Method for predicting in-vehicle wind noise through mixed model of graph neural network and multi-layer perceptron
By using a hybrid model of graph neural networks and multilayer perceptrons, the problem of high cost and low efficiency in in-vehicle wind noise prediction is solved, enabling rapid evaluation and accurate prediction of early design schemes, and reducing the resource consumption and cost of vehicle wind noise development.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-23
- Publication Date
- 2026-03-24
AI Technical Summary
Existing in-vehicle wind noise prediction technologies suffer from high costs and low efficiency. In particular, during the development of whole-vehicle wind noise, the numerical simulation and wind tunnel verification stages consume huge resources and the design modification costs are high.
A hybrid model combining graph neural networks and multilayer perceptrons is adopted. By establishing a vehicle simulation geometric model, point cloud data and sound source information are obtained, and fluid dynamics and acoustic coupling simulation is performed. A dataset is established and the hybrid model is trained to predict the wind noise response spectrum.
In the early stages or optimization phase of vehicle development, rapid evaluation of design schemes can avoid the need for whole-vehicle wind noise CFD calculations and wind tunnel tests, thereby reducing costs and improving development efficiency and prediction accuracy.
Smart Images
Figure CN121723583A_ABST
Abstract
Description
Technical Field
[0001] This specification relates to the field of vehicle in-vehicle wind noise prediction technology, and in particular to a method for predicting vehicle in-vehicle wind noise using a hybrid model of graph neural networks and multilayer perceptrons. Background Technology
[0002] With the rapid iteration of automotive technology, the trends of electrification, intelligentization, and lightweighting are becoming increasingly prominent, posing greater challenges to aerodynamic design. Meanwhile, consumers' demands for driving comfort have expanded beyond basic quietness to include a comprehensive pursuit of a healthy cabin environment and a high-quality auditory experience. In-vehicle wind noise levels not only directly affect ride comfort but also profoundly impact the clarity of in-vehicle voice interaction, the performance of the audio system, and even fatigue during long-distance driving. Therefore, it has become one of the key indicators for measuring the overall quality, technological sophistication, and brand competitiveness of a vehicle.
[0003] Currently, the industry generally follows a sequential process of "numerical simulation optimization → model wind tunnel verification → prototype wind tunnel calibration" in the development of wind noise in complete vehicles. This process faces significant efficiency bottlenecks and cost pressures in practical applications: In the early stage of numerical simulation, high-fidelity fluid dynamics and acoustic coupling analysis consumes huge amounts of computing resources and is highly dependent on the experience of senior engineers for model setup and result interpretation, resulting in long cycles and high investment; in the mid-stage model wind tunnel verification stage, not only is it necessary to make high-precision, high-rigidity scaled-down or full-size complete vehicle models, but it also relies on extremely scarce and expensive professional aerodynamics-acoustic wind tunnel facilities, making this stage high-barrier and inflexible; what is particularly prominent is that problems are often not fully exposed until the later stage of prototype wind tunnel testing, at which point the vehicle design is basically frozen, and any modifications involving styling, structure, or sealing systems will trigger a chain reaction, causing the cost of "remedial" measures such as mold restart and component re-verification to rise exponentially, resulting in high costs. Summary of the Invention
[0004] This specification provides a method for predicting in-vehicle wind noise using a hybrid model of graph neural networks and multilayer perceptrons, in order to address the high cost problem in existing technologies.
[0005] This specification adopts the following technical solution: a method for predicting in-vehicle wind noise using a hybrid model of graph neural networks and multilayer perceptrons, comprising the following steps: S1: Establish a simulation geometric model of the vehicle, perform data preprocessing on the simulation geometric model, and obtain the whole vehicle point cloud data of each whole vehicle model as the first point cloud data; obtain the point cloud data of the preset sound source area components corresponding to each whole vehicle model as the second point cloud data; S2: Establish a wind noise package model based on the acoustic cabin structure and acoustic characteristics of the vehicle; A fluid dynamics and acoustic coupling simulation calculation is performed on the vehicle's simulation geometric model to obtain the temporal sound source information of the preset sound source region; S3: Perform data processing on the time-series sound source information and the second point cloud data to obtain the sound source spectrum, and input the sound source spectrum into the wind noise packet model to obtain the response spectrum; S4: Establish a dataset based on the first point cloud data, the second point cloud data, the acoustic characteristics, and the response spectrum; A hybrid model based on graph neural networks and multilayer perceptrons was established. S5: Train and optimize the hybrid model based on the data in the dataset to obtain the final prediction model; S6: Input the first point cloud data, the second point cloud data, and the acoustic characteristics of the vehicle to be predicted into the final prediction model to predict the wind noise response spectrum of the vehicle to be predicted.
[0006] Based on the aforementioned technical means, for a new vehicle to be predicted, only its first point cloud data, second point cloud data, and acoustic characteristics need to be input. The wind noise response spectrum can be quickly inferred through the model. In the early or optimization stage of vehicle development, engineers can use this prediction model to quickly evaluate and screen a large number of design schemes, avoiding the huge expenses of whole vehicle wind noise CFD calculation and whole vehicle model wind tunnel testing, improving development efficiency and reducing development costs.
[0007] In this invention, a hybrid model based on graph neural networks and multilayer perceptrons is established. This model can fully utilize the advantages of graph neural networks in processing unstructured geometric data to extract local and global spatial geometric features. At the same time, by utilizing the powerful function fitting capability of multilayer perceptrons, the complex and nonlinear mapping relationship between the geometric features and sound source information of these vehicles and the wind noise response spectrum inside the vehicle is learned, thereby improving the accuracy of the prediction model.
[0008] Furthermore, the preprocessing in S1 specifically includes: removing surfaces in the vehicle's simulation geometric model that are unrelated to wind noise calculation, and performing a stitching operation on the free edges of the vehicle's simulation geometric model to ensure that the simulation geometric model is in a completely closed state.
[0009] Based on the above technical means, by eliminating surfaces in the vehicle's simulation geometry model that are irrelevant to wind noise calculation (such as interior parts, engine compartment interiors, and other components that have little impact on the external flow field and side window area sound sources), the model can be significantly simplified, reducing the consumption of simulation computing resources and the computation time; by stitching together the free edges of the vehicle's simulation geometry model, it is possible to ensure that the model is completely closed for reliable computational fluid dynamics analysis.
[0010] Furthermore, the preset sound source area is the side window area of the vehicle model.
[0011] Furthermore, the time-series sound source information includes time-series turbulent pressure pulsation information and time-series acoustic pressure pulsation information.
[0012] Furthermore, S2 specifically includes: S21: At a pre-set vehicle speed, establish the transient incompressible flow equation and the acoustic disturbance equation; S22: Solve the transient incompressible flow equation and acoustic disturbance equation to obtain the time-series turbulent pressure pulsation information and time-series acoustic pressure pulsation information of the preset sound source region.
[0013] Furthermore, S3 specifically includes the following steps: S31: Perform time-to-frequency conversion on the time-series turbulent pressure pulsation information and the time-series acoustic pressure pulsation information respectively to obtain the frequency domain turbulent pressure pulsation information and the frequency domain acoustic pressure pulsation information; S32: Perform voxel grid downsampling on the first point cloud data and the second point cloud data to obtain the one-third octave band sound source spectrum of the turbulent sound source and the one-third octave band sound source spectrum of the turbulent sound pressure. S33: Input the turbulent pulsating one-third octave band sound source spectrum and the turbulent sound pressure one-third octave band sound source spectrum into the wind noise packet model to obtain the response spectrum.
[0014] By using the above-mentioned technical means, voxel mesh downsampling of the point cloud data reduces the data size and computational complexity, saving storage and computing resources.
[0015] Furthermore, the construction of the hybrid model in S4 specifically involves using the output data of the pooling layer of the graph neural network as the input data of the multilayer perceptron.
[0016] Based on the aforementioned technical means, graph neural networks focus on extracting spatially related features from unstructured geometric data, and then transform them into structured representations through pooling layers; multilayer perceptrons, on the other hand, complete end-to-end high-dimensional nonlinear mapping, combining the spatial structure understanding ability of graph neural networks with the high-dimensional nonlinear fitting ability of multilayer perceptrons, which can increase the prediction accuracy and precision of the model.
[0017] Furthermore, training the hybrid model in S5 specifically includes: using the first point cloud data, the second point cloud data, and the acoustic characteristics in the dataset as inputs to the hybrid model, and using the response spectrum in the dataset as the target output for training the hybrid model.
[0018] Furthermore, the optimization of the hybrid model in S5 specifically includes: using a grid search method to optimize the hyperparameters in the hybrid model.
[0019] Based on the above technical means, the grid search method can avoid local optima or suboptimal solutions caused by randomness or heuristic search by traversing all possible combinations in the predefined hyperparameter space. This ensures that the hybrid model is fully trained to its best performance state, thereby maximizing the potential of the model and ensuring that the trained prediction model has stable and reliable performance, meeting the accuracy requirements of practical engineering applications.
[0020] Furthermore, the hyperparameters in the hybrid model specifically include: the number of neighbors in the hybrid model, the number of hidden layers in the hybrid model, the dimension of the hidden layers in the hybrid model, and the loss function of the hybrid model.
[0021] Based on the aforementioned technical means, these four parameters, from four interrelated yet independent core levels—local structure perception, model architecture depth and width, and optimization criteria—completely cover the control factors that determine the final performance of this hybrid model, avoiding resource waste on irrelevant or secondary parameters and significantly improving the efficiency and targeting of model development and optimization.
[0022] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: 1. For a new vehicle to be predicted, only its first point cloud data, second point cloud data and acoustic characteristics need to be input. The wind noise response spectrum can be quickly inferred through the model. In the early or optimization stage of vehicle development, engineers can use this prediction model to quickly evaluate and screen a large number of design schemes, avoiding the huge expenses of whole vehicle wind noise CFD calculation and whole vehicle model wind tunnel test, improving development efficiency and reducing development costs.
[0023] 2. In this invention, a hybrid model based on graph neural networks and multilayer perceptrons is established, which can make full use of the advantages of graph neural networks in processing unstructured geometric data and extract local and global spatial geometric features. At the same time, by utilizing the powerful function fitting capability of multilayer perceptrons, the complex and nonlinear mapping relationship between the geometric features and sound source information of these vehicles and the wind noise response spectrum inside the vehicle is learned, thereby improving the accuracy of the prediction model. Attached Figure Description
[0024] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall process of this embodiment; Figure 2 This is a schematic diagram illustrating the process of obtaining temporal turbulent pressure and acoustic pressure pulsation information in this embodiment; Figure 3This is a schematic diagram of the data processing flow in this embodiment; Figure 4 This is a schematic diagram of the hybrid model in this embodiment; Figure 5 This is a schematic diagram of the prediction and test results for vehicle A in this embodiment; Figure 6 This is a schematic diagram of the prediction and test results for vehicle B in this embodiment; Figure 7 This is a schematic diagram of the prediction and test results for vehicle C in this embodiment.
[0025] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings. The same or similar reference numerals correspond to the same or similar components. The terms describing positional relationships in the drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.
[0027] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.
[0028] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.
[0029] In the embodiments 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 technical features indicated. Thus, a feature defined with "first" and "second" may explicitly or implicitly include one or more of that feature.
[0030] In the embodiments of this application, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral part; it can be a direct connection or an indirect connection through an intermediate medium.
[0031] In embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0032] The technical solutions provided in the various embodiments of this specification are described in detail below with reference to the accompanying drawings.
[0033] like Figure 1 As shown, this embodiment provides a method for predicting in-vehicle wind noise using a hybrid model of graph neural networks and multilayer perceptrons, including the following steps: S1: Establish a simulation geometric model of the vehicle, and obtain the point cloud data of each vehicle model as the first point cloud data; obtain the point cloud data of the preset sound source area components corresponding to each vehicle model as the second point cloud data.
[0034] In this embodiment, the preprocessing in S1 specifically includes: removing surfaces in the simulation geometric model that are not related to wind noise calculation, and performing a stitching operation on the free edges of the simulation geometric model to ensure that the simulation geometric model is in a completely closed state.
[0035] By eliminating surfaces in the vehicle's simulation geometry that are irrelevant to wind noise calculation (such as interior parts, engine compartment interiors, and other components that have little impact on the external flow field and sound sources in the side window area), the model can be significantly simplified, reducing simulation computational resource consumption and computation time. Stitching together the free edges of the vehicle's simulation geometry ensures that the model is completely closed for reliable computational fluid dynamics analysis.
[0036] In this preferred embodiment, the data preprocessing further includes: encrypting the preset sound source area and the side window glass area with a 2mm grid to obtain the point cloud data of the preset sound source area component after encryption, which is the second point cloud data.
[0037] In this preferred embodiment, the preset sound source area is the side window area of the vehicle model.
[0038] In this embodiment, the components of the preset sound source area include the A-pillar, the rearview mirror, and the side window trim.
[0039] S2: Establish a wind noise package model based on the acoustic cabin structure and acoustic characteristics of the vehicle; The fluid dynamics and acoustic coupling simulation calculations are performed on each of the vehicle simulation geometric models to obtain the temporal sound source information of the preset sound source region; In this embodiment, the acoustic characteristics of the acoustic chamber include the internal volume, internal surface area, glass thickness, glass type, and reverberation time.
[0040] like Figure 2 As shown, in this embodiment, S2 specifically includes the following steps: S21: At a pre-set vehicle speed, establish the transient incompressible flow equation and the acoustic disturbance equation; In this embodiment, the preset vehicle speed is 140 kph.
[0041] S22: Solve the transient incompressible flow equation and acoustic disturbance equation to obtain the time-series turbulent pressure fluctuation information and time-series acoustic pressure fluctuation information of the preset sound source region; In this embodiment, the location information of the sound source and the side window, the temporal turbulent pressure pulsation information of the preset sound source area, and the temporal acoustic pressure pulsation information are output at time intervals of 0.00002s.
[0042] S3: Perform data processing on the time-series sound source information and the point cloud data to obtain the sound source spectrum, and input the sound source spectrum into the wind noise packet model to obtain the response spectrum.
[0043] like Figure 3 As shown, in this embodiment, S2 specifically includes the following steps: S31: Perform time-to-frequency conversion on the time-series turbulent pressure pulsation information and the time-series acoustic pressure pulsation information respectively to obtain the frequency domain turbulent pressure pulsation information and the frequency domain acoustic pressure pulsation information; S32: Perform voxel grid downsampling on the first point cloud data and the second point cloud data to obtain the one-third octave band sound source spectrum of the turbulent sound source and the one-third octave band sound source spectrum of the turbulent sound pressure. S33: Input the turbulent pulsating one-third octave band sound source spectrum and the turbulent sound pressure one-third octave band sound source spectrum into the wind noise packet model to obtain the response spectrum.
[0044] In this preferred embodiment, the first point cloud data and the second point cloud data are downsampled using a voxel grid to obtain downsampled point cloud data, and the information of each node in the side window region of the downsampled point cloud data is converted into frequency domain data to obtain a one-third octave band spectrum of each node from 200Hz to 10000Hz.
[0045] S4: Establish a dataset based on the first point cloud data, the second point cloud data, the acoustic characteristics, and the response spectrum; A hybrid model based on graph neural networks and multilayer perceptrons was established. like Figure 4 As shown, in this embodiment, the hybrid model is established by using the output data of the pooling layer of the graph neural network as the input data of the multilayer perceptron.
[0046] In this embodiment, the establishment of the hybrid model specifically includes the following steps: The basic idea of a graph neural network is to use the point-by-point spectral curve of the side window portion as input data and treat it as a set of nodes in a graph structure. Edges between nodes can be constructed using the K-nearest neighbor strategy to form a local topological structure. Multi-layer graph convolutional modules are introduced to update node features layer by layer. Each convolutional operation aggregates information from neighboring nodes, capturing geometric changes and spectral correlation patterns within the local area of the point cloud. As the number of layers increases, the model gradually acquires the ability to extract global structural features. The node update operation uses a multi-layer perceptron, consisting of an input layer, hidden layers, and an output layer. Each layer consists of multiple neurons, and each layer is connected to the previous layer through a fully connected layer. Finally, the graph features of the graph neural network module are converted into the driver's sound pressure level curve output. The mathematical form of a single-layer perceptron is as follows: , in, As the input source, ~ The parameters to be learned For activation function, This is the output item.
[0047] Multilayer perceptrons learn the complex mapping relationship between unstructured information on the car window surface and the acoustic response inside the car by introducing an activation function (ReLU function) in the hidden layer. The model parameters are optimized by backpropagation algorithm, which enables the model to gradually adjust weights and biases to minimize the loss function and improve the model's fitting ability on the training data.
[0048] The hyperparameter values of the mixture model are determined, and the core parameters include the number of neighbors, the number of hidden layers, the dimension of the hidden layers, and the loss function. For curve fitting quality, [further details are needed]. The evaluation is performed using the following formula: , in, This represents the true value of the i-th sample. This represents the predicted value of the i-th sample. This represents the average value of the sample.
[0049] In this embodiment, a grid search method is used to optimize the hyperparameters in the hybrid model.
[0050] The number of neighbors determines the local receptive field of each node in the graph neural network, controlling the model's precision in capturing local geometric features and serving as a fundamental parameter affecting the representational ability of the graph neural network. The number and dimension of hidden layers jointly determine the capacity and complexity of the multilayer perceptron. The number of hidden layers affects the depth at which the model learns nonlinear mappings, while the dimension affects the information carrying capacity of each layer. Together, they regulate the model's ability to fit and generalize the complex functional relationship between "geometry / sound source to in-vehicle response" extracted from the features of the graph neural network. The loss function, acting as the "guide" for model training, directly defines the optimization objective. For example, choosing between mean squared error and mean absolute error directly affects the model's sensitivity to prediction errors, thus guiding the model's learning direction and being crucial in balancing prediction accuracy and robustness.
[0051] S5: Train the hybrid model based on the data in the dataset; In this embodiment, the training process is as follows: the first point cloud data, the second point cloud data, and the acoustic characteristics in the dataset are used as inputs to the hybrid model, and the response spectrum in the dataset is used as the target output for training the hybrid model.
[0052] In this example, the dataset is divided into training, validation, and test sets in a ratio of 90:7:3.
[0053] In this embodiment, the training process specifically includes: conducting group-by-group controlled variable experiments on hyperparameters to adjust the parameters and obtain an inference model for wind noise source prediction.
[0054] When the above prediction model has high prediction accuracy, only the first point cloud data and the second point cloud data need to be input to infer the wind noise response spectrum of the vehicle driver.
[0055] The loss function used was the mean absolute error (MAE) to evaluate the deviation between the model's predictions and the actual observations. The trained inference model had a correlation coefficient ≥ 0.9 and a mean absolute error ≤ 5%, and the validation set further verified that the model's accuracy met the prediction requirements.
[0056] S6: Input the first point cloud data, the second point cloud data, and the acoustic characteristics of the acoustic chamber of the vehicle to be predicted into the trained hybrid model, which can predict the wind noise response spectrum of the vehicle to be predicted.
[0057] like Figures 5-7 As shown in the figure, in this embodiment, the first point cloud data and the second point cloud data of the test set vehicles are input into the trained model to predict the response spectrum. The comparison between the prediction results and the actual results for some vehicles is shown in Figure 5. Figure 7 As shown.
[0058] In other embodiments, the accuracy of the prediction model can be changed by altering the size of the training set.
[0059] In summary, in this embodiment, for a new vehicle to be predicted, only its first point cloud data, second point cloud data, and acoustic characteristics need to be input. The wind noise response spectrum can be quickly inferred through the model. In the early or optimization stage of vehicle development, engineers can use this prediction model to quickly evaluate and screen a large number of design schemes, avoiding the huge costs of whole vehicle wind noise CFD calculation and whole vehicle model wind tunnel testing, improving development efficiency and reducing development costs.
[0060] In this embodiment, a hybrid model based on graph neural networks and multilayer perceptrons is established. This model can fully utilize the advantages of graph neural networks in processing unstructured geometric data to extract local and global spatial geometric features. At the same time, by leveraging the powerful function fitting capabilities of multilayer perceptrons, the complex and nonlinear mapping relationship between the geometric features and sound source information of these vehicles and the wind noise response spectrum inside the vehicle is learned, thereby improving the accuracy of the prediction model.
[0061] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention.
Claims
1. A method for predicting in-vehicle wind noise using a hybrid model of graph neural networks and multilayer perceptrons, characterized in that, Includes the following steps: S1: Establish a simulation geometric model of the vehicle, perform data preprocessing on the simulation geometric model, and obtain the whole vehicle point cloud data of each whole vehicle model as the first point cloud data; The point cloud data of the preset sound source area components corresponding to each of the vehicle models is obtained as the second point cloud data; S2: Establish a wind noise package model based on the acoustic cabin structure and acoustic characteristics of the vehicle; A fluid dynamics and acoustic coupling simulation calculation is performed on the vehicle's simulation geometric model to obtain the temporal sound source information of the preset sound source region; S3: Perform data processing on the time-series sound source information and the second point cloud data to obtain the sound source spectrum, and input the sound source spectrum into the wind noise packet model to obtain the response spectrum; S4: Establish a dataset based on the first point cloud data, the second point cloud data, the acoustic characteristics, and the response spectrum; A hybrid model based on graph neural networks and multilayer perceptrons was established. S5: Train and optimize the hybrid model based on the data in the dataset to obtain the final prediction model; S6: Input the first point cloud data, the second point cloud data, and the acoustic characteristics of the vehicle to be predicted into the final prediction model to predict the wind noise response spectrum of the vehicle to be predicted.
2. The method for predicting in-vehicle wind noise using a hybrid model of graph neural network and multilayer perceptron as described in claim 1, characterized in that, The preprocessing in S1 specifically includes: removing surfaces in the vehicle's simulation geometry model that are irrelevant to wind noise calculation, and performing a stitching operation on the free edges of the vehicle's simulation geometry model to ensure that the simulation geometry model is in a completely closed state.
3. The method for predicting in-vehicle wind noise using a hybrid model of graph neural network and multilayer perceptron as described in claim 1, characterized in that, The preset sound source area is the side window area of the vehicle model.
4. The method for predicting in-vehicle wind noise using a hybrid model of graph neural network and multilayer perceptron as described in claim 1, characterized in that, The time-series sound source information includes time-series turbulent pressure pulsation information and time-series acoustic pressure pulsation information.
5. The method for predicting in-vehicle wind noise using a hybrid model of graph neural network and multilayer perceptron as described in claim 4, characterized in that, S2 specifically includes: S21: At a pre-set vehicle speed, establish the transient incompressible flow equation and the acoustic disturbance equation; S22: Solve the transient incompressible flow equation and acoustic disturbance equation to obtain the time-series turbulent pressure pulsation information and time-series acoustic pressure pulsation information of the preset sound source region.
6. The method for predicting in-vehicle wind noise using a hybrid model of graph neural network and multilayer perceptron as described in claim 5, characterized in that, S3 specifically includes the following steps: S31: Perform time-to-frequency conversion on the time-series turbulent pressure pulsation information and the time-series acoustic pressure pulsation information respectively to obtain the frequency domain turbulent pressure pulsation information and the frequency domain acoustic pressure pulsation information; S32: Perform voxel grid downsampling on the first point cloud data and the second point cloud data to obtain the one-third octave band sound source spectrum of the turbulent sound source and the one-third octave band sound source spectrum of the turbulent sound pressure. S33: Input the turbulent pulsating one-third octave band sound source spectrum and the turbulent sound pressure one-third octave band sound source spectrum into the wind noise packet model to obtain the response spectrum.
7. The method for predicting in-vehicle wind noise using a hybrid model of graph neural network and multilayer perceptron according to claim 1, characterized in that, The construction of the hybrid model in S4 specifically involves using the output data of the pooling layer of the graph neural network as the input data of the multilayer perceptron.
8. The method for predicting in-vehicle wind noise using a hybrid model of graph neural network and multilayer perceptron according to claim 1, characterized in that, The training of the hybrid model in S5 specifically includes: using the first point cloud data, the second point cloud data, and the acoustic characteristics in the dataset as inputs to the hybrid model, and using the response spectrum in the dataset as the target output for training the hybrid model.
9. The method for predicting in-vehicle wind noise using a hybrid model of graph neural network and multilayer perceptron as described in claim 1 or 8, characterized in that, The optimization of the hybrid model in S5 specifically includes: using a grid search method to optimize the hyperparameters in the hybrid model.
10. The method for predicting in-vehicle wind noise using a hybrid model of graph neural network and multilayer perceptron according to claim 9, characterized in that, The hyperparameters in the hybrid model specifically include: the number of neighbors in the hybrid model, the number of hidden layers in the hybrid model, the dimension of the hidden layers in the hybrid model, and the loss function of the hybrid model.