A bearing fault diagnosis method under variable load based on subdomain adaptive adversarial network

Abstract

The present invention discloses a bearing fault diagnosis method under variable load based on a subdomain adaptive adversarial network, which belongs to the field of rotating machinery fault diagnosis. The present invention utilizes a one-dimensional convolutional subdomain adaptive adversarial network model to extract domain-invariant features, establishes an effective cross-load rolling bearing fault diagnosis model, realizes unsupervised fault diagnosis on the target domain, alleviates the problem of scarce fault data labeling information in practical applications, and improves the effectiveness and accuracy of rolling bearing fault diagnosis, thereby ensuring the safety and reliability of modern rotating machinery operation and the rapidity and accuracy of equipment component supply.

Description

Technical Field

[0001] The present invention belongs to the field of rotating machinery fault diagnosis, and in particular relates to a bearing fault diagnosis method under variable load based on a subdomain adaptive adversarial network. Background Art

[0002] Rolling bearings are common parts of rotating machinery. They are prone to wear and tear after long-term operation in harsh environments, which can lead to failures. Research on rolling bearing fault identification helps to timely discover the location and extent of the fault, ensure the safe and stable operation of the equipment, and at the same time put forward clear and rapid requirements for manufacturing companies.

[0003] The running status information of rolling bearings is often reflected through their vibration signals. With the development of data-driven fault diagnosis methods, research based on rolling bearing vibration signals has become a hot topic. There have been a lot of studies on diagnosis based on traditional signal processing methods. The vibration data is processed using time-frequency domain analysis technology to extract weak signals that can reflect faults. The main methods include adaptive modal decomposition, Fourier transform, short-time Fourier transform, wavelet transform, envelope spectrum analysis, etc. However, such methods have their limitations and require prior information on fault characteristic frequencies. Machine learning has been widely used in the field of fault diagnosis. Traditional machine learning methods are mainly based on the manual selection and construction of vibration signal feature vectors, and the selected features are used to train the diagnosis model. The main models used are support vector machines, K nearest neighbors, decision trees, etc. From the mechanism of action, it can be seen that the diagnostic effect of such methods is highly related to the composition of feature vectors. Deep learning methods can establish an end-to-end model of vibration signals, autonomously learn features and then identify health status. Convolutional neural networks (CNNs), which are widely used in deep learning, have the characteristics of parameter sharing, sparse connections, and translation invariance. They have small data processing volume, low training difficulty, and significant effects. At the same time, based on the one-dimensional representation of the vibration signal, the use of one-dimensional CNN is more suitable for the application scenario of the present invention.

[0004] Deep learning methods require that training samples and test samples have the same feature space and probability distribution. However, in actual fault diagnosis scenarios, due to the working environment and load changes of rotating machinery, the training data and test data often have distribution differences. In addition, after a fault occurs, it is difficult to obtain labeled data, and the fault data under certain operating loads is unsupervised. This reduces the diagnostic effect of the model, and the generalization ability of the model cannot be guaranteed, which brings challenges to fault diagnosis. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies of the prior art and propose a bearing fault diagnosis method under variable load based on a subdomain adaptive adversarial network. The method introduces a transfer learning framework to study the establishment and optimization of fault classification models and improve the universality of fault information between different loads. First, the bearing vibration signals under different operating loads are collected as source domain and target domain data, and variational mode decomposition (VMD) is performed on the vibration signal characteristics as signal preprocessing; secondly, the source domain and target domain data after VMD are input into a one-dimensional convolutional subdomain adaptive adversarial network to shorten the feature space distance between the source domain and the target domain, extract the invariant features between the domains, realize the subdomain feature alignment between the source domain and the target domain, and realize the knowledge transfer from the supervised source domain to the unsupervised target domain; finally, the trained model is used to perform fault diagnosis of the unsupervised signal in the target domain.

[0006] The object of the present invention is achieved through the following technical solution: a bearing fault diagnosis method under variable load based on a subdomain adaptive adversarial network, comprising the following steps:

[0007] (1) Collect vibration signals of rolling bearings in multiple healthy states under different operating loads as a bearing fault vibration signal database, use vibration signals under loads with known fault information as source domain signals, and use vibration signals under loads with unknown fault information as target domain signals;

[0008] (2) Perform non-overlapping equal-length divisions on the multiple health status signals in the source and target domains;

[0009] (3) The divided source domain signal and target domain signal are subjected to variational mode decomposition respectively, and the intrinsic mode function obtained by decomposition constitutes the source domain dataset D S With the target domain dataset D T ;

[0010] (4) Establish a fault diagnosis model based on a one-dimensional convolutional subdomain adaptive adversarial network; the model includes a feature extractor G F , label classifier G Y and domain discriminator G D ; Feature extractor G F Used to implicitly extract data features, using a one-dimensional convolutional neural network; label classifier G Y Used to classify source domain data and connect feature extractor G F ; Domain discriminator G D Used to determine whether the data comes from the source domain or the target domain, and also connected to the feature extractor G F ; Feature extractor G F With domain discriminator G D There is a gradient reversal layer in between;

[0011] (5) Input the source domain signal of known fault information and the target domain signal of unknown fault information into the one-dimensional convolutional subdomain adaptive adversarial network for model training, and fix the model parameters;

[0012] (6) During the online diagnosis process, a vibration signal under a target domain load is collected and subjected to variational modal decomposition, which is then input into the trained model and the fault diagnosis result is output.

[0013] Furthermore, the health status vibration signal in step (1) includes vibration acceleration signals of the bearing having multiple health statuses, including normal, inner ring fault, outer ring fault, rolling element fault, and different degrees of degradation.

[0014] Furthermore, in step (3), the source domain signal and the target domain signal are subjected to variational mode decomposition respectively, and the VMD parameters are set. To ensure the consistency of data dimension, the number of decompositions is the same when VMD is performed on signals under different health states. The intrinsic mode functions (IMFs) obtained by decomposition are used as samples to form the source domain dataset D S With the target domain dataset D T :

[0015]

[0016]

[0017] in and denote the i-th source domain sample and target domain sample respectively, Represents the sample data label of the source domain, n s and n t Represent the number of samples in the source domain and the target domain respectively; the feature spaces of the two domains are the same as the category space, that is, X S =X T With Y S =Y T , but the data distribution of the two domains is different, that is, P(X S )≠P(X T ).

[0018] Furthermore, in step (4), the fault diagnosis model based on the one-dimensional convolutional subdomain adaptive adversarial network is established as follows:

[0019] (4-1) The feature extractor consists of three convolutional pooling modules. The first two pooling layers use maximum pooling operations, and the last pooling layer uses average pooling. To prevent network overfitting, a dropout layer is introduced. The source domain samples and the target domain samples share the feature extractor.

[0020] (4-2) The label classifier contains two fully connected layers. The number of neurons in the last layer is the same as the number of categories. The softmax function is used to map the output of the last fully connected layer to the (0,1) interval, which can be expressed as:

[0021]

[0022] where μ i is the output value of the i-th node, C is the number of output nodes, that is, the number of categories;

[0023] (4-3) The domain discriminator confuses the source domain and target domain data. There is a gradient reversal layer (GRL) between the feature extractor and the domain discriminator, which is expressed as follows:

[0024] R λ (x) = x

[0025]

[0026] R λ (x) is a pseudo function that defines the gradient reversal layer, x is a network parameter, λ is a constant parameter that controls the weight of the domain discriminator in the network, and is generally selected based on experience. I is the unit matrix. During the back-propagation process, the parameters of the domain discriminator are automatically reversed before the loss is propagated to the feature extractor, so that the training objectives of the feature extractor and the domain discriminator are opposite, thus forming an adversarial relationship between the two. The domain discriminator contains two fully connected layers, and the number of neurons in the last layer is 1, which determines whether the sample comes from the source domain or the target domain.

[0027] Furthermore, in step (5), known fault information source domain samples and unknown fault information target domain samples are input simultaneously, and hyperparameters such as the number of iterations, optimizer, learning rate, and sample batch size are set to perform model training; the model optimization objectives during the training process are as follows:

[0028] (5-1) The first optimization goal is to minimize the fault classification loss function on the source domain data set, hoping that the label classifier outputs the correct predicted label; the label classifier loss function uses the cross entropy loss function, which is defined as follows:

[0029]

[0030] where p=[p0,...,p C-1 ] is a probability distribution, element p i represents the probability that the sample belongs to the i-th category, y=[y0,...,y C-1 ] is the one-hot representation of the sample label. When the sample belongs to the i-th category, y i =1, otherwise y i =0; this part of the loss function is:

[0031] J F,Y =J ce (G Y (G F (x s )),y s )

[0032] where x s is the source domain sample, y s is the source domain sample label J ce (·) is the cross entropy loss function;

[0033] (5-2) The second optimization goal is that the domain discriminator hopes to distinguish samples from the source domain and the target domain, but the feature extractor hopes that the extracted features can confuse the domain discriminator, so that it cannot determine the source of the sample. The two are in a confrontational game. Given that the domain label of the source domain sample is 1 and the domain label of the target domain sample is 0, the optimization goal can be written as:

[0034]

[0035] Among them, n s and n t are the number of samples in the source domain and the target domain respectively, and are source domain and target domain samples respectively;

[0036] (5-3) The third optimization goal is to minimize the feature representation distance between the source domain and the target domain dataset after feature extraction, hoping to learn domain-independent features; the local maximum mean difference of the fused sample-level weight (sample-level weight LMMD, slw-LMMD) is used to measure the feature representation distance. The slw-LMMD is defined as follows:

[0037]

[0038] where p (c) and q (c) are the distribution of source domain data category c and target domain data set category c, respectively, s and z t are the outputs of the source domain and target domain instances after the feature extractor, E c Refers to the expectation on category c, According to the distribution p (c) Seek hope, According to the distribution q (c) Seek hope, To define the notation, H is the reproducing kernel Hilbert space, φ(·) is the function that maps the original data to H; since the data in the target domain is unsupervised, a soft label method is used to predict the probability that the sample belongs to category c; slw-LMMD considers the sample-level weight of the target domain on the basis of LMMD. The degree of association and similarity between each sample in the target domain and the source domain sample are different, so when calculating the weight, different weights are considered for each sample in the target domain, which is more conducive to inter-domain alignment; then each sample is assigned a weight according to the weight ω c belongs to each class, so the unbiased estimator of slw-LMMD is:

[0039]

[0040]

[0041]

[0042]

[0043] in and are the outputs of the feature extractor for the i-th instance in the source domain and the j-th instance in the target domain, respectively. and Respectively represent and The weight of belonging to category c, is the output of the target domain sample after the domain discriminator. Since the domain label of the source domain sample is 1, The output range is 0 to 1. The larger the value, the greater the similarity between the target domain sample and the source domain sample, and the higher the weight should be in cross-domain alignment; ic is the vector y i The cth element of , the source domain data uses the real one-hot label to calculate the weight, and the target domain data uses the softmax output of the label classifier to calculate the weight; slw-LMMD considers the sample-level weight, calculates the instance mapping value on each category, calculates the distance on each subdomain separately, and then finds the expected value on the category; this part of the loss function term is:

[0044]

[0045] (5-4) Due to the existence of the gradient reversal layer between the feature extractor and the domain discriminator, the training process is simplified, and the purpose of adversarial training can be achieved within one cycle, so that the overall optimization goal can be written as a schematic formula:

[0046] minJ=J F,Y +αJ F -λJ F,D

[0047] Among them, α is the weight hyperparameter of the feature representation distance metric slw-LMMD, and λ is actually the parameter of the gradient reversal layer. Due to the existence of GRL, the optimization target of the part of the network that actually belongs to the domain discriminator is still minJ F,D .

[0048] The beneficial effects of the present invention are:

[0049] (1) The present invention proposes a bearing fault diagnosis method under variable load based on a subdomain adaptive adversarial network, which is more in line with the actual fault occurrence scenario, can face the challenge of unsupervised rolling bearing fault data, and utilize the fault marking information under different loads to realize cross-domain bearing fault diagnosis.

[0050] (2) The present invention can utilize the fine-grained information of fault data and use the soft label method to divide the unlabeled target domain fault data into subdomains, thereby achieving subdomain alignment of the feature representation of the source domain and target domain data.

[0051] (3) The present invention can take into account the sample-level weight of each sample in the target domain, so that they have different importance levels in the process of alignment between sub-domains, which is more conducive to inter-domain feature alignment.

[0052] (4) As a type of time series data, vibration signals retain their signal-level periodicity, impact and other characteristics, and have time-frequency domain properties. The present invention uses variational mode decomposition as data preprocessing, which can adaptively decompose the signal into finite intrinsic mode functions, realize the frequency domain division of the signal, and the decomposition has good sparsity and effectiveness, which is beneficial to the subsequent feature extraction step. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of the bearing fault diagnosis method under variable load based on the subdomain adaptive adversarial network in the present invention.

[0054] Figure 2 This is the structure diagram of the one-dimensional convolutional subdomain-adapted adversarial network.

[0055] Figure 3 This is the convergence curve of the loss function in the network training process.

[0056] Figure 4 Confusion matrix for model testing results. DETAILED DESCRIPTION

[0057] The present invention is a bearing fault diagnosis method under variable load based on a subdomain adaptive adversarial network, and the method is aimed at the fault diagnosis problem of rotating machinery. The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0058] The embodiment combines the bearing failure data of Western Reserve University to illustrate the effectiveness of the method of the present invention. The main components of the test bench are: 1.5KW motor, torque sensor / decoder, power tester, electronic controller. Figure 1 As shown, the implementation steps of the present invention are described in detail as follows:

[0059] Step 1: Collect vibration acceleration signals at specific positions of the bearing to be tested under different loads that can reflect various health states of the bearing. The data is the vibration acceleration data of the drive-end bearing collected by the acceleration sensor installed above the bearing seat, with a sampling frequency of 12kHz; the motor has four load conditions of 0, 1, 2, and 3HP. The vibration signal under 0HP is used as the source domain signal with known label information, and the vibration signal under 3HP is used as the target domain signal with unknown label information; the bearing has four faults: normal, inner ring fault, rolling element fault, and outer ring fault, and the inner ring, rolling element, and outer ring contain three faults with fault diameters of 7mil, 14mil, and 21mil respectively, so the bearing has a total of 10 health states.

[0060] Step 2: Set the length of each sample to 1024, and use non-overlapping sampling to divide the vibration signal of each health state into 100 samples. Then the source domain and the target domain each have 1000 samples containing 10 health states.

[0061] Step 3: Perform variational mode decomposition (VMD) on the divided source domain signal and target domain signal respectively. Set VMD parameters, set penalty factor α to 2000, noise tolerance τ to 0, mode number K to 5, DC component DC to 0, convergence accuracy ε to 1*10 -7 The initialization center frequency is uniform initialization. To ensure the consistency of data dimension, the parameter K takes the same value of 5 when performing VMD on signals under different health states. The decomposed IMFs are used as samples to form the source domain dataset D S With the target domain dataset D T :

[0062]

[0063]

[0064] in and denote the i-th source domain sample and target domain sample respectively, Represents the sample data label of the source domain, n s and n t Represent the number of samples in the source domain and the target domain respectively; the feature spaces of the two domains are the same as the category space, that is, X S =X T With YS =Y T , but the data distribution of the two domains is different, that is, P(X S )≠P(X T ).

[0065] Step 4: Dataset D S and D T Divided into training set and test set, divided according to the ratio of 7:3, that is, in the source domain and target domain, 70 samples of each health state belong to the training set, and 30 samples belong to the test set. The source domain contains 700 training set samples and 300 test set samples, and the target domain also contains 700 training set samples and 300 test set samples. The training set is used for model training, and the test set is used to predict the diagnostic effect of the model. The details of the single data domain dataset are shown in Table 1.

[0066] Table 1 Rolling bearing data set information

[0067]

[0068] Step 5: Establish a fault diagnosis model based on one-dimensional convolutional subdomain adaptive adversarial network. The model can be divided into three parts: feature extractor G F , label classifier G Y , domain discriminator G D , the network structure is as follows Figure 2 As shown, specifically:

[0069] The feature extractor is used to implicitly extract data features and uses a one-dimensional convolutional neural network. The feature extractor consists of three convolutional pooling modules. The first two pooling layers use maximum pooling operations, and the last pooling layer uses average pooling. To prevent the network from overfitting, a dropout layer is introduced. The source domain and target domain samples share the feature extractor.

[0070] The label classifier is used to classify the source domain data, connected to the feature extractor, and contains two fully connected layers. The number of neurons in the last layer is the same as the number of categories, which is set to 10. The softmax function is used to map the output of the last fully connected layer to the (0,1) interval, which can be expressed as:

[0071]

[0072] where μ i is the output value of the i-th node, and C is the number of output nodes, that is, the number of categories.

[0073] The domain discriminator is used to determine whether the data comes from the source domain or the target domain. It is also connected to the feature extractor. However, since the purpose of the present invention is to confuse the source domain and target domain data by the domain discriminator, there is a gradient reversal layer (GRL) between the feature extractor and the domain discriminator, which is expressed as follows:

[0074] R λ (x) = x

[0075]

[0076] R λ (x) is a pseudo function that defines the gradient reversal layer, x is a network parameter, λ is a constant parameter that controls the role of the domain discriminator in the network, which is generally selected based on experience, and I is the unit matrix. During the back propagation process, the parameters of the domain discriminator are automatically reversed before the loss is propagated to the feature extractor, so that the training objectives of the feature extractor and the domain discriminator are opposite, thus forming an adversarial relationship between the two. The domain discriminator contains two fully connected layers, and the number of neurons in the last layer is 1, which determines whether the sample comes from the source domain or the target domain.

[0077] The network parameters are shown in Table 2.

[0078] Table 2 Parameters of one-dimensional convolution subdomain adaptation adversarial network

[0079]

[0080] Step 6: Input the labeled source domain training set and the unlabeled target domain training set, set the number of iterations to 50, the sample batch size to 64, select the Adam optimizer, set the initial learning rate to 0.005, adjust the learning rate according to exponential decay during the iteration process, and perform model training. The model optimization goals during the training process include three components, as follows:

[0081] The first optimization goal is to minimize the fault classification loss function on the source domain dataset, hoping that the label classifier outputs the correct predicted label. The label classifier loss function uses the cross entropy loss function, which is defined as follows:

[0082]

[0083] where p=[p0,...,p C-1 ] is a probability distribution, element p i represents the probability that the sample belongs to the i-th category, y=[y0,...,y C-1 ] is the one-hot representation of the sample label. When the sample belongs to the i-th category, y i =1, otherwise y i = 0. This part of the loss function is:

[0084] J F,Y =J ce (G Y (G F (x s )),y s )

[0085] where x s is the source domain sample, y s is the source domain sample label J ce (·) is the cross entropy loss function.

[0086] The second optimization goal is that the domain discriminator hopes to distinguish samples from the source domain and the target domain, but the feature extractor hopes that the extracted features can confuse the domain discriminator so that it cannot determine the source of the sample, and the two are in a confrontational game. Given that the domain label of the source domain sample is 1 and the domain label of the target domain sample is 0, the optimization goal can be written as:

[0087]

[0088] Among them, n s and n t are the number of samples in the source domain and the target domain respectively, and are source domain and target domain samples respectively.

[0089] The third optimization goal is to minimize the feature representation distance between the source domain and the target domain dataset after feature extraction, hoping to learn domain-independent features. The feature representation distance is measured using the local maximum mean difference of the fused sample-level weight (sample-level weight LMMD, slw-LMMD), and slw-LMMD is defined as follows:

[0090]

[0091] where p (c) and q (c) are the distribution of source domain data category c and target domain data set category c, respectively, z s and z t are the outputs of the source domain and target domain instances after the feature extractor, E c Refers to the expectation on category c, According to the distribution p (c) Seek hope, According to the distribution q (c) Seek hope, To define the notation, H is the reproducing kernel Hilbert space, and φ(·) is the function that maps the original data to H. Since the target domain data is unsupervised, a soft label method is used to predict the probability that the sample belongs to category c. slw-LMMD considers the sample-level weight of the target domain on the basis of LMMD. The degree of association and similarity between each sample in the target domain and the source domain sample are different. Therefore, when calculating the weight, different weights are considered for each sample in the target domain, which is more conducive to inter-domain alignment. Then each sample is given a weight ω. cbelongs to each class, so the unbiased estimator of slw-LMMD is:

[0092]

[0093]

[0094]

[0095]

[0096] in and are the outputs of the feature extractor for the i-th instance in the source domain and the j-th instance in the target domain, respectively. and Respectively represent and The weight of belonging to category c, is the output of the target domain sample after the domain discriminator. Since the domain label of the source domain sample is 1, The output range is 0 to 1. The larger the value, the greater the similarity between the target domain sample and the source domain sample, and the higher the weight should be when aligning between domains. ic is the vector y i The cth element of , the source domain data uses the real one-hot label to calculate the weight, and the target domain data uses the softmax output of the label classifier to calculate the weight. slw-LMMD considers the sample-level weight, calculates the instance mapping value on each category, calculates the distance on each subdomain separately, and then calculates the expected value on the category. The kernel technique is introduced for actual calculation:

[0097]

[0098] Where k(,) selects the Gaussian kernel function. The loss function term for this part is:

[0099]

[0100] Therefore, due to the existence of the gradient reversal layer between the feature extractor and the domain discriminator, the training process is simplified, and the purpose of adversarial training can be achieved in one cycle, so that the overall optimization goal can be written as a schematic formula:

[0101] minJ=J F,Y +αJ F -λJ F,D

[0102] Among them, α is the weight hyperparameter of the feature representation distance metric slw-LMMD, and λ is actually the parameter of the gradient reversal layer in the fifth step. Due to the existence of GRL, the optimization target of the part of the network that actually belongs to the domain discriminator is still minJ F,D . Parameters α and λ increase progressively as the number of network iterations increases, as follows:

[0103]

[0104]

[0105] Wherein, τ is a coefficient for controlling the change speed of the parameters α and λ, and is set to 10 in this embodiment; epoch is the current iteration number, and Nepoch is the total number of iterations.

[0106] The convergence curve of the loss function during the training process is as follows Figure 3 shown.

[0107] Step 7: Input the target domain test set into the trained model for testing and output the fault diagnosis results. At this time, we only care about the output results of the target domain test set data after passing through the feature extractor and label classifier. The confusion matrix of the model test results is as follows: Figure 4 As shown, it shows that the present invention can obtain relatively accurate diagnosis results. In order to illustrate the feasibility and advantages of the present invention, two groups of schemes are compared. One is to train the network using source domain data without subdomain adaptation; the other is to use the classic transfer learning method deep adaptation network (DAN). The comparison of diagnostic accuracy is shown in Table 3.

[0108] Table 3 Accuracy comparison

[0109] No domain adaptation DAN Method of the present invention Accuracy 70.67％ 83.67％ 98.67％

[0110] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A bearing fault diagnosis method under variable load based on subdomain adaptive adversarial network, characterized in that: The following steps are involved: (1) Collect vibration signals of rolling bearings in multiple healthy states under different operating loads as a bearing fault vibration signal database, use vibration signals under loads with known fault information as source domain signals, and use vibration signals under loads with unknown fault information as target domain signals; (2) Perform non-overlapping equal-length divisions on the multiple health status signals in the source and target domains; (3) The divided source domain signal and target domain signal are subjected to variational mode decomposition respectively, and the intrinsic mode function obtained by decomposition constitutes the source domain dataset D S With the target domain dataset D T ; (4) Establish a fault diagnosis model based on a one-dimensional convolutional subdomain adaptive adversarial network; the model includes a feature extractor G F , label classifier G Y and domain discriminator G D ; Feature extractor G F Used to implicitly extract data features, using a one-dimensional convolutional neural network; label classifier G Y Used to classify source domain data and connect feature extractor G F ; Domain discriminator G D Used to determine whether the data comes from the source domain or the target domain, and also connected to the feature extractor G F ; Feature extractor G F With domain discriminator G D There is a gradient reversal layer in between; (5) Input the source domain signal of known fault information and the target domain signal of unknown fault information into the one-dimensional convolutional subdomain adaptive adversarial network for model training, and fix the model parameters; (6) During the online diagnosis process, a vibration signal under a target domain load is collected and subjected to variational modal decomposition, which is then input into the trained model and the fault diagnosis result is output.

2. The bearing fault diagnosis method under variable load based on subdomain adaptive adversarial network according to claim 1 is characterized in that: The health status vibration signal in step (1) includes vibration acceleration signals of the bearing having multiple health statuses, including normal, inner ring fault, outer ring fault, rolling element fault, and different degradation degrees.

3. The bearing fault diagnosis method under variable load based on subdomain adaptive adversarial network according to claim 1 is characterized in that: In step (3), the source domain signal and the target domain signal are respectively subjected to variational mode decomposition, and the VMD parameters are set. To ensure the consistency of data dimension, the number of decompositions is the same when VMD is performed on signals under different health states. The intrinsic mode functions (IMFs) obtained by decomposition are used as samples to form the source domain dataset D S With the target domain dataset D T : in and denote the i-th source domain sample and target domain sample respectively, Represents the sample data label of the source domain, n s and n t Represent the number of samples in the source domain and the target domain respectively; the feature spaces of the two domains are the same as the category space, that is, X S =X T With Y S =Y T , but the data distribution of the two domains is different, that is, P(X S )≠P(X T ).

4. The bearing fault diagnosis method under variable load based on subdomain adaptive adversarial network according to claim 1 is characterized in that: In step (4), the fault diagnosis model based on the one-dimensional convolutional subdomain adaptive adversarial network is established as follows: (4-1) The feature extractor consists of three convolutional pooling modules. The first two pooling layers use maximum pooling operations, and the last pooling layer uses average pooling. To prevent network overfitting, a dropout layer is introduced. The source domain samples and the target domain samples share the feature extractor. (4-2) The label classifier contains two fully connected layers. The number of neurons in the last layer is the same as the number of categories. The softmax function is used to map the output of the last fully connected layer to the (0,1) interval, which can be expressed as: where μ i is the output value of the i-th node, C is the number of output nodes, that is, the number of categories; (4-3) The domain discriminator confuses the source domain and target domain data. There is a gradient reversal layer (GRL) between the feature extractor and the domain discriminator, which is expressed as follows: R λ (x)=x R λ (x) is a pseudo function that defines the gradient reversal layer, x is a network parameter, λ is a constant parameter that controls the weight of the domain discriminator in the network, and is generally selected based on experience. I is the unit matrix. During the back-propagation process, the parameters of the domain discriminator are automatically reversed before the loss is propagated to the feature extractor, so that the training objectives of the feature extractor and the domain discriminator are opposite, thus forming an adversarial relationship between the two. The domain discriminator contains two fully connected layers, and the number of neurons in the last layer is 1, which determines whether the sample comes from the source domain or the target domain.

5. The bearing fault diagnosis method under variable load based on subdomain adaptive adversarial network according to claim 1 is characterized in that: In step (5), the known fault information source domain samples and the unknown fault information target domain samples are input simultaneously, and the hyperparameters such as the number of iterations, optimizer, learning rate, and sample batch size are set to perform model training. The model optimization objectives during the training process are as follows: (5-1) The first optimization goal is to minimize the fault classification loss function on the source domain data set, hoping that the label classifier outputs the correct predicted label; the label classifier loss function uses the cross entropy loss function, which is defined as follows: where p=[p0,...,p C-1 ] is a probability distribution, element p i represents the probability that the sample belongs to the i-th category, y=[y0,...,y C-1 ] is the one-hot representation of the sample label. When the sample belongs to the i-th category, y i =1, otherwise y i =0; this part of the loss function is: J F,Y =J ce (G Y (G F (x s )),y s ) where x s is the source domain sample, y s is the source domain sample label J ce (·) is the cross entropy loss function; (5-2) The second optimization goal is that the domain discriminator hopes to distinguish samples from the source domain and the target domain, but the feature extractor hopes that the extracted features can confuse the domain discriminator, so that it cannot determine the source of the sample. The two are in a confrontational game. Given that the domain label of the source domain sample is 1 and the domain label of the target domain sample is 0, the optimization goal can be written as: Among them, n s and n t are the number of samples in the source domain and the target domain respectively, and are source domain and target domain samples respectively; (5-3) The third optimization goal is to minimize the feature representation distance between the source domain and the target domain dataset after feature extraction, hoping to learn domain-independent features; the local maximum mean difference of the fused sample-level weight (sample-level weight LMMD, slw-LMMD) is used to measure the feature representation distance. The slw-LMMD is defined as follows: where p (c) and q (c) are the distribution of source domain data category c and target domain data set category c, respectively, s and z t are the outputs of the source domain and target domain instances after the feature extractor, E c Refers to the expectation on category c, According to the distribution p (c) Seek hope, According to the distribution q (c) Seek hope, To define the notation, H is the reproducing kernel Hilbert space, φ(·) is the function that maps the original data to H; since the data in the target domain is unsupervised, a soft label method is used to predict the probability that the sample belongs to category c; slw-LMMD considers the sample-level weight of the target domain on the basis of LMMD. The degree of association and similarity between each sample in the target domain and the source domain sample are different, so when calculating the weight, different weights are considered for each sample in the target domain, which is more conducive to inter-domain alignment; then each sample is assigned a weight according to the weight ω c belongs to each class, so the unbiased estimator of slw-LMMD is: in and are the outputs of the feature extractor for the i-th instance in the source domain and the j-th instance in the target domain, respectively. and Respectively represent and The weight of belonging to category c, is the output of the target domain sample after the domain discriminator. Since the domain label of the source domain sample is 1, The output range is 0 to 1. The larger the value, the greater the similarity between the target domain sample and the source domain sample, and the higher the weight should be in cross-domain alignment; ic is the vector y i The cth element of , the source domain data uses the real one-hot label to calculate the weight, and the target domain data uses the softmax output of the label classifier to calculate the weight; slw-LMMD considers the sample-level weight, calculates the instance mapping value on each category, calculates the distance on each subdomain separately, and then finds the expected value on the category; this part of the loss function term is: (5-4) Due to the existence of the gradient reversal layer between the feature extractor and the domain discriminator, the training process is simplified, and the purpose of adversarial training can be achieved within one cycle, so that the overall optimization goal can be written as a schematic formula: min J=J F,Y +αJ F -λJ F,D Where α is the weight hyperparameter of the feature representation distance metric slw-LMMD, and λ is actually the parameter of the gradient reversal layer. Due to the existence of GRL, the optimization target of the part of the network that actually belongs to the domain discriminator is still min J F,D .

Images