CN116106012A - A Domain Adaptive Fault Diagnosis Method for Rolling Bearings Based on Attention Mechanism - Google Patents

Abstract

The invention relates to a fault diagnosis method for rolling bearing domain adaptation based on an attention mechanism, which is characterized in that a feature extractor which is formed by one-dimensional separable convolution and embedded with a channel attention mechanism and a length attention mechanism extracts deep fault features from collected rolling bearing vibration monitoring signals; the local attention domain adaptation module and the global attention domain adaptation module are constructed to screen signals and signal fragments with good mobility, so that the generalization capability of the model is improved, and the model can better cope with the problem of fault diagnosis under variable working conditions; compared with an intelligent fault diagnosis algorithm for transfer learning, the algorithm considers the different signals and the different migratability of the signal fragments, and improves the interpretability of the model; experiments are carried out by applying various bearing vibration data, so that the algorithm is verified to have good performance stability, and excellent diagnosis results can be still maintained under various working condition change conditions.

Description

A domain-adaptive fault diagnosis method for rolling bearings based on attention mechanism

Technical Field

The present invention belongs to the field of variable operating condition bearing fault diagnosis, and specifically is a rolling bearing domain adaptive fault diagnosis method based on an attention mechanism.

Background Art

As a key component of rotating machinery, rolling bearings operate in a relatively harsh environment and are extremely susceptible to damage during service, resulting in unexpected industrial accidents that are difficult to predict and control. At the very least, they may affect the operation of local equipment and reduce production efficiency. At worst, they may cause serious production safety accidents and lead to irreversible consequences.

With the rapid rise of technologies such as the Industrial Internet of Things, mechanical fault diagnosis has gradually entered the "big data" era, and intelligent fault diagnosis algorithms based on deep learning and other methods have gradually become a research hotspot in recent years. However, existing fault diagnosis algorithms based on deep learning are generally based on the assumption that the training set and the test set are independent and identically distributed, which is inconsistent with the conditions of a large number of noisy environments and complex variable working conditions in actual work. Therefore, transfer learning algorithms that can cope with this have gradually become a research hotspot for intelligent fault diagnosis algorithms.

However, the existing transfer learning intelligent fault diagnosis algorithms lack interpretability and do not take into account the different transferability that different signals and signal fragments may have. Therefore, it is crucial to propose an intelligent fault diagnosis technology with stronger generalization ability and certain interpretability.

Summary of the invention

In view of the shortcomings of the prior art, the present invention provides a rolling bearing domain adaptive fault diagnosis method based on the attention mechanism, which improves the generalization ability of the model and can cope with various fault diagnosis problems, while also improving the interpretability of the model.

The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

A rolling bearing domain adaptation fault diagnosis method based on attention mechanism includes the following steps:

For the same bearing system, vibration monitoring data of the health status of various types of bearings under different working conditions are collected respectively, and the data are labeled according to the fault type to obtain bearing data sets for different working conditions;

The data in the bearing dataset is divided into a source domain dataset and a target domain dataset, and then divided into a training set and a validation set respectively;

Construct a rolling bearing domain adaptive fault diagnosis model, and train the model using the training set of the source domain dataset and the training set of the target domain dataset;

Optimize the domain-adaptive fault diagnosis model for rolling bearings;

The test set of the target domain data is input into the optimized rolling bearing domain adaptive fault diagnosis model to obtain the fault type of the target domain data.

The source domain data is vibration monitoring data with a quantity greater than a threshold and complete labels, and the target domain data is vibration monitoring data under the working condition to be detected.

The construction of the rolling bearing domain adaptive fault diagnosis model comprises:

A feature extractor is used to extract fault features in the bearing dataset and divide the fault features into multiple segments;

The local attention domain adaptation module is used to calculate the local attention value of each segment fault feature, perform weighted processing on it, and dimensionally merge the weighted local attention value of each segment;

The global attention domain adaptation module is used to calculate the global attention value based on the local attention values after dimension merging, and weight it into a classification loss to complete the fault type classification of the input data.

The feature extractor includes sequentially connected spatial convolution, length attention module, channel convolution, channel attention module and pooling layer.

The local attention domain adaptation module includes: a maximum mean difference module, multiple domain classifiers and a residual connection module, wherein the fault features of multiple fragments output by the feature extractor are processed by the maximum mean difference module, and the fault features of each fragment are input into a domain classifier, the probability that the corresponding fragment belongs to the source domain is calculated, and the local attention value of the fault feature of each fragment is calculated using an entropy function to weight the fault features, and the weighted fault features of all fragments are dimensionally merged and output through the residual connection module.

The global attention domain adaptation module includes: a domain classifier, inputting the weighted fault feature into the domain classifier, calculating the probability that the fault feature belongs to the source domain, and using an entropy function to calculate the global attention value of the fault feature to weight the fault feature.

In the construction of the rolling bearing domain adaptive fault diagnosis model, four types of loss functions are set, specifically:

1) Bearing fault classification loss of source domain dataset:

为交叉熵损失函数，G_y为分类器，h_i为局部注意力加权特征，y_i为样本标签；Among them, n _S is the number of source domain samples, D _S is the source domain sample space,

is the cross entropy loss function, _Gy is the classifier, _hi is the local attention weighted feature, and _yi is the sample label;

2) MMD loss used to reduce the classification difference between the source domain and the target domain:

和

来自源域与目标域中的信号样本；φ(·)是用于将源域及目标域样本映射到同一Hilbert空间的非线性映射，k(·,·)为核函数；Among them, n _t is the number of samples in the target domain,

and

Signal samples from the source domain and the target domain; φ(·) is a nonlinear mapping used to map the source domain and target domain samples to the same Hilbert space, and k(·,·) is a kernel function;

及全局注意力域适应损失

3) Local attention domain adaptation loss for extracting local domain and global domain invariant features from signal segments and signal samples respectively

and global attention domain adaptation loss

是用于域分类的交叉熵损失；Where D _S represents the source domain sample space; D _T represents the target domain sample space, and d _i represents the domain label of the signal sample _xi ;

is the cross entropy loss for domain classification;

4) Global weighted entropy loss for auxiliary classification:

Where c represents the number of fault types; pi _,j represents the probability value of signal sample _xi being classified as label j.

The total loss function is:

分别为特征提取器、标签分类器、全局注意力域适应模块及局部注意力域适应模块的模型参数，η，γ和λ分别为修正权重。where, θ _f , θ _y , θ _d and

are the model parameters of the feature extractor, label classifier, global attention domain adaptation module and local attention domain adaptation module, respectively. η, γ and λ are the correction weights respectively.

The rolling bearing domain adaptive fault diagnosis model is optimized using back propagation and Adam optimization algorithm, that is, the total loss function is minimized.

The model parameters are updated using the following formula:

Among them, ε is the learning rate.

The present invention has the following beneficial effects and advantages:

1. The present invention can autonomously select signals and signal segments with better transferability, enhance the domain adaptation capability of the model, and improve the accuracy of fault diagnosis.

2. Compared with many currently popular transfer learning intelligent fault diagnosis algorithms, the present invention has the best performance and robustness and can handle various variable operating condition fault diagnosis problems.

3. The present invention has a certain degree of explainability in the selection of transferable features.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a model structure diagram of the rolling bearing domain adaptive fault diagnosis method based on the attention mechanism of the present invention.

DETAILED DESCRIPTION

The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

A rolling bearing domain adaptation fault diagnosis method based on attention mechanism includes the following steps:

Step 1: For the same bearing system, collect vibration monitoring signals of various bearing health conditions under different working conditions, and label the data according to the fault type to obtain bearing data sets for different working conditions;

Step 2: Select two of the data sets under different working conditions obtained in step 1 as the source domain and target domain, and divide them into corresponding training sets and test sets;

Step 3: Use the training set data of the source domain and the target domain as model input to train the model, use the grid search method to optimize the hyperparameters, select the training model with the highest diagnostic accuracy as the final diagnostic model, and fix the model;

Step 4: Perform a performance test on the final model obtained in step 3 on the test set data of the target domain.

The one-dimensional separable convolution embedded with channel attention mechanism and length attention mechanism is: the present invention applies one-dimensional separable convolution to replace traditional convolution, and embeds two attention mechanisms therein. The model mainly includes the following parts: filtering the signal samples in each channel, and spatial convolution with a specific convolution kernel size; a length attention mechanism module that obtains the importance of each signal segment in fault diagnosis by aggregating the feature information of each channel; a channel convolution that uses a set of convolution kernels to learn the relationship between channels and is used to improve the feature dimension; a channel attention module that captures the interdependence between channels to judge the importance of features and promotes the network to pay attention to important features.

The specific convolution kernel size of the spatial convolution is equivalent to the convolution kernel size of the replaced traditional convolution.

The transfer learning module composed of the local attention domain adaptation module and the global attention domain adaptation module is:

以计算相应片段属于源域的概率，从而判断该片段的可迁移能力，并使用熵函数计算局部注意力值，实现对特征进行加权；同时为保证源域及目标域间特征差异过大而导致负迁移，在输入局部注意力域适应模块前，本发明使用最大均值差异(MMD)来减小源域及目标域的分布差异；局部注意力加权特征会被输入到全局注意力域适应模块，以衡量那些信号的可迁移性更强，全局注意力域适应模块由一个域分类器G_D构成，其输出代表相应输入信号属于源域的概率，同样用过熵函数计算注意力值，并加权到一个分类损失中，以便辅助分类。The basic structure of the local attention domain adaptation module and the global attention domain adaptation module used in the present invention refers to the domain classifier in the DANN model; the features output from the feature extractor will be divided into K segments and input into the local attention domain adaptation module, which contains K domain classifiers, that is, each signal segment corresponds to a domain classifier

The probability that the corresponding segment belongs to the source domain is calculated to judge the transferability of the segment, and the entropy function is used to calculate the local attention value to achieve weighting of the features; at the same time, in order to ensure that the feature difference between the source domain and the target domain is too large to cause negative transfer, before inputting the local attention domain adaptation module, the present invention uses the maximum mean difference (MMD) to reduce the distribution difference between the source domain and the target domain; the local attention weighted features will be input into the global attention domain adaptation module to measure which signals are more transferable. The global attention domain adaptation module is composed of a domain classifier G _D , whose output represents the probability that the corresponding input signal belongs to the source domain. The attention value is also calculated by the entropy function and weighted to a classification loss to assist classification.

The number of blocks K for the input features in the local attention domain adaptation module is determined based on the output size of the last layer of the feature extractor, which should be the product of the height and width of the feature map. Since the present invention is aimed at a one-dimensional vibration signal, the height is 1, and the number of blocks K is equal to the width of the input feature map.

The global parameters in the process of training the model are set as follows: learning rate is 0.001, batch size is 32, and the number of training times is 500.

As shown in FIG1 , this is a model structure diagram of the rolling bearing domain adaptive fault diagnosis method based on the attention mechanism of the present invention.

The present invention proposes a rolling bearing domain adaptation fault diagnosis method based on attention mechanism, and the network structure consists of two parts: a feature extractor and a domain adaptation module; the feature extractor is composed of a one-dimensional separable convolution stack embedded with a channel attention mechanism and a length attention mechanism, which is used to extract fault-related features; the domain adaptation module is composed of a local attention domain adaptation module and a global attention domain adaptation module, which screens signals and signal fragments with good transferability, thereby improving the generalization ability of the model, so that the model can better cope with variable working condition fault diagnosis problems.

The specific steps are as follows:

Step 1: Data acquisition and preprocessing.

For the same bearing system, vibration monitoring signals of various bearing health conditions under different working conditions are collected respectively, and the data are labeled according to the fault type, so as to obtain bearing data sets for different working conditions. Two of the data sets under different working conditions are selected as the source domain and target domain respectively, and the corresponding training set and test set are divided;

Step 2: Extract fault features through the optimized feature extractor.

Applying one-dimensional separable convolution to replace traditional convolution can reduce the number of model parameters without affecting the diagnostic accuracy. At the same time, the length attention mechanism and the channel attention mechanism are distributed and embedded after the two steps of the one-dimensional separable convolution. The reason is that the spatial convolution in the one-dimensional separable convolution uses independent convolution kernels to filter different channels respectively, so that each channel has its own unique characteristics. The length attention mechanism mainly summarizes the feature information of different channels, and it can play a richer role when placed after the spatial convolution layer. At the same time, spatial convolution does not change the feature dimension. In the present invention, the channel convolution in the one-dimensional separable convolution is not only used to learn the relationship between channels by aggregating the output of the spatial convolution, but also to increase the dimension of the feature space. The present invention places the channel attention mechanism after the channel convolution to ensure that this attention module is applied to a higher-dimensional feature space. The input signal sample x _i can obtain the fault-related features well after passing through the feature extractor G _f .

Step 3: Extract domain invariant features through the optimized domain adaptation module.

将其输入对应的局部域分类器

后，可得到该片段属于源域的概率

域分类器与特征提取器形成对抗关系，通过训练使得域分类器无法分辨特征提取器生成的特征属于源域还是目标域，从而实现对域不变性特征的提取。当概率值

接近0或1时，表示该信号片段可以被域分类器分类，其可迁移性较差。该概率值不能直接用于对特征进行加权，故而使用熵函数

计算每个片段局部注意力值，其中x为离散随机变量X 中的变量值，p为离散随机变量X概率分布函数，每个片段局部注意力值的最终计算公式如下：The features of the signal sample x _i after passing through the feature extractor G _f are divided into K segments, and the features of each segment are

Input it into the corresponding local domain classifier

After that, the probability that the segment belongs to the source domain can be obtained.

The domain classifier and the feature extractor form an adversarial relationship. Through training, the domain classifier cannot distinguish whether the features generated by the feature extractor belong to the source domain or the target domain, thereby realizing the extraction of domain invariant features.

When it is close to 0 or 1, it means that the signal segment can be classified by the domain classifier, and its transferability is poor. This probability value cannot be used directly to weight the feature, so the entropy function is used

Calculate the local attention value of each fragment, where x is the variable value in the discrete random variable X, and p is the probability distribution function of the discrete random variable X. The final calculation formula for the local attention value of each fragment is as follows:

At the same time, the local domain classifier adds a residual connection module to prevent negative transfer caused by incorrect attention values. Therefore, the final features of each segment after local attention weighting are:

同样使用熵函数计算注意力值，得到全局注意力值为：Then, the corresponding segments after local attention weighting are merged according to the positions of the segments before segmentation, and the dimension of the obtained feature _hi is consistent with the dimension of the feature before segmentation. After the local attention weighted feature _hi is input into the global domain classifier G _D , the probability that the signal belongs to the source domain can be obtained.

The entropy function is also used to calculate the attention value, and the global attention value is:

Step 4: Loss function setting.

The loss functions of the present invention include the following four categories:

1) Bearing fault classification loss of source domain dataset:

为交叉熵损失函数，G_y为分类器，h_i为局部注意力加权特征，y_i为样本标签；Among them, n _S is the number of source domain samples, D _S is the source domain sample space,

is the cross entropy loss function, _Gy is the classifier, _hi is the local attention weighted feature, and _yi is the sample label;

2) MMD loss used to reduce the classification difference between the source domain and the target domain:

和

来自源域与目标域中的信号样本；φ(·)是用于将源域及目标域样本映射到同一Hilbert空间的非线性映射，k(·,·)为核函数，用于计算两映射的内积，如上式中

为源域样本和目标域样本经映射后的内积，即

Among them, n _t is the number of samples in the target domain,

and

Signal samples from the source domain and the target domain; φ(·) is a nonlinear mapping used to map the source domain and target domain samples to the same Hilbert space, and k(·,·) is a kernel function used to calculate the inner product of the two mappings, as shown in the above formula.

is the inner product of the source domain sample and the target domain sample after mapping, that is,

及全局注意力域适应损失

3) Local attention domain adaptation loss for extracting local and global domain invariant features from signal segments and signal samples respectively

and global attention domain adaptation loss

The signal segments are samples obtained by evenly dividing the signal samples according to a certain number, and the number of divisions is consistent with the number of blocks when the feature extractor divides the fault features.

是用于域分类的交叉熵损失；Where D _S represents the source domain sample space; D _T represents the target domain sample space; d _i represents the domain label of the vibration signal sample _xi ;

is the cross entropy loss for domain classification;

4) Global weighted entropy loss for auxiliary classification:

Where c represents the number of fault types; pi _,j represents the probability value of signal sample _xi being classified as label j.

Therefore, the total loss function of the present invention is:

分别为特征提取器、标签分类器、全局域分类器及局部域分类器的模型参数，η，γ和λ分别为修正权重。Among them, θ _f , θ _y , θ _d and

are the model parameters of feature extractor, label classifier, global domain classifier and local domain classifier respectively, and η, γ and λ are the correction weights respectively.

Step 5: Optimize policy settings.

The network uses standard back propagation and Adam optimization algorithms to complete the training process while minimizing the total loss function. The parameters of each network structure are updated according to the following formula:

Among them, ε is the learning rate.

Step 6: Perform fault diagnosis on the target domain test set.

The grid search method is used to find the optimal hyper-parameter combination, and the training model is fixed under the optimal hyper-hyperparameter combination. The fault diagnosis model performance is tested on the target domain test set.

Claims (10)

1. A rolling bearing domain adaptation fault diagnosis method based on attention mechanism, characterized by comprising the following steps: For the same bearing system, vibration monitoring data of the health status of various types of bearings under different working conditions are collected respectively, and the data are labeled according to the fault type to obtain bearing data sets for different working conditions; The data in the bearing dataset is divided into a source domain dataset and a target domain dataset, and then divided into a training set and a validation set respectively; Construct a rolling bearing domain adaptive fault diagnosis model, and train the model using the training set of the source domain dataset and the training set of the target domain dataset; Optimize the domain-adaptive fault diagnosis model for rolling bearings; The test set of the target domain data is input into the optimized rolling bearing domain adaptive fault diagnosis model to obtain the fault type of the target domain data. 2. According to a rolling bearing domain adaptive fault diagnosis method based on an attention mechanism according to claim 1, it is characterized in that the source domain data is vibration monitoring data with a quantity greater than a threshold and complete labels, and the target domain data is vibration monitoring data under the working conditions to be tested. 3. According to the method for rolling bearing domain adaptation fault diagnosis based on attention mechanism in claim 1, it is characterized in that the construction of rolling bearing domain adaptation fault diagnosis model comprises: A feature extractor is used to extract fault features in the bearing dataset and divide the fault features into multiple segments; The local attention domain adaptation module is used to calculate the local attention value of each segment fault feature, perform weighted processing on it, and dimensionally merge the weighted local attention value of each segment; The global attention domain adaptation module is used to calculate the global attention value based on the local attention values after dimension merging, and weight it into a classification loss to complete the fault type classification of the input data. 4. According to a rolling bearing domain adaptive fault diagnosis method based on attention mechanism according to claim 3, it is characterized in that the feature extractor includes sequentially connected spatial convolution, length attention module, channel convolution, channel attention module and pooling layer. 5. According to claim 3, a rolling bearing domain adaptation fault diagnosis method based on an attention mechanism is characterized in that the local attention domain adaptation module includes: a maximum mean difference module, multiple domain classifiers and a residual connection module, wherein the fault features of multiple fragments output by the feature extractor are processed by the maximum mean difference module, and the fault features of each fragment are input into a domain classifier, and the probability that the corresponding fragment belongs to the source domain is calculated, and the entropy function is used to calculate the local attention value of the fault feature of each fragment to weight the fault feature, and the weighted fault features of all fragments are dimensionally merged and output through the residual connection module. 6. According to a rolling bearing domain adaptation fault diagnosis method based on an attention mechanism as described in claim 3, it is characterized in that the global attention domain adaptation module includes: a domain classifier, which inputs the weighted fault feature into the domain classifier, calculates the probability that the fault feature belongs to the source domain, and uses an entropy function to calculate the global attention value of the fault feature to weight the fault feature. 7. A rolling bearing domain adaptation fault diagnosis method based on attention mechanism according to claim 3, characterized in that in the construction of the rolling bearing domain adaptation fault diagnosis model, four types of loss functions are set, specifically: 1) Bearing fault classification loss of source domain dataset:

Among them, n _S is the number of source domain samples, D _S is the source domain sample space,

is the cross entropy loss function, _Gy is the classifier, _hi is the local attention weighted feature, and _yi is the sample label; 2) MMD loss used to reduce the classification difference between the source domain and the target domain:

Among them, n _t is the number of samples in the target domain,

and

Signal samples from the source domain and the target domain; φ(·) is a nonlinear mapping used to map the source domain and target domain samples to the same Hilbert space, and k(·,·) is a kernel function; 3) Local attention domain adaptation loss for extracting local domain and global domain invariant features from signal segments and signal samples respectively

and global attention domain adaptation loss

Where D _S represents the source domain sample space; D _T represents the target domain sample space, and d _i represents the domain label of the signal sample _xi ;

is the cross entropy loss for domain classification; 4) Global weighted entropy loss for auxiliary classification:

Where c represents the number of fault types; pi _,j represents the probability value of signal sample _xi being classified as label j. 8. The rolling bearing domain adaptation fault diagnosis method based on attention mechanism according to claim 7, characterized in that the total loss function is:

where, θ _f , θ _y , θ _d and

are the model parameters of the feature extractor, label classifier, global attention domain adaptation module and local attention domain adaptation module, respectively. η, γ and λ are the correction weights respectively. 9. A rolling bearing domain adaptation fault diagnosis method based on an attention mechanism according to claim 1 or 8, characterized in that the rolling bearing domain adaptation fault diagnosis model is optimized using back propagation and Adam optimization algorithms, that is, the total loss function is minimized. 10. A rolling bearing domain adaptation fault diagnosis method based on attention mechanism according to claim 8, characterized in that the model parameters are updated using the following formula:

Among them, ε is the learning rate.

Images