Nuclear power equipment operation and maintenance dataset enhancement method, system, medium and device

Abstract

A nuclear power equipment operation and maintenance data set enhancement method, system, medium and device, in the method, signals of different health states of nuclear power equipment are collected; a residual shrinkage construction unit RSBU-CW is introduced in a network generator to prevent the network from paying attention to insignificant features; a residual channel attention mask self-attention module RCA-SA is constructed to extract global and local features of the signals; a multi-scale feature cascade module MSFC is constructed to reorganize the features and enhance the representation ability and generalization ability of the model; a generator and a discriminator of a generative adversarial network are constructed, a Wasserstein distance is introduced, pseudo samples similar to the original samples are output, and nuclear power equipment operation and maintenance data enhancement is realized.

Description

Technical Field

[0001] This invention relates to the field of nuclear power equipment operation and maintenance data processing technology, and in particular to a method, system, medium and equipment for enhancing nuclear power equipment operation and maintenance datasets. Background Technology

[0002] Generally, the more abundant and diverse the operation and maintenance data of nuclear power equipment, the higher the accuracy of the intelligent fault diagnosis model. However, in reality, mechanical equipment failures are relatively rare. For highly reliable equipment like nuclear power plants, components are usually replaced before failure occurs, making the probability of failure even lower. This results in a situation where, even though the condition monitoring system continuously collects a large amount of data, the number of fault samples is relatively small. This imbalance between the number of normal and fault samples leads to an imbalance in the training of the intelligent diagnostic model, meaning that the overall classification accuracy is satisfied at the expense of the classification accuracy of a few samples.

[0003] Generative adversarial networks (GANs) offer a feasible solution to the sample imbalance problem in fault diagnosis from a data generation perspective. GANs can extract features from the original data and generate new samples with the same distribution as the original data. However, GANs face problems such as vanishing gradients and mode collapse during model training, making the model difficult to train, and the generated samples lack diversity. Furthermore, existing methods rarely comprehensively consider the global interactions and local dependencies of sensor signals during data generation, resulting in poor sample quality. Therefore, how to improve sample quality while ensuring sample diversity is a pressing issue that needs to be addressed.

[0004] The information disclosed in the background section is only for enhancing the understanding of the background of this invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] This invention provides a method, system, medium, and equipment for enhancing nuclear power equipment operation and maintenance datasets, generating samples under fault conditions and achieving a balance between positive and negative samples in the dataset, thereby improving the accuracy of fault diagnosis.

[0006] A method for enhancing nuclear power equipment operation and maintenance datasets includes:

[0007] The first step is to collect signals from different health states of nuclear power equipment;

[0008] The second step is to introduce the residual shrinkage building block RSBU-CW into the network generator to prevent the network from focusing on irrelevant features.

[0009] The third step is to construct the residual channel attention mask self-attention module RCA-SA to extract global and local features of the signal;

[0010] The fourth step is to construct a multi-scale feature concatenation module (MSFC) to reorganize features and enhance the model's representation and generalization capabilities.

[0011] The fifth step involves constructing a generator and discriminator for a generative adversarial network, introducing the Wasserstein distance, and outputting pseudo-samples similar to the original samples to enhance the operation and maintenance data of nuclear power equipment.

[0012] In the aforementioned method for enhancing the operation and maintenance data set of nuclear power equipment, in the first step, sensors are used to collect signals from the nuclear power equipment under normal and fault conditions to obtain multi-dimensional operating signals of the nuclear power equipment. The sensors include vibration sensors, displacement sensors, temperature sensors, pressure sensors, and flow sensors.

[0013] In the aforementioned method for enhancing nuclear power equipment operation and maintenance datasets, the second step includes:

[0014] Perform feature forward processing, input feature matrix B represents the batch size, C represents the number of channels, and N represents the feature dimension. The feature is passed through two consecutive 1×3 convolutional layers, each followed by a batch normalization layer (BN) and a ReLU activation function, outputting an intermediate feature matrix X2.

[0015]

[0016] Where X1 is the output after one convolutional layer, X2 is the output after two consecutive convolutional layers, ReLU(·) is the activation function operation, BN(·) is the batch normalization operation, and f 1×3 (·) represents a 1×3 convolutional layer operation.

[0017] Channel attention is generated, and global average pooling is performed on the absolute value X2 to obtain the feature matrix. The feature matrix A is then input into the attention module to mine the correlation information of the channels, resulting in the feature matrix. The attention module consists of two fully connected layers, with a Batch Normalization (BN) layer and a ReLU activation function in between. The feature matrix S is normalized by a Sigmoid activation layer, and the resulting value is denoted as α. i :

[0018]

[0019] Where, α i The activation value calculated for each channel; s i Let be the eigenvalue of the i-th channel of the feature matrix S.

[0020] Adaptive threshold calculation is performed by multiplying the channel attention weights α with the feature matrix A channel by channel to generate a dynamic threshold matrix T.

[0021] Τ=α⊙A

[0022] Here, ⊙ represents the element-wise dot product operation.

[0023] Soft thresholding is performed on feature matrix X2 to obtain feature matrix X3. The soft thresholding process is as follows:

[0024]

[0025] Where y is the output of the soft thresholding process, x is the input before the soft thresholding process, and τ is the threshold, which is determined by the statistics of the corresponding channel in the dynamic threshold matrix T.

[0026] Perform a residual connection, adding the input feature matrix X to the soft-thresholding output matrix X3 element-wise to obtain the final output matrix Y:

[0027] Y = X + X3

[0028] The residual shrinkage building unit RSBU-CW achieves differentiated weighting of feature channels through a channel attention mechanism, eliminates redundant information by combining soft thresholding, and ensures gradient propagation stability by utilizing the residual structure.

[0029] In the aforementioned method for enhancing nuclear power equipment operation and maintenance datasets, the third step includes:

[0030] Perform multi-branch feature mapping, input feature matrix B represents the batch size, C represents the number of channels, and N represents the feature dimension. The query matrix Q, key matrix K, and value matrix V are generated using three independent 1×1 convolutional kernels.

[0031]

[0032] Among them, f 1×1 (·) represents a 1×1 convolutional layer operation.

[0033] Multi-head segmentation and encoding are performed, dividing the query matrix, key matrix, and value matrix into p different attention heads. Simultaneously, the mask input M is also split into p heads, each with dimensions B×C / p×N. The segmented representation is as follows:

[0034]

[0035] Contact(·) is the concatenation operation, where Q1, Q2, ..., Q p For different query matrix headers, K1, K2, ..., K pFor different key matrix headers, V1, V2, ..., V p For different value matrix headers, M1, M2, ..., M p For different mask input headers,

[0036] Perform multi-head self-attention calculation, and use the query matrix Q j Bond matrix K j Each row of the multiplication result is normalized using the Softmax function. Before performing the Softmax operation, relative position encoding is added to allow the self-attention mechanism to more effectively capture the relative position information between elements in the sequence. Simultaneously, each value matrix V... j Header and mask matrix M j Element-wise multiplication is performed to incorporate the weight differences of different channels into the value matrix:

[0037]

[0038] Where Encode(·) is the relative position encoding operation, Softmax(·) is the normalization operation, p is the number of attention heads, and Q is the relative position encoding operation. j T For query matrix Q j The transpose of the matrix,

[0039] Feature fusion and output are performed by concatenating the outputs of p attention heads along the channel dimension and then integrating them through 1×1 convolution to obtain the final output X. o :

[0040]

[0041] Among them, f 1×1 (·) represents a 1×1 convolutional layer operation.

[0042] By implementing a multi-head self-attention mechanism to achieve differentiated weight allocation in the spatiotemporal dimensions, the cross-channel and cross-temporal correlation of key features is strengthened while suppressing noisy features.

[0043] In the aforementioned method for enhancing nuclear power equipment operation and maintenance datasets, the fourth step includes:

[0044] Multi-scale feature segmentation and hierarchical extraction, input feature matrix B represents the batch size, C represents the number of channels, and N represents the feature dimension. An initial mapping is performed using a 1×1 convolution kernel to generate an intermediate feature matrix P1. Then, P1 is divided into n equal parts according to the channel dimension. Next, a 1×1 depthwise separable convolution operation is used for hierarchical feature extraction. The above process can be represented as:

[0045]

[0046] Split(·) is a channel-based partitioning operation. For 1×1 depth-separable convolution operations, For the i-th output sub-feature, P1 i For the i-th intermediate sub-feature,

[0047] Perform cross-scale feature fusion and output each sub-feature. The input features are concatenated along the channel dimension and integrated across channels using a 1×1 convolutional kernel. Finally, the original input features and the concatenated features are added element-wise through residual connections to obtain the final output feature P. o :

[0048]

[0049] Contact(·) is the concatenation operation, f 1×1 (·) represents a 1×1 convolutional layer operation.

[0050] By using multi-scale channel partitioning and hierarchical residual fusion, combined with depthwise separable convolution to reduce the number of parameters, and by cross-scale cascading recombination to improve the model's ability to generalize the time-frequency features of the signal.

[0051] In the aforementioned method for enhancing nuclear power equipment operation and maintenance datasets, the fifth step includes:

[0052] Construct a generator, assuming a noise vector B represents the batch size, d z Representing the vector dimension, the noise vector first passes through a fully connected layer to obtain the output vector F0:

[0053]

[0054] Where LeakyReLU(·) represents the activation function operation, BN(·) represents the batch normalization operation, W0 represents the fully connected weight matrix, and b0 represents the bias vector.

[0055] After passing through the first deconvolution module and the RSBU-CW module, the output vector F1 is obtained. The RSBU-CW module first removes redundant information from the noise vector to address the task of generating spurious signals.

[0056] F1=RSBU_CW(LeakyReLU(BN(ConvT(F0))))

[0057] Where RSBU_CW(·) represents the RSBU-CW module operation, and ConvT(·) represents the deconvolution operation.

[0058] The RCA-SA module is applied after the second deconvolution module to obtain the output vector F2, which captures the weight relationship between different positions of the same signal sample from different channel dimensions:

[0059] F2=RCA_SA(LeakyReLU(BN(ConvT(F1))))

[0060] Where RCA_SA(·) represents the RCA_SA module operation,

[0061] The MSFC module is applied after the third deconvolution module to obtain the output vector F3. The information obtained after denoising by the RSBU-CW module and extraction by the RCA-SA module is then subjected to multi-scale feature concatenation for information integration.

[0062] F3=MSFC(LeakyReLU(BN(ConvT(F2))))

[0063] In this context, MSFC(·) represents the MSFC module operation.

[0064] After the final deconvolution module and reshaping, the spurious vector X is finally generated and input into the discriminator along with the original sample:

[0065] X = Reshape(ConvT(F3))

[0066] Where Reshape(·) represents the reshaping operation,

[0067] Construct a discriminator, which consists of a series of convolutional modules. Each convolutional module includes a convolutional layer, followed by batch normalization (BN) and the LeakyReLU activation function. Each convolutional module is then followed by a max-pooling layer.

[0068] X i =MaxPool(LeakyReLU(BN(Conv)) i (X i-1 ))))

[0069] Where MaxPool(·) represents the global average pooling operation, X i ,X i-1 Conv represents the intermediate feature vector of the discriminator. i This represents the i-th convolution operation.

[0070] Finally, a global average pooling layer is applied to reduce the feature dimension to 1, in order to identify the authenticity of the input samples.

[0071]

[0072] Where GAP(·) represents the global average pooling operation. For the discriminator output,

[0073] Introducing the Wasserstein distance, we construct the loss function for the model. The loss functions for the generator and discriminator are as follows:

[0074]

[0075] Where E is the mathematical expectation operator, x is the real sample data, z is the random noise vector in the latent space, G(z) is the synthetic sample generated by the generator based on the noise z, D(x) is the score of the discriminator for the real sample x, min(·) is to take the minimum value, and max(·) is to take the maximum value.

[0076] In the aforementioned method for enhancing the operation and maintenance dataset of nuclear power equipment, the nuclear power equipment includes a nuclear power steam-driven feedwater pump.

[0077] A system for implementing the method includes:

[0078] The acquisition module collects signals from different health states of nuclear power equipment.

[0079] The first building block is used to introduce the residual shrinkage building block RSBU-CW into the network generator to prevent the network from focusing on irrelevant features;

[0080] The second building unit is used to construct the residual channel attention mask self-attention module RCA-SA, which is used to extract global and local features of the signal.

[0081] The third building block is used to construct the Multi-Scale Feature Concatenation Module (MSFC) to reorganize features and enhance the model's representation and generalization capabilities.

[0082] The enhancement unit is used to construct the generator and discriminator of the generative adversarial network. It introduces the Wasserstein distance and outputs pseudo-samples similar to the original samples to enhance the operation and maintenance data of nuclear power equipment.

[0083] A computer storage medium including computer instructions that, when run on a computer, cause the computer to perform the method.

[0084] An electronic device, the electronic device comprising:

[0085] Memory, processor, and computer programs stored in memory and executable on the processor, wherein,

[0086] The processor implements the method when executing the program.

[0087] Compared with existing technologies, this invention has the following advantages: Firstly, by introducing Wasserstein distance into the generator and discriminator losses of the GAN network, this invention minimizes the gradient vanishing and mode collapse problems in GAN networks. Secondly, utilizing a self-attention mechanism and a novel channel attention mechanism, a multi-head self-attention module with residual channel attention masks is designed and applied to the generator to capture the fault characteristics of the next-dimensional time-series signals of different fault types in nuclear power equipment at both global and local levels. This invention can also serve as a pre-processing step for fault diagnosis of other high-reliability equipment, expanding its practical value in engineering applications. Attached Figure Description

[0088] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0089] In the attached diagram:

[0090] Figure 1 This is a schematic diagram illustrating the steps of one embodiment of the present invention;

[0091] Figures 2(a) and 2(b) are structural diagrams of a generator and a discriminator according to one embodiment of the present invention, wherein Figures 2(a) and 2(b) are structural diagrams of the generator and the discriminator, respectively.

[0092] Figures 3(a), 3(b), 3(c), and 3(d) are time-domain comparison diagrams of the original signal and the generated signal according to one embodiment of the present invention. Figures 3(a), 3(b), 3(c), and 3(d) are the time-domain diagrams of the original signal, the time-domain diagrams obtained by applying the method of the present invention, the time-domain diagrams obtained by applying the WGAN method, and the time-domain diagrams obtained by applying the FAWGAN method, respectively.

[0093] Figures 4(a), 4(b), 4(c), and 4(d) are spectrum comparison diagrams of the original signal and the generated signal according to one embodiment of the present invention. Figures 4(a), 4(b), 4(c), and 4(d) are the spectrum diagrams of the original signal, the spectrum diagram obtained by applying the method of the present invention, the spectrum diagram obtained by applying the WGAN method, and the spectrum diagram obtained by applying the FAWGAN method, respectively.

[0094] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0095] Specific embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While specific embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0096] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0097] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0098] like Figures 1 to 4(d) As shown, the method for enhancing the nuclear power equipment operation and maintenance dataset includes the following steps:

[0099] The first step is to collect signals from different health states of nuclear power equipment;

[0100] The second step is to introduce the residual shrinkage building block RSBU-CW into the network generator to prevent the network from focusing on irrelevant features.

[0101] The third step is to construct the residual channel attention mask self-attention module RCA-SA to extract global and local features of the signal;

[0102] The fourth step is to construct a multi-scale feature concatenation module (MSFC) to reorganize features and enhance the model's representation and generalization capabilities.

[0103] The fifth step involves constructing a generator and discriminator for a generative adversarial network, introducing the Wasserstein distance, and outputting pseudo-samples similar to the original samples to enhance the operation and maintenance data of nuclear power equipment.

[0104] In a preferred embodiment of the nuclear power equipment operation and maintenance data augmentation method, the first step involves using sensors to collect signals from the nuclear power equipment under normal and fault conditions to obtain multi-dimensional operating signals. These sensors include vibration sensors, displacement sensors, temperature sensors, pressure sensors, and flow sensors. Utilizing sensors (such as vibration, displacement, and temperature sensors) to acquire multi-dimensional operating signals from the nuclear power equipment under normal and fault conditions provides a realistic sample basis for subsequent data augmentation. This helps ensure that the generated data has practical significance and reflects the actual equipment status. The key is to ensure the authenticity and representativeness of the data.

[0105] In a preferred embodiment of the nuclear power equipment operation and maintenance dataset enhancement method, the second step includes:

[0106] Perform feature forward processing, input feature matrix B represents the batch size, C represents the number of channels, and N represents the feature dimension. The feature is passed through two consecutive 1×3 convolutional layers, each followed by a batch normalization layer (BN) and a ReLU activation function, outputting an intermediate feature matrix X2.

[0107]

[0108] Where X1 is the output after one convolutional layer, X2 is the output after two consecutive convolutional layers, ReLU(·) is the activation function operation, BN(·) is the batch normalization operation, and f 1×3 (·) represents a 1×3 convolutional layer operation.

[0109] Channel attention is generated, and global average pooling is performed on the absolute value X2 to obtain the feature matrix. The feature matrix A is then input into the attention module to mine the correlation information of the channels, resulting in the feature matrix. The attention module consists of two fully connected layers, with a Batch Normalization (BN) layer and a ReLU activation function in between. The feature matrix S is normalized by a Sigmoid activation layer, and the resulting value is denoted as α. i :

[0110]

[0111] Where, α i The activation value calculated for each channel; s i Let be the eigenvalue of the i-th channel of the feature matrix S.

[0112] Adaptive threshold calculation is performed by multiplying the channel attention weights α with the feature matrix A channel by channel to generate a dynamic threshold matrix T.

[0113] Τ=α⊙A

[0114] Here, ⊙ represents the element-wise dot product operation.

[0115] Soft thresholding is performed on feature matrix X2 to obtain feature matrix X3. The soft thresholding process is as follows:

[0116]

[0117] Where y is the output of the soft thresholding process, x is the input before the soft thresholding process, and τ is the threshold, which is determined by the statistics of the corresponding channel in the dynamic threshold matrix T.

[0118] Perform a residual connection, adding the input feature matrix X to the soft-thresholding output matrix X3 element-wise to obtain the final output matrix Y:

[0119] Y = X + X3

[0120] The RSBU-CW module implements differentiated weighting of feature channels through a channel attention mechanism, combines soft thresholding to eliminate redundant information, and utilizes residual structures to ensure gradient propagation stability. The residual shrinkage construction unit RSBU-CW is introduced: through feature forward processing, channel attention generation, adaptive threshold calculation, soft thresholding, and residual connections, it prevents the network from focusing on irrelevant features while ensuring gradient propagation stability. The key focus is on eliminating redundant information, strengthening key features, and ensuring gradient propagation stability.

[0121] In a preferred embodiment of the nuclear power equipment operation and maintenance data enhancement method, the third step includes:

[0122] Perform multi-branch feature mapping, input feature matrix B represents the batch size, C represents the number of channels, and N represents the feature dimension. The query matrix Q, key matrix K, and value matrix V are generated using three independent 1×1 convolutional kernels.

[0123]

[0124] Among them, f 1×1 (·) represents a 1×1 convolutional layer operation.

[0125] Multi-head segmentation and encoding are performed, dividing the query matrix, key matrix, and value matrix into p different attention heads. Simultaneously, the mask input M is also split into p heads, each with dimensions B×C / p×N. The segmented representation is as follows:

[0126]

[0127] Contact(·) is the concatenation operation, where Q1, Q2, ..., Q p For different query matrix headers, K1, K2, ..., Kp For different key matrix headers, V1, V2, ..., V p For different value matrix headers, M1, M2, ..., M p For different mask input headers,

[0128] Perform multi-head self-attention calculation, and use the query matrix Q j Bond matrix K j Each row of the multiplication result is normalized using the Softmax function. Before performing the Softmax operation, relative position encoding is added to allow the self-attention mechanism to more effectively capture the relative position information between elements in the sequence. Simultaneously, each value matrix V... j Header and mask matrix M j Element-wise multiplication is performed to incorporate the weight differences of different channels into the value matrix:

[0129]

[0130] Where Encode(·) is the relative position encoding operation, Softmax(·) is the normalization operation, p is the number of attention heads, and Q is the relative position encoding operation. j T For query matrix Q j The transpose of the matrix,

[0131] Feature fusion and output are performed by concatenating the outputs of p attention heads along the channel dimension and then integrating them through 1×1 convolution to obtain the final output X. o :

[0132]

[0133] Among them, f 1×1 (·) represents a 1×1 convolutional layer operation.

[0134] By implementing a multi-head self-attention mechanism to achieve differentiated weight allocation in the spatiotemporal dimensions, the cross-channel and cross-temporal correlation of key features is strengthened while suppressing noisy features.

[0135] In a preferred embodiment of the nuclear power equipment operation and maintenance dataset enhancement method, the fourth step includes:

[0136] Multi-scale feature segmentation and hierarchical extraction, input feature matrix B represents the batch size, C represents the number of channels, and N represents the feature dimension. An initial mapping is performed using a 1×1 convolution kernel to generate an intermediate feature matrix P1. Then, P1 is divided into n equal parts according to the channel dimension. Next, a 1×1 depthwise separable convolution operation is used for hierarchical feature extraction. The above process can be represented as:

[0137]

[0138] Split(·) is a channel-based partitioning operation. For 1×1 depth-separable convolution operations, For the i-th output sub-feature, P1 i For the i-th intermediate sub-feature,

[0139] Perform cross-scale feature fusion and output each sub-feature. The input features are concatenated along the channel dimension and integrated across channels using a 1×1 convolutional kernel. Finally, the original input features and the concatenated features are added element-wise through residual connections to obtain the final output feature P. o :

[0140]

[0141] Contact(·) is the concatenation operation, f 1×1 (·) represents a 1×1 convolutional layer operation.

[0142] By using multi-scale channel partitioning and hierarchical residual fusion, combined with depthwise separable convolution to reduce the number of parameters, and by cross-scale cascading recombination to improve the model's ability to generalize the time-frequency features of the signal.

[0143] In a preferred embodiment of the method for enhancing the operation and maintenance dataset of nuclear power equipment, the fifth step includes:

[0144] Construct a generator, assuming a noise vector B represents the batch size, d z Representing the vector dimension, the noise vector first passes through a fully connected layer to obtain the output vector F0:

[0145]

[0146] Where LeakyReLU(·) represents the activation function operation, BN(·) represents the batch normalization operation, W0 represents the fully connected weight matrix, and b0 represents the bias vector.

[0147] After passing through the first deconvolution module and the RSBU-CW module, the output vector F1 is obtained. The RSBU-CW module first removes redundant information from the noise vector to address the task of generating spurious signals.

[0148] F1=RSBU_CW(LeakyReLU(BN(ConvT(F0))))

[0149] Where RSBU_CW(·) represents the RSBU-CW module operation, and ConvT(·) represents the deconvolution operation.

[0150] The RCA-SA module is applied after the second deconvolution module to obtain the output vector F2, which captures the weight relationship between different positions of the same signal sample from different channel dimensions:

[0151] F2=RCA_SA(LeakyReLU(BN(ConvT(F1))))

[0152] Where RCA_SA(·) represents the RCA_SA module operation,

[0153] The MSFC module is applied after the third deconvolution module to obtain the output vector F3. The information obtained after denoising by the RSBU-CW module and extraction by the RCA-SA module is then subjected to multi-scale feature concatenation for information integration.

[0154] F3=MSFC(LeakyReLU(BN(ConvT(F2))))

[0155] In this context, MSFC(·) represents the MSFC module operation.

[0156] After the final deconvolution module and reshaping, the spurious vector X is finally generated and input into the discriminator along with the original sample:

[0157] X = Reshape(ConvT(F3))

[0158] Where Reshape(·) represents the reshaping operation,

[0159] Construct a discriminator, which consists of a series of convolutional modules. Each convolutional module includes a convolutional layer, followed by batch normalization (BN) and the LeakyReLU activation function. Each convolutional module is then followed by a max-pooling layer.

[0160] X i =MaxPool(LeakyReLU(BN(Conv)) i (X i-1 ))))

[0161] Where MaxPool(·) represents the global average pooling operation, X i ,X i-1 Conv represents the intermediate feature vector of the discriminator. i This represents the i-th convolution operation.

[0162] Finally, a global average pooling layer is applied to reduce the feature dimension to 1, in order to identify the authenticity of the input samples.

[0163]

[0164] Where GAP(·) represents the global average pooling operation. For the discriminator output,

[0165] Introducing the Wasserstein distance, we construct the loss function for the model. The loss functions for the generator and discriminator are as follows:

[0166]

[0167] Where E is the mathematical expectation operator, x is the real sample data, z is the random noise vector in the latent space, G(z) is the synthetic sample generated by the generator based on the noise z, D(x) is the score of the discriminator for the real sample x, min(·) is to take the minimum value, and max(·) is to take the maximum value.

[0168] In a preferred embodiment of the method for enhancing the operation and maintenance dataset of nuclear power equipment, the nuclear power equipment includes a nuclear power steam-driven feedwater pump.

[0169] A system for implementing the method includes:

[0170] The acquisition module collects signals from different health states of nuclear power equipment.

[0171] The first building block is used to introduce the residual shrinkage building block RSBU-CW into the network generator to prevent the network from focusing on irrelevant features;

[0172] The second building unit is used to build the residual channel attention mask self-attention module RCA-SA, which is used to extract global and local features of the signal.

[0173] The third building block is used to construct the Multi-Scale Feature Concatenation Module (MSFC) to reorganize features and enhance the model's representation and generalization capabilities.

[0174] The enhancement unit is used to construct the generator and discriminator of the generative adversarial network. It introduces the Wasserstein distance and outputs pseudo-samples similar to the original samples to enhance the operation and maintenance data of nuclear power equipment.

[0175] In one embodiment, a method for enhancing the operation and maintenance dataset of nuclear power equipment, such as Figure 1 As shown, it includes the following steps:

[0176] The first step is to collect signals from different health states of nuclear power equipment;

[0177] The second step is to introduce the residual shrinkage building block RSBU-CW into the network generator to prevent the network from focusing on irrelevant features.

[0178] The third step is to design a residual channel attention mask self-attention module RCA-SA to extract global and local features of the signal;

[0179] The fourth step is to construct a multi-scale feature concatenation module (MSFC) to reorganize features and enhance the model's representation and generalization capabilities.

[0180] The fifth step involves constructing the basic structure of the generator and discriminator of the generative adversarial network, introducing the Wasserstein distance to improve the stability of the model and the quality of data generation.

[0181] In a preferred embodiment of the method, in the first step, signals from the nuclear power equipment under normal and fault conditions are collected using multiple types of sensors, including but not limited to vibration sensors, displacement sensors, temperature sensors, pressure sensors, and flow sensors, to obtain multi-dimensional operating parameters of the nuclear power equipment.

[0182] In a preferred embodiment of the method, in the second step, the RSBU-CW in the network generator prevents the network from focusing on irrelevant features. This includes a residual shrinkage building unit (RSBU-CW) structure consisting of five parts: feature forward processing, channel attention generation, adaptive threshold calculation, soft thresholding, and residual connection. The input is the feature matrix obtained after processing the noise vector through a fully connected layer and the first deconvolution module.

[0183] (1) Perform feature forward processing and input the feature matrix. B represents the batch size, C represents the number of channels, and N represents the feature dimension. The feature is passed through two consecutive 1×3 convolutional layers, each followed by a batch normalization layer (BN) and a ReLU activation function, outputting an intermediate feature matrix X2.

[0184]

[0185] Where X1 is the output after one convolutional layer, X2 is the output after two consecutive convolutional layers, ReLU(·) is the activation function operation, BN(·) is the batch normalization operation, and f 1×3 (·) represents a 1×3 convolutional layer operation.

[0186] (2) Perform channel attention generation, and perform global average pooling (expansion) on the absolute value X2 to obtain the feature matrix. The feature matrix A is then input into the attention module to mine the correlation information of the channels, resulting in the feature matrix. The attention module consists of two fully connected layers, with a Batch Normalization (BN) layer and a ReLU activation function in between. The feature matrix S is normalized by a Sigmoid activation layer, and the resulting value is denoted as α. i :

[0187]

[0188] Where, α i The activation value calculated for each channel; s i Let be the eigenvalue of the i-th channel of the feature matrix S.

[0189] (3) Perform adaptive threshold calculation by multiplying the channel attention weight α with the feature matrix A channel by channel to generate the dynamic threshold matrix T:

[0190] Τ=α⊙A

[0191] Here, ⊙ represents the element-wise dot product operation.

[0192] (4) Soft thresholding: Apply soft thresholding to feature matrix X2 to obtain feature matrix X3. The soft thresholding process is as follows:

[0193]

[0194] Where y is the output of the soft thresholding process, x is the input before the soft thresholding process, and τ is the threshold, which is determined by the statistics of the corresponding channel in the dynamic threshold matrix T.

[0195] (5) Residual connection: The input feature matrix X is added element-wise to the soft-thresholding output matrix X3 to obtain the final output matrix Y:

[0196] Y = X + X3

[0197] The RSBU-CW module implements differentiated weighting of feature channels through a channel attention mechanism, eliminates redundant information by combining soft thresholding, and ensures gradient propagation stability by utilizing residual structures.

[0198] In a preferred embodiment of the method, the third step involves designing a residual channel attention mask self-attention module (RCA-SA) to extract global and local features of the signal, including:

[0199] (1) Multi-branch feature mapping, input feature matrix B represents the batch size, C represents the number of channels, and N represents the feature dimension. The query matrix Q, key matrix K, and value matrix V are generated using three independent 1×1 convolutional kernels.

[0200]

[0201] Among them, f 1×1 (·) represents a 1×1 convolutional layer operation.

[0202] (2) Multi-head segmentation and encoding: The query matrix, key matrix, and value matrix are segmented into p different attention heads. At the same time, the mask input M is also split into p heads, each head having a dimension of B×C / p×N. The segmented representation is as follows:

[0203]

[0204] Contact(·) is the concatenation operation, where Q1, Q2, ..., Q p For different query matrix headers, K1, K2, ..., K p For different key matrix headers, V1, V2, ..., V p For different value matrix headers, M1, M2, ..., M p For different mask input headers,

[0205] (3) Multi-head self-attention computation, specifically, this process involves processing the query matrix Q j Bond matrix K j Each row of the multiplication result is normalized using the Softmax function. Before performing the Softmax operation, relative position encoding is added. This is done to allow the self-attention mechanism to more effectively capture the relative position information between elements in the sequence. Simultaneously, each value matrix V... j Header and mask matrix M j Element-wise multiplication is performed to incorporate the weight differences of different channels into the value matrix:

[0206]

[0207] Where Encode(·) is the relative position encoding operation, Softmax(·) is the normalization operation, p is the number of attention heads, and Q is the relative position encoding operation. j T For query matrix Q j The transpose of the matrix,

[0208] (4) Feature fusion and output: The outputs of p attention heads are concatenated along the channel dimension and then integrated through 1×1 convolution to obtain the final output X. o :

[0209]

[0210] Among them, f 1×1 (·) represents a 1×1 convolutional layer operation.

[0211] The building unit achieves differentiated weight allocation in the spatiotemporal dimensions through a multi-head self-attention mechanism, which ultimately strengthens the cross-channel and cross-temporal correlation of key features while suppressing noise features.

[0212] In a preferred embodiment of the method, the fourth step involves constructing a multi-scale feature concatenation module (MSFC) to reorganize features and enhance the model's representational and generalization capabilities, including:

[0213] (1) Multi-scale feature segmentation and hierarchical extraction, input feature matrix B represents the batch size, C represents the number of channels, and N represents the feature dimension. An initial mapping is performed using a 1×1 convolution kernel to generate an intermediate feature matrix P1. Then, P1 is divided into n equal parts according to the channel dimension. Next, a 1×1 depthwise separable convolution operation is used for hierarchical feature extraction. The above process can be represented as:

[0214]

[0215] Split(·) is a channel-based partitioning operation. For 1×1 depth-separable convolution operations, For the i-th output sub-feature, P1 i For the i-th intermediate sub-feature,

[0216] (2) Cross-scale feature fusion, outputting each sub-feature The input features are concatenated along the channel dimension and integrated across channels using a 1×1 convolutional kernel. Finally, the original input features and the concatenated features are added element-wise through residual connections to obtain the final output feature P. o :

[0217]

[0218] Contact(·) is the concatenation operation, f 1×1 (·) represents a 1×1 convolutional layer operation.

[0219] The RSBU-CW module reduces the number of parameters by using multi-scale channel partitioning and hierarchical residual fusion, combined with depthwise separable convolution, and improves the model's ability to generalize the time-frequency features of signals through cross-scale cascade recombination.

[0220] In a preferred embodiment of the method, the fifth step involves constructing the basic structure of the generator and discriminator of the generative adversarial network, introducing Wasserstein distance to improve the stability of the model and the quality of data generation, including:

[0221] (1) Construct a generator, assuming a noise vector B represents the batch size, d z Representing the vector dimension, the noise vector first passes through a fully connected layer to obtain the output vector F0:

[0222]

[0223] Where LeakyReLU(·) represents the activation function operation, BN(·) represents the batch normalization operation, W0 represents the fully connected weight matrix, and b0 represents the bias vector.

[0224] (2) After passing through the first deconvolution module and the RSBU-CW module, the output vector F1 is obtained. The RSBU-CW module first removes redundant information from the noise vector for the task of generating spurious signals:

[0225] F1=RSBU_CW(LeakyReLU(BN(ConvT(F0))))

[0226] Where RSBU_CW(·) represents the RSBU-CW module operation, and ConvT(·) represents the deconvolution operation.

[0227] (3) The RCA-SA module is applied after the second deconvolution module to obtain the output vector F2, which captures the weight relationship between different positions of the same signal sample from different channel dimension perspectives:

[0228] F2=RCA_SA(LeakyReLU(BN(ConvT(F1))))

[0229] Here, RCA_SA(·) represents the RCA_SA module operation, constructing the residual channel attention mask self-attention module RCA-SA: through multi-branch feature mapping, multi-head segmentation and encoding, multi-head self-attention calculation, and feature fusion and output, it achieves differentiated weight allocation in the spatiotemporal dimensions, strengthening the cross-channel and cross-temporal correlation of key features while suppressing noisy features. The focus is on improving the model's ability to capture global and local features, and suppressing noisy features.

[0230] (4) The MSFC module is applied after the third deconvolution module to obtain the output vector F3. The information after denoising by the RSBU-CW module and extraction by the RCA-SA module is cascaded at multiple scales to integrate the information:

[0231] F3=MSFC(LeakyReLU(BN(ConvT(F2))))

[0232] Here, MSFC(·) represents the MSFC module operation, which constructs a multi-scale feature cascade module (MSFC). This module reduces the number of parameters through multi-scale feature partitioning and hierarchical extraction, cross-scale feature fusion, and depthwise separable convolution. Furthermore, it enhances the model's generalization ability to the time-frequency features of signals through cross-scale cascade recombination. The key focus is on improving the model's ability to represent and generalize signal features while reducing the number of parameters.

[0233] (5) After the last deconvolution module and reshaping, the spurious vector X is finally generated and input into the discriminator along with the original sample:

[0234] X = Reshape(ConvT(F3))

[0235] Where Reshape(·) represents the reshaping operation,

[0236] (6) Construct a discriminator, which consists of a series of convolutional modules. Each convolutional module includes a convolutional layer followed by batch normalization (BN) and the LeakyReLU activation function. Each convolutional module is followed by a max pooling layer.

[0237] X i =MaxPool(LeakyReLU(BN(Conv)) i (X i-1 ))))

[0238] Where MaxPool(·) represents the global average pooling operation, X i ,X i-1 Conv represents the intermediate feature vector of the discriminator. i This represents the i-th convolution operation.

[0239] (7) Connect the global average pooling layer to reduce the feature dimension to 1 in order to identify the authenticity of the input samples.

[0240]

[0241] Where GAP(·) represents the global average pooling operation. For the discriminator output,

[0242] (8) Introducing the Wasserstein distance, the loss function of the model is constructed. The loss functions (optimization objectives) of the generator and discriminator are as follows:

[0243]

[0244] Where E is the mathematical expectation operator, x is the real sample data, z is the random noise vector in the latent space, G(z) is the synthetic sample generated by the generator based on the noise z, D(x) is the score of the discriminator for the real sample x, min(·) is the minimum value, and max(·) is the maximum value. A generator and discriminator of a generative adversarial network are constructed, and Wasserstein distance is introduced: by constructing the basic structure of the generator and discriminator and introducing Wasserstein distance as part of the loss function, the stability of the model and the quality of data generation are improved. The key is to ensure that the generated pseudo-samples can maintain the characteristics of the original samples, increase the diversity of the dataset, and improve the stability of the model.

[0245] To further illustrate the method of the present invention, a specific embodiment and accompanying drawings are provided below to explain and describe a method for enhancing the operation and maintenance dataset of nuclear power equipment, which is not intended to limit the present invention.

[0246] Figure 1 Figure 2(a) and Figure 2(b) are structural diagrams of the generator and discriminator according to one embodiment of the present invention, where Figure 2(a) and Figure 2(b) are structural diagrams of the generator and discriminator, respectively. Figure 3(a), Figure 3(b), Figure 3(c), and Figure 3(d) are time-domain comparison diagrams of the original signal and the generated signal according to one embodiment of the present invention, where Figure 3(a), Figure 3(b), Figure 3(c), and Figure 3(d) are time-domain diagrams of the original signal, time-domain diagrams obtained by applying the method of the present invention, time-domain diagrams obtained by applying the WGAN method, and time-domain diagrams obtained by applying the FAWGAN method, respectively. Figure 4(a), Figure 4(b), Figure 4(c), and Figure 4(d) are spectral comparison diagrams of the original signal and the generated signal according to one embodiment of the present invention, where Figure 4(a), Figure 4(b), Figure 4(c), and Figure 4(d) are spectral diagrams of the original signal, spectral diagrams obtained by applying the method of the present invention, spectral diagrams obtained by applying the WGAN method, and spectral diagrams obtained by applying the FAWGAN method, respectively.

[0247] According to the nuclear power equipment operation and maintenance data enhancement method of the present invention, an exemplary implementation is provided as follows:

[0248] In the first step, three acceleration sensors were placed on the pump cover of the nuclear power plant steam-driven feedwater pump test bench to collect vibration signals of the impeller in the x-axis, y-axis and z-axis directions. The steam-driven feedwater pump speed was 3000 r / min and the number of impeller blades was 4.

[0249] The network structures of the generator and discriminator are constructed using the modules proposed in the second, third, and fourth steps, as shown in Figures 2(a) and 2(b). The generator consists of a series of deconvolution modules and design modules, mainly implementing upsampling operations. Each deconvolution module includes a deconvolution layer, followed by batch normalization (BN) and the LeakyReLU activation function. The noise vector first passes through a fully connected layer and the first deconvolution module, and then through the RSBU-CW module. This module first removes redundant information from the noise vector to avoid the RCA-SA module focusing on meaningless signal components. The RCA-SA module is applied after the second deconvolution module to deeply capture the weight relationships between different positions of the same signal sample from different channel dimensions. The MSFC module is applied after the third deconvolution module to perform multi-scale feature concatenation on the information denoised by the RSBU-CW module and extracted by the RCA-SA module, integrating the information. After passing through the last deconvolution module and undergoing reshaping, the spurious vector is finally generated and input into the discriminator along with the original sample. The discriminator consists of a series of convolutional modules, similar to the deconvolutional modules of the generator. Each convolutional module includes a convolutional layer, followed by batch normalization (BN) and the activation function LeakyReLU. Each convolutional module is followed by a max pooling layer, and the last convolutional layer is followed by a global average pooling layer to reduce the feature dimension to 1 in order to identify the authenticity of the input samples. The model parameters of the generator and discriminator are shown in Table 1.

[0250] Table 1. Generator and Discriminator Model Parameters

[0251]

[0252] In the fifth step, the data augmentation model is used to generate samples for the collected vibration signals in the z-axis direction. The WGAN model and the FAWGAN model are selected as comparison models. WGAN is a Wasserstein generative adversarial network, and FAWGAN is a fully attention Wasserstein generative adversarial network. The hyperparameter settings of each model are shown in Table 2.

[0253] Table 2 Model Hyperparameter Settings

[0254] Hyperparameters generator Discriminator Learning rate 0.002 0.002 Optimizer RMSProp RMSProp Number of training rounds 3000 3000 Batch size 32 32

[0255] Using z-axis vibration signals, WGAN, FAWGAN, and the proposed method were trained respectively. A generated sample from each model was randomly selected, resulting in time-domain comparison diagrams (Figures 3(a), 3(b), 3(c), and 3(d)) and spectral comparison diagrams (Figures 4(a), 4(b), 4(c), and 4(d)). Ignoring phase differences, the FAWGAN method generated the worst sample quality, with its distribution completely deviating from the original sample distribution. While the WGAN method's generated samples showed an overall similarity to the original samples, there were deviations at certain time points. Compared to other methods, the proposed method generated samples with higher time-domain similarity to the original samples, especially in regions with larger amplitudes. This indicates a significant advantage in capturing the features of the original samples and effectively reconstructing the dynamic changes of the original signal. Similar conclusions can be drawn from the spectral comparison diagrams (Figures 4(a), 4(b), 4(c), and 4(d). Both methods showed significant differences in preserving the main characteristic frequencies and peak frequencies of the real samples compared to the proposed method.

[0256] To more intuitively compare the impact of generated sample quality on fault diagnosis results, samples generated by different models were used, and training sets were constructed based on different numbers of generated samples. The fault diagnosis results were then compared and analyzed. To reduce the influence of randomness, all experiments were repeated 10 times. The diagnostic accuracy and standard deviation are shown in Table 3. It can be seen that as the number of training samples increases, the average fault diagnosis accuracy of each method also increases accordingly. Under different numbers of training samples, the diagnostic accuracy of the method of this invention is significantly higher than other methods, indicating that the method of this invention can generate high-quality pseudo-samples that are highly similar to the original samples.

[0257] Table 3 Comparison of fault diagnosis accuracy (%) under different training samples

[0258] Number of training samples WGAN FAWGAN Method of the present invention 50 74.70±1.66 66.13±1.00 79.67±1.13 100 78.60±2.80 67.5±2.27 80.30±1.54 200 80.23±2.64 64.40±1.07 84.43±2.93

[0259] Channel attention generation is performed by applying global average pooling to the absolute value to obtain a feature matrix. This feature matrix is ​​then input into the attention module to mine channel correlation information, resulting in a new feature matrix. The attention module contains two fully connected layers with a Batch Normalization (BN) layer and a ReLU activation function in between. The feature matrix is ​​normalized using a Sigmoid activation layer. Relationships between channels are mined, and Sigmoid normalization is used to highlight important channels, thereby strengthening key features. Multi-head self-attention computation is then performed. Specifically, this process involves normalizing each row of the result of multiplying the query matrix and the key matrix using a Softmax function. Before performing the Softmax operation, relative position encoding is added to allow the self-attention mechanism to more effectively capture the relative position information between elements in the sequence. Simultaneously, the head and mask matrices of each value matrix are multiplied element-wise to incorporate the weight differences of different channels into the value matrix. This improves the model's ability to capture global and local features and enhances its understanding of relationships between elements within the sequence through relative position encoding.

[0260] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.

Claims

1. A method for enhancing nuclear power equipment operation and maintenance datasets, characterized in that, Includes the following steps: The first step is to collect signals from different health states of nuclear power equipment; The second step is to introduce the residual shrinkage building block RSBU-CW into the network generator to prevent the network from focusing on irrelevant features. The third step is to construct the residual channel attention mask self-attention module RCA-SA to extract global and local features of the signal; The fourth step is to construct a multi-scale feature concatenation module (MSFC) to reorganize features and enhance the model's representation and generalization capabilities. The fifth step involves constructing a generator and discriminator for a generative adversarial network, introducing the Wasserstein distance, and outputting pseudo-samples similar to the original samples to enhance the operation and maintenance data of nuclear power equipment.

2. The method for enhancing nuclear power equipment operation and maintenance datasets according to claim 1, characterized in that, Preferably, in the first step, sensors are used to collect signals from the nuclear power equipment under normal and fault conditions to obtain multi-dimensional operating signals of the nuclear power equipment. The sensors include vibration sensors, displacement sensors, temperature sensors, pressure sensors, and flow sensors.

3. The method for enhancing nuclear power equipment operation and maintenance datasets according to claim 1, characterized in that, The second step includes, Perform feature forward processing, input feature matrix B represents the batch size, C represents the number of channels, and N represents the feature dimension. The feature is passed through two consecutive 1×3 convolutional layers, each followed by a batch normalization layer (BN) and a ReLU activation function, outputting an intermediate feature matrix X2. Where X1 is the output after one convolutional layer, X2 is the output after two consecutive convolutional layers, ReLU(·) is the activation function operation, BN(·) is the batch normalization operation, and f 1×3 (·) represents a 1×3 convolutional layer operation. Channel attention is generated, and global average pooling is performed on the absolute value X2 to obtain the feature matrix. The feature matrix A is then input into the attention module to mine the correlation information of the channels, resulting in the feature matrix. The attention module consists of two fully connected layers, with a Batch Normalization (BN) layer and a ReLU activation function in between. The feature matrix S is normalized by a Sigmoid activation layer, and the resulting value is denoted as α. i : Where, α i The activation value calculated for each channel; s i Let be the eigenvalue of the i-th channel of the feature matrix S. Adaptive threshold calculation is performed by multiplying the channel attention weights α with the feature matrix A channel by channel to generate a dynamic threshold matrix T. Τ=α⊙A Here, ⊙ represents the element-wise dot product operation. Soft thresholding is performed on feature matrix X2 to obtain feature matrix X3. The soft thresholding process is as follows: Where y is the output of the soft thresholding process, x is the input before the soft thresholding process, and τ is the threshold, which is determined by the statistics of the corresponding channel in the dynamic threshold matrix T. Perform a residual connection, adding the input feature matrix X to the soft-thresholding output matrix X3 element-wise to obtain the final output matrix Y: Y = X + X3 The residual shrinkage building unit RSBU-CW achieves differentiated weighting of feature channels through a channel attention mechanism, eliminates redundant information by combining soft thresholding, and ensures gradient propagation stability by utilizing the residual structure.

4. The method for enhancing nuclear power equipment operation and maintenance datasets according to claim 1, characterized in that, The third step includes, Perform multi-branch feature mapping, input feature matrix B represents the batch size, C represents the number of channels, and N represents the feature dimension. The query matrix Q, key matrix K, and value matrix V are generated using three independent 1×1 convolutional kernels. Among them, f 1×1 (·) represents a 1×1 convolutional layer operation. Multi-head segmentation and encoding are performed, dividing the query matrix, key matrix, and value matrix into p different attention heads. Simultaneously, the mask input M is also split into p heads, each with dimensions B×C / p. The segmented representation is as follows: Contact(·) is the concatenation operation, where Q1, Q2, ..., Q p For different query matrix headers, K1, K2, ..., K p For different key matrix headers, V1, V2, ..., V p For different value matrix headers, M1, M2, ..., M p For different mask input headers, Perform multi-head self-attention calculation, and use the query matrix Q j Bond matrix K j Each row of the multiplication result is normalized using the Softmax function. Before performing the Softmax operation, relative position encoding is added to allow the self-attention mechanism to more effectively capture the relative position information between elements in the sequence. Simultaneously, each value matrix V... j Header and mask matrix M j Element-wise multiplication is performed to incorporate the weight differences of different channels into the value matrix: Where Encode(·) is the relative position encoding operation, Softmax(·) is the normalization operation, p is the number of attention heads, and Q is the relative position encoding operation. j T For query matrix Q j The transpose of the matrix, Feature fusion and output are performed by concatenating the outputs of p attention heads along the channel dimension and then integrating them through 1×1 convolution to obtain the final output X. o : Among them, f 1×1 (·) represents a 1×1 convolutional layer operation. By implementing a multi-head self-attention mechanism to achieve differentiated weight allocation in the spatiotemporal dimensions, the cross-channel and cross-temporal correlation of key features is strengthened while suppressing noisy features.

5. The method for enhancing nuclear power equipment operation and maintenance datasets according to claim 1, characterized in that, The fourth step includes, Multi-scale feature segmentation and hierarchical extraction, input feature matrix B represents the batch size, C represents the number of channels, and N represents the feature dimension. An initial mapping is performed using a 1×1 convolution kernel to generate an intermediate feature matrix P1. Then, P1 is divided into n equal parts according to the channel dimension. Next, a 1×1 depthwise separable convolution operation is used for hierarchical feature extraction. The above process can be represented as: Split(·) is a channel-based partitioning operation. For 1×1 depth-separable convolution operations, For the i-th output sub-feature, For the i-th intermediate sub-feature, Perform cross-scale feature fusion and output each sub-feature. The input features are concatenated along the channel dimension and integrated across channels using a 1×1 convolutional kernel. Finally, the original input features and the concatenated features are added element-wise through residual connections to obtain the final output feature P. o : Contact(·) is the concatenation operation, f 1×1 (·) represents a 1×1 convolutional layer operation. By using multi-scale channel partitioning and hierarchical residual fusion, combined with depthwise separable convolution to reduce the number of parameters, and by cross-scale cascading recombination to improve the model's ability to generalize the time-frequency features of the signal.

6. The method for enhancing nuclear power equipment operation and maintenance datasets according to claim 1, characterized in that, The fifth step includes: Construct a generator, assuming a noise vector B represents the batch size, d z Representing the vector dimension, the noise vector first passes through a fully connected layer to obtain the output vector F0: Where LeakyReLU(·) represents the activation function operation, BN(·) represents the batch normalization operation, W0 represents the fully connected weight matrix, and b0 represents the bias vector. After passing through the first deconvolution module and the RSBU-CW module, the output vector F1 is obtained. The RSBU-CW module first removes redundant information from the noise vector to address the task of generating spurious signals. F1=RSBU_CW(LeakyReLU(BN(ConvT(F0)))) Where RSBU_CW(·) represents the RSBU-CW module operation, and ConvT(·) represents the deconvolution operation. The RCA-SA module is applied after the second deconvolution module to obtain the output vector F2, which captures the weight relationship between different positions of the same signal sample from different channel dimensions: F2=RCA_SA(LeakyReLU(BN(ConvT(F1)))) Where RCA_SA(·) represents the RCA_SA module operation, The MSFC module is applied after the third deconvolution module to obtain the output vector F3. The information obtained after denoising by the RSBU-CW module and extraction by the RCA-SA module is then subjected to multi-scale feature concatenation for information integration. F3=MSFC(LeakyReLU(BN(ConvT(F2)))) In this context, MSFC(·) represents the MSFC module operation. After the final deconvolution module and reshaping, the spurious vector X is finally generated and input into the discriminator along with the original sample: X = Reshape(ConvT(F3)) Where Reshape(·) represents the reshaping operation, Construct a discriminator, which consists of a series of convolutional modules. Each convolutional module includes a convolutional layer, followed by batch normalization (BN) and the LeakyReLU activation function. Each convolutional module is then followed by a max-pooling layer. X i =MaxPool(LeakyReLU(BN(Conv i (X i-1 )))) Where MaxPool(·) represents the global average pooling operation, X i ,X i-1 Conv represents the intermediate feature vector of the discriminator. i This represents the i-th convolution operation. Finally, a global average pooling layer is applied to reduce the feature dimension to 1, in order to identify the authenticity of the input samples. Where GAP(·) represents the global average pooling operation. For the discriminator output, Introducing the Wasserstein distance, we construct the loss function for the model. The loss functions for the generator and discriminator are as follows: Where E is the mathematical expectation operator, x is the real sample data, z is the random noise vector in the latent space, G(z) is the synthetic sample generated by the generator based on the noise z, D(x) is the score of the discriminator for the real sample x, min(·) is to take the minimum value, and max(·) is to take the maximum value.

7. The method for enhancing nuclear power equipment operation and maintenance datasets according to claim 1, characterized in that, Nuclear power equipment includes nuclear power steam-driven feedwater pumps.

8. A system for implementing the method of any one of claims 1-7, characterized in that, It includes: The acquisition module collects signals from different health states of nuclear power equipment. The first building block is used to introduce the residual shrinkage building block RSBU-CW into the network generator to prevent the network from focusing on irrelevant features; The second building unit is used to construct the residual channel attention mask self-attention module RCA-SA, which is used to extract global and local features of the signal. The third building block is used to construct the Multi-Scale Feature Concatenation Module (MSFC) to reorganize features and enhance the model's representation and generalization capabilities. The enhancement unit is used to construct the generator and discriminator of the generative adversarial network. It introduces the Wasserstein distance and outputs pseudo-samples similar to the original samples to enhance the operation and maintenance data of nuclear power equipment.

9. A computer storage medium, characterized in that, The storage medium includes computer instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any one of claims 1-7.

Images