Single-source-domain generalization intelligent identification method based on flow model feature enhancement

Abstract

The invention provides a single-source-domain generalization intelligent identification method based on flow model feature enhancement, and relates to the technical field of mechanical fault diagnosis, and the method comprises the steps: carrying out the interception, length unification and amplitude normalization preprocessing of a collected mechanical vibration time domain signal, converting the signal into frequency domain data through fast Fourier transform, and carrying out the analysis of the frequency domain data; dividing into a single-source domain data set and a target domain data set according to working conditions; constructing a fault diagnosis training model; training a fault diagnosis training model according to a set loss function and an optimization algorithm by using the single-source domain data set to obtain a trained fault diagnosis training model; and constructing a fault recognition model based on the trained fault diagnosis training model, and inputting the target domain data set into the fault recognition model for fault category recognition. According to the method, the problems of high dimension of generated data, interference information and the like of a traditional method are solved, and the model generalization ability and the target working condition fault recognition precision are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and in particular to a single-source domain generalization intelligent recognition method based on flow model feature enhancement. Background Art

[0002] In recent years, deep learning methods have been widely used in mechanical fault diagnosis due to their advantages in automatic feature extraction. However, traditional methods essentially rely on the independent and identically distributed assumption. When there are differences in the working conditions (such as load, rotational speed, or sensor deployment) between the training data and the target data, domain distribution shift will occur in the target domain data, resulting in a sharp decline in the accuracy of mechanical fault recognition. To alleviate the domain shift, transfer learning adapts the source domain and the target domain through domain alignment. Some studies introduce the adversarial concept to obtain domain-invariant features through adversarial training, but this method requires some labeled target domain data to participate in the training, which is unrealistic in industrial environments, and discriminative fault features may be distorted under large domain shifts.

[0003] Therefore, domain generalization methods that do not require target domain data have received attention. Multi-source domain generalization constructs a strong generalization model by aggregating multi-source knowledge. Its core is to extract domain-invariant features and suppress domain-specific interference. However, the complex distribution differences between multi-source domains (such as device type, working environment, sampling frequency changes) make it difficult to align the feature space. The low correlation between some source domains and the target domain complicates the weighting, and there is a risk of negative transfer. Moreover, in industrial scenarios, the cost of collecting multi-source data is high, and it is restricted by privacy or protocols, resulting in limited availability of the dataset, which restricts its application in complex industrial environments.

[0004] Single-source domain generalization constructs a generalizable model only using single-source data, and performs fault recognition through single-source feature extraction, reducing the demand for the number of source domains and the manual and time costs of data collection and labeling, which is more in line with actual industrial needs. The core of the existing single-source domain generalization fault diagnosis technology is to generate an extended domain dataset through data augmentation, and then combine feature extraction and transfer learning to extract domain-invariant features to generalize to different target working conditions. However, its performance depends on the diversity and consistency of the extended domain dataset relative to the source domain. The high-dimensional data samples generated by existing methods not only contain health state information but also contain interference information unrelated to the state, affecting the quality of the generated data, resulting in problems such as high dimension of the generated data, complex enhancement network, weak state representation ability, poor generalization performance, and low recognition accuracy of target working conditions. Summary of the Invention

[0005] Therefore, an embodiment of the present invention provides a single-source domain generalization intelligent recognition method based on flow model feature enhancement, which is used to solve the problems of high dimension of the generated data, containing interference information, low quality, poor generalization performance, and low recognition accuracy of target working condition faults in the prior art.

[0006] To solve the above problems, an embodiment of the present invention provides a single-source domain generalization intelligent recognition method based on enhanced flow model features, including: S1: Intercept, unify the length, and perform amplitude normalization preprocessing on the collected mechanical vibration time-domain signal, convert it into frequency-domain data through fast Fourier transform, and divide it into a single-source domain dataset and a target domain dataset according to working conditions; S2: Construct a fault diagnosis training model, which is composed of a deep feature extraction module, a feature generation module, and a fault diagnosis module, where: The deep feature extraction module is used to extract deep features of the source domain; The feature generation module is used to generate extended domain features with different distributions from the deep features of the source domain; The fault diagnosis module is used to extract domain-invariant features and perform category recognition and domain discrimination; S3: Use the single-source domain dataset to train the fault diagnosis training model according to the set loss function and optimization algorithm to obtain a trained fault diagnosis training model; S4: Construct a fault recognition model based on the trained fault diagnosis training model, and input the target domain dataset into the fault recognition model for fault category recognition.

[0007] Preferably, the deep feature extraction module is composed of a feature extractor F1 and a classifier C1, where the feature extractor F1 is used to input the preprocessed single-source domain dataset and output deep features of the source domain; The classifier C1 is used to input the deep features of the source domain output by the feature extractor F1 and output fault category labels.

[0008] Preferably, the feature generation module is composed of a binary masker, a low-frequency perturbation module, and a flow model generation module, where the binary masker is used to input the deep features of the source domain output by the feature extractor F1, rearrange the deep features of the source domain with the low-frequency components in the middle and the high-frequency components at both ends, generate a binary mask of the same length as the feature vector, multiply the rearranged feature vector element by element with the binary mask, and separate the low-frequency and high-frequency components of the deep features of the source domain; The low-frequency perturbation module is used to input the low-frequency components output by the binary masker, calculate the statistical distribution of the low-frequency components, and superimpose Gaussian perturbations to generate a new statistical distribution; The flow model generation module is used to input the new statistical distribution output by the low-frequency perturbation module, map the low-frequency components to the new statistical distribution, and generate new low-frequency components; The feature generation module combines the new low-frequency components with the high-frequency components to generate extended domain features with different distributions from the deep features of the source domain.

[0009] Preferably, the central region of the mask matrix of the binary masker is 1 and the two ends are 0, and the length of the central region accounts for 50% of the total length.

[0010] Preferably, the operations of the low-frequency perturbation module include: Determine the number of sub-distributions according to the number of categories, and each category has a set of statistical distributions, that is, sub-distributions; Calculate the variance of the mean and variance of the low-frequency components, multiply by the perturbation coefficient and then superimpose it on the original statistical distribution to obtain a new statistical distribution after random perturbation.

[0011] Preferably, the flow model generation module adopts an affine coupling layer structure, optimizes the model by maximizing the log-likelihood loss function, and the flow model generation module is constructed by at least one of the NICE model, the Real-NVP model, and the GLOW model.

[0012] Preferably, the fault diagnosis module consists of a feature extractor F2, a classifier C2, and a domain discriminator Y. The feature extractor F2 is used to input the source domain deep features output by the feature extractor F1 and the extended domain features generated by the feature generation module, and output domain-invariant features; The classifier C2 is used to input the domain-invariant features output by the feature extractor F2 and output the predicted fault category; The domain discriminator Y is used to input the domain-invariant features output by the feature extractor F2 and output the predicted domain category.

[0013] Preferably, the process of training the fault diagnosis training model using the single-source domain dataset according to the set loss function and optimization algorithm to obtain the trained fault diagnosis training model includes: S31: Input the single-source domain dataset into the deep feature extraction module, and minimize the cross-entropy loss function of the classifier C1 Optimize the feature extractor F1 so that the source domain deep features extracted by the feature extractor F1 Satisfy: ; In the formula, is the source domain data, is the corresponding label; S32: Input the source domain deep features into the feature generation module: S321: Separate the low-frequency components from the high-frequency components : ; Among them, represents the feature reordering operation, represents the binary mask matrix, denotes element-wise dot multiplication; S322: In the low-frequency perturbation module, calculate the mean and variance of the low-frequency features by category to generate a new statistical distribution: ; In the formula, is the fault category index, is the low-frequency feature of the class-conditional mean, is the low-frequency feature of the class-conditional variance, is the mean of each category of the variance, is the variance of each category of the variance, is sampled from a normal distribution with a mean of 0 and a specified variance, is the perturbation coefficient, is the new class-conditional mean after perturbation, is the new class-conditional variance after perturbation; S323: In the flow model generation module, train the flow model generation module by maximizing the log-likelihood loss : ; In the formula, is the flow model generation module, is the number of batch training samples, is the flow model generation module defined probability density function; Generate a new low-frequency component and combine it with the high-frequency component to form an extended-domain feature: ; In the formula, is the extended-domain feature, is the feature concatenation operation; S33: Input the source domain deep feature and the extended-domain feature into the fault diagnosis module, and minimize the total loss function : ; In the formula, is the supervised contrast loss, which promotes the compactness of intra-class features and the separation of inter-class features; is the binary cross-entropy loss of the domain discriminator Y, which distinguishes the source domain / extended-domain features; is the cross-entropy loss of the classifier C2, which optimizes the fault category recognition accuracy; , is the weighting coefficient for weighing different losses; S34: Repeat the execution of S31 - S33, update the parameters using an optimization algorithm, and stop training when the number of iterations reaches the preset number of iterations or the total loss converges to a stable state, to obtain a trained fault diagnosis training model.

[0014] Preferably, the optimization algorithm is at least one of the root mean square propagation algorithm, the stochastic gradient descent algorithm, or the adaptive moment estimation algorithm.

[0015] Preferably, the method for constructing a fault identification model based on the trained fault diagnosis training model is as follows: Based on the trained fault diagnosis training model, obtain the trained feature extractor F1, feature extractor F2, and classifier C2, and construct a fault identification model with the trained feature extractor F1, feature extractor F2, and classifier C2.

[0016] From the above technical solutions, it can be seen that the present invention application has the following beneficial effects: (1) Reduce data dimension and computational complexity: Abandon the traditional way of generating high-dimensional samples through data augmentation. First, extract the low-dimensional deep features of the source domain data through the deep feature extraction module, and then generate the extended domain features through the feature generation module, reducing the computational complexity and avoiding the interference information in high-dimensional data, making the model training and inference more efficient.

[0017] (2) Improve the feature representation ability: The feature generation module separates the low-frequency and high-frequency components through the binary masker, and combines the low-frequency perturbation module and the flow model generation module to perturb and map the low-frequency components. The generated extended domain features can more accurately represent the mechanical health state, excluding the interference information irrelevant to the state and enhancing the discriminability of the features.

[0018] (3) Enhance feature diversity and consistency: The low-frequency perturbation module simulates the feature offset caused by real working condition changes by adjusting the perturbation coefficient, generating diverse statistical distributions; the flow model generation module aligns the low-frequency features to the sub-distributions of the corresponding categories, ensuring that the generated features are diverse and consistent at the same time, and improving the quality of the extended domain features.

[0019] (4) Improve the model generalization ability: The fault diagnosis module extracts the domain-invariant features through the collaborative action of the supervised contrast loss, the domain discriminator, and the classifier, enabling the model to adapt to the domain distribution shift of unknown target working conditions and having strong generalization ability for mechanical faults under different working conditions, with a wide range of applications.

[0020] (5) Improve the accuracy of fault identification: Under the dual effects of feature enhancement and extraction of domain-invariant features, the model can accurately identify mechanical faults under target working conditions. The experimental results show that it has a high diagnostic accuracy for various health states and can achieve reliable identification especially in the case of complex compound faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. By referring to the drawings, the features and advantages of the present invention will be more clearly understood. The drawings are schematic and should not be construed as limiting the present invention in any way. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. Among them: Figure 1 It is a flowchart of a single-source domain generalization intelligent identification method based on flow model feature enhancement provided by the present invention; Figure 2 It is a schematic structural diagram of a fault diagnosis training model provided by the present invention; Figure 3 It is a schematic structural diagram of a fault identification model for applying the trained fault diagnosis training model to an unknown target working condition that has not participated in the training; Figure 4 It is a schematic diagram of the visualization clustering result of the high-level hidden features of the target domain output by the feature extractor F2 in the fault identification model provided by the present invention; Figure 5 It is a confusion matrix of the target domain bearing health state results predicted by the fault identification model provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0023] Embodiment 1:

[0024] To solve the problems in the prior art such as high dimensionality of generated data, containing interference information, low quality, poor generalization performance, and low accuracy of fault identification for target working conditions as Figure 1 shown, the present invention proposes a single-source domain generalization intelligent identification method based on flow model feature enhancement, including: S1: Intercept the collected mechanical vibration time-domain signal, unify the length, and perform amplitude normalization preprocessing. Convert it into frequency-domain data through fast Fourier transform, and divide it into a single-source domain dataset and a target domain dataset according to working conditions; S2: Construct a fault diagnosis training model. The fault diagnosis training model consists of a deep feature extraction module, a feature generation module, and a fault diagnosis module. The deep feature extraction module is used to extract deep features of the source domain; the feature generation module is used to generate extended domain features with different distributions from the deep features of the source domain; the fault diagnosis module is used to extract domain-invariant features and perform class recognition and domain discrimination; S3: Use the single-source domain dataset to train the fault diagnosis training model according to the set loss function and optimization algorithm to obtain a trained fault diagnosis training model; S4: Construct a fault recognition model based on the trained fault diagnosis training model, and input the target domain dataset into the fault recognition model for fault class recognition.

[0025] As can be seen from the above technical solutions, the present invention proposes a single-source domain generalization intelligent recognition method based on flow model feature enhancement. In step S1, by intercepting, unifying the length, performing amplitude normalization preprocessing on the mechanical vibration time-domain signal, and converting it into frequency-domain data through fast Fourier transform, and dividing it into a single-source domain and a target domain dataset according to working conditions, it can eliminate noise interference and data format differences in the signal acquisition process, provide standardized and structured data input for subsequent model training, and ensure the independence and distinguishability of data under different working conditions. Step S2 constructs a fault diagnosis training model consisting of a deep feature extraction module, a feature generation module, and a fault diagnosis module, realizing the modular integration of functions from deep feature extraction of the source domain to extended domain feature generation, and then to extraction of domain-invariant features and class / domain discrimination, laying a model architecture foundation for generalization recognition. Step S3 uses the single-source domain dataset and trains the model based on the set loss function and optimization algorithm, enabling the model to learn the distribution laws of source domain features and extended domain features under single-source data drive, enhancing the feature representation ability and domain invariance through loss function optimization, and forming a fault diagnosis model with generalization ability. Step S4 constructs a fault recognition model based on the trained fault diagnosis training model, inputs the target domain dataset into the fault recognition model for fault class recognition, and can utilize the feature extraction and generalization capabilities obtained by the model in single-source domain training to achieve accurate classification of mechanical faults under unknown working conditions, verifying the effectiveness of cross-working condition diagnosis of the model. The present invention solves problems such as high data dimension and interference information in traditional methods, and improves the model generalization ability and target working condition fault recognition accuracy.

[0026] In step S1, the mechanical vibration time-domain signal collected by the present invention is intercepted, unified in length, and amplitude-normalized for preprocessing, and then converted into frequency-domain data through fast Fourier transform, and divided into a single-source domain dataset and a target domain dataset according to working conditions.

[0027] It should be noted that the dataset is divided according to different working conditions of the machine. The same rotational speed and load are in the same working condition. Multiple health state samples under the same working condition are placed in the same domain, and the same mechanical health state category is included in different domains. Only the single-source domain dataset is used for model training; the target domain is not visible during the model training stage and is only used for model testing, so it does not participate in model training.

[0028] In step S2, the present invention constructs a fault diagnosis training model, as Figure 2 shown. The fault diagnosis training model is composed of a deep feature extraction module E, a feature generation module G, and a fault diagnosis module D. Among them, the deep feature extraction module E is used to extract deep features of the source domain; the feature generation module G is used to generate extended domain features with different distributions from the deep features of the source domain; the fault diagnosis module D is used to extract invariant features between domains and perform class recognition and domain discrimination.

[0029] Specifically, the deep feature extraction module E is composed of a feature extractor F1 and a classifier C1. Among them, the feature extractor F1 is used to input the preprocessed single-source domain dataset and output the deep features of the source domain; the classifier C1 is used to input the deep features of the source domain output by the feature extractor F1 and output the fault category label.

[0030] It should be noted that the feature extractor F1 is constructed by at least one of a fully connected network, a deep convolutional network, a deep belief network, or a deep residual network. The classifier C1 is composed of a fully connected layer and a Softmax classifier, and its label classification loss is the cross-entropy loss of predicting the category label from the samples in the labeled source domain .

[0031] Furthermore, the feature generation module G consists of a binary masker M, a low-frequency perturbation module A, and a flow model generation module R. The binary masker M is used to input the deep features of the source domain output by the feature extractor F1, rearrange the deep features of the source domain with the low-frequency components centered and the high-frequency components at both ends, generate a binary mask of the same length as the feature vector, and multiply the rearranged feature vector element-wise with the binary mask to separate the low-frequency and high-frequency components of the deep features of the source domain. The low-frequency perturbation module A is used to input the low-frequency components output by the binary masker M, calculate the statistical distribution of the low-frequency components, and superimpose Gaussian perturbations to generate a new statistical distribution (for example, assuming that the mean and variance of a batch of data both conform to a Gaussian distribution, calculate the variance of the mean and the variance of the variance, multiply them by the perturbation coefficients respectively, and then add them to the mean and variance respectively to obtain a new statistical distribution after random perturbation). The flow model generation module R is used to input the new statistical distribution output by the low-frequency perturbation module A, map the low-frequency components to the new statistical distribution, and generate new low-frequency components. Finally, the feature generation module G combines the new low-frequency components and the high-frequency components to generate extended domain features with a different distribution from the deep features of the source domain.

[0032] It should be noted that the central region of the mask matrix of the binary masker M is 1 and the two ends are 0, and the length of the central region accounts for 50% of the total length. The operations of the low-frequency perturbation module A include: determining the number of sub-distributions according to the number of categories, and each category has a set of statistical distributions, that is, sub-distributions; calculating the variances of the mean and variance of the low-frequency components, multiplying them by the perturbation coefficients and then superimposing them on the original statistical distribution to obtain a new statistical distribution after random perturbation, where the perturbation coefficients can be adjusted to obtain different statistical distributions. The low-frequency perturbation module A fits the distribution by statistical methods, so no training is required, and at the same time, the number of model parameters is reduced. The flow model generation module R adopts an affine coupling layer structure to enhance the non-linear representation ability and optimizes the model by maximizing the log-likelihood loss function. The flow model generation module R is constructed by at least one of the NICE model, the Real-NVP model, and the GLOW model.

[0033] Furthermore, the fault diagnosis module D consists of a feature extractor F2, a classifier C2, and a domain discriminator Y. The feature extractor F2 is used to input the deep features of the source domain output by the feature extractor F1 and the extended domain features generated by the feature generation module G, and output domain-invariant features. The classifier C2 is used to input the domain-invariant features output by the feature extractor F2 and output the predicted fault categories. The domain discriminator Y is used to input the domain-invariant features output by the feature extractor F2 and output the predicted domain categories.

[0034] It should be noted that the feature extractor F2 is constructed by at least one of a fully connected network, a deep convolutional network, a deep belief network, or a deep residual network. The domain discriminator Y consists of a fully connected layer and a Sigmoid activation function, and its domain discrimination loss is the binary cross-entropy loss of the predicted domain labels of the source domain and extended domain samples. The classifier C2 consists of a fully connected layer and a Softmax classifier, and its label classification loss is the cross-entropy loss of the predicted class labels of the samples in the labeled source domain and the extended domain. The features projected by the feature extractor F2 use supervised contrastive learning to establish clear class boundaries, promote intra-class compactness and inter-class separability, and its loss is the supervised contrastive loss. .

[0035] In step S3, the present invention uses a single-source domain dataset to train a fault diagnosis training model according to the set loss function and optimization algorithm, and obtains a trained fault diagnosis training model, including: S31: Input the single-source domain dataset into the deep feature extraction module to minimize the cross-entropy loss function of the classifier C1. Optimize the feature extractor F1 so that the deep features of the source domain extracted by the feature extractor F1 Satisfy: ; In the formula, is the source domain data, is the corresponding label.

[0036] S32: Input the deep features of the source domain into the feature generation module: S321: Separate the low-frequency component and the high-frequency component : ; Among them, represents the feature reordering operation, represents the binary mask matrix, represents the element-wise dot product; S322: In the low-frequency perturbation module, calculate the mean and variance of the low-frequency features by category to generate a new statistical distribution: ; In the formula, is the fault category index, is the low-frequency feature of the class-conditional mean, is the low-frequency feature of the class-conditional variance, is the mean of each category of the variance, is the variance of each category of the variance, is sampled from a normal distribution with a mean of 0 and a specified variance, is the perturbation coefficient, is the new class conditional mean after perturbation, is the new class conditional variance after perturbation; S323: In the flow model generation module, by maximizing the log-likelihood loss train the flow model generation module: ; wherein, is the flow model generation module, is the number of batch training samples, is the probability density function defined by the flow model generation module ; Generate a new low-frequency component , and combine it with the high-frequency component to form an extended domain feature: ; wherein, is the extended domain feature, is the feature splicing operation.

[0037] S33: Input the source domain deep feature and the extended domain feature into the fault diagnosis module, and minimize the total loss function : ; wherein, is the supervised contrast loss, which promotes the compactness of intra-class features and the separability of inter-class features; is the binary cross-entropy loss of the domain discriminator Y, which distinguishes the source domain / extended domain features; is the cross-entropy loss of the classifier C2, which optimizes the fault category recognition accuracy; , are the weighting coefficients for weighing different losses.

[0038] S34: Repeat S31 - S33, update the parameters using an optimization algorithm, and stop training when the number of iterations reaches the preset number of iterations or the total loss converges to a stable state, and obtain the trained fault diagnosis training model.

[0039] It should be noted that the optimization algorithm is at least one of the root mean square propagation algorithm, the stochastic gradient descent algorithm, or the adaptive moment estimation algorithm.

[0040] In step S4, the present invention constructs a fault recognition model based on the trained fault diagnosis training model, and inputs the target domain data set into the fault recognition model for fault category recognition.

[0041] Specifically, based on the trained fault diagnosis training model, the trained feature extractors F1, F2, and classifier C2 are obtained, and the trained feature extractors F1, F2, and classifier C2 are constructed into a fault recognition model.

[0042] To more clearly understand the technical solution and its effects of the present invention, a specific experiment will be described in detail below.

[0043] 1. Experimental data: An experiment was carried out on the train axle box bearing dataset. This dataset covers 11 types of health state data, specifically including normal state (N), inner race fault (I), outer race fault (O), rolling element fault (R), cage fault (C), inner and outer race compound fault (IO), outer race and rolling element compound fault (OR), outer race and cage compound fault (OC), inner and outer race and rolling element compound fault (IOR), inner and outer race and cage compound fault (IOC), inner and outer race and rolling element and cage compound fault (IORC), which respectively correspond to fault class labels from 0 to 10. The experiment used data under nine working conditions. The data under the same working condition form a domain, and the number of samples in different health states in each domain is 312. During the experiment, one marked source domain in Table 1 was selected as the training dataset, and any one of the remaining domains was used as the target domain test dataset.

[0044] Table 1 Description of the experimental bearing dataset

[0045] 2. Experimental steps: Step 1: First, intercept the collected mechanical vibration time-domain signal, unify the sample length to 2048 points, and perform normalization processing on the sample amplitude in the interval [0,1]; then convert the time-domain signal to frequency-domain data through the fast Fourier transform, and the sample length becomes 1024 after conversion; finally, divide the dataset according to the working conditions, and classify the frequency-domain data under the same rotational speed and load conditions into the same domain. One domain is used as the single-source domain training dataset, and the other domain is used as the target domain test dataset. The preprocessed frequency-domain data is directly used as the model input sample.

[0046] Step 2: Construct a fault diagnosis training model, which is composed of a deep feature extraction module, a feature generation module, and a fault diagnosis module, as Figure 2 shown. Among them The deep feature extraction module E consists of a feature extractor F1 and a classifier C1. The feature extractor F1 uses four one-dimensional convolutional layers, followed by a batch normalization layer and a ReLU activation function in sequence for each layer. Taking the preprocessed 1024-dimensional frequency domain signal as input, it outputs a 128-dimensional deep feature vector. The classifier C1 is a fully connected network followed by a Softmax activation function. Taking the output of the feature extractor F1 as input, it outputs an 11-dimensional health state category vector, and calculates the cross-entropy loss of the fault category of the source domain samples based on the label output probability 。

[0047] The feature generation module G includes a binary masker M, a low-frequency perturbation module A, and a flow model generation module R. The central region (accounting for 50% of the total length) of the binary masker M is set to 1, and the two ends are 0. Multiply the deep features of the re-ordered and centered low-frequency components by M to separate the low-frequency and high-frequency components. The low-frequency perturbation module A calculates the mean and variance of the low-frequency components, and generates a new statistical distribution through the perturbation coefficient (each class corresponds to a set of sub-distributions). The flow model generation module R adopts a Real-NVP structure with four affine coupling layers, maps the low-frequency components to the target sub-distribution, and finally combines them with the original high-frequency components to generate extended domain features

[0048] The fault diagnosis module D consists of a feature extractor F2, a classifier C2, and a domain discriminator Y. The feature extractor F2 outputs 128-dimensional domain-invariant features through four one-dimensional convolutional upsampling (from 128 dimensions to 2048 dimensions) and downsampling (from 2048 dimensions to 128 dimensions). The classifier C2 is a fully connected network followed by a Softmax activation function, outputs an 11-dimensional health state category vector, and calculates the cross-entropy loss of the source domain and extended domain samples ; The domain discriminator Y is a two-layer fully connected network (the hidden layer dimensions are 64 and 1), followed by ReLU and Sigmoid activation functions in sequence, outputs a one-dimensional domain category vector, and calculates the binary cross-entropy loss between the source domain and the extended domain 。

[0049] Step 3: Train the fault diagnosis training model with a single source domain dataset. The specific steps are as follows: Input the single source domain dataset into the constructed model, and perform training according to the set loss function and optimization algorithm. First, in the deep feature extraction module, for all labeled data in the source domain, optimize the feature extractor F1 by minimizing the cross-entropy loss function of the classifier C1 to optimize the feature extractor F1; then, in the feature generation module, pass the source domain deep features output by the deep feature extraction module through the binary masker M and the low-frequency perturbation module A in sequence to determine the sub-distributions of all classes, and then through maximum likelihood estimation Optimize the convection model generation module R; in the fault diagnosis module, the total model loss function is set to , where the supervised contrastive loss promotes the aggregation of similar features through the comparison of positive and negative sample pairs. The classifier C2 uses the cross-entropy loss to ensure the accuracy of class recognition. The domain discriminator Y uses the binary cross-entropy loss to distinguish the domain to which the features belong. The goal of the total loss is to map the features to a specific feature space so that the classifier C2 can distinguish the classes of the features in the source domain and the unknown target domain. The optimization algorithm uses the Adaptive Moment Estimation algorithm (Adam), sets the learning rate to 0.001, and after 200 iterations of training, when the loss of the model objective function tends to be stable, the training process ends.

[0050] Step 4: Construct the trained feature extractor F1, feature extractor F2, and classifier C2 into a fault recognition model, input the target domain dataset into this model, and complete the recognition of fault categories. The specific process is shown in Figure 3.

[0051] 3. Experimental analysis: As Figure 4 shown, the t-SNE technology is used to visualize the feature extraction results of the target test domain rolling bearing data samples, where different shapes represent 11 fault categories (the category information is marked in Figure 4 the upper right corner). The visualization results show that the features of the same category in the target domain show significant aggregation, the boundaries between different category features are clear, and the discrimination is strong, verifying that the method of the present invention can learn feature representations independent of data distribution and extract more effective features conducive to classification, realizing the efficient extraction of domain-invariant features.

[0052] The confusion matrix of the diagnosis results of the method of the present invention for the target domain test dataset is as Figure 5 shown. The experimental data shows that the diagnosis accuracy of the method for the target domain data of 8 health states reaches 100%, only a small number of misjudgments occur for the outer ring plus cage compound fault (OC, label 7), and the recognition accuracies for the bearing rolling element fault (R, label 3) and the inner ring plus outer ring compound fault (IO, label 5) both exceed 90%, fully demonstrating its excellent fault recognition performance under unknown target working conditions.

[0053] In summary, by constructing the deep feature extraction module E and the feature generation module G, the present invention realizes low-dimensional feature enhancement. While reducing the computational complexity and excluding noise interference, it effectively generates diverse extended domain features containing category information, significantly improving the model generalization ability. By establishing the fault diagnosis module D, it further realizes the efficient extraction of domain-invariant features, and finally enables the single-source domain generalization model to accurately identify mechanical faults under unknown working conditions.

[0054] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0055] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for realizing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0056] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured product including an instruction device, and the instruction device realizes the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks. These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for realizing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0057] Obviously, the above embodiments are merely examples given for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or alterations can be made based on the above description. It is not necessary and impossible to list all implementation manners here. And the obvious changes or alterations derived therefrom still fall within the protection scope of this invention.

Claims

1. A single-source domain generalization intelligent recognition method based on enhanced flow model features, characterized in that Including: S1: Intercept, unify the length, and perform amplitude normalization preprocessing on the collected mechanical vibration time-domain signal, convert it into frequency-domain data through fast Fourier transform, and divide it into a single-source domain dataset and a target domain dataset according to the working conditions; S2: Construct a fault diagnosis training model, which consists of a deep feature extraction module, a feature generation module, and a fault diagnosis module, where: The deep feature extraction module is used to extract deep features of the source domain; The feature generation module is used to generate extended domain features with different distributions from the deep features of the source domain; The fault diagnosis module is used to extract domain-invariant features and perform category recognition and domain discrimination; S3: Use the single-source domain dataset to train the fault diagnosis training model according to the set loss function and optimization algorithm to obtain a trained fault diagnosis training model; S4: Construct a fault recognition model based on the trained fault diagnosis training model, and input the target domain dataset into the fault recognition model for fault category recognition.

2. The single-source domain generalization intelligent recognition method based on enhanced flow model features according to claim 1, wherein The deep feature extraction module consists of a feature extractor F1 and a classifier C1, where the feature extractor F1 is used to input the preprocessed single-source domain dataset and output deep features of the source domain; The classifier C1 is used to input the deep features of the source domain output by the feature extractor F1 and output fault category labels.

3. The single-source domain generalization intelligent recognition method based on enhanced flow model features according to claim 1, wherein The feature generation module consists of a binary masker, a low-frequency perturbation module, and a flow model generation module, where the binary masker is used to input the deep features of the source domain output by the feature extractor F1, rearrange the deep features of the source domain with the low-frequency components in the middle and the high-frequency components at both ends, generate a binary mask of the same length as the feature vector, and multiply the rearranged feature vector element by element with the binary mask to separate the low-frequency and high-frequency components of the deep features of the source domain; The low-frequency perturbation module is used to input the low-frequency components output by the binary masker, calculate the statistical distribution of the low-frequency components, and superimpose Gaussian perturbations to generate a new statistical distribution; The flow model generation module is used to input the new statistical distribution output by the low-frequency perturbation module, map the low-frequency components to the new statistical distribution, and generate new low-frequency components; The feature generation module combines the new low-frequency components with the high-frequency components to generate extended domain features with different distributions from the deep features of the source domain.

4. The single-source domain generalization intelligent recognition method based on enhanced flow model features according to claim 3, wherein, The central region of the mask matrix of the binary masker is 1 and the two ends are 0, and the length of the central region accounts for 50% of the total length.

5. The single-source domain generalization intelligent recognition method based on enhanced flow model features according to claim 3, wherein The operations of the low-frequency perturbation module include: Determine the number of sub-distributions according to the number of categories, and each category has a set of statistical distributions, that is, sub-distributions; Calculate the variance of the mean and variance of the low-frequency components, multiply by the perturbation coefficient, and superimpose it on the original statistical distribution to obtain a new statistical distribution after random perturbation.

6. The single-source domain generalization intelligent recognition method based on flow model feature enhancement according to claim 3, wherein The flow model generation module adopts an affine coupling layer structure, optimizes the model by maximizing the log-likelihood loss function, and the flow model generation module is constructed by at least one of the NICE model, the Real-NVP model, and the GLOW model.

7. The single-source domain generalization intelligent recognition method based on enhanced flow model features according to claim 1, characterized in that The fault diagnosis module consists of a feature extractor F2, a classifier C2, and a domain discriminator Y. The feature extractor F2 is used to input the source domain deep features output by the feature extractor F1 and the extended domain features generated by the feature generation module, and output the inter-domain invariant features. The classifier C2 is used to input the inter-domain invariant features output by the feature extractor F2 and output the predicted fault categories. The domain discriminator Y is used to input the inter-domain invariant features output by the feature extractor F2 and output the predicted domain categories.

8. The single-source domain generalization intelligent recognition method based on enhanced flow model features according to claim 1, characterized in that The process of training the fault diagnosis training model using the single-source domain dataset according to the set loss function and optimization algorithm to obtain the trained fault diagnosis training model includes: S31: Input the single-source domain dataset into the deep feature extraction module, and minimize the cross-entropy loss function of classifier C1 Optimize the feature extractor F1 so that the deep features of the source domain extracted by the feature extractor F1 Meet the following conditions: ； wherein, is the source domain data, is the corresponding label; S32: Input the deep features of the source domain Input feature generation module: S321: Separate the low-frequency components through a binary masker and the high-frequency components : ； Among them, represents a feature reordering operation, represents a binary mask matrix, represents an element-wise dot product; S322: In the low-frequency perturbation module, calculate the mean and variance of the low-frequency features by category to generate a new statistical distribution: ​ ； Wherein, is the fault category index, is the class conditional mean of the low-frequency feature , is the class conditional variance of the low-frequency feature , is the variance of the mean values of each category , is the variance of the variances of each category , is sampled from a normal distribution with a mean of 0 and a specified variance, is the perturbation coefficient, is the new class conditional mean after perturbation, is the new class conditional variance after perturbation; S323: In the flow model generation module, maximize the log-likelihood loss Train the flow model generation module: ； In the formula, is the flow model generation module, is the number of batch training samples, is the flow model generation module defines the probability density function; Generate a new low-frequency component , and combine it with the high-frequency component to form an extended-domain feature: ； In the formula, is the extended domain feature, is the feature splicing operation; S33: Input the deep features of the source domain and the features of the extended domain into the fault diagnosis module to minimize the total loss function : ； In the formula, is the supervised contrastive loss, which promotes the intra-class feature compactness and inter-class separability; is the binary cross-entropy loss of the domain discriminator Y, which distinguishes the source domain / extended domain features; is the cross-entropy loss of the classifier C2, which optimizes the fault class recognition accuracy; and are the weighting coefficients for balancing different losses; S34: Repeat S31 - S33, update the parameters using the optimization algorithm, and stop training when the number of iterations reaches the preset number of iterations or the total loss converges to a stable state to obtain the trained fault diagnosis training model.

9. The single-source domain generalization intelligent recognition method based on enhanced flow model features according to claim 8, characterized in that, The optimization algorithm is at least one of the root mean square propagation algorithm, the stochastic gradient descent algorithm, or the adaptive moment estimation algorithm.

10. The single-source domain generalization intelligent recognition method based on flow model feature enhancement according to claim 1, characterized in that The method for constructing a fault recognition model based on the trained fault diagnosis training model is: Based on the trained fault diagnosis training model, obtain the trained feature extractor F1, feature extractor F2, and classifier C2, and construct the trained feature extractor F1, feature extractor F2, and classifier C2 into a fault recognition model.

Images