Rolling bearing equipment fault diagnosis sample feature enhancement method

Abstract

The present application relates to the technical field of feature enhancement, in particular to a rolling bearing equipment fault diagnosis sample feature enhancement method, comprising: obtaining original data under a multi-physical field of a nuclear power unit; extracting high-dimensional features of different signals in the original data by using a deep convolutional neural network; constructing and training a joint model based on a Transformer and a graph fusion network; the Transformer is used for splicing the high-dimensional features of different signals to obtain fused preliminary features, and the extracted high-dimensional features of different signals and the fused preliminary features are respectively subjected to sequence enhancement to obtain enhanced features of different modalities; the graph fusion network is used for fusing the enhanced features of different modalities to generate multi-modal fusion features; and the trained joint model is used to generate global multi-modal fusion features. Through the sample feature enhancement method, under the condition of small sample data, high-reliability and high-dimension multi-modal fusion features can be generated.

Description

Technical Field

[0001] This invention relates to the field of sample feature enhancement technology, and in particular to a method for enhancing the features of rolling bearing equipment fault diagnosis samples. Background Technology

[0002] Circulating water pumps are typical rotating equipment in the mechanical field. In nuclear power plants, they mainly include two types: concrete volute centrifugal pumps and vertical metal mixed-flow pumps. The bearings at the non-drive end of the circulating water pump are cylindrical roller bearings, located in the radial bearing chamber at the bottom of the pump body. Compared to bearings in other positions, they bear more radial loads and have a higher probability of failure. Furthermore, because they are located at the bottom of the detachable and removable rotor assembly, maintenance is more difficult and costly.

[0003] The stable and reliable operation of circulating water pumps is crucial for the stable operation of nuclear power equipment. The reliability and safety requirements for this equipment are extremely high, and intelligent operation and maintenance methods can effectively improve the safe and reliable operation of power generation equipment. Traditional intelligent fault diagnosis algorithms require a large amount of data for training and are highly dependent on data samples, especially fault sample data. However, in practical engineering, due to the extremely high safety requirements of circulating water pumps, fault samples are extremely rare. For safety reasons, it is also impossible to artificially generate and obtain fault sample data on actual equipment.

[0004] In the prior art, a Chinese invention patent document with publication number CN114818961A and publication date of July 29, 2022, was proposed to solve the above-mentioned technical problems. The technical solution disclosed in this patent document is as follows: a multi-modal information multi-level fusion monitoring method and system for rolling bearings; the method acquires multi-modal signals of vibration, sound, temperature and speed during bearing operation without changing the original structure of the bearing, and transmits the original signal data to a multi-modal multi-level fusion monitoring model deployed on a host computer. Through multi-level fusion of data-level fusion, feature-level fusion and decision-level fusion of the original signal data, the current working state of the bearing is accurately determined.

[0005] In practical applications, the above-mentioned technical solutions mainly rely on Transformer networks to achieve multi-level fusion. They only utilize the differences between feature-level and decision-level fusion and do not specifically address the robustness issues of feature extraction and diagnosis under small sample data. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a feature enhancement method for fault diagnosis samples in rolling bearing equipment. Under conditions of small sample data, it can generate highly reliable and high-dimensional multimodal fusion features, which can significantly improve the accuracy and robustness of subsequent fault diagnosis. This solves the technical bottleneck of traditional methods that rely on a large number of fault samples and whose performance drops sharply under small sample conditions.

[0007] This invention is achieved by adopting the following technical solution: A method for enhancing the features of rolling bearing equipment fault diagnosis samples includes the following steps: Step S1. Obtain raw data of the nuclear power unit under multi-physics field conditions, including the distribution data of three physical fields: vibration signal, temperature signal and pressure signal generated during the operation of the circulating water pump; Step S2. Use a deep convolutional neural network to extract high-dimensional features of different signals in the original data; Step S3. Construct and train a joint model based on Transformer and graph fusion network; the Transformer is used to concatenate the high-dimensional features of different signals to obtain preliminary fused features, and to perform sequence enhancement on the extracted high-dimensional features of different signals and the preliminary fused features to obtain enhanced features of different modalities; the graph fusion network is used to fuse the enhanced features of different modalities to generate multimodal fused features. Step S4. Use the trained joint model to generate global multimodal fusion features.

[0008] Step S2 specifically refers to: using a deep convolutional neural network to extract high-dimensional features of the vibration signal from the original data. High-dimensional characteristics of temperature signals and the high-dimensional characteristics of pressure signals : , , , In the formula, It is a vibration signal. Indicates the time step. Indicates the signal dimension; It is a vibration signal. Indicates the time step. Indicates the signal dimension; It is a vibration signal. Indicates the time step. The dimension of the signal is represented by ; CNN is a deep convolutional neural network that extracts local features of a signal.

[0009] A modal category embedding layer is added to the input of the Transformer to distinguish different physical field signals through modal labels; LayerNorm operations are introduced before and after the input and output of the Transformer's self-attention layer and feedforward layer.

[0010] Dropout layers are set after the self-attention layer and feedforward layer of the Transformer, and L2 weight decay is used as an additional regularization method; an early stopping strategy is applied during training.

[0011] By concatenating the high-dimensional features of different signals, a preliminary fused feature is obtained, specifically: , In the formula, concat Indicates a splicing operation; These are the initial characteristics of integration; For the high-dimensional characteristics of vibration signals, This represents a high-dimensional feature of the temperature signal. This represents the high-dimensional characteristics of the pressure signal.

[0012] Enhancement features of different modes include single-mode enhancement features of vibration. Single-mode enhancement characteristics of temperature Single-mode enhancement characteristics of pressure and the multimodal features spliced ​​from the three. The sequence enhancement method is as follows: , , , .

[0013] The graph fusion network is used to fuse and represent enhanced features from different modalities to generate multimodal fusion features, specifically including the following steps: Step S 31 The enhancement features of different modalities are used as initial graph nodes, represented as follows: , , , , In the formula, This is a single-mode enhancement feature of vibration. For single-mode enhancement features of temperature, For single-mode enhancement characteristics of pressure, The multimodal features are spliced ​​from the three components; Step S 32 Construct the graph node set: , In the formula, n Representing the n The graph after the next iteration; Step S 33A dynamic weight learning mechanism based on attention is used to calculate the relationship weights between different modalities. , In the formula, and Indicates the first Nodes in the layer diagram i With nodes j Features W For learnable weight matrix, For attention vectors, This indicates a splicing operation. , ; Step S 34 By utilizing graph fusion networks to optimize feature representations layer by layer, nodes... i The features are updated by a weighted sum of the features of its neighboring nodes: , in, Represents a node i The set of neighboring nodes, For the first The weight matrix of the layer graph, For the first l The layer's bias vector, after L After the feature maps of the layers are fused, the final feature representation of each node is obtained. ; Step S 35 Aggregate the features of all modal nodes to obtain multimodal fusion features: , in, This is a multimodal fusion feature.

[0014] During the training of the joint model based on Transformer and graph fusion network, a joint loss function is introduced to optimize the parameters of the joint model.

[0015] The joint loss function is: , , , , In the formula, For multimodal fusion features, For the sample size, K Number of fault categories For the sample i The true category, For the sample i The predicted probability; L 1 represents the specific loss function. L 2 represents the modal inconsistency loss. L is the balance coefficient, and L is the joint loss function; For the process L After the feature maps of the layers are fused, the final features of each node are obtained. .

[0016] Optimizing the parameters of the joint model specifically refers to using the Adam optimizer to optimize the parameters of the joint model.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention addresses the extremely scarce fault samples in the practical application of rolling bearings in circulating water pumps of nuclear power plants, proposing a novel feature enhancement and fusion method suitable for small-sample scenarios. This method constructs a multi-level, multi-mechanism feature extraction and fusion process by combining convolutional neural networks (CNN), Transformer, and graph fusion networks (GFN): CNN is used to extract high-dimensional features of vibration, temperature, and pressure signals; Transformer is used to model and enhance time-series characteristics, strengthening intra-modal temporal representation; and the graph fusion network (GFN) is combined to further optimize inter-modal interactions, achieving deep fusion and enhancement of multi-modal features.

[0018] By leveraging the synergistic effects of CNNs, Transformers, and graph fusion networks, this invention can fully exploit multi-physics features under extremely limited sample conditions, generating reliable high-dimensional multimodal fusion features and significantly improving fault diagnosis capabilities in small-sample scenarios. This invention is particularly suitable for practical engineering scenarios involving high safety in nuclear power plants, challenging data acquisition, and extremely scarce samples.

[0019] 2. In the multimodal feature fusion stage, this invention introduces a graph fusion mechanism based on graph neural networks (GFN) for the first time. GFN can use features from different modalities as graph nodes and dynamically learn the correlation and weights between modalities through an attention mechanism, so that the relationship between multimodal features can be adaptively adjusted according to actual data, maximizing the synergistic effect of various modal features.

[0020] By leveraging graph fusion networks to fully mine and optimize features from small sample data, adaptive optimization and global aggregation of multimodal nodes are achieved. This not only significantly enhances the complementarity of information between modalities and the global expressive power, but also strengthens the model's robustness and discriminative power under complex operating conditions, abnormal patterns, and small sample conditions, overcoming the limitations of traditional multimodal fusion methods under static structures and single mechanisms.

[0021] 3. This invention designs a joint loss function that comprehensively considers classification error and modality consistency, introducing a global optimization objective into the model training process. This effectively avoids the problem of feature imbalance in traditional models when samples are insufficient. This design enables multimodal features to maintain discriminative power while possessing stronger collaborative expression and generalization capabilities. The joint loss function ensures consistency and overall optimization among multimodal features, avoiding performance loss caused by imbalance in a single modality or feature, and further improving the stability and accuracy of the model under conditions of scarce samples.

[0022] 4. Abundant, high-quality data is the foundation for ensuring the performance of intelligent fault diagnosis methods. This invention fuses and augments small sample data of unit components under three physical field distributions for each operating condition. The generated data can characterize the state of equipment components from different dimensions, thus compensating for the low accuracy of intelligent fault diagnosis of circulating water pump equipment under small sample conditions.

[0023] 5. In existing fault diagnosis of rolling bearing equipment, commonly used feature extraction methods mainly include manual extraction based on statistical features (such as mean, variance, kurtosis, peak value, etc.), feature statistics after frequency domain transformation (such as FFT, wavelet transform, etc.), and preliminary feature dimensionality reduction and transformation using shallow machine learning models (such as PCA, LDA). These methods mainly rely on human experience and prior knowledge, and have limited processing capabilities. They have certain limitations in extracting high-dimensional, multimodal, and nonlinearly correlated raw signal information, especially under conditions of scarce samples and complex working conditions, making it difficult to capture deep-level discriminative features.

[0024] This invention employs a deep convolutional neural network (CNN) to automatically extract high-dimensional features from multi-physics field signals. It can directly act on the original physical signals such as vibration, temperature, and pressure. Through multi-layer convolution and nonlinear transformation, it effectively mines the complex temporal, spatial, and modal correlations and local detailed features in the signals, avoiding the problems of feature loss and limited expression.

[0025] For small sample data, this extraction method exhibits stronger generalization ability. By sharing parameters and extracting features at multiple scales, it not only enhances feature diversity but also alleviates overfitting and feature sparsity issues caused by data scarcity in small sample scenarios. Furthermore, the local receptive field and weight sharing mechanism inherent in deep convolutional neural networks can fully utilize the patterns contained in limited samples, thereby improving feature representation ability and model robustness under small sample conditions.

[0026] By independently extracting high-dimensional features for each type of physical field using CNNs, more discriminative and malleable atomic features can be provided for subsequent steps such as feature concatenation, multimodal fusion, and graph structure optimization. These automatically extracted high-dimensional features not only contain the original discriminative information of the signal but also possess multi-level and hierarchical abstraction capabilities, making them more suitable for further deep fusion and global optimization.

[0027] 6. In the comparison files, the Transformer is mainly used as a feature extraction network. It concatenates multimodal signals and directly feeds them into the Transformer layer for unified feature fusion, or processes each modality separately only during feature-level / decision-level fusion. It lacks dedicated intra-modal enhancement and few-shot adaptation mechanisms. Multi-level fusion focuses on the "fusion level" rather than deep enhancement and adaptation mechanisms for each modality.

[0028] In contrast, in this application, the Transformer adopts a hierarchical enhancement strategy, that is, each single modality is first enhanced by an independent Transformer, and then globally fused by a dedicated Transformer after splicing, which significantly improves the intramodal discriminative power and intermodal collaborative expressive power.

[0029] Furthermore, high-order, nonlinear, and dynamic correlations between modalities are subsequently established through Graph Fusion Network (GFN), enabling adaptive optimal path selection and feature synthesis for intermodal information flow. Essentially, this upgrades static feature concatenation to a learnable and interpretable deep modal collaboration mechanism, significantly improving the utilization rate and diagnostic discriminative power of multi-source information.

[0030] 7. Conventional Transformer structures typically include multi-head self-attention, feedforward networks, residual connections, and layer normalization. This invention, based on this classic structure, makes the following improvements for multimodal few-shot scenarios: Adding a modality category embedding layer at the input enables the model to automatically perceive and utilize the semantic information of different signals, improving fusion efficiency. The number of layers and multi-heads in the Transformer are optimized to balance expressiveness and generalization ability, preventing overfitting.

[0031] To address the differences in the distribution of heterogeneous features from multiple sources, layer normalization and residual paths are optimized to ensure the consistency of the numerical range and distribution of signals from each modality during feature fusion. Specifically, this includes introducing LayerNorm operations before and after the input and output of the Transformer's self-attention layer and feedforward layer.

[0032] To improve the robustness of the model in scenarios with noise interference and insufficient samples, stronger Dropout and regularization strategies are introduced. Specifically, Dropout layers are set after the self-attention layer and feedforward layer of the Transformer, and L2 weight decay is used as an additional regularization method; an early stopping strategy is applied during training.

[0033] Through the aforementioned improvements, the Transformer in this application is more suitable for industrial scenarios with extremely scarce samples and high noise interference. Combined with subsequent graph neural networks to achieve end-to-end deep fusion, feature enhancement is not only reflected in the Transformer itself, but also in the entire process of collaborative optimization with GFN.

[0034] 8. Conventional graph fusion networks typically use sensor nodes, spatial points, or only the data samples themselves as nodes, resulting in a single node type and focusing on modeling spatial structure, topology, or relationships within the same mode. This invention uses both "single-modal enhancement features" and "multi-modal features from the splicing of the three" as initial graph nodes. That is, the features of each physical field (such as vibration, temperature, and pressure), as well as the preliminary features of splicing and fusion, are modeled separately in the graph. The node representation is more detailed and richer, taking into account both intra-modal and inter-modal information, thus improving the modeling granularity and fusion accuracy of the graph structure.

[0035] 9. Conventional graph fusion networks often use statically set adjacency matrices, based on spatial distance, relevance thresholds, or prior physical structure, with fixed edge weights, resulting in poor flexibility and adaptability. This invention employs a dynamic weight learning mechanism based on attention, whereby the network automatically learns the relevance weights between nodes (modalities, fusion features) through an end-to-end training process. The relation matrix can be dynamically and adaptively updated with data and training, greatly improving modal collaboration and global representation capabilities.

[0036] This design enables flexible adjustment of information flow between modalities based on the actual distribution and health status characteristics of the samples. It adapts to changes in information structure under small sample and uncertain operating conditions. It also allows different physical field features to retain their individual expressions in the graph structure, while actively complementing and removing redundancy through dynamic edge weights. This achieves deep fusion of high-order information between modalities and enhances the overall feature discrimination power.

[0037] The graph fusion network of this invention is specifically designed for global optimization and robustness improvement of multimodal features. Combined with a preceding CNN-Transformer to enhance features, and a subsequent joint loss function (classification + modality consistency), it achieves end-to-end adaptive enhancement and discriminative power improvement, avoiding local information loss or fragmented optimization, and ultimately achieving optimal feature representation and classification. Attached Figure Description

[0038] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments, wherein: Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a confusion matrix diagram obtained through the present invention. Detailed Implementation

[0039] Example 1 As a basic embodiment of the present invention, the present invention includes a method for enhancing the features of rolling bearing equipment fault diagnosis samples, comprising the following steps: Step S1. Obtain raw data of the nuclear power unit under multi-physics field conditions, including the distribution data of three physical fields: vibration signal, temperature signal and pressure signal generated during the operation of the circulating water pump.

[0040] Step S2. Use a deep convolutional neural network to extract high-dimensional features of different signals in the original data.

[0041] Step S3. Construct and train a joint model based on Transformer and graph fusion network. The Transformer is used to concatenate the high-dimensional features of different signals to obtain preliminary fused features. Sequence enhancement is then performed on the extracted high-dimensional features of different signals and the preliminary fused features to obtain enhanced features of different modalities. The graph fusion network is used to fuse the enhanced features of different modalities to generate multimodal fused features.

[0042] Step S4. Use the trained joint model to generate global multimodal fusion features.

[0043] Example 2 As a preferred embodiment of the present invention, the present invention includes a method for enhancing the features of rolling bearing equipment fault diagnosis samples, comprising the following steps: Step S1. Obtain raw data of the nuclear power unit under multi-physics field conditions, including the distribution data of three physical fields: vibration signal, temperature signal and pressure signal generated during the operation of the circulating water pump.

[0044] Step S2. Extract high-dimensional features of different signals from the original data using a deep convolutional neural network. Specifically, extract high-dimensional features of the vibration signals from the original data using a deep convolutional neural network. High-dimensional characteristics of temperature signals and the high-dimensional characteristics of pressure signals .

[0045] Step S3. Construct and train a joint model based on Transformer and graph fusion network.

[0046] The Transformer is used to concatenate the high-dimensional features of different signals to obtain preliminary fused features. High-dimensional features of different extracted signals ( , , ) and the initial characteristics of integration Sequence enhancement was performed separately to obtain enhanced features for different modalities. , , , .

[0047] The graph fusion network is used to fuse enhanced features from different modalities to generate multimodal fusion features. Specifically, it includes the following steps: Step S 31 Enhancement features of different modalities , , , These are used as initial graph nodes, represented as follows: , , , .

[0048] Step S 32 Construct the graph node set: , In the formula, n Representing the n The graph after the next iteration.

[0049] Step S 33 A dynamic weight learning mechanism based on attention is used to calculate the relationship weights between different modalities. , In the formula, and Indicates the first Nodes in the layer diagram i With nodes j Features W For learnable weight matrix, For attention vectors, This indicates a splicing operation. , .

[0050] Step S 34 By utilizing graph fusion networks to optimize feature representations layer by layer, nodes... i The features are updated by a weighted sum of the features of its neighboring nodes: , in, Represents a node i The set of neighboring nodes, For the first The weight matrix of the layer graph, For the first l The layer's bias vector, after L After the feature maps of the layers are fused, the final feature representation of each node is obtained. .

[0051] Step S 35 Aggregate the features of all modal nodes to obtain multimodal fusion features: , in, This is a multimodal fusion feature.

[0052] Step S4. Use the trained joint model to generate global multimodal fusion features.

[0053] Example 3 In another preferred embodiment of the present invention, the present invention includes a method for enhancing the features of rolling bearing equipment fault diagnosis samples, comprising the following steps: Step S1. Obtain raw data of the nuclear power unit under multi-physics field conditions, including the distribution data of three physical fields: vibration signal, temperature signal and pressure signal generated during the operation of the circulating water pump.

[0054] Step S2. Use a deep convolutional neural network to extract high-dimensional features of different signals in the original data.

[0055] Step S3. Construct and train a joint model based on Transformer and graph fusion network.

[0056] The Transformer is used to concatenate the high-dimensional features of different signals into vectors to obtain preliminary fused features. The single-modal feature enhancement module based on the Transformer performs sequence enhancement on the extracted high-dimensional features of different signals and the preliminary fused features, respectively, to optimize the time series characteristics and the interaction relationship between modes, and obtain enhanced features of different modes.

[0057] The graph fusion network is used to fuse enhanced features from different modalities to generate multimodal fused features. Specifically, it includes: graph node representation, which uses the enhanced multimodal features as graph node input; graph edge weight learning, which optimizes the weighted connections between graph nodes by utilizing the dynamic relationships between modalities; and multi-layer fusion, which constructs an optimized multimodal feature representation through layer-by-layer feature fusion.

[0058] In training the joint model based on Transformer and graph fusion network, a joint loss function is introduced to optimize the parameters of the joint model. The joint loss function is: , , , , , In the formula, For multimodal fusion features, For the sample size, K Number of fault categories For the sample i The true category, For the sample i The predicted probability; L 1 represents the specific loss function. L 2 represents the modal inconsistency loss. is the balance coefficient, and L is the joint loss function. For the process L After the feature maps of the layers are fused, the final features of each node are obtained. .

[0059] Step S4. Use the trained joint model to generate global multimodal fusion features.

[0060] Example 4 In another preferred embodiment of the present invention, the present invention includes a method for enhancing the features of rolling bearing equipment fault diagnosis samples, wherein the equipment is a circulating water pump. (See attached specification.) Figure 1 The method includes the following steps: Step S1. Obtain raw data of the nuclear power unit under multi-physics field conditions, including the distribution data of three physical fields: vibration signal, temperature signal and pressure signal generated during the operation of the circulating water pump.

[0061] Specifically, the hardware includes a circulating water system, motor, circulating water pump (vertical centrifugal pump), regulating valves, water storage manifold, instrumentation and control platform, and piping. To acquire operational data, measuring equipment includes flow meters, thermometers, pressure gauges, vibration sensors, and signal acquisition and demodulation devices. These devices collect vibration, temperature, and pressure signals generated during the operation of the circulating water pump as raw data, which is then used as a small sample for diagnostic tasks.

[0062] Step S2. Use a deep convolutional neural network to extract high-dimensional features of different signals, namely vibration signals, temperature signals, and pressure signals, from the original data to obtain the high-dimensional features of the vibration signals. High-dimensional characteristics of temperature signals and the high-dimensional characteristics of pressure signals : , , , In the formula, It is a vibration signal. Indicates the time step. Indicates the signal dimension; It is a vibration signal. Indicates the time step. Indicates the signal dimension; It is a vibration signal. Indicates the time step. Indicates the signal dimension.

[0063] CNNs are deep convolutional neural networks that extract local features from signals. They possess powerful end-to-end automatic feature learning capabilities, directly applying them to raw physical signals such as vibration, temperature, and pressure. Through multiple layers of convolution and nonlinear transformations, they effectively uncover complex temporal, spatial, and modal correlations and local details within signals, avoiding feature loss and representational limitations. Through parameter sharing and multi-scale feature extraction, CNNs not only enhance feature diversity but also mitigate overfitting and feature sparsity issues caused by data scarcity in small-sample scenarios. The inherent local receptive field and weight sharing mechanism of the convolutional structure fully utilize the patterns contained in limited samples, improving feature representation and model robustness in small-sample scenarios.

[0064] Step S3. Construct and train a joint model based on Transformer and graph fusion network.

[0065] Specifically, existing Transformer structures include multi-head self-attention, feedforward networks, residual connections, and layer normalization. Based on this, the Transformer of this invention makes the following improvements for multimodal few-shot scenarios: The Transformer input layer adds a modal category embedding layer to distinguish different physical field signals through modal labels, enabling the model to automatically perceive and utilize the semantic information of different signals, thereby improving fusion efficiency.

[0066] Considering the small sample size problem in real-world scenarios, the number of layers and the number of multi-head heads in the Transformer were optimized to balance expressiveness and generalization ability while preventing overfitting. Specifically, optimizing the number of layers and multi-head heads involved conducting multiple experiments during model development to test the model's performance on small sample datasets with different combinations of layer numbers (e.g., 2, 3, 4 layers) and the number of multi-head self-attention heads (e.g., 4, 8 heads). The optimal parameter configuration was selected based on the diagnostic accuracy and generalization ability on the validation set.

[0067] To address the differences in the distribution of heterogeneous features from multiple sources, layer normalization and residual paths were optimized to ensure consistency in the numerical range and distribution of signals from different modalities during feature fusion. Specifically, LayerNorm operations were introduced before and after the input and output of the Transformer's self-attention layer and feedforward layer to ensure the consistency of numerical distribution of features from different modalities during fusion, which helps stabilize model training. Secondly, residual connections adopted the standard "input + sub-layer output" approach to ensure smooth gradient flow in deep networks and improve model robustness. In multimodal scenarios, normalization parameters can be set separately or shared according to the actual data distribution to adapt to differences between different modalities.

[0068] To address engineering noise and insufficient sample size scenarios, stronger Dropout and regularization strategies are introduced to further enhance model robustness. Specifically, Dropout layers are added after the self-attention layer and feedforward layer of the Transformer, with Dropout ratios typically set to 0.3 and 0.5, respectively, and the optimal parameters are selected based on actual training performance. Simultaneously, L2 weight decay (typically 1e-45e-4) is used as an additional regularization method to apply regularization constraints to all trainable parameters. Furthermore, an early stopping strategy is applied during training to prevent overfitting under small sample conditions.

[0069] The Transformer is used to concatenate the high-dimensional features of vibration, temperature, and pressure signals into vectors to obtain preliminary fused features: , In the formula, concat Indicates a splicing operation; This is a preliminary characteristic of integration.

[0070] The Transformer is also used to model and enhance the time-series characteristics of the extracted high-dimensional features and the preliminary fused features of different signals, optimize the time-series characteristics and the interaction between modes, and obtain enhanced features of different modes, including single-mode enhancement features of vibration. Single-mode enhancement characteristics of temperature Single-mode enhancement characteristics of pressure and the multimodal features spliced ​​from the three. : , , , .

[0071] Specifically, by utilizing the Transformer's multi-head self-attention mechanism to model the global correlation and dependency relationships of signals across different time steps, it can capture complex dynamic features such as signal trends, periodicity, outliers, and mutations. The Transformer can adaptively mine short-term and long-term correlations in sequences, overcoming the limitations of traditional convolutional / recurrent networks on sequence length and dependency range. Positional encoding further helps the model understand the temporal information of the sequence, enhancing its responsiveness to changes in operating conditions and equipment anomalies. The overall goal is to enable the model to have a higher ability to recognize and express sensitive patterns in time distribution, state transitions, and changes in operating conditions, thereby improving the accuracy and reliability of equipment health diagnosis.

[0072] The graph fusion network is used to optimize and fuse enhanced features from different modalities to generate multimodal fusion features. Specifically, it includes the following steps: Step S 31 The single-mode enhancement features obtained in step S2 Single-mode enhancement characteristics of temperature Single-mode enhancement characteristics of pressure and the multimodal features spliced ​​from the three. These are used as initial graph nodes, represented as follows: , , , .

[0073] Step S 32 Construct the graph node set: , In the formula, n Representing the n The graph after the next iteration.

[0074] Step S 33 A dynamic weight learning mechanism based on attention is used to calculate the relationship weights between different modalities. , In the formula, and Indicates the first Nodes in the layer diagrami With nodes j Features W For learnable weight matrix, For attention vectors, This indicates a splicing operation. , .

[0075] Step S 34 By utilizing graph fusion networks to optimize feature representations layer by layer, nodes... i The features are updated by a weighted sum of the features of its neighboring nodes: , in, Represents a node i The set of neighboring nodes, For the first The weight matrix of the layer graph, For the first l The layer's bias vector, after L After the feature maps of the layers are fused, the final feature representation of each node is obtained. .

[0076] Step S 35 Aggregate the features of all modal nodes to obtain multimodal fusion features: , in, This is a multimodal fusion feature.

[0077] More specifically, during the training of the joint model based on Transformer and graph fusion network, a joint loss function is introduced to optimize the parameters of the joint model through backpropagation, taking into account the classification error of fault prediction and the interaction consistency between modal features.

[0078] The joint loss function is: , , , , In the formula, For multimodal fusion features, For the sample size, K Number of fault categories For the sample i The true category, For the sample i The predicted probability; L 1 represents the specific loss function. L 2 represents the modal inconsistency loss. is the balance coefficient, and L is the joint loss function.

[0079] Optimizing the parameters of the joint model specifically refers to using the Adam optimizer to optimize the parameters of the joint model. , In the formula, The parameters to be learned for the updated joint model. The parameters required to learn the joint model before the update. For learning rate, This is the gradient of the joint loss function.

[0080] Step S4. Utilize the trained joint model to generate global multimodal fusion features, which are then fed into the downstream fault diagnosis model for classification / regression, etc.

[0081] Specifically, multimodal fusion features are used as input to the fault diagnosis model to classify and predict samples. The fault categories predicted by the fault diagnosis model are compared one-to-one with the true categories of the test samples. The frequency of each combination of actual and predicted categories is counted, ultimately forming a confusion matrix. This confusion matrix is ​​then displayed in a visual chart format, as shown in the appendix to the instruction manual. Figure 2 As shown. The values ​​on the diagonal represent the number of samples that were accurately classified, while the values ​​off-diagonal represent the confusion between different categories. (See attached instruction manual.) Figure 2 It can be seen that, under the condition of small sample multimodal data, the multimodal fusion features generated by the joint model of the present invention can achieve high accuracy in identifying different types of faults after being fed into the fault diagnosis model, with low misjudgment and confusion rates.

[0082] In summary, any other corresponding modifications made by those skilled in the art after reading this invention document, without requiring creative mental effort, based on the technical solutions and concepts of this invention, are all within the scope of protection of this invention.

Claims

1. A rolling bearing device fault diagnosis sample feature enhancement method, characterized by: Includes the following steps: Step S1. Obtain raw data of the nuclear power unit under multi-physics field conditions, including the distribution data of three physical fields: vibration signal, temperature signal and pressure signal generated during the operation of the circulating water pump; Step S2. Use a deep convolutional neural network to extract high-dimensional features of different signals in the original data; step S3. Construct and train a joint model based on Transformer and graph fusion network; the Transformer is used to concatenate the high-dimensional features of different signals to obtain the preliminary fused features, and to perform sequence enhancement on the extracted high-dimensional features of different signals and the preliminary fused features to obtain enhanced features of different modalities. The graph fusion network is used to fuse enhanced features from different modalities to generate multimodal fusion features, specifically including the following steps: Step S 31 . The enhanced features of different modalities are respectively taken as initial graph nodes, represented as: ， ， ， ， wherein is a single modal enhancement feature of vibration, is a single modal enhancement feature of temperature, is a single modal enhancement feature of pressure, is a multi-modal feature of the three. Step S 32 . Constructing graph node set: ， In the formula, n represent the first n iteration of the map; Step S 33 . Use a dynamic weight learning mechanism based on attention mechanism to calculate the relationship weight between different modalities: ， wherein and denotes the node in the layer graph i characteristic of the node j , W is a learnable weight matrix, is an attention vector, denotes a concatenation operation, , ; Step S 34 . The feature representation is optimized layer by layer using the graph fusion network, the node i feature is updated by the weighted sum of its neighbor node features: ， in, Represents a node i The set of neighboring nodes, For the first The weight matrix of the layer graph, For the first l The layer's bias vector, after L After the feature maps of the layers are fused, the final feature representation of each node is obtained. ; Step S 35 . Aggregate the features of all modal nodes to obtain a multi-modal fusion feature: ， wherein, is a multi-modal fusion feature; Step S4. Use the trained joint model to generate global multimodal fusion features.

2. The method for enhancing the features of rolling bearing equipment fault diagnosis samples according to claim 1, characterized in that: The step S2 specifically refers to: using a deep convolutional neural network to extract high-dimensional features of the vibration signal, high-dimensional features of the temperature signal and high-dimensional features of the pressure signal in the original data respectively. :​​ ， ， ， In the formula, It is a vibration signal. Indicates the time step. Indicates the signal dimension; It is a vibration signal. Indicates the time step. Indicates the signal dimension; It is a vibration signal. Indicates the time step. The dimension of the signal is represented by ; CNN is a deep convolutional neural network that extracts local features of a signal.

3. The method of claim 1, wherein: A modal category embedding layer is added to the input of the Transformer to distinguish different physical field signals through modal labels; LayerNorm operations are introduced before and after the input and output of the Transformer's self-attention layer and feedforward layer.

4. The method for enhancing the features of rolling bearing equipment fault diagnosis samples according to claim 3, characterized in that: Dropout layers are set after the self-attention layer and feedforward layer of the Transformer, and L2 weight decay is used as an additional regularization method. The early stop strategy is applied during training.

5. The method of claim 4, wherein: By concatenating the high-dimensional features of different signals, a preliminary fused feature is obtained, specifically: ， In the formula, concat Indicates a splicing operation; These are the initial characteristics of integration; For the high-dimensional characteristics of vibration signals, This represents a high-dimensional feature of the temperature signal. This represents the high-dimensional characteristics of the pressure signal.

6. The method for enhancing the features of rolling bearing equipment fault diagnosis samples according to claim 5, characterized in that: Different modalities of augmentation features include single modality augmentation features of vibration , temperature , pressure , and multi-modality features of the three ; the sequence augmentation method is: ， ， ， 。 7. The method according to claim 1 or 6, characterized in that: During the training of the joint model based on Transformer and graph fusion network, a joint loss function is introduced to optimize the parameters of the joint model.

8. The method of claim 7, wherein the method further comprises: The joint loss function is: ， ， ， ， In the formula, For multimodal fusion features, For the sample size, K Number of fault categories For the sample i The true category, For the sample i The predicted probability; L 1 represents the specific loss function. L 2 represents the modal inconsistency loss. L is the balance coefficient, and L is the joint loss function; For the process L After the feature maps of the layers are fused, the final features of each node are obtained. .

9. The method for enhancing the features of rolling bearing equipment fault diagnosis samples according to claim 7, characterized in that: Optimizing the parameters of the joint model specifically refers to using the Adam optimizer to optimize the parameters of the joint model.

Images