A method and system for enhancing vibration signal data

By using temporal coding with multi-scale convolution and attention mechanisms, combined with generative adversarial networks, the problem of insufficient feature encoding and fusion in existing vibration signal data enhancement methods is solved, generating more realistic and diverse vibration signal samples that are suitable for fault diagnosis of rotating machinery.

CN123116153APending Publication Date: 2026-11-17INST OF ENGINEERING THERMOPHYSICS - CHINESE ACAD OF SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611348301.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing vibration signal data enhancement methods based on generative adversarial networks suffer from problems such as limited feature encoding methods, insufficient fusion of multi-source features, and incomplete evaluation systems, resulting in incomplete feature representation, insufficient information fusion, and inadequate quality evaluation of generated samples.

Method used

We employ multi-scale convolution and attention-based temporal coding, and train generators and discriminators alternately. The generator has parallel multi-scale coding branches and temporal coding branches. We use adaptive weights for feature fusion and improve training stability through generative adversarial loss training.

Benefits of technology

It achieves simultaneous characterization of multi-scale local features and long-range temporal dependence features of vibration signals, improves the authenticity and diversity of generated samples, and is applicable to fault diagnosis models of rotating machinery equipment, alleviating the problems of data scarcity and class imbalance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116153A_ABST
    Figure CN123116153A_ABST
Patent Text Reader

Abstract

This invention discloses a vibration signal data enhancement method and system, comprising: acquiring vibration signal samples under different health state categories of rotating machinery and preprocessing them to obtain random noise; for each health state category, inputting the random noise into a generator, which maps it to initial noise features through a noise mapping layer within the generator; performing multi-scale feature encoding and attention-based temporal feature encoding on the initial noise features to obtain multiple scale features and temporal features; determining adaptive weights for each scale feature based on the temporal features and weighted aggregating multiple scale features to obtain fused features; inputting the random noise into the generator to generate generated samples for the corresponding category; inputting the generated samples and initial noise features into a discriminator to train the generator and discriminator; and using the trained generator to generate enhanced samples based on random noise. This improves the authenticity, diversity, and diagnostic usability of generated samples under conditions of scarce fault samples and class imbalance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment condition monitoring and intelligent fault diagnosis technology, and specifically discloses a vibration signal data enhancement method and system. Background Technology

[0002] Vibration signals are the most commonly used information source for monitoring the health status and diagnosing faults in rotating machinery. In engineering practice, collecting vibration signals by installing sensors at key measuring points of equipment and using signal analysis and pattern recognition methods to determine the equipment's operating status has become a widely adopted technical approach in industry. In recent years, with the rapid development of deep learning technology, intelligent fault diagnosis methods based on deep models such as convolutional neural networks and long short-term memory networks have achieved significant results in fault identification of key equipment such as rolling bearings, gearboxes, and aero engines.

[0003] However, the effectiveness of deep learning methods heavily relies on large-scale, high-quality, class-balanced training datasets. In real-world industrial scenarios, rotating machinery operates normally most of the time, with low frequency and short duration of various faults. Furthermore, the identification and recording of fault states require expert judgment and offline detection methods. This results in a significant imbalance between normal and faulty samples in the actual collected datasets. Data scarcity and class imbalance together limit the accuracy and generalization ability of deep learning-based fault diagnosis models.

[0004] To address the issues of data scarcity and imbalance, generative model-based data augmentation techniques have become a mainstream research direction, including variational autoencoders, diffusion models, and generative adversarial networks (GANs). Variational autoencoders learn the latent distribution of data and sample new samples from it; however, their training relies on assumed loss functions and KL divergence to approximate the true distribution, often resulting in overly smooth samples lacking high-frequency details. Diffusion models learn the data distribution through parameterized Markov chains that progressively add and remove noise, but their application in one-dimensional vibration signals is still immature, and training requires significant computational resources. GANs learn the data distribution characteristics of input samples through alternating adversarial interactions between generators and discriminators, enabling the generation of more realistic samples. However, the original GANs used JS divergence as a distance metric between distributions, making them prone to gradient vanishing and mode collapse during training.

[0005] Existing data augmentation methods based on generative adversarial networks still have the following technical shortcomings: First, the feature encoding methods are too simplistic and cannot fully characterize the complex characteristics of vibration signals. Most existing methods use only a single network structure for feature encoding. However, vibration signals are one-dimensional time series with complex structures, containing both local impact features reflecting different frequency components and periodic temporal features reflecting the signal's evolution. A single network structure cannot simultaneously capture both types of features, resulting in incomplete feature representation of the generated samples.

[0006] Second, the methods for multi-source feature fusion are limited and fail to effectively integrate different features. When using multiple networks to extract multi-source features, existing methods typically employ simple feature concatenation or feature summation for fusion. This simple fusion approach cannot capture the nonlinear relationships between different features, nor can it achieve adaptive weighted interactions between features, resulting in insufficient fusion of multi-scale and temporal information.

[0007] Third, the evaluation system is not comprehensive enough, making it difficult to measure the quality of generated data from multiple dimensions. Existing studies often focus on a single indicator when evaluating generated data, lacking a comprehensive evaluation system that considers multiple dimensions such as distribution fitting ability, preservation of physical properties, sample authenticity and diversity. Summary of the Invention

[0008] To overcome the above deficiencies, the present invention provides a vibration signal data enhancement method and system that can simultaneously characterize the multi-scale local features and long-range temporal dependence features of vibration signals, thereby improving the authenticity, diversity, and diagnostic usability of generated samples under conditions of scarce fault samples and class imbalance.

[0009] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: In one aspect, the present invention provides a method for enhancing vibration signal data, comprising: Vibration signal samples of rotating machinery under different health state categories are acquired, and the vibration signal samples are preprocessed to obtain random noise for the corresponding health state category. For each health state category, the random noise is input into a generator, and the random noise is mapped to initial noise features through a noise mapping layer in the generator. The initial noise features are subjected to multi-scale feature encoding and attention-based temporal feature encoding to obtain multiple scale features and temporal features. Based on the temporal features, adaptive weights corresponding to each scale feature are determined, and the multiple scale features are weighted and aggregated according to the adaptive weights to obtain fused features; Random noise is input into a generator corresponding to the health state category to be generated, generating samples of the health state category; the generated samples and the initial noise features, which are used as real samples, are input into a discriminator, and the generator and the discriminator are trained based on generative adversarial loss; Using a post-trained generator corresponding to the health state category to be enhanced, random noise is input into the post-trained generator to generate generated samples of the health state category to be enhanced. The generated samples are then de-normalized according to the normalization parameters corresponding to the health state category to be enhanced to obtain enhanced samples of the health state category to be enhanced.

[0010] Optionally, the preprocessing of the vibration signal samples includes: The vibration signal sample is filtered to obtain a filtered vibration signal sample. The filtered vibration signal sample is segmented using a sliding window to obtain multiple sample segments of equal length. The multiple equal-length sample segments are normalized according to the health status category, so that the sample segments under the same health status category have zero mean and unit variance. Record the mean and standard deviation used for each health status category during the normalization process, and use them as the normalization parameters for each health status category; The normalized sample fragments are labeled with their corresponding health status categories to obtain the initial noise features.

[0011] Optionally, the multi-scale feature encoding of the initial noise features includes: The initial noise features are subjected to a first-level multi-scale convolution process to obtain the first-scale features; The first-scale features are downsampled, and the downsampled features are then subjected to a second-level multi-scale convolution to obtain the second-scale features. Continue downsampling and multi-scale convolution processing in a cascaded manner to obtain at least one subsequent scale feature; Each level of multi-scale convolution processing includes: performing multi-path convolution processing on the input features to obtain intermediate features under different receptive fields; concatenating the intermediate features along the channel dimension and performing convolution fusion on the concatenated features to obtain convolution fusion features; and performing residual connection between the convolution fusion features and the input features to obtain the scale features corresponding to that level of multi-scale convolution processing.

[0012] Optionally, the multi-path convolution processing includes: One-dimensional convolution processing with different kernel sizes was employed. One-dimensional dilated convolution with different dilation rates is used for processing; Continuous one-dimensional convolution with the same kernel size is used for processing; Feature encoding is performed using a combination of convolution and pooling processes. The intermediate features obtained from different paths correspond to different receptive fields.

[0013] Optionally, the downsampling process includes max pooling, average pooling, or stride convolution. The scale feature obtained from each level of multi-scale convolution processing is used as one of the multiple scale features; Except for the last stage of multi-scale convolution processing, the scale features obtained from each stage of multi-scale convolution processing are downsampled and then used as the input features for the next stage of multi-scale convolution processing.

[0014] Optionally, the temporal feature encoding of the initial noise features based on an attention mechanism includes: The initial noise features are mapped to a feature space of a preset dimension through linear mapping, and positional encoding is added to obtain the embedded features; The embedded features are encoded using a multi-head self-attention mechanism to model the dependencies between different time steps in the initial noise features; The attention output obtained by the encoding process is subjected to a feedforward transformation, and residual connection and layer normalization are performed after the multi-head self-attention mechanism and the feedforward transformation, respectively. The features obtained from each time step are weighted and fused to obtain a global time series representation. The global time series representation is subjected to feature mapping to obtain the time series features.

[0015] Optionally, the step of determining the adaptive weights corresponding to each scale feature based on the temporal features, and weighting and aggregating the multiple scale features according to the adaptive weights to obtain the fused features includes: Perform global pooling on each scale feature to obtain the global feature vector corresponding to each scale feature; Each of the global feature vectors is mapped to the same feature dimension as the temporal feature; Calculate the similarity between the temporal features and each of the global feature vectors, and normalize the similarity to obtain the adaptive weights corresponding to each scale feature; The feature transformations of the multiple scale features are performed respectively to obtain the fusion vector corresponding to each scale feature; The weighted summation of each fusion vector is performed according to the adaptive weights to obtain the weighted aggregated features; The weighted aggregated features are fused with the temporal features to obtain the fused features.

[0016] Optionally, fusing the weighted aggregated features with the temporal features to obtain the fused features includes: The weighted aggregated features are concatenated, added, or residual-connected with the temporal features to obtain the fused features; The fused features are subjected to layer normalization and feedforward transformation to obtain the fused features. The similarity includes cosine similarity, dot product similarity, or a relevance score calculated by a neural network.

[0017] Optionally, training the generator and the discriminator based on the generative adversarial loss includes: For each health status category, the initial noise features of that health status category are used as real samples, and random noise is input into the generator to obtain the generated samples; Based on the discriminator's output on the initial noise features and the generated samples, a discriminator loss function and a generator loss function are constructed. Based on the discriminator loss function and the generator loss function, the network parameters of the discriminator and the generator are updated alternately; The discriminator loss function includes a gradient penalty term constructed based on random interpolated samples between the initial noise features and the generated samples.

[0018] In another aspect, the present invention provides a vibration signal data enhancement system, comprising: The preprocessing module is used to acquire vibration signal samples of rotating machinery under different health state categories, and preprocess the vibration signal samples to obtain random noise for the corresponding health state category; for each health state category, the random noise is input into the generator, and the random noise is mapped to initial noise features through the noise mapping layer in the generator; The feature encoding module is used to perform multi-scale feature encoding and attention-based temporal feature encoding on the initial noise features to obtain multiple scale features and temporal features. The weighted aggregation module is used to determine the adaptive weights corresponding to each scale feature based on the temporal features, and to perform weighted aggregation on the multiple scale features based on the adaptive weights to obtain the fused features; The training module is used to set up corresponding generators and discriminators for different health status categories. Random noise is input into the generator corresponding to the health status category to be generated to generate generated samples of the health status category. The generated samples and the initial noise features belonging to the health status category are input into the corresponding discriminator, and the generator and the discriminator are trained based on the generative adversarial loss. The enhanced sample generation module is used to input random noise into the trained generator corresponding to the health state category to be enhanced, generate generated samples of the health state category to be enhanced, and perform inverse normalization processing on the generated samples according to the normalization parameters corresponding to the health state category to be enhanced, so as to obtain the enhanced samples of the health state category to be enhanced.

[0019] Compared with the closest existing technology, the present invention has the following advantages: This invention proposes a vibration signal data enhancement method and system. By setting up parallel multi-scale coding branches and temporal coding branches inside the generator, the initial noise features obtained by noise mapping are encoded in dual streams. The multi-scale local correlation structure and long-range temporal dependency structure of the vibration signal are injected as structural priors into the generation process, enabling the generator to simultaneously simulate local impacts, multi-frequency components and long-range temporal dependencies in the vibration signal.

[0020] This invention determines the adaptive weights corresponding to each scale feature based on the temporal features, so that different scale features can participate in the fusion according to their correlation with the temporal features, avoiding insufficient information utilization caused by simple splicing or fixed weight fusion.

[0021] This invention trains or sets corresponding generators and discriminators for different health status categories. The generator corresponding to the health status category to be generated is used to process random noise, thereby generating generated samples for the corresponding health status category. The generator does not accept labels or fused features as input, thus avoiding inconsistencies between the technical solution description and the actual network structure.

[0022] This invention improves training stability and alleviates gradient vanishing and mode collapse problems by training the generator and discriminator with a generative adversarial loss that includes a gradient penalty term.

[0023] This invention is applicable to scenarios where fault data for rotating machinery is scarce and the categories are unbalanced, and can provide more balanced and richer initial noise features for subsequent fault diagnosis models. Attached Figure Description

[0024] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0025] Figure 1 This is a schematic diagram of the overall process of the vibration signal data enhancement method provided in the embodiments of the present invention; Figure 2This is a schematic diagram of the multi-scale convolution processing flow provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a multi-level, multi-scale feature encoding process provided in an embodiment of the present invention; Figure 4 A schematic diagram of the temporal feature encoding process based on the attention mechanism provided in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the weighted aggregation of multiple scale features based on temporal features, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram comparing the time-domain waveforms of the generated sample and the real sample provided in an embodiment of the present invention. Figure 7 This is a schematic diagram comparing the time-domain waveforms of generated samples and real samples under different datasets provided in this embodiment of the invention. Figure 8 This is a schematic diagram illustrating the quality evaluation results of generated samples from different methods in the bearing dataset provided in this embodiment of the invention. Figure 9 This diagram illustrates the quality evaluation results of generated samples from different methods in the gear dataset provided in this embodiment of the invention.

[0026] Figure 10 This is a schematic diagram of the vibration signal data enhancement system provided in an embodiment of the present invention; Figure 11 This is a diagram of the internal structure of the electronic device provided by the present invention. Detailed Implementation

[0027] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention.

[0028] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0029] This application provides a vibration signal data enhancement method and system, which can be applied to vibration signal enhancement, fault sample expansion, intelligent diagnostic model training, and equipment health status identification of rotating machinery such as aero-engines, wind turbine generators, rolling bearings, gearboxes, and motor transmission systems. Embodiments of the invention are described below with reference to the accompanying drawings.

[0030] Example 1: Example 1 of the present invention provides a method for enhancing vibration signal data, such as... Figure 1 As shown, the method includes the following steps: S101, acquire vibration signal samples of rotating machinery under different health state categories, and preprocess the vibration signal samples to obtain random noise for the corresponding health state category; for each health state category, input the random noise into the generator, and map the random noise into initial noise features through the noise mapping layer in the generator; Specifically, accelerometers are installed at measurement points on rotating machinery to collect one-dimensional vibration signal samples under different health status categories. The rotating machinery can be a bearing, gearbox, aircraft engine, wind turbine generator set, or motor drive system. The health status categories include a normal state and at least one fault state, which can include bearing inner ring fault, bearing outer ring fault, rolling element fault, gear cracks, broken teeth, spalling, etc.

[0031] When preprocessing vibration signal samples, the samples are first filtered to reduce noise interference. Then, a sliding window is used to segment the filtered vibration signal samples, resulting in multiple sample segments of equal length. The sliding window can use overlapping or non-overlapping sampling. For example, the window length can be set to 4096 sampling points or 3600 sampling points.

[0032] Subsequently, the sample segments were normalized according to the health status category, so that sample segments within the same health status category had zero mean and unit variance. The normalization process can be expressed as:

[0033] In the formula, μ and σ represent the mean and standard deviation of all initial noise features under this health state category, respectively. The original input sequence data, This represents the initial noise characteristics after normalization.

[0034] Finally, the normalized sample fragments are labeled with their corresponding health status categories to obtain the initial noise features.

[0035] S102, for each health state category, random noise is input into the generator, and the initial noise features are obtained by mapping the random noise through the two-stream coding network in the generator.

[0036] Specifically, such as Figure 2 and Figure 3 As shown, when encoding the initial noise features at multiple scales, multi-level convolutional processing can be used.

[0037] In the first-level multi-scale convolution processing, the initial noise features are processed by multi-path convolution to obtain intermediate features under different receptive fields; the intermediate features are concatenated in the channel dimension, and the concatenated features are convolved and fused to obtain convolutional fused features; the convolutional fused features are residually connected with the input features to obtain the first-scale features.

[0038] The first-scale features are downsampled, and then a second-level multi-scale convolution is performed on the downsampled features to obtain the second-scale features. Continuing with downsampling and multi-scale convolution in a cascaded manner yields at least one subsequent-scale feature.

[0039] In one specific implementation, multi-scale feature encoding includes three levels of multi-scale convolution processing to obtain first-scale, second-scale, and third-scale features, respectively. Except for the last level of multi-scale convolution processing, the scale features obtained from each level of multi-scale convolution processing are downsampled and used as input features for the next level of multi-scale convolution processing. The downsampling processing can be max pooling, average pooling, or stride convolution.

[0040] Multi-path convolution processing can include at least two of the following: one-dimensional convolution with different kernel sizes; one-dimensional dilated convolution with different dilation rates; continuous one-dimensional convolution with the same kernel size; and feature encoding using a combination of convolution and pooling. The intermediate features obtained from different paths correspond to different receptive fields.

[0041] For example, when using consecutive one-dimensional convolutions with the same kernel size, let the input feature be I. After passing through three one-dimensional convolutions, batch normalization, and activation, intermediate features F1, F2, and F3 are obtained:

[0042]

[0043]

[0044] F1, F2, and F3 are concatenated along the channel dimension, then fused by convolution, and finally residual-connected with the input feature I to obtain the corresponding scale feature. :

[0045] in, This represents one-dimensional convolution processing, BN represents batch normalization processing, ReLU represents activation processing, and Concat represents concatenation along the channel dimension.

[0046] When performing temporal feature encoding on the initial noise features based on an attention mechanism, the initial noise features are first mapped to a feature space of a preset dimension through a linear mapping, and positional encoding is added to obtain embedded features. Then, the embedded features are encoded using a multi-head self-attention mechanism to model the dependencies between different time steps in the initial noise features.

[0047] like Figure 4 As shown, in the encoding process, the attention output obtained by the multi-head self-attention mechanism is subjected to residual connection and layer normalization. Then, the processing result is subjected to feedforward transformation, and residual connection and layer normalization are performed after the feedforward transformation. After at least one of the above encoding processes, the features of each time step are weighted and fused to obtain a global temporal representation. The global temporal representation is then subjected to feature mapping to obtain temporal features.

[0048] In one specific implementation, the weighted fusion of features from each time step can be expressed as:

[0049]

[0050]

[0051] in, For the first in the input sequence Feature vectors at each time step ( ), , , For learnable parameters, For the first Unnormalized attention score at each time step For the first The contribution ratio of each time step to the global representation satisfies , This is the output of the adaptive weighted fusion module.

[0052] Finally, it passes through the fully connected layer. Output time series characteristics : ; in, These are temporal features extracted based on an attention mechanism. Represents the global timing representation Perform linear mapping / fully connected transformation to convert... Transformed into time-series features of preset dimensions .

[0053] S103. Determine the adaptive weights corresponding to each scale feature based on the temporal features, and then perform weighted aggregation of multiple scale features based on the adaptive weights to obtain the fused features.

[0054] Specifically, such as Figure 5 As shown, global pooling is performed on each scale feature to obtain the global feature vector corresponding to each scale feature; each global feature vector is mapped to the same feature dimension as the temporal feature; the similarity between the temporal feature and each global feature vector is calculated, and the similarity is normalized to obtain the adaptive weight corresponding to each scale feature.

[0055] In one specific implementation, let multiple scale features be... , , Global average pooling is performed on each scale feature, and a global feature vector is obtained through a fully connected mapping. :

[0056] in, l Indicates the scale number. For the first The layer depth features correspond to the global features, and GAP represents global average pooling. , For the first Learnable parameters corresponding to layer depth features.

[0057] Then calculate the time series features. With each global feature vector Cosine similarity between them:

[0058] The similarity is normalized to obtain adaptive weights:

[0059] in, For time series features and the first Cosine similarity of global depth features at each layer. For the first The similarity between scale features and time-series features; For scale features, =1,2,…,L; For the first The adaptive weights corresponding to the global depth features of the layer are 1, and the sum of the adaptive weights corresponding to each scale feature is 1. L represents the number of scale features. In this embodiment, L=3. In addition to cosine similarity, similarity can also be expressed as dot product similarity or correlation score calculated by neural network.

[0060] Subsequently, as Figure 5 As shown, feature transformations are performed on features at multiple scales to obtain the fusion vector corresponding to each scale feature. For example, one-dimensional convolutional dimensionality reduction and global pooling can be performed on each scale feature:

[0061] The weighted aggregate features are obtained by weighting and summing the fusion vectors according to the adaptive weights.

[0062] in, For the first The fusion vector of layer depth features. The multi-scale fusion feature is obtained by weighted summation of the multi-scale fusion vectors.

[0063] Finally, the weighted aggregate features With time series characteristics The features are then fused to obtain the fused features. The fusion method can be concatenation, addition, or residual connection. The fused features are further processed through layer normalization and feedforward transformation to obtain the final fused features. :

[0064] in, For the final fusion features, Indicates feedforward transform, Presentation layer normalization processing.

[0065] S104, input random noise into the generator corresponding to the health state category to be generated, generate a sample of the health state category; input the generated sample and the initial noise feature as the real sample into the discriminator, and train the generator and the discriminator based on the generative adversarial loss.

[0066] Specifically, random noise z is sampled from a standard normal distribution or a uniform distribution, and mapped to initial noise features through a fully connected process. The initial noise features are then input into a generator corresponding to the health status category to be generated.

[0067] In one specific implementation, the random noise z is 100-dimensional Gaussian noise. After being processed by a multi-layer fully connected layer and mapped to initial noise features, it is input into a generator corresponding to the health state category to be generated. The generator performs multi-scale feature encoding, temporal feature encoding and weighted aggregation on the initial noise features to generate generated samples with the same length as the initial noise features.

[0068] The generated samples are generated by a generator corresponding to the target health state category. The generator is trained adversarially using the real initial noise features of the health state category, and thus can learn the vibration signal distribution features of the initial noise features of the health state category.

[0069] Furthermore, the initial noise features are used as input to the discriminator from real samples, and the generated samples output by the generator are used as input to the discriminator from generated samples. The discriminator outputs the score corresponding to the input sample. Based on the discriminator's output results for real samples and generated samples, the discriminator loss function and the generator loss function are constructed.

[0070] In one specific implementation, a generative adversarial training method with a gradient penalty term is employed. Let the distribution of the real samples be... The generated sample distribution is The discriminator is D, the generator is G, and the discriminator loss function is... for:

[0071] in, Random interpolation points on the line connecting the real sample and the generated sample ( ε follows a uniform distribution between 0 and 1; This indicates that the sample follows a generated sample distribution. samples x Seeking expectations, This represents the expected value of the discriminator score on the real samples. Indicates interpolated samples The distribution, The gradient penalty coefficient is... This indicates that the discriminator model is effective for samples. The gradient of the generator. The generator's loss function. for:

[0072] During training, the network parameters of the discriminator and generator are updated alternately based on the discriminator loss function and the generator loss function. For example, the generator can be updated once after every 5 discriminator updates, until a preset number of training epochs is reached or the loss function meets the convergence condition.

[0073] S105, using the post-trained generator corresponding to the health state category to be enhanced, random noise is input into the post-trained generator to generate generated samples of the health state category to be enhanced, and the generated samples are de-normalized according to the normalization parameters corresponding to the health state category to be enhanced to obtain enhanced samples of the health state category to be enhanced.

[0074] Specifically, such as Figure 1As shown, for health status categories to be enhanced that have insufficient sample quantity or are class imbalanced, the post-training generator corresponding to the health status category is first determined, then random noise is sampled and input into the post-training generator to generate enhanced samples for the health status category.

[0075] The generated augmented samples can be combined with the original initial noise features to form an augmented training set, which can be used to train subsequent fault diagnosis models, thereby improving the classification performance and generalization ability of the fault diagnosis model under conditions of small sample size and class imbalance. The evaluation results of the augmented samples can be as follows: Figure 8 and Figure 9 As shown, where Figure 8 Used to illustrate the quality evaluation results of generated samples from different methods in the bearing dataset. Figure 9 This is used to show the quality evaluation results of generated samples from different methods in the gear dataset.

[0076] The inverse normalization formula is:

[0077] in, Generate samples for the normalized domain output by the generator. For the enhanced samples after denormalization, and These are the mean and standard deviation corresponding to the target health status category c, respectively.

[0078] The method further includes: The maximum mean difference is used to evaluate the distributional difference between the generated samples and the real samples; The difference in temporal statistical characteristics between the generated sample and the real sample is evaluated using temporal statistical feature similarity. The classification performance of the generated samples is evaluated by training and testing the classifier based on the generated samples to assess the class distinguishability and diagnostic usability of the generated samples. The diversity of the generated samples is evaluated using the feature distance between generated samples within the same health status category; By performing a joint analysis of the above evaluation results, the combined quality evaluation results of the generated samples are obtained.

[0079] Example 2: This example uses the vibration signal of a rolling bearing to illustrate the specific application of the present invention.

[0080] Vibration signals of rolling bearings under normal conditions, rolling element faults, inner ring faults, and outer ring faults were collected. Each type of fault can include different fault sizes. The collected vibration signals were divided into training data and test data according to a preset ratio. The training data was segmented using a sliding window with a length of 4096 sampling points, and the sample segments under each health condition category were normalized to obtain initial noise characteristics.

[0081] In multi-scale feature encoding, such as Figure 2 and Figure 3 As shown, a three-stage multi-scale convolution processing method is employed. The first stage of multi-scale convolution processing can have 32 output channels, the second stage can have 64 output channels, and the third stage can have 128 output channels. Max pooling is used for downsampling between adjacent stages of multi-scale convolution processing.

[0082] In attention-based temporal feature encoding, such as Figure 4 As shown, the initial noise features are mapped to a 64-dimensional feature space, and after adding positional encoding, they are input into a two-layer encoding process based on a multi-head self-attention mechanism. Each encoding layer includes eight attention heads and a feedforward transform, with the intermediate dimension of the feedforward transform being 256. Finally, the features at each time step are weighted, fused, and mapped to obtain the temporal features.

[0083] In time-guided multi-scale feature fusion, such as Figure 5 As shown, the scale features output by the three-level multi-scale convolution are subjected to global average pooling and mapped to the same dimension as the temporal features. Cosine similarity is used to calculate the similarity between the temporal features and the global feature vectors of each scale. The similarity is normalized to obtain the adaptive weights corresponding to each scale feature. Then, the fusion vectors of each scale are weighted and summed according to the adaptive weights, and residual connection, layer normalization and feedforward transformation are performed with the temporal features to obtain the fusion features.

[0084] During the sample generation process, such as Figure 1 As shown, the random noise can be 100-dimensional Gaussian noise. The random noise is mapped to initial noise features through fully connected processing, and the initial noise features are input into the generator corresponding to the health state category to be generated. Through multi-scale feature encoding, temporal feature encoding and weighted aggregation, generated samples of length 4096 are generated.

[0085] During training, such as Figure 1As shown, the discriminator employs multi-layer one-dimensional convolution, global pooling, and fully connected processing to output sample scores. During training, initial noise features are used as real samples, and generated samples are used as generated samples. A generative adversarial loss with a gradient penalty term is used to train the generator and discriminator alternately. After training, for the healthy state category with insufficient sample numbers, augmented samples are generated based on random noise.

[0086] The time-domain waveforms of the generated samples and the real samples can be compared as follows: Figure 6 As shown in the figure. In the experiment, the maximum mean difference, temporal statistical feature similarity, and methods based on training and testing classifiers using generated samples can be used to evaluate the distribution fitting ability, physical property preservation ability, authenticity, and diversity of the generated samples. The evaluation results can be as follows: Figure 8 As shown. Among them, time-domain statistical characteristics can include mean, standard deviation, skewness, kurtosis, peak factor, and impulse factor.

[0087] Example 3: This embodiment uses gearbox vibration signal as an example to illustrate the applicability of the present invention.

[0088] Vibration signals of the gearbox were collected under various conditions, including healthy, missing teeth, cracks, spalling, and different degrees of sharpening faults. Each type of vibration signal was divided into training and testing data according to a preset ratio. Segmented sampling was performed using a sliding window with 3600 sampling points, and normalization was performed according to the health status category.

[0089] Then, following steps S102 to S105 in Example 1, multi-scale feature encoding, attention-based temporal feature encoding, fusion feature generation, sample generation, generator and discriminator training, and augmented sample generation are performed respectively. The multi-scale feature encoding process can be as follows: Figure 2 and Figure 3 As shown, the temporal feature encoding process based on the attention mechanism can be as follows: Figure 4 As shown, the feature fusion generation process can be as follows: Figure 5 As shown, the time-domain waveforms of the generated sample and the real sample can be compared as follows: Figure 7 As shown, the sample quality assessment results can be generated as follows: Figure 9 As shown.

[0090] Because the generator is trained independently for each health state category and learns the data distribution of the initial noise features for that category, it can generate corresponding augmented samples for different health state categories of the gearbox.

[0091] This embodiment demonstrates that the present invention is not dependent on a specific type of rotating machinery, nor is it limited to a specific sampling frequency or sample length. For vibration signals from different devices, with different fault types and different sampling lengths, the present invention can achieve data augmentation by adjusting the sliding window length, the number of convolution channels, the number of coding layers, and the training parameters.

[0092] Example 4: Based on the same technical concept, Example 4 of the present invention also provides a vibration signal data enhancement system, such as... Figure 10 As shown, it includes: a preprocessing module 210, a feature encoding module 220, a weighted aggregation module 230, a training module 240, and an enhanced sample generation module 250. Wherein: The preprocessing module 210 is used to acquire vibration signal samples of rotating machinery under different health state categories, and preprocess the vibration signal samples to obtain the initial noise features of the corresponding health state category; for each health state category, random noise is input into the generator, and the random noise is mapped to the initial noise features through the noise mapping layer in the generator. The feature encoding module 220 is used to perform multi-scale feature encoding and attention-based temporal feature encoding on the initial noise features to obtain multiple scale features and temporal features. The weighted aggregation module 230 is used to determine the adaptive weights corresponding to each scale feature based on the temporal features, and to perform weighted aggregation on the multiple scale features based on the adaptive weights to obtain fused features; The training module 240 is used to set up corresponding generators and discriminators for different health state categories. Random noise is input into the generator corresponding to the health state category to be generated to generate generated samples of the health state category. The generated samples and the initial noise features belonging to the health state category are input into the corresponding discriminator, and the generator and the discriminator are trained based on the generative adversarial loss. The enhanced sample generation module 250 is used to generate samples of the health state category to be enhanced by inputting random noise into the trained generator corresponding to the health state category to be enhanced, and to perform inverse normalization processing on the generated samples according to the normalization parameters corresponding to the health state category to be enhanced, so as to obtain the enhanced samples of the health state category to be enhanced.

[0093] Example 5: In one embodiment, Example 5 of the present invention also provides an electronic device; the electronic device may be an edge control terminal or an array master control device, and its internal structure diagram may be as follows. Figure 11As shown. The electronic device includes a processor, memory, communication interface, input device, and display unit connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. When the computer program is executed by the processor, it implements the vibration signal data enhancement method described in any one of steps S101 to S105.

[0094] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the electronic device to which the present invention is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0095] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0096] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0097] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0098] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0099] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A method for enhancing vibration signal data, characterized in that, The method includes: Vibration signal samples of rotating machinery under different health state categories are acquired, and the vibration signal samples are preprocessed to obtain random noise for the corresponding health state category. For each health state category, the random noise is input into a generator, and the random noise is mapped to initial noise features through a noise mapping layer in the generator. The initial noise features are subjected to multi-scale feature encoding and attention-based temporal feature encoding to obtain multiple scale features and temporal features. Based on the temporal features, adaptive weights corresponding to each scale feature are determined, and the multiple scale features are weighted and aggregated according to the adaptive weights to obtain fused features; Random noise is input into a generator corresponding to the health state category to be generated, generating samples of the health state category; the generated samples and the initial noise features, which are used as real samples, are input into a discriminator, and the generator and the discriminator are trained based on generative adversarial loss; Using a post-trained generator corresponding to the health state category to be enhanced, random noise is input into the post-trained generator to generate generated samples of the health state category to be enhanced. The generated samples are then de-normalized according to the normalization parameters corresponding to the health state category to be enhanced to obtain enhanced samples of the health state category to be enhanced.

2. The method according to claim 1, characterized in that, The preprocessing of the vibration signal samples includes: The vibration signal sample is filtered to obtain a filtered vibration signal sample; The filtered vibration signal sample is segmented using a sliding window to obtain multiple sample segments of equal length. The multiple equal-length sample segments are normalized according to the health status category, so that the sample segments under the same health status category have zero mean and unit variance. Record the mean and standard deviation used for each health status category during the normalization process, and use them as the normalization parameters for each health status category; The normalized sample fragments are labeled with their corresponding health status categories to obtain the initial noise features.

3. The method according to claim 1, characterized in that, The multi-scale feature encoding of the initial noise features includes: The initial noise features are subjected to a first-level multi-scale convolution process to obtain the first-scale features; The first-scale features are downsampled, and the downsampled features are then subjected to a second-level multi-scale convolution to obtain the second-scale features. Continue downsampling and multi-scale convolution processing in a cascaded manner to obtain at least one subsequent scale feature; Each level of multi-scale convolution processing includes: performing multi-path convolution processing on the input features to obtain intermediate features under different receptive fields; concatenating the intermediate features along the channel dimension and performing convolution fusion on the concatenated features to obtain convolution fusion features; and performing residual connection between the convolution fusion features and the input features to obtain the scale features corresponding to that level of multi-scale convolution processing.

4. The method according to claim 3, characterized in that, The multi-path convolution process includes: One-dimensional convolution processing with different kernel sizes was employed. One-dimensional dilated convolution with different dilation rates is used for processing; Continuous one-dimensional convolution with the same kernel size is used for processing; Feature encoding is performed using a combination of convolution and pooling processes. The intermediate features obtained from different paths correspond to different receptive fields.

5. The method according to claim 3, characterized in that, The downsampling process includes max pooling, average pooling, or stride convolution. The scale feature obtained from each level of multi-scale convolution processing is used as one of the multiple scale features; Except for the last stage of multi-scale convolution processing, the scale features obtained from each stage of multi-scale convolution processing are downsampled and then used as the input features for the next stage of multi-scale convolution processing.

6. The method according to claim 1, characterized in that, The step of encoding the initial noise features using a temporal feature encoding based on an attention mechanism includes: The initial noise features are mapped to a feature space of a preset dimension through linear mapping, and positional encoding is added to obtain the embedded features; The embedded features are encoded using a multi-head self-attention mechanism to model the dependencies between different time steps in the initial noise features; The attention output obtained by the encoding process is subjected to a feedforward transformation, and residual connection and layer normalization are performed after the multi-head self-attention mechanism and the feedforward transformation, respectively. The features obtained from each time step are weighted and fused to obtain a global time series representation. The global time series representation is subjected to feature mapping to obtain the time series features.

7. The method according to claim 1, characterized in that, The step of determining adaptive weights corresponding to each scale feature based on the temporal features, and then weighting and aggregating the multiple scale features according to the adaptive weights to obtain the fused features includes: Perform global pooling on each scale feature to obtain the global feature vector corresponding to each scale feature; Each of the global feature vectors is mapped to the same feature dimension as the temporal feature; Calculate the similarity between the temporal features and each of the global feature vectors, and normalize the similarity to obtain the adaptive weights corresponding to each scale feature; The feature transformations of the multiple scale features are performed respectively to obtain the fusion vector corresponding to each scale feature; The weighted summation of each fusion vector is performed according to the adaptive weights to obtain the weighted aggregated features; The weighted aggregated features are fused with the temporal features to obtain the fused features.

8. The method according to claim 7, characterized in that, The step of fusing the weighted aggregated features with the temporal features to obtain the fused features includes: The weighted aggregated features are concatenated, added, or residual-connected with the temporal features to obtain the fused features; The fused features are subjected to layer normalization and feedforward transformation to obtain the fused features. The similarity includes cosine similarity, dot product similarity, or a relevance score calculated by a neural network.

9. The method according to claim 1, characterized in that, The training of the generator and the discriminator based on generative adversarial loss includes: For each health status category, the initial noise features of that health status category are used as real samples, and random noise is input into the generator to obtain the generated samples; Based on the discriminator's output on the initial noise features and the generated samples, a discriminator loss function and a generator loss function are constructed. Based on the discriminator loss function and the generator loss function, the network parameters of the discriminator and the generator are updated alternately; The discriminator loss function includes a gradient penalty term constructed based on random interpolated samples between the initial noise features and the generated samples.

10. A vibration signal data enhancement system, characterized in that, include: The preprocessing module is used to acquire vibration signal samples of rotating machinery under different health state categories, and preprocess the vibration signal samples to obtain random noise for the corresponding health state category; for each health state category, the random noise is input into the generator, and the random noise is mapped to initial noise features through the noise mapping layer in the generator; The feature encoding module is used to perform multi-scale feature encoding and attention-based temporal feature encoding on the initial noise features to obtain multiple scale features and temporal features. The weighted aggregation module is used to determine the adaptive weights corresponding to each scale feature based on the temporal features, and to perform weighted aggregation on the multiple scale features based on the adaptive weights to obtain the fused features; The training module is used to set up corresponding generators and discriminators for different health status categories, and input random noise into the generator corresponding to the health status category to generate generated samples of the health status category. The generated sample and the initial noise features belonging to the health state category are input into the corresponding discriminator, and the generator and the discriminator are trained based on the generative adversarial loss. The enhanced sample generation module is used to input random noise into the trained generator corresponding to the health state category to be enhanced, generate generated samples of the health state category to be enhanced, and perform inverse normalization processing on the generated samples according to the normalization parameters corresponding to the health state category to be enhanced, so as to obtain the enhanced samples of the health state category to be enhanced.