Method for evaluating sound insulation performance of acoustic materials based on machine learning

Abstract

The application discloses an acoustic material sound insulation performance evaluation method based on machine learning, which comprises the following steps: acquiring the first apparent characteristics of the acoustic material through image recognition, constructing an acoustic material graph neural network, predicting the key parameters of the finite element according to the acoustic material graph neural network and obtaining the predicted basic sound insulation performance through acoustic simulation, obtaining the predicted deep sound insulation performance according to the deep sound insulation performance prediction model, determining the first performance weight according to the acoustic material design target, correcting the first performance weight according to the environmental state to obtain the second performance weight, determining the sound insulation score according to the second performance weight, the predicted deep sound insulation performance and the predicted basic sound insulation performance, and determining the sound insulation grade of the acoustic material to be evaluated according to the sound insulation score. The method can accurately and efficiently evaluate the sound insulation performance of the acoustic material, and provides strong technical support for the research and development, selection and application of the acoustic material.

Description

Technical Field

[0001] This invention relates to the field of performance evaluation, and more particularly to a method for evaluating the sound insulation performance of acoustic materials based on machine learning. Background Technology

[0002] In fields such as architectural acoustics, traffic noise reduction, and industrial noise control, the sound insulation performance of acoustic materials is directly related to the quality of the acoustic environment and the effectiveness of noise control. As the requirements for acoustic environment increase, the evaluation technology continues to develop. Accurate and efficient evaluation of the sound insulation performance of acoustic materials can accelerate the research and development of new materials, optimize acoustic engineering design, and is of great significance to promoting the development of industries such as green building and high-end transportation equipment.

[0003] Traditional acoustic material sound insulation performance evaluation relies on single laboratory tests, which has significant limitations: First, the testing cycle is long, requiring repeated setup and testing of newly developed materials while waiting for physical test results, making it difficult to quickly respond to R&D needs; second, it fails to fully connect the multi-dimensional characteristics of materials, relying only on limited physical parameters for evaluation, ignoring the influence of the material's microstructure on sound insulation performance, resulting in large deviations between the evaluation results and actual scenarios; third, it does not establish a deep correlation between material properties and simulated parameters, and lacks adaptability to complex real-world scenarios (multi-material mixtures, environmental noise interference). This invention proposes a machine learning-based acoustic material sound insulation performance evaluation method. By integrating multimodal image recognition and direct measurement technologies, it comprehensively and accurately acquires the apparent features of acoustic materials. It leverages graph neural networks to deeply mine the intrinsic correlation between apparent features and sound insulation performance, and through deep temporal feature extraction and dynamic performance weight correction, it fully considers the impact of complex environmental factors in real-world scenarios on sound insulation performance, making the evaluation results highly consistent with real-world application scenarios and effectively overcoming the shortcomings of traditional technologies. This not only improves the accuracy of sound insulation performance evaluation of acoustic materials, but also broadens the applicability of evaluation scenarios, promoting the intelligent and precise evaluation of sound insulation performance of acoustic materials, which is of great significance to the upgrading of related industries. Summary of the Invention

[0004] The purpose of this invention is to provide a method for evaluating the sound insulation performance of acoustic materials based on machine learning.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0006] This invention includes the following steps:

[0007] Acoustic tests were conducted on acoustic materials in a laboratory environment to obtain basic sound insulation performance. Multimodal images of acoustic materials were identified to obtain the first apparent characteristics of the acoustic materials, and the second apparent characteristics of the acoustic materials were directly measured. The first apparent characteristics include porosity, porosity, fiber orientation, layered structure state, and RGB reflectivity. The second apparent characteristics include flow resistance, size, and density.

[0008] Based on the first apparent feature and the second apparent feature, finite element models of acoustic materials of different material categories are constructed to perform acoustic simulations and obtain simulated basic sound insulation performance. Based on the simulated basic sound insulation performance and the experimental basic sound insulation performance, the finite element models of acoustic materials are optimized to output key finite element parameters, and an acoustic material graph neural network is constructed.

[0009] The first and second apparent features of the acoustic material to be evaluated are input into the acoustic material graph neural network to obtain the first predictive finite element key parameters. The first predictive finite element key parameters are adjusted and acoustic simulation is performed to obtain the predicted basic sound insulation performance.

[0010] Collect mixed sound insulation data of acoustic materials and environmental conditions in actual scenarios, extract deep temporal features from the mixed sound insulation data to obtain deep sound insulation features, directly measure the deep sound insulation performance of acoustic materials, and construct a deep sound insulation performance prediction model.

[0011] The mixed sound insulation data of the acoustic material to be evaluated in the actual scene and the environmental conditions are input into the deep sound insulation performance prediction model to obtain the predicted deep sound insulation performance. The first performance weight is determined according to the acoustic material design goal, and the first performance weight is corrected according to the environmental conditions to obtain the second performance weight.

[0012] A sound insulation score is determined based on the second performance weight, the predicted depth sound insulation performance, and the predicted basic sound insulation performance. The sound insulation level of the acoustic material to be evaluated is then determined based on the sound insulation score.

[0013] Furthermore, the method for obtaining the first apparent feature of the acoustic material includes:

[0014] Acoustic material multimodal images are acquired, and the SIFT-PSO nonlinear registration algorithm is used to fuse the acoustic material multimodal images to obtain acoustic material registration images. Physical constraint enhancement is applied to the acoustic material registration images according to different regions to obtain acoustic material registration enhancement images. Feature orientation extraction is performed on the acoustic material registration enhancement images to obtain first apparent features. The acoustic material multimodal images include surface microstructure micrographs, material cross-sectional CT scan images, and multispectral surface reflectance maps.

[0015] The steps for fusing multimodal images of acoustic materials using the SIFT-PSO nonlinear registration algorithm are as follows:

[0016] Traditional SIFT feature points are extracted from microscopic images of surface microstructures, edge curvature feature points are extracted from CT scan images of material cross sections, and multi-band gradient feature points are extracted from multispectral surface reflectance images. The extracted feature points are then combined into a hybrid feature point set.

[0017] The particle parameter space range is determined based on the mixed feature point set as the mixed descriptor. Particles are defined and the population size is set according to the particle parameter space indices. The parameter space indices include rotation angles. Scaling factor Coordinate translation and ;

[0018] The feature matching terms and physical constraint terms of the mixed feature point set are determined. The physical constraint fitness function is calculated based on the feature matching terms and physical constraint terms. A hierarchical PSO iteration is performed based on the physical constraint fitness to output the initial transformation matrix. The hierarchical PSO iteration includes coarse optimization and fine optimization. The optimization objective of the coarse optimization is to maximize the feature matching terms. The optimization objective of the fine optimization is to maximize the physical constraint fitness.

[0019] The physical constraint fitness function is expressed as follows:

[0020]

[0021] in For population physical constraint fitness, For the weights of the feature matching terms, For the weights of the physical constraint terms, The penalty coefficient is... To match the number of point pairs, For the image to be registered The mixed descriptor vector of the matching points, To register the reference image The mixed descriptor vector of the matching points, The curvature of the thin plate under the current transformation. The bending curvature is based on material theory.

[0022] A thin-plate distortion model is constructed to calculate the physical distortion of the control points. Based on the physical distortion of the control points and the initial transformation matrix output by PSO iteration, nonlinear fine correction is performed on the control points to output an acoustic material registration image. The expression is:

[0023]

[0024]

[0025] in , Control points Physical distortion in the x and y directions of pixel coordinates. The distortion coefficients control the deformation amplitude, including uniform expansion / contraction distortion coefficient, x-direction bending distortion coefficient, and y-direction bending distortion coefficient. The basis functions describe the deformation modes, including translational components. curvature in the x direction y-direction curvature , Control points Corrected pixel coordinates Scaling factor For rotation angle, , These represent the translation amounts in the x and y directions;

[0026] The primary apparent characteristics of acoustic materials consist of porosity, porosity, fiber orientation, layered structure, and RGB reflectivity.

[0027] Furthermore, the method for constructing an acoustic material graph neural network includes the following steps:

[0028] A finite element model of the acoustic material is constructed based on the first and second apparent features. Boundary conditions are set according to multiple sets of specific frequencies and incident angles, and randomly set finite element parameters are designed to conduct acoustic simulations to obtain the simulated basic sound insulation performance. The finite element parameters are continuously adjusted according to the error between the simulated and experimental basic sound insulation performance until the error is minimized, at which point the key finite element parameters are output. The basic sound insulation performance includes the sound insulation coefficient, sound absorption coefficient, weighted sound insulation, and transmission loss at specific frequencies. The key finite element parameters include Poisson's ratio, Young's modulus, and damping coefficient.

[0029] The first apparent feature, the second apparent feature, and the key parameters of the finite element are used as a set of graph neural data to obtain graph neural data of acoustic materials with different material categories and different apparent features.

[0030] An acoustic material graph neural network is constructed using graph neural network data; the acoustic material graph neural network includes a graph data structure, a physically regularized graph convolutional layer, and a physically decoded output layer;

[0031] The specific steps for determining the graph data structure include: defining micro nodes based on the first apparent feature, defining macro nodes based on the second apparent feature, setting cross-scale physical edges to connect micro-macro nodes, and setting initial features based on feature values ​​and scale types.

[0032] The physically regularized graph convolutional layer is used to constrain feature propagation through physical prior vector knowledge, and includes an aggregation convolutional layer, a physical convolutional layer, and a scale convolutional layer, expressed as:

[0033]

[0034] in For nodes In the The feature vector of the layer, For nodes The set of neighboring nodes, For the neighborhood node index, For the first The graph convolution weight matrix of the layer, For neighboring nodes In the The feature vector of the layer, The physical constraint strength coefficient. It is a multilayer perceptron. These are trainable parameters, including the weight matrix and bias vector. For nodes The physical prior vector;

[0035] The physical decoding output layer performs scale separation, pooling, and physical constraint mapping on the feature output of the final node of the graph data structure to output the first prediction finite element key parameters and prediction confidence.

[0036] Furthermore, the method for obtaining the predicted basic sound insulation performance includes:

[0037] The first and second apparent features of the acoustic material to be evaluated are input into the acoustic material graph neural network to obtain the first predicted finite element key parameters and prediction confidence. The parameter adjustment strategy is selected based on the prediction confidence to obtain the second predicted finite element key parameters.

[0038] The parameter adjustment strategy includes:

[0039] When the prediction confidence is greater than the first confidence threshold, the first prediction finite element key parameter is directly defined as the second prediction finite element key parameter.

[0040] When the prediction confidence level is between the first confidence level threshold and the second confidence level threshold, acoustic simulation is performed based on the first prediction finite element key parameters to obtain the simulated basic sound insulation performance, and acoustic tests are conducted at a fixed frequency to obtain the experimental basic sound insulation performance. The first prediction finite element key parameters are then fine-tuned based on the error between the simulated basic sound insulation performance and the experimental basic sound insulation performance of the acoustic material to be evaluated to obtain the second prediction finite element key parameters.

[0041] When the prediction confidence is less than the second confidence threshold, multiple sets of acoustic tests and acoustic simulations at specific frequencies are repeated to obtain the simulated basic sound insulation performance and the experimental basic sound insulation performance. The finite element parameters of the acoustic material finite element model are continuously adjusted according to the performance error until the performance error is minimized, at which point the second prediction finite element key parameters are output.

[0042] The acoustic material graph neural network is updated using the first apparent feature, the second apparent feature, and the second predicted finite element key parameters.

[0043] Based on the first apparent feature, the second apparent feature, and the second predicted finite element key parameters, a finite element model of the acoustic material is constructed to obtain the predicted basic sound insulation performance.

[0044] Furthermore, the method for obtaining deep sound insulation characteristics includes:

[0045] A distributed fiber optic acoustic sensor array is deployed on the surface of an acoustic material using a grid topology to collect mixed sound insulation data and environmental conditions in a real-world scenario. The deep sound insulation performance is obtained through direct measurement and data inversion based on the mixed sound insulation data. The mixed sound insulation data includes sound pressure signals, three-dimensional vibration acceleration signals, sound pressure differences under broadband noise excitation, and multi-angle incident sound wave data. The environmental conditions include time-series records of temperature, humidity, air pressure, and dust concentration. The deep sound insulation performance includes vibration isolation efficiency, transmissivity, mixed-frequency transmission loss, equivalent complex wavenumber, and equivalent acoustic impedance.

[0046] The mixed sound insulation data is input into the depth sound insulation performance prediction model to extract depth time-series features, thereby obtaining depth sound insulation features and depth environmental features. The specific steps include:

[0047] The sound pressure signal is subjected to frequency domain Wiener filtering and inverse Fourier transform to obtain a denoised signal. Standard empirical mode decomposition (SEM) is then performed on the denoised signal to obtain candidate IMFs. The Hilbert correlation between each IMF component and the vibration signal is calculated. Candidate IMFs are then filtered according to the Hilbert correlation threshold to obtain an IMF set. Finally, sound-vibration IMF reconstruction is performed based on the IMF set to obtain the dominant structural vibration component. Harmony and acoustic transmission dominant components ;

[0048] Determine the environmental state vector based on the environmental state. The environmental state vector, the dominant component of structural vibration, the dominant component of acoustic transmission, the sound pressure difference, and the frequency components of sound waves are input into a multi-scale physical feature extraction network to obtain deep sound insulation features and deep environmental features. The multi-scale physical feature extraction network includes a time-frequency analysis layer, an environmental modulator, and a physical feature fusion layer.

[0049] The time-frequency analysis layer performs a Wigner-Ville distribution on the dominant structural vibration component to obtain the time-frequency energy matrix, and performs a physically constrained wavelet packet transform on the dominant acoustic transmission component to obtain the sub-band energy proportion, expressed as follows:

[0050]

[0051]

[0052] in for Time Frequency The dominant components of structural vibration Time-frequency analysis, For time delay variables, To Complex conjugate operation It is an exponentially decaying window function. The window function attenuation coefficient, It is a complex exponential kernel function. The imaginary unit, Angular frequency, For the first The normalized energy percentage of the subband Let b be the wavelet packet coefficient vector of the b-th sub-band. This represents the total number of sub-bands.

[0053] The environment modulator is based on the environment state vector Environmental feature encoding is performed to generate a physical modulation vector. Based on the physical modulation vector, feature dynamic modulation is performed to obtain the modulation time-frequency features, expressed as:

[0054]

[0055]

[0056] in For physical modulation vector, Encoding vector for the environment, This is the environmental weight matrix. For bias vectors, For modulation time-frequency characteristics, The original time-frequency characteristics;

[0057] The physical feature fusion layer is used to construct acoustic-vibration coupling features based on the time-frequency energy matrix and sub-band energy ratio, and based on the environmental state vector. The deep environmental features are constructed using the physical modulation vector, and the sound-vibration coupling features, sound pressure difference, and sound frequency components are concatenated to output the deep sound insulation features, expressed as follows:

[0058]

[0059]

[0060] in It is a sound-vibration coupling characteristic. As a deep environment feature, The time-frequency energy matrix, The subband energy vector is the partial derivative of the subband energy with respect to frequency. Energy percentage of sub-bands constitute, For physical modulation vector It is a diagonal matrix.

[0061] Furthermore, the method for obtaining the predicted depth sound insulation performance includes:

[0062] The deep sound insulation features, deep environmental features, first apparent features, second apparent features and corresponding deep sound insulation performance are combined into a deep sound insulation set. The deep sound insulation set is randomly divided into a training set and a test set in a 6:4 ratio using the random forest algorithm. A deep sound insulation performance prediction model is constructed. The deep sound insulation performance prediction model is trained using the training set and tested using the test set.

[0063] The deep sound insulation performance prediction model includes a feature fusion layer, a physical enhancement layer, and a performance prediction layer; the accuracy of the deep sound insulation performance prediction model is evaluated using a composite loss function, the expression of which is:

[0064]

[0065] in For composite loss function, The mean squared error loss function is used to evaluate the difference between predicted and true values. , To lose weight, The equivalent impedance predicted by the model. For the density of acoustic materials, The Young's modulus of the material. The equivalent wavenumber predicted by the model. For frequency, For the speed of sound, Model weight matrix The square of the Frobenius norm, For physical constraint regularization terms;

[0066] The feature fusion layer is used to obtain deep environmental representation features by cross-domain feature interaction of deep sound insulation features, deep environmental features, first appearance features and second appearance features according to the feature weight matrix.

[0067] The physical enhancement layer is used to inject physical prior knowledge into the deep environment representation features according to the physical weight matrix to obtain physical enhancement features;

[0068] The performance prediction layer is used to learn the mapping relationship between physical enhancement features and deep sound insulation performance to predict deep sound insulation performance.

[0069] The predicted deep sound insulation performance is obtained by inputting the mixed sound insulation data of the acoustic material to be evaluated in the actual scene and the environmental conditions into the deep sound insulation performance prediction model.

[0070] Furthermore, the method for determining the sound insulation level of the acoustic material to be evaluated includes:

[0071] The first performance weight vector is obtained by directly matching a preset weight reference library according to the acoustic material design objectives; the acoustic material design objectives include design service life and building application.

[0072] The skewness of each environmental indicator is calculated based on the environmental status. The average skewness of each environmental indicator is taken as the environmental correction coefficient. The environmental correction coefficient is used to correct the first performance weight vector to obtain the second performance weight vector.

[0073] The sound insulation performance coefficient is obtained by calculating the ratio of the predicted basic sound insulation performance, the predicted deep sound insulation performance, and the corresponding performance standard reference value. The sound insulation score is calculated based on the sound insulation coefficient and the second performance weight vector. The sound insulation level of the acoustic material to be evaluated is determined based on the sound insulation score.

[0074] The beneficial effects of this invention are:

[0075] This invention is a machine learning-based method for evaluating the sound insulation performance of acoustic materials. Compared with existing technologies, this invention has the following technical advantages:

[0076] This invention, through image recognition, acoustic simulation, graph neural network construction, deep temporal feature extraction, and model building, enhances data preprocessing capabilities in the evaluation of acoustic material sound insulation performance. It reduces the cost and speed of model training, thereby improving the efficiency and accuracy of acoustic material sound insulation performance evaluation. Optimizing this technology significantly saves resources, increases work efficiency, and enables the evaluation of acoustic material sound insulation performance. This provides strong technical support for the research, selection, and application of acoustic materials, and is of great significance for promoting the intelligent and precise evaluation of acoustic material sound insulation performance, as well as for the upgrading of related industries. Attached Figure Description

[0077] Figure 1 This is a flowchart illustrating the steps of the machine learning-based acoustic material sound insulation performance evaluation method of the present invention. Detailed Implementation

[0078] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0079] The machine learning-based method for evaluating the sound insulation performance of acoustic materials of this invention includes the following steps:

[0080] like Figure 1 As shown, this embodiment includes the following steps:

[0081] Acoustic tests were conducted on acoustic materials in a laboratory environment to obtain basic sound insulation performance. Multimodal images of acoustic materials were identified to obtain the first apparent characteristics of the acoustic materials, and the second apparent characteristics of the acoustic materials were directly measured. The first apparent characteristics include porosity, porosity, fiber orientation, layered structure state, and RGB reflectivity. The second apparent characteristics include flow resistance, size, and density.

[0082] Based on the first apparent feature and the second apparent feature, finite element models of acoustic materials of different material categories are constructed to perform acoustic simulations and obtain simulated basic sound insulation performance. Based on the simulated basic sound insulation performance and the experimental basic sound insulation performance, the finite element models of acoustic materials are optimized to output key finite element parameters, and an acoustic material graph neural network is constructed.

[0083] The first and second apparent features of the acoustic material to be evaluated are input into the acoustic material graph neural network to obtain the first predictive finite element key parameters. The first predictive finite element key parameters are adjusted and acoustic simulation is performed to obtain the predicted basic sound insulation performance.

[0084] Collect mixed sound insulation data of acoustic materials and environmental conditions in actual scenarios, extract deep temporal features from the mixed sound insulation data to obtain deep sound insulation features, directly measure the deep sound insulation performance of acoustic materials, and construct a deep sound insulation performance prediction model.

[0085] The mixed sound insulation data of the acoustic material to be evaluated in the actual scene and the environmental conditions are input into the deep sound insulation performance prediction model to obtain the predicted deep sound insulation performance. The first performance weight is determined according to the acoustic material design goal, and the first performance weight is corrected according to the environmental conditions to obtain the second performance weight.

[0086] A sound insulation score is determined based on the second performance weight, the predicted depth sound insulation performance, and the predicted basic sound insulation performance. The sound insulation level of the acoustic material to be evaluated is then determined based on the sound insulation score.

[0087] In this embodiment, the method for obtaining the first apparent feature of the acoustic material includes:

[0088] Acoustic material multimodal images are acquired, and the SIFT-PSO nonlinear registration algorithm is used to fuse the acoustic material multimodal images to obtain acoustic material registration images. Physical constraint enhancement is applied to the acoustic material registration images according to different regions to obtain acoustic material registration enhancement images. Feature orientation extraction is performed on the acoustic material registration enhancement images to obtain first apparent features. The acoustic material multimodal images include surface microstructure micrographs, material cross-sectional CT scan images, and multispectral surface reflectance maps.

[0089] The steps for fusing multimodal images of acoustic materials using the SIFT-PSO nonlinear registration algorithm are as follows:

[0090] Traditional SIFT feature points are extracted from microscopic images of surface microstructures, edge curvature feature points are extracted from CT scan images of material cross sections, and multi-band gradient feature points are extracted from multispectral surface reflectance images. The extracted feature points are then combined into a hybrid feature point set.

[0091] The particle parameter space range is determined based on the mixed feature point set as the mixed descriptor. Particles are defined and the population size is set according to the particle parameter space indices. The parameter space indices include rotation angles. Scaling factor Coordinate translation and ;

[0092] The feature matching terms and physical constraint terms of the mixed feature point set are determined. The physical constraint fitness function is calculated based on the feature matching terms and physical constraint terms. A hierarchical PSO iteration is performed based on the physical constraint fitness to output the initial transformation matrix. The hierarchical PSO iteration includes coarse optimization and fine optimization. The optimization objective of the coarse optimization is to maximize the feature matching terms. The optimization objective of the fine optimization is to maximize the physical constraint fitness.

[0093] The physical constraint fitness function is expressed as follows:

[0094]

[0095] in For population physical constraint fitness, For the weights of the feature matching terms, For the weights of the physical constraint terms, The penalty coefficient is... To match the number of point pairs, For the image to be registered The mixed descriptor vector of the matching points, To register the reference image The mixed descriptor vector of the matching points, The curvature of the thin plate under the current transformation. The bending curvature is based on material theory.

[0096] A thin-plate distortion model is constructed to calculate the physical distortion of the control points. Based on the physical distortion of the control points and the initial transformation matrix output by PSO iteration, nonlinear fine correction is performed on the control points to output an acoustic material registration image. The expression is:

[0097]

[0098]

[0099] in , Control points Physical distortion in the x and y directions of pixel coordinates. The distortion coefficients control the deformation amplitude, including uniform expansion / contraction distortion coefficient, x-direction bending distortion coefficient, and y-direction bending distortion coefficient. The basis functions describe the deformation modes, including translational components. curvature in the x direction y-direction curvature , Control points Corrected pixel coordinates Scaling factor For rotation angle, , These represent the translation amounts in the x and y directions;

[0100] The primary apparent characteristics of acoustic materials consist of porosity, void ratio, fiber orientation, layered structure, and RGB reflectivity.

[0101] In practical evaluation, the extracted feature points are combined into a hybrid feature point set. Based on this hybrid feature point set as a hybrid descriptor, the particle parameter space range is determined. , , Define each particle Set population size Perform particle initialization and determine the curvature of the thin plate under the current transformation. By inversely inferring the material's theoretical bending curvature through the displacement of characteristic points. Calculated using Young's modulus, the weights of the feature matching terms are taken. Weights of physical constraint terms Penalty coefficient During coarse optimization, a global search strategy is used to update particles (with an inertia weight of 0.9), while during fine optimization, a local development strategy is used to update particles (with an inertia weight of 0.4).

[0102] The physical constraint enhancement of the acoustic material registration image according to different regions specifically includes using a morphological watershed improved algorithm to process the pore region of the acoustic material registration image (constraining the material to conform to the pore size distribution model), using directional Gabor filtering to process the fiber region of the acoustic material registration image (limiting the fiber curvature radius to ≥5μm), and using a multi-scale edge sharpening algorithm to process the interlayer bonding region of the acoustic material registration image (performing layer thickness continuity constraints).

[0103] The method for obtaining the first apparent feature through feature orientation extraction includes: using a three-dimensional connected component labeling algorithm and an ellipsoid fitting algorithm to process the pore region of the acoustic material registration and enhancement image to obtain the porosity and porosity ratio of the acoustic material; using a Fourier frequency domain directional spectrum analysis algorithm (specific steps include extracting the Fourier transform amplitude spectrum of the ROI region, detecting the principal energy axis direction in the frequency domain to obtain the fiber principal orientation angle, and calculating the radial energy distribution entropy to obtain the fiber orientation dispersion) to process the fiber region of the acoustic material registration and enhancement image to obtain the fiber orientation of the acoustic material (including the fiber principal orientation angle and fiber orientation dispersion); using a cross-layer grayscale projection curve fitting algorithm to process the interlayer bonding region of the acoustic material registration and enhancement image to obtain the layered structure state of the acoustic material; and using a spectral unmixing reflectance inversion algorithm to process the acoustic material registration and enhancement image to obtain the RGB reflectance; the specific steps of the spectral unmixing reflectance inversion include calibrating the reflectance mapping matrix between the multispectral image and the standard color chart, and calculating the equivalent reflectance of the RGB three channels based on the reflectance mapping matrix to obtain the RGB reflectance.

[0104] The specific steps for fitting the cross-layer grayscale projection curve include generating a grayscale projection curve along the stacking direction, using the difference of Gaussians (DoG) to detect interlayer peaks and valleys to obtain the number of fiber layers, and calculating the interlayer bonding index. The layered structure state includes the number of fiber layers and the corresponding interlayer bonding index, expressed as follows:

[0105]

[0106] in This is the interlayer bonding index. The number of fiber layers, This represents the average gray level within the layer. For the first Layer interface grayscale gradient;

[0107] Feature orientation extraction was performed on the A1 acoustic material registration enhancement image to obtain the first apparent features: porosity 0.85, porosity ratio 1.2, fiber principal orientation angle 30°, fiber orientation dispersion 0.2, number of fiber layers 5, interlayer bonding index 0.7, and RGB reflectance (120, 150, 180).

[0108] In this embodiment, the method for constructing an acoustic material graph neural network includes the following steps:

[0109] A finite element model of the acoustic material is constructed based on the first and second apparent features. Boundary conditions are set according to multiple sets of specific frequencies and incident angles, and randomly set finite element parameters are designed to conduct acoustic simulations to obtain the simulated basic sound insulation performance. The finite element parameters are continuously adjusted according to the error between the simulated and experimental basic sound insulation performance until the error is minimized, at which point the key finite element parameters are output. The basic sound insulation performance includes the sound insulation coefficient, sound absorption coefficient, weighted sound insulation, and transmission loss at specific frequencies. The key finite element parameters include Poisson's ratio, Young's modulus, and damping coefficient.

[0110] The first apparent feature, the second apparent feature, and the key parameters of the finite element are used as a set of graph neural data to obtain graph neural data of acoustic materials with different material categories and different apparent features.

[0111] An acoustic material graph neural network is constructed using graph neural network data; the acoustic material graph neural network includes a graph data structure, a physically regularized graph convolutional layer, and a physically decoded output layer;

[0112] The specific steps for determining the graph data structure include: defining micro nodes based on the first apparent feature, defining macro nodes based on the second apparent feature, setting cross-scale physical edges to connect micro-macro nodes, and setting initial features based on feature values ​​and scale types.

[0113] The physically regularized graph convolutional layer is used to constrain feature propagation through physical prior vector knowledge, and includes an aggregation convolutional layer, a physical convolutional layer, and a scale convolutional layer, expressed as:

[0114]

[0115] in For nodes In the The feature vector of the layer, For nodes The set of neighboring nodes, For the neighborhood node index, For the first The graph convolution weight matrix of the layer, For neighboring nodes In the The feature vector of the layer, The physical constraint strength coefficient. It is a multilayer perceptron. These are trainable parameters, including the weight matrix and bias vector. For nodes The physical prior vector;

[0116] The physical decoding output layer performs scale separation pooling and physical constraint mapping on the feature output of the final node of the graph data structure to output the first prediction finite element key parameters and prediction confidence.

[0117] In practical evaluation, taking the determination of key finite element parameters for A1 acoustic material as an example (the second apparent characteristics obtained by measurement are: flow resistance 3000 Pa・s / m², thickness 0.05 m, density 25 kg / m³), the boundary conditions are set as follows: specific frequencies 100 / 500 / 1000 / 2000 Hz, incident angle 0°, and finite element parameters are set as follows (Poisson's ratio 0.3, Young's modulus 3 × 10⁻⁶). 4 With Pa and damping coefficient 0.05, acoustic simulations were performed to calculate the mean square error of the simulated and experimental basic sound insulation performance. The finite element parameters were continuously adjusted (iterative optimization). When Poisson's ratio was adjusted to 0.32 and Young's modulus to 3.2 × 10⁻⁶, the results were obtained. 4 When Pa and damping coefficient are adjusted to 0.06, the mean square error of the basic sound insulation performance in the simulation and experiment is minimized. At this point, the key finite element parameters of the A1 acoustic material are output (Poisson's ratio 0.32, Young's modulus 3×10⁻⁶). 4 Pa, damping coefficient 0.06);

[0118] Acoustic material graph neural network is constructed by acquiring graph neural network data of acoustic materials with different material categories and appearance characteristics;

[0119] When determining the graph data structure, porosity, porosity, fiber principal orientation angle, fiber orientation dispersion, number of fiber layers, interlayer bonding index, and RGB reflectivity are used as micro-nodes, while flow resistance, density, thickness, and size are used as macro-nodes. The weight of the physical edge across scales is calculated as exp{-|node i feature scale level-node j feature scale level-| / 0.5}, where the feature scale level of macro-nodes is 1 and the feature scale level of micro-nodes is 0.

[0120] In the physically regularized graph convolutional layer, the aggregation convolutional layer is used to aggregate the mean of neighborhood features, the physical convolutional layer adds fiber-pore association constraints through the injection of physical regularization terms, and the scale convolutional layer forces micro-features to match macro-rheological properties through cross-scale feature crossing.

[0121] In the physical decoding output layer, the expressions for scale-separated pooling and physical constraint mapping operations are:

[0122]

[0123]

[0124]

[0125] in For nodes In the feature vectors of the 3rd layer, For microscale feature vectors, through the set of micro nodes China The maximum pooling operation is performed to obtain the result. As a macroscopic-scale feature vector, it is obtained through the macroscopic node set. China The result is obtained by performing mean pooling. , , These are intermediate results for Poisson's ratio, Young's modulus, and damping coefficient after processing with the Softplus activation function. This is the weight matrix of the output layer;

[0126] Poisson's ratio correction Intermediate results from Poisson's ratio Substitution Young's modulus correction obtained Intermediate results from Young's modulus Substitution Obtain, damping coefficient correction Intermediate results from damping coefficient Substitution get.

[0127] In this embodiment, the method for obtaining the predicted basic sound insulation performance includes:

[0128] The first and second apparent features of the acoustic material to be evaluated are input into the acoustic material graph neural network to obtain the first predicted finite element key parameters and prediction confidence. The parameter adjustment strategy is selected based on the prediction confidence to obtain the second predicted finite element key parameters.

[0129] The parameter adjustment strategy includes:

[0130] When the prediction confidence is greater than the first confidence threshold, the first prediction finite element key parameter is directly defined as the second prediction finite element key parameter.

[0131] When the prediction confidence level is between the first confidence level threshold and the second confidence level threshold, acoustic simulation is performed based on the first prediction finite element key parameters to obtain the simulated basic sound insulation performance, and acoustic tests are conducted at a fixed frequency to obtain the experimental basic sound insulation performance. The first prediction finite element key parameters are then fine-tuned based on the error between the simulated basic sound insulation performance and the experimental basic sound insulation performance of the acoustic material to be evaluated to obtain the second prediction finite element key parameters.

[0132] When the prediction confidence is less than the second confidence threshold, multiple sets of acoustic tests and acoustic simulations at specific frequencies are repeated to obtain the simulated basic sound insulation performance and the experimental basic sound insulation performance. The finite element parameters of the acoustic material finite element model are continuously adjusted according to the performance error until the performance error is minimized, at which point the second prediction finite element key parameters are output.

[0133] The acoustic material graph neural network is updated using the first apparent feature, the second apparent feature, and the second predicted finite element key parameters.

[0134] Based on the first apparent feature, the second apparent feature, and the second predicted finite element key parameters, an acoustic material finite element model is constructed to obtain the predicted basic sound insulation performance.

[0135] In practical evaluation, taking a certain acoustic material as an example, the corresponding first and second apparent features are input into the acoustic material graph neural network to obtain the first predicted finite element key parameters and prediction confidence. The first confidence threshold is 0.9 and the second confidence threshold is 0.6. If the prediction confidence is 0.92, which is greater than the first confidence threshold, the first predicted finite element key parameters are directly used as the second predicted finite element key parameters. If the prediction confidence is 0.77, which is between the first and second confidence thresholds, the basic sound insulation performance of the test is obtained through acoustic testing at a fixed frequency (1000Hz). The first predicted finite element key parameters are fine-tuned according to the error between the simulated basic sound insulation performance and the experimental basic sound insulation performance of the acoustic material to be evaluated to obtain the second predicted finite element key parameters. If the prediction confidence is 0.68, which is less than the second confidence threshold, steps S1 and S2 are repeated to obtain the optimal finite element key parameters as the second predicted finite element key parameters.

[0136] The first and second apparent features of the A2 acoustic material to be evaluated are input into the acoustic material graph neural network to obtain the first predictive finite element key parameters (Poisson's ratio 0.33, Young's modulus 3×10⁻⁶). 4 Using Pa (damping coefficient 0.05) and a prediction confidence level of 0.85, basic sound insulation performance was obtained through acoustic testing at a fixed frequency (1000Hz). The first predictive finite element key parameters were fine-tuned based on the error between the simulated and experimental basic sound insulation performance of the acoustic material to be evaluated, resulting in the second predictive finite element key parameters (Poisson's ratio 0.34, Young's modulus 3.1 × 10⁻⁶). 4 Pa, damping coefficient 0.052);

[0137] A finite element model of acoustic materials is constructed to obtain the predicted basic sound insulation performance (including sound insulation coefficient, sound absorption coefficient, weighted sound insulation amount, and transmission loss at frequencies of 100Hz / 1000Hz / 2000Hz).

[0138] In this embodiment, the method for obtaining deep sound insulation characteristics includes:

[0139] A distributed fiber optic acoustic sensor array is deployed on the surface of an acoustic material using a grid topology to collect mixed sound insulation data and environmental conditions in a real-world scenario. The deep sound insulation performance is obtained through direct measurement and data inversion based on the mixed sound insulation data. The mixed sound insulation data includes sound pressure signals, three-dimensional vibration acceleration signals, sound pressure differences under broadband noise excitation, and multi-angle incident sound wave data. The environmental conditions include time-series records of temperature, humidity, air pressure, and dust concentration. The deep sound insulation performance includes vibration isolation efficiency, transmissivity, mixed-frequency transmission loss, equivalent complex wavenumber, and equivalent acoustic impedance.

[0140] The mixed sound insulation data is input into the depth sound insulation performance prediction model to extract depth time-series features, thereby obtaining depth sound insulation features and depth environmental features. The specific steps include:

[0141] The sound pressure signal is subjected to frequency domain Wiener filtering and inverse Fourier transform to obtain a denoised signal. Standard empirical mode decomposition (SEM) is then performed on the denoised signal to obtain candidate IMFs. The Hilbert correlation between each IMF component and the vibration signal is calculated. Candidate IMFs are then filtered according to the Hilbert correlation threshold to obtain an IMF set. Finally, sound-vibration IMF reconstruction is performed based on the IMF set to obtain the dominant structural vibration component. Harmony and acoustic transmission dominant components ;

[0142] Determine the environmental state vector based on the environmental state. The environmental state vector, the dominant component of structural vibration, the dominant component of acoustic transmission, the sound pressure difference, and the frequency components of sound waves are input into a multi-scale physical feature extraction network to obtain deep sound insulation features and deep environmental features. The multi-scale physical feature extraction network includes a time-frequency analysis layer, an environmental modulator, and a physical feature fusion layer.

[0143] The time-frequency analysis layer performs a Wigner-Ville distribution on the dominant structural vibration component to obtain the time-frequency energy matrix, and performs a physically constrained wavelet packet transform on the dominant acoustic transmission component to obtain the sub-band energy proportion, expressed as follows:

[0144]

[0145]

[0146] in for Time Frequency The dominant components of structural vibration Time-frequency analysis, For time delay variables, To Complex conjugate operation It is an exponentially decaying window function. The window function attenuation coefficient, It is a complex exponential kernel function. The imaginary unit, Angular frequency, For the first The normalized energy percentage of the subband Let b be the wavelet packet coefficient vector of the b-th sub-band. This represents the total number of sub-bands.

[0147] The environment modulator is based on the environment state vector Environmental feature encoding is performed to generate a physical modulation vector. Based on the physical modulation vector, feature dynamic modulation is performed to obtain the modulation time-frequency features, expressed as:

[0148]

[0149]

[0150] in For physical modulation vector, Encoding vector for the environment, This is the environmental weight matrix. For bias vectors, For modulation time-frequency characteristics, The original time-frequency characteristics;

[0151] The physical feature fusion layer is used to construct acoustic-vibration coupling features based on the time-frequency energy matrix and sub-band energy ratio, and based on the environmental state vector. The deep environmental features are constructed using the physical modulation vector, and the sound-vibration coupling features, sound pressure difference, and sound frequency components are concatenated to output the deep sound insulation features, expressed as follows:

[0152]

[0153]

[0154] in It is a sound-vibration coupling characteristic. As a deep environment feature, The time-frequency energy matrix, The subband energy vector is the partial derivative of the subband energy with respect to frequency. Energy percentage of sub-bands constitute, For physical modulation vector A matrix that is diagonal;

[0155] In practical assessments, a frequency-varying noise spectrum model is constructed based on dust concentration. For sound pressure signals Perform frequency domain Wiener filtering to obtain the filtered signal For the filtered signal The denoised signal is obtained by performing an inverse Fourier transform. For noise reduction signals Perform standard empirical mode decomposition to obtain a set of candidate IMFs, and calculate the relationship between each IMF component and the vibration signal. Hilbert correlation, screening The IMF set is obtained from the IMF components, expressed as:

[0156]

[0157] in For the first Each IMF component and Vibration signal at all times Hilbert correlation, for Time of the first The result of performing Hilbert transform on each IMF component;

[0158] Based on the IMF set, acoustic-vibration IMF recombination is performed, expressed as follows:

[0159]

[0160]

[0161] in As the dominant component of structural vibration, For a high vibration-related IMF index set, No. One IMF component, It is the dominant component of acoustic transmission. The IMF index set is dominated by acoustic transmission.

[0162] Environment state vector ,in , , These are the mean environmental noise, the maximum-minimum noise deviation, and the standard deviation of environmental noise, respectively. , , These represent the mean ambient temperature, the maximum-minimum temperature deviation, and the standard deviation of ambient temperature, respectively. , , These represent the mean ambient humidity, the maximum-minimum humidity deviation, and the standard deviation of ambient humidity, respectively. , These represent the mean ambient air pressure and the maximum-minimum air pressure deviation, respectively. , The average environmental dust concentration and the maximum-minimum dust concentration deviation are used to obtain the sound wave frequency components by performing empirical mode decomposition on the multi-angle incident sound wave data; the sound wave frequency components include the sound wave period and the sound wave frequency band ratio.

[0163] In the time-frequency analysis layer of the multi-scale physical feature extraction network, the wavelet basis function is dynamically selected according to air pressure / temperature: the Meyer wavelet is used when the temperature is high (>35℃), the Biorthogonal 3.1 wavelet is used when the humidity is high (RH>80%), and the Db4 wavelet is used when the pressure is low (<90kPa).

[0164] Collect mixed sound insulation data and environmental conditions of A2 acoustic materials in a real-world scenario. Obtain the depth sound insulation performance by directly measuring and inverting the mixed sound insulation data. Determine the environmental state vector based on the environmental conditions. By inputting the environmental state vector, the dominant component of structural vibration, the dominant component of acoustic transmission, the sound pressure difference, and the frequency components of sound waves into a multi-scale physical feature extraction network, we can obtain deep sound insulation features and deep environmental features.

[0165] In this embodiment, the method for obtaining the predicted depth sound insulation performance includes:

[0166] The deep sound insulation features, deep environmental features, first apparent features, second apparent features and corresponding deep sound insulation performance are combined into a deep sound insulation set. The deep sound insulation set is randomly divided into a training set and a test set in a 6:4 ratio using the random forest algorithm. A deep sound insulation performance prediction model is constructed. The deep sound insulation performance prediction model is trained using the training set and tested using the test set.

[0167] The deep sound insulation performance prediction model includes a feature fusion layer, a physical enhancement layer, and a performance prediction layer; the accuracy of the deep sound insulation performance prediction model is evaluated using a composite loss function, the expression of which is:

[0168]

[0169] in For composite loss function, The mean squared error loss function is used to evaluate the difference between predicted and true values. , To lose weight, The equivalent impedance predicted by the model. For the density of acoustic materials, The Young's modulus of the material. The equivalent wavenumber predicted by the model. For frequency, For the speed of sound, Model weight matrix The square of the Frobenius norm, For physical constraint regularization terms;

[0170] The feature fusion layer is used to obtain deep environmental representation features by cross-domain feature interaction of deep sound insulation features, deep environmental features, first appearance features and second appearance features according to the feature weight matrix.

[0171] The physical enhancement layer is used to inject physical prior knowledge into the deep environment representation features according to the physical weight matrix to obtain physical enhancement features;

[0172] The performance prediction layer is used to learn the mapping relationship between physical enhancement features and deep sound insulation performance to predict deep sound insulation performance.

[0173] The predicted deep sound insulation performance is obtained by inputting the mixed sound insulation data of the acoustic material to be evaluated in the actual scenario and the environmental conditions into the deep sound insulation performance prediction model.

[0174] In practical evaluation, the loss weights of the composite loss function in the deep sound insulation performance prediction model are... , With values ​​of 0.3 and 0.1, a physical constraint regularization term is added to the feature fusion layer to penalize the invalid interaction between material properties and acoustic vibration features. The physical enhancement layer applies acoustic energy conservation constraints to the physical weight matrix and injects deep environmental representation features through the Swish activation function to obtain physical enhancement features.

[0175] The predicted deep sound insulation performance is obtained by inputting the mixed sound insulation data of the acoustic material to be evaluated in the actual scene and the environmental conditions into the deep sound insulation performance prediction model.

[0176] In this embodiment, the method for determining the sound insulation level of the acoustic material to be evaluated includes:

[0177] The first performance weight vector is obtained by directly matching a preset weight reference library according to the acoustic material design objectives; the acoustic material design objectives include design service life and building application.

[0178] The skewness of each environmental indicator is calculated based on the environmental status. The average skewness of each environmental indicator is taken as the environmental correction coefficient. The environmental correction coefficient is used to correct the first performance weight vector to obtain the second performance weight vector.

[0179] The sound insulation performance coefficient is obtained by calculating the ratio of the predicted basic sound insulation performance, the predicted deep sound insulation performance, and the corresponding performance standard reference value. The sound insulation score is calculated based on the sound insulation coefficient and the second performance weight vector. The sound insulation level of the acoustic material to be evaluated is determined based on the sound insulation score.

[0180] In the actual assessment, the first performance weight vector is obtained by directly matching the preset weight reference library according to the design goals of the A2 acoustic material to be evaluated (service life of 10 years, office building project). The skewness of each environmental indicator is calculated as (mean - mode) / standard deviation based on the environmental conditions. The mean skewness of each environmental indicator, 0.89, is taken as the environmental correction coefficient. The second performance weight vector is obtained by multiplying the environmental correction coefficient by the first performance weight vector. The ratios of the predicted basic sound insulation performance and the predicted deep sound insulation performance to the corresponding performance standard reference values ​​are calculated to obtain the sound insulation performance coefficient (for basic...). In terms of sound insulation performance, the ratios of sound insulation coefficient, sound absorption coefficient, weighted sound insulation amount, and transmission loss at a specific frequency to the corresponding performance standard reference value are calculated at frequencies of 100 / 500 / 1000 / 2000Hz, and then the average value is taken to obtain the corresponding sound insulation performance coefficient. The sound insulation coefficient is multiplied by the weight corresponding to the second performance weight vector to obtain a sound insulation score of 8.6. According to the correspondence between sound insulation level and sound insulation score {Level 3 - [0,5], Level 2 - (5,8], Level 1 (8,9], Special Level (9,+∞)}, the A2 acoustic material to be evaluated is determined to be a Level 1 sound insulation material.

[0181] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the sound insulation performance of acoustic materials based on machine learning, characterized in that, Includes the following steps: S1. Acoustic tests are conducted on acoustic materials in a laboratory environment to obtain basic sound insulation performance. Multimodal images of acoustic materials are identified to obtain the first apparent characteristics of the acoustic materials, and the second apparent characteristics of the acoustic materials are directly measured. The first apparent characteristics include porosity, porosity, fiber orientation, layered structure state, and RGB reflectivity. The second apparent characteristics include flow resistance, size, and density. S2. Based on the first apparent feature and the second apparent feature, construct finite element models of acoustic materials of different material categories to perform acoustic simulation and obtain simulated basic sound insulation performance. Based on the simulated basic sound insulation performance and the experimental basic sound insulation performance, optimize the finite element model of acoustic materials, output key finite element parameters, and construct a graph neural network for acoustic materials. S3. Input the first and second apparent features of the acoustic material to be evaluated into the acoustic material graph neural network to obtain the first predictive finite element key parameters, adjust the predictive finite element key parameters and perform acoustic simulation to obtain the predicted basic sound insulation performance. S4. Collect mixed sound insulation data of acoustic materials and environmental conditions in actual scenarios, extract deep temporal features from the mixed sound insulation data to obtain deep sound insulation features, directly measure the deep sound insulation performance of acoustic materials, and construct a deep sound insulation performance prediction model. S5. Input the mixed sound insulation data of the acoustic material to be evaluated in the actual scene and the environmental conditions into the deep sound insulation performance prediction model to obtain the predicted deep sound insulation performance. Determine the first performance weight according to the acoustic material design goal, and correct the first performance weight according to the environmental conditions to obtain the second performance weight. S6. Determine the sound insulation score based on the second performance weight, the predicted depth sound insulation performance, and the predicted basic sound insulation performance, and determine the sound insulation level of the acoustic material to be evaluated based on the sound insulation score.

2. The method for evaluating the sound insulation performance of acoustic materials based on machine learning according to claim 1, characterized in that, The method for obtaining the first apparent feature of an acoustic material includes: Acoustic material multimodal images are acquired, and the SIFT-PSO nonlinear registration algorithm is used to fuse the acoustic material multimodal images to obtain acoustic material registration images. Physical constraint enhancement is applied to the acoustic material registration images according to different regions to obtain acoustic material registration enhancement images. Feature orientation extraction is performed on the acoustic material registration enhancement images to obtain first apparent features. The acoustic material multimodal images include surface microstructure micrographs, material cross-sectional CT scan images, and multispectral surface reflectance maps. The steps for fusing multimodal images of acoustic materials using the SIFT-PSO nonlinear registration algorithm are as follows: Traditional SIFT feature points are extracted from microscopic images of surface microstructures, edge curvature feature points are extracted from CT scan images of material cross sections, and multi-band gradient feature points are extracted from multispectral surface reflectance images. The extracted feature points are then combined into a hybrid feature point set. The particle parameter space range is determined based on the mixed feature point set as the mixed descriptor. Particles are defined and the population size is set according to the particle parameter space indices. The parameter space indices include rotation angles. Scaling factor Coordinate translation and ; The feature matching terms and physical constraint terms of the mixed feature point set are determined. The physical constraint fitness function is calculated based on the feature matching terms and physical constraint terms. A hierarchical PSO iteration is performed based on the physical constraint fitness to output the initial transformation matrix. The hierarchical PSO iteration includes coarse optimization and fine optimization. The optimization objective of the coarse optimization is to maximize the feature matching terms. The optimization objective of the fine optimization is to maximize the physical constraint fitness. The physical constraint fitness function is expressed as follows: ； in For population physical constraint fitness, For the weights of the feature matching terms, For the weights of the physical constraint terms, The penalty coefficient is... To match the number of point pairs, For the image to be registered The mixed descriptor vector of the matching points, To register the reference image The mixed descriptor vector of the matching points, The curvature of the thin plate under the current transformation. The bending curvature is based on material theory. A thin-plate distortion model is constructed to calculate the physical distortion of the control points. Based on the physical distortion of the control points and the initial transformation matrix output by PSO iteration, nonlinear fine correction is performed on the control points to output an acoustic material registration image. The expression is: ； ； in , Control points Physical distortion in the x and y directions of pixel coordinates. The distortion coefficients control the deformation amplitude, including uniform expansion / contraction distortion coefficient, x-direction bending distortion coefficient, and y-direction bending distortion coefficient. The basis functions describe the deformation modes, including translational components. curvature in the x direction y-direction curvature , Control points Corrected pixel coordinates Scaling factor For rotation angle, , These represent the translation amounts in the x and y directions; The primary apparent characteristics of acoustic materials consist of porosity, porosity, fiber orientation, layered structure, and RGB reflectivity.

3. The method for evaluating the sound insulation performance of acoustic materials based on machine learning according to claim 1, characterized in that, The method for constructing an acoustic material graph neural network includes the following steps: A finite element model of the acoustic material is constructed based on the first and second apparent features. Boundary conditions are set according to multiple sets of specific frequencies and incident angles, and randomly set finite element parameters are designed to conduct acoustic simulations to obtain the simulated basic sound insulation performance. The finite element parameters are continuously adjusted according to the error between the simulated and experimental basic sound insulation performance until the error is minimized, at which point the key finite element parameters are output. The basic sound insulation performance includes the sound insulation coefficient, sound absorption coefficient, weighted sound insulation, and transmission loss at specific frequencies. The key finite element parameters include Poisson's ratio, Young's modulus, and damping coefficient. The first apparent feature, the second apparent feature, and the key parameters of the finite element are used as a set of graph neural data to obtain graph neural data of acoustic materials with different material categories and different apparent features. An acoustic material graph neural network is constructed using graph neural network data; the acoustic material graph neural network includes a graph data structure, a physically regularized graph convolutional layer, and a physically decoded output layer; The specific steps for determining the graph data structure include: defining micro nodes based on the first apparent feature, defining macro nodes based on the second apparent feature, setting cross-scale physical edges to connect micro-macro nodes, and setting initial features based on feature values ​​and scale types. The physically regularized graph convolutional layer is used to constrain feature propagation through physical prior vector knowledge, and includes an aggregation convolutional layer, a physical convolutional layer, and a scale convolutional layer, expressed as: ； in For nodes In the The feature vector of the layer, For nodes The set of neighboring nodes, For the neighborhood node index, For the first The graph convolution weight matrix of the layer, For neighboring nodes In the The feature vector of the layer, The physical constraint strength coefficient. It is a multilayer perceptron. These are trainable parameters, including the weight matrix and bias vector. For nodes The physical prior vector; The physical decoding output layer performs scale separation, pooling, and physical constraint mapping on the feature output of the final node of the graph data structure to output the first prediction finite element key parameters and prediction confidence.

4. The method for evaluating the sound insulation performance of acoustic materials based on machine learning according to claim 3, characterized in that, The method for obtaining the predicted basic sound insulation performance includes: The first and second apparent features of the acoustic material to be evaluated are input into the acoustic material graph neural network to obtain the first predicted finite element key parameters and prediction confidence. The parameter adjustment strategy is selected based on the prediction confidence to obtain the second predicted finite element key parameters. The parameter adjustment strategy includes: When the prediction confidence is greater than the first confidence threshold, the first prediction finite element key parameter is directly defined as the second prediction finite element key parameter. When the prediction confidence level is between the first confidence level threshold and the second confidence level threshold, acoustic simulation is performed based on the first prediction finite element key parameters to obtain the simulated basic sound insulation performance, and acoustic tests are conducted at a fixed frequency to obtain the experimental basic sound insulation performance. The first prediction finite element key parameters are then fine-tuned based on the error between the simulated basic sound insulation performance and the experimental basic sound insulation performance of the acoustic material to be evaluated to obtain the second prediction finite element key parameters. When the prediction confidence is less than the second confidence threshold, multiple sets of acoustic tests and acoustic simulations at specific frequencies are repeated to obtain the simulated basic sound insulation performance and the experimental basic sound insulation performance. The finite element parameters of the acoustic material finite element model are continuously adjusted according to the performance error until the performance error is minimized, at which point the second prediction finite element key parameters are output. The acoustic material graph neural network is updated using the first apparent feature, the second apparent feature, and the second predicted finite element key parameters. Based on the first apparent feature, the second apparent feature, and the second predicted finite element key parameters, a finite element model of the acoustic material is constructed to obtain the predicted basic sound insulation performance.

5. The method for evaluating the sound insulation performance of acoustic materials based on machine learning according to claim 3, characterized in that, The method for obtaining deep sound insulation characteristics includes: A distributed fiber optic acoustic sensor array is deployed on the surface of an acoustic material using a grid topology to collect mixed sound insulation data and environmental conditions in a real-world scenario. The deep sound insulation performance is obtained through direct measurement and data inversion based on the mixed sound insulation data. The mixed sound insulation data includes sound pressure signals, three-dimensional vibration acceleration signals, sound pressure differences under broadband noise excitation, and multi-angle incident sound wave data. The environmental conditions include time-series records of temperature, humidity, air pressure, and dust concentration. The deep sound insulation performance includes vibration isolation efficiency, transmissivity, mixed-frequency transmission loss, equivalent complex wavenumber, and equivalent acoustic impedance. The mixed sound insulation data is input into the depth sound insulation performance prediction model to extract depth time-series features and depth environmental features. The specific steps include: The sound pressure signal is subjected to frequency domain Wiener filtering and inverse Fourier transform to obtain a denoised signal. Standard empirical mode decomposition (SEM) is then performed on the denoised signal to obtain candidate IMFs. The Hilbert correlation between each IMF component and the vibration signal is calculated. Candidate IMFs are then filtered according to the Hilbert correlation threshold to obtain an IMF set. Finally, sound-vibration IMF reconstruction is performed based on the IMF set to obtain the dominant structural vibration component. Harmony and acoustic transmission dominant components ; Determine the environmental state vector based on the environmental state. The environmental state vector, the dominant component of structural vibration, the dominant component of acoustic transmission, the sound pressure difference, and the frequency components of sound waves are input into a multi-scale physical feature extraction network to obtain deep sound insulation features and deep environmental features. The multi-scale physical feature extraction network includes a time-frequency analysis layer, an environmental modulator, and a physical feature fusion layer. The time-frequency analysis layer performs a Wigner-Ville distribution on the dominant structural vibration component to obtain the time-frequency energy matrix, and performs a physically constrained wavelet packet transform on the dominant acoustic transmission component to obtain the sub-band energy proportion, expressed as follows: ； ； in for Time Frequency The dominant components of structural vibration Time-frequency analysis, For time delay variables, To Complex conjugate operation It is an exponentially decaying window function. The window function attenuation coefficient, It is a complex exponential kernel function. The imaginary unit, Angular frequency, For the first The normalized energy percentage of the subband Let b be the wavelet packet coefficient vector of the b-th sub-band. This represents the total number of sub-bands. The environment modulator is based on the environment state vector Environmental feature encoding is performed to generate a physical modulation vector. Based on the physical modulation vector, feature dynamic modulation is performed to obtain the modulation time-frequency features, expressed as: ； ； in For physical modulation vector, Encoding vector for the environment, This is the environmental weight matrix. For bias vectors, For modulation time-frequency characteristics, The original time-frequency characteristics; The physical feature fusion layer is used to construct acoustic-vibration coupling features based on the time-frequency energy matrix and sub-band energy ratio, and based on the environmental state vector. The deep environmental features are constructed using the physical modulation vector, and the sound-vibration coupling features, sound pressure difference, and sound frequency components are concatenated to output the deep sound insulation features, expressed as follows: ； ； in It is a sound-vibration coupling characteristic. As a deep environment feature, The time-frequency energy matrix, The subband energy vector is the partial derivative of the subband energy with respect to frequency. Energy percentage of sub-bands constitute, For physical modulation vector It is a diagonal matrix.

6. The method for evaluating the sound insulation performance of acoustic materials based on machine learning according to claim 1, characterized in that, The method for obtaining the predicted depth sound insulation performance includes: The deep sound insulation features, deep environmental features, first apparent features, second apparent features and corresponding deep sound insulation performance are combined into a deep sound insulation set. The deep sound insulation set is randomly divided into a training set and a test set in a 6:4 ratio using the random forest algorithm. A deep sound insulation performance prediction model is constructed. The deep sound insulation performance prediction model is trained using the training set and tested using the test set. The deep sound insulation performance prediction model includes a feature fusion layer, a physical enhancement layer, and a performance prediction layer; the accuracy of the deep sound insulation performance prediction model is evaluated using a composite loss function, the expression of which is: ； in For composite loss function, The mean squared error loss function is used to evaluate the difference between predicted and true values. , To lose weight, The equivalent impedance predicted by the model. For the density of acoustic materials, The Young's modulus of the material. The equivalent wavenumber predicted by the model. For frequency, For the speed of sound, Model weight matrix The square of the Frobenius norm, For physical constraint regularization terms; The feature fusion layer is used to obtain deep environmental representation features by cross-domain feature interaction of deep sound insulation features, deep environmental features, first appearance features and second appearance features according to the feature weight matrix. The physical enhancement layer is used to inject physical prior knowledge into the deep environment representation features according to the physical weight matrix to obtain physical enhancement features; The performance prediction layer is used to learn the mapping relationship between physical enhancement features and deep sound insulation performance to predict deep sound insulation performance. The predicted deep sound insulation performance is obtained by inputting the mixed sound insulation data of the acoustic material to be evaluated in the actual scene and the environmental conditions into the deep sound insulation performance prediction model.

7. The method for evaluating the sound insulation performance of acoustic materials based on machine learning according to claim 1, characterized in that, The method for determining the sound insulation level of the acoustic material to be evaluated includes: The first performance weight vector is obtained by directly matching a preset weight reference library according to the acoustic material design objectives; the acoustic material design objectives include design service life and building application. The skewness of each environmental indicator is calculated based on the environmental status. The average skewness of each environmental indicator is taken as the environmental correction coefficient. The environmental correction coefficient is used to correct the first performance weight vector to obtain the second performance weight vector. The sound insulation performance coefficient is obtained by calculating the ratio of the predicted basic sound insulation performance, the predicted deep sound insulation performance, and the corresponding performance standard reference value. The sound insulation score is calculated based on the sound insulation coefficient and the second performance weight vector. The sound insulation level of the acoustic material to be evaluated is determined based on the sound insulation score.

Images