Pump cavitation state intelligent identification method based on multi-scale fusion DCNN

Abstract

The invention discloses a pump cavitation state intelligent identification method based on multi-scale fusion DCNN, and the method comprises the steps: (1) collecting noise signals when a centrifugal pump is in different cavitation states, and segmenting the noise signals into a plurality of noise signal samples in a non-overlapping manner; (2) acquiring a wavelet time-frequency diagram of each noise signal sample, constructing a time-frequency diagram data set, and dividing the time-frequency diagram data set into a training set, a verification set and a test set; (3) constructing a time-frequency enhanced wavelet fusion multi-scale interpretable DCNN model, wherein the model comprises a time-frequency attention module, a fusion multi-scale module and a fusion class activation mapping module; (4) training the constructed model based on the training set, and storing optimal parameters of the model based on the verification set; and (5) testing the model based on the test set, and outputting the predicted cavitation state and the class activation diagram corresponding to each sample. According to the method, high classification accuracy can be obtained in the cavitation state recognition task of the centrifugal pump, and high interpretability is achieved.

Description

Technical Field

[0001] This invention belongs to the field of centrifugal pump cavitation state recognition, and in particular relates to an intelligent recognition method for pump cavitation state based on multi-scale fusion DCNN. Background Technology

[0002] Cavitation is a common abnormal flow phenomenon in centrifugal pumps, leading to decreased hydraulic performance, increased vibration and noise, and even impeller cavitation. Therefore, accurately identifying the cavitation state of centrifugal pumps is of great engineering significance. Since sensor monitoring signals can sensitively reflect the occurrence and development of cavitation, some researchers analyze monitoring signals such as vibration, noise, and pressure pulsations to determine the cavitation state; however, this relies heavily on signal processing techniques and expert knowledge. In recent years, the continuous development of artificial intelligence technology has sparked a surge of research into cavitation identification based on deep learning. Deep convolutional neural networks (DCNNs) are a representative type of deep learning network model, possessing advantages such as local receptive fields and parameter sharing. They have been widely adopted by researchers in the field of cavitation state identification for automatic feature extraction and classification of one-dimensional signals or two-dimensional images.

[0003] For example, Chinese patent document CN117009870A discloses a pump cavitation state identification method based on frequency domain improved SDP diagram. By using deep learning to detect pump cavitation state, it automatically learns complex features in vibration signals, thereby improving the accuracy of cavitation phenomenon detection.

[0004] Chinese patent document CN111401136A discloses a method for detecting the cavitation degree of a plunger pump. The method uses the conversion of vibration signals into RGB images and combined with a convolutional neural network to identify the cavitation degree of the plunger pump.

[0005] However, existing cavitation state recognition methods based on DCNN still have two major shortcomings: (1) Most DCNNs use single-scale convolution kernels, so the feature extraction effect of cavitation signals is not good, and the accuracy of cavitation recognition still needs to be improved; (2) Since DCNN is essentially a black box model, the process of its output prediction results is difficult for humans to understand, that is, the interpretability is poor. Summary of the Invention

[0006] This invention discloses an intelligent identification method for pump cavitation state based on multi-scale fusion DCNN, which not only improves the accuracy of centrifugal pump cavitation identification, but also highlights the cavitation-sensitive frequency band, thereby enhancing the interpretability of the model.

[0007] A method for intelligent identification of pump cavitation state based on multi-scale fusion DCNN includes the following steps:

[0008] (1) Collect noise signals of the centrifugal pump under different cavitation states and divide them into several noise signal samples without overlap;

[0009] (2) Obtain the wavelet time-frequency plot of each noise signal sample, construct the time-frequency plot dataset and divide it into training set, validation set and test set;

[0010] (3) Construct a time-frequency enhanced wavelet fusion multi-scale interpretable DCNN model. The model includes a time-frequency attention module, a fusion multi-scale module, and a fusion class activation mapping module. The time-frequency attention module is used to adaptively enhance the time-frequency positions related to cavitation, the fusion multi-scale module is used to enhance the feature extraction and feature fusion capabilities of the model, and the fusion class activation mapping module is used to reveal the cavitation sensitive frequency band. The former two will jointly improve the cavitation recognition accuracy, and the latter will enhance the interpretability of the model.

[0011] (4) Train the constructed model based on the training set and save the optimal parameters of the model based on the validation set;

[0012] (5) Test the model based on the test set and output the predicted cavitation state and class activation map for each sample.

[0013] In step (2), the wavelet time-frequency plot of each noise signal sample is obtained. The specific process is as follows:

[0014] (2-1) Calculate the wavelet coefficients cwt(s,τ) of each noise signal sample x(t) based on continuous wavelet transform:

[0015]

[0016] In the formula, s is the scaling parameter; τ is the translation parameter; ψ(t) is the wavelet basis function; * denotes complex conjugation;

[0017] (2-2) Construct a two-dimensional time-frequency graph based on wavelet coefficients, with the horizontal axis representing time, the vertical axis representing frequency, and the color representing the magnitude of the wavelet coefficients.

[0018] In step (2), the time-frequency graph dataset is randomly divided into training set, validation set and test set in a ratio of 7:2:1.

[0019] In step (3), the working process of the time-frequency attention module is as follows:

[0020] (3-1-1) Average pooling is used to compress the input feature map along the time and frequency directions respectively, generating frequency-aware aggregated feature maps z. f and time-aware aggregated feature map z t , respectively The value at frequency f of the c-th channel is represented by... This represents the value at time t of the c-th channel:

[0021]

[0022] In the formula, x c (j,i) represents the two-dimensional feature map of the c-th channel; W represents the width of the input feature map; i is the index in the width direction; H represents the height of the input feature map; j is the index in the height direction;

[0023] (3-1-2) Aggregate frequency-aware feature maps z f and time-aware aggregated feature map z t The data is then concatenated, followed by 1x1 convolution to reduce the number of channels, and an activation function is used to add non-linearity, resulting in the intermediate feature map IF.

[0024] IF=δ(1*1Conv([z f ,z t ]))

[0025] In the formula, [·,·] represents the splicing operation along the spatial dimension; 1*1Conv(·) represents the 1*1 convolution operation; δ(·) represents the ReLU activation function;

[0026] Next, the intermediate feature map IF is split along the spatial dimension into two separate tensors, namely the frequency-aware tensor F. f and time-aware tensor F t Each tensor is processed by a 1x1 convolution to recover the number of channels, and then normalized using the Sigmoid function to obtain the frequency attention weights w. f And time attention weight w t :

[0027] w f =σ(1*1Conv(F) f ))

[0028] w t =σ(1*1Conv(F) t ))

[0029] In the formula, σ(·) represents the Sigmoid function;

[0030] (3-1-3) The frequency attention weight w f And time attention weight w t Applying this to the input feature map yields the output feature map, denoted by y. c (j,i) represents the two-dimensional output feature map corresponding to the c-th channel:

[0031]

[0032] In the formula, This represents the j-th value of the frequency attention weight corresponding to the c-th channel; This represents the i-th value of the temporal attention weight corresponding to the c-th channel.

[0033] In step (3), the multi-scale fusion module consists of a multi-scale convolution sub-module and a wavelet fusion sub-module;

[0034] The multi-scale convolutional submodule comprises three branches of different scales: branch 1 consists of two convolutional layers with a kernel size of 1*1, a ReLU activation function, and a max pooling layer with a kernel size of 2*2; branch 2 consists of two convolutional layers with a kernel size of 3*3, a ReLU activation function, and a max pooling layer with a kernel size of 2*2; and branch 3 consists of two convolutional layers with a kernel size of 5*5, a ReLU activation function, and a max pooling layer with a kernel size of 2*2.

[0035] The working process of the wavelet fusion submodule is as follows:

[0036] (3-2-1) Perform two-dimensional discrete wavelet transform on the output feature map of each scale branch in the multi-scale convolutional submodule to obtain four sub-matrices:

[0037] DWT(F s ) = LL s ,LH s HL s ,HH s ,s=1,2,3

[0038] In the formula, DWT(·) represents the two-dimensional discrete wavelet transform; F s This represents the output feature map of the s-th scale branch; LL s LH represents the horizontal and vertical low-frequency submatrix of the s-th scale branch. s HL represents the horizontal low-frequency and vertical high-frequency submatrix of the s-th scale branch; s HH represents the horizontal high-frequency and vertical low-frequency submatrix of the s-th scale branch; s This represents the horizontal and vertical high-frequency submatrix of the s-th scale branch;

[0039] (3-2-2) Feature fusion is performed on the sub-matrices of the same type in the three different scale branches. The horizontal low-frequency and vertical low-frequency sub-matrices use a maximum fusion strategy, meaning the fused eigenvalues ​​are the maximum values ​​in the corresponding sub-matrices of each branch. The other three sub-matrices use an average fusion strategy, meaning the fused eigenvalues ​​are the average values ​​in the corresponding sub-matrices of each branch. The calculation formula is as follows:

[0040] LL f =Max(LL1,LL2,LL3),

[0041] LH f =Mean(LH1,LH2,LH3),

[0042] HL f =Mean(HL1,HL2,HL3),

[0043] HH f =Mean(HH1,HH2,HH3)

[0044] In the formula, LL f ,LH f HL f and HH f These are the fused horizontal low-frequency and vertical low-frequency sub-matrices, horizontal low-frequency and vertical high-frequency sub-matrices, horizontal high-frequency and vertical low-frequency sub-matrices, and horizontal high-frequency and vertical high-frequency sub-matrices, respectively; Max(·) represents taking the maximum value at the corresponding position; Mean(·) represents taking the average value at the corresponding position;

[0045] (3-2-3) The four fused sub-matrices are synthesized using inverse discrete wavelet transform to obtain the output feature map I of wavelet fusion:

[0046] I = IDWT(LL) f ,LH f HL f ,HH f )

[0047] In the formula, IDWT(·) represents the inverse discrete wavelet transform.

[0048] In step (3), the operation of the fusion class activation mapping module is as follows:

[0049] (3-3-1) Obtain the output feature map A of the l-th layer through forward propagation. l And obtain the class score y of the network output for the target class c. c ;

[0050] (3-3-2) Obtain the category score y through backpropagation c Relative to feature map A l The gradient of the k-th channel is then used to perform global average pooling to obtain the weights for each channel.

[0051]

[0052] In the formula, Representing feature map A l The k-th channel, where i and j are indices in the height and width directions respectively, and H and W are its height and width respectively;

[0053] (3-3-3) Feature Map The class activation map (CAM) corresponding to the l-th layer is obtained by weighted summation of each channel and nonlinear activation operation. l :

[0054]

[0055] In the formula, C represents the output feature map A of the l-th layer. l The number of channels; δ(·) represents the ReLU activation function;

[0056] (3-3-4) Mutual information is used to measure the class activation map (CAM) at different layers. l The correlation between the grayscale image X and the original input image is used as a weight to apply to the class activation map CAM of different layers. l Weighted summation is performed, and the time dimension is compressed to obtain the weighted fusion class activation map (CAM). w :

[0057]

[0058] In the formula, MI(X; CAM) l ) represents CAM l The mutual information value between x and X; x i Indicates the value of X; cam lj CAM l The value of P(x) i ,cam lj ) represents the joint probability; P(x) i ) and P(cam lj ) represents the marginal probability; L represents the number of layers that need to be fused.

[0059] The specific process of step (4) is as follows:

[0060] (4-1) Divide the training set into several batches of data and input them into the model in sequence. Perform forward propagation to obtain the predicted class distribution. Use cross-entropy loss as the classification loss for backpropagation and gradient descent. Iteratively update the model parameters. After all batches of data have been input, it is considered that one round of training has been completed. Record the classification accuracy of the model in the validation set for this round.

[0061] (4-2) Shuffle the order of the images in the training set and then perform the next round of training and validation until the number of training rounds reaches the preset value. After all rounds of training are completed, save the model parameters corresponding to the round with the highest accuracy on the validation set.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] 1. This invention proposes a time-frequency attention module, which enhances the time-frequency positions related to cavitation recognition and reduces interference from irrelevant components by weighting the time and frequency directions of the time-frequency graph.

[0064] 2. This invention proposes a wavelet fusion multi-scale convolution module, which enhances the ability of cavitation feature extraction and feature fusion through multi-scale convolution and discrete wavelet transform.

[0065] 3. This invention proposes a fusion-type activation mapping module, which can highlight regions in the feature map that are sensitive to cavitation state, reveal cavitation feature frequency bands, and improve the interpretability of cavitation recognition. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the intelligent identification method for pump cavitation state based on multi-scale fusion DCNN of the present invention.

[0067] Figure 2 This is a schematic diagram of a time-frequency enhanced wavelet fusion multi-scale interpretable DCNN in an embodiment of the present invention.

[0068] Figure 3 This is a schematic diagram of the time-frequency attention module in an embodiment of the present invention.

[0069] Figure 4 This is a schematic diagram of the fusion of multi-scale modules in an embodiment of the present invention.

[0070] Figure 5 This is a schematic diagram of the fusion class activation mapping module in an embodiment of the present invention.

[0071] Figure 6 This is a schematic diagram of a centrifugal pump unit visualization test bench in an embodiment of the present invention.

[0072] Figure 7 This is a fusion class activation mapping diagram in an embodiment of the present invention. Detailed Implementation

[0073] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0074] like Figure 1 As shown, a method for intelligent identification of pump cavitation state based on multi-scale fusion DCNN includes the following steps:

[0075] S01 collects noise signals from the centrifugal pump when it is in different cavitation states and divides them into several signal samples without overlap.

[0076] S02, obtain the wavelet time-frequency plot of each noise signal sample, construct the time-frequency plot dataset and randomly divide it into training set, validation set and test set in a ratio of 7:2:1.

[0077] (2-1) Calculate the wavelet coefficients cwt(s,τ) of each noise signal sample x(t) based on continuous wavelet transform:

[0078]

[0079] In the formula, s is the scaling parameter; τ is the translation parameter; ψ(t) is the wavelet basis function; * denotes complex conjugation;

[0080] (2-2) Construct a two-dimensional time-frequency graph based on wavelet coefficients, with the horizontal axis representing time, the vertical axis representing frequency, and the color representing the magnitude of the wavelet coefficients.

[0081] S03, Construct a time-frequency enhanced wavelet fusion multi-scale interpretable DCNN model.

[0082] like Figure 2 As shown, the constructed time-frequency enhanced wavelet fusion multi-scale interpretable DCNN model includes three modules: a time-frequency attention module, a fusion multi-scale module, and a fusion class activation mapping module.

[0083] Time-frequency attention module, such as Figure 3 As shown, it specifically includes:

[0084] (3-1-1) Average pooling is used to compress the input feature map along the time and frequency directions respectively, generating frequency-aware aggregated feature maps z. f and time-aware aggregated feature map z t , respectively The value at frequency f of the c-th channel is represented by... This represents the value at time t of the c-th channel:

[0085]

[0086] In the formula, x c (j,i) represents the two-dimensional feature map of the c-th channel; W represents the width of the input feature map; i is the index in the width direction; H represents the height of the input feature map; j is the index in the height direction.

[0087] (3-1-2) The frequency-aware aggregated feature map and the time-aware aggregated feature map are concatenated, then dimensionality is reduced by channel number using 1*1 convolution, and an activation function is used to add non-linearity to obtain the intermediate feature map IF:

[0088] IF=δ(1*1Conv([z f ,z t ]))

[0089] In the formula, [·,·] represents the splicing operation along the spatial dimension; 1*1Conv(·) represents the 1*1 convolution operation; δ(·) represents the ReLU activation function.

[0090] Next, the intermediate feature map IF is split along the spatial dimension into two separate tensors, namely the frequency-aware tensor F. f and time-aware tensor F t Each tensor is processed by a 1x1 convolution to recover the number of channels, and then normalized using the Sigmoid function to obtain the frequency attention weights w. f And time attention weight w t :

[0091] w f =σ(1*1Conv(F) f ))

[0092] w t =σ(1*1Conv(F) t ))

[0093] In the formula, σ(·) represents the Sigmoid function.

[0094] (3-1-3) The frequency attention weight w f And time attention weight w t Applying this to the input feature map yields the output feature map, denoted by y. c (j,i) represents the two-dimensional output feature map corresponding to the c-th channel:

[0095]

[0096] In the formula, This represents the j-th value of the frequency attention weight corresponding to the c-th channel; This represents the i-th value of the temporal attention weight corresponding to the c-th channel.

[0097] The multi-scale fusion module consists of a multi-scale convolutional submodule and a wavelet fusion submodule, such as... Figure 4 As shown, the multi-scale convolutional submodule contains three branches of different scales: branch 1 consists of two convolutional layers with a kernel size of 1*1, a ReLU activation function, and a max pooling layer with a kernel size of 2*2; branch 2 consists of two convolutional layers with a kernel size of 3*3, a ReLU activation function, and a max pooling layer with a kernel size of 2*2; and branch 3 consists of two convolutional layers with a kernel size of 5*5, a ReLU activation function, and a max pooling layer with a kernel size of 2*2.

[0098] The wavelet fusion submodule includes three steps:

[0099] (3-2-1) Perform two-dimensional discrete wavelet transform on the output feature map of each scale branch in the multi-scale convolutional submodule to obtain four sub-matrices:

[0100] DWT(F s ) = LL s ,LH s HL s ,HH s ,s=1,2,3

[0101] In the formula, DWT(·) represents the two-dimensional discrete wavelet transform; F s This represents the output feature map of the s-th scale branch; LL s LH represents the horizontal and vertical low-frequency submatrix of the s-th scale branch. s HL represents the horizontal low-frequency and vertical high-frequency submatrix of the s-th scale branch; s HH represents the horizontal high-frequency and vertical low-frequency submatrix of the s-th scale branch; s This represents the horizontal and vertical high-frequency submatrix of the s-th scale branch.

[0102] (3-2-2) Feature fusion is performed on the sub-matrices of the same type in the three different scale branches. The horizontal low-frequency and vertical low-frequency sub-matrices use a maximum fusion strategy, meaning the fused eigenvalues ​​are the maximum values ​​in the corresponding sub-matrices of each branch. The other three sub-matrices use an average fusion strategy, meaning the fused eigenvalues ​​are the average values ​​in the corresponding sub-matrices of each branch. The calculation formula is as follows:

[0103] LL f =Max(LL1,LL2,LL3),

[0104] LH f =Mean(LH1,LH2,LH3),

[0105] HL f =Mean(HL1,HL2,HL3),

[0106] HH f =Mean(HH1,HH2,HH3)

[0107] In the formula, LL f ,LH f HL f and HH f These are the fused horizontal low-frequency and vertical low-frequency sub-matrices, horizontal low-frequency and vertical high-frequency sub-matrices, horizontal high-frequency and vertical low-frequency sub-matrices, and horizontal high-frequency and vertical high-frequency sub-matrices, respectively; Max(·) represents taking the maximum value at the corresponding position; Mean(·) represents taking the average value at the corresponding position.

[0108] (3-2-3) The four fused sub-matrices are synthesized using inverse discrete wavelet transform to obtain the output feature map I of wavelet fusion:

[0109] I = IDWT(LL) f ,LH f HL f ,HH f )

[0110] In the formula, IDWT(·) represents the inverse discrete wavelet transform.

[0111] Fusion-type activation mapping modules, such as Figure 5 As shown, it specifically includes:

[0112] (3-3-1) Obtain the output feature map A of the l-th layer through forward propagation. l And obtain the class score y of the network output for the target class c. c (logit value before softmax)

[0113] (3-3-2) Obtain the category score y through backpropagation c Relative to feature map A l The gradient of the k-th channel is then used to perform global average pooling to obtain the weights for each channel.

[0114]

[0115] In the formula, Representing feature map A l The k-th channel, where i and j are indices in the height and width directions respectively, and H and W are its height and width respectively.

[0116] (3-3-3) Feature Map The class activation map (CAM) corresponding to the l-th layer is obtained by weighted summation of each channel and nonlinear activation operation. l :

[0117]

[0118] In the formula, C represents the output feature map A of the l-th layer. l The number of channels; δ(·) represents the ReLU activation function.

[0119] (3-3-4) Mutual information is used to measure the class activation map (CAM) at different layers. l The correlation between the grayscale image X and the original input image is used as a weight to apply to the class activation map CAM of different layers. lWeighted summation is performed, and the time dimension is compressed to obtain the weighted fusion class activation map (CAM). w :

[0120]

[0121] In the formula, MI(X; CAM) l ) represents CAM l The mutual information value between x and X; x i Indicates the value of X; cam lj CAM l The value of P(x) i ,cam lj ) represents the joint probability; P(x) i ) and P(cam lj ) represents the marginal probability; L represents the number of layers that need to be fused.

[0122] S04: Train the model based on the training set and save the model's optimal parameters based on the validation set.

[0123] (4-1) Divide the training set into several batches of data and input them into the model in sequence. Perform forward propagation to obtain the predicted class distribution. Use cross-entropy loss as the classification loss for backpropagation and gradient descent. Iteratively update the model parameters. After all batches of data have been input, it is considered that one round of training has been completed. Record the classification accuracy of the model in the validation set for this round.

[0124] (4-2) Shuffle the order of the images in the training set and then perform the next round of training and validation until the number of training rounds reaches the preset value. After all rounds of training are completed, save the model parameters corresponding to the round with the highest accuracy on the validation set.

[0125] S05, test the model based on the test set, output the predicted cavitation state and class activation map for each sample, and calculate the classification accuracy.

[0126] (5-1) Input the samples in the test set into the model in sequence for identification, and output the corresponding predicted cavitation state and class activation map.

[0127] (5-2) Calculate the classification accuracy and confusion matrix based on the actual cavitation state and predicted cavitation state of the test set samples.

[0128] To verify the effectiveness of this invention, the cavitation state identification effect of the centrifugal pump was analyzed, as follows:

[0129] Cavitation simulation experiments and noise signal acquisition were conducted on a centrifugal pump unit visualization test bench. A schematic diagram of the test bench is shown below. Figure 6As shown in Table 1, different cavitation states were simulated by reducing the inlet pressure of the centrifugal pump while keeping the flow rate constant. Based on the head descent rate and flow field images from a high-speed camera, the cavitation states were classified into four categories: no cavitation (NC), initial cavitation (IC), significant cavitation (OC), and severe cavitation (SC). Data acquisition from the test bench was conducted under six different operating conditions (speed, flow rate), as shown in Table 1. The sampling frequency of the noise signal was 40kHz, and the duration of each sample after slicing was 0.2s, resulting in 600 samples for each cavitation state under each operating condition.

[0130] Table 1

[0131] Operating condition number Rotational speed (rpm) <![CDATA[Flow rate (m 3 / h)]]> Total number of samples A 923 6 4*600 B 923 5 4*600 C 1029 6 4*600 D 1029 5 4*600 E 1131 6 4*600 F 1131 5 4*600

[0132] To fully demonstrate the superiority of this invention, four existing classic methods (Modified-LeNet, AlexNet, InceptionV1, and InceptionV4) and three advanced methods (InceptionV4-CBAM, OWPLFDM, and MAFResCNN) are compared with the method of this invention (MSF-DCNN) in comparative experiments. Note: "Modified-LeNet" represents an improved LeNet model; "AlexNet" is a classic convolutional neural network; "InceptionV1" and "InceptionV4" represent different versions of the Inception model; "InceptionV4-CBAM" represents InceptionV4 with the introduction of the CBAM hybrid attention mechanism; "OWPLFDM" is a residual convolutional network with attribute and spatial attention mechanisms; "MAFResCNN" is a residual convolutional network with multi-attention fusion.

[0133] Table 2 shows the cavitation recognition accuracy (%) of different methods under different operating conditions. The method of this invention (MSF-DCNN) achieved the highest classification accuracy in all operating conditions, demonstrating the effectiveness and superiority of the time-frequency attention module and the wavelet fusion multi-scale convolution module in the method of this invention.

[0134] Table 2

[0135]

[0136]

[0137] To further demonstrate the effectiveness of the fusion-type activation mapping module proposed in this invention, wavelet time-frequency diagrams of the four cavitation states in operating condition A are analyzed, such as... Figure 7As shown, a1-a4 represent the wavelet time-frequency plots of NC, IC, OC, and SC, respectively; b1-b4 represent the fused activation mapping plots of NC, IC, OC, and SC, respectively; and c1-c4 represent the wavelet time-frequency plots after the superposition of the activation mapping plots of NC, IC, OC, and SC, respectively. The horizontal axis of each subplot represents time, and the vertical axis represents frequency. From figures a1-a4, it can be observed that the amplitude of the 1kHz-3kHz frequency band gradually increases with the deepening of cavitation. Figures b1-b4 similarly reveal the model's focus on the 1kHz-3kHz frequency band, thus increasing the interpretability of the DCNN model and proving the effectiveness of the proposed fused activation mapping module.

[0138] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-scale fusion DCNN-based intelligent identification method for pump cavitation state, characterized in that, Includes the following steps: (1) Collect noise signals of the centrifugal pump under different cavitation states and divide them into several noise signal samples without overlap; (2) Obtain the wavelet time-frequency plot of each noise signal sample, construct the time-frequency plot dataset and divide it into training set, validation set and test set; (3) Construct a time-frequency enhanced wavelet fusion multi-scale interpretable DCNN model. The model includes a time-frequency attention module, a fusion multi-scale module and a fusion class activation mapping module. The time-frequency attention module is used to adaptively enhance the time-frequency positions related to cavitation, the fusion multi-scale module is used to enhance the feature extraction and feature fusion capabilities of the model, and the fusion class activation mapping module is used to reveal the cavitation sensitive frequency band. (4) Train the constructed model based on the training set and save the optimal parameters of the model based on the validation set; (5) Test the model based on the test set and output the predicted cavitation state and class activation map for each sample.

2. The multi-scale fusion DCNN-based intelligent identification method for pump cavitation state according to claim 1, characterized in that, In step (2), the wavelet time-frequency plot of each noise signal sample is obtained. The specific process is as follows: (2-1) Calculate the wavelet coefficients cwt(s,τ) of each noise signal sample x(t) based on continuous wavelet transform: In the formula, s is the scaling parameter; τ is the translation parameter; ψ(t) is the wavelet basis function; * denotes complex conjugation; (2-2) Construct a two-dimensional time-frequency graph based on wavelet coefficients, with the horizontal axis representing time, the vertical axis representing frequency, and the color representing the magnitude of the wavelet coefficients.

3. The multi-scale fusion DCNN-based intelligent identification method for pump cavitation state according to claim 1, characterized in that, In step (2), the time-frequency graph dataset is randomly divided into training set, validation set and test set in a ratio of 7:2:

1.

4. The multi-scale fusion DCNN-based intelligent identification method for pump cavitation state according to claim 1, characterized in that, In step (3), the working process of the time-frequency attention module is as follows: (3-1-1) The input feature map is compressed along the time direction and the frequency direction respectively by using average pooling to generate frequency-aware aggregated feature map z f and time-aware aggregated feature map z t , respectively, where represents the value at frequency f of the cth channel, and represents the value at time t of the cth channel. wherein x c (j,i) denotes the two-dimensional feature map of the cth channel; W denotes the width of the input feature map; i is the index in the width direction; H denotes the height of the input feature map; j is the index in the height direction; (3-1-2) Concatenate the frequency-aware aggregated feature maps z f and the time-aware aggregated feature maps z t and reduce the channel number by 1*1 convolution, and increase the nonlinearity by using the activation function to obtain the intermediate feature map IF: IF = delta(1 * 1Conv([z f ,z t ])) In the formula, [·,·] represents the splicing operation along the spatial dimension; 1*1Conv(·) represents the 1*1 convolution operation; δ(·) represents the ReLU activation function; Then, the intermediate feature map IF is split along the spatial dimension into two separate tensors, i.e. a frequency-aware tensor F f and a time-aware tensor F t Each tensor recovers the number of channels by a 1*1 convolution and is normalized with a Sigmoid function, resulting in a frequency attention weight w f and a time attention weight w t respectively: w f = σ(1 * 1Conv(F f )) w t = σ(1 * 1Conv(F t )) In the formula, σ(·) represents the Sigmoid function; (3-1-3) applying the frequency attention weight w f and the time attention weight w t to the input feature map to obtain an output feature map, denoted as y c (j,i) represents a two-dimensional output feature map corresponding to the c-th channel: In the formula, This represents the j-th value of the frequency attention weight corresponding to the c-th channel; This represents the i-th value of the temporal attention weight corresponding to the c-th channel.

5. The intelligent identification method for pump cavitation state based on multi-scale fusion DCNN according to claim 1, characterized in that, In step (3), the multi-scale fusion module consists of a multi-scale convolution sub-module and a wavelet fusion sub-module; The multi-scale convolutional submodule comprises three branches of different scales: branch 1 consists of two convolutional layers with a kernel size of 1*1, a ReLU activation function, and a max pooling layer with a kernel size of 2*2; branch 2 consists of two convolutional layers with a kernel size of 3*3, a ReLU activation function, and a max pooling layer with a kernel size of 2*2; and branch 3 consists of two convolutional layers with a kernel size of 5*5, a ReLU activation function, and a max pooling layer with a kernel size of 2*2.

6. The intelligent identification method for pump cavitation state based on multi-scale fusion DCNN according to claim 5, characterized in that, The working process of the wavelet fusion submodule is as follows: (3-2-1) Perform two-dimensional discrete wavelet transform on the output feature map of each scale branch in the multi-scale convolutional submodule to obtain four sub-matrices: DWT(F s ) = LL s , LH s , HL s , HH s s = 1, 2, 3 where DWT(·) denotes a two-dimensional discrete wavelet transform; F s denotes the output feature map of the s-th scale branch; LL s denotes the horizontal low-frequency vertical low-frequency sub-matrix of the s-th scale branch; LH s denotes the horizontal low-frequency vertical high-frequency sub-matrix of the s-th scale branch; HL s denotes the horizontal high-frequency vertical low-frequency sub-matrix of the s-th scale branch; HH s denotes the horizontal high-frequency vertical high-frequency sub-matrix of the s-th scale branch; (3-2-2) Feature fusion is performed on the sub-matrices of the same type in the three different scale branches. The horizontal low-frequency and vertical low-frequency sub-matrices use a maximum fusion strategy, meaning the fused eigenvalues ​​are the maximum values ​​in the corresponding sub-matrices of each branch. The other three sub-matrices use an average fusion strategy, meaning the fused eigenvalues ​​are the average values ​​in the corresponding sub-matrices of each branch. The calculation formula is as follows: LL f = Max(LL1, LL2, LL3), LH f = Mean(LH1, LH2, LH3), HL f = Mean(HL1, HL2, HL3), HH f = Mean(HH1, HH2, HH3) wherein LL, LH, HL and HH are the fused horizontal low-frequency vertical low-frequency sub-matrix, the horizontal low-frequency vertical high-frequency sub-matrix, the horizontal high-frequency vertical low-frequency sub-matrix and the horizontal high-frequency vertical high-frequency sub-matrix, respectively; Max(·) represents taking the maximum value at the corresponding position; Mean(·) represents taking the average value at the corresponding position; and f f f f wherein LL, LH, HL and HH are the fused horizontal low-frequency vertical low-frequency sub-matrix, the horizontal low-frequency vertical high-frequency sub-matrix, the horizontal high-frequency vertical low-frequency sub-matrix and the horizontal high-frequency vertical high-frequency sub-matrix, respectively; Max(·) represents taking the maximum value at the corresponding position; Mean(·) represents taking the average value at the corresponding position; and​​​ (3-2-3) The four fused sub-matrices are synthesized using inverse discrete wavelet transform to obtain the output feature map I of wavelet fusion: I = IDWT(LL f , LH f , HL f , HH f ) In the formula, IDWT(·) represents the inverse discrete wavelet transform.

7. The intelligent identification method for pump cavitation state based on multi-scale fusion DCNN according to claim 1, characterized in that, In step (3), the operation of the fusion class activation mapping module is as follows: (3-3-1) Obtain the output feature map A of the lth layer by forward propagation l and obtain the class score y for the target class c of the network output c ; (3-3-2) Obtain the category score y through backpropagation c Relative to feature map A l The gradient of the k-th channel is then used to perform global average pooling to obtain the weights for each channel. In the formula, Representing feature map A l The k-th channel, where i and j are indices in the height and width directions respectively, and H and W are its height and width respectively; (3-3-3) Feature Map The class activation map (CAM) corresponding to the l-th layer is obtained by weighted summation of each channel and nonlinear activation operation. l : In the formula, C represents the output feature map A of the first layer l ; δ(·) represents a ReLu activation function. (3-3-4) Using mutual information to measure the class activation map CAM of different layers l correlation with the grayscale image X of the original input image, which is used as a weight to the class activation map CAM of different layers l weighted sum and compression of the time dimension, and finally obtain the weighted fused class activation map CAM w : where MI(x; CAM l ) denotes the mutual information value between CAM l and X; x i denotes the value of X; cam lj denotes the value of CAM l ; P(x i , cam lj ) denotes the joint probability; P(x i ) and P(cam lj ) denote the marginal probability; and L denotes the number of layers that need to be fused.

8. The intelligent identification method for pump cavitation state based on multi-scale fusion DCNN according to claim 1, characterized in that, The specific process of step (4) is as follows: (4-1) Divide the training set into several batches of data and input them into the model in sequence. Perform forward propagation to obtain the predicted class distribution. Use cross-entropy loss as the classification loss for backpropagation and gradient descent. Iteratively update the model parameters. After all batches of data have been input, it is considered that one round of training has been completed. Record the classification accuracy of the model in the validation set for this round. (4-2) Shuffle the order of the images in the training set and then perform the next round of training and validation until the number of training rounds reaches the preset value. After all rounds of training are completed, save the model parameters corresponding to the round with the highest accuracy on the validation set.

Images