Roadbed compaction quality real-time evaluation method and system based on multi-domain signal processing

Abstract

The invention provides a roadbed compaction quality real-time evaluation method and system for multi-domain signal processing, and relates to the technical field of roadbed compaction detection, and the method comprises the steps: obtaining a vibration wheel acceleration signal, a roadbed vibration signal and a noise signal; obtaining a vibration wheel excitation signal according to the vibration wheel acceleration signal and the noise signal; according to the vibration wheel excitation signal and the roadbed vibration signal after noise reduction, obtaining a wave velocity dispersion characteristic and an attenuation coefficient; obtaining an energy-compactness quantitative relation according to the wave velocity dispersion characteristic and the attenuation coefficient; generating a depth feature vector according to a wave velocity dispersion characteristic and an energy-compactness quantitative relation; and outputting a real-time compaction degree evaluation result according to the depth feature vector. According to the invention, three-dimensional space global continuous sensing of a vibration wave field is realized through the hexagonal geophone array; a time-frequency-space three-domain cascade noise reduction framework is combined, so that the purity of a vibration signal is improved; based on an online incremental learning mechanism, adaptive modeling during working condition change is realized.

Description

Technical Field

[0001] This invention relates to the field of roadbed compaction detection technology, and more specifically, to a method and system for real-time evaluation of roadbed compaction quality using multi-domain signal processing. Background Technology

[0002] In highway and railway engineering construction, subgrade compaction is a crucial step in ensuring project quality, as its quality directly affects the road's load-bearing capacity, stability, and service life. Therefore, accurate testing of subgrade compaction quality is of paramount importance.

[0003] Currently, while roadbed compaction testing technology is continuously developing, several technical bottlenecks remain. Traditional destructive testing methods, such as sand cone and ring sampler methods, require on-site sampling and rely on laboratory analysis. These methods are not only cumbersome and inefficient, but also lack adaptability to coarse-grained soils and rockfill roadbeds. While non-destructive methods like nuclear density analyzers shorten testing time, their widespread application is hampered by radioactive risks. The intelligent compaction technology developed in recent years, although enabling process monitoring through positioning systems and sensors, still faces three core bottlenecks: reliance on a single sensor (vibration wheel accelerometer) only allows for localized vibration data acquisition, failing to capture the spatial distribution characteristics of vibration waves across the entire road section, resulting in localized blind spots in the test results; insufficient suppression of co-frequency noise interference, failing to address co-frequency mechanical vibration interference, leading to a low signal-to-noise ratio and signal distortion to some extent; and weak model generalization ability, with fixed parameter designs unable to adapt to changes in soil type and dynamic adjustments to working conditions, resulting in cross-condition failure.

[0004] Based on the shortcomings of the existing technologies, there is an urgent need for a real-time evaluation method and system for roadbed compaction quality using multi-domain signal processing. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for real-time evaluation of roadbed compaction quality using multi-domain signal processing, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0006] Firstly, this application provides a real-time evaluation method for roadbed compaction quality using multi-domain signal processing, including:

[0007] The system acquires the vibration wheel acceleration signal of the road roller, the roadbed vibration signal collected by a hexagonal array, and the noise signal, wherein the noise signal includes a reference noise signal and an ambient noise signal;

[0008] The vibration wheel excitation signal is obtained by performing joint noise reduction based on the vibration wheel acceleration signal and noise signal.

[0009] Based on the noise-reduced vibration wheel excitation signal and the roadbed vibration signal, calculations are performed to reconstruct the wave field propagation path and energy distribution by solving the wave equation, thereby obtaining the wave velocity dispersion characteristics and attenuation coefficient related to the compaction degree.

[0010] Based on the wave velocity dispersion characteristics and attenuation coefficient, energy attenuation analysis was performed to obtain a quantitative relationship between energy and compaction.

[0011] Multimodal feature fusion is performed based on the wave velocity dispersion characteristics and the energy-compaction degree quantitative relationship to generate a deep feature vector;

[0012] Based on the deep feature vector, the compaction degree is dynamically predicted, and the dynamic parameter library and network weights are updated through an online incremental learning mechanism to output the real-time compaction degree evaluation result.

[0013] Secondly, this application also provides a real-time evaluation system for roadbed compaction quality using multi-domain signal processing, comprising:

[0014] The acquisition module is used to acquire the vibration wheel acceleration signal of the road roller, the roadbed vibration signal collected by the hexagonal array, and the noise signal, wherein the noise signal includes a reference noise signal and an ambient noise signal;

[0015] The noise reduction module is used to perform joint noise reduction based on the acceleration signal and noise signal of the vibration wheel to obtain the noise-reduced excitation signal of the vibration wheel;

[0016] The calculation module is used to perform calculations based on the noise-reduced vibration wheel excitation signal and the roadbed vibration signal, and to reconstruct the wave field propagation path and energy distribution by solving the wave equation, thereby obtaining the wave velocity dispersion characteristics and attenuation coefficient related to the compaction degree.

[0017] The analysis module is used to perform energy attenuation analysis based on the wave velocity dispersion characteristics and attenuation coefficient to obtain a quantitative relationship between energy and compaction degree.

[0018] The fusion module is used to perform multimodal feature fusion based on the wave velocity dispersion characteristics and the energy-compaction degree quantitative relationship to generate a deep feature vector;

[0019] The prediction module is used to dynamically predict the compaction degree based on the depth feature vector, update the dynamic parameter library and network weights through an online incremental learning mechanism, and output the real-time compaction degree evaluation result.

[0020] The beneficial effects of this invention are as follows:

[0021] This invention achieves continuous three-dimensional spatial perception of vibration wave fields through a hexagonal seismic detector array, breaking through the limitations of traditional single-point sensing and eliminating blind spots in roadbed detection; combined with a time-frequency-space three-domain cascaded noise reduction architecture, it significantly improves the purity of effective vibration signals; and based on a physically constrained deep network and an online incremental learning mechanism, it achieves adaptive modeling under changing working conditions. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the process for a real-time evaluation method of roadbed compaction quality using multi-domain signal processing, as described in an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the structure of a real-time evaluation system for roadbed compaction quality using multi-domain signal processing, as described in an embodiment of the present invention.

[0025] Figure 3 A schematic diagram of a hexagonal close-packed sensor arrangement;

[0026] Figure 4 Flowchart of parameter identification algorithm;

[0027] Figure 5 Here is a flowchart of the frequency domain subspace filtering algorithm;

[0028] Figure 6 Flowchart for modeling wave field propagation;

[0029] Figure 7 Here is a flowchart of the parameter inversion algorithm;

[0030] Figure 8 A flowchart of a deep neural network architecture;

[0031] Figure 9 This is a flowchart of the deep learning algorithm for the model.

[0032] The diagram is labeled as follows: 901, Acquisition module; 902, Noise reduction module; 903, Calculation module; 904, Analysis module; 905, Fusion module; 906, Prediction module. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0034] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0035] Example 1:

[0036] This embodiment provides a method for real-time evaluation of roadbed compaction quality using multi-domain signal processing.

[0037] See Figure 1 The figure shows that the method includes steps S100 to S600.

[0038] Step S100: Acquire the vibration wheel acceleration signal of the road roller, the roadbed vibration signal and noise signal collected by the hexagonal array, the noise signal including the reference noise signal and the ambient noise signal;

[0039] Understandably, this step achieves three-dimensional full-domain sampling of the vibration wave field through the spatial topology deployment of a hexagonal seismic detector array, breaking through the dimensional limitations of traditional single-point sensing. It simultaneously collects two key signals: vibration wheel acceleration signal and reference noise (coherent mechanical sources such as engine / hydraulic pump) and environmental noise (incoherent interference such as wind noise), thus constructing a space-signal collaborative sensing network.

[0040] Specifically, three waterproof accelerometers (range ±50g, frequency response 0-1000Hz) are installed at a 120° angle on the outer surface of the vibrating wheel and fixed using high-temperature welding to ensure that the sampling direction is aligned with the tangent, radial, and vertical of the wheel axle; Figure 3As shown, seismic detectors are embedded in a hexagonal grid within the compacted area. The energy distribution of seismic waves generated during roadbed compaction propagates through the medium and is related to the direction of propagation. Unlike traditional single-point sensors, this invention employs a closely spaced hexagonal array to ensure wavefield sampling integrity across 360°, satisfying the spatial sampling theorem requirements. The spacing calculation formula is as follows:

[0041] ;

[0042] in, Indicates the spacing between adjacent sensors; This indicates the estimated minimum wave velocity (in m / s). Indicates the highest analysis frequency (in Hz); This indicates the thickness of the compacted layer (in meters).

[0043] The reference noise signal is obtained by installing accelerometers in the engine compartment and hydraulic pump of the road roller. The core purpose is to collect and eliminate the vibration noise generated by these mechanical components, thereby improving the signal-to-noise ratio of the roadbed compaction vibration signal and ensuring that the system can accurately identify the effective vibration waves and energy generated by the compactor's vibrating wheel.

[0044] During the roadbed compaction process, the effective vibration signal generated by the vibratory roller will be subject to the following interferences:

[0045] a. Engine vibration: periodic high-frequency vibration;

[0046] b. Hydraulic system shock: transient pulse noise;

[0047] c. Mechanical transmission: Vibration is transmitted through the frame to the vibrating wheel, contaminating the effective signal;

[0048] These noises overlap with the effective vibration signal frequency band, making them difficult to distinguish using traditional filtering methods, thus increasing the error in compaction degree detection. Engine vibration exhibits periodic and harmonic characteristics, with its fundamental frequency... By installing an accelerometer in the engine compartment, the vibration characteristics of rotating components such as the engine crankshaft and gears can be directly collected as a reference noise signal. Hydraulic shock has short-term high peak value (duration < 1 ms) and wide bandwidth characteristics (energy distribution up to 1 kHz or more). By installing an accelerometer in the hydraulic pump, the transient shock signal generated by pressure pulsation and valve core switching in the hydraulic system can be captured.

[0049] Ambient noise signals were collected by deploying ambient noise microphones around the construction area. The vibration noise generated by the vibrating wheel attenuates with distance; according to the inverse square law, the noise decreases with distance. However, at the construction site, other machinery (such as trucks) may generate noise from further away. Deploying microphones only in the near field may not effectively capture distant ambient noise, or the noise from the vibrating wheel itself may dominate. It is necessary to determine the location of the primary noise sources. If the ambient noise sources are far away, it may be necessary to adjust the microphone positions or deploy them at multiple distances to cover different noise sources. The arrangement of the ambient noise microphones is shown in Table 1.

[0050] Table 1. Relevant Arrangement of Ambient Noise Microphones

[0051] ;

[0052] Step S200: Perform joint noise reduction based on the vibration wheel acceleration signal and noise signal to obtain the noise-reduced vibration wheel excitation signal;

[0053] It should be noted that time-domain processing adaptively eliminates mechanical periodic interference by utilizing the coherence between the reference noise and the main signal; frequency-domain processing removes broadband random noise through subspace orthogonal projection; and spatial-domain processing combines the geometric advantages of the hexagonal array to inversely separate wave field aliasing components, ultimately outputting a clean excitation signal for the vibrating wheel.

[0054] Step S300: Calculate based on the noise-reduced vibration wheel excitation signal and the roadbed vibration signal, and reconstruct the wave field propagation path and energy distribution by solving the wave equation to obtain the wave velocity dispersion characteristics and attenuation coefficient related to the compaction degree.

[0055] Understandably, this step involves analyzing the processed clean signal to establish a propagation model of vibration waves in the layered roadbed medium. By solving the wave equation, the wave field propagation path and energy distribution are reconstructed, thereby determining the propagation characteristics of the vibration wave from the vibration source (vibration wheel) to each sensor, and establishing a quantitative relationship between compaction degree, wave velocity, and attenuation coefficient.

[0056] Step S400: Perform energy attenuation analysis based on wave velocity dispersion characteristics and attenuation coefficient to obtain a quantitative relationship between energy and compaction degree;

[0057] It should be noted that this step establishes an energy-compaction degree mapping relationship based on the attenuation characteristics of vibration wave energy with propagation distance, and evaluates the compaction uniformity and local defects by inverting the material damping coefficient through the energy attenuation rate.

[0058] Step S500: Perform multi-modal feature fusion based on wave velocity dispersion characteristics and energy-compaction quantitative relationship to generate a deep feature vector;

[0059] Understandably, this step generates deep feature vectors that are both interpretable and discriminative through feature fusion, overcoming the redundancy of manual features and the defects of nonlinear mapping.

[0060] Step S600: Dynamically predict compaction degree based on deep feature vectors, update dynamic parameter library and network weights through online incremental learning mechanism, and output real-time compaction degree evaluation results.

[0061] It should be noted that this step is based on the physical constraint network outputting the initial compaction distribution; when the soil conditions change, the weights of the fully connected layers are finely adjusted through hierarchical transfer learning, while the Lamé constant library is dynamically updated; combined with a timed feedback calibration mechanism, a closed-loop optimization is formed to achieve cross-scenario adaptation.

[0062] Further, step S200 includes steps S210 to S230.

[0063] Step S210: Perform time-domain coherent noise cancellation processing based on the vibration wheel acceleration signal and the reference noise signal. Separate the periodic noise components generated by the engine and hydraulic pump by calculating the correlation between the reference noise and the main signal and solving the noise transfer function to obtain the time-domain signal after mechanical noise cancellation.

[0064] Specifically, by utilizing the coherence between the reference noise signal (such as engine or hydraulic pump vibration) and the noise components in the main signal (vibrating wheel vibration), the noise transfer function is estimated through an adaptive filter, and the interference components related to the reference noise are eliminated from the main signal. This eliminates periodic or impact noise directly related to the mechanical equipment and improves the signal-to-noise ratio (SNR).

[0065] The noise component in the main signal can be represented as a linear combination of the reference noise:

[0066] ;

[0067] in, This indicates the received acceleration signal of the vibrating wheel; This indicates the acceleration signal of the cleanroom vibrating wheel; Indicates the first The transfer function from each noise source to the main sensor; Indicates the reference noise signal; This is a time variable, representing the current time point (unit: s), indicating the moment when the cleaning signal is being calculated; Let be the integration variable (unit: seconds), representing the time delay of the reference noise signal in the convolution operation. The integral ranges from 0 to 1. This indicates that all delays from historical time 0 to the current time t are considered; Index representing the noise source; This indicates the number of noise sources.

[0068] The goal is to adjust the transfer function This makes the noise estimation To approximate real noise, the main algorithm flow is as follows: Figure 4 As shown in the figure, noise_ref represents the reference noise signal (such as engine / hydraulic pump vibration), main_signal represents the noisy main signal (vibrating wheel acceleration signal); max_lag represents the maximum number of delay points (impulse response length); cross_corr represents the time delay correlation between the reference noise and the main signal, and the peak position reflects the noise propagation delay; max_lag=len(noise_ref)-1 indicates the use of the maximum possible delay, i.e., the reference signal length minus 1. This algorithm estimates the transfer function of a linear system using the least squares method. It uses the Wiener-Hoppt equation to solve for the transfer function based on the reference noise (noise_ref) and the main signal (main_signal). Its core principle is as follows:

[0069] a. Cross-correlation function: calculates the correlation between the reference noise and the main signal;

[0070] b. Autocorrelation function: calculates the autocorrelation of the reference itself;

[0071] c. Wiener-Hoppt equation: The transfer function is solved using the least squares method with the autocorrelation matrix and cross-correlation vector. Its core formula is as follows:

[0072] ;

[0073] in, Indicates the first The filter coefficients for the next iteration; Indicates the first The filter coefficients for the next iteration; Indicates the error signal. , is the filtered output signal, and is the estimated noise component; This represents the step size factor, which controls the convergence speed. This indicates a regularization term to prevent division by zero. This is a time index, representing the time offset of signal sampling;

[0074] d. Toeplitz matrix: The value of the autocorrelation function is used to construct a Toeplitz matrix, which has a special structure suitable for solving linear equations.

[0075] Step S220: Perform frequency domain random noise suppression processing on the time domain signal, and filter out non-coherent broadband interference by separating the signal-dominant subspace components through feature decomposition to obtain a frequency domain clean signal;

[0076] Specifically, the signal is converted to the frequency domain, and the differences in the subspace distribution of signal and noise are utilized to separate the signal-dominated subspace component through feature decomposition, suppressing the noise subspace and eliminating incoherent interference such as broadband random noise (e.g., ambient wind noise, people walking). The overall process is handled by algorithms (such as...) Figure 5 The implementation (as shown) is carried out, and the relevant theoretical formulas are derived as follows:

[0077] The signal is divided into frames (typically 256ms / frame) and the STFT (Short Time Fourier Transform) is calculated to construct the time-frequency matrix:

[0078] ;

[0079] in, Represents the time-frequency matrix; For frequency index, indicating the number of... One frequency point; For time frame index, indicating the first... One time frame; Indicates the first Frame number The complex spectrum of a frequency point.

[0080] For the framed signal The covariance matrix is:

[0081] ;

[0082] in, This represents the transpose of a matrix.

[0083] For covariance matrix Perform eigenvalue decomposition Retain the eigenvectors corresponding to the first m largest eigenvalues. The pure signal is reconstructed into

[0084] ;

[0085] The rank k is determined by the proportion of eigenvalue energy:

[0086] ;

[0087] in, The eigenvalues ​​represent the features arranged in descending order; Indicates the first The eigenvectors corresponding to each eigenvalue.

[0088] By performing time-domain processing, strongly correlated mechanical noise (such as that from engines and hydraulic pumps) is eliminated, effectively improving the signal-to-noise ratio (SNR). Then, by performing frequency-domain processing, residual random noise (such as ambient wind noise) is further suppressed, further improving the SNR. The two processes work synergistically to achieve the desired noise reduction effect.

[0089] Step S230: Perform spatial wave field separation based on the frequency domain pure signal, and separate the aliased signal through the wave field inverse algorithm to obtain the noise-reduced vibration wheel excitation signal.

[0090] Furthermore, in the spatial domain, the aliased signals are separated using a wave field inverse algorithm, ultimately achieving a significant improvement in the signal-to-noise ratio.

[0091] Further, step S300 includes steps S310 to S330.

[0092] Step S310: Based on the noise-reduced vibration wheel excitation signal and the roadbed vibration signal, perform layered medium wave modeling processing. By establishing the elastic dynamic control equation of the layered roadbed and discretizing the spatial domain, the layered medium wave model is obtained.

[0093] Step S320: Perform time-domain wave field propagation calculation based on the layered medium wave model, and use the implicit time integration method to simultaneously calculate the vibration wave propagation path and energy attenuation process to obtain the dynamic wave field distribution characteristics.

[0094] Step S330: Perform compaction parameter inversion processing based on dynamic wave field distribution characteristics. By analyzing the wave velocity dispersion curve and attenuation gradient variation law, obtain the wave velocity dispersion characteristics and attenuation coefficient directly related to the degree of compaction.

[0095] Specifically, based on the theory of elastic dynamics, a propagation model of vibration waves in layered roadbed media is established. By solving the wave equation, the wave field propagation path and energy distribution are reconstructed, thereby determining the propagation characteristics of the vibration wave from the vibration source (vibration wheel) to each sensor. A quantitative relationship between compaction degree, wave velocity, and attenuation coefficient is established. The overall process is as follows: Figure 6 As shown, its core dynamic equations are as follows:

[0096] ;

[0097] in, This indicates the density of the material and is related to its compaction degree. ; Lamé constant represents the stiffness of a material. ; The term representing the vibration source is obtained by integrating the acceleration of the vibrating wheel. Represents the medium displacement vector (unit: m); Represents depth coordinates (unit: m); Indicates radial distance (unit: m).

[0098] The numerical solution method is as follows:

[0099] a. Spatial discretization: The roadbed is discretized in layers using the finite difference method (FDM);

[0100] b. Time integration: Newmark-β method is used (β=0.25 to ensure numerical stability);

[0101] c. Boundary conditions: Free boundary on the ground: The bottom boundary is suppressed using a PML layer.

[0102] Further, step S400 includes steps S410 to S430.

[0103] Step S410: Extract and process vibration energy observations based on wave velocity dispersion characteristics and attenuation coefficients. Calculate vibration energy density at different propagation distances using data from hexagonal array sensors. Correct geometric attenuation effects based on wave velocity dispersion characteristics to obtain spatially distributed vibration energy observations.

[0104] Step S420: Based on the observed values ​​of vibration energy in spatial distribution, a physical model for attenuation is constructed. By establishing an energy flow equation that includes material damping attenuation and geometric diffusion effects, a parameterized energy attenuation model is obtained.

[0105] Step S430: Perform material parameter inversion processing based on the parameterized energy decay model. By minimizing the difference between the observed energy and the model prediction, solve for the optimal combination of decay parameters and establish a quantitative relationship between energy and compaction.

[0106] Specifically, an energy-compaction degree mapping relationship is established based on the attenuation characteristics of vibration wave energy with propagation distance. Energy attenuation is jointly determined by material damping and geometric diffusion. The material damping coefficient is inverted by the energy attenuation rate to evaluate compaction uniformity and local defects.

[0107] Energy flow model:

[0108] ;

[0109] in, The initial energy of the vibration source is represented by the integral of the acceleration of the vibrating wheel; This represents the material attenuation coefficient, reflecting the material's damping characteristics, and is negatively correlated with compaction degree (unit: ); Represents the natural base; Indicates the geometric decay exponent (spherical wave) =2, cylindrical wave =1); This represents the interference modulation factor, reflecting the layered interface. Indicates the wavelength of the vibration wave; This indicates the propagation distance (the straight-line distance from the vibration source to the sensor).

[0110] like Figure 7 As shown, the specific process of the parameter inversion algorithm is as follows:

[0111] a. Multi-source data acquisition: Acceleration signals are acquired via a vibrating wheel accelerometer. Hexagonal array sensors collect roadbed vibration signals. RTK-GPS acquires the coordinates of the vibration source and sensors, etc.

[0112] b. Observational energy extraction and correction: Calculate the energy for each sensor's data. To eliminate transient interference and obtain reliable energy observations;

[0113] c. Problem Formulation for Optimization: Define a loss function to measure the difference between predicted and actual measured values, reducing the impact of measurement errors and local disturbances.

[0114] ;

[0115] in, Representing the energy model exist The predicted value at that location.

[0116] d. Constrained optimization solution: The L-BFGS-B algorithm is used to solve the nonlinear least squares problem and output the optimal set of parameters.

[0117] Further, step S500 includes steps S510 to S530.

[0118] Step S510: Based on the wave velocity dispersion characteristics, perform waveform time-frequency feature extraction processing, and capture the local correlation and frequency-varying propagation characteristics of the vibration signal through a multi-scale gated convolutional layer to generate waveform depth features;

[0119] Step S520: Based on the quantitative relationship between energy and compaction, perform energy parameter stability enhancement processing, eliminate dimensional differences of material parameters through spectral normalization, and extract energy parameter enhancement characteristics;

[0120] Step S530: Perform gated fusion processing based on waveform depth features and energy parameter enhancement features, and dynamically combine wave characteristics and energy characteristics through an adaptive weight allocation mechanism to generate a multimodal depth feature vector representing the compaction state.

[0121] Specifically, by integrating the waveform and energy characteristics of vibration signals, accurate prediction of compaction degree is achieved through multimodal deep learning. Key time-frequency domain features are extracted from the vibration signals, and a nonlinear mapping relationship between multi-sensor data and compaction degree is established. The deep neural network architecture flowchart is shown below. Figure 8 As shown.

[0122] The core module design is as follows:

[0123] Waveform feature branches

[0124] Convolutional Neural Networks (CNNs) are used to extract features from waveform data through convolutional layers, normalization, activation functions, and pooling layers. Convolutional layers capture local features in the waveform, while pooling layers reduce the feature dimensionality by flattening the extracted features into a two-dimensional tensor.

[0125] Gated convolutional layers:

[0126] ;

[0127] in, This represents the output feature map of the i-th layer; Indicates the gated convolution kernel and the standard convolution kernel; This represents the Sigmoid function, which controls the flow of information. Let represent the hyperbolic tangent function, which normalizes the eigenvalues ​​to [-1, 1] and stabilizes the gradient.

[0128] Multi-Scale Attention:

[0129] We use convolutional kernels of different scales (3, 5, 7) in parallel to weighted fuse multi-resolution features.

[0130] Energy characteristic branch:

[0131] Spectral Normalized Fully Connected Layer (Spectral Norm FC) is used to process energy parameters, which are usually physical quantities associated with waveforms. This can prevent overfitting and enhance robustness to input energy parameters (such as attenuation coefficient and quality factor).

[0132] Feature fusion layer:

[0133] A gated fusion mechanism is adopted to stitch together waveform and energy features, which are then fused and decided upon through a fully connected layer, ultimately outputting the compaction percentage.

[0134] ;

[0135] in, Indicates the percentage of compaction degree; This represents the fusion weight matrix, which learns the correlation between two types of features; The feature vector representing the waveform branch output; This represents the feature vector output by the energy branch.

[0136] Further, step S600 includes steps S610 to S630.

[0137] Step S610: Perform initial prediction processing of compaction state based on depth feature vector, and generate preliminary compaction distribution values ​​for each area of ​​the subgrade through forward propagation calculation of physical constraint deep neural network.

[0138] Step S620: Based on the preliminary compaction distribution value, perform incremental update processing of model parameters, selectively update the weights of the fully connected layers of the network through a hierarchical transfer learning strategy, and adaptively correct the dynamic parameter library in combination with changes in material properties to obtain the updated prediction model.

[0139] Step S630: Perform compaction quality assessment based on the updated prediction model, calibrate the model prediction accuracy through a real-time feedback closed-loop mechanism, and output the subgrade compaction assessment results.

[0140] Specifically, this process utilizes online learning technology, an incremental learning method that updates model parameters using only one or a small number of samples at a time, rather than using the entire dataset all at once (batch learning). Therefore, an online update function needs to be designed for incremental training of the deep learning model. This function receives new data samples and updates the model's parameters using a stochastic gradient descent (SGD) optimizer.

[0141] Preferably, the flowchart of the deep learning algorithm for the model is as follows: Figure 9 As shown. The prediction error is minimized by calculating the gradient of the loss function and adjusting the model parameters according to the learning rate. The mean squared error is used as the loss function to measure the difference between the model's predictions and the true values. The mixed loss function is shown below:

[0142] Main loss:

[0143] ;

[0144] Parameter Explanation

[0145] in, This represents the mean squared error loss, which measures the average squared difference between the model's predicted values ​​and the actual values. This indicates the batch size, which is the number of samples used in the current loss calculation. Indicates the first The true target value of each sample; Indicates the first The predicted value for each sample.

[0146] Gradient matching loss:

[0147] ;

[0148] Parameter Explanation

[0149] in, This represents the gradient matching loss value; Equivalent to , indicating the first The true target value of each sample; Indicates the predicted gradient. Represents the actual physical gradient.

[0150] Total loss:

[0151] ;

[0152] Parameter Explanation

[0153] in, This represents the total loss value that needs to be minimized during model training. This represents the weighting coefficient, which controls the relative importance of gradient matching loss in the total loss.

[0154] For adjusting the dynamic learning rate, the initial learning rate is set to... Then, the validation loss is adaptively adjusted:

[0155] ;

[0156] Parameter Explanation

[0157] in, This represents the learning rate at the current time step; This indicates the learning rate to be used in the next time step; The loss value calculated on the validation set represents the model's generalization ability and is a key indicator for judging whether the model is overfitting.

[0158] Example 2:

[0159] like Figure 2 As shown in the figure, this embodiment provides a real-time evaluation system for roadbed compaction quality using multi-domain signal processing. The system includes:

[0160] The acquisition module 901 is used to acquire the vibration wheel acceleration signal of the road roller, the roadbed vibration signal collected by the hexagonal array, and the noise signal, including the reference noise signal and the ambient noise signal.

[0161] The noise reduction module 902 is used to perform joint noise reduction based on the vibration wheel acceleration signal and the noise signal to obtain the noise-reduced vibration wheel excitation signal.

[0162] The calculation module 903 is used to perform calculations based on the noise-reduced vibration wheel excitation signal and the roadbed vibration signal. By solving the wave equation, the wave field propagation path and energy distribution are reconstructed, and the wave velocity dispersion characteristics and attenuation coefficient related to the compaction degree are obtained.

[0163] Analysis module 904 is used to perform energy attenuation analysis based on wave velocity dispersion characteristics and attenuation coefficient to obtain a quantitative relationship between energy and compaction degree;

[0164] The fusion module 905 is used to perform multimodal feature fusion based on wave velocity dispersion characteristics and energy-compaction quantitative relationship to generate a deep feature vector.

[0165] The prediction module 906 is used to dynamically predict compaction degree based on the deep feature vector. It updates the dynamic parameter library and network weights through an online incremental learning mechanism and outputs real-time compaction degree evaluation results.

[0166] In one specific embodiment of the present invention, the noise reduction module 902 includes:

[0167] The first noise reduction unit is used to perform time-domain coherent noise cancellation processing based on the vibration wheel acceleration signal and the reference noise signal. It separates the periodic noise components generated by the engine and hydraulic pump by calculating the correlation between the reference noise and the main signal and solving the noise transfer function, and obtains the time-domain signal after mechanical noise cancellation.

[0168] The second noise reduction unit is used to perform frequency domain random noise suppression processing based on the time domain signal. It separates the signal-dominant subspace components through feature decomposition to filter out non-coherent broadband interference and obtain a frequency domain clean signal.

[0169] The third noise reduction unit is used to separate the spatial wave field based on the frequency domain pure signal, and to separate the aliased signal through the wave field inverse algorithm to obtain the noise-reduced vibration wheel excitation signal.

[0170] In one specific embodiment of the present invention, the calculation module 903 includes:

[0171] The first calculation unit is used to perform layered medium wave modeling based on the noise-reduced vibration wheel excitation signal and the roadbed vibration signal. By establishing the elastic dynamic control equation of the layered roadbed and discretizing the spatial domain, the layered medium wave model is obtained.

[0172] The second calculation unit is used to perform time-domain wave field propagation calculation based on the layered medium wave model. It uses the implicit time integration method to simultaneously calculate the vibration wave propagation path and energy attenuation process to obtain the dynamic wave field distribution characteristics.

[0173] The third calculation unit is used to perform compaction parameter inversion processing based on the dynamic wave field distribution characteristics. By analyzing the wave velocity dispersion curve and the attenuation gradient variation law, the wave velocity dispersion characteristics and attenuation coefficient directly related to the degree of compaction are obtained.

[0174] In one specific embodiment of the present invention, the analysis module 904 includes:

[0175] The first analysis unit is used to extract and process vibration energy observations based on wave velocity dispersion characteristics and attenuation coefficients. It calculates vibration energy density at different propagation distances using data from a hexagonal array sensor and corrects for geometric attenuation effects based on wave velocity dispersion characteristics to obtain spatially distributed vibration energy observations.

[0176] The second analysis unit is used to construct a physical model of attenuation based on the observed values ​​of vibration energy in spatial distribution. By establishing an energy flow equation that includes material damping attenuation and geometric diffusion effects, a parameterized energy attenuation model is obtained.

[0177] The third analysis unit is used to perform material parameter inversion processing based on the parameterized energy decay model. By minimizing the difference between the observed energy and the model prediction, it solves for the optimal combination of decay parameters and establishes a quantitative relationship between energy and compaction.

[0178] In one specific embodiment of the present invention, the fusion module 905 includes:

[0179] The first fusion unit is used to extract waveform time-frequency features based on wave velocity dispersion characteristics. It captures the local correlation and frequency-varying propagation characteristics of vibration signals through multi-scale gated convolutional layers to generate waveform depth features.

[0180] The second fusion unit is used to enhance the stability of energy parameters based on the quantitative relationship between energy and compaction. It eliminates the dimensional differences of material parameters through spectral normalization and extracts the energy parameter enhancement features.

[0181] The third fusion unit is used to perform gated fusion processing based on waveform depth features and energy parameter enhancement features. It dynamically combines wave characteristics and energy characteristics through an adaptive weight allocation mechanism to generate a multimodal depth feature vector representing the compaction state.

[0182] In one specific embodiment of the present invention, the prediction module 906 includes:

[0183] The first prediction unit is used to perform initial prediction processing of compaction state based on the depth feature vector, and to generate the preliminary compaction degree distribution values ​​of each area of ​​the subgrade through forward propagation calculation of the physical constraint deep neural network.

[0184] The second prediction unit is used to incrementally update the model parameters based on the initial compaction distribution value. It selectively updates the weights of the fully connected layers of the network through a hierarchical transfer learning strategy and adaptively corrects the dynamic parameter library in combination with changes in material properties to obtain the updated prediction model.

[0185] The third prediction unit is used to perform compaction quality assessment based on the updated prediction model, calibrate the model prediction accuracy through a real-time feedback closed-loop mechanism, and output the subgrade compaction assessment results.

[0186] 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 changes 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 real-time evaluation of roadbed compaction quality using multi-domain signal processing, characterized in that, include: The system acquires the vibration wheel acceleration signal of the road roller, the roadbed vibration signal collected by a hexagonal array, and the noise signal, wherein the noise signal includes a reference noise signal and an ambient noise signal; The vibration wheel excitation signal is obtained by performing joint noise reduction based on the vibration wheel acceleration signal and noise signal. Based on the noise-reduced vibration wheel excitation signal and the roadbed vibration signal, calculations are performed to reconstruct the wave field propagation path and energy distribution by solving the wave equation, thereby obtaining the wave velocity dispersion characteristics and attenuation coefficient related to the compaction degree. Based on the wave velocity dispersion characteristics and attenuation coefficient, energy attenuation analysis was performed to obtain a quantitative relationship between energy and compaction. Multimodal feature fusion is performed based on the wave velocity dispersion characteristics and the energy-compaction degree quantitative relationship to generate a deep feature vector; Based on the deep feature vector, the compaction degree is dynamically predicted, and the dynamic parameter library and network weights are updated through an online incremental learning mechanism to output the real-time compaction degree evaluation result.

2. The method for real-time evaluation of roadbed compaction quality using multi-domain signal processing according to claim 1, characterized in that, The vibration wheel excitation signal is obtained by jointly denoising the acceleration signal and noise signal of the vibration wheel, including: Based on the acceleration signal of the vibrating wheel and the reference noise signal, time-domain coherent noise cancellation processing is performed. By calculating the correlation between the reference noise and the main signal and solving the noise transfer function, the periodic noise components generated by the engine and hydraulic pump are separated to obtain the time-domain signal after mechanical noise cancellation. Based on the time-domain signal, frequency-domain random noise suppression processing is performed, and non-coherent broadband interference is filtered out by separating the signal-dominant subspace components through feature decomposition to obtain a frequency-domain clean signal. Based on the frequency domain pure signal, spatial wave field separation is performed, and the aliased signal is separated by the wave field inverse algorithm to obtain the noise-reduced vibration wheel excitation signal.

3. The method for real-time evaluation of roadbed compaction quality using multi-domain signal processing according to claim 1, characterized in that, Based on the noise-reduced vibration wheel excitation signal and the roadbed vibration signal, calculations are performed to reconstruct the wave field propagation path and energy distribution by solving the wave equation, thereby obtaining the wave velocity dispersion characteristics and attenuation coefficient related to compaction degree, including: Based on the noise-reduced vibration wheel excitation signal and the roadbed vibration signal, a layered medium wave modeling process is performed. By establishing the elastic dynamic control equation of the layered roadbed and discretizing the spatial domain, a layered medium wave model is obtained. Based on the layered medium wave model, time-domain wave field propagation is calculated, and the implicit time integration method is used to simultaneously calculate the vibration wave propagation path and energy attenuation process to obtain the dynamic wave field distribution characteristics. Based on the dynamic wave field distribution characteristics, compaction parameters are inverted. By analyzing the wave velocity dispersion curve and the attenuation gradient variation law, the wave velocity dispersion characteristics and attenuation coefficient directly related to the degree of compaction are obtained.

4. The method for real-time evaluation of roadbed compaction quality using multi-domain signal processing according to claim 1, characterized in that, Based on the wave velocity dispersion characteristics and attenuation coefficient, energy attenuation analysis is performed to obtain a quantitative relationship between energy and compaction degree, including: The vibration energy observation values ​​are extracted and processed based on the wave velocity dispersion characteristics and attenuation coefficient. The vibration energy density at different propagation distances is calculated using data from a hexagonal array sensor. The geometric attenuation effect is corrected based on the wave velocity dispersion characteristics to obtain the spatially distributed vibration energy observation values. Based on the observed vibration energy values ​​of the spatial distribution, a physical model for attenuation is constructed. By establishing an energy flow equation that includes material damping attenuation and geometric diffusion effects, a parameterized energy attenuation model is obtained. Material parameter inversion is performed based on the parameterized energy decay model. By minimizing the difference between the observed energy and the model prediction, the optimal combination of decay parameters is solved and a quantitative relationship between energy and compaction degree is established.

5. The method for real-time evaluation of roadbed compaction quality using multi-domain signal processing according to claim 1, characterized in that, Based on the wave velocity dispersion characteristics and the energy-compaction degree quantitative relationship, multimodal feature fusion is performed to generate a deep feature vector, including: Based on the wave velocity dispersion characteristics, waveform time-frequency features are extracted, and the local correlation and frequency-varying propagation characteristics of the vibration signal are captured by a multi-scale gated convolutional layer to generate waveform depth features. Based on the quantitative relationship between energy and compaction, energy parameter stability enhancement processing is performed. The dimensional differences of material parameters are eliminated by spectral normalization, and the energy parameter enhancement characteristics are extracted. Gated fusion processing is performed based on the waveform depth features and the energy parameter enhancement features. A multimodal depth feature vector representing the compaction state is generated by dynamically combining the wave characteristics and energy characteristics through an adaptive weight allocation mechanism.

6. A real-time evaluation system for roadbed compaction quality using multi-domain signal processing, characterized in that, include: The acquisition module is used to acquire the vibration wheel acceleration signal of the road roller, the roadbed vibration signal collected by the hexagonal array, and the noise signal, wherein the noise signal includes a reference noise signal and an ambient noise signal; The noise reduction module is used to perform joint noise reduction based on the acceleration signal and noise signal of the vibration wheel to obtain the noise-reduced excitation signal of the vibration wheel; The calculation module is used to perform calculations based on the noise-reduced vibration wheel excitation signal and the roadbed vibration signal, and to reconstruct the wave field propagation path and energy distribution by solving the wave equation, thereby obtaining the wave velocity dispersion characteristics and attenuation coefficient related to the compaction degree. The analysis module is used to perform energy attenuation analysis based on the wave velocity dispersion characteristics and attenuation coefficient to obtain a quantitative relationship between energy and compaction degree. The fusion module is used to perform multimodal feature fusion based on the wave velocity dispersion characteristics and the energy-compaction degree quantitative relationship to generate a deep feature vector; The prediction module is used to dynamically predict the compaction degree based on the depth feature vector, update the dynamic parameter library and network weights through an online incremental learning mechanism, and output the real-time compaction degree evaluation result.

7. The real-time evaluation system for roadbed compaction quality using multi-domain signal processing according to claim 6, characterized in that, The noise reduction module includes: The first noise reduction unit is used to perform time-domain coherent noise cancellation processing based on the vibration wheel acceleration signal and the reference noise signal. By calculating the correlation between the reference noise and the main signal and solving the noise transfer function, the periodic noise components generated by the engine and hydraulic pump are separated to obtain the time-domain signal after mechanical noise cancellation. The second noise reduction unit is used to perform frequency domain random noise suppression processing on the time domain signal, and filter out non-coherent broadband interference by separating the signal-dominant subspace components through feature decomposition to obtain a frequency domain clean signal. The third noise reduction unit is used to perform spatial wave field separation based on the frequency domain pure signal, and to separate the aliased signal through the wave field inverse algorithm to obtain the noise-reduced vibration wheel excitation signal.

8. The real-time evaluation system for roadbed compaction quality using multi-domain signal processing according to claim 6, characterized in that, The computing module includes: The first calculation unit is used to perform layered medium wave modeling processing based on the noise-reduced vibration wheel excitation signal and the roadbed vibration signal. By establishing the elastic dynamic control equation of the layered roadbed and discretizing the spatial domain, the layered medium wave model is obtained. The second calculation unit is used to perform time-domain wave field propagation calculation based on the layered medium wave model. It uses the implicit time integration method to simultaneously calculate the vibration wave propagation path and energy attenuation process to obtain the dynamic wave field distribution characteristics. The third calculation unit is used to perform compaction parameter inversion processing based on the dynamic wave field distribution characteristics. By analyzing the wave velocity dispersion curve and the attenuation gradient variation law, the wave velocity dispersion characteristics and attenuation coefficient directly related to the compaction degree are obtained.

9. A real-time evaluation system for roadbed compaction quality using multi-domain signal processing according to claim 6, characterized in that, The analysis module includes: The first analysis unit is used to extract and process vibration energy observations based on the wave velocity dispersion characteristics and attenuation coefficients, calculate vibration energy density at different propagation distances using data from hexagonal array sensors, and correct for geometric attenuation effects based on wave velocity dispersion characteristics to obtain spatially distributed vibration energy observations. The second analysis unit is used to construct a physical model of attenuation based on the observed vibration energy values ​​of the spatial distribution. By establishing an energy flow equation that includes material damping attenuation and geometric diffusion effects, a parameterized energy attenuation model is obtained. The third analysis unit is used to perform material parameter inversion processing based on the parameterized energy decay model. By minimizing the difference between the observed energy and the model prediction, it solves for the optimal combination of decay parameters and establishes a quantitative relationship between energy and compaction.

10. A real-time evaluation system for roadbed compaction quality using multi-domain signal processing according to claim 6, characterized in that, The fusion module includes: The first fusion unit is used to perform waveform time-frequency feature extraction processing based on the wave velocity dispersion characteristics, and to capture the local correlation and frequency-varying propagation characteristics of the vibration signal through a multi-scale gated convolutional layer to generate waveform depth features. The second fusion unit is used to perform energy parameter stability enhancement processing based on the energy-compaction degree quantitative relationship, eliminate the dimensional differences of material parameters through spectrum normalization, and extract the energy parameter enhancement features. The third fusion unit is used to perform gated fusion processing based on the waveform depth features and the energy parameter enhancement features, and dynamically combines the wave characteristics and energy characteristics through an adaptive weight allocation mechanism to generate a multimodal depth feature vector representing the compaction state.

Images