Centrifugal pump cavitation migration identification method based on multi-adversarial attention network

Through the method of multi-adversarial attention network, the problem of performance degradation across working conditions in centrifugal pump cavitation recognition is solved, efficient cavitation recognition under different working conditions is achieved, and the recognition accuracy and feature extraction ability of the model are improved.

CN120744664APending Publication Date: 2025-10-03ZHEJIANG UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202510830915.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-06-20
Publication Date
2025-10-03

AI Technical Summary

Technical Problem

Existing deep learning models have problems in cavitation identification under multiple working conditions of centrifugal pumps, such as performance degradation across working conditions, poor recognition performance due to distribution differences, and information compression, and lack effective deep domain adaptation technology.

Method used

A method based on multi-adversarial attention network is adopted to generate time-frequency maps through continuous wavelet transform. Combined with the residual convolutional attention feature extractor and the multi-adversarial distribution alignment module, the loss function is dynamically adjusted to achieve accurate extraction and classification of cavitation noise features.

Benefits of technology

The accuracy of cavitation identification under different working conditions is improved, the feature expression and distribution alignment capabilities are enhanced, and the generalization performance of the model is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120744664A_ABST
    Figure CN120744664A_ABST
Patent Text Reader

Abstract

The invention discloses a centrifugal pump cavitation migration identification method based on a multi-adversarial attention network, and the method comprises the steps: collecting centrifugal pump sound signals of different cavitation severity degrees under different working conditions, and constructing a corresponding source domain and a target domain; preprocessing the sound signals from the source domain and the target domain based on continuous wavelet transform to generate a time-frequency map; constructing a multi-adversarial depth field adaptive network model comprising a residual convolution attention feature extractor, a multi-adversarial distribution alignment module and a cavitation state classifier; simultaneously inputting marked source domain data and unmarked target domain data into the constructed model, and carrying out gradient updating on network parameters by utilizing a back propagation algorithm; and when the total loss function converges or the network is subjected to iteration updating for a certain number of times, exporting the whole model, and inputting target domain data into the model to obtain a cavitation recognition result. According to the invention, accurate extraction and classification of cavitation noise time-frequency features can be realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of centrifugal pump cavitation state recognition, and in particular relates to a centrifugal pump cavitation migration recognition method based on a multi-adversarial attention network. Background Art

[0002] Cavitation is a common abnormal flow phenomenon in centrifugal pumps, which may cause hazards such as decreased hydraulic performance and increased vibration and noise. Therefore, accurately identifying the cavitation state of a centrifugal pump is of great significance to ensuring the safe, stable and efficient operation of the centrifugal pump.

[0003] In recent years, deep learning-based cavitation state recognition methods have made significant progress, achieving automatic feature extraction and accurate classification and recognition of cavitation signals. Since centrifugal pumps often operate under a variety of operating conditions, changes in operating parameters such as flow rate and speed place higher demands on the generalization performance of cavitation recognition models. Due to differences in signal distribution between different operating conditions, models trained under existing operating conditions (source domain) will experience a significant decline in cavitation recognition performance when faced with new operating conditions (target domain). Furthermore, because signals in the target domain may lack cavitation category labels, direct supervised training is not possible.

[0004] Deep domain adaptation is a promising transfer learning method that can effectively address this issue. It maps source and target domain data into the same feature space, enabling cavitation recognition in the unlabeled target domain. For example, Chinese patent publication CN114611617A discloses a deep domain adaptation image classification method based on a prototype network; and Chinese patent publication CN109766921A discloses a vibration data fault classification method based on deep domain adaptation.

[0005] However, there is currently no mature deep domain adaptation technology for cavitation identification in centrifugal pumps. At the same time, existing deep learning and transfer learning recognition models still have the following limitations: (1) Due to distribution differences, traditional deep learning methods show performance degradation in cross-operating recognition tasks; (2) Most domain adaptation methods ignore conditional or joint distribution alignment, resulting in poor recognition performance in scenarios with significant distribution differences; (3) The classic attention mechanism has the limitation of information compression, which hinders the effective extraction of discriminant features in the time-frequency graph after signal processing.

[0006] In order to solve the problems of complex domain differences caused by the multi-operating conditions of centrifugal pumps and the susceptibility of key cavitation discrimination features to interference, it is a worthwhile attempt to adopt advanced signal processing methods and transfer learning methods. Summary of the Invention

[0007] The present invention provides a centrifugal pump cavitation migration identification method based on a multi-adversarial attention network, which can accurately extract and classify the time-frequency characteristics of cavitation noise.

[0008] A centrifugal pump cavitation migration recognition method based on a multi-adversarial attention network includes the following steps:

[0009] (1) Collect the acoustic signals of centrifugal pumps with different cavitation severity under different working conditions and construct the corresponding source and target domains;

[0010] (2) Preprocessing the acoustic signals from the source and target domains based on continuous wavelet transform to generate time-frequency spectra;

[0011] (3) Construct a multi-adversarial deep domain adaptation network model including a residual convolutional attention feature extractor, a multi-adversarial distribution alignment module, and a cavitation state classifier;

[0012] (4) The labeled source domain data and the unlabeled target domain data are simultaneously input into the model constructed in step (3), and the total loss function composed of the classification loss function, the joint loss function and the conditional loss function is calculated by the forward propagation algorithm, and the network parameters are updated using the gradient of the backpropagation algorithm;

[0013] (5) When the total loss function converges or the network has been iteratively updated a certain number of times, the entire model is exported and the target domain data is input into the model to obtain the cavitation recognition result.

[0014] The present invention uses continuous wavelet transform to preprocess the cavitation noise signal to generate a time-frequency diagram, inputs the result into the constructed network model, combines the residual convolution attention feature extractor with the multi-adversarial distribution alignment module, and dynamically measures the joint probability distribution and conditional probability distribution to achieve accurate extraction and classification of cavitation noise time-frequency features.

[0015] Furthermore, the specific process of step (2) is:

[0016] First calculate the wavelet coefficient cwt(s,τ), the formula is:

[0017]

[0018] Where t is time; s is the scale parameter; τ is the translation parameter; * is the conjugate symbol; x(·) is the acoustic signal; ψ(·) is the wavelet function;

[0019] By adjusting the scale parameter and translation parameter, a wavelet coefficient matrix containing different time-frequency positions is obtained. The modulus value of the wavelet coefficient matrix is ​​used as the intensity value. A two-dimensional heat map is drawn with time as the horizontal axis and the scale-converted frequency as the vertical axis, thus obtaining a time-frequency spectrum.

[0020] Furthermore, in step (3), the residual convolutional attention feature extractor includes a multi-scale global hybrid attention module and a deep residual network;

[0021] The multi-scale global hybrid attention module achieves enhanced feature expression through sequentially constructed channel attention mechanism and spatial attention mechanism;

[0022] The deep residual network is composed of a ResNet-50 network, which includes five sub-modules. The first sub-module consists of a convolutional layer, a normalization layer, an activation function, and a maximum pooling layer; the second to fifth sub-modules each contain a residual block I and several residual blocks II;

[0023] The multi-scale global hybrid attention module is embedded between the second and third submodules of the deep residual network to form a residual convolutional attention feature extractor.

[0024] Furthermore, the channel attention mechanism first splits the input feature map F into three scales of 1×1, 2×2 and 4×4 respectively;

[0025] Then, average pooling and maximum pooling are performed simultaneously along the spatial direction in the segmented area, where the feature vectors AP of three scales are obtained by average pooling 1×1 (F), AP 2×2 (F), AP 4×4 (F), the feature vectors MP of the other three scales are obtained through maximum pooling 1×1 (F), MP 2×2 (F), MP 4×4 (F);

[0026] The feature vectors that have undergone the same pooling operation are concatenated to form two 21-dimensional combined feature vectors; the two combined feature vectors are processed by a shared multi-layer perceptron to generate corresponding attention logarithms; the attention logarithms are summed element-by-element through the sigmoid activation function to generate attention weights.

[0027] Furthermore, the spatial attention mechanism uses three independent 1×1 convolution operations to project the feature map F into three compressed embeddings to generate K matrix, Q matrix and V matrix; then flatten the spatial dimension to reshape the projected vector into a two-dimensional matrix, perform matrix multiplication on the transpose of the Q matrix and the K matrix, calculate the global similarity between each pair of spatial points, normalize it by the row softmax function, and obtain the spatial attention weight; then multiply the spatial attention weight with the V matrix to obtain the attention feature matrix V ′ ; Finally, the attention feature matrix is ​​reconstructed through 1×1 convolution to restore the original channel dimension.

[0028] Furthermore, in step (4), the predicted probability distribution of the high-dimensional feature vector is generated by the cavitation state classifier, and the difference between the source domain true label and the predicted distribution is calculated, that is, the classification loss L y ;

[0029] By calculating the joint loss L in the multi-adversarial distribution alignment module j and conditional loss L c , aligning the joint distribution with the conditional distribution.

[0030] Furthermore, in step (4), the formula of the total loss function is as follows:

[0031] L=L y +αL da ;

[0032] Where, L is the total loss function; L y is the classification loss; α is the hyperparameter; L da is the overall migration loss function, and the formula is:

[0033] L da =μL j +(1-μ)L c ;

[0034] Where, L j is the joint loss; L c is the conditional loss; μ is the dynamic adaptive factor.

[0035] The calculation formula of the dynamic adaptive factor μ is as follows:

[0036]

[0037] Where, A distance representing the joint probability; A represents the distance of the conditional probability distribution; C represents the total number of cavitation categories.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. This paper proposes a residual convolutional attention feature extractor to extract highly discriminative features from cavitation time-frequency maps. A multi-scale global hybrid attention module is embedded in a deep residual network to selectively enhance the representation of features related to cavitation.

[0040] 2. The present invention proposes a multi-adversarial distribution alignment module, which uses the category information provided by the cavitation state classifier to adaptively align the joint distribution and conditional distribution.

[0041] 3. The present invention proposes an adaptive weighting factor to dynamically weight the joint loss and conditional loss to further narrow the difference between the source domain and the target domain and improve the performance of the network model in recognizing cavitation states. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of a centrifugal pump cavitation migration identification method based on a multi-adversarial attention network in the present invention.

[0043] Figure 2 This is a structural diagram of the multi-adversarial deep domain adaptive network model in an embodiment of the present invention.

[0044] Figure 3 This is a structural diagram of the multi-scale global hybrid attention module in an embodiment of the present invention.

[0045] Figure 4 This is a structural diagram of the residual convolutional attention feature extractor in an embodiment of the present invention.

[0046] Figure 5 Schematic diagram of the centrifugal pump group experimental platform in an embodiment of the present invention.

[0047] Figure 6 These are time-frequency diagrams corresponding to four different cavitation states of a centrifugal pump under two different working conditions in an embodiment of the present invention.

[0048] Figure 7 1 is a confusion matrix diagram of three comparison methods in the embodiments of the present invention and the method of the present invention. DETAILED DESCRIPTION

[0049] The present invention will be described in further detail below with reference to the accompanying drawings and examples. It should be noted that the following examples are intended to facilitate understanding of the present invention and do not have any limiting effect on the present invention.

[0050] like Figure 1 As shown, a centrifugal pump cavitation migration recognition method based on a multi-adversarial attention network includes the following steps:

[0051] S01, collect the acoustic signals of centrifugal pumps with different cavitation severity under different working conditions and construct the corresponding source domain and target domain.

[0052] S02, preprocessing the acoustic signal based on continuous wavelet transform to generate a time-frequency spectrum.

[0053] First calculate the wavelet coefficient cwt(s,τ):

[0054]

[0055] Where t is time, s is the scale parameter, τ is the translation parameter, * is the conjugate symbol, x(·) is the acoustic signal, and ψ(·) is the wavelet function. By adjusting the scale parameter and translation parameter, a wavelet coefficient matrix containing different time-frequency positions is obtained. The modulus of the wavelet coefficient matrix is ​​used as the intensity value. A two-dimensional heat map is drawn with time as the horizontal axis and the scale-converted frequency as the vertical axis, thus obtaining a time-frequency spectrum.

[0056] S03, build a multi-adversarial deep domain adaptation network including residual convolutional attention feature extractor, multi-adversarial distribution alignment module and cavitation state classifier, such as Figure 2 shown.

[0057] The residual convolutional attention feature extractor consists of a multi-scale global mixed attention module and a deep residual network.

[0058] (3-1) Figure 3 As shown in Figure 2, the multi-scale global hybrid attention module achieves enhanced feature expression through the sequentially constructed channel attention mechanism and spatial attention mechanism. The channel attention mechanism first divides the input feature map into three scales of 1×1, 2×2 and 4×4, and then obtains AP by averaging pooling along the spatial direction in the divided area. 1×1 (F), AP 2×2 (F), AP 4×4 (F) and maximum pooling to get MP 1×1 (F), MP 2×2 (F), MP 4×4 (F) Concatenate the feature vectors from different scales that have undergone the same pooling operation to form two 21-dimensional combined feature vectors. Use a shared multi-layer perceptron to process the combined feature vectors after the two different pooling operations to generate the corresponding attention logarithms. Use the sigmoid activation function to sum the attention logarithms element by element to generate the attention weight. Calculate the channel attention weight W c (F):

[0059] W c (F) = σ{MLP[cat(AP 1×1 (F),AP 2×2 (F),AP 4×4 (F))]+

[0060] MLP[cat(MP 1×1 (F),MP 2×2 (F),MP 4×4 (F))]}

[0061] Where F represents the input feature map; σ represents the sigmoid activation function; cat represents the concatenation operation.

[0062] (3-2) The spatial attention mechanism uses three independent 1×1 convolution operations to project the feature map into three compressed embeddings, generating the K matrix, the Q matrix, and the V matrix. The spatial dimensions are then flattened to reshape the projected vector into a two-dimensional matrix. The Q matrix is ​​multiplied by the transpose of the K matrix, and the global similarity between each pair of spatial points is calculated. The spatial attention weights are obtained by normalizing them using the row-wise softmax function.

[0063] Multiply the spatial attention weights by the V matrix to obtain the attention feature matrix V′:

[0064] V ′ =softmax(Q·K T )·V;

[0065] Finally, the attention feature matrix is ​​reconstructed through 1×1 convolution to restore the original channel dimension.

[0066] (3-3) The deep residual network is composed of the ResNet-50 network, which includes five sub-modules. The first sub-module consists of a convolutional layer, a normalization layer, an activation function and a maximum pooling layer; modules two to five are composed of a residual block I and several residual blocks II.

[0067] (3-4) The multi-scale global hybrid attention module is embedded between the second and third submodules of the deep residual network to form a residual convolutional attention feature extractor, such as Figure 4 shown.

[0068] S04: Input the labeled source domain data and the unlabeled target domain data into the proposed model simultaneously, calculate the cross entropy loss of the source domain in the cavitation state classifier and the transfer loss function of the multi-adversarial distribution alignment module, and perform gradient update and backpropagation to update the network parameters. Specifically, it includes:

[0069] (4-1) The cavitation state classifier consists of a fully connected layer and a softmax function to generate a predicted probability distribution of a high-dimensional feature vector. The difference between the true source domain label and the predicted distribution is calculated, i.e., the classification loss L y :

[0070]

[0071] Where n s represents the number of source domain samples; y i represents the label of the i-th source domain sample; p j represents the softmax value of the jth cavitation state; C represents the total number of cavitation categories; I(·) represents the indicator function, when y i =j, I = 1, when y i When ≠j, I=0.

[0072] (4-2) The multi-adversarial distribution alignment module consists of multiple basic domain adversarial networks, and aligns the joint distribution with the conditional distribution through the loss function. Calculate the joint loss L in the multi-adversarial distribution alignment module j :

[0073]

[0074] Where n s Indicates the number of source domain samples; n t represents the number of samples in the target domain; ω i represents the weight of the i-th feature; BCE[·] represents the cross entropy function; JDD(·) represents the joint domain discriminator; J i A joint variable representing the feature vector and the class; represents the domain label of the i-th sample, when Represented as the source domain, when Represented as the target domain.

[0075] (4-3) Calculate the conditional loss L in the multi-adversarial distribution alignment module c :

[0076]

[0077] Where n s Indicates the number of source domain samples; n t represents the number of target domain samples; P ij represents the predicted probability of the i-th sample for the j-th category; BCE[·] represents the cross entropy function; CDD j (·) represents the j-th conditional domain discriminator; f i represents i eigenvectors; represents the domain label of the i-th sample, when Represented as the source domain, when Represented as the target domain.

[0078] (4-4) Calculate the dynamic adaptive factor μ used to measure the alignment joint probability distribution and the alignment conditional probability distribution:

[0079]

[0080] Where, Represents the joint probability distance; Represents the conditional probability distribution Distance; L j represents joint loss; L c represents the condition loss; C represents the total number of cavitation categories.

[0081] (4-5) Calculate the overall migration loss L using the dynamic adaptive factor da :

[0082] L da =μL j +(1-μ)L c ;

[0083] (4-6) Calculate the total loss function L of the network, including the cross entropy loss function L of the source domain data c And the overall transfer loss function L da , the specific formula is:

[0084] L=L y +αL da ;

[0085] Where α is a hyperparameter.

[0086] (4-7) The model is back-propagated based on the total loss function, and the network parameters are updated using the stochastic gradient descent algorithm.

[0087] S05, when the loss function converges or the network has been iteratively updated a certain number of times, the entire model is exported, and the target domain data is input into the aforementioned model to obtain the cavitation recognition result.

[0088] To verify the effectiveness of the present invention, the cavitation state identification task of a centrifugal pump under different working conditions is analyzed as follows:

[0089] A centrifugal pump cavitation simulation experiment was conducted on a closed test bench in the laboratory. The schematic diagram of the test bench is shown in the figure. Figure 5 As shown in the figure, the pump inlet pressure is lowered by reducing the inlet valve opening. If necessary, a vacuum pump is used to simulate cavitation. As the inlet pressure decreases, the NPSH gradually decreases. High-speed photography can be used to observe the generation and development of cavitation bubbles at the impeller inlet, and acoustic signals related to various cavitation states are collected. The key parameters of the centrifugal pump are shown in Table 1.

[0090] Table 1

[0091] parameter value Rated speed 1450rpm Specific speed 74.93 Design flow <![CDATA[30m 3 / h]]> Design head 12m Outlet diameter 65mm Inlet diameter 80mm

[0092] According to the head drop rate, there are four different cavitation states: none (NC), incipient cavitation (IC), obvious cavitation (OC), and severe cavitation (SC). The data collection of the centrifugal pump cavitation test bench was carried out under four different working conditions (speed and flow): 923 rpm and 5 m 3 / h, 923rpm and 6m 3 / h, 1029rpm and 5m 3 / h, 1029rpm and 6m 3 / h. The high-speed camera frame rate was 5000 fps, and the acoustic signal acquisition frequency was 40 kHz. Each sample was sampled for 0.3 s, with 200 samples collected for each cavitation state under different operating conditions. Dataset construction information is shown in Table 2.

[0093] Table 2

[0094]

[0095]

[0096] There are 12 migration tasks constructed: A→B, A→C, A→D, B→A, B→C, B→D, C→A, C→B, C→D, D→A, D→B, D→C; A→B is used as an example for explanation: A represents the source domain dataset, and B represents the target domain dataset for identification.

[0097] The continuous wavelet transform method is used to transform the acoustic signal into a time-frequency spectrum. The time-frequency spectrum of different cavitation states of datasets B and D is shown in Figure 2. Figure 6 shown.

[0098] In order to fully demonstrate the superiority of the present invention, six existing classic methods (ResNet-50, DAN, JAN, DSAN, DCORAL, DANN) were compared with the method of the present invention (MAAN) in an experiment.

[0099] The target domain recognition accuracy of the seven methods under 12 cavitation recognition tasks is shown in Table 3. ResNet-50 shows the worst recognition performance in almost all tasks. This is because ResNet-50 does not integrate any distribution alignment module, resulting in poor performance in cross-condition recognition tasks. Among the four explicit metric-based methods, DSAN has the best recognition effect, followed by JAN, DAN the worst, and DCORAL the worst. Although DCORAL calculates the second-order statistical differences between domains, it lacks kernel mapping of features, resulting in the worst performance in terms of metric distribution differences. By considering the distribution differences between subdomains, DSAN makes more in-depth use of category information to align conditional distributions, and therefore generally outperforms JAN in recognition accuracy. Among the implicit metric-based methods, the proposed MAAN achieves the highest recognition accuracy in all tasks, highlighting the superiority of the feature extraction and distribution alignment modules in MAAN.

[0100] Table 3

[0101]

[0102]

[0103] In order to compare the judgment ability of each model on the specific cavitation state category in more detail, taking the migration task B→D as an example, the target domain confusion matrices of the three comparison methods (ResNet-50, DAN, JAN) and the proposed method are shown as follows: Figure 7 As shown in Figure 3, different methods perform well in identifying SC, while their recognition capabilities for the other three categories vary. In contrast, MAAN has a clear advantage in identifying early cavitation states.

[0104] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A centrifugal pump cavitation migration recognition method based on multi-adversarial attention network, characterized by: The following steps are involved: (1) Collect the acoustic signals of centrifugal pumps with different cavitation severity under different working conditions and construct the corresponding source and target domains; (2) Preprocessing the acoustic signals from the source and target domains based on continuous wavelet transform to generate time-frequency spectra; (3) Construct a multi-adversarial deep domain adaptation network model including a residual convolutional attention feature extractor, a multi-adversarial distribution alignment module, and a cavitation state classifier; (4) The labeled source domain data and the unlabeled target domain data are simultaneously input into the model constructed in step (3), and the total loss function composed of the classification loss function, the joint loss function and the conditional loss function is calculated by the forward propagation algorithm, and the network parameters are updated using the gradient of the backpropagation algorithm; (5) When the total loss function converges or the network has been iteratively updated a certain number of times, the entire model is exported and the target domain data is input into the model to obtain the cavitation recognition result.

2. The centrifugal pump cavitation migration identification method based on multi-adversarial attention network according to claim 1 is characterized in that: The specific process of step (2) is: First calculate the wavelet coefficient cwt(s,τ) , The formula is: Where t is time; s is the scale parameter; τ is the translation parameter; * is the conjugate symbol; x(·) is the acoustic signal; ψ(·) is the wavelet function; By adjusting the scale parameter and translation parameter, a wavelet coefficient matrix containing different time-frequency positions is obtained. The modulus value of the wavelet coefficient matrix is ​​used as the intensity value. A two-dimensional heat map is drawn with time as the horizontal axis and the scale-converted frequency as the vertical axis, thus obtaining a time-frequency spectrum.

3. The centrifugal pump cavitation migration identification method based on multi-adversarial attention network according to claim 1 is characterized in that: In step (3), the residual convolutional attention feature extractor includes a multi-scale global hybrid attention module and a deep residual network; The multi-scale global hybrid attention module achieves enhanced feature expression through sequentially constructed channel attention mechanism and spatial attention mechanism; The deep residual network is composed of a ResNet-50 network, which includes five sub-modules. The first sub-module consists of a convolutional layer, a normalization layer, an activation function, and a maximum pooling layer; the second to fifth sub-modules each contain a residual block I and several residual blocks II; The multi-scale global hybrid attention module is embedded between the second and third submodules of the deep residual network to form a residual convolutional attention feature extractor.

4. The centrifugal pump cavitation migration identification method based on multi-adversarial attention network according to claim 3 is characterized in that: The channel attention mechanism first splits the input feature map F into three scales: 1×1, 2×2, and 4×4; Then, average pooling and maximum pooling are performed simultaneously along the spatial direction in the segmented area, where the feature vectors AP of three scales are obtained by average pooling 1×1 (F), AP 2×2 (F), AP 4×4 (F), the feature vectors MP of the other three scales are obtained through maximum pooling 1×1 (F), MP 2×2 (F), MP 4×4 (F); The feature vectors that have undergone the same pooling operation are concatenated to form two 21-dimensional combined feature vectors; the two combined feature vectors are processed by a shared multi-layer perceptron to generate corresponding attention logarithms; the attention logarithms are summed element-by-element through the sigmoid activation function to generate attention weights.

5. The centrifugal pump cavitation migration identification method based on multi-adversarial attention network according to claim 3 is characterized in that: The spatial attention mechanism uses three independent 1×1 convolution operations to project the feature map F into three compressed embeddings, generating K matrix, Q matrix and V matrix; then flattening the spatial dimensions to reshape the projected vector into a two-dimensional matrix, performing matrix multiplication on the transpose of the Q matrix and the K matrix, calculating the global similarity between each pair of spatial points, and normalizing it with the row softmax function to obtain the spatial attention weight; Then multiply the spatial attention weight by the V matrix to obtain the attention feature matrix V ′ ; Finally, the attention feature matrix is ​​reconstructed through 1×1 convolution to restore the original channel dimension.

6. The centrifugal pump cavitation migration identification method based on multi-adversarial attention network according to claim 1 is characterized in that: In step (4), the predicted probability distribution of the high-dimensional feature vector is generated by the cavitation state classifier, and the difference between the source domain true label and the predicted distribution is calculated, that is, the classification loss L y ; By calculating the joint loss L in the multi-adversarial distribution alignment module j and conditional loss L c , aligning the joint distribution with the conditional distribution.

7. The centrifugal pump cavitation migration identification method based on multi-adversarial attention network according to claim 1 is characterized in that: In step (4), the formula of the total loss function is as follows: L=L y +αL da ; Where, L is the total loss function; L y is the classification loss; α is the hyperparameter; L da is the overall migration loss function, and the formula is: L da =μL j +(1-μ)L c ; Where, L j is the joint loss; L c is the conditional loss; μ is the dynamic adaptive factor.

8. The centrifugal pump cavitation migration identification method based on multi-adversarial attention network according to claim 7 is characterized in that: The calculation formula of the dynamic adaptive factor μ is as follows: Where, A distance representing the joint probability; A represents the distance of the conditional probability distribution; C represents the total number of cavitation categories.

Citation Information

Patent Citations

  • A vibration data fault classification method based on depth domain adaptation

    CN109766921A

  • Depth field adaptive image classification method based on prototype network

    CN114611617A