Local sound field control method based on self-defined loss function and multi-feature constraint

Abstract

The application provides a local sound field regulation method based on a self-defined loss function and multi-feature constraints, comprising the following steps: determining the geometric features of an acoustic metasurface structure, designing structure parameter variables, laying the metasurface on a backing plate, and constructing a forward simulation model of the metasurface; performing finite element analysis by the forward simulation model to obtain the far-field sound pressure level intensity distribution corresponding to the input metasurface physical structure parameters, and taking the metasurface physical structure parameters and the corresponding far-field sound pressure level intensity data as sample data; extracting multi-dimensional feature information and local sound field intensity information of the far-field sound pressure level intensity distribution as input, and constructing a multi-feature constrained deep neural network model to predict the metasurface physical structure parameters; comparing the far-field sound pressure level intensity distribution obtained by forward simulation of the metasurface physical structure parameter prediction value with the local target sound field intensity, verifying the accuracy of the deep neural network model prediction result, and then realizing effective regulation of the local sound field.

Description

Technical Field

[0001] This invention relates to the field of acoustic metasurface technology, and in particular to a method for local acoustic field manipulation based on a custom loss function and multi-feature constraints. Background Technology

[0002] Acoustic metamaterials are artificially designed materials with properties not found in natural materials, such as negative reflection and planar focusing. They can be used in fields such as sonar stealth for underwater ships and non-destructive testing in industry.

[0003] Traditional metasurface design methods rely on experienced experts, resulting in low design efficiency and complex fabrication processes. In recent years, with the development of deep learning and big data technologies, they have played a significant role in the design of acoustic metasurfaces. This has made it possible to construct the mapping relationship between acoustic metasurface structural parameters and sound fields by designing or optimizing deep neural network models, thereby deriving the metasurface structural parameters from the sound field. However, deep learning methods suffer from high computational cost and long processing time.

[0004] Therefore, it is necessary to provide a novel local sound field control method based on a custom loss function and multi-feature constraints to overcome the above-mentioned shortcomings. Summary of the Invention

[0005] The purpose of this invention is to provide a local sound field control method based on a custom loss function and multi-feature constraints. By introducing a kernel loss function and extracting sound field feature information when designing the network, the input and computation of the network model can be reduced, providing a new approach for the control of local sound fields.

[0006] To achieve the above objectives, this invention provides a local sound field modulation method based on a custom loss function and multi-feature constraints, comprising the following steps:

[0007] S1: Determine the geometric characteristics of the acoustic metasurface structure and design the structural parameter variables; lay the metasurface on the backing plate and construct the forward simulation model of the metasurface;

[0008] S2: The far-field sound pressure level intensity distribution corresponding to the input metasurface property structural parameters is obtained by finite element analysis of the forward simulation model, and the metasurface property structural parameters and the corresponding far-field sound pressure level intensity data are used as sample data.

[0009] S3: Extract multi-dimensional feature information of far-field sound pressure level intensity distribution and local sound field intensity information as input, and construct a deep neural network model with multi-feature constraints to predict the metasurface material property structure parameters;

[0010] S4: By comparing the far-field sound pressure level intensity distribution obtained by forward modeling the predicted values ​​of metasurface property structural parameters with the local target sound field intensity, the accuracy of the prediction results of the deep neural network model is verified, thereby realizing the effective control of the local sound field.

[0011] Preferably, step S1 further includes step S11: determining the normal thickness H and length L of the acoustic metasurface, discretizing the metasurface into 25 structural units with the same geometric shape and size, and the material density ρ of each unit. i They are all different and ρ i The value of ρ1 varies linearly with the position of the structural unit in the length direction, i = 1, 2, ..., 25. The material density ρ1 of the first metasurface structural unit and the material density gradient g of the two adjacent metasurface structural units are used as the material parameters to be designed, and the range of the density ρ1 of the first metasurface structural unit is limited to [1 / 3, 3].

[0012] Preferably, the input to the forward simulation in step S2 is the physical and structural parameters of the metasurface, and the output is the far-field sound pressure level intensity distribution P. The metasurface is composed of subwavelength structural units. With other parameters such as the metasurface length L and normal thickness H fixed, the material density ρ of the metasurface structural units is changed. i To achieve control over the reflected sound field;

[0013] Preferably, in step S2, COMSOL Multiphysics simulation software is used to perform finite element analysis to establish a one-to-one mapping relationship from metasurface material property parameters to far-field sound pressure intensity distribution. A large amount of sample data is generated by forward modeling to prepare sample datasets for training and testing deep neural network models.

[0014] Preferably, the multi-feature constrained deep neural network model in step S3 includes two sub-neural network models, which are trained on the deep neural network by taking the global feature information and local intensity information of far-field sound pressure as inputs, respectively.

[0015] Preferably, the sound field in the range of 0° to 180° is divided into 6 equally spaced intervals. The peak, trough values ​​and angles of each interval are taken. Two sound field intensity values ​​and angles on each side of the peak and trough, the principal reflection angle and the number of the maximum peaks are taken to characterize the global feature information of the sound field.

[0016] In the multi-feature constrained deep neural network model, the input x1 of the constrained subnetwork is the peak, valley values ​​and angles of 6 equally spaced intervals. The sound field intensity values ​​and angles on both sides of the peak and valley, the main reflection angle, and the number of the maximum peaks are taken.

[0017] The local sound field intensity value is set as x2, which is the input of the prediction sub-network in the multi-feature constrained deep neural network model, between 85° and 95°. The MSE function is usually used as the loss function for fitting training, but this loss function cannot accurately evaluate the nonlinear characteristics of the error and is sensitive to outliers. Therefore, a kernel function is introduced to modify the loss function.

[0018] Preferably, φ is set from the original feature space R n The nonlinear mapping function to the reproducing kernel Hilbert space has the form φ: R n →H, the inner product in the reproducing kernel Hilbert space is defined as follows:

[0019]

[0020] In equation (1-5) above, y t For the true value, For predicted values, The kernel function is represented by φ, which is a nonlinear mapping function.

[0021] The square of the distance between data points in the reproducing kernel Hilbert space is represented by a Mercer kernel as follows:

[0022]

[0023] d H The distance between data points within H is determined using the radial basis function as follows:

[0024] ;

[0025] In equation (1-7) above, γ is a parameter, which is set here to... n feature The number of network features;

[0026] Substituting formula (1-7) into formula (1-6) yields:

[0027]

[0028] The kernel loss function can be written in the following form:

[0029]

[0030] In formula (1-9), N is the total number of samples, and yt is the true value. This is a predicted value;

[0031] The kernel loss function L1 of the constrained subnetwork is:

[0032]

[0033] The kernel loss function L2 for the prediction subnetwork is:

[0034]

[0035] In the above formula (1-10) In the above formula (1-11) The final target kernel function L is set as follows:

[0036] L=αL1+(1-α)L2 (1-12);

[0037] Here, α is a weighting factor, which determines the proportion of a certain neural network in the whole model. In fact, it is the degree of influence of global feature values ​​on the predicted local feature values. In the whole deep neural network model, local feature values ​​have a greater impact on the whole model. Here, α is set to 0.4.

[0038] Preferably, in step S4, given the sound field intensity distribution P, data processing is performed to extract global feature information and local sound field intensity values, which are then input into the trained constraint subnetwork and prediction subnetwork, respectively, to obtain the predicted values ​​of the material parameters to be designed for the metasurface structure. The prediction results of the deep neural network model are then verified by forward simulation using COMSOL Multiphysics simulation software established with the previously generated sample data.

[0039] Compared with existing technologies, the beneficial effects are as follows: the local sound field control method based on custom loss function and multi-feature constraints simplifies the network model by extracting features from the sound field distribution and reducing the input of the deep neural network model. By introducing a custom kernel loss function, the computational load of the network model is further reduced, providing a new approach for the control of local sound fields. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 The flowchart shows the local sound field control method based on a custom loss function and multiple feature constraints provided by this invention.

[0042] Figure 2 This is a schematic diagram of the global far-field sound pressure level distribution in an embodiment of the local sound field modulation method based on a custom loss function and multiple feature constraints provided by the present invention.

[0043] Figure 3This is a schematic diagram of the linearly increasing material density distribution of the metasurface structure unit in an embodiment of the local sound field control method based on a custom loss function and multiple feature constraints provided by the present invention.

[0044] Figure 4 This is a schematic diagram of the linearly decreasing material density distribution of the metasurface structure unit in an embodiment of the local sound field control method based on a custom loss function and multiple feature constraints provided by the present invention.

[0045] Figure 5 This is a schematic diagram of a metasurface laid on a backing plate in an embodiment of the local sound field control method based on a custom loss function and multiple feature constraints provided by the present invention.

[0046] Figure 6 This is a schematic diagram of the basic network topology of the deep neural network model in an embodiment of the local sound field modulation method based on a custom loss function and multiple feature constraints provided by the present invention.

[0047] Figure 7 This is a schematic diagram of the basic network topology of the overall multi-feature constraint deep neural network model in the embodiment of the local sound field modulation method based on a custom loss function and multi-feature constraints provided by the present invention.

[0048] Figure 8 This is a schematic diagram of the loss function result of the first metasurface structure parameter density ρ1 in an embodiment of the local sound field control method based on a custom loss function and multiple feature constraints provided by the present invention.

[0049] Figure 9 This is a schematic diagram of the loss function result of the material density gradient g in an embodiment of the local sound field control method based on a custom loss function and multiple feature constraints provided by the present invention.

[0050] Figure 10 This is a schematic diagram of the desired sound field distribution and the predicted sound field distribution in an embodiment of the local sound field modulation method based on a custom loss function and multiple feature constraints provided by the present invention.

[0051] Figure 11 This is a schematic diagram of the local sound field intensity values ​​between 85° and 95° in an embodiment of the local sound field control method based on a custom loss function and multiple feature constraints provided by the present invention.

[0052] Figure 12 This is a schematic diagram illustrating the difference in local sound field intensity between the desired and predicted sound fields in the range of 85° to 95°, as provided in the embodiment of the local sound field control method based on a custom loss function and multiple feature constraints provided by the present invention.

[0053] Figure 13This diagram illustrates the comparison between the predicted sound field distribution obtained by using untrained sample data as input to the network model in an embodiment of the local sound field modulation method based on a custom loss function and multiple feature constraints provided by the present invention, and the local sound field intensity of the sample data between 85° and 95°. Detailed Implementation

[0054] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described in this specification are merely for explaining the invention and are not intended to limit the invention.

[0055] Please see Figure 1 This invention provides a local sound field modulation method based on a custom loss function and multi-feature constraints, comprising the following steps:

[0056] S1: Determine the geometric characteristics of the acoustic metasurface structure and design the structural parameter variables; lay the metasurface on the backing plate and construct the forward simulation model of the metasurface;

[0057] S2: The far-field sound pressure level intensity distribution corresponding to the input metasurface property structural parameters is obtained by finite element analysis of the forward simulation model, and the metasurface property structural parameters and the corresponding far-field sound pressure level intensity data are used as sample data.

[0058] S3: Extract multi-dimensional feature information of far-field sound pressure level intensity distribution and local sound field intensity information as input, and construct a multi-feature constrained deep neural network model to predict the metasurface physical property structure parameters;

[0059] S4: By comparing the far-field sound pressure level intensity distribution obtained by forward modeling the predicted values ​​of metasurface property structural parameters with the local target sound field intensity, the accuracy of the prediction results of the deep neural network model is verified, thereby realizing the effective control of the local sound field intensity.

[0060] Specifically, step S1 further includes step S11: determining the normal thickness H and length L of the acoustic metasurface, discretizing the metasurface into 25 structural units with the same geometric shape and size, and the material density ρ of each unit. i They are all different and ρ i The value of ρ1 varies linearly with the position of the structural unit in the length direction, i = 1, 2, ..., 25. The material density ρ1 of the first metasurface structural unit and the material density gradient g of the two adjacent metasurface structural units are used as the material parameters to be designed, and the range of the density ρ1 of the first metasurface structural unit is limited to [1 / 3, 3].

[0061] Specifically, the input to the forward simulation described in step S2 is the physical and structural parameters of the metasurface (e.g., the metasurface density ρ, length L, normal thickness H, etc.), and the output is the far-field sound pressure level intensity distribution P. The metasurface is composed of subwavelength structural units. With other parameters such as the metasurface length L and normal thickness H fixed, the material density ρ of the metasurface structural units is changed. i (i = 1, 2, ..., 25) can be used to control the reflected sound field.

[0062] To obtain a large amount of sample data for subsequent training and testing of the inversion model, COMSOL Multiphysics simulation software was used in step S2 for finite element analysis to establish a one-to-one mapping relationship from metasurface material property parameters to far-field sound pressure intensity distribution. A large amount of sample data was generated by forward simulation to prepare sample datasets for training and testing of deep neural network models.

[0063] Specifically, the multi-feature constrained deep neural network model in step S3 includes two sub-neural network models. The two sub-neural networks are trained by taking the global feature information and local intensity information of far-field sound pressure as inputs, respectively.

[0064] In step S3, when constructing the multi-feature constrained deep neural network model, the multi-feature constrained deep neural network model uses the global feature information of the far-field sound pressure level distribution and the local sound field intensity information as input. Since the entire reflected sound field is the result of the incident sound wave passing through 25 metasurface structural units, when focusing on the control of the local sound field intensity, it is considered that the local sound field is a subset of the entire sound field distribution, and its distribution is also affected by the trend of the entire sound field distribution.

[0065] Therefore, while using local sound field intensity values ​​as input to a deep neural network, it is also necessary to combine the feature information of the global sound field to constrain the local sound field intensity, thereby further improving the accuracy of the prediction model for the parameters of the material to be designed.

[0066] Depend on Figure 2 As shown in the global far-field sound pressure level distribution, the global characteristic information of the sound field is mainly concentrated in and around the peaks and troughs. Therefore, the sound field in the range of 0° to 180° can be divided into 6 equally spaced intervals. The peak and trough values ​​and angles of each interval can be taken, and two sound field intensity values ​​and angles on each side of the peak and trough, the principal reflection angle, and the number of maximum peaks (the threshold between the maximum peaks is set to 1 dB) can be taken to characterize the global characteristic information of the sound field.

[0067] Therefore, in the multi-feature constrained deep neural network model, the input x1 of the constrained sub-network is the peak, valley and angle of 6 equally spaced intervals (0° to 180° is divided into 6 intervals, each 30° is an interval). The sound field intensity value and angle on both sides of the peak and valley, the main reflection angle and the number of the maximum peaks are taken (the threshold between the maximum peaks is set to 1dB).

[0068] The local sound field intensity value is set as the input x2 of the prediction sub-network in the multi-feature constrained deep neural network model, which is set to a sound field intensity value between 85° and 95° in this invention;

[0069] When using deep neural networks for fitting training, the MSE function is usually used as the loss function. However, the MSE loss function cannot accurately assess the non-linear characteristics of the error, and it is sensitive to outliers. It will amplify errors with a difference greater than 1 and shrink errors with a difference less than 1, thus reducing the stability of the network.

[0070] To avoid this problem, a kernel function is introduced to modify the loss function. The custom kernel loss function iteratively updates by mapping the solution of the deep neural network to a high-dimensional and smooth space to find the optimal solution. At the same time, kernel tricks are used to solve the problem of increased computational complexity caused by moving from a low-dimensional space to a high-dimensional space.

[0071] Specifically, in step S4, given the sound field intensity distribution P, data processing is performed to extract global feature information and local sound field intensity values, which are then input into the trained constraint subnetwork and prediction subnetwork, respectively, to obtain the predicted values ​​of the material parameters to be designed for the metasurface structure. The prediction results of the deep neural network model are then verified by forward simulation using COMSOL Multiphysics simulation software established with the previously generated sample data.

[0072] The specific implementation method is as follows:

[0073] In step S1 above, the acoustic metasurface has a normal thickness H = 0.12 m and a length L = 1 m. Each metasurface structural unit has a length of 0.04 m. The density ρ1 of the first metasurface structural unit ranges from [1 / 3, 3]. Since the material density values ​​of the metasurface structural units are arranged in a linear gradient, if the material density gradient g between two adjacent metasurface structural units is greater than 0, it indicates that the material density gradient of the entire metasurface structural unit is linearly increasing. Figure 3 As shown;

[0074] Conversely, it exhibits a decreasing distribution, such as Figure 4 As shown, the value of the material density gradient g should satisfy a constraint condition, namely, that the material density ρ of all metasurface structural units should be such that... i∈[1 / 3,3],i=1,2,...,25.

[0075] In step S2 above, finite element analysis is performed using COMSOL Multiphysics simulation software. The density of an ideal metasurface material is continuous, but artificial periodic structures cannot achieve continuous changes in material density. Therefore, it is necessary to divide the metasurface into multiple metasurface structural units with the same density gradient along the length direction.

[0076] In large-scale environments, the finite element model of a single metasurface structural unit can be considered as a point mass. Therefore, when designing it, the material density at its center point is taken as the material density of the entire structural unit for subsequent calculations, establishing a one-to-one mapping relationship from the metasurface's physical properties and structural parameters to the far-field sound pressure intensity distribution, such as... Figure 5 As shown, the metasurface is laid on a backing plate, and the incident wave is incident perpendicularly. The metasurface is divided into 25 metasurface structural units with the same size and linearly varying material density along its length. Through forward modeling, sample data can be obtained as a sample dataset for training and testing subsequent deep neural network models.

[0077] In step S3 above, the basic network topology of the deep neural network model used for inversion is as follows: Figure 6 As shown, it is divided into an input layer x i Hidden layer h i and output layer y i The formulas for calculating input and output are as follows:

[0078]

[0079] In equation (1-1) above, ω represents the neural network parameters between the input and output layers, x represents the input of the neural network, y represents the output of the neural network, and the ReLU function is used for the activation functions of the hidden and output layers of the neural network model.

[0080]

[0081] When updating the parameter ω, the Adam optimization algorithm was chosen, and its expression is as follows:

[0082]

[0083]

[0084] In formulas (1-3) and (1-4) above, α is the learning rate, t is the time step initialized to 0, β1 and β2 are the exponential decay rates of the first and second moments, respectively, and m... t v tThese are the first-order moment estimates and second-order moment estimates of the gradient, respectively. ε is a constant to prevent the denominator from being zero, and ω is updated accordingly.

[0085] Overall multi-feature constrained deep neural network model, such as Figure 7 As shown, it includes two sub-neural network models, which divide the sound field in the range of 0° to 180° into 6 equally spaced intervals. The peak, trough values ​​and angles of each interval are taken, and the sound field intensity values ​​and angles on both sides of the peak and trough, the main reflection angle and the number of maximum peaks (the threshold between the maximum peaks is set to 1dB) are taken to characterize the global feature information of the sound field.

[0086] The input x1 of the constrained subnetwork is the sound field ranging from 0° to 180° divided into 6 equally spaced intervals, taking the peak, trough values ​​and angles of each interval, taking two sound field intensity values ​​and angles on each side of the peak and trough, the principal reflection angle, and the number of maximum peaks (the threshold between maximum peaks is set to 1dB). The output y1 is the density ρ of the first metasurface structural unit. m1 The gradient g1; the input x2 of the prediction subnetwork is the local sound field intensity value between 85° and 95°, and the output y2 is the density ρ of the first metasurface structural unit. m2 And gradient g2.

[0087] When modifying the loss function, a kernel function from deep learning is introduced to modify the loss function, and φ is set to be derived from the original feature space R. n The nonlinear mapping function to the reproducing kernel Hilbert space (RKHS) is of the form φ: R n →H, the inner product in the reproducing kernel Hilbert space can be defined as follows:

[0088]

[0089] In equation (1-5) above, y t For the true value, For predicted values, The kernel function is represented by φ, which is a nonlinear mapping function.

[0090] The square of the distance between data points in the reproducing kernel Hilbert space is represented by a Mercer kernel as follows:

[0091]

[0092] d H Let H be the distance between data points. Commonly used kernel functions include radial basis function (RBF) kernel function and polynomial kernel function. In this application, the radial basis function is used. The radial basis function is as follows:

[0093]

[0094] In equation (1-7) above, γ is a parameter, which is set here to... n feature The number of network features;

[0095] Substituting formula (1-7) into formula (1-6) yields:

[0096]

[0097] The kernel loss function can be written in the following form:

[0098]

[0099] In formula (1-9), N is the total number of samples, and yt is the true value. This is a predicted value;

[0100] In this application, the neural network model consists of two sub-neural networks, and the kernel loss function L1 constrains the sub-neural networks as follows:

[0101]

[0102] The kernel loss function L2 for the prediction subnetwork is:

[0103]

[0104] In the above formula (1-10) In the above formula (1-11) The final target kernel function L is set as follows:

[0105] L=αL1+(1-α)L2 (1-12)

[0106] Here, α is a weighting factor, which determines the proportion of a certain neural network in the whole model. In fact, it is the degree of influence of global feature values ​​on the predicted local feature values. In the whole deep neural network model, local feature values ​​have a greater impact on the whole model. Here, α is set to 0.4.

[0107] In step S4 above, data processing is performed on the given sound field intensity distribution P to extract the peak, trough values ​​and angles within six equally spaced intervals. Two sound field intensity values ​​and angles on each side of the peak and trough, the principal reflection angle, and the number of maximum peaks (with a threshold of 1 dB between maximum peaks) are taken as global feature values. These, along with the local sound field intensity values ​​corresponding to 85°–95°, are used as inputs to a multi-feature constrained deep neural network. This allows for the prediction of the density and gradient of the first metasurface structural unit. and Using this as input to a forward simulation model, predicted values ​​can be obtained. and Corresponding predicted sound field intensity distribution The error between the two is then compared with the given desired sound field intensity distribution P.

[0108] Using the local sound field intensity between 85° and 95° as the control target, the loss function result of the first metasurface structure parameter density ρ1 in the trained multi-feature constrained deep neural network model is as follows: Figure 8 As shown, the loss function result for gradient g is as follows: Figure 9 As shown, the desired sound field distribution and the predicted sound field distribution are as follows: Figure 10 As shown, its local sound field intensity values ​​between 85° and 95° are as follows: Figure 11 As shown, the local sound field intensity difference between the desired sound field and the predicted sound field between 85° and 95° is as follows: Figure 12 As shown.

[0109] Depend on Figure 10 It can be seen that the overall distribution trends of the expected sound field and the predicted sound field are basically consistent, indicating that the multi-feature constrained prediction network model has learned the global sound field features from the input. Furthermore, because the prediction network is given a relatively large weight in the network model...

[0110] Therefore, the difference between the expected local sound field intensity and the predicted sound field intensity is relatively small between 85° and 95°, such as... Figure 12 As shown, to further verify the generalization ability of the multi-feature constrained network model, this invention extracts 1000 sets of sample data from the untrained sample data as input to the network model, then uses a forward simulation model to obtain the predicted sound field distribution, and then compares it with the local sound field intensity of the sample data between 85° and 95°. The distribution of the difference is shown in the figure. Figure 13 As shown, 92.9% of the sample data had an error of less than 5 dB compared to the predicted sound field data. This invention introduces a kernel loss function in the network design to reduce the computational load of the model. Experimental verification shows that the distribution trends of the sample data and the predicted sound field are basically consistent. Furthermore, within the target control range (i.e., 90°±5°), the predicted sound field intensity is very close to the true value. This demonstrates a good control effect on the local sound field, providing a new approach for controlling the local sound field.

[0111] The present invention is not limited to the description in the specification and embodiments, and thus other advantages and modifications can be readily realized by those skilled in the art. Therefore, the present invention is not limited to the specific details, representative devices and examples shown and described herein without departing from the spirit and scope of the general concept as defined by the claims and their equivalents.

Claims

1. A local sound field control method based on a custom loss function and multi-feature constraints, characterized in that, Includes the following steps: S1: Determine the geometric characteristics of the acoustic metasurface structure and design the structural parameter variables; lay the metasurface on the backing plate and construct the forward simulation model of the metasurface; S2: The far-field sound pressure level intensity distribution corresponding to the input metasurface material property structural parameters is obtained by finite element analysis of the forward simulation model. The material property structural parameters of the metasurface and the corresponding far-field sound pressure level intensity data are used as sample data. S3: Extract multi-dimensional feature information of far-field sound pressure level intensity distribution and local sound field information as input, and construct a deep neural network model with multi-feature constraints to predict the metasurface material property structure parameters; S4: By comparing the far-field sound pressure level intensity distribution obtained by forward modeling the predicted values ​​of metasurface material property structural parameters with the local target sound field intensity, the accuracy of the prediction results of the deep neural network model is verified, thereby realizing the effective control of the local sound field. The sound field in the range of 0° to 180° is divided into 6 equally spaced intervals. The peak, trough values ​​and angles of each interval are taken. Two sound field intensity values ​​and angles on each side of the peak and trough, the principal reflection angle and the number of the maximum peaks are taken to characterize the global feature information of the sound field. Constraint subnetwork input in a deep neural network model with multiple feature constraints The values ​​and angles of the peaks and troughs in 6 equally spaced intervals, the two sound field intensity values ​​and angles on each side of the peaks and troughs, the principal reflection angle, and the number of the maximum peaks; Predicting subnetwork input in multi-feature constrained deep neural network models The local sound field intensity value is set to a value between 85° and 95°; a kernel function is introduced to modify the commonly used MSE loss function, and the modified loss function is used for fitting and training; In step S4, given the sound field intensity distribution Data processing is performed, and global feature information and local sound field intensity values ​​are extracted and input into the trained constraint subnetwork and prediction subnetwork respectively to obtain the predicted values ​​of the material parameters to be designed for the metasurface structure. Forward simulation is used to verify the prediction results of the deep neural network model.

2. The local sound field control method based on a custom loss function and multiple feature constraints as described in claim 1, characterized in that, Step S1 further includes step S11: determining the normal thickness of the acoustic metasurface. ,length The metasurface is discretized into 25 structural units with the same geometry and size, and the material density of each unit is... They are all different and The value of varies linearly with the position of the structural element in the length direction. =1, 2, …, 25, with the material density of the first metasurface structural unit. Material density gradient between two adjacent metasurface structural units As material parameters to be designed, and limiting the density of the first metasurface structural unit. The range of values ​​is .

3. The local sound field control method based on a custom loss function and multiple feature constraints as described in claim 2, characterized in that, The input to the forward simulation described in step S2 is the physical and structural parameters of the metasurface, and the output is the far-field sound pressure level intensity distribution. The metasurface is composed of subwavelength structural units arranged in a specific pattern. L Normal thickness With the material density of the metasurface structural unit fixed, the method can be changed. To achieve the control of the reflected sound field.

4. The local sound field control method based on a custom loss function and multiple feature constraints as described in claim 3, characterized in that, In step S2, finite element analysis is performed using COMSOL Multiphysics simulation software to establish a one-to-one mapping relationship between metasurface material properties and structural parameters and far-field sound pressure intensity distribution. A large amount of sample data is generated by forward modeling to prepare sample datasets for training and testing deep neural network models.

5. The local sound field control method based on a custom loss function and multiple feature constraints as described in claim 3, characterized in that, The multi-feature constrained deep neural network model in step S3 includes two sub-neural networks, which are trained by taking the global feature information and local intensity information of far-field sound pressure as inputs, respectively.

6. The local sound field control method based on a custom loss function and multiple feature constraints as described in claim 1, characterized in that, set up From the original feature space R n The nonlinear mapping function to the reproducing kernel Hilbert space has the form: : R n ， The inner product in the reproducing kernel Hilbert space is defined as follows: (1-5)； In the above formula (1-5) For the true value, For predicted values, Represents the kernel function. It is a nonlinear mapping function; The square of the distance between data points in the reproducing kernel Hilbert space is represented by a Mercer kernel as follows: (1-6)； for H The distance between internal data points is determined using the radial basis function as follows: （1-7）； In the above formula (1-7) For a parameter, set it to [value] here. , The number of network features; Substituting formula (1-7) into formula (1-6) yields: （1-8）； The kernel loss function can be written in the following form: （1-9）； In formula (1-9), N The total number of samples. For the true value, This is a predicted value; Kernel loss function of constrained subnetwork for: （1-10）； Kernel loss function of the prediction subnetwork for: （1-11）； In the above formula (1-10) In the above formula (1-11) Set the final target kernel loss function. for: （1-12）； in, As the weighting factor, use The weight of a particular sub-neural network in the overall model is determined here; in reality, it reflects the degree of influence of global feature values ​​on the predicted local feature values. .

Images