Impedance composite silencer modeling method and system based on weight frequency finite element method neural network

Abstract

The invention provides an impedance composite silencer modeling method and system based on a weight frequency finite element method neural network, and belongs to the field of internal combustion engine silencers. The invention aims to solve the problems of sharp increase of high-frequency calculation amount and low multi-objective optimization efficiency of an existing finite element prediction method for silencer performance. According to the method, the weight frequency finite element neural network is constructed, and the advantages of a physical model and a data driving model are combined, so that full-band efficient performance prediction and optimization design of the silencer are realized. According to the method, the multi-objective optimization efficiency can be improved, a dynamic parameter adjustment mechanism is introduced, frequency weight distribution is optimized, and a scientific and efficient solution is provided for design and optimization of the impedance composite silencer.

Description

Technical Field

[0001] This invention relates to the field of mufflers, and more specifically, to a modeling method and system for impedance composite mufflers based on a weighted frequencyized finite element method neural network. Background Technology

[0002] Impedance composite mufflers are crucial noise reduction devices for marine internal combustion engines, and their performance directly impacts the ship's noise level. Traditional methods for predicting muffler performance primarily rely on physical models (such as the finite element method, FEM) and empirical formulas. While these methods offer high accuracy in the low-frequency range, FEM calculations at high frequencies require significantly more meshes, leading to a surge in computational complexity. Furthermore, although traditional experimental design methods can reduce the number of experiments, they struggle to fully cover design variables in a multi-dimensional parameter space, resulting in limitations in the optimization results.

[0003] In recent years, with the development of machine learning technology, data-driven models (such as neural networks) have shown significant advantages in modeling and optimizing complex systems. However, relying solely on data-driven models also has its problems, especially when data is scarce or noise is high, making it difficult to guarantee the model's generalization ability and physical rationality. Therefore, how to organically combine physical models with data-driven models has become an important research direction in the field of muffler design. Summary of the Invention

[0004] The technical problem to be solved by this invention is:

[0005] To address the problems of the surge in computational load and low efficiency of multi-parameter objective optimization in existing finite element prediction methods for muffler performance prediction.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0007] This invention provides a method for modeling impedance composite mufflers based on weighted frequency-modulated finite element method neural networks, comprising the following steps:

[0008] S100, Constructing the structural parameter space for impedance composite silencers;

[0009] S200. Design orthogonal experiments and generate initial samples through orthogonal experiments;

[0010] S300. Perform stratified sampling on the initial sample obtained in step S200, and then construct a large sample database using Latin hypercube sampling.

[0011] S400: Construct a weighted frequency-based finite element method neural network to output the insertion loss performance index in a large sample database;

[0012] S500: Based on the weighted frequencyized finite element method neural network constructed in step S400, parameter optimization and iteration are performed, namely, the global sensitivity index of Sobol is calculated based on the surrogate model, a subset of sensitive parameters is selected, the parameter configuration is optimized by applying the reinforcement learning algorithm deep Q network, and Pareto optimal solution set is generated.

[0013] S600, design a hybrid loss function, including a finite element model for low-frequency bands and a data-driven model for high-frequency bands.

[0014] Further, in step S100,

[0015] Considering the performance factors affecting impedance composite silencers, including the sound-absorbing material filling density ξ, perforation rate θ, inlet pipe length L1, silencer cavity length L2, outlet pipe length L3, perforated pipe inner diameter d, and perforated pipe wall thickness t. w aperture d h The silencer chamber has an inner diameter D0 and a frequency f. The design parameter space Ω = {ξ, θ, L1, L2, L3, d, t} is defined. w d h ,D0,f}.

[0016] Further, in step S200,

[0017] Construct L n (m l A type () orthogonal array is used to select experimental samples with multiple parameter combinations, choose parameter combinations with significant interactions and appropriate frequency ranges, and generate n representative experimental samples that satisfy:

[0018]

[0019] Where m represents the number of levels for each factor; l represents the number of independent variables in the experiment; m l The orthogonal array can accommodate l factors, each factor having m levels; min represents the minimum, and max represents the maximum.

[0020] The performance parameters of each sample, namely the transfer loss TL, are obtained through orthogonal experiments.

[0021] Further, in step S300,

[0022] For design variable space Perform stratified sampling, and include ξ, θ, L1, L2, L3, d, t w d h D0 is divided into S at equal intervals. ξ S θ , S d , S DSpecifically, including,

[0023] Divide each parameter into equal intervals:

[0024]

[0025] Where k represents the number of layers;

[0026] Then, Latin hypercube sampling is used to uniformly draw samples within each layer, that is, uniformly distributed random numbers are independently generated within each layer. Where ζ represents the distribution of independently generated uniformly distributed random numbers, and U represents the zero-one distribution, the sample values ​​are calculated.

[0027]

[0028] Where S is the total number of layers, used to ensure that the combination of parameters is not repeated and covers the design space; Define the baseline position of the currently extracted sample as the lower limit value of the i-th parameter in the k-th layer (determined in advance by the stratification strategy);

[0029] The experimental database is structured as follows:

[0030]

[0031] Downsampling is performed using a data compression algorithm to generate a large sample database:

[0032]

[0033] in, Represents each orthogonal experimental group in f∈[f min f max ]

[0034] All large sample datasets of transmission loss within the frequency range; q represents the sequence of the number of all large sample datasets of transmission loss; Q represents the maximum value of the sequence of the number of large sample datasets of transmission loss; Ω q (f) represents the input sample space of the training sample space; TL q The label sample space represents the training sample space.

[0035] Further, in step S400, the input layer of the weighted frequency-modulated finite element method neural network contains data from a large sample database, including the sound-absorbing material filling density ξ, perforation rate θ, inlet pipe length L1, muffler cavity length L2, outlet pipe length L3, perforated pipe inner diameter d, and perforated pipe wall thickness t. w aperture d h The inner diameter D0 of the muffler cavity and its frequency f.

[0036] Further, in step S500, the following are included:

[0037] S510. Based on the weighted frequencyized finite element method neural network model, the gradient weights of each parameter with respect to the performance index are generated, and the gradient weights of each parameter with respect to the performance index propagation loss TL are calculated using the backpropagation algorithm.

[0038]

[0039] S520. Visualize the gradient weights, i.e., the impact of sensitive parameters, using visualization tools;

[0040] S530 employs a reinforcement learning algorithm to dynamically adjust parameter combinations to approximate the optimal performance of the propagation loss TL.

[0041] Further, in step S530, the following is included:

[0042] S531. Define the state space s = (ξ, θ, L1, L2, L3, d, t) w d h ,D0);

[0043] S532. Define the action space a = (Δξ0, Δθ0, ΔL1, ΔL2, ΔL3, Δd, Δt w Δd h , ΔD0); among them, Δξ0, Δθ0, ΔL1, ΔL2, ΔL3, Δd, Δt w Δd h ΔD0 represents the sound-absorbing material filling density ξ, perforation rate θ, inlet pipe length L1, silencer cavity length L2, outlet pipe length L3, perforated pipe inner diameter d, and perforated pipe wall thickness t. w aperture d h Changes in parameters such as the inner diameter D0 of the muffler chamber;

[0044] S533, Design the reward function r:

[0045] r = TL(s) - TL(s) prevous )

[0046] Among them, s prevous This represents the state space generated in the previous step;

[0047] S534, Policy Update: Update the policy network parameters using Q-learning.

[0048]

[0049] Where η is the learning rate and γ is the discount factor; s' represents the newly generated state space and a' represents the newly generated action space; Q(s, a) represents the Q-learning network learned using the old state space and old action space; Q(s', a') represents the Q-learning network learned using the new state space and new action space.

[0050] Further, in step S600, the following are included:

[0051] The equations of the finite element acoustic system are:

[0052]

[0053] Where, N i Here, S is the interpolation function, V is the volume of the acoustic domain, and S is the interpolation function. e p represents the unit area; 'p' represents the pressure. Represents the area S of the unit e The normal vector of the boundary; V e h represents the unit volume; h represents the wavenumber; B is the area of ​​the acoustic domain.

[0054] The sound pressure level of the unknown node at the desired frequency is obtained by solving the problem.

[0055] Calculation of transmission loss TL using the finite element method at low frequencies FEM for:

[0056]

[0057] Where, p i and are the sound pressure levels at the inlet and outlet of the muffler, respectively;

[0058] Define the loss function L:

[0059]

[0060] Among them, TL test The transfer loss obtained from the experiment; TL pred This represents the propagation loss obtained from neural network predictions.

[0061] The weighting parameters α and β are:

[0062]

[0063] β(f) = 1 - α(f)

[0064] Low frequency band (f << f c ): α→1, physical model dominates;

[0065] High frequency band (f>>f c ): β→1, data-driven model dominates;

[0066] Transition section (f≈f) c ): Smooth weight transition to avoid prediction jumps;

[0067] Among them, f c Represents the cutoff frequency:

[0068] Based on the applicability of the physical model, when the finite element method assumes plane wave propagation, the excitation frequency of its higher-order modes, i.e., the cutoff frequency, is:

[0069]

[0070] Where c is the speed of sound and D is the characteristic dimension of the muffler;

[0071] The smoothing coefficient is g: g is calculated using an empirical formula, and the relationship between the transition bandwidth Δf and g is:

[0072]

[0073] An impedance composite muffler modeling system based on weighted frequencyized finite element method neural network is provided. The system has program modules corresponding to the above steps and executes the steps in the above-mentioned impedance composite muffler modeling method based on weighted frequencyized finite element method neural network when running.

[0074] A computer-readable storage medium storing a computer program configured to, when invoked by a processor, implement steps of a method for modeling an impedance composite muffler based on a weighted frequencyized finite element method neural network.

[0075] Compared with the prior art, the beneficial effects of the present invention are:

[0076] This invention utilizes a weighted frequency-modulated finite element method neural network (WF-FEM-PINN) to ensure prediction accuracy in the low-frequency range using the finite element method (FEM), while reducing computational load and improving prediction efficiency in the high-frequency range using the neural network, balancing the rigor of the physical model with adaptability to real-world data. By combining gradient weight sensitivity analysis and reinforcement learning (DQN), sensitive parameters are quickly selected and a Pareto optimal solution set is generated, significantly shortening the design iteration cycle and improving optimization efficiency. Orthogonal experimental design and stratified sampling (Latin hypercube sampling) are employed to generate a large subset dataset (Q≥10n) with a small number of experiments, comprehensively covering the multi-parameter design space and enhancing the robustness of model training. The experimental database supports multi-dimensional parameter label indexing and dynamic updates, and data compression algorithms reduce redundancy and improve storage efficiency; smooth transition of weights across frequency bands avoids prediction jumps and improves model stability.

[0077] This invention dynamically balances the contributions of the physical model and data-driven approaches through frequency-related weighting. The low-frequency band is dominated by the physical model, while the high-frequency band is dominated by data-driven approaches, with a smooth transition. A closed-loop design process is formed, from parameter identification and stratified sampling to hybrid surrogate model training, covering the entire process of "parameter selection - data generation - model training - optimization iteration." Reinforcement learning (DQN) is introduced to dynamically adjust parameter combinations, and the reward function is based on the propagation loss increment (ΔTL), enabling autonomous optimization of parameter configuration. Orthogonal array design reduces the number of experiments, and stratified sampling ensures parameter independence and uniform coverage, balancing efficiency and data quality. Attached Figure Description

[0078] Figure 1 This is a flowchart of an impedance composite muffler modeling method based on a weighted frequencyized finite element method neural network in an embodiment of the present invention.

[0079] Figure 2 In the embodiments of the present invention, the weight functions α(f) and β(f) are... f A graph showing the change in frequency.

[0080] Figure 3 This is a design drawing of a muffler according to an embodiment of the present invention;

[0081] Figure 4 This is a flowchart illustrating how reinforcement learning algorithms are used to dynamically adjust parameter combinations to approximate the optimal performance of the transitive loss (TL) in an embodiment of the present invention. Detailed Implementation

[0082] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0083] Specific Implementation Plan 1: Combining Figures 1 to 4 As shown, this invention provides a method for modeling impedance composite mufflers based on a weighted frequency-modulated finite element method neural network, comprising the following steps:

[0084] S100, Constructing the structural parameter space for impedance composite silencers.

[0085] Considering the performance factors affecting impedance composite silencers, including the sound-absorbing material filling density ξ, perforation rate θ, inlet pipe length L1, silencer cavity length L2, outlet pipe length L3, perforated pipe inner diameter d, and perforated pipe wall thickness t. w aperture d h The silencer chamber has an inner diameter D0 and a frequency f. The design parameter space Ω = {ξ, θ, L1, L2, L3, d, t} is defined. w d h ,D0,f};

[0086] S200. Design orthogonal experiments and generate initial samples through orthogonal experiments.

[0087] Construct L n (m l A type () orthogonal array is used to select experimental samples with multiple parameter combinations, choose parameter combinations with significant interactions and appropriate frequency ranges, and generate n representative experimental samples that satisfy:

[0088]

[0089] Where m represents the number of levels for each factor; l represents the number of factors (columns), i.e., the number of independent variables in the experiment; m l The orthogonal array can accommodate l factors, each factor having m levels; min represents the minimum, and max represents the maximum.

[0090] The performance parameters of each sample, i.e., the propagation loss TL, are obtained through orthogonal experiments. Here, f is not involved in the design of the orthogonal experiment table because all performance parameters within f∈Hz must be measured in each orthogonal experiment group.

[0091] S300: Perform stratified sampling and construct a large subsample dataset from the initial samples obtained in step S200.

[0092] For design variable space Perform stratified sampling, and include ξ, θ, L1, L2, L3, d, t w d h D0 is divided into S at equal intervals. ξ S θ , S d , S D Layer (S) ζ S θ , S d , S D

[0093] ≥5), specifically including,

[0094] Divide each parameter into equal intervals:

[0095]

[0096] Where k represents the number of layers;

[0097] Then, Latin hypercube sampling (LHS) is used to uniformly sample within each layer, ensuring that the structural parameters of each muffler are independent and compatible with the frequency dataset f∈[f min f maxThe corresponding performance parameters (transmission loss TL) cover the design space; that is, uniformly distributed random numbers are generated independently within each layer. Where ζ represents the distribution of independently generated uniformly distributed random numbers, and U represents the zero-one distribution, the sample values ​​are calculated.

[0098]

[0099] Where S is the total number of layers, ensuring that the combination of parameters is unique and covers the design space; Define the baseline position of the currently extracted sample as the lower limit value of the i-th parameter in the k-th layer (determined in advance by the stratification strategy);

[0100] The experimental database is established as follows:

[0101]

[0102] The experimental database adopts a parameter-labeled index structure, supporting fast retrieval and dynamic updates based on multi-dimensional parameters. It also reduces storage redundancy through a data compression algorithm, specifically a downsampling algorithm, resulting in a large-sample database:

[0103]

[0104] in, Represents each orthogonal experimental group in f∈[f min f max ] All large sample datasets of transmission loss within the frequency range; q represents the sequence of the number of all large sample datasets of transmission loss; Q represents the maximum value of the sequence of the number of large sample datasets of transmission loss; Ω q (f) represents the input sample space of the training sample space; TL q The label sample space represents the training sample space;

[0105] S400. Construct a weighted frequency-modulated transfer matrix neural network (WF-FEM-PINN). Its input layer contains the sound-absorbing material filling density ξ, perforation rate θ, inlet pipe length L1, muffler cavity length L2, outlet pipe length L3, perforated pipe inner diameter d, and perforated pipe wall thickness t from a large sample database. w aperture d h The silencer chamber has an inner diameter D0 and a frequency f. The design parameter space Ω = {ξ, θ, L1, L2, L3, d, t} is defined. w d h ,D0,f},output layer is the insertion loss performance index in a large sample database;

[0106] S500: Based on the weighted frequency-based finite element neural network constructed in step S400, parameter optimization and iteration are performed. Specifically, the Sobol global sensitivity index is calculated based on the surrogate model, a subset of sensitive parameters is selected, and the deep Q-network (DQN) reinforcement learning algorithm is applied to optimize the parameter configuration, generating a Pareto optimal solution set. This includes...

[0107] S510. Generate the gradient weights of each parameter with respect to the performance index based on the weighted frequencyized finite element neural network (WF-FEM-PINN) model, and calculate the gradient weights of each parameter with respect to the performance index propagation loss TL using the backpropagation algorithm.

[0108]

[0109] S520. Visualize the gradient weights, i.e., the impact of sensitive parameters, using visualization tools;

[0110] S530, combined Figure 4 As shown, a reinforcement learning algorithm is used to dynamically adjust the parameter combination to approximate the optimal performance of the propagation loss TL. Specifically, this includes:

[0111] S531. Define the state space s = (ξ, θ, L1, L2, L3, d, t) w d h ,D0);

[0112] S532. Define the action space a = (Δξ0, Δθ0, ΔL1, ΔL2, ΔL3, Δd, Δt w Δd h , ΔD0); among them, Δξ0, Δθ0, ΔL1, ΔL2, ΔL3, Δd, Δt w Δd h ΔD0 represents the sound-absorbing material filling density ξ, perforation rate θ, inlet pipe length L1, silencer cavity length L2, outlet pipe length L3, perforated pipe inner diameter d, and perforated pipe wall thickness t. w aperture d h Changes in parameters such as the inner diameter D0 of the muffler chamber;

[0113] S533, Design the reward function r:

[0114] r = TL(s) - TL(s) prevous )

[0115] Among them, s prevous This represents the state space generated in the previous step;

[0116] S534, Policy Update: Update the policy network parameters using Q-learning.

[0117]

[0118] Where η is the learning rate, γ is the discount factor; s' represents the newly generated state space, a' represents the newly generated action space; Q(s, a) represents the Q-learning network learned using the old state space and the old action space; Q(s', a') represents the Q-learning network learned using the new state space and the new action space.

[0119] S600, design a hybrid loss function, including a physical model for low-frequency bands and a data-driven model for high-frequency bands.

[0120] The equations of the finite element acoustic system are:

[0121]

[0122] Where, N i Let S be the interpolation function, V be the volume of the acoustic domain, and solving for S yields the sound pressure level at the unknown node at the desired frequency; e p represents the unit area; 'p' represents the pressure. Represents the area S of the unit e The normal vector of the boundary; V e h represents the unit volume; h represents the wavenumber; B is the area of ​​the acoustic domain.

[0123] Commonly used acoustic performance evaluation indicators for silencers include insertion loss, transmission loss, and noise reduction.

[0124] Calculation of transmission loss TL using the finite element method at low frequencies FEM for:

[0125]

[0126] Where, p i ,p t These are the sound pressure levels at the inlet and outlet of the muffler, respectively.

[0127] Define the loss function L:

[0128]

[0129] Among them, TL test The transfer loss obtained from the experiment; TL pred This represents the propagation loss obtained from neural network predictions.

[0130] The weighting parameters α and β are:

[0131]

[0132] β(f) = 1 - α(f)

[0133] Low frequency band (f << fc ): α→1, physical model dominates (transfer matrix method ensures theoretical rationality);

[0134] High frequency band (f>>f c ): β→1, data-driven model dominant (WF-TMM-PINN captures nonlinear high-frequency noise);

[0135] Transition section (f≈f) c ): Smooth weight transition to avoid prediction jumps;

[0136] Among them, f c Representative cutoff frequency: combined with Figure 2 and Figure 3 The figure shows the variation of weights α and β.

[0137] The frequency of higher-order mode excitation, i.e., the cutoff frequency, is determined based on the applicability of the physical model. For example, the finite element method assumes plane wave propagation, and its cutoff frequency is:

[0138]

[0139] Where c is the speed of sound, and D is the characteristic dimension of the muffler (such as the diameter of the expansion cavity);

[0140] The smoothing coefficient is g: g is calculated using an empirical formula, and the relationship between the transition bandwidth Δf (typically 20% of the cutoff frequency) and g is as follows:

[0141]

[0142] This invention overcomes the following drawbacks:

[0143] To address the surge in computational load for high-frequency predictions using traditional finite element methods, a frequency weighting mechanism is introduced. In the low-frequency range, the physical model takes precedence, while in the high-frequency range, the data-driven model takes precedence, achieving a smooth transition and avoiding prediction jumps.

[0144] Improve the efficiency of multi-objective optimization: Employ hierarchical sampling and Latin hypercube sampling (LHS) methods to fully cover the multidimensional parameter space, and combine reinforcement learning algorithms (such as deep Q-networks, DQN) to dynamically adjust parameter combinations to achieve the global optimal solution.

[0145] A dynamic parameter adjustment mechanism is introduced: Based on reinforcement learning algorithm, the design parameters of the muffler are dynamically adjusted to achieve real-time optimization of parameters and improve design efficiency.

[0146] Optimize frequency weight allocation: By defining the cutoff frequency and smoothing coefficient, the frequency weight is dynamically allocated to ensure the prediction accuracy and stability of the model in different frequency bands.

[0147] Specific Implementation Scheme 2: The present invention provides an impedance composite muffler modeling system based on a weighted frequencyized finite element method neural network. The system has a program module corresponding to the above steps, and executes the steps in the above-mentioned impedance composite muffler modeling method based on a weighted frequencyized finite element method neural network when running.

[0148] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0149] Specific Implementation Scheme 3: The present invention provides a computer-readable storage medium storing a computer program configured to implement, when called by a processor, the steps of an impedance composite muffler modeling method based on a weighted frequencyized finite element method neural network.

[0150] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0151] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for modeling impedance composite mufflers based on weighted frequency-modulated finite element method neural networks, characterized in that, Includes the following steps: S100, Constructing the structural parameter space for impedance composite silencers; S200. Design orthogonal experiments and generate initial samples through orthogonal experiments; S300. Perform stratified sampling on the initial sample obtained in step S200, and then construct a large sample database using Latin hypercube sampling. S400: Construct a weighted frequency-based finite element method neural network to output the insertion loss performance index in a large sample database; S500: Based on the weighted frequencyized finite element method neural network constructed in step S400, parameter optimization and iteration are performed, namely, the global sensitivity index of Sobol is calculated based on the surrogate model, a subset of sensitive parameters is selected, the parameter configuration is optimized by applying the reinforcement learning algorithm deep Q network, and Pareto optimal solution set is generated. S600, design a hybrid loss function, including a finite element model for low-frequency bands and a data-driven model for high-frequency bands.

2. The impedance composite muffler modeling method based on weighted frequency-modulated finite element method neural network according to claim 1, characterized in that: In step S100, Considering the performance factors affecting impedance composite silencers, including the sound-absorbing material filling density ξ, perforation rate θ, inlet pipe length L1, silencer cavity length L2, outlet pipe length L3, perforated pipe inner diameter d, and perforated pipe wall thickness t. w aperture d h The silencer chamber has an inner diameter D0 and a frequency f. The design parameter space Ω = {ξ, θ, L1, L2, L3, d, t} is defined. w d h ,D0,f}.

3. The impedance composite muffler modeling method based on weighted frequency-modulated finite element method neural network according to claim 1, characterized in that: In step S200, Construct L n (m l A type () orthogonal array is used to select experimental samples with multiple parameter combinations, choose parameter combinations with significant interactions and appropriate frequency ranges, and generate n representative experimental samples that satisfy: Where m represents the number of levels for each factor; l represents the number of independent variables in the experiment; m l The orthogonal array can accommodate l factors, each factor having m levels; min represents the minimum, and max represents the maximum. The performance parameters of each sample, namely the transfer loss TL, are obtained through orthogonal experiments.

4. The impedance composite muffler modeling method based on weighted frequency-modulated finite element method neural network according to claim 1, characterized in that: In step S300, For design variable space Perform stratified sampling, and include ξ, θ, L1, L2, L3, d, t w d h D0 is divided into S at equal intervals. ξ S θ , S d , S D Specifically, including, Divide each parameter into equal intervals: Where k represents the number of layers; Then, Latin hypercube sampling is used to uniformly draw samples within each layer, that is, to independently generate uniformly distributed random numbers ζ∽U(0,1) within each layer, where ζ represents the distribution of independently generated uniformly distributed random numbers, and U represents the zero-one distribution. The sample values ​​are then calculated. Where S is the total number of layers, used to ensure that the combination of parameters is not repeated and covers the design space; Define the baseline position of the currently extracted sample as the lower limit value of the i-th parameter in the k-th layer (determined in advance by the stratification strategy); The experimental database is structured as follows: Downsampling is performed using a data compression algorithm to generate a large sample database: in, Represents each orthogonal experimental group in f∈[f min f max ] All large sample datasets of transmission loss within the frequency range; q represents the sequence of the number of all large sample datasets of transmission loss; Q represents the maximum value of the sequence of the number of large sample datasets of transmission loss; Ω q (f) represents the input sample space of the training sample space; TL q The label sample space represents the training sample space.

5. The impedance composite muffler modeling method based on weighted frequency-modulated finite element method neural network according to claim 1, characterized in that: In step S400, the input layer of the weighted frequency-modulated finite element method neural network contains data from a large sample database, including the sound-absorbing material filling density ξ, perforation rate θ, inlet pipe length L1, muffler cavity length L2, outlet pipe length L3, perforated pipe inner diameter d, and perforated pipe wall thickness t. w aperture d h The inner diameter D0 of the muffler cavity and its frequency f.

6. The impedance composite muffler modeling method based on weighted frequency-modulated finite element method neural network according to claim 1, characterized in that: In step S500, the following are included: S510. Based on the weighted frequencyized finite element method neural network model, the gradient weights of each parameter with respect to the performance index are generated, and the gradient weights of each parameter with respect to the performance index propagation loss TL are calculated using the backpropagation algorithm. S520. Visualize the gradient weights, i.e., the impact of sensitive parameters, using visualization tools; S530 employs a reinforcement learning algorithm to dynamically adjust parameter combinations to approximate the optimal performance of the propagation loss TL.

7. The impedance composite muffler modeling method based on weighted frequency-modulated finite element method neural network according to claim 6, characterized in that: Step S530 includes, S531. Define the state space s = {ξ, θ, L1, L2, L3, d, t}. w d h ,D0}; S532. Define the action space a = (Δξ0, Δθ0, ΔL1, ΔL2, ΔL3, Δd, Δt w Δd h , ΔD0); among them, Δξ0, Δθ0, ΔL1, ΔL2, ΔL3, Δd, Δt w Δd h ΔD0 represents the sound-absorbing material filling density ξ, perforation rate θ, inlet pipe length L1, silencer cavity length L2, outlet pipe length L3, perforated pipe inner diameter d, and perforated pipe wall thickness t. w aperture d h Changes in parameters such as the inner diameter D0 of the muffler chamber; S533, Design the reward function r: r=TL(s)-TL(s prevous ) Among them, s prevous This represents the state space generated in the previous step; S534, Policy Update: Update the policy network parameters using Q-learning. Where η is the learning rate and γ is the discount factor; s' represents the newly generated state space and a' represents the newly generated action space; Q(s,a) represents the Q-learning network learned using the old state space and old action space; Q(s',a') represents the Q-learning network learned using the new state space and new action space.

8. The impedance composite muffler modeling method based on weighted frequency-modulated finite element method neural network according to claim 1, characterized in that: Step S600 includes, The equations of the finite element acoustic system are: Where, N i Here, S is the interpolation function, V is the volume of the acoustic domain, and S is the interpolation function. e p represents the unit area; 'p' represents the pressure. Represents the area S of the unit e The normal vector of the boundary; V e h represents the unit volume; h represents the wavenumber; B is the area of ​​the acoustic domain. The sound pressure level of the unknown node at the desired frequency is obtained by solving the problem. Calculation of transmission loss TL using the finite element method at low frequencies FEM for: Where, p i and are the sound pressure levels at the inlet and outlet of the muffler, respectively; Define the loss function L: Among them, TL test The transfer loss obtained from the experiment; TL pred This represents the propagation loss obtained from neural network predictions. The weighting parameters α and β are: Low frequency band (f< <f c ):α→1, physical model dominates; High frequency band (f>>f c ):β→1, data-driven model dominates; Transition section (f≈f) c ): Smooth weight transition to avoid prediction jumps; Among them, f c Represents the cutoff frequency: Based on the applicability of the physical model, when the finite element method assumes plane wave propagation, the excitation frequency of its higher-order modes, i.e., the cutoff frequency, is: Where c is the speed of sound and D is the characteristic dimension of the muffler; The smoothing coefficient is g: g is calculated using an empirical formula, and the relationship between the transition bandwidth Δf and g is:

9. A modeling system for impedance composite silencers based on weighted frequency-modulated finite element method neural networks, characterized in that: The system has a program module corresponding to the steps of any one of the claims 1-8 above, and executes the steps in the above-described impedance composite muffler modeling method based on weighted frequencyized finite element method neural network when running.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the impedance composite muffler modeling method based on a weighted frequencyized finite element method neural network according to any one of claims 1-8.

Images