A rotating machinery fault diagnosis method based on self-supervised hypergraph representation learning

Abstract

The application discloses a rotating machinery fault diagnosis method based on self-supervised hypergraph representation learning. k The method comprises the following steps: acquiring a vibration signal and slicing and dividing a data set; extracting original attribute features through a CNN; constructing a hypergraph topology as an original structure feature based on a near neighbor; generating enhanced attribute features and enhanced structure features, and inputting the enhanced attribute features and the enhanced structure features into a hypergraph autoencoder to extract multi-view features; constructing a double-path variance-invariance-covariance regularization mechanism, and combining a global reconstruction loss to form a compound objective function to perform self-supervised pre-training; loading pre-training weights, performing bidirectional interactive fusion on the attribute features and the structure features through a mutual attention fusion module, and outputting a fault diagnosis result by using a small amount of labels to fine-tune a classifier. The application fully excavates high-order topological correlations among samples, enhances the robustness of a model to data loss and topological disturbance, and realizes high-precision and strong-robustness fault diagnosis in a small-label scene.

Description

Technical Field

[0001] This invention relates to a self-supervised hypergraph representation learning method and applies it to rotating machinery fault diagnosis, thus relating to the field of rotating machinery fault diagnosis technology. Background Technology

[0002] Rotating machinery is a key piece of equipment in industrial production and energy development. Under complex operating environments, rotating machinery systems are prone to problems such as rotor imbalance, misalignment, loosening, and bearing failure. Therefore, ensuring their safety is a crucial task in research and application. Fault diagnosis, as a key means of ensuring the safety of rotating machinery, has profound significance. Under complex operating conditions, there is a complex nonlinear relationship between the types and characteristics of rotating machinery faults. However, relying solely on signal characteristics from a single perspective often fails to comprehensively depict the operating state of the equipment, resulting in a low fault identification rate under complex conditions. Therefore, how to mine data information from multiple perspectives and effectively fuse it is key to improving diagnostic accuracy.

[0003] In current signal processing and feature extraction research, traditional analysis methods typically process time-series data using signal decomposition algorithms. However, these methods often rely on manual parameter setting and struggle to extract effective fault features in high-dimensional space. With the development of deep learning, convolutional neural networks (CNNs) are widely used for automatic attribute feature extraction. However, existing methods have the following shortcomings: 1. The feature extraction scale is singular, making it difficult to take into account both global trends and local details. Most existing CNN methods use fixed-scale convolutional kernels, making it difficult to simultaneously capture both the long-term global trend and the weak fault impact components in vibration signals. Single-scale feature extraction is prone to losing key information, resulting in incomplete representation of attribute features and limiting the discriminative ability of subsequent classifiers.

[0004] 2. Ignoring higher-order topological relationships between samples and underutilizing structural features. Conventional CNN methods primarily focus on the physical properties of samples (i.e., attribute features), neglecting the potential topological relationships (i.e., structural features) between samples. In fact, similarities and higher-order correlations exist between different fault samples, and this structural information is crucial for distinguishing similar fault types. Existing methods fail to effectively utilize the topological relationships between samples, leading to a decrease in recognition rate when fault types are similar.

[0005] 3. Lacks robustness to missing data and topology disturbances, resulting in significant performance degradation under real-world conditions. In real-world industrial scenarios, issues such as partial sensor failure, data transmission packet loss, and unstable connections frequently occur, leading to missing attribute features or topological disturbances. Existing methods lack enhancement strategies for these abnormal conditions. Once the input data is incomplete or the correlation is distorted, the model performance will drop sharply, making it difficult to meet the reliability requirements of practical engineering applications.

[0006] 4. The fusion method of attribute features and structural features is simple and lacks in-depth interaction. To uncover higher-order relationships between samples, Hypergraph Convolutional Networks (HGCNs) were developed. They support many-to-many node connections and can characterize complex relationships between data in greater detail than ordinary Graph Convolutional Networks (GCNs). However, most existing methods use simple concatenation to combine attribute features with structural features, lacking deep information interaction, resulting in insufficient feature fusion and failing to leverage the complementary advantages of multiple views.

[0007] 5. Label data is difficult to obtain, and overfitting is prone to occur in scenarios with few samples. In real-world industrial scenarios, obtaining accurately labeled fault data is extremely difficult and costly, leaving a large amount of data unlabeled. Supervised learning methods heavily rely on a large number of labels and are prone to overfitting in scenarios with few samples. Although self-supervised learning methods based on contrastive learning can be pre-trained on unlabeled data, these methods are prone to model collapse (i.e., all outputs converge to a constant) or require a large number of negative sample pairs, resulting in insufficient generalization ability in scenarios with scarce labels.

[0008] 6. Incomplete coverage of fault types and weak ability to diagnose complex faults. Existing research on fault diagnosis of rotating machinery, such as AUV propellers, mostly focuses on single fault types (such as entanglement, loosening, and imbalance), while paying insufficient attention to complex faults (such as entanglement + imbalance, and multiple concurrent faults) that are more common in actual operations. The characteristics of complex faults are mutually coupled and mask each other, making diagnosis far more difficult than for single faults. Existing methods have low accuracy in identifying such complex operating conditions.

[0009] To address this, this invention proposes a self-supervised hypergraph representation learning method and applies it to fault diagnosis of rotating machinery. This diagnostic model utilizes a CNN to map the original vibration signal into low-dimensional attribute features and constructs a hypergraph structure based on nearest neighbor relationships to obtain structural features. Enhanced features are generated through random masking and random edge dropping strategies, improving the model's robustness to data loss and topological perturbations. A parallel graph autoencoder architecture is designed, and a mutual attention fusion module achieves deep quality control and fusion of attribute and structural information. A variance-invariance-covariance regularization mechanism is introduced for self-supervised pre-training, effectively improving the model's fault diagnosis accuracy and generalization ability in low-sample scenarios while fully utilizing a large number of unlabeled samples. Furthermore, the model covers various single and compound faults, better meeting the needs of practical engineering applications. Summary of the Invention

[0010] To address the shortcomings of existing fault diagnosis methods, which focus solely on the physical properties of samples while neglecting the implicit higher-order topological relationships between samples, lack deep interactive quality checks during feature fusion, and suffer from high feature redundancy and weak model generalization ability in scenarios with scarce labeled data, a self-supervised hypergraph representation learning method is proposed and applied to the fault diagnosis of rotating machinery. This diagnostic model includes the following steps: S1: Obtain the raw vibration signal of the object to be diagnosed, perform slice processing, and divide the dataset. Vibration signals of rotating machinery at constant speed were collected using a data acquisition system and a vibration acceleration sensor. The collected vibration signals were sliced ​​using a sliding window technique to generate a standardized time-series sample sequence. The generated sample sequence was then randomly divided into a training set, a validation set, and a test set in a 6:2:2 ratio.

[0011] S2: Input the high-dimensional vibration signal into a convolutional neural network to construct the original attribute features, and construct a hypergraph topology based on the attribute features as the original structural features. S2-1: Constructing a deep temporal feature mapping network: Construct a feature extractor consisting of a cascaded four-layer one-dimensional convolutional neural network. The first convolutional layer uses a large convolutional kernel to capture global trend features, while the second to fourth layers use small convolutional kernels to extract local fine features. Each convolutional layer is followed by a batch normalization layer, a ReLU activation function, and a max pooling layer.

[0012] S2-2: Perform the mapping calculation from high-dimensional time series signals to low-dimensional attribute features: Input the sample sequence into the feature extractor described above, and perform convolution operations, feature dimensionality reduction, and matrix construction in sequence to output the low-dimensional feature matrix Z. attr , is defined as a primitive attribute feature.

[0013] S2-3: Based on kThe nearest neighbor algorithm constructs the hypergraph topology: after obtaining the attribute feature matrix Z... attr Then, the Euclidean distance between samples is calculated, and the distance between each sample is selected. k The nearest neighbor samples are used to construct hyperedges using hypergraph theory and generate an association matrix H, which, after normalization, is defined as the original structural feature Z. struct .

[0014] S3: Generates enhanced attribute and structure views, and inputs different combinations of original and enhanced features into a parallel hypergraph autoencoder to extract features. S3-1: Constructing Enhanced Attribute Features Based on Energy Compensation: For the original attribute feature Z attr Random mask perturbation is performed to generate a mask matrix M. An energy compensation coefficient λ = 1 / (1-p) is introduced to calculate the enhanced attribute features. .

[0015] S3-2: Constructing Enhanced Structural Features Based on Topological Perturbation: Targeting the original structural feature Z struct A random edge dropping strategy is used to generate a Boolean mask vector V. mask And construct a diagonal mask matrix Λ mask Calculate enhanced structural features .

[0016] S3-3: Input attribute features and structural features into the correlation encoder: Construct a parallel hypergraph autoencoder consisting of an attribute encoder and a structural encoder. The attribute encoder inputs the original attribute features Z... attr Mapped to latent representation H attr The structure encoder receives three sets of attribute-structure pairing data (the first set {Z...). attr Z struct}、Second group {Z ' attr Z struct}、The third group {Z attr Z ' struct}, using hypergraph convolutional layers to extract latent representations H struct H aug1 H aug2 .

[0017] S4: Construct a dual-path variance-invariance-covariance regularization mechanism for feature alignment and structural robustness, and combine it with the global reconstruction loss to form a composite optimization objective function. This mechanism iterates through backpropagation to extract pre-trained feature weights. S4-1: Constructing a self-supervised objective function S4-1-1: Perform projective feature mapping: transform the latent representation H output by the encoder... attr H struct Haug1 H aug2 Input the VICReg projection head separately to obtain the projection feature P. attr P struct P aug1 P aug2 .

[0018] S4-1-2: Calculating the robustness loss of the structural encoder : Projecting feature P aug1 With P aug2 The input VICReg loss function is used to constrain the robustness of the structure encoder to different input perturbations.

[0019] S4-1-3: Calculate the feature alignment loss : Projecting feature P attr With P struct Input the VICReg loss function to narrow the distance between attribute features and structural features in the latent space.

[0020] S4-1-4: Calculate the global reconstruction loss The input features are reconstructed using an attribute decoder, and the mean squared error between the reconstructed features and the original input features is calculated. The overall optimization objective function is: .

[0021] S4-2: Applying VICReg regularization to enhance feature alignment and structural robustness: Variance, invariance, and covariance regularization mechanisms are deployed on two optimization paths. The VICReg loss function is defined as follows: The first path performs feature alignment between pure attributes and pure topological features, aligning P... attr With P struct Input the VICReg loss function; the second path performs structural robustness enhancement between masking and discarding features, and P aug1 With P aug2 Input the VICReg loss function.

[0022] S5: Load upstream task parameters and perform quality checks and fusion of attribute features and structural features through the mutual attention fusion module. Finally, fine-tune the classifier using a small number of labels to output the fault diagnosis results of the rotating machinery. The mutual attention fusion module consists of a dual-channel attention interaction layer, a residual normalization layer, a feature concatenation layer, and a fusion projection layer. The input is the attribute feature P. attr and structural features P struct All dimensions d model .

[0023] S5-1: Define the mutual attention mechanism: The scaling dot product attention formula is as follows The attribute feature P output by the VICReg projection head. attr To query vector Q, the structural feature P output by the VICReg projection head is used. struct Let K be the key vector and V be the key vector. The goal is to enhance the attribute features using the topological information of the structure. The resulting enhanced feature is denoted as Z. attr_final Based on structural feature P struct For the query vector Q, with attribute features P attr Given a key vector K and a value vector V, the goal is to supplement the details of structural features using the local physical properties of the attributes. The resulting enhanced feature is denoted as Z. struct_final .

[0024] S5-2: Perform residual connection and layer normalization: After performing residual connection between the interactive features and the original input features, input them into the layer normalization module.

[0025] S5-3: Perform feature concatenation and final fusion: After concatenating the two sets of enhanced features along the channel dimension, input them into the fusion projection layer to obtain the final fused features: .

[0026] S5-4: Execute downstream task training: Load pre-trained weights, freeze the underlying network parameters, reinitialize the classifier, input a small number of labeled samples into the network, send them to the classifier through the mutual attention fusion module to calculate the cross-entropy loss, fine-tune the classifier parameters, and finally output the fault diagnosis results of the rotating machinery.

[0027] Compared with the prior art, the present invention has the following beneficial effects: 1. High-order topological relationships are fully explored, multi-scale attribute features are completely extracted, and feature representation capabilities are significantly improved. This invention utilizes a hypergraph topology structure to construct a hypergraph topology. k Nearest neighbor algorithm ( k =10) Each sample is connected to its nearest neighbor as a hyperedge, forming a many-to-many high-order association. Compared to traditional graph convolutional networks (GCNs), which can only describe pairwise relationships, hypergraph convolutional networks (HGCNs) can more meticulously characterize the complex similarities between fault samples, especially beneficial for distinguishing fault types with similar attribute features but different categories. Meanwhile, the four-layer one-dimensional convolutional neural network feature extractor constructed in this invention uses a large convolutional kernel in the first layer to capture global trend features, and small convolutional kernels in the subsequent three layers to extract local fine features. Combined with batch normalization, ReLU activation function, and max pooling layers, it achieves multi-scale feature extraction of "global first, local later," ensuring that attribute features contain both macroscopic working condition information and retain microscopic fault details. The synergistic effect of both lays a high-quality feature foundation for subsequent feature fusion and fault classification.

[0028] 2. The model's robustness and generalization ability are significantly improved through the combined enhancement of random masking and edge dropping, and the input of multi-view encoders. This invention applies random masking perturbation to the original attribute features, resulting in a masking rate. p =0.2, and an energy compensation coefficient λ=1 / (1-p)=1.25 is introduced to ensure the consistency of the statistical distribution of features before and after masking, forcing the model to infer global features using the residual attribute information that has not been masked. Simultaneously, a random edge dropping strategy and hyperedge occlusion rate are adopted for the hypergraph association matrix to simulate unstable sensor connections or missing partial associations, prompting the model to learn more robust topological structure information. Based on this, this invention constructs three sets of "attribute-structure" paired data parallel input structure encoders: {original attribute + original structure}, {enhanced attribute + original structure}, and {original attribute + enhanced structure}. Through various combinations of complete / incomplete attributes and complete / incomplete structures, the model can simultaneously learn feature representations under multiple operating conditions. The three enhancement strategies work synergistically, enabling the model to maintain stable diagnostic performance under various adverse conditions such as attribute loss and topological perturbation.

[0029] 3. VICReg dual-path regularization and global reconstruction loss are jointly optimized, achieving both feature alignment and structural robustness. This invention deploys a variance-invariance-covariance (VICReg) regularization mechanism across two co-optimization paths: the feature alignment path projects pure attributes onto feature P. attr With pure topological projection feature P struct The input VICReg loss function narrows the semantic distance between the two complementary features in the latent space, bridging the representation gap; the structurally robust path combines incomplete attributes with the projected features P of the complete structure. aug1 Projected features P of complete attributes + incomplete structure aug2 Input the VICReg loss function to constrain the consistency of the two sets of projections, enabling the model to resist topological disturbances caused by missing local information and dynamic operating conditions.

[0030] The VICReg loss function comprises three synergistic objectives: an invariance term narrows the feature distance between positive sample pairs; a variance term prevents model collapse; and a covariance term forces feature dimension decoupling and reduces redundancy. By adjusting these three hyperparameters, an optimal balance can be achieved between feature discriminability, diversity, and independence. Simultaneously, the global reconstruction loss forces the encoder to retain key physical attribute information. The composite objective function achieves joint optimization of the three losses, ensuring that pre-trained features possess both discriminative power and completeness.

[0031] 4. Mutual attention and bidirectional interactive fusion, residual and layer normalization ensure training stability, and feature fusion depth surpasses simple splicing. This invention abandons the traditional simple splicing method and introduces a multi-head attention mechanism to achieve bidirectional dynamic interaction between attribute features and structural features: using P attr For Q and P struct For K and V, topological information is used to enhance attribute features; conversely, for P... struct For Q and P attr For K and V, structural features are supplemented using attribute details. This bidirectional enhancement mechanism enables the fused features to possess both physical details and high-order correlation information, achieving deep quality inspection and fusion.

[0032] Meanwhile, residual connections and layer normalization alleviate gradient degradation and accelerate convergence; 10% Dropout effectively prevents overfitting in scenarios with few labels; scaling factors prevent gradient vanishing and ensure training stability. Finally, the fusion feature compresses and fuses complementary information from attribute and structural features to output high-order discriminative features for subsequent classification.

[0033] 5. Freeze-and-fine adjustments adapt to scenarios with few samples, and sliding window sampling improves data utilization with minimal computational cost. This invention employs a freeze-tuning strategy in downstream classification tasks: after loading self-supervised pre-trained weights, the underlying parameters of the basic feature extraction network and each encoder are locked, and only the linear classifier is reinitialized and fine-tuned. The underlying feature extraction network has already learned general fault feature representations on unlabeled data, and can quickly adapt to downstream tasks without a large amount of labeled data, avoiding overfitting problems caused by insufficient labeled data, with minimal computational cost.

[0034] In addition, the sliding window slicing process uses a sample length of 1024, a step size of 512, and an overlap rate of 0.5, which not only ensures the continuity of fault features at the sample boundaries, but also greatly expands the diversity of training samples under limited data conditions, providing sufficient data support for deep learning models. It is especially suitable for scenarios in actual engineering where it is difficult to collect a large amount of fault data.

[0035] 6. Comprehensive coverage of both single and complex faults, demonstrating significant practical engineering value. This invention covers rotating machinery fault types including normal conditions, single faults, and compound faults. Compound faults more closely resemble the complex operating conditions in real-world industrial scenarios. The model not only performs excellently in identifying single faults but also possesses the ability to distinguish compound faults. Through a technical approach of "hypergraph structure modeling, self-supervised VICReg pre-training, and mutual attention fusion," this invention achieves high accuracy, strong robustness, and low label dependency in rotating machinery fault diagnosis tasks, significantly enhancing the practical value of the diagnostic model in real-world complex environments. Attached Figure Description

[0036] Figure 1 Typical failures of AUVs and their propulsion systems; Figure 2 Data preprocessing and enhancement; Figure 3 Overall framework and flow of the model; Figure 4 Mutual attention fusion module; Figure 5 AUV thruster fault diagnosis results. Detailed Implementation

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment aims to fully demonstrate the specific implementation process of a self-supervised hypergraph representation learning method in rotating machinery fault diagnosis, but the implementation of the present invention is not limited thereto.

[0038] A method for fault diagnosis of rotating machinery based on self-supervised hypergraph representation learning includes the following steps: S1: Obtain the original vibration signal of the object to be diagnosed, perform slicing processing, and divide the dataset; The vibration signals of rotating machinery (using an LMS data acquisition system and a vibration acceleration sensor as an example in this implementation) at a speed of 3000 r / min were collected, with the sampling frequency set to 12.8 kHz. A sliding window technique was used to slice the collected vibration signals, with a sample length of 1024 data points and a sliding step size of 512 data points (i.e., an overlap rate of 0.5), thereby generating a standardized time-series sample sequence.

[0039] The generated sample sequences are randomly divided into training, validation, and test sets in a 6:2:2 ratio. The training set is used for pre-training model features, the validation set is used for parameter optimization and model accuracy monitoring, and the test set is used for final fault diagnosis and assessment. Besides normal conditions, the AUV thruster fault types covered in this embodiment include single faults and compound faults. Single faults include, but are not limited to, slipway loosening, entanglement, single-blade imbalance, and double-blade imbalance. Compound faults include, but are not limited to, entanglement + imbalance, slipway loosening + imbalance + entanglement, slipway loosening + entanglement, slipway loosening + single-blade imbalance, and slipway loosening + double-blade imbalance.

[0040] S2: Input the high-dimensional vibration signal into the convolutional neural network to construct the original attribute features, and construct the hypergraph topology based on the attribute features as the original structural features; S2-1: Constructing a deep temporal feature mapping network A feature extractor consisting of four cascaded one-dimensional convolutional neural networks is constructed to map the raw vibration signal to a low-dimensional feature space. The network structure is designed as follows: the first convolutional layer uses a large convolutional kernel (kernel_size=64 in this embodiment) to capture global trend features in long-term signals; the second to fourth layers use small convolutional kernels (kernel_size=3 in this embodiment) to extract local fine features. After each convolutional operation, a batch normalization (BN) layer is sequentially connected to accelerate convergence, a ReLU nonlinear activation function is used to enhance nonlinear expressive power, and a max pooling layer is used to reduce the feature dimensionality.

[0041] S2-2: Perform the mapping calculation from time-series signals to low-dimensional attribute features. The high-dimensional sample sequence processed in step S1 (set to 1024 dimensions in this embodiment) is input into the feature extractor described above, and the input sample is denoted as . x in The original attribute features are constructed through feature mapping using the following steps: S2-2-1: Perform convolution operation: Perform a convolution operation on the input signal, the formula is denoted as follows. y k =σ (W k *x in +b k ), where W k For the first k The convolutional kernel weights of the layer, b k For bias, σ This is the ReLU activation function, and the symbol * represents the convolution operation.

[0042] S2-2-2: Perform feature dimensionality reduction: Gradually compress the time dimension through multi-layer pooling operations, and fix the feature dimension using adaptive average pooling in the last layer. Then, map the features to the target dimension through a fully connected layer. D (This embodiment is set to 128 dimensions).

[0043] S2-2-3: Perform matrix construction: for those containing N A batch of samples is generated, and the final output is a single sample with dimension 1. N × D The low-dimensional characteristic matrix, denoted as Z attr Each row in this matrix represents a physical attribute feature vector of a sample, defined as the original attribute features of the data.

[0044] S2-3: Based onk Nearest neighbor algorithm for constructing hypergraph topology In obtaining the attribute feature matrix Z attr Then, the hypergraph topology is constructed using the distribution relationship of samples in the feature space. The specific steps are as follows: S2-3-1: Perform distance metric: Calculate attribute feature matrix Z attr Any two samples x i and x j The Euclidean distance between them is denoted as d ij =||x i -x j || 2.

[0045] S2-3-2: Perform nearest neighbor selection: For each central sample x i According to distance d ij Sort the values ​​in ascending order and select the nearest neighbor. k One sample (set in this embodiment) k =10) is its set of neighboring nodes, denoted as N( x i ).

[0046] S2-3-3: Performing Hyperedge Construction: Using hypergraph theory to construct associations and define each hyperedge. e i Includes central node x i and all its neighboring nodes N ( x i A hyperedge connects not just two nodes, but also all nodes including the central node. k +1 A set of nodes is used to capture local group relationships.

[0047] S2-3-4: Perform Incidence Matrix Generation: Based on the hyperedge construction results, generate the hypergraph's incidence matrix H (dimension 1). N × N If sample node v j Belongs to v i The superedge centered e i Then matrix element H ji =1, otherwise 0. Finally, H is normalized, and the correlation matrix is ​​defined as the original structural feature Z of the data. struct .

[0048] S3: Generates enhanced attribute and structure views, and inputs different combinations of original and enhanced features into a parallel hypergraph autoencoder to extract features. S3-1: Constructing Enhanced Attribute Features Based on Energy Compensation To enhance the model's robustness to missing attribute features, the original attribute features Z obtained in step S2 are... attr Perform random mask perturbation (e.g.) Figure 2 (Yellow arrow). Specifically, generate a [something] with Z. attr A mask matrix M of the same dimension, where each element independently follows a Bernoulli distribution with parameter 1-p, i.e., M ij ~Bernoulli(1-p), where p is the preset masking rate (in this embodiment, p=0.2). To maintain the consistency of the statistical distribution of data before and after the masking operation (i.e., energy conservation), an energy compensation coefficient λ=1 / (1-p) is introduced. Enhanced attribute feature Z ' attr The calculation formula is as follows:

[0049] in, This represents the Hadamard product (element-wise multiplication). This step forces the model to use unmasked residual attribute information to infer global features.

[0050] S3-2: Constructing Enhanced Structural Features Based on Topological Perturbations For the original structural features, a random edge-dropping strategy is used to generate enhanced structural features (such as...). Figure 2 (Larger green arrow). For N in the hypergraph e Generate a super edge of length N. e Boolean mask vector V mask The vector elements follow a parameter of 1-p. edge Bernoulli distribution (in this embodiment, p is set) edge =0.2). To achieve random removal of hyperedges, construct a structure with V mask A diagonal mask matrix Λ with diagonal elements mask =diag(V mask Enhanced structural feature Z) ' struct The calculation formula is as follows:

[0051] Here, × represents matrix multiplication, and diag(·) represents the operation of converting a vector into a diagonal matrix. This step sets the associated columns corresponding to the occluded hyperedges to zero, simulating the condition of unstable sensor connections or missing partial associations, thus enabling the model to learn more robust topological information.

[0052] S3-3: Input the attribute features and structural features into the relevant encoder (e.g., ... Figure 3 (As shown) A parallel graph autoencoder consisting of an attribute encoder and a structure encoder is constructed, which takes the original and enhanced attribute features and structure features as input.

[0053] In this step, both the attribute encoder and the structure encoder output latent space representations (before they have passed through the VICReg projection head). To distinguish them from the projected features in subsequent steps, the encoder output is denoted as H. attr H struct H aug1 H aug2 .

[0054] S3-3-1: Run the attribute encoder: convert the original attribute features Z... attr The direct input attribute encoder (composed of multiple fully connected layers) extracts high-order physical attribute features H of the sample. attr .

[0055] S3-3-2: Running the Structure Encoder: To achieve multi-view structural feature extraction, three sets of "attribute-structure" paired data are constructed and input into the structure encoder respectively. The first set of inputs is {Z attr Z struct}, extract feature H struct The second input group {Z ' attr Z struct By aggregating incomplete attribute information using the complete topological structure, latent representation H is extracted. aug1 The third input group {Z attr Z ' struct By aggregating complete attribute information using incomplete topological structures, latent representation H is extracted. aug2 The structural encoder utilizes hypergraph convolutional layers to weight and aggregate attribute features based on the input correlation matrix, outputting the corresponding latent space representation.

[0056] S4: Construct a dual-path variance-invariance-covariance regularization mechanism for feature alignment and structural robustness, and combine it with the global reconstruction loss to form a composite optimization objective function. The network parameters are iterated through backpropagation to extract pre-trained feature weights. S4-1: Constructing a self-supervised objective function To ensure the robustness of the structural encoder, this invention designs a composite loss function with regularization constraints. Specifically, it includes: S4-1-1: Perform projective feature mapping: Before calculating the VICReg loss, project the latent representation H from the encoder output. attr H struct H aug1 H aug2 Inputting the VICReg projection head (composed of multiple fully connected layers) into each layer yields the projection feature P. attr P struct P aug1 P aug2 .

[0057] S4-1-2: Calculating the robustness loss of the structural encoder The aim is to constrain the robustness of the structural encoder to different input perturbations. The encoded data of "enhanced attribute features + original structural features" is projected onto P. aug1 The encoding projection P of "original attribute features + enhanced structural features" aug2 Input the VICReg loss function. Consistency constraints are achieved by minimizing the mean squared errors of both, variance regularization is used to prevent model collapse, and covariance regularization is used to ensure orthogonality and independence between feature dimensions.

[0058] S4-1-3: Calculate the feature alignment loss The aim is to bridge the gap between attribute features and structural features in the latent space. The output of the attribute encoder is projected onto P. attr The output projection P of the structure encoder struct The input VICReg loss function forces the physical property features of the samples to remain semantically aligned with the hypergraph topology features, while reducing feature redundancy.

[0059] S4-1-4: Calculate the global reconstruction loss The mean squared error between the reconstructed structural features and the original input features is calculated by the attribute decoder (FC+ReLU+FC), ensuring that the encoder retains key physical attribute information. The overall optimization objective function is to jointly optimize the above three losses.

[0060]

[0061] in, The weighting factor for the reconstruction loss.

[0062] S4-2: Applying the VICReg regularization mechanism to perform feature alignment and enhance structural robustness. This step deploys regularization mechanisms, including variance, invariance, and covariance, in two separate optimization paths to constrain the features.

[0063] The VICReg mechanism is built upon three mutually synergistic optimization objectives. Given any pair of projected feature matrices... and Its joint loss function is defined as It includes calculations of three terms: invariance, variance, and covariance.

[0064] in, , , These are the weighted hyperparameters that control the intensity of the three loss terms.

[0065] Perform feature alignment between pure attributes and pure topological features: project the output of the attribute encoder P attr The output projection P of the structure encoder struct By inputting the VICReg loss function, the semantic distance between the two complementary features in the latent space is narrowed by minimizing the joint loss of variance, invariance and covariance between them.

[0066] Structural robustness enhancement between masking and discarding features: Projecting the encoding of "enhanced attribute features + original structural features" onto P aug1 The encoding projection P of "original attribute features + enhanced structural features" aug2 By inputting the VICReg loss function, the consistency between the two sets of projections is constrained. At the same time, the variance regularization term is used to prevent model collapse and the covariance regularization term is used to reduce feature redundancy, enabling the model to resist the topological disturbances caused by local information loss and dynamic operating conditions.

[0067] S5: Load upstream task parameters and perform quality checks and fusion of attribute features and structural features through the mutual attention fusion module. Finally, use a small number of labels to fine-tune the classifier and output the fault diagnosis results of the rotating machinery, i.e., the AUV propulsion.

[0068] The mutual attention fusion module mainly consists of four parts: a dual-channel attention interaction layer, a residual normalization layer, a feature concatenation layer, and a fusion projection layer. The core interaction layer abandons the traditional simple concatenation and instead introduces a multi-head attention mechanism. This mechanism maps heterogeneous features to unified feature tokens and achieves bidirectional dynamic interaction of features through multi-head self-attention (MHSA). The input is the attribute features P output by the attribute encoder. attr and the structural feature P output by the structural encoder struct All dimensions d model(This embodiment uses 128 dimensions). The specific quality inspection and fusion calculation process can be divided into the following steps: S5-1: Define the mutual attention mechanism Before building the interaction layer, we first define the core attention formula. For the query vector Q, key vector K, and value vector V, the formula for calculating ScaledDot-ProductAttention is:

[0069] Where d is the dimension of the key vector. This is a scaling factor used to prevent the gradient from vanishing due to an excessively large dot product result. Softmax This is a normalized exponential function. After obtaining the feature sequence, it is input into a multi-head self-attention module for dynamic interaction. The attribute features P output by the VICReg projection head are used as the basis for this process. attr To query vector Q, the structural feature P output by the VICReg projection head is used. struct Let K be the key vector and V be the key vector. The goal is to enhance the attribute features using the topological information of the structure. The resulting enhanced feature is denoted as Z. attr_final Based on structural feature P struct For the query vector Q, with attribute features P attr Given a key vector K and a value vector V, the goal is to supplement the details of structural features using the local physical properties of the attributes. The resulting enhanced feature is denoted as Z. struct_final .

[0070] S5-2: Residual Connectivity and Layer Normalization To prevent network degradation and accelerate convergence, the interacted features are residually concatenated with the original input features and then input to the LayerNorm module. The corresponding formula is:

[0071] in, Dropout This is a random deactivation operation (the ratio is set to 0.1 in this embodiment).

[0072] S5-3: Feature splicing and final fusion The two sets of features, after interactive quality inspection and normalization, are concatenated along the channel dimension (CAT) to form a joint feature vector of dimension 2×d_model. This vector is then fed into a fusion projection layer, which consists of a fully connected layer (Linear), normalization, and a ReLU activation function. The final fused feature Z... final The calculation formula is:

[0073] Among them, W fusionThis is the weight matrix of the fusion layer. b fusion This is the bias vector. This step compresses and fuses complementary information from attribute features and structural features, outputting high-order discriminative features for subsequent classification.

[0074] S5-4: Training to perform downstream tasks After pre-training, the model enters the downstream task training phase. The model extracts the underlying network from the pre-training phase and loads the corresponding weights. First, a freeze mechanism is used to lock the underlying parameters of the basic feature extraction network and each encoder, and the downstream linear classifier is reinitialized. Subsequently, a small number of labeled one-dimensional vibration signals are input into the network, and the extracted attributes and topological features are quality checked and fused by the mutual attention fusion module (MHSA). Then, the global features are fed into the classifier. To calculate the cross-entropy fine-tuning loss During backpropagation, the gradient flow only travels along... Backpropagation in the direction of the classifier allows for fine-tuning of its learnable parameters with a very small number of labeled samples. Finally, output the 10-category fault diagnosis results of the AUV thruster (e.g., Figure 5 (As shown).

[0075] Table 1. Dimensional changes of each module in the model ; The above are preferred embodiments of the present invention. Those skilled in the art can make various modifications or improvements based on these embodiments. Without departing from the general concept of the present invention, such modifications or improvements should fall within the scope of protection claimed by the present invention.

Claims

1. A method for fault diagnosis of rotating machinery based on self-supervised hypergraph representation learning, characterized in that, Includes the following steps: S1: Obtain the original vibration signal of the object to be diagnosed, perform slicing processing, and divide the dataset; S2: Input the high-dimensional vibration signal into the convolutional neural network to construct the original attribute features, and construct the hypergraph topology based on the attribute features as the original structural features; S3: Generate enhanced attribute features and structural features, and input different combinations of the original features and enhanced features into a parallel hypergraph autoencoder to extract features; S4: Construct a dual-path variance-invariance-covariance regularization mechanism for feature alignment and structural robustness, and combine it with the global reconstruction loss to form a composite optimization objective function. The network parameters are iterated through backpropagation to extract pre-trained feature weights. S5: Load upstream task parameters and perform quality checks and fusion of attribute features and structural features through the mutual attention fusion module. Finally, use a small number of labels to fine-tune the classifier and output the fault diagnosis results of the rotating machinery.

2. The method according to claim 1, characterized in that, Step S1 specifically involves: using a data acquisition system and a vibration acceleration sensor to acquire vibration signals of rotating machinery at a constant rotation speed; using a sliding window technique to slice the acquired vibration signals; setting the sample length to 1024 data points and the sliding step size to 512 data points to generate a standardized time-series sample sequence; and randomly dividing the generated sample sequence into a training set, a validation set, and a test set in a 6:2:2 ratio.

3. The method according to claim 1, characterized in that, Step S2 specifically includes: S2-1: Construct a feature extractor consisting of four cascaded one-dimensional convolutional neural networks. The first convolutional layer uses a large convolutional kernel to capture global trend features, while the second to fourth layers use small convolutional kernels to extract local fine features. Each convolutional layer is followed by a batch normalization layer, a ReLU activation function, and a max pooling layer. S2-2: Input the high-dimensional sample sequence into the feature extractor, and perform convolution operation, feature dimensionality reduction and matrix construction in sequence to output a low-dimensional feature matrix Z. attr , is defined as a primitive attribute feature; S2-3: Calculate the Euclidean distance between samples, selecting each sample... k The nearest neighbor samples are used to construct hyperedges using hypergraph theory and generate an association matrix H, which, after normalization, is defined as the original structural feature Z. struct .

4. The method according to claim 3, characterized in that, In step S2-2, feature dimensionality reduction is achieved by progressively compressing the time dimension through multi-layer pooling operations. In the last layer, adaptive average pooling is used to fix the feature dimension. Subsequently, the features are mapped to the target dimension D=128 through a fully connected layer.

5. The method according to claim 3, characterized in that, In step S2-3, k=10.

6. The method according to claim 1, characterized in that, Step S3 specifically includes: S3-1: For the original attribute feature Z attr Random mask perturbation is performed to generate a mask matrix M. An energy compensation coefficient λ = 1 / (1-p) is introduced to calculate the enhanced attribute features. The shading rate is p=0.2; S3-2: Targeting the original structural feature Z struct A random edge dropping strategy is used to generate a Boolean mask vector V. mask， And construct a diagonal mask matrix Λ mask Calculate enhanced structural features Among them, the super-edge occlusion rate p edge =0.2; S3-3: Construct a parallel hypergraph autoencoder consisting of an attribute encoder and a structure encoder. The attribute encoder will convert Z... attr Mapped to latent representation H attr The structure encoder receives three sets of paired data: the first set {Z attr Z struct }、Second group {Z ' attr Z struct }、The third group {Z attr Z ' struct }, using hypergraph convolutional layers to extract latent representations H struct H aug1 H aug2 .

7. The method according to claim 1, characterized in that, Step S4 specifically includes: S4-1: Construct a self-supervised objective function, which is the latent representation H output by the encoder. attr H struct H aug1 H aug2 Inputting the VICReg projection head (composed of multiple fully connected layers) into each layer yields the projection feature P. attr P struct P aug1 P aug2 ; Calculate the robustness loss of the structural encoder Feature alignment loss and global reconstruction loss The composite objective function is ; S4-2: Applying the VICReg regularization mechanism to perform feature alignment and structural robustness enhancement, the VICReg loss function is defined as follows: The first path will P attr With P struct The input VICReg loss function performs feature alignment, and the second path will P aug1 With P aug2 The input VICReg loss function enhances the robustness of the execution structure.

8. The method according to claim 1, characterized in that, Step S5 specifically includes: S5-1: Define the mutual attention mechanism, the scaling dot product attention formula is as follows: The attribute feature P output by the VICReg projection head attr To query vector Q, the structural feature P output by the VICReg projection head is used. struct Given key vectors K and V, the calculated enhanced feature is denoted as Z. attr_final； With structural feature P struct For the query vector Q, with attribute features P attr Given a key vector K and a value vector V, the calculated enhanced feature is denoted as Z. struct_final ; S5-2: Perform residual connectivity and layer normalization, the formula is as follows: Dropout is a random deactivation operation, with a ratio set to 0.

1. S5-3: Perform feature concatenation and final fusion. After concatenating the two sets of enhanced features along the channel dimension, input them into the fusion projection layer to obtain the final fused features. ; S5-4: Perform downstream task training, load pre-trained weights, freeze all network parameters except the classifier, reinitialize the classifier, input a small number of labeled samples into the network, send them to the classifier after passing through the mutual attention fusion module to calculate the cross-entropy loss, update only the classifier parameters, and output the fault diagnosis results of the rotating machinery.

9. The method according to claim 3, characterized in that, The structural encoder in the hypergraph autoencoder uses the hypergraph convolutional layer to perform weighted aggregation of attribute features based on the input correlation matrix, and outputs the corresponding latent space representation.

Images