A method for non-destructive estimation of physical simulation parameters of a material based on an acoustic response

Abstract

The application discloses a kind of based on acoustic response material physics simulation parameter nondestructive estimation method.This application contains four steps:1.to the acoustic simulation parameter of material in real world is collected, so as to obtain the probability distribution that acoustic simulation parameter obeys.By sampling and hierarchical filtering to the probability distribution, obtain several approximate feasible material acoustic simulation parameter.2.based on pre-computed boundary element method obtains the acoustic response signal of different acoustic simulation parameter.3.using one-dimensional convolutional neural network to establish the corresponding relationship between acoustic response signal and simulation parameter, and in the training process, utilize frequency domain background noise to carry out data enhancement.4.acquisition obtains the impulse response signal of material to be measured, and the signal is input into neural network, complete material acoustic simulation parameter estimation.In the process of simulation parameter measurement in the application, material sample will not be damaged, can nondestructively estimate material property, with higher efficiency.

Description

Technical Field

[0001] This invention relates to the fields of computer graphics, acoustics, and physical simulation, and particularly to a non-destructive estimation method for material physics simulation parameters based on acoustic response. Background Technology

[0002] Realistic acoustic simulation is one of the more active research areas in computer graphics. In recent years, researchers in the field have proposed a series of methods to improve the computational efficiency of acoustic simulation systems, but research on the realism of acoustic simulations is relatively limited.

[0003] The setting of physical simulation parameters has a significant impact on the realism of acoustic simulations. Traditional destructive physical parameter measurement methods based on tensile and compressive loading have significant limitations: these methods damage material samples and can only measure elastic information such as Young's modulus, not the internal friction coefficient of the material. In recent years, researchers have proposed vision-based object classification methods (see Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H. Adelson, and William T. Freeman. "Visually indicated sounds." In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 2405-2413. 2016.), and subsequent researchers have further extended object classification methods to material physical parameter estimation methods. Current deep learning-based physical parameter measurement and estimation methods can estimate the acoustic parameters of some materials, but these methods all rely on the collision process of falling objects as the main basis for acoustic simulation parameter measurement. The collision process of objects varies greatly and is easily affected by external interference, making it difficult to guarantee the stability, accuracy, and robustness of parameter measurements. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a non-destructive estimation method for material physics simulation parameters based on acoustic response, which can estimate the acoustic simulation parameters of materials without damaging the material sample.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a non-destructive estimation method for material physics simulation parameters based on acoustic response, the method comprising the following steps:

[0006] (1) Acoustic simulation parameters of materials in the real world are collected, including density, Poisson's ratio and Young's modulus. The probability distribution of acoustic simulation parameters is approximated by the Gaussian mixture model, and a simulation parameter subspace is constructed. Several sets of acoustic simulation parameters are sampled from it.

[0007] (2) The acoustic response signals corresponding to the different acoustic simulation parameters sampled in step (1) are obtained by using the boundary element method based on pre-computation acceleration;

[0008] (3) Using the acoustic simulation parameters sampled in step (1) and the acoustic response signal calculated in step (2) as the training set, collect n noise samples N in the working environment. i For each t, ​​i = 1, 2, ..., n, calculate the time-domain noise samples N using the short-time Fourier transform. i The spectrum W of (t) i Randomly select the noise spectrum W i It is then converted into a time-domain signal and superimposed with the input signal of the neural network to complete data augmentation. A one-dimensional convolutional neural network is trained to extract features from the acoustic response signal. While the output of the one-dimensional convolutional neural network is normalized, the acoustic simulation parameters are logarithmically operated on, and a nonlinear relationship between the features in the acoustic response signal and the acoustic simulation parameters is established.

[0009] (4) Obtain the impulse acoustic response signal of the material under test, and input the impulse acoustic response signal into the neural network trained in step (3) to obtain the features in the acoustic response signal of the material under test, thereby non-destructively estimating the physical simulation parameters of the material under test.

[0010] Furthermore, in step (1), when collecting acoustic simulation parameters, for anisotropic materials, the physical quantities in their dissimilar directions are taken as average values.

[0011] Furthermore, the specific process of sampling several sets of acoustic simulation parameters is as follows:

[0012] (1.1) By sampling the Gaussian mixture model, several approximate material densities, Poisson's ratios and Young's moduli are obtained; the damping parameters a and b are uniformly sampled and combined with the material density, Poisson's ratio and Young's moduli to form a complete set of acoustic simulation parameters.

[0013] (1.2) Using low-resolution and second-low-resolution grids as carriers, the response signals of several acoustic simulation parameters in (1.1) are quickly calculated, and illegal sound simulation parameters are filtered out according to the response signals, so as to obtain several more reasonable sets of acoustic simulation parameters and reduce the computational cost.

[0014] Furthermore, the specific steps for obtaining the acoustic response signals corresponding to different acoustic simulation parameters using the pre-computation-accelerated boundary element method are as follows:

[0015] (2.1) Set a thin plate as the output result of the test vibration, and set the position of the microphone in the space;

[0016] (2.2) Discretize the thin plate into a tetrahedral mesh with N vertices. Given the density, Young's modulus, Poisson's ratio, and damping coefficient of the thin plate, calculate the mass matrix M, stiffness matrix K, and Rayleigh damping matrix D = aM + bK of the material using the neo-Hookean constitutive model. The range of damping coefficient a is [1e-8, 1e-5], and the range of damping coefficient b is [0, 5e3]. Further, the dynamic equation of the material is obtained:

[0017]

[0018] Where x is the world coordinate of the tetrahedron vertex and f is the external impact on the material;

[0019] (2.3) By performing generalized eigenvalue decomposition KU=MUS on the mass matrix and stiffness matrix, and using the eigenma matrix U to establish the coordinate transformation relationship x=Uq for the eigenspace vector q, the vibration on the modal is calculated:

[0020]

[0021] By solving the modal vibration equations and superimposing the vibrations on multiple modes, the x-value at the i-th vertex of the thin plate is calculated. i vibration function

[0022] (2.4) Through the solid-gas boundary condition, the vibration at the vertex of the thin plate is transformed into the sound intensity p(x,t)=p(x)e on the surface of the thin plate. jwt , where w is the frequency of the sound wave; based on the sensitivity analysis of the human ear to the audible frequency range, the continuous audible frequency range is discretized into several frequency intervals, and for each frequency interval, the Helmholtz equation satisfied by the propagation of the sound wave in the air is pre-calculated:

[0023]

[0024] in Let be the Laplace operator, k be the wavenumber, and p be the sound pressure at a point in space. Taking the geometric center of the thin plate as the origin, the Helmholtz equation can be solved using the fast multipole source method, yielding numerical solutions for the Helmholtz equation in several frequency ranges.

[0025] (2.5) Based on the modal analysis in step (2.3), several vibration frequencies of the thin plate corresponding to the current acoustic simulation parameters are calculated. Using these vibration frequencies as a basis, corresponding numerical solutions are selected from the numerical solutions of the Helmholtz equation pre-calculated in step (2.4) and superimposed to calculate the microphone x in space. mic Acoustic response signal p(x) at the location mic ,t).

[0026] The beneficial effects of this invention are as follows:

[0027] 1. Since this invention estimates the physical parameters of materials based on their acoustic impact response, compared with traditional methods based on tensile and compressive loads, the technology in this invention does not damage the material sample during the simulation parameter measurement process, and can estimate the material properties non-destructively; at the same time, the entire parameter estimation process can be completed within 3-5 seconds, which has higher efficiency.

[0028] 2. Compared with previous studies, this invention uses a Gaussian mixture model to approximate the feasible physical parameter distribution of the material and obtains feasible acoustic simulation parameters by sampling the distribution. This eliminates unreasonable density, Young's modulus and Poisson ratio, greatly reducing the pre-computation time.

[0029] 3. Compared with previous studies, this invention proposes a hierarchical acoustic simulation parameter filtering method, which removes unreasonable combinations of damping coefficients and further reduces the pre-calculation time.

[0030] 4. This invention proposes a data augmentation technique based on spectral decomposition to model indoor noise under operating conditions and intervene in the input signal during network training. This enables the parameter estimation system in this invention to stably complete measurement tasks under different indoor environments, exhibiting stronger robustness.

[0031] 5. This invention proposes a pre-calculated boundary element method for solving the sound distribution at a target location. This method discretizes the continuous audible range (20Hz-20kHz) into 1000 frequency groups by performing sensitivity analysis on the audible range. By pre-calculating the numerical solutions of the Helmholtz equations on these frequency groups, the sound solution speed of the boundary element method is improved, and the pre-calculation cost of acoustic simulation parameter estimation is reduced. Attached Figure Description

[0032] Figure 1 This is a diagram of the overall network framework of the method of the present invention. Detailed Implementation

[0033] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The present invention provides a non-destructive estimation method for material physics simulation parameters based on acoustic response, which can estimate the acoustic simulation parameters of materials without damaging the material sample. The present invention has high application efficiency; the estimation of a single sample can be completed within three to five seconds.

[0034] like Figure 1 As shown, the present invention provides a non-destructive estimation method for material physics simulation parameters based on acoustic response, comprising the following steps:

[0035] (1) Acoustic simulation parameters of real-world materials were collected, including density, Poisson's ratio, and Young's modulus, and their probability distribution was approximated using a Gaussian mixture model. This invention first measured the acoustic simulation parameters of real-world materials, including common metals and polymers. The measured parameters included density, Young's modulus, and Poisson's ratio. To further expand the sample size, this invention used publicly available materials physics datasets from the internet. For anisotropic materials, the values ​​in each direction were averaged. Ultimately, this invention collected the physical properties of 73 common materials (including ceramics, metals, and polymers). This invention used a Gaussian mixture model with a cluster size of 5 to approximate the probability density distribution of the above physical properties and sampled 1024 different physical simulation parameters from it. The specific steps for sampling several sets of acoustic simulation parameters are as follows: By sampling the Gaussian mixture model, several approximate material densities, Poisson's ratios, and Young's moduli are obtained; damping parameters a and b are uniformly sampled and combined with material density, Poisson's ratio, and Young's modulus to form several complete acoustic simulation parameters; using low-resolution grids with 537 vertices and second-lowest resolution grids with 1026 vertices as carriers, the response signals of several acoustic simulation parameters are quickly calculated, and illegal sound simulation parameters are filtered based on the response signals, thereby obtaining several more reasonable sets of acoustic simulation parameters and reducing computational costs.

[0036] In real life, due to friction between molecules within an object, the vibration of an excited object gradually decays until it comes to rest. In acoustic simulations, this decay phenomenon is often approximated by the Rayleigh damping model. This model includes two damping coefficients, a and b, which are difficult to measure. Through simulation numerical experiments, this invention determines the effective range of the damping coefficients: the range of damping coefficient a is [1e-8, 1e-5], and the range of damping coefficient b is [0, 5e3]. By uniformly sampling 25 properties on these two channels and performing a Cartesian product with the previous 1024 physical parameters, a total of 25600 sets of material parameters were obtained. To reduce pre-calculation costs and improve sampling efficiency, this invention proposes a hierarchical material parameter filtering method. This method pre-solves the sound response signals of the 25600 sets of material parameters on several low-resolution thin plates. If the solved sound signal exhibits overdamping or a continuous sound exceeding 0.3 seconds, the set of material parameters is deemed unrealistic and is no longer subject to further finer-grained sampling. In this invention, two types of low-resolution thin plates with N=537 and N=1026 are used for hierarchical filtering, and 7900 feasible material acoustic simulation parameters are obtained from 25600 sets of parameters, reducing the amount of computation by 70%.

[0037] (2) The acoustic response signals corresponding to different acoustic simulation parameters sampled in step (1) are obtained by using the boundary element method based on pre-computation acceleration. In this invention, a thin plate with a width of 3cm, a length of 5cm, and a thickness of 0.5cm is selected as the output result of the test vibration. The thin plate is discretized into a tetrahedral mesh composed of N=1537 vertices. Given the density, Young's modulus, and Poisson's ratio of the material, this invention uses the neo-Hookean constitutive model to calculate the mass matrix M, stiffness matrix K, and Rayleigh damping matrix D=αM+βK of the material, and further obtains the dynamic equation of the material:

[0038]

[0039] Where x represents the world coordinates of the tetrahedron vertices, and f represents the external impact on the material. By performing generalized eigenvalue decomposition KU = MUS on the mass and stiffness matrices, and using the eigenma matrix U to establish the coordinate transformation relationship x = Uq for the eigenspace vector q, the above formula can be transformed into modal vibration:

[0040]

[0041] By solving the above modal vibration equations and superimposing the vibrations on multiple modes, this invention can calculate the value of x at any vertex of the thin plate. i vibration function By applying the solid-gas boundary condition, the vibration at the apex of the thin plate can be converted into the sound intensity p(x,t) = p(x)e on the surface of the thin plate. jwt, where w is the frequency of the sound wave. Based on the sensitivity analysis of the human ear to the audible frequency range, the continuous audible frequency range is discretized into several frequency intervals. For each frequency interval, the propagation of the sound wave in a uniform medium is pre-calculated to satisfy the Helmholtz equation:

[0042]

[0043] in Let k be the wavenumber and p be the sound pressure at a point in space, and let the geometric center of the thin plate be the origin. The Helmholtz equation can be solved using the fast multipole source method to obtain numerical solutions over several frequency ranges. Based on modal analysis, several vibration frequencies of the thin plate corresponding to the current acoustic simulation parameters are calculated. Using these vibration frequencies as a basis, corresponding numerical solutions are selected from the pre-calculated numerical solutions of the Helmholtz equation and superimposed to calculate the microphone x in space. mic The sound signal p(x) at the location mic To accelerate the solution of sound signals, this invention proposes a pre-calculation and rapid synthesis method based on the Helmholtz equation within the audible range of the human ear. Based on the sensitivity analysis of the human ear to the audible frequency range (20Hz-20kHz), this invention discretizes the continuous audible frequency range into 1000 frequency intervals. For each frequency interval, a numerical solution to the Helmholtz equation satisfied by sound wave propagation in air is pre-calculated, and the numerical solution for each frequency interval is stored on disk. For a given set of acoustic simulation parameters, 60 sets of natural frequencies corresponding to the parameters are calculated through generalized eigenvalue decomposition. Furthermore, the closest frequency numerical solution is extracted from the pre-calculated numerical solutions of the Helmholtz equation at the 1000 frequencies, and the sound signal at the thin plate is rapidly synthesized. Compared with directly calculating the sound signal, the pre-calculation method proposed in this invention has a computational efficiency that is an order of magnitude higher.

[0044] The microphone position is determined based on the microphone position in the acoustic acquisition system in step (4). In this invention, the sound sampling signal duration is 1 second, the sampling rate is 44.1 kHz, and the sampling frequency range is the human ear-friendly frequency range of 20 Hz to 20 kHz.

[0045] The above sampling process took 288 hours to complete on an experimental platform equipped with an Intel i7-7700K CPU.

[0046] (3) This invention uses a one-dimensional convolutional neural network as the basic network structure to extract features from acoustic signals and establish a nonlinear relationship between acoustic features and physical properties. To more simply describe the neural network structure, this invention defines "convolution-batch normalization-max pooling" as the basic block. The network structures used in this invention are shown in Table 1:

[0047] Table 1

[0048] Layer name Layer parameters Input normalization layer The mean and variance of the input training data (audio signal) Basic Block 1 Kernel size = 64, stride = 2; normalized momentum = 0.1, pooling stride = 8 Basic Block 2 Kernel size = 32, stride = 2; normalized momentum = 0.1, pooling stride = 8 Basic Block 3 Kernel size = 16, stride = 2; normalized momentum = 0.1, pooling stride = 4 Basic Block 4 Kernel size = 8, stride = 2; normalized momentum = 0.1, pooling stride = 4 Basic Block 5 Kernel size = 4, stride = 2; normalized momentum = 0.1, pooling stride = 4 Output normalization layer Output the mean and variance of the training data (logarithmic acoustic simulation parameters).

[0049] In this invention, the Rectified Linear Function (RELU) is used as the nonlinear activation function. Due to the significant differences in units and value ranges among the different components of the acoustic simulation parameters, to ensure training stability, this invention performs a logarithmic operation on the acoustic simulation parameters while normalizing the output. By establishing a L2 loss function between the network's predicted acoustic parameters and the actual parameters, and setting the batch size to 32, the network training is completed after 60 iterations.

[0050] This invention incorporates three types of data augmentation during network training. Experiments show that after adding the data augmentation designed in this invention, the neural network can provide stable outputs in noisy environments and under different recording devices, significantly improving its robustness. The three data augmentation methods will be described in detail below.

[0051] 1. Variable speed data enhancement: The input signal is subjected to frequency-preserving time stretching, and the time stretching ratio is determined by uniformly generated random numbers between 0.9 and 1.1.

[0052] 2. Spectral data enhancement based on indoor noise: Acquisition of ten noise samples N from common operating environments of the system. i (t), i = 1, 2, ..., 10, each noise sample is 60s long, and the sampling frequency is 44.1kHz. The time-domain noise signal N is calculated using Discrete Fourier Transform. i The spectrum W of (t) i And save it to disk. During network training, a noise spectrum W is randomly selected. i It is then converted into a time-domain signal and superimposed on the input signal of the neural network.

[0053] 3. Amplitude data enhancement: The amplitude of the input signal is uniformly and randomly scaled proportionally, with the scaling factor determined by a random number between 0.5 and 2.

[0054] (4) The impulse response signal of the material under test is obtained by acoustic acquisition and the signal is input into the neural network trained in step (3) to deduce the material physical simulation parameters.

[0055] The component information used in the acoustic acquisition process is shown in Table 2.

[0056] Table 2

[0057] serial number name parameter 1 microphone SE Electronic 2200C 2 Sample Pine wood sample 3 sound card Steinberg UR44C sound card 4 hammer Total weight 3kg, rubber hammer head 5 Computer for data collection Lenovo ThinkPad laptops

[0058] This invention uses REAPER software to acquire and pre-edit sound signals, ensuring that the frequency of the sampled signal is consistent with the frequency of the training input signal for the network in step (3). The specific pre-editing operation involves locating the start time of the impulse signal in the software, and then clipping a 1-second portion of the signal from that moment as the input to the neural network trained in step (3). Rapid inference is achieved through the neural network, thereby obtaining the acoustic simulation parameters of the material.

[0059] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A non-destructive estimation method for material physics simulation parameters based on acoustic response, characterized in that, The method includes the following steps: (1) Acoustic simulation parameters of materials in the real world are collected, including density, Poisson's ratio and Young's modulus. Then, using the probability distribution of acoustic simulation parameters of Gaussian mixture model, a simulation parameter subspace is constructed, from which several sets of acoustic simulation parameters are sampled. (2) The acoustic response signals corresponding to the different acoustic simulation parameters sampled in step (1) are obtained by using the boundary element method based on pre-computation acceleration; the specific steps of obtaining the acoustic response signals corresponding to the different acoustic simulation parameters by the boundary element method based on pre-computation acceleration are as follows: (2.1) Set up a flat plate as the output of the test vibration, and set the position of the microphone in the space; (2.2) Discretize the plate into a tetrahedral mesh with N vertices. Given the plate's density, Young's modulus, Poisson's ratio, and damping coefficient, calculate the material's mass matrix M, stiffness matrix K, and Rayleigh damping matrix D = aM + bK using the neo-Hookean constitutive model. The damping coefficient a ranges from [1e-8, 1e-5], and the damping coefficient b ranges from [0, 5e3]. Further, the material's dynamic equations are obtained: Where x is the world coordinate of the tetrahedron vertex and f is the external impact on the material; (2.3) By performing generalized eigenvalue decomposition KU=MUS on the mass matrix and stiffness matrix, and using the eigenma matrix U to establish the coordinate transformation relationship x=Uq for the eigenspace vector q, the vibration on the modal is calculated: By solving the modal vibration equations and superimposing the vibrations on multiple modes, the x-th vertex of the plate is calculated. i vibration function (2.4) Through the solid-gas boundary condition, the vibration at the top of the plate is transformed into the sound intensity p(x,t)=p(x)e on the plate surface. jwt , where w is the frequency of the sound wave; based on the sensitivity analysis of the human ear to the audible frequency range, the continuous audible frequency range is discretized into several frequency intervals, and for each frequency interval, the Helmholtz equation satisfied by the propagation of the sound wave in the air is pre-calculated: in Let be the Laplace operator, k be the wave number, and p be the sound pressure at a point in space; assuming the geometric center of the plate is the origin, the Helmholtz equation can be solved using the fast multipole source method to obtain numerical solutions of the Helmholtz equation in several frequency ranges. (2.5) Based on the modal analysis in step (2.3), several vibration frequencies of the plate corresponding to the current acoustic simulation parameters are calculated. Using these vibration frequencies as a basis, corresponding numerical solutions are selected from the pre-calculated numerical solutions of the Helmholtz equation in step (2.4) and superimposed to calculate the microphone x in space. mic Acoustic response signal p(x) at the location mic ,t); (3) Using the acoustic simulation parameters sampled in step (1) and the acoustic response signal calculated in step (2) as the training set, collect n noise samples N in the working environment. i For each t, ​​i = 1, 2, ..., n, calculate the time-domain noise samples N using the short-time Fourier transform. i The spectrum W of (t) i Randomly select the noise spectrum W i It is then converted into a time-domain signal and superimposed with the input signal of the neural network to complete data augmentation. A one-dimensional convolutional neural network is trained to extract features from the acoustic response signal. While the output of the one-dimensional convolutional neural network is normalized, the acoustic simulation parameters are logarithmically operated on, and a nonlinear relationship between the features in the acoustic response signal and the acoustic simulation parameters is established. (4) Obtain the impulse acoustic response signal of the material under test, and input the impulse acoustic response signal into the neural network trained in step (3) to obtain the features in the acoustic response signal of the material under test, thereby non-destructively estimating the physical simulation parameters of the material under test.

2. The non-destructive estimation method for material physics simulation parameters based on acoustic response according to claim 1, characterized in that, In step (1), when collecting acoustic simulation parameters, for anisotropic materials, the physical quantities in their dissimilar directions are taken as average values.

3. The non-destructive estimation method for material physics simulation parameters based on acoustic response according to claim 1, characterized in that, The specific process of sampling several sets of acoustic simulation parameters is as follows: (1.1) By sampling the Gaussian mixture model, several material densities, Poisson's ratios and Young's moduli are obtained; the damping parameters a and b are uniformly sampled and combined with the material densities, Poisson's ratios and Young's moduli to form a complete set of acoustic simulation parameters. (1.2) Using low-resolution and second-low-resolution grids as carriers, the response signals of several acoustic simulation parameters in (1.1) are quickly calculated, and illegal sound simulation parameters are filtered out according to the response signals, so as to obtain several more reasonable sets of acoustic simulation parameters and reduce the computational cost.

Images