Acoustic metamaterial performance analysis method and system for railway sound barriers

Abstract

The present invention relates to the field of material analysis technology, and in particular to a method and system for analyzing the performance of acoustic metamaterials for railway sound barriers. An acoustic metamaterial performance analysis system for railway sound barriers comprises an acoustic metamaterial X-ray projection image set acquisition module and an acoustic metamaterial performance analysis module. The present invention obtains X-ray projection images of acoustic metamaterials at different angles, and sends all the X-ray projection images of acoustic metamaterials at different angles to an acoustic metamaterial performance analysis model for processing, and outputs an acoustic metamaterial performance data set. Since the performance of acoustic metamaterials is mainly determined by the internal microstructure characteristics, the microstructure of the acoustic metamaterial is reflected by the X-ray projection images of the acoustic metamaterials at different angles, and then the microstructure characteristics inside the acoustic metamaterial are analyzed by the acoustic metamaterial performance analysis model, so as to realize automatic performance analysis of the acoustic metamaterial without the need for additional detection operations.

Description

Technical Field

[0001] The present invention relates to the technical field of material analysis, and in particular to a method and system for analyzing the performance of acoustic metamaterials for railway sound barriers. Background Art

[0002] Railway sound barriers are important facilities used to reduce the impact of noise generated during railway operation on the surrounding environment. In recent years, in order to improve the noise reduction performance of sound barriers, acoustic metamaterials have gradually been studied and applied due to their special structural design and excellent acoustic performance. However, the current analysis and evaluation of the performance of acoustic metamaterials mainly rely on traditional physical methods, such as building complex test platforms in acoustic laboratories to measure physical properties. These methods require complex equipment, cumbersome operating procedures, and are accompanied by certain environmental restrictions and cost issues, and are inconvenient to operate. Summary of the invention

[0003] The present invention obtains X-ray projection images of acoustic metamaterials at different angles, and sends all the X-ray projection images of acoustic metamaterials at different angles to the acoustic metamaterial performance analysis model for processing, and outputs an acoustic metamaterial performance data set. Since the performance of the acoustic metamaterial is mainly determined by the internal microstructure characteristics, the microstructure of the acoustic metamaterial is reflected by the X-ray projection images of the acoustic metamaterial at different angles, and then the microstructure characteristics inside the acoustic metamaterial are analyzed by the acoustic metamaterial performance analysis model to realize automatic performance analysis of the acoustic metamaterial without additional detection operations, which is convenient to operate; and in the process of using the acoustic metamaterial performance analysis model, the influence of redundant information is reduced by image alignment operations and self-attention mechanism operations, and the global feature information across angles is captured by graph structure feature extraction, and the self-attention mechanism operation is strengthened.

[0004] The present invention provides a method for analyzing the performance of an acoustic metamaterial for a railway sound barrier, comprising:

[0005] Acquire an acoustic metamaterial X-ray projection image set, where the acoustic metamaterial X-ray projection image set includes a plurality of acoustic metamaterial X-ray projection images at different angles;

[0006] Sending the acoustic metamaterial X-ray projection image set to the acoustic metamaterial performance analysis model for processing, and outputting an acoustic metamaterial performance data set, wherein the acoustic metamaterial performance data set includes acoustic metamaterial performance data corresponding to the acoustic metamaterial;

[0007] The acoustic metamaterial performance analysis model is established based on the Transformer model, and specifically includes a preprocessing layer, an image alignment layer, an image feature extraction layer, an image feature splicing layer, a graph structure feature extraction layer, an encoder layer, a decoder layer and an output layer. The preprocessing layer is used to perform a preprocessing operation on each acoustic metamaterial X-ray projection image in the acoustic metamaterial X-ray projection image set to construct a corresponding acoustic metamaterial X-ray projection image; the image alignment layer is used to perform an image alignment operation on all acoustic metamaterial X-ray projection images to construct a corresponding acoustic metamaterial X-ray projection standard image; the image feature extraction layer is used to perform a feature extraction operation on each acoustic metamaterial X-ray projection standard image to construct a corresponding acoustic metamaterial feature vector; the image feature splicing layer is used to rotate all acoustic metamaterial feature vectors from small to large The acoustic metamaterial feature map is spliced ​​from top to bottom based on the acoustic metamaterial feature map; the graph structure feature extraction layer is used to construct the acoustic metamaterial feature map structure data based on the acoustic metamaterial feature map, and perform feature extraction based on the acoustic metamaterial feature map structure data to construct the acoustic metamaterial graph structure features; the encoder layer includes six encoding blocks, and the encoder layer is used to receive the acoustic metamaterial feature map, and perform encoding operations through the self-attention mechanism, the decoder layer includes six decoding blocks, and the decoder layer is used to perform decoding operations through the self-attention mechanism, and the self-attention mechanism in the encoder layer and the decoder layer is based on the acoustic metamaterial graph structure features, and the acoustic metamaterial feature map is processed by the encoder layer and the decoder layer in turn to output the acoustic metamaterial performance analysis vector; the output layer is used to fully connect the acoustic metamaterial performance analysis vector to output the acoustic metamaterial performance data set.

[0008] As a preferred aspect, an image alignment layer is used to perform an image alignment operation on all acoustic metamaterial X-ray projection images to construct a corresponding acoustic metamaterial X-ray projection standard image, which specifically includes the following steps:

[0009] Traverse all acoustic metamaterial X-ray projection images, and for each acoustic metamaterial X-ray projection image, execute the following: send the selected acoustic metamaterial X-ray projection image to the built-in transformation matrix generation network in the image alignment layer for processing to construct a transformation matrix; then perform affine transformation and pixel value sampling on the selected acoustic metamaterial X-ray projection image through the transformation matrix to obtain the corresponding acoustic metamaterial X-ray projection standard image.

[0010] As a preferred aspect, a graph structure feature extraction layer is used to construct acoustic metamaterial feature graph structure data based on the acoustic metamaterial feature graph, and feature extraction is performed based on the acoustic metamaterial feature graph structure data to construct acoustic metamaterial graph structure features, specifically including the following steps:

[0011] Each acoustic metamaterial feature vector in the acoustic metamaterial feature graph is regarded as a feature node, and then two feature nodes are randomly selected from all feature nodes. For the two selected feature nodes, the following operations are performed: the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is calculated, and it is determined whether the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is greater than the similarity threshold. If the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is greater than the similarity threshold, a feature edge is constructed between the two selected feature nodes, and the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is used as the weight corresponding to the feature edge. If the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is not greater than the similarity threshold, two feature nodes are continuously randomly selected from all feature nodes; until any two feature nodes have been selected, the acoustic metamaterial feature graph structure data is constructed based on all feature nodes and feature edges;

[0012] Feature extraction is performed using the following formula: H(e+1)=σ(A b H(e)w(e)), where H(e+1) is the aggregate feature map output at the e+1th iteration, e=0, 1,…, E, E is the maximum number of iterations, H(e) is the aggregate feature map output at the eth iteration, H(0) is the acoustic metamaterial feature map, σ() is the activation function, and A b is the normalized adjacency matrix, and A b =D -0.5 AD -0.5 , D is a diagonal matrix, and the values ​​on the diagonal are the degrees corresponding to each feature node, A is the adjacency matrix, the values ​​in the adjacency matrix are the weights of the corresponding feature edges, w(e) is the graph structure weight matrix at the e-th iteration; and H(E+1) is recorded as the acoustic metamaterial graph structure feature.

[0013] As a preferred aspect, the self-attention mechanism in the encoder layer and the decoder layer is performed based on the structural features of the acoustic metamaterial graph, specifically comprising the following steps:

[0014] When the self-attention mechanism is executed in the encoder layer and the decoder layer, the input data is recorded as the input vector, and the output data is recorded as the output vector. The input vector is multiplied with the value weight matrix and the key weight matrix respectively to construct the corresponding value vector V and key vector K; the acoustic metamaterial graph structure feature is multiplied with the query weight matrix to construct the query vector Q; the output vector Z=softmax(QK T / d 0.5 ) V, T is the matrix transpose operation, and d is the dimension size of the key vector K.

[0015] As a preferred aspect, training the acoustic metamaterial performance analysis model specifically includes the following steps:

[0016] A plurality of training samples are obtained, wherein the training samples include an acoustic metamaterial X-ray projection image set corresponding to the same acoustic metamaterial, and the training samples are annotated through an acoustic metamaterial performance data set, and all annotated training samples are combined into a training set, and an acoustic metamaterial performance analysis model is trained through the training set, and a comprehensive loss value is calculated, wherein the comprehensive loss value includes a loss value corresponding to the acoustic material performance analysis and a loss value corresponding to the feature alignment, and it is determined whether the comprehensive loss value is within a preset range. If the comprehensive loss value is within the preset range, the trained acoustic metamaterial performance analysis model is output; otherwise, the acoustic metamaterial performance analysis model is continuously trained through the training set.

[0017] As a preferred aspect, when calculating the similarity between two acoustic metamaterial feature vectors corresponding to two selected feature nodes, a Euclidean distance calculation method is used.

[0018] The present invention also provides an acoustic metamaterial performance analysis system for railway sound barriers, comprising:

[0019] An acoustic metamaterial X-ray projection image set acquisition module is used to acquire an acoustic metamaterial X-ray projection image set, wherein the acoustic metamaterial X-ray projection image set includes a plurality of acoustic metamaterial X-ray projection images at different angles;

[0020] An acoustic metamaterial performance analysis module, used for sending the acoustic metamaterial X-ray projection image set to the acoustic metamaterial performance analysis model for processing, and outputting an acoustic metamaterial performance data set, wherein the acoustic metamaterial performance data set includes acoustic metamaterial performance data corresponding to the acoustic metamaterial;

[0021] The acoustic metamaterial performance analysis model is established based on the Transformer model, and specifically includes a preprocessing layer, an image alignment layer, an image feature extraction layer, an image feature splicing layer, a graph structure feature extraction layer, an encoder layer, a decoder layer and an output layer. The preprocessing layer is used to perform a preprocessing operation on each acoustic metamaterial X-ray projection image in the acoustic metamaterial X-ray projection image set to construct a corresponding acoustic metamaterial X-ray projection image; the image alignment layer is used to perform an image alignment operation on all acoustic metamaterial X-ray projection images to construct a corresponding acoustic metamaterial X-ray projection standard image; the image feature extraction layer is used to perform a feature extraction operation on each acoustic metamaterial X-ray projection standard image to construct a corresponding acoustic metamaterial feature vector; the image feature splicing layer is used to rotate all acoustic metamaterial feature vectors from small to large The acoustic metamaterial feature map is spliced ​​from top to bottom based on the acoustic metamaterial feature map; the graph structure feature extraction layer is used to construct the acoustic metamaterial feature map structure data based on the acoustic metamaterial feature map, and perform feature extraction based on the acoustic metamaterial feature map structure data to construct the acoustic metamaterial graph structure features; the encoder layer includes six encoding blocks, and the encoder layer is used to receive the acoustic metamaterial feature map, and perform encoding operations through the self-attention mechanism, the decoder layer includes six decoding blocks, and the decoder layer is used to perform decoding operations through the self-attention mechanism, and the self-attention mechanism in the encoder layer and the decoder layer is based on the acoustic metamaterial graph structure features, and the acoustic metamaterial feature map is processed by the encoder layer and the decoder layer in turn to output the acoustic metamaterial performance analysis vector; the output layer is used to fully connect the acoustic metamaterial performance analysis vector to output the acoustic metamaterial performance data set.

[0022] The present invention has the following advantages:

[0023] The present invention obtains X-ray projection images of acoustic metamaterials at different angles, and sends all the X-ray projection images of acoustic metamaterials at different angles to the acoustic metamaterial performance analysis model for processing, and outputs an acoustic metamaterial performance data set. Since the performance of the acoustic metamaterial is mainly determined by the internal microstructure characteristics, the microstructure of the acoustic metamaterial is reflected by the X-ray projection images of the acoustic metamaterial at different angles, and then the microstructure characteristics inside the acoustic metamaterial are analyzed by the acoustic metamaterial performance analysis model to realize automatic performance analysis of the acoustic metamaterial without additional detection operations; and in the process of using the acoustic metamaterial performance analysis model, the influence of redundant information is reduced by image alignment operations and self-attention mechanism operations, and the global feature information across angles is captured by graph structure feature extraction, and the self-attention mechanism operation is strengthened. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the structure of the acoustic metamaterial performance analysis model adopted in an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the structure of the acoustic metamaterial performance analysis system for railway sound barriers used in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0027] Embodiment 1, a method for analyzing the performance of an acoustic metamaterial for a railway sound barrier, comprising:

[0028] Acquiring an acoustic metamaterial X-ray projection image set, which includes several acoustic metamaterial X-ray projection images at different angles. It should be noted that when performing performance analysis on acoustic metamaterials used for railway sound barriers, the acoustic metamaterials will be photographed by an X-ray imager, and during the photographing process, the acoustic metamaterial to be analyzed will be placed on a rotating platform, and the rotating platform will rotate at a certain angle (for example, 1° or 0.5°). After each rotation, the acoustic metamaterial will be photographed by the X-ray imager to obtain an acoustic metamaterial X-ray projection image, and then all the acoustic metamaterial X-ray projection images at different angles will be combined into an acoustic metamaterial X-ray projection image set;

[0029] The acoustic metamaterial X-ray projection image set is sent to the acoustic metamaterial performance analysis model for processing, and an acoustic metamaterial performance data set is output. The acoustic metamaterial performance data set includes acoustic metamaterial performance data corresponding to the acoustic metamaterial, such as transmission loss, sound insulation and sound absorption coefficient, etc., to complete the performance analysis of the acoustic metamaterial and provide a reference for the quality inspection and use of the acoustic metamaterial;

[0030] See also Figure 1The acoustic metamaterial performance analysis model is established based on the Transformer model, which specifically includes a preprocessing layer, an image alignment layer, an image feature extraction layer, an image feature splicing layer, a graph structure feature extraction layer, an encoder layer, a decoder layer and an output layer. The preprocessing layer is used to perform preprocessing operations on each acoustic metamaterial X-ray projection image in the acoustic metamaterial X-ray projection image set, such as pixel value normalization, which can eliminate the intensity difference between pixel values ​​and construct the corresponding acoustic metamaterial X-ray projection image; the image alignment layer is used to perform image alignment operations on all acoustic metamaterial X-ray projection images to ensure that the acoustic metamaterial X-ray projection images at different angles are consistent. The geometric center of the image is consistent, avoiding feature loss due to the rotation center of the acoustic metamaterial, reducing feature noise caused by geometric deviation, so as to construct the corresponding acoustic metamaterial X-ray projection standard image; the image feature extraction layer is established based on the CNN network (such as VGG16), which is used to perform feature extraction operations on each acoustic metamaterial X-ray projection standard image, and can make full use of the geometric and density information in the acoustic metamaterial X-ray projection image, extract features related to acoustic performance, and construct the corresponding acoustic metamaterial feature vector; the image feature splicing layer is used to splice all acoustic metamaterial feature vectors from top to bottom according to the rotation angle from small to large. , construct an acoustic metamaterial feature map; the graph structure feature extraction layer is used to construct acoustic metamaterial feature map structure data based on the acoustic metamaterial feature map, and perform feature extraction based on the acoustic metamaterial feature map structure data, which can realize the learning of acoustic metamaterial feature vectors at different angles, and capture global feature information across angles, thereby improving the utilization efficiency of multi-angle acoustic metamaterial X-ray projection images, thereby improving the accuracy of subsequent performance analysis, in order to construct acoustic metamaterial graph structure features; the encoder layer includes six encoding blocks, and the encoder layer is used to receive the acoustic metamaterial feature map and perform encoding operations through the self-attention mechanism, the decoder layer includes six decoding blocks, and the decoder The encoder layer is used to perform decoding operations through the self-attention mechanism. It should be noted that the encoder layer and the decoder layer here are both set with reference to the Transformer model. The self-attention mechanism can capture the interactive relationship between different angles, and then perform different degrees of feature enhancement for different features to further reduce the impact of redundant information. The self-attention mechanism in the encoder layer and the decoder layer is based on the structural characteristics of the acoustic metamaterial graph. The acoustic metamaterial feature graph is processed by the encoder layer and the decoder layer in turn to output the acoustic metamaterial performance analysis vector; the output layer is used to fully connect the acoustic metamaterial performance analysis vector and output the acoustic metamaterial performance data set.

[0031] The present application obtains X-ray projection images of acoustic metamaterials at different angles, and sends all the X-ray projection images of acoustic metamaterials at different angles to the acoustic metamaterial performance analysis model for processing, and outputs an acoustic metamaterial performance data set. Since the performance of acoustic metamaterials is mainly determined by the internal microstructure characteristics, the microstructure of the acoustic metamaterial is reflected by the X-ray projection images of the acoustic metamaterials at different angles, and then the microstructure characteristics inside the acoustic metamaterial are analyzed by the acoustic metamaterial performance analysis model to realize automatic performance analysis of the acoustic metamaterial without additional detection operations; and in the process of using the acoustic metamaterial performance analysis model, the influence of redundant information is reduced by image alignment operations and self-attention mechanism operations, and the global feature information across angles is captured by graph structure feature extraction, and the self-attention mechanism operation is strengthened.

[0032] The image alignment layer is used to perform image alignment operations on all acoustic metamaterial X-ray projection images to construct corresponding acoustic metamaterial X-ray projection standard images, which specifically includes the following steps:

[0033] Traverse all acoustic metamaterial X-ray projection images, and for each acoustic metamaterial X-ray projection image, execute the following contents: send the selected acoustic metamaterial X-ray projection image to the built-in transformation matrix generation network in the image alignment layer for processing. The transformation matrix generation network is established with reference to the convolutional neural network, and generally includes a convolution layer, a pooling layer and two fully connected layers to construct a transformation matrix; then perform affine transformation and pixel value sampling on the selected acoustic metamaterial X-ray projection image through the transformation matrix to obtain the corresponding acoustic metamaterial X-ray projection standard image. It should be noted that when performing affine transformation and pixel value sampling on the selected acoustic metamaterial X-ray projection image through the transformation matrix, a target matrix of a corresponding size is first set according to the selected acoustic metamaterial X-ray projection image, and then the coordinates of the target matrix are mapped to the selected acoustic metamaterial X-ray projection image through the transformation matrix for sampling. Since the coordinates processed by the transformation matrix are not necessarily integer values, they also need to be sampled by bilinear interpolation; the parameters in the transformation matrix generation network are adjusted with the overall training of the acoustic metamaterial performance analysis model;

[0034] The graph structure feature extraction layer is used to construct acoustic metamaterial feature graph structure data based on the acoustic metamaterial feature graph, and feature extraction is performed based on the acoustic metamaterial feature graph structure data to construct acoustic metamaterial graph structure features, specifically including the following steps:

[0035] Each acoustic metamaterial feature vector in the acoustic metamaterial feature graph is regarded as a feature node, and then two feature nodes are randomly selected from all feature nodes. For the two selected feature nodes, the following operations are performed: the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is calculated. The similarity calculation method can be used to calculate the Euclidean distance to determine whether the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is greater than the similarity threshold. The similarity threshold is set manually. If the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is greater than the similarity threshold, the two selected feature nodes are selected. A feature edge is constructed between the points, and the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is used as the weight corresponding to the feature edge. If the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is not greater than the similarity threshold, two feature nodes are randomly selected from all feature nodes; until any two feature nodes have been selected, the acoustic metamaterial feature graph structure data is constructed based on all feature nodes and feature edges; the construction of feature edges is controlled by the similarity threshold, which can effectively control the sparsity of graph structure data, reduce the computational complexity, and provide a high-quality foundation for the construction of global features;

[0036] Feature extraction is performed using the following formula: H(e+1)=σ(A b H(e)w(e)), where H(e+1) is the aggregate feature map output at the e+1th iteration, e=0, 1,…, E, E is the maximum number of iterations, generally 3, H(e) is the aggregate feature map output at the eth iteration, H(0) is the acoustic metamaterial feature map, σ() is the activation function, generally the sigmoid activation function is used, A b is the normalized adjacency matrix, and A b =D -0.5 AD -0.5 , D is a diagonal matrix, and the values ​​on the diagonal are the degrees corresponding to each feature node, that is, the number of feature nodes connected to the feature node, A is the adjacency matrix, and the values ​​in the adjacency matrix are the weights of the corresponding feature edges. If the feature nodes are not connected, the weight is recorded as 0, w(e) is the graph structure weight matrix at the e-th iteration, which is adjusted following the overall training of the acoustic metamaterial performance analysis model; and H(E+1) is recorded as the acoustic metamaterial graph structure feature.

[0037] The self-attention mechanism in the encoder layer and the decoder layer is based on the structural features of the acoustic metamaterial graph, which specifically includes the following steps:

[0038] When the self-attention mechanism is executed in the encoder layer and the decoder layer, the input data is recorded as the input vector, and the output data is recorded as the output vector. The input vector is multiplied by the value weight matrix and the key weight matrix respectively to construct the corresponding value vector V and key vector K; the acoustic metamaterial graph structure feature is multiplied by the query weight matrix to construct the query vector Q. It should be noted that the value weight matrix, key weight matrix and query weight matrix are all set with reference to the Transformer model and adjusted following the overall training of the acoustic metamaterial performance analysis model; the output vector Z=softmax(QK T / d 0.5 ) V, T is the matrix transpose operation, and d is the dimension size of the key vector K.

[0039] Training the acoustic metamaterial performance analysis model includes the following steps:

[0040] A number of training samples are obtained, wherein the training samples include an acoustic metamaterial X-ray projection image set corresponding to the same acoustic metamaterial, and the training samples are annotated through an acoustic metamaterial performance data set. It should be noted that the acoustic metamaterial performance data set here is data obtained by an operator through actual measurement or through finite element analysis. All annotated training samples are composed of a training set, and the acoustic metamaterial performance analysis model is trained through the training set to calculate a comprehensive loss value, which includes a loss value corresponding to the acoustic material performance analysis and a loss value corresponding to the feature alignment, wherein the loss value corresponding to the acoustic material performance analysis is the mean square error between the acoustic metamaterial performance data set predicted by the acoustic metamaterial performance analysis model and the annotated acoustic metamaterial performance data set, the loss value corresponding to the feature alignment is the loss value in the feature alignment process, and the loss value corresponding to the feature alignment is the loss value in the feature alignment process. A method for calculating the loss value is: obtaining all acoustic metamaterial X-ray projection standard images, and then calculating the feature center corresponding to the acoustic metamaterial X-ray projection standard image through the centroid of all pixels in the acoustic metamaterial X-ray projection standard image, and performing cluster analysis on the feature centers corresponding to all acoustic metamaterial X-ray projection standard images to obtain the cluster center, and calculating the mean square error between the cluster center and the feature center corresponding to all acoustic metamaterial X-ray projection standard images, which is the loss value corresponding to the feature alignment. The loss value corresponding to the feature alignment represents the feature noise caused by the geometric deviation, and determines whether the comprehensive loss value is within a preset range. The preset range is set artificially. If the comprehensive loss value is within the preset range, the trained acoustic metamaterial performance analysis model is output; otherwise, the acoustic metamaterial performance analysis model is continuously trained through the training set.

[0041] Example 2, an acoustic metamaterial performance analysis system for railway sound barriers, such as Figure 2 As shown, including:

[0042] An acoustic metamaterial X-ray projection image set acquisition module is used to acquire an acoustic metamaterial X-ray projection image set, which includes several acoustic metamaterial X-ray projection images at different angles. It should be noted that when the performance of the acoustic metamaterial used for the railway sound barrier is analyzed, the acoustic metamaterial will be photographed by an X-ray imager, and during the photographing process, the acoustic metamaterial to be analyzed will be placed on a rotating platform, and the rotating platform will rotate at a certain angle (for example, 1° or 0.5°). After each rotation, the acoustic metamaterial will be photographed by the X-ray imager to obtain an acoustic metamaterial X-ray projection image, and then all the acoustic metamaterial X-ray projection images at different angles will be combined into an acoustic metamaterial X-ray projection image set;

[0043] The acoustic metamaterial performance analysis module is used to send the acoustic metamaterial X-ray projection image set to the acoustic metamaterial performance analysis model for processing, and output the acoustic metamaterial performance data set. The acoustic metamaterial performance data set includes the acoustic metamaterial performance data corresponding to the acoustic metamaterial, such as transmission loss, sound insulation and sound absorption coefficient, so as to complete the performance analysis of the acoustic metamaterial and provide a reference for the quality inspection and use of the acoustic metamaterial;

[0044] The acoustic metamaterial performance analysis model is established based on the Transformer model, which specifically includes a preprocessing layer, an image alignment layer, an image feature extraction layer, an image feature splicing layer, a graph structure feature extraction layer, an encoder layer, a decoder layer, and an output layer. The preprocessing layer is used to perform preprocessing operations on each acoustic metamaterial X-ray projection image in the acoustic metamaterial X-ray projection image set, such as pixel value normalization, which can eliminate the intensity difference between pixel values ​​and construct the corresponding acoustic metamaterial X-ray projection image; the image alignment layer is used to perform image alignment operations on all acoustic metamaterial X-ray projection images to ensure that the acoustic metamaterial X-ray projection images at different angles are aligned. The geometric center of the image is consistent, avoiding feature loss due to the rotation center of the acoustic metamaterial, reducing feature noise caused by geometric deviation, so as to construct the corresponding acoustic metamaterial X-ray projection standard image; the image feature extraction layer is established based on the CNN network (such as VGG16), which is used to perform feature extraction operations on each acoustic metamaterial X-ray projection standard image, and can make full use of the geometric and density information in the acoustic metamaterial X-ray projection image, extract features related to acoustic performance, and construct the corresponding acoustic metamaterial feature vector; the image feature splicing layer is used to splice all acoustic metamaterial feature vectors from top to bottom according to the rotation angle from small to large. , construct an acoustic metamaterial feature map; the graph structure feature extraction layer is used to construct acoustic metamaterial feature map structure data based on the acoustic metamaterial feature map, and perform feature extraction based on the acoustic metamaterial feature map structure data, which can realize the learning of acoustic metamaterial feature vectors at different angles, and capture global feature information across angles, thereby improving the utilization efficiency of multi-angle acoustic metamaterial X-ray projection images, thereby improving the accuracy of subsequent performance analysis, in order to construct acoustic metamaterial graph structure features; the encoder layer includes six encoding blocks, and the encoder layer is used to receive the acoustic metamaterial feature map and perform encoding operations through the self-attention mechanism, the decoder layer includes six decoding blocks, and the decoder The encoder layer is used to perform decoding operations through the self-attention mechanism. It should be noted that the encoder layer and the decoder layer here are both set with reference to the Transformer model. The self-attention mechanism can capture the interactive relationship between different angles, and then perform different degrees of feature enhancement for different features to further reduce the impact of redundant information. The self-attention mechanism in the encoder layer and the decoder layer is based on the structural characteristics of the acoustic metamaterial graph. The acoustic metamaterial feature graph is processed by the encoder layer and the decoder layer in turn to output the acoustic metamaterial performance analysis vector; the output layer is used to fully connect the acoustic metamaterial performance analysis vector and output the acoustic metamaterial performance data set.

[0045] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A method for analyzing the performance of acoustic metamaterials for railway sound barriers, characterized in that: include: Acquire an acoustic metamaterial X-ray projection image set, where the acoustic metamaterial X-ray projection image set includes a plurality of acoustic metamaterial X-ray projection images at different angles; Sending the acoustic metamaterial X-ray projection image set to the acoustic metamaterial performance analysis model for processing, and outputting an acoustic metamaterial performance data set, wherein the acoustic metamaterial performance data set includes acoustic metamaterial performance data corresponding to the acoustic metamaterial; The acoustic metamaterial performance analysis model is established based on the Transformer model, and specifically includes a preprocessing layer, an image alignment layer, an image feature extraction layer, an image feature splicing layer, a graph structure feature extraction layer, an encoder layer, a decoder layer, and an output layer. The preprocessing layer is used to perform a preprocessing operation on each acoustic metamaterial X-ray projection image in the acoustic metamaterial X-ray projection image set to construct a corresponding acoustic metamaterial X-ray projection image; the image alignment layer is used to perform an image alignment operation on all acoustic metamaterial X-ray projection images to construct a corresponding acoustic metamaterial X-ray projection standard image; The image feature extraction layer is used to perform feature extraction operations on each acoustic metamaterial X-ray projection standard image to construct a corresponding acoustic metamaterial feature vector; The image feature concatenation layer is used to concatenate all acoustic metamaterial feature vectors from top to bottom according to the rotation angles from small to large to construct an acoustic metamaterial feature map; The graph structure feature extraction layer is used to construct acoustic metamaterial feature graph structure data based on the acoustic metamaterial feature graph, and perform feature extraction based on the acoustic metamaterial feature graph structure data to construct acoustic metamaterial graph structure features; the encoder layer includes six encoding blocks, and the encoder layer is used to receive the acoustic metamaterial feature graph, and perform encoding operations through a self-attention mechanism, the decoder layer includes six decoding blocks, and the decoder layer is used to perform decoding operations through a self-attention mechanism, and the self-attention mechanism in the encoder layer and the decoder layer is based on the acoustic metamaterial graph structure features, and the acoustic metamaterial feature graph is processed by the encoder layer and the decoder layer in turn to output an acoustic metamaterial performance analysis vector; the output layer is used to perform a full connection operation on the acoustic metamaterial performance analysis vector, and output an acoustic metamaterial performance data set; The image alignment layer is used to perform image alignment operations on all acoustic metamaterial X-ray projection images to construct corresponding acoustic metamaterial X-ray projection standard images, which specifically includes the following steps: Traverse all acoustic metamaterial X-ray projection images, and for each acoustic metamaterial X-ray projection image, perform the following: send the selected acoustic metamaterial X-ray projection image to the built-in transformation matrix generation network in the image alignment layer for processing to construct a transformation matrix; then perform affine transformation and pixel value sampling on the selected acoustic metamaterial X-ray projection image through the transformation matrix to obtain the corresponding acoustic metamaterial X-ray projection standard image; The graph structure feature extraction layer is used to construct acoustic metamaterial feature graph structure data based on the acoustic metamaterial feature graph, and feature extraction is performed based on the acoustic metamaterial feature graph structure data to construct acoustic metamaterial graph structure features, specifically including the following steps: Each acoustic metamaterial feature vector in the acoustic metamaterial feature graph is regarded as a feature node, and then two feature nodes are randomly selected from all feature nodes. For the two selected feature nodes, the following operations are performed: the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is calculated, and it is determined whether the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is greater than the similarity threshold. If the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is greater than the similarity threshold, a feature edge is constructed between the two selected feature nodes, and the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is used as the weight corresponding to the feature edge. If the similarity between the two acoustic metamaterial feature vectors corresponding to the two selected feature nodes is not greater than the similarity threshold, two feature nodes are continuously randomly selected from all feature nodes; until any two feature nodes have been selected, the acoustic metamaterial feature graph structure data is constructed based on all feature nodes and feature edges; Feature extraction is performed using the following formula: H(e+1)=σ(A b H(e)w(e)), where H(e+1) is the aggregate feature map output at the e+1th iteration, e=0, 1,…, E, E is the maximum number of iterations, H(e) is the aggregate feature map output at the eth iteration, H(0) is the acoustic metamaterial feature map, σ() is the activation function, and A b is the normalized adjacency matrix, and A b =D -0.5 AD -0.5 , D is a diagonal matrix, and the values ​​on the diagonal are the degrees corresponding to each feature node, A is the adjacency matrix, the values ​​in the adjacency matrix are the weights of the corresponding feature edges, and w(e) is the graph structure weight matrix at the e-th iteration; And H(E+1) is recorded as the structural feature of the acoustic metamaterial graph; The self-attention mechanism in the encoder layer and the decoder layer is based on the structural features of the acoustic metamaterial graph, which specifically includes the following steps: When the self-attention mechanism is executed in the encoder layer and the decoder layer, the input data is recorded as the input vector, and the output data is recorded as the output vector. The input vector is multiplied with the value weight matrix and the key weight matrix respectively to construct the corresponding value vector V and key vector K; the acoustic metamaterial graph structure feature is multiplied with the query weight matrix to construct the query vector Q; the output vector Z=softmax(QK T / d 0.5 ) V, T is the matrix transpose operation, and d is the dimension size of the key vector K.

2. The method for analyzing the performance of acoustic metamaterials for railway sound barriers according to claim 1, characterized in that: Training the acoustic metamaterial performance analysis model includes the following steps: A plurality of training samples are obtained, wherein the training samples include an acoustic metamaterial X-ray projection image set corresponding to the same acoustic metamaterial, and the training samples are annotated through an acoustic metamaterial performance data set, and all annotated training samples are combined into a training set, and an acoustic metamaterial performance analysis model is trained through the training set, and a comprehensive loss value is calculated, wherein the comprehensive loss value includes a loss value corresponding to the acoustic material performance analysis and a loss value corresponding to the feature alignment, and it is determined whether the comprehensive loss value is within a preset range. If the comprehensive loss value is within the preset range, the trained acoustic metamaterial performance analysis model is output; otherwise, the acoustic metamaterial performance analysis model is continuously trained through the training set.

3. The method for analyzing the performance of acoustic metamaterials for railway sound barriers according to claim 2, characterized in that: When calculating the similarity between two acoustic metamaterial feature vectors corresponding to two selected feature nodes, the Euclidean distance calculation method is used.

4. An acoustic metamaterial performance analysis system for railway sound barriers, characterized in that: The system applies the acoustic metamaterial performance analysis method for railway sound barriers according to any one of claims 1 to 3, including: An acoustic metamaterial X-ray projection image set acquisition module is used to acquire an acoustic metamaterial X-ray projection image set, wherein the acoustic metamaterial X-ray projection image set includes a plurality of acoustic metamaterial X-ray projection images at different angles; An acoustic metamaterial performance analysis module, used for sending the acoustic metamaterial X-ray projection image set to the acoustic metamaterial performance analysis model for processing, and outputting an acoustic metamaterial performance data set, wherein the acoustic metamaterial performance data set includes acoustic metamaterial performance data corresponding to the acoustic metamaterial; The acoustic metamaterial performance analysis model is established based on the Transformer model, and specifically includes a preprocessing layer, an image alignment layer, an image feature extraction layer, an image feature splicing layer, a graph structure feature extraction layer, an encoder layer, a decoder layer and an output layer. The preprocessing layer is used to perform a preprocessing operation on each acoustic metamaterial X-ray projection image in the acoustic metamaterial X-ray projection image set to construct a corresponding acoustic metamaterial X-ray projection image; the image alignment layer is used to perform an image alignment operation on all acoustic metamaterial X-ray projection images to construct a corresponding acoustic metamaterial X-ray projection standard image; the image feature extraction layer is used to perform a feature extraction operation on each acoustic metamaterial X-ray projection standard image to construct a corresponding acoustic metamaterial feature vector; the image feature splicing layer is used to rotate all acoustic metamaterial feature vectors from small to large The acoustic metamaterial feature map is spliced ​​from top to bottom based on the acoustic metamaterial feature map; the graph structure feature extraction layer is used to construct the acoustic metamaterial feature map structure data based on the acoustic metamaterial feature map, and perform feature extraction based on the acoustic metamaterial feature map structure data to construct the acoustic metamaterial graph structure features; the encoder layer includes six encoding blocks, and the encoder layer is used to receive the acoustic metamaterial feature map, and perform encoding operations through the self-attention mechanism, the decoder layer includes six decoding blocks, and the decoder layer is used to perform decoding operations through the self-attention mechanism, and the self-attention mechanism in the encoder layer and the decoder layer is based on the acoustic metamaterial graph structure features, and the acoustic metamaterial feature map is processed by the encoder layer and the decoder layer in turn to output the acoustic metamaterial performance analysis vector; the output layer is used to fully connect the acoustic metamaterial performance analysis vector to output the acoustic metamaterial performance data set.

Images