Method for identifying adhesion characteristics of thin-sand pavement

Abstract

This invention discloses a method for identifying the adhesion characteristics of thin sand pavements. The method includes collecting acoustic signals generated by the interaction between a tire and thin sand pavements with different adhesion characteristics, reflecting the differences in adhesion characteristics between the different thin sand pavements; preprocessing the acoustic signals and extracting multi-dimensional signal features; organizing the signal features from multiple consecutive time windows into a multi-time-step sequence; performing generative time-series data augmentation on the multi-time-step sequence to simultaneously preserve the signal statistical distribution and temporal sequence relationship, resulting in an augmented dataset; using the augmented dataset as input and the corresponding thin sand pavement adhesion characteristics as labels, training a pre-constructed time-series neural network model; and using the trained time-series neural network model to identify the adhesion characteristics of thin sand pavements.

Description

Technical Field

[0001] This invention belongs to the field of vehicle engineering technology, and in particular relates to a method for identifying the adhesion characteristics of thin sand pavement. Background Technology

[0002] Road surface adhesion characteristics are core input parameters for active safety control systems such as anti-lock braking systems (ABS) and electronic stability programs (ESCs). The accuracy of their identification directly impacts vehicle driving safety and control stability. In desert sections, suburban construction zones, and coastal desertified sections, road surfaces are often covered with a thin layer of sand (2-5 mm). The adhesion characteristics of such surfaces fluctuate significantly with sand particle size and distribution, easily leading to risks such as increased braking distance and steering skidding. Therefore, accurate identification of the adhesion characteristics of thin sand road surfaces is of significant engineering importance.

[0003] Existing methods for identifying road surface adhesion characteristics are mostly designed for hard pavements such as asphalt and concrete. Based on the rigid contact mechanism of tires and roads in hard pavements, they generally adopt an identification mode that combines physical signal interception with fixed feature matching. However, the sand layer structure of thin sand pavements is loose. During tire rolling, the tread rubber undergoes elastic deformation, which in turn causes the sand particles to undergo discrete motion behaviors such as impact and rolling. The tire-road interaction mechanism is fundamentally different from that of hard pavements. Directly applying existing identification modes for hard pavements cannot accurately capture the unique particle impact acoustic characteristics of thin sand pavements, making it difficult to effectively distinguish between thin sand pavements with different adhesion states.

[0004] At the signal processing level, most existing methods for identifying pavement adhesion characteristics directly extract features after windowing the original acquired signals, without specifically filtering and optimizing features for the impact-type signal characteristics of thin sand pavements. The resulting feature sets contain a large number of redundant dimensions unrelated to pavement adhesion characteristics, with the overall feature dimensions reaching over a thousand levels. Excessive feature redundancy not only significantly increases the computational cost of subsequent identification models and reduces overall identification efficiency, but also weakens effective discriminative information due to interference from irrelevant features, further reducing the accuracy of identifying the adhesion characteristics of thin sand pavements.

[0005] In terms of utilizing temporal information, existing methods for identifying road surface adhesion characteristics mostly rely on static features within a single time window. They fail to fully explore the temporal evolution of the gravel impact process, and the temporal dependence between features in adjacent windows is not effectively utilized. This makes it impossible to fully characterize the continuous change of gravel movement over time under thin sand pavements, and it is difficult to accurately depict the differences in adhesion characteristics corresponding to sand layers of different particle sizes, thus limiting the reliability of the identification results.

[0006] In addition, the cost of collecting test data on real thin sand pavement is high and the applicable working conditions are limited. Insufficient sample size can easily lead to overfitting of the identification model. Traditional data augmentation methods such as random noise addition, amplitude shifting, and window clipping can only change the statistical distribution of the signal, which will destroy the inherent time dependence of the sand impact process. The generated samples have poor temporal authenticity and cannot effectively improve the generalization ability of the model.

[0007] In summary, existing technologies cannot adapt to the special tire-road action mechanism of thin sand pavement, and have many defects such as high feature redundancy, insufficient utilization of time-series information, and low identification accuracy under small sample conditions, making it difficult to meet the engineering application requirements for accurate and efficient identification of the adhesion characteristics of thin sand pavement. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention proposes a method for identifying the adhesion characteristics of thin sand pavement.

[0009] The technical solution of the present invention is as follows:

[0010] A method for identifying the adhesion characteristics of thin sand pavement includes:

[0011] Acoustic signals generated by the interaction between the tire and thin sand pavement with different adhesion characteristics were collected, which can reflect the differences in adhesion characteristics of different thin sand pavement.

[0012] The acoustic signal is preprocessed and multi-dimensional signal features are extracted, which cover the temporal morphological characteristics, frequency domain structural characteristics and transient change characteristics of the signal;

[0013] The signal features of multiple consecutive time windows are organized into a multi-time step sequence to characterize the process of the movement of gravel after contact with tires under different thin sand pavement adhesion characteristics over time.

[0014] Generative time-series data augmentation is performed on multi-time-step sequences to simultaneously preserve the signal statistical distribution and temporal sequence relationship, resulting in an augmented dataset.

[0015] Using the augmented dataset as input and the corresponding thin sand road surface adhesion characteristics as labels, a pre-constructed temporal neural network model is trained, and the trained temporal neural network model is used to identify the adhesion characteristics of thin sand road surfaces.

[0016] Furthermore, the preprocessing includes bandpass filtering, differential noise reduction, short-time window segmentation, and amplitude normalization. Bandpass filtering removes low-frequency structural vibrations and high-frequency environmental noise unrelated to the interaction between the tire and the thin sand pavement. Differential noise reduction addresses the different representations of the acoustic signal in the time and frequency domains. In the time domain, short-window median filtering extracts impact response variation features equivalent to structural vibrations. In the frequency domain, a slight spectral subtraction method is used to reduce waveform distortion while preserving key impact features, thereby improving the separability of tire-gravel contact events.

[0017] Furthermore, the multi-dimensional signal features are extracted from five dimensions: time-domain statistics, frequency-domain statistics, difference and derivative, envelope and morphology, and time-frequency perception.

[0018] Furthermore, after extracting signal features, principal component analysis is used to reduce the dimensionality of the high-dimensional features.

[0019] Furthermore, the multi-time-step sequence is composed of signal features from 10 consecutive time windows spliced ​​together in chronological order, with each time window lasting 0.06 seconds and adjacent windows having an overlap rate of 50%.

[0020] Furthermore, the generative temporal data augmentation is implemented using the TimeGAN temporal generative adversarial network.

[0021] Furthermore, the TimeGAN adopts an architecture that coordinates the training of five parts: an embedding network, a recovery network, a supervisor, a generator, and a discriminator, and is trained through joint optimization of the reconstruction loss.

[0022] Furthermore, the temporal neural network model is a CNN-LSTM model, in which the convolutional neural network layer extracts short-term impact features within a local time window, and the long short-term memory network layer captures long-range temporal dependencies across time steps.

[0023] Furthermore, before training the CNN-LSTM model, the augmented dataset is divided into training, validation, and test sets.

[0024] Furthermore, acoustic signals from thin sand pavements with at least five different particle sizes were collected and classified into three roughness levels—coarse, medium, and fine—according to different pavement adhesion characteristic ranges, based on engineering application requirements.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] This invention proposes a method for identifying the adhesion characteristics of thin sand pavement. This method addresses the special working conditions of thin sand pavement and the shortcomings of existing technologies. It effectively adapts to the tire-road action characteristics of thin sand pavement through acoustic signal acquisition, multi-dimensional feature extraction and screening, multi-time step sequence construction, generative time series data augmentation and time series neural network identification. It solves the problems of feature redundancy, insufficient utilization of time series dependence and low identification accuracy under small sample conditions in existing solutions.

[0027] This invention addresses the unique mechanism of loose sand layers on thin sand pavements, where tire rolling causes sand particles to impact and move. Based on the acoustic signals generated by the interaction between the tire and the thin sand pavement, this invention identifies adhesion characteristics. The signal processing and identification logic are designed around the acoustic laws of particle impact on thin sand pavements. Unlike traditional physical interception identification methods for hard paved surfaces, this invention can accurately capture the acoustic signal differences corresponding to sand layers of different particle sizes, effectively distinguishing thin sand pavements with different adhesion states, and adapting to the specific identification needs of thin sand pavements.

[0028] To address the issues of high feature redundancy and low identification efficiency caused by directly extracting signals in existing solutions, this invention extracts multi-dimensional signal features from five dimensions: time-domain statistics, frequency-domain statistics, difference and derivative, envelope and morphology, and time-frequency sensing. This comprehensively covers the signal's time-domain morphological characteristics, frequency-domain structural characteristics, and transient change characteristics, and eliminates thousands of redundant feature dimensions through targeted feature filtering. This design effectively removes irrelevant and redundant features while fully preserving core discrimination information related to road surface adhesion characteristics, reducing the model's computational overhead, improving overall identification efficiency, and avoiding interference from redundant features on effective information, thus enhancing the discriminative power of the feature set.

[0029] This invention employs a collaborative design of multi-dimensional feature extraction and multi-time-step sequence construction. First, it extracts multi-dimensional features encompassing transient changes from a single-window signal, preserving the temporal and frequency domain details of a single impact process. Then, it organizes the signal features from multiple consecutive time windows into a multi-time-step sequence to characterize the evolution of the motion of gravel after contact with a tire under different thin sand pavement adhesion characteristics. This addresses the problem of insufficient utilization of signal time dependence in existing technologies. This design fully explores the temporal evolution of the gravel impact process, strengthens the temporal dependence between adjacent features, and enables the feature sequence to fully reflect the dynamic changes of gravel rolling and impacting under different adhesion states. It better aligns with the generation mechanism of acoustic signals on thin sand pavements, providing richer temporal dimension information for subsequent identification models and effectively improving the accuracy of adhesion characteristic identification.

[0030] This invention employs TimeGAN (Generative Adversarial Network) to achieve generative temporal data augmentation. Through a collaborative training architecture consisting of five parts—embedding network, recovery network, supervisor, generator, and discriminator—and combined optimization of reconstruction loss, supervisory loss, and adversarial loss, the model is trained, addressing the challenges of difficult and limited real-world thin sand pavement sample collection. This augmentation method differs from traditional data augmentation methods that only alter the statistical distribution; it simultaneously preserves the statistical distribution and temporal sequence of the signal, generating synthetic sequences that closely approximate real samples in both statistical distribution and temporal structure. This effectively expands the dataset size, compensates for the insufficient number of real sand samples, alleviates overfitting issues in identification models under small sample conditions, and significantly improves the model's generalization ability.

[0031] At the identification model level, the method of this invention employs a CNN-LSTM temporal neural network model. It utilizes convolutional neural network layers to extract short-term impact features within local time windows and long short-term memory network layers to capture long-range temporal dependencies across time steps. This hybrid architecture is compatible with the aforementioned multi-time-step sequence design, enabling it to fully learn the dynamic evolution characteristics of gravel impact sequences, accurately capture the acoustic signal differences corresponding to thin sand of different particle sizes, and further improve the identification accuracy of thin sand pavement adhesion characteristics.

[0032] Furthermore, this invention employs a combined processing strategy of bandpass filtering, differential noise reduction, short-time window segmentation, and amplitude normalization in the signal preprocessing stage, and further reduces the dimensionality of high-dimensional features using principal component analysis after feature extraction. Differential noise reduction uses appropriate noise reduction methods for different signal types, reducing waveform distortion while preserving key impact features; short-time window segmentation matches the duration of gravel impacts to avoid loss of impact information; amplitude normalization eliminates the influence of amplitude differences between different signal segments; and principal component analysis further compresses feature dimensions while retaining key discriminative information. This series of designs effectively purifies signal quality, reduces subsequent computational overhead, and further improves the training efficiency and operational stability of the generative and identification models. Attached Figure Description

[0033] Figure 1 A schematic diagram of the hardware platform construction;

[0034] Figure 2 This is a magnified view of a section of the test road surface;

[0035] Figure 3 The waveforms are shown in the time domain before and after filtering.

[0036] Figure 4 This is a structural diagram of the features;

[0037] Figure 5 Here is a flowchart illustrating the operating principle of TimeGAN;

[0038] Figure 6 This is a diagram of the CNN-LSTM network structure.

[0039] Figure 7 This is a schematic diagram of the prediction results before and after augmentation. Detailed Implementation

[0040] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0041] Example 1:

[0042] This invention discloses a method for identifying the adhesion characteristics of thin sand pavements. The thin sand pavement referred to in this invention is a pavement covered with a sand layer with a thickness of 2mm to 5mm. Since the thickness of the thin sand is much smaller than the tire contact depth, the tire remains in direct contact with the pavement substrate during rolling. Therefore, the tire-pavement interaction is mainly affected by the local distribution and particle size of the sand particles. The tire contact depth refers to the effective vertical contact depth between the tire tread rubber and the pavement substrate after compression under load when the tire tread contacts the pavement under normal driving conditions. The method of this invention specifically includes the following steps:

[0043] Acoustic signals generated by the interaction between the tire and thin sand pavement with different adhesion characteristics were collected, which can reflect the differences in adhesion characteristics of different thin sand pavement.

[0044] The acoustic signal is preprocessed and multi-dimensional signal features are extracted, which cover the signal’s temporal morphological characteristics, frequency domain structural characteristics and transient change characteristics.

[0045] The signal features of multiple consecutive time windows are organized into a multi-time step sequence to characterize the process of the movement of gravel after contact with tires under different thin sand pavement adhesion characteristics over time.

[0046] Generative time-series data augmentation is performed on multi-time-step sequences to simultaneously preserve the statistical distribution and temporal sequence relationship of the signals, resulting in an augmented dataset. Generative time-series data augmentation retains both the distribution fitting ability of GANs and the time dependency modeling ability of RNNs, enabling the generated sequences to approximate real samples in both statistical distribution and temporal structure.

[0047] Using the augmented dataset as input and the corresponding thin sand road surface adhesion characteristics as labels, a pre-built temporal neural network model is trained, and the trained temporal neural network model is used to identify the adhesion characteristics of thin sand road surfaces.

[0048] Furthermore, the acoustic signal (also known as the acoustic vibration signal or sound vibration signal) is acquired by an acoustic sensor. The acoustic sensor can be set at the bottom of the frame near the door of the vehicle near the tire, or it can be set inside the wheel arch or inside the fender. The location of the acoustic sensor is not limited to the typical locations mentioned above. Any location that can capture a clear acoustic signal without affecting the normal operation of the vehicle is a feasible implementation of the present invention.

[0049] Furthermore, the different adhesion characteristics of thin sand pavement can be distinguished by thin sand pavement with gravel of different particle sizes. Alternatively, different particle sizes can be further classified into specific categories such as coarse, medium, and fine, with the classified categories being equivalent to the different adhesion characteristics of thin sand pavement.

[0050] The following example illustrates the acquisition of acoustic signals:

[0051] This example first sets up and connects as follows: Figure 1 The hardware platform of the intelligent tire system shown was installed on the test vehicle, which then drove over test sand surfaces with five different particle sizes: 6-8 mesh, 16-26 mesh, 26-40 mesh, 40-70 mesh, and 100-200 mesh. These five particle sizes were then grouped into three categories: coarse, medium, and fine. Subsequently, a data acquisition program jointly developed using LabVIEW and DAQami was used to upload data and monitor signals in real time, acquiring a sound signal dataset.

[0052] Example 2:

[0053] This embodiment, based on Embodiment 1, further designs the preprocessing, which includes bandpass filtering, differential noise reduction, short-time window segmentation, and amplitude normalization. Bandpass filtering removes low-frequency structural vibrations and high-frequency environmental noise unrelated to the interaction between the tire and the thin sand pavement. Differential noise reduction addresses the different representations of the acoustic signal in the time and frequency domains. In the time domain, short-window median filtering extracts impact response variation features equivalent to structural vibrations. In the frequency domain, a slight spectral subtraction method is used to reduce waveform distortion while preserving key impact features, thereby improving the separability of tire-gravel contact events.

[0054] The purpose of the above design is to create a clean signal environment for subsequent feature extraction. It cannot perform signal differentiation or classification on its own; it only ensures that the useful impact modes in the signal are as complete as possible with minimal distortion, so that the subsequent extraction of time-domain, frequency-domain, MFCC, and other features can more accurately reflect the tire-sand contact characteristics. The main goal is to remove transient noise spikes or wind noise interference, while retaining the amplitude changes of the transient impact signal generated by tire-sand contact.

[0055] Mild spectral subtraction reduces environmental noise by subtracting the estimated background noise spectrum in the frequency domain. "Mild" indicates a limited noise reduction margin, aiming to minimize distortion of the impact signal's spectral characteristics. This invention employs a mild spectral subtraction algorithm for acoustic signal denoising, with the over-subtraction coefficient α set to 1.2 and the spectral base coefficient β set to 0.02. This results in a background noise suppression of approximately 3–6 dB, ensuring that the amplitude change of the main impact peak after denoising does not exceed 5%. This approach reduces environmental noise while preserving the transient time-frequency characteristics generated by tire-gravel contact to the greatest extent possible.

[0056] A short window means a very small filtering window to avoid smoothing out impact details; the short window segmentation matches the duration of the gravel impact; a median filter with a window length of 5 sampling points (approximately 1 ms) is applied to the acoustic time-domain vibration signal to suppress random spike noise while avoiding over-smoothing of impact edges. The denoised signal is then segmented into 60 ms short-time windows with a 50% window overlap. Since the duration of gravel impacts is typically less than 30 ms, a 60 ms analysis window can completely cover an impact event and ensure good temporal continuity between adjacent windows. Amplitude normalization eliminates the influence of amplitude differences between different signal segments.

[0057] Furthermore, the above method will be illustrated with an example below:

[0058] In this example, the acoustic signal is acquired using an acoustic sensor (or simply sensor). First, the acquired signal is cleaned. The sensor output is then filtered using a Butterworth bandpass filter in the 200–6000 Hz range, effectively suppressing structural vibrations (below 200 Hz) and high-frequency noise (above 6000 Hz). To mitigate transient spikes and wind noise interference, a short-window median filter is applied to the acoustic signal in the time domain, while mild spectral attenuation is used for noise reduction in the frequency domain. This preserves the impact morphology characteristics while minimizing waveform distortion. The time-domain waveforms before and after filtering are shown below. Figure 3 As shown in the figure, the original signal contains significant low-frequency fluctuations and random noise interference, with obvious baseline oscillations and large amplitude changes around 100 ms, 195 ms, and 290 ms. After bandpass filtering in the 200–6000 Hz range, the low-frequency trend term and high-frequency random noise are effectively suppressed, the signal fluctuates around zero, the baseline is more stable, and the main waveform corresponding to the impact event is still preserved.

[0059] The magnified view further illustrates the filtering effect. Within the 140–180 ms timeframe, the original signal exhibits significant low-frequency drift, making the true impact response easily masked by background fluctuations. After bandpass filtering, the slowly changing low-frequency components are effectively removed, resulting in a smoother waveform and a clearer local oscillation structure. This indicates that the filtering primarily preserves the effective vibration frequency components generated during the tire-gravel contact process without causing significant phase distortion or impact waveform distortion.

[0060] The filtered signal is divided into short time windows of 0.06 s and spliced ​​together with 50% overlap. Since the impact duration of sand and gravel is often less than 30 ms, using short windows and overlapping patterns can more fully capture the propagation process of the impact in the waveform and avoid information loss due to window boundaries. To reduce the impact of amplitude differences between different segments on subsequent feature statistics, this example performs RMS normalization on each segment, making the model focus on the pattern itself rather than amplitude differences.

[0061] ;

[0062] in, The normalized acoustic signal; This refers to the raw time-series signal (i.e., the coarse, raw acoustic signal directly imported from the sensor measurement). For acoustic signals The amplitude of each sampling point; This represents the signal length (number of sampling points).

[0063] Example 3:

[0064] This embodiment is a further design based on Embodiment 1. In this example, the multi-dimensional signal features are extracted from five dimensions: time domain statistics, frequency domain statistics, difference and derivative, envelope and morphology, and time-frequency perception. More than 1,000 redundant features can be eliminated through targeted feature screening.

[0065] The following example illustrates the extraction of multi-dimensional signal features. This example designs a multi-time-step engineering feature system for impact processes. Each window in this system extracts 166 dimensions describing the signal's temporal morphology, spectral structure, and transient changes. Continuous windows are organized into a multi-time-step sequence to reflect the true temporal evolution pattern during gravel rolling. The feature composition is as follows: Figure 4 As shown in Figure 1 and Table 1:

[0066] Table 1 Characteristic Composition Table

[0067]

[0068] Example 4:

[0069] This embodiment is a further design based on Embodiment 1. In this example, after extracting signal features, principal component analysis is used to reduce the dimensionality of high-dimensional features. While retaining the main discriminative information, the feature dimension is further reduced, thereby improving the training efficiency and stability of the generative model and the identification model.

[0070] Example 5:

[0071] This embodiment is further designed based on the first embodiment in that the multi-time step sequence in this example is composed of signal features of 10 consecutive time windows spliced ​​together in chronological order, with each time window having a duration of 0.06 seconds and an overlap rate of 50% between adjacent windows.

[0072] Example 6:

[0073] This embodiment is a further design based on Embodiment 1. In this example, generative time-series data augmentation is implemented using the TimeGAN temporal generative adversarial network. The TimeGAN temporal generative adversarial network retains the distribution fitting ability of GAN and the time dependency modeling ability of RNN. Unlike traditional data augmentation methods such as simple copying, translation, and adding noise, it enables the generated sequence to approximate the real sample in both statistical distribution and temporal structure, and makes up for the problem of the limited number of real sand and gravel samples.

[0074] Example 7:

[0075] This embodiment, based on Embodiment Six, further designs TimeGAN using a five-part collaborative training architecture consisting of an embedding network, a recovery network, a supervisor, a generator, and a discriminator. (TimeGAN employs a hybrid architecture of adversarial, self-supervised, and autoencoder loss.) Training is performed through joint optimization of the autoencoder reconstruction loss, supervised loss, and adversarial loss. The training process is as follows: Figure 5 As shown, these multiple constraints ensure that the model can generate high-quality synthetic sequences while maintaining the true temporal structure.

[0076] Furthermore, reconstruction loss This constraint is used to ensure that the Embedding network E and the Recovery network R form an invertible mapping, so that the hidden space representation can faithfully preserve the main features of the input sequence. During training, the real sequence X is first mapped to the hidden space representation through the constraint Embedding network E. Then reconstructed from R The reconstruction loss measures the difference between the two in terms of time and feature dimensions, and is defined as follows:

[0077] ;

[0078] in, For expectation calculation; This is a true sequence; To reconstruct the sequence. This loss ensures that the dynamic structure learned in the latent space is invertible, thus preventing the generator from learning pseudo-structures in the latent space that are unrelated to the real data.

[0079] Monitoring losses The core objective is to enhance the temporal dependence of the hidden space. Specifically, the Supervisor network S approximates the dynamic transitions of the true sequence by learning the next prediction of the true hidden vector:

[0080] ;

[0081] in, The loss is predicted by the Supervisor network. This loss ensures that the hidden space possesses the temporal smoothness and dynamic regularity of the real sequence, even when random noise is input during the generation stage. It can also generate elements with natural temporal variations. ; This represents the value of the true hidden vector at time t + 1; The prediction obtained by the Supervisor network S in The hidden vector at time step T is the total number of steps in the entire time series. When summing to calculate the loss, the calculation range is... step.

[0082] Combat losses This scenario involves an adversarial game between generator G and discriminator D. The discriminator is input with the true hidden vector. and generating hidden sequences It outputs the binary classification probability of "Real / Fake". The adversarial target is described by the following loss function:

[0083] ;

[0084] This indicates that the discriminator has a good understanding of the true hidden vector. The probability of it being judged as true; This indicates that the discriminator generates hidden sequences. The probability of it being judged as true;

[0085] Under this mechanism, the generator continuously improves the latent space distribution of the generated sequence to approximate the real data distribution; the discriminator continuously enhances its discrimination ability, and the two form a classic GAN adversarial process to improve the realism of the final generated sequence.

[0086] Based on the three types of losses mentioned above, the final training objective of TimeGAN can be written as a comprehensive optimization function:

[0087] ;

[0088] in, and Hyperparameters are used to balance the weights of each component.

[0089] Example 8:

[0090] This embodiment is a further design based on embodiment seven, in which the temporal neural network model is a CNN-LSTM model, wherein the convolutional neural network layer extracts short-term impact features within the local time window, and the long short-term memory network layer captures long-term temporal dependencies across time steps.

[0091] Furthermore, the CNN-LSTM model will be explained in detail below. The purpose of choosing the CNN-LSTM model is to improve the accuracy of road surface recognition by extracting local and periodic features of the signal. Using the CNN model for spatial expansion, the shared kernel parameters and sparsity of inter-layer connections within its hidden layers allow the convolutional neural network to perform gridding with relatively low computational cost, enabling it to extract local features from the data. However, its elements are independent, resulting in independent input and output. The LSTM model can perform temporal expansion processing on the data; the hidden layer input includes not only the input layer output but also the previous hidden layer output, thus the current data sequence output is also influenced by the previous output. Combining CNN and LSTM as a hybrid CNN-LSTM neural network for time series prediction allows the CNN part of the model to extract local trend features and common features that repeat at different intervals in the time series. For long sequence data, the CNN model suffers from gradient divergence. However, due to the presence of gate logic units, the LSTM model has a faster learning speed and higher accuracy. Combining the two models can integrate the advantages of each, improve the data training effect, and obtain high-dimensional features of the data. The convolutional layer first extracts short-term patterns (such as impact spikes, energy diffusion, etc.) within a local time window, and then the LSTM layer captures long-range dependencies across time steps, thereby learning the dynamic evolution characteristics of the gravel impact sequence.

[0092] The structure of the CNN-LSTM model used in this invention is as follows: Figure 6 Let the output of the convolutional layer be... Then, a one-dimensional convolution operation can be represented as:

[0093] ;

[0094] in, The kernel length is 1. Input the number of channels. It is a non-linear activation function. In the In a convolutional layer, the corresponding input channel convolution kernel position Weight parameters; At time step Input Channel Input values ​​on; No. The bias parameters corresponding to each convolutional layer. Convolutional layers are mainly responsible for extracting short-time patterns, such as impact peaks and energy diffusion within a local time window.

[0095] Based on the output of the convolutional layers, LSTM further models long-term temporal dependencies. Its gated update process is defined as follows:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] in, , , These are the forget gate, input gate, and output gate, respectively; It refers to the candidate cell state obtained after processing the input information. In cellular state, It is in a hidden state. It is the weight matrix of the forget gate. This indicates that the hidden state from the previous moment will be restored. and the output of the convolutional layer at the current time Piecing them together, It is the bias vector of the forget gate. It is the input gate weight matrix. It is the input gate bias vector. It is a weight matrix used to calculate the candidate cell state. It is the corresponding bias vector. It is the output gate weight matrix. It is the output gate bias vector. Current cell state.

[0103] Through this gating mechanism, LSTM can effectively retain dynamic information at key time points, thereby better describing the temporal evolution of the contact process between fine sand particles.

[0104] Example 9:

[0105] This embodiment is a further design based on embodiment eight. In this example, before training the CNN-LSTM model, the augmented dataset is divided into training set, validation set and test set to avoid data leakage caused by time slicing and ensure the reliability and reproducibility of the model evaluation results.

[0106] Example 10:

[0107] This embodiment is a further design based on Embodiment 1, in which at least five different particle sizes of thin sand pavement acoustic signals are collected and classified into three roughness levels: coarse, medium, and fine, according to engineering application requirements, each corresponding to different pavement adhesion characteristic ranges.

[0108] Application Examples:

[0109] This example verifies the feasibility of a tire-to-sand pavement adhesion characteristic identification method based on acoustic sensors. It investigates the influence of different gravel particle sizes on the acoustic characteristics of tire-to-pavement contact, designs a real-vehicle road test, collects tire-to-pavement acoustic signals from thin sand pavements of different particle sizes, and establishes corresponding datasets. Based on this, by constructing multi-timestep feature mapping, data augmentation, and classification models, it achieves the identification of thin sand pavements with different roughnesses and verifies the recognition performance of the proposed method under small sample conditions.

[0110] This example includes three sets of comparative tests. The first set compares different road surfaces: such as... Figure 2 The test sand pavement with five different particle sizes is shown, corresponding to mesh sizes of 6-8 mesh, 16-26 mesh, 26-40 mesh, 40-70 mesh, and 100-200 mesh. These five particle sizes are further categorized into three groups: coarse (6-8 mesh), medium (16-26 mesh, 26-40 mesh), and fine (40-70 mesh, 100-200 mesh). The second group compares the model accuracy before and after augmentation, aiming to verify whether TimeGAN augmentation can improve model performance. The third group compares different models, selecting Balanced Random Forest and CNN-LSTM neural networks for accuracy comparison to identify the superior classification model.

[0111] The acoustic sensor is mounted on the vehicle chassis, near the bottom of the frame below the right front tire door, approximately 5cm from the tire tread. The sensor is connected to a constant current adapter already connected to a micro-battery. The collected signal is displayed on the host computer software via a data acquisition card. The hardware connection platform is as follows: Figure 1 As shown.

[0112] The test was conducted in a closed road environment. To avoid the influence of external traffic noise and complex working conditions, the vehicle was kept at a constant speed of 30 km / h during the test, and data collection for each group was completed using the same tire pressure, load, and driving method. Gravel of different particle sizes was evenly laid on the test surface of the same length and thickness to ensure that different groups of tests had consistent testing conditions.

[0113] After acquiring the raw signal, preprocessing measures such as filtering are required. The acquired signal is cleaned, and the sensor output undergoes a Butterworth bandpass filter in the 200–6000 Hz range to effectively suppress structural vibrations (below 200 Hz) and high-frequency noise (above 6000 Hz). To mitigate transient spikes and wind noise interference, a short-window median filter is used in the time domain of the acoustic signal, while a slight spectral attenuation method is applied in the frequency domain to reduce noise. This preserves the impact morphology characteristics while minimizing waveform distortion. The time-domain waveforms before and after filtering are shown below. Figure 3 As shown.

[0114] The overall waveform shows that the original signal contains significant low-frequency fluctuations and random noise interference, with obvious baseline oscillations and large amplitude changes around 100 ms, 195 ms, and 290 ms. After bandpass filtering in the 200–6000 Hz range, the low-frequency trend term and high-frequency random noise are effectively suppressed, the signal fluctuates around zero, the baseline is more stable, and the main waveform corresponding to the impact event is still preserved.

[0115] The magnified view further illustrates the filtering effect. Within the 140–180 ms timeframe, the original signal exhibits significant low-frequency drift, making the true impact response easily masked by background fluctuations. After bandpass filtering, the slowly changing low-frequency components are effectively removed, resulting in a smoother waveform and a clearer local oscillation structure. This indicates that the filtering primarily preserves the effective vibration frequency components generated during the tire-gravel contact process without causing significant phase distortion or impact waveform distortion.

[0116] After obtaining the preprocessed data, this invention designs a multi-time-step engineering feature system for the impact process. In this system, each window extracts 166 dimensions of features describing the signal's temporal morphology, spectral structure, and transient changes. Continuous windows are organized into a multi-time-step sequence to reflect the true temporal evolution pattern during gravel rolling. The feature composition is as follows: Figure 4 As shown.

[0117] This invention employs a TimeGAN (Generative Adversarial Network) to compensate for the insufficient number of real sand samples, alleviate the overfitting problem of the identification model under small sample conditions, and significantly improve the model's generalization ability. This TimeGAN retains both the distribution fitting ability of GANs and the temporal dependency modeling ability of RNNs. Unlike traditional data augmentation methods such as simple copying, translation, and adding noise, it enables the generated sequences to approximate real samples in both statistical distribution and temporal structure, thus overcoming the limitation of the limited number of real sand samples.

[0118] The balanced RF and CNN-LSTM models were selected for a final accuracy comparison, and the comparison results are as follows: Figure 7 As shown in the figure, the accuracy of the balanced random forest (RF) before augmentation (65.02%) is higher than that of the CNN-LSTM (60.35%). After augmentation, the CNN-LSTM performs best, with an accuracy of 85.56%, which is higher than the 74.43% accuracy of the balanced RF. Under the condition of small sample size before augmentation, the balanced random forest can directly utilize the extracted tabular engineered features and obtain more robust generalization performance through balanced sampling and ensemble mechanisms; while the CNN-LSTM has a larger parameter scale and relies more on sufficient temporal sample coverage, making it difficult to learn stable temporal discrimination patterns when data is insufficient, thus performing worse than the balanced random forest. After introducing TimeGAN augmentation, the sample coverage and temporal variation are significantly enhanced, and the learning advantage of the CNN-LSTM for dynamic patterns can be brought into play, thus achieving the best performance.

[0119] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying the adhesion characteristics of thin sand pavement, characterized in that, include: Acoustic signals generated by the interaction between the tire and thin sand pavement with different adhesion characteristics were collected, which can reflect the differences in adhesion characteristics of different thin sand pavement. The acoustic signal is preprocessed and multi-dimensional signal features are extracted, which cover the temporal morphological characteristics, frequency domain structural characteristics and transient change characteristics of the signal; The signal features of multiple consecutive time windows are organized into a multi-time step sequence to characterize the process of the movement of gravel after contact with tires under different thin sand pavement adhesion characteristics over time. Generative time-series data augmentation is performed on multi-time-step sequences to simultaneously preserve the signal statistical distribution and temporal sequence relationship, resulting in an augmented dataset. Using the augmented dataset as input and the corresponding thin sand road surface adhesion characteristics as labels, a pre-constructed temporal neural network model is trained, and the trained temporal neural network model is used to identify the adhesion characteristics of thin sand road surfaces.

2. The method for identifying the adhesion characteristics of thin sand pavement according to claim 1, characterized in that, The preprocessing includes bandpass filtering, differential noise reduction, short-time window segmentation, and amplitude normalization. Bandpass filtering removes low-frequency structural vibrations and high-frequency environmental noise unrelated to the tire-gravel pavement interaction. Differential noise reduction addresses the different representations of the acoustic signal in the time and frequency domains. In the time domain, short-window median filtering extracts impact response variation features equivalent to structural vibrations. In the frequency domain, a slight spectral subtraction method is used to reduce waveform distortion while preserving key impact features, thereby improving the separability of tire-gravel contact events.

3. The method for identifying the adhesion characteristics of thin sand pavement according to claim 1, characterized in that, The multi-dimensional signal features are extracted from five dimensions: time-domain statistics, frequency-domain statistics, difference and derivative, envelope and morphology, and time-frequency perception.

4. The method for identifying the adhesion characteristics of thin sand pavement according to claim 1, characterized in that, After extracting signal features, principal component analysis is used to reduce the dimensionality of the high-dimensional features.

5. The method for identifying the adhesion characteristics of thin sand pavement according to claim 1, characterized in that, The multi-time-step sequence is composed of signal features from 10 consecutive time windows spliced ​​together in chronological order. Each time window has a duration of 0.06 seconds and the overlap rate between adjacent windows is 50%.

6. The method for identifying the adhesion characteristics of thin sand pavement according to claim 1, characterized in that, The generative temporal data augmentation is implemented using the TimeGAN temporal generative adversarial network.

7. The method for identifying the adhesion characteristics of thin sand pavement according to claim 6, characterized in that, The TimeGAN adopts an architecture that coordinates the training of five parts: an embedding network, a recovery network, a supervisor, a generator, and a discriminator. It is trained by jointly optimizing the reconstruction loss.

8. The method for identifying the adhesion characteristics of thin sand pavement according to claim 7, characterized in that, The temporal neural network model is a CNN-LSTM model, in which the convolutional neural network layer extracts short-term impact features within a local time window, and the long short-term memory network layer captures long-range temporal dependencies across time steps.

9. The method for identifying the adhesion characteristics of thin sand pavement according to claim 8, characterized in that, Before training the CNN-LSTM model, the augmented dataset is divided into training, validation, and test sets.

10. The method for identifying the adhesion characteristics of thin sand pavement according to claim 1, characterized in that, Acoustic signals from thin sand pavements with at least five different particle sizes were collected and classified into three roughness levels—coarse, medium, and fine—according to different pavement adhesion characteristic ranges, based on engineering application requirements.

Images