Shoemaking machine and method for multi-angle waterproof monitoring in rain shoe production

Abstract

The present application discloses a shoe-making machine and a method for multi-angle waterproof monitoring in rain shoe production. First, it calculates the time-domain enhancement maps of the first sound signal of the rain shoe to be detected that is not soaked and the second sound signal of the rain shoe to be detected that is soaked and then dried to obtain a first time-domain enhancement map and a second time-domain enhancement map. Then, the first time-domain enhancement map and the second time-domain enhancement map are passed through a siamese network model to obtain a first feature map and a second feature map. Next, it calculates the differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map obtained by passing the first feature map and the second feature map through a multi-branch perception domain module respectively. Finally, the differential feature map is passed through a classifier to obtain a classification result for indicating whether the waterproof performance of the rain shoe to be detected meets a predetermined standard. In this way, the detection time can be shortened and the accuracy of rain shoe waterproof performance detection can be improved.

Description

Technical Field

[0001] The present application relates to the field of intelligent monitoring technology, and more specifically, to a shoe-making machine and method for multi-angle waterproof monitoring for rain boot production. Background Art

[0002] Rain boots are a kind of waterproof shoes, mostly worn when it rains. Some workers who work in places with a lot of water will also wear rain boots. The main function of rain boots is waterproof. Therefore, after the production of rain boots is completed, the rain boots need to be tested for waterproofness.

[0003] Currently, when testing the waterproofness of rain boots, most people manually hold the rain boots, place the rain boots in a water tank, and then observe the inside of the rain boots to see if there is water accumulation inside the rain boots. This waterproof testing mode is not conducive to observation and the test results are not intuitive.

[0004] After research, the inventors of this application found that when conducting factory waterproof tests on rain boots, if the quality of the rain boots is poor, it is often not detected during the factory waterproof test. The reason is that it takes a long time for the relatively inferior rain boots to penetrate into the rain boots, and it is difficult to observe in a short time.

[0005] Therefore, an optimized rain boots waterproof monitoring solution is expected. Summary of the Invention

[0006] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a shoe-making machine and method for multi-angle waterproof monitoring in the production of rain boots. It first calculates the time domain enhancement map of the first sound signal of the rain boots to be detected that have not been soaked and the second sound signal of the rain boots to be detected that have been soaked and wiped dry to obtain the first time domain enhancement map and the second time domain enhancement map, then, the first time domain enhancement map and the second time domain enhancement map are passed through the twin network model to obtain the first feature map and the second feature map, then, the differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map obtained by passing the first feature map and the second feature map through the multi-branch perception domain module is calculated, and finally, the differential feature map is passed through a classifier to obtain a classification result for indicating whether the waterproof performance of the rain boots to be detected meets the predetermined standard. In this way, the detection time can be shortened and the accuracy of the rain boots waterproof performance detection can be improved.

[0007] According to one aspect of the present application, a shoe-making machine with multi-angle waterproof monitoring for rain boot production is provided, comprising:

[0008] A detection signal acquisition unit is used to obtain a first sound signal of the rain boots to be detected that have not been soaked and a second sound signal of the rain boots to be detected that have been soaked and wiped dry;

[0009] A time-domain conversion unit for calculating time-domain enhancement maps of the first sound signal and the second sound signal to obtain a first time-domain enhancement map and a second time-domain enhancement map;

[0010] A siamese encoding unit for passing the first time-domain enhancement map and the second time-domain enhancement map through a siamese network model including a first convolutional neural network and a second convolutional neural network to obtain a first feature map and a second feature map, where the first convolutional neural network and the second convolutional neural network have the same network structure;

[0011] A multi-scale perception unit for passing the first feature map and the second feature map through a multi-branch perception domain module respectively to obtain a first multi-scale perception feature map and a second multi-scale perception feature map;

[0012] A difference evaluation unit for calculating a differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map; and

[0013] A waterproof monitoring result generation unit for passing the differential feature map through a classifier to obtain a classification result, where the classification result is used to indicate whether the waterproof performance of the rain shoes to be detected meets a predetermined standard.

[0014] According to another aspect of the present application, there is provided a multi-angle waterproof monitoring method for rain shoe production, including:

[0015] Obtaining a first sound signal of the rain shoes to be detected that are not soaked and a second sound signal of the rain shoes to be detected that are soaked and then dried;

[0016] Calculating time-domain enhancement maps of the first sound signal and the second sound signal to obtain a first time-domain enhancement map and a second time-domain enhancement map;

[0017] Passing the first time-domain enhancement map and the second time-domain enhancement map through a siamese network model including a first convolutional neural network and a second convolutional neural network to obtain a first feature map and a second feature map, where the first convolutional neural network and the second convolutional neural network have the same network structure;

[0018] Passing the first feature map and the second feature map through a multi-branch perception domain module respectively to obtain a first multi-scale perception feature map and a second multi-scale perception feature map;

[0019] Calculating a differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map; and

[0020] Passing the differential feature map through a classifier to obtain a classification result, where the classification result is used to indicate whether the waterproof performance of the rain shoes to be detected meets a predetermined standard.

[0021] Compared with the prior art, the present application provides a shoe-making machine and a method for multi-angle waterproof monitoring in rain shoe production. First, a time-domain enhancement map of a first sound signal of a to-be-detected rain shoe that is not soaked and a second sound signal of the to-be-detected rain shoe that is soaked and then dried is calculated to obtain a first time-domain enhancement map and a second time-domain enhancement map. Then, the first time-domain enhancement map and the second time-domain enhancement map are passed through a siamese network model to obtain a first feature map and a second feature map. Next, a differential feature map is calculated between a first multi-scale perception feature map and a second multi-scale perception feature map obtained by passing the first feature map and the second feature map through a multi-branch perception domain module respectively. Finally, the differential feature map is passed through a classifier to obtain a classification result for indicating whether the waterproof performance of the to-be-detected rain shoe meets a predetermined standard. In this way, the detection time can be shortened and the accuracy of rain shoe waterproof performance detection can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] By describing the embodiments of the present application in more detail in conjunction with the drawings, the above and other objects, features, and advantages of the present application will become more obvious. The drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification, and are used to explain the present application together with the embodiments of the present application, and do not constitute a limitation to the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0023] Figure 1 The application scenario diagram of the shoe-making machine for multi-angle waterproof monitoring in rain shoe production according to the embodiment of the present application is illustrated.

[0024] Figure 2 The block diagram schematic diagram of the shoe-making machine for multi-angle waterproof monitoring in rain shoe production according to the embodiment of the present application is illustrated.

[0025] Figure 3 The block diagram schematic diagram of the siamese coding unit in the shoe-making machine for multi-angle waterproof monitoring in rain shoe production according to the embodiment of the present application is illustrated.

[0026] Figure 4 The block diagram schematic diagram of the multi-scale perception unit in the shoe-making machine for multi-angle waterproof monitoring in rain shoe production according to the embodiment of the present application is illustrated.

[0027] Figure 5 The block diagram schematic diagram of the multi-branch perception sub-unit in the shoe-making machine for multi-angle waterproof monitoring in rain shoe production according to the embodiment of the present application is illustrated.

[0028] Figure 6 The block diagram schematic diagram of the training module further included in the shoe-making machine for multi-angle waterproof monitoring in rain shoe production according to the embodiment of the present application is illustrated.

[0029] Figure 7 The flowchart of the multi - angle waterproof monitoring method for rain shoe production according to an embodiment of the present application is illustrated.

[0030] Figure 8 The schematic diagram of the system architecture of the multi - angle waterproof monitoring method for rain shoe production according to an embodiment of the present application is illustrated. Detailed implementation manners

[0031] Hereinafter, exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described herein.

[0032] Scene Overview

[0033] As described above, currently when conducting a waterproof test on rain shoes, mostly a person manually holds the rain shoes, then places the rain shoes in a water tank, and then observes the inside of the rain shoes to check if there is water accumulation inside. This waterproof detection mode is not conducive to observation and the detection result is not intuitive.

[0034] Through research, the inventors of the present application found that when conducting a waterproof test for rain shoes before leaving the factory, if the quality of the rain shoes is poor, it is often not detected during the factory waterproof test. The reason is that relatively inferior rain shoes take a long time to penetrate into the rain shoes, and it is difficult to observe within a short time. Therefore, an optimized rain shoe waterproof monitoring scheme is expected.

[0035] Correspondingly, the applicant of the present application found that if the rain shoes are relatively inferior, after being immersed in the water tank, even if the water stains on the surface of the rain shoes are wiped dry later, some water molecules will still penetrate into the rain shoes, resulting in a change in the acoustic characteristics of the rain shoes. Therefore, the waterproof performance of the rain shoes can be judged by the acoustic characteristics of the rain shoes after being immersed and wiped dry. In this way, it is possible to avoid the need to soak the rain shoes for a long time to observe the state inside the shoes in traditional waterproof detection, thereby shortening the detection time and improving the accuracy of the waterproof performance detection of the rain shoes.

[0036] Specifically, in the technical solution of the present application, an artificial intelligence detection technology based on deep learning is used to extract the feature of the sound signal of the to - be - detected rain shoes that have not been immersed and the to - be - detected rain shoes that have been immersed and wiped dry respectively, and the waterproof performance of the to - be - detected rain shoes is evaluated by the difference features between the two. That is, a waterproof performance detection scheme is constructed by comparing the difference features between the acoustic characteristics of the rain shoes after immersion and the acoustic characteristics of the rain shoes that have not been immersed.

[0037] Specifically, in the technical solution of the present application, first, a first sound signal of the rain shoes to be detected that have not been soaked and a second sound signal of the rain shoes to be detected that have been soaked and then dried are obtained. Next, considering that when detecting the waterproof performance of rain shoes through the acoustic characteristics of rain shoes, during the acquisition process of the sound signal, the sound signal of the rain shoes to be detected may be weak due to interference from other factors such as external environmental noise. Therefore, it is further necessary to perform time-domain enhancement on the sound signal. That is, calculate the time-domain enhancement maps of the first sound signal and the second sound signal to obtain a first time-domain enhancement map and a second time-domain enhancement map.

[0038] Then, considering that the convolutional neural network model has excellent performance in feature mining of images, in the technical solution of the present application, the convolutional neural network model is used to perform deep feature mining on the first time-domain enhancement map and the second time-domain enhancement map. Specifically, the first time-domain enhancement map and the second time-domain enhancement map are processed through a siamese network model including a first convolutional neural network and a second convolutional neural network to respectively extract the feature distributions of the local features in the first time-domain enhancement map and the second time-domain enhancement map in the high-dimensional space, so as to obtain a first feature map and a second feature map. Here, the first convolutional neural network and the second convolutional neural network have the same network structure.

[0039] It should be understood that in the standard deep convolutional neural network, convolution and pooling are common downsampling operations. However, while downsampling expands the receptive field, it also reduces the scale of the feature map, resulting in information loss. And because the standard deep convolutional neural network model uses a convolutional kernel with a fixed size, during feature extraction, it cannot learn multi-scale feature information, leading to the problem of local information loss caused by the grid effect.

[0040] To address the above problems, in the technical solution of the present application, a multi-branch receptive field module is used to perform feature enhancement on the first feature map and the second feature map obtained by using the traditional convolutional neural network model as a feature extractor, so as to obtain a first multi-scale receptive feature map and a second multi-scale receptive feature map. Compared with the traditional convolutional neural network model, the multi-branch receptive field module has the following advantages: 1) The multi-branch receptive field module uses dilated convolution to replace the traditional convolutional kernel. By using its unique dilation rate, the original convolutional kernel has a larger receptive field with the same number of parameters. That is, the multi-branch receptive field module can expand the receptive field by using dilated convolution, thus eliminating the need for downsampling, avoiding information loss, and achieving the same scale of the input and output of the feature map.

[0041] 2) The multi-branch perception field module designs a parallel dilated convolution structure with different dilation rates, which allows the network to learn multi-scale feature information, thus solving the problem of local information loss caused by the grid effect. Moreover, this structure increases the amount of small target information available for object detection, thereby solving the problem that the small target information cannot be reconstructed by using the pooling layer in traditional convolutional neural networks.

[0042] Further, calculate the differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map. In a specific example, the differential feature map can be obtained by calculating the position-wise difference between the first multi-scale perception feature map and the second multi-scale perception feature map. In this way, classification can be performed based on the differential features between the acoustic features of the rain shoes to be detected that are not soaked and the rain shoes to be detected that are soaked and then dried, so as to evaluate the waterproof performance of the rain shoes to be detected.

[0043] Particularly, in the technical solution of this application, since the differential feature map as the classification feature map is obtained by calculating the difference between the first multi-scale perception feature map and the second multi-scale perception feature map, during the training of the model, when calculating the gradient of the loss function and backpropagating from the classifier to the model, the gradient will pass through the corresponding convolutional neural network and multi-branch perception field module that obtain the first multi-scale perception feature map and the second multi-scale perception feature map respectively. At this time, the feature patterns extracted by the cascade branches of the corresponding convolutional neural network and multi-branch perception field module may be eliminated due to abnormal gradient branches.

[0044] Therefore, in addition to the classification loss function, a classification pattern elimination suppression loss function is further introduced to solve the elimination of the extracted feature patterns. Specifically, the classification pattern elimination suppression loss function is expressed as:

[0045]

[0046] V1 and V2 are the feature vectors obtained after unfolding the first multi-scale perception feature map and the second multi-scale perception feature map respectively, and M1 and M2 are the weight matrices of the classifier for V1 and V2 respectively, represents the square of the two-norm of the vector.

[0047] That is, by introducing a classification pattern cancellation suppression loss function, the pseudo-difference of the classifier weights is pushed towards the true feature distribution difference between the first multi-scale perception feature map and the second multi-scale perception feature map, that is, the feature distribution of the differential feature map, so as to ensure that the directional derivative during gradient backpropagation is regularized near the gradient branch point. That is, the gradient is over-weighted between the corresponding cascaded branches of the convolutional neural network and the multi-branch perception domain module, so as to suppress the classification pattern cancellation of the features, improve the extraction ability of the classification features of the corresponding cascaded branches of the convolutional neural network and the multi-branch perception domain module, and accordingly improve the accuracy of the classification result of the differential feature map. In this way, the waterproof performance of the rain boots can be evaluated and detected through the acoustic features of the rain boots to be detected, avoiding the problem of too long detection time caused by long-term soaking in traditional processes, and improving the accuracy of the waterproof performance detection of the rain boots.

[0048] Based on this, the present application provides a shoe-making machine for multi-angle waterproof monitoring in rain boot production, which includes: a detection signal acquisition unit for acquiring a first sound signal of a rain boot to be detected that has not been soaked and a second sound signal of the rain boot to be detected that has been soaked and dried; a time-domain conversion unit for calculating time-domain enhancement maps of the first sound signal and the second sound signal to obtain a first time-domain enhancement map and a second time-domain enhancement map; a twin coding unit for passing the first time-domain enhancement map and the second time-domain enhancement map through a twin network model including a first convolutional neural network and a second convolutional neural network to obtain a first feature map and a second feature map, where the first convolutional neural network and the second convolutional neural network have the same network structure; a multi-scale perception unit for respectively passing the first feature map and the second feature map through a multi-branch perception domain module to obtain a first multi-scale perception feature map and a second multi-scale perception feature map; a difference evaluation unit for calculating a differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map; and a waterproof monitoring result generation unit for passing the differential feature map through a classifier to obtain a classification result, where the classification result is used to indicate whether the waterproof performance of the rain boot to be detected meets a predetermined standard.

[0049] Figure 1 The figure shows an application scenario diagram of a shoe-making machine for multi-angle waterproof monitoring in rain boot production according to an embodiment of the present application. As Figure 1 shown, in this application scenario, by acquiring a first sound signal (for example, D1 as shown in Figure 1 ) of a rain boot to be detected that has not been soaked (for example, F1 as shown in Figure 1 ) and a second sound signal (for example, D2 as shown in Figure 1 ) of the rain boot to be detected that has been soaked and dried (for example, F2 as shown in Figure 1as shown in D2). Then, the obtained first sound signal and the second sound signal are input into a server deployed with a multi-angle waterproof monitoring algorithm for rain shoe production (e.g., Figure 1 as shown in S), wherein the server can use the multi-angle waterproof monitoring algorithm for rain shoe production to process the first sound signal and the second sound signal to generate a classification result indicating whether the waterproof performance of the rain shoes to be detected meets a predetermined standard.

[0050] After introducing the basic principle of the present application, various non-limiting embodiments of the present application will be specifically introduced with reference to the accompanying drawings.

[0051] Exemplary Shoe-making Machine

[0052] Figure 2 The block diagram of a shoe-making machine for multi-angle waterproof monitoring in rain shoe production according to an embodiment of the present application is illustrated. As Figure 2 shown, a shoe-making machine 100 for multi-angle waterproof monitoring in rain shoe production according to an embodiment of the present application includes: a detection signal acquisition unit 110 for acquiring a first sound signal of a rain shoe to be detected that is not soaked and a second sound signal of the rain shoe to be detected that is soaked and then dried; a time-domain conversion unit 120 for calculating time-domain enhancement diagrams of the first sound signal and the second sound signal to obtain a first time-domain enhancement diagram and a second time-domain enhancement diagram; a siamese encoding unit 130 for passing the first time-domain enhancement diagram and the second time-domain enhancement diagram through a siamese network model including a first convolutional neural network and a second convolutional neural network to obtain a first feature map and a second feature map, where the first convolutional neural network and the second convolutional neural network have the same network structure; a multi-scale perception unit 140 for respectively passing the first feature map and the second feature map through a multi-branch perception domain module to obtain a first multi-scale perception feature map and a second multi-scale perception feature map; a difference evaluation unit 150 for calculating a differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map; and a waterproof monitoring result generation unit 160 for passing the differential feature map through a classifier to obtain a classification result, where the classification result is used to indicate whether the waterproof performance of the rain shoes to be detected meets a predetermined standard.

[0053] More specifically, in the embodiments of the present application, the detection signal acquisition unit 110 is configured to obtain a first sound signal of the rain shoes to be detected that are not soaked and a second sound signal of the rain shoes to be detected that are soaked and then dried. If the rain shoes are of relatively poor quality, after being soaked in the water tank, even if the water stains on the surface of the rain shoes are dried subsequently, some water molecules will still penetrate into the rain shoes, resulting in a change in the acoustic characteristics of the rain shoes. Therefore, the waterproof performance of the rain shoes can be judged by the acoustic characteristics of the soaked and dried rain shoes, that is, a waterproof performance detection scheme is constructed by comparing the differential characteristics between the acoustic characteristics of the soaked rain shoes and the acoustic characteristics of the non-soaked rain shoes, thereby avoiding the need to soak the rain shoes for a long time to observe the internal state of the shoes in traditional waterproof detection.

[0054] More specifically, in the embodiments of the present application, the time-domain conversion unit 120 is configured to calculate time-domain enhancement maps of the first sound signal and the second sound signal to obtain a first time-domain enhancement map and a second time-domain enhancement map. Further considering that when detecting the waterproof performance of rain shoes through the acoustic characteristics of the rain shoes, during the acquisition process of the sound signal, the sound signal of the rain shoes to be detected may be weak due to interference from other factors such as external environmental noise. Therefore, it is further necessary to perform time-domain enhancement on the sound signal, so as to more obviously obtain the differences in subsequent feature mining and comparison.

[0055] More specifically, in the embodiments of the present application, the siamese coding unit 130 is configured to pass the first time-domain enhancement map and the second time-domain enhancement map through a siamese network model including a first convolutional neural network and a second convolutional neural network to obtain a first feature map and a second feature map, and the first convolutional neural network and the second convolutional neural network have the same network structure. The convolutional neural network model has excellent performance in feature mining of images. Therefore, in the technical solution of the present application, a convolutional neural network model with the same network structure is used to perform deep feature mining on the first time-domain enhancement map and the second time-domain enhancement map respectively to obtain their deep features.

[0056] Correspondingly, as Figure 3 shown, in a specific example, the siamese coding unit 130 includes: a first convolutional coding subunit 131, configured to use each layer of the first convolutional neural network to perform convolutional processing, pooling processing, and non-linear activation processing on the input data respectively during the forward propagation of the layer, and output the first feature map from the last layer of the first convolutional neural network; and a second convolutional coding subunit 132, configured to use each layer of the second convolutional neural network to perform convolutional processing, pooling processing, and non-linear activation processing on the input data respectively during the forward propagation of the layer, and output the second feature map from the last layer of the second convolutional neural network.

[0057] It should be understood that in the standard deep convolutional neural network, both convolution and pooling are common downsampling operations. However, while downsampling expands the receptive field, it also reduces the scale of the feature map, resulting in information loss. Moreover, since the standard deep convolutional neural network model uses a convolutional kernel of a fixed size, during feature extraction, it is unable to learn multi-scale feature information, leading to the problem of local information loss caused by the grid effect. To address the above problems, in the technical solution of this application, a multi-branch receptive field module is used to enhance the features of the first feature map and the second feature map obtained by using a traditional convolutional neural network model as a feature extractor, thereby obtaining a first multi-scale receptive feature map and a second multi-scale receptive feature map.

[0058] More specifically, in the embodiment of this application, the multi-scale receptive unit 140 is used to respectively pass the first feature map and the second feature map through the multi-branch receptive field module to obtain a first multi-scale receptive feature map and a second multi-scale receptive feature map. Compared with the traditional convolutional neural network model, the multi-branch receptive field module has the following advantages: 1) The multi-branch receptive field module uses dilated convolution to replace the traditional convolutional kernel. By using its unique dilation rate, the original convolutional kernel has a larger receptive field with the same number of parameters. That is, the multi-branch receptive field module can expand the receptive field using dilated convolution, thus eliminating the need for downsampling, avoiding information loss, and achieving consistent input and output scales of the feature map; 2) The multi-branch receptive field module designs a parallel dilated convolution structure with different dilation rates, enabling the network to learn multi-scale feature information. Therefore, the problem of local information loss caused by the grid effect is solved. This structure increases the amount of small target information available for object detection, thereby solving the problem that the traditional convolutional neural network cannot reconstruct small target information using the pooling layer.

[0059] Correspondingly, as Figure 4As shown, in a specific example, the multi-scale perception unit 140 includes: a first point convolution subunit 141, configured to input the first feature map and the second feature map into the first point convolution layer of the multi-branch perception domain module respectively to obtain a first convolution feature map and a second convolution feature map; a multi-branch perception subunit 142, configured to input the first convolution feature map and the second convolution feature map through the first branch perception domain unit, the second branch perception domain unit, and the third branch perception domain unit of the multi-branch perception domain module respectively to obtain a first branch perception feature map, a second branch perception feature map, and a third branch perception feature map, and a fourth branch perception feature map, a fifth branch perception feature map, and a sixth branch perception feature map, wherein the first branch perception domain unit, the second branch perception domain unit, and the third branch perception domain unit have a parallel structure; a fusion subunit 143, configured to cascade the first branch perception feature map, the second branch perception feature map, and the third branch perception feature map to obtain a first fusion perception feature map, and cascade the fourth branch perception feature map, the fifth branch perception feature map, and the sixth branch perception feature map to obtain a second fusion perception feature map; a second point convolution subunit 144, configured to input the first fusion perception feature map and the second fusion perception feature map into the second point convolution layer of the multi-branch perception domain module respectively to obtain a first channel-corrected fusion perception feature map and a second channel-corrected fusion perception feature map; and a residual cascade subunit 145, configured to calculate the point-by-point addition of the first channel-corrected fusion perception feature map and the first convolution feature map to obtain the first multi-scale perception feature map, and calculate the point-by-point addition of the second channel-corrected fusion perception feature map and the second convolution feature map to obtain the second multi-scale perception feature map.

[0060] Accordingly, as Figure 5As shown, in a specific example, the multi-branch perception subunit 142 includes: a first one-dimensional convolutional encoding secondary subunit 14201, configured to obtain a first one-dimensional convolutional feature map by passing the first convolutional feature map through the first one-dimensional convolutional layer of the first branch perception domain unit; a first dilated convolutional encoding secondary subunit 14202, configured to obtain the first branch perception feature map by passing the first one-dimensional convolutional feature map through the first two-dimensional convolutional layer with a first dilation rate; a second one-dimensional convolutional encoding secondary subunit 14203, configured to obtain a second one-dimensional convolutional feature map by passing the first convolutional feature map through the second one-dimensional convolutional layer of the second branch perception domain unit; a second dilated convolutional encoding secondary subunit 14204, configured to obtain the second branch perception feature map by passing the second one-dimensional convolutional feature map through the second two-dimensional convolutional layer with a second dilation rate; a third one-dimensional convolutional encoding secondary subunit 14205, configured to obtain a third one-dimensional convolutional feature map by passing the first convolutional feature map through the third one-dimensional convolutional layer of the third branch perception domain unit; a third dilated convolutional encoding secondary subunit 14206, configured to obtain the third branch perception feature map by passing the third one-dimensional convolutional feature map through the third two-dimensional convolutional layer with a third dilation rate; a fourth one-dimensional convolutional encoding secondary subunit 14207, configured to obtain a fourth one-dimensional convolutional feature map by passing the second convolutional feature map through the fourth one-dimensional convolutional layer of the fourth branch perception domain unit; a fourth dilated convolutional encoding secondary subunit 14208, configured to obtain the fourth branch perception feature map by passing the fourth one-dimensional convolutional feature map through the fourth two-dimensional convolutional layer with a fourth dilation rate; a fifth one-dimensional convolutional encoding secondary subunit 14209, configured to obtain a fifth one-dimensional convolutional feature map by passing the second convolutional feature map through the fifth one-dimensional convolutional layer of the fifth branch perception domain unit; a fifth dilated convolutional encoding secondary subunit 14210, configured to obtain the fifth branch perception feature map by passing the fifth one-dimensional convolutional feature map through the fifth two-dimensional convolutional layer with a fifth dilation rate; a sixth one-dimensional convolutional encoding secondary subunit 14211, configured to obtain a sixth one-dimensional convolutional feature map by passing the second convolutional feature map through the sixth one-dimensional convolutional layer of the sixth branch perception domain unit; and a sixth dilated convolutional encoding secondary subunit 14212, configured to obtain the sixth branch perception feature map by passing the sixth one-dimensional convolutional feature map through the sixth two-dimensional convolutional layer with a sixth dilation rate.

[0061] Correspondingly, in a specific example, the first dilation rate, the second dilation rate, and the third dilation rate are not equal to each other, and the third dilation rate, the fourth dilation rate, and the fifth dilation rate are not equal to each other.

[0062] More specifically, in the embodiments of the present application, the difference evaluation unit 150 is configured to calculate a differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map. Classification is performed based on the differential features between the acoustic features of the to-be-detected rain shoes that are not soaked and the to-be-detected rain shoes that are soaked and then dried, so as to evaluate the waterproof performance of the to-be-detected rain shoes. In a specific example, the differential feature map can be obtained by calculating the position-wise difference between the first multi-scale perception feature map and the second multi-scale perception feature map.

[0063] Correspondingly, in a specific example, the difference evaluation unit 150 is further configured to: calculate the differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map using the following formula; where the formula is:

[0064]

[0065] where, F d represents the differential feature map, F1 represents the first multi-scale perception feature map, F2 represents the second multi-scale perception feature map, represents position-wise subtraction.

[0066] More specifically, in the embodiments of the present application, the waterproof monitoring result generation unit 160 is configured to pass the differential feature map through a classifier to obtain a classification result, and the classification result is used to indicate whether the waterproof performance of the to-be-detected rain shoes meets a predetermined standard.

[0067] Correspondingly, in a specific example, the waterproof monitoring result generation unit 160 is further configured to: use the classifier to process the differential feature map with the following formula to generate a classification result; where the formula is: softmax{(M c , B c )|Project(F)}, where Project(F) represents projecting the differential feature map into a vector, M c is the weight matrix of the fully connected layer, and B c represents the bias vector of the fully connected layer.

[0068] Correspondingly, in a specific example, the shoe-making machine for multi-angle waterproof monitoring for rain shoe production further includes a training module for training the twin network, the multi-branch perception domain module, and the classifier; where, as Figure 6As shown, the training module 200 includes: a training detection signal acquisition unit 210, configured to obtain training data, where the training data includes a first training sound signal of a rain shoe to be detected that is not soaked and a second training sound signal of the rain shoe to be detected that is soaked and then dried, and a true value indicating whether the waterproof performance of the rain shoe to be detected meets a predetermined standard; a training time-domain conversion unit 220, configured to calculate time-domain enhancement maps of the first training sound signal and the second training sound signal to obtain a first training time-domain enhancement map and a second training time-domain enhancement map; a training siamese encoding unit 230, configured to pass the first training time-domain enhancement map and the second training time-domain enhancement map through the siamese network model including a first convolutional neural network and a second convolutional neural network to obtain a first training feature map and a second training feature map, where the first convolutional neural network and the second convolutional neural network have the same network structure; a training multi-scale perception unit 240, configured to pass the first training feature map and the second training feature map through the multi-branch perception domain module respectively to obtain a first training multi-scale perception feature map and a second training multi-scale perception feature map; a training difference evaluation unit 250, configured to calculate a training difference feature map between the first training multi-scale perception feature map and the second training multi-scale perception feature map; a classification loss function value calculation unit 260, configured to pass the training difference feature map through the classifier to obtain a classification loss function value; a classification pattern cancellation suppression loss function value calculation unit 270, configured to calculate a classification pattern cancellation suppression loss function value of the first training multi-scale perception feature map and the second training multi-scale perception feature map; and a training unit 280, configured to use a weighted sum of the classification pattern cancellation suppression loss function value and the classification loss function value as a loss function value to train the siamese network, the multi-branch perception domain module, and the classifier.

[0069] Specifically, in the technical solution of the present application, since the difference feature map as the classification feature map is obtained by calculating the difference between the first multi-scale perception feature map and the second multi-scale perception feature map, during the training of the model, when calculating the gradient of the loss function and backpropagating from the classifier to the model, the gradient will pass through the corresponding convolutional neural network and multi-branch perception domain module that obtain the first multi-scale perception feature map and the second multi-scale perception feature map respectively. At this time, the feature patterns extracted by the cascaded branches of the corresponding convolutional neural network and multi-branch perception domain module may be cancelled due to abnormal gradient branches. Therefore, in addition to the classification loss function, a classification pattern cancellation suppression loss function is further introduced to solve the cancellation of the extracted feature patterns.

[0070] Correspondingly, in a specific example, the classification pattern cancellation suppression loss function value calculation unit 270 is further configured to: where the formula is:

[0071]

[0072] Where V1 and V2 are the feature vectors obtained after unfolding the first training multi-scale perception feature map and the second training multi-scale perception feature map respectively, and M1 and M2 are the weight matrices of the classifier for the feature vectors obtained after unfolding the first training multi-scale perception feature map and the second training multi-scale perception feature map respectively. Denotes the square of the two-norm of a vector, ‖·‖ F Denotes the Frobenius norm of a matrix. Denotes subtraction by position, exp(·) denotes the exponential operation of a vector and the exponential operation of a matrix. The exponential operation of the vector means calculating the natural exponential function values with the eigenvalues at each position in the vector as the exponents, and the exponential operation of the matrix means calculating the natural exponential function values with the eigenvalues at each position in the matrix as the exponents.

[0073] That is, by introducing a classification pattern cancellation and suppression loss function, the pseudo-difference of the classifier weights is pushed towards the true feature distribution difference between the first multi-scale perception feature map and the second multi-scale perception feature map, that is, the feature distribution of the differential feature map, so as to ensure that the directional derivative during gradient backpropagation is regularized near the gradient branch point. That is, the gradient is over-weighted between the cascaded branches of the corresponding convolutional neural network and the multi-branch perception domain module, so as to suppress the classification pattern cancellation of the features, improve the extraction ability of the classification features of the cascaded branches of the corresponding convolutional neural network and the multi-branch perception domain module, and thus correspondingly improve the accuracy of the classification result of the differential feature map. In this way, the waterproof performance of the rain shoes can be evaluated and detected through the acoustic features of the rain shoes to be detected, so as to avoid the problem of too long detection time caused by long-term soaking in the traditional process, and can also improve the accuracy of the waterproof performance detection of the rain shoes.

[0074] In summary, the shoe-making machine 100 for multi-angle waterproof monitoring in rain shoe production according to the embodiments of the present application is elucidated. First, it calculates the time-domain enhancement maps of the first sound signal of the rain shoe to be detected that is not soaked and the second sound signal of the rain shoe to be detected that is soaked and then dried to obtain the first time-domain enhancement map and the second time-domain enhancement map. Then, the first time-domain enhancement map and the second time-domain enhancement map are passed through a siamese network model to obtain a first feature map and a second feature map. Next, it calculates the differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map obtained by passing the first feature map and the second feature map through a multi-branch perception domain module respectively. Finally, the differential feature map is passed through a classifier to obtain a classification result for indicating whether the waterproof performance of the rain shoe to be detected meets a predetermined standard. In this way, the detection time can be shortened and the accuracy of rain shoe waterproof performance detection can be improved.

[0075] As described above, the shoe-making machine 100 for multi-angle waterproof monitoring in rain shoe production according to the embodiments of the present application can be implemented in various terminal devices, such as a server with an algorithm for multi-angle waterproof monitoring in rain shoe production. In one example, the shoe-making machine 100 for multi-angle waterproof monitoring in rain shoe production can be integrated into the terminal device as a software module and / or a hardware module. For example, the shoe-making machine 100 for multi-angle waterproof monitoring in rain shoe production can be a software module in the operating system of the terminal device, or can be an application program developed for the terminal device; of course, the shoe-making machine 100 for multi-angle waterproof monitoring in rain shoe production can also be one of the many hardware modules of the terminal device.

[0076] Alternatively, in another example, the shoe-making machine 100 for multi-angle waterproof monitoring in rain shoe production and the terminal device can also be separate devices, and the shoe-making machine 100 for multi-angle waterproof monitoring in rain shoe production can be connected to the terminal device through a wired and / or wireless network and transmit interaction information according to a predefined data format.

[0077] Exemplary Method

[0078] Figure 7 The flowchart of the method for multi-angle waterproof monitoring in rain shoe production according to the embodiments of the present application is illustrated. As Figure 7As shown, the multi-angle waterproof monitoring method for rain shoe production according to an embodiment of the present application includes: S110, obtaining a first sound signal of the rain shoe to be detected that is not soaked and a second sound signal of the rain shoe to be detected that is soaked and then dried; S120, calculating time-domain enhancement maps of the first sound signal and the second sound signal to obtain a first time-domain enhancement map and a second time-domain enhancement map; S130, passing the first time-domain enhancement map and the second time-domain enhancement map through a siamese network model including a first convolutional neural network and a second convolutional neural network to obtain a first feature map and a second feature map, where the first convolutional neural network and the second convolutional neural network have the same network structure; S140, respectively passing the first feature map and the second feature map through a multi-branch perception domain module to obtain a first multi-scale perception feature map and a second multi-scale perception feature map; S150, calculating a differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map; and S160, passing the differential feature map through a classifier to obtain a classification result, where the classification result is used to indicate whether the waterproof performance of the rain shoe to be detected meets a predetermined standard.

[0079] Figure 8 The figure shows a schematic diagram of the system architecture of the multi-angle waterproof monitoring method for rain shoe production according to an embodiment of the present application. As Figure 8 shown, in the system architecture of the multi-angle waterproof monitoring method for rain shoe production, first, a first sound signal of the rain shoe to be detected that is not soaked and a second sound signal of the rain shoe to be detected that is soaked and then dried are obtained; then, time-domain enhancement maps of the first sound signal and the second sound signal are calculated to obtain a first time-domain enhancement map and a second time-domain enhancement map; next, the first time-domain enhancement map and the second time-domain enhancement map are passed through a siamese network model including a first convolutional neural network and a second convolutional neural network to obtain a first feature map and a second feature map, where the first convolutional neural network and the second convolutional neural network have the same network structure; then, the first feature map and the second feature map are respectively passed through a multi-branch perception domain module to obtain a first multi-scale perception feature map and a second multi-scale perception feature map; next, a differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map is calculated; finally, the differential feature map is passed through a classifier to obtain a classification result, where the classification result is used to indicate whether the waterproof performance of the rain shoe to be detected meets a predetermined standard.

[0080] In a specific example, in the above multi-angle waterproof monitoring method for rain shoe production, the step of obtaining a first feature map and a second feature map by passing the first time-domain enhanced map and the second time-domain enhanced map through a siamese network model including a first convolutional neural network and a second convolutional neural network includes: using each layer of the first convolutional neural network to perform convolution processing, pooling processing, and non-linear activation processing on the input data respectively during the forward pass of the layer, so as to output the first feature map from the last layer of the first convolutional neural network; and using each layer of the second convolutional neural network to perform convolution processing, pooling processing, and non-linear activation processing on the input data respectively during the forward pass of the layer, so as to output the second feature map from the last layer of the second convolutional neural network.

[0081] In a specific example, in the above multi-angle waterproof monitoring method for rain shoe production, the step of obtaining a first multi-scale perception feature map and a second multi-scale perception feature map by passing the first feature map and the second feature map through a multi-branch perception domain module respectively includes: inputting the first feature map and the second feature map into the first point convolutional layer of the multi-branch perception domain module respectively to obtain a first convolutional feature map and a second convolutional feature map; passing the first convolutional feature map and the second convolutional feature map through the first branch perception domain unit, the second branch perception domain unit, and the third branch perception domain unit of the multi-branch perception domain module respectively to obtain a first branch perception feature map, a second branch perception feature map, and a third branch perception feature map, as well as a fourth branch perception feature map, a fifth branch perception feature map, and a sixth branch perception feature map, wherein the first branch perception domain unit, the second branch perception domain unit, and the third branch perception domain unit have a parallel structure; concatenating the first branch perception feature map, the second branch perception feature map, and the third branch perception feature map to obtain a first fusion perception feature map, and concatenating the fourth branch perception feature map, the fifth branch perception feature map, and the sixth branch perception feature map to obtain a second fusion perception feature map; inputting the first fusion perception feature map and the second fusion perception feature map into the second point convolutional layer of the multi-branch perception domain module respectively to obtain a first channel-corrected fusion perception feature map and a second channel-corrected fusion perception feature map; and calculating the point-by-point addition of the first channel-corrected fusion perception feature map and the first convolutional feature map to obtain the first multi-scale perception feature map, and calculating the point-by-point addition of the second channel-corrected fusion perception feature map and the second convolutional feature map to obtain the second multi-scale perception feature map.

[0082] In a specific example, in the above multi-angle waterproof monitoring method for rain shoe production, the steps of respectively passing the first convolutional feature map and the second convolutional feature map through the first branch perception domain unit, the second branch perception domain unit, and the third branch perception domain unit of the multi-branch perception domain module to obtain the first branch perception feature map, the second branch perception feature map, and the third branch perception feature map, as well as the fourth branch perception feature map, the fifth branch perception feature map, and the sixth branch perception feature map, include: passing the first convolutional feature map through the first one-dimensional convolutional layer of the first branch perception domain unit to obtain a first one-dimensional convolutional feature map; passing the first one-dimensional convolutional feature map through a first two-dimensional convolutional layer with a first dilation rate to obtain the first branch perception feature map; passing the first convolutional feature map through the second one-dimensional convolutional layer of the second branch perception domain unit to obtain a second one-dimensional convolutional feature map; passing the second one-dimensional convolutional feature map through a second two-dimensional convolutional layer with a second dilation rate to obtain the second branch perception feature map; passing the first convolutional feature map through the third one-dimensional convolutional layer of the third branch perception domain unit to obtain a third one-dimensional convolutional feature map; passing the third one-dimensional convolutional feature map through a third two-dimensional convolutional layer with a third dilation rate to obtain the third branch perception feature map; passing the second convolutional feature map through the fourth one-dimensional convolutional layer of the fourth branch perception domain unit to obtain a fourth one-dimensional convolutional feature map; passing the fourth one-dimensional convolutional feature map through a fourth two-dimensional convolutional layer with a fourth dilation rate to obtain the fourth branch perception feature map; passing the second convolutional feature map through the fifth one-dimensional convolutional layer of the fifth branch perception domain unit to obtain a fifth one-dimensional convolutional feature map; passing the fifth one-dimensional convolutional feature map through a fifth two-dimensional convolutional layer with a fifth dilation rate to obtain the fifth branch perception feature map; passing the second convolutional feature map through the sixth one-dimensional convolutional layer of the sixth branch perception domain unit to obtain a sixth one-dimensional convolutional feature map; and passing the sixth one-dimensional convolutional feature map through a sixth two-dimensional convolutional layer with a sixth dilation rate to obtain the sixth branch perception feature map.

[0083] In a specific example, in the above multi-angle waterproof monitoring method for rain shoe production, the first dilation rate, the second dilation rate, and the third dilation rate are not equal to each other, and the third dilation rate, the fourth dilation rate, and the fifth dilation rate are not equal to each other.

[0084] In a specific example, in the above multi-angle waterproof monitoring method for rain shoe production, calculating the differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map further includes: calculating the differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map using the following formula; where the formula is:

[0085]

[0086] Among them, F d represents the differential feature map, F1 represents the first multi-scale perception feature map, and F2 represents the second multi-scale perception feature map. represents subtraction by position.

[0087] In a specific example, in the above multi-angle waterproof monitoring method for rain shoe production, the step of obtaining the classification result by passing the differential feature map through a classifier further includes: using the classifier to process the differential feature map according to the following formula to generate a classification result; where the formula is: softmax{(M c , B c )|Project(F)}, where Project(F) represents projecting the differential feature map into a vector, M c is the weight matrix of the fully connected layer, and B c represents the bias vector of the fully connected layer.

[0088] In a specific example, in the above multi-angle waterproof monitoring method for rain shoe production, the multi-angle waterproof monitoring method for rain shoe production further includes: training the siamese network, the multi-branch perception domain module, and the classifier; where training the siamese network, the multi-branch perception domain module, and the classifier includes: obtaining training data, where the training data includes a first training sound signal of the to-be-detected rain shoe that is not soaked and a second training sound signal of the to-be-detected rain shoe that is soaked and then dried, and a true value indicating whether the waterproof performance of the to-be-detected rain shoe meets a predetermined standard; calculating time-domain enhanced maps of the first training sound signal and the second training sound signal to obtain a first training time-domain enhanced map and a second training time-domain enhanced map; passing the first training time-domain enhanced map and the second training time-domain enhanced map through the siamese network model including a first convolutional neural network and a second convolutional neural network to obtain a first training feature map and a second training feature map, where the first convolutional neural network and the second convolutional neural network have the same network structure; respectively passing the first training feature map and the second training feature map through the multi-branch perception domain module to obtain a first training multi-scale perception feature map and a second training multi-scale perception feature map; calculating a training differential feature map between the first training multi-scale perception feature map and the second training multi-scale perception feature map; passing the training differential feature map through the classifier to obtain a classification loss function value; calculating a classification pattern cancellation and suppression loss function value of the first training multi-scale perception feature map and the second training multi-scale perception feature map; and training the siamese network, the multi-branch perception domain module, and the classifier with the weighted sum of the classification pattern cancellation and suppression loss function value and the classification loss function value as the loss function value.

[0089] In a specific example, in the above multi-angle waterproof monitoring method for rain shoe production, calculating the classification pattern cancellation and suppression loss function values of the first training multi-scale perception feature map and the second training multi-scale perception feature map further includes: where the formula is:

[0090]

[0091] where V1 and V2 are the feature vectors obtained after unfolding the first training multi-scale perception feature map and the second training multi-scale perception feature map respectively, and M1 and M2 are the weight matrices of the classifier for the feature vectors obtained after unfolding the first training multi-scale perception feature map and the second training multi-scale perception feature map respectively, represents the square of the two-norm of the vector, ‖·‖ F represents the Frobenius norm of the matrix, represents element-wise subtraction, exp(·) represents the exponential operation of the vector and the exponential operation of the matrix. The exponential operation of the vector means calculating the natural exponential function values with the eigenvalues at each position in the vector as the exponents, and the exponential operation of the matrix means calculating the natural exponential function values with the eigenvalues at each position in the matrix as the exponents.

[0092] Here, those skilled in the art can understand that the specific operations of each step in the above multi-angle waterproof monitoring method for rain shoe production have been introduced in detail in the description of the shoe-making machine for multi-angle waterproof monitoring of rain shoe production above with reference to Figures 1 to 6 and thus, the repeated description thereof will be omitted.

Claims

1. A shoe-making machine for multi-angle waterproof monitoring in rain shoe production, characterized in that, Including: A detection signal acquisition unit, configured to obtain a first sound signal of a rain shoe to be detected that is not soaked and a second sound signal of the rain shoe to be detected that is soaked and then dried. A time-domain conversion unit, configured to calculate time-domain enhancement maps of the first sound signal and the second sound signal to obtain a first time-domain enhancement map and a second time-domain enhancement map. A siamese encoding unit, configured to pass the first time-domain enhancement map and the second time-domain enhancement map through a siamese network model including a first convolutional neural network and a second convolutional neural network to obtain a first feature map and a second feature map, where the first convolutional neural network and the second convolutional neural network have the same network structure. A multi-scale perception unit, configured to respectively pass the first feature map and the second feature map through a multi-branch perception domain module to obtain a first multi-scale perception feature map and a second multi-scale perception feature map. A difference evaluation unit, configured to calculate a differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map. And A waterproof monitoring result generation unit, configured to pass the differential feature map through a classifier to obtain a classification result, where the classification result is used to indicate whether the waterproof performance of the rain shoe to be detected meets a predetermined standard.

2. The shoemaking machine for multi-angle waterproof monitoring used in the production of rain shoes according to claim 1, characterized in that The siamese encoding unit includes: A first convolutional encoding sub-unit, configured to respectively perform convolution processing, pooling processing, and non-linear activation processing on input data by using each layer of the first convolutional neural network in the forward pass of the layer, and output the first feature map from the last layer of the first convolutional neural network; and A second convolutional encoding sub-unit, configured to respectively perform convolution processing, pooling processing, and non-linear activation processing on input data by using each layer of the second convolutional neural network in the forward pass of the layer, and output the second feature map from the last layer of the second convolutional neural network.

3. The shoe-making machine for multi-angle waterproof monitoring used in rain shoe production according to claim 2, characterized in that, The multi-scale perception unit includes: A first point convolution sub-unit, configured to respectively input the first feature map and the second feature map into a first point convolution layer of the multi-branch perception domain module to obtain a first convolutional feature map and a second convolutional feature map. A multi-branch perception sub-unit, configured to respectively pass the first convolutional feature map and the second convolutional feature map through a first branch perception domain unit, a second branch perception domain unit, and a third branch perception domain unit of the multi-branch perception domain module to obtain a first branch perception feature map, a second branch perception feature map, and a third branch perception feature map, and a fourth branch perception feature map, a fifth branch perception feature map, and a sixth branch perception feature map, where the first branch perception domain unit, the second branch perception domain unit, and the third branch perception domain unit have a parallel structure. A fusion sub-unit, configured to concatenate the first branch perception feature map, the second branch perception feature map, and the third branch perception feature map to obtain a first fusion perception feature map, and concatenate the fourth branch perception feature map, the fifth branch perception feature map, and the sixth branch perception feature map to obtain a second fusion perception feature map. The second point convolution sub-unit is used to input the first fusion perception feature map and the second fusion perception feature map into the second point convolution layer of the multi-branch perception domain module respectively to obtain a first channel-corrected fusion perception feature map and a second channel-corrected fusion perception feature map; and The residual concatenation sub-unit is used to calculate the element-wise addition of the first channel-corrected fusion perception feature map and the first convolution feature map to obtain the first multi-scale perception feature map, and calculate the element-wise addition of the second channel-corrected fusion perception feature map and the second convolution feature map to obtain the second multi-scale perception feature map.

4. The shoe-making machine for multi-angle waterproof monitoring used in rain shoe production according to claim 3, characterized in that, The multi-branch perception sub-unit includes: The first one-dimensional convolution encoding secondary sub-unit is used to pass the first convolution feature map through the first one-dimensional convolution layer of the first branch perception domain unit to obtain a first one-dimensional convolution feature map; The first dilated convolution encoding secondary sub-unit is used to pass the first one-dimensional convolution feature map through the first two-dimensional convolution layer with the first dilation rate to obtain the first branch perception feature map; The second one-dimensional convolution encoding secondary sub-unit is used to pass the first convolution feature map through the second one-dimensional convolution layer of the second branch perception domain unit to obtain a second one-dimensional convolution feature map; The second dilated convolution encoding secondary sub-unit is used to pass the second one-dimensional convolution feature map through the second two-dimensional convolution layer with the second dilation rate to obtain the second branch perception feature map; The third one-dimensional convolution encoding secondary sub-unit is used to pass the first convolution feature map through the third one-dimensional convolution layer of the third branch perception domain unit to obtain a third one-dimensional convolution feature map; The third dilated convolution encoding secondary sub-unit is used to pass the third one-dimensional convolution feature map through the third two-dimensional convolution layer with the third dilation rate to obtain the third branch perception feature map; The fourth one-dimensional convolution encoding secondary sub-unit is used to pass the second convolution feature map through the fourth one-dimensional convolution layer of the fourth branch perception domain unit to obtain a fourth one-dimensional convolution feature map; The fourth dilated convolution encoding secondary sub-unit is used to pass the fourth one-dimensional convolution feature map through the fourth two-dimensional convolution layer with the fourth dilation rate to obtain the fourth branch perception feature map; The fifth one-dimensional convolution encoding secondary sub-unit is used to pass the second convolution feature map through the fifth one-dimensional convolution layer of the fifth branch perception domain unit to obtain a fifth one-dimensional convolution feature map; The fifth dilated convolution encoding secondary sub-unit is used to pass the fifth one-dimensional convolution feature map through the fifth two-dimensional convolution layer with the fifth dilation rate to obtain the fifth branch perception feature map; The sixth one-dimensional convolution encoding secondary sub-unit is used to pass the second convolution feature map through the sixth one-dimensional convolution layer of the sixth branch perception domain unit to obtain a sixth one-dimensional convolution feature map; and The sixth dilated convolution encoding secondary sub-unit is used to pass the sixth one-dimensional convolution feature map through the sixth two-dimensional convolution layer with the sixth dilation rate to obtain the sixth branch perception feature map.

5. The shoemaking machine for multi-angle waterproof monitoring used in rain shoe production according to claim 4, characterized in that, The first dilation rate, the second dilation rate, and the third dilation rate are not equal to each other, and the third dilation rate, the fourth dilation rate, and the fifth dilation rate are not equal to each other.

6. The shoemaking machine for multi-angle waterproof monitoring used in the production of rain boots according to claim 5, wherein The difference evaluation unit is further configured to: Calculate a differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map by using the following formula; Wherein, the formula is: Among them, represents the differential feature map, represents the first multi-scale perception feature map, represents the second multi-scale perception feature map, represents subtraction by position.

7. The shoe-making machine for multi-angle waterproof monitoring used in rain shoe production according to claim 6, characterized in that, The waterproof monitoring result generation unit is further configured to: Use the classifier to process the differential feature map according to the following formula to generate a classification result; where the formula is: , where represents projecting the differential feature map into a vector, is the weight matrix of the fully connected layer, represents the bias vector of the fully connected layer.

8. The shoe-making machine for multi-angle waterproof monitoring used in the production of rain boots according to claim 1, characterized in that, It further includes a training module for training the siamese network, the multi-branch perception domain module and the classifier; Wherein, the training module includes: A training detection signal acquisition unit for acquiring training data, the training data including a first training sound signal of the to-be-detected rain shoes not soaked and a second training sound signal of the to-be-detected rain shoes soaked and dried, and a true value of whether the waterproof performance of the to-be-detected rain shoes meets a predetermined standard; A training time-domain conversion unit for calculating time-domain enhancement maps of the first training sound signal and the second training sound signal to obtain a first training time-domain enhancement map and a second training time-domain enhancement map; A training siamese encoding unit for passing the first training time-domain enhancement map and the second training time-domain enhancement map through the siamese network model including a first convolutional neural network and a second convolutional neural network to obtain a first training feature map and a second training feature map, and the first convolutional neural network and the second convolutional neural network have the same network structure; A training multi-scale perception unit for respectively passing the first training feature map and the second training feature map through the multi-branch perception domain module to obtain a first training multi-scale perception feature map and a second training multi-scale perception feature map; A training difference evaluation unit for calculating a training differential feature map between the first training multi-scale perception feature map and the second training multi-scale perception feature map; A classification loss function value calculation unit for passing the training differential feature map through the classifier to obtain a classification loss function value; A classification mode cancellation and suppression loss function value calculation unit for calculating a classification mode cancellation and suppression loss function value of the first training multi-scale perception feature map and the second training multi-scale perception feature map; and A training unit for training the siamese network, the multi-branch perception domain module and the classifier by using the weighted sum of the classification mode cancellation and suppression loss function value and the classification loss function value as the loss function value.

9. The shoe-making machine for multi-angle waterproof monitoring used in rain shoe production according to claim 8, characterized in that, The classification mode cancellation and suppression loss function value calculation unit is further configured to: Wherein, the formula is: where and are the feature vectors obtained after unfolding the first training multi-scale perception feature map and the second training multi-scale perception feature map respectively, and and are the weight matrices of the classifier for the feature vectors obtained after unfolding the first training multi-scale perception feature map and the second training multi-scale perception feature map respectively, represents the square of the two-norm of a vector, represents the Frobenius norm of a matrix, represents element-wise subtraction, represents the exponential operation of a vector and the exponential operation of a matrix. The exponential operation of the vector means calculating the values of the natural exponential function with the eigenvalues at each position in the vector as the exponents, and the exponential operation of the matrix means calculating the values of the natural exponential function with the eigenvalues at each position in the matrix as the exponents.

10. A multi-angle waterproof monitoring method for rain shoe production, characterized in that, It includes: Obtain a first sound signal of the to-be-detected rain shoes not soaked and a second sound signal of the to-be-detected rain shoes soaked and dried; Calculate time-domain enhancement maps of the first sound signal and the second sound signal to obtain a first time-domain enhancement map and a second time-domain enhancement map; Pass the first time-domain enhancement map and the second time-domain enhancement map through a siamese network model including a first convolutional neural network and a second convolutional neural network to obtain a first feature map and a second feature map, and the first convolutional neural network and the second convolutional neural network have the same network structure; Respectively pass the first feature map and the second feature map through the multi-branch perception domain module to obtain a first multi-scale perception feature map and a second multi-scale perception feature map; Calculate the differential feature map between the first multi-scale perception feature map and the second multi-scale perception feature map; And Pass the differential feature map through a classifier to obtain a classification result, and the classification result is used to indicate whether the waterproof performance of the rain shoes to be detected meets a predetermined standard.

Images