Vehicle Wind Noise Prediction Using Auto-Encoder-Based Point Cloud Compression and GWO-ResNet

Abstract

In response to the inability to quickly assess wind noise performance during the early stages of automotive styling design, this paper proposes a method for predicting interior wind noise by integrating automotive point cloud models with the Gray Wolf Optimization Residual Network model (GWO-ResNet). Based on wind tunnel test data under typical operating conditions, the point cloud model of the test vehicle is compressed using an auto-encoder and used as input features to construct a nonlinear mapping model between the whole vehicle point cloud and the wind noise level at the driver’s left ear. Through adaptive optimization of key hyperparameters of the ResNet model using the gray wolf optimization algorithm, the accuracy and generalization of the prediction model are improved. The prediction results on the test set indicate that the proposed GWO-ResNet model achieves prediction results that are consistent with the actual measured values for the test samples, thereby validating the effectiveness of the proposed method. A comparative analysis with traditional ResNet models, GWO-LSTM models, and LSTM models revealed that the GWO-ResNet model achieved Mean Absolute Percentage Error (MAPE) and mean squared error (MSE) of 9.72% and 20.96, and 9.88% and 19.69, respectively, on the sedan and SUV test sets, significantly outperforming the other comparison models. The prediction results on the independent validation set also demonstrate good generalization ability and stability (MAPE of 10.14% and 10.15%, MSE of 23.97 and 29.15), further proving the reliability of this model in practical applications. The research results provide an efficient and feasible technical approach for the rapid evaluation of wind noise performance in vehicles and provide a reference for wind noise control in the early design stage of vehicles. At the same time, due to the limitations of the current test data, it is impossible to predict the wind noise during the actual driving of the vehicle. Subsequently, the wind noise during actual driving can be predicted by the test data of multiple working conditions.

1. Introduction

With the continuous development of the automotive industry, consumers’ focus on overall vehicle performance is gradually shifting from power and economy to higher requirements for comprehensive performance such as ride comfort and quietness [1]. Among these, interior noise, as one of the key factors affecting driving comfort, is receiving increasing attention in research and engineering practice. The causes of vehicle noise are complex, with the main sources including powertrain noise [2,3], tire–road contact noise [4,5], aerodynamic noise [6,7], and chassis noise [8,9]. Among these, aerodynamic noise is primarily caused by high-speed airflow flowing around the vehicle body surface, resulting in flow separation and turbulent structures at complex curved surfaces or irregular structures of the vehicle body. Research indicates that when vehicle speed exceeds 100 km/h, the strong shear force between the vehicle body and the air significantly increases the sound pressure level of aerodynamic noise, with its energy contribution in the mid-to-high frequency range (800–5000 Hz) accounting for over 60% of the total interior noise [10]. At the same time, with the acceleration of the electrification of automobiles, the proportion of noise from traditional powertrains in overall vehicle noise has decreased significantly [11,12], further highlighting the adverse impact of aerodynamic noise on the interior acoustic environment. Related studies have pointed out that strong aerodynamic noise may not only reduce concentration, thereby increasing traffic safety risks [13], but may also affect the physical and mental health of drivers [14]. In addition, areas with high noise levels are often accompanied by significant airflow disturbances and increased air resistance, which in turn adversely affect the energy consumption and range of electric vehicles [15]. Therefore, effective control of aerodynamic noise during vehicle development has become a key technical challenge in improving overall vehicle comfort and energy efficiency.

The evaluation of current wind noise performance in vehicles mainly relies on a combination of wind tunnel experiments and actual road tests, using multi-channel acoustic data collection and aerodynamic parameter measurements to improve testing accuracy and efficiency [16]. Among these, wind tunnel testing has become an important means of studying the contribution of external structural noise to vehicle noise due to its standardization and high controllability. With the help of microphone array beamforming algorithms, it is possible to accurately identify the dominant noise sources of sensitive structures such as A-pillars and exterior rearview mirrors, while surface pressure sensors help capture flow excitation characteristics. Related studies, such as Duell et al. [17], conducted measurements on full-scale vehicles in a DNW acoustic wind tunnel to identify the wind noise contributions of components such as side mirrors, A-pillars, and antennas, and proposed noise reduction optimization strategies that reduced the interior noise level by approximately 3 dB. He et al. [18] deployed a 120-channel planar spiral array and a 48-channel spherical array at the Shanghai Automotive Wind Tunnel Center to locate noise sources outside and inside the vehicle and quantify their acoustic coupling relationships. Chen et al. [19] compared the acoustic performance of biomimetic designs and traditional structures through rearview mirror wake noise analysis and numerical modeling. The results showed that the former had better noise reduction effects in the frequency range above 500 Hz. However, wind tunnel tests are typically conducted under single operating conditions (such as constant wind speed, yaw angle, and turbulent flow), making it difficult to fully replicate the complex, non-steady aerodynamic environment found on actual roads. In contrast, road testing can utilize vehicle-mounted multi-channel data acquisition systems synchronized with GPS to capture the combined effects of actual factors such as road surface excitation and natural wind interference on wind noise. Barden et al. [20] conducted comparative tests on heavy-duty trucks under wind tunnel and road conditions and found significant differences in wind noise results under the same operating conditions, verifying the idealization of the wind tunnel test environment.

With the development of computing resources and numerical simulation technology, computational fluid dynamics (CFD) has become an important supplementary method for studying aerodynamic noise. Talay et al. [21] used CFD simulation to study changes in surface pressure at different vehicle speeds, investigated the impact of door structure deformation on interior noise, and proposed evaluating the air intake performance of the sealing system during the initial design stage to reduce wind noise. Oettle et al. [22] proposed a CFD solver based on the Lattice Boltzmann Method and combined it with Statistical Energy Analysis (SEA) to predict the acoustic response inside a vehicle, providing an efficient tool for acoustic design. Jiang et al. [23] constructed a simulation model to reproduce the low-drag and moving floor scenarios and compared it with wind tunnel test results to verify the accuracy of the aerodynamic simulation. Chen et al. [24] combined the SEA method with unsteady CFD to identify the vibration sound power of the car body structure and assess the sensitivity of each subsystem to the noise inside the car, further supporting multi-source acoustic optimization design. Although CFD technology has effectively reduced physical testing costs and improved R&D efficiency, it still has limitations in terms of turbulence modeling and acoustic transmission accuracy, especially when performing wind noise analysis in the mid-to-high frequency range, where it is highly dependent on computing resources [25]. Therefore, how to efficiently extract key wind noise characteristics based on actual modeling data and combine artificial intelligence methods to construct a high-precision wind noise performance prediction model has become an important research direction in the field of interior aerodynamic noise control. The data-driven method is significantly less dependent on computing resources than the CFD method, and the prediction of the results can be quickly achieved after the model training is completed.

In recent years, data-driven methods have received widespread attention in the field of automotive interior noise suppression. With machine learning and big data analysis, these methods can effectively uncover complex nonlinear relationships between noise characteristics and vehicle parameters [26,27], significantly improving the accuracy of noise prediction models and speeding up the acoustic performance development process [28,29]. Huang et al. [30] used convolutional neural networks (CNNs), Residual Network (ResNet), and long short-term memory networks (LSTMs) to build a side window wind noise prediction model based on wind tunnel test data, revealing a highly nonlinear relationship between wind noise and design features. Sun et al. [31] introduced the support vector regression method to construct a machine learning model for predicting the sound insulation performance of electric vehicle body systems, and validated it on specific vehicle models, confirming its reliability in sound insulation performance prediction. Wang et al. [32] used machine learning methods to establish a mapping relationship between material parameters and sound insulation performance, avoiding complex acoustic transmission path modeling and enabling rapid performance evaluation of new sound insulation materials. Wu et al. [33] constructed a multi-objective prediction and optimization model for acoustic packages based on ResNet, achieving simultaneous improvements in prediction accuracy and optimization efficiency. Chen et al. [34] optimized the acoustic performance and structural weight of porous materials using a particle swarm optimization algorithm. Simulation results showed that sound absorption capacity could be maintained or even improved while reducing weight. Huang et al. [28] proposed a method combining adaptive learning rate forests with improved LSTM, integrating sound absorption and insulation theory with data-driven strategies, effectively enhancing the performance of sound absorption and insulation materials. It can be seen from the above research that the data-driven method has been widely used in the prediction of vehicle interior noise, but the data-driven method, especially the neural network-based method, has a strong correlation between the speed of model training and its input and output dimensions. In order to improve the training speed of the model, the dimension of the input data should be compressed as much as possible.

In the development of automotive noise, vibration, and harshness, data-driven methods have been widely adopted due to their ability to efficiently replace traditional development processes [35]. Taking wind noise performance prediction as an example, deep learning models trained based on wind tunnel or real-vehicle road test data can achieve rapid assessment of wind noise performance, significantly reducing redundant testing and improving development efficiency. At the same time, such methods have advantages in terms of computing resource utilization [36], enabling parallel evaluation of multiple design schemes during the conceptual design stage, thereby providing strong support for early decision-making. Compared to traditional methods that rely on physical experiments, data-driven methods can effectively reduce testing costs and resource investment [37,38] by modeling and predicting existing simulation or test data [39]. In the aerodynamic design process, wind noise prediction models can serve as preliminary validation tools to quickly determine the wind noise response trends of new designs under specific conditions, thereby facilitating rapid design iteration and optimization.

Based on the aforementioned research foundation, this paper proposes a ResNet-based method for predicting wind noise inside vehicles. The method uses vehicle body point cloud data as input and the 1/3 octave noise level at the driver’s left ear position as the output indicator to establish a nonlinear mapping relationship between wind noise performance inside the vehicle. By utilizing deep neural networks to uncover the coupling patterns between point cloud geometric features and in-vehicle wind noise, this method provides an efficient data-driven tool for wind noise assessment during the early design stage. It demonstrates excellent generalization capability and engineering adaptability, significantly reducing the prediction process and aiding in the optimization of vehicle aerodynamic noise design. The main technical contributions of this study are as follows:

(1)

We propose a dimension reduction method for vehicle point cloud data based on Auto-Encoder (AE), which can effectively compress data dimensions while retaining geometric feature information to the greatest extent possible, thereby improving the training efficiency and stability of subsequent prediction models.

(2)

Based on the AE dimension reduction results, an efficient method for predicting wind noise inside vehicles was developed. Combining experimental and simulation data, a prediction model was established with point cloud files as input and the 1/3 octave wind noise level at the right ear of the driver as output, enabling rapid assessment and advanced control of wind noise inside vehicles during the design stage.

The structure of this paper is organized as follows: Section 2 introduces point cloud AE dimension reduction methods, ResNet, and the improved model GWO-ResNet obtained by combining ResNet with the GWO algorithm; Section 3 describes the testing methods and data collection process for in-vehicle wind noise, and analyzes the characteristics of the experimental data; Section 4 details the construction and training process of the AE dimension reduction model, and compares and analyzes the dimension reduction effects; Section 5 proposes an indoor wind noise prediction strategy based on GWO-ResNet, verifies the model’s prediction performance and generalization ability, and compares it with other typical models; Section 6 summarizes the entire work and research conclusions. The overall research process of this paper is shown in Figure 1.

2. Proposed Methods

2.1. FoldingNetcF

Auto-encoders are a classic unsupervised learning method that can automatically extract key features from input data in the absence of labeled data [40]. Its core mechanism lies in capturing the essential representation of data through compression and reconstruction processes: the encoder first maps the input in the original data space (X) to a low-dimensional hidden space (H), forming a compact latent representation (h); then, the decoder attempts to accurately reconstruct the input data in the original space (X) based on this latent representation (h) [41], as show in Figure 2. By minimizing the reconstruction error between the input and the reconstructed output, the model effectively extracts the most representative feature information from the data during compression and restoration.

Currently, there are three main methods for processing 3D point cloud data: projection method [42], voxelization method [43], and direct point-based processing method [44,45]. Among them, the projection method maps the 3D point cloud onto a 2D plane and then uses traditional image processing techniques for feature extraction and analysis [46], which has the advantages of low computational cost and the ability to use mature image models. However, this method is prone to losing spatial depth information during the mapping process, leading to feature occlusion or information sparsity. The voxelization method converts point clouds into regular three-dimensional voxel grids, making them suitable for processing by three-dimensional convolutional neural networks [47]. Although this method has advantages in terms of structural standardization, it inevitably introduces quantization errors [48]. While increasing voxel resolution can reduce information loss, it also significantly increases the computational overhead of the model. In contrast, point-based direct processing methods do not require preprocessing of the point cloud, enabling the maximum retention of original geometric information [49,50], but typically require more complex network structures to achieve effective encoding. Considering the need for precise retention of automotive exterior features and the dependency of subsequent prediction models on input data quality, this paper adopts a point cloud AE architecture based on the FoldingNet framework.

Point cloud is a non-structured data format composed of a number of three-dimensional points in space. The order of the points is not fixed and may contain information such as coordinates, normal vectors, and colors. Due to its non-Euclidean structure, traditional grid-based data processing methods (such as CNN) are difficult to apply directly. To address this issue, PointNet was proposed, which for the first time adopted a mult-layer perceptron (MLP) with shared parameters and maximum pooling operations to extract global features from point clouds, effectively solving the problem of point cloud disorder [51]. However, PointNet is mainly aimed at classification and segmentation tasks, and its ability to express geometric details is still limited in unsupervised learning scenarios (such as point cloud reconstruction). To improve point cloud reconstruction accuracy, FoldingNet introduces a “folding” mechanism based on PointNet, achieving adaptive mapping between structured grids and unstructured point clouds [52]. The FoldingNet structure consists of an encoder, decoder, and loss function, enabling effective feature compression and reconstruction while maintaining geometric consistency [53].

The FoldingNet encoder is based on the PointNet architecture and mainly consists of three parts: encoder, decoder, and loss function. The FoldingNet is based on the PointNet architecture and aims to extract global features from the input point cloud. It mainly extracts and compresses global features through three steps. Assume that the input point cloud is a set of (n) points $P=\left\{p_i\in R^3{\mid }_{i=1,2,3,\dots ,n}\left.\right\}\right.$, where $R^3$ refers to the three-dimensional coordinates represented by each point $p_i=\left(x_i,y_i,z_i\right)$. First, perform point-by-point feature extraction. For each point $p_i$, extract local features through an $MLP$ with shared weights. Assuming that the parameters of the $MLP$ are $\varphi$, then for each point, we have the following:

$$f_i=ML{P}_{\varphi }\left(p_i\right)\in R^d$$

(1)

where d is the dimension of local features.

Then, through maximum pooling, global features are extracted from the local features of all points:

$$g=MaxPool\left(\left\{f_i\left|i=1,2,3,\dots ,n\right.\right\}\right)\in R^d$$

(2)

Max pooling ensures that the global feature $g$ remains unchanged in the point cloud arrangement, which effectively addresses issues caused by point cloud disorder.

The final step is feature compression, which further compresses the global features $g$ through an additional $MLP$ to satisfy the latent vector $z\in R^k$ set by the auto-encoder, where $k\ll d$:

$$z=ML{P}_{\varphi }\left(g\right)\in R^k$$

(3)

where $\varphi$ is the parameter of $MLP$ and $k$ is the dimension of the latent space.

The encoder maps the input point cloud to a low-dimensional latent vector $z$ while preserving global features as much as possible.

The decoder of FoldingNet restores point clouds from low-dimensional latent vectors to the original space through a “folding” operation. Traditional autoencoders typically use fully connected layers or convolutional layers for decoding, but neither of these methods is suitable for the disorder and geometric characteristics of point clouds. The decoder’s input includes a latent vector $z\in R^k$ and a predefined two-dimensional grid $G=\left\{g_j\in R^2{\mid }_{j=1,2,3, \dots ,m}\right\}$, which is typically a uniformly sampled 2D grid. The decoding process consists of two stages of folding operations:

The first fold first copies the latent vector $z$ $m$ times, concatenates it with the two-dimensional grid points $g_j$, and forms the input features:

$$h_j=\left[z,g_j\right]\in R^{k+2}$$

(4)

Then, map $h_j$ to three-dimensional space through an $MLP$ (parameter ${\psi }_1$):

$$q_j=ML{P}_{{\psi }_1}\left(h_j\right)\in R^3$$

(5)

This step “folds” the two-dimensional grid points into three-dimensional space, generating an intermediate point cloud $Q=\left\{q_j\left|j=1,2,3,\dots ,m\right.\right\}$.

The second folding first concatenates the middle point cloud $Q$ with each point $q_j$ and the potential vector $z$ again:

$$h_j^{\prime }=\left[z,q_j\right]\in R^{k+3}$$

(6)

Further adjust the position of the points through another $MLP$ (parameter ${\psi }_2$):

$$p_j^{\prime }=ML{P}_{{\psi }_2}\left(h_j^{\prime }\right)\in R^3$$

(7)

This step generates the final reconstructed point cloud $P_j^{\prime }=\left\{p_j^{\prime }\left|j=1,2,3,\dots ,m\right.\right\}$.

The last component is the loss function. Since FoldingNet is an unsupervised learning model, its training objective is to minimize the difference between the input point cloud $P$ and the reconstructed point cloud $P^{\prime }$. The loss function used in the self-encoder constructed in this paper is the chamfer distance, which is commonly used in point cloud reconstruction tasks. It is mainly used to measure the difference or distance between two point clouds and is robust to the disorder of point clouds [54]. Its expression is as follows:

$$L(P,P^{\prime })={\displaystyle \sum _{p\in P}}\underset{p^{\prime }\in P^{\prime }}{\min }\Vert p-p^{\prime }{\Vert }_2^2+\sum _{p^{\prime }\in P^{\prime }}\underset{p\in P}{\min }\Vert p^{\prime }-p{\Vert }_2^2$$

(8)

2.2. GWO-Resnet

2.2.1. Resnet

ResNet address the “degradation phenomenon” inherent in convolutional neural networks by introducing “short-circuit” or “skip connections.” This prevents a decline in model accuracy as the number of neural network layers increases, facilitating the training of deeper networks. ResNet primarily consists of convolutional layers and residual blocks [55]. The convolution layers in residual networks are consistent with those in convolutional neural networks, and residual blocks are the structure that implements “shortcuts” or “skip connections” in residual neural networks [56]. The residual block is defined by the optimal function $H_{\left(x\right)}$, the residual function $f_{\left(x\right)}$, and the identity function $X$. Taking ResNet18 as an example, the residual function $f_{\left(x\right)}$ consists of two 3 × 3 convolution layers and ReLU activation functions connecting them (the function expression is shown in Formula (9)), while the optimal function $H_{\left(x\right)}$ is associated with the residual function $f_{\left(x\right)}$ and the identity function $X$ through Formula (10). Its specific structure is shown in Figure 3:

$$f\left(x\right)=\left\{\begin{array}{l}x,x\ge 0\hfill \\ 0,x<0\hfill \end{array}=\max \left(0,x\right)\right.$$

(9)

From the residual block structure, it can be seen that after the convolution layer calculation, the data is processed through the residual function $f_{\left(x\right)}$ on one hand and the identity function $X$ on the other. Since the two resulting matrices may have different sizes and channel numbers, the solution is as follows: If the final two matrices have different sizes, they are padded to the same size using the padding method. If the final two matrices have different channel numbers, a 1 × 1 convolution kernel is used to unify the channel numbers of the two matrices [57]. After unifying the size and channel numbers of the resulting matrices, they are added together to obtain the optimal function for that layer. After passing through the ReLU activation function, the output can be input to the next layer. The mathematical expression for the residual block can be written as shown in Formula (11):

$$H\left(x\right)=F\left(x\right)+x$$

(10)

$$y=\sigma \left(F\left(x,W\right)+x\right)$$

(11)

In this formula, $y$ represents the output of the residual block, $\sigma \left(.\right)$ represents the activation function, $x$ represents the input of the layer, and $W$ represents all the weights within the residual block. For machine learning, nonlinear transformations are key operations that analyze the characteristics of data by mapping it to a high-dimensional space. As the number of network layers increases, so does the number of nonlinear transformations, and more activation functions are introduced into the model. Data is mapped to a more discrete space, and subsequent layers often limit features to only a certain part, leading to a “degradation phenomenon.” ResNet residual neural networks achieve linear transformations through the identity function of residual blocks, ensuring that the learning effectiveness of each layer is at least as good as the previous layer. This allows for more layers to be stacked, improving model training effectiveness. The residual learning mechanism of ResNet makes the network easier to train and achieves better performance in deeper network structures. Compared with traditional networks, ResNet has significantly improved recognition performance on ImageNet and also achieved good results in object detection [58,59] and semantic segmentation [60,61]. Its network structure is simple and easy to implement, making it one of the foundational models for many deep learning tasks.

2.2.2. Introduction to GWO-Resnet

ResNet effectively alleviates the gradient vanishing and model degradation problems in deep neural networks through a residual connection mechanism, significantly improving the training stability and performance of the model. However, its performance is highly dependent on the reasonable configuration of hyperparameters. Key hyperparameters such as network depth, number of convolution kernels, learning rate, and batch size directly affect the training efficiency and final accuracy of the model [62]. Inappropriate parameter configuration not only significantly prolongs training time but may also lead to a decline in model performance. Traditional hyperparameter optimization methods, such as grid search and random search, are inefficient and prone to getting stuck in local optima, making it difficult to efficiently explore ideal parameter combinations. Therefore, adopting more efficient optimization algorithms to tune the hyperparameters of ResNet has become a key direction for further improving model performance [63].

The GWO is an intelligent optimization algorithm that simulates the social hierarchy and hunting behavior of gray wolves. It has the advantages of fast convergence speed, simple parameter settings, and easy implementation, and performs well in global optimization problems [64]. This group intelligence-based hierarchical collaboration mechanism enables GWO to demonstrate strong optimization performance and stability when solving complex optimization problems. The position update of gray wolves is determined by the following formula:

$$\overrightarrow{X}\left(t\right)={\overrightarrow{X}}_p\left(t\right)\cdot \overrightarrow{A}\times \overrightarrow{D}$$

(12)

$$\overrightarrow{D}=\left|\overrightarrow{C}\times {\overrightarrow{X}}_p\left(t\right)-\overrightarrow{X}\left(t\right)\right|$$

(13)

In this formula, $t$ represents the current iteration number, $\overrightarrow{A}$ and $\overrightarrow{C}$ represent the coefficient vectors controlling convergence and exploration, ${\overrightarrow{X}}_p$ represents the position vector of the prey, and $\overrightarrow{X}$ represents the current position vector of the gray wolf individuals.

Parameter vectors $\overrightarrow{A}$ and $\overrightarrow{C}$ are determined by the following formulas:

$$\overrightarrow{A}=2\overrightarrow{a}\times {\overrightarrow{r}}_1-\overrightarrow{a}$$

(14)

$$\overrightarrow{C}=2{\overrightarrow{r}}_2$$

(15)

$$\overrightarrow{a}=2\cdot 2{\displaystyle \frac{t}{T}}$$

(16)

In this formula, vector $\overrightarrow{a}$ is a control parameter that decreases linearly with the number of iterations, ${\overrightarrow{r}}_1$ and ${\overrightarrow{r}}_2$ are random vectors uniformly distributed between 0 and 1, and $\mathrm{T}$ is the maximum number of iterations. GWO’s adaptive adjustment factor and information feedback mechanism can dynamically adjust the search direction and step size during the optimization process, thereby improving optimization efficiency and accuracy.

When applying GWO to optimize the ResNet model, the fitness function of GWO can be defined as the prediction error of the ResNet model on the validation set, usually measured by Mean Square Error (MSE) or Root Mean Square Error(RMSE):

$$MSE={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}$$

(17)

In this formula, $y_i$ is the actual value, and ${\widehat{y}}_i$ is the predicted value of the ResNet model. The optimization process includes the following steps:

First, initialize the positions of the gray wolf population, i.e., randomly set the hyperparameters of the ResNet model; second, search for the optimal hyperparameter combination by iteratively updating the positions of the gray wolves; finally, select the optimal hyperparameter combination based on the fitness function value and apply it to the ResNet model. The global search capability of GWO enables it to effectively avoid getting stuck in local optima, thereby improving the generalization ability and prediction accuracy of the ResNet model. The computational logic of the ResNet model based on the gray wolf optimization algorithm is shown in Figure 4:

3. In-Vehicle Master Driving Noise Test

3.1. Test Platform

A wind tunnel is a specialized testing facility designed to simulate the operational state of vehicles in natural wind conditions, widely used in the research and development of vehicle aerodynamic performance and wind noise performance [65]. Based on the airflow circulation method, wind tunnels are typically classified into direct-flow and recirculating-flow types, with recirculating-flow wind tunnels being widely adopted in modern automotive testing due to their excellent airflow stability and high energy efficiency. Depending on the configuration of the test section, wind tunnels can also be categorized into open and closed structures. In recent years, newly constructed wind tunnels have generally adopted a 3/4 open recirculating structure and integrated a ground effect simulation system at the bottom of the test section to more accurately reproduce the relative motion state of the road surface during actual vehicle operation. Based on their functional purposes, automotive wind tunnels can be further subdivided into aerodynamic and acoustic, and environmental wind tunnels. Among these, aerodynamic acoustic wind tunnels are specifically designed for testing and evaluating wind noise inside and outside vehicles, serving as an indispensable platform in wind noise performance development.

The test platform used in this study is an aerodynamic and acoustic wind tunnel constructed by a certain wind tunnel center. The wind tunnel has the following key performance parameters: test section length of 16 m, nozzle area of 28 m², maximum wind speed of 250 km/h, temperature control range of 20–25 °C, boundary layer thickness not exceeding 5 mm, low-frequency pressure pulsation less than 0.01, axial static pressure gradient below 0.001, and turbulence controlled within 0.2%.To optimize the acoustic performance of the wind tunnel, the fan, flow channels, corner structures, and plenum chamber have all undergone acoustic absorption treatment, ensuring that background noise is kept below 59 dB (A) at a wind speed of 150 km/h. The plenum chamber cutoff frequency is 50 Hz, and both the overall flow field and acoustic environment meet international advanced standards. Additionally, the wind tunnel is equipped with an advanced acoustic testing system, including a 576-channel sound source identification system and a fourth-generation artificial head testing system, supporting acoustic testing tasks for various vehicle types such as passenger cars, commercial vehicles, and unmanned aerial vehicles (UAVs), enabling high-precision wind noise identification and quantitative analysis. Figure 5 shows the plan layout diagram of the aerodynamic acoustic wind tunnel.

3.2. Test Methods

The test selected one suburban utility vehicle (SUV) and one sedan model each from the mass-produced vehicles of a certain automobile group company as research subjects. The interior noise test setup is shown in Figure 6. The microphone was installed on the left side of the headrest of the driver’s seat, with a sampling frequency set to 48 kHz and a single sampling duration of 15 s.

To ensure the stability of the aerodynamic environment in the test area, the floor area of the wind tunnel was standardized before the test began. This included installing movable covers over the sinking balance area, closing the floor suction and moving belt systems, and ensuring smooth airflow. The test vehicle was precisely positioned at the center of the wind tunnel test section, with the longitudinal symmetry plane of the vehicle body maintained within ±0.1° of the wind tunnel center axis to ensure symmetry conditions.

The vehicle seats were adjusted to the standard mid-position, with the seatbacks kept vertical, and the microphone fixed to the left side of the headrest. No additional weights were added to the vehicle interior. The vehicle was secured using the parking brake, the cooling fan was turned off but allowed to rotate freely, and the air conditioning system was set to internal circulation mode with all air vents closed. Rearview mirrors were kept open to maintain consistency with the test conditions.

After the wind tunnel system parameters are set, the data acquisition system is first calibrated to ensure measurement accuracy and data consistency. The wind speed is then set to 120 km/h, the yaw angle to 0°, and data collection begins once the flow field stabilizes. Each test lasts for at least 15 s, and each test vehicle is subjected to three or more repeat tests to verify data stability and repeatability. A total of 50 vehicle test samples were collected in this test. After the test, the system was recalibrated to ensure the traceability and validity of the data throughout the process.

3.3. Test Results for Noise in the Driver’s Left Ear Inside the Vehicle

The post-processing frequency range for the collected in-vehicle noise data was set to 200–8000 Hz. For feature analysis, one SUV and one sedan were randomly selected from the test samples for comparison.

Figure 7 shows the 1/3 octave spectrum curve for the left ear area of the driver’s seat. The results indicate that neither vehicle type exhibits sharp peaks in the spectrum curves, suggesting that the overall aerodynamic acoustic condition of the test vehicles is good and no significant abnormal noise sources are present.

Further comparison revealed that the wind noise levels at the driver’s seat were generally higher in SUVs than in sedans. This phenomenon is primarily attributed to the taller and wider body profile of SUVs, which is more prone to flow separation and turbulence under high-speed airflow, thereby generating stronger aerodynamic noise. Additionally, the larger interior space of SUVs results in more complex sound propagation paths, facilitating echoes and cavity resonance, which further amplify the perceived noise intensity in localized areas of the cabin.

4. Auto-Encoder Modeling and Analysis

4.1. Car Point Cloud AE Modeling

The dataset used in this study was provided by an automotive group and contains 3D model files of 200 different vehicle models in STereoLithography (STL) format. Each STL file is approximately 9 MB in size and contains approximately 9 million discrete points in the point cloud model. Considering the large number of points in the point cloud, directly inputting the complete point cloud data into the auto-encoder model would result in significant computational resource consumption. Therefore, this paper preprocesses the original point cloud by reducing the data dimension through a downsampling strategy. The specific processing flow is as follows: first, the original point cloud is preliminarily thinned using equidistant sampling. Although this method can preserve geometric distribution uniformity to a certain extent, it is difficult to accurately control the number of output points. To meet the model’s requirement for a fixed input size, a random sampling mechanism is introduced after equidistant sampling. If the number of points in the initial sampling result exceeds the model’s required input point count, random downsampled is performed; otherwise, random upscaling is used to supplement the points to achieve the desired dimension. This approach preserves the geometric contour information of the point cloud while ensuring the consistency and standardization of the network input data structure. After multiple rounds of testing and comparative analysis, the standard input point count for the auto-encoder was finally determined to be 8192. Figure 8 shows the point cloud comparison results before and after downsampled. Although the point density has decreased, the global contours of the vehicle’s exterior are still effectively preserved.

4.2. Point Cloud Dimension Reduction Analysis

In the point cloud dimensionality reduction modeling process, to balance model performance and computational efficiency, three sets of different encoding output dimensions were selected for comparison and analysis, namely 1024 × 3, 512 × 3, and 256 × 3. For the above three dimensions configurations, self-encoder models were constructed, and the same dataset, training rounds, and batch size were used for training. Model performance was evaluated by calculating the chamfer distance between the reconstructed point cloud and the original point cloud, reflecting the reconstruction fidelity of the model after dimensionality reduction. Figure 9 shows the point cloud reconstruction results for the three encoding dimensions. The comparison analysis indicates that as the encoding dimension decreases, the geometric similarity between the reconstructed point cloud and the original point cloud gradually decreases. When the encoding dimension is 1024 × 3, the model can reproduce the geometric details of the original point cloud relatively well; however, when the dimension is reduced to 256 × 3, the completeness and accuracy of the reconstructed point cloud significantly decrease.

Table 1. Chamfer distance of reconstructed point clouds under different encoding dimensions.

Encoding Dimensions	Chamfer Distance
1024 × 3	0.18
512 × 3	0.23
256 × 3	1.41

lists the chamfering distances corresponding to different coding dimensions. The results show that when the coding dimension is 1024 × 3, the chamfering distance is the smallest (0.18), indicating that the reconstruction performance is the best. However, when the dimension is reduced to 256 × 3, the chamfer distance increases to 1.41, resulting in a significant increase in reconstruction error. Considering the reconstruction accuracy and the training efficiency of the subsequent wind noise prediction model, 1024 × 3 and 512 × 3 are selected as the point cloud feature vector dimensions of this study. In the follow-up study, the prediction model is constructed based on these two feature vector dimensions, and the training time of the two prediction models is compared. The input dimension with the shortest training time is selected as the output dimension of AE.

4.3. Model Training Effects Under Different AE Output Dimensions

In order to select the output dimension of AE, this paper first builds two prediction models with different input dimensions based on ResNet. Except for different input dimensions, the other hyperparameters of the two models are consistent. After training with the same training set, the error performance of the training set and the time required for training are shown in

Table 2. The performance of different output dimensions of AE in the same training set.

AE Output Dimension	MAPE	MSE	Training Duration
1024 × 3	14.37%	63.25	5 min
512 × 3	12.43%	48.51	3 min

.

By comparing the training time and error performance of models with different input dimensions, it can be seen that although the 1024 × 3 dimension is closer to the original model in geometry, too many input parameters greatly increase the training time of the model. At the same time, too many input parameters have higher requirements for the training set, and the training effect under the same size of the training set will be worse than the model after input compression. Therefore, 512 × 3 is selected as the output dimension of AE.

5. In-Vehicle Driver Noise Prediction and Verification Based on GWO-ResNet

5.1. Establishment of the GWO-ResNet Model

In order to reduce the high costs and resource consumption associated with wind tunnel experiments and real-vehicle testing, a method for predicting interior driver noise based on coded point cloud input has been developed. This method uses compressed three-dimensional point clouds of the vehicle as input and combines them with 1/3 octave noise data measured in a wind tunnel for the driver’s left ear to establish a prediction model for assessing wind noise inside the vehicle.

The input features for the prediction model are generated by the FoldingNet encoder, which reduces the dimensionality of the original point cloud and ultimately outputs a feature set containing 512 three-dimensional points. The model output is the 1/3 octave noise level measured at the position of the driver’s right ear at a wind speed of 120 km/h. This output has completed spectral and sound pressure level information, which can comprehensively reflect the aerodynamic noise characteristics inside the vehicle. The entire dataset consists of vehicle point cloud models and corresponding wind noise test data. The main structure of the model is based on a ResNet and combines the GWO to optimize key hyperparameters, thereby minimizing prediction errors and improving model stability.

Figure 10 shows the structural framework of the GWO-ResNet model, where the learning rate and batch size are set as optimization targets. GWO is used to identify the optimal hyperparameter combination within a predefined search space: the learning rate search range is set to 0.001–0.01, and the batch size range is set to 16–128. The parameters related to ResNet and GWO are set as shown in

Table 3. Parameter settings within the model.

Algorithm Type	Parameter Name	Parameter Value
ResNet	Input dimension	512 × 3
	Maximum number of iterations	100
	Number of ResNet layers	50
	Learning rate optimization range	[0.001, 0.01]
	Batch size optimization range	[16–128]
GWO	Maximum number of iterations	13
	Population size	5

.

The data used for model training includes wind tunnel test data described in Section 3 and additional data provided by an automobile group, totaling 300 samples. To improve training accuracy, all input features were normalized before modeling to reduce the impact of different dimensions on the learning process. The dataset was divided into a training set and a test set in an 80:20 ratio, containing 240 and 60 samples, respectively. Each sample consists of 512 three-dimensional point features and their corresponding right ear noise outputs of the main driver.

During training, the GWO algorithm dynamically adjusts network hyperparameters by simulating the hierarchical structure and cooperative hunting mechanism of gray wolves, ultimately obtaining the optimal combination: a learning rate of 0.009 and a batch size of 64. Modeling and simulation were both performed on the Pycharm 2024.1 platform, with a hardware configuration of an AMD Ryzen 7 7700 processor and 32 GB of memory. To evaluate model performance, the Mean Absolute Percentage Error (MAPE) and Mean Square Error (MSE) were used as the primary evaluation metrics.

5.2. GWO-ResNet Model Predictive Analysis

After optimizing the hyperparameters of the ResNet network using the GWO optimization algorithm, the GWO-ResNet model was used to predict the interior noise levels in the driver’s seat area of passenger vehicles. Figure 11 shows a comparison of the model prediction results and actual measurement data for a typical sedan sample and an SUV sample. The prediction results indicate that the GWO-ResNet model can well fit the trend of the actual measurement data, with the prediction curve highly consistent with the actual measurement curve. To further quantify the prediction accuracy of the model on the test set, a random sample was selected, and the MAPE and MSE results are as follows: The MAPE for the sedan sample is 9.72%, and the MSE is 20.96; for SUV samples, the MAPE is 9.88%, and the MSE is 19.69. These results indicate that the GWO-ResNet model not only possesses strong nonlinear fitting capabilities but also exhibits good generalization performance and robustness.

To further validate the reliability of the model in real-world scenarios, an independent validation set was constructed to test the model’s generalization ability. Figure 12 shows the comparison between the prediction results of the GWO-ResNet model and the actual measurement values on the validation set. The calculation results show that the MAPE of the GWO-ResNet model on the sedan and SUV validation samples are 10.14% and 10.15%, respectively, and the MSEs are 23.97 and 29.15, respectively, further proving the robustness and practicality of the model under different sample conditions.

5.3. Comparison with Other Models

To further validate the effectiveness of the GWO-ResNet model, a comparative analysis was conducted with three typical neural network models: the traditional ResNet model, the GWO-LSTM model, and the unoptimized LSTM model. All models were trained using the same training set and evaluated on the same test set for prediction accuracy. The hyperparameters of the ResNet and LSTM models were optimized via grid search, while the key parameters of the GWO-LSTM model were obtained through GWO algorithm optimization. Figure 13 shows the prediction comparison results of the four models on the sedan and SUV test sets, and

Table 4. MAPE and MSE metrics of the four models on the test set.

Model	Sedan		SUV
	MAPE	MSE	MAPE	MSE
GWO-ResNet	9.72%	20.96	9.88%	19.69
ResNet	12.43%	48.51	12.76%	64.61
GWO-LSTM	11.43%	31.90	11.56%	45.51
LSTM	14.96%	67.41	15.10%	57.36

lists the MAPE and MSE evaluation results of each model on randomly sampled data.

The results show that the GWO-ResNet model, which uses the GWO optimization algorithm for hyperparameter optimization, significantly outperforms the unoptimized ResNet model, the GWO-LSTM model, and the traditional LSTM model in terms of prediction accuracy on the test set. This advantage is particularly evident in key evaluation metrics such as the MAPE and MSE. Specifically, compared with the unoptimized ResNet model, the GWO-ResNet model reduced the MAPE of sedan and SUV samples by 2.71 and 2.88 percentage points, respectively, and reduced the MSE by 27.55 and 44.92, respectively. This significant performance improvement fully verifies the effectiveness and superiority of the GWO algorithm in the parameter optimization process of deep neural networks.

Compared with traditional grid search and other parameter optimization methods, the GWO algorithm simulates the collaborative search mechanism of a gray wolf population, demonstrating stronger global optimization capabilities and convergence efficiency. It can effectively avoid the interference of local optima in complex high-dimensional parameter spaces, thereby enhancing the model’s ability to model the complex nonlinear relationship between in-vehicle wind noise and point cloud data, and ultimately significantly improving prediction accuracy. After comparing and analyzing the prediction results of different neural network models, it can be found that the GWO optimization strategy not only has a good optimization effect on the ResNet structure but also shows significant performance improvement in the LSTM network. The GWO-LSTM model outperforms the traditional LSTM model in terms of MAPE and MSE on the test set, further validating the broad adaptability and robustness of GWO as a general-purpose efficient optimizer. Based on the performance results of all comparison models, the GWO-ResNet model, which integrates the GWO optimization algorithm and the ResNet structure, demonstrates the best performance in terms of prediction accuracy, stability, and generalization ability, and has good engineering application potential.

6. Conclusions

This paper addresses the problem of predicting wind noise at the driver’s seat during passenger car driving. We propose an efficient prediction method based on deep learning. The method first uses an auto-encoder to encode and compress the three-dimensional point cloud data of the car’s exterior, effectively reducing the data dimension while retaining key geometric features. Then, combining wind tunnel test data collected at the driver’s left ear, we construct an end-to-end prediction model with point cloud encoding as input and interior noise level as output. To enhance the model’s ability to model complex nonlinear mapping relationships, the key hyperparameters of ResNet are further optimized using GWO, resulting in the GWO-ResNet prediction model. The prediction results of the GWO-ResNet model on the test set show high consistency with the actual measurement data, demonstrating the effectiveness of the proposed method in modeling wind noise inside vehicles. Comparison results with traditional ResNet, GWO-LSTM, and LSTM models show that the GWO-ResNet model achieves MAPE values of 9.72% and 9.88%, and MSE values of 20.96 and 19.69, respectively, on passenger car and SUV test samples, demonstrating significantly superior prediction performance compared to other models. Further testing on an independent validation set also demonstrated good generalization ability and stability, with MAPE of 10.14% and 10.15%, and MSE of 23.97 and 29.15% for sedan and SUV samples, respectively. Since the test data only contains a single working condition, it is impossible to simulate the complex working conditions during actual driving. Therefore, the prediction results of the model only represent the wind noise under the test conditions, but this method can still effectively improve the wind noise performance evaluation speed in the early stage of automobile styling design. In the future, more complex test conditions will be introduced to simulate the airflow state during actual driving, and the model prediction results will be improved according to the test data.