the data set enhanced by the four methods are respectively subjected to fault diagnosis of three network structures, namely one-dimensional convolutional neural network (1D-CNN), support vector machine (SVM) and dense convolutional network (DENSE), and the diagnostic accuracy of each model is compared to verify the accuracy of the samples generated by the model. The comparison chart of the diagnostic accuracy of the unbalanced training set and the four methods is shown in Figure 3.

As shown in Figure 3, the PDA-WGANGP method based on the three fault diagnosis models shows stable characteristics. The highest diagnostic accuracy is achieved on DENSE, which can reach 93.35% accuracy, and the lowest is 92.25% on 1D-CNN, with a difference of only 1.1%. The AC-WGANGP data enhancement method has the largest fluctuation, with the maximum fault diagnosis accuracy of 91.63% and the minimum of 87.13%, with a difference of 4.5%, indicating that the quality of samples generated by AC-WGANGP is not high. For unbalanced training data sets, the diagnostic effect is the worst because the model lacks the input of training data. Although SMOTE and ADASYN have a strong correlation with real samples in terms of generated samples, they rely heavily on data features and do not consider the real distribution characteristics of minority class samples. Their fault diagnosis effect is not as good as PDA-WGANGP.

Thus, the application process of the present invention was implemented through data analysis, and the effectiveness and superiority of the method of the present invention were verified.

Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims (7)

1. A rolling bearing enhanced diagnosis method based on PDA-WGANGP, which is characterized in that: the method comprises the following steps: step 1, a sensor collects health signals and fault signal data, preprocesses the health signals and the fault signal data based on a PDA method and pretrains a network model; the method specifically comprises the following steps:

step 1.1, converting a time domain signal of data into a frequency domain signal through fast Fourier transform to enhance signal characteristics, and carrying out normalization processing on the signal;

step 1.2, dividing the data into a pre-training data set, an unbalanced training set and a test set;

step 1.3, improving a discriminator and a Dense classifier: pre-training the one-dimensional convolution network of the discriminator and the Dense classifier by using the pre-training data set in the step 1.2, and adding the structure of the high-dimensional convolution network;

step 2, taking the fault signal data as input quantity, calling a pre-trained network model, generating a sample based on a WGANGP method, and evaluating the quality of the generated sample to output a high-quality fault model, wherein the specific steps are as follows: step 2.1, establishing a WGANGP model and initializing parameters;

step 2.2, training a WGANGP model: introducing a residual network to improve a generator, inputting random noise and fault class labels into the improved generator to produce labeled composite samples; attaching a class label to the unbalanced training set as an input original sample; inputting the synthesized sample and the original sample into an improved discriminator and a Dense classifier to respectively judge the authenticity and classify the samples;

step 2.3, if the improved discriminator judges that the data is a synthesized sample, a false discrimination result is given, and if the improved discriminator judges that the data is an original sample, a true discrimination result is given; the Dense classifier outputs the fault category of the synthesized sample and the original sample finally;

step 2.4, when the improved discriminator cannot accurately judge whether a certain sample is an original sample or a synthesized sample, the whole model achieves Nash equilibrium, and a generated sample based on Nash equilibrium is output;

step 2.5, evaluating the quality of the generated sample output in the step 2.4, and taking the high-quality generated sample as input data of bearing fault diagnosis under the condition of the balance training set in the step 3;

and 3, after the unbalance training set is expanded by the high-quality generated sample, training a fault diagnosis model by taking the high-quality generated sample as an input sample under the condition of the balance training set, inputting the test set divided in the step 1.2 into the trained fault diagnosis model, outputting a fault diagnosis result, verifying the diagnosis capability of the network model, and finally obtaining the rolling bearing enhanced diagnosis result based on the PDA-WGANGP.

2. The PDA-WGANGP based rolling bearing enhanced diagnostic method according to claim 1, wherein: the method for improving the generator in the step 2.2 is to introduce a residual network, and add a residual block a to the network structure of the generator to solve the problem that the gradient of the convolutional neural network disappears due to the increase of the convolutional layer, wherein the output of the residual block is G (a), and F (a) is the mapping of the residual block: g (a) =f (a) +a.

3. The PDA-WGANGP based rolling bearing enhanced diagnostic method according to claim 2, wherein: the network of generators uses an IN algorithm instead of a BN algorithm for model optimization to speed up model convergence and maintain independence between each signal instance.

4. A PDA-WGANGP-based rolling bearing enhanced diagnostic method according to claim 3, wherein: the step 2.2 is to input random noise and fault class labels into an improved generator, specifically: set of random noise vectors z= (Z) ₁ ，z ₂ ，z ₃ ...z _m ) M= (1, 2 …) and tag set y= (Y) ¹ ，y ² ，y ³ …y ^k ) K= (1, 2 …) are simultaneously input into the modified generator such that the generator produces labeled composite samples

Where m is the number of samples to be generated and k is the number of fault categories. />

5. The PDA-WGANGP based rolling bearing enhanced diagnostic method according to claim 1, wherein: the Dense classifier adopts an alternate pooling method and an activation function to reduce the influence of noise in the training process.

6. The PDA-WGANGP-based rolling bearing enhanced diagnostic method according to claim 5, wherein: the principle of the improved discriminator and the Dense classifier is as follows:

wherein L is _D ，

L _C The method is an improved discriminator, classification of real data by a Dense classifier, classification of generated data by the Dense classifier and a loss function of the Dense classifier; p (P) _r (s) is the distribution of real data, P _z (z) is an a priori distribution of noise vector z; θ _D 、θ _G 、θ _C Is a parameter of an improved discriminator, generator and Dense classifier; c _g Is the category sampled from the conditional noise, c _r Is a category of real data; ρ is a superparameter for controlling the accuracy of the improved generator for true data classification and the importance of the accuracy for generating data classification;

For interpolation of the real data and the generated data λ is the weight of the gradient penalty term to satisfy the 1-lipschitz condition.

7. The PDA-WGANGP-based rolling bearing enhanced diagnostic method according to claim 6, wherein: and in the step 2.5, two evaluation indexes of cosine similarity and pearson correlation coefficient are adopted to evaluate the similarity of the generated sample and the real sample.

Images