Pavement disease detection method and system based on vibration signals

Abstract

The invention discloses a pavement disease detection method and system based on vibration signals, and relates to the technical field of pavement engineering detection. The method comprises the steps that a system is initialized, detection equipment is calibrated, and a preset multi-scale SVM-Transform hybrid recognition model, a disease evolution trend prediction model and a cross-module linkage control center are loaded; based on a disease vibration propagation model and a targeted acquisition mechanism adapted to an evolution stage, acquiring a road surface vibration signal and corresponding acquisition position information, and outputting vibration original data and position data; and preprocessing the original vibration data by adopting a time domain and frequency domain linkage adaptive filtering algorithm, and outputting a preprocessed signal. According to the method, comprehensive capturing and accurate distinguishing of the diseases are achieved, the signal processing quality and recognition reliability are improved, the method is adaptive to the diseases in different evolution stages and various pavement types, comprehensive and practical reference information is provided for pavement maintenance, and scientific maintenance decision making is assisted.

Description

Technical Field

[0001] This invention relates to the field of pavement engineering inspection technology, specifically to a method and system for detecting pavement defects based on vibration signals. Background Technology

[0002] Road surface distress detection is a crucial aspect of transportation infrastructure maintenance. Accurately identifying the type, location, and evolution of distress is essential for ensuring traffic safety, reducing maintenance costs, and extending road surface lifespan. Vibration signal-based detection methods have become one of the mainstream technologies for road surface distress detection due to their ease of operation and high efficiency. The core idea is to collect road surface vibration signals through sensors, and then detect distress through signal preprocessing, feature extraction, and model recognition.

[0003] Existing methods for detecting pavement defects based on vibration signals mostly employ fixed acquisition parameters, filtering strategies, and feature weights. They fail to dynamically adjust key parameters based on the vibration propagation characteristics at different stages of defect evolution, and lack deep inter-module linkage mechanisms. This results in current technologies struggling to accurately capture characteristic signals of defects at different evolution stages, exhibiting insufficient adaptability to defects in varying developmental states, and failing to meet the practical needs for accurate detection of pavement defects throughout their entire lifecycle. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for detecting pavement defects based on vibration signals, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting pavement defects based on vibration signals, comprising the following steps: Step 100: System initialization, calibration of detection equipment and loading of preset multi-scale SVM-Transformer hybrid recognition model, disease evolution trend prediction model and cross-module linkage control center; Step 200: Based on the pathological vibration propagation model and the targeted acquisition mechanism adapted to the evolution stage, road vibration signals and corresponding acquisition location information are acquired, and raw vibration data and location data are output. The targeted acquisition mechanism receives filtering feedback parameters through the cross-module linkage control center and dynamically adjusts the sampling frequency, oversampling factor, and sensor gain. The pathological vibration propagation model is... In the formula, To distance from disease The vibration amplitude at that location, This represents the initial amplitude at the site of the disease. The vibration attenuation coefficient is... It is a natural constant. The distance between the detection point and the disease; Step 300: The original vibration data is preprocessed using a time-domain and frequency-domain linked adaptive filtering algorithm, and the preprocessed signal is output. The filtering algorithm includes an improved adaptive Kalman filter and a disease frequency domain adaptive mask filter. The cross-module linkage control center receives feature discrimination feedback and dynamically adjusts the filtering threshold and process noise variance. The process noise variance of the improved adaptive Kalman filter satisfies In the formula, for Time-process noise variance This is the initial value of the process noise variance. The mean of the feature discrimination index; The mask for the frequency domain adaptive masking filter of the disease Satisfy: When Disease frequency band and power spectral density hour ,otherwise In the formula, For frequency domain masking functions, For frequency, For dynamic thresholds, Let be the power spectral density function of the vibration signal. This represents the maximum value of the power spectral density function; Step 400: Based on the dynamic feedback weight feature system driven by the disease mechanism, feature extraction and dimensionality reduction are performed on the preprocessed signal to output a feature vector. The feature system includes time-domain features, frequency-domain features, time-frequency-domain features, and disease evolution features. The weights are dynamically updated through the cross-module linkage control center receiving classification confidence feedback. Each feature weight is evaluated using a discriminative evaluation function. After calculation and allocation, adjustments are made based on classification confidence levels, where, For the first The discriminative power of each feature For the sample containing disease The mean of each feature, For disease-free samples The mean of each feature, For the sample containing disease The standard deviation of each feature For disease-free samples Standard deviation of each feature; Step 500: Input the feature vector into the multi-scale SVM-Transformer hybrid recognition model, and output the preliminary disease identification result and classification confidence. The hybrid recognition model includes a multi-scale SVM branch and a Transformer encoder branch, which are fused using a weighted attention formula. To achieve output fusion, in the formula, , LocalScore represents the local feature matching score, and GlobalScore represents the global feature matching score. The output results for identifying SVM branches. The recognition output of the Transformer encoder branch. This is the final recognition output after fusion; Step 600: The validity of the preliminary disease identification results is confirmed by an evolutionary adaptation spatiotemporal consistency verification mechanism, and the final disease identification results are output. The verification mechanism dynamically adjusts the verification parameters, including the number of time verification frames and the number of spatial valid points, according to the disease evolution stage and the classification confidence level. Step 700: Combining the location data, the final identification result of the defects, and the defect evolution stage output by the defect evolution trend prediction model, output and store a three-dimensional distribution map of road defects and an evolution trend report.

[0006] Preferably, step 200 includes: Step 210: Divide the sensor array into a main acquisition unit and an auxiliary acquisition unit. The main acquisition unit is installed in the strong vibration area on both sides of the front axle of the vehicle, and the auxiliary acquisition unit is installed on both sides of the rear axle of the vehicle. Step 220: The disease evolution trend prediction model is based on the LSTM algorithm. It takes historical vibration signal characteristics, disease type and environmental data as input and outputs the disease evolution stage, which includes the budding stage, development stage and maturity stage. Step 230, the vibration attenuation coefficient Adjusted according to the disease's evolution stage: budding stage =0.03-0.05, development period =0.05-0.1, maturity stage =0.1-0.2; Step 240: The cross-module linkage control center adjusts the acquisition parameters according to the disease evolution stage: sampling frequency 6kHz and oversampling factor 3 times during the budding stage; sampling frequency 4kHz and oversampling factor 2 times during the development stage; and sampling frequency 3kHz and oversampling factor 1 time during the maturity stage. Simultaneously, it receives the residual interference from the filtered feedback. When the residual interference is >10%, the sensor gain is increased by 20%. Step 250: Synchronously collect the acquisition location information corresponding to the original vibration data through the positioning module. The positioning accuracy of the positioning module is ±0.8m. Associate the original vibration data and the location data according to the timestamp, and output the original vibration data and the location data.

[0007] Preferably, step 300 includes: Step 310, the initial value of the process noise variance =0.05, the state equation of the improved adaptive Kalman filter is: In the formula, for The system state vector at any given time. Here is the system state transition matrix. for The system state vector at any given time. For noise driving matrix, for Time-of-flight noise; observation noise variance ; Step 320: Dynamically adjust the target frequency range according to the disease evolution stage: 5-150Hz for the budding stage, 5-300Hz for the development stage, and 5-500Hz for the maturity stage. Step 330: Perform FFT transformation on the time-domain denoised signal, with 1024 FFT points, to generate a power spectral density map. The cross-module linkage control center receives feature discrimination feedback to obtain the average feature discrimination value. When the mean of the feature discrimination is At that time, the dynamic threshold =0.25; when the mean of the feature discrimination is... At that time, the dynamic threshold =0.3, shielding the corresponding narrowband interference; Step 340: Fuse the time-domain denoising result with the disease frequency-domain adaptive mask filtering result, and output the preprocessed signal with a signal-to-noise ratio ≥ 55dB.

[0008] Preferably, step 400 includes: Step 410: Extract the time-domain features, frequency-domain features, time-frequency-domain features, and disease evolution features of the preprocessed signal; the time-domain features include peak value, root mean square value, kurtosis, impulse factor, and waveform factor; the frequency-domain features include peak power spectral density, dominant frequency, spectral centroid, and spectral bandwidth; the time-frequency-domain features are the energy entropy after wavelet packet decomposition; the disease evolution features include signal rise time smoothness and spectral harmonic growth rate. Step 420: Calculate the discriminative power of each feature using the discriminative power evaluation function. ,according to Assign initial weights, where, For the first The initial weights of each feature, The sum of the discriminative power of all features is used; then, adjustments are made based on the classification confidence: the weights of features with a classification confidence < 0.8 are reduced by 20%, and the weights of features with a classification confidence ≥ 0.95 are increased by 30%. Step 430: The PCA algorithm is used to reduce the dimensionality of the extracted multidimensional features, retaining 99% of the feature information, and outputting the 12-dimensional feature vector.

[0009] Preferably, step 500 includes: Step 510: The multi-scale SVM branch takes into account the time-domain basic features and frequency-domain features to capture local features; the Transformer encoder branch contains multiple encoders, takes into account the feature vector, captures global features, and outputs a global feature vector. Step 520: Calculate the LocalScore and the GlobalScore, and substitute them into the weighted attention fusion formula to obtain the fusion output; Step 530: Train a basic model based on a core sample library containing labeled samples of asphalt pavement and cement pavement; adapt to niche pavement using a meta-learning algorithm, achieving a high recognition rate with only a small number of labeled samples of niche pavement; output the preliminary identification results of the defects and the classification confidence score, where the classification confidence score ranges from 0 to 1.

[0010] Preferably, step 600 includes: Step 610: Adjust the time verification frame number according to the disease evolution stage: 4 consecutive frames in the budding stage, 3 consecutive frames in the development stage, and 2 consecutive frames in the maturity stage, and determine whether the consecutive frames are the same preliminary identification result of the disease. Step 620: Adjust the spatial verification parameters according to the disease evolution stage and the classification confidence level: ≥3 valid points within a preset range during the budding stage, ≥5 valid points within a preset range during the development stage, and ≥4 valid points within a preset range during the maturity stage; when the classification confidence level is <0.8, increase the number of valid spatial points by 1. Step 630: If both time verification and spatial verification conditions are met, output the final disease identification result; otherwise, it is determined to be an interference signal and the final disease identification result is not output.

[0011] Preferably, step 700 includes: Step 710: Associate the final identification result of the disease, the location data, and the disease evolution stage to generate a three-dimensional distribution map of the road surface disease and an evolution trend report containing the disease location, type, severity, disease evolution stage, and detection time. The latitude and longitude accuracy of the disease location is ±0.8m. Step 720: The three-dimensional distribution map of pavement defects and the evolution trend report are transmitted to the cloud backend management system through the communication module. The data transmission rate of the communication module is ≥100Mbps. The cloud backend management system stores, queries, and statistically analyzes the three-dimensional distribution map of pavement defects and the evolution trend report, and automatically generates maintenance suggestions.

[0012] The present invention also provides a pavement distress detection system based on vibration signals, comprising: The system initialization module is used to calibrate the detection equipment and load the preset multi-scale SVM-Transformer hybrid recognition model, disease evolution trend prediction model, and cross-module linkage control center; The targeted acquisition module is used to acquire road vibration signals and corresponding acquisition location information based on the targeted acquisition mechanism adapted to the vibration propagation model and evolution stage of the disease, and output the original vibration data and location data. The targeted acquisition module includes a sensor component for acquiring vibration signals and a positioning component for acquiring location information. The cross-module linkage control center receives filtering feedback parameters and dynamically adjusts the sampling frequency, oversampling factor and sensor gain. The dual-domain adaptive preprocessing module is used to preprocess the original vibration data using a time-domain and frequency-domain linked adaptive filtering algorithm and output the preprocessed signal. The dual-domain adaptive preprocessing module receives feature discrimination feedback through the cross-module linkage control center and dynamically adjusts the filtering threshold and process noise variance. The dynamic feedback feature module is used to extract features and reduce dimensions of the preprocessed signal based on the dynamic feedback weight feature system driven by the disease mechanism, and output feature vectors. The feature system includes time-domain features, frequency-domain features, time-frequency-domain features and disease evolution features. The weights are dynamically updated by receiving classification confidence feedback through the cross-module linkage control center. The hybrid recognition module is used to input the feature vector into the multi-scale SVM-Transformer hybrid recognition model and output the preliminary disease recognition result and classification confidence. The hybrid recognition model includes a multi-scale SVM branch and a Transformer encoder branch, which are fused using a weighted attention formula. In the formula, , ; An evolutionary adaptation verification module is used to confirm the validity of the preliminary disease identification results using an evolutionary adaptation spatiotemporal consistency verification mechanism and output the final disease identification results; the verification mechanism dynamically adjusts the verification parameters according to the disease evolution stage and the classification confidence level. The linkage output module is used to combine the location data, the final identification result of the disease, and the disease evolution stage to output and store a three-dimensional distribution map of road surface diseases and an evolution trend report.

[0013] The present invention also provides an electronic device, which is a physical device, comprising: The processor and the memory are communicatively connected. The memory is used to store at least one executable instruction executed by the processor, which executes the executable instruction to implement a road surface distress detection method based on vibration signals as described above.

[0014] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for detecting road surface defects based on vibration signals.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By employing a targeted acquisition mechanism adapted to the vibration propagation model and evolution stage of road defects, a time-domain and frequency-domain linked adaptive filtering algorithm, a dynamic feedback weight feature system driven by the defect mechanism, a multi-scale SVMTransformer hybrid recognition model, and an evolution-adaptive spatiotemporal consistency verification mechanism, comprehensive defect capture and accurate differentiation are achieved, improving signal processing quality and recognition reliability. This system is adaptable to defects at different evolution stages and various road surface types, providing comprehensive and practical reference information for road maintenance and assisting in scientific maintenance decision-making. Attached Figure Description

[0016] Figure 1 The main flowchart of a road surface distress detection method based on vibration signals provided in an embodiment of the present invention; Figure 2 A structural block diagram of a road surface defect detection system based on vibration signals provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0017] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1This invention provides a method for detecting pavement defects based on vibration signals, the method being applied to, including: Step 100: System initialization, calibration of detection equipment and loading of preset multi-scale SVM-Transformer hybrid recognition model, disease evolution trend prediction model and cross-module linkage control center.

[0019] Step 200: Based on the pathological vibration propagation model and the targeted acquisition mechanism adapted to the evolution stage, road vibration signals and corresponding acquisition location information are acquired, and raw vibration data and location data are output. The targeted acquisition mechanism receives filtering feedback parameters through the cross-module linkage control center and dynamically adjusts the sampling frequency, oversampling factor, and sensor gain. The pathological vibration propagation model is... In the formula, To distance from disease The vibration amplitude at that location, This represents the initial amplitude at the site of the disease. The vibration attenuation coefficient is... It is a natural constant. This refers to the distance between the detection point and the disease.

[0020] Specifically, step 200 includes: Step 210: Divide the sensor array into a main acquisition unit and an auxiliary acquisition unit. The main acquisition unit is installed in the strong vibration area on both sides of the front axle of the vehicle, and the auxiliary acquisition unit is installed on both sides of the rear axle of the vehicle. Step 220: The disease evolution trend prediction model is based on the LSTM algorithm. It takes historical vibration signal characteristics, disease type and environmental data as input and outputs the disease evolution stage, which includes the budding stage, development stage and maturity stage. Step 230, the vibration attenuation coefficient Adjusted according to the disease's evolution stage: budding stage =0.03-0.05, development period =0.05-0.1, maturity stage =0.1-0.2; Step 240: The cross-module linkage control center adjusts the acquisition parameters according to the disease evolution stage: sampling frequency 6kHz and oversampling factor 3 times during the budding stage; sampling frequency 4kHz and oversampling factor 2 times during the development stage; and sampling frequency 3kHz and oversampling factor 1 time during the maturity stage. Simultaneously, it receives the residual interference from the filtered feedback. When the residual interference is >10%, the sensor gain is increased by 20%. Step 250: Synchronously collect the acquisition location information corresponding to the original vibration data through the positioning module. The positioning accuracy of the positioning module is ±0.8m. Associate the original vibration data and the location data according to the timestamp, and output the original vibration data and the location data.

[0021] It should be noted that in this embodiment, the division of the sensor array into primary and secondary acquisition units is based on the core design of the disease vibration propagation model. The primary acquisition unit focuses on areas of strong disease vibration to capture accurate signals, while the secondary acquisition unit is used to assist in offsetting environmental interference. Together, they achieve targeted acquisition of vibration signals. The core idea of ​​this design is to utilize the spatial distribution differences between disease vibration and environmental interference to improve the signal-to-noise ratio of the acquired signal. The LSTM algorithm, as the core algorithm for trend prediction of time-series data, can effectively mine the temporal correlation between historical vibration signal characteristics, disease types, and environmental data (such as temperature, humidity, and traffic flow), thereby accurately outputting the disease evolution stage and providing a core basis for the dynamic adjustment of subsequent acquisition and filtering parameters. The vibration attenuation coefficient... The design is adjusted according to the evolutionary stage because the vibration propagation characteristics of diseases differ at different stages. Diseases in the budding stage experience faster vibration decay, requiring smaller... To clearly define the area of ​​strong signal acquisition, the vibration propagation of mature diseases is more stable, allowing for the use of larger acquisition methods. Balancing acquisition efficiency and accuracy, the cross-module linkage control center dynamically adjusts the acquisition parameters, achieving deep linkage between the acquisition process and the disease evolution status and filtering feedback. This avoids signal distortion or redundancy caused by fixed parameter acquisition. Furthermore, the positioning module is linked to the timestamp of the vibration data, ensuring that each set of vibration data can be accurately traced to the specific road surface location, providing a foundation for subsequent disease location.

[0022] In one possible implementation, the number of main acquisition units and auxiliary acquisition units of the sensor array can be adjusted according to the detection accuracy requirements. The main acquisition units can be set to 2-4, and the auxiliary acquisition units can be set to 2-3. The installation height can be within the range of 30-40cm to adapt to different types of detection vehicles. The input data of the LSTM algorithm can also include basic information such as road material type and service life to further improve the prediction accuracy of the evolution stage. The positioning module can adopt Beidou + GPS + GLONASS tri-mode positioning, and the positioning accuracy can be optimized to ±0.5m.

[0023] For example, in one feasible implementation, when a vehicle is detected driving on an asphalt road, the sensor array includes three main acquisition units (installed on both sides and in the middle of the front axle, 35cm above the ground) and two auxiliary acquisition units (installed on both sides of the rear axle, 35cm above the ground). The damage evolution trend prediction model is input with the characteristics of historical vibration signals over the past three months, the type of damage (cracks), and environmental data (average daily temperature 25℃, humidity 60%, average daily traffic flow 5000 vehicles). The output indicates that the crack is in the development stage. Based on this, the cross-module linkage control center adjusts the sampling frequency of the main acquisition units to 4kHz and the oversampling factor to 2 times, and the sampling frequency of the auxiliary acquisition units to 2kHz. At the same time, the residual interference from the received filtered feedback is 8% (≤10%), and the sensor gain remains at the default. The positioning module uses Beidou + GPS dual-mode positioning, collects location information at a sampling frequency of 1Hz, associates the original vibration data with latitude and longitude information by timestamp, and finally outputs a dataset containing 1000 sets of vibration data and corresponding location information.

[0024] Step 300: The original vibration data is preprocessed using a time-domain and frequency-domain linked adaptive filtering algorithm, and the preprocessed signal is output. The filtering algorithm includes an improved adaptive Kalman filter and a disease frequency domain adaptive mask filter. The cross-module linkage control center receives feature discrimination feedback and dynamically adjusts the filtering threshold and process noise variance. The process noise variance of the improved adaptive Kalman filter satisfies In the formula, for Time-process noise variance This is the initial value of the process noise variance. The mean of the feature discrimination index; The mask for the frequency domain adaptive masking filter of the disease Satisfy: When Disease frequency band and power spectral density hour ,otherwise In the formula, For frequency domain masking functions, For frequency, For dynamic thresholds, Let be the power spectral density function of the vibration signal. This represents the maximum value of the power spectral density function.

[0025] Specifically, step 300 includes: Step 310, the initial value of the process noise variance =0.05, the state equation of the improved adaptive Kalman filter is: In the formula, for The system state vector at any given time. Here is the system state transition matrix. for The system state vector at any given time. For noise driving matrix, for Time-of-flight noise; observation noise variance ; Step 320: Dynamically adjust the target frequency range according to the disease evolution stage: 5-150Hz for the budding stage, 5-300Hz for the development stage, and 5-500Hz for the maturity stage. Step 330: Perform FFT transformation on the time-domain denoised signal, with 1024 FFT points, to generate a power spectral density map. The cross-module linkage control center receives feature discrimination feedback to obtain the average feature discrimination value. When the mean of the feature discrimination is At that time, the dynamic threshold =0.25; when the mean of the feature discrimination is... At that time, the dynamic threshold =0.3, shielding the corresponding narrowband interference; Step 340: Fuse the time-domain denoising result with the disease frequency-domain adaptive mask filtering result, and output the preprocessed signal with a signal-to-noise ratio ≥ 55dB.

[0026] It should be noted that in this embodiment, the improved adaptive Kalman filter is an optimized design of the traditional Kalman filter. Its core idea is to introduce the mean of feature discrimination. Dynamic adjustment process noise variance This allows the filtering process to adapt to signals with different characteristic qualities, avoiding over- or under-denoising caused by the fixed noise variance in traditional Kalman filtering; system state transition matrix and noise driving matrix The settings are based on the non-stationary characteristics of road vibration signals to ensure that the state prediction can accurately track signal changes. The design of adjusting the target frequency range according to the stage of disease evolution is because the vibration frequency distribution of diseases varies at different stages. Vibration energy of diseases in the budding stage is concentrated in the low frequency band, while vibration energy of diseases in the mature stage covers a wider frequency band. Targeted adjustment of the target frequency range can effectively preserve disease characteristics and filter out irrelevant interference. The FFT transform is used to convert the time-domain vibration signal into a frequency-domain signal to generate a power spectral density map. This provides a frequency distribution basis for frequency domain masking filtering, and a dynamic threshold. The threshold is flexibly adjusted based on the mean of feature discrimination. When the feature discrimination is low, the threshold is lowered to retain more potential disease features. When the feature discrimination is high, the threshold is raised to strengthen interference shielding. Finally, by fusing time-domain and frequency-domain filtering, the signal-to-noise ratio of the preprocessed signal is ≥55dB, providing a high-quality data foundation for subsequent feature extraction.

[0027] In one possible implementation, the number of FFT points can be adjusted to 512, 2048, etc., depending on the sampling frequency. A higher sampling frequency allows for a larger number of FFT points to improve frequency domain resolution. The observation noise variance of the improved adaptive Kalman filter... It can be dynamically adjusted according to the intensity of environmental interference; when the interference is strong... It can be set to 0.15-0.2 when the interference is weak. It can be set to 0.05-0.1; the specific range of the target frequency range can be fine-tuned according to the road surface type. The target frequency range of gravel road surface can be increased by 10-20Hz.

[0028] For example, in one feasible implementation, for the raw vibration data (sampling frequency 4kHz) of pothole defects in the development stage, an improved adaptive Kalman filter is first used for time-domain denoising, and the system state transition matrix is... Let [[1,0.00025],[0,1]] (based on a sampling period of 0.00025s), and let the noise driving matrix be... Let [[0.000125], [0.00025]] be the initial value of the process noise variance. =0.05, Observation noise variance The target frequency range was adjusted to 5-300Hz; a 1024-point FFT transformation was performed on the time-domain denoised signal to generate a power spectral density map. The average feature discrimination of the cross-module linkage control center. , dynamic threshold Setting it to 0.25, the 100Hz narrowband signal corresponding to engine interference is shielded; after fusing the time-domain and frequency-domain filtering results, the preprocessed signal with a signal-to-noise ratio of 58dB is output. The defect characteristics of this signal are clear, and the interference components are significantly reduced.

[0029] Step 400: Based on the dynamic feedback weight feature system driven by the disease mechanism, feature extraction and dimensionality reduction are performed on the preprocessed signal to output a feature vector. The feature system includes time-domain features, frequency-domain features, time-frequency-domain features, and disease evolution features. The weights are dynamically updated through the cross-module linkage control center receiving classification confidence feedback. Each feature weight is evaluated using a discriminative evaluation function. After calculation and allocation, adjustments are made based on classification confidence levels, where, For the first The discriminative power of each feature For the sample containing disease The mean of each feature, For disease-free samples The mean of each feature, For the sample containing disease The standard deviation of each feature For disease-free samples The standard deviation of each feature.

[0030] Specifically, step 400 includes: Step 410: Extract the time-domain features, frequency-domain features, time-frequency-domain features, and disease evolution features of the preprocessed signal; the time-domain features include peak value, root mean square value, kurtosis, impulse factor, and waveform factor; the frequency-domain features include peak power spectral density, dominant frequency, spectral centroid, and spectral bandwidth; the time-frequency-domain features are the energy entropy after wavelet packet decomposition; the disease evolution features include signal rise time smoothness and spectral harmonic growth rate. Step 420: Calculate the discriminative power of each feature using the discriminative power evaluation function. ,according to Assign initial weights, where, For the first The initial weights of each feature, The sum of the discriminative power of all features is used; then, adjustments are made based on the classification confidence: the weights of features with a classification confidence < 0.8 are reduced by 20%, and the weights of features with a classification confidence ≥ 0.95 are increased by 30%. Step 430: The PCA algorithm is used to reduce the dimensionality of the extracted multidimensional features, retaining 99% of the feature information, and outputting the 12-dimensional feature vector.

[0031] It should be noted that in this embodiment, the dynamic feedback weighted feature system driven by the disease mechanism is the core design. Its feature selection is closely integrated with the occurrence and evolution mechanism of pavement diseases. Time-domain features are used to capture macroscopic characteristics such as the amplitude and morphology of vibration signals; frequency-domain features are used to capture characteristics such as the frequency distribution and energy concentration areas of vibration signals; time-frequency domain features (energy entropy after wavelet packet decomposition) are used to capture the time-frequency distribution uniformity of vibration signals; and disease evolution features specifically capture the unique vibration patterns of diseases at different evolution stages. These four types of features synergistically cover the multi-dimensional characteristics of disease vibration, ensuring the comprehensiveness and discriminative power of the features; the discriminative power evaluation function... The PCA algorithm is used to quantify the ability of each feature to distinguish between diseased and non-diseased features. Features with higher discriminative power are assigned higher initial weights, enabling them to play a greater role in subsequent identification. The dynamic weight adjustment based on classification confidence achieves a feedback linkage between weight and identification effect. When the classification confidence of a feature is low, its weight is reduced to reduce interference; when the classification confidence is high, its weight is increased to enhance its discriminative power. The PCA algorithm is used to reduce the dimensionality of multi-dimensional features. While retaining 99% of the feature information, the feature dimension is reduced to 12 dimensions. This reduces the computational complexity of the subsequent identification model, improves inference efficiency, and avoids the decline in identification accuracy caused by the curse of dimensionality. The final output 12-dimensional feature vector has both discriminative power and computational efficiency.

[0032] In one possible implementation, the time-domain features may also include mean, peak factor, kurtosis factor, etc., and the frequency-domain features may also include spectral skewness, spectral kurtosis, etc. The disease evolution features may also include vibration decay rate, spectral energy growth rate, etc. The calculation of the discrimination evaluation function may introduce a weighting factor, assigning a weight coefficient of 1.2-1.5 times to the disease evolution features. The proportion of feature information retained by the PCA algorithm may be adjusted to 98%-99.5% according to the recognition accuracy requirements, corresponding to an output of 8-15 dimensional feature vectors.

[0033] For example, in one feasible implementation, five time-domain features (such as peak value and root mean square value), four frequency-domain features (such as peak power spectral density and dominant frequency), eight time-frequency domain features (energy entropy) after five-level decomposition of the db4 wavelet packet, and two disease evolution features (such as signal rise edge smoothness and spectral harmonic growth rate) are extracted from the preprocessed development-stage crack disease signal, for a total of 19 features; the discriminative power of each feature is calculated using a discriminative evaluation function. Among them, the main frequency =0.9, signal rising edge smoothness =0.85, according to Initial weights were assigned: the weight of the main frequency was 0.12, and the weight of the signal rising edge smoothness was 0.11. Based on the classification confidence feedback, the classification confidence corresponding to the main frequency was 0.96 ≥ 0.95, so the weight was increased by 30% to 0.156; the classification confidence corresponding to a certain time-domain feature (mean) was 0.75 < 0.8, so the weight was decreased by 20% to 0.08. The PCA algorithm was used to reduce the dimensionality of the 19 features, retaining 99% of the feature information, and finally outputting a 12-dimensional feature vector, which can accurately characterize the core characteristics of crack disease in the development stage.

[0034] Step 500: Input the feature vector into the multi-scale SVM-Transformer hybrid recognition model, and output the preliminary disease identification result and classification confidence. The hybrid recognition model includes a multi-scale SVM branch and a Transformer encoder branch, which are fused using a weighted attention formula. To achieve output fusion, in the formula, , LocalScore represents the local feature matching score, and GlobalScore represents the global feature matching score. The output results for identifying SVM branches. The recognition output of the Transformer encoder branch. This is the final recognition output after fusion.

[0035] Specifically, step 500 includes: Step 510: The multi-scale SVM branch takes into account the time-domain basic features and frequency-domain features to capture local features; the Transformer encoder branch contains multiple encoders, takes into account the feature vector, captures global features, and outputs a global feature vector. Step 520: Calculate the LocalScore and the GlobalScore, and substitute them into the weighted attention fusion formula to obtain the fusion output; Step 530: Train a basic model based on a core sample library containing labeled samples of asphalt pavement and cement pavement; adapt to niche pavement using a meta-learning algorithm, achieving a high recognition rate with only a small number of labeled samples of niche pavement; output the preliminary identification results of the defects and the classification confidence score, where the classification confidence score ranges from 0 to 1.

[0036] It should be noted that, in this embodiment, the core design of the multi-scale SVM-Transformer hybrid recognition model is to capture both local fine-grained features and global correlation features. The multi-scale SVM branch excels at processing high-dimensional small sample data and can accurately capture local details (such as the peak value and dominant frequency of disease vibration) in the time domain and frequency domain. The Transformer encoder branch, through a self-attention mechanism, can mine global correlation information in the feature vector (such as the continuous vibration correlation of crack diseases and the regional vibration concentration correlation of pit diseases). The fusion of the two achieves comprehensive feature utilization of local details and global correlations, improving recognition accuracy. LocalScore is used to quantify the degree of matching between the SVM branch output and the local features of the sample library, and GlobalScore is used to quantify the degree of matching between the Transformer branch output and the global features of the sample library. The weighted attention fusion formula uses dynamic weights. , The system achieves adaptive fusion of two types of features, enabling the model to flexibly adjust the contribution of local and global features based on the feature advantages of different disease types. The core sample library is used to train the basic model based on labeled samples of mainstream pavements such as asphalt and cement, ensuring the model's ability to identify diseases of mainstream pavements. The meta-learning algorithm is used for niche pavements such as gravel roads and composite roads, and quickly fine-tunes the model parameters with a small number of labeled samples (only about 500 sets are needed) to solve the adaptation problem caused by insufficient samples of niche pavements. The final output of the preliminary disease identification results includes information such as disease type and severity, and the classification confidence quantifies the reliability of the identification results, providing a basis for subsequent spatiotemporal verification.

[0037] In one possible implementation, the number of layers in the Transformer encoder branch can be set to 4-8 layers. The more layers there are, the stronger the ability to capture global features, but the higher the computational complexity. The meta-learning algorithm can be MAML, Reptile, etc., and the number of niche road surface annotation samples can be adjusted to 300-800 groups according to the road surface complexity. The kernel function of the multi-scale SVM branch can be selected from RBF kernel, polynomial kernel, etc., and adapted according to the type of disease features.

[0038] For example, in one feasible implementation, the multi-scale SVM branch takes time-domain basic features (peak value, root mean square value) and frequency-domain features (dominant frequency, spectral centroid) as input, and uses the RBF kernel function to capture local features; the Transformer encoder branch contains 6 encoder layers, takes a 12-dimensional feature vector as input, captures global features through a self-attention mechanism, and outputs a 256-dimensional global feature vector; the calculated LocalScore = 0.88 and GlobalScore = 0.92 are substituted into the weighted attention fusion formula, , Resulting in a fused output The basic model was trained based on labeled samples of asphalt pavement (5000 sets) and cement pavement (4600 sets). For gravel pavement distress, the MAML meta-learning algorithm was used to fine-tune the model parameters by inputting 500 sets of labeled samples. The final output of the preliminary distress identification result was "cracks in the development stage of gravel pavement", with a classification confidence of 0.93.

[0039] Step 600: The validity of the preliminary disease identification results is confirmed by an evolutionary adaptation spatiotemporal consistency verification mechanism, and the final disease identification results are output. The verification mechanism dynamically adjusts the verification parameters, including the number of time verification frames and the number of spatial valid points, according to the disease evolution stage and the classification confidence level.

[0040] Specifically, step 600 includes: Step 610: Adjust the time verification frame number according to the disease evolution stage: 4 consecutive frames in the budding stage, 3 consecutive frames in the development stage, and 2 consecutive frames in the maturity stage, and determine whether the consecutive frames are the same preliminary identification result of the disease. Step 620: Adjust the spatial verification parameters according to the disease evolution stage and the classification confidence level: ≥3 valid points within a preset range during the budding stage, ≥5 valid points within a preset range during the development stage, and ≥4 valid points within a preset range during the maturity stage; when the classification confidence level is <0.8, increase the number of valid spatial points by 1. Step 630: If both time verification and spatial verification conditions are met, output the final disease identification result; otherwise, it is determined to be an interference signal and the final disease identification result is not output.

[0041] It should be noted that, in this embodiment, the evolution-adaptive spatiotemporal consistency verification mechanism is the core design for reducing the misjudgment rate. Its core idea is to combine the temporal stability and spatial distribution characteristics of the disease to perform secondary verification on the initial disease identification results, avoiding misjudging interference signals as diseases. The design of adjusting the number of temporal verification frames according to the disease's evolutionary stage is because the vibration stability of the disease varies at different stages. Diseases in the budding stage have weaker vibrations and poorer stability, requiring more consecutive frame verifications to ensure reliability. Diseases in the mature stage have stable and persistent vibrations, allowing for fewer consecutive frame verifications to improve efficiency. Spatial verification parameters... The adjustments are based on the spatial distribution patterns of diseases. In the budding stage, the disease range is small and there are few effective identification points. In the development stage, the disease range expands and the number of effective identification points increases. In the mature stage, the disease range is stable and the number of effective identification points is moderate. At the same time, the number of effective points in space is adjusted in conjunction with the classification confidence. When the classification confidence is low, the number of effective points is increased to strengthen the rigor of verification. When the classification confidence is high, the default number of effective points is maintained to balance accuracy and efficiency. The dual conditions of time verification and spatial verification ensure the reliability of the final disease identification results, control the false judgment rate within 0.05%, and avoid missing effective diseases.

[0042] In one possible implementation, the number of time verification frames can be finely adjusted according to the sampling frequency. When the sampling frequency is 3kHz, the budding stage can be set to 5 frames and the mature stage can be set to 1 frame. The preset range of spatial verification can be adjusted according to the type of disease. The preset range for crack diseases can be set to 8-12m and the preset range for pit diseases can be set to 2-4m. The threshold of classification confidence can be adjusted to 0.75-0.85 according to the detection accuracy requirements.

[0043] For example, in one feasible implementation, for the preliminary identification result of budding crack disease (classification confidence = 0.78 < 0.8), the time verification frame number is set to 4 frames. If the identification result of 4 consecutive frames is "budding crack", the time verification condition is met. The spatial verification parameter is adjusted to ≥ 4 valid points within a preset range of 8m (3 by default + 1 more if the classification confidence is insufficient). The location data confirms that there are 4 valid identification points within 8m, and the distance between adjacent valid points is ≤ 1.5m, which meets the spatial verification condition. The final output disease identification result is "budding crack (width 0.25mm)". If a preliminary identification result only meets the time verification (consistent for 3 consecutive frames), but there is only 1 valid point within 3m in the spatial verification, it is determined to be an interference signal and the identification result is not output.

[0044] Step 700: Combining the location data, the final identification result of the defects, and the defect evolution stage output by the defect evolution trend prediction model, output and store a three-dimensional distribution map of road defects and an evolution trend report.

[0045] Specifically, step 700 includes: Step 710: Associate the final identification result of the disease, the location data, and the disease evolution stage to generate a three-dimensional distribution map of the road surface disease and an evolution trend report containing the disease location, type, severity, disease evolution stage, and detection time. The latitude and longitude accuracy of the disease location is ±0.8m. Step 720: The three-dimensional distribution map of pavement defects and the evolution trend report are transmitted to the cloud backend management system through the communication module. The data transmission rate of the communication module is ≥100Mbps. The cloud backend management system stores, queries, and statistically analyzes the three-dimensional distribution map of pavement defects and the evolution trend report, and automatically generates maintenance suggestions.

[0046] It should be noted that, in this embodiment, the generation of a three-dimensional distribution map and evolution trend report of pavement defects is the core link in the practical application of the detection results. The three-dimensional distribution map of pavement defects uses latitude and longitude (accuracy ±0.8m) and defect depth as coordinates to intuitively present the spatial distribution location, range, and severity of defects, facilitating maintenance personnel to quickly locate defects. The evolution trend report, based on the defect evolution stage and historical detection data, predicts the future development speed of defects (e.g., crack width increases by 0.1mm per month) and potential risks (e.g., potholes in the mature stage may cause vehicle bumps and tire damage), providing a basis for selecting the timing of maintenance; communication The module is designed with a data transmission rate of ≥100Mbps to ensure that the 3D distribution map (containing a large amount of spatial coordinate data) and evolution trend report can be quickly and in real time transmitted to the cloud back-end management system, avoiding data transmission delays from affecting maintenance decisions. The cloud back-end management system not only has data storage, query, and statistical functions, but can also automatically generate maintenance suggestions based on the type of damage, evolution stage, and pavement type. For example, it recommends sealing maintenance for cracks in the budding stage, filling maintenance for potholes in the development stage, and reconstruction maintenance for subsidence in the mature stage. This provides full-process support from detection to maintenance decision-making, improving the accuracy and efficiency of pavement maintenance.

[0047] In addition, it should be noted that the design of this step breaks through the limitation of traditional detection that only outputs the location of the disease. It realizes spatial visualization of the disease through a three-dimensional distribution map, predicts the time dimension of the disease through an evolution trend report, and realizes the practical application of the detection results through maintenance suggestions. The three work together to provide complete data support for road maintenance in terms of spatial location, time trend, and action plan, which significantly improves the scientific nature and pertinence of maintenance work.

[0048] In one possible implementation, the communication module can be selected from 5G, WiFi 6, etc., and the data transmission rate of the 5G module can be optimized to ≥300Mbps, which is suitable for data transmission of large-scale pavement inspection; the three-dimensional distribution map of pavement defects can support online annotation, distance measurement, area calculation and other functions; the prediction period of the evolution trend report can be selected from 1 month, 3 months, 6 months, etc., to adapt to the maintenance plan cycle; the cloud back-end management system can be connected to the highway maintenance management platform to realize data sharing and collaborative office.

[0049] For example, in one feasible implementation, by associating the final identification result of the pavement defect ("potholes in the mature stage of cement pavement, depth 5mm"), location data (latitude and longitude 116.4°E, 39.9°N), and evolution stage (maturity stage), a three-dimensional distribution map of the pavement defect is generated. This map uses Google Maps as the base map, marking the latitude and longitude range of the potholes (116.4001°E-116.4003°E, 39.9002°N-39.9004°N) and the depth of 5mm, and marking them in red (representing high risk). The evolution trend report predicts that if the pothole is not maintained, its depth may increase to 8mm within one month, posing a risk of tire blowout. The three-dimensional distribution map and evolution trend report are transmitted to the cloud backend management system through the Huawei ME909s-8215G communication module (transmission rate 200Mbps). The system automatically generates a maintenance suggestion: "Use hot asphalt for maintenance within 7 days, with a filling thickness of 6mm". This suggestion is also synchronized to the highway maintenance management platform. Maintenance personnel can query the information on the defect and the maintenance suggestion through the platform and arrange maintenance work.

[0050] In this embodiment, a targeted acquisition mechanism adapted to the vibration propagation model and evolution stage of the road surface is used, along with a time-domain and frequency-domain linked adaptive filtering algorithm, a dynamic feedback weight feature system driven by the road surface mechanism, a multi-scale SVMTransformer hybrid recognition model, and an evolution-adaptive spatiotemporal consistency verification mechanism. This enables comprehensive capture and accurate differentiation of road surface defects, improves signal processing quality and recognition reliability, adapts to defects at different evolution stages and various road surface types, provides comprehensive and practical reference information for road surface maintenance, and assists in scientific maintenance decision-making.

[0051] Based on the above embodiments, such as Figure 2 As shown, the present invention also provides a pavement distress detection system based on vibration signals to support the pavement distress detection method based on vibration signals described in the above embodiments. The pavement distress detection system based on vibration signals includes: The system initialization module is used to calibrate the detection equipment and load the preset multi-scale SVM-Transformer hybrid recognition model, disease evolution trend prediction model, and cross-module linkage control center; The targeted acquisition module is used to acquire road vibration signals and corresponding acquisition location information based on the targeted acquisition mechanism adapted to the vibration propagation model and evolution stage of the disease, and output the original vibration data and location data. The targeted acquisition module includes a sensor component for acquiring vibration signals and a positioning component for acquiring location information. The cross-module linkage control center receives filtering feedback parameters and dynamically adjusts the sampling frequency, oversampling factor and sensor gain. The dual-domain adaptive preprocessing module is used to preprocess the original vibration data using a time-domain and frequency-domain linked adaptive filtering algorithm and output the preprocessed signal. The dual-domain adaptive preprocessing module receives feature discrimination feedback through the cross-module linkage control center and dynamically adjusts the filtering threshold and process noise variance. The dynamic feedback feature module is used to extract features and reduce dimensions of the preprocessed signal based on the dynamic feedback weight feature system driven by the disease mechanism, and output feature vectors. The feature system includes time-domain features, frequency-domain features, time-frequency-domain features and disease evolution features. The weights are dynamically updated by receiving classification confidence feedback through the cross-module linkage control center. The hybrid recognition module is used to input the feature vector into the multi-scale SVM-Transformer hybrid recognition model and output the preliminary disease recognition result and classification confidence. The hybrid recognition model includes a multi-scale SVM branch and a Transformer encoder branch, which are fused using a weighted attention formula. In the formula, , ; An evolutionary adaptation verification module is used to confirm the validity of the preliminary disease identification results using an evolutionary adaptation spatiotemporal consistency verification mechanism and output the final disease identification results; the verification mechanism dynamically adjusts the verification parameters according to the disease evolution stage and the classification confidence level. The linkage output module is used to combine the location data, the final identification result of the disease, and the disease evolution stage to output and store a three-dimensional distribution map of road surface diseases and an evolution trend report.

[0052] In this embodiment, a targeted acquisition mechanism adapted to the vibration propagation model and evolution stage of the road surface is used, along with a time-domain and frequency-domain linked adaptive filtering algorithm, a dynamic feedback weight feature system driven by the road surface mechanism, a multi-scale SVMTransformer hybrid recognition model, and an evolution-adaptive spatiotemporal consistency verification mechanism. This enables comprehensive capture and accurate differentiation of road surface defects, improves signal processing quality and recognition reliability, adapts to defects at different evolution stages and various road surface types, provides comprehensive and practical reference information for road surface maintenance, and assists in scientific maintenance decision-making.

[0053] Furthermore, the pavement distress detection system based on vibration signals can run the aforementioned pavement distress detection method based on vibration signals. For specific implementation details, please refer to the method embodiment, which will not be repeated here.

[0054] Based on the above embodiments, such as Figure 3 As shown, the present invention also provides an electronic device, the electronic device comprising: The processor 22 includes at least one processor 22, at least one memory 21, a communication interface 23, and a communication bus 24, wherein the processor 22 is communicatively connected to the memory 21. In this embodiment, the memory 21 can be implemented in any suitable manner, for example, the memory 21 can be a read-only memory, a hard disk drive, a solid-state drive, or a USB flash drive, etc.; the memory 21 is used to store at least one executable instruction executed by the processor; In this embodiment, the processor 22 can be implemented in any suitable manner. For example, the processor 22 can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) that can be executed by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc.; the processor is used to execute the executable instructions to implement a road surface distress detection method based on vibration signals as described above.

[0055] Based on the above embodiments, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for detecting road surface defects based on vibration signals.

[0056] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0057] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, equipment, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0058] In the several embodiments provided in this application, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or units may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or equipment, and may be electrical, mechanical, or other forms.

[0059] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0060] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0061] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program instructions, such as USB flash drives, portable hard drives, read-only storage servers, random access storage servers, magnetic disks, or optical disks.

[0062] Furthermore, it should be noted that the combination of the various technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0063] It should be noted that the above examples are merely specific embodiments of the present invention, and the present invention is obviously not limited to the above embodiments, with many similar variations. All modifications that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should fall within the protection scope of this invention.

[0064] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting pavement defects based on vibration signals, characterized in that, Includes the following steps: Step 100: System initialization, calibration of detection equipment and loading of preset multi-scale SVM-Transformer hybrid recognition model, disease evolution trend prediction model and cross-module linkage control center; Step 200: Based on the pathological vibration propagation model and the targeted acquisition mechanism adapted to the evolution stage, road vibration signals and corresponding acquisition location information are acquired, and raw vibration data and location data are output. The targeted acquisition mechanism receives filtering feedback parameters through the cross-module linkage control center and dynamically adjusts the sampling frequency, oversampling factor, and sensor gain. The pathological vibration propagation model is... In the formula, To distance from disease The vibration amplitude at that location, This represents the initial amplitude at the site of the disease. The vibration attenuation coefficient is... It is a natural constant. The distance between the detection point and the disease; Step 300: The original vibration data is preprocessed using a time-domain and frequency-domain linked adaptive filtering algorithm, and the preprocessed signal is output. The filtering algorithm includes an improved adaptive Kalman filter and a disease frequency domain adaptive mask filter. The cross-module linkage control center receives feature discrimination feedback and dynamically adjusts the filtering threshold and process noise variance. The process noise variance of the improved adaptive Kalman filter satisfies In the formula, for Time-process noise variance This is the initial value of the process noise variance. The mean of the feature discrimination index; The mask for the frequency domain adaptive masking filter of the disease Satisfy: When Disease frequency band and power spectral density hour ,otherwise In the formula, For frequency domain masking functions, For frequency, For dynamic thresholds, Let be the power spectral density function of the vibration signal. This represents the maximum value of the power spectral density function; Step 400: Based on the dynamic feedback weight feature system driven by the disease mechanism, feature extraction and dimensionality reduction are performed on the preprocessed signal to output a feature vector. The feature system includes time-domain features, frequency-domain features, time-frequency-domain features, and disease evolution features. The weights are dynamically updated through the cross-module linkage control center receiving classification confidence feedback. Each feature weight is evaluated using a discriminative evaluation function. After calculation and allocation, adjustments are made based on classification confidence levels, where, For the first The discriminative power of each feature For the sample containing disease The mean of each feature, For disease-free samples The mean of each feature, For the sample containing disease The standard deviation of each feature For disease-free samples Standard deviation of each feature; Step 500: Input the feature vector into the multi-scale SVM-Transformer hybrid recognition model, and output the preliminary disease identification result and classification confidence. The hybrid recognition model includes a multi-scale SVM branch and a Transformer encoder branch, which are fused using a weighted attention formula. To achieve output fusion, in the formula, , LocalScore represents the local feature matching score, and GlobalScore represents the global feature matching score. The output results for identifying SVM branches. The recognition output of the Transformer encoder branch. This is the final recognition output after fusion; Step 600: The validity of the preliminary disease identification results is confirmed by an evolutionary adaptation spatiotemporal consistency verification mechanism, and the final disease identification results are output. The verification mechanism dynamically adjusts the verification parameters, including the number of time verification frames and the number of spatial valid points, according to the disease evolution stage and the classification confidence level. Step 700: Combining the location data, the final identification result of the defects, and the defect evolution stage output by the defect evolution trend prediction model, output and store a three-dimensional distribution map of road defects and an evolution trend report.

2. The method for detecting pavement defects based on vibration signals according to claim 1, characterized in that, Step 200 includes: Step 210: Divide the sensor array into a main acquisition unit and an auxiliary acquisition unit. The main acquisition unit is installed in the strong vibration area on both sides of the front axle of the vehicle, and the auxiliary acquisition unit is installed on both sides of the rear axle of the vehicle. Step 220: The disease evolution trend prediction model is based on the LSTM algorithm. It takes historical vibration signal characteristics, disease type and environmental data as input and outputs the disease evolution stage, which includes the budding stage, development stage and maturity stage. Step 230, the vibration attenuation coefficient Adjusted according to the disease's evolution stage: budding stage =0.03-0.05, development period =0.05-0.1, maturity stage =0.1-0.2; Step 240: The cross-module linkage control center adjusts the acquisition parameters according to the disease evolution stage: sampling frequency 6kHz and oversampling factor 3 times during the budding stage; sampling frequency 4kHz and oversampling factor 2 times during the development stage; and sampling frequency 3kHz and oversampling factor 1 time during the maturity stage. Simultaneously, it receives the residual interference from the filtered feedback. When the residual interference is >10%, the sensor gain is increased by 20%. Step 250: Synchronously collect the acquisition location information corresponding to the original vibration data through the positioning module. The positioning accuracy of the positioning module is ±0.8m. Associate the original vibration data and the location data according to the timestamp, and output the original vibration data and the location data.

3. The method for detecting pavement defects based on vibration signals according to claim 1, characterized in that, Step 300 includes: Step 310, the initial value of the process noise variance =0.05, the state equation of the improved adaptive Kalman filter is: In the formula, for The system state vector at any given time. Here is the system state transition matrix. for The system state vector at any given time. For noise driving matrix, for Time-of-flight noise; observation noise variance ; Step 320: Dynamically adjust the target frequency range according to the disease evolution stage: 5-150Hz for the budding stage, 5-300Hz for the development stage, and 5-500Hz for the maturity stage. Step 330: Perform FFT transformation on the time-domain denoised signal, with 1024 FFT points, to generate a power spectral density map. The cross-module linkage control center receives feature discrimination feedback to obtain the average feature discrimination value. When the mean of the feature discrimination is At that time, the dynamic threshold =0.25; when the mean of the feature discrimination is... At that time, the dynamic threshold =0.3, shielding the corresponding narrowband interference; Step 340: Fuse the time-domain denoising result with the disease frequency-domain adaptive mask filtering result, and output the preprocessed signal with a signal-to-noise ratio ≥ 55dB.

4. The method for detecting pavement defects based on vibration signals according to claim 1, characterized in that, Step 400 includes: Step 410: Extract the time-domain features, frequency-domain features, time-frequency-domain features, and disease evolution features of the preprocessed signal; the time-domain features include peak value, root mean square value, kurtosis, impulse factor, and waveform factor; the frequency-domain features include peak power spectral density, dominant frequency, spectral centroid, and spectral bandwidth; the time-frequency-domain features are the energy entropy after wavelet packet decomposition; the disease evolution features include signal rise time smoothness and spectral harmonic growth rate. Step 420: Calculate the discriminative power of each feature using the discriminative power evaluation function. ,according to Assign initial weights, where, For the first The initial weights of each feature, The sum of the discriminative power of all features is used; then, adjustments are made based on the classification confidence: the weights of features with a classification confidence < 0.8 are reduced by 20%, and the weights of features with a classification confidence ≥ 0.95 are increased by 30%. Step 430: The PCA algorithm is used to reduce the dimensionality of the extracted multidimensional features, retaining 99% of the feature information, and outputting the 12-dimensional feature vector.

5. The method for detecting pavement defects based on vibration signals according to claim 1, characterized in that, Step 500 includes: Step 510: The multi-scale SVM branch takes into account the time-domain basic features and frequency-domain features to capture local features; the Transformer encoder branch contains multiple encoders, takes into account the feature vector, captures global features, and outputs a global feature vector. Step 520: Calculate the LocalScore and the GlobalScore, and substitute them into the weighted attention fusion formula to obtain the fusion output; Step 530: Train a basic model based on a core sample library containing labeled samples of asphalt pavement and cement pavement; adapt to niche pavement using a meta-learning algorithm, achieving a high recognition rate with only a small number of labeled samples of niche pavement; output the preliminary identification results of the defects and the classification confidence score, where the classification confidence score ranges from 0 to 1.

6. The method for detecting pavement defects based on vibration signals according to claim 1, characterized in that, Step 600 includes: Step 610: Adjust the time verification frame number according to the disease evolution stage: 4 consecutive frames in the budding stage, 3 consecutive frames in the development stage, and 2 consecutive frames in the maturity stage, and determine whether the consecutive frames are the same preliminary identification result of the disease. Step 620: Adjust the spatial verification parameters according to the disease evolution stage and the classification confidence level: ≥3 valid points within a preset range during the budding stage, ≥5 valid points within a preset range during the development stage, and ≥4 valid points within a preset range during the maturity stage; when the classification confidence level is <0.8, increase the number of valid spatial points by 1. Step 630: If both time verification and spatial verification conditions are met, output the final disease identification result; otherwise, it is determined to be an interference signal and the final disease identification result is not output.

7. The method for detecting pavement defects based on vibration signals according to claim 1, characterized in that, Step 700 includes: Step 710: Associate the final identification result of the disease, the location data, and the disease evolution stage to generate a three-dimensional distribution map of the road surface disease and an evolution trend report containing the disease location, type, severity, disease evolution stage, and detection time. The latitude and longitude accuracy of the disease location is ±0.8m. Step 720: The three-dimensional distribution map of pavement defects and the evolution trend report are transmitted to the cloud backend management system through the communication module. The data transmission rate of the communication module is ≥100Mbps. The cloud backend management system stores, queries, and statistically analyzes the three-dimensional distribution map of pavement defects and the evolution trend report, and automatically generates maintenance suggestions.

8. A pavement distress detection system based on vibration signals, characterized in that, Includes the following steps: The system initialization module is used to calibrate the detection equipment and load the preset multi-scale SVM-Transformer hybrid recognition model, disease evolution trend prediction model, and cross-module linkage control center; The targeted acquisition module is used to acquire road vibration signals and corresponding acquisition location information based on the targeted acquisition mechanism adapted to the vibration propagation model and evolution stage of the disease, and output the original vibration data and location data. The targeted acquisition module includes a sensor component for acquiring vibration signals and a positioning component for acquiring location information. The cross-module linkage control center receives filtering feedback parameters and dynamically adjusts the sampling frequency, oversampling factor and sensor gain. The dual-domain adaptive preprocessing module is used to preprocess the original vibration data using a time-domain and frequency-domain linked adaptive filtering algorithm and output the preprocessed signal. The dual-domain adaptive preprocessing module receives feature discrimination feedback through the cross-module linkage control center and dynamically adjusts the filtering threshold and process noise variance. The dynamic feedback feature module is used to perform feature extraction and dimensionality reduction on the preprocessed signal based on the dynamic feedback weight feature system driven by the disease mechanism, and output a feature vector. The feature system includes time-domain features, frequency-domain features, time-frequency-domain features, and disease evolution features. The weights are dynamically updated by receiving classification confidence feedback through the cross-module linkage control center. The hybrid recognition module is used to input the feature vector into the multi-scale SVM-Transformer hybrid recognition model and output the preliminary disease recognition result and classification confidence. The hybrid recognition model includes a multi-scale SVM branch and a Transformer encoder branch, which are fused using a weighted attention formula. In the formula, , ; An evolutionary adaptation verification module is used to confirm the validity of the preliminary disease identification results using an evolutionary adaptation spatiotemporal consistency verification mechanism and output the final disease identification results; the verification mechanism dynamically adjusts the verification parameters according to the disease evolution stage and the classification confidence level. The linkage output module is used to combine the location data, the final identification result of the disease, and the disease evolution stage to output and store a three-dimensional distribution map of road surface diseases and an evolution trend report.

9. An electronic device, characterized in that, The electronic device includes: The processor and the memory are communicatively connected. The memory is used to store at least one executable instruction executed by the processor, the processor being used to execute the executable instruction to implement a road surface distress detection method based on vibration signals as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements a method for detecting pavement defects based on vibration signals as described in any one of claims 1 to 7.

Images