Bearing fault feature extraction model based on field generalization and bearing fault diagnosis method

Abstract

The invention belongs to the field of bearing fault diagnosis, and particularly relates to a bearing fault feature extraction model based on field generalization and a bearing fault diagnosis method. A training mode of the model comprises the following steps: 1) inputting phase component information in a Fourier spectrum of a multi-source domain bearing vibration signal into a feature extraction module of the model to extract preliminary features; dividing the preliminary features into two parts, calculating internal invariant features through an internal invariant feature module according to the part corresponding to the product of the feature separation coefficient and the preliminary features, and calculating mutual invariant features through a mutual invariant feature module according to the remaining part; the feature separation coefficient is obtained according to the ratio of loss function values corresponding to the internal invariant feature module and the mutual invariant feature module; 2) performing parameter updating on the model by taking minimization of a total loss function as a target; iterating 1)-2) until a stop condition is met; the total loss function comprises a bearing fault diagnosis result determined by the bearing fault feature and a loss item corresponding to the bearing fault real label.

Description

Technical Field

[0001] The present invention belongs to the field of bearing fault diagnosis, and specifically relates to a bearing fault feature extraction model based on domain generalization and a bearing fault diagnosis method. Background Art

[0002] In recent years, domain adaptation technology has been proposed to solve the problem of bearing fault diagnosis when the target domain samples are unlabeled. Through a lot of research, domain adaptation technology has achieved great success in the field of bearing fault diagnosis. However, domain adaptation technology is essentially unable to solve the problem of unknown target domain samples. Therefore, domain generalization technology is proposed, which can achieve the purpose of state diagnosis of the target bearing when the target domain information is completely missing during model training.

[0003] Compared with the extensive application of domain adaptation technology in the field of bearing fault diagnosis, there are only a few studies on bearing fault diagnosis technology based on domain generalization. In the existing domain generalization network for bearing fault feature extraction by extracting domain invariant features, the feature extractor G used for preliminary feature extraction is usually replaced by the f The output features are divided into two parts, which are used to calculate internal invariant features (also called intra-domain invariant features) and mutual invariant features (also called inter-domain invariant features). However, when the internal invariant features or any of the mutual invariant features in the domain generalization network play a very limited role in the domain generalization process, the domain generalization process will find it difficult to utilize the information contained in this part of the features, thereby reducing the generalization performance of the network and causing a negative impact on the accuracy of the diagnostic results obtained by using the extracted bearing fault features for bearing fault diagnosis. Summary of the invention

[0004] The purpose of the present invention is to provide a bearing fault feature extraction model and a bearing fault diagnosis method based on domain generalization, which are used to solve the problem in the prior art that the generalization performance of the bearing fault feature extraction network is reduced due to the extremely limited role played by any feature part of the internal invariant features or mutually invariant features in the domain generalization process.

[0005] In order to achieve the above object, the present invention provides a bearing fault feature extraction model based on domain generalization, and the training method of the model includes:

[0006] 1) Inputting the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal used for training into the feature extraction module of the model to extract preliminary features; dividing the preliminary features into two parts, calculating the internal invariant features according to the part corresponding to the product of the feature separation coefficient and the features through the internal invariant feature module in the model, and calculating the mutual invariant features according to the remaining part through the mutual invariant feature module in the model; the feature separation coefficient is obtained according to the ratio of the value of the loss function corresponding to the internal invariant feature module to the sum of the values ​​of the loss functions corresponding to the mutual invariant feature module and the internal invariant feature module;

[0007] 2) updating the parameters of the model with the goal of minimizing the total loss function; iterating steps 1)-2) until the iteration stop condition is met; the total loss function includes a loss term corresponding to the bearing fault diagnosis result determined according to the bearing fault characteristics obtained by the internal invariant features and the mutual invariant features and the true label of the bearing fault.

[0008] Furthermore, the total loss function includes a loss function corresponding to an internal invariant feature module as a regularization term; the internal invariant feature module is obtained by teaching a student network through knowledge distillation by an independently trained teacher network; the teacher network is obtained by independent training with the goal of minimizing the loss function corresponding to the bearing fault diagnosis result obtained by the teacher network from the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal and the true label of the bearing fault; the loss function corresponding to the internal invariant feature module includes the loss function corresponding to the bearing fault diagnosis result obtained by the student network from the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal and the true label of the bearing fault;

[0009] The methods of obtaining the internal invariant feature module by teaching the student network through knowledge distillation by the independently trained teacher network include:

[0010] A loss term of the loss function obtained by adding a loss function corresponding to the internal invariant feature module in the total loss function as a regularization term according to the number of source domains corresponding to the multi-source domain bearing vibration signal and the difference between the features extracted from the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal input to the teacher network and the features extracted from the multi-source domain bearing vibration signal input to the student network is added to the loss function corresponding to the internal invariant feature module in the total loss function.

[0011] Furthermore, the mutually invariant feature module includes a label classification submodule and a domain identification submodule; the loss function corresponding to the mutually invariant feature module includes a loss function obtained by superimposing a loss term corresponding to the sum of local maximum mean differences between each pair of source domains as a regularization term on the bearing fault diagnosis result obtained by using the label classification submodule according to the remaining preliminary feature part input to the mutually invariant feature module and the true label of the bearing fault, and subtracting a loss term corresponding to the sum of adversarial loss functions between each pair of source domains as a regularization term;

[0012] The adversarial loss function is determined according to the number of samples corresponding to the input multi-source domain bearing vibration signals, the results of identification using the domain identification submodule according to the features extracted from the bearing vibration signals of two different source domains, and the domain labels of the samples corresponding to the multi-source domain bearing vibration signals;

[0013] The total loss function also includes the loss function corresponding to the mutually invariant feature modules as a regularization term.

[0014] Furthermore, the total loss function also includes a loss term corresponding to the difference between internal and mutual invariant features as a regularization term.

[0015] Furthermore, the loss term corresponding to the bearing fault diagnosis result obtained by the label classification submodule using the remaining preliminary feature part input to the mutually invariant feature module and the true label of the bearing fault is obtained according to the cross entropy loss function of the bearing fault diagnosis result obtained by the label classification submodule using the remaining preliminary feature part input to the mutually invariant feature module and the true label of the bearing fault.

[0016] Furthermore, the local maximum mean difference between the two source domains is:

[0017]

[0018] in, is the local maximum mean difference; and x i and x j The weight of the bearing fault category c; C is the total number of bearing fault categories in each source domain; and are two different source domains; x i and x j They belong to and The samples corresponding to the bearing vibration signal; is the -2 norm of the reproducible Hilbert eigenspace; φ(x i ),φ(x j ) are x i、x j Mapping function to a reproducible Hilbert eigenspace.

[0019] Further, a method of adding a loss term of a loss function obtained according to the number of source domains corresponding to the multi-source domain bearing vibration signal and the difference between the features extracted by the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal input to the teacher network and the features extracted by the multi-source domain bearing vibration signal input to the student network as a regularization term to the loss function corresponding to the internal invariant feature module in the total loss function includes:

[0020]

[0021] in, It is the loss function corresponding to the internal invariant feature module in the total loss function; A loss term of a loss function obtained according to the number of source domains corresponding to the multi-source domain bearing vibration signal and the difference between the features extracted by the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal input to the teacher network and the features extracted by the multi-source domain bearing vibration signal input to the student network; The loss function corresponding to the bearing fault diagnosis result obtained by inputting the features extracted from the multi-source domain bearing vibration signal into the student network and the true label of the bearing fault is: The bearing fault diagnosis result is obtained by inputting the features extracted from the multi-source domain bearing vibration signal into the student network; y is the true label of the bearing fault; λ 0 is the hyperparameter corresponding to the regularization term; N is the size of the feature of the part corresponding to the product of the feature separation coefficient and the feature; is the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal corresponding to the i-th sample; is the multi-source domain bearing vibration signal corresponding to the i-th sample; Represents the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal input to the trained teacher network The extracted features, Indicates that the multi-source domain bearing vibration signal is input to the student network The extracted features.

[0022] Furthermore, according to the number of samples corresponding to the input multi-source domain bearing vibration signals, the results of identification using the domain identification submodule based on the features extracted from the bearing vibration signals of two different source domains, and the domain labels of the samples corresponding to the multi-source domain bearing vibration signals, the adversarial loss function between the two source domains is determined as follows:

[0023]

[0024] in, is the adversarial loss function; M represents the number of samples corresponding to the bearing vibration signal in the multi-source domain, G f (x im θ f ) represents the preliminary features according to the source domain The mth sample x corresponding to the bearing vibration signal im The extracted features, G f (x jm θ f ) indicates the source in the preliminary feature The mth sample x corresponding to the bearing vibration signal jm The extracted features, θ f ,θ d are the differentiable parameters in the feature extraction module and the domain identification submodule of the mutually invariant feature module of the model, G d represents the result of identification using the domain identification submodule, d m represents the domain label of the mth sample.

[0025] Furthermore, the loss term corresponding to the difference between the internal and mutually invariant features as the regularization term with negative weights included in the total loss function is the loss term corresponding to the cosine distance loss function between the internal and mutually invariant features

[0026] The above technical solution of the present invention provides a new bearing fault feature extraction model based on domain generalization, and its beneficial effects include:

[0027] By setting a feature separation coefficient that changes dynamically with the ratio of the value of the loss function corresponding to the internal invariant feature module to the value of the loss function corresponding to the mutual invariant feature module, the number of preliminary features used to calculate the internal invariant feature and the number of preliminary features used to calculate the mutual invariant feature are dynamically divided; since the part corresponding to the product of the feature separation coefficient and the feature is used to calculate the internal invariant feature, when the value of the loss function corresponding to the internal invariant feature module is small (indicating that the impact on the domain generalization process is small), the number of preliminary features allocated to the internal invariant feature module for calculating the internal invariant feature is also small, and accordingly, since the remaining preliminary features are all used to calculate the mutual invariant features, the number of preliminary features allocated to the mutual invariant feature module for calculating the mutual invariant features is more; when the value of the loss function corresponding to the mutual invariant feature module is small, the opposite is true; it can be seen that through the feature separation coefficient, more preliminary features can be allocated to the calculation of the invariant features that have a greater impact on the domain generalization process, thereby enabling the bearing fault feature extraction model to dynamically adjust the feature sizes used to calculate the internal invariance and mutual invariance, respectively, so as to ensure that the features extracted initially are more fully utilized and maximize the impact of the features on the domain generalization process.

[0028] The present invention also provides a bearing fault diagnosis method based on domain generalization, which determines the bearing fault diagnosis result by inputting the bearing fault feature into the classification module corresponding to the bearing fault; the bearing fault feature is obtained by inputting the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal into the bearing fault feature extraction model;

[0029] The bearing fault feature extraction model adopts the above-mentioned bearing fault feature extraction model.

[0030] The bearing fault diagnosis method based on domain generalization of the present invention can achieve the same beneficial effects as the technical solution of the bearing fault feature extraction model mentioned above. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flowchart of a training method of a bearing fault feature extraction model based on domain generalization in an embodiment of the bearing fault feature extraction model based on domain generalization of the present invention;

[0032] Figure 2 It is a structural schematic diagram of a bearing fault feature extraction model based on domain generalization in an embodiment of the bearing fault feature extraction model based on domain generalization of the present invention;

[0033] Figure 3 It is a specific structural schematic diagram of a mutually invariant feature module for calculating mutually invariant features in a bearing fault feature extraction model based on domain generalization in an embodiment of the bearing fault feature extraction model based on domain generalization of the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0035] Implementation example of bearing fault feature extraction model based on domain generalization

[0036] This embodiment provides a technical solution for a bearing fault feature extraction model based on domain generalization. The solution dynamically adjusts the allocation ratio of preliminary features used to calculate internal invariant features and mutual invariant features according to the influence of the internal invariance of features (i.e., internal invariant features) and the mutual invariance of features (i.e., mutual invariant features) on the domain generalization process. That is, when the internal invariance of features has less influence on the domain generalization process than the mutual invariance of features, more features are allocated to calculate the mutual invariant features; when the invariance of mutual features has less influence on the domain generalization process than the internal invariance of features, more features are allocated to calculate the internal invariant features. Thus, the bearing fault feature extraction model can dynamically adjust the feature sizes used to calculate the internal invariance and mutual invariance, respectively, so as to ensure that the preliminarily extracted features are more fully utilized and the maximum influence of the features on the domain generalization process is maximized.

[0037] Reference Figure 1 , the training method of the bearing fault feature extraction model specifically includes:

[0038] 1) Input the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal used for training into the feature extraction module of the bearing fault feature extraction model to extract preliminary features; divide the preliminary features into two parts, calculate the internal invariant features according to the part corresponding to the product of the feature separation coefficient and the preliminary features through the internal invariant feature module in the bearing fault feature extraction model, and calculate the mutual invariant features according to the remaining part through the mutual invariant feature module in the bearing fault feature extraction model; the feature separation coefficient is obtained according to the ratio of the value of the loss function corresponding to the internal invariant feature module to the sum of the values ​​of the loss functions corresponding to the mutual invariant feature module and the internal invariant feature module; for bearing fault diagnosis and bearing fault feature extraction, different source domains correspond to different working conditions, and the bearing vibration signals of different source domains are the bearing vibration signals under the corresponding working conditions; each source domain corresponds to the same number of fault types.

[0039] The specific method for extracting the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal is as follows:

[0040]

[0041] in, The sample corresponding to the bearing vibration signal The phase component information in the Fourier spectrum; j represents an imaginary number, and N represents the source domain The total number of for The sample instance in The samples corresponding to the bearing vibration signal, t=0,1,…,n-1, stands for Fast Fourier Transform.

[0042] 2) The parameters of the bearing fault feature extraction model are updated with the goal of minimizing the total loss function; steps 1)-2) are iterated until the iteration stop condition is met; the total loss function includes the loss term corresponding to the bearing fault diagnosis result determined by the bearing fault features obtained by the internal invariant features and the mutual invariant features and the true label of the bearing fault The iteration stop condition here is usually set to that the number of iterations reaches a set number threshold or the value of the total loss function is less than a set upper limit threshold, which is a conventional setting and will not be repeated in this embodiment.

[0043] Therefore, by setting a feature separation coefficient that changes dynamically with the ratio of the value of the loss function corresponding to the internal invariant feature module to the value of the loss function corresponding to the mutual invariant feature module, the number of preliminary features used to calculate the internal invariant feature and the number of preliminary features used to calculate the mutual invariant feature are dynamically divided; since the part corresponding to the product of the feature separation coefficient and the preliminary feature is used to calculate the internal invariant feature, when the value of the loss function corresponding to the internal invariant feature module is small (indicating that the impact on the domain generalization process is small), the number of preliminary features allocated to the internal invariant feature module for calculating the internal invariant feature is also small, and accordingly, since the remaining preliminary features are all used to calculate the mutual invariant features, the number of preliminary features allocated to the mutual invariant feature module for calculating the mutual invariant features is more; when the value of the loss function corresponding to the mutual invariant feature module is small, the opposite is true; it can be seen that through the feature separation coefficient, more preliminary features can be allocated to the calculation of the invariant features that have a greater impact on the domain generalization process, thereby enabling the bearing fault feature extraction model to dynamically adjust the feature sizes used to calculate the internal invariance and mutual invariance, respectively, so as to ensure that the features extracted initially are more fully utilized and maximize the impact of the features on the domain generalization process.

[0044] In this embodiment, the structure of the bearing fault feature extraction model is as follows: Figure 2As shown; wherein the internal invariant feature module is obtained by teaching the student network through knowledge distillation by the teacher network after independent training; the teacher network is obtained by independent training with the goal of minimizing the loss function corresponding to the true label of the bearing fault obtained by the teacher network from the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal; since the phase component information in the Fourier spectrum of the bearing vibration signal itself contains internal invariant features, the teacher network independently trained by the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal naturally has the ability to extract internal invariant features; and then the ability of the teacher network to extract internal invariant features is taught to the student network through knowledge distillation, and the obtained student network also has a certain ability to extract internal invariant features, so that the student network can be used as an internal invariant feature module to calculate internal invariant features (the features output after the preliminary features of the internal invariant feature module are input are used as internal invariant features). In the teacher network, by minimizing the cross entropy loss function of the teacher network To achieve independent training through the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal, the cross entropy loss function is specifically:

[0045]

[0046] in, is the cross entropy loss function of the teacher network, that is, the loss function corresponding to the bearing fault diagnosis result obtained by the teacher network from the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal and the true label of the bearing fault; y i is the true label of bearing fault; Predict labels for the teacher network, that is, the bearing fault diagnosis results obtained based on the features extracted by the teacher network from the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal.

[0047] The value of the loss function corresponding to the internal invariant feature module is obtained according to the number of source domains corresponding to the multi-source domain bearing vibration signal and the difference between the features extracted by the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal input to the teacher network and the features extracted by the multi-source domain bearing vibration signal input to the student network; the total loss function also includes the loss function corresponding to the internal invariant feature module as a regularization term. The bearing fault feature extraction model of this embodiment takes into account that although the teacher network has the ability to extract internal invariant features in the source domain through learning the phase component information, after analysis, in order to achieve this goal, there are a large number of parameters in the teacher network, which will reduce the network training efficiency and increase the training time, and further require a higher hardware configuration during the training process; thus, by independently training the teacher network that has the ability to extract internal invariant features in the source domain and using the student network taught by the teacher network as the internal invariant feature module, the entire bearing fault feature extraction model is made more lightweight, the differentiable parameters in the bearing fault feature extraction model are reduced, and the training efficiency of the model is improved.

[0048] Specifically, the method of obtaining the value of the loss function corresponding to the internal invariant feature module in the total loss function according to the above parameters includes:

[0049]

[0050] in, It is the loss function corresponding to the internal invariant feature module in the total loss function; A loss term of a loss function obtained according to the number of source domains corresponding to the multi-source domain bearing vibration signal and the difference between the features extracted by the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal input to the teacher network and the features extracted by the multi-source domain bearing vibration signal input to the student network; The loss function corresponding to the bearing fault diagnosis result obtained by inputting the features extracted from the multi-source domain bearing vibration signal into the student network and the true label of the bearing fault is: The bearing fault diagnosis result is obtained by inputting the features extracted from the multi-source domain bearing vibration signal into the student network; y is the true label of the bearing fault; λ 0 is the hyperparameter corresponding to the regularization term; N is the size of the feature of the part corresponding to the product of the feature separation coefficient and the feature; is the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal corresponding to the i-th sample; is the multi-source domain bearing vibration signal corresponding to the i-th sample; Represents the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal input to the trained teacher network The extracted features are obtained by the feature extractor in the teacher network; Indicates that the multi-source domain bearing vibration signal is input to the student network The extracted features are obtained by the feature extractor in the student network.

[0051] Reference Figure 3 In this embodiment, the mutually invariant feature module includes a label classification submodule (label classifier) ​​and a domain identification submodule (domain identifier); the loss function corresponding to the mutually invariant feature module includes the loss term corresponding to the bearing fault diagnosis result obtained by the label classification submodule according to the remaining preliminary feature part (which is directly used as the mutually invariant feature) input to the mutually invariant feature module and the true label of the bearing fault The local maximum mean difference between the source and target domains is superimposed as a regularization term The corresponding loss term, minus the adversarial loss function as a regularization term The loss function is obtained by minimizing the loss term corresponding to the local maximum mean difference, which can effectively align the features of the same category in different fields and consider the fine-grained information of the categories in different fields; maximizing the loss term corresponding to the adversarial loss function can reduce the global distribution difference and learn domain-invariant features; minimizing the loss term corresponding to the bearing fault diagnosis result obtained by the label classification submodule based on the remaining preliminary features input to the mutually invariant feature module and the true label of the bearing fault can ensure that the fault diagnosis (fault classification) result obtained by the mutually invariant feature module is closer to the actual fault condition of the bearing (true label of the bearing fault).

[0052] Therefore, the loss function corresponding to the mutually invariant feature modules is as follows:

[0053]

[0054] Among them, λ and μ are the weight parameters of the local maximum mean difference and adversarial loss function respectively. To ensure the overall training effect, the total loss function also includes the loss function corresponding to the mutually invariant feature module as a regularization term; by minimizing the total loss function including the loss function corresponding to the mutually invariant feature module, when the model converges, a mutually invariant feature module that can capture the mutually invariant features in multiple source domains will be obtained.

[0055] Among them, the local maximum mean difference between the two source domains is:

[0056]

[0057] in, is the local maximum mean difference; and x i and x jThe weight of the bearing fault category c; C is the total number of bearing fault categories in each source domain; and are two different source domains; x i and x j They belong to and The samples corresponding to the bearing vibration signal; is the -2 norm of the reproducible Hilbert eigenspace; φ(x i ),φ(x j ) are x i 、x j Mapping function to reproducible Hilbert eigenspace. i For example, the weight The calculation is as follows:

[0058]

[0059] where y ic is the vector y i The cth entry of .

[0060] Adversarial loss function between two source domains The adversarial loss function is determined according to the number of samples corresponding to the input multi-source domain bearing vibration signals, the results of identification using the domain identification submodule based on the features extracted from the bearing vibration signals of two different source domains, and the domain labels of the samples corresponding to the multi-source domain bearing vibration signals; specifically, the adversarial loss function is determined as:

[0061]

[0062] in, is the adversarial loss function; M represents the number of samples corresponding to the bearing vibration signal in the multi-source domain, G f (x im θ f ) represents the preliminary features according to the source domain The mth sample x corresponding to the bearing vibration signal im The extracted features, G f (x jm θ f ) indicates the source in the preliminary feature The mth sample x corresponding to the bearing vibration signal jm The extracted features, θ f ,θ d are the differentiable parameters in the feature extraction module and the domain identification submodule of the mutually invariant feature module of the model, G d represents the result of identification using the domain identification submodule, d mrepresents the domain label of the mth sample.

[0063] On this basis, the specific calculation formula of the feature separation coefficient δ obtained according to the ratio of the value of the loss function corresponding to the internal invariant feature module to the sum of the values ​​of the loss functions corresponding to the mutual invariant feature modules and the internal invariant feature modules is:

[0064]

[0065] In the above formula, the smaller Indicates that domain-invariant features have better internal invariance and smaller Indicates better mutual invariance. When the internal invariance of samples in the domain is much better than the mutual invariance, δ is close to 1. Correspondingly, when the internal invariance alignment is much worse than the mutual invariance, δ is close to 0. It is worth noting that when the impact of internal invariance on the domain generalization process in domain invariant features is equal to mutual invariance, δ is equal to 0.5, which is the strategy adopted by the existing technology.

[0066] Then δ is used as the weight for dynamic separation of features:

[0067]

[0068] in, is the feature size of the total output of the feature extractor, that is, the size of the preliminary feature, is the feature size used to calculate the internal invariance, that is, the feature size of the part corresponding to the product of the feature separation coefficient and the preliminary feature; correspondingly, the preliminary feature size used to calculate the mutual invariance feature is That is, the feature size of the remaining part. When the preliminary features are divided into two parts, the preliminary features are directly divided into the front and back parts according to the ratio of δ to maintain the coherence and specific order between the features, rather than dividing by randomly selecting a certain number of features.

[0069] In this embodiment, in order to increase the diversity of features and reduce the duplication between internal invariant features and mutual invariant features, the difference between internal invariant features and mutual invariant features must be maximized. and They correspond to the internal invariant features and mutual invariant features of the kth source domain sample respectively. This embodiment maximizes the cosine distance loss function. and The difference between . Cosine distance loss The specific formula is as follows:

[0070]

[0071] Where ∈ is an infinitesimal quantity, the purpose is to avoid the sum of the elements of the vector to be 0. Since the optimization goal of the cosine distance loss function between the internal and mutually invariant features is to maximize, the regularization weight of the loss term corresponding to the cosine distance loss function between the internal and mutually invariant features as a regularization term should be a negative value.

[0072] In summary, in this embodiment, the expression of the total loss function is:

[0073]

[0074] in, is the loss term corresponding to the bearing fault diagnosis result determined according to the bearing fault features obtained by the internal invariant features and the mutual invariant features and the true label of the bearing fault; 1 ,λ 2 and λ 3 They are the loss functions corresponding to the internal invariant feature modules. Loss function corresponding to mutually invariant feature modules and the loss term corresponding to the difference between internal and mutual invariant features The positive weight of .

[0075] In fact, in this embodiment, whether it is the feature extraction module for extracting preliminary features or the feature extractors in the teacher network and the student network, they are all implemented by a feature extractor including a deep residual neural network (which has significant improvements over the deep convolutional neural network in feature extraction). The deep residual neural network is composed of multiple residual blocks, each of which is composed of a convolutional layer, a nonlinear activation function, and a batch normalization layer. The principle of implementing feature extraction is basically the same, and the overall mathematical expression is as follows:

[0076] y=G f (x,w)+b

[0077] Among them, x and y are the input and output of the feature extractor respectively; G f Represents the mapping of the feature extractor, including convolution operation, batch normalization operation and ReLU function, w and b are differentiable parameters in the feature extractor.

[0078] As for the method of updating the parameters in the model, the stochastic gradient descent method is used in each module to update the parameters in the model. The mathematical expression is:

[0079]

[0080] Where θ is the differentiable parameter in the model and η is the learning rate.

[0081] Embodiment of bearing fault diagnosis method based on domain generalization

[0082] This embodiment provides a technical solution for a bearing fault diagnosis method based on domain generalization. The method determines a bearing fault diagnosis result by inputting bearing fault features into a classification module corresponding to the bearing fault; the bearing fault features are obtained by inputting phase component information in the Fourier spectrum of a multi-source domain bearing vibration signal into a bearing fault feature extraction model;

[0083] The bearing fault feature extraction model adopts the bearing fault feature extraction model in the above bearing fault feature extraction model embodiment. Since the training method and functional principle of the bearing fault feature extraction model have been described in detail in the above bearing fault feature extraction model embodiment, they will not be repeated here.

[0084] The present invention has the following characteristics:

[0085] 1) By setting a feature separation coefficient that changes dynamically with the ratio of the value of the loss function corresponding to the internal invariant feature module to the value of the loss function corresponding to the mutual invariant feature module, the number of preliminary features used to calculate the internal invariant feature and the number of preliminary features used to calculate the mutual invariant feature are dynamically divided; since the part corresponding to the product of the feature separation coefficient and the preliminary feature is used to calculate the internal invariant feature, when the value of the loss function corresponding to the internal invariant feature module is small (indicating that the impact on the domain generalization process is small), the number of preliminary features allocated to the internal invariant feature module for calculating the internal invariant feature is also small, and accordingly, since the remaining preliminary features are all used to calculate the mutual invariant features, the number of preliminary features allocated to the mutual invariant feature module for calculating the mutual invariant features is more; when the value of the loss function corresponding to the mutual invariant feature module is small, the opposite is true; it can be seen that through the feature separation coefficient, more preliminary features can be allocated to the calculation of the invariant features that have a greater impact on the domain generalization process, thereby enabling the bearing fault feature extraction model to dynamically adjust the feature sizes used to calculate the internal invariance and mutual invariance, respectively, so as to ensure that the features extracted preliminarily are more fully utilized and maximize the impact of the features on the domain generalization process.

[0086] 2) Considering that although the teacher network has the ability to extract internal invariant features in the source domain through learning the phase component information, after analysis, in order to achieve this goal, there are a large number of parameters in the teacher network, which will reduce the network training efficiency and increase the training time, and further require a higher hardware configuration during the training process; therefore, by independently training the teacher network that has the ability to extract internal invariant features in the source domain and using the student network taught by the teacher network as the internal invariant feature module, the entire bearing fault feature extraction model is made more lightweight, the differentiable parameters in the bearing fault feature extraction model are reduced, and the training efficiency of the model is improved.

[0087] It should be understood that the above specific embodiments of the present invention are only used to illustrate or explain the principles of the present invention, but do not constitute a limitation to the present invention.

Claims

1. A bearing fault feature extraction model based on domain generalization, characterized in that: The training method of the model includes: 1) Inputting the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal used for training into the feature extraction module of the model to extract preliminary features; dividing the preliminary features into two parts, calculating the internal invariant features according to the part corresponding to the product of the feature separation coefficient and the features through the internal invariant feature module in the model, and calculating the mutual invariant features according to the remaining part through the mutual invariant feature module in the model; the feature separation coefficient is obtained according to the ratio of the value of the loss function corresponding to the internal invariant feature module to the sum of the values ​​of the loss functions corresponding to the mutual invariant feature module and the internal invariant feature module; 2) updating the parameters of the model with the goal of minimizing the total loss function; iterating steps 1)-2) until the iteration stop condition is met; the total loss function includes a loss term corresponding to the bearing fault diagnosis result determined according to the bearing fault characteristics obtained by the internal invariant features and the mutual invariant features and the true label of the bearing fault.

2. The bearing fault feature extraction model based on domain generalization according to claim 1 is characterized in that: The total loss function includes a loss function corresponding to an internal invariant feature module as a regularization term; the internal invariant feature module is obtained by teaching a student network through knowledge distillation by an independently trained teacher network; the teacher network is obtained by independent training with the goal of minimizing the loss function corresponding to the true label of the bearing fault obtained by the teacher network from the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal; the loss function corresponding to the internal invariant feature module includes a loss function corresponding to the true label of the bearing fault diagnosis result obtained by the student network from the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal; The methods of obtaining the internal invariant feature module by teaching the student network through knowledge distillation by the independently trained teacher network include: A loss term of the loss function obtained by adding a loss function corresponding to the internal invariant feature module in the total loss function as a regularization term according to the number of source domains corresponding to the multi-source domain bearing vibration signal and the difference between the features extracted from the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal input to the teacher network and the features extracted from the multi-source domain bearing vibration signal input to the student network is added to the loss function corresponding to the internal invariant feature module in the total loss function.

3. The bearing fault feature extraction model based on domain generalization according to claim 2 is characterized in that: The mutually invariant feature module includes a label classification submodule and a domain identification submodule; the loss function corresponding to the mutually invariant feature module includes a loss function obtained by superimposing the loss term corresponding to the sum of the local maximum mean differences between each pair of source domains as a regularization term on the bearing fault diagnosis result obtained by using the label classification submodule according to the remaining preliminary feature part input to the mutually invariant feature module and the true label of the bearing fault, and subtracting the loss term corresponding to the sum of the adversarial loss functions between each pair of source domains as a regularization term; The adversarial loss function is determined according to the number of samples corresponding to the input multi-source domain bearing vibration signals, the results of identification using the domain identification submodule according to the features extracted from the bearing vibration signals of two different source domains, and the domain labels of the samples corresponding to the multi-source domain bearing vibration signals; The total loss function also includes the loss function corresponding to the mutually invariant feature modules as a regularization term.

4. The bearing fault feature extraction model based on domain generalization according to claim 1 is characterized in that: The total loss function also includes a loss term corresponding to the difference between internal and mutual invariant features as a regularization term with negative weights.

5. The bearing fault feature extraction model based on domain generalization according to claim 3 is characterized in that: The loss term corresponding to the bearing fault diagnosis result obtained by the label classification submodule using the remaining preliminary feature part input to the mutually invariant feature module and the true label of the bearing fault is obtained according to the cross entropy loss function of the bearing fault diagnosis result obtained by the label classification submodule using the remaining preliminary feature part input to the mutually invariant feature module and the true label of the bearing fault.

6. The bearing fault feature extraction model based on domain generalization according to claim 3 is characterized in that: The local maximum mean difference between the two source domains is: in, is the local maximum mean difference; and x i and x j The weight of the bearing fault category c; C is the total number of bearing fault categories in each source domain; and are two different source domains; x i and x j They belong to and The samples corresponding to the bearing vibration signal; is the -2 norm of the reproducible Hilbert eigenspace; φ(x i ),φ(x j ) are x i 、x j Mapping function to a reproducible Hilbert eigenspace.

7. The bearing fault feature extraction model based on domain generalization according to claim 2 is characterized in that: The method of adding a loss term of the loss function obtained by the difference between the number of source domains corresponding to the multi-source domain bearing vibration signal and the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal input to the teacher network and the feature extracted from the multi-source domain bearing vibration signal input to the student network as a regularization term to the loss function corresponding to the internal invariant feature module in the total loss function includes: in, It is the loss function corresponding to the internal invariant feature module in the total loss function; A loss term of a loss function obtained according to the number of source domains corresponding to the multi-source domain bearing vibration signal and the difference between the features extracted by the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal input to the teacher network and the features extracted by the multi-source domain bearing vibration signal input to the student network; The loss function corresponding to the bearing fault diagnosis result obtained by inputting the features extracted from the multi-source domain bearing vibration signal into the student network and the true label of the bearing fault is: The bearing fault diagnosis result obtained by inputting the features extracted from the multi-source domain bearing vibration signal into the student network; y is the true label of the bearing fault; λ0 is the hyperparameter corresponding to the regularization term; N is the size of the feature corresponding to the product of the feature separation coefficient and the feature; is the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal corresponding to the i-th sample; is the multi-source domain bearing vibration signal corresponding to the i-th sample; Represents the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal input to the trained teacher network The extracted features, Indicates that the multi-source domain bearing vibration signal is input to the student network The extracted features.

8. The bearing fault feature extraction model based on domain generalization according to claim 3 is characterized in that: According to the number of samples corresponding to the input multi-source domain bearing vibration signals, the results of identification using the domain identification submodule based on the features extracted from the bearing vibration signals of two different source domains, and the domain labels of the samples corresponding to the multi-source domain bearing vibration signals, the adversarial loss function between the two source domains is determined as follows: in, is the adversarial loss function; M represents the number of samples corresponding to the bearing vibration signal in the multi-source domain, G f (x im θ f ) represents the preliminary features according to the source domain The mth sample x corresponding to the bearing vibration signal im The extracted features, G f (x jm θ f ) indicates the source in the preliminary feature The mth sample x corresponding to the bearing vibration signal jm The extracted features, θ f ,θ d are the differentiable parameters in the feature extraction module and the domain identification submodule of the mutually invariant feature module of the model, G d represents the result of identification using the domain identification submodule, d m represents the domain label of the mth sample.

9. The bearing fault feature extraction model based on domain generalization according to claim 4 is characterized in that: The loss term corresponding to the difference between the internal and mutual invariant features included in the total loss function as a regularization term with a negative weight is the loss term corresponding to the cosine distance loss function between the internal and mutual invariant features.

10. A bearing fault diagnosis method based on domain generalization, characterized in that: The bearing fault diagnosis result is determined by inputting the bearing fault feature into the classification module corresponding to the bearing fault; the bearing fault feature is obtained by inputting the phase component information in the Fourier spectrum of the multi-source domain bearing vibration signal into the bearing fault feature extraction model; The bearing fault feature extraction model adopts the bearing fault feature extraction model of any one of claims 1-9.

Images