CN114429152A - Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption - Google Patents

Abstract

The invention discloses a fault diagnosis method for a rolling bearing based on dynamic index confrontation and self-adaptation, comprising the following steps: collecting vibration data of the bearing during operation under different working conditions; and domain discriminator and optimize the feature extractor, calculate the loss; use the loss to build the objective function of the bearing fault diagnosis model, find the best parameters, until the bearing fault diagnosis model is completed, use the dynamic index adjustment factor to reduce the source during the training process. The marginal distribution and conditional distribution of domain samples and target domain samples are different; the target domain samples are input into the bearing fault diagnosis model, and the bearing fault diagnosis results are output. The present invention can accurately and quantitatively measure the proportions of edge distribution and conditional distribution in the overall data distribution, so that the model can migrate data sets under different working conditions in a more targeted manner and achieve accurate fault diagnosis.

Description

A Fault Diagnosis Method for Rolling Bearings Based on Dynamic Index Adversarial Adaptive

technical field

The invention relates to the technical field of mechanical fault diagnosis, in particular to a rolling bearing fault diagnosis method based on dynamic index confrontation and self-adaptation.

Background technique

With the development of industry, more and more rotating mechanical machines are used in production and life. Rolling bearing is one of the most important key components in rotating machinery, and its state is directly related to the normal operation of such rotating machinery. Fault diagnosis is a comprehensive technology and an important measure to ensure the safe and reliable operation of mechanical equipment. Therefore, the diagnosis of rolling bearing faults, especially the analysis of early failures, and the realization of fast and accurate bearing fault monitoring are of great significance for the normal operation of mechanical equipment and safe production. Traditional fault diagnosis methods, such as time-domain statistical analysis, wavelet transform, sparse representation and Fourier spectrum analysis, can achieve accurate fault diagnosis. However, traditional fault diagnosis requires extensive experience and deep prior knowledge of engineers. For example, in a bearing fault diagnosis method based on wavelet denoising, a suitable wavelet base must be selected manually. To overcome this limitation, coupled with the emergence and continuous advancement of artificial intelligence technology, many researchers have turned their attention to intelligent fault diagnosis technology.

Intelligent fault diagnosis is the application of machine learning theory such as artificial neural network, support vector machine and deep neural network in machine fault diagnosis. Among them, deep neural networks are widely used due to their excellent performance in feature extraction. However, the accuracy of fault diagnosis based on deep neural networks largely depends on the number of samples involved in training, but in actual practice, the fault signals collected by rolling bearings under different speeds and loads are also different. While it is possible to obtain a complete marked fault signal under some operating conditions in the laboratory, it is impractical to obtain a large number of marked fault samples covering all operating conditions when the bearing is operating under variable operating conditions. Therefore, it has become a new challenge to use the existing labeled fault data to diagnose data different from its operating state. In this context, transfer learning becomes a new solution.

Transfer learning can exploit similarities between data, tasks or models to apply models and knowledge learned in an old domain to a new domain. Domain adaptation is an important research direction of transfer learning, which is aimed at scenarios with the same tasks in different fields, and is a direct promotion of transfer learning. The data processed by the unsupervised domain adaptation problem does not contain the target domain label, so how to align the data distribution of the source domain and the target domain to realize the fault diagnosis of the target domain unlabeled data with the help of the source domain labeled data becomes a new difficulty.

Most of the existing fault diagnosis methods focus on aligning the marginal distribution of features, while ignoring the alignment of within-class distributions, i.e., conditional distributions. These methods assume that conditional distributions are automatically aligned during the alignment of marginal distributions. This assumption is invalid and has a negative effect on the model. Few methods that consider conditional distribution, such as joint distribution alignment method, assume that marginal distribution alignment and conditional distribution alignment have the same weight in the overall data distribution alignment, which obviously cannot be applied to all data distributions.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a fault diagnosis method for rolling bearings based on dynamic index adversarial self-adaptation, which can accurately and quantitatively measure the proportion of edge distribution and conditional distribution in the overall data distribution, so that the model can be more targeted. Data sets under different working conditions are migrated to achieve accurate fault diagnosis.

In order to solve the above technical problems, the present invention provides a method for diagnosing faults of rolling bearings based on dynamic index confrontation and self-adaptation, which includes the following steps:

S1: Collect the vibration data of the bearing during operation under different working conditions, and obtain the source domain sample and the target domain sample;

S2: Input the source domain samples into the feature extractor to obtain the source domain features; input the source domain samples and the target domain samples together into the feature extractor to obtain the mixed domain sample features;

S3: Take the source domain features and the mixed domain sample features as input, train the classifier and the domain discriminator against adversarial and optimize the feature extractor, and calculate the loss of the classifier and the domain discriminator; where the domain discriminator includes the global domain discriminator and the local domain discriminator;

计算源域样本与目标域样本之间分布距离的全局度量和局部度量得到指数动态调节因子，利用指数动态调节因子重新定义域鉴别器损失，其中，全局度量和局部度量分别对应边缘分布和条件分布差异；S4: For global domain discriminator loss and local domain discriminator loss by similarity measure

Calculate the global measure and local measure of the distribution distance between the source domain samples and the target domain samples to obtain the exponential dynamic adjustment factor, and use the exponential dynamic adjustment factor to redefine the loss of the domain discriminator, where the global measure and the local measure correspond to the marginal distribution and the conditional distribution, respectively difference;

S5: Use the classifier loss and the redefined domain discriminator loss to construct the objective function of the bearing fault diagnosis model, and find the optimal parameters of the objective function through the training of the labeled source domain samples and the unlabeled target domain samples, until The bearing fault diagnosis model is completed, wherein in the training process, a dynamic index adjustment factor is used to reduce the marginal distribution and conditional distribution differences between the source domain samples and the target domain samples;

S6: Input the target domain sample into the completed bearing fault diagnosis model, and output the bearing fault diagnosis result.

As a further improvement of the present invention, in the step S1, the bearing health status under each working condition is different, and the bearing vibration data of different health status under each working condition is used as a transferable data field, and the data field is attached with Domain labels, the source domain samples and target domain samples are selected from the data domain, and the source domain samples are attached with fault type labels.

As a further improvement of the present invention, an acceleration sensor is used to collect vibration signals of the bearing during each working condition, a source domain data set and a target domain data set are constructed, and short-time Fourier transform is used to analyze the source domain data set and all The target domain data set is processed, and the two-dimensional processing is performed, and the processed multi-source domain samples and target domain samples are output.

y_i是数据真实故障标签，C(G(x_i))是分类器预测的故障标签，L_c计算两者的交叉熵损失。As a further improvement of the present invention, in the step S3, the source domain features are input into the classifier for supervised training to obtain the predicted source domain label and the classifier loss, wherein the supervised training is to calculate the predicted source domain label and The cross-entropy loss of the true labels in the source domain gets the classifier loss

y _i is the true fault label of the data, C(G( _xi )) is the fault label predicted by the classifier, and L _c calculates the cross-entropy loss of both.

As a further improvement of the present invention, in the step S3:

其中，d_k为数据真实的域标签，D_g(G(x_k))代表预测的域标签，L_Dg计算两者的交叉熵损失；The mixed domain samples are fed into the global domain discriminator for training, resulting in predicted domain labels and global domain loss

Among them, d _k is the real domain label of the data, D _g (G(x _k )) represents the predicted domain label, and L _Dg calculates the cross entropy loss of the two;

计算局部域损失，其中，H代表故障标签种类数量，

是与第h类相关的域鉴别器，

是

对应的交叉熵损失，

代表第k个样本在h类上存在的概率分布，d_k为数据真实的域标签。Input the mixed sample features into the classifier to obtain the probability distribution of the target domain fault type prediction labels; input the mixed sample features into multiple local domain discriminators to obtain multiple domain identification prediction labels and calculate the cross entropy loss of each domain with the real domain label, Multiply and sum the cross-entropy loss and the predicted label probability distribution of the target domain fault type to obtain the final local domain loss calculation formula:

Calculate the local domain loss, where H represents the number of fault label types,

is the domain discriminator associated with class h,

Yes

The corresponding cross-entropy loss,

represents the probability distribution of the kth sample on the h class, and _dk is the real domain label of the data.

As a further improvement of the present invention, the step S4 specifically includes:

计算，即利用全局差异度量公式

和局部差异度量

计算两域之间分布距离的全局度量和局部度量；For global domain loss and local domain loss, after

Calculation, that is, using the global difference measure formula

and local difference measure

Calculate the global measure and local measure of the distribution distance between two domains;

利用指数动态调节因子调节两个域边缘分布和条件分布的差异；Converted to an exponential dynamic adjustment factor ω, expressed as

Use an exponential dynamic adjustment factor to adjust the difference between the marginal and conditional distributions of the two domains;

Considering the exponential dynamic adjustment factor, the final domain discriminator loss is defined as:

As a further improvement of the present invention, the step S5 specifically includes:

即最终的总损失计算如式L＝L_y-λL_D；According to the classifier loss L _y and the domain discriminator loss L _D , the objective function of the bearing fault diagnosis model is established, that is, the total loss is calculated, and the proportional change of the two follows the formula

That is, the final total loss is calculated as formula L=L _{y -λLD} ;

Calculate the total loss once in each epoch, and use the Adam algorithm to optimize the established feature extractor, classifier, global domain discriminator and local domain discriminator;

得到每一步的步长，将整个模型依据步长循环所述epoch数次，得到训练好的模型。Use a predefined number of epochs to determine the number of model training times, according to a predefined step size and step size decay formula

The step size of each step is obtained, and the entire model is cycled through the epoch several times according to the step size to obtain a trained model.

As a further improvement of the present invention, the step S6 specifically includes:

Use short-time Fourier transform to process the target domain data set to obtain the target domain sample picture;

Input the target domain sample picture into the trained fault diagnosis model to obtain the final fault diagnosis result.

A rolling bearing fault diagnosis system based on dynamic index adversarial self-adaptation, comprising:

The acquisition module is used to collect the vibration data of the bearing during operation under different working conditions, and obtain the source domain samples and the target domain samples;

The feature extraction module is used to input the source domain samples into the feature extractor to obtain the source domain features; input the source domain samples and the target domain samples together into the feature extractor to obtain the mixed domain sample features;

The classification calculation module is used to take the source domain features and the mixed domain sample features as input, train the classifier and the domain discriminator against adversarial and optimize the feature extractor, and calculate the loss of the classifier and the domain discriminator; wherein, the domain discriminator includes global domain discriminator and local domain discriminator;

计算源域样本与目标域样本之间分布距离的全局度量和局部度量得到指数动态调节因子，利用指数动态调节因子重新定义域鉴别器损失，其中，全局度量和局部度量分别对应边缘分布和条件分布差异；Exponential dynamics module for passing similarity measure for global domain discriminator loss and local domain discriminator loss

Calculate the global measure and local measure of the distribution distance between the source domain samples and the target domain samples to obtain the exponential dynamic adjustment factor, and use the exponential dynamic adjustment factor to redefine the loss of the domain discriminator, where the global measure and the local measure correspond to the marginal distribution and the conditional distribution, respectively difference;

The training module is used to construct the objective function of the bearing fault diagnosis model using the classifier loss and the redefined domain discriminator loss, and find the optimal value of the objective function through the training of the labeled source domain samples and the unlabeled target domain samples parameters until the bearing fault diagnosis model is completed, wherein in the training process, a dynamic index adjustment factor is used to reduce the marginal distribution and conditional distribution differences between the source domain samples and the target domain samples;

The testing module is used for inputting the target domain samples into the completed bearing fault diagnosis model, and outputting bearing fault diagnosis results.

As a further improvement of the present invention, the classification calculation module includes:

The classification module is used to obtain the predicted label by using the feature extraction module to obtain the predicted label to classify the fault type, and calculate the cross-entropy loss with the real label;

The global domain identification module is used to perform domain classification on the samples by using the mixed sample features to obtain the predicted label of each sample feature, and calculate the cross entropy loss of the global domain discriminator using the predicted label and the real label of each sample feature ;

The local domain identification module is used to perform domain classification on the samples by using the mixed sample features to obtain the predicted label of each sample feature, and use the classifier to classify all the sample features to obtain the pseudo-label probability distribution of the target domain samples, The cross-entropy loss of the local domain discriminator is computed using the predicted labels of each sample feature together with the true label and pseudo-label probability distributions.

Beneficial effects of the present invention: the present invention can independently screen features that are more suitable for migration through the confrontation between the feature extractor, the classifier and the domain discriminator, and obtain better fault diagnosis results; the method can not explicitly specify the source The distance measurement between the domain and the target domain, using the loss to indirectly measure the distribution difference between the source domain and the target domain, can better measure the distribution difference between the source domain and the target domain, so as to obtain a better transfer effect. ; This method uses exponential dynamic factors to stably and quantitatively adjust the proportion of marginal distribution and conditional distribution differences in the overall data distribution differences. The exponential function can effectively reduce the data collapse caused by computational defects (division by zero) in the quantitative calculation process, making the model More stable; dynamic quantitative adjustment can make the model better align the data distribution between the source domain and the target domain, so that the model still has a good diagnostic effect in the face of variable working conditions, and can be widely used in machinery, Fault diagnosis tasks of complex systems such as electrical, chemical, and aviation under variable working conditions.

Description of drawings

Fig. 1 is the method flow chart of bearing fault diagnosis of the present invention;

2 is a flowchart of a specific method according to an embodiment of the present invention;

FIG. 3 is a test diagram of a test bench for generating rolling bearing data according to an embodiment of the present invention;

4 is a diagram of a fault diagnosis confusion matrix under a certain migration task according to an embodiment of the present invention;

FIG. 5 is a diagram showing the visualization of fault diagnosis features under a certain migration task according to an embodiment of the present invention;

FIG. 6 is a schematic block diagram of the system structure of the present invention.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

Referring to FIG. 1, the present invention provides a method for diagnosing faults of rolling bearings based on dynamic index adversarial self-adaptation, including the following steps:

S1: Collect the vibration data of the bearing during operation under different working conditions, and obtain the source domain sample and the target domain sample;

S2: Input the source domain samples into the feature extractor to obtain the source domain features; input the source domain samples and the target domain samples together into the feature extractor to obtain the mixed domain sample features;

S3: Take the source domain features and the mixed domain sample features as input, train the classifier and the domain discriminator against adversarial and optimize the feature extractor, and calculate the loss of the classifier and the domain discriminator; where the domain discriminator includes the global domain discriminator and the local domain discriminator;

计算源域样本与目标域样本之间分布距离的全局度量和局部度量得到指数动态调节因子，利用指数动态调节因子重新定义域鉴别器损失，其中，全局度量和局部度量分别对应边缘分布和条件分布差异；S4: For global domain discriminator loss and local domain discriminator loss by similarity measure

Calculate the global measure and local measure of the distribution distance between the source domain samples and the target domain samples to obtain the exponential dynamic adjustment factor, and use the exponential dynamic adjustment factor to redefine the loss of the domain discriminator, where the global measure and the local measure correspond to the marginal distribution and the conditional distribution, respectively difference;

S5: Use the classifier loss and the redefined domain discriminator loss to construct the objective function of the bearing fault diagnosis model, and find the optimal parameters of the objective function through the training of the labeled source domain samples and the unlabeled target domain samples, until The bearing fault diagnosis model is completed, wherein in the training process, a dynamic index adjustment factor is used to reduce the marginal distribution and conditional distribution differences between the source domain samples and the target domain samples;

S6: Input the target domain sample into the completed bearing fault diagnosis model, and output the bearing fault diagnosis result.

The method of the invention uses the acceleration sensor to collect the vibration signal of the bearing under each working condition, constructs the source domain data set and the target domain data set, and uses the short-time Fourier transform and two-dimensional processing to output the source domain sample and the target domain sample , which can help the model to better extract the required features; this method can autonomously screen features that are more suitable for migration through the confrontation between feature extractors, classifiers and domain discriminators, and get better fault diagnosis results . Adversarial adaptation can not explicitly specify the distance measure between the source domain and the target domain, and use the loss to indirectly measure the distribution difference between the source domain and the target domain, which can better analyze the distribution difference between the source domain and the target domain. Measure to get better migration effect; use exponential dynamic factor to stably and quantitatively adjust the proportion of marginal distribution and conditional distribution difference in the overall data distribution difference, exponential function can effectively reduce the quantitative calculation process due to calculation defects (divide by zero) The resulting data collapse makes the model more stable; dynamic quantitative adjustment can make the model better align the data distribution between the source domain and the target domain, so that the model still has a good diagnostic effect in the face of variable working conditions.

Example

In this embodiment, based on the above method and in combination with specific embodiments, a more detailed description is given. The collected vibration signals are used to construct a transferable data set of bearings in seven health states under six different conditions, and the bearing fault model is trained. Please refer to Figure 2, the specific operation steps are as follows:

Step S101: use an acceleration sensor to collect vibration signals of the bearing during operation under six working conditions, and construct a source domain data set and a target domain data set;

Using the signals collected by the test rig shown in Fig. 3, a transferable dataset of bearings with seven health states under six different operating conditions was constructed, and each dataset had different conditional and marginal distributions. The test bearings were set up for three single faults (inner ring fault, roller fault and outer ring fault) and four composite faults (inner ring + outer ring fault, inner ring + roller fault, roller + outer ring fault, rolling sub + inner ring + outer ring fault). Thus, seven health states were obtained, as listed in Table 1. Each health state has 200 training samples and 200 test samples. Each sample consists of 1024 sampling points.

Seven kinds of fault bearing information for each domain in Table 1:

According to six different working conditions, the specific settings are shown in Table 2, and 12 different migration tasks are established. Table 3 provides the migration tasks and their abbreviations.

Table 2 Migration task condition settings:

Table 3 All migration tasks and abbreviations:

Step S102: using short-time Fourier transform and two-dimensional processing to output source domain samples and target domain samples;

A short-time Fourier transform (STFT) is performed on the samples. In STFT, the samples are converted from time-domain signals to time-frequency domain signals. The converted signals contain both time-domain information and frequency-domain information. The rich information can help the model to better diagnose faults.

The generated samples are reshaped from 3×1024 to 3×32×32. For the source domain, the fault type label is attached according to the fault type it belongs to. Add domain labels to all datasets containing fault category labels;

Step S103: Use a feature extractor and a classifier to input the source domain samples into it, and use a supervised method to train the feature extractor; input the target domain samples and the source domain samples into the feature together After the extractor, the mixed sample features are obtained;

The feature extractor is constructed according to the parameters in Table 4, and its essence is a deep residual network mainly composed of multiple convolution blocks, pooling layers and a neural network connected by a fully connected layer.

Table 4 Structured parameters of the feature learner:

The source domain samples are used as the input of the feature extractor, and the source domain feature representation is obtained through the feature extractor.

其中y_i是数据真实故障标签，C(G(x_i))是分类器预测的故障标签，L_c计算两者的交叉熵损失。Build a classifier, which is essentially a neural network consisting of a fully connected layer and a softmax layer. The supervised method trains the feature extractor, which means that the classifier loss is obtained by calculating the cross-entropy loss of the predicted source domain label and the source domain real label.

where y _i is the true fault label of the data, C(G( _xi )) is the fault label predicted by the classifier, and L _c calculates the cross-entropy loss of both.

After the target domain samples and the source domain samples are jointly input into the feature extractor, mixed sample features are obtained.

Step S104: Utilize the global domain classifier and the local domain classifier, use the mixed sample feature as the input of the global domain classifier and the local domain classifier, and utilize a supervised method to confront the feature extractor and the domain discriminator. train, and output the predicted domain label and domain loss;

计算全局域损失。Using the global domain classifier, the mixed sample feature and the domain label, according to the global domain loss calculation formula

Calculate the global domain loss.

计算两者的交叉熵损失。where d _k is the real domain label of the data, D _g (G(x _k )) represents the predicted domain label,

Calculate the cross-entropy loss for both.

计算局部域损失。Use the mixed sample features to input the classifier to obtain the target domain fault type predicted label probability distribution; input the mixed sample features to the multiple local domain discriminators to obtain multiple domain identification predicted labels and real domain labels Calculate the cross-entropy loss of each domain, and multiply and sum the cross-entropy loss and the predicted label probability distribution of the target domain fault type to obtain the final local domain loss calculation formula

Compute the local domain loss.

是与第h类相关的域鉴别器，

是

对应的交叉熵损失，

代表第k个样本在h类上存在的概率分布，d_k为数据真实的域标签。Where H represents the number of types of fault labels,

is the domain discriminator associated with class h,

Yes

The corresponding cross-entropy loss,

represents the probability distribution of the kth sample on the h class, and _dk is the real domain label of the data.

Step S105: utilize the exponential dynamic factor to adjust the global domain loss and the local domain loss, and then adjust the influence of the marginal distribution and the conditional distribution on the overall training of the model, so as to realize the adjustment of the domain loss;

计算，具体利用全局差异度量公式

和局部差异度量

计算对两域之间分布距离的全局度量和局部度量。For the global domain loss and local domain loss, after

Calculation, specifically using the global difference metric formula

and local difference measure

Computes a global measure and a local measure of the distribution distance between two domains.

由于边缘分布差异可间接用全局距离度量来衡量，条件分布差异可间接用局部距离度量来衡量，因此，在调节两个域边缘分布和条件分布的差异对模型参数影响的时候，可以用指数动态调节因子调节输入数据情况下两个分布的重要性。Using the exponential dynamic adjustment factor ω, it can be expressed as

Since the marginal distribution difference can be measured indirectly by the global distance metric, and the conditional distribution difference can be measured indirectly by the local distance metric, therefore, when adjusting the influence of the difference between the marginal distribution and the conditional distribution of the two domains on the model parameters, the exponential dynamic The adjustment factor adjusts the importance of the two distributions given the input data.

Considering the exponential dynamic adjustment factor, the final domain discriminator loss can be defined as

Step S106: constructing the objective function of the bearing fault diagnosis model by using classifier loss and domain loss, and reducing the marginal distribution and conditional distribution difference between the source domain sample feature and the target domain sample feature by using an adversarial adaptive training strategy, The Adam algorithm is used to optimize the model, and iterative is performed according to the predetermined epoch number and step size until the bearing fault diagnosis model is trained;

即最终的总损失计算如式L＝L_y-λL_D。在每一个epoch计算一次总损失，并将总损失利用Adam算法对建立的特征提取器、分类器、全局域鉴别器和局部域鉴别器进行优化。在优化过程中，由于域鉴别器与特征提取器、分类器优化目标相反，因此利用梯度翻转层使得模型可以在一次训练中同时完成相反的训练目的。The total loss is calculated according to the classification loss _Ly and the domain discriminator loss _LD , and the proportional change of the two follows the formula

That is, the final total loss is calculated as formula L=L _{y -λLD} . The total loss is calculated once in each epoch, and the Adam algorithm is used to optimize the total loss for the established feature extractor, classifier, global domain discriminator and local domain discriminator. In the optimization process, since the domain discriminator and the feature extractor and classifier have the opposite optimization goals, the gradient flip layer is used to enable the model to accomplish the opposite training purpose in one training.

得到每一步的步长。将整个模型依据步长循环所述100次，得到训练好的模型。Use the predefined number of epochs = 100 to determine the number of model training times, according to the predefined step size = 0.001 and the step size decay formula

Get the step size for each step. The entire model is cycled 100 times according to the step size to obtain a trained model.

Step S107: inputting the target domain data set into the bearing fault diagnosis model, performing fault type matching on the fault diagnosis feature, and outputting the bearing fault diagnosis result of the target domain data set;

Using the short-time Fourier transform to process the target domain data set to obtain the target domain sample picture;

Input the target domain sample picture into the trained fault diagnosis model to obtain the final fault diagnosis result.

Table 5 The fault diagnosis accuracy rate of the method of the present invention and each variant method under each migration task:

Migration tasks this method Variant 1 Variant 2 variant 3 D1->D2 100 100 100 100 D2->D1 100 99.78 100 100 D1->D3 100 97.21 100 100 D3->D1 100 91.43 88.64 100 D1->D4 100 100 100 99.93 D4->D1 100 99.78 100 100 D1->D5 100 100 100 85.36 D5->D1 100 66.86 99.93 100 D1->D6 100 100 68 100 D6->D1 100 82.64 100 100 average accuracy 100 93.77 95.66 98.53

Figure 4, Figure 5 and Table 5 respectively show the fault diagnosis confusion matrix of this method under a certain migration task, the visualization of fault diagnosis features, and the fault diagnosis accuracy of each variant method under each migration task. Through experimental data verification, a fault diagnosis method for rolling bearings based on dynamic index adversarial self-adaptation of the present invention is used for fault diagnosis according to the above process. Under the data conditions of 1400 source domain samples and 1400 target domain samples, the In the case of changing, rotating speed and load rotating speed, this method can achieve 100% fault diagnosis accuracy after 100 iterations, which shows that this method can stably realize the migration fault diagnosis of rolling bearings under variable working conditions. , this classification accuracy can meet the practical application requirements.

In summary, the present invention discloses a method for diagnosing rolling bearing faults based on dynamic index adversarial self-adaptation, which uses a feature extractor to extract features, and uses a distance metric based on domain loss to evaluate domain differences. The superiority and robustness of the method are confirmed by experimental validation and comparison with several methods. Some results of the present invention are summarized as follows: 1) Through the confrontation between the feature extractor, the classifier and the domain discriminator, the features that are more suitable for migration can be screened autonomously, and better fault diagnosis results can be obtained. 2) Do not explicitly specify the distance measure between the source domain and the target domain, and use the loss to indirectly measure the distribution difference between the source domain and the target domain, which can better measure the distribution difference between the source domain and the target domain, 3) The proposed method uses the exponential dynamic factor to stably and quantitatively adjust the proportion of marginal distribution and conditional distribution difference in the overall data distribution difference, and the exponential function can effectively reduce the quantitative calculation process due to computational defects ( The data collapse caused by dividing by zero) makes the model more stable; dynamic quantitative adjustment can make the model better align the data distribution between the source domain and the target domain, so that the model still has good performance in the face of variable working conditions. diagnostic effect.

In the method provided in this embodiment, data sets under six different working conditions are collected as source domain data sets and target domain data sets to train and test the bearing fault diagnosis model, and feature extractors, classifiers and domain discriminators are used. The training strategy against each other, using the gradient flip layer to optimize the three at the same time, can effectively improve the training efficiency and reduce the possibility of model collapse; the exponential dynamic adaptive factor is used to stably and dynamically adjust the importance of edge distribution and conditional distribution to model optimization. Adjustment to better meet the requirements of the distribution adaptation of transfer learning, so that the knowledge learned in the source domain can be better utilized by the target domain, and the egg knife has a good fault diagnosis effect. Compared with other variant methods, the present invention has an average fault diagnosis accuracy rate as high as 100%, better feature extraction effect and good stability, and can handle fault diagnosis under variable working conditions.

The present invention also provides a rolling bearing fault diagnosis system based on dynamic index confrontation and self-adaptation, comprising:

The acquisition module 100 is used to collect vibration data of the bearing during operation under different working conditions, and obtain source domain samples and target domain samples;

The feature extraction module 300 is used to input the source domain samples into the feature extractor to obtain the source domain features; the source domain samples and the target domain samples are jointly input to the feature extractor to obtain the mixed domain sample features;

The classification calculation module 400 is used to take the source domain feature and the mixed domain sample feature as input, train the classifier and the domain discriminator against the adversarial and optimize the feature extractor, and calculate the loss of the classifier and the domain discriminator; wherein, the domain discriminator is Including global domain discriminator and local domain discriminator;

计算源域样本与目标域样本之间分布距离的全局度量和局部度量得到指数动态调节因子，利用指数动态调节因子重新定义域鉴别器损失，其中，全局度量和局部度量分别对应边缘分布和条件分布差异；Exponential Dynamics Module 500 for Passing Similarity Metrics for Global Domain Discriminator Loss and Local Domain Discriminator Loss

Calculate the global measure and local measure of the distribution distance between the source domain samples and the target domain samples to obtain the exponential dynamic adjustment factor, and use the exponential dynamic adjustment factor to redefine the loss of the domain discriminator, where the global measure and the local measure correspond to the marginal distribution and the conditional distribution, respectively difference;

The training module 600 is used for constructing the objective function of the bearing fault diagnosis model by using the classifier loss and the redefined domain discriminator loss, and finding the most optimal value of the objective function through the training of the labeled source domain samples and the unlabeled target domain samples. optimize parameters until the bearing fault diagnosis model is completed, wherein in the training process, a dynamic index adjustment factor is used to reduce the marginal distribution and conditional distribution differences between the source domain samples and the target domain samples;

The testing module 700 is configured to input the target domain sample into the completed bearing fault diagnosis model, and output the bearing fault diagnosis result.

Further, the classification calculation module 400 includes:

The classification module 401 is used to obtain the predicted label by using the feature extraction module to obtain the predicted label so as to classify the fault type, and calculate the cross-entropy loss with the real label;

The global domain identification module 402 is used to perform domain classification on the sample by using the mixed sample feature, obtain the predicted label of each sample feature, and calculate the cross entropy of the global domain discriminator using the predicted label and the real label of each sample feature. loss;

The local domain identification module 403 is used to perform domain classification on the sample by using the mixed sample feature to obtain the predicted label of each sample feature, and use the classifier to classify all the sample features to obtain the pseudo-label probability distribution of the target domain sample , using the predicted label of each sample feature together with the true label and pseudo-label probability distribution to calculate the cross-entropy loss of the local domain discriminator.

This system is used to implement the aforementioned bearing fault diagnosis method, so the specific implementation of the bearing fault diagnosis system can be found in the embodiment part of the bearing fault diagnosis method above, for example, the acquisition module 100 and the processing module 200, and the feature extraction module 300 and the classification module 400, the global domain identification module 500 and the local domain identification module 600, the exponential dynamic module 700, and the testing module 800 are respectively used to realize steps S101-S107 in the above-mentioned bearing fault diagnosis method, so the specific implementation can be Reference is made to the descriptions of the corresponding respective partial embodiments, which will not be repeated here.

The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments may be referred to each other. As for the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part can be referred to the description of the method.

Professionals may further realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two, in order to clearly illustrate the possibilities of hardware and software. Interchangeability, the above description has generally described the components and steps of each example in terms of function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each particular application, but such implementations should not be considered beyond the scope of the present invention. The steps of a method or algorithm described in conjunction with the embodiments disclosed herein may be directly implemented in hardware, a software module executed by a processor, or a combination of the two. A software module can be placed in random access memory (RAM), internal memory, read only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other in the technical field. in any other known form of storage medium.

The above-mentioned embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention is subject to the claims.

Claims (10)

1. A rolling bearing fault diagnosis method based on dynamic index antagonism self-adaptation is characterized in that: the method comprises the following steps:

s1: acquiring vibration data of a bearing in operation under different working conditions to obtain a source domain sample and a target domain sample;

s2: inputting the source domain samples into a feature extractor to obtain source domain features; inputting the source domain sample and the target domain sample into a feature extractor together to obtain mixed domain sample features;

s3: taking the source domain characteristics and the mixed domain sample characteristics as input, training a classifier and a domain discriminator in an antagonistic way, optimizing a characteristic extractor, and calculating the loss of the classifier and the domain discriminator; wherein, the domain discriminator comprises a global domain discriminator and a local domain discriminator;

s4: passing similarity measures for global discriminator loss and local discriminator loss

Calculating global measurement and local measurement of distribution distance between the source domain sample and the target domain sample to obtain an index dynamic adjustment factor, and redefining the loss of the domain discriminator by using the index dynamic adjustment factor, wherein the global measurement and the local measurement respectively correspond to edge distribution and condition distribution difference;

s5: constructing an objective function of a bearing fault diagnosis model by using classifier loss and redefined domain discriminator loss, and searching optimal parameters of the objective function through training of a source domain sample with a label and a target domain sample without a label until the bearing fault diagnosis model is completed, wherein the edge distribution and condition distribution difference of the source domain sample and the target domain sample is reduced by using a dynamic index adjustment factor in the training process;

s6: and inputting the target domain sample into the finished bearing fault diagnosis model, and outputting a bearing fault diagnosis result.

2. The rolling bearing fault diagnosis method based on dynamic index antagonism adaptation according to claim 1, characterized in that: in step S1, the bearing health status in each condition is different, the bearing vibration data in the different health status in each condition is used as a transferable data field, the data field is attached with a field label, the source field sample and the target field sample are selected from the data field, and the source field sample is attached with a fault type label.

3. The rolling bearing fault diagnosis method based on dynamic index antagonism adaptation as claimed in claim 2, characterized in that: the method comprises the steps of collecting vibration signals of a bearing in operation under each working condition by using an acceleration sensor, constructing a source domain data set and a target domain data set, processing the source domain data set and the target domain data set by using short-time Fourier transform, performing two-dimensional processing, and outputting processed multi-source domain samples and target domain samples.

4. The rolling bearing fault diagnosis method based on dynamic index antagonism adaptation as claimed in claim 2, characterized in that: inputting the source domain features into a classifier to perform supervised training in the step S3 to obtain predicted source domain labels and classifier losses, wherein the supervised training is to obtain the classifier losses by calculating cross entropy losses of the predicted source domain labels and the source domain true labels

y_iIs a data true failure tag, C (G (x)_i) Is a fault label predicted by the classifier, L_cThe cross entropy loss of the two is calculated.

5. The rolling bearing fault diagnosis method based on dynamic index antagonism adaptation according to claim 4, characterized in that: in the step S3:

inputting the mixed domain sample into a global area discriminator for training to obtain a predicted domain label and a predicted global area loss

Wherein d is_kDomain tags for data trueness, D_g(G(x_k) Is) represents a predicted domain label,

both are calculatedCross entropy loss;

inputting the characteristics of the mixed samples into a classifier to obtain the probability distribution of the target domain fault category prediction labels; inputting the characteristics of the mixed samples into a plurality of local domain discriminators to obtain a plurality of domain discrimination prediction labels, calculating cross entropy loss of each domain with a real domain label, and multiplying and summing the cross entropy loss and the probability distribution of the target domain fault type prediction label to obtain a final local domain loss calculation formula:

calculating local area loss, wherein H represents the number of fault label types,

is a domain discriminator associated with class h,

is that

The corresponding cross-entropy loss is then taken into account,

representing the probability distribution of the presence of the kth sample over the h class, d_kIs a domain label of the data reality.

6. The rolling bearing fault diagnosis method based on dynamic index antagonism adaptation according to claim 5, characterized in that: the step S4 specifically includes:

for global and local losses, passing through

Calculating, i.e. using global difference metric formulae

And local variance measure

Calculating global measurement and local measurement of distribution distance between two domains;

is converted into an exponential dynamic adjustment factor omega expressed as

Adjusting the difference between the edge distribution and the condition distribution of the two domains by using an index dynamic adjustment factor;

considering the exponential dynamics adjustment factor, the final domain discriminator loss is defined as:

7. the rolling bearing fault diagnosis method based on dynamic index antagonism adaptation according to claim 6, characterized in that: the step S5 specifically includes:

according to classifier loss L_ySum-field discriminator loss L_DEstablishing an objective function of a bearing fault diagnosis model, namely calculating the total loss, wherein the proportional change of the two follows a formula

I.e. the final total loss calculation is given by the formula L ═ L_y-λL_D；

Calculating the total loss once in each epoch, and optimizing the established feature extractor, classifier, global domain discriminator and local domain discriminator by using the Adam algorithm for the total loss;

determining the number of times of model training by using a predefined epoch number, and according to a predefined step length and a step length attenuation formula

Obtaining the step length of each step, and circulating the whole model for a plurality of times according to the step length to obtain the trainedThe model of (1).

8. The rolling bearing fault diagnosis method based on dynamic index antagonism adaptation as claimed in claim 2, characterized in that: the step S6 specifically includes:

processing the target domain data set by using short-time Fourier transform to obtain a target domain sample picture;

and inputting the target domain sample picture into the trained fault diagnosis model to obtain a final fault diagnosis result.

9. A rolling bearing fault diagnosis system based on dynamic index antagonism self-adaptation is characterized in that: the method comprises the following steps:

the acquisition module is used for acquiring vibration data of the bearing during operation under different working conditions to obtain a source domain sample and a target domain sample;

the characteristic extraction module is used for inputting the source domain samples into the characteristic extractor to obtain source domain characteristics; inputting the source domain sample and the target domain sample into a feature extractor together to obtain mixed domain sample features;

the classification calculation module is used for taking the source domain characteristics and the mixed domain sample characteristics as input, training a classifier and a domain discriminator in an antagonistic way, optimizing a characteristic extractor and calculating the loss of the classifier and the domain discriminator; wherein, the domain discriminator comprises a global domain discriminator and a local domain discriminator;

an exponential dynamic module for passing similarity measures for global discriminator loss and local discriminator loss

Calculating global measurement and local measurement of distribution distance between the source domain sample and the target domain sample to obtain an index dynamic adjustment factor, and redefining the loss of the domain discriminator by using the index dynamic adjustment factor, wherein the global measurement and the local measurement respectively correspond to edge distribution and condition distribution difference;

the training module is used for constructing an objective function of the bearing fault diagnosis model by using classifier loss and redefined domain discriminator loss, searching the optimal parameters of the objective function through the training of a source domain sample with a label and a target domain sample without a label until the bearing fault diagnosis model is completed, wherein the edge distribution and condition distribution difference of the source domain sample and the target domain sample are reduced by using a dynamic index adjustment factor in the training process;

and the test module is used for inputting the target domain sample into the finished bearing fault diagnosis model and outputting a bearing fault diagnosis result.

10. The dynamic index antagonism-based adaptive rolling bearing fault diagnosis system according to claim 9, wherein: the classification calculation module includes:

the classification module is used for obtaining a prediction label by utilizing the characteristics obtained by the characteristic extraction module so as to classify fault types and calculating cross entropy loss with a real label;

the global area identification module is used for carrying out domain classification on the samples by utilizing the mixed sample characteristics to obtain a prediction label of each sample characteristic, and calculating the cross entropy loss of the global area identifier by utilizing the prediction label and the real label of each sample characteristic;

and the local area identification module is used for carrying out domain classification on the samples by utilizing the mixed sample characteristics to obtain a prediction label of each sample characteristic, classifying all the sample characteristics by utilizing the classifier to obtain the pseudo label probability distribution of the target domain sample, and calculating the cross entropy loss of the local area identifier by utilizing the prediction label of each sample characteristic, the real label and the pseudo label probability distribution.

Images