Building outer wall sound insulation effect optimization method and system based on deep learning

Abstract

The application discloses a deep learning-based building outer wall sound insulation effect optimization method and system, relates to the technical field of building energy-saving sound insulation materials, and comprises the following steps: establishing a material-acoustic performance database; constructing and training a process influence correction model; constructing a joint prediction model comprising a generator, a discriminator and an auxiliary predictor which are coupled, and performing joint supervised training based on the material-acoustic performance database; obtaining initial material composition parameters of a target scene, and iteratively optimizing the initial material composition parameters based on the process influence correction model and the joint prediction model to obtain optimal material composition parameters. The application solves the technical problem that the existing building outer wall sound insulation design cannot predict the acoustic performance under different material compositions and process conditions, leading to great blindness in material selection and structure design, and the global optimal configuration of material composition cannot be achieved.

Description

Technical Field

[0001] This application relates to the field of building energy-saving and sound-insulating materials technology, specifically to a method and system for optimizing the sound insulation effect of building exterior walls based on deep learning. Background Technology

[0002] With the acceleration of urbanization and the improvement of people's living standards, the acoustic performance of the built environment, especially the sound insulation effect of the exterior walls, has become one of the indicators for measuring building quality.

[0003] However, traditional sound insulation design for building exterior walls relies heavily on experience and trial-and-error methods, making it difficult to accurately predict acoustic performance under different material compositions and manufacturing processes. This leads to significant uncertainty in material selection and structural design. Furthermore, optimizing sound insulation for building exterior walls is often limited to adjusting single material parameters or simple structural improvements. It lacks the ability to systematically model and accurately predict the complex nonlinear relationships between material composition, manufacturing parameters, and acoustic performance, making it difficult to achieve globally optimal material configuration. Consequently, it fails to meet increasingly stringent sound insulation design requirements and the development trend of green buildings. Summary of the Invention

[0004] This application provides a method and system for optimizing the sound insulation effect of building exterior walls based on deep learning. This solves the technical problem that existing building exterior wall sound insulation designs are difficult to predict the acoustic performance under different material compositions and process conditions, resulting in a large degree of blindness in material selection and structural design, and the inability to achieve the globally optimal configuration of material composition.

[0005] The technical solution to the above-mentioned technical problems in this application is as follows: In a first aspect, this application provides a method for optimizing the sound insulation effect of building exterior walls based on deep learning, the method comprising: Establish a material-acoustic performance database, wherein the material-acoustic performance database includes associated storage of material composition parameters and acoustic performance data of building exterior wall materials; Based on the material-acoustic property database, a process influence correction model is constructed and trained, which is used to correct the material composition parameters. A joint prediction model comprising a coupled generator, discriminator, and auxiliary predictor is constructed and trained under joint supervision based on the aforementioned material-acoustic performance database. The initial material composition parameters of the target scenario are obtained, and the initial material composition parameters are iteratively optimized based on the process influence correction model and the joint prediction model to obtain the optimal material composition parameters.

[0006] Secondly, this application provides a deep learning-based system for optimizing the sound insulation effect of building exterior walls, including: The database creation module is used to create a material-acoustic performance database, wherein the material-acoustic performance database includes associated storage of material composition parameters and acoustic performance data of building exterior wall materials; The correction model training module is used to construct and train a process influence correction model based on the material-acoustic performance database, and the process influence correction model is used to correct the material composition parameters; The prediction model training module is used to construct a joint prediction model containing a coupled generator, discriminator and auxiliary predictor, and to perform joint supervised training based on the material-acoustic performance database. The parameter iteration module obtains the initial material composition parameters of the target scenario, and iteratively optimizes the initial material composition parameters based on the process influence correction model and the joint prediction model to obtain the optimal material composition parameters.

[0007] This application provides one or more technical solutions, which have at least the following technical effects or advantages: This application provides a method and system for optimizing the sound insulation effect of building exterior walls based on deep learning. First, a material-acoustic performance database is established. Second, a process influence correction model is constructed and trained based on this database, effectively correcting the impact of process factors on material composition parameters and improving parameter accuracy. Third, a joint prediction model is constructed, and then the generator and auxiliary predictor are jointly trained, fully utilizing the complex nonlinear relationships in the data to improve the accuracy of acoustic performance prediction. Finally, the initial material composition parameters of the target scene are obtained, resulting in a deterministic prediction model. The initial parameters are iteratively optimized using a comprehensive evaluation model and a multi-objective optimization algorithm to obtain the optimal material composition parameters.

[0008] Through the above technical solution, this method, based on deep learning, captures the complex relationship between material composition, process parameters and acoustic performance, and realizes the global optimization configuration of building exterior wall material composition, effectively solving the problem of blindness in traditional design. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the method for optimizing the sound insulation effect of building exterior walls based on deep learning, as provided in an embodiment of this application. Figure 2This is a schematic diagram of the structure of the deep learning-based building exterior wall sound insulation effect optimization system provided in the embodiments of this application.

[0011] The components represented by each number in the attached diagram are explained below: Database creation module 11, model training correction module 12, prediction model training module 13, parameter iteration module 14. Detailed Implementation

[0012] This application provides a method and system for optimizing the sound insulation effect of building exterior walls based on deep learning. This method addresses the technical problem that existing sound insulation designs for building exterior walls are unable to predict the acoustic performance under different material compositions and process conditions, leading to significant blindness in material selection and structural design, and the inability to achieve globally optimal configuration of material composition.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0015] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid unnecessarily obscuring the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0016] Example 1, as Figure 1 As shown in the embodiments of this application, a method for optimizing the sound insulation effect of building exterior walls based on deep learning is provided, including: S10: Establish a material-acoustic performance database, wherein the material-acoustic performance database includes associated stored material composition parameters and acoustic performance data of building exterior wall materials; In this embodiment, a material-acoustic performance database is established. First, data on commonly used building exterior wall materials are collected, including cement, sand and gravel, thermal insulation cotton, sound insulation felt, and lightweight partition boards. Further, the material composition parameters include the proportions of each raw material, the stoichiometry of key components, and the physical property parameters of the materials.

[0017] Meanwhile, acoustic performance data are obtained through standardized acoustic testing, and the impact of test environment conditions on test results is recorded to establish a material-acoustic performance database.

[0018] Specifically, step S10 in the method includes: Obtain the parametric design space for the building exterior wall materials in the target scene; Based on the parametric design space, high-fidelity acoustic samples are obtained using big data methods. The high-fidelity acoustic samples include at least a high-fidelity acoustic spectrum and associated material composition parameters. Calculate the macroscopic sound insulation index value of the high-fidelity acoustic sample, and combine it with the high-fidelity acoustic spectrum to output the acoustic performance data; The acoustic performance data and the material composition parameters are stored in a structured manner to obtain the material-acoustic performance database.

[0019] In this embodiment of the application, firstly, parametric analysis is performed on the building exterior wall materials of the target scenario to clarify the value range and constraints of each material composition parameter, thereby determining the parametric design space. The parametric design space covers all key material parameter dimensions that may affect acoustic performance.

[0020] Secondly, based on this parametric design space, high-fidelity acoustic samples are obtained using existing big data technology. These high-fidelity acoustic samples include high-fidelity acoustic spectra and corresponding material composition parameters, ensuring the comprehensiveness and accuracy of the samples.

[0021] Then, for the obtained high-fidelity acoustic samples, their macroscopic sound insulation index values, such as weighted sound insulation, are calculated according to relevant acoustic standards and calculation formulas. The macroscopic sound insulation index values ​​are then integrated with the high-fidelity acoustic spectrum to form complete acoustic performance data.

[0022] Finally, a structured data storage method is adopted to associate and store acoustic performance data with corresponding material composition parameters, thereby establishing a material-acoustic performance database.

[0023] S20: Based on the material-acoustic performance database, construct and train a process influence correction model, which is used to correct the material composition parameters; In this embodiment of the application, a process influence correction model is constructed and trained. First, process-related characteristic parameters are extracted from the material-acoustic performance database, including process variables such as stirring speed, stirring time, curing temperature, curing humidity, and molding pressure, as well as the deviation data between the actual measured values ​​and theoretical design values ​​of material composition parameters under the influence of process variables.

[0024] A process influence correction model based on a multilayer perceptron is constructed, using process variables as input features and deviations in material composition parameters as output targets. The hidden layers of this model employ ReLU activation functions, while the output layer uses linear activation functions to accommodate the prediction requirements of continuous deviation values. During the training phase, the Adam optimizer is used, with mean squared error as the loss function, to iteratively optimize the model parameters.

[0025] After training, the process influence correction model can predict the correction amount of material composition parameters under actual production process conditions based on the input process parameters. In other words, the process influence correction model can correct the apparent particle size distribution in the material composition parameters to the true particle size distribution.

[0026] Specifically, step S20 in the method includes: Based on the material-acoustic performance database, the apparent particle size distribution, actual particle size distribution and stirring parameters of the modified material samples are screened and extracted to form the first training dataset. Construct a process impact correction model based on regression analysis; Using the apparent particle size distribution of the sample and the sample stirring parameters as training inputs, and the actual particle size distribution of the sample as supervision, the process influence correction model is iteratively trained until the preset training termination constraint is met. The actual particle size distribution refers to the particle size distribution of the modified material in the building exterior wall material after stirring with the stirring parameters.

[0027] In this embodiment of the application, firstly, sample data involving modified materials are selected from the material-acoustic performance database, and the apparent particle size distribution, actual particle size distribution, and corresponding sample stirring parameters, such as stirring rate, stirring time, and stirring blade type, are extracted and integrated to form the first training dataset.

[0028] The actual particle size distribution refers to the particle size distribution of the modified material in the building exterior wall material after being treated with specific mixing parameters, while the apparent particle size distribution is the particle size distribution of the modified material measured without actual mixing process or standard conditions.

[0029] Furthermore, to obtain the true particle size distribution, a technique combining the Monte Carlo ultrasonic attenuation model and particle swarm optimization algorithm was adopted. This technique, by analyzing the attenuation spectrum of ultrasound in the suspension, can simultaneously deduce the characteristic diameter, distribution width, and mixing ratio of the particles. The calculation deviation for the mixing ratio can be less than 1%, and the deviation for the characteristic diameter is less than 3%.

[0030] Secondly, a process influence correction model based on regression analysis is constructed, such as using a multilayer perceptron model structure. The core objective is to establish a mapping relationship from apparent particle size distribution and stirring parameters to actual particle size distribution.

[0031] During model training, the apparent particle size distribution of the sample and the sample stirring parameters are used together as the training input of the model, and the actual particle size distribution of the sample is used as the supervision signal, i.e. the expected output. By continuously adjusting the model parameters, the error between the particle size distribution predicted by the model and the actual particle size distribution is minimized.

[0032] The training process will continue to iterate until the preset training termination constraints are met. For example, when the prediction error of the model on the validation set is lower than the set threshold, such as the root mean square error (RMSE) or the number of training iterations reaches the preset upper limit, it is considered that the process influence correction model has been trained to maturity and can predict the actual particle size distribution of the modified material after the actual stirring process based on the given apparent particle size distribution and stirring parameters, thereby realizing the process influence correction of key parameters in the material composition.

[0033] For example, the following steps are taken to construct a process impact correction model based on regression analysis: First, data preparation involves selecting sample data related to the modified material from the material-acoustic performance database, including the apparent particle size distribution of the modified material, the actual particle size distribution obtained under specific stirring parameters, and detailed records of the corresponding stirring parameters.

[0034] Secondly, regarding model construction, a multilayer perceptron (MLP) was chosen as the basic model architecture. This MLP model includes an input layer, several hidden layers, and an output layer. The number of neurons in the input layer corresponds to the feature dimension of the apparent particle size distribution. For example, if there are 13 features in the particle size distribution and the number of stirring parameters, then the total number of input neurons is 13. Two to three hidden layers can be set, and the number of neurons in each layer can be determined empirically or through cross-validation. For example, the first layer can have 128 neurons, and the second layer can have 64 neurons. The activation function used in both layers is ReLU to enhance the model's nonlinear fitting ability and alleviate the gradient vanishing problem. The number of neurons in the output layer is consistent with the feature dimension of the true particle size distribution, and a linear activation function is used because the percentage of each interval in the particle size distribution is a continuous value.

[0035] Next, the model is trained by dividing the first training dataset into training, validation, and test sets, for example, in a 7:2:1 ratio. The apparent particle size distribution and mixing parameters of the samples in the training set are used as inputs to the model, while the actual particle size distribution of the samples is used as the supervised training objective. The Adam optimizer is used to optimize the model parameters, with an initial learning rate set to 0.001, which can be dynamically adjusted based on the performance of the validation set during training. The loss function is the mean squared error, which is the mean of the sum of squares of the differences between corresponding elements in the predicted particle size distribution and the actual particle size distribution.

[0036] Furthermore, during training, the model performance is evaluated on the validation set after each iteration, and the parameters of the model with the best performance on the validation set are saved. The training iteration process continues until a preset training termination constraint is met. For example, the preset training termination constraint is that the MSE of the model on the validation set remains unchanged for 50 consecutive iterations, at which point the trained process impact correction model is obtained.

[0037] S30: Construct a joint prediction model that includes a coupled generator, discriminator, and auxiliary predictor, and perform joint supervised training based on the material-acoustic performance database; In this embodiment, a joint prediction model is constructed, which couples a generator, a discriminator, and an auxiliary predictor. The generator employs a deep convolutional neural network architecture to generate acoustic spectra based on material parameters.

[0038] The discriminator also employs a convolutional neural network structure to distinguish between genuine and fake input acoustic spectra. By learning the distribution characteristics of real acoustic spectra, the discriminator evaluates the quality of the generator's output spectrum, and its output probability value represents the confidence level that the input spectrum is a genuine sample.

[0039] The auxiliary predictor is used to predict sound insulation performance indicators based on the characteristics of the generator. The auxiliary predictor can predict macroscopic indicators by utilizing the temporal characteristics of the acoustic spectrum, and generates additional loss signals by comparing the predicted macroscopic indicators with the real macroscopic indicators in the database, thereby constraining and guiding the training of the generator.

[0040] Specifically, step S30 in the method includes: Construct the generator and the discriminator based on a conditional generative adversarial network, wherein the input-output paradigm of the generator includes: a conditional vector and a random noise vector generated by material composition parameters as input, and a synthetic acoustic spectrum as output; A second training dataset is constructed based on the material-acoustic performance database and the input-output paradigm of the generator, and the generator and the discriminator are subjected to adversarial training in combination with a preset alternating training strategy. Construct the auxiliary predictor based on a convolutional neural network, wherein the input-output paradigm of the auxiliary predictor includes: taking the intermediate features of the generator as input and the predicted value of the macroscopic sound insulation index as output; With the parameters of the generator fixed, and the goal of minimizing the residual between the predicted value and the actual value of the macroscopic sound insulation index as the auxiliary training objective, a third training dataset is constructed based on the material-acoustic performance database and the input-output paradigm of the auxiliary predictor to perform auxiliary training of the auxiliary predictor. Unfreeze the parameters of the generator and jointly train the generator and the auxiliary predictor, wherein the goal of the joint training is a weighted sum of the adversarial training objective and the auxiliary training objective.

[0041] In this embodiment, a generator and discriminator based on a conditional generative adversarial network (GAN) are first constructed. The generator employs a deep convolutional neural network architecture. Its input includes a conditional vector generated from material composition parameters and a random noise vector. These two vectors are concatenated as input to the generator, which then undergoes multi-layer convolution, batch normalization, and activation function processing to finally output a synthesized acoustic spectrum. The discriminator also employs a convolutional neural network structure. Its input is the acoustic spectrum, such as a real acoustic spectrum or a synthesized acoustic spectrum generated by the generator. Features are extracted through layer-by-layer convolution, and a probability value is finally output to determine whether the input acoustic spectrum comes from a sample in a real database or a sample synthesized by the generator.

[0042] Secondly, a second training dataset is constructed based on the material-acoustic performance database and the generator's input-output paradigm. This dataset contains material composition parameters, random noise vectors, and corresponding real acoustic spectra.

[0043] Then, the generator and discriminator are subjected to adversarial training using a pre-defined alternating training strategy. Specifically, during training, the generator's parameters are first fixed, and the discriminator is trained to accurately distinguish between the real spectrum and the generated spectrum. Subsequently, the discriminator's parameters are fixed, and the generator is trained to make its generated spectrum as close as possible to the real spectrum, thus "deceiving" the discriminator. This alternating iterative process continues until the quality of the spectrum generated by the generator reaches the preset requirements or the training reaches a state of equilibrium.

[0044] Next, an auxiliary predictor based on a convolutional neural network is constructed. The input to the auxiliary predictor is the intermediate feature map generated by the generator during the acoustic spectrum generation process. The intermediate features contain detailed information about spectrum generation. The auxiliary predictor outputs predicted values ​​of macroscopic sound insulation indicators, such as the weighted sound insulation R, by performing operations such as convolution and pooling on the intermediate features. W wait.

[0045] Specifically, during the training phase of the auxiliary predictor, the generator's parameters are first fixed to ensure the stability of its intermediate features. Then, a third training dataset is constructed based on the material-acoustic performance database and the input-output paradigm of the auxiliary predictor. This dataset includes the generator's intermediate features and the corresponding true values ​​of macroscopic sound insulation indicators. The auxiliary predictor is trained with the goal of minimizing the residual between the predicted and true values ​​of the macroscopic sound insulation indicators, enabling it to accurately predict these indicators based on the generator's intermediate features.

[0046] Finally, the generator parameters are unfrozen, and the generator and auxiliary predictor are jointly trained. At this point, the objective function for joint training is a weighted sum of the adversarial training objective and the auxiliary training objective, i.e., a weighted sum of the adversarial loss of the generator and discriminator and the prediction loss of the auxiliary predictor.

[0047] S40: Obtain the initial material composition parameters of the target scenario, and iteratively optimize the initial material composition parameters based on the process influence correction model and the joint prediction model to obtain the optimal material composition parameters.

[0048] In this embodiment of the application, the initial material composition parameters of the target scenario are obtained, and then the initial material composition parameters are input into the process influence correction model that has been trained.

[0049] The process influence correction model modifies key variables in the initial material composition parameters based on the actual available production process parameters for the target scenario, resulting in modified material parameters. These modified material parameters are then input into the generator of the joint prediction model for iterative optimization of the initial material composition parameters. The adjusted material composition parameters are then input back into the process influence correction model for further process correction, continuously updating the material composition parameters until the optimal material composition parameters are obtained.

[0050] Specifically, step S40 in the method includes: The trained generator is converted into a deterministic model and connected in series with the process influence correction model to obtain a deterministic prediction model. The output of the deterministic prediction model is used as the sound insulation evaluation factor, and a comprehensive evaluation model is constructed by combining the multi-dimensional additional evaluation factors of the target scene. The initial material composition parameters are iteratively optimized based on a multi-objective optimization algorithm and the comprehensive evaluation model.

[0051] In this embodiment, the trained generator is first converted into a deterministic model. During the adversarial training phase, the generator's input includes a random noise vector to enhance the diversity of generated samples. However, in practical optimization applications, to ensure the stability and reproducibility of the prediction results, the influence of random noise needs to be removed.

[0052] Specifically, the input noise vector of the generator is fixed as a zero vector or a fixed constant vector, so that the generator outputs a unique acoustic spectrum under the same material parameter input, thereby converting it into a deterministic model with a definite mapping relationship between input and output. Subsequently, this deterministic generator is connected in series with a pre-trained process influence correction model to form an end-to-end deterministic prediction model.

[0053] The output of the process influence correction model is used as the input of the deterministic generator. That is, the initial material composition parameters are first processed by the process influence correction model to obtain the corrected real material parameters, and then input into the deterministic generator to generate the corresponding acoustic spectrum.

[0054] Secondly, a comprehensive evaluation model is constructed. Based on the acoustic spectrum output by the deterministic prediction model, sound insulation evaluation factors are extracted. These factors include the weighted sound insulation quantity R. W Key indicators include the area under the spectral characteristic curve and the average sound insulation value in a specific frequency band. Additionally, multi-dimensional evaluation factors related to the target scenario are considered, such as material cost, material density, construction feasibility, and environmental indicators.

[0055] Furthermore, the comprehensive evaluation model forms a comprehensive score function by weighting and combining sound insulation evaluation factors with multi-dimensional additional evaluation factors. The weight of each evaluation factor can be determined by expert scoring according to the specific needs of the target scenario, thereby achieving a comprehensive evaluation of the material composition scheme.

[0056] Finally, iterative optimization is performed based on a multi-objective optimization algorithm. Starting with the initial material composition parameters, the scoring function of the comprehensive evaluation model is used as the optimization objective. The multi-objective optimization algorithm generates an initial population by encoding the variables of each dimension of the material composition parameters. Each individual in the population represents a set of material composition parameter schemes. These schemes are input into a deterministic prediction model to obtain sound insulation evaluation factors, and the comprehensive score is calculated by combining these factors with additional evaluation factors.

[0057] The multi-objective optimization algorithm evolves the population through genetic operations such as selection, crossover, and mutation. In each generation, individuals are evaluated and selected based on their comprehensive scores, retaining the best solutions. During the iteration process, the material composition parameters are continuously updated, and the above "parameter correction-performance prediction-comprehensive evaluation" process is repeated until the generated solution set converges or the preset number of iterations is reached.

[0058] Finally, from the Pareto optimal solution set obtained through optimization, the most suitable material composition parameters are selected as the optimal solution based on the specific preferences of the target scenario.

[0059] The trained generator is converted into a deterministic model and concatenated with the process influence correction model to obtain a deterministic prediction model, including: On the validation set, a small perturbation within a preset range is introduced through the auxiliary predictor, and a preset number of iterative predictions are performed to obtain the first perturbation validation set; Analyze the first perturbation verification set, calculate and obtain the first output fluctuation indicator, and compare it with the preset fluctuation significance threshold; If the first output fluctuation indicator is less than the fluctuation significance threshold, then the random noise input in the generator is set to a fixed value, and the generator is converted into a deterministic model accordingly.

[0060] In this embodiment, firstly, a certain number of samples are selected on the validation set, and a small perturbation within a preset range is applied to the intermediate features or input condition vector of the generator through an auxiliary predictor. For example, small noise following a normal distribution is added to each element in the input material composition parameter vector, and the standard deviation of the noise is set to 0.1% of the original feature value to simulate parameter measurement errors or small fluctuations that may occur in actual applications.

[0061] Subsequently, using the input with added perturbation, a preset number of iterative predictions are performed, such as 100 predictions. Each prediction records the acoustic spectrum output by the generator and the macroscopic sound insulation index output by the auxiliary predictor, thereby constructing the first perturbation validation set, which contains the fluctuation of the model output under small perturbations.

[0062] Secondly, the first perturbation validation set is analyzed to calculate the first output fluctuation indicator. Specifically, for the generated acoustic spectrum, its fluctuation amplitude at each frequency point is calculated. For example, the standard deviation of the sound pressure level at the corresponding frequency point after multiple perturbations of the same input sample is calculated, and the average of the standard deviations of all frequency points is taken as the spectrum fluctuation indicator; for macroscopic sound insulation indicators, such as weighted sound insulation R... W Then, the coefficient of variation (COP) of multiple disturbance prediction results is calculated, which is the ratio of the standard deviation to the mean. The COP of the spectral volatility index and the macroeconomic index COP are weighted and summed to obtain a comprehensive first output volatility indicator. This indicator is then compared with a preset volatility significance threshold.

[0063] Through the above operations, the generator will always output a unique acoustic spectrum under the same material composition parameters, thereby realizing the transformation from a generative model containing random factors to a deterministic model with a deterministic mapping of input and output.

[0064] Furthermore, the trained generator is converted into a deterministic model and concatenated with the process influence correction model to obtain a deterministic prediction model, which also includes: If the first output fluctuation indicator is greater than the fluctuation significance threshold, then according to the preset cumulative step size, the number of samplings is iteratively accumulated to obtain the sampling count set; Based on the sampling set, perform a preset number of iterative predictions on the validation set, and calculate the prediction variance of the iterative prediction results for each sampling number. The expected approximate number of samplings for the generator is determined based on the multiple prediction variances under multiple sampling counts. The mean of the prediction results after approximately a certain number of sampling iterations of the generator is defined as the deterministic model.

[0065] In this embodiment, firstly, if the first output fluctuation indicator is greater than the fluctuation significance threshold, it indicates that the generator's output stability under small perturbations does not meet the requirements. To improve the approximate accuracy of the deterministic model for the generator's true output distribution, a multi-sampling averaging strategy can be adopted. Specifically, based on a preset cumulative step size, for example, an initial sampling count of 5 times and a cumulative step size of 5 times, the sampling count is iteratively configured to generate a sampling count set, such as {5, 10, 15, 20, 25}. This sampling count set contains different sampling levels from low to high, used to explore the impact of different sampling counts on the prediction variance.

[0066] Secondly, for each sampling number in the sampling set, a preset number of iterative predictions are performed on the validation set, such as 20 iterations. For each iterative prediction, the same material composition parameters are used as input, but the random noise vector of the generator is sampled independently to generate a corresponding number of acoustic spectrum samples.

[0067] Subsequently, the prediction variance between the acoustic spectra generated by multiple samplings is calculated for each sampling number. For example, for a specific material parameter input and sampling number N, N acoustic spectra are obtained through N independent samplings. The variance of the sound pressure level at each frequency point of the N spectra is calculated, and then the average of the variances of all frequency points is taken as the prediction variance for that sampling number.

[0068] Subsequently, based on multiple prediction variances across various sampling counts, the trend of prediction variance as the number of sampling counts increases is analyzed. Generally, as the number of sampling counts increases, the prediction variance gradually decreases and tends to stabilize. The expected approximate sampling count for the generator is determined using a threshold method. The expected approximate sampling count refers to the number of sampling counts at which the decrease in prediction variance is less than a preset minimum improvement threshold when the number of sampling counts continues to increase. This sampling count is considered sufficient to approximate the generator's true output expectation with high accuracy, while also considering computational efficiency.

[0069] For example, when the prediction variance decreases to less than 5% of the initial variance, the minimum number of samples is set as the expected approximate number of samples, where the initial variance is the variance when the number of samples is 1.

[0070] Finally, the mean of the expected approximate sampling number of iterations of prediction results of the generator is defined as the deterministic model. That is, for any given material composition parameters as input, the deterministic model will generate multiple acoustic spectra by the generator through the expected approximate sampling number of independent random samplings, and then take the arithmetic mean of the sound pressure level at each frequency point of the spectrum as the final deterministic output.

[0071] For example, if the desired number of approximate samplings is determined to be 20, then for a certain modified material composition parameter, the generator will input 20 different random noise vectors while keeping the material parameter fixed, generating 20 acoustic spectra. Subsequently, the sound pressure level values ​​at corresponding frequency points in the 20 spectra are arithmetically averaged to obtain the average acoustic spectrum. This average spectrum serves as the final output of the deterministic model under the given material parameter input.

[0072] Further, obtaining the initial material composition parameters of the target scenario, and iteratively optimizing the initial material composition parameters based on the process influence correction model and the joint prediction model to obtain the optimal material composition parameters, also includes: By concatenating the output layer of the process influence correction model with the output layer of the joint prediction model, a volatility prediction model is obtained, wherein the volatility prediction model takes the initial material composition parameters and stirring parameters as inputs and the synthesized acoustic spectrum as output; Based on the aforementioned volatility prediction model, the density distribution center of the synthetic acoustic spectrum set output by the volatility prediction model is defined as the volatility prediction result. The macroscopic sound insulation index prediction value calculated based on the volatility prediction result is defined as the optimization target. The initial material composition parameters are iteratively optimized based on random variation, and the optimal macroscopic sound insulation index prediction value is output as the optimal material composition parameter.

[0073] In this embodiment, firstly, the output layer of the process influence correction model and the output layer of the joint prediction model are concatenated to construct a volatility prediction model. The input of this volatility prediction model includes the initial material composition parameters of the target scenario and the stirring parameters in the actual production process of the target scenario, as process variables. By integrating the material parameter correction capability of the process influence correction model with the acoustic spectrum generation capability of the joint prediction model, the volatility prediction model outputs a set of synthetic acoustic spectra reflecting the fluctuation range of the material's acoustic performance under different process conditions.

[0074] Secondly, based on the synthetic acoustic spectrum set output by the wave prediction model, the density distribution center of this spectrum set is defined as the wave prediction result. Specifically, for each frequency point in the synthetic acoustic spectrum set, the average sound pressure level of all spectra at that frequency point is calculated, and the spectrum curve formed by the average values ​​is determined as the density distribution center, i.e., the wave prediction result. The central spectrum comprehensively reflects the average level and central tendency of the material's acoustic performance under the influence of different stirring parameters.

[0075] Then, the predicted value of the macroscopic sound insulation index, calculated based on the volatility prediction result, is defined as the optimization objective. For example, the weighted sound insulation quantity R corresponding to the volatility prediction result is used as the optimization objective. W As a core optimization indicator, a random mutation mechanism is introduced when iteratively optimizing the initial material composition parameters. In each iteration, the variables of each dimension of the current material composition parameters, such as the mass percentage and density of each component, are randomly adjusted according to the preset mutation probability and mutation amplitude to generate a new material composition parameter scheme.

[0076] Furthermore, the new material composition parameter schemes and different combinations of stirring parameters are input into the wave-like prediction model to obtain the corresponding synthetic acoustic spectrum set and its density distribution center, i.e., the wave-like prediction result, which is then used to calculate the predicted value of the macroscopic sound insulation index. The predicted values ​​of the macroscopic sound insulation index of each scheme are compared, and the scheme with the best performance is retained for the next iteration. By continuously repeating the process of "parameter variation - wave-like prediction - index evaluation - scheme selection", the material composition parameters are gradually guided to evolve towards a better macroscopic sound insulation index, and finally, the material composition parameters with the best predicted value of the macroscopic sound insulation index are output as the optimal material composition parameters.

[0077] In summary, compared to existing technologies, this application introduces a process influence correction step to establish an independent process influence prediction model, correcting the apparent material parameters designed in the laboratory to more realistic parameters that conform to actual construction conditions. Secondly, it constructs a generative joint prediction model and internalizes the microscopic uncertainties of material performance through an adversarial training mechanism. Finally, by employing strategies such as fixing the noise vector or approximating the expected value, the stochastic generative model in the training phase is transformed into a deterministic high-performance prediction function, thereby searching for the global optimal solution. Considering the influence of actual construction processes and the inherent uncertainties of materials, this application achieves rapid and accurate prediction of the sound insulation performance of vitrified microsphere concrete exterior walls and intelligent and reliable reverse design of material proportions.

[0078] In summary, the embodiments of this application have at least the following technical effects: This application provides a deep learning-based method for optimizing the sound insulation effect of building exterior walls. First, a material-acoustic performance database is established. Second, a process influence correction model is constructed and trained based on this database, effectively correcting the impact of process factors on material composition parameters and improving parameter accuracy. Third, a joint prediction model is constructed, and the generator and auxiliary predictor are jointly trained to fully utilize the complex nonlinear relationships in the data, improving the accuracy of acoustic performance prediction. Finally, the initial material composition parameters of the target scene are obtained, resulting in a deterministic prediction model. The initial parameters are then iteratively optimized using a comprehensive evaluation model and a multi-objective optimization algorithm to obtain the optimal material composition parameters.

[0079] Through the above technical solution, this method, based on deep learning, captures the complex relationship between material composition, process parameters and acoustic performance, and realizes the global optimization configuration of building exterior wall material composition, effectively solving the problem of blindness in traditional design.

[0080] Example 2, as Figure 2 As shown, based on the same inventive concept as the deep learning-based method for optimizing the sound insulation effect of building exterior walls provided in Embodiment 1, this application also provides a deep learning-based system for optimizing the sound insulation effect of building exterior walls, including: Database creation module 11 is used to create a material-acoustic performance database, wherein the material-acoustic performance database includes associated storage of material composition parameters and acoustic performance data of building exterior wall materials; The correction model training module 12 is used to construct and train a process influence correction model based on the material-acoustic performance database, and the process influence correction model is used to correct the material composition parameters. The prediction model training module 13 is used to construct a joint prediction model containing a coupled generator, discriminator and auxiliary predictor, and to perform joint supervised training based on the material-acoustic performance database. The parameter iteration module 14 obtains the initial material composition parameters of the target scenario, and iteratively optimizes the initial material composition parameters based on the process influence correction model and the joint prediction model to obtain the optimal material composition parameters.

[0081] In one embodiment, the database creation module 11 is specifically used for: Obtain the parametric design space for the building exterior wall materials in the target scene; Based on the parametric design space, high-fidelity acoustic samples are obtained using big data methods. The high-fidelity acoustic samples include at least a high-fidelity acoustic spectrum and associated material composition parameters. Calculate the macroscopic sound insulation index value of the high-fidelity acoustic sample, and combine it with the high-fidelity acoustic spectrum to output the acoustic performance data; The acoustic performance data and the material composition parameters are stored in a structured manner to obtain the material-acoustic performance database.

[0082] In one embodiment, the modified model training module 12 is specifically used for: Based on the material-acoustic performance database, the apparent particle size distribution, actual particle size distribution and stirring parameters of the modified material samples are screened and extracted to form the first training dataset. Construct a process impact correction model based on regression analysis; Using the apparent particle size distribution of the sample and the sample stirring parameters as training inputs, and the actual particle size distribution of the sample as supervision, the process influence correction model is iteratively trained until the preset training termination constraint is met. The actual particle size distribution refers to the particle size distribution of the modified material in the building exterior wall material after stirring with the stirring parameters.

[0083] In one embodiment, the prediction model training module 13 is specifically used for: Construct the generator and the discriminator based on a conditional generative adversarial network, wherein the input-output paradigm of the generator includes: a conditional vector and a random noise vector generated by material composition parameters as input, and a synthetic acoustic spectrum as output; A second training dataset is constructed based on the material-acoustic performance database and the input-output paradigm of the generator, and the generator and the discriminator are subjected to adversarial training in combination with a preset alternating training strategy. Construct the auxiliary predictor based on a convolutional neural network, wherein the input-output paradigm of the auxiliary predictor includes: taking the intermediate features of the generator as input and the predicted value of the macroscopic sound insulation index as output; With the parameters of the generator fixed, and the goal of minimizing the residual between the predicted value and the actual value of the macroscopic sound insulation index as the auxiliary training objective, a third training dataset is constructed based on the material-acoustic performance database and the input-output paradigm of the auxiliary predictor to perform auxiliary training of the auxiliary predictor. Unfreeze the parameters of the generator and jointly train the generator and the auxiliary predictor, wherein the goal of the joint training is a weighted sum of the adversarial training objective and the auxiliary training objective.

[0084] In one embodiment, the parameter iteration module 14 is specifically used for: The trained generator is converted into a deterministic model and connected in series with the process influence correction model to obtain a deterministic prediction model. The output of the deterministic prediction model is used as the sound insulation evaluation factor, and a comprehensive evaluation model is constructed by combining the multi-dimensional additional evaluation factors of the target scene. The initial material composition parameters are iteratively optimized based on a multi-objective optimization algorithm and the comprehensive evaluation model.

[0085] Further, in one embodiment, the trained generator is converted into a deterministic model and concatenated with the process influence correction model to obtain a deterministic prediction model, including: On the validation set, a small perturbation within a preset range is introduced through the auxiliary predictor, and a preset number of iterative predictions are performed to obtain the first perturbation validation set; Analyze the first perturbation verification set, calculate and obtain the first output fluctuation indicator, and compare it with the preset fluctuation significance threshold; If the first output fluctuation indicator is less than the fluctuation significance threshold, then the random noise input in the generator is set to a fixed value, and the generator is converted into a deterministic model accordingly.

[0086] Furthermore, in one embodiment, the trained generator is converted into a deterministic model and concatenated with the process influence correction model to obtain a deterministic prediction model, further comprising: If the first output fluctuation indicator is greater than the fluctuation significance threshold, then according to the preset cumulative step size, the number of samplings is iteratively accumulated to obtain the sampling count set; Based on the sampling set, perform a preset number of iterative predictions on the validation set, and calculate the prediction variance of the iterative prediction results for each sampling number. The expected approximate number of samplings for the generator is determined based on the multiple prediction variances under multiple sampling counts. The mean of the prediction results after approximately a certain number of sampling iterations of the generator is defined as the deterministic model.

[0087] Further, obtaining the initial material composition parameters of the target scenario, and iteratively optimizing the initial material composition parameters based on the process influence correction model and the joint prediction model to obtain the optimal material composition parameters, also includes: By concatenating the output layer of the process influence correction model with the output layer of the joint prediction model, a volatility prediction model is obtained, wherein the volatility prediction model takes the initial material composition parameters and stirring parameters as inputs and the synthesized acoustic spectrum as output; Based on the aforementioned volatility prediction model, the density distribution center of the synthetic acoustic spectrum set output by the volatility prediction model is defined as the volatility prediction result. The macroscopic sound insulation index prediction value calculated based on the volatility prediction result is defined as the optimization target. The initial material composition parameters are iteratively optimized based on random variation, and the optimal macroscopic sound insulation index prediction value is output as the optimal material composition parameter.

[0088] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0089] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0090] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for optimizing the sound insulation effect of building exterior walls based on deep learning, characterized in that, include: Establish a material-acoustic performance database, wherein the material-acoustic performance database includes associated storage of material composition parameters and acoustic performance data of building exterior wall materials; Based on the material-acoustic property database, a process influence correction model is constructed and trained, which is used to correct the material composition parameters. A joint prediction model comprising a coupled generator, discriminator, and auxiliary predictor is constructed and trained under joint supervision based on the aforementioned material-acoustic performance database. The initial material composition parameters of the target scenario are obtained, and the initial material composition parameters are iteratively optimized based on the process influence correction model and the joint prediction model to obtain the optimal material composition parameters.

2. The method for optimizing the sound insulation effect of building exterior walls based on deep learning as described in claim 1, characterized in that, The material composition parameters of the building exterior wall material include at least the modified material dosage, the modified material pre-wetting water dosage, and the modified material particle size distribution.

3. The method for optimizing the sound insulation effect of building exterior walls based on deep learning as described in claim 1, characterized in that, Establish a material-acoustic performance database, wherein the material-acoustic performance database includes associated stored material composition parameters and acoustic performance data of building exterior wall materials, including: Obtain the parametric design space for the building exterior wall materials in the target scene; Based on the parametric design space, high-fidelity acoustic samples are obtained using big data methods. The high-fidelity acoustic samples include at least a high-fidelity acoustic spectrum and associated material composition parameters. Calculate the macroscopic sound insulation index value of the high-fidelity acoustic sample, and combine it with the high-fidelity acoustic spectrum to output the acoustic performance data; The acoustic performance data and the material composition parameters are stored in a structured manner to obtain the material-acoustic performance database.

4. The method for optimizing the sound insulation effect of building exterior walls based on deep learning as described in claim 1, characterized in that, Based on the aforementioned material-acoustic property database, a process influence correction model is constructed and trained. This model is used to correct the material composition parameters, including: Based on the material-acoustic performance database, the apparent particle size distribution, actual particle size distribution and stirring parameters of the modified material samples are screened and extracted to form the first training dataset. Construct a process impact correction model based on regression analysis; Using the apparent particle size distribution of the sample and the sample stirring parameters as training inputs, and the actual particle size distribution of the sample as supervision, the process influence correction model is iteratively trained until the preset training termination constraint is met. The actual particle size distribution refers to the particle size distribution of the modified material in the building exterior wall material after stirring with the stirring parameters.

5. The method for optimizing the sound insulation effect of building exterior walls based on deep learning as described in claim 1, characterized in that, A joint prediction model is constructed, comprising a coupled generator, discriminator, and auxiliary predictor, and jointly supervised training is performed based on the aforementioned material-acoustic property database, including: Construct the generator and the discriminator based on a conditional generative adversarial network, wherein the input-output paradigm of the generator includes: a conditional vector and a random noise vector generated by material composition parameters as input, and a synthetic acoustic spectrum as output; A second training dataset is constructed based on the material-acoustic performance database and the input-output paradigm of the generator, and the generator and the discriminator are subjected to adversarial training in combination with a preset alternating training strategy. Construct the auxiliary predictor based on a convolutional neural network, wherein the input-output paradigm of the auxiliary predictor includes: taking the intermediate features of the generator as input and the predicted value of the macroscopic sound insulation index as output; With the parameters of the generator fixed, and the goal of minimizing the residual between the predicted value and the actual value of the macroscopic sound insulation index as the auxiliary training objective, a third training dataset is constructed based on the material-acoustic performance database and the input-output paradigm of the auxiliary predictor to perform auxiliary training of the auxiliary predictor. Unfreeze the parameters of the generator and jointly train the generator and the auxiliary predictor, wherein the goal of the joint training is a weighted sum of the adversarial training objective and the auxiliary training objective.

6. The method for optimizing the sound insulation effect of building exterior walls based on deep learning as described in claim 1, characterized in that, Obtain the initial material composition parameters of the target scenario, and iteratively optimize the initial material composition parameters based on the process influence correction model and the joint prediction model to obtain the optimal material composition parameters, including: The trained generator is converted into a deterministic model and connected in series with the process influence correction model to obtain a deterministic prediction model. The output of the deterministic prediction model is used as the sound insulation evaluation factor, and a comprehensive evaluation model is constructed by combining the multi-dimensional additional evaluation factors of the target scene. The initial material composition parameters are iteratively optimized based on a multi-objective optimization algorithm and the comprehensive evaluation model.

7. The method for optimizing the sound insulation effect of building exterior walls based on deep learning as described in claim 6, characterized in that, The trained generator is converted into a deterministic model and concatenated with the process influence correction model to obtain a deterministic prediction model, including: On the validation set, a small perturbation within a preset range is introduced through the auxiliary predictor, and a preset number of iterative predictions are performed to obtain the first perturbation validation set; Analyze the first perturbation verification set, calculate and obtain the first output fluctuation indicator, and compare it with the preset fluctuation significance threshold; If the first output fluctuation indicator is less than the fluctuation significance threshold, then the random noise input in the generator is set to a fixed value, and the generator is converted into a deterministic model accordingly.

8. The method for optimizing the sound insulation effect of building exterior walls based on deep learning as described in claim 7, characterized in that, The trained generator is converted into a deterministic model and concatenated with the process influence correction model to obtain a deterministic prediction model, which also includes: If the first output fluctuation indicator is greater than the fluctuation significance threshold, then according to the preset cumulative step size, the number of samplings is iteratively accumulated to obtain the sampling count set; Based on the sampling set, perform a preset number of iterative predictions on the validation set, and calculate the prediction variance of the iterative prediction results for each sampling number. The expected approximate number of samplings for the generator is determined based on the multiple prediction variances under multiple sampling counts. The mean of the prediction results after approximately a certain number of sampling iterations of the generator is defined as the deterministic model.

9. The method for optimizing the sound insulation effect of building exterior walls based on deep learning as described in claim 1, characterized in that, The method further includes obtaining initial material composition parameters for the target scenario, and iteratively optimizing these parameters based on the process influence correction model and the joint prediction model to obtain optimal material composition parameters. By concatenating the output layer of the process influence correction model with the output layer of the joint prediction model, a volatility prediction model is obtained, wherein the volatility prediction model takes the initial material composition parameters and stirring parameters as inputs and the synthesized acoustic spectrum as output; Based on the aforementioned volatility prediction model, the density distribution center of the synthetic acoustic spectrum set output by the volatility prediction model is defined as the volatility prediction result. The macroscopic sound insulation index prediction value calculated based on the volatility prediction result is defined as the optimization target. The initial material composition parameters are iteratively optimized based on random variation, and the optimal macroscopic sound insulation index prediction value is output as the optimal material composition parameter.

10. A deep learning-based system for optimizing the sound insulation effect of building exterior walls, characterized in that, The method for optimizing the sound insulation effect of building exterior walls based on deep learning as described in any one of claims 1-9 includes: The database creation module is used to create a material-acoustic performance database, wherein the material-acoustic performance database includes associated storage of material composition parameters and acoustic performance data of building exterior wall materials; The correction model training module is used to construct and train a process influence correction model based on the material-acoustic performance database, and the process influence correction model is used to correct the material composition parameters; The prediction model training module is used to construct a joint prediction model containing a coupled generator, discriminator and auxiliary predictor, and to perform joint supervised training based on the material-acoustic performance database. The parameter iteration module obtains the initial material composition parameters of the target scenario, and iteratively optimizes the initial material composition parameters based on the process influence correction model and the joint prediction model to obtain the optimal material composition parameters.

Images