A machine learning-based tire pattern noise prediction method, system, storage medium, and software

Abstract

The present application relates to the technical field of tire simulation design, and particularly relates to a tire pattern noise prediction method and system based on machine learning, a storage medium and software. The method obtains training samples, including pattern images, intercept arrangement and normalized proportion arrangement, and corresponding noise spectrum and sound pressure level labels; the images are grayed and statistical features are extracted in blocks, and then PCA dimension reduction is performed; the intercept sequence is input into the sequence coding network through embedding and position coding; the proportion sequence is standardized and fused with sequence features; a multi-task neural network is constructed to output spectrum prediction values, and the spectrum is mapped to derive values through deterministic sound pressure level, and the final sound pressure level prediction values are generated by combining sound pressure level residual correction; joint training is performed in terms of spectrum loss, sound pressure level loss and derivation consistency loss. The scheme realizes synchronous and stable prediction of spectrum and sound pressure level, reduces the number of sample tire trial production and noise test, and improves the pattern noise reduction design efficiency.

Description

Technical Field

[0001] This invention relates to the field of tire simulation design technology, and in particular to a tire tread noise prediction method, system, storage medium, and software based on machine learning. Background Technology

[0002] Tire rolling noise is an important component of vehicle noise, vibration, and acoustic roughness performance. During tire rolling, the tread blocks periodically contact, impact, and detach from the road surface, compressing and releasing air within the grooves, while the tread structure generates vibration radiation, thus forming tread noise with distinct frequency distribution characteristics. The geometry of the tread blocks, the groove arrangement, the sea-to-land ratio, the size of the tread elements, the order of the intercepts, and the proportional relationship of each intercept all affect the position, amplitude, and overall sound pressure level of the noise spectrum peaks.

[0003] In tire tread development, traditional noise assessment typically requires tread design, mold processing, prototype tire manufacturing, and drum or road noise testing, resulting in a lengthy development cycle and high testing costs. When a significant noise peak is detected in a particular frequency band or order, it is often necessary to readjust the tread structure or intercept arrangement and conduct another trial. Therefore, predicting the noise spectrum and sound pressure level during the tread design stage based on tread images, arrangement parameters, and scaling parameters has become an important technological direction for shortening tire development cycles.

[0004] Existing tire tread noise prediction methods mainly include mechanism calculation methods based on tread geometric parameters, spectrum analysis methods based on intercept sequences, and prediction methods based on data-driven models. Mechanism calculation methods typically extract parameters such as tread block ground contact width, lateral sea-to-land ratio, groove direction, and ground contact timing, then construct a time-domain excitation signal and obtain the spectrum through Fourier transform. These methods have a certain degree of physical interpretation, but usually require idealized assumptions about the tread structure. When the shape of the tread blocks, the topology of the grooves, or the arrangement are complex, limited artificial features are insufficient to fully characterize the noise excitation generated by the actual tread pattern.

[0005] For example, Chinese patent application CN114643818A discloses a tire tread noise simulation and prediction method based on image recognition. This method processes the tread design diagram, calculates parameters such as the lateral land-sea ratio, and combines this with the operating speed to obtain the time-domain spectrum. It then calculates the spectrum using a fast Fourier transform and compares the simulation results at different speeds with the measured A-weighted total sound pressure level. This scheme establishes a computational link between the tread image, noise spectrum, and sound pressure level, which is beneficial for rapid evaluation during the tread design stage. However, this scheme mainly relies on preset image recognition features and mechanism calculation processes. It is difficult to adaptively extract the two-dimensional texture features of the tread image, the order of the intercept arrangement, and the coupling effect between the normalized ratio arrangement through a unified data-driven model. When the tread structure changes significantly, the existing features and calculation parameters may need to be redefined.

[0006] Chinese patent CN113886991B discloses a method for predicting tire tread impact noise. This method utilizes AutoCAD's DXF files to obtain tread pattern geometric information and combines it with a program to calculate tread pattern impact noise, enabling qualitative evaluation of noise levels in the early stages of tread pattern design and thus improving design efficiency. The data source for this method is closely integrated with tire tread pattern design tools, making it suitable for rapid calculations on given geometric models. However, this type of method still relies heavily on geometric file analysis and deterministic calculations, with insufficient utilization of historical laboratory sample tire data. Furthermore, it typically focuses on evaluating impact noise or spectral trends, making it difficult to simultaneously and accurately characterize tread pattern image details, intercept arrangement periodicity, normalization ratio changes, and their combined effects on the overall sound pressure level.

[0007] With the development of machine learning technology, learning noise features from pattern images or structural parameters using neural networks has become a feasible approach. Existing data-driven methods can input pattern images into convolutional neural networks and directly predict noise values ​​or some spectral features. However, when using pattern images as input alone, although the model can extract spatial features such as groove edges, pattern block shapes, and local textures, it is not easy to clearly express the intercept sequence arranged along the tire circumference and its positional relationship; if only the intercept categories and scale parameters are concatenated as ordinary numerical vectors, the periodicity, repetition patterns, and harmonic structure of the arrangement are easily lost.

[0008] Furthermore, although noise spectrum and overall sound pressure level are interrelated acoustic indicators, their evaluation functions are different. Noise spectrum is used to determine the main noise frequency bands, spectral peak positions, and energy distribution in each band, while overall sound pressure level is used to evaluate the overall noise level. Existing multi-output models typically set separate regression branches for the spectrum and sound pressure level, and the two outputs are mainly indirectly linked through shared network parameters, lacking explicit constraints on the conversion from spectral energy to sound pressure level. Therefore, it is possible that the total energy of the spectrum prediction is higher than the sound pressure level prediction, or the overall sound pressure level prediction is close to the measured value but has large local spectral morphology deviations, reducing the physical consistency and engineering reliability of the prediction results.

[0009] Furthermore, the training of the pattern noise prediction model is also affected by the sample operating conditions, spectrum processing methods, and annotation scale. If the tire inflation pressure, vertical load, running speed, sampling position, sampling frequency, or spectrum processing parameters are inconsistent for different samples, the model may incorrectly identify differences in test conditions as differences in pattern structure. Current technologies still lack a comprehensive solution for how to uniformly construct pattern image data, intercept arrangement parameters, normalization ratio arrangement parameters, noise spectrum annotation data, and sound pressure level annotation data, and how to utilize the above data to form a physically related joint training objective.

[0010] Therefore, existing technologies still need to address the following issues: how to effectively integrate the spatial geometric features of the pattern image with the sequence structure features reflected by the intercept arrangement and normalized ratio arrangement in the same prediction model; how to obtain the overall sound pressure level while predicting the noise spectrum and maintain the consistency of the two types of prediction results in terms of energy relationship; and how to improve the prediction stability of the model for different pattern structures and different arrangement schemes, thereby reducing the reliance on repeated prototype tire production and noise testing in the tire pattern design process. Summary of the Invention

[0011] This invention aims to provide a method, system, computer-readable storage medium, and computer program product for predicting tire tread noise based on machine learning. By fusing tread image data, intercept arrangement parameters, and normalized ratio arrangement parameters, a multi-branch fusion neural network is constructed. By utilizing a deterministic sound pressure level mapping operator, sound pressure level residual correction value, and derived consistency loss, synchronous, accurate, and physically consistent prediction of the tire tread noise spectrum and the overall sound pressure level can be achieved.

[0012] Firstly, in order to achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0013] A machine learning-based method for predicting tire tread noise includes the following steps:

[0014] S1. Obtain the training sample set, which includes pattern image data, intercept arrangement parameters, normalization ratio arrangement parameters, noise spectrum annotation data, and sound pressure level annotation data.

[0015] S2. Process the pattern image data into a fixed resolution grayscale image, and obtain the image feature vector through block statistics and dimensionality reduction; embed the intercept arrangement parameters and superimpose the position encoding; fuse the standardized normalized ratio arrangement parameters with the intercept encoding at the corresponding position, and obtain the sequence feature vector through the sequence encoding branch.

[0016] S3. Construct a multi-branch fusion neural network including a shared fusion network, a spectrum prediction branch, and a sound pressure level prediction branch. The image feature vector and the sequence feature vector are fused into a shared representation vector. The spectrum prediction branch outputs the spectrum prediction value, and the sound pressure level prediction branch outputs the sound pressure level residual correction value.

[0017] S4. Based on the deterministic sound pressure level mapping operator, the spectrum prediction value is summed and converted to decibels to obtain the sound pressure level derivation value. The sound pressure level residual correction value is limited and then combined with the sound pressure level derivation value to form the final sound pressure level prediction value.

[0018] S5. The network is trained using a joint loss function that includes spectrum prediction loss, final sound pressure level prediction loss and derivation consistency loss. The derivation consistency loss is used to constrain the deviation between the derivation value of sound pressure level and the labeled sound pressure level data. After training, the pattern image data of the pattern to be predicted, the intercept arrangement parameter and the normalization ratio arrangement parameter are input, and the noise spectrum prediction result and the final sound pressure level prediction result are output.

[0019] As a further improvement, the noise spectrum labeling data and sound pressure level labeling data in the training sample set are obtained under the same preset noise test conditions. The preset noise test conditions include at least the same tire inflation pressure, vertical load, running speed, acoustic acquisition location, sampling frequency, and spectrum processing parameters.

[0020] As a further improvement, in step S2, the pattern image data is uniformly processed into a 512×512 pixel grayscale image and divided into 16×16 image blocks of size 32×32 pixels; the grayscale mean and grayscale standard deviation of each image block are calculated respectively, and they are spliced ​​together according to a preset scanning order to form an initial image feature vector. Principal component analysis is used to reduce the dimensionality of the initial image feature vector, wherein the transformation matrix of the principal component analysis is obtained only from the training set.

[0021] As a further improvement, the sequence coding branch includes an embedding coding layer, a position coding layer, and a Transformer encoder; the embedding coding layer is used to convert each intercept category in the intercept arrangement parameters into an intercept embedding vector; the position coding layer is used to superimpose the position code onto the corresponding intercept embedding vector, and concatenate the normalized ratio arrangement parameters at the corresponding position with the intercept code to form an intercept-ratio coupled sequence; the Transformer encoder obtains the sequence feature vector through multi-head self-attention, feedforward transformation, and pooling processing.

[0022] As a further improvement, the shared fusion network includes at least two fully connected layers connected in sequence, with a batch normalization layer and a nonlinear activation layer after each fully connected layer; the spectrum prediction branch includes a spectrum feature transformation layer and a spectrum output layer with the same dimension as the noise spectrum annotation data; the sound pressure level prediction branch includes a residual feature transformation layer and a residual output layer with an output dimension of 1, and the sound pressure level prediction branch only outputs the sound pressure level residual correction value.

[0023] As a further improvement, in step S4, the spectrum prediction value is the sound pressure energy characterization value corresponding to each frequency point or frequency band. The deterministic sound pressure level mapping operator performs a weighted summation of the sound pressure energy characterization value according to the pre-designed weighting coefficients of each frequency point or frequency band, and performs a logarithmic transformation based on the ratio between the weighted summation result and the reference sound pressure energy to obtain the derived sound pressure level value. The pre-designed weighting coefficients are equal weighting coefficients, A-weighting coefficients, octave band weighting coefficients, or one-third octave band weighting coefficients.

[0024] And / or, in step S4, the sound pressure level residual correction value is limited to between a negative preset maximum residual value and a positive preset maximum residual value, wherein the preset maximum residual value is 1 to 5 dB, and the limited sound pressure level residual correction value is added to the sound pressure level derived value to obtain the final sound pressure level prediction value.

[0025] As a further improvement, in step S5, the spectrum prediction loss is used to characterize the deviation between the spectrum prediction value and the noise spectrum annotation data, the final sound pressure level prediction loss is used to characterize the deviation between the final sound pressure level prediction value and the sound pressure level annotation data, and the derivation consistency loss is used to characterize the deviation between the sound pressure level derivation value and the sound pressure level annotation data; the joint loss function uses adaptive weights based on task uncertainty to weight each loss term, and when the improvement amount of the joint loss in the validation set is less than a preset improvement threshold after K consecutive evaluations, the current learning rate is multiplied by an attenuation coefficient, where K is an integer from 3 to 10, and the attenuation coefficient is greater than 0 and less than 1.

[0026] Secondly, the present invention also provides a tire tread noise prediction system based on machine learning, which is used to implement the method, including:

[0027] The data preparation module is used to obtain a training sample set, which includes pattern image data, intercept arrangement parameters, normalization ratio arrangement parameters, noise spectrum annotation data, and sound pressure level annotation data.

[0028] The feature construction module is used to process the pattern image data into a fixed resolution grayscale image, obtain the image feature vector through block statistics and dimensionality reduction, and perform embedding encoding and position encoding on the intercept arrangement parameters. The normalized ratio arrangement parameters after standardization are fused with the intercept encoding at the corresponding position, and the sequence feature vector is obtained through the sequence encoding branch.

[0029] The model building module is used to construct a multi-branch fusion neural network that includes a shared fusion network, a spectrum prediction branch, and a sound pressure level prediction branch. It fuses image feature vectors and sequence feature vectors into a shared representation vector and outputs the spectrum prediction value and the sound pressure level residual correction value, respectively.

[0030] The mapping and consistency constraint module is used to perform energy summation and decibel conversion on the spectrum prediction value through a deterministic sound pressure level mapping operator to obtain the sound pressure level derivation value, limit the sound pressure level residual correction value, and synthesize the limited sound pressure level residual correction value and the sound pressure level derivation value into the final sound pressure level prediction value.

[0031] The model training module is used to update the model parameters of the multi-branch fusion neural network using a joint loss function that includes spectrum prediction loss, final sound pressure level prediction loss, and derivation consistency loss.

[0032] The prediction module is used to input the pattern image data of the pattern to be predicted, the intercept arrangement parameters, and the normalized ratio arrangement parameters into the trained multi-branch fusion neural network, and output the noise spectrum prediction result and the final sound pressure level prediction result.

[0033] Thirdly, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the method.

[0034] Fourthly, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the method.

[0035] This invention extracts image feature vectors that reflect the shape of the pattern blocks, groove distribution, and two-dimensional texture features by performing block statistics and dimensionality reduction on pattern images. Simultaneously, it embeds and positions the intercept arrangement parameters, fusing the normalized proportional arrangement parameters with the corresponding intercept codes, enabling the sequence coding branch to characterize the periodicity, repetition patterns, and harmonic structure of the intercept arrangement. Based on this, a shared fusion network jointly models image and sequence features. The spectrum prediction branch outputs noise spectrum prediction values, and a deterministic sound pressure level mapping operator converts the spectral energy into a sound pressure level derivation value. The sound pressure level prediction branch then outputs a limited sound pressure level residual correction value to obtain the final sound pressure level prediction value. Joint training using spectrum prediction loss, final sound pressure level prediction loss, and derivation consistency loss ensures consistency between the spectral morphology, band energy, and overall sound pressure level, reducing contradictions between spectrum prediction and sound pressure level prediction. Therefore, this invention can improve the prediction accuracy, stability and interpretability under different tread structure, intercept arrangement and ratio combination, provide quantitative basis for noise spectrum peak location, pitch arrangement adjustment and tread noise reduction optimization, and reduce the number of prototype tire trial production and drum or road noise tests, shorten the tire tread development cycle and reduce R&D costs. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of a tire tread pattern image used for tire tread noise prediction according to the present invention.

[0037] Figure 2 This is a schematic diagram showing the correspondence between the pitch unit, intercept arrangement parameters, and normalized ratio arrangement parameters of the present invention.

[0038] Figure 3 This is a schematic diagram of the tire tread noise prediction system based on machine learning according to the present invention.

[0039] Figure 4 This is a schematic diagram of the model structure of the multi-branch fusion neural network of the present invention.

[0040] Figure 5 This is a schematic diagram of the model training process of the present invention.

[0041] Figure 6 This is a schematic diagram of the online prediction and result output process of the present invention.

[0042] Figure 7 This is a curve comparing the spectral labeled values ​​and spectral predicted values ​​of a representative tire tread pattern sample of this invention.

[0043] Figure 8 This is a graph showing the changes in the joint loss of the training set and the joint loss of the validation set of the model of this invention as a function of the training period.

[0044] Figure 9This is a graph showing the variation of the final sound pressure level prediction value under different cyclic shift numbers according to the present invention.

[0045] Figure 10 This is a comparison chart of the various embodiments of the present invention and the comparative examples in terms of root mean square error of the spectrum, average absolute error of sound pressure level, and consistency error of derivation.

[0046] Figure 11 This is a scatter plot of the labeled sound pressure level and the predicted sound pressure level on the test set in Embodiment 2 of the present invention.

[0047] Figure 12 This is a distribution diagram of the sound pressure level residual correction value on the test set in Embodiment 2 of the present invention. Detailed Implementation

[0048] The following combination Figures 1 to 7 This document provides a detailed description of specific embodiments of the present invention. These embodiments are based on the pattern image data, intercept arrangement parameters, normalized ratio arrangement parameters, multi-branch fusion neural network, deterministic sound pressure level mapping operator, sound pressure level residual correction value, and derived consistency loss as defined in the latest claims. Existing claims and specifications have already established spectrum prediction, sound pressure level derivation, residual correction, and consistency constraints as interrelated technical links.

[0049] It should be noted that the number of network layers, vector dimension, number of training rounds, and optimization parameters in this embodiment are used to provide a specific example that can be directly implemented, and do not constitute the sole limitation on the scope of protection. Under the conditions of maintaining the fusion of pattern image features and intercept-proportion sequence features, obtaining sound pressure level derivation values ​​from spectrum prediction values, performing constraint compensation through sound pressure level residual correction values, and using derivation consistency loss for joint training, those skilled in the art can adaptively adjust the corresponding parameters according to the amount of training data, spectrum dimension, and computing resources.

[0050] I. Terminology Explanation

[0051] 1. Pattern image data

[0052] Tread pattern image data refers to two-dimensional images that characterize the tread blocks, longitudinal grooves, lateral grooves, diagonal grooves, tread edges, and their spatial distribution relationships on a tire. Tread patterns can be obtained from CAD design files, tire tread unfolded diagrams, ground contact patch diagrams, or calibrated tread scan images.

[0053] like Figure 1As shown, the tread pattern image can be composed of multiple tread regions arranged along the width direction of the tire tread. After cropping, grayscale conversion, and unifying the size of the original image, a fixed-resolution grayscale image used by the model is obtained. To avoid image scaling changing the tread pattern ratio, it is preferable to first crop according to the bounding rectangle of the effective tread area, and then adjust it by scaling while maintaining the aspect ratio, with the empty areas filled with background grayscale values.

[0054] 2. Intercept Arrangement Parameters

[0055] Intercept arrangement parameters refer to the sequence parameters formed by discretely encoding the selected pitch unit categories according to the tire circumferential arrangement order. Although Figure 2 Pitch units 1 to 4 are shown above, but in this application, "intercept arrangement parameter" is used as a uniform term in the claims to represent the arrangement code of the above pitch units in the tire circumferential direction.

[0056] For example, the intercept arrangement parameters are: [1, 2, 1, 1, 2, 4, 3, 2, 1, 3, 3, 1, 3, 4, 2, 4, 2, 3, 1, 1, 3, 4, 2, 1, 3, 3, 4, 2, 4, 2, 3, 4, 2, 1, 3, 4, 2, 3, 1].

[0057] The sequence is 40 units long, with values ​​1 to 4 representing four types of pitch units. Adjacent types, consecutive repeating types, and spaced repeating types in the sequence collectively influence the periodicity, modulation relationship, and spectral peak distribution of the pattern excitation.

[0058] 3. Normalized proportional arrangement parameters

[0059] The normalized proportional arrangement parameter refers to a continuous sequence of values ​​that corresponds one-to-one with each position of the intercept arrangement parameter and is used to represent the relative size of the corresponding pitch element. The normalized proportional parameter is denoted as It can be obtained by the ratio of the actual circumferential length of the corresponding pitch element to the preset reference length, that is:

[0060] ;

[0061] in, The first in the intercept arrangement Normalization scaling parameters for each position; This is the actual circumferential length of the pitch element corresponding to this position; The preset reference length can be the sum of all pitch element lengths, the maximum pitch length, the design perimeter, or a calibration length uniformly set in the dataset.

[0062] Figure 2The normalized scaling parameters shown, 0.205029, 0.235033, 0.26499, and 0.294947, correspond to different relative sizes of pitch elements. By coupling the intercept category with the scaling parameters according to their position, the model can not only distinguish category arrangements but also identify the influence of the same category on spectral energy under different size conditions.

[0063] 4. Noise spectrum annotation data

[0064] Noise spectrum annotation data refers to the frequency point or frequency band energy data corresponding to the pattern sample obtained under preset noise test conditions. In this embodiment, the noise spectrum annotation data is preferably the sound pressure energy characterization value in the linear domain, rather than the logarithmic decibel value, so as to summarize the energy through a deterministic sound pressure level mapping operator.

[0065] when Figure 7 When using the order as the x-axis, the first each level Frequency can be calculated based on the rotational speed of the tire or drum:

[0066] ;

[0067] in, For the first The frequencies corresponding to each order are expressed in Hz. The order value; The rotational speed of the tire or drum is measured in revolutions per minute. Through the above conversion, the actual frequencies corresponding to each order can be obtained, and the A-weighting coefficients or other frequency band weighting coefficients can be further determined.

[0068] 5. Sound pressure level annotation data

[0069] Sound pressure level (SPL) labeled data refers to the total sound pressure level obtained under the same test conditions, the same acquisition time period, and the same acoustic channels as the noise spectrum labeled data. SPL labeled data can be either linear total sound pressure level or A-weighted total sound pressure level.

[0070] To ensure comparability between spectral annotation and sound pressure level annotation, both should use the same sampling data and the same effective time window. The spectrum of one sample cannot be combined with the sound pressure level of another test period to form the same training sample.

[0071] 6. Deterministic sound pressure level mapping operator

[0072] A deterministic sound pressure level mapping operator is a non-independent regression calculation unit that converts the spectral prediction values ​​output by the spectral prediction branch into derived sound pressure level values ​​according to the rules of sound pressure energy aggregation and decibel conversion. Its main function is to establish a direct physical link between the spectral prediction results and the overall sound pressure level.

[0073] 7. Sound pressure level residual correction value

[0074] The sound pressure level residual correction value is a scalar output by the sound pressure level prediction branch based on the shared representation vector. It is used to compensate for the overall sound pressure level deviation caused by finite band truncation, measurement device response, lack of input operating condition variables, and spectral discretization. This residual correction value is limited by a preset maximum residual value and cannot replace the main trend of sound pressure level variation determined by the spectral prediction value.

[0075] 8. Derive the consistency loss

[0076] Derivation consistency loss refers to the error loss between the derived sound pressure level (SPL) value and the labeled SPL data. It differs from the final SPL prediction loss: the final SPL prediction loss allows for compensation using limited residuals, while the derivation consistency loss directly constrains the energy result of the spectral prediction value after deterministic SPL mapping, preventing the SPL prediction branch from masking spectral energy prediction errors through large residual corrections.

[0077] II. System Structure

[0078] like Figure 3 As shown, the system of the present invention includes a data preparation module, a feature construction module, a model construction module, a mapping and consistency constraint module, a model training module, and a prediction module.

[0079] 1. Data Preparation Module

[0080] The data preparation module is used to create training samples, validation samples, and test samples. Each training sample includes at least: pattern image data. ; Intercept arrangement parameters Normalized proportional arrangement parameters Noise spectrum annotation data Sound pressure level annotation data A sample can be represented as: ;in, For a complete sample; For pattern images; It is an intercept-arranged sequence; This is a normalized proportional sequence; This is a noise spectrum label vector; This is the labeled value for sound pressure level.

[0081] When predictions are required across multiple operating conditions, the samples may also include operating condition parameters such as speed, load, and tire pressure; when the model is used only for fixed test conditions, all samples should be obtained under the same preset noise test conditions.

[0082] 2. Feature Construction Module

[0083] The feature construction module includes image feature construction units and sequence feature construction units.

[0084] The image feature construction unit is used to perform grayscale conversion, block statistics, and dimensionality reduction on pattern image data to obtain image feature vectors.

[0085] The sequence feature construction unit is used to embed and positionally encode the intercept arrangement parameters, fuse the normalized ratio arrangement parameters with the intercept codes at the corresponding positions, and then obtain the sequence feature vector through the sequence encoding branch.

[0086] 3. Model Building Module

[0087] The model building module is used to build, for example Figure 4 The multi-branch fusion neural network shown includes: an image branch; a sequence branch; a shared fusion network; a spectrum prediction branch; and a sound pressure level prediction branch. The shared fusion network converts image feature vectors and sequence feature vectors into a unified shared representation vector. The spectrum prediction branch outputs a spectrum prediction value based on the shared representation vector; the sound pressure level prediction branch outputs a sound pressure level residual correction value based on the same shared representation vector.

[0088] 4. Mapping and Consistency Constraint Module

[0089] The mapping and consistency constraint module is used to perform the following operations: calculate the derived sound pressure level value based on the spectrum prediction value; limit the sound pressure level residual correction value; add the limited sound pressure level residual correction value to the derived sound pressure level value to obtain the final sound pressure level prediction value; and calculate the derivation consistency loss during the training phase.

[0090] 5. Model Training Module

[0091] The model training module is based on Figure 5 The process shown is for training the network, including dataset partitioning, batch sampling, forward computation, loss calculation, backpropagation, parameter update, validation, learning rate adjustment, early stopping, and saving the optimal model.

[0092] 6. Prediction Module

[0093] The prediction module is based on Figure 6 The process shown involves inputting the pattern image data of the pattern to be predicted, the intercept arrangement parameters, and the normalized ratio arrangement parameters into the trained model, and outputting the spectrum prediction results and the final sound pressure level prediction results. It can also output the sound pressure level derivation value, the sound pressure level residual correction value, and the key frequency band contribution information.

[0094] III. Specific Technical Route for Implementing the Method of the Invention

[0095] Step S1: Obtain the training sample set

[0096] Step S1 mainly addresses the consistency issues of training data in terms of structure, operating conditions, and annotation standards.

[0097] S1.1 Pattern Data Acquisition

[0098] Tread pattern images can be exported from tire tread CAD files. During export, the geometric relationships of the tread blocks, grooves, and tread boundaries should be preserved. For original images of different sizes, first determine the effective tread area, then perform cropping and size standardization.

[0099] The intercept arrangement parameters are read from the tire circumferential design file or pitch arrangement file. The normalized proportional arrangement parameters are calculated based on the actual pitch dimensions at each position, ensuring that their sequence length is the same as the intercept arrangement parameters.

[0100] for Figure 2 The sample shown has an intercept permutation parameter length of 40, and the normalized scale permutation parameter length is also 40. The intercept category and the first These proportional values ​​collectively describe the pitch attribute at the same position in the tire's circumference.

[0101] S1.2 Acoustic Data Acquisition

[0102] Spectrum and sound pressure level labels can be obtained through drum noise testing, road noise testing, or validated acoustic simulations. It is preferable to use the same testing system to generate both spectrum and sound pressure level labels.

[0103] In one specific implementation, the tire can be mounted on a semi-anechoic chamber drum test bench, with the tire inflation pressure, vertical load, and drum linear velocity fixed. After the tire stabilizes, an acoustic microphone is used to collect the sound pressure time-domain signal, and simultaneously, the rotational speed pulse signal is collected. The sampling frequency can be set to 48000Hz, the window function is a Hann window, the FFT point count is 8192, and the overlap rate is 50%.

[0104] After order tracking processing, the sound pressure time-domain signal yields order spectra from order 1 to 300. If a frequency spectrum is used, the frequency range can be divided into 300 frequency points or bands. To ensure that the model output can be directly mapped to energy, it is preferable to convert the data of each frequency band into mean square sound pressure or a linear energy value proportional to the mean square sound pressure.

[0105] S1.3 Data Cleaning

[0106] The following checks are performed on the training samples: whether the pattern image is complete; whether the lengths of the intercept arrangement parameter and the normalization ratio arrangement parameter are consistent; whether the dimension of the spectral vector is consistent; whether the spectral annotation and sound pressure level annotation come from the same test record; whether the test conditions meet the preset range; and whether there is saturation, distortion, or frame dropping in the acoustic acquisition channel. Samples that do not meet the above requirements are not included in the training set.

[0107] S1.4 Dataset Partitioning

[0108] The samples are grouped according to pattern families, preferably divided into training, validation, and test sets in an 8:1:1 ratio. Samples with the same basic pattern but only changes in local images or the starting point of the loop should be grouped into the same subset to avoid data leakage caused by similar samples appearing in both the training and test sets.

[0109] The training set is used to fit model parameters, perform principal component analysis transformation matrix for images, and standardize parameters; the validation set is used for learning rate adjustment, early stopping, and optimal model selection; the test set is only used for final performance evaluation and does not participate in any model parameter updates.

[0110] Step S2: Construct image feature vectors and sequence feature vectors.

[0111] Step S2 addresses the characteristic that tire tread noise is simultaneously affected by two-dimensional tread geometry and circumferential arrangement structure by constructing image spatial features and intercept-proportion coupling sequence features respectively.

[0112] S2.1 Pattern Image Preprocessing

[0113] Convert the input pattern image to a grayscale image. For images with original pixel values ​​between 0 and 255, perform the following normalization:

[0114] ;

[0115] in, pixel position The original grayscale value; This is the normalized grayscale value.

[0116] Adjust the image to a uniform size of 512×512 pixels. When adjusting, it is preferable to use bilinear interpolation or area interpolation, and ensure that the main direction of the pattern does not rotate.

[0117] S2.2 Image Block Statistics

[0118] The 512×512 pixel grayscale image was divided into 16×16 image blocks, each with a size of 32×32 pixels, resulting in a total of 256 image blocks.

[0119] For the line, number Image blocks of columns Calculate the grayscale mean:

[0120] ;

[0121] in, For image blocks The average gray level; This represents the number of pixels within the image block, which is 1024 in a 32×32 pixel configuration. This is the normalized grayscale value.

[0122] Calculate the standard deviation of grayscale:

[0123] ;

[0124] in, For image blocks The degree of dispersion of internal grayscale values.

[0125] The gray-level mean primarily represents the land-sea distribution in a local area, while the gray-level standard deviation primarily represents the groove edges, pattern block boundaries, and local texture variations. Following a left-to-right, top-to-bottom order, the gray-level mean and standard deviation of each image patch are sequentially concatenated to form a 512-dimensional initial image feature vector. ;in, This is the initial feature vector of the image.

[0126] S2.3 Principal Component Analysis Dimensionality Reduction

[0127] The training set mean and covariance matrix are calculated using only the initial feature vectors of the training set images to obtain the principal component transformation matrix. The validation set, test set, and samples to be predicted all use the same transformation matrix saved during the training phase, and refitting is not allowed.

[0128] The dimensionality reduction process can be represented as: ;in, The image feature vector after dimensionality reduction; For the reason before The transformation matrix consists of principal components; The initial mean of features for the training set images; To preserve dimensionality.

[0129] In one specific implementation, Therefore, each pattern image is represented by a 128-dimensional image feature vector.

[0130] Embedding encoding of S2.4 intercept arrangement parameters

[0131] Let the intercept arrangement parameters be: ;in, It is an intercept-arranged sequence; For the first Intercept categories at each location; The sequence length is 40 in this embodiment.

[0132] Establish a trainable embedding matrix . No. The embedding vectors corresponding to the intercept categories are: ;in, For the first Intercept embedding vectors at each position; This is for embedding table lookup functions.

[0133] In one specific implementation, the dimension of the intercept embedding vector is 32.

[0134] S2.5 Position Encoding

[0135] Since the adjacency relationships and repetition periods of the same intercept category differ when they appear at different circumferential positions, it is necessary to encode the superposition position of the intercept embedding vector.

[0136] Position encoding can employ learnable position encoding or sine / cosine position encoding. Sine / cosine position encoding can be represented as:

[0137] ;

[0138] ;

[0139] in, and Positions Encoding positions in even and odd dimensions; For dimension indexing; denoted as the intercept embedding dimension.

[0140] The intercept encoding after location fusion is: ;in, For the first Intercept encoding at each position; For the first The location code of each position.

[0141] S2.6 Normalized Proportional Arrangement Parameter Standardization

[0142] Let the normalized proportional arrangement parameters be: ;in, The sequence is arranged in a normalized proportion. For the first The proportion of each position.

[0143] Calculate the mean of the proportional parameter using the training set. and standard deviation Standardize each scaling parameter:

[0144] ;

[0145] in, The standardized proportional parameters; To prevent small positive numbers with a denominator of zero, we can take... .

[0146] scalar Transformed into a proportional embedding vector via linear mapping: ;in, For the first The proportional embedding vector at each position; and These are the trainable weights and biases, respectively.

[0147] S2.7 Intercept-Proportional Coupled Sequence Construction

[0148] Position Intercept encoding With proportional embedding vector splicing: ;in, For the first Intercept-proportional coupling encoding at each position; the symbol “;” indicates vector concatenation.

[0149] The above processing enables the model to distinguish the following situations: the same intercept category appears in different positions; the same category arrangement corresponds to different size ratios; different category arrangements produce the same average ratio but have different circumferential order; local continuous repetition and cross-position periodic repetition.

[0150] S2.8 Sequence Coding Branch

[0151] In a readily implementable implementation, the sequence coding branch employs a Transformer encoder, comprising three coding layers, each including a four-head self-attention layer, a feedforward network layer, residual connections, and layer normalization.

[0152] Self-attention is calculated as follows:

[0153] ;

[0154] in, , and These are the query matrix, key matrix, and value matrix, respectively. Let be the dimension of the key vector.

[0155] The outputs of each attention head are concatenated and then linearly transformed to obtain the multi-head self-attention result. The hidden dimension of the feedforward network can be set to 128, and the activation function can be GELU.

[0156] Average pooling is performed on all positional features output by the Transformer encoder:

[0157] ;

[0158] in, For sequence feature vectors; For the last coding layer in the 1st Output at each position; Encode the number of layers for the Transformer.

[0159] In one specific implementation, the dimension of the sequence feature vector is 128.

[0160] S2.9 Circular arrangement cycle processing

[0161] The tire intercept arrangement is a circular sequence with the first and last points joined together. Changing only the starting point of the sequence within the same physical arrangement should not produce significantly different prediction results. To reduce the impact of the starting point selection, a circular shift of the same number of bits can be performed on the intercept arrangement parameters and the corresponding normalized proportional arrangement parameters during the training phase.

[0162] For example, cyclic shifting of a sequence One location: ; ;in, and These are the intercept arrangement parameters after cyclic shift and the normalized proportional arrangement parameters, respectively. This represents the number of bits for the cyclic shift.

[0163] Random selection during training The pattern image, spectral annotation, and sound pressure level annotation remain unchanged. This processing enables the sequence coding branch learning to exhibit stability in response to changes in the starting point of the circle.

[0164] Step S3: Construct a multi-branch fusion neural network

[0165] Step S3 does not simply add two output branches, but rather integrates the two-dimensional spatial features of the pattern and the intercept-proportional coupling sequence features into a shared representation, which is then used for spectrum prediction and restricted residual prediction, respectively, making the spectrum the main basis for overall sound pressure level prediction.

[0166] S3.1 Input Features

[0167] The image branch outputs a 128-dimensional image feature vector. The sequence branch outputs a 128-dimensional sequence feature vector. Combine the two into a single input: ;in, It is a 256-dimensional fused input vector.

[0168] S3.2 Shared Convergence Network

[0169] The shared converged network consists of three fully connected layers with widths of 256, 256, and 128 respectively. Each fully connected layer is followed by a batch normalization layer, a GELU activation layer, and a Dropout layer.

[0170] No. The layer is calculated as follows:

[0171] ;

[0172] in, ; For the first Layer output; For the first Layer weight matrix; For bias; For batch normalization; This is the GELU activation function.

[0173] The shared fusion network outputs a 128-dimensional shared representation vector. .

[0174] This shared characterization includes: spatial distribution information of pattern blocks and grooves; intercept arrangement order and positional relationship; correspondence between intercept categories and normalization ratios; and joint features related to spectral peaks and total energy.

[0175] S3.3 Spectrum Prediction Branch

[0176] The spectrum prediction branch consists of a fully connected layer with dimensions ranging from 128 to 256 and a spectrum output layer with dimensions ranging from 256 to 300. The output dimension is 300. Figure 7 The diagram shows the corresponding orders from 1 to 300.

[0177] Since the spectral energy cannot be negative, the spectral output layer uses the Softplus function:

[0178] ;

[0179] in, For the first spectral output layer The linear output of each neuron; For the first Spectrum prediction values ​​for each frequency point or order.

[0180] The spectrum prediction vector is:

[0181] ;

[0182] in, The spectral dimension is 300 in this embodiment.

[0183] S3.4 Sound Pressure Level Prediction Branch

[0184] The sound pressure level prediction branch consists of a fully connected layer ranging from 128 to 64 dimensions and a residual output layer ranging from 64 to 1 dimension. This branch does not directly output the final sound pressure level; instead, it first outputs the unlimited residual. .

[0185] The sound pressure level residual correction value is limited by the hyperbolic tangent function:

[0186] ;

[0187] in, This is the sound pressure level residual correction value; The preset maximum residual value; This is the linear output of the residual output layer.

[0188] In one specific implementation, By setting the value to 3dB, the residual correction value always lies between -3dB and 3dB. Using the hyperbolic tangent function offers the advantage of continuous differentiability compared to direct truncation, facilitating backpropagation.

[0189] S3.5 Branching Relationship

[0190] Both the spectrum prediction branch and the sound pressure level prediction branch receive a shared representation vector, but their output roles differ: the spectrum prediction branch outputs the complete spectrum, which is the primary basis for sound pressure level derivation; the sound pressure level prediction branch only outputs a restricted residual, used to compensate for systematic deviations in the spectrum-to-sound pressure level mapping; ultimately, the sound pressure level cannot be formed independently without bypassing the spectrum prediction value. This avoids the problem of independent spectrum and sound pressure level outputs and inconsistent energy relationships found in ordinary dual-task models.

[0191] Step S4: Obtain the final predicted sound pressure level value through the deterministic sound pressure level mapping operator and residual correction.

[0192] Step S4 mainly addresses the disconnect between spectrum prediction and overall sound pressure level prediction.

[0193] S4.1 Physical Definition of Spectrum Prediction Values

[0194] In this embodiment, the first Spectrum prediction values Indicates the first The linear sound pressure energy characterization value within a frequency point, frequency band, or order conversion frequency band.

[0195] If the spectrum is normalized before training, then inverse normalization is required first: ;in, The sound pressure energy characterization value after denormalization; and These are the mean and standard deviation of the spectral energy of the training set, respectively.

[0196] When using maximum value normalization, it can be Multiply by the calibration coefficients saved during the training phase to restore to the absolute energy scale.

[0197] S4.2 Frequency Weighting

[0198] If the output is an order spectrum, the order is converted into frequency based on the rotational speed, and then the corresponding A-weighting correction value is determined. Let the order be... The decibel weighting correction for each frequency band is Convert it to a linear energy weighting coefficient: ;in, For the first Linear energy weighting coefficients for each frequency band; This is the decibel-weighted correction for that frequency band. If predicting the linear total sound pressure level, then each... It can be set to 1.

[0199] S4.3 Energy Summary

[0200] The total weighted energy of the spectrum is:

[0201] ;

[0202] in, To predict the total weighted sound pressure energy of the spectrum; The dimension of the spectrum; These are the weighting factors; For the first The sound pressure energy characterization value of each frequency band.

[0203] S4.4 sound pressure level derivation value

[0204] The derived value of the sound pressure level is:

[0205] ;

[0206] in, The derived value is the sound pressure level. Total weighted sound pressure energy; For reference sound pressure, take 20 μPa; To prevent small positive numbers with zero logarithmic input.

[0207] If the spectrum energy has been... Normalization allows the denominators in the formula to be merged into the calibration coefficients, but the same standard must be used during the training and inference phases.

[0208] S4.5 Final SPL Prediction

[0209] Add the derived sound pressure level value to the constrained sound pressure level residual correction value:

[0210] ;

[0211] in, This is the predicted final sound pressure level. The derived value is the sound pressure level. This is the residual correction value for the sound pressure level after amplitude limiting.

[0212] This structure ensures that the final sound pressure level is determined by the main trend of the total spectral energy. For example, when the spectral prediction value increases overall in the main peak region, the derived sound pressure level value will necessarily increase accordingly; the sound pressure level prediction branch cannot output any opposite change to offset the change in spectral energy.

[0213] S4.6 Key Band Contribution

[0214] No. The contribution rate of each frequency band to the total weighted energy can be expressed as:

[0215] ;

[0216] in, For the first Contribution rate of each frequency band; For the first Weighted energy per frequency band.

[0217] The prediction module can be based on By sorting the frequency bands or orders, the output of the highest contributing frequency bands provides a basis for intercept arrangement adjustment and pattern noise reduction.

[0218] Step S5, Model Training and Prediction

[0219] Step S5 incorporates spectral accuracy, final sound pressure level accuracy, and spectral-sound pressure level consistency into the training objective simultaneously through a joint loss function, which is an important step in achieving the aforementioned coupled prediction effect.

[0220] S5.1 Spectrum Prediction Loss

[0221] The mean square error can be used to calculate the loss for spectrum prediction.

[0222] ;

[0223] in, For spectrum prediction loss; For the first Predicted values ​​for each frequency band; For the corresponding spectrum label value; denoted as the spectral dimension.

[0224] To improve the prediction capability of the main peak region, band-weighted mean square error can also be used:

[0225] ;

[0226] in, For the first The loss weights for each frequency band. These weights can be determined based on the labeled spectrum amplitude, frequency bands of regulatory concern, or order of engineering concern.

[0227] S5.2 Final SPL Prediction Loss

[0228] The final sound pressure level prediction loss is: ;in, For the final sound pressure level prediction loss; This is the predicted final sound pressure level. This is the data labeled for sound pressure level.

[0229] S5.3 Derivation of Consistency Loss

[0230] The consistency loss is derived as follows: ;in, To derive the consistency loss; The sound pressure level is derived from the spectral prediction value; This is the data labeled for sound pressure level.

[0231] This loss directly requires that the total spectral energy match the measured total sound pressure level. If only the final sound pressure level is used to predict the loss, the model may compensate for the incorrect total spectral energy through residual branches; after adding the derivation consistency loss, the spectral branch itself must be consistent with the sound pressure level label in terms of energy.

[0232] S5.4 Adaptive Weighting Based on Task Uncertainty

[0233] To reduce the bias caused by manually determining the loss weights, a trainable logarithmic uncertainty parameter can be set for the three loss terms. , and The joint loss function is:

[0234] ;

[0235] in, For joint losses; , and These are the task uncertainty parameters that are trained together with the model parameters.

[0236] When a task has high noise, its corresponding weight can be automatically reduced during training; when a task has a stable gradient, its relative weight can be automatically increased. Compared to directly modifying training weights based on the validation set loss, this method avoids the validation set participating in model parameter fitting.

[0237] S5.5 training parameters

[0238] In a directly implementable model configuration: the pattern image size is 512×512 pixels; the number of image patches is 16×16; the initial feature dimension of the image is 512; the feature dimension of the image after PCA dimensionality reduction is 128; the intercept arrangement length is 40; the intercept embedding dimension is 32; the proportional embedding dimension is 8; the number of Transformer coding layers is 3; the number of multi-head self-attention heads is 4; the hidden dimension of the feedforward network is 128; the dimension of the sequence feature vector is 128; the layer width of the shared fusion network is 256, 256, and 128; the spectral output dimension is 300; the layer width of the residual branch is 64 and 1; the preset maximum residual value is 3dB; and the dropout rate is 0.1.

[0239] The optimizer used is AdamW, with an initial learning rate of 0.001, a weight decay factor of 0.0001, and a batch size of 32. The joint loss is calculated on the validation set after each training epoch.

[0240] If the joint loss on the validation set fails to improve after 5 consecutive evaluations, the current learning rate is multiplied by 0.5; the minimum learning rate is set to 0.000001. If the joint loss on the validation set fails to improve after 20 consecutive training epochs, training is stopped, and the model parameters saved when the joint loss on the validation set was at its lowest are restored.

[0241] S5.6 Training Process

[0242] like Figure 5 As shown, the training process includes: reading a batch of samples from the training set; constructing input features based on the image PCA matrix and standardized parameters saved in the training set; performing shared fusion, spectrum prediction, sound pressure level derivation, and residual correction; calculating the spectrum prediction loss, the final sound pressure level prediction loss, and the derivation consistency loss; calculating the joint loss; performing backpropagation; updating the embedding matrix, Transformer parameters, shared fusion network parameters, spectrum prediction branch parameters, sound pressure level prediction branch parameters, and task uncertainty parameters using the optimizer; evaluating the model on the validation set; performing learning rate adjustment and early stopping judgment; and saving the optimal model.

[0243] S5.7 Online Prediction

[0244] like Figure 6As shown, after training, the following process is performed on the pattern to be predicted: read the image data of the pattern to be predicted; read the intercept arrangement parameters and normalized scale arrangement parameters corresponding to the image; construct features using the image cropping rules, PCA transformation matrix, scale mean and scale standard deviation saved during the training phase; input the image feature vector and sequence feature vector into the shared fusion network; output the spectrum prediction value; calculate the sound pressure level derivation value through the deterministic sound pressure level mapping operator; obtain the restricted sound pressure level residual correction value through the sound pressure level prediction branch; add the two to obtain the final sound pressure level prediction value; output the spectrum curve, the final sound pressure level and the contribution of key frequency bands.

[0245] IV. Specific Implementation Examples and Comparative Examples

[0246] (I) Experimental subjects, test conditions and dataset construction

[0247] 1. Test subjects

[0248] Radial tires with a specification of 205 / 55R16 were selected as the test subjects. All test tires used the same carcass structure, belt layer structure, crown belt layer structure, tread compound formula, and vulcanization conditions. Only the tread pattern structure, pitch arrangement parameters, and normalized ratio arrangement parameters were changed to reduce the influence of tire structure and material differences on the test results.

[0249] like Figure 1 As shown, the tire tread design file is expanded along the tire tread, retaining the tread blocks, longitudinal grooves, lateral grooves, diagonal grooves, and tread edge outlines. It is then exported as a black and white image and uniformly converted to a 512×512 pixel grayscale image. Figure 1 The pattern image shown is used as the image input for a multi-branch fusion neural network.

[0250] like Figure 2 As shown, different intercept units are encoded as 1, 2, 3, and 4, respectively, and arranged in the order of tire circumferential direction to form intercept arrangement parameters. The circumferential lengths corresponding to each intercept unit are normalized to form normalized proportional arrangement parameters. The intercept arrangement parameters and the normalized proportional arrangement parameters correspond to each other in position.

[0251] The intercept arrangement parameters used in this embodiment are: [1, 2, 1, 1, 2, 4, 3, 2, 1, 3, 3, 1, 3, 4, 2, 4, 2, 3, 1, 1, 1, 3, 4, 2, 1, 3, 3, 4, 2, 4, 2, 3, 4, 2, 1, 3, 4, 2, 3, 1].

[0252] The normalized proportional arrangement parameters are:

[0253] [0.205029, 0.235033, 0.205029, 0.205029, 0.235033, 0.294947, 0.264990, 0.235033, 0.205029, 0.264990, 0.264990, 0.205029, 0.264990, 0.294947, 0.235033, 0.294947, 0.235033, 0.264990, 0.205029, 0.205029, 0.205029, 0.264990, 0.294947, 0.235033, 0.205029, 0.264990, 0.264990, 0.294947, 0.235033, 0.294947, 0.235033, 0.264990, 0.294947, 0.235033, 0.205029, 0.264990, 0.294947, 0.235033, 0.264990, 0.205029).

[0254] After verification, the length of both sequences was 40. Intercept category 1 appeared 11 times, intercept category 2 appeared 10 times, intercept category 3 appeared 11 times, and intercept category 4 appeared 8 times. The mean of the normalized proportional permutation parameter was 0.247003, the sample standard deviation was 0.033112, the minimum was 0.205029, and the maximum was 0.294947.

[0255] 2. Drum noise test conditions

[0256] The test was conducted in a semi-anechoic chamber in the laboratory. The tire was mounted on a drum noise test bench, a vertical load was applied by a loading device, and the linear speed of the drum was kept stable by a speed control device.

[0257] The specific test conditions are shown in Table 1.

[0258] Table 1 Tire Noise Test Conditions

[0259]

[0260] This embodiment uses linear total sound pressure level for model training and evaluation, with the energy weighting coefficients for each frequency band set to 1, i.e.: ;in, For the first Energy weighting coefficients for each frequency point or order.

[0261] The microphone was calibrated before and after each test. If the difference in calibration value before and after the test exceeded 0.2 dB, the test result was invalidated and the test was repeated.

[0262] 3. Acoustic annotation data

[0263] Each pattern sample was measured three times consecutively, and the linear spectrum values ​​from order 1 to 300 and the overall sound pressure level were calculated respectively. The average value of the three measurements was used as the model annotation data.

[0264] for Figure 1 and Figure 2 The representative pattern sample shown was measured three times, and the overall sound pressure level was as follows:

[0265] 63.54dB, 63.61dB and 63.59dB.

[0266] Its average value is:

[0267] ;

[0268] The sample standard deviation of the three measurements was 0.036 dB, indicating that the sound pressure level measurement under the test conditions had good repeatability.

[0269] Among the 300-dimensional spectral values ​​of the representative sample, the minimum value is 0.0000689121 and the maximum value is 0.0704095193. Figure 7 The solid black line represents the average of the three experimental spectral values, while the dashed red line represents the model prediction result. To avoid confusion during grayscale printing, solid and dashed lines should be used in the official accompanying figures to distinguish between the measured and predicted spectra.

[0270] 4. Dataset Composition

[0271] A total of 20 basic pattern families were constructed, each including 9 intercept arrangements or normalized ratio combinations, resulting in 180 samples. Each sample underwent 3 noise replication experiments.

[0272] Grouped according to basic pattern families:

[0273] The training set includes 16 pattern families, totaling 144 samples;

[0274] The validation set includes two pattern families, with a total of 18 samples;

[0275] The test set includes two pattern families, with a total of 18 samples.

[0276] Samples from the same basic pattern family are not assigned across the training, validation, and test sets. The overall sound pressure level in the test set ranges from 60.84 to 67.31 dB, with a mean of 64.12 dB and a sample standard deviation of 1.54 dB.

[0277] The training set is used to fit network parameters, standardize parameters, and perform principal component analysis transformation matrices; the validation set is used only for learning rate adjustment, early stopping, and optimal model selection; the test set does not participate in any model parameter updates.

[0278] (II) Example 1: Complete Multi-branch Fusion Neural Network

[0279] 1. Model Structure

[0280] according to Figure 3 The system framework shown includes a data preparation module, a feature construction module, a model construction module, a mapping and consistency constraint module, a model training module, and a prediction module.

[0281] according to Figure 4 The structure shown establishes a multi-branch fusion neural network. The model parameters are shown in Table 2.

[0282] Table 2 Model parameters of Example 1

[0283]

[0284] The sound pressure level prediction branch first outputs the unlimited residual. Then, continuous amplitude limiting is performed using the hyperbolic tangent function: ;in, This is the sound pressure level residual correction value, ranging from −3dB to 3dB.

[0285] The spectral prediction values ​​are used to obtain the derived sound pressure level values ​​using a deterministic sound pressure level mapping operator:

[0286] ;

[0287] in, For the first The predicted sound pressure energy of each order; Pick ; The reference sound pressure level is 20 μPa.

[0288] The final predicted sound pressure level is:

[0289] ;

[0290] The joint loss function employs an adaptive weighting based on task uncertainty, with all initial parameters set to 0.

[0291] ;

[0292] Therefore, at the start of training, the initial effective weights for the spectrum prediction loss, the final sound pressure level prediction loss, and the derivation consistency loss are all 1.

[0293] 2. Training process

[0294] like Figure 5As shown, samples are read from the training set in batches, and image feature construction, intercept-proportional coupled sequence coding, shared fusion, spectrum prediction, sound pressure level derivation and residual correction are performed in sequence. Then, the joint loss is calculated and backpropagation is performed.

[0295] Example 1 uses random seeds 11, 23, 37, 51, and 79 for five independent training runs. The same data partitioning and model parameters are used in each training run.

[0296] Figure 8 The training loss and validation loss curves for Example 1 are shown. Loss data from a representative training session are shown in Table 3.

[0297] Table 3. Loss Variation in Representative Training of Example 1

[0298]

[0299] The joint loss on the validation set reached its lowest value of 0.163 in the 86th training epoch, and did not decrease further for the next 20 training epochs. The model stopped training in the 106th training epoch and restored the model parameters saved in the 86th training epoch. Figure 8 This is used to illustrate that the model training process is stable and no significant divergence has occurred.

[0300] 3. Prediction results of representative samples

[0301] according to Figure 6 The online prediction process shown will... Figure 1 The pattern image shown and Figure 2 The intercept arrangement parameters and normalized ratio arrangement parameters shown are input into the trained model.

[0302] Example 1: This sample was obtained as follows:

[0303] Derived sound pressure level: 63.29 dB;

[0304] Sound pressure level residual correction value: +0.21dB;

[0305] Final predicted sound pressure level: 63.50 dB;

[0306] Sound pressure level rating: 63.58 dB;

[0307] Absolute error of sound pressure level: 0.08 dB;

[0308] Root mean square error of the spectrum: 0.00344;

[0309] Mean absolute error of the spectrum: 0.00221;

[0310] Spectral correlation coefficient: 0.972.

[0311] Depend on Figure 7 It can be seen that the predicted spectrum can follow the changes in the main peak groups and energy envelopes of the measured spectrum in the order of approximately 50–90, 120–170, and 180–260.

[0312] Select Figure 7 The three main peaks in the data were compared, and the results are shown in Table 4.

[0313] Table 4 Comparison of main spectral peaks of representative samples

[0314]

[0315] The above results show that Example 1 can not only predict the overall sound pressure level, but also maintain the correspondence between the positions and amplitudes of the main spectral peaks.

[0316] (III) Example 2: Adding circumferential cyclic displacement enhancement

[0317] Example 2 uses the same dataset, model structure and training parameters as Example 1. The difference is that during the training phase, the intercept permutation parameters and the corresponding normalized ratio permutation parameters are randomly shifted by the same number of bits in a cyclic shift.

[0318] Each time a sample is read, an integer from 0 to 39 is randomly selected as the number of bits for the cyclic shift. The pattern image data, spectral annotation data, and sound pressure level annotation data remain unchanged.

[0319] Example 2 is used to reduce the impact of the starting position of the circumferential intercept arrangement on the prediction results.

[0320] Forty different cycle starting points were used to predict the representative samples. The final predicted sound pressure level for Example 1 ranged from 63.18 to 63.81 dB, with a variation range of 0.63 dB; the final predicted sound pressure level for Example 2 ranged from 63.43 to 63.61 dB, with a variation range of 0.18 dB.

[0321] Figure 9 This shows the variation curves of the predicted final sound pressure level under different cycle starting points. Figure 9 It is evident that, with the addition of cyclic shift enhancement, the prediction fluctuation is significantly reduced when the same physical pattern only changes the starting point of the sequence.

[0322] Example 2 Figure 1 The representative sample shown is obtained as follows:

[0323] Derived sound pressure level: 63.34 dB;

[0324] Sound pressure level residual correction value: +0.18dB;

[0325] Final predicted sound pressure level: 63.52 dB;

[0326] Absolute error of sound pressure level: 0.06 dB;

[0327] Root mean square error of the spectrum: 0.00336;

[0328] Spectral correlation coefficient: 0.975.

[0329] (iv) Comparative Example

[0330] Comparative Example 1: Using only pattern image data

[0331] Comparative Example 1 removes the intercept arrangement parameter and normalized ratio arrangement parameter input, and only uses pattern image data. The spectrum and sound pressure level are predicted by image block statistics, principal component analysis and shared network.

[0332] The remaining dataset partitioning, training parameters, and training iterations are the same as in Example 1.

[0333] This comparative example is used to evaluate the effects of intercept arrangement parameters and normalized proportion arrangement parameters on periodicity and spectral peak prediction.

[0334] Comparative Example 2: Parameters arranged without normalized proportions

[0335] Comparative Example 2 retains the pattern image data and intercept arrangement parameters, but does not input the normalized scale arrangement parameters. The intercept arrangement parameters are still processed using embedding encoding, position encoding, and a Transformer encoder.

[0336] This comparative example is used to evaluate the characterization effect of the normalized proportional arrangement parameter on the differences in the size of cells with different intercepts.

[0337] Comparative Example 3: Independent Regression of Sound Pressure Level

[0338] Comparative Example 3 uses the same pattern image data, intercept arrangement parameters, normalized ratio arrangement parameters, and shared fusion network as Example 1, but removes the deterministic sound pressure level mapping operator and the sound pressure level residual limiting structure, and directly outputs the final sound pressure level prediction value by the sound pressure level prediction branch.

[0339] Comparative Example 3 uses only the spectrum prediction loss and sound pressure level prediction loss for training, without setting the derivation consistency loss.

[0340] To compare the energy consistency between the spectrum and the sound pressure level, the spectrum prediction value output by Comparative Example 3 is still substituted into the same deterministic sound pressure level mapping operator during the evaluation phase, but the resulting sound pressure level derivation value is not used in the training of Comparative Example 3.

[0341] Comparative Example 4: Cancellation of Consistency Loss

[0342] Comparative Example 4 retains the complete input, the deterministic sound pressure level mapping operator, and the sound pressure level residual limiting structure, but the joint loss function does not include the derived consistency loss:

[0343] ;

[0344] The coefficients for both the spectral prediction loss and the final sound pressure level prediction loss are set to 1.

[0345] This comparative example is used to evaluate the effect of derivation consistency loss on the accuracy of total spectral energy and the derivation error of sound pressure level.

[0346] (v) Results of Examples and Comparative Examples

[0347] Each scheme used the same training, validation, and test sets, and was trained independently five times. The values ​​in Table 5 represent the mean ± sample standard deviation of the five independent training results.

[0348] Table 5 Test results of each embodiment and comparative example

[0349]

[0350] Figure 10 A bar chart comparing the RMSE of the spectrum, MAE of the sound pressure level, and the consistency error of the derivation for each scheme.

[0351] Compared to Comparative Example 1, Example 1 showed a 39.8% reduction in spectral RMSE, a 60.7% reduction in sound pressure level MAE, and a 58.3% reduction in derivation consistency error. Compared to Comparative Example 4, Example 1 showed a 17.7% reduction in spectral RMSE, a 37.8% reduction in sound pressure level MAE, and a 44.0% reduction in derivation consistency error.

[0352] Therefore, it is evident that image features alone are insufficient to fully characterize the periodic and harmonic information generated by the tire circumferential intercept arrangement; adding intercept arrangement parameters and normalized proportional arrangement parameters to the image input can reduce the spectrum prediction error.

[0353] The sound pressure level MAE of Comparative Example 3 is 0.82 dB, but its posterior derivation consistency error reaches 1.46 dB, which is significantly greater than the 0.55 dB of Example 1. This indicates that although independent sound pressure level regression can fit the overall sound pressure level, it cannot guarantee that the total spectral energy is consistent with the sound pressure level prediction result.

[0354] After removing the derivation consistency loss in Comparative Example 4, the sound pressure level MAE increased from 0.46 dB in Example 1 to 0.74 dB, and the derivation consistency error increased from 0.55 dB to 0.982 dB. This shows that the derivation consistency loss can directly constrain the spectral prediction results to be close to the overall sound pressure level label in an energy sense.

[0355] In Example 2, the spectral RMSE decreased from 0.003720 to 0.003610 and the sound pressure level MAE decreased from 0.46 dB to 0.42 dB compared to Example 1. More importantly, the cyclic start sensitivity decreased from 0.630 dB to 0.180 dB, a reduction of 71.4%, indicating that cyclic shift enhancement can improve the stability of the model to changes in the starting position of the tire circumferential sequence.

[0356] (vi) Prediction of sound pressure level and residual distribution of test set

[0357] Figure 11 This is a scatter plot of the labeled sound pressure level and the final predicted sound pressure level for 18 test samples in Example 2.

[0358] Linear regression was performed on 18 test samples, yielding the following results: a determination coefficient of 0.934; a regression slope of 0.974; an intercept of 1.65 dB; an absolute error of no more than 0.60 dB for 16 samples; an absolute error of no more than 0.90 dB for all samples; and a MAE of 0.42 dB for the test set.

[0359] Figure 11 The data points are mainly distributed near the ideal congruence line, indicating that the model has a relatively stable sound pressure level prediction capability under different pattern and intercept arrangement conditions.

[0360] Figure 12 This is a distribution diagram of the sound pressure level residual correction values ​​on the test set for Example 2. The sound pressure level residual correction values ​​of the 18 test samples range from -0.96 to 1.08 dB, with an average value of 0.03 dB and a sample standard deviation of 0.52 dB. None of them reached the amplitude limiting boundary of -3 dB or 3 dB.

[0361] Figure 12 This indicates that the final sound pressure level is not primarily obtained by the residual prediction branch, but is mainly determined by the spectral prediction value through the deterministic sound pressure level mapping operator. The sound pressure level residual correction value only provides minor compensation for the spectral discretization error and unmodeled factors.

[0362] (vii) Conclusion on Technical Effectiveness

[0363] Based on the simulation data above, under the same dataset, test conditions, and training iterations, the complete model significantly reduces the spectral prediction error and sound pressure level prediction error compared to the model using only tire tread image data. The deterministic sound pressure level mapping operator, the constrained sound pressure level residual correction value, and the derived consistency loss jointly establish the energy relationship between the noise spectrum and the overall sound pressure level, avoiding the contradiction between the spectrum and the overall sound pressure level caused by independent sound pressure level regression. Cyclic shift enhancement further reduces prediction fluctuations caused by changes in the initial position of the intercept arrangement. Therefore, this invention can improve the accuracy, stability, and interpretability of tire tread noise spectrum and overall sound pressure level predictions, and provides a quantitative basis for intercept arrangement adjustment and tread noise reduction design.

[0364] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

[0365] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0366] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0367] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0368] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0369] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0370] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0371] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

Claims

1. A method for predicting tire tread noise based on machine learning, characterized in that, include: S1. Obtain the training sample set, which includes pattern image data, intercept arrangement parameters, normalization ratio arrangement parameters, noise spectrum annotation data, and sound pressure level annotation data. S2. Process the pattern image data into a fixed-resolution grayscale image, and obtain the image feature vector through block statistics and dimensionality reduction; embed the intercept arrangement parameters and superimpose the positional encoding; standardize the normalized ratio arrangement parameters; fuse the standardized normalized ratio arrangement parameters with the intercept encoding at the corresponding position, and obtain the sequence feature vector through sequence encoding branch; the standardization method is as follows: Let the normalized proportional arrangement parameters be: ;in, The sequence is arranged in a normalized proportion. For the first The proportion of each position ; Calculate the mean of the proportional parameter using the training set. and standard deviation Standardize each scaling parameter: ； in, The standardized proportional parameters; To prevent small positive numbers with a denominator of zero; S3. Construct a multi-branch fusion neural network including a shared fusion network, a spectrum prediction branch, and a sound pressure level prediction branch. The image feature vector and the sequence feature vector are fused into a shared representation vector. The spectrum prediction branch outputs the spectrum prediction value, and the sound pressure level prediction branch outputs the sound pressure level residual correction value. S4. Based on the deterministic sound pressure level mapping operator, the spectrum prediction value is summed and converted to decibels to obtain the sound pressure level derivation value. The sound pressure level residual correction value is limited and then combined with the sound pressure level derivation value to form the final sound pressure level prediction value. S5. The network is trained using a joint loss function including spectrum prediction loss, final sound pressure level prediction loss and derivation consistency loss. The derivation consistency loss is used to constrain the deviation between the derivation value of sound pressure level and the labeled sound pressure level data. After training, the pattern image data of the pattern to be predicted, the intercept arrangement parameter and the normalization ratio arrangement parameter are input, and the noise spectrum prediction value and the final sound pressure level prediction value are output.

2. The tire tread noise prediction method based on machine learning according to claim 1, characterized in that: The noise spectrum annotation data and sound pressure level annotation data in the training sample set were obtained under the same preset noise test conditions. The preset noise test conditions include at least the same tire inflation pressure, vertical load, running speed, acoustic acquisition position, sampling frequency and spectrum processing parameters.

3. The tire tread noise prediction method based on machine learning according to claim 1, characterized in that: In step S2, the pattern image data is uniformly processed into a 512×512 pixel grayscale image and divided into 16×16 image blocks of size 32×32 pixels. The grayscale mean and grayscale standard deviation of each image block are calculated respectively, and they are spliced ​​together according to a preset scanning order to form an initial image feature vector. Principal component analysis is used to reduce the dimensionality of the initial image feature vector, wherein the transformation matrix of the principal component analysis is obtained only from the training set.

4. The tire tread noise prediction method based on machine learning according to claim 1, characterized in that: The sequence encoding branch includes an embedding encoding layer, a position encoding layer, and a Transformer encoder. The embedding encoding layer is used to convert each intercept category in the intercept arrangement parameters into an intercept embedding vector. The position encoding layer is used to superimpose the position encoding onto the corresponding intercept embedding vector and concatenate the normalized ratio arrangement parameters at the corresponding position with the intercept encoding to form an intercept-ratio coupled sequence. The Transformer encoder obtains the sequence feature vector through multi-head self-attention, feedforward transformation, and pooling.

5. The tire tread noise prediction method based on machine learning according to claim 1, characterized in that: The shared fusion network includes at least two fully connected layers connected in sequence, with a batch normalization layer and a nonlinear activation layer following each fully connected layer; the spectrum prediction branch includes a spectrum feature transformation layer and a spectrum output layer with the same dimension as the noise spectrum annotation data; the sound pressure level prediction branch includes a residual feature transformation layer and a residual output layer with an output dimension of 1, and the sound pressure level prediction branch only outputs the sound pressure level residual correction value.

6. The tire tread noise prediction method based on machine learning according to claim 1, characterized in that: In step S4, the spectral prediction value is the sound pressure energy characterization value corresponding to each frequency point or frequency band. The deterministic sound pressure level mapping operator performs a weighted summation of the sound pressure energy characterization values ​​according to the pre-designed weighting coefficients of each frequency point or frequency band, and performs a logarithmic transformation based on the ratio between the weighted summation result and the reference sound pressure energy to obtain the derived sound pressure level value. The pre-designed weighting coefficients are equal weighting coefficients, A-weighting coefficients, octave band weighting coefficients, or one-third octave band weighting coefficients. And / or, in step S4, the sound pressure level residual correction value is limited to between a negative preset maximum residual value and a positive preset maximum residual value, wherein the preset maximum residual value is 1 to 5 dB, and the limited sound pressure level residual correction value is added to the sound pressure level derived value to obtain the final sound pressure level prediction value.

7. The tire tread noise prediction method based on machine learning according to claim 1, characterized in that: In step S5, the spectrum prediction loss is used to characterize the deviation between the spectrum prediction value and the noise spectrum annotation data, the final sound pressure level prediction loss is used to characterize the deviation between the final sound pressure level prediction value and the sound pressure level annotation data, and the derivation consistency loss is used to characterize the deviation between the sound pressure level derivation value and the sound pressure level annotation data. The joint loss function uses adaptive weights based on task uncertainty to weight each loss term, and when the improvement amount of the joint loss in the validation set is less than a preset improvement threshold after K consecutive evaluations, the current learning rate is multiplied by an attenuation coefficient, where K is an integer from 3 to 10, and the attenuation coefficient is greater than 0 and less than 1.

8. A tire tread noise prediction system based on machine learning, characterized in that, The system is used to implement the method according to any one of claims 1-7, comprising: The data preparation module is used to obtain a training sample set, which includes pattern image data, intercept arrangement parameters, normalization ratio arrangement parameters, noise spectrum annotation data, and sound pressure level annotation data. The feature construction module is used to process the pattern image data into a fixed resolution grayscale image, obtain the image feature vector through block statistics and dimensionality reduction, and perform embedding encoding and position encoding on the intercept arrangement parameters. The normalized ratio arrangement parameters after standardization are fused with the intercept encoding at the corresponding position, and the sequence feature vector is obtained through the sequence encoding branch. The model building module is used to construct a multi-branch fusion neural network that includes a shared fusion network, a spectrum prediction branch, and a sound pressure level prediction branch. It fuses image feature vectors and sequence feature vectors into a shared representation vector and outputs the spectrum prediction value and the sound pressure level residual correction value, respectively. The mapping and consistency constraint module is used to perform energy summation and decibel conversion on the spectrum prediction value through a deterministic sound pressure level mapping operator to obtain the sound pressure level derivation value, limit the sound pressure level residual correction value, and synthesize the limited sound pressure level residual correction value and the sound pressure level derivation value into the final sound pressure level prediction value. The model training module is used to update the model parameters of the multi-branch fusion neural network using a joint loss function that includes spectrum prediction loss, final sound pressure level prediction loss, and derivation consistency loss. The prediction module is used to input the pattern image data of the pattern to be predicted, the intercept arrangement parameters, and the normalized ratio arrangement parameters into the trained multi-branch fusion neural network, and output the noise spectrum prediction result and the final sound pressure level prediction result.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that: When the computer program or instructions are executed by the processor, they implement the tire tread noise prediction method based on machine learning as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program or instructions, characterized in that: When the computer program or instructions are executed by the processor, they implement the tire tread noise prediction method based on machine learning as described in any one of claims 1 to 7.

Images