Multi-source domain rolling bearing fault diagnosis method and system based on diffusion model

Abstract

The invention discloses a multi-source domain rolling bearing fault diagnosis method and system based on a diffusion model. The diagnosis method comprises the steps that rolling bearing vibration signals under different rotating speed working conditions are collected; in the forward process of the diffusion model, converting a source domain label into a one-dimensional vector, and adding noise to destroy original distribution to obtain yT approximate to Gaussian distribution; replacing yT with a Gaussian distribution variable # imgabs0 # in a reverse process, learning noise distribution characteristics added in forward distribution through a noise prediction network, denoising # imgabs1 # step by step to obtain # imgabs2 # in the same distribution as y0, inputting source domain data into the noise prediction network to extract data characteristics, inputting target domain data into the noise prediction network, and outputting the target domain data into the noise prediction network; global noise priori and local noise priori corresponding to the target domain data are obtained; and inputting source domain data and target domain data into a noise prediction network, mapping source domain data features to a target domain by using a multi-source domain adaptive algorithm, learning noise distribution features related to the target domain data, and gradually denoising # imgabs3 # to obtain a fault type of the target domain data.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and in particular to a multi-source domain rolling bearing fault diagnosis method and system based on a diffusion model. Background Art

[0002] Rolling bearings are one of the most critical components of rotating machinery. They operate in harsh environments and need to work in changing conditions for a long time. Collecting fault data under different conditions not only requires a large economic investment, but also makes it difficult to obtain accurate label information. This limits deep learning diagnostic methods based on large amounts of training data and label information.

[0003] Domain adaptation (DA) is an important research direction in the field of transfer learning. The premise is that the source domain and the target domain have similar feature distributions. The source domain data is used to train the deep learning model, and feature transfer is performed between the source domain and the target domain, so as to achieve relevant tasks in the target domain. Domain adaptation algorithms include single-source domain adaptation algorithms and multi-source domain adaptation algorithms. The feature distribution of the single source domain and the target domain is quite different, which is prone to overfitting in the process of transfer learning. In contrast, multi-source domain data has richer features and better generalization. The general domain adaptation method is to directly align the source domain features and the target domain features one by one to minimize the feature space difference between the two. It is difficult to effectively align multiple source domains and target domains. In addition, the receptive field of the feature extraction network is fixed, and it is difficult to extract features of different details. When the speed difference is large, it is difficult to extract effective features, and the generalization is poor.

[0004] The diffusion model is a generative model based on the Markov chain. It destroys the distribution of the original data by adding Gaussian noise and restores the original distribution of the data in the reverse process to achieve related tasks. Compared with the GAN and VAE architectures, the generated data has better generalization and the model is easier to converge. It can fuse features of multiple scales and depths to make the extracted features richer and make up for the problem of monotonous information extraction in the multi-source domain adaptation feature network. However, the diffusion model has the problem of large computational complexity, and redundant sampling in the reverse reasoning process exacerbates this problem.

[0005] In summary, in order to solve the problem of rolling bearing fault diagnosis across working conditions and improve the accuracy and generalization of fault diagnosis, on the one hand, it is necessary to strengthen the feature extraction ability of deep learning. Using the diffusion model as a feature extraction model can effectively solve this problem, but it has the problems of slow calculation speed and sampling redundancy. On the other hand, it is necessary to better align the source domain data features and the target domain features. Choosing a feature alignment algorithm on the joint probability distribution can more effectively migrate the source domain features to the target domain. Summary of the invention

[0006] In response to the problems and needs raised above, this solution proposes a multi-source domain rolling bearing fault diagnosis method and system based on a diffusion model. Due to the adoption of the following technical features, the above technical objectives can be achieved and many other technical effects can be brought about.

[0007] An object of the present invention is to provide a multi-source domain rolling bearing fault diagnosis method based on a diffusion model, comprising the following steps:

[0008] S10, data collection: collecting rolling bearing vibration signals under different speed conditions, taking a certain unlabeled speed condition data as target domain data, and other speed condition fault data as source domain data;

[0009] S20, build a diffusion model framework: including forward process and reverse process. In the forward process of the diffusion model, the source domain label is converted into a one-dimensional vector y0, and the noise prediction network is pre-trained to obtain the global noise prior. and local noise prior Add noise to y0 T times to destroy the original distribution of y0 until a variable y with an approximate Gaussian distribution is obtained. T ; Use Gaussian distributed variables in the reverse process Instead of y T , and use the noise prediction network to learn the noise distribution features added in the forward distribution. Gradually denoise, and finally get the same distribution as y0

[0010] S30, input the source domain data into the noise prediction network to fully extract the data features, input the target domain data into the noise prediction network, use the multi-source domain adaptation algorithm to map the source domain features to the target domain, and obtain the global noise prior corresponding to the target domain data and local noise prior

[0011] S40, inputting the source domain data and the target domain data into the noise prediction network, and using the multi-source domain adaptation algorithm to map the noise prediction network features to the target domain, the noise prediction network fuses the fault data features after feature migration, and learns the noise features related to the target domain data;

[0012] S50, bringing the learned target domain-related Gaussian noise and the noise prior after migration into the reverse process in step S20, and performing the Gaussian distribution of random variables De-noising step by step, predicting the fault data type of the target domain.

[0013] In addition, the multi-source domain rolling bearing fault diagnosis method based on the diffusion model according to the present invention may also have the following technical features:

[0014] In one example of the present invention, the noise prediction network includes a feature extraction network, a conditional guidance network and a U-Net network, the feature extraction network is configured to extract data features based on a Ghost convolutional network; the conditional guidance network is configured to extract noise priors; and the U-Net network is configured to fuse the data features and the noise priors.

[0015] In one example of the present invention, the feature extraction network includes:

[0016] A first convolutional layer, a first BN layer, a first ReLu layer, a Ghost layer and a fully connected layer are sequentially connected in series, wherein the Ghost layer includes a plurality of layers and is sequentially connected in series between the first ReLu layer and the fully connected layer, the first convolutional layer is configured to perform a convolution operation on the input data, the first BN layer and the first ReLu are configured to accelerate the convergence speed of the model and increase the generalization of the model; the plurality of Ghost layers are configured to reduce the parameters of the model while fully extracting the data features; the fully connected layer is configured to adjust the dimension of the output features to 1024.

[0017] In one example of the present invention, the conditional guidance network includes:

[0018] ResNet18 network and CNN+Transform multi-feature fusion network. The ResNet18 network is configured to extract the global features of the fault signal and output the noise prior. The CNN+Transform multi-feature fusion network is configured to learn the fault data features of different scales, concatenate the results of convolution operations of different scales at the channel scale, and then input them into the Transformer to learn the relationship between features of different scales, and finally output the noise prior.

[0019] In one example of the present invention, the CNN+Transform multi-feature fusion network includes: a convolutional unit layer, a feature fusion layer and a Transformer module connected in series in sequence, the convolutional unit layer includes a second convolutional layer, a third convolutional layer and a fourth convolutional layer connected in parallel in sequence, the third convolutional layer and the fourth convolutional layer are respectively connected in series with the fifth convolutional layer and the sixth convolutional layer, the fifth convolutional layer, the sixth convolutional layer and the second convolutional layer are connected in parallel and connected to the feature fusion layer; wherein the convolutional unit layer is configured to learn fault data features of different scales, the feature splicing layer is configured to splice the learned features of different scales to obtain the feature size required by the Transformer; the Transformer module is configured to learn the correlation and dependency between different features to make the learned fault data features more complete.

[0020] In one example of the present invention, the conditional guidance network uses weighted cross entropy loss and multi-source domain adaptation loss as the loss function of the conditional guidance network, and its expression is as follows:

[0021] L gl =L CE +γL MIgl .

[0022] in, is the cross entropy loss, is a multi-source domain adaptation loss function; the former is used to learn the classification loss in the source domain, and the latter is used to learn the joint probability distribution difference between the feature distributions of multiple source domains and the target domain; γ is used to balance the two types of losses. It is initially 0. After learning the classification features of the source domain data, it increases with the increase in the number of training times, and the maximum value does not exceed 1.

[0023] In one example of the present invention, the U-Net network uses weighted noise difference loss and multi-source domain adaptation loss as the loss function of the U-Net network, wherein the noise difference loss includes numerical difference and distribution difference after fusion of different noise priors, and its expression is as follows:

[0024]

[0025] L gl =L diff +γ1L MI .

[0026] In the formula, is the loss function between the noise predicted by the U-Net network and the real noise, which includes the mean square difference And the distribution difference after fusing different noise priors and λ balances the two types of losses and takes a fixed value of 0.5; It is the difference loss function of the feature distribution of multiple source domains and target domain data. γ1 balances the two types of losses. It is initially 0. After learning the features of the source domain data, it gradually increases, and the maximum value does not exceed 1.

[0027] In one example of the present invention, in step S20,

[0028] The expression of the forward noise addition process of the diffusion model is as follows:

[0029]

[0030] Where ∈ is random noise, is a parameter related to the number of noise additions, y0 is the one-dimensional vector into which the label is transformed, and y t is the variable obtained by adding noise t times;

[0031] The expression of the denoising process of the inverse process of the diffusion model is as follows:

[0032]

[0033] Among them, ρ(x) is the feature vector learned by the feature extraction network, y t is the variable obtained by adding noise to the prediction t times, p θ is the distribution characteristic of Gaussian noise.

[0034] In an example of the present invention, in step S30 and step S40, the multi-source domain adaptation algorithm is configured to perform feature alignment on multiple source domain and target domain data on a joint probability distribution, and transform the problem of solving feature alignment into the problem of finding the maximum value of a convex function. The specific derivation process is as follows:

[0035]

[0036]

[0037]

[0038]

[0039]

[0040] Where MI(h(X), Y; L) is the multi-source domain loss function, h(x), y, l, p and m are the output features, the original labels of the data, the number of source domains, the feature probability distribution, and the number of all data, respectively. Formula (2) is substituted into formula (1), and the similar features contained in each source domain and target domain are approximately equal. The sample mean is used to replace the convolution to simplify the formula (3); r(h(x), y, l; θ) is a linear function about θ, where k(h(x), h(x i )) is the Gaussian kernel function, δ(y,y i ) is the indicator kernel, when y=y i 1 when δ(l,l i ) is also the indicator core, when l = l i 1 when , and 0 otherwise; Substitute formula (4) into formula (3) to obtain the final loss function, where the two sub-formulas in formula (5) are about Convex function of .

[0041] Another object of the present invention is to provide a diagnosis system for the multi-source domain rolling bearing fault diagnosis method based on the diffusion model as described above, comprising the following steps:

[0042] A data acquisition module is configured to collect rolling bearing vibration signals under different speed conditions, using a certain unlabeled speed condition data as target domain data and other speed condition fault data as source domain data;

[0043] Construct a model module, configured to convert the source domain label into a one-dimensional vector y0 in the forward process of the diffusion model, pre-train the noise prediction network, and obtain the global noise prior and local noise prior Add noise to y0 T times to destroy the original distribution of y0 until a variable y with an approximate Gaussian distribution is obtained. T ; Use Gaussian distributed variables in the reverse process Instead of y T , and use the noise prediction network to learn the noise distribution features added in the forward distribution. Gradually denoise, and finally get the same distribution as y0

[0044] The pre-training module is configured to input the source domain data into the noise prediction network to fully extract the data features, input the target domain data into the noise prediction network, and use the multi-source domain adaptation algorithm to map the source domain features to the target domain to obtain the global noise prior corresponding to the target domain data. and local noise prior

[0045] The noise migration module is configured to input the source domain data and the target domain data into the noise prediction network, and use the multi-source domain adaptation algorithm to map the noise prediction network features to the target domain. The noise prediction network fuses the fault data features after feature migration and learns the noise features related to the target domain data.

[0046] The fault prediction module is configured to bring the learned target domain-related Gaussian noise and the noise prior after migration into the reverse process of the model building module, and perform the Gaussian distribution of random variables De-noising step by step, predicting the fault data type of the target domain.

[0047] The beneficial effects of the present invention are: improving the noise prediction network in the diffusion model, designing a feature extraction network based on Ghost convolution and a conditional guidance network of multi-scale feature fusion, the feature extraction network based on Ghost convolution reduces the computational complexity of the diffusion model under the premise of fully learning the fault data features, the conditional guidance network of multi-scale feature fusion can learn fault data features of different scales, extract richer features, better guide the diffusion process, achieve convergence within fewer noise addition times, and further reduce the computational complexity of the diffusion model. The diffusion model framework and the multi-source domain adaptation algorithm are effectively combined to effectively extract fault data features and achieve effective alignment of the source domain and the target domain.

[0048] The best embodiment for carrying out the present invention will be described in more detail below with reference to the accompanying drawings so that the features and advantages of the present invention can be easily understood. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0050] Figure 1 is a flow chart of a multi-source domain rolling bearing fault diagnosis method based on a diffusion model according to an embodiment of the present invention;

[0051] Figure 2 is an overall schematic diagram of a diffusion model according to an embodiment of the present invention;

[0052] Figure 3 Schematic diagram of Ghost convolution according to an embodiment of the present invention;

[0053] Figure 4 A schematic diagram of a feature extraction network according to an embodiment of the present invention;

[0054] Figure 5 A schematic diagram of multi-scale feature fusion according to an embodiment of the present invention;

[0055] Figure 6 is a TSNE diagram of CWRU data features according to an embodiment of the present invention;

[0056] Figure 7 A TSNE diagram of data features collected in a laboratory according to an embodiment of the present invention;

[0057] Figure 8 A diagram of a denoising process of a diffusion model according to an embodiment of the present invention;

[0058] Fig. 9 FIG. 4 is a fault diagnosis confusion matrix diagram of a data set collected in a laboratory according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the technical solution of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described in conjunction with the drawings of specific embodiments of the present invention. The same figure marks in the drawings represent the same parts. It should be noted that the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] Unless otherwise defined, the technical terms or scientific terms used herein shall have the usual meanings understood by persons of ordinary skill in the field to which the present invention belongs. The words "first", "second" and similar words used in the patent application specification and claims of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not necessarily indicate a quantity limitation. Words such as "include" or "comprise" mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0061] According to the first aspect of the present invention, a multi-source domain rolling bearing fault diagnosis method based on a diffusion model comprises the following steps:

[0062] S10, data collection: collecting rolling bearing vibration signals under different speed conditions, taking a certain unlabeled speed condition data as target domain data, and other speed condition fault data as source domain data;

[0063] S20, build a diffusion model framework: including forward process and reverse process. In the forward process of the diffusion model, the source domain label is converted into a one-dimensional vector y0, and the noise prediction network is pre-trained to obtain the global noise prior. and local noise prior Add noise to y0 T times to destroy the original distribution of y0 until a variable y with an approximate Gaussian distribution is obtained. T ; Use Gaussian distributed variables in the reverse process Instead of y T , and use the noise prediction network to learn the noise distribution features added in the forward distribution. Gradually denoise, and finally get the same distribution as y0

[0064] S30, input the source domain data into the noise prediction network to fully extract the data features, input the target domain data into the noise prediction network, use the multi-source domain adaptation algorithm to map the source domain features to the target domain, and obtain the global noise prior corresponding to the target domain data and local noise prior

[0065] S40, inputting the source domain data and the target domain data into the noise prediction network, and using the multi-source domain adaptation algorithm to map the noise prediction network features to the target domain, the noise prediction network fuses the fault data features after feature migration, and learns the noise features related to the target domain data;

[0066] S50, bringing the learned target domain-related Gaussian noise and the noise prior after migration into the reverse process in step S20, and performing the Gaussian distribution of random variables De-noising step by step, predicting the fault data type of the target domain.

[0067] This fault diagnosis method improves the noise prediction network in the diffusion model, designs a feature extraction network based on Ghost convolution and a conditional guidance network of multi-scale feature fusion. The feature extraction network based on Ghost convolution reduces the computational complexity of the diffusion model while ensuring that the fault data features are fully learned. The conditional guidance network of multi-scale feature fusion can learn fault data features of different scales, extract richer features, better guide the diffusion process, achieve convergence in fewer noise addition times, and further reduce the computational complexity of the diffusion model. The diffusion model framework and the multi-source domain adaptation algorithm are effectively combined to effectively extract fault data features and achieve effective alignment of the source domain and the target domain.

[0068] In one example of the present invention, the noise prediction network includes a feature extraction network, a conditional guidance network and a U-Net network, the feature extraction network is configured to extract data features based on a Ghost convolutional network; the conditional guidance network is configured to extract noise priors; and the U-Net network is configured to fuse the data features and the noise priors.

[0069] It should be noted that the U-Net network is a four-layer convolutional network, including a two-layer encoder and a two-layer decoder structure. After the first layer of convolution passes through the BN layer and the Relu activation function, the features of the network are extracted by fusion of features through arithmetic operations, and the features of the network are guided by the fusion of conditions through splicing operations.

[0070] In one example of the present invention, the feature extraction network includes:

[0071] A first convolutional layer, a first BN layer, a first ReLu layer, a Ghost layer and a fully connected layer are sequentially connected in series, wherein the Ghost layer includes a plurality of layers and is sequentially connected in series between the first ReLu layer and the fully connected layer, the first convolutional layer is configured to perform a convolution operation on the input data, the first BN layer and the first ReLu are configured to accelerate the convergence speed of the model and increase the generalization of the model; the plurality of Ghost layers are configured to reduce the parameters of the model while fully extracting the data features; the fully connected layer is configured to adjust the dimension of the output features to 1024.

[0072] In order to fully extract the features of the data, the feature extraction network first uses a 5*5 convolution kernel with 32 channels to perform preliminary feature extraction on the input data, followed by a BN layer and ReLu activation function to speed up the convergence of the model and increase the generalization of the model; three Ghost convolutions are followed by the maximum pooling layer to fully extract the data features while reducing the parameters of the model; finally, a fully connected layer is used to adjust the dimension of the output features to 1024.

[0073] In one example of the present invention, the Ghost convolution structure in the feature extraction network first uses a 1*1 convolution kernel to perform a convolution operation on the input data to map the data to the feature domain, then performs a 3*3 convolution kernel operation on the obtained features to obtain more feature maps, and finally concatenates the results of the two convolution operations on the channel, normalizes the obtained features, and uses the ReLu function as the activation function to obtain the final result.

[0074] In one example of the present invention, the conditional guidance network includes:

[0075] ResNet18 network and CNN+Transform multi-feature fusion network. The ResNet18 network is configured to extract the global features of the fault signal and output the noise prior. The CNN+Transform multi-feature fusion network is configured to learn the fault data features of different scales, concatenate the results of convolution operations of different scales at the channel scale, and then input them into the Transformer to learn the relationship between features of different scales, and finally output the noise prior.

[0076] In one example of the present invention, the CNN+Transform multi-feature fusion network includes: a convolutional unit layer, a feature fusion layer and a Transformer module connected in series in sequence, the convolutional unit layer includes a second convolutional layer, a third convolutional layer and a fourth convolutional layer connected in parallel in sequence, the third convolutional layer and the fourth convolutional layer are respectively connected in series with the fifth convolutional layer and the sixth convolutional layer, the fifth convolutional layer, the sixth convolutional layer and the second convolutional layer are connected in parallel and connected to the feature fusion layer; wherein the convolutional unit layer is configured to learn fault data features of different scales, the feature splicing layer is configured to splice the learned features of different scales to obtain the feature size required by the Transformer; the Transformer module is configured to learn the correlation and dependency between different features to make the learned fault data features more complete.

[0077] Among them, the fifth convolution layer, the sixth convolution layer and the second convolution layer are connected in parallel and connected to the feature fusion layer; among them, the second convolution layer is connected in series to the second BN layer and the second ReLu layer, the third convolution layer is connected in series to the third BN layer and the third ReLu layer, the fourth convolution layer is connected in series to the fourth BN layer and the fourth ReLu layer, the fifth convolution layer is connected in series to the fifth BN layer and the fifth ReLu layer, and the sixth convolution layer is connected in series to the sixth BN layer and the sixth ReLu layer.

[0078] Specifically, the conditional guidance network includes two branches: the ResNet18 network and the CNN+Transform multi-feature fusion network. The ResNet18 network extracts the global features of the fault signal and outputs the noise prior. The multi-feature fusion network of CNN+Transform enhances the feature extraction capability of fault data by learning the features of fault data at different scales. The network includes three branches. The first branch passes through the 1*1 convolution layer with 32 channels, the BN layer and the Relu module. The second branch first passes through the 3*3 convolution operation with 8 channels, the BN layer and the Relu module, and then passes through the 5*5 convolution operation with 32 channels, the BN layer and the Relu module. The third branch first passes through the 5*5 convolution operation with 8 channels, the BN layer and the Relu module, and then passes through the 3*3 convolution operation with 32 channels, the BN layer and the Relu module. The results of the convolution operations at different scales are spliced ​​at the channel scale and then input into the Transformer to learn the relationship between the features at different scales, and finally output the noise prior.

[0079] In one example of the present invention, in order to fully integrate the characteristics of the fault data of the target domain into the noise prior, the conditional guidance network uses weighted cross entropy loss and multi-source domain adaptation loss as the loss function of the conditional guidance network, and its expression is as follows:

[0080] L gl =L CE +γL MIgl .

[0081] in, is the cross entropy loss, is a multi-source domain adaptation loss function; the former is used to learn the classification loss in the source domain, and the latter is used to learn the joint probability distribution difference between the feature distributions of multiple source domains and the target domain; γ is used to balance the two types of losses. It is initially 0. After learning the classification features of the source domain data, it increases with the increase in the number of training times, and the maximum value does not exceed 1.

[0082] In one example of the present invention, in order to enable the U-Net network to fully learn the distribution characteristics of random Gaussian noise related to the target domain data in the forward noisy process, the U-Net network uses weighted noise difference loss and multi-source domain adaptation loss as the loss function of the U-Net network, wherein the noise difference loss includes numerical difference and distribution difference after fusion of different noise priors, and its expression is as follows:

[0083]

[0084]

[0085] In the formula, is the loss function between the noise predicted by the U-Net network and the real noise, which includes the mean square difference And the distribution difference after fusing different noise priors and λ balances the two types of losses and takes a fixed value of 0.5; It is the difference loss function of the feature distribution of multiple source domains and target domain data. γ1 balances the two types of losses. It is initially 0. After learning the features of the source domain data, it gradually increases, and the maximum value does not exceed 1.

[0086] In one example of the present invention, in step S20,

[0087] The expression of the forward noise addition process of the diffusion model is as follows:

[0088]

[0089] Where ∈ is random noise, is a parameter related to the number of noise additions, y0 is the one-dimensional vector into which the label is transformed, and y t is the variable obtained by adding noise t times;

[0090] The expression of the denoising process of the inverse process of the diffusion model is as follows:

[0091]

[0092] Among them, ρ(x) is the feature vector learned by the feature extraction network, y t is the variable obtained by adding noise to the prediction t times, p θ is the distribution characteristic of Gaussian noise.

[0093] In an example of the present invention, in step S30 and step S40, the multi-source domain adaptation algorithm is configured to perform feature alignment on multiple source domain and target domain data on a joint probability distribution, and transform the problem of solving feature alignment into the problem of finding the maximum value of a convex function. The specific derivation process is as follows:

[0094]

[0095]

[0096]

[0097]

[0098]

[0099] Where MI(h(X), Y; L) is the multi-source domain loss function, h(x), y, l, p and m are the output features, the original labels of the data, the number of source domains, the feature probability distribution, and the number of all data, respectively. Formula (2) is substituted into formula (1), and the similar features contained in each source domain and target domain are approximately equal. The sample mean is used to replace the convolution to simplify the formula (3); r(h(x), y, l; θ) is a linear function about θ, where k(h(x), h(x i )) is the Gaussian kernel function, δ(y,y i ) is the indicator kernel, when y=y i 1 when δ(l,l i ) is also the indicator core, when l = l i 1 when , and 0 otherwise; Substitute formula (4) into formula (3) to obtain the final loss function, where the two sub-formulas in formula (5) are about Convex function of .

[0100] In one example of the present invention, the U-Net network is a four-layer convolution, including a two-layer encoder plus a two-layer decoder structure. After the first layer of convolution passes through the BN layer and the Relu activation function, the features of the network are extracted by fusion of features through arithmetic operations, and the features of the network are guided by fusion of conditions through splicing operations.

[0101] According to the second aspect of the present invention, a diagnostic system for the multi-source domain rolling bearing fault diagnosis method based on the diffusion model as described above comprises the following steps:

[0102] A data acquisition module is configured to collect rolling bearing vibration signals under different speed conditions, using a certain unlabeled speed condition data as target domain data and other speed condition fault data as source domain data;

[0103] Construct a model module, configured to convert the source domain label into a one-dimensional vector y0 in the forward process of the diffusion model, pre-train the noise prediction network, and obtain the global noise prior and local noise prior Add noise to y0 T times to destroy the original distribution of y0 until a variable y with an approximate Gaussian distribution is obtained.T ; Use Gaussian distributed variables in the reverse process Instead of y T , and use the noise prediction network to learn the noise distribution features added in the forward distribution. Gradually denoise, and finally get the same distribution as y0

[0104] The pre-training module is configured to input the source domain data into the noise prediction network to fully extract the data features, input the target domain data into the noise prediction network, and use the multi-source domain adaptation algorithm to map the source domain features to the target domain to obtain the global noise prior corresponding to the target domain data. and local noise prior

[0105] The noise migration module is configured to input the source domain data and the target domain data into the noise prediction network, and use the multi-source domain adaptation algorithm to map the noise prediction network features to the target domain. The noise prediction network fuses the fault data features after feature migration and learns the noise features related to the target domain data.

[0106] The fault prediction module is configured to bring the learned target domain-related Gaussian noise and the noise prior after migration into the reverse process of the model building module, and perform the Gaussian distribution of random variables De-noising step by step, predicting the fault data type of the target domain.

[0107] The fault diagnosis system improves the noise prediction network in the diffusion model, designs a feature extraction network based on Ghost convolution and a conditional guidance network of multi-scale feature fusion. The feature extraction network based on Ghost convolution reduces the computational complexity of the diffusion model while ensuring that the fault data features are fully learned. The conditional guidance network of multi-scale feature fusion can learn fault data features of different scales, extract richer features, better guide the diffusion process, achieve convergence in fewer noise addition times, and further reduce the computational complexity of the diffusion model. The diffusion model framework and the multi-source domain adaptation algorithm are effectively combined to effectively extract fault data features and achieve effective alignment of the source domain and the target domain.

[0108] Specific examples

[0109] The process of multi-source domain adaptation method based on diffusion model is as follows Figure 1 The detailed process is as follows:

[0110] S10: Collect the vibration signals of rolling bearings with labels under various speed conditions as source domain data; and take another speed load without labels as the target domain. The training set ratio of the source domain and the target domain is 1:1, and the data set is collected by randomly selecting fault data samples through a sliding window function.

[0111] S20: Construct a diffusion model framework based on the Markov process, including the forward denoising process and the reverse denoising process. In the forward denoising process, the original label is transformed into a one-dimensional vector y0, and the variable y with an approximate Gaussian distribution is obtained by denoising the one-dimensional label vector T times. T , so that it can be used with random Gaussian distribution in the reverse process Instead of y T , completing the reverse process and domain adaptation task of the diffusion model. In the reverse process, the noise distribution features added in the forward distribution are learned through the noise prediction network. Step by step, we can finally get the same distribution as the label variable y0. To achieve fault diagnosis tasks. Design a diffusion model noise prediction network, which consists of three parts: feature extraction network, conditional guidance network and U-Net network. The noise prediction network structure and the noise addition and denoising process are shown in the figure below. Figure 2 As shown in the figure, the noise coefficient is 0.0001-0.002. The number of noise additions needs to be fine-tuned to achieve the best effect as the data set changes. The number of noise additions for the CWRU data set is 80 times, and the number of noise additions for the laboratory-made bearing data set is 100 times. Linear noise addition is selected as the noise addition method. In order to fully extract the characteristics of the source domain data, a multi-layer Ghost convolution layer is set. The specific network structure and parameters are shown in Table 1. The structure is as follows Figure 3 and Figure 4 As shown, Figure 3 This is a schematic diagram of Ghost convolution. Figure 4 This is a schematic diagram of the feature extraction network. The conditional guidance network includes ResNet18 and CNN+Transform multi-feature fusion network. The output of the ResNet18 network is the global noise prior. The CNN+Transform multi-feature fusion network consists of three branches. Different branches are used to learn fault features of different scales. Finally, the features are fused and output to the Transformer to further learn the relationship between the features. The network structure is as follows: Figure 5 As shown, the output local noise prior

[0112] Table 1 shows the network structure and parameters of the diffusion model framework

[0113]

[0114] S30: Input the source domain data into the conditional guidance network to fully extract the source domain data features, then input the target domain fault data, and use the multi-source domain adaptation algorithm to map the source domain features to the target domain to obtain the global noise prior and local noise prior The pre-training was performed for 20 epochs. The first 10 epochs only learned the source domain data features, and the domain adaptation loss function was added from the 11th epoch.

[0115] S40: Input the source domain data into the feature extraction network to fully learn the source domain data features, then input the target domain data, use the multi-source domain adaptation algorithm to map its features to the target domain, and use the U-Net network to fuse the migrated feature extraction network features and the pre-trained conditional guidance network features to predict the Gaussian noise related to the target domain data.

[0116] S50: Bring the Gaussian noise and noise prior related to the target domain into the inverse process of the diffusion model in step 2, and perform random Gaussian noise Step by step, denoising is performed to obtain the same distribution as the target domain fault data type y0

[0117] Experimental verification:

[0118] In order to verify the effectiveness and generalization of the proposed model, the public bearing dataset of Case Western Reserve University and the bearing dataset collected by the laboratory were selected as experimental objects, and compared with some mainstream domain adaptation methods, including ADACL, MSSA, MFSAN, DAN, CORAL, DANN and ACDANN. In order to reduce the impact of accidental errors on fault diagnosis, the average of 5 fault diagnoses was selected as the final result.

[0119] Experimental Dataset 1:

[0120] The data collected by the driver of the CWRU dataset was selected as the object of this experiment, and the data sampling frequency was 12kHz. It includes four working conditions of 1730rpm, 1750rmp, 1772rpm and 1797rpm, represented by letters AD respectively. The fault location of each working condition has three fault types: inner ring, outer ring and rolling element. Each fault type contains three fault diameters of 0.1778cm, 0.3556cm and 0.5334cm respectively. The labels and fault types are shown in Table 2 below. 100 fault types of each type are selected in the training set and the test set. The experimental results are shown in Table 3.

[0121] Table 2 shows the CWRU data labels and fault types.

[0122]

[0123] Table 3 shows the accuracy of data processed by different models

[0124]

[0125]

[0126] Experimental Dataset 2:

[0127] In order to further verify the generalization of the proposed model, a laboratory-made data set with a larger speed difference is selected to verify the effectiveness of the model. The bearing data set collected in the laboratory uses SKF1007ECP cylindrical roller bearings with a sampling frequency of 102.4kHz. It includes three working conditions of 1000rpm, 2000rmp and 3000rpm, represented by the letters EG respectively. It includes four fault types: normal data, inner ring peeling, outer ring crack and ball peeling. The fault and label types are shown in Table 4 below. 150 of each fault type are selected in the training set and test set. The experimental results are shown in Table 5 below.

[0128] Table 4 shows the bearing faults and label types

[0129]

[0130] Table 5 shows the accuracy of data processed by different models

[0131]

[0132] As shown in Table 3, the average accuracy on the Case Western Reserve University bearing dataset is 99.57, and the accuracy of the comparison algorithms are 96.44%, 96.53%, 93.72%, 95.07%, 95.07%, 94.26%, and 95.53% respectively. The accuracy of the proposed method in this paper is 3.08%-5.85% higher than that of other methods. As shown in Table 4, on the bearing dataset collected in the laboratory with a larger speed difference, the accuracy of the proposed method is 96.46%, and the comparison methods are 93.05%, 91.96%, 90.81%, 90.67%, 90.82%, 87.87%, and 87.81% respectively. The proposed method is 3.43%-8.65% higher than the comparison algorithm. Due to the larger speed difference, the accuracy of various methods has decreased, but the multi-source domain adaptation method based on the diffusion model still has the highest accuracy.

[0133] In order to more intuitively demonstrate the feature transfer effect and classification ability of the method proposed in this paper, a t-distributed stochastic neighborhood embedding (T-SNE) visualization analysis is performed on the feature fusion layer in the U-Net network. Figures 6 to 8 They are domain adaptation graph and diffusion model denoising process graph respectively. Figure 6 and Figure 7 As can be seen, multiple source domain features and target domain features can be well aligned, and the classification effect is very obvious. Since the domain adaptation diagram has already demonstrated the classification effect of the model, only the highest speed condition in the two sets of data is selected as the model denoising process for demonstration, as shown in Figure 2. Figure 8As shown in the figure, t represents the predicted result obtained by adding noise t times. It can be seen that as the number of denoising increases, the noise added to the data features is gradually removed, and a clear and visible classification effect is obtained.

[0134] In order to further verify the generalization performance of the proposed method, the confusion matrix diagram is used to analyze the experimental results, such as Fig. 9 As shown in the figure. Due to the high diagnostic accuracy of CWRU, only the data set produced in the laboratory is analyzed. As can be seen from the figure, most of the fault data are diagnosed correctly, and only a small part of the outer ring crack data is diagnosed as other fault types.

[0135] The exemplary implementation scheme of the multi-source domain rolling bearing fault diagnosis method and system based on the diffusion model proposed in the present invention is described in detail above with reference to the preferred embodiments. However, those skilled in the art will understand that, without departing from the concept of the present invention, various modifications and variations can be made to the above-mentioned specific embodiments, and various technical features and structures proposed in the present invention can be combined in various ways without exceeding the protection scope of the present invention. The protection scope of the present invention is determined by the attached claims.

Claims

1. A multi-source domain rolling bearing fault diagnosis method based on diffusion model, characterized in that: The steps include: S10, data collection: collecting rolling bearing vibration signals under different speed conditions, taking a certain unlabeled speed condition data as target domain data, and other speed condition fault data as source domain data; S20, build a diffusion model framework: including forward process and reverse process. In the forward process of the diffusion model, the source domain label is converted into a one-dimensional vector y0, and the noise prediction network is pre-trained to obtain the global noise prior. and local noise prior Add noise to y0 T times to destroy the original distribution of y0 until a variable y with an approximate Gaussian distribution is obtained. T ; Use Gaussian distributed variables in the reverse process Instead of y T , and use the noise prediction network to learn the noise distribution features added in the forward distribution. Gradually denoise, and finally get the same distribution as y0 S30, input the source domain data into the noise prediction network to fully extract the data features, input the target domain data into the noise prediction network, use the multi-source domain adaptation algorithm to map the source domain features to the target domain, and obtain the global noise prior corresponding to the target domain data and local noise prior S40, inputting the source domain data and the target domain data into the noise prediction network, and using the multi-source domain adaptation algorithm to map the noise prediction network features to the target domain, the noise prediction network fuses the fault data features after feature migration, and learns the noise features related to the target domain data; S50, bringing the learned target domain-related Gaussian noise and the noise prior after migration into the reverse process in step S20, and performing the Gaussian distribution on the random variable De-noising step by step, predicting the fault data type of the target domain.

2. The multi-source domain rolling bearing fault diagnosis method based on diffusion model according to claim 1 is characterized in that: The noise prediction network includes a feature extraction network, a conditional guidance network and a U-Net network, and the feature extraction network is configured to extract data features based on a Ghost convolutional network; The conditional guidance network is configured to extract noise priors; the U-Net network is configured to fuse the data features and the noise priors.

3. The multi-source domain rolling bearing fault diagnosis method based on diffusion model according to claim 2 is characterized in that: The feature extraction network includes: A first convolutional layer, a first BN layer, a first ReLu layer, a Ghost layer and a fully connected layer are sequentially connected in series, wherein the Ghost layer includes a plurality of layers and is sequentially connected in series between the first ReLu layer and the fully connected layer, the first convolutional layer is configured to perform a convolution operation on the input data, the first BN layer and the first ReLu are configured to accelerate the convergence speed of the model and increase the generalization of the model; the plurality of Ghost layers are configured to reduce the parameters of the model while fully extracting the data features; the fully connected layer is configured to adjust the dimension of the output features to 1024.

4. The multi-source domain rolling bearing fault diagnosis method based on diffusion model according to claim 2 is characterized in that: The conditional guidance network comprises: ResNet18 network and CNN+Transform multi-feature fusion network. The ResNet18 network is configured to extract the global features of the fault signal and output the noise prior. The CNN+Transform multi-feature fusion network is configured to learn the fault data features of different scales, concatenate the results of convolution operations of different scales at the channel scale, and then input them into the Transformer to learn the relationship between features of different scales, and finally output the noise prior.

5. The multi-source domain rolling bearing fault diagnosis method based on diffusion model according to claim 4 is characterized in that: The CNN+Transform multi-feature fusion network includes: a convolution unit layer, a feature fusion layer and a Transformer module connected in series in sequence, the convolution unit layer includes a second convolution layer, a third convolution layer and a fourth convolution layer connected in parallel in sequence, the third convolution layer and the fourth convolution layer are respectively connected in series with the fifth convolution layer and the sixth convolution layer, the fifth convolution layer, the sixth convolution layer and the second convolution layer are connected in parallel and connected to the feature fusion layer; wherein the convolution unit layer is configured to learn fault data features of different scales, the feature splicing layer is configured to splice the learned features of different scales to obtain the feature size required by the Transformer; the Transformer module is configured to learn the correlation and dependency between different features to make the learned fault data features more complete.

6. The multi-source domain rolling bearing fault diagnosis method based on diffusion model according to claim 2 is characterized in that: The conditional guidance network uses weighted cross entropy loss and multi-source domain adaptation loss as the loss function of the conditional guidance network, which is expressed as follows: THE gl =L CE +γL MIgl . in, is the cross entropy loss, is a multi-source domain adaptation loss function; the former is used to learn the classification loss in the source domain, and the latter is used to learn the joint probability distribution difference between the feature distributions of multiple source domains and the target domain; γ is used to balance the two types of losses. It is initially 0. After learning the classification features of the source domain data, it increases with the increase in the number of training times, and the maximum value does not exceed 1.

7. The multi-source domain rolling bearing fault diagnosis method based on diffusion model according to claim 2 is characterized in that: The U-Net network uses weighted noise difference loss and multi-source domain adaptation loss as the loss function of the U-Net network, where the noise difference loss includes numerical difference and distribution difference after fusing different noise priors, and its expression is as follows: L gl =L diff +γ1L MI . In the formula, is the loss function between the noise predicted by the U-Net network and the real noise, which includes the mean square difference And the distribution difference after fusing different noise priors and λ balances the two types of losses and takes a fixed value of 0.5; It is the difference loss function of the feature distribution of multiple source domains and target domain data. γ1 balances the two types of losses. It is initially 0. After learning the features of the source domain data, it gradually increases, and the maximum value does not exceed 1.

8. The multi-source domain rolling bearing fault diagnosis method based on diffusion model according to claim 1 is characterized in that: In step S20, The expression of the forward noise addition process of the diffusion model is as follows: Where ∈ is random noise, is a parameter related to the number of noise additions, y0 is the one-dimensional vector into which the label is transformed, and y t is the variable obtained by adding noise t times; The expression of the denoising process of the inverse process of the diffusion model is as follows: Among them, ρ(x) is the feature vector learned by the feature extraction network, y t is the variable obtained by adding noise t times to the prediction, p θ is the distribution characteristic of Gaussian noise.

9. The multi-source domain rolling bearing fault diagnosis method based on diffusion model according to claim 1 is characterized in that: In step S30 and step S40, the multi-source domain adaptation algorithm is configured to perform feature alignment on multiple source domain and target domain data on the joint probability distribution, and transform the problem of solving feature alignment into the problem of finding the maximum value of a convex function. The specific derivation process is as follows: Where MI(h(X), Y; L) is the multi-source domain loss function, h(x), y, l, p and m are the output features, the original labels of the data, the number of source domains, the feature probability distribution, and the number of all data, respectively. Formula (2) is substituted into formula (1), and the similar features contained in each source domain and target domain are approximately equal. The sample mean is used to replace the convolution to simplify the formula (3); r(h(x), y, l; θ) is a linear function about θ, where k(h(x), h(x i )) is the Gaussian kernel function, δ(y,y i ) is the indicator kernel, when y=y i 1 when δ(l,l i ) is also the indicator core, when l = l i 1 when , and 0 otherwise; Substitute formula (4) into formula (3) to obtain the final loss function, where the two sub-formulas in formula (5) are about Convex function of .

10. A diagnostic system for a rolling bearing fault diagnosis method based on a multi-source domain diffusion model according to any one of claims 1 to 9, characterized in that: The steps include: A data acquisition module is configured to collect rolling bearing vibration signals under different speed conditions, using a certain unlabeled speed condition data as target domain data and other speed condition fault data as source domain data; Construct a model module, configured to convert the source domain label into a one-dimensional vector y0 in the forward process of the diffusion model, pre-train the noise prediction network, and obtain the global noise prior and local noise prior Add noise to y0 T times to destroy the original distribution of y0 until a variable y with an approximate Gaussian distribution is obtained. T ; Use Gaussian distributed variables in the reverse process Instead of y T , and use the noise prediction network to learn the noise distribution features added in the forward distribution. Gradually denoise, and finally get the same distribution as y0 The pre-training module is configured to input the source domain data into the noise prediction network to fully extract the data features, input the target domain data into the noise prediction network, and use the multi-source domain adaptation algorithm to map the source domain features to the target domain to obtain the global noise prior corresponding to the target domain data. and local noise prior The noise migration module is configured to input the source domain data and the target domain data into the noise prediction network, and use the multi-source domain adaptation algorithm to map the noise prediction network features to the target domain. The noise prediction network fuses the fault data features after feature migration and learns the noise features related to the target domain data. The fault prediction module is configured to bring the learned target domain-related Gaussian noise and the noise prior after migration into the reverse process of the model building module, and perform the Gaussian distribution of random variables De-noising step by step, predicting the fault data type of the target domain.

Images