Transformer-based multi-modal porous asphalt mixture noise prediction method and system

Abstract

The application discloses a multi-modal porous asphalt mixture noise prediction method and system based on a Transformer, and the method comprises the following steps: S1, collecting a section image of the porous asphalt mixture, performing image processing to extract a void feature of the porous asphalt mixture and calculate a depth feature, and simultaneously acquiring physical parameters of the porous asphalt mixture; S2, performing weighted fusion processing on the void feature, the depth feature and the physical parameters to generate a unified feature vector; and S3, inputting the unified feature vector into a Transformer model, processing the relationship among the void feature, the depth feature and the physical parameters by using a self-attention mechanism of the Transformer model, and respectively outputting prediction values of sound absorption coefficients of corresponding frequency bands through multiple mutually independent output ends of the Transformer model. The application can realize deep fusion of multi-source features and intelligent prediction of full-band sound absorption performance, and provides a scientific basis for noise control and material optimization design of the porous asphalt pavement.

Description

Technical Field

[0001] This invention belongs to the field of road material acoustic performance prediction technology, specifically, it relates to a method and system for predicting noise in porous asphalt mixtures based on Transformer multimodal technology. Background Technology

[0002] Porous asphalt mixtures, due to their abundant interconnected pores that effectively dissipate sound wave energy, have been widely used in low-noise pavements to reduce traffic noise pollution. Currently, the prediction of the sound absorption coefficient—the core evaluation indicator of the noise reduction performance of this type of material—mainly relies on direct measurement using the standing wave tube method in the laboratory. However, this traditional method has inherent drawbacks: the specimen preparation and testing process is costly and time-consuming, severely restricting the efficiency of rapid screening and optimization of new material proportions. To overcome the limitations of experimental methods, several predictive models have been developed in existing technologies, mainly including empirical formulas, traditional machine learning models (such as BP neural networks), and numerical simulation methods (such as the finite element method). However, all of these methods have significant shortcomings:

[0003] (1) Limited prediction accuracy: Empirical formulas and simple neural networks are difficult to capture the deep relationship between the complex multi-scale void structure of materials and nonlinear acoustic response. The model determination coefficient R² is usually low, and the prediction reliability is insufficient.

[0004] (2) Single feature representation: Existing models mostly use only limited physical parameters or simple two-dimensional image features, failing to fully consider the key influence of the geometry of the cross-sectional gap and its gradient change in the depth direction (three-dimensional connectivity) on the absorption of sound waves, especially low-frequency sound waves, resulting in weak model generalization ability.

[0005] (3) Poor interpretability and guidance: Existing methods effectively quantify the contribution of different features to noise reduction performance, making it difficult to derive targeted material ratio optimization schemes in reverse, thus limiting their practical engineering value.

[0006] Furthermore, while some patents involve using deep learning for road surface condition detection, none address the prediction of the intrinsic acoustic properties of materials. Finite element simulation-based methods are computationally complex and fail to achieve effective fusion of multimodal data. Summary of the Invention

[0007] The primary objective of this invention is to overcome the shortcomings and deficiencies of existing technologies and provide a Transformer-based multimodal method for predicting the noise of porous asphalt mixtures. This invention can achieve deep fusion of multi-source features and intelligent prediction of sound absorption performance across the entire frequency band, providing a scientific basis for noise control and material optimization design of porous asphalt pavements.

[0008] The second objective of this invention is to provide a Transformer-based multimodal porous asphalt mixture noise prediction system.

[0009] The objective of this invention is achieved through the following technical solution: a multimodal, porous asphalt mixture noise prediction method based on Transformer, comprising the following steps:

[0010] S1. Feature Extraction: Collect cross-sectional images of porous asphalt mixtures, perform image processing on the cross-sectional images to extract the pore features of the porous asphalt mixtures and calculate the depth features, and at the same time obtain the physical parameters of the porous asphalt mixtures.

[0011] S2. Multimodal feature fusion: The gap features, depth features and physical parameters are weighted and fused to generate a unified feature vector;

[0012] S3. Frequency band prediction: The unified feature vector is input into the Transformer model. The self-attention mechanism of the Transformer model is used to process the relationship between the gap features, depth features and physical parameters. The predicted values ​​of the sound absorption coefficients for the corresponding frequency bands are output through multiple independent output terminals of the Transformer model. Each output terminal corresponds to a different frequency band.

[0013] Preferably, in step S1, the void characteristics include cross-sectional void ratio, average effective area of ​​a single void, and average aspect ratio; the depth characteristics include connectivity parameters of voids in adjacent cross-sections and void ratio gradient along the depth direction; and the physical parameters include the maximum nominal particle size of the aggregate and the asphalt-aggregate ratio.

[0014] Preferably, in step S1, the image processing includes: converting the acquired cross-sectional image to grayscale and denoising it, extracting the gap region based on the threshold segmentation method, and performing morphological optimization on the gap region.

[0015] Preferably, step S2 specifically includes:

[0016] S21. Normalization process: Normalize the void features and depth features respectively, and standardize the physical parameters;

[0017] S22. Feature stitching: The normalized gap features, depth features and physical parameters are stitched together.

[0018] S23. The spliced ​​gap features, depth features and physical parameters are weighted and fused according to the preset feature importance weights to generate the unified feature vector.

[0019] Preferably, the preset feature importance weights are determined using a random forest or gradient decision tree model.

[0020] Preferably, in step S3, the Transformer model adopts a TransformerEncoder structure, which includes a multi-head self-attention mechanism and a feedforward neural network. Both the multi-head self-attention mechanism and the feedforward neural network include residual connections and layer normalization structures.

[0021] Preferably, step S3 specifically includes:

[0022] S31. Feature Sequence Input: Organize the unified feature vector into a feature sequence and input it into the Transformer model;

[0023] S32. Feature Encoding: Using the multi-head self-attention mechanism in the Transformer model, the dependency relationship between each feature in the feature sequence is calculated, and a high-dimensional feature representation is obtained by nonlinear transformation through a feedforward neural network.

[0024] S33. Frequency band output mapping: The high-dimensional feature representation is mapped to three independent output terminals of the Transformer model, including low-frequency output terminal, mid-frequency output terminal and high-frequency output terminal, respectively, to predict the sound absorption coefficient of the corresponding frequency band.

[0025] Preferably, after step S3, the following is also included:

[0026] Step S4, Model Validation and Optimization: Calculate the error between the predicted value and the measured value to optimize the parameters of the Transformer model and obtain the Transformer model with the desired prediction accuracy.

[0027] A Transformer-based multimodal porous asphalt mixture noise prediction system includes:

[0028] The feature extraction module is used to acquire cross-sectional images of porous asphalt mixtures, perform image processing on the cross-sectional images to extract the pore features of the porous asphalt mixtures and calculate the depth features, and at the same time obtain the physical parameters of the porous asphalt mixtures.

[0029] The multimodal feature fusion module is used to perform weighted fusion processing on the gap features, depth features and physical parameters to generate a unified feature vector;

[0030] The frequency band prediction module is used to input the unified feature vector into the Transformer model, use the self-attention mechanism of the Transformer model to process the relationship between gap features, depth features and physical parameters, and output the predicted value of the sound absorption coefficient of the corresponding frequency band through multiple independent output terminals of the Transformer model, wherein each output terminal corresponds to a different frequency band.

[0031] Preferably, the prediction system further includes a model validation and optimization module, used to calculate the error between the predicted value and the measured value, so as to optimize the parameters of the Transformer model and obtain a Transformer model with the desired prediction accuracy.

[0032] The present invention has the following advantages and effects compared with the prior art:

[0033] (1) The proposed Transformer-based multimodal noise prediction method for porous asphalt mixtures constructs a multimodal feature system with clear physical meaning and weights and fuses void features, depth features, and material physical parameters to achieve coordinated characterization of the macroscopic performance and microstructure information of porous asphalt mixtures within a unified feature space. Furthermore, utilizing the self-attention mechanism of the Transformer model, adaptive modeling of the correlation strength between different features is performed, enabling the model to distinguish different types of features and their relative contributions to sound absorption performance prediction, thereby improving the interpretability of the model's prediction results. Simultaneously, the unified feature vector input format and frequency-band output structure design allow the model to flexibly introduce new structural features or physical parameters without changing the overall architecture, and to expand frequency band divisions or output indicators as needed, exhibiting good scalability and engineering adaptability. This method improves prediction accuracy while maintaining the universality of the model structure, providing a stable and scalable intelligent technical path for noise performance modeling and material optimization design of porous asphalt pavements.

[0034] (2) Experimental results show that the coefficient of determination R of the Transformer model of the present invention is high across the entire frequency band. 2 The mean square error (MSE) is 0.939, the mean square error (MSE) is 0.0019, and the mean relative error (RE) is 17.1%. Compared with existing technologies, the average prediction error is significantly reduced, overcoming the problem that traditional empirical models or single machine learning models are unable to capture the complex nonlinear relationship between the gap structure and the acoustic response, resulting in limited prediction accuracy. At the same time, compared with traditional laboratory standing wave tube testing methods, the method of this invention uses model inference to replace a large amount of repetitive standing wave tube measurement work, and only performs standardized testing and verification on representative samples, reducing the overall experimental workload and testing cost, and shortening the evaluation cycle.

[0035] (3) Based on the modeling approach of multimodal feature fusion and frequency-band Transformer, this invention sets up three independent output terminals to address the differences in sound absorption mechanisms at different frequency bands. These terminals jointly predict the sound absorption coefficients of the low-frequency, mid-frequency, and high-frequency bands, thereby directly obtaining the frequency-band sound absorption curves for engineering design. Compared with traditional methods that rely solely on empirical formulas or single regression models, the multi-output prediction framework of this invention can simultaneously capture the interdependencies between different frequency bands, avoiding the information fragmentation problem caused by independent modeling of each frequency band, thus improving prediction consistency and stability.

[0036] (4) This invention encodes cross-sectional void morphology (void features, two-dimensional or three-dimensional images), depth distribution features (depth features) and traditional physical parameters to form a multi-scale, multi-modal redundant representation. This fusion strategy enhances the model's ability to express the microstructure-macro performance mapping, thereby improving the prediction accuracy and generalization ability of the sound absorption performance of complex porous systems. Attached Figure Description

[0037] Figure 1 This is a schematic flowchart of the Transformer-based multimodal porous asphalt mixture noise prediction method of the present invention.

[0038] Figure 2 This is a schematic diagram of the image processing process in an embodiment of the present invention, wherein (a) is the original image, (b) is the grayscale conversion image, (c) is the nonlocal filtering denoising image, (d) is the gap extraction image, (e) is the morphological denoising processing image, and (f) is the median filtering denoising image.

[0039] Figure 3 This is a schematic diagram of the multimodal feature fusion process in an embodiment of the present invention.

[0040] Figure 4 This is a network topology diagram predicted by the Transformer model in an embodiment of the present invention.

[0041] Figure 5 This is a comparison chart of predicted and measured values ​​in the model verification and optimization steps of this invention.

[0042] Figure 6 This is a graph showing the relationship between frequency and sound absorption coefficient (sound absorption coefficient curves at different frequency bands) in an embodiment of the present invention. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0044] Example 1

[0045] like Figure 1The diagram shows a flowchart of a Transformer-based multimodal porous asphalt mixture noise prediction method, including the following steps:

[0046] S1. Feature Extraction: Acquire cross-sectional images of porous asphalt mixtures, perform image processing on the cross-sectional images to extract the void features and calculate the depth features of the porous asphalt mixtures, and simultaneously obtain the physical parameters of the porous asphalt mixtures; the void features include cross-sectional porosity, average effective area of ​​a single void, and average aspect ratio; the depth features include connectivity parameters of voids in adjacent cross-sections and the porosity gradient along the depth direction; the physical parameters include the maximum nominal particle size of the aggregates and the asphalt-aggregate ratio.

[0047] S2. Multimodal feature fusion: The gap features, depth features and physical parameters are weighted and fused to generate a unified feature vector;

[0048] S3. Frequency band prediction: The unified feature vector is input into the Transformer model. The self-attention mechanism of the Transformer model is used to process the relationship between the gap features, depth features and physical parameters. The predicted values ​​of the sound absorption coefficients for the corresponding frequency bands are output through multiple independent output terminals of the Transformer model. Each output terminal corresponds to a different frequency band.

[0049] Specifically, the proposed Transformer-based multimodal noise prediction method for porous asphalt mixtures achieves multi-scale, multimodal information fusion by weightedly fusing cross-sectional void morphology (geometric information from two-dimensional high-resolution images or three-dimensional imaging), depth distribution characteristics (such as connectivity characterization parameters and porosity gradients), and conventional physical quantities (such as the maximum nominal particle size and asphalt-aggregate ratio of aggregates). This method simultaneously reflects the microscopic geometric features and macroscopic physical properties of porous asphalt mixtures, thus overcoming the shortcomings of traditional single measurement methods that cannot comprehensively describe the multi-scale structure of materials. Furthermore, by employing the self-attention mechanism of the Transformer model, weights are adaptively allocated along the feature dimensions, capturing the complex nonlinear coupling relationship between cross-sectional void morphology and physical parameters in response to acoustic responses at different frequency bands. Simultaneously, by setting three independent output terminals for low, mid, and high frequencies, joint prediction of the sound absorption coefficients across frequency bands is achieved, effectively avoiding information fragmentation caused by independent modeling for each frequency, thereby improving the consistency and accuracy of the prediction. This method builds a bridge connecting data-driven prediction and physical mechanism-based analysis, and can provide support for subsequent research on acoustic model mechanisms.

[0050] The specific steps of step S1 include:

[0051] S11. Image Acquisition: Acquire cross-sectional images of multiple sets of porous asphalt mixture specimens of different specifications. To ensure image quality and accuracy, a high-resolution digital camera will be used for shooting. During shooting, ensure that the cross-section of each specimen is clearly visible and free from external interference. The specimens can be placed on a standardized background board to ensure image consistency.

[0052] S12. Perform image preprocessing using MATLAB and extract void feature parameters from cross-sections at different depths:

[0053] Regarding image preprocessing: Images may contain noise and need to be processed using methods such as denoising (including median filtering or Gaussian filtering) and edge detection (such as the Canny edge detection algorithm).

[0054] Regarding gap feature extraction: Image segmentation methods (such as thresholding or edge detection) are used to separate gap regions from other parts of the image. Then, morphological processing (such as dilation, erosion, etc.) is used to extract the shape and size of each gap.

[0055] like Figure 2 The diagram shows the image processing process, where (a) is the original image, (b) is the grayscale conversion image, (c) is the nonlocal filtering denoising image, (d) is the gap extraction image, (e) is the morphological denoising processing image, and (f) is the median filtering denoising image.

[0056] In this embodiment, in order to extract the void region of the cross-section of the porous asphalt mixture specimen, an image recognition method based on grayscale threshold segmentation, nonlocal mean denoising, morphological processing, and median filtering denoising was mainly adopted. The specific steps are as follows:

[0057] (1) Grayscale conversion: The acquired color image of the porous asphalt mixture cross-section is dark due to the voids, so it needs to be converted into a single-channel grayscale image. The data processing is simplified by extracting the pixel brightness, while the brightness difference between the voids and other areas is preserved, laying the foundation for subsequent segmentation and recognition. Then, a non-local mean filter with a smoothing parameter of 30 is used to denoise the grayscale image, which effectively removes background texture noise, smooths the image and preserves edge details.

[0058] (2) Gray-scale value range analysis and threshold setting: In gray-scale images, gaps have low gray-scale values, while aggregate and asphalt areas have higher gray-scale values. By statistically analyzing the minimum and maximum gray-scale values ​​of all pixels, a gap extraction threshold is dynamically set: a certain percentage above the lowest gray-scale value is selected to accurately extract the main gap areas while avoiding excessive background noise. This dynamic threshold setting can adapt to different image brightness variations, effectively improving the applicability of gap extraction.

[0059] (3) Preliminary extraction of void regions: The image is segmented into void and solid regions by setting a grayscale threshold: regions with grayscale values ​​below the threshold are identified as voids (white), and the rest are considered as solid (black). The binary image generated by threshold segmentation clearly separates voids from other parts, providing a clear basis for region division for subsequent analysis.

[0060] (4) Morphological processing and median filtering noise reduction: To address the problem of isolated noise points and irregular boundaries in binarized images, mathematical morphological operations are adopted: opening operations are used to remove isolated small noise points and repair gap boundaries, closing operations are used to fill small holes in the gaps, and circular structural elements are combined to simulate the geometric characteristics of the gaps, effectively improving the purity and boundary regularity of the extraction results.

[0061] Based on morphological processing, a 3×3 window median filtering method is applied to replace the center value with the median of the pixel neighborhood. This method preserves the details of the gap edges, removes residual small noise points, and smooths irregular edges, avoiding excessive blurring of geometric features and providing a high-quality image foundation for subsequent statistical analysis.

[0062] Specifically, the void characteristics include cross-sectional void ratio, average effective area of ​​a single void, and average aspect ratio; the cross-sectional void ratio is calculated by statistically analyzing the proportion of the void area to the total cross-sectional area in each cross-section; the effective area of ​​each void is calculated and its average value is obtained to get the average effective area of ​​a single void; the aspect ratio of each void is calculated and its average value is obtained to get the average aspect ratio of the void.

[0063] S13. Calculate depth features:

[0064] The depth features include connectivity parameters for voids between adjacent cross sections and a porosity gradient along the depth direction. The connectivity parameters for voids between adjacent cross sections characterize the equivalent connectivity of the void structure between adjacent cross sections along the depth direction, and their calculation is based on the spatial overlap and continuity of voids between adjacent cross sections. For any adjacent i-th and i+1-th cross sections, spatial registration is performed on the void regions in the two cross sections to identify void regions that maintain spatial overlap or continuous distribution in adjacent cross sections; the area of ​​the void region in the spatially overlapping portion of the i-th and i+1-th cross sections is denoted as the connected void area A. c,i Let A be the total area of ​​the voids in the i-th cross section. v,i The connectivity characterization parameter C of the gaps between adjacent cross sections i Expressed as: C i =A c,i / A v,iThe porosity gradient is defined as follows: porosity gradient = (|surface porosity - subsurface porosity|) / total depth (% / mm). That is, by calculating the difference between the porosity of each cross section and the porosity of adjacent cross sections, the porosity change rate (porosity gradient) along the depth direction is obtained. This feature helps to describe the deep structural changes of porous asphalt mixtures.

[0065] S14. Obtain physical parameters:

[0066] The physical parameters include the maximum nominal aggregate size and the asphalt-aggregate ratio. The aggregate size is extracted and measured from porous asphalt mixture specimens, where the maximum nominal aggregate size can be obtained through image measurement tools or physical measurement data of the sample; the asphalt-aggregate ratio is obtained by calculating the mass ratio of asphalt to aggregate.

[0067] Specifically, such as Figure 3 The diagram shows a flowchart of the multimodal feature fusion process of the present invention. Step S2 specifically includes:

[0068] S21. Normalize each feature separately: Normalize the gap feature and depth feature separately (e.g., Min-Max normalization), and standardize the physical parameters (e.g., Z-score standardization) to ensure that features from different sources are within their appropriate scales and will not affect the subsequent weight allocation due to excessive differences in numerical range.

[0069] Specifically, the image feature vectors (i.e., gap features and depth features) are processed by the ViT (Vision Transformer) encoder to output a 1024-dimensional vector. This 1024-dimensional output vector means that after the acquired cross-sectional images are processed by the ViT encoder, each image is mapped into a 1024-dimensional high-dimensional feature vector. This is the result of image feature abstraction, and together with the gap features and depth features, it forms the image feature vector. The Min-Max normalization method is used to compress the image feature vector to the [0,1] interval to preserve the relative differences in features. The formula is as follows:

[0070] ,

[0071] Where, x i Here, x represents the original feature values, and min(x) and max(x) represent the minimum and maximum values ​​in the image feature vector, respectively. The normalized value. ;

[0072] For the physical parameter vector (maximum nominal particle size, oil-aggregate ratio, etc.): Z-score standardization is used to make the physical parameters follow a standard normal distribution, enhancing cross-sample comparability. The formula is as follows:

[0073] ,

[0074] Where, x j The original physical parameter value, μ j σ is the mean of this parameter. j Standard deviation, The values ​​are standardized; in this embodiment, the mean is 0 and the standard deviation is 1.

[0075] S22, Feature splicing:

[0076] The normalized void features and depth features, along with the standardized physical parameters, are concatenated to obtain a feature vector that contains all void structure features, depth features, and physical features.

[0077] S23, Weighted Fusion:

[0078] Different importance weights are assigned to different feature categories (including void features, depth features, and physical parameters). The preset feature importance weights are determined through feature importance analysis, specifically using random forest or gradient decision tree models. In this embodiment, a random forest model is used to calculate the Gini importance of each feature to the sound absorption coefficient, and the normalized value is used as the weight. For example, when the void feature category weight is 0.4, the weights of key parameters such as void ratio are allocated according to their importance proportion within the void feature category, based on the random forest model calculation results. When the physical feature category weight is 0.3, the weights of parameters such as oil-stone ratio are allocated according to their importance proportion within the physical feature category. This method ensures consistency and rationality between individual feature weights and feature category weights.

[0079] Based on the preset feature importance weights, the spliced ​​gap features, depth features, and physical parameters are weighted and fused to obtain a comprehensive feature vector.

[0080] S24, Unified Feature Vector:

[0081] Since weighted fusion changes the distribution of different features, it is necessary to perform normalization again (usually Min-Max) to ensure that the numerical range of the final unified feature vector is consistent, which is beneficial for subsequent input into the model.

[0082] This invention forms a multi-scale, multi-modal redundant representation by weightedly fusing cross-sectional void morphology (geometric information from two-dimensional high-resolution images or three-dimensional imaging), depth distribution characteristics (such as connectivity characterization parameters and porosity gradient), and conventional physical quantities (such as the maximum nominal particle size of aggregates and the asphalt-aggregate ratio). This fusion strategy enhances the model's ability to express the microstructure-macro performance mapping, thereby improving the prediction accuracy and generalization ability of the sound absorption performance of complex porous systems.

[0083] Figure 4 The figure shows the network topology diagram for the Transformer model prediction of this invention. The Transformer encoder is responsible for extracting input features, and then predicts the low-frequency, mid-frequency, and high-frequency absorption coefficients through three independent fully connected prediction heads. The three outputs are task prediction layers constructed based on the output features of the Transformer encoder. Step S3 specifically includes:

[0084] S31. Feature Sequence Input: Organize the unified feature vector into a feature sequence and input it into the Transformer model;

[0085] Specifically, the unified feature vector contains information from all input features, and the values ​​are on the same scale, ready for Transformer encoding. This unified feature vector X is used as input and is typically organized into a feature sequence X=[x1,x2,...,x...]. n ], where X∈R d x i Let represent the input vector at the i-th time step (or feature), where n is the length of the feature vector, d is the dimension of each feature vector, and R represents the set of real numbers. d It represents the set of all d-dimensional real vectors, that is, it indicates that each component of the eigenvector x is a real number.

[0086] S32. Feature Encoding: Using the self-attention mechanism in the Transformer model, the dependency relationship between each feature in the feature sequence is calculated, and a high-dimensional feature representation is obtained by performing a nonlinear transformation through a feedforward neural network.

[0087] The Transformer model is a TransformerEncoder structure, which includes a multi-head self-attention mechanism and a feedforward neural network. Both the multi-head self-attention mechanism and the feedforward neural network include residual connections and layer normalization structures.

[0088] Specifically, self-attention is the core of the Transformer, capable of calculating the correlations or dependencies between elements in a sequence. In capturing complex relationships between features through self-attention, each feature x... i The correlation between each input feature and other features will be calculated using the following steps: for each input feature x i The query vector Q is generated by learning the weight matrix. i Key vector K i Sum vector V iQuery vector Q i and bond vector K i The dimension is d k Value vector V i The dimension is d v Q i =W q x i +b q K i =W k x i +b k V i =W v x i +b v Among them, W q W k and W v It is a trainable weight matrix, b q b k and b v It is a bias term.

[0089] The attention weights are primarily calculated by measuring the similarity between the query vector and the key vector in the representation space learned by the Transformer model. This similarity is then scaled and softmaxed to form the attention weights, as shown in the following formula:

[0090] ,

[0091] Where Attention(Q,K,V) represents the output of the attention mechanism, QK T This represents the matrix product of the query vector matrix and the key vector matrix; d k Represents the query vector Q i With key vector K i The common vector dimension; the above formula means that after the softmax operation, the scaled similarity results are normalized to obtain the weighting coefficients, and the value vector matrix V is weighted and summed accordingly to generate the attention output result.

[0092] Multi-Head Attention: Transformer models typically use multi-head self-attention mechanisms to compute multiple attention distributions in parallel, allowing the model to focus on features in different subspaces. In multi-head self-attention, each head computes an independent attention distribution, and then the outputs of all heads are concatenated and a linear transformation is applied to obtain the final output.

[0093] Multhead(Q,K,V)=Concat(head1,L,head h )Wo,

[0094] Where h is the number of heads, W O It is the final linear transformation matrix.

[0095] Feed-Forward Network: Each self-attention layer is typically followed by a feed-forward network for further feature processing. This feed-forward network consists of two linear layers and an activation function, expressed as follows:

[0096] FFN(x)=W2[max(0,xW1+b1)]+b2,

[0097] Where W1 represents the weight matrix of the first layer, which is used to project the input vector x into a higher or lower dimension latent space; W2 represents the weight matrix of the second layer, which is used to project the activated latent representation back to the original dimension; b1 represents the bias vector of the first layer, which, together with W1, determines the translation amount of the linear transformation; b2 represents the bias vector of the second layer, which adjusts the translation of the second layer linear transformation.

[0098] Feedforward neural networks provide a non-linear mapping for each input feature, thereby enhancing the expressive power of the model.

[0099] Layer Normalization and Residual Connection: To prevent the vanishing gradient problem, layer normalization is performed after each self-attention layer and feedforward neural network, and residual connection is used, that is, input plus output: LayerNorm(x+Attention Output). This can speed up training and make the model more stable.

[0100] S33. Frequency band output mapping: The high-dimensional feature representation is mapped to three independent output terminals of the Transformer model, including low-frequency output terminal, mid-frequency output terminal and high-frequency output terminal, respectively, to predict the sound absorption coefficient of the corresponding frequency band.

[0101] Specifically, since the sound absorption mechanisms of sound waves differ across frequency bands, this invention designs three independent output terminals based on these different absorption mechanisms to predict the sound absorption coefficients for the low-frequency, mid-frequency, and high-frequency bands, respectively. Each output terminal receives a high-dimensional feature representation encoded by a Transformer and generates the corresponding frequency band's sound absorption coefficient through a fully connected layer.

[0102] y low =FC low (Transformer output)

[0103] ymid =FC mid (Transformer output)

[0104] y high =FC high (Transformer output)

[0105] Among them, y low y mid and y high Let represent the predicted absorption coefficients for the low, mid, and high frequency bands, respectively. FC represents the fully connected layer, which is a linear mapping function. The fully connected layer has the form: FC(x) = xW + b, where W is the weight matrix and b is the bias term; C low C mid and C high Let C represent the parameter sets (including weight matrices and bias vectors) of the fully connected layers for low-frequency, mid-frequency, and high-frequency bands, respectively. low =(W low ,b low ), used to predict the sound absorption coefficient in the low-frequency range; C mid =(W mid ,b mid ), used to predict the sound absorption coefficient in the mid-frequency range; C high =(W high ,b high ), used to predict the sound absorption coefficient in the high-frequency band.

[0106] This invention, based on multimodal feature fusion and frequency-band Transformer modeling, sets up three independent outputs to jointly predict the absorption coefficients of low-frequency, mid-frequency, and high-frequency bands, taking into account the differences in sound absorption mechanisms across different frequency bands. Compared to traditional methods that rely solely on empirical formulas or single regression models, this invention's multi-output prediction framework can simultaneously capture the interdependencies between different frequency bands, avoiding the information fragmentation problem caused by independently modeling each frequency band, thereby improving prediction consistency and stability.

[0107] Step S3 is followed by:

[0108] Step S4, Model Validation and Optimization: Calculate the error between the predicted value and the measured value to optimize the parameters of the Transformer model and obtain the Transformer model with the desired prediction accuracy.

[0109] Specifically, the error between the predicted and measured (true) sound absorption coefficients is calculated for each frequency band (low, mid, and high) to evaluate the model's accuracy. For example, measured data from laboratory standing wave tubes are used as a control to verify the prediction results. Commonly used evaluation metrics include the coefficient of determination (R²), mean square error (MSE), and mean relative error (RE), which characterize the fitting accuracy and error. The parameters of the Transformer model are then optimized based on the error to ensure the model accurately captures the complex relationships between features, thereby generating precise sound absorption coefficients.

[0110] This invention can plot sound absorption curves for different frequency bands based on the final predicted and accurate sound absorption coefficient output for each frequency band, demonstrating the noise absorption effect of different porous asphalt mixture proportions in each frequency band, thereby proposing the optimal low-noise porous asphalt mixture proportion scheme, which has important engineering application value.

[0111] Specifically, the method of the present invention will be further explained below using specific experimental data.

[0112] First, a TransformerEncoder model is established using cross-sectional images and physical parameters, as follows:

[0113] Step 1: Prepare the specimen and obtain cross-sectional images:

[0114] Porous asphalt mixture specimens with different gradations (PAC-10, PAC-13, and PAC-16, with corresponding maximum nominal particle sizes of 9.5 mm, 13.2 mm, and 16 mm, respectively; PAC refers to porous asphalt mixture type) and different aggregate maximum nominal particle sizes (9.5 mm, 13.2 mm, and 16 mm) were prepared. The porosity was controlled within the range of 17% to 23% to cover the porosity characteristic range of common low-noise porous asphalt pavement structures. The specimens were prepared using the standard Marshall process, the main steps of which are as follows:

[0115] (1) Select aggregate crushed stone, mineral powder and asphalt material of different particle sizes, weigh and mix them according to the design gradation ratio;

[0116] (2) Heat and stir at 180±5 ℃ to ensure that the asphalt uniformly coats the aggregate;

[0117] (3) Use a Marshall compactor to compact the material on both sides at 150±3℃, compacting each side 75 times to obtain a standard cylindrical specimen (diameter 100 mm, height 63.5 mm).

[0118] (4) The porosity P was determined by weighing and vacuum saturation. By adjusting the oilstone ratio and aggregate gradation, the porosity of the prepared specimens was distributed between 17% and 23%.

[0119] (5) Using digital image technology, we conducted a study on the void characteristics of 55 samples of porous asphalt mixtures in five groups: PAC-10 (target void ratio of 20%), PAC-13 (target void ratios of 17%, 20%, and 23%), and PAC-16 (target void ratio of 20%).

[0120] This embodiment uses PAC-13 (target porosity of 20%) as an example to detail the process of extracting the porosity structure characteristics of porous asphalt mixtures: First, PAC-13 porous asphalt mixture specimens with a target porosity of 20% are prepared in the laboratory. The specimens have a diameter of 100 mm and a thickness of 50 mm. Next, the specimens are cut at 8 mm intervals, and the cross-sections at both ends are removed. Finally, digital cameras are used to photograph the specimens to obtain cross-sectional images at different depths.

[0121] Then, MATLAB is used for image preprocessing, including grayscale conversion, nonlocal mean filtering for noise reduction, dynamic thresholding, morphological operations, and median filtering, to obtain a clear binary image of the gap region.

[0122] Extract the cross-sectional porosity, average effective area of ​​a single void, and average aspect ratio; calculate the connectivity parameters and porosity gradient of voids in adjacent cross-sections; simultaneously obtain the physical parameters: maximum nominal particle size D of the aggregate. max The optimal oilstone ratio is 13.2 mm. b It is 4.9%.

[0123] Step 2: Normalize and fuse the above features: Gap features and depth features are normalized using Min-Max, and physical parameters are normalized using Z-score. Then, weighted fusion is performed according to feature importance weights. Comparative verification shows that, considering the importance of gap features, depth features, and physical features, a weighted fusion strategy based on feature importance allocation is adopted. Specifically, the weight of gap features is set to 0.4, the weight of depth features to 0.3, and the weight of physical features to 0.3. Under this weight configuration, the model exhibits superior prediction accuracy in low, medium, and high frequency bands, achieving optimal overall performance. Finally, a unified 12-dimensional feature vector is generated.

[0124] Step 3: Input the unified 12-dimensional feature vector generated in Step 2 into the TransformerEncoder model. The model is set up with a 4-layer encoder and 8 multi-head attention mechanisms. The complex relationship between gap features, depth features and physical parameters is captured by the self-attention mechanism. After processing by feedforward neural network, residual connection and layer normalization, a high-dimensional feature representation is obtained.

[0125] Step 4: Map the encoded features to three independent output terminals: low frequency band (200~500Hz), mid frequency band (500~1000Hz), and high frequency band (1000~2000Hz), and predict the sound absorption coefficient of the corresponding frequency band.

[0126] Step 5: Measure the sound absorption coefficient y using a standing wave tube. i As control data, the model fitting accuracy and prediction error were calculated. A joint spectral model was used, predicting the entire spectrum as a whole, and outputting the coefficient of determination R0. 2 Mean square error (MSE), mean relative error (RE).

[0127] ,

[0128] ,

[0129] ,

[0130] Among them, y i , represents the measured and predicted values ​​of the sound absorption coefficient of porous asphalt mixture specimens, respectively, where n represents the number of asphalt mixture specimens. The measured value of the sound absorption coefficient y i The average.

[0131] like Figure 5 The chart shows a comparison between the predicted and measured values ​​in this verification and optimization step. The sample points in both the training and test sets are generally distributed around the y=x reference line, without any significant systematic deviation, indicating that the model has good fit to the measured data and possesses a certain generalization ability. The coefficient of determination R0 is... 2 The model's fitting accuracy is indicated by the mean square error (MSE), which measures the average deviation between predicted and measured values, i.e., the average degree to which the model's predicted values ​​deviate from the measured values. The coefficient of determination (R²) is also mentioned. 2The closer the MSE value is to 1, the smaller it is, indicating that the model's predicted value is closer to the measured value. During model training and optimization, measured sample data of porous asphalt mixtures obtained from laboratory experiments are used as the dataset, and randomly divided into training and test sets at a ratio of 70% to 30%. This invention uses prediction accuracy as the main criterion for model optimization, and adjusts and optimizes model parameters through comprehensive analysis of evaluation indicators such as the coefficient of determination (R²), mean square error (MSE), and mean relative error (RE). Specifically, during model training, the changes in prediction performance on the training and test sets are continuously monitored. When, with training rounds or parameter adjustments, the model's R² on the test set no longer significantly increases, and the MSE and RE show a stable downward trend, the model is considered to have reached the desired prediction accuracy. Under these conditions, the model optimization process is stopped, and the current model is determined as the final sound absorption coefficient prediction model. The resulting model, while ensuring high prediction accuracy, avoids overfitting and maintains stable prediction performance on unknown samples, thus achieving accurate prediction of the sound absorption coefficient of porous asphalt mixtures. After final calculation, the coefficient of determination R of the model of this invention is [value missing] across the entire frequency band. 2 =0.939, mean square error MSE=0.0019, mean relative error RE=17.1%.

[0132] It can be seen that the method of the present invention can well reflect the overall trend of the sound absorption coefficient of porous asphalt mixture with frequency, and the predicted results maintain a high degree of consistency with the measured results. Meanwhile, the prediction error is within the allowable error range for engineering testing of the sound absorption performance of porous asphalt mixture, and can meet the requirements of engineering applications for the accuracy of sound absorption performance prediction.

[0133] It should be noted that, due to the non-uniformity of the internal pore structure of porous asphalt mixtures, and the differences in the propagation and dissipation mechanisms of sound waves at different frequency bands, certain prediction deviations may still occur at local frequency points. However, this does not affect the effective characterization of the overall change law of sound absorption performance by the method of this invention. Therefore, the method of this invention, while ensuring prediction stability, also takes into account a reasonable error level under actual engineering conditions, and has good engineering applicability and practical value.

[0134] Step 6: Output the predicted frequency band sound absorption curves, and propose the optimal low-noise porous asphalt mixture mix ratio scheme based on the model results, which can directly guide engineering applications.

[0135] Specifically, such as Figure 6The figure shows the relationship between frequency and sound absorption coefficient (sound absorption coefficient curves at different frequency bands) in an embodiment of the present invention. As can be seen from the figure, the overall trend of the low-frequency, mid-frequency, and high-frequency sound absorption coefficient curves output by the model is consistent with the actual acoustic response law of porous asphalt mixtures. The sound absorption coefficient exhibits obvious frequency-segmented characteristics as it changes with frequency. The peak positions and variation amplitudes of the sound absorption coefficients of different materials show significant differences, reflecting the varying influences of the internal pore structure of the material and the sound wave interaction mechanism within different frequency ranges. This invention, based on the sound absorption coefficients of each frequency band, outputs frequency-segmented sound absorption coefficient curves. This can be further combined with parameter sensitivity analysis and optimization algorithms to achieve parameter inversion and proportion optimization under the target sound absorption performance, providing a scientific guiding basis for the design of low-noise porous asphalt mixtures.

[0136] Example 2

[0137] A Transformer-based multimodal porous asphalt mixture noise prediction system includes:

[0138] The feature extraction module is used to acquire cross-sectional images of porous asphalt mixtures, perform image processing on the cross-sectional images to extract the pore features of the porous asphalt mixtures and calculate the depth features, and at the same time obtain the physical parameters of the porous asphalt mixtures.

[0139] The multimodal feature fusion module is used to perform weighted fusion processing on the gap features, depth features and physical parameters to generate a unified feature vector;

[0140] The frequency band prediction module is used to input the unified feature vector into the Transformer model, use the self-attention mechanism of the Transformer model to process the relationship between gap features, depth features and physical parameters, and output the predicted value of the sound absorption coefficient of the corresponding frequency band through multiple independent output terminals of the Transformer model, wherein each output terminal corresponds to a different frequency band.

[0141] The model validation and optimization module is used to calculate the error between the predicted value and the measured value in order to optimize the parameters of the Transformer model and obtain the Transformer model with the desired prediction accuracy.

[0142] Specifically, this invention acquires image features and performs multimodal feature fusion. Based on the Transformer model, it utilizes a self-attention mechanism to handle the nonlinear relationships between features and maps them to three output terminals: low-frequency, mid-frequency, and high-frequency bands. Ultimately, it predicts the sound absorption coefficient for each frequency band, addressing the problems of high experimental costs and low prediction accuracy in existing technologies. Based on the output sound absorption coefficient values, sound absorption coefficient curves for each frequency band can be plotted. These curves can be directly used to guide the mix design of low-noise porous asphalt mixtures, enabling intelligent material design and maximizing noise reduction benefits. This invention does not replace physical testing but rather establishes a quantitative mapping relationship between the void structure characteristics and frequency band sound absorption performance by jointly modeling the cross-sectional void structure characteristics and physical parameters of porous asphalt mixtures. In the early stages of material mix design, by preparing a small number of representative specimens and obtaining their void structure characteristics, the sound absorption performance of different material structure schemes in each frequency band can be quickly predicted, thereby clarifying key structural targets such as void size distribution and connectivity characteristics. Based on the material structure objectives, engineers can adjust the mix proportion parameters such as aggregate gradation, asphalt content, and maximum nominal particle size in reverse order to achieve optimized design of the noise reduction performance of porous asphalt mixtures.

[0143] The above embodiments are preferred embodiments of the present invention and are not intended to limit the present invention. Any other changes or equivalent substitutions made without departing from the technical solution of the present invention are included within the protection scope of the present invention.

Claims

1. A multimodal noise prediction method for porous asphalt mixtures based on Transformer, characterized in that, Including the following steps: S1. Feature Extraction: Collect cross-sectional images of porous asphalt mixtures, perform image processing on the cross-sectional images to extract the pore features of the porous asphalt mixtures and calculate the depth features, and at the same time obtain the physical parameters of the porous asphalt mixtures. S2. Multimodal feature fusion: The gap features, depth features and physical parameters are weighted and fused to generate a unified feature vector; S3. Frequency band prediction: The unified feature vector is input into the Transformer model. The self-attention mechanism of the Transformer model is used to process the relationship between the gap features, depth features and physical parameters. The predicted values ​​of the sound absorption coefficients for the corresponding frequency bands are output through multiple independent output terminals of the Transformer model. Each output terminal corresponds to a different frequency band.

2. The method for predicting noise in porous asphalt mixtures based on Transformer multimodal operation according to claim 1, characterized in that, In step S1, the void characteristics include cross-sectional void ratio, average effective area of ​​a single void, and average aspect ratio; the depth characteristics include connectivity parameters of voids in adjacent cross-sections and void ratio gradient along the depth direction; the physical parameters include the maximum nominal particle size of the aggregate and the asphalt-aggregate ratio.

3. The method for predicting noise in porous asphalt mixtures based on Transformer multimodal operation according to claim 1, characterized in that, In step S1, the image processing includes: converting the acquired cross-sectional image to grayscale and denoising it, extracting the gap region based on the threshold segmentation method, and performing morphological optimization on the gap region.

4. The method for predicting noise in porous asphalt mixtures based on Transformer multimodal operation according to claim 1, characterized in that, Step S2 specifically includes: S21. Normalization process: Normalize the void features and depth features respectively, and standardize the physical parameters; S22. Feature stitching: The normalized gap features, depth features and physical parameters are stitched together. S23. The spliced ​​gap features, depth features and physical parameters are weighted and fused according to the preset feature importance weights to generate the unified feature vector.

5. The method for predicting noise in porous asphalt mixtures based on Transformer multimodal operation according to claim 4, characterized in that, The preset feature importance weights are determined using random forest or gradient decision tree models.

6. The method for predicting noise in porous asphalt mixtures based on Transformer multimodal operation according to claim 1, characterized in that, In step S3, the Transformer model adopts a TransformerEncoder structure, which includes a multi-head self-attention mechanism and a feedforward neural network. Both the multi-head self-attention mechanism and the feedforward neural network include residual connections and layer normalization structures.

7. The method for predicting noise in porous asphalt mixtures based on Transformer multimodal operation according to claim 6, characterized in that, Step S3 specifically includes: S31. Feature Sequence Input: Organize the unified feature vector into a feature sequence and input it into the Transformer model; S32. Feature Encoding: Using the multi-head self-attention mechanism in the Transformer model, the dependency relationship between each feature in the feature sequence is calculated, and a high-dimensional feature representation is obtained by nonlinear transformation through a feedforward neural network. S33. Frequency band output mapping: The high-dimensional feature representation is mapped to three independent output terminals of the Transformer model, including low-frequency output terminal, mid-frequency output terminal and high-frequency output terminal, respectively, to predict the sound absorption coefficient of the corresponding frequency band.

8. The method for predicting noise in porous asphalt mixtures based on Transformer multimodal operation according to claim 1, characterized in that, Step S3 is followed by: Step S4, Model Validation and Optimization: Calculate the error between the predicted value and the measured value to optimize the parameters of the Transformer model and obtain the Transformer model with the desired prediction accuracy.

9. A Transformer-based multimodal porous asphalt mixture noise prediction system, characterized in that, include: The feature extraction module is used to acquire cross-sectional images of porous asphalt mixtures, perform image processing on the cross-sectional images to extract the pore features of the porous asphalt mixtures and calculate the depth features, and at the same time obtain the physical parameters of the porous asphalt mixtures. The multimodal feature fusion module is used to perform weighted fusion processing on the gap features, depth features and physical parameters to generate a unified feature vector; The frequency band prediction module is used to input the unified feature vector into the Transformer model, use the self-attention mechanism of the Transformer model to process the relationship between gap features, depth features and physical parameters, and output the predicted value of the sound absorption coefficient of the corresponding frequency band through multiple independent output terminals of the Transformer model, wherein each output terminal corresponds to a different frequency band.

10. The Transformer-based multimodal porous asphalt mixture noise prediction system according to claim 9, characterized in that, It also includes a model validation and optimization module, which is used to calculate the error between the predicted value and the measured value in order to optimize the parameters of the Transformer model and obtain a Transformer model with the desired prediction accuracy.

Images