Acoustic covering layer design method based on improved series neural network model

Abstract

The invention discloses an acoustic covering layer design method based on an improved series neural network model. The method comprises the following steps: constructing a data set comprising design parameters of an acoustic covering layer and corresponding spectral response; constructing a forward neural network, and obtaining the optimal initial weight and bias of the forward neural network by adopting a PSO optimization algorithm; taking the obtained weight and bias as initial values to be substituted into a forward neural network for pre-training; constructing a series neural network model in a manner that a reverse neural network is cascaded in front of a forward neural network; training a series neural network model through the data set, and taking the output of a reverse neural network as the input of a pre-trained forward neural network; and inputting the expected frequency spectrum into the trained series neural network model, and predicting structure parameters and material parameters meeting conditions. According to the method, the defects of huge occupied time cost and calculation resources are overcome, errors of design parameters are reduced, and efficient reverse design of the sound absorption performance of the acoustic covering layer is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of acoustics, and particularly relates to a design method for an acoustic covering layer based on an improved tandem neural network model. Background Art

[0002] An acoustic covering layer is a sound-absorbing layer used on the surface of a submarine, aiming to reduce the target strength of the submarine and weaken the external radiation noise. Designing an acoustic covering layer structure with high-efficiency low-frequency broadband sound absorption is a key issue. Usually, the design of a low-frequency broadband acoustic covering layer relies on numerical methods, such as the finite element method. However, due to its complexity and difficulty, traditional numerical methods often cannot meet the requirements.

[0003] To achieve the desired performance and goals, researchers have proposed various design methods, such as genetic algorithms, reverse neural networks, and topology optimization. However, these methods have some limitations in practical applications. For example, the nature of random search requires a large amount of time, cost, and computing resources for design. Therefore, in practical applications, these methods cannot achieve the efficient design of the acoustic covering layer.

[0004] On the other hand, inverse design based on deep learning has attracted more and more attention in the field of acoustics. However, there is usually a one-to-many mapping problem between the functions and structures of the acoustic covering layer, rather than a deterministic injective or surjective relationship. Traditional artificial neural network models cannot solve this multi-valued problem. At the same time, the first assignment of the weights of the neural network and the first determined threshold have a greater impact on the result accuracy. If the model parameters are randomly assigned in a conventional manner, the phenomenon of falling into local optima often occurs during model training. In actual situations, once the design parameters predicted by the tandem neural network do not match the expected ones, some parameters are difficult to implement in engineering. For example, if the cavity is too large and exceeds the height of the matrix, it is difficult to achieve inverse design. The neural network may miss the optimal solution and even generate unreasonable structures.

[0005] In summary, the design of the acoustic covering layer is a challenging problem. There are certain limitations in both traditional numerical methods and inverse design based on deep learning. Therefore, it is necessary to further research and develop new methods to achieve the efficient design of acoustic structures. Summary of the Invention

[0006] Object of the Invention: The present invention proposes a design method for an acoustic covering layer based on an improved tandem neural network model, which overcomes the disadvantages of huge time cost and computing resources occupation, reduces the error of design parameters, and realizes the inverse design of the sound absorption performance of the high-efficiency acoustic covering layer.

[0007] Technical Solution: A design method for an acoustic covering layer based on an improved tandem neural network model according to the present invention includes the following steps:

[0008] (1) Construct a dataset including the design parameters of the acoustic covering layer and the corresponding spectral responses;

[0009] (2) Construct a forward neural network, and use the PSO optimization algorithm to obtain the optimal initial weights and biases of the forward neural network;

[0010] (3) Substitute the obtained weights and biases as initial values into the forward neural network for pre-training;

[0011] (4) Construct a cascaded neural network model in the way that the reverse neural network is in the front and the forward neural network is in the back; train the cascaded neural network model with the dataset, and the output of the reverse neural network is used as the input of the pre-trained forward neural network;

[0012] (5) Input the desired spectrum into the trained cascaded neural network model to predict the structural parameters and material parameters that meet the conditions.

[0013] Further, the forward neural network model in step (2) includes an input layer, a hidden layer, and an output layer. The layers are fully connected, and the activation function ReLu and the optimizer Sdg are used.

[0014] Further, the implementation process of obtaining the optimal initial weights and biases of the forward neural network by using the PSO optimization algorithm in step (2) is as follows:

[0015] S1: Initialize the particle swarm: Randomly generate a group of particles, and each particle represents the initial weights and biases of the forward neural network; the position of the particle represents the values of the weights and biases, and the velocity represents the change rate of the weights and biases;

[0016] S2: Initialize the velocity of the particles: Randomly initialize the velocity of each particle within a suitable range;

[0017] S3: Calculate the fitness: Use the weight and bias configuration of the current particle to construct a forward neural network, and calculate the loss function or error of the neural network through the training data;

[0018] S4: Update the individual optimal position: For each particle, update the individual optimal position according to its own historical best fitness;

[0019] S5: Update the global optimal position: Update the global optimal position according to the best fitness among all particles;

[0020] S6: Update the velocity and position: For each particle, update the velocity and position according to the current velocity, individual optimal position, and global optimal position;

[0021] S7: Iteratively execute S3 to S6 until the stop condition is reached;

[0022] S8: Output result: Obtain the optimal initial weights and bias configurations according to the optimal particles, and use them as the initial weights and biases of the forward neural network.

[0023] Furthermore, the implementation process of step (3) is as follows:

[0024] Pre-train the forward neural network using the dataset to establish a forward prediction from design parameters to spectral response. The dataset is normalized before being input into the forward neural network; during the training process, use the design parameters as the input of the forward neural network model, and use the spectral response as the output of the forward neural network model. Stop training when the loss function converges, and fix the weights of the forward neural network.

[0025] Furthermore, the loss function of the forward neural network is the error between the predicted spectral response and the true spectral response, and MAE is used as the measurement index.

[0026] Furthermore, the inverse neural network in step (4) includes an input layer, a hidden layer, and an output layer. The layers are fully connected, and the activation function ReLu and the optimizer Sdg are used.

[0027] Furthermore, the implementation process of step (4) is as follows:

[0028] Fix the weights of the forward neural network and cascade it with the inverse neural network to construct a cascaded neural network model; then, use the dataset to train this cascaded neural network model to ensure that the forward neural network and the inverse neural network can work together; specifically, the output of the inverse neural network, that is, the deterministic predicted design parameters, is used as the input of the forward neural network; considering the difference in spectral response, the difference between the spectral response input to the cascaded neural network model and the predicted spectral response output by the model is used as a construction element of the loss function; during the training process, stop training when the loss function converges, and save the parameters of the cascaded neural network model.

[0029] Furthermore, the loss function of the cascaded neural network in step (4) is:

[0030]

[0031] where P i is the expected spectral response, T i is the predicted spectral response of the cascaded network, N is the number of training samples, D i ' and D i are the design parameters corresponding to the expected spectral response and the design parameters predicted by the cascaded network respectively; λ is D i ' and Di The weight ratio of the loss.

[0032] Further, the fitness is:

[0033]

[0034] Where N represents the number of particles and m is the number of samples; is the predicted spectral response, that is, the output value of the neural network, y i (k) is the expected value in the model.

[0035] Further, λ is set to 1, that is, the spectrum and the design parameters adopt the same normalization method.

[0036] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention designs an acoustic covering layer based on a cascaded neural network model, which not only overcomes the problem of one-to-many mapping from spectral response to structural parameters in the traditional acoustic design process, but also can improve the accuracy of the overall cascaded neural network model by improving the performance of the forward prediction neural network; at the same time, by adding restrictive conditions to the design parameters before training the cascaded neural network model, the reliability of the design parameters is ensured, and the solution can be obtained immediately after training the model, significantly improving the calculation time and work efficiency. In practical engineering applications, it can reduce human intervention, and for any complex structure, the optimal parameters that meet the conditions can be obtained. Brief Description of the Drawings

[0037] Figure 1 The flowchart of the present invention;

[0038] Figure 2 is the structure of the acoustic covering layer and its spectral response;

[0039] Figure 3 is the schematic diagram of the improved cascaded neural network structure;

[0040] Figure 4 is the comparison diagram of the design results of the improved cascaded neural network. Detailed Embodiment

[0041] The present invention will be further described in detail below with reference to the accompanying drawings.

[0042] As Figure 1 shown, an acoustic covering layer design method based on an improved cascaded neural network model. The method includes the following steps:

[0043] Step 1: Construct a data set of the design parameters of the acoustic covering layer and the corresponding spectral responses.

[0044] In this embodiment, the acoustic covering layer is an acoustic absorption structure composed of a rubber matrix with cavities and a steel backing. For the cylindrical cavity type acoustic covering layer, its main acoustic absorption mechanisms include cavity resonance, waveform conversion, impedance mismatch, etc. The physical parameters affecting these mechanisms mainly include: matrix density, elastic modulus, loss factor, Poisson's ratio, and the geometric parameters of the cavity. Different combinations of these parameters will result in different acoustic absorption performances. As Figure 2 shown, the mapping from design parameters to spectral response is forward prediction, and the mapping from spectral response to design parameters is inverse design. 8750 data samples are generated through random design parameters. These data samples are divided into three different groups: 7750 for training, 500 for validation, and 500 for final testing. The frequency range of the acoustic absorption coefficient corresponding to the design parameters output by each group of samples is between 1000 Hz and 5000 Hz, with an interval of 100 Hz, and there are a total of 41 frequency point responses.

[0045] Step 2: Construct a forward neural network model, and use the PSO optimization algorithm to obtain the optimal initial weights and biases of the forward neural network.

[0046] The forward neural network model includes an input layer, a hidden layer, and an output layer. The layers are connected in a fully connected manner, and the activation function ReLu and the optimizer Sdg are used.

[0047] The steps for optimizing the initial weights and biases of the forward neural network by the PSO algorithm are as follows:

[0048] 1) Initialize the particle swarm: Randomly generate a group of particles, and each particle represents the initial weights and biases of the forward neural network. The position of the particle represents the values of the weights and biases, and the velocity represents the change rate of the weights and biases.

[0049] 2) Initialize the velocity of the particles: Randomly initialize the velocity for each particle, limited within a suitable range.

[0050] 3) Calculate the fitness: Use the weight and bias configuration of the current particle to construct a forward neural network, and calculate the loss function or error of the neural network through the training data.

[0051] 4) Update the individual optimal position: For each particle, update the individual optimal position according to its own historical best fitness:

[0052]

[0053] where N represents the number of particles, and m is the number of samples; is the predicted spectral response, that is, the output value of the neural network, y i (k) is the expected value in the model.

[0054] 5) Update the global best position: Update the global best position according to the best fitness among all particles.

[0055] 6) Update the velocity and position: For each particle, update the velocity and position according to the current velocity, the personal best position, and the global best position. The update of the velocity takes into account the influence of the particle's inertia, personal experience, and global experience.

[0056] 7) Repeat steps 3 to 6: Iteratively execute steps 3 to 6 until the stopping condition is reached (such as reaching the maximum number of iterations or the fitness meets the predetermined threshold).

[0057] 8) Output the result: Obtain the optimal initial weights and bias configuration according to the optimal particle as the initial weights and bias of the forward neural network.

[0058] First, determine the structure of the forward neural network according to the number of inputs and outputs in the dataset. Secondly, obtain the population dimension in the PSO algorithm according to the network structure to get the initial population. Then, calculate the fitness value of the particle according to the following formula, determine the extreme values of the individual and the global, and update its position and velocity:

[0059]

[0060] where N represents the number of particles and m is the number of samples; is the predicted spectral response, that is, the output value of the neural network, y i (k) is the expected value in the model.

[0061] Finally, judge whether the termination condition is satisfied. If the condition is not reached, continue the loop; if the condition is reached, obtain the best weights and thresholds.

[0062] Step 3: Use the weights and biases obtained in step 2 as the initial values and bring them into the forward neural network for pre-training.

[0063] Considering the non-uniqueness of the acoustic covering layer, use the dataset to pre-train the forward neural network to establish a forward prediction from the design parameters to the spectral response. The dataset is normalized before being input into the forward neural network; during the training process, use the design parameters as the input of the forward neural network model and the spectral response as the output of the forward neural network model. Stop training when the loss function converges and fix the weights of the model. The loss function of the forward neural network model is the error between the predicted spectral response and the true spectral response, and use MAE as the measurement index.

[0064] Step 4: As Figure 3As shown, a cascaded neural network model is constructed in the way that the reverse prediction neural network model is in the front and the forward prediction neural network model is in the back; the cascaded neural network model is trained with the said data set, and the output of the reverse prediction neural network model is used as the input of the pre-trained forward neural network model.

[0065] The reverse neural network model includes an input layer, a hidden layer and an output layer. The connection between layers is fully connected, and the activation function ReLu and the optimizer Sdg are used.

[0066] To ensure the convergence of the reverse neural network training, the following steps are taken: First, fix the weights of the prediction model and cascade it with the reverse design model to construct a cascaded neural network model. Then, use the data set to train this cascaded model to ensure that the forward prediction model and the reverse design model can work together. Specifically, the output of the reverse design model, that is, the deterministic prediction design parameters, is used as the input of the forward neural network model. The difference between the spectral response input to the cascaded neural network model and the predicted spectral response of the model output also needs to be considered as an element in constructing the loss function. During the training process, when the loss function converges, stop the training and save the model parameters.

[0067] The improved loss function of the cascaded neural network is:

[0068]

[0069] Where, P i is the expected spectral response, T i is the predicted spectral response of the cascaded network, N is the number of training samples, D i ' and D i are the design parameters corresponding to the expected spectrum and the design parameters predicted by the cascaded network respectively. λ is the weight ratio of the loss of D i ' and D i Considering different normalization methods, here the spectrum and the design parameters adopt the same normalization method, and λ is set to 1.

[0070] Step 5: Input the expected spectrum into the cascaded neural network model trained in Step 4, and this cascaded neural network model predicts the design parameters to achieve fast reverse design.

[0071] As Figure 4 shown in (a) below, to verify the accuracy of the model, an absorption spectrum is customized from 1000Hz to 5000Hz. Input this expected spectrum into the model, obtain the design parameters, and then simulate the actual spectrum through finite element and compare it with the expected spectrum. It can be found that the accuracy is very good, proving the effectiveness of this method. Figure 4Figure (b) is the error histogram of the cascaded model optimized by PSO, and its average error is about 1.8%. Figure 4 Figure (c) is a traditional cascaded neural network, and its average error is about 2.8%. Obviously, the cascaded neural network model optimized by PSO has better prediction performance.

Claims

1. An acoustic coverage layer design method based on an improved tandem neural network model, characterized in that, It includes the following steps: (1) Construct a data set including the design parameters of the acoustic covering layer and the corresponding spectral responses; (2) Construct a forward neural network and use the PSO optimization algorithm to obtain the optimal initial weights and biases of the forward neural network; (3) Substitute the obtained weights and biases as initial values into the forward neural network for pre-training; (4) Construct a cascaded neural network model in the way that the reverse neural network is in the front and the forward neural network is in the back; train the cascaded neural network model with the data set, and the output of the reverse neural network is used as the input of the pre-trained forward neural network; (5) Input the desired spectrum into the trained cascaded neural network model to predict the structural parameters and material parameters that meet the conditions.

2. The acoustic coverage layer design method based on an improved tandem neural network model according to claim 1, wherein The forward neural network model in step (2) includes an input layer, a hidden layer, and an output layer. The layers are fully connected, and the activation function ReLu and the optimizer Sdg are used.

3. The acoustic coverage layer design method based on an improved tandem neural network model according to claim 1, wherein The process of obtaining the optimal initial weights and biases of the forward neural network by using the PSO optimization algorithm in step (2) is as follows: S1: Initialize the particle swarm: Randomly generate a group of particles, and each particle represents the initial weights and biases of the forward neural network; the position of the particle represents the values of the weights and biases, and the velocity represents the change rate of the weights and biases; S2: Initialize the velocity of the particles: Randomly initialize the velocity of each particle within a suitable range; S3: Calculate the fitness: Use the weight and bias configuration of the current particle to construct a forward neural network, and calculate the loss function or error of the neural network through the training data; S4: Update the individual optimal position: For each particle, update the individual optimal position according to its own historical best fitness; S5: Update the global optimal position: Update the global optimal position according to the best fitness among all particles; S6: Update the velocity and position: For each particle, update the velocity and position according to the current velocity, individual optimal position, and global optimal position; S7: Iteratively execute S3 to S6 until the stop condition is reached; S8: Output the result: Obtain the optimal initial weight and bias configuration according to the optimal particle as the initial weights and biases of the forward neural network.

4. The acoustic coverage layer design method based on an improved tandem neural network model according to claim 1, characterized in that The implementation process of step (3) is as follows: Use the data set to pre-train the forward neural network to establish a forward prediction from the design parameters to the spectral response. The data set is normalized before being input into the forward neural network; during the training process, the design parameters are used as the input of the forward neural network model, and the spectral response is used as the output of the forward neural network model. When the loss function converges, stop training and fix the weights of the forward neural network.

5. The acoustic coverage layer design method based on an improved tandem neural network model according to claim 4, wherein The loss function of the forward neural network is the error between the predicted spectral response and the true spectral response, and MAE is used as the measurement index.

6. The acoustic coverage layer design method based on an improved tandem neural network model according to claim 1, characterized in that The reverse neural network in step (4) includes an input layer, a hidden layer, and an output layer. The layers are fully connected, and the activation function ReLu and the optimizer Sdg are used.

7. An acoustic coverage layer design method based on an improved tandem neural network model according to claim 1, characterized in that The implementation process of step (4) is as follows: Fix the weights of the forward neural network and cascade it with the inverse neural network to construct a cascaded neural network model. Then, use the dataset to train this cascaded neural network model to ensure that the forward neural network and the inverse neural network can work together. Specifically, the output of the inverse neural network, that is, the deterministic prediction design parameters, is used as the input of the forward neural network. Considering the difference in spectral response, the difference between the spectral response input to the cascaded neural network model and the predicted spectral response of the model output is used as an element for constructing the loss function. During the training process, when the loss function converges, stop the training and save the parameters of the cascaded neural network model.

8. An acoustic coverage layer design method based on an improved tandem neural network model according to claim 1, characterized in that, The loss function of the cascaded neural network described in step (4) is: Among them, P i is the expected spectral response, T i is the predicted spectral response of the series network, N is the number of training samples, D i ' and D i are the design parameters corresponding to the expected spectrum and the design parameters predicted by the series network respectively; λ is the weight ratio of the loss of D i ' and D i respectively.

9. The acoustic coverage layer design method based on an improved tandem neural network model according to claim 3, characterized in that The fitness is: Among them, N represents the number of particles, and m is the number of samples; is the predicted spectral response, that is, the output value of the neural network, y i (k) is the expected value in the model.

10. A method for designing an acoustic coverage layer based on an improved tandem neural network model according to claim 8, characterized in that, The λ is set to 1, that is, the same normalization method is used for both the spectrum and the design parameters.

Images