Bearing fault diagnosis method based on multi-scale progressive feature fusion

Abstract

The invention belongs to the field of signal processing technology and aero-engine bearing fault diagnosis, and discloses a bearing fault diagnosis method based on multi-scale progressive feature fusion, which comprises the following steps of: decomposing an original acceleration signal into feature components with different scales by using wavelet transform; components with high correlation with the original signals are screened out through correlation coefficients for reconstruction, and noise interference in the original signals is effectively reduced. And meanwhile, a time-frequency graph is constructed by using the denoised signal, feature extraction is performed by dividing the time-frequency graph into RGB channels, and finally, features of different channels are enhanced and fused through a multi-scale progressive feature fusion module MSIFFM, so that the problem of feature interference during fusion of features of different scales can be effectively solved. Compared with an existing method, the bearing fault diagnosis method based on multi-scale progressive feature fusion can still obtain high accuracy in a strong noise environment.

Description

Technical Field

[0001] The present invention belongs to the fields of signal processing technology and aero-engine bearing fault diagnosis, and specifically relates to a bearing fault diagnosis method based on multi-scale progressive feature fusion. Background Art

[0002] Aero-engine bearings usually work in a strong noise environment. Facing samples containing strong noise, traditional spectrum analysis is difficult to accurately distinguish fault signals, affecting the diagnostic accuracy. Wavelet packet analysis is an extension and expansion of wavelet analysis. It can not only analyze low-frequency components, but also further subdivide high-frequency components, ensuring the accuracy and comprehensiveness of the decomposition process, and then more accurately extracting the key features of the signal. Then, according to the distribution characteristics of noise and signals in different sub-bands, an appropriate threshold is selected to suppress the noise, and finally the denoised signal is obtained through reconstruction. This method can effectively improve the signal-to-noise ratio of the signal and provide high-quality signals for subsequent feature extraction and fault diagnosis.

[0003] Bearing fault signals usually have multi-scale characteristics, that is, signals with different frequency components may contain different fault information. The current mainstream feature extraction method is a multi-scale processing method, and the existing method for fusing the processing results only sums up the features of different channels after strengthening, and cannot effectively distinguish similar and confusing features in different channels. Therefore, the effective matching of different-channel and different-scale features when fusing features is a difficult problem. Summary of the Invention

[0004] Aiming at the deficiencies in the prior art, the purpose of the present invention is to provide a bearing fault diagnosis method based on multi-scale progressive feature fusion. Among them, wavelet transform decomposes the original acceleration signal into different-scale feature components, and the components with a relatively high correlation with the original signal are selected through correlation coefficient screening for reconstruction. In the feature extraction stage, features are extracted through different RGB channels respectively. In the feature fusion stage, due to the multi-level nature of fault features, different features need to be extracted by convolution kernels of different scales. The proposed Multi-Scale Incremental Feature Fusion Module (MSIFFM) can highlight and fuse the key features of different channels, provide richer and more accurate information for the diagnosis task, and thus improve the accuracy and efficiency of classification.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A bearing fault diagnosis method based on multi-scale progressive feature fusion, the steps are as follows:

[0007] S1: Measure the dataset using an aeroengine bearing fault test bench. Use the acceleration signals in the dataset as diagnostic data, and segment the acceleration signals of different fault types into data segments of the same length as the original input signal X.

[0008] S2: Use the Wavelet Packet Transform (WPT) to adaptively decompose the original input signal X into different frequency subbands. Select the three components with the highest correlation coefficient with X through the Pearson correlation coefficient for reconstruction, remove noise interference, and generate a time-frequency feature map.

[0009] S3: Divide the time-frequency feature map of the signal after noise reduction in S2 into a training set and a test set according to a certain proportion.

[0010] S4: Build a fault diagnosis model.

[0011] S5: Calculate the error between the output label of the fault diagnosis model and the actual label, calculate the value of the loss function Loss, and repeat the training process of the fault diagnosis model in S4. Among them, the loss function selects the Cross-Entropy Loss, and the parameter optimizer uses the Adam Optimizer.

[0012] S6: Use the trained model to test the test set in S3, obtain the fault type classification result of the test set, and compare the result with the actual fault type to realize the diagnostic classification of the fault type.

[0013] The specific steps of step S2 are as follows:

[0014] S2-1: Perform WPT on X and decompose it into eight components in three layers to realize the decomposition of the original input signal X into different frequency subbands. The decomposition form is as follows: Among them,

[0015]

[0016] In the formula, represents the wavelet packet coefficient of the j-th layer and the n-th frequency band; h and g represent the low-pass and high-pass filter coefficients respectively; i and l represent discrete time indices.

[0017] S2-2: By setting other components to zero, calculate the reconstructed signal of each component respectively. The reconstruction formula is as follows:

[0018]

[0019] And calculate the Pearson correlation coefficient between the reconstructed signal of each component and the original signal respectively. The calculation method is as follows:

[0020]

[0021] In the formula, represents the signal reconstructed from the nth component of the third layer of wavelet packet decomposition, cov(*,*) represents covariance, and σ * represents the variance value, and E(*) represents taking the expectation.

[0022] Three components with a relatively high degree of correlation with the original input signal are selected for reconstruction. At the same time, a time-frequency feature map is generated based on the reconstructed signal, and the time-frequency feature map contains the time-domain information and frequency-domain information of the signal.

[0023] The specific implementation process of the fault diagnosis model in step S4 is as follows:

[0024] S4.1: The input time-frequency feature image is first subjected to normalization and size adjustment processing. The processed image is divided into RGB three primary color channels and subjected to convolution and pooling operations respectively for preliminary feature extraction to generate a feature map;

[0025] Furthermore, the normalization process is as follows:

[0026] The time-frequency feature map is normalized to ensure the consistency of the input of the fault diagnosis model. The normalization method is:

[0027]

[0028] where m and m′ are the time-frequency image pixel matrices before and after normalization respectively, and min(*) and max(*) represent taking the minimum value and the maximum value respectively;

[0029] Furthermore, the preliminary feature extraction is specifically: Select two-dimensional convolution (2D-CDD) calculation and max-pooling operation:

[0030]

[0031] where C(i,j) represents the value at the position (i,j) of the output feature map, K(m,n) represents the convolution kernel with a size of k×k, where m,n∈[0,k], and I(i+m,j+n) represents the value at the position (i+m,j+n) of the input image;

[0032]

[0033] where O(i,j) represents the value at the position (i,j) of the output feature map, D represents the pooling window size, and S represents the window movement step size.

[0034] S4.2: Input the feature map obtained in step S4.1 into the feature fusion module, which includes a convolutional layer and a Sigmoid layer. The implementation process of the feature fusion module is as follows: First, perform the first feature weighted fusion: The feature map is convolved using different-scale convolutional kernels in three channels. The three different-scale features are passed through the Sigmoid layer to obtain the weight W1, which serves as the feature weight of the R image, and 1 - W1 serves as the weight of the G image. The two images are subjected to the first feature weighted fusion. The feature map after the first feature weighted fusion is passed through the feature fusion module again to obtain the fusion weight W2 of the two images after weighted fusion, and 1 - W2 serves as the weight of the B image. The features of the three images are subjected to the second feature weighted fusion to obtain the final feature.

[0035] Further, the specific steps of the first feature weighted fusion are as follows:

[0036] Use three different-scale convolutional kernels to decompose and analyze features at different scales. At the same time, the output of the large-scale convolution is not only used as the final feature weight but also input into the subsequent small-scale convolution for further feature extraction. The steps of the first feature weighted fusion are as follows:

[0037] W = S([η3(η3(X)+η5(η5(X)+η7(X)))]*X)

[0038] where W is the fusion feature weight, S(*) represents the Sigmoid operation, and η i (*) represents the CNN with a convolutional kernel size of i×i and the Sigmoid operation.

[0039] Further, use the weighted fusion feature image obtained in the first feature fusion and the image to be fused as the input for the next feature weighted fusion, repeat the process of S4.2, and the final fusion result is as follows:

[0040]

[0041] where, represents the feature map after the final feature weighted fusion, R, G, and B respectively represent the three primary color images, and W1 and W2 respectively represent the weights after the first and second feature weighted fusions.

[0042] S4.3: Pass the time-frequency feature map input in S2 through the channel attention mechanism to obtain the weights of different channels, and finally complete the output of the fault type through the fully connected layer and the Softmax layer.

[0043] Further, the steps of the channel attention mechanism are as follows:

[0044] The time-frequency feature map is passed through the channel attention mechanism SE Block, and weights are assigned to the feature map after feature fusion in step S4.2. The SE Block includes two steps, Squeeze and Excitation. Among them, the Squeeze step is specifically as follows: The size of the original feature map is H×W×C, where H is the height, W is the width, and C is the number of channels. The purpose of Squeeze is to obtain global information embedding (feature vectors) for each channel of the feature map by calculating the average value within each channel. The global average pooling (GAP) operation is expressed by the following formula:

[0045]

[0046] The Excitation step is specifically as follows: Two fully connected layers (FC Layer) are used to calculate the feature weights of each channel. The ReLU activation function is used in the middle to control the complexity of the fault diagnosis model, and finally the Sigmoid activation function is used. The Excitation step is expressed by the following formula:

[0047] s = σ(w2δ(w1z))

[0048] Among them, σ(*) represents the Sigmoid operation, δ(*) represents the ReLU activation function, w1 and w2 represent the weights obtained through the first fully connected layer and the second fully connected layer respectively, and the final dimension of s is 1×1×C. The output of the Excitation step scales the channels of the feature map obtained in step S4.2 through multiplication.

[0049] Compared with the prior art, the present invention has at least the following beneficial effects:

[0050] The present invention proposes a multi-scale progressive fault feature fusion method. By decomposing and reconstructing the noise reduction signal through wavelet packet decomposition and constructing a time-frequency diagram, the key features in the fault signal can be retained, and this multi-dimensional information representation method helps the model to more comprehensively understand the signal characteristics; RGB three channels are used to initially extract features, and the interference problem in multi-scale feature fusion can be solved through the feature fusion module. Compared with other feature fusion methods, the multi-scale progressive fault feature fusion method proposed by the present invention has a large improvement in diagnostic accuracy, has certain application value, and provides a reference solution for the fault diagnosis of rolling bearings in the current big data environment.

[0051] The additional aspects and advantages of the present invention will be given in the following specific embodiments, and some will become obvious from the following description, or will be understood through the specific embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a schematic flow chart of the present invention.

[0053] Figure 2 It is a schematic diagram of the feature fusion module of the present invention.

[0054] Figure 3 It is a graph of the experimental results of a specific embodiment of the present invention. Specific Embodiments

[0055] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The embodiments are exemplary and are intended to explain the present invention, but should not be construed as limiting the present invention.

[0056] Example Experiment 1:

[0057] The experimental data of this example comes from the dataset measured by an aero-engine bearing fault test bench. The acceleration signal in the fifth column of the dataset is used as the diagnostic dataset. The 20,480 data points in each group are segmented into data segments with a length of 5,120 as the input signal, and a total of 9,648 data are obtained. The dataset is divided into a 70% training set and a 30% test set. The accuracy of the model on the test set is recorded under different training times, as Figure 3 shown. It can be seen from Figure 3 that the trained neural network can achieve a diagnostic accuracy of 99.75% on the test set and is more stable.

[0058] As Figure 1 , 2 shown, the present invention proposes a bearing fault diagnosis method, including the following steps:

[0059] S1: Use an aero-engine bearing fault test bench to measure the dataset, use the acceleration signal in the dataset as the diagnostic data, segment the acceleration signals of different fault types into data segments with the same length as the original input signal X, and set "0", "1", "2", "3" as the labels of "healthy bearing", "0.5 mm inner ring fracture of the bearing", "1.0 mm inner ring fracture of the bearing", and "outer ring fault of the bearing" respectively;

[0060] S2: Use the wavelet packet transform (WPT) to adaptively decompose the original input signal X into different frequency sub-bands, select the three components with the highest correlation coefficient with X through the Pearson correlation coefficient for reconstruction, remove noise interference, and generate a time-frequency feature map;

[0061] S3: Divide the time-frequency feature map of the signal after noise reduction in S2 into a training set and a test set according to a ratio of 7:3;

[0062] S4: Construct a fault diagnosis model;

[0063] S5: Calculate the error between the output label of the fault diagnosis model and the actual label, calculate the value of the loss function Loss, and repeat the training process of the fault diagnosis model in S4. During training, the batch-size (the number of data passed to the network for training at one time) is set to 16, the cross-entropy loss function is selected as the loss function, the Adam Optimizer is used as the parameter optimizer, the learning rate is set to 0.0001, and the number of iterations is 100;

[0064] S6: Use the trained model to test the test set in S3, obtain the fault type classification result of the test set, and compare the result with the actual fault type to achieve the diagnostic classification of the fault type.

[0065] The specific steps of the above-mentioned step S2 are as follows:

[0066] S2-1: Perform WPT on X, decompose it into three layers to obtain eight components, and realize the decomposition of the original input signal X into different frequency sub-bands. The decomposition form is as follows:

[0067]

[0068] In the formula, represents the wavelet packet coefficient of the j-th layer and the n-th frequency band; h and g respectively represent the low-pass and high-pass filter coefficients; i and l represent the discrete time indices.

[0069] S2-2: By setting other components to zero, calculate the reconstructed signal of each component respectively. The reconstruction formula is as follows:

[0070]

[0071] And calculate the Pearson correlation coefficient between the reconstructed signal of each component and the original signal respectively. The calculation method is as follows:

[0072]

[0073] In the formula, represents the signal after reconstruction of the n-th component in the third layer of wavelet packet decomposition, cov(*,*) represents the covariance, σ * represents the variance value, and E(*) represents taking the expectation.

[0074] Select three components with a relatively high degree of correlation with the original input signal for reconstruction. At the same time, generate a time-frequency feature map based on the reconstructed signal. The time-frequency feature map contains the time-domain information and frequency-domain information of the signal.

[0075] The specific implementation process of the fault diagnosis model in the above-mentioned step S4 is as follows:

[0076] S4.1: The input time-frequency feature image is first normalized and resized to 128×128. The processed image is divided into three RGB primary color channels and subjected to convolution and pooling operations respectively for preliminary feature extraction to generate feature maps.

[0077] Further, the normalization process is as follows:

[0078] The time-frequency feature map is normalized to ensure the consistency of the input to the fault diagnosis model. The normalization method is:

[0079]

[0080] where m and m′ are the time-frequency image pixel matrices before and after normalization respectively, and min(*) and max(*) represent taking the minimum value and the maximum value respectively.

[0081] Further, the preliminary feature extraction is specifically: Select two-dimensional convolution (2D-CDD) calculation and max-pooling operations:

[0082]

[0083] where C(i,j) represents the value at the position (i,j) of the output feature map, K(m,n) represents the convolution kernel with a size of k×k. Here, the sizes in the two convolutions are 5×5 and 3×3 respectively, and the convolution kernel moving step is 1. Where m,n∈[0,k], and I(i+m,j+n) represents the value at the position (i+m,j+n) of the input image.

[0084]

[0085] where O(i,j) represents the value at the position (i,j) of the output feature map, F represents the pooling window size, which is set to 2 here, and S represents the window moving step, which is set to 2 here.

[0086] S4.2: Input the feature map obtained in step S4.1 into the feature fusion module. The feature fusion module includes a convolution layer and a Sigmoid layer. The implementation process of the feature fusion module is as follows: First, perform the first feature weighted fusion: The feature map is convolved in three channels using different scale convolution kernels. The three different scale features pass through the Sigmoid layer to obtain the weight W1, which is used as the feature weight of the R image, and 1 - W1 is used as the weight of the G image. The two images are subjected to the first feature weighted fusion. The feature map after the first feature weighted fusion passes through the feature fusion module again to obtain the fusion weight W2 of the two images after weighted fusion, and 1 - W2 is used as the weight of the B image. The features of the three images are subjected to the second feature weighted fusion to obtain the final features.

[0087] Further, the specific steps of the first feature weighted fusion are as follows:

[0088] Use three different scales of convolutional kernels to decompose and analyze features of different scales. At the same time, the output of the large-scale convolution is not only used as the final feature weight but also input into the subsequent small-scale convolution to further extract features. The steps of the first feature weighted fusion are as follows:

[0089] W = S([η3(η3(X)+η5(η5(X)+η7(X)))]*X)

[0090] where W is the fused feature weight, S(*) represents the Sigmoid operation, and η i (*) represents the CNN with a convolutional kernel size of i×i and the Sigmoid operation.

[0091] Further, take the feature image after weighted fusion obtained in the first feature fusion and the image to be fused as the input for the next feature weighted fusion, and repeat the process of S4.2. The final fusion result is as follows:

[0092]

[0093] where, represents the feature map after the final feature weighted fusion, R, G, and B respectively represent the three primary color images, and W1 and w2 respectively represent the weights after the first and second feature weighted fusions.

[0094] S4.3: Pass the time-frequency feature map input in S2 through the channel attention mechanism to obtain the weights of different channels, and finally complete the output of the fault type through the fully connected layer and the Softmax layer;

[0095] Further, the steps of the channel attention mechanism are as follows:

[0096] Pass the time-frequency feature map through the channel attention mechanism SE Block and assign the weights to the feature map after feature fusion in step S4.2. The SE Block includes two steps, Squeeze and Excitation. Among them, the Squeeze step is specifically as follows: The size of the original feature map is H×W×C, where H is the height, W is the width, and C is the number of channels. The purpose of Squeeze is to obtain the global information embedding (feature vector) for each channel of the feature map by calculating the average value within each channel. The global average pooling (GAP) operation is expressed by the formula as follows:

[0097]

[0098] The Excitation step is specifically as follows: Use two fully connected layers (FC Layer) to calculate the feature weights of each channel, use the ReLU activation function in the middle to control the complexity of the fault diagnosis model, and finally use the Sigmoid activation function. The Excitation step is expressed by the formula as follows:

[0099] s = σ(w2δ(w1z))

[0100] Among them, σ(*) represents the Sigmoid operation, δ(*) represents the ReLU activation function, w1 and w2 respectively represent the weights obtained through the first fully connected layer and the second fully connected layer, and the final dimension of s is 1×1×C. The output of the Excitation step scales the channels of the feature map obtained in step S4.2 through multiplication.

[0101] Comparing with the cross-fusion, attention feature fusion, and dynamic feature fusion algorithms, the diagnostic accuracies of each feature fusion algorithm are shown in Table 1.

[0102] Table 1

[0103]

[0104] It can be seen from the comparative experiments in Table 1 that the diagnostic accuracy of the proposed multi-scale progressive feature fusion method is significantly higher than that of other feature fusion algorithm models, indicating the superiority of the proposed multi-scale progressive feature fusion.

[0105] The embodiments of the present invention have been described above in conjunction with the accompanying drawings, but the present invention is not limited to the above specific embodiments. The above embodiments are merely illustrative and guiding, rather than restrictive. Those of ordinary skill in the art can also make various forms under the inspiration of this specification and without departing from the scope protected by the claims of the present invention, and these all belong to the scope of protection of the present invention.

Claims

1. A bearing fault diagnosis method based on multi-scale progressive feature fusion, characterized in that The steps are as follows: S1: Use an aeroengine bearing fault test bench to measure a dataset. Use the acceleration signals in the dataset as diagnostic data, and segment the acceleration signals of different fault types into data segments of the same length as the original input signal X; S2: Use wavelet packet transform to adaptively decompose the original input signal X into different frequency sub-bands. Select the three components with the highest correlation coefficient with X through Pearson correlation coefficient for reconstruction, remove noise interference, and generate a time-frequency feature map; S3: Divide the time-frequency feature map of the signal after noise reduction in S2 into a training set and a test set according to a ratio; S4: Build a fault diagnosis model; S5: Calculate the error between the output label and the actual label of the fault diagnosis model, calculate the Loss value of the loss function, and repeat the training process of the fault diagnosis model in S4. Among them, the loss function selects the cross-entropy loss function, and the parameter optimizer uses AdamOptimizer; S6: Use the trained model to test the test set in S3, obtain the fault type classification result of the test set, and compare the result with the actual fault type to realize the diagnosis and classification of the fault type.

2. The bearing fault diagnosis method based on multi-scale progressive feature fusion according to claim 1, wherein The specific steps of the step S2 are as follows: S2-1: Perform WPT on X, decompose it into three layers to obtain eight components, and achieve the decomposition of the original input signal X into different frequency sub-bands. The decomposition form is as follows: wherein, represents the wavelet packet coefficient of the j-th layer and the n-th frequency band; h and g respectively represent the low-pass and high-pass filter coefficients; i and l represent discrete-time indices; S2-2: By setting other components to zero, calculate the reconstruction signals of each component respectively. The reconstruction formula is as follows: And calculate the Pearson correlation coefficient between the reconstruction signal of each component and the original signal respectively. The calculation method is as follows: Wherein, represents the signal reconstructed from the n-th component of the third layer of wavelet packet decomposition, cov(*, *) represents covariance, and σ * represents the variance value, and E(*) represents taking the expectation; Select the three components with the highest correlation with the original input signal for reconstruction; at the same time, generate a time-frequency feature map according to the reconstructed signal. The time-frequency feature map contains the time-domain information and frequency-domain information of the signal.

3. A bearing fault diagnosis method based on multi-scale progressive feature fusion according to claim 1, characterized in that The specific implementation process of the fault diagnosis model in the step S4 is as follows: S4.1: The input time-frequency feature image is first normalized and resized. The processed image is divided into three RGB primary color channels for convolution and pooling operations respectively for preliminary feature extraction to generate a feature map; S4. ​ ​ 4. A bearing fault diagnosis method based on multi-scale progressive feature fusion according to claim 3, characterized in that ​ ​ Among them, m and m' are the time-frequency image pixel matrices before and after normalization respectively, and min(*) and max(*) represent taking the minimum value and the maximum value respectively.

5. A bearing fault diagnosis method based on multi-scale progressive feature fusion according to claim 3, characterized in that In the step S4.1, the preliminary feature extraction is specifically as follows: Select two-dimensional convolution calculation and max pooling operation: Among them, C(i,j) represents the value at the position (i,j) of the output feature map, K(m,n) represents the convolution kernel with a size of k×k, where m,n∈[0,k], and I(i+m,j+n) represents the value at the position (i+m,j+n) of the input image; Among them, O(i,j) represents the value at the position (i,j) of the output feature map, F represents the size of the pooling window, and S represents the step size of the window movement.

6. A bearing fault diagnosis method based on multi-scale progressive feature fusion according to claim 3, characterized in that, In the step S4.2, the specific steps of the first feature weighted fusion are as follows: Use three different scales of convolution kernels to decompose and analyze features of different scales. At the same time, the output of the large-scale convolution is not only used as the final feature weight but also input into the subsequent small-scale convolution for further feature extraction. The steps of the first feature weighted fusion are as follows: W = S([η3(η3(X)+η5(η5(X)+η7(X)))]*X) where W is the weight of the fused feature, S(*) represents the Sigmoid operation, and η i (*) represents a CNN with a convolutional kernel size of i×i and the Sigmoid operation.

7. A bearing fault diagnosis method based on multi-scale progressive feature fusion according to claim 3, characterized in that In the step S4.2, take the weighted fusion feature image obtained in the first feature fusion and the image to be fused as the input for the next feature weighted fusion, and repeat the process of S4.

2. The final fusion result is as follows: Among them, represents the feature map after the final feature weighted fusion. R, G, and B respectively represent the three primary color images, and W1 and W2 respectively represent the weights after the first and second feature weighted fusions.

8. A bearing fault diagnosis method based on multi-scale progressive feature fusion according to claim 3, characterized in that In the step S4.3, the steps of the channel attention mechanism are as follows: Pass the time-frequency feature map through the channel attention mechanism SE Block and assign weights to the feature map after the feature fusion in step S4.2; The SE Block includes two steps, Squeeze and Excitation; Among them, the Squeeze step is specifically as follows: The size of the original feature map is H×W×C, where H is the height, W is the width, and C is the number of channels. The purpose of Squeeze is to obtain global information embedding for each channel of the feature map by calculating the average value within each channel; The global average pooling operation is expressed by the formula as follows: The Excitation step is specifically as follows: Use two fully connected layers to calculate the feature weights of each channel, use the ReLU activation function in the middle to control the complexity of the fault diagnosis model, and finally use the Sigmoid activation function. The Excitation step is expressed by the formula as follows: s = σ(w2δ(w1z)) Among them, σ(*) represents the Sigmoid operation, δ(*) represents the ReLU activation function, w1 and w2 represent the weights obtained through the first fully connected layer and the second fully connected layer respectively, and the final dimension of s is 1×1×C; The output of the Excitation step scales the channels of the feature map obtained in step S4.2 through multiplication.

Images