Aero-engine bearing fault diagnosis method based on clustering enhancement domain generalization

Abstract

The invention belongs to the technical field of intelligent fault diagnosis, and discloses an aero-engine bearing fault diagnosis method based on clustering enhancement domain generalization. Firstly, a time domain vibration signal is converted into frequency domain representation through fast Fourier transform, a two-stage convolutional neural network architecture comprising a domain alignment encoder and a classification encoder is constructed, and cross-working-condition feature alignment under adversarial training is realized by using a domain discriminator; introducing a multi-source domain maximum mean difference statistical alignment strategy and a clustering enhanced triple loss mechanism, and synchronously optimizing intra-class feature compactness and inter-class separability through pseudo label clustering center anchoring and dynamic interval constraint; and gradient inversion adversarial training and multi-modal feature fusion are combined. The method provided by the invention can effectively improve the accuracy and robustness of generalization judgment when the aero-engine bearing fault diagnosis lacks target domain data.

Description

Technical Field

[0001] The present invention belongs to the field of aero-engine fault detection, relates to the technology of engine rolling bearing fault diagnosis, and particularly relates to a method for diagnosing aero-engine bearing faults based on clustering-enhanced domain generalization. Background Art

[0002] As the core power system of modern aircraft, the accuracy and reliability of aero-engine fault diagnosis are the keys to ensuring flight safety. With the evolution of engines towards high thrust-to-weight ratio and long life, the structural complexity and operating condition diversity have increased significantly, and the fault modes show characteristics of multi-source, coupling and gradualness. Traditional diagnostic methods rely on expert experience and threshold alarm mechanisms, and it is difficult to meet the real-time analysis requirements of massive multi-source heterogeneous data; while the intelligent diagnostic technology based on deep learning can mine the deep features of data, but it is limited by practical problems such as scarce fault samples and cross-domain data distribution differences, resulting in insufficient generalization performance of the model in actual deployment, especially the diagnostic accuracy drops sharply under new models of engines or extreme operating conditions.

[0003] The current mainstream cross-domain diagnostic methods mainly improve the adaptability of the model by reducing the data distribution difference between the source domain and the target domain, but these methods still face the following bottlenecks: First, the global alignment strategy pays too much attention to the overall distribution matching between domains and ignores the category characteristics of fault features, resulting in an unclear clustering structure of the same type of faults in the feature space, thus reducing the cross-domain classification accuracy; Second, in domain adversarial training, the adversarial optimization between the domain discriminator and the feature extractor is prone to fall into local optimal solutions, which not only exacerbates the contradiction between domain invariance and class separability in the feature space, but also further weakens the reliability of the model.

[0004] In view of the above technical defects, future research should focus on the following directions: In the process of feature alignment, not only the global distribution matching between domains should be achieved, but also the feature alignment at the category level should be strengthened to ensure that the same type of faults has a clear clustering structure in the feature space; at the same time, optimize the adversarial training mechanism to avoid the domain discriminator and the feature extractor falling into local optimal solutions, so as to alleviate the contradiction between domain invariance and class separability. Summary of the Invention

[0005] Aiming at the problems existing in the fault diagnosis model: ignoring the category characteristics of fault features leads to a fuzzy clustering structure of the same type of faults in the feature space, and being prone to fall into local optimal solutions during the adversarial optimization process, the present invention provides a method for diagnosing aero-engine bearing faults based on clustering-enhanced domain generalization.

[0006] The technical solution of the present invention:

[0007] A method for diagnosing aero-engine bearing faults based on clustering-enhanced domain generalization includes the following steps:

[0008] Step 1, vibration signal data processing;

[0009] Collect time-series vibration signals and use to represent the dataset of time-series vibration signals, to represent the time-series vibration signals of the th source domain, to represent the total number of source domains; segment the time-series vibration signals to obtain time-series vibration signal samples, to represent the set of time-series vibration signal samples of the th source domain after segmentation, where the th source domain set contains time-series vibration signal samples with corresponding health status annotations , to represent the domain label of the th source domain;

[0010] Subsequently, perform a fast Fourier transform on the time-series vibration signal samples to convert the time-series vibration signal samples into frequency-domain signal samples :

[0011]

[0012] where, represents the frequency index, is the total number of frequency points, is the sampling point index in the time domain, is the total number of sampling points of the time-series signal samples; decompose the time-domain signal samples into frequency-domain signal samples of different frequency components ;

[0013] Step 2, construction of a fault diagnosis model clustering-enhanced domain generalization network;

[0014] The constructed fault diagnosis model clustering-enhanced domain generalization network consists of two cascaded feature encoders, a classifier, and a domain discriminator;

[0015] Step 2.1, construction of a feature encoder;

[0016] The feature encoder uses a convolutional neural network to extract features with rich time-frequency information from the frequency-domain signal samples , and its expression is:

[0017]

[0018] Among them, is the value of the output feature map of the -th layer at position . represents the input of the -th layer at position . For the first convolutional neural network layer of the feature encoder, that is, when , its input is the frequency-domain signal sample, that is, is the convolution operation, is the -th convolution kernel of the -th layer, is the width of the convolution kernel, is the -th weight value of the

[0019] The feature encoder consists of two parts connected in series, namely the domain alignment encoder and the classification encoder . The domain alignment encoder extracts the cross-condition domain alignment features of the frequency-domain signal samples, and uses the MMD method and the domain discriminator to achieve distribution alignment between different domains; the classification encoder further extracts the high-level abstract classification features of the domain alignment features on the basis of the domain alignment features , and directly uses them for the classifier to achieve accurate classification of the fault state;

[0020] Step 2.2, construct the classifier and the domain discriminator;

[0021] The classifier is located at the backend of the fault diagnosis model clustering enhanced domain generalization network architecture, and adopts a fully connected network, expressed as:

[0022]

[0023] Among them, is the value of the -th output neuron of the -th layer, is the value of the -th input neuron of the -th layer, is the -th input neuron of the -th layer and the -th The weights between the output neurons, is the bias value of all neurons in the layer for the th output neuron in the is the total number of neurons in the layer;

[0024] Classifier Performs classification prediction on the frequency-domain signal samples based on the classification features extracted by the classification encoder :

[0025]

[0026] Among them, represents the probability that the category belongs to , represents the category prediction value of the frequency-domain signal sample corresponding to the classification feature , represents the logit value of the frequency-domain signal sample corresponding to the classification feature belonging to the th category. The classification loss is obtained by calculating the cross-entropy loss based on the category prediction value ;

[0027] Domain discriminator Based on a fully connected network, it determines the source domain from which the frequency-domain signal sample corresponding to the input domain alignment feature comes;

[0028]

[0029] Among them, represents the probability that the source domain belongs to , represents the domain prediction value of the frequency-domain signal sample, represents the logit value of the frequency-domain signal sample corresponding to the domain alignment feature belonging to the th source domain. The domain classification loss is obtained by calculating the cross-entropy loss based on the domain prediction value ;

[0030] A gradient reversal layer is set before the domain discriminator. The gradient reversal layer acts as an identity mapping during forward propagation, while reversing the gradient direction during backpropagation, thereby promoting the feature extractor to learn a domain-invariant domain alignment feature representation through adversarial training and achieving implicit alignment of the domain alignment feature distributions of each source domain;

[0031] Step 2.3, construct multi-source domain feature alignment;

[0032] The feature encoder outputs domain - aligned features After that, the frequency - domain signal samples in different source domains are grouped according to their source domains, and the maximum mean discrepancy (MMD) between their domain - aligned features is calculated; MMD is used to measure the difference between different distributions, and its basic idea is to calculate the distance between the mean embeddings of two distributions in the reproducing kernel Hilbert space; if the means of the two source - domain feature distributions are significantly different, it indicates that there are large differences in their distributions; on the contrary, it indicates that the two source - domain distributions are relatively close; finally, the sum of MMDs between all source domains is used as the multi - source domain confusion loss;

[0033]

[0034] Among them, represents the MMD calculation; represents the number of different source domains in this batch of data; 、 denote the sets of domain - aligned features from two different source domains;

[0035] Step 2.4, construct the clustering - enhanced triplet loss;

[0036] During the fault diagnosis process, to enhance the intra - class compactness and inter - class separability, the fault diagnosis model clustering - enhanced domain generalization network adopts the clustering - enhanced triplet loss; first, use the clustering algorithm to cluster the classification features After clustering, generate corresponding pseudo - labels for the classification features of the frequency - domain signal samples. Take the label with the largest proportion of the true labels of the frequency - domain signal samples in this cluster as the same pseudo - label for the cluster, so as to initially reflect the clustering structure and potential class information of the classification features, that is:

[0037]

[0038] Among them, is the generated shared pseudo - label for the cluster , is the set of all possible true labels, the newly generated clustering sample is , represents the th cluster sample in the th cluster, represents the pseudo - label of the

[0039] Select the classification features of the clustering samples whose pseudo - labels are consistent with the true labels, that is , and use the clustering center of the classification features of these high - confidence samples as the anchor points, represents the The number of high-confidence samples in each cluster is used to construct the triplet loss;

[0040]

[0041] Among them, , , represents the cluster The -th high-confidence sample's classification feature of the cluster, is a hyperparameter used to control the minimum margin between positive and negative samples;

[0042] Finally, the cluster-enhanced triplet loss is expressed as:

[0043]

[0044] Among them, represents the number of high-confidence samples;

[0045] Step 2.5, construct the total loss function;

[0046] The classification loss, domain classification loss, multi-source domain confusion loss, and cluster-enhanced triplet loss are weighted and summed to obtain the final comprehensive loss function:

[0047]

[0048] Among them , , represent weight parameters, and the model parameters are updated through the Adam optimizer.

[0049] In this way, the sample features of the same category are pulled closer, making them tightly clustered around the cluster center, enhancing the intra-class compactness; at the same time, the sample features of different categories are pushed farther apart, increasing the inter-class distance, improving the distinguishability of the feature space, and making the fault categories more clearly separable in feature representation.

[0050] The beneficial effects of the present invention: Feature extraction is performed through a multi-layer convolutional neural network, combined with the cluster-enhanced triplet loss mechanism, to accurately capture the features in the vibration signal, thereby significantly improving the accuracy of fault diagnosis. At the same time, domain adversarial training and multi-source domain feature alignment strategies are introduced to effectively reduce the differences in data distributions under different working conditions, further enhancing the generalization ability of the model in various working condition environments.

[0051] The present invention proposes an aircraft engine bearing fault diagnosis method based on clustering enhanced domain generalization. The method proposed in the present invention realizes the time-frequency domain conversion of the signal by collecting vibration signal data under different working conditions and performing fast Fourier transform on it. In terms of model construction, a multi-layer convolutional neural network is constructed as a feature encoder to deeply explore the time-frequency characteristics of the vibration signal; at the same time, a classifier is constructed to achieve accurate fault classification, and a domain discriminator is set to carry out domain adversarial training to enhance the cross-working condition adaptability of the model. In addition, a triple loss mechanism of multi-source domain feature alignment and clustering enhancement is introduced to further optimize the model performance and improve the accuracy and robustness of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of the fault diagnosis method for aircraft engine rolling bearings based on clustering enhanced domain generalization.

[0053] Figure 2 Schematic diagram of the clustering enhanced domain generalization network structure.

[0054] Figure 3 is the loss curve during model training.

[0055] Figure 4 This is the accuracy curve during model training.

[0056] Figure 5 This is a visualization of the clustering of samples during model training.

[0057] Figure 6 Visualization of reliable sample clustering and its cluster centers during model training. DETAILED DESCRIPTION

[0058] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0059] Aircraft engine bearing fault diagnosis process Figure 1 As shown, the clustering enhanced domain generalization network structure is as follows Figure 2 As shown, the specific implementation steps of each step are described in detail below.

[0060] Step 1: Use Fast Fourier Transform to process vibration signal data

[0061] First, we collected time series vibration signal data from the same engine bearing under different working conditions. These data cover the vibration characteristics of the machine under various operating conditions. In order to facilitate subsequent analysis and classification, the fault conditions are clearly divided into three categories, namely normal condition (NM), inner race fault (IF) and outer race fault (OF). At the same time, since the data comes from different working conditions, in order to distinguish the sources of these data, the data from different working conditions are labeled with domain labels.

[0062] Use to represent a data set, where represents the time series vibration signal data of the th source domain, which contains time series vibration signals , and the working conditions corresponding to the signals . is the set of all possible true labels. Among them has three values of 0, 1, and 2, corresponding to three types of faults: NM, IF, and OF respectively. represents the domain label of the th domain data.

[0063] The original time series vibration signal data is segmented into multiple windows with a fixed length of 4096. The data within each window is used as an independent time series vibration signal sample unit. For each window, corresponding fault labels and domain labels are generated to ensure the integrity and accuracy of the samples.

[0064] Segment the time series vibration signal to obtain time series vibration signal samples, represents the set of time series vibration signal samples of the th domain after segmentation. Among them, the th domain contains time series vibration signal samples with corresponding health status annotations .

[0065] Subsequently, perform FFT transformation on the time series vibration signal samples to decompose the complex time series vibration signal samples into frequency domain signal samples of different frequency components :

[0066]

[0067] Among them, represents the frequency index, is the total number of frequency points, is the sampling point index in the time domain, is the total number of sampling points of the time series signal sample.

[0068] Step 2, construct a feature encoder based on a convolutional neural network

[0069] The core architecture of the feature encoder adopts a Convolutional Neural Network (CNN) to extract the time-frequency features of vibration signals. CNN can automatically learn the local features and global structure of signals to enhance the feature expression ability. The feature encoder consists of cascaded multiple convolutional layers and pooling layers, gradually extracting deep features. The convolutional layer extracts local features, such as vibration patterns within a specific frequency range, which can be expressed as

[0070]

[0071] where is the value of the output feature map of the -th layer at position , represents the input of the -th layer at position . For the first-layer convolutional neural network of the feature encoder, that is when , its input is the frequency-domain signal sample, that is , is the convolution operation, is the -th convolutional kernel of the -th layer, is the width of the convolutional kernel, is the -th weight of the

[0072] The activation layer enhances the network's expression ability through non-linear mapping and consists of activation functions. The convolutional layer and the fully connected layer perform linear operations, and then the activation layer is added for non-linear transformation to improve the network's learning and processing ability for complex data. The present invention adopts Leaky ReLU as the activation function:

[0073]

[0074] where is the activation value of the input feature at position , is the slope of the negative part.

[0075] The pooling layer effectively compresses the feature dimension through downsampling, achieving key feature screening while reducing the model's computational complexity and the number of parameters. The present invention uses the max pooling operation

[0076]

[0077] where is the input feature in the pooling window Activation value

[0078] The feature encoder of the present invention consists of a 6-layer convolutional neural network. The domain alignment encoder is composed of the first 4 convolutional layers, and outputs domain alignment features across working conditions for the domain discriminator to judge the feature source, realize domain alignment, and reduce the influence of working condition differences on fault diagnosis. The classification encoder is composed of the last 2 convolutional layers, and extracts high-level classification features for the classifier to perform fault identification. The classifier is based on and combines a preset classification algorithm to divide the vibration signal into fault states

[0079] Step 3: Construct a classifier based on a fully connected layer

[0080] Classifier is located at the backend of the model and outputs the probability distribution of the fault category according to the features extracted by the feature encoder Its structure is a fully connected network, composed of multiple layers of neurons, fully connected between layers, and learns the mapping relationship between input features and fault categories

[0081]

[0082] Among them, is the value of the th output neuron of the th layer, is the value of the th input neuron of the th layer, is the weight value between the th input neuron of the th layer and the th output neuron of the

[0083] The classifier calculates through the activation function and the weight matrix, and classifies and predicts the input classification features for classification prediction

[0084]

[0085] Among them, represents the probability that the category belongs to , represents the classification feature The predicted class value of the corresponding frequency-domain signal sample represents the logit value that the frequency-domain signal sample corresponding to the classification feature belongs to the th class.

[0086] During the training process, the classifier parameters are adjusted by an optimization algorithm to minimize the classification error and improve the diagnostic accuracy. In this design, the cross-entropy loss

[0087]

[0088] where is the logits processed by the Softmax function, and its target distribution is a one-hot type vector , with only the position corresponding to the target class being 1 and the rest being 0, that is, when the class of the target is , otherwise . is the input batch size.

[0089] After the training is completed, the classifier outputs the probability of each fault class, providing a basis for the fault diagnosis decision.

[0090] Table 1 Parameters of each part of the model

[0091]

[0092] Step 4, construct the domain discriminator

[0093] The domain discriminator is a component of the gradient reversal adversarial training, with a structure similar to the classifier, responsible for judging the source domain of the input domain-aligned features . Among them, is the input batch size, and the output is the domain label prediction of the frequency-domain signal sample corresponding to the domain-aligned features . represents the probability that its source domain belongs to . The domain classification loss uses the cross-entropy loss:

[0094]

[0095] where is the logits processed by the Softmax function, and its target distribution is a one-hot type vector .

[0096] In the present invention, a gradient reversal layer is provided at the front end of the discriminator to achieve implicit alignment of the domain-aligned feature distributions. The domain discriminator is used to distinguish the source domain of the features. During forward propagation, the gradient reversal layer directly passes the input; during backpropagation, the gradient sign is reversed, enabling the feature encoder to learn domain-invariant features that are helpful for fault classification and can "fool" the domain discriminator.

[0097] During the training process, the domain discriminator continuously learns the differences in features from different source domains to improve its discrimination ability; at the same time, its reverse gradient acts through the gradient reversal layer to guide the feature encoder to learn domain-invariant features, promoting the alignment of features between the source domain and the target domain, and enhancing the cross-condition adaptation ability of the model.

[0098] Step 5, construct a multi-source domain feature confusion loss

[0099] The present invention adopts a multi-dimensional statistical alignment strategy, calculates the MMD between source domains as the alignment method, and quantifies the distribution difference based on hypothesis testing and kernel methods:

[0100]

[0101] where and represent the sets of domain-aligned features from two different source domains, and represent the numbers of domain-aligned features from the and domains respectively in this batch of data, represents the th domain-aligned feature from represents the kernel function, and in this design, the kernel functions are all Gaussian kernel functions.

[0102] During training, data from 3 source domains are input, mixed, and then sent to the feature extractor. After passing through the domain-aligned encoder the domain-aligned features are extracted. Subsequently, according to the domain labels they are grouped, and the MMD between different source domains is calculated. The sum of the MMDs between pairwise source domains is used as the multi-source domain confusion loss

[0103]

[0104] where represents the number of source domains included in this batch of data.

[0105] Step 6, construct a clustering-enhanced triplet loss

[0106] During the fault diagnosis process, to enhance the intra-class compactness and inter-class separability, the present invention adopts a clustering-enhanced triplet loss mechanism.

[0107] First, the KMeans algorithm is used to encode the categorical features Clustering. The process of KMeans algorithm is: randomly select The classification features are used as the centers of the initial clusters , is the number of categories contained in the classification feature, that is, the number of clusters. , assigning it to the cluster center closest to it ,Right now: ,in Indicates Classification features To cluster center In each iteration, the center of each cluster is recalculated , i.e. cluster The mean of all points in , repeat the iteration until the cluster center no longer changes or changes very little, and the process ends.

[0108] After the Kmeans algorithm, the corresponding pseudo labels are generated for the classification features of the frequency domain signal samples after clustering. The labels with the largest proportion of real labels of the classification features in the cluster are regarded as the same cluster to share the same pseudo labels, so as to preliminarily reflect the clustering structure and potential category information of the classification features, that is,

[0109]

[0110] in, For clustering Generate shared pseudo labels, is a set of all possible true labels, and the newly generated clustering samples are , Indicates The first The classification features of cluster samples, Indicates Pseudo labels for clusters.

[0111] Next, the classification features of cluster samples whose pseudo labels are consistent with the true labels are screened, that is, , with the cluster centers of the classification features of these high-confidence samples As an anchor point, Indicates The number of high confidence samples in a cluster, constructing triple loss

[0112]

[0113] in , , Indicates the cluster to which it belongs The classification features of the th high-confidence sample,

[0114] Finally, the cluster-enhanced triplet loss is expressed as

[0115]

[0116] where represents the number of high-confidence samples.

[0117] Step 7, construct the loss function

[0118] During the forward propagation of the model, the classification loss, domain classification loss, multi-source domain confusion loss, and cluster-enhanced triplet loss are weighted and summed to obtain the final comprehensive loss function

[0119]

[0120] Step 8, model training

[0121] The training process of the present invention mainly includes the following key steps to ensure that the model can efficiently and stably learn the feature representation and classification ability suitable for cross-condition fault diagnosis:

[0122] The training process consists of multiple epochs, and each epoch processes a batch of data. In each training step, the following operations are specifically performed:

[0123] (1) Data preparation: Randomly shuffle the source domain data, obtain a new batch of data, and pass it together with the corresponding labels into the model for training.

[0124] (2) Forward propagation: The data passes through the encoder and classifier in sequence to generate prediction results.

[0125] (3) Loss calculation: Calculate the cross-entropy loss to evaluate the accuracy of classification prediction; the clustering triplet loss to make samples of the same class gather more closely in the feature space and separate samples of different classes; the domain confusion loss to introduce adversarial training between source domains to enhance the cross-domain adaptation ability of the model; the domain classification loss to train the adversarial learning of the model.

[0126] (4) Backward propagation: Use the Adam optimizer to update the parameters of each network module according to the calculated gradients, and the gradient reversal layer reverses the gradient direction of the domain discriminator during backward propagation.

[0127] Dataset composition:

[0128] The training data is from the bearing and gear dataset of Shandong University of Science and Technology. The experiments were carried out on the fault implantation test bench for rotating machinery transmission components at Shandong University of Science and Technology. Fault diagnosis experiments of rolling bearings and gears under different working conditions were carried out. The test components were 6205 bearings and planetary gearboxes. Two three-axis piezoelectric acceleration sensors were fixed to the fixed brackets at both ends of the bearing housing and the planetary gearbox through magnetic seats respectively to obtain the vibration signals of the bearings and the gearbox. In the experiment, a steady-state test mode was set, the sampling duration was 40 s each time, and the sampling frequency was 25.6 kHz. In each sampling, vibration signals of a total of 6 channels including two three-axis piezoelectric acceleration sensors were obtained. The experiment selected part of the data of the 6205 bearing dataset under constant working conditions as the experimental data. Each working condition included a set of healthy data, and one set of inner race fault data of the bearing with 0.2 mm, 0.4 mm, and 0.6 mm respectively, and one set of outer race fault data of the bearing with 0.2 mm, 0.4 mm, and 0.6 mm respectively. The vibration signals were segmented, and the length of each segmented vibration signal sample was 4096, and 256 samples could be obtained for each data. 1792 samples could be obtained for each working condition. 500 samples were selected as the training set according to the health status ratio for the data used as the source domain, and 200 samples were selected as the test set. 200 samples were selected as the test set according to the health status ratio for the data used as the target domain.

[0129] Model Training and Test Result Analysis

[0130] The experimental settings are shown in Table 3, including two groups of constant load variable speed generalization experiments, two groups of constant speed variable load generalization experiments, and two groups of variable speed variable load generalization experiments. Among them, the parameter was set to 0.5, 1, and 3 respectively, and the learning rate was 0.0001. Table 4 shows the accuracy comparison results of the CEDGN method and the other two methods.

[0131] Among them, the ERM method conducts generalization experiments based on the CNN network alone. The MMD method introduces multi-source domain MMD calculation into the CNN structure, and the DANN method introduces domain adversarial training into the CNN structure. All experiments are repeated 30 times, and the classification accuracy of the optimal model is selected for comparison. The experimental results show that the introduction of multi-source domain MMD loss and domain adversarial training does not necessarily lead to an improvement in accuracy. For example, in groups A, C, D, and F of the experiments, the accuracy decreases. This is because MMD is mainly based on mean matching and fails to fully capture the complex characteristics of the data distribution. In groups E and F of the experiments, domain adversarial training mainly aims at the overall alignment of feature distributions and may ignore the fine-grained matching of local features. In contrast, the CEDGN method combines the domain adversarial mechanism and learns a more flexible non-linear mapping through end-to-end adversarial training to optimize the consistency of feature distributions and improve the adaptability of the model in complex scenarios to make up for the deficiencies of the MMD method. The MMD method ensures that the feature distributions in different domains are closer in both local and global aspects through kernel methods, thereby enhancing the stability of adversarial training.

[0132] Table 2 Test conditions of 6205 bearing dataset

[0133]

[0134] Table 3 Experimental condition settings

[0135]

[0136] Table 4 Experimental accuracy results

[0137]

[0138] Figure 3 shows the , , loss curves during model training. Among them, the curve is magnified 10 times. Due to the calculation involves sample screening and is affected by the fluctuations in the number of misclassified samples and screened samples, the curve fluctuates greatly in the early stage and gradually stabilizes in the later stage. The main role of the clustering-enhanced triplet loss is to enhance the clustering of high-quality samples and eliminate samples that are easy to confuse and affect the model. In the initial stage of model training, it helps the model converge more quickly. The experimental results show that the accuracy of the model improves rapidly in the early stage. Figure 4 shows the accuracy curve during model training. The experimental results show that the model can converge quickly in the early stage of iteration. Figure 5 shows the visualization results of the samples after clustering during model training, Figure 6It shows the clustering of reliable samples and the visualization of their clustering centers. After clustering, the samples located at the clustering boundary are more likely to be confused, which may interfere with the iterative process of the model. By screening reliable samples, the clustering boundary can be made clearer, which helps the model converge faster in the early stage of training.

[0139] The present invention proposes a method for diagnosing aero-engine bearing faults based on clustering-enhanced domain generalization. The results show that the constructed clustering-enhanced domain generalization network model can effectively achieve a high fault recognition rate under the condition of different distributions of aero-engine fault data. In summary, the model proposed by the present invention has achieved a good application effect.

Claims

1. A method for diagnosing faults of an aero-engine bearing based on clustering-enhanced domain generalization, characterized in that It includes the following steps: Step 1, vibration signal data processing; Step 2, construction of a fault diagnosis model clustering-enhanced domain generalization network; The constructed fault diagnosis model clustering-enhanced domain generalization network consists of two cascaded feature encoders, a classifier, and a domain discriminator; Step 2.1, construction of a feature encoder; The feature encoder consists of two parts connected in sequence, namely the domain alignment encoder and the classification encoder ; The domain alignment encoder extracts the cross-condition domain alignment features of the frequency-domain signal samples , and uses the MMD method and the domain discriminator to achieve distribution alignment between different domains; The classification encoder further extracts the high-level abstract classification features of the domain alignment features on the basis of the domain alignment features , and directly uses them for the classifier to achieve accurate classification of the fault state; Step 2.2, construction of a classifier and a domain discriminator; Classifier According to the classification encoder The extracted classification features Perform classification prediction on the frequency-domain signal samples: (4) Among them, represents the probability that the category belongs to ; represents the predicted category value of the frequency-domain signal sample corresponding to the classification feature ; represents the logit value that the frequency-domain signal sample corresponding to the classification feature belongs to the th category, and the classification loss is obtained by calculating the cross-entropy loss based on the predicted category value ; Domain discriminator Based on a fully connected network, determine the source domain from which the frequency domain signal samples corresponding to the input domain alignment features are from; (5) Among them, represents the probability that the source domain belongs to , represents the domain prediction value of the frequency-domain signal sample, represents the logit value that the frequency-domain signal sample corresponding to the domain alignment feature belongs to the th source domain, and the domain classification loss is obtained by calculating the cross-entropy loss based on the domain prediction value ; A gradient reversal layer is set before the domain discriminator. The gradient reversal layer exists as an identity mapping during forward propagation, while reversing the gradient direction during backpropagation, so as to promote the feature extractor to learn domain-aligned feature representations with domain invariance through adversarial training, and achieve implicit alignment of the domain-aligned feature distributions of each source domain; Step 2.3, construction of multi-source domain feature alignment; The feature encoder outputs domain-aligned features After that, the frequency-domain signal samples in different source domains are grouped according to their source domains, and the maximum mean difference (MMD) between their domain-aligned features is calculated; finally, the MMDs between all source domains are summed as the multi-source domain confusion loss; (6) Among them, represents MMD calculation; represents the number of different source domains included in this batch of data; , represents the domain alignment feature sets from two different source domains; Step 2.4, construction of a clustering-enhanced triplet loss; During the fault diagnosis process, to enhance the intra-class compactness and inter-class separability, the clustering-enhanced domain generalization network of the fault diagnosis model adopts a clustering-enhanced triplet loss. First, a clustering algorithm is used to cluster the classification features and generate corresponding pseudo-labels for the classification features of the frequency-domain signal samples after clustering. The label with the largest proportion of the true labels of the frequency-domain signal samples in the clustering is used as the same pseudo-label for the same clustering, so as to initially reflect the clustering structure and potential class information of the classification features, that is: (7) Among them, is for clustering to generate shared pseudo-labels, is the set of all possible true labels, and the newly generated clustering samples are , represents the classification feature of the th clustering sample in the th clustering, represents the pseudo-label of the th clustering; Screen the classification features of the clustering samples whose pseudo-labels are consistent with the true labels, that is , and use the clustering center of the classification features of these high-confidence samples as the anchor point, represents the number of high-confidence samples in the th cluster, and construct a triplet loss; (8) Among them, , , represents the classification feature of the th high-confidence sample in the cluster to which it belongs, is a hyperparameter used to control the minimum margin between positive and negative samples; Finally, the clustering-enhanced triplet loss is expressed as: (9) Among them, represents the number of high-confidence samples; Step 2.5, construction of a total loss function; The classification loss, domain classification loss, multi-source domain confusion loss, and clustering-enhanced triplet loss are weighted and summed to obtain the final comprehensive loss function: (10) Among them , , represent weight parameters, and the model parameters are updated by the Adam optimizer.

2. The method for diagnosing faults of an aero-engine bearing based on clustering-enhanced domain generalization according to claim 1, wherein Step 1, the specific implementation process of vibration signal data processing is as follows: Collect time series vibration signals and use to represent the data set of time series vibration signals, to represent the time series vibration signals of the th source domain, where represents the total number of source domains; segment the time series vibration signals to obtain time series samples, to represent the time series vibration signals of the th source domain after segmentation, where the th source domain contains time series vibration signal samples with corresponding health condition annotations, to represent the domain label of the th source domain; Subsequently, the time series vibration signal samples are subjected to a fast Fourier transform to convert the time series vibration signal samples into frequency domain signal samples : (1) Among them, represents the frequency index, is the total number of frequency points, is the sampling point index in the time domain, is the total number of sampling points of the time series signal sample; the time domain signal sample is decomposed into frequency domain signal samples of different frequency components through fast Fourier transform .​ 3. The method for diagnosing faults of an aero-engine bearing based on clustering-enhanced domain generalization according to claim 1, wherein The feature encoder uses a convolutional neural network to extract features with rich time-frequency information from the frequency-domain signal samples and its expression is as follows: (2) Among them, is the value of the output feature map of the th layer at position . represents the input of the th layer at position . For the first convolutional neural network layer of the feature encoder, that is, when , its input is the frequency-domain signal sample, that is, is the convolution operation, is the th convolution kernel of the th layer, is the width of the convolution kernel, is the th weight value of the convolution kernel.

4. The method for diagnosing faults of an aero-engine bearing based on clustering-enhanced domain generalization according to claim 1, wherein Classifier It is located at the backend of the clustering enhancement domain generalization network architecture of the fault diagnosis model and adopts a fully connected network, expressed as: (3) Among them, is the value of the -th output neuron in the -th layer, is the value of the -th input neuron in the -th layer, is the weight between the -th input neuron in the -th layer and the -th output neuron in the -th layer, is the bias value of all neurons in the -th layer with respect to the -th output neuron in the -th layer, is the total number of neurons in the -th layer.

Images