Acoustic superstructure design method and device based on automatic encoder

Abstract

The invention discloses an acoustic superstructure design method and device based on an automatic encoder, and the method comprises the steps: sequentially and independently training a decoder and an encoder, artificially constraining an output result of a bottleneck layer, and overcoming a common non-uniqueness problem of a conventional reverse design. And meanwhile, each detail in the neural network is not deeply studied excessively, the training process of the automatic encoder is changed from a complete'black box 'problem to a'grey box' problem, a decision-making process is further deeply understood, and the interpretability of the network is improved to a certain extent. The ultra-thin broadband sound absorption superstructure is successfully designed according to the method, numerical simulation verification and experimental verification are carried out on the ultra-thin broadband sound absorption superstructure, and the simulation result and the experimental result are well matched with the target curve. When the method is widely applied to on-demand reverse design of other types of acoustic superstructures, the design efficiency of the acoustic superstructures can be greatly improved, and the practical application of the acoustic superstructures is further promoted.

Description

Technical Field

[0001] The present invention belongs to the technical field of superstructure design, and in particular relates to an acoustic superstructure design method and device based on an automatic encoder. Background Art

[0002] Sound is one of the earliest physical phenomena studied by scientists, and acoustics is also one of the oldest and still developing branches of classical physics. In 2000, researchers from the Hong Kong University of Science and Technology proposed a phononic crystal based on the idea of ​​local resonance, and manufactured a phononic crystal with a 5mm lead ball as the inner layer and a 2.5mm silicone rubber layer as the outer layer. The lattice constant of the phononic crystal is two orders of magnitude smaller than the wavelength of the relevant sound wave. In recent years, the acoustic superstructures developed from this have shown broad application prospects in low-frequency noise control, acoustic imaging, and multifunctional acoustic devices, which has greatly promoted the development of acoustic superstructures.

[0003] Given that traditional acoustic superstructure design requires designers to have extensive experience, even if various heuristic algorithms are used for on-demand reverse design, in order to approach the goal, parameters must be adjusted on a multiple-iteration basis and each result must be evaluated through simulation software. It is undeniable that under the guidance of these methods, fruitful results have been achieved in the early stages of the rapid development of superstructures. However, with the continuous development of manufacturing technology, the complexity of the designed superstructures has also increased. The manpower, computing power, and time costs of using traditional methods are constantly increasing. A more efficient and convenient design method is urgently needed, and deep learning technology is expected to achieve this goal. Summary of the invention

[0004] In order to overcome the "non-uniqueness" and difficult convergence problems when using neural networks for reverse design of acoustic superstructures, an acoustic superstructure design method and device based on an autoencoder is proposed to realize on-demand reverse design of acoustic superstructures. The "black box" problem is transformed into a "gray box" problem by training the decoder and encoder separately. After training, the decoder and encoder can respectively realize accurate and rapid performance prediction of input structural parameters and input target curves to obtain the structural parameters of the superstructure to be designed.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] An acoustic superstructure design method based on an autoencoder comprises the following steps:

[0007] Step 1), obtaining sufficient and reliable data of the type of superstructure to be designed; the data includes structural parameters and corresponding performance curves;

[0008] Step 2), preprocess the data and randomly divide the processed data into training set and test set according to a set ratio;

[0009] Step 3), build an autoencoder neural network model, set the number of neural network layers, number of neurons, activation function and optimizer hyperparameters;

[0010] Step 4), the decoder is trained separately based on the preprocessed training set data, and the decoder training effect is detected based on the preprocessed test set data. After the training is completed, the decoder will have the ability to predict the performance of the super structure;

[0011] Step 5), freeze the weight parameters of the decoder part, train the encoder based on the preprocessed training set data, and detect the training effect based on the preprocessed test set data. After the training, the encoder has the ability to reversely design the super structure according to the performance curve.

[0012] A further improvement of the present invention is that the detailed steps of step 1) are as follows:

[0013] Step 1.1), deriving the performance formula based on the superstructure characteristics;

[0014] Step 1.2), verify the accuracy of the derived performance formula through modeling and simulation;

[0015] In step 1.3), the performance formula is written in matrix form, and the powerful matrix operation capability of Matlab is used to generate enough initial data for training in a short time.

[0016] A further improvement of the present invention is that the detailed steps of step 2) are as follows:

[0017] Step 2.1), arrange the structural parameters in a set order to make the neural network converge more easily;

[0018] Step 2.2), discretizing the performance curve within the target design interval to set intervals and obtain multiple sequence values ​​on the performance curve;

[0019] Step 2.3), normalizing the structural parameters and their corresponding sequence values;

[0020] Step 2.4), the processed structural parameters are divided into training set and test set according to the set ratio.

[0021] A further improvement of the present invention is that the detailed steps of step 3) are as follows:

[0022] Step 3.1), construct an autoencoder neural network, set the number of neurons in the encoder input layer to be the same as the number of discrete sequence values ​​of the performance curve, set the number of neurons in the bottleneck layer to be the same as the number of parameters to be designed of the superstructure to be designed, and set the number of neurons in the decoder output layer to be the same as the number of discrete sequence values ​​of the performance curve;

[0023] Step 3.2), according to the difficulty of the problem, set the number of hidden layers and the number of neurons. Usually, the encoder sets 4 to 8 hidden layers, and the number of neurons is distributed in a spindle shape. Select a suitable activation function and optimizer. The encoder part also sets 4 to 8 hidden layers, and the number of neurons is distributed in a spindle shape. Select a suitable activation function and optimizer. The specific formula of the ReLU activation function is as follows:

[0024]

[0025] A further improvement of the present invention is that the detailed steps of step 4) are as follows:

[0026] Step 4.1), the loss function is set as follows when training the decoder:

[0027]

[0028] Where n is the number of samples, y is the true value of the sequence value of the i-th data, is the decoder's prediction of the value of the i-th data sequence;

[0029] Step 4.2), take the structural parameter part of the data as the input of the decoder, and the corresponding sequence value as the ideal output, train the decoder, and set the appropriate learning rate and number of training times according to the difficulty setting of the problem and the required target accuracy;

[0030] Step 4.3), when the number of training times is reached or the loss function value is small enough and no longer decreases, the training can be stopped and the decoder is considered to have completed training;

[0031] Step 4.4), input the super structure parameters to be predicted into the decoder, and the decoder outputs the corresponding sequence values. A line graph is drawn based on the sequence values ​​to obtain the predicted performance curve.

[0032] A further improvement of the present invention is that the detailed steps of step 5) are as follows:

[0033] Step 5.1), the loss function is set as follows when training the encoder:

[0034]

[0035] Where α+β=1, x is the true value of the i-th data structure parameter, is the encoder's predicted value of the i-th data structure parameter, y is the true value of the sequence value of the i-th data, is the decoder's prediction of the structural parameters The predicted value of the corresponding sequence value;

[0036] Step 5.2), using the discretized sequence value of the performance curve as the input of the encoder, and comprehensively considering the error value between the overall output and the overall input of the autoencoder and the error value between the structural parameters corresponding to the encoder output and the overall input according to the loss function, the encoder is trained, and the learning rate and the number of training times are set according to the difficulty setting of the problem and the required target accuracy;

[0037] Step 5.3), construct a target curve whose trend and shape are close to the real performance curve;

[0038] In step 5.4), the target curve is discretized into sequence values ​​as the input of the trained encoder, and the output of the bottleneck layer is extracted to obtain the structural parameters of the reverse design. Based on the parameter modeling, the super structure of reverse design on demand can be obtained.

[0039] A further improvement of the present invention is that the performance curves should include a transmission coefficient curve, a reflection coefficient curve, a sound absorption coefficient curve and a sound insulation curve.

[0040] A further improvement of the present invention is that the data set is divided according to the following ratios: (0.6:0.4), (0.7:0.3), (0.8:0.2) and (0.9:0.1).

[0041] An acoustic superstructure design device based on an automatic encoder, comprising:

[0042] A data acquisition module, for acquiring sufficient and reliable data of the superstructure to be designed; the data includes structural parameters and corresponding performance curves;

[0043] The data preprocessing module preprocesses the data and randomly divides the processed data into training sets and test sets according to a set ratio;

[0044] Model building module, build the autoencoder neural network model, set the number of neural network layers, number of neurons, activation function and optimizer hyperparameters;

[0045] The decoder training module trains the decoder based on the preprocessed training set data and detects the decoder training effect based on the preprocessed test set data. After the training is completed, the decoder will have the ability to predict the performance of the super structure.

[0046] The encoder training module freezes the weight parameters of the decoder part, trains the encoder based on the preprocessed training set data, and detects the training effect based on the preprocessed test set data. After the training, the encoder has the ability to reversely design the super structure according to the performance curve.

[0047] A further improvement of the present invention is that, in the data preprocessing module, the data is preprocessed, and the processed data is randomly divided into a training set and a test set according to a set ratio, including:

[0048] Derive the performance formula based on the superstructure characteristics;

[0049] Verify the accuracy of the derived performance formula through modeling and simulation;

[0050] The performance formula is written in matrix form, and the powerful matrix operation capability of Matlab is used to generate enough initial data for training in a short time.

[0051] Compared with the prior art, the present invention has at least the following beneficial technical effects:

[0052] 1. Compared with the traditional method of researchers' meticulous design, this method has relatively loose requirements on the professional knowledge reserve and experience accumulation of designers;

[0053] 2. It can realize reverse design and accurate and rapid prediction of superstructure performance based on structural parameters.

[0054] 3. Ability to achieve performance design with adjustable target range within a larger range.

[0055] 4. Able to realize reverse design of single performance band and multiple performance bands.

[0056] 5. The designed acoustic superstructure has a relatively stable performance curve and no obvious defects within the target range.

[0057] 6. This method is simple to implement and has strong versatility.

[0058] The present invention proposes an acoustic superstructure design method based on an autoencoder. By training the decoder and the encoder independently in sequence, the output result of the bottleneck layer is artificially constrained to overcome the common "non-uniqueness" problem of traditional reverse design. At the same time, every detail inside the neural network is not studied too deeply, and the autoencoder training process is changed from a complete "black box" problem to a "gray box" problem. At the same time, the decision-making process has a further in-depth understanding, which improves the interpretability of the network to a certain extent. First, the decoder is trained separately. The decoder that has completed the training can be used alone to achieve accurate and rapid prediction of structural performance, and can accurately predict the performance of acoustic superstructures with multiple strong parameter couplings. Subsequently, the encoder is trained based on the decoder that has completed the training. The encoder that has completed the training can reverse design a single performance band and multiple performance bands on demand within a large range of target intervals. The reverse design based on the autoencoder can effectively reduce the possible over-coupling and under-coupling units during manual design, making the entire performance curve relatively flat, avoiding the generation of obvious defects in the target interval, and also greatly improving the space utilization.

[0059] The wide application of this method to on-demand reverse design of other types of acoustic superstructures can greatly improve the design efficiency of acoustic superstructures and further promote the practical application of acoustic superstructures. At the same time, this method has strong versatility and is expected to be applied to other types of superstructures that can be parameterized. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Top view of the reverse-engineered ultrathin acoustic superstructure.

[0061] Figure 2 Schematic diagram of the air domain inside a single cavity.

[0062] Figure 3 Schematic diagram of the autoencoder neural network structure.

[0063] Figure 4 The figure is a comparison between the analytical solution and the simulation results.

[0064] Figure 5 Decoder training results and their prediction results.

[0065] Figure 6 Schematic diagram of encoder training results and its reverse design results.

[0066] Figure 7 is the target curve finally adopted.

[0067] Figure 8 Schematic diagram of COMSOL simulation finite element numerical model.

[0068] Fig. 9Comparison chart between the target curve and simulation results for reverse design of broadband and efficient sound absorption structure.

[0069] Fig.10 Comparison chart between the target curve and simulation results for the reverse design of a dual-band high-efficiency sound-absorbing structure.

[0070] Fig.11 Schematic diagram of the experimental test device.

[0071] Fig.12 Schematic diagram of the experimental results of reverse design of broadband sound absorption structure. DETAILED DESCRIPTION

[0072] Hereinafter, the present invention will be described in detail with reference to the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited to the following embodiments, but shall include all the contents in the claims. Moreover, those skilled in the art can realize all the contents in the claims from the following embodiments. Based on the implementation cases in the present invention, all other implementation cases obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention.

[0073] Example 1

[0074] The present invention provides an acoustic superstructure design method based on an autoencoder, which artificially controls the training process of the autoencoder neural network, and trains the decoder and the encoder respectively in sequence. After the training, the decoder and the encoder can respectively input structural parameters for accurate and rapid performance prediction and input target performance curves to obtain structural parameters of the superstructure to be designed. The method comprises the following steps:

[0075] Step 1), obtaining sufficient and reliable data of the type of superstructure to be designed; the data includes structural parameters and corresponding performance curves;

[0076] Step 2), preprocess the data and randomly divide the processed data into training set and test set according to a set ratio;

[0077] Step 3), build an autoencoder neural network model, set the number of neural network layers, number of neurons, activation function and optimizer hyperparameters;

[0078] Step 4), the decoder is trained separately based on the preprocessed training set data, and the decoder training effect is detected based on the preprocessed test set data. After the training is completed, the decoder will have the ability to predict the performance of the super structure;

[0079] Step 5), freeze the weight parameters of the decoder part, train the encoder based on the preprocessed training set data, and detect the training effect based on the preprocessed test set data. After the training, the encoder has the ability to reversely design the super structure according to the performance curve.

[0080] In this embodiment, the detailed steps of step 1) are as follows:

[0081] Step 1.1), deriving the performance formula based on the superstructure characteristics;

[0082] Step 1.2), verify the accuracy of the derived performance formula through modeling and simulation;

[0083] In step 1.3), the performance formula is written in matrix form, and the powerful matrix operation capability of Matlab is used to generate enough initial data for training in a short time.

[0084] In this embodiment, the detailed steps of step 2) are as follows:

[0085] Step 2.1), arrange the structural parameters in a set order to make the neural network converge more easily;

[0086] Step 2.2), discretizing the performance curve within the target design interval to set intervals and obtain multiple sequence values ​​on the performance curve;

[0087] Step 2.3), normalizing the structural parameters and their corresponding sequence values;

[0088] Step 2.4), the processed structural parameters are divided into training set and test set according to the set ratio.

[0089] In this embodiment, the detailed steps of step 3) are as follows:

[0090] Step 3.1), construct an autoencoder neural network, set the number of neurons in the encoder input layer to be the same as the number of discrete sequence values ​​of the performance curve, set the number of neurons in the bottleneck layer to be the same as the number of parameters to be designed of the superstructure to be designed, and set the number of neurons in the decoder output layer to be the same as the number of discrete sequence values ​​of the performance curve;

[0091] Step 3.2), according to the difficulty of the problem, set the number of hidden layers and the number of neurons. Usually, the encoder sets 4 to 8 hidden layers, and the number of neurons is distributed in a spindle shape. Select a suitable activation function and optimizer (usually ReLU and Adam). The encoder also sets 4 to 8 hidden layers, and the number of neurons is distributed in a spindle shape. Select a suitable activation function and optimizer (usually ReLU and Adam). The specific formula of the ReLU activation function is as follows:

[0092]

[0093] In this embodiment, the detailed steps of step 4) are as follows:

[0094] Step 4.1), the loss function is set as follows when training the decoder:

[0095]

[0096] Where n is the number of samples, y is the true value of the sequence value of the i-th data, is the decoder's prediction of the value of the i-th data sequence;

[0097] Step 4.2), take the structural parameter part of the data as the input of the decoder, and the corresponding sequence value as the ideal output, train the decoder, and set the appropriate learning rate and number of training times according to the difficulty setting of the problem and the required target accuracy;

[0098] Step 4.3), when the number of training times is reached or the loss function value is small enough and no longer decreases, the training can be stopped and the decoder is considered to have completed training;

[0099] Step 4.4), input the super structure parameters to be predicted into the decoder, and the decoder outputs the corresponding sequence values. A line graph is drawn based on the sequence values ​​to obtain the predicted performance curve.

[0100] In this embodiment, the detailed steps of step 5) are as follows:

[0101] Step 5.1), the loss function is set as follows when training the encoder:

[0102]

[0103] Where α+β=1, x is the true value of the i-th data structure parameter, is the encoder's predicted value of the i-th data structure parameter, y is the true value of the sequence value of the i-th data, is the decoder's prediction of the structural parameters The predicted value of the corresponding sequence value;

[0104] Step 5.2), the sequence value after discretization of the performance curve is used as the input of the encoder, and the error value between the overall output of the autoencoder (the output of the decoder) and the overall input (the input of the encoder) and the error value between the structural parameters corresponding to the encoder output and the overall input are comprehensively considered according to the loss function, and the encoder is trained, and the appropriate learning rate and number of training times are set according to the difficulty setting of the problem and the required target accuracy;

[0105] Step 5.3), construct a target curve whose trend and shape are close to the real performance curve;

[0106] In step 5.4), the target curve is discretized into sequence values ​​as the input of the trained encoder, and the output of the bottleneck layer is extracted to obtain the structural parameters of the reverse design. Based on the parameter modeling, the super structure of reverse design on demand can be obtained.

[0107] In this embodiment, the performance curves should include but are not limited to a transmission coefficient curve, a reflection coefficient curve, a sound absorption coefficient curve, a sound insulation curve, and the like.

[0108] In this embodiment, the data set is divided according to the following ratios: (0.6:0.4), (0.7:0.3), (0.8:0.2), and (0.9:0.1).

[0109] Artificial control of the training process can make the neural network converge more easily, overcoming the key "non-uniqueness" problem of reverse design.

[0110] The loss function when training the encoder takes into account both the inverse design structure parameters and their corresponding performance curves, making the overall performance of the network more stable.

[0111] The present invention can realize performance reverse design with adjustable target interval within a relatively large range.

[0112] The present invention can realize reverse design of single performance frequency band and multiple performance frequency bands.

[0113] The designed acoustic superstructure has a relatively stable performance curve and no obvious defects within the target range.

[0114] Example 2

[0115] The present invention provides an acoustic superstructure design device based on an automatic encoder, comprising:

[0116] A data acquisition module, for acquiring sufficient and reliable data of the superstructure to be designed; the data includes structural parameters and corresponding performance curves;

[0117] The data preprocessing module preprocesses the data and randomly divides the processed data into training sets and test sets according to a set ratio;

[0118] Model building module, build the autoencoder neural network model, set the number of neural network layers, number of neurons, activation function and optimizer hyperparameters;

[0119] The decoder training module trains the decoder based on the preprocessed training set data and detects the decoder training effect based on the preprocessed test set data. After the training is completed, the decoder will have the ability to predict the performance of the super structure.

[0120] The encoder training module freezes the weight parameters of the decoder part, trains the encoder based on the preprocessed training set data, and detects the training effect based on the preprocessed test set data. After the training, the encoder has the ability to reversely design the super structure according to the performance curve.

[0121] Example 3

[0122] As an embodiment, a spatially folded 14-cavity ultra-thin sound-absorbing superstructure is selected as a target object.

[0123] (1) Geometry model and theoretical support of reverse-designed sound-absorbing superstructure

[0124] Implementation examples Figure 1 As shown, it is formed by 14 independent Helmholtz resonant cavities with holes at both ends. The cross-sectional width of a single cavity is W = 8 mm, the cross-sectional height is H = 20 mm, the diameter of the small holes on both sides is d = 2 mm, the depth of the small holes is t = 1 mm, and the distance between the small holes on both sides and the boundary is B = 3 mm.

[0125] For a single resonant cavity, the surface impedance Z a It consists of two parts:

[0126] Z a =Z h +Z c

[0127] Where Z h is the impedance of the Helmholtz resonant cavity aperture, Z c is the impedance of the Helmholtz resonant cavity. Therefore, the relative surface impedance of a single resonant cavity is:

[0128]

[0129] Where S0=W×L is the surface area of ​​the cavity, ρ0 and c0 are the air density and the speed of sound in the air respectively.

[0130] Using the Euler equation, we know that the impedance of the small hole is

[48] :

[0131]

[0132] Where n is the number of small holes, S h is the area of ​​the small hole, is the small hole perforation constant, B0 and B1 are the first-kind Bessel functions of the 0th and 1st order, respectively, μ is the aerodynamic viscosity, Δt1, Δt2 are corrections for the pressure radiation at the discontinuity from the neck to the cavity and the radiation from the waveguide to the neck, respectively.

[0133] Surface impedance Z c The calculation formula is as follows:

[0134]

[0135] in are the effective characteristic impedance and effective wave number of the air in the cavity respectively.

[0136] The structure studied is composed of multiple double-ended open-hole Helmholtz resonant cavities connected in parallel. The impedance of the entire surface of the parallel Helmholtz resonant cavities can be expressed as:

[0137]

[0138] Wherein, n is the number of parallel structural units.

[0139] According to the above formula, a more accurate sound absorption coefficient can be obtained. Taking a single cavity and five cavities as examples, the analytical solution obtained according to the formula is compared with the COMSOL simulation results. Figure 4 As shown. The absorption peak frequency of a single cavity simulated by software is 710Hz, and the amplitude is 0.65. The absorption peak frequency calculated by analytical formula is 718Hz, and the amplitude is 0.77. For the 5-cavity superstructure, according to the COMSOL simulation results, it has a high sound absorption coefficient in the range of 720-920Hz, and the average sound absorption coefficient is close to 0.8. The simulation results are highly consistent with the analytical results. The corresponding frequencies and amplitudes of each peak of the two are almost the same. It can be seen that the above formula has a high accuracy.

[0140] The performance formula is rewritten into the form of matrix operation and compiled into Matlab code. The 14 structural parameters of each data are randomly generated within the specified range through the Monte Carlo sampling method. The corresponding sound absorption coefficient curve can be obtained after calculation. Using the powerful matrix operation ability of Matlab, tens of thousands of data can be generated in a few seconds.

[0141] The obtained raw data were processed by first arranging the 14 structural parameters of each data in descending order, and then discretizing the sound absorption coefficient curve in the range of 200-1700Hz, with 5Hz as the interval, and obtaining 300 sequence values ​​on the sound absorption coefficient curve. The structural parameters in descending order and their corresponding 300 sequence values ​​were linearly normalized, and the calculation formula is as follows:

[0142]

[0143] Where x is the original data, x min is the minimum value in the data set, x max is the maximum value in the data set, x newis the normalized data. The normalized descending structural parameters and their corresponding sequence values ​​constitute new data, which will be used for subsequent training.

[0144] (2) Autoencoder neural network structure and training method, process and results

[0145] Based on Figure 3 The autoencoder neural network shown is used to perform the inverse design of the structural parameters.

[0146] First, train the decoder. The decoder consists of 5 hidden layers. The number of neurons is 128, 256, 512, 1024, and 512, respectively. The ReLU activation function is used. The formula is as follows:

[0147]

[0148] The output layer uses a linear activation function, and the optimizer selects Adam. MSE is selected as the loss function, and the specific formula is as follows:

[0149]

[0150] Where n is the number of samples, y is the true value of the sequence value of the i-th data, is the decoder's prediction value for the i-th data sequence value. The initial learning rate is set to 0.004, and the number of training times is set to 1500. At the beginning of training, the 14 cavity length parameters of each set of data and the corresponding 300 sequence values ​​are used as the input and ideal output of the decoder to train the decoder. The training is considered completed when the decoder can quickly and accurately predict the sequence value based on the cavity length. The training results are shown in Figure 5 As shown, Figure 5 (a) is the loss (MSE) curve of the decoder training. It can be seen that fast convergence can be achieved in the first 400 trainings. In order to visually observe the accuracy of the decoder prediction, several curves are randomly generated according to the formula, and their structural parameters are input into the decoder to check the gap between the prediction results and the theoretical solution. Figure 5 As shown in (b)-(d), it can be seen that the decoder can achieve relatively accurate prediction in the low frequency band, high frequency band and wide frequency band, and the number of peaks, peak frequency and change trend are highly consistent. This shows that the training of the decoder has been completed, and the next step is to train the encoder.

[0151] The encoder part is also composed of 5 hidden layers, with the number of neurons being 512, 1024, 512, 256, and 128, respectively. The ReLU activation function is used in all layers, and the output layer uses a linear activation function. The Adam optimizer is used. MSE is used as the loss function, and the specific formula is as follows:

[0152]

[0153] Where α = β = 0.5, n is the number of samples, x is the true value of the i-th data structure parameter, is the encoder's predicted value of the i-th data structure parameter, y is the true value of the sequence value of the i-th data, is the decoder's prediction of the structural parameters The predicted value of the corresponding sequence value. The initial learning rate is set to 0.004, and the number of training times is set to 1500. The decoder weights are frozen before training to ensure that only the encoder weights are updated during training. The two trainings need to be performed on the same data set. When the error value is less than an ideal value or the number of cycles is reached, the training is considered to be completed. The results are as follows Figure 6 As shown, Figure 6 (a) is the encoder training loss (MSE) curve. In order to more intuitively observe the accuracy of the encoder reverse design, several curves were randomly generated according to the formula. The sound absorption coefficient curve was discretized into 300 sequence values ​​and input into the encoder to obtain a set of structural parameters at the bottleneck layer. The structural parameters were then brought into the analytical formula to preliminarily check the accuracy of the reverse design results. The results are shown in Figure 6 As shown in (b)-(d), we can see that good reverse design effects can be achieved in low frequency band, high frequency band and wide frequency band, and the number of peaks, peak frequency and change trend are highly consistent. This shows that the encoder has been trained and can be applied according to actual needs.

[0154] (3) Performance simulation analysis and experimental verification of reverse-designed acoustic superstructures

[0155] By taking the rising and falling parts of the existing sound absorption coefficient curve as the basis, splicing a line segment between the two parts and then superimposing multiple sine functions of different periods and amplitudes on the line segment, a target sound absorption coefficient curve with a shape close to the actual situation is fitted. The results are as follows: Figure 7 shown.

[0156] Two target curves were generated using the above method, representing broadband efficient sound absorption and dual-band efficient sound absorption, respectively. Fig. 9 , Fig.10 The two target curves are discretized and input into the autoencoder neural network in turn, and the output of the bottleneck layer is extracted as the corresponding structural parameters. The acoustic performance of the inverse design results is numerically analyzed using the pressure acoustics-frequency domain and solid mechanics modules of the commercial finite element software COMSOL Multiphysics. The established numerical model is shown in Figure 8 shown.

[0157] Two domain point probes are set in the air domain of the numerical model to detect the incident signal p1 and the reflected signal p2. The two domain point probes are placed at z1 = 130 mm and z2 = 126 mm from the structure surface. According to the sound pressure values ​​detected by the domain point probes, the incident sound pressure and the reflected sound pressure can be obtained by substituting them into the formula:

[0158]

[0159] Wherein, k=ω / c0 represents the sound wave number, ω represents the angular frequency, and c0 represents the speed of sound in air.

[0160] The sound absorption coefficient of the structure can be obtained by further calculation:

[0161] α=1-(|p ref / p in |) 2

[0162] In the numerical model, the speed of sound c0 of air is 343 m / s and the density ρ0 is 1.29 kg / m 3 The Young's modulus, density and Poisson's ratio of the superstructure material are (2.2+0.11i)Gpa, 1320kg / m 3 and 0.375, where the imaginary part of the Young's modulus of the material represents the viscoelastic damping loss of the material.

[0163] The corresponding three-dimensional model is established according to the reverse design results, and the simulation verification is carried out by COMSOL according to the above method. The two target curves are compared with the simulation results of the reverse design as shown below. Fig. 9 , Fig.10 As shown. Fig. 9 It can be seen that the target sound absorption range is 520Hz-1000Hz, and the average sound absorption coefficient in the range is greater than 0.8. The simulated sound absorption coefficient curve is highly similar to the target curve. The starting frequency, ending frequency and sound absorption effect all meet the expected design requirements, proving that the autoencoder has the ability to reverse design broadband and efficient sound absorption superstructures. Fig.10 It can be seen that the target curve has two target sound absorption intervals, which are 600Hz-800Hz and 1000Hz-1200Hz, respectively, 200Hz apart, and the average sound absorption coefficient is about 0.8. The reverse design simulation results also have two independent peaks. Although the end frequency of the first peak and the overall shift of the second peak to the low frequency direction are slightly offset, the key information is realized and the distance between the two peaks is about 200Hz, and the width and average sound absorption coefficient of the two peaks can meet the requirements. It proves that the autoencoder has the ability to realize the reverse design of multi-band efficient sound absorption superstructures.

[0164] The broadband high-efficiency sound absorption superstructure was processed, and its sound absorption performance was tested using an impedance tube with a size of 100mm*100mm. Fig.11 As shown in the figure, three sound wave collection sensors are installed on the impedance tube, which are used to measure the incident, reflected and transmitted sound pressures from left to right. After the signal is collected to the computer and calculated according to the formula, the sound absorption coefficient curve of the sample can be obtained.

[0165] For the comparison between the target curve and experimental results of the reverse-designed broadband high-efficiency sound absorption superstructure, Fig.12 As shown, it can be observed that the experimental result curve is highly consistent with the target curve, especially the sound absorption peaks at 550Hz and 950Hz at both ends of the target curve can be found in the experimental results. In addition, the peaks at 670Hz and 810Hz of the target curve can also find corresponding trends in the experimental results. A relatively smooth sound absorption effect is achieved in the entire target range of 520Hz-1000Hz, and there is no obvious trough in sound absorption performance.

[0166] In summary, the present invention proposes an acoustic superstructure design method based on an autoencoder. By training the decoder and encoder independently in sequence, the output result of the bottleneck layer is artificially constrained to overcome the common "non-uniqueness" problem of traditional reverse design. At the same time, every detail inside the neural network is not studied too deeply, and the autoencoder training process is changed from a complete "black box" problem to a "gray box" problem. At the same time, the decision-making process has a further in-depth understanding, which improves the interpretability of the network to a certain extent. The reverse design based on the autoencoder can effectively reduce the possible over-coupling and under-coupling units in manual design, so that the entire sound absorption coefficient curve is relatively flat, avoiding the generation of obvious sound absorption troughs in the target range, and greatly improving the space utilization. Based on this, an ultra-thin broadband sound absorption superstructure was successfully designed, and numerical simulation and experimental verification were performed on it. The simulation results and experimental results are in good agreement with the target curve. The wide application of this method to the on-demand reverse design of other types of acoustic superstructures can greatly improve the design efficiency of acoustic superstructures and further promote the practical application of acoustic superstructures. At the same time, this method is highly versatile and can be applied to other types of superstructures that can be parameterized.

[0167] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved.

[0168] In addition, it should be understood that although this specification is described in accordance with the implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation modes that can be understood by those skilled in the art. The above content is only to illustrate the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution according to the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. An acoustic superstructure design method based on an autoencoder, characterized in that: The steps include: Step 1), obtaining sufficient and reliable data of the type of superstructure to be designed; the data includes structural parameters and corresponding performance curves; Step 2), preprocess the data and randomly divide the processed data into training set and test set according to a set ratio; Step 3), build an autoencoder neural network model, set the number of neural network layers, number of neurons, activation function and optimizer hyperparameters; Step 4), the decoder is trained separately based on the preprocessed training set data, and the decoder training effect is detected based on the preprocessed test set data. After the training is completed, the decoder will have the ability to predict the performance of the super structure; Step 5), freeze the weight parameters of the decoder part, train the encoder based on the preprocessed training set data, and detect the training effect based on the preprocessed test set data. After the training, the encoder has the ability to reversely design the super structure according to the performance curve.

2. The acoustic superstructure design method based on an autoencoder according to claim 1, characterized in that: The detailed steps of step 1) are as follows: Step 1.1), deriving the performance formula based on the superstructure characteristics; Step 1.2), verify the accuracy of the derived performance formula through modeling and simulation; In step 1.3), the performance formula is written in matrix form, and the powerful matrix operation capability of Matlab is used to generate enough initial data for training in a short time.

3. The acoustic superstructure design method based on an autoencoder according to claim 1, characterized in that: The detailed steps of step 2) are as follows: Step 2.1), arrange the structural parameters in a set order to make the neural network converge more easily; Step 2.2), discretizing the performance curve within the target design interval to set intervals and obtain multiple sequence values ​​on the performance curve; Step 2.3), normalizing the structural parameters and their corresponding sequence values; Step 2.4), the processed structural parameters are divided into training set and test set according to the set ratio.

4. The acoustic superstructure design method based on an autoencoder according to claim 1, characterized in that: The detailed steps of step 3) are as follows: Step 3.1), construct an autoencoder neural network, set the number of neurons in the encoder input layer to be the same as the number of discrete sequence values ​​of the performance curve, set the number of neurons in the bottleneck layer to be the same as the number of parameters to be designed of the superstructure to be designed, and set the number of neurons in the decoder output layer to be the same as the number of discrete sequence values ​​of the performance curve; Step 3.2), according to the difficulty of the problem, set the number of hidden layers and the number of neurons. Usually, the encoder sets 4 to 8 hidden layers, and the number of neurons is distributed in a spindle shape. Select a suitable activation function and optimizer. The encoder part also sets 4 to 8 hidden layers, and the number of neurons is distributed in a spindle shape. Select a suitable activation function and optimizer. The specific formula of the ReLU activation function is as follows:

5. The acoustic superstructure design method based on an autoencoder according to claim 1, characterized in that: The detailed steps of step 4) are as follows: Step 4.1), the loss function is set as follows when training the decoder: Where n is the number of samples, y is the true value of the sequence value of the i-th data, is the decoder's prediction of the value of the i-th data sequence; Step 4.2), take the structural parameter part of the data as the input of the decoder, and the corresponding sequence value as the ideal output, train the decoder, and set the appropriate learning rate and number of training times according to the difficulty setting of the problem and the required target accuracy; Step 4.3), when the number of training times is reached or the loss function value is small enough and no longer decreases, the training can be stopped and the decoder is considered to have completed training; Step 4.4), input the super structure parameters to be predicted into the decoder, and the decoder outputs the corresponding sequence values. A line graph is drawn based on the sequence values ​​to obtain the predicted performance curve.

6. The acoustic superstructure design method based on an autoencoder according to claim 1, characterized in that: The detailed steps of step 5) are as follows: Step 5.1), the loss function is set as follows when training the encoder: Where α+β=1, x is the true value of the i-th data structure parameter, is the encoder's predicted value of the i-th data structure parameter, y is the true value of the sequence value of the i-th data, is the decoder's prediction of the structural parameters The predicted value of the corresponding sequence value; Step 5.2), using the discretized sequence value of the performance curve as the input of the encoder, and comprehensively considering the error value between the overall output and the overall input of the autoencoder and the error value between the structural parameters corresponding to the encoder output and the overall input according to the loss function, the encoder is trained, and the learning rate and the number of training times are set according to the difficulty setting of the problem and the required target accuracy; Step 5.3), construct a target curve whose trend and shape are close to the real performance curve; In step 5.4), the target curve is discretized into sequence values ​​as the input of the trained encoder, and the output of the bottleneck layer is extracted to obtain the structural parameters of the reverse design. Based on the parameter modeling, the super structure of reverse design on demand can be obtained.

7. The acoustic superstructure design method based on an autoencoder according to claim 1, characterized in that: The performance curves should include a transmission coefficient curve, a reflection coefficient curve, a sound absorption coefficient curve and a sound insulation curve.

8. The acoustic superstructure design method based on an autoencoder according to claim 1, characterized in that: The data set is divided into the following ratios: (0.6:0.4), (0.7:0.3), (0.8:0.2), and (0.9:0.1).

9. An acoustic superstructure design device based on an automatic encoder, characterized in that: include: A data acquisition module, for acquiring sufficient and reliable data of the superstructure to be designed; the data includes structural parameters and corresponding performance curves; The data preprocessing module preprocesses the data and randomly divides the processed data into training sets and test sets according to a set ratio; Model building module, build the autoencoder neural network model, set the number of neural network layers, number of neurons, activation function and optimizer hyperparameters; The decoder training module trains the decoder based on the preprocessed training set data and detects the decoder training effect based on the preprocessed test set data. After the training is completed, the decoder will have the ability to predict the performance of the super structure. The encoder training module freezes the weight parameters of the decoder part, trains the encoder based on the preprocessed training set data, and detects the training effect based on the preprocessed test set data. After the training, the encoder has the ability to reversely design the super structure according to the performance curve.

10. The acoustic superstructure design device based on an automatic encoder according to claim 9, characterized in that: In the data preprocessing module, the data is preprocessed and randomly divided into training sets and test sets according to a set ratio, including: Derive the performance formula based on the superstructure characteristics; Verify the accuracy of the derived performance formula through modeling and simulation; The performance formula is written in matrix form, and the powerful matrix operation capability of Matlab is used to generate enough initial data for training in a short time.

Images