A method for fault diagnosis of aero-engine bearings based on clustering-enhanced domain generalization

Abstract

The present invention belongs to the field of intelligent fault diagnosis technology, and discloses a method for diagnosing aircraft engine bearing faults based on clustering-enhanced domain generalization. First, the time domain vibration signal is converted into a frequency domain representation through fast Fourier transform, and a two-stage convolutional neural network architecture including a domain alignment encoder and a classification encoder is constructed. The domain discriminator is used to achieve cross-working condition feature alignment under adversarial training; the multi-source domain maximum mean difference statistical alignment strategy and the clustering-enhanced triplet loss mechanism are introduced, and the intra-class feature compactness and inter-class separability are simultaneously optimized through pseudo-label cluster center anchoring and dynamic interval constraints; gradient reversal adversarial training is combined with multimodal feature fusion. The method of the present invention can effectively improve the accuracy and robustness of generalization judgment of aircraft engine bearing fault diagnosis in the absence of target domain data.

Description

Technical Field

[0001] The present invention belongs to the field of aircraft engine fault detection, relates to engine rolling bearing fault diagnosis technology, and in particular to an aircraft engine bearing fault diagnosis method based on clustering enhanced domain generalization. Background Art

[0002] As the core power system of modern aircraft, the accuracy and reliability of aircraft engine fault diagnosis are key to ensuring flight safety. As engines evolve towards higher thrust-to-weight ratios and longer lifespans, their structural complexity and operating condition diversity have increased significantly, and their failure modes are characterized by multi-source, coupled, and progressive nature. Traditional diagnostic methods rely on expert experience and threshold alarm mechanisms, making it difficult to cope with the real-time analysis needs of massive multi-source heterogeneous data. While deep learning-based intelligent diagnostic technology can mine deep data features, it is limited by practical issues such as the scarcity of fault samples and differences in cross-domain data distribution. This results in insufficient generalization performance in actual deployments, especially with new engine models or under extreme operating conditions, where diagnostic accuracy drops sharply.

[0003] Current mainstream cross-domain diagnosis 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 focuses too much on matching the overall distribution between domains and ignores the category characteristics of fault features, resulting in unclear clustering structure of similar faults in the feature space, thereby reducing the accuracy of cross-domain classification; 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 category separability in the feature space, but also further weakens the reliability of the model.

[0004] In response to the above technical defects, future research should focus on the following directions: In the feature alignment process, it is necessary not only to achieve global distribution matching between domains, but also to strengthen feature alignment at the category level to ensure that similar faults have a clear clustering structure in the feature space; at the same time, optimize the adversarial training mechanism to prevent the domain discriminator and feature extractor from falling into local optimality, thereby alleviating the contradiction between domain invariance and category 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 fuzzy clustering structure of similar faults in the feature space, and easily falling into local optimum during the adversarial optimization process, the present invention provides an aircraft engine bearing fault diagnosis method 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 comprises the following steps:

[0008] Step 1, vibration signal data processing;

[0009] Collect time series vibration signals and use A dataset representing a time series vibration signal, Indicates the A time series vibration signal in the source domain, Represents the total number of source domains; segment the time series vibration signal to obtain time series vibration signal samples, Indicates the first A set of time series vibration signal samples in the source domain, where The source domain set contains with corresponding health conditions Annotated time series vibration signal samples , Indicates the Domain labels of source domains;

[0010] Then, the time series vibration signal samples Perform fast Fourier transform to transform the time series vibration signal samples Convert to frequency domain signal samples :

[0011]

[0012] in, 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 samples are transformed into Frequency domain signal samples decomposed into different frequency components ;

[0013] Step 2: Fault diagnosis model clustering enhancement domain generalization network construction;

[0014] The constructed fault diagnosis model clustering enhanced domain generalization network consists of two cascaded feature encoders, classifiers and domain discriminators;

[0015] Step 2.1, construct feature encoder;

[0016] The feature encoder uses a convolutional neural network to extract the frequency domain signal samples The features with rich time-frequency information are extracted from it, which is expressed as:

[0017]

[0018] in, For the The output feature map of the layer is at position The value at Indicates the Layer input In position The value at , for the first layer of the convolutional neural network of the feature encoder, that is, When its input is the frequency domain signal sample, that is , is the convolution operation, For the Layer convolution kernels, is the width of the convolution kernel, For the The convolution kernel weights;

[0019] The feature encoder consists of two parts: the domain alignment encoder and classification encoder ;Domain aligned encoder Extracting cross-case domain alignment features of frequency domain signal samples , and use the MMD method and domain discriminator to achieve distribution alignment between different domains; classification encoder Align features in domains Based on this, domain alignment features are further extracted High-level abstract classification features , directly used in the classifier to achieve accurate classification of fault status;

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

[0021] Classifier Located at the back end of the fault diagnosis model clustering enhanced domain generalization network architecture, a fully connected network is used, which is expressed as:

[0022]

[0023] in, For the Tier The value of the output neuron, For the Tier The value of the input neuron, For the Tier The input neuron Tier The weights between the output neurons, For the All neurons in the layer Tier The bias value of the output neuron, For the The total number of neurons in the layer;

[0024] Classifier Based on classification encoder The extracted classification features Perform classification predictions on frequency-domain signal samples:

[0025]

[0026] in, Indicates that the category belongs to The probability of Represents categorical features The corresponding category prediction value of the frequency domain signal sample, Indicates that the frequency domain signal sample corresponding to the classification feature belongs to The logit value of each category is used to calculate the cross entropy loss based on the category prediction value to obtain the classification loss. ;

[0027] Domain Discriminator Based on the fully connected network, determine the input domain alignment features The source domain from which the corresponding frequency domain signal samples come;

[0028]

[0029] in, Indicates that the source domain belongs to The probability of represents the domain prediction value of the frequency domain signal sample, The frequency domain signal sample corresponding to the domain alignment feature belongs to the The logit value of the source domain is used to calculate the cross entropy loss based on the domain prediction value to obtain the domain classification loss. ;

[0030] A gradient reversal layer is placed before the domain discriminator. The gradient reversal layer acts as an identity mapping during forward propagation and reverses the gradient direction during backward propagation. This allows the feature extractor to learn domain-invariant domain-aligned feature representations through adversarial training, achieving implicit alignment of the domain-aligned feature distributions of each source domain.

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

[0032] Feature encoder outputs domain aligned features Finally, 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 alignment features is calculated. MMD is used to measure the differences between different distributions. 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 feature distributions of the two source domains differ significantly in [1], it means that there is a large difference in their distributions; otherwise, it indicates that the distributions of the two source domains are relatively close. Finally, the sum of the MMDs between all source domains is used as the multi-source domain confusion loss.

[0033]

[0034] in, Represents MMD calculation; Represents the number of different source domains in this batch of data; 、 Represents a set of domain-aligned features from two different source domains;

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

[0036] In the fault diagnosis process, in order to enhance the intra-class compactness and inter-class separability, the fault diagnosis model clustering enhancement domain generalization network adopts clustering enhanced triple loss; first, the clustering algorithm is used to classify the classification features. Perform clustering and generate corresponding pseudo labels for the classification features of the frequency domain signal samples after clustering. The labels with the largest proportion of true labels of the frequency domain signal samples in the cluster are regarded as the same cluster and share the same pseudo labels. This can preliminarily reflect the clustering structure and potential category information of the classification features, namely:

[0037]

[0038] in, For clustering Generate shared pseudo labels, is the set of all possible true labels, and the newly generated cluster samples are , Indicates the The first cluster The classification features of cluster samples, Indicates the Pseudo labels for clusters;

[0039] Filter the classification features of cluster samples whose pseudo labels are consistent with the true labels, that is, , with the cluster centers of the classification features of these high confidence samples As an anchor point, Indicates the The number of high-confidence samples in a cluster is used to construct the triplet loss;

[0040]

[0041] in, , , Indicates the cluster No. The classification features of high-confidence samples, Is a hyperparameter used to control the minimum interval between positive and negative samples;

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

[0043]

[0044] in, Indicates 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 clustering enhanced triple loss are weighted summed to obtain the final comprehensive loss function:

[0047]

[0048] in 、 、 Represents the weight parameter, and the model parameters are updated through the Adam optimizer.

[0049] In this way, the features of samples of the same category are brought closer together, so that they are compactly clustered around the cluster center, enhancing the compactness within the class; at the same time, the features of samples of different categories are pushed further apart, increasing the distance between classes and improving the feature space discrimination, making the fault categories clearer and more separable in feature representation.

[0050] The present invention achieves the following beneficial effects: It uses a multi-layer convolutional neural network for feature extraction, combined with a cluster-enhanced triplet loss mechanism, to accurately capture features in vibration signals, significantly improving the accuracy of fault diagnosis. Furthermore, the introduction of domain adversarial training and a multi-source domain feature alignment strategy effectively reduces differences in data distribution under different operating conditions, further enhancing the model's generalization capabilities across a wide range of operating environments.

[0051] The present invention proposes a method for diagnosing aircraft engine bearing faults based on clustering-enhanced domain generalization. The method proposed in the present invention collects vibration signal data under different working conditions and performs fast Fourier transform on it to achieve time-frequency domain conversion of the signal. 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 triplet 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 Flowchart of the fault diagnosis method for rolling bearings of aircraft engines 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 operating conditions. This data covers the vibration characteristics of the machine under various operating conditions. To facilitate subsequent analysis and classification, we clearly divided the fault conditions into three categories: normal condition (NM), inner race fault (IF), and outer race fault (OF). Furthermore, since the data originated from different operating conditions, domain labels were assigned to distinguish the data sources.

[0062] use represents a dataset, where Indicates the The source domain time series vibration signal data contains Time series vibration signal , and the working condition corresponding to the signal , is the set of all possible true labels. There are three values: 0, 1, and 2, corresponding to NM, IF, and OF faults respectively. Indicates the Domain labels for domain data.

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

[0064] Segment the time series vibration signal to obtain time series vibration signal samples. Indicates the first The time series vibration signal sample set of the field The fields include with corresponding health conditions Annotated time series vibration signal samples .

[0065] Subsequently, the time series vibration signal samples are transformed by FFT to convert the complex time series vibration signal samples into Frequency domain signal samples decomposed into different frequency components :

[0066]

[0067] in, 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 timing signal.

[0068] Step 2: Build a feature encoder based on convolutional neural network

[0069] The core architecture of the feature encoder uses a convolutional neural network (CNN) to extract the time-frequency features of the vibration signal. CNN can automatically learn the local features and global structure of the signal to enhance the feature expression ability. The feature encoder consists of multiple layers of convolutional layers and pooling layers in cascade to gradually extract deep features. The convolutional layer extracts local features, such as the vibration pattern within a specific frequency range, which can be expressed as

[0070]

[0071] in, For the The output feature map of the layer is at position The value at Indicates the Layer input In position The value at , for the first layer of the convolutional neural network of the feature encoder, that is, When its input is the frequency domain signal sample, that is , is the convolution operation, For the Layer convolution kernels, is the width of the convolution kernel, For the The convolution kernel A weight.

[0072] The activation layer enhances the network's expressiveness through nonlinear mapping and is composed of activation functions. Convolutional layers and fully connected layers perform linear operations, and activation layers are subsequently added to perform nonlinear transformations to improve the network's ability to learn and process complex data. This paper uses Leaky ReLU as the activation function:

[0073]

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

[0075] The pooling layer effectively compresses the feature dimension by downsampling, reducing the computational complexity and number of parameters of the model while achieving key feature screening. This invention uses the maximum pooling operation

[0076]

[0077] in, is the input feature In the pooling window The activation value of .

[0078] The feature encoder of the present invention is composed of 6 layers of convolutional neural network. It consists of the first 4 convolutional layers, which output domain-aligned features across working conditions. , which is used by the domain discriminator to determine the source of features, achieve domain alignment, and reduce the impact of working condition differences on fault diagnosis. It is composed of the last two convolutional layers to extract high-level classification features. , for the classifier to identify faults. The classifier is based on ,Combined with the preset classification algorithm, the vibration signal is divided into fault states.

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

[0080] Classifier Located at the back end of the model, based on the features extracted by the feature encoder Output the probability distribution of fault categories. Its structure is a fully connected network consisting of multiple layers of neurons, each layer is fully connected, and learns the mapping relationship between input features and fault categories.

[0081]

[0082] in, For the Tier The value of the output neuron, For the Tier The value of the input neuron, For the Tier The input neuron Tier The weights between the output neurons, For the All neurons in the layer Tier The bias value of the output neuron, For the The total number of neurons in the layer.

[0083] The classifier calculates the classification features of the input through activation function and weight matrix Make classification predictions

[0084]

[0085] in, Indicates that the category belongs to The probability of Represents categorical features The corresponding category prediction value of the frequency domain signal sample, Indicates that the frequency domain signal sample corresponding to the classification feature belongs to The logit value of each category.

[0086] During the training process, the classifier parameters are adjusted with the help of optimization algorithms to minimize the classification error and improve the diagnostic accuracy.

[0087]

[0088] in is the logits processed by the Softmax function, and its target distribution is a one-hot vector , only the position corresponding to the target category is 1, and the rest are 0, that is, when the target category is hour, ,otherwise , is the batch size of the input.

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

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

[0091]

[0092] Step 4: Build a domain discriminator

[0093] Domain Discriminator It is a component of gradient reversal adversarial training. Its structure is similar to that of the classifier and it is responsible for determining the alignment features of the input domain. The source domain of The input batch size is the domain label prediction of the frequency domain signal sample corresponding to the domain alignment feature. , Indicates that its source domain belongs to The domain classification loss uses cross entropy loss:

[0094]

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

[0096] This paper achieves implicit alignment of domain-aligned feature distributions by placing a gradient reversal layer before the discriminator. The domain discriminator distinguishes the source domain of features. During forward propagation, the gradient reversal layer directly passes the input; during backward propagation, the gradient sign is reversed, allowing the feature encoder to learn domain-invariant features that both aid fault classification and "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, through the gradient inversion layer, guides the feature encoder to learn domain-invariant features, promotes the alignment of source and target domain features, and enhances the model's ability to adapt across working conditions.

[0098] Step 5: Construct multi-source domain feature confusion loss

[0099] This paper adopts a multidimensional statistical alignment strategy, calculates the MMD between source domains as the alignment method, and quantifies the distribution differences based on hypothesis testing and kernel methods:

[0100]

[0101] in 、 represents the domain alignment feature set from two different source domains, and Indicates that the data in this batch come from and The number of domain-aligned features of the domain, Indicates from No. Domain alignment features, Represents the kernel function. In this design, the kernel functions are all Gaussian kernel functions.

[0102] During training, data from three source domains are input, mixed and fed into the feature extractor, and then passed through the domain alignment encoder. Extracting domain alignment features . Then, according to the domain label Group and calculate the MMD between different source domains, and take the sum of the MMD between each source domain as the multi-source domain confusion loss

[0103]

[0104] in, Represents the number of different source domains included in this batch of data.

[0105] Step 6: Construct cluster-enhanced triplet loss

[0106] In the fault diagnosis process, in order 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 classification features Clustering. The process of KMeans algorithm is: randomly select Classification features as the center of the initial cluster , is the number of categories contained in the classification feature, that is, the number of clusters. , assign it to the cluster center closest to it ,Right now: ,in Indicates the classification features To cluster center In each iteration, the center of each cluster is recalculated , that is, cluster The mean of all points in , iterate repeatedly 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 true labels of the classification features in the cluster are regarded as the same cluster and 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 the set of all possible true labels, and the newly generated cluster samples are , Indicates the The first cluster The classification features of cluster samples, Indicates the 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 The number of high confidence samples in a cluster, constructing triple loss

[0112]

[0113] in , , Indicates the cluster No. The classification features of high-confidence samples, is a hyperparameter that controls the minimum margin between positive and negative samples.

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

[0115]

[0116] in, Indicates the number of high confidence samples.

[0117] Step 7: Construct loss function

[0118] During the forward propagation of the model, the classification loss, domain classification loss, multi-source domain confusion loss, and clustering enhanced triple loss are weighted 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 capabilities suitable for cross-operating fault diagnosis:

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

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

[0124] (2) Forward propagation: 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 predictions; cluster triplet loss to make samples of the same class cluster more closely in the feature space, while separating samples of different classes; domain confusion loss to introduce adversarial training between source domains to enhance the cross-domain adaptability of the model; domain classification loss to train the adversarial learning of the model.

[0126] (4) Back propagation: The Adam optimizer is used to update the parameters of each network module according to the calculated gradient. The gradient reversal layer reverses the gradient direction of the domain discriminator during back propagation.

[0127] Dataset composition:

[0128] The training data comes from the Shandong University of Science and Technology's bearing and gear dataset. Experiments were conducted on a rotating machinery transmission component fault implantation test bench at Shandong University. Fault diagnosis tests for rolling bearings and gears were conducted under different operating conditions. The test components were a 6205 bearing and a planetary gearbox. Two three-axis piezoelectric accelerometers, magnetically fixed to fixed brackets at the bearing housing and the planetary gearbox, were used to acquire vibration signals from the bearing and gearbox. The experiments were conducted in steady-state mode, with a sampling period of 40 seconds and a sampling frequency of 25.6 kHz. During each sampling period, six channels of vibration signals were acquired from the two three-axis piezoelectric accelerometers. Data from the 6205 bearing dataset under steady-state operating conditions were selected as experimental data. Each operating condition included a set of healthy data, one set of data for inner race faults with diameters of 0.2 mm, 0.4 mm, and 0.6 mm, and one set of data for outer race faults with diameters of 0.2 mm, 0.4 mm, and 0.6 mm. The vibration signals were segmented, each with a sample length of 4096, resulting in 256 samples per data set. For each working condition, 1792 samples were obtained. For the data in the source domain, 500 samples were selected as the training set according to the health status ratio, and 200 samples were selected as the test set. For the data in the target domain, 200 samples were selected as the test set according to the health status ratio.

[0129] Model training and test result analysis

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

[0131] The ERM method conducted generalization experiments based solely on a CNN network, the MMD method incorporated multi-source domain MMD calculations into the CNN architecture, and the DANN method incorporated domain adversarial training into the CNN architecture. All experiments were trained 30 times, and the classification accuracy of the best models was compared. The experimental results show that the introduction of multi-source domain MMD loss and domain adversarial training does not necessarily lead to improved accuracy. For example, in experiments in groups A, C, D, and F, the accuracy decreased. This is because MMD is primarily based on mean matching and fails to fully capture the complex characteristics of the data distribution. In experiments in groups E and F, domain adversarial training primarily targets global alignment of feature distributions, potentially overlooking fine-grained matching of local features. In contrast, the CEDGN method incorporates a domain adversarial mechanism and learns a more flexible nonlinear mapping through end-to-end adversarial training. This optimizes the consistency of feature distributions and improves the model's adaptability in complex scenarios, compensating for the shortcomings of the MMD method. MMD uses kernel methods to ensure that features from different domains are more closely distributed locally and globally, thereby enhancing the stability of adversarial training.

[0132] Table 2 Test conditions of the 6205 bearing dataset

[0133]

[0134] Table 3 Experimental conditions settings

[0135]

[0136] Table 4 Experimental accuracy results

[0137]

[0138] Figure 3 Shows the model training process 、 、 Loss curve. Among them, The curve is magnified 10 times. The calculation involves sample screening and is affected by fluctuations in the number of misclassified samples and screened samples. The curve fluctuates significantly in the early stages but gradually stabilizes in the later stages. The main function of the cluster-enhanced triplet loss is to enhance the clustering of high-quality samples and eliminate samples that are easily confused and affect the model. This helps the model converge more quickly in the early stages of model training. Experimental results show that the model's accuracy improves rapidly in the early stages. Figure 4 The accuracy curve during model training is shown. The experimental results show that the model can converge quickly in the early stages of iteration. Figure 5 It shows the visualization results of sample clustering during model training. Figure 6This shows a visualization of clusters of reliable samples and their cluster centers. After clustering, samples at the cluster boundaries are more likely to be confused, potentially interfering with the model's iteration process. By screening reliable samples, cluster boundaries can be made clearer, helping the model converge faster in the early stages of training.

[0139] This paper proposes a method for diagnosing aircraft engine bearing faults based on clustering-enhanced domain generalization. Results show that the constructed clustering-enhanced domain generalization network model can effectively achieve a high fault recognition rate even when aircraft engine fault data has a different distribution. In summary, the proposed model achieves excellent application results.

Claims

1. A method for diagnosing aircraft engine bearing faults based on clustering-enhanced domain generalization, characterized in that: The following steps are involved: Step 1, vibration signal data processing; Step 2: Fault diagnosis model clustering enhancement domain generalization network construction; The constructed fault diagnosis model clustering enhanced domain generalization network consists of two cascaded feature encoders, classifiers and domain discriminators; Step 2.1, construct feature encoder; The feature encoder consists of two parts: the domain alignment encoder and classification encoder ;Domain aligned encoder Extracting cross-case domain alignment features of frequency domain signal samples , and use the MMD method and domain discriminator to achieve distribution alignment between different domains; classification encoder Align features in domains Based on this, domain alignment features are further extracted High-level abstract classification features , directly used in the classifier to achieve accurate classification of fault status; Step 2.2, build the classifier and domain discriminator; Classifier Based on classification encoder The extracted classification features Perform classification predictions on frequency-domain signal samples: (4)； in, Indicates that the category belongs to The probability of Represents categorical features The corresponding category prediction value of the frequency domain signal sample, Indicates that the frequency domain signal sample corresponding to the classification feature belongs to The logit value of each category is used to calculate the cross entropy loss based on the category prediction value to obtain the classification loss. ; Domain Discriminator Based on the fully connected network, determine the input domain alignment features The source domain from which the corresponding frequency domain signal samples come; (5)； in, Indicates that the source domain belongs to The probability of represents the domain prediction value of the frequency domain signal sample, The frequency domain signal sample corresponding to the domain alignment feature belongs to the The logit value of the source domain is used to calculate the cross entropy loss based on the domain prediction value to obtain the domain classification loss. ; A gradient reversal layer is placed before the domain discriminator. The gradient reversal layer acts as an identity mapping during forward propagation and reverses the gradient direction during backward propagation. This allows the feature extractor to learn domain-invariant domain-aligned feature representations through adversarial training, achieving implicit alignment of the domain-aligned feature distributions of each source domain. Step 2.3: construct multi-source domain feature alignment; Feature encoder outputs domain aligned features Finally, 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 alignment features is calculated; finally, the MMD between all source domains is summed as the multi-source domain confusion loss; (6)； in, Represents MMD calculation; Represents the number of different source domains in this batch of data; 、 Represents a set of domain-aligned features from two different source domains; Step 2.4, construct the cluster-enhanced triplet loss; In the fault diagnosis process, in order to enhance the intra-class compactness and inter-class separability, the fault diagnosis model clustering enhancement domain generalization network adopts clustering enhanced triple loss; first, the clustering algorithm is used to classify the classification features. Perform clustering and generate corresponding pseudo labels for the classification features of the frequency domain signal samples after clustering. The labels with the largest proportion of true labels of the frequency domain signal samples in the cluster are regarded as the same cluster and share the same pseudo labels. This can preliminarily reflect the clustering structure and potential category information of the classification features, namely: (7)； in, For clustering Generate shared pseudo labels, is the set of all possible true labels, and the newly generated cluster samples are , Indicates the The first cluster The classification features of cluster samples, Indicates the Pseudo labels for clusters; Filter the classification features of cluster samples whose pseudo labels are consistent with the true labels, that is, , with the cluster centers of the classification features of these high confidence samples As an anchor point, Indicates the The number of high-confidence samples in a cluster is used to construct the triplet loss; (8)； in, , , Indicates the cluster No. The classification features of high-confidence samples, Is a hyperparameter used to control the minimum interval between positive and negative samples; Finally, the clustering-enhanced triplet loss is expressed as: (9)； in, Indicates the number of high confidence samples; Step 2.5, construct the total loss function; The classification loss, domain classification loss, multi-source domain confusion loss and clustering enhanced triple loss are weighted summed to obtain the final comprehensive loss function: (10)； in 、 、 Represents the weight parameter, and the model parameters are updated through the Adam optimizer.

2. The method for diagnosing aero-engine bearing faults based on clustering-enhanced domain generalization according to claim 1, characterized in that: Step 1: The specific implementation process of vibration signal data processing is as follows: Collect time series vibration signals and use A dataset representing a time series vibration signal, Indicates the A time series vibration signal in the source domain, Represents the total number of source domains; segment the time series vibration signal to obtain time series samples, Indicates the first The time series vibration signal of the source domain is The source domain contains with corresponding health conditions Annotated time series vibration signal samples , Indicates the Domain labels of source domains; Then, the time series vibration signal samples Perform fast Fourier transform to transform the time series vibration signal samples Convert to frequency domain signal samples : (1)； in, 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 samples are transformed into Frequency domain signal samples decomposed into different frequency components .

3. The method for diagnosing aircraft engine bearing faults based on clustering-enhanced domain generalization according to claim 1, characterized in that: The feature encoder uses a convolutional neural network to extract the signal from the frequency domain The features with rich time-frequency information are extracted from it, which is expressed as: (2)； in, For the The output feature map of the layer is at position The value at Indicates the Layer input In position The value at , for the first layer of the convolutional neural network of the feature encoder, that is, When its input is the frequency domain signal sample, that is , is the convolution operation, For the Layer convolution kernels, is the width of the convolution kernel, For the The convolution kernel A weight.

4. The method for diagnosing aero-engine bearing faults based on clustering-enhanced domain generalization according to claim 1, characterized in that: Classifier Located at the back end of the fault diagnosis model clustering enhanced domain generalization network architecture, a fully connected network is used, which is expressed as: (3)； in, For the Tier The value of the output neuron, For the Tier The value of the input neuron, For the Tier The input neuron Tier The weights between the output neurons, For the All neurons in the layer Tier The bias value of the output neuron, For the The total number of neurons in the layer.

Images