Single-source domain generalized intelligent recognition method based on flow model feature enhancement

Abstract

The present invention provides a single-source domain generalized intelligent recognition method based on flow model feature enhancement, which relates to the technical field of mechanical fault diagnosis. The method includes preprocessing the collected mechanical vibration time domain signal by truncation, length unification, and amplitude normalization, converting it into frequency domain data via fast Fourier transform, and dividing it into a single-source domain dataset and a target domain dataset according to the operating condition; constructing a fault diagnosis training model; using the single-source domain dataset, training the fault diagnosis training model according to a set loss function and optimization algorithm to obtain a trained fault diagnosis training model; constructing a fault recognition model based on the trained fault diagnosis training model, and inputting the target domain dataset into the fault recognition model to identify the fault category. The present invention solves the problems of high-dimensional data generation and interference information in traditional methods, thereby improving the model's generalization ability and the accuracy of target operating condition fault recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and in particular to a single-source domain generalized 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 inherently rely on the independent and identically distributed (IID) assumption. When the operating conditions of the training data and the target data (such as load, speed, or sensor deployment) differ, the target domain data will experience a domain distribution shift, resulting in a sharp decline in mechanical fault identification accuracy. To mitigate domain shift, transfer learning adapts the source and target domains through domain alignment. Some research has introduced adversarial training to obtain domain-invariant features. However, this method requires partially labeled target domain data for training, which is impractical in industrial environments. Moreover, discriminative fault features may be distorted under large domain shifts.

[0003] To this end, domain generalization methods that do not require target domain data have attracted 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 equipment type, working environment, and sampling frequency changes) make feature space alignment difficult. The low correlation between some source domains and target domains leads to weighted complexity and the risk of negative transfer. In addition, the cost of collecting multi-source data in industrial scenarios is high, subject to privacy or protocol restrictions, and the availability of data sets is limited, which restricts its application in complex industrial environments.

[0004] Single-source domain generalization uses only single-source data to build a generalizable model and performs fault identification through single-source feature extraction. This reduces the number of source domains required, reduces the labor and time costs of data collection and labeling, and better meets actual industrial needs. The core of 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 inter-domain invariant features for generalization to different target operating 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 status information but also interference information unrelated to the status, which affects the quality of the generated data. This leads to problems such as high dimensionality of the generated data, complex augmentation networks, weak state representation capabilities, poor generalization performance, and low target operating condition identification accuracy. Summary of the Invention

[0005] To this end, an embodiment of the present invention provides a single-source domain generalized intelligent recognition method based on flow model feature enhancement, which is used to solve the problems in the existing technology of high-dimensional generated data, interference information, low quality, poor generalization performance and low target working condition fault recognition accuracy.

[0006] To solve the above problems, an embodiment of the present invention provides a single-source domain generalized intelligent recognition method based on flow model feature enhancement, including:

[0007] S1: The collected mechanical vibration time domain signal is preprocessed by truncation, length unification and amplitude normalization, converted into frequency domain data through fast Fourier transform, and divided into a single source domain data set and a target domain data set according to the working conditions;

[0008] S2: Construct a fault diagnosis training model, which consists of a deep feature extraction module, a feature generation module, and a fault diagnosis module, wherein:

[0009] The deep feature extraction module is used to extract deep features of the source domain;

[0010] The feature generation module is used to generate extended domain features that are different from the deep feature distribution of the source domain;

[0011] The fault diagnosis module is used to extract inter-domain invariant features and perform category recognition and domain discrimination;

[0012] S3: Using a single-source domain dataset, the fault diagnosis training model is trained according to a set loss function and optimization algorithm to obtain a trained fault diagnosis training model;

[0013] S4: 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 to perform fault category recognition.

[0014] Preferably, the deep feature extraction module is composed of a feature extractor F1 and a classifier C1, wherein the feature extractor F1 is used to input a preprocessed single-source domain dataset and output source domain deep features;

[0015] The classifier C1 is used to input the source domain deep features output by the feature extractor F1 and output a fault category label.

[0016] Preferably, the feature generation module is composed of a binary masker, a low-frequency perturbation module, and a flow model generation module, wherein the binary masker is used to input the source domain deep features output by the feature extractor F1, rearrange the source domain deep features with the low-frequency component 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 reordered feature vector and the binary mask element by element to separate the low-frequency component and the high-frequency component of the source domain deep features;

[0017] The low-frequency perturbation module is used to input the low-frequency component output by the binary masker, calculate the statistical distribution of the low-frequency component, and superimpose Gaussian perturbations to generate a new statistical distribution;

[0018] The flow model generation module is used to input the new statistical distribution output by the low-frequency perturbation module, map the low-frequency component to the new statistical distribution, and generate a new low-frequency component;

[0019] The feature generation module combines the new low-frequency components with the high-frequency components to generate extended domain features that are different from the distribution of deep features in the source domain.

[0020] Preferably, the mask matrix of the binary masker has a central area of ​​1 and two ends of 0, and the length of the central area accounts for 50% of the total length.

[0021] Preferably, the operation of the low-frequency disturbance module includes:

[0022] Determine the number of sub-distributions by the number of categories, and each category has a set of statistical distributions, namely sub-distributions;

[0023] Calculate the variance of the mean and variance of the low-frequency component, multiply it by the disturbance coefficient, and then add it to the original statistical distribution to obtain the new statistical distribution after random disturbance.

[0024] Preferably, the flow model generation module adopts an affine coupling layer structure and optimizes the model by maximizing the log-likelihood loss function. The flow model generation module is constructed by at least one of the NICE model, the Real-NVP model and the GLOW model.

[0025] Preferably, the fault diagnosis module is composed of a feature extractor F2, a classifier C2, and a domain discriminator Y, wherein 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 inter-domain invariant features;

[0026] The classifier C2 is used to input the inter-domain invariant features output by the feature extractor F2 and output the predicted fault category;

[0027] The domain discriminator Y is used to input the inter-domain invariant features output by the feature extractor F2 and output the predicted domain category.

[0028] Preferably, the process of using the single-source domain dataset to train the fault diagnosis training model according to a set loss function and optimization algorithm to obtain a trained fault diagnosis training model includes:

[0029] S31: Input the single-source domain dataset into the deep feature extraction module to 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 are satisfy:

[0030] ;

[0031] Where, is the source domain data, is the corresponding label;

[0032] S32: The source domain deep features Input feature generation module:

[0033] S321: Separate low-frequency components through a binary masker With high frequency components :

[0034] ;

[0035] in, represents the feature reordering operation, represents a binary mask matrix, represents element-wise dot product;

[0036] S322: In the low-frequency perturbation module, low-frequency features are calculated by category The mean and variance of , generate a new statistical distribution:

[0037] ;

[0038] Where, is the fault category index, Low-frequency features The class conditional mean of Low-frequency features The class conditional variance of The mean of each category The variance of is the variance of each category The variance of To sample from a normal distribution with mean 0 and specified variance, is the disturbance coefficient, is the new class conditional mean after disturbance, is the new class conditional variance after disturbance;

[0039] S323: In the flow model generation module, by maximizing the log-likelihood loss Training flow model generation module:

[0040] ;

[0041] Where, Generate modules for streaming models, is the number of batch training samples, Generate modules for streaming models Define the probability density function;

[0042] Generate new low-frequency components , and combined with the high-frequency components to form the extended domain features:

[0043] ;

[0044] Where, is the extended domain feature, It is the feature splicing operation;

[0045] S33: Deep features in the source domain and extended domain features Input fault diagnosis module to minimize the total loss function :

[0046] ;

[0047] Where, To supervise the contrast loss, it promotes the compactness of intra-class features and the separation between classes; It is the binary cross entropy loss of the domain discriminator Y, which distinguishes the source domain / extended domain features; The cross entropy loss of classifier C2 is used to optimize the fault category recognition accuracy; 、 The weighting coefficient for weighing different losses;

[0048] S34: Repeat S31-S33, use the optimization algorithm to update the parameters, 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.

[0049] Preferably, the optimization algorithm is at least one of a root mean square transfer algorithm, a stochastic gradient descent algorithm or an adaptive moment estimation algorithm.

[0050] Preferably, the method for constructing a fault identification model based on the trained fault diagnosis training model is:

[0051] Based on the trained fault diagnosis training model, the trained feature extractor F1, feature extractor F2 and classifier C2 are obtained, and the trained feature extractor F1, feature extractor F2 and classifier C2 are constructed into a fault recognition model.

[0052] It can be seen from the above technical solutions that the present invention has the following beneficial effects:

[0053] (1) Reduce data dimension and computational complexity: Abandon the traditional data enhancement method of generating high-dimensional samples. First, the low-dimensional deep features of the source domain data are extracted through the deep feature extraction module, and then the extended domain features are generated through the feature generation module. This reduces the computational complexity, avoids interference information in high-dimensional data, and makes model training and inference more efficient.

[0054] (2) Improving feature representation capabilities: The feature generation module separates low-frequency and high-frequency components through a 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 health status of the machine, eliminate interference information irrelevant to the status, and enhance the discriminability of the features.

[0055] (3) Enhanced feature diversity and consistency: The low-frequency perturbation module simulates the feature offset caused by changes in real working conditions by adjusting the perturbation coefficient, generating diverse statistical distributions; the flow model generation module aligns the low-frequency features to the sub-distribution of the corresponding category, ensuring that the generated features are diverse and consistent, thereby improving the quality of the extended domain features.

[0056] (4) Improve the generalization ability of the model: The fault diagnosis module extracts inter-domain invariant features through the synergy of supervised contrast loss, domain discriminator and classifier, so that the model can adapt to the domain distribution offset of unknown target working conditions. It has strong generalization ability for mechanical faults under different working conditions and has a wide range of applications.

[0057] (5) Improved fault identification accuracy: Through the dual effects of feature enhancement and inter-domain invariant feature extraction, the model can accurately identify mechanical faults under target operating conditions. Experimental results show that the diagnosis accuracy for various health states is high, and reliable identification can be achieved in complex and compound fault situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the implementation cases of the present invention or the technical solutions in the prior art, the following is a brief description of the drawings required for use 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 understood as limiting the present invention in any way. Those skilled in the art can derive other drawings based on these drawings without inventive effort. Among them:

[0059] Figure 1 A flowchart of a single-source domain generalized intelligent recognition method based on stream model feature enhancement provided by the present invention;

[0060] Figure 2 A schematic diagram of the structure of the fault diagnosis training model provided by the present invention;

[0061] Figure 3 A structural diagram of applying a trained fault diagnosis training model to an unknown target working condition fault recognition model that has not participated in the training is provided by the present invention;

[0062] Figure 4 A schematic diagram of the visualized clustering results of the target domain high-level implicit features output by the feature extractor F2 in the fault recognition model provided by the present invention;

[0063] Figure 5 This is the confusion matrix of the target domain bearing health status results predicted by the fault identification model provided by the present invention. DETAILED DESCRIPTION

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0065] Example 1:

[0066] In order to solve the problems of high dimension, interference information, low quality, poor generalization performance and low accuracy of target working condition fault identification in the existing technology, Figure 1 As shown, the present invention proposes a single-source domain generalized intelligent recognition method based on flow model feature enhancement, including:

[0067] S1: The collected mechanical vibration time domain signal is preprocessed by truncation, length unification and amplitude normalization, converted into frequency domain data through fast Fourier transform, and divided into a single source domain data set and a target domain data set according to the working conditions;

[0068] 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 from the source domain; the feature generation module is used to generate extended domain features that have a different distribution from the deep features in the source domain; and the fault diagnosis module is used to extract inter-domain invariant features and perform category recognition and domain discrimination.

[0069] S3: Using the single-source domain dataset, the fault diagnosis training model is trained according to the set loss function and optimization algorithm to obtain a trained fault diagnosis training model;

[0070] S4: Build a fault recognition model based on the trained fault diagnosis training model, and input the target domain dataset into the fault recognition model to identify the fault category.

[0071] As can be seen from the above technical solution, the present invention proposes a single-source domain generalized intelligent recognition method based on flow model feature enhancement. Step S1 converts the mechanical vibration time-domain signal into frequency-domain data by truncation, length unification, amplitude normalization, and fast Fourier transform. The signal is then divided into single-source and target domain datasets based on the operating conditions. This eliminates noise interference and data format differences during signal acquisition, provides standardized and structured data input for subsequent model training, and ensures the independence and distinguishability of data from different operating 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. This modular integration of deep feature extraction from the source domain, feature generation in the extended domain, and then to inter-domain invariant feature extraction and category / domain discrimination functions lays the model architecture foundation for generalized recognition. Step S3 utilizes a single-source domain dataset and trains the model based on a set loss function and optimization algorithm. Driven by the single-source data, the model learns the distribution patterns of source and extended domain features. Loss function optimization enhances feature representation capabilities and inter-domain invariance, resulting in a generalizable fault diagnosis model. Step S4 constructs a fault recognition model based on the trained fault diagnosis training model. The target domain dataset is then fed into the fault recognition model to identify fault categories. This approach leverages the feature extraction and generalization capabilities acquired during single-source domain training to accurately classify mechanical faults under unknown operating conditions, validating the model's effectiveness in cross-operating condition diagnosis. This approach addresses the challenges of high-dimensional data generation and the inclusion of interfering information in traditional methods, improving both the model's generalization capabilities and the accuracy of target operating condition fault identification.

[0072] In step S1, the present invention performs preprocessing of truncation, length unification and amplitude normalization on the collected mechanical vibration time domain signal, converts it into frequency domain data through fast Fourier transform, and divides it into a single source domain data set and a target domain data set according to the working conditions.

[0073] It should be noted that the dataset is divided according to the different operating conditions of the machinery. The same speed and load are considered the same operating condition. Multiple health status samples under the same operating condition are placed in the same domain, and different domains contain the same machinery health status category. The model is trained using only a single source domain dataset; the target domain is invisible during the model training phase and is only used for model testing, so it does not participate in model training.

[0074] In step S2, the present invention constructs a fault diagnosis training model, such as Figure 2 As shown in the figure, the fault diagnosis training model consists of a deep feature extraction module E, a feature generation module G and a fault diagnosis module D. 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 a different distribution from the deep features of the source domain; the fault diagnosis module D is used to extract inter-domain invariant features and perform category recognition and domain discrimination.

[0075] Specifically, the deep feature extraction module E consists of a feature extractor F1 and a classifier C1. The feature extractor F1 is used to input the preprocessed single-source domain dataset and output the source domain deep features; the classifier C1 is used to input the source domain deep features output by the feature extractor F1 and output the fault category label.

[0076] 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 consists of a fully connected layer and a Softmax classifier, and its label classification loss is the cross entropy loss of the predicted class label of the sample in the labeled source domain. .

[0077] 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 takes the source domain deep features output by the feature extractor F1 and rearranges them with the low-frequency component centered and the high-frequency components at both ends, generating a binary mask of the same length as the feature vector. The rearranged feature vector is element-wise multiplied by the binary mask to separate the low-frequency and high-frequency components of the source domain deep features. The low-frequency perturbation module A takes the low-frequency components output by the binary masker M, calculates their statistical distribution, and superimposes Gaussian perturbations to generate a new statistical distribution (for example, assuming that the mean and variance of a batch of data follow a Gaussian distribution, the variance of the mean and the variance of the variance are calculated, multiplied by the perturbation coefficient, and then added to the mean and variance, respectively, to obtain a new statistical distribution after random perturbation). The flow model generation module R takes the new statistical distribution output by the low-frequency perturbation module A and maps the low-frequency components to the new statistical distribution to generate a new low-frequency component. Finally, the feature generation module G combines the new low-frequency components with the high-frequency components to generate extended domain features that are different from the deep feature distribution of the source domain.

[0078] It should be noted that the mask matrix of the binary masker M has a central region of 1 and two ends of 0, with the central region length accounting for 50% of the total length. The operations of the low-frequency perturbation module A include: determining the number of subdistributions based on the number of categories, with each category having a set of statistical distributions, namely subdistributions; calculating the variance of the mean and variance of the low-frequency component, multiplying it by the perturbation coefficient, and superimposing it on the original statistical distribution to obtain a new statistical distribution after random perturbation. The perturbation coefficient can be adjusted to obtain different statistical distributions. The low-frequency perturbation module A fits the distribution through statistical methods, so no training is required, while reducing the number of model parameters. The flow model generation module R uses an affine coupling layer structure to enhance nonlinear representation capabilities and optimizes the model by maximizing the log-likelihood loss function. The flow model generation module R is constructed from at least one of the NICE model, the Real-NVP model, and the GLOW model.

[0079] Furthermore, the fault diagnosis module D consists of a feature extractor F2, a classifier C2, and a domain discriminator Y. Feature extractor F2 takes in the source domain deep features output by feature extractor F1 and the extended domain features generated by feature generation module G, and outputs inter-domain invariant features. Classifier C2 takes in the inter-domain invariant features output by feature extractor F2 and outputs a predicted fault category. Domain discriminator Y takes in the inter-domain invariant features output by feature extractor F2 and outputs a predicted domain category.

[0080] 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 domain labels predicted by the source domain and the extended domain samples. 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 category labels of 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 separation, and its loss is supervised contrastive loss .

[0081] In step S3, the present invention uses a single-source domain dataset to train a fault diagnosis training model according to a set loss function and optimization algorithm to obtain a trained fault diagnosis training model, including:

[0082] S31: Input the single-source domain dataset into the deep feature extraction module to 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 are satisfy:

[0083] ;

[0084] Where, is the source domain data, For the corresponding label.

[0085] S32: The source domain deep features Input feature generation module:

[0086] S321: Separate low-frequency components through a binary masker With high frequency components :

[0087] ;

[0088] in, represents the feature reordering operation, represents a binary mask matrix, represents element-wise dot product;

[0089] S322: In the low-frequency perturbation module, low-frequency features are calculated by category The mean and variance of , generate a new statistical distribution:

[0090] ;

[0091] Where, is the fault category index, Low-frequency features The class conditional mean of Low-frequency features The class conditional variance of The mean of each category The variance of is the variance of each category The variance of To sample from a normal distribution with mean 0 and specified variance, is the disturbance coefficient, is the new class conditional mean after disturbance, is the new class conditional variance after disturbance;

[0092] S323: In the flow model generation module, by maximizing the log-likelihood loss Training flow model generation module:

[0093] ;

[0094] Where, Generate modules for streaming models, is the number of batch training samples, Generate modules for streaming models Define the probability density function;

[0095] Generate new low-frequency components , and combined with the high-frequency components to form the extended domain features:

[0096] ;

[0097] Where, is the extended domain feature, It is a feature splicing operation.

[0098] S33: Deep features in the source domain and extended domain features Input fault diagnosis module to minimize the total loss function :

[0099] ;

[0100] Where, To supervise the contrast loss, it promotes the compactness of intra-class features and the separation between classes; It is the binary cross entropy loss of the domain discriminator Y, which distinguishes the source domain / extended domain features; The cross entropy loss of classifier C2 is used to optimize the fault category recognition accuracy; 、 is the weighting coefficient for weighing different losses.

[0101] S34: Repeat S31-S33, use the optimization algorithm to update the parameters, 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.

[0102] It should be noted that the optimization algorithm is at least one of a root mean square transfer algorithm, a stochastic gradient descent algorithm, or an adaptive moment estimation algorithm.

[0103] 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 to perform fault category recognition.

[0104] Specifically, based on the trained fault diagnosis training model, the trained feature extractor F1, feature extractor F2 and classifier C2 are obtained, and the trained feature extractor F1, feature extractor F2 and classifier C2 are constructed into a fault recognition model.

[0105] In order to more clearly understand the technical solution and effects of the present invention, a detailed description is given below in conjunction with a specific experiment.

[0106] 1. Experimental data:

[0107] Experimental validation was conducted on a train axlebox bearing dataset. This dataset covers 11 health status data, including normal state (N), inner race fault (I), outer race fault (O), rolling element fault (R), cage fault (C), inner race and outer race combined fault (IO), outer race and rolling element combined fault (OR), outer race and cage combined fault (OC), inner race and outer race rolling element combined fault (IOR), inner race and outer race cage combined fault (IOC), and inner race and outer race rolling element and cage combined fault (IORC). These faults are assigned fault category labels ranging from 0 to 10. The experiment used data from nine operating conditions, with data from the same operating condition constituting a domain. The number of samples from each domain across different health statuses was 312. During the experiment, one of the labeled source domains listed in Table 1 was selected as the training dataset, and any of the remaining domains was used as the target domain test dataset.

[0108] Table 1 Description of experimental bearing dataset

[0109]

[0110] 2. Experimental steps:

[0111] Step 1: First, the collected mechanical vibration time domain signal is intercepted, the sample length is unified to 2048 points, and the sample amplitude is normalized in the interval [0,1]. Then, the time domain signal is converted into frequency domain data through fast Fourier transform. After conversion, the sample length becomes 1024. Finally, the data set is divided according to the working conditions. The frequency domain data under the same speed and load conditions are classified into the same domain. One domain is used as the single-source domain training data set, and the other domain is used as the target domain test data set. The preprocessed frequency domain data is directly used as the model input sample.

[0112] 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. Figure 2 As shown.

[0113] The deep feature extraction module E consists of a feature extractor F1 and a classifier C1. The feature extractor F1 uses four layers of one-dimensional convolutional layers, each followed by a batch normalization layer and a ReLU activation function. It takes the preprocessed 1024-dimensional frequency domain signal as input and outputs a 128-dimensional deep feature vector. The classifier C1 is a fully connected network with a Softmax activation function. It takes the output of the feature extractor F1 as input and outputs an 11-dimensional health status category vector. It also calculates the fault category cross entropy loss of the source domain sample based on the label output probability. .

[0114] 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 center region (50% of the total length) of the binary masker M is set to 1, with both ends set to 0. The deep features of the reordered, centered low-frequency components are dot-producted with 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 (one sub-distribution for each category) using the perturbation coefficients. The flow model generation module R uses a Real-NVP structure with four affine coupling layers to map the low-frequency components to the target sub-distribution, which is then combined with the original high-frequency components to generate extended domain features.

[0115] The fault diagnosis module D consists of a feature extractor F2, a classifier C2, and a domain discriminator Y. Feature extractor F2 uses four layers of one-dimensional convolution to perform upsampling (from 128 dimensions to 2048 dimensions) and downsampling (from 2048 dimensions to 128 dimensions), outputting 128-dimensional domain-invariant features. Classifier C2 is a fully connected network with a Softmax activation function, outputting an 11-dimensional health status category vector and calculating the cross-entropy loss between source and extended domain samples. The domain discriminator Y is a two-layer fully connected network (hidden layer dimensions 64 and 1), which is connected to the 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. .

[0116] Step 3: Use the single source domain dataset to train the fault diagnosis training model. The specific steps are as follows: Input the single source domain dataset into the constructed model and implement the 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, minimize the cross entropy loss function of the classifier C1. To optimize the feature extractor F1; then, in the feature generation module, the source domain deep features output by the deep feature extraction module are sequentially passed through the binary masker M and the low-frequency perturbation module A to determine the sub-distribution of all categories, and then the maximum likelihood estimation is used. The convection model generation module R is optimized; in the fault diagnosis module, the total loss function of the model is set as , where the supervised contrast loss By comparing positive and negative sample pairs, similar features are clustered together, and the classifier C2 is trained through cross entropy loss. To ensure the accuracy of category recognition, the domain discriminator Y is trained through binary cross entropy loss The goal of the total loss is to map features to a specific feature space, enabling classifier C2 to distinguish features from the source domain and the unknown target domain. The optimization algorithm uses the Adaptive Moment Estimation (Adam) algorithm with a learning rate of 0.001. After 200 iterations of training, the training process ends when the model objective function loss stabilizes.

[0117] 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 the model, and complete the fault category identification. The specific process is shown in Figure 3.

[0118] 3. Experimental analysis:

[0119] like Figure 4As shown in the figure, the feature extraction results of the rolling bearing data samples in the target test domain are visualized using t-SNE technology, where different shapes represent 11 fault categories (category information is marked in Figure 4 The visualization results show that features of the same category in the target domain exhibit significant clustering, while features of different categories have clear boundaries and strong discrimination. This verifies that the proposed method can learn feature representations that are independent of data distribution and extract effective features that are more conducive to classification, achieving efficient extraction of invariant features between domains.

[0120] The confusion matrix of the diagnosis results of the target domain test data set by the method of the present invention is as follows: Figure 5 As shown in the figure, experimental data show that the method has a 100% diagnostic accuracy for target domain data in eight health states, with only a few misjudgments for the outer ring plus cage composite fault (OC, label 7). The recognition accuracy for bearing rolling element fault (R, label 3) and inner ring plus outer ring composite fault (IO, label 5) exceeds 90%, fully demonstrating its excellent fault recognition performance under unknown target conditions.

[0121] In summary, the present invention realizes low-dimensional feature enhancement by constructing a deep feature extraction module E and a feature generation module G. While reducing computational complexity and eliminating noise interference, it effectively generates diverse extended domain features containing category information, significantly improving the model generalization ability. By establishing a fault diagnosis module D, the efficient extraction of inter-domain invariant features is further realized, ultimately enabling the single-source domain generalization model to accurately identify mechanical faults under unknown working conditions.

[0122] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0123] The present application is described with reference to the flowcharts and / or block diagrams of the 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 implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0124] These computer program instructions may 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 produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 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 produce a computer-implemented process, thereby providing instructions for implementing the process in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0125] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications derived therefrom remain within the scope of protection of the present invention.

Claims

1. A single-source domain generalized intelligent recognition method based on flow model feature enhancement, characterized by: include: S1: The collected mechanical vibration time domain signal is preprocessed by truncation, length unification and amplitude normalization, converted into frequency domain data through fast Fourier transform, and divided into a single source domain data set and a target domain data set 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, wherein: 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 that are different from the deep feature distribution of the source domain; The fault diagnosis module is used to extract inter-domain invariant features and perform category recognition and domain discrimination; S3: Using the single-source domain dataset, training the fault diagnosis training model according to a set loss function and optimization algorithm to obtain 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 classifier C1 Optimize the feature extractor F1 so that the deep features of the source domain extracted by the feature extractor F1 are satisfy: ； Where, is the source domain data, is the corresponding label; S32: The source domain deep features Input feature generation module: S321: Separate low-frequency components through a binary masker With high frequency components : ； in, represents the feature reordering operation, represents a binary mask matrix, represents element-wise dot product; S322: In the low-frequency perturbation module, low-frequency features are calculated by category The mean and variance of , generate a new statistical distribution: ； Where, is the fault category index, Low-frequency features The class conditional mean of Low-frequency features The class conditional variance of The mean of each category The variance of is the variance of each category The variance of To sample from a normal distribution with mean 0 and specified variance, is the disturbance coefficient, is the new class conditional mean after disturbance, is the new class conditional variance after disturbance; S323: In the flow model generation module, by maximizing the log-likelihood loss Training flow model generation module: ； Where, Generate modules for streaming models, is the number of batch training samples, Generate modules for streaming models Define the probability density function; Generate new low-frequency components , and combined with the high-frequency components to form the extended domain features: ； Where, is the extended domain feature, It is the feature splicing operation; S33: Deep features in the source domain and extended domain features Input fault diagnosis module to minimize the total loss function : ； Where, To supervise the contrast loss, it promotes the compactness of intra-class features and the separation between classes; It is the binary cross entropy loss of the domain discriminator Y, which distinguishes the source domain / extended domain features; The cross entropy loss of classifier C2 is used to optimize the fault category recognition accuracy; 、 The weighting coefficient for weighing different losses; S34: Repeat S31-S33, use the optimization algorithm to update the parameters, 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; S4: 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 to perform fault category recognition.

2. The single-source domain generalized intelligent recognition method based on flow model feature enhancement according to claim 1 is characterized in that: The deep feature extraction module is composed of a feature extractor F1 and a classifier C1, wherein the feature extractor F1 is used to input a preprocessed single source domain dataset and output source domain deep features; The classifier C1 is used to input the source domain deep features output by the feature extractor F1 and output a fault category label.

3. The single-source domain generalized intelligent recognition method based on flow model feature enhancement according to claim 1, characterized in that: The feature generation module consists of a binary masker, a low-frequency perturbation module, and a flow model generation module, wherein the binary masker is used to input the source domain deep features output by the feature extractor F1, rearrange the source domain deep features with the low-frequency component in the middle and the high-frequency components at both ends, and generate a binary mask of the same length as the feature vector. The rearranged feature vector is element-wise multiplied by the binary mask to separate the low-frequency component and the high-frequency component of the source domain deep features; The low-frequency perturbation module is used to input the low-frequency component output by the binary masker, calculate the statistical distribution of the low-frequency component, 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 component to the new statistical distribution, and generate a new low-frequency component; The feature generation module combines the new low-frequency components with the high-frequency components to generate extended domain features that are different from the distribution of deep features in the source domain.

4. The single-source domain generalized intelligent recognition method based on flow model feature enhancement according to claim 3 is characterized in that: The center area of ​​the mask matrix of the binary masker is 1 and both ends are 0, and the length of the center area accounts for 50% of the total length.

5. The single-source domain generalized intelligent recognition method based on flow model feature enhancement according to claim 3 is characterized in that: The operation of the low-frequency disturbance module includes: Determine the number of sub-distributions by the number of categories, and each category has a set of statistical distributions, namely sub-distributions; Calculate the variance of the mean and variance of the low-frequency component, multiply it by the disturbance coefficient, and then add it to the original statistical distribution to obtain the new statistical distribution after random disturbance.

6. The single-source domain generalized intelligent recognition method based on flow model feature enhancement according to claim 3 is characterized in that: The flow model generation module adopts an affine coupling layer structure and optimizes the model by maximizing the log-likelihood loss function. 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 generalized intelligent recognition method based on flow model feature enhancement according to claim 1 is characterized in that: The fault diagnosis module is composed of a feature extractor F2, a classifier C2, and a domain discriminator Y, wherein 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 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 category; The domain discriminator Y is used to input the inter-domain invariant features output by the feature extractor F2 and output the predicted domain category.

8. The single-source domain generalized intelligent recognition method based on flow model feature enhancement according to claim 1 is characterized in that: The optimization algorithm is at least one of a root mean square transfer algorithm, a stochastic gradient descent algorithm, or an adaptive moment estimation algorithm.

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

Images