Aerodynamic Noise Reduction System and Method Based on Semi-Supervised Adversarial Neural Network

Abstract

The present invention relates to a pneumatic noise reduction system and method based on a semi-supervised adversarial neural network, which uses a neural network to autonomously train a noise reduction structure diagram. As long as the required initial boundary conditions are input, it can automatically obtain a noise reduction structure diagram that is superior to the manually designed one, with stable effects and a short cycle. Moreover, the present invention is based on a semi-supervised neural network. When training the discriminative neural network, only a small part of the noise reduction structure diagrams are randomly selected for finite element calculation and the neural network is trained with this, reducing the computational amount of the entire system and improving the overall training speed at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of pneumatic noise reduction, and particularly to a pneumatic noise reduction system and method based on a semi-supervised adversarial neural network. Background Art

[0002] Pneumatic noise is a sound with chaotic amplitudes and frequencies and statistically irregular characteristics directly generated by airflows. In traditional noise reduction structure designs, most rely on the personal experience and intuition of designers for pre-optimization, and then use computers to obtain the simulation effects of multiple envisioned structures. However, this traditional designer mode relying on individuals is costly, has unstable effects, and a long cycle, and can no longer meet the increasingly stringent noise reduction requirements nowadays. Summary of the Invention

[0003] Aiming at the problems existing in the prior art, the purpose of the present invention is to provide a semi-supervised adversarial neural network-based autonomous learning noise reduction system and method, which can output a more optimized noise reduction structure.

[0004] After adopting the above solution, the technical solution adopted by the present invention is as follows:

[0005] A pneumatic noise reduction system based on a semi-supervised adversarial neural network, comprising:

[0006] An initial structure input module, configured to input the external contour of a preset target product and the internal components of a determined structure, and store them in the system; the external contour includes a three-dimensional drawing with material properties, a flow inlet, and a flow outlet; the internal components include a three-dimensional drawing with material properties and the operating parameters of moving components;

[0007] An initial structure constraint condition input module, configured to input additional constraint conditions required by a preset target product and store them in the system; the input external contour includes a three-dimensional drawing of the external contour with material properties, on which a flow inlet and a flow outlet are provided; the internal components include functional components and moving components located in the air duct, the input functional components include a three-dimensional drawing of the functional components with material properties; the input moving components include a three-dimensional drawing of the moving components and operating parameters;

[0008] A neural network generating structure boundary condition module, connected to the initial structure input module and the initial acceptance or rejection constraint condition input module, configured to receive the external contour of a preset target product and the internal components of a determined structure, as well as the additional constraint conditions required by the preset target product, and output a boundary condition feature vector through the neural network;

[0009] A neural network structure generation module is connected to the neural network generation structure boundary condition module for receiving boundary condition feature vectors. This neural network structure generation module is also used to input random noise. The neural network generates an initial generated structure vector according to the input random noise and the boundary condition feature vectors.

[0010] A neural network generated structure transformation module is connected to the neural network generated structure boundary condition module and the generated network neural network structure generation module for receiving the boundary condition feature vectors and the initial generated structure vector, and determining whether the initial generated structure vector meets the requirements of the boundary condition feature vectors through the neural network. If it meets the requirements, a noise reduction structure diagram is output; if it does not meet the requirements, the process terminates.

[0011] A discriminant neural network generated structure scoring module is connected to the neural network generated structure transformation module for receiving the noise reduction structure diagram. The neural network calculates the corresponding score according to the noise reduction structure diagram; and selects the noise reduction structure diagram with the highest score as the final noise reduction structure diagram for output.

[0012] A neural network generated structure simplification module is connected to the neural network generated structure transformation module for periodically and randomly receiving some of the noise reduction structure diagrams. The neural network outputs a simplified noise reduction structure diagram according to the selected noise reduction structure diagrams.

[0013] A finite element aerodynamic noise simulation module is connected to the neural network generated structure simplification module for receiving the simplified noise reduction structure diagram, analyzing it through finite element analysis software, and outputting the actual noise of the noise reduction structure under various initial conditions.

[0014] A discriminant neural network generated structure scoring module retraining module is connected to the finite element noise simulation module and the discriminant neural network generated structure scoring module for receiving the actual noise of some of the noise reduction structure diagrams randomly selected periodically under various initial conditions, and the scores output by them through the discriminant neural network generated structure scoring module, and training the network weights of the discriminant neural network generated structure scoring module in combination with the current network weights of the discriminant neural network generated structure scoring module, and overwriting the original network weights.

[0015] The noise reduction system further includes:

[0016] A neural network structure generation module retraining module is connected to the neural network structure generation module, the neural network generated structure transformation module, and the discriminant neural network generated structure scoring module for receiving whether the current initial generated structure vector passes the structure transformation determination and its score if it passes the determination, and training the network weights of the neural network structure generation module in combination with the current network weights of the neural network structure generation module, and overwriting the original network weights.

[0017] A noise reduction method based on a semi - supervised adversarial neural network, which is implemented based on the above - mentioned noise reduction system, and includes a training part and a noise reduction structure diagram generation part;

[0018] The training part includes the following steps:

[0019] S11. Input various different initial condition target external contours, internal components, and additional constraint conditions;

[0020] S12. Convert the input initial conditions into boundary condition feature vectors recognizable by the neural network through the neural network - generated structure boundary condition module;

[0021] S13. Input the boundary condition feature vectors into the neural network - generated structure generation module. Then, in this module, the generation neural network outputs an initial generation structure vector according to the input boundary condition feature vectors;

[0022] S14. Input the initial generation structure vector and the boundary condition feature vectors into the neural network - generated structure transformation module. The neural network - generated structure transformation module pre - determines whether the current generation meets the basic requirements according to the boundary condition feature vectors. If not, end the current generation and start the next generation. If so, output the noise reduction structure diagram;

[0023] S15. Input the noise reduction structure diagram into the discriminant neural network - generated structure scoring module for scoring;

[0024] S16. Repeat steps S11 to S15 until all training data is trained;

[0025] S17. Regularly and randomly select some noise reduction structure diagrams, and simplify their structures through the neural network - generated structure simplification module to obtain simplified noise reduction structure diagrams, so as to avoid excessive finite - element simulation calculation amounts;

[0026] S18. Analyze the simplified noise reduction structure diagrams through finite - element analysis software to obtain the actual noise of the simplified noise reduction structures under various initial conditions;

[0027] S19. According to the actual noise of the regularly and randomly selected partial noise reduction structure diagrams under various initial conditions, the scores output by them through the discriminant neural network - generated structure scoring module, and combined with the network weights of the current discriminant neural network - generated structure scoring module, train the network weights of the discriminant neural network - generated structure scoring module and overwrite the original network weights;

[0028] Noise reduction structure diagram generation part:

[0029] S21. Input the preset target product external contour, internal components, and additional constraint conditions;

[0030] S22. The structure boundary condition module generated by the neural network converts the input external contour of the preset target product, internal components, and additional constraint conditions into a boundary condition feature vector recognizable by the neural network;

[0031] S23. Input the boundary condition feature vector into the generated neural network structure generation module. Then, in this generated neural network structure generation module, the generated neural network will output an initial generated structure vector according to the input;

[0032] S24. Input the generated initial generated structure vector into the neural network generated structure conversion module. This neural network generated structure conversion module pre-determines whether the current initial generated structure vector meets the requirements according to the boundary condition feature vector. If not, end the current generation and perform the next generation. If it meets the requirements, output the noise reduction structure diagram;

[0033] S25. Input the noise reduction structure diagram into the discriminant neural network generated structure scoring module for scoring.

[0034] S26. Repeat S22 to S25 until the number of repetitions meets the set number of times, and output the noise reduction structure diagram with the highest score as the final noise reduction structure diagram.

[0035] After adopting the above solution, the present invention uses a neural network for autonomous training of the noise reduction structure diagram. As long as the required initial boundary conditions are input, it can automatically obtain a noise reduction structure diagram that is superior to the manually designed one, with stable effects and a short cycle. Moreover, the present invention is based on a semi-supervised neural network. When training the discriminant neural network, only a small number of randomly selected noise reduction structure diagrams are used for finite element calculation and the neural network is trained with this, reducing the computational amount of the entire system and improving the overall training speed at the same time. Brief Description of the Drawings

[0036] Figure 1 is the principle block diagram of the present invention;

[0037] Figure 2 is a schematic diagram of the turbulence calculation model of the fluid simulation software;

[0038] Figure 3 is a schematic diagram of the sub-model of the turbulence calculation model of the fluid simulation software;

[0039] Figure 4 is a schematic diagram of the initial parameters of the turbulence calculation model;

[0040] Figure 5 is a schematic diagram of the aerodynamic noise model of the fluid simulation software. Detailed Embodiments

[0041] As Figure 1As shown in the figure, the present invention discloses a pneumatic noise reduction system based on a semi-supervised adversarial neural network, which includes:

[0042] An initial structure input module, configured to input the external contour of a preset target product and the internal components of a determined structure, and store them in the system. Among them, the input external contour includes a three-dimensional drawing of the external contour with material attributes, and the three-dimensional drawing of the external contour is provided with a flow inlet and a flow outlet; the internal components include functional components and moving components located in the air duct, the input functional components include a three-dimensional drawing of the functional components with material attributes; the input moving components include a three-dimensional drawing of the moving components and operating parameters.

[0043] The external contours and internal components input for different target products will be slightly different. For example, in the noise reduction system of an ultraviolet disinfection machine, assume the goal is the noise condition, such as below 55 dB; the outlet air volume, such as 1000 m³ / h. The external contour is a three-dimensional drawing of the external contour of the ultraviolet disinfection machine (a CAD drawing or a drawing in other formats is acceptable), and the positions of the flow inlet and the flow outlet, as well as the material attributes of the external contour scan, need to be marked on the three-dimensional drawing of the external contour. Among the input internal components, the functional components are three-dimensional drawings of the ultraviolet lamp, the ultraviolet lamp cover, and the photocatalytic plate; the moving components are a three-dimensional drawing of the fan and the operating parameters of the fan (such as rotation speed, wind direction, etc.). The three-dimensional drawings of the functional components and the moving components mark the wiring areas of the components.

[0044] In the noise reduction system of a hair dryer, the external contour is a three-dimensional drawing of the external contour of the hair dryer. Similarly, the positions of the flow inlet and the flow outlet, as well as the material attributes of the external contour scan, need to be marked on the three-dimensional drawing of the external contour. Among the input internal components, the functional components are three-dimensional drawings of the heating wire and the dust-proof net; the moving components are a three-dimensional drawing of the fan and the operating parameters of the fan (such as rotation speed, wind direction, etc.).

[0045] In the noise reduction system of an air purifier, the external contour is a three-dimensional drawing of the external contour of the air purifier. Among the input internal components, the functional components are three-dimensional drawings of the coarse-effect green net and the high-efficiency filter cartridge; the moving components are a three-dimensional drawing of the fan and the operating parameters of the fan (such as rotation speed, wind direction, etc.).

[0046] In the noise reduction system of an air conditioner, the external contour is a three-dimensional drawing of the external contour of the air conditioner. Among the input internal components, the functional components are three-dimensional drawings of the condenser tube; the moving components are a three-dimensional drawing of the fan and the operating parameters of the fan (such as rotation speed, wind direction, etc.).

[0047] The initial structure constraint condition input module is used to input the additional constraint conditions required for the preset target product and store them in the system. The additional constraint conditions include but are not limited to the required turbulence calculation model, aerodynamic noise model, and model initial parameters, the material used for generating the structure, the complexity of the generated structure, and the upper limit value of the final noise under multiple initial conditions. In this embodiment, a fluid simulation software (Fluent) is selected to determine the turbulence calculation model, aerodynamic noise model, and the corresponding model parameters. Currently, the turbulence calculation models included in Fluent are as Figure 2 shown, Figure 3 which is Figure 2 the subdivision model under the selected turbulence calculation model, Figure 4 being the initial parameters of this model. The aerodynamic noise models included in Fluent are as Figure 5 shown, Figure 5 where the Model part in it refers to the aerodynamic noise model, and the right part is the model initial parameters. Different models will have different model initial parameters.

[0048] The neural network generated structure boundary condition module is connected to the initial structure input module and the initial acceptance constraint condition input module, and is used to receive the external contour of the preset target product and the internal components of the determined structure, as well as the additional constraint conditions required for the preset target product, and output the boundary condition feature vector through the neural network.

[0049] The neural network structure generation module is connected to the neural network generated structure boundary condition module, and is used to receive the boundary condition feature vector. This neural network structure generation module is also used to input random noise, and the generated neural network outputs the initial generated structure vector according to the input random noise and the boundary condition feature vector.

[0050] The neural network generated structure transformation module is connected to the neural network generated structure boundary condition module and the neural network structure generation module, and is used to receive the boundary condition feature vector and the initial generated structure vector, and determine whether the initial generated structure vector meets the requirements of the boundary condition feature vector through the neural network. If it meets the requirements, it outputs the noise reduction structure diagram; if it does not meet the requirements, it terminates.

[0051] The discriminant neural network generated structure scoring module is connected to the neural network generated structure transformation module, and is used to receive the noise reduction structure diagram. The neural network calculates the corresponding score according to the noise reduction structure diagram; and selects the noise reduction structure diagram with the highest score as the final noise reduction structure diagram for output.

[0052] The neural network generated structure simplification module is connected to the neural network generated structure transformation module, and is used to randomly receive some noise reduction structure diagrams regularly. The neural network outputs the simplified noise reduction structure diagram according to the selected noise reduction structure diagrams.

[0053] The finite element pneumatic noise simulation module is connected to the neural network generation structure simplification module, and is used to receive the simplified noise reduction structure diagram, analyze it through finite element analysis software, and output the actual noise of the noise reduction structure under each initial condition.

[0054] The discriminant neural network generation structure scoring module retraining module is connected to the finite element pneumatic noise simulation module and the discriminant neural network generation structure scoring module, and is used to receive the actual noise of the periodically randomly selected part of the noise reduction structure diagrams under each initial condition, and the scores output by them through the discriminant neural network generation structure scoring module, and combine the network weights of the current discriminant neural network generation structure scoring module to train the network weights of the discriminant neural network generation structure scoring module, and overwrite the original network weights.

[0055] On the above basis, the noise reduction system of the present invention further includes:

[0056] The generating neural network structure generating module retraining module is connected to the generating neural network structure generating module, the neural network generation structure transformation module, and the discriminant neural network generation structure scoring module, and is used to receive whether the initial generated structure vector passes the structure transformation determination this time and its score if it passes the determination, and combine the network weights of the current generating neural network structure generating module to train the network weights of the generating neural network structure generating module, and overwrite the original network weights.

[0057] Based on the above noise reduction system, the present invention also discloses a pneumatic noise reduction method based on a semi-supervised adversarial neural network, which includes a training part and a noise reduction structure diagram generation part.

[0058] Since the generating neural network structure generating module and the discriminant neural network generation structure scoring module under the initial conditions have very poor effects, adversarial training is needed to improve the performance of both. Among them, the neural network generation structure boundary condition module and the neural network generation structure transformation module are pre-trained modules and do not require additional training. Therefore, the training part includes the following steps:

[0059] S11. Input various different initial condition target external contours, internal components, and additional constraint conditions. Among them, the external contour includes information such as a three-dimensional diagram with material properties, a flow inlet, and a flow outlet, and the internal components include a three-dimensional diagram with material properties, and operating parameters of moving components such as a rotating fluid area (i.e., a fan commonly understood); the additional constraint conditions include but are not limited to the required turbulence calculation model, pneumatic noise model, and model initial parameters, the material used for the generated structure, the complexity of the generated structure, and the upper limit value of the final noise under multiple initial conditions.

[0060] S12. Convert the input initial conditions into boundary condition feature vectors recognizable by the neural network through the neural network generation structure boundary condition module.

[0061] S13. Input the boundary condition feature vector into the generation neural network structure generation module. Then, in this module, the generation neural network outputs an initial generation structure vector according to the input boundary condition feature vector.

[0062] S14. Input the initial generation structure vector and the boundary condition feature vector into the neural network generation structure transformation module. This neural network generation structure transformation module pre-determines whether the current generation meets the basic requirements according to the boundary condition feature vector. If not, end the current generation and perform the next generation. If so, output the noise reduction structure diagram.

[0063] S15. Input the noise reduction structure diagram into the discriminant neural network generation structure scoring module for scoring.

[0064] S16. Repeat steps S11 to S15 until all the training data is trained.

[0065] S17. Regularly and randomly select some of the noise reduction structure diagrams, and simplify their structures through the neural network generation structure simplification module to obtain simplified noise reduction structure diagrams, so as to avoid excessive finite element simulation calculation volume.

[0066] S18. Analyze the simplified noise reduction structure diagrams through finite element analysis software to obtain the actual noise of the simplified noise reduction structures under various initial conditions.

[0067] S19. According to the actual noise of the regularly and randomly selected part of the noise reduction structure diagrams under various initial conditions, the scores output by them through the discriminant neural network generation structure scoring module, and combined with the network weights of the current discriminant neural network generation structure scoring module, train the network weights of the discriminant neural network generation structure scoring module and overwrite the original network weights.

[0068] Among them, S17 and S18 are the semi-supervised part of the present invention. Since it takes a great deal of time to perform finite element simulation on all the noise reduction structure diagrams, the discriminant neural network generation structure scoring module is used to replace the finite element simulation. Only a small number of noise reduction structure diagrams are randomly selected by the system for calculation and training of the discriminant neural network. The random selection is relatively frequent in the initial stage of the system operation, such as at a frequency of 1 / 3, to improve the overall training speed. After the system runs stably, only occasional spot checks are required, such as at a frequency of 1 / 100.

[0069] Noise reduction structure diagram generation part:

[0070] S21. Input the external contour, internal components of the preset target product, and additional limiting conditions.

[0071] S22. The external contour, internal components, and additional constraint conditions of the preset target product input are transformed into boundary condition feature vectors recognizable by the neural network through the neural network generating structure boundary condition module;

[0072] S23. Input the boundary condition feature vectors into the generating neural network structure generating module. Then, in this generating neural network structure generating module, the generating neural network will output an initial generated structure vector according to the input.

[0073] S24. Input the generated initial generated structure vector into the neural network generating structure transformation module. This neural network generating structure transformation module pre-determines whether the current initial generated structure vector meets the requirements according to the boundary condition feature vectors. If it does not meet the requirements, end the current generation and perform the next generation. If it meets the requirements, output the noise reduction structure diagram.

[0074] S25. Input the noise reduction structure diagram into the discriminant neural network generating structure scoring module for scoring.

[0075] S26. Repeat S22 to S25 until the number of repetitions meets the set number of times, and output the noise reduction structure diagram with the highest score as the final noise reduction structure diagram.

[0076] As described above, it is only an embodiment of the present invention and does not impose any limitation on the technical scope of the present invention. Therefore, any minor modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A pneumatic noise reduction system based on a semi-supervised adversarial neural network, characterized in that: The noise reduction system includes: An initial structure input module, which is used to input the external contour of a preset target product and the internal components of a determined structure, and store them in the system; the input external contour includes a three-dimensional external contour diagram with material properties, and a flow inlet and a flow outlet are provided on the three-dimensional external contour diagram; the internal components include functional components and moving components located in the air duct, the input functional components include a three-dimensional functional component diagram with material properties; the input moving components include a three-dimensional moving component diagram and operating parameters; An initial structure constraint condition input module, which is used to input additional constraint conditions required for a preset target product and store them in the system; the additional constraint conditions include but are not limited to the required turbulence calculation model, aerodynamic noise model, and model initial parameters, the material used for generating the structure, the complexity of the generated structure, and the final noise upper limit value under multiple initial conditions; A neural network generating structure boundary condition module, which is connected to the initial structure input module and the initial structure constraint condition input module, and is used to receive the external contour of a preset target product and the internal components of a determined structure, as well as the additional constraint conditions required for a preset target product, and output a boundary condition feature vector through the neural network; A generating neural network structure generating module, which is connected to the neural network generating structure boundary condition module, and is used to receive the boundary condition feature vector. The generating neural network structure generating module is also used to input random noise, and the generating neural network outputs an initial generated structure vector according to the input random noise and the boundary condition feature vector; A neural network generating structure transformation module, which is connected to the neural network generating structure boundary condition module and the generating neural network structure generating module, and is used to receive the boundary condition feature vector and the initial generated structure vector, and determine whether the initial generated structure vector meets the requirements of the boundary condition feature vector through the neural network. If it meets the requirements, a noise reduction structure diagram is output; if it does not meet the requirements, the process terminates; A discriminant neural network generating structure scoring module, which is connected to the neural network generating structure transformation module, and is used to receive the noise reduction structure diagram, and the neural network calculates the corresponding score according to the noise reduction structure diagram; and selects the noise reduction structure diagram with the highest score as the final noise reduction structure diagram for output; A neural network generating structure simplification module, which is connected to the neural network generating structure transformation module, and is used to randomly receive some noise reduction structure diagrams regularly, and the neural network outputs a simplified noise reduction structure diagram according to the selected noise reduction structure diagrams; A finite element aerodynamic noise simulation module, which is connected to the neural network generating structure simplification module, and is used to receive the simplified noise reduction structure diagram, analyze it through finite element analysis software, and output the actual noise of the noise reduction structure under each initial condition; A discriminant neural network generating structure scoring module retraining module, which is connected to the finite element noise simulation module and the discriminant neural network generating structure scoring module, and is used to receive the actual noise of some randomly selected noise reduction structure diagrams under each initial condition, and the score output by it through the discriminant neural network generating structure scoring module, and combine the network weights of the current discriminant neural network generating structure scoring module to train the network weights of the discriminant neural network generating structure scoring module and overwrite the original network weights.

2. The pneumatic noise reduction system based on the semi-supervised adversarial neural network according to claim 1, characterized in that: The noise reduction system further includes: A re-training module for the neural network structure generation module, connected to the neural network structure generation module, the neural network generation structure transformation module, and the discriminant neural network generation structure scoring module, which is used to receive whether the current initial generation structure vector passes the structure transformation determination, the score when it passes the determination, and combine the network weights of the current neural network structure generation module to train the network weights of the neural network structure generation module and overwrite the original network weights.

3. A pneumatic noise reduction method based on a semi-supervised adversarial neural network, characterized in that: The noise reduction method is implemented based on the noise reduction system described in claim 1 or 2, and includes a training part and a noise reduction structure diagram generation part; The training part includes the following steps: S11. Input the external contours, internal components, and additional constraint conditions of various different initial condition target products; S12. Convert the input initial conditions into boundary condition feature vectors that can be recognized by the neural network through the neural network generation structure boundary condition module; S13. Input the boundary condition feature vectors into the neural network structure generation module, and in this module, the neural network generates an initial generation structure vector according to the input boundary condition feature vectors; S14. Input the initial generation structure vector and the boundary condition feature vectors into the neural network generation structure transformation module. The neural network generation structure transformation module pre-determines whether the current generation meets the basic requirements according to the boundary condition feature vectors. If not, end the current generation and perform the next generation. If so, output the noise reduction structure diagram; S15. Input the noise reduction structure diagram into the discriminant neural network generation structure scoring module for scoring; S16. Repeat steps S11 to S15 until all training data is trained; S17. Regularly and randomly select some noise reduction structure diagrams, and simplify their structures through the neural network generation structure simplification module to obtain simplified noise reduction structure diagrams to avoid excessive finite element simulation calculation amounts; S18. Analyze the simplified noise reduction structure diagrams through finite element analysis software to obtain the actual noise of the simplified noise reduction structures under various initial conditions; S19. According to the actual noise of the regularly and randomly selected partial noise reduction structure diagrams under various initial conditions, the scores output by the discriminant neural network generation structure scoring module, and combine the network weights of the current discriminant neural network generation structure scoring module to train the network weights of the discriminant neural network generation structure scoring module and overwrite the original network weights; Noise reduction structure diagram generation part: S21. Input the external contours, internal components, and additional constraint conditions of the preset target product; S22. Convert the input external contours, internal components, and additional constraint conditions of the preset target product into boundary condition feature vectors that can be recognized by the neural network through the neural network generation structure boundary condition module; S23. Input the boundary condition feature vectors into the neural network structure generation module, and in this neural network structure generation module, the neural network will output an initial generation structure vector according to the input; S24. Input the generated initial generated structure vector into the neural network generated structure conversion module. This neural network generated structure conversion module pre-determines whether the current initial generated structure vector meets the requirements according to the boundary condition feature vector. If it does not meet the requirements, end the current generation and perform the next generation. If it meets the requirements, output the noise reduction structure diagram; S25. Input the noise reduction structure diagram into the discriminant neural network generated structure scoring module for scoring; S26. Repeat S22 to S25 until the number of repetitions meets the set number of times, and output the noise reduction structure diagram with the highest score as the final noise reduction structure diagram.

Images