Deep Learning-Based Method for Extracting Impact Tabletop Simulation Image Features

Abstract

The present invention belongs to the technical field of neural network feature extraction, and specifically relates to a method for extracting features of impact table simulation images based on deep learning. The method includes: obtaining an impact table simulation image data set, calculating the feature capture index of a feature map through the feature map and the symmetric feature map, obtaining the discrimination degree between feature maps according to the feature capture index of the feature map and the information entropy of the feature map, obtaining the reconstruction coefficient of the target input feature map group according to the discrimination degree between the feature maps in the target input feature map group, performing grouped reconstruction according to the reconstruction coefficient, extracting features of the impact table simulation image through the neural network after grouped reconstruction, and judging whether the structure of the impact table is reasonable. The present invention improves the feature extraction ability of grouped convolution through grouped reconstruction, which is beneficial to the judgment of the rationality of the impact table structure.

Description

Technical Field

[0001] The present invention relates to the technical field of neural network feature extraction. More specifically, the present invention relates to a method for extracting features of an impact table simulation image based on deep learning. Background Art

[0002] When a ship encounters a non-contact explosion during its use, it is the main form of damage to the ship's equipment system failure. If there are defects in the ship's production and manufacturing process, it will seriously affect the ship's life cycle. In order to standardize the design, manufacturing and acceptance of ship equipment and improve the ability to resist non-contact explosion, an impact machine can be used as a test device for ship equipment impact tests. In the test, the impact table of the impact machine is a key component for simulating the force response of ship equipment in a non-contact explosion impact environment. Its function is to effectively transfer the impact load generated by the impact hammer to the device under test. The impact table needs to have sufficient structural strength and stiffness. An unreasonable structure of the impact table will affect the stability and repeatability of the impact loading process. Therefore, it is necessary to judge the rationality of the impact table structure.

[0003] By the method of producing impact tables of various structures for load testing, it is necessary to consume certain human and material resources, which will affect the judgment progress of the load capacity of impact tables of various structures. Therefore, in the design stage, it is usually necessary to combine finite element analysis to perform dynamic response analysis and structural optimization on the impact table to ensure the reliability of the impact table.

[0004] In the related art, a convolutional neural network is used to extract features of the impact table simulation image to judge whether the structure of the impact table is reasonable. In order to reduce the calculation burden, the grouped convolution method can be used to reduce the model parameters. However, in the prior art, the grouping of the grouped convolution is fixed. The grouped convolution with fixed grouping may cause information fragmentation between the feature maps of different channels. At the same time, the grouped convolution with fixed grouping cannot dynamically respond to the structural characteristics of the impact table, which may cause some key information to be lost, resulting in insufficient extraction of important features in the impact table simulation image, thereby affecting the analysis of the impact table load and further affecting the judgment of whether the structure of the impact table is reasonable. Summary of the Invention

[0005] To solve the technical problems of information fragmentation between feature maps caused by the above-mentioned grouped convolution and the inability to fully utilize the structural characteristics of the impact table, the present invention provides a method for extracting features of an impact table simulation image based on deep learning, including:

[0006] Classify the impact table simulation images using a convolutional neural network to identify abnormal impact table loads. During the training process of the convolutional neural network, perform grouped reconstruction on the grouped convolution module, including: using the input of the grouped convolution module as the input feature map, the output of the grouped convolution module as the output feature map, and several input feature maps corresponding to any output feature map as the target input feature map group; using any input feature map in the target input feature map group as the target feature map, determining the feature capture index of the target feature map according to the similarity between the target feature map and the symmetric feature map, where the symmetric feature map is the feature map obtained by rotating the target feature map around its center point; obtaining the distinguishability between any two input feature maps in the target input feature map group according to the difference between the feature capture indices of the two input feature maps and the difference between the information entropies; obtaining the reconstruction coefficient of the target input feature map group according to the standard deviation of the feature capture indices of the input feature maps in the target input feature map group, the mean of the distinguishabilities between the input feature maps, and the stage at which the neural network training is located; and performing grouped reconstruction on the input feature maps according to the reconstruction coefficient of the target input feature map group.

[0007] The present invention obtains the feature capture index through the structural symmetry characteristics of the impact table, strengthening the model's perception ability of the symmetry of the impact table load distribution; comprehensively considering the difference in feature expression ability and information richness difference between feature maps, obtaining the distinguishability between feature maps, which helps to divide complementary feature maps to prevent multiple feature maps in the input feature map group from expressing redundant information or missing key information; through the grouped reconstruction of the input feature map group, a dynamic adjustment grouping mechanism for feature maps is realized, taking into account the symmetry of the impact table structure and the feature complementary requirements of feature maps, overcoming the problem of information fragmentation between feature maps caused by fixed grouping of grouped convolution, improving the feature extraction ability of grouped convolution, making grouped convolution more adaptable to changes in the local structure of the impact table, thereby improving the accuracy of identifying abnormal impact table loads and the accuracy of judging the impact table structure.

[0008] Preferably, the similarity between the target feature map and the symmetric feature map is obtained according to the cosine similarity.

[0009] Preferably, the similarity between the target feature map and the symmetric feature map is obtained according to the Pearson correlation coefficient.

[0010] Preferably, the symmetric feature map is the feature map obtained by rotating the target feature map around its center point, including: using the feature map obtained by rotating the target feature map around its center point by 90 degrees, 180 degrees, or 270 degrees as the symmetric feature map of the target feature map.

[0011] Preferably, the method for obtaining the feature capture index of the target feature map is to map the similarity between the target feature map and the symmetric feature map to the range of [0, 1], and use the obtained result as the feature capture index of the target feature map.

[0012] The present invention utilizes the symmetric characteristics of the impact tabletop to evaluate the feature capture ability of the feature map, quantifies whether the feature map effectively captures the features related to the symmetric structure of the impact tabletop; makes the feature capture index of the feature map focus on the symmetry features of the impact tabletop, guides the model to learn the features related to the structure of the impact tabletop, reduces the attention to irrelevant features, and improves the pertinence and efficiency of feature extraction.

[0013] Preferably, the discrimination degree satisfies the relational expression: ; in the formula, is the discrimination degree between the feature map and the feature map in the target input feature map group, is the feature capture index of the feature map , is the feature capture index of the feature map , is the information entropy of the feature map , is the information entropy of the feature map , is and the maximum value among.

[0014] The present invention uses the discrimination degree to identify the feature redundancy or feature irrelevance of the feature maps within the target input feature map group, thereby supporting subsequent grouped reconstruction to achieve the complementarity of the features of the feature maps within the target input feature map group, making the feature maps within the target input feature map group neither redundant nor having a certain correlation; through the discrimination degree, the feature maps with repeated features extracted by the model can be separated, or the feature maps lacking a common structure can be reallocated, thereby optimizing the feature extraction ability of grouped convolution.

[0015] Preferably, the reconstruction coefficient of the target input feature map group satisfies the relational expression: ; in the formula, is the reconstruction coefficient of the th input feature map group of the grouped convolution module, is the standard deviation of the feature capture indices of all the feature maps in the th input feature map group, is the mean value of the discrimination degrees between the respective feature maps in the th input feature map group, is the number of input feature groups of the grouped convolution module, is the ratio between the number of trained iterations and the preset number of training iterations, is the exponential function with the natural constant as the base, is a hyperparameter.

[0016] Through the measurement in the training stage and the feature differences between input feature map groups, the present invention obtains the reconstruction coefficient, avoiding the problem of unstable parameter update caused by frequent recombination and reconstruction in the initial stage of neural network training. At the same time, in the later stage of training, grouped reconstruction is moderately carried out to optimize the feature map grouping, so that the grouped convolution improves the feature extraction ability of the impact table while retaining the advantage of reducing model parameters.

[0017] Preferably, the grouping and reconstruction of the input feature map according to the reconstruction coefficient of the target input feature map group includes: when the reconstruction coefficient of the target input feature map group is greater than the preset threshold, the input feature map with the smallest distinguishability between the target input feature map group and other input feature maps is removed from the target input feature map group, and the input feature map removed from the target input feature map group is added to the remaining input feature map groups.

[0018] Preferably, the grouping and reconstruction of the input feature map according to the reconstruction coefficient of the target input feature map group includes: when the reconstruction coefficient of the target input feature map group is greater than the preset threshold, any input feature map in the target input feature map group is removed from the target input feature map group, and the input feature map removed from the target input feature map group is added to the remaining input feature map groups.

[0019] Through grouped reconstruction, the present invention enables the model to adjust the interaction range of the convolutional kernel according to the distinguishability of the feature maps and the discrete state of the feature maps in the input feature map group, avoiding the extraction of duplicate features or irrelevant features, and enhancing the expression ability of complex features of the impact table.

[0020] Preferably, using the convolutional neural network to classify the impact table simulation image to identify the abnormal load of the impact table includes: using the convolutional neural network based on grouped convolution to classify the impact table simulation image. If the prediction result is 1, the impact table load is abnormal; if the prediction result is 0, the impact table load is normal.

[0021] The beneficial effects of the present invention are as follows: The present invention makes full use of the structural characteristics of the impact table simulation image, effectively solves the problem of fragmented feature map information existing in the prior art of grouped convolution through the method of grouped reconstruction, improves the accuracy and integrity of the feature extraction of the impact table by grouped convolution, and reduces the loss of key information; the present invention enhances the adaptability of the model to the structural characteristics of the impact table and dynamically responds to the diversity of the content of the impact table simulation image; the present invention optimizes the feature extraction effect in the neural network training process to make the feature extraction adapt to different requirements in different training stages; the present invention improves the recognition accuracy of abnormal loads on the impact table and provides a more reliable basis for judging the rationality of the impact table structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 FIG. is a flowchart schematically showing a method for feature extraction of an impact table simulation image based on deep learning in the present invention;

[0023] Figure 2 FIG. is a schematic diagram showing an impact table simulation image;

[0024] Figure 3 FIG. is a schematic structural diagram showing grouped convolution;

[0025] Figure 4 FIG. is a schematic structural diagram of grouped convolution before grouped reconstruction;

[0026] Figure 5 FIG. is a schematic structural diagram of grouped convolution after grouped reconstruction. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0028] Next, the specific embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0029] The embodiments of the present invention disclose a method for feature extraction of an impact table simulation image based on deep learning. Referring to Figure 1 , it includes steps S1 to S5:

[0030] S1. Obtain an impact table simulation image dataset.

[0031] Specifically, various structural impact tabletop simulation images are collected. During the collection process, the impact tabletop needs to be located at the center position of the image. The images in the dataset are labeled. That is, for each image in the dataset, if there is an area with abnormal load in the image, the image is labeled as 1; if there is no area with abnormal load in the image, the image is labeled as 0.

[0032] S2. Calculate the feature capture index of the feature map through the feature map and the symmetric feature map.

[0033] Specifically, the images in the impact tabletop simulation image dataset are input into a convolutional neural network for neural network training. The convolutional neural network contains a grouped convolution module, and the structure of the convolutional neural network is an image classification convolutional neural network, such as MobileNet.

[0034] Furthermore, the image obtained by convolving the impact tabletop simulation image is used as the feature map. At the same time, the output obtained by convolving the feature map is also the feature map. Taking any grouped convolution module in the neural network as the target grouped convolution module, and taking any input feature map of the target grouped convolution module as the target feature map, the target feature map is flattened into a one-dimensional vector to obtain the target vector; the target feature map is rotated around the center point to obtain the symmetric feature map, and the symmetric feature map is flattened into a one-dimensional vector to obtain the symmetric vector, where the rotation angle can be 90 degrees, 180 degrees, or 270 degrees.

[0035] Obtain the similarity between the target feature map and the symmetric feature map according to the target vector and the symmetric vector, and obtain the feature capture index of the target feature map according to the similarity between the target feature map and the symmetric feature map.

[0036] It should be noted that Figure 2 For the impact tabletop simulation image, in order to make the load on the impact tabletop relatively evenly distributed, the structure of the impact tabletop usually has a certain symmetry. The symmetric structure can evenly disperse the force generated by impact or vibration to each support point to avoid local stress concentration, thereby reducing the risk of structural deformation or damage. Since the symmetric structure in the impact tabletop should have similar loads, there should also be similar features at the symmetric positions in the impact tabletop simulation image. When the target feature map fully extracts a certain feature in the impact tabletop simulation image, the target feature map should also have a certain symmetry. Therefore, the present invention calculates the feature capture index of the feature map through the feature map and the symmetric feature map.

[0037] In one embodiment, the similarity between the target feature map and the symmetric feature map is the cosine similarity between the target vector and the symmetric vector, and the feature capture index of the target feature map is the value after normalizing the cosine similarity between the target vector and the symmetric vector, where the way of normalizing the cosine similarity is to add 1 to the value of the cosine similarity and then divide by 2.

[0038] In another embodiment, the similarity between the target feature map and the symmetric feature map is the Pearson correlation coefficient between the target vector and the symmetric vector, and the feature capture index of the target feature map is the value after normalizing the Pearson correlation coefficient between the target vector and the symmetric vector, where the way to normalize the Pearson correlation coefficient is to add 1 to the value of the Pearson correlation coefficient and then divide by 2.

[0039] It should be further noted that the greater the similarity between the target vector and the symmetric vector, the more similar the target feature map and the symmetric target map are, the more likely the target feature map has symmetric characteristics, the more likely the target feature map captures a certain feature of the impact table simulation image, and the greater the feature capture index of the target feature map; the smaller the similarity between the target vector and the symmetric vector, the greater the difference between the target feature map and the symmetric target map, the more likely the target feature map does not have symmetric characteristics, the more likely the target feature map does not fully capture the features of the impact table simulation image, and the smaller the feature capture index of the target feature map.

[0040] S3. Obtain the discrimination degree between feature maps according to the feature capture index of the feature map and the information entropy of the feature map.

[0041] It should be noted that when extracting image features through grouped convolution, the output feature map of the input feature map group is obtained by inputting the input feature map group into the convolution. If the feature maps in the input feature map group are too similar, it means that the information they extract is highly redundant and they only learn repetitive features. If the difference between the feature maps in the input feature map group is too large, it means that there is a lack of common relevant structure between them. Therefore, the feature maps in the input feature map group in the grouped convolution module should have sufficient complementarity, and there should be both a certain information difference and a certain relevant structure between the feature maps in the input feature map group. Therefore, the present invention obtains the discrimination degree between feature maps according to the feature capture index of the feature map and the information entropy of the feature map.

[0042] Specifically, taking the set of several input feature maps corresponding to any output feature map in the target grouped convolution module as the target input feature map group, calculate the information entropy of the feature maps in the target input feature map group, and calculate the discrimination degree between the feature maps through the difference in the feature capture index between the feature maps in the target input feature map group and the difference in the information entropy between the feature maps in the target input feature map group.

[0043] Exemplarily, Figure 3 is a schematic structural diagram of grouped convolution, Figure 3Among them, 1, 2, and 3 are output feature maps, 4, 5, and 6 are all input feature maps corresponding to 1, and the combination of 4, 5, and 6 is input feature map group 1; 7, 8, and 9 are all input feature maps corresponding to 2, and the combination of 7, 8, and 9 is input feature map group 2; 10, 11, and 12 are all input feature maps corresponding to 3, and the combination of 10, 11, and 12 is input feature map group 3.

[0044] Preferably, the distinguishability between the feature maps satisfies the relational expression: ; in the formula, is the distinguishability between the feature map and the feature map in the target input feature map group, is the feature capture index of the feature map , is the feature capture index of the feature map , is the information entropy of the feature map , is the information entropy of the feature map , is and the maximum value of.

[0045] Among them, represents the difference in the feature capture index between the feature map and the feature map . The larger this value is, the greater the difference in the integrity of feature capture between the two feature maps, and the greater the distinguishability between the feature map and the feature map ; the smaller this value is, the smaller the difference in the integrity of feature capture between the two feature maps, and the smaller the distinguishability between the feature map and the feature map ; represents the difference in the information entropy between the feature map and the feature map . The larger this value is, the greater the difference in the amount of information contained in the features extracted by the two feature maps, and the greater the distinguishability between the feature map and the feature map ; the smaller this value is, the smaller the difference in the amount of information contained in the features extracted by the two feature maps, and the smaller the distinguishability between the feature map and the feature map . For the convenience of subsequent calculations, in the formula, is divided by to normalize

[0046] S4. Obtain the reconstruction coefficient of the target input feature map group according to the distinguishability between the feature maps in the target input feature map group.

[0047] Specifically, calculate the standard deviation of the feature capture indices of each input feature map within each input feature map group corresponding to the target grouped convolution module, calculate the distinguishability between each input feature map within each input feature map group and other input feature maps within the same input feature map group, and obtain the reconstruction coefficient of each input feature map group based on the difference in the standard deviation of the capture indices between each input feature map group of the target grouped convolution module and other input feature map groups, the difference in the distinguishability between feature maps within the input feature map group and other input feature map groups, and the stage of neural network training.

[0048] It is easy to understand that, as Figure 3 shown, take the standard deviation of the capture indices of 4, 5, and 6 as the standard deviation of the capture indices of input feature map group 1 , take the standard deviation of the capture indices of 7, 8, and 9 as the standard deviation of the capture indices of input feature map group 2 , take the standard deviation of the capture indices of 10, 11, and 12 as the standard deviation of the capture indices of input feature map group 3 , for the difference in the standard deviation of the capture indices between input feature map group 1 and other input feature map groups is the difference between and

[0049] Furthermore, the distinguishability between feature maps within input feature map group 1 is the average of the distinguishability between 4 and 5, the distinguishability between 4 and 6, and the distinguishability between 5 and 6 , the distinguishability between feature maps within input feature map group 2 is the average of the distinguishability between 7 and 8, the distinguishability between 7 and 9, and the distinguishability between 8 and 9 , the distinguishability between feature maps within input feature map group 3 is the average of the distinguishability between 10 and 11, the distinguishability between 10 and 12, and the distinguishability between 11 and 12 , then the difference in the distinguishability between feature maps within input feature map group 1 and other input feature map groups is the difference between and

[0050] In one embodiment, the reconstruction coefficient of the target input feature map group satisfies the relational expression: ; where is the reconstruction coefficient of the th input feature map group of the grouped convolution module, is the standard deviation of the feature capture indices of all feature maps in the th input feature map group, is the average of the distinguishability between each feature map in the th input feature map group, is the number of input feature groups of the grouped convolution module, is the ratio between the number of trained iterations and the preset number of training iterations, is the exponential function with the natural constant as the base, is a hyperparameter. In this embodiment , the implementer can select according to the actual situation value.

[0051] Among them, represents the difference in distinguishability between the th input feature map group and other input feature map groups in the target convolution module. The larger this value, the more the differences between the feature maps within the th input feature map group deviate from the overall level of the differences between the feature maps within each input feature map group. The feature maps within the th input feature map group are more likely to be highly similar or highly distinguishable, and the more the th input feature map group should be reorganized, then the reconstruction coefficient of the th input feature map group is larger; The smaller it is, the closer the differences between the feature maps within the th input feature map group are to the overall level of the differences between the feature maps within each input feature map group. The distinguishability between the feature maps within the th input feature map group is closer to the distinguishability between the feature maps within each input feature map group, and the less the th input feature map group needs to be reorganized, and the reconstruction coefficient of the th input feature map group is smaller.

[0052] is the difference in capturing the exponential standard deviation between the feature maps within the th input feature map group and other input feature map groups, represents the difference between the discrete situation of the feature capture indices of the feature maps within the th input feature map group and the discrete situation of the feature capture indices of the feature maps within all input feature map groups. The larger this value, the more the discrete state of the feature capture indices of the feature maps within the th input feature map group deviates from the discrete state of the feature capture indices of each input feature map group, and the more the th input feature map group should be reorganized, then the reconstruction coefficient of the th input feature map group is larger; The smaller this value, the closer the discrete state of the feature capture indices of the feature maps within the th input feature map group is to the discrete state of the overall feature capture indices of each input feature map group, then the less the th input feature map group needs to be reorganized, then the The smaller the reconstruction coefficient of the input feature map group.

[0053] It should be noted that in the initial stage of neural network training, due to the random initialization of network parameters, there may be significant differences in the feature maps within the same group. To stabilize the parameter update of the neural network in the initial stage of neural network training, the reconstruction coefficients of each input feature map group need to be at a relatively low level to avoid unstable network parameter updates caused by frequent grouped reconstructions. As the network training and feature capture progress, grouped reconstruction should be appropriately carried out to improve the diversity of model feature expression. Therefore, by performing correction, when the training process is closer to the initial stage, the smaller is, the more likely is greater than 1. At this time, performs a downward correction on to make the reconstruction coefficient smaller, so as to avoid excessive grouped reconstruction of input feature map groups in the initial stage of training. When the training process is closer to the completion stage, the larger is, the more likely is less than 1. At this time,

[0054] S5. Perform grouped reconstruction according to the reconstruction coefficient, extract the features of the impact tabletop simulation image through the neural network after grouped reconstruction, and judge whether the structure of the impact tabletop is reasonable.

[0055] Specifically, during the training process of the neural network, calculate the reconstruction coefficients of each input feature map group in the grouped convolution module, and perform grouped reconstruction on the input feature map groups according to the reconstruction coefficients until the training of the neural network reaches the preset number of iterations.

[0056] In one embodiment, when the reconstruction coefficient of the target input feature map group is greater than the preset threshold, a randomly selected feature map within the target input feature map group is removed from the target input feature map group.

[0057] If the number of feature maps in the remaining input feature map groups is the same, the input feature map removed from the target input feature map group is added to a randomly selected input feature map group outside the target input feature map group. If the number of feature maps in the remaining input feature map groups is different, the input feature map removed from the target input feature map group is added to the feature map group with the least number of feature maps.

[0058] In another embodiment, when the reconstruction coefficient of the target input feature map group is greater than a preset threshold, the input feature map with the least distinguishability from other input feature maps within the target input feature map group is removed from the target input feature map group, and the input feature map removed from the target input feature map group is added to the input feature map group with the least number of feature maps among the remaining input feature map groups.

[0059] Exemplarily, the preset threshold can be 0.5, and the implementer can adjust the threshold according to the actual situation.

[0060] Exemplarily, Figure 4 is a schematic diagram of the grouped convolution structure before grouped reconstruction, Figure 5 is a schematic diagram of the grouped convolution structure after grouped reconstruction.

[0061] It should be noted that, Figure 5 The grouped convolution structure after grouped reconstruction shown in Figure 5 is only one case. In actual application, the grouped convolution structure after grouped reconstruction may not be limited to

[0062] shown.

[0063] Furthermore, taking the grouped state of the grouped convolution at the end of training as the grouped state of the grouped convolution in the final neural network, the impact tabletop simulation image is classified through the trained neural network, and then whether there is a load anomaly in the structure of the impact tabletop is judged according to the classification result, so as to judge the structural rationality of the impact tabletop.

[0063] For any impact tabletop simulation image to be detected, if the prediction result is 1, it means that the impact tabletop has a load anomaly and there may be an unreasonable structure on the impact tabletop. If the prediction result is 0, it means that the impact tabletop has a normal load and there may be no unreasonable structure on the impact tabletop.

Claims

1. A method for extracting features of a shock table simulation image based on deep learning, which uses a convolutional neural network to classify the shock table simulation image to identify abnormal shock table loads, is characterized in that, During the training process of a convolutional neural network, perform grouped reconstruction on the grouped convolution module, including: Use the input of the grouped convolution module as the input feature map, the output of the grouped convolution module as the output feature map, and several input feature maps corresponding to any output feature map as the target input feature map group; Use any input feature map in the target input feature map group as the target feature map, and determine the feature capture index of the target feature map according to the similarity between the target feature map and the symmetric feature map, where the symmetric feature map is the feature map obtained by rotating the target feature map around its center point; Obtain the distinguishability between any two input feature maps in the target input feature map group according to the difference between the feature capture indices of the two input feature maps and the difference between the information entropies, satisfying the relationship: , is the distinguishability between the feature map and the feature map in the target input feature map group, is the feature capture index of the feature map , is the feature capture index of the feature map , is the information entropy of the feature map , is the information entropy of the feature map , is and the maximum value among them; Obtain the reconstruction coefficient of the target input feature map group according to the standard deviation of the feature capture index of the input feature maps in the target input feature map group, the mean value of the distinguishability between the input feature maps, and the stage at which the neural network training is in, satisfying the relational expression: , is the reconstruction coefficient of the th input feature map group of the grouped convolution module, is the standard deviation of the feature capture indexes of all the feature maps in the th input feature map group, is the mean value of the distinguishability between each pair of feature maps in the th input feature map group, is the number of input feature groups of the grouped convolution module, is the ratio of the number of trained iterations to the preset number of training iterations, is the exponential function with the natural constant as the base, is a hyperparameter; Perform grouped reconstruction on the input feature map according to the reconstruction coefficient of the target input feature map group.

2. The method for extracting impact tabletop simulation image features based on deep learning according to claim 1, wherein The similarity between the target feature map and the symmetric feature map is obtained according to the cosine similarity.

3. The method for extracting impact tabletop simulation image features based on deep learning according to claim 1, wherein The similarity between the target feature map and the symmetric feature map is obtained according to the Pearson correlation coefficient.

4. The method for extracting impact tabletop simulation image features based on deep learning according to claim 1, wherein The symmetric feature map is the feature map obtained by rotating the target feature map around its center point, including: Use the feature map obtained by rotating the target feature map 90 degrees, 180 degrees, or 270 degrees around its center point as the symmetric feature map of the target feature map.

5. The method for extracting impact tabletop simulation image features based on deep learning according to claim 1, wherein The method for obtaining the feature capture index of the target feature map is: Map the similarity between the target feature map and the symmetric feature map to the range [0, 1], and use the obtained result as the feature capture index of the target feature map.

6. The method for extracting impact tabletop simulation image features based on deep learning according to claim 1, wherein, The performing grouped reconstruction on the input feature map according to the reconstruction coefficient of the target input feature map group includes: When the reconstruction coefficient of the target input feature map group is greater than the preset threshold, remove the input feature map with the smallest distinguishability from the other input feature maps in the target input feature map group, and add the input feature map removed from the target input feature map group to the remaining input feature map groups.

7. The method for extracting impact tabletop simulation image features based on deep learning according to claim 1, wherein The performing grouped reconstruction on the input feature map according to the reconstruction coefficient of the target input feature map group includes: When the reconstruction coefficient of the target input feature map group is greater than the preset threshold, remove any input feature map in the target input feature map group, and add the input feature map removed from the target input feature map group to the remaining input feature map groups.

8. The method for extracting impact tabletop simulation image features based on deep learning according to claim 1, characterized in that, Using a convolutional neural network to classify the impact tabletop simulation image to identify abnormal impact tabletop load, including: Use a convolutional neural network based on grouped convolution to classify the impact tabletop simulation image. If the prediction result is 1, the impact tabletop load is abnormal; if the prediction result is 0, the impact tabletop load is normal.

Images