A turbo-machine fault diagnosis method based on shaft center trajectory data enhancement

Abstract

This invention relates to the field of turbine machinery fault diagnosis technology, and provides a turbine machinery fault diagnosis method based on shaft center trajectory data augmentation, comprising: step 1, augmenting shaft center trajectory data existing in the form of grayscale images of shaft center trajectories; step 2, establishing a diagnostic model based on a convolutional neural network; step 3, training the diagnostic model based on the convolutional neural network; and step 4, using the trained diagnostic model based on the convolutional neural network to diagnose faults in the turbine machinery. This invention includes a shaft center trajectory data augmentation method and a diagnostic model based on a convolutional neural network. By rotating, scaling, translating, and mirroring the shaft center trajectory to unify geometric properties, it reduces the data distribution differences between similar fault samples, thereby improving the generalization performance of the model.

Description

Technical Field

[0001] This invention relates to the field of turbine machinery fault diagnosis technology, and in particular to a turbine machinery fault diagnosis method based on shaft center trajectory data augmentation. Background Technology

[0002] Centrifugal compressors, steam turbines, and gas turbines are complex pieces of equipment in industries such as petrochemicals, metallurgy, energy, and defense. Manufacturing defects, improper use, or harsh working environments inevitably lead to turbine failures, which, if not detected early, can cause economic losses or even serious accidents. Therefore, accurate fault identification is crucial to ensuring the long-term reliable, efficient, and stable operation of turbine machinery.

[0003] The shaft center trajectory integrates rotor displacement information in two orthogonal directions, providing a direct reflection of the dynamic behavior of the rotor-bearing system. Different types of rotor faults typically correspond to distinctly different trajectory patterns, giving it strong discriminative ability in diagnosing faults such as imbalance, misalignment, oil film whirl, rubbing, and surge. Although the shaft center trajectory contains rich fault information, the differences between shaft center trajectories of the same type of fault are significant. When there are insufficient fault samples, these differences can significantly reduce diagnostic accuracy. Therefore, introducing shaft center trajectory data augmentation techniques is of great importance in mitigating the impact of insufficient samples and improving fault diagnosis accuracy.

[0004] Currently, axis trajectory data augmentation techniques can be divided into two main categories. The first category involves rotating, scaling, and translating the image to increase the amount of data. The second category uses neural network models such as Variational Autoencoder (VAE), Generative Adversarial Network (GAN), and Diffusion Models (DM) to generate images.

[0005] Most fault diagnosis models based on axis trajectory are built using convolutional neural networks (CNNs), among which commonly used models include ResNet, EfficientNet, and Vision Transformer (ViT).

[0006] Diagnostic models such as ResNet, EfficientNet, and ViT possess a certain degree of rotation, scaling, and translation invariance. Therefore, combining them with the first type of axisymmetric trajectory data augmentation methods can easily lead to model overfitting, thus limiting generalization performance. Similarly, the second type of axisymmetric trajectory data augmentation methods are also constrained by insufficient sample size, making it difficult to fully improve the performance of the diagnostic models.

[0007] Furthermore, the aforementioned model structure is quite complex, making it difficult to achieve ideal results under small sample conditions, and its computational efficiency is also low, further limiting its application in practical engineering scenarios. Summary of the Invention

[0008] This invention primarily addresses the technical problems of existing fault diagnosis models based on shaft center trajectory, which are prone to overfitting, resulting in limited generalization performance, difficulty in fully improving diagnostic model performance, and low accuracy in turbine machinery fault diagnosis under small sample conditions. It proposes a turbine machinery fault diagnosis method based on shaft center trajectory data augmentation, comprising a shaft center trajectory data augmentation method and a diagnostic model based on a convolutional neural network. By rotating, scaling, translating, and mirroring the shaft center trajectory to unify geometric properties, the method reduces the data distribution differences between similar fault samples, thereby improving the model's generalization performance.

[0009] This invention provides a turbine machinery fault diagnosis method based on shaft center trajectory data augmentation, comprising the following steps:

[0010] Step 1 involves enhancing the axis center trajectory data, which exists in the form of a grayscale image of the axis center trajectory. Step 1 includes the following steps 101 to 106:

[0011] Step 101, process the grayscale image of the axis trajectory. Extract the coordinates and grayscale values ​​of all non-white pixels. Represent natural numbers;

[0012] Step 102, for the extracted pixel coordinate matrix Perform principal component analysis to obtain the rotated matrix. ;

[0013] Step 103, for the rotated matrix Perform a translation operation to obtain the translated matrix. ;

[0014] Step 104, for the translated matrix Perform a mirroring operation to obtain the mirrored matrix. ;

[0015] Step 105, for the mirrored matrix Scaling is performed to obtain the scaled matrix. ;

[0016] Step 106, based on the scaled matrix Reconstruct a new grayscale image ;

[0017] Step 2: Establish a diagnostic model based on a convolutional neural network;

[0018] Step 3: Train the diagnostic model based on the convolutional neural network;

[0019] Step 4: Use the trained diagnostic model based on convolutional neural networks to diagnose faults in the turbine machinery.

[0020] Furthermore, step 101 includes:

[0021] pixel coordinate matrix This contains the spatial locations of pixels with grayscale values ​​less than 255, and the corresponding grayscale values ​​are stored in the grayscale value vector. middle;

[0022] ;

[0023] ;

[0024] Where i represents the horizontal axis coordinate value of the grayscale image of the axis-centered trajectory, and j represents the vertical axis coordinate value of the grayscale image of the axis-centered trajectory. This indicates the number of pixels with a grayscale value less than 255.

[0025] Furthermore, step 102 includes:

[0026] Through pixel coordinate matrix Center the coordinates by subtracting the mean of the corresponding column from each value:

[0027] ;

[0028] For covariance matrix Perform eigenvalue decomposition:

[0029] ;

[0030] in, Represents the eigenvalue matrix. Represents the diagonal matrix of eigenvalue decomposition. Represents the first eigenvalue. Indicates the second eigenvalue;

[0031] Projecting the centered coordinates onto the principal axis yields a new pixel coordinate matrix, which is then rotated. :

[0032] .

[0033] Furthermore, step 103 includes:

[0034] For each dimension, calculate the midpoint between the maximum and minimum values ​​and use it as the translation offset. Apply the resulting translation offset to the rotated matrix. Perform a translation operation to shift the axis trajectory to a point centered at the origin, resulting in the translated matrix. :

[0035] First offset The calculation formula is as follows:

[0036] ;

[0037] Second offset The calculation formula is as follows:

[0038] ;

[0039] Translation operation formula:

[0040] .

[0041] Furthermore, step 104 includes:

[0042] According to the following formula Perform a mirroring operation to obtain the mirrored matrix. :

[0043] ;

[0044] In the above formula, sgn represents the step sign. When less than 0, sgn( )=-1, When equal to 0, sgn( )=0, When greater than 0, sgn( )=1; This indicates element-wise multiplication.

[0045] Furthermore, step 105 includes:

[0046] Calculate the scaling factor using the following formula. scaling factor Determined by the larger of the two coordinate dimensions:

[0047] ;

[0048] Then, the normalized coordinates are mapped to the interval. Adding an offset of 64, we obtain the scaled matrix. :

[0049] .

[0050] Furthermore, step 106 includes:

[0051] Initialize the image with all pixel values ​​set to 255, then use the grayscale vector... The grayscale value stored in the image is assigned to the corresponding pixel position:

[0052] ;

[0053] use Mini-pooling operation:

[0054] ;

[0055] in, , .

[0056] Furthermore, the diagnostic model based on convolutional neural networks includes: a feature extraction module and a diagnostic module;

[0057] The feature extraction module includes a convolutional layer, an activation function, a max pooling layer, and a feature unrolling layer;

[0058] The diagnostic module is a multilayer perceptron; the multilayer perceptron consists of two linear layers and two activation functions, wherein the number of neurons in the hidden layers is defined by the following formula:

[0059] .

[0060] Furthermore, step 3 includes: employing the Adam optimizer with a learning rate of The cross-entropy loss function is used; the validation set loss is used to control early stopping of training, and the model parameters corresponding to the lowest validation loss during the training process are selected.

[0061] Furthermore, step 4 includes the following steps 401 to 404:

[0062] Step 401: Obtain vibration signals in two perpendicular directions from any end of the rotor for turbine machinery monitoring;

[0063] Step 402: Determine whether to trigger an alarm based on the peak-to-peak value and peak-to-peak value threshold; if there is no alarm, return to "healthy"; otherwise, perform subsequent diagnostics.

[0064] Step 403: Combine the two vibration waveforms into a grayscale image of the axis trajectory, and perform enhancement processing on the grayscale image of the axis trajectory according to step 1.

[0065] Step 404: Load the model parameters obtained in step 3 into the diagnostic model of the convolutional neural network established in step 2, and perform fault diagnosis on the grayscale image of the axis trajectory.

[0066] This invention provides a turbine machinery fault diagnosis method based on shaft center trajectory data augmentation. By rotating, scaling, translating, and mirroring the shaft center trajectory to unify geometric properties, it reduces the data distribution differences between similar fault samples, thereby improving the model's generalization performance. Furthermore, the constructed convolutional neural network-based diagnostic model employs a lightweight design, enabling it to exhibit good generalization ability under small sample conditions and facilitating engineering deployment. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating the implementation of the turbine machinery fault diagnosis method based on shaft trajectory data enhancement provided by the present invention.

[0068] Figure 2 This is a flowchart of the geometric attribute transformation of the axis trajectory provided by the present invention;

[0069] Figure 3 This is a schematic diagram of the diagnostic model based on convolutional neural networks provided by the present invention. Detailed Implementation

[0070] To make the technical problems solved by this invention, the technical solutions adopted, and the technical effects achieved clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings, not all of them.

[0071] like Figure 1-2 As shown in the figure, an embodiment of the present invention provides a turbine machinery fault diagnosis method based on shaft center trajectory data augmentation, which includes the following process:

[0072] Step 1: Enhance the axis center trajectory data, which exists in the form of a grayscale image of the axis center trajectory.

[0073] Step 1 includes the following steps 101 to 106:

[0074] Step 101, process the grayscale image of the axis trajectory. Extract the coordinates and grayscale values ​​of all non-white pixels. Represents natural numbers.

[0075] The axis trajectory data exists in the form of a grayscale image. Specifically, it is a pixel coordinate matrix. This contains the spatial locations of pixels with grayscale values ​​less than 255, and the corresponding grayscale values ​​are stored in the grayscale value vector. middle.

[0076] ;

[0077] ;

[0078] Where i represents the horizontal axis coordinate value of the grayscale image of the axis-centered trajectory, and j represents the vertical axis coordinate value of the grayscale image of the axis-centered trajectory. This indicates the number of pixels with a grayscale value less than 255.

[0079] Step 102, for the extracted pixel coordinate matrix Perform principal component analysis to obtain the rotated matrix. .

[0080] To eliminate the effect of rotation, the extracted pixel coordinate matrix... Perform principal component analysis (PCA).

[0081] First, through the pixel coordinate matrix Center the coordinates by subtracting the mean of the corresponding column from each value:

[0082] ;

[0083] Subsequently, the covariance matrix Perform eigenvalue decomposition:

[0084] ;

[0085] in, Represents the eigenvalue matrix. Represents the diagonal matrix of eigenvalue decomposition. Represents the first eigenvalue. This represents the second eigenvalue.

[0086] Finally, the centered coordinates are projected onto the principal axis to obtain a new pixel coordinate matrix, which is then rotated. :

[0087] ;

[0088] This transformation aligns the axis trajectory with its principal direction, thereby achieving rotational invariance.

[0089] Step 103, for the rotated matrix Perform a translation operation to obtain the translated matrix. .

[0090] To further standardize spatial distribution, Perform a translation operation. For each dimension, calculate the midpoint between the maximum and minimum values, and use this midpoint as the translation offset. Apply the resulting translation offset to the rotated matrix. Perform a translation operation to shift the axis trajectory to a point centered at the origin, resulting in the translated matrix. :

[0091] First offset The calculation formula is as follows:

[0092] ;

[0093] Second offset The calculation formula is as follows:

[0094] ;

[0095] Translation operation formula:

[0096] ;

[0097] Step 103 allows the axis trajectory to be translated to a position centered on the origin.

[0098] Step 104, for the translated matrix Perform a mirroring operation to obtain the mirrored matrix. .

[0099] To ensure consistent directionality, follow the formula below. Perform a mirroring operation to obtain the mirrored matrix. :

[0100] ;

[0101] In the above formula, sgn represents the step sign. When less than 0, sgn( )=-1, When equal to 0, sgn( )=0, When greater than 0, sgn( )=1; This represents element-wise multiplication. By adjusting the signs of each coordinate dimension, the sum of the coordinates along both coordinate axes is made positive.

[0102] Step 105, for the mirrored matrix Scaling is performed to obtain the scaled matrix. To adapt to the target image size.

[0103] After mirroring, the coordinates are scaled to fit the target image size. The scaling factor is calculated using the following formula. scaling factor Determined by the larger of the two coordinate dimensions:

[0104] ;

[0105] Then, the normalized coordinates are mapped to the interval. Adding an offset of 64, we obtain the scaled matrix. :

[0106] ;

[0107] Step 106, based on the scaled matrix Reconstruct a new grayscale image .

[0108] First, initialize the image with all pixel values ​​set to 255, then generate the grayscale vector. The grayscale value stored in the image is assigned to the corresponding pixel position:

[0109] ;

[0110] Finally, adopt Mini-pooling operation is used to suppress isolated noise and enhance the robustness of the resulting image:

[0111] ;

[0112] in, , .

[0113] The above processing yields a new grayscale image. .

[0114] Step 2: Establish a diagnostic model based on convolutional neural networks.

[0115] The diagnostic model based on convolutional neural networks includes a feature extraction module and a diagnostic module.

[0116] like Figure 3 As shown, the feature extraction module includes a convolutional layer (Conv), an activation function (LeakyReLU, abbreviated as LRelu), a max pooling layer (MaxPool), and a feature unfolding layer (Reshape). The convolutional kernel size is set to 3×3, with a stride of 2 and padding of 1. These parameters of the convolutional kernel are widely used in image classification models.

[0117] The diagnostic module is a multilayer perceptron (MLP). The MLP consists of two linear layers and two activation functions (LeakyReLU, abbreviated as LReLU). The number of neurons in the hidden layers (non-input and non-output layers within the linear layers) is defined by the following formula:

[0118] .

[0119] Step 3: Train the diagnostic model based on the convolutional neural network.

[0120] The Adam optimizer is used, with a learning rate of The cross-entropy loss function is used. To prevent overfitting, validation set loss is used to control early training termination; training stops when the validation loss exceeds the historical minimum for 10 consecutive times. The model parameters corresponding to the lowest validation loss during training are then used. A batch size of 16 is chosen for model training.

[0121] Step 4: Using the trained diagnostic model based on a convolutional neural network, perform fault diagnosis on the turbine machinery. Step 4 includes the following steps 401 to 404:

[0122] Step 401: Obtain vibration signals from two perpendicular directions at any end of the rotor used for turbine machinery monitoring, i.e., the vibration waveforms after edge processing. The vibration waveforms typically correspond to 32 rotor rotation cycles, with a discrete value of 1024 units in length.

[0123] Step 402: Determine whether to trigger an alarm based on the peak-to-peak value (the maximum value minus the minimum value of the vibration waveform) and the peak-to-peak value threshold. If there is no alarm, return to "healthy"; otherwise, proceed with subsequent diagnostics.

[0124] Step 403: Combine the two vibration waveforms into a grayscale image of the axis trajectory, and perform enhancement processing on the grayscale image of the axis trajectory according to step 1.

[0125] Step 404: Load the model parameters obtained in step 3 into the diagnostic model of the convolutional neural network established in step 2, and perform fault diagnosis on the grayscale image of the axis trajectory.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions for some or all of the technical features, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for fault diagnosis of turbine machinery based on shaft center trajectory data augmentation, characterized in that, The process includes the following: Step 1 involves enhancing the axis center trajectory data, which exists in the form of a grayscale image of the axis center trajectory. Step 1 includes the following steps 101 to 106: Step 101, process the grayscale image of the axis trajectory. Extract the coordinates and grayscale values ​​of all non-white pixels. Represent natural numbers; Step 102, for the extracted pixel coordinate matrix Perform principal component analysis to obtain the rotated matrix. ; Step 103, for the rotated matrix Perform a translation operation to obtain the translated matrix. ; Step 104, for the translated matrix Perform a mirroring operation to obtain the mirrored matrix. ; Step 105, for the mirrored matrix Scaling is performed to obtain the scaled matrix. ; Step 106, based on the scaled matrix Reconstruct a new grayscale image ; Step 2: Establish a diagnostic model based on a convolutional neural network; Step 3: Train the diagnostic model based on the convolutional neural network; Step 4: Use the trained diagnostic model based on convolutional neural networks to diagnose faults in the turbine machinery.

2. The turbine machinery fault diagnosis method based on shaft center trajectory data augmentation according to claim 1, characterized in that, Step 101 includes: pixel coordinate matrix This contains the spatial locations of pixels with grayscale values ​​less than 255, and the corresponding grayscale values ​​are stored in the grayscale value vector. middle; ； ； Where i represents the horizontal axis coordinate value of the grayscale image of the axis-centered trajectory, and j represents the vertical axis coordinate value of the grayscale image of the axis-centered trajectory. This indicates the number of pixels with a grayscale value less than 255.

3. The turbine machinery fault diagnosis method based on shaft center trajectory data augmentation according to claim 2, characterized in that, Step 102 includes: Through pixel coordinate matrix Center the coordinates by subtracting the mean of the corresponding column from each value: ； For covariance matrix Perform eigenvalue decomposition: ； in, Represents the eigenvalue matrix. Represents the diagonal matrix of eigenvalue decomposition. Represents the first eigenvalue. Indicates the second eigenvalue; Projecting the centered coordinates onto the principal axis yields a new pixel coordinate matrix, which is then rotated. : 。 4. The turbine machinery fault diagnosis method based on shaft center trajectory data augmentation according to claim 3, characterized in that, Step 103 includes: For each dimension, calculate the midpoint between the maximum and minimum values ​​and use it as the translation offset. Apply the resulting translation offset to the rotated matrix. Perform a translation operation to shift the axis trajectory to a point centered at the origin, resulting in the translated matrix. : First offset The calculation formula is as follows: ； Second offset The calculation formula is as follows: ； Translation operation formula: 。 5. The turbine machinery fault diagnosis method based on shaft center trajectory data augmentation according to claim 4, characterized in that, Step 104 includes: According to the following formula Perform a mirroring operation to obtain the mirrored matrix. : ； In the above formula, sgn represents the step sign. When less than 0, sgn( )=-1, When equal to 0, sgn( )=0, When greater than 0, sgn( )=1; This indicates element-wise multiplication.

6. The turbine machinery fault diagnosis method based on shaft center trajectory data augmentation according to claim 5, characterized in that, Step 105 includes: Calculate the scaling factor using the following formula. scaling factor Determined by the larger of the two coordinate dimensions: ； Then, the normalized coordinates are mapped to the interval. Adding an offset of 64, we obtain the scaled matrix. : 。 7. The turbine machinery fault diagnosis method based on shaft center trajectory data augmentation according to claim 6, characterized in that, Step 106 includes: Initialize the image with all pixel values ​​set to 255, then use the grayscale vector... The grayscale value stored in the image is assigned to the corresponding pixel position: ； use Mini-pooling operation: ； in, , .

8. The turbine machinery fault diagnosis method based on shaft center trajectory data augmentation according to claim 1, characterized in that, The diagnostic model based on convolutional neural networks includes: a feature extraction module and a diagnostic module; The feature extraction module includes a convolutional layer, an activation function, a max pooling layer, and a feature unrolling layer; The diagnostic module is a multilayer perceptron; the multilayer perceptron consists of two linear layers and two activation functions, wherein the number of neurons in the hidden layers is defined by the following formula: 。 9. The turbine machinery fault diagnosis method based on shaft center trajectory data augmentation according to claim 1, characterized in that, Step 3 includes: employing the Adam optimizer with a learning rate of The cross-entropy loss function is used; the validation set loss is used to control early stopping of training, and the model parameters corresponding to the lowest validation loss during the training process are selected.

10. The turbine machinery fault diagnosis method based on shaft center trajectory data augmentation according to claim 1, characterized in that, Step 4 includes the following steps 401 to 404: Step 401: Obtain vibration signals in two perpendicular directions from any end of the rotor for turbine machinery monitoring; Step 402: Determine whether to trigger an alarm based on the peak-to-peak value and peak-to-peak value threshold; if there is no alarm, return to "healthy"; otherwise, perform subsequent diagnostics. Step 403: Combine the two vibration waveforms into a grayscale image of the axis trajectory, and perform enhancement processing on the grayscale image of the axis trajectory according to step 1. Step 404: Load the model parameters obtained in step 3 into the diagnostic model of the convolutional neural network established in step 2, and perform fault diagnosis on the grayscale image of the axis trajectory.

Images