A structural health assessment method based on road multi-dimensional detection data

Abstract

This invention relates to the field of data fusion technology, specifically to a structural health assessment method based on multidimensional road detection data. The method includes: simultaneously acquiring road surface images and ground-penetrating radar data and performing spatiotemporal alignment; extracting surface and structural features; learning the causal relationship between surface and structural defects based on historical data, using a causal graph to guide weight adjustment, and outputting a road feature vector by combining weights and features; calculating a health index and predicting defect evolution trends. This invention ensures that surface images and radar data have consistent locations and timestamps through spatiotemporal alignment, avoiding alignment errors for different defects; by performing confidence calculations simultaneously with feature extraction, noise interference is reduced, ensuring that all image or radar data is usable and preventing defect omissions; by analyzing and mining high-precision defect detection data, the distribution and evolution trends of road defects can be revealed.

Description

Technical Field

[0001] This invention relates to the field of data fusion technology, and specifically to a structural health assessment method based on multidimensional road detection data. Background Technology

[0002] Currently, the main methods for detecting apparent road defects fall into two categories: manual inspection and multi-functional road inspection vehicles. While manual inspection can comprehensively analyze the types, distribution, and dimensions of road defects, it suffers from low efficiency, requires traffic closures, and results are susceptible to subjective judgment. Multi-functional inspection vehicles, while largely unaffected by traffic, still require manual assistance for defect identification, location, and measurement, resulting in continued low efficiency and inaccuracy. Therefore, current methods for detecting road defects cannot meet the requirements for high efficiency, intelligence, and precision in road inspection. Particularly in the detection of latent defects within the structure, although ground-penetrating radar (GPR) technology has been gradually applied to non-destructive testing of road structures, the non-uniformity and strong attenuation of road structure layers, along with the complex and variable electromagnetic wave propagation environment of GPR, make the radar data characteristics of road defects easily susceptible to severe interference. Furthermore, differences in the size, severity, and development stage of defects can lead to distortions in their morphological features in GPR profile images, making the identification of latent road defects a significant challenge, facing strong noise interference, complex morphologies, and a scarcity of samples.

[0003] Accurate detection, evaluation, and management of surface and internal structural defects in highways have become urgent problems to be solved in highway construction and maintenance. Therefore, this application provides a structural health assessment method based on multi-dimensional road inspection data, which utilizes automated detection technology that integrates pavement appearance and structural data. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a structural health assessment method based on multidimensional road inspection data, which is an automated detection technology for fusing pavement appearance and structural data.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A structural health assessment method based on multidimensional road detection data includes:

[0007] Simultaneously acquire the road's apparent image I and ground-penetrating radar data R and perform spatiotemporal alignment;

[0008] Apparent feature extraction: Using a pre-trained CNN to extract image features and obtain the apparent feature vector. And calculate the confidence level of the apparent data. ; ;in, The variance of image noise;

[0009] Structural Feature Extraction: Using 3D CNN to process radar data to obtain structural feature vectors. And calculate the confidence level of the structured data. ; ;in, The variance of radar signal noise;

[0010] Constructing a cause-effect graph: Based on historical data, learn the causal relationship between apparent diseases and structural diseases, and use the cause-effect graph to guide the adjustment of weights; , ;where softmax() is a function; Corr() is the correlation coefficient between appearance and structural features, calculated using the Pearson correlation coefficient, reflecting the causal relationship of the disease; As weight;

[0011] By combining weights and features, a road feature vector is output. , ;

[0012] Calculate the health index , And predict the evolution trend of the disease. , Where sigmoid() is a function, W is the weight matrix, and b is the bias term; For time intervals;

[0013] Health Index The output is the health score of the current road area, combined with the predicted trend of disease evolution. Provide maintenance recommendations.

[0014] Preferably, the image noise variance is obtained based on wavelet transform.

[0015] Preferably, the image noise variance obtained based on wavelet transform further includes:

[0016] Choose a wavelet basis function;

[0017] A two-dimensional discrete wavelet transform is performed on the appearance image to decompose it into four sub-bands:

[0018] LL: Low frequency in both horizontal and vertical directions, containing the main approximate information of the appearance image;

[0019] LH: Horizontal low frequency, vertical high frequency, containing edge details in the vertical direction;

[0020] HL: Horizontal high frequencies, vertical low frequencies, including edge details in the horizontal direction;

[0021] HH: Both horizontal and vertical frequencies are high frequencies, containing noise and details in the diagonal direction;

[0022] Perform multi-level decomposition;

[0023] Noise figure extraction: Extract wavelet coefficients of the HH subband after the first level of decomposition;

[0024] Robust variance estimation:

[0025] The standard deviation is calculated using the absolute deviation of the median.

[0026] Preferably, the standard deviation is calculated using the absolute deviation of the median, including:

[0027] ;in, For each wavelet coefficient in the HH subband; median This is the median of the absolute values ​​of all wavelet coefficients.

[0028] Preferably, the wavelet basis function is db4 or sym8 in the Daubechies wavelet system.

[0029] Preferably, the radar signal noise variance is obtained based on the variation of S.

[0030] Preferably, the radar signal noise variance obtained based on the S-variance further includes:

[0031] Ground-penetrating radar data includes a one-dimensional GPR signal s(t);

[0032] S-transform: Performing an S-transform on a one-dimensional GPR signal s(t) yields a two-dimensional complex time-frequency matrix. ,in, f represents time, and 'f' represents depth; 'f' represents frequency.

[0033] Modulus of complex time-frequency matrix Indicates the signal in time and energy density at frequency f;

[0034] Noise region identification: Select a band where no effective signal is generated as the noise region;

[0035] Calculate the variance of the noise region: , where Var() is used to calculate the variance of this set.

[0036] Preferably, the extended Kalman filter algorithm is used to perform spatiotemporal alignment of the apparent image I and the ground penetrating radar data R.

[0037] Preferably, when At that time, the apparent image I and the ground penetrating radar data R are output as independent events.

[0038] Preferably, the separately output visual image I and ground-penetrating radar data R are manually analyzed.

[0039] Preferably, when If the value is less than 0.6, the current road area is considered unhealthy, and the maintenance recommendation for the current road area is to take immediate action.

[0040] when =0, and If the value is less than or equal to 0.7, the disease is considered a stable disease, and the maintenance recommendation for the corresponding road section is to not take any action for the time being and to continue monitoring.

[0041] When -1 < When <0, and If the value is less than or equal to 0.7, the disease is considered to be a slowly developing disease, and the maintenance recommendation for the corresponding road section is to pay attention to the trend and then take action.

[0042] when When ≤-1, and If the value is less than or equal to 0.7, the disease is identified as a rapidly deteriorating disease, and the maintenance recommendation for the corresponding road section is to prioritize its treatment.

[0043] The beneficial effects of this invention are as follows: By combining surface images acquired by a depth camera with ground-penetrating radar data acquired by a 3D ground-penetrating radar, and using an extended Kalman filter algorithm to achieve spatiotemporal alignment of multi-source data, this invention ensures that the surface images and radar data have a unified location and timestamp, guaranteeing correct time alignment when establishing causal relationships for disease, and avoiding alignment errors between different diseases. By modeling the causal relationships and evolution trends between surface and structural diseases, dynamic causal diagnosis is achieved, enabling the prediction of disease evolution trends. Furthermore, by extracting features while simultaneously calculating confidence levels, the introduction of a confidence level formula reduces noise interference with the data; that is, the greater the noise, the lower the confidence level, while ensuring that all images or radar data can be used, avoiding the omission of diseases due to noise removal. The existence of confidence levels enhances the weighting of the image or structural features calculated. Or β is low, and +β=1 shifts the data's reliance towards relatively reliable apparent or structural data. Based on historical data, it learns the causal relationship between apparent defects (such as cracks) and structural defects (such as voids) (e.g., cracks may accelerate internal moisture penetration, leading to voids). This can be used for highway defect detection, timely identifying hidden internal defects on roads, reducing traffic accidents, and improving road efficiency. By analyzing and mining high-precision defect detection data, the distribution and evolution trends of road defects can be revealed. This provides relevant maintenance departments and decision-makers with scientific management basis, optimizes resource allocation and maintenance plans, reduces costs and increases efficiency, improves the service life and economic benefits of highways, supports preventive maintenance, and avoids the extensive nature of static assessment. Attached Figure Description

[0044] Figure 1 This is a schematic diagram illustrating a structural health assessment method based on multidimensional road detection data, as described in a specific embodiment of the present invention.

[0045] Figure 2 This is an example diagram of the first type of appearance feature extraction and structural feature extraction in a structural health assessment method based on multidimensional road detection data according to a specific embodiment of the present invention.

[0046] Figure 3 This is a second example diagram illustrating the appearance feature extraction and structural feature extraction of a structural health assessment method based on multidimensional road detection data, as described in a specific embodiment of the present invention.

[0047] Figure 4 This diagram illustrates the output maintenance recommendations of a structural health assessment method based on multidimensional road detection data, as described in a specific embodiment of the present invention. Detailed Implementation

[0048] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0049] Please refer to Figures 1 to 4 A structural health assessment method based on multidimensional road detection data, comprising:

[0050] Simultaneously acquire the road's apparent image I and ground-penetrating radar data R and perform spatiotemporal alignment;

[0051] Apparent feature extraction: Using a pre-trained CNN to extract image features and obtain the apparent feature vector. And calculate the confidence level of the apparent data. ; ;in, The variance of image noise;

[0052] Structural Feature Extraction: Using 3D CNN to process radar data to obtain structural feature vectors. And calculate the confidence level of the structured data. ; ;in, The variance of radar signal noise;

[0053] Constructing a cause-effect graph: Based on historical data, learn the causal relationship between apparent diseases and structural diseases, and use the cause-effect graph to guide the adjustment of weights; , ;where, softmax() is a function; Corr() is the correlation coefficient between appearance and structural features, calculated using the Pearson correlation coefficient, reflecting the causal relationship of the disease; As weight;

[0054] By combining weights and features, a road feature vector is output. , ;

[0055] Calculate the health index , And predict the evolution trend of the disease. , Where sigmoid() is a function, W is the weight matrix, and b is the bias term; For time intervals;

[0056] Health Index The output is the health score of the current road area, combined with the predicted trend of disease evolution. Provide maintenance recommendations.

[0057] As described above, by combining surface images acquired by depth cameras with ground-penetrating radar data acquired by 3D ground-penetrating radar, and using an extended Kalman filter algorithm to achieve spatiotemporal alignment of multi-source data, the surface images and radar data are ensured to have consistent location and timestamps. This guarantees correct time alignment when establishing causal relationships for disease, avoiding alignment errors between different diseases. By modeling the causal relationships and evolution trends between surface and structural diseases, dynamic causal diagnosis is achieved, enabling the prediction of disease evolution trends. Simultaneously, confidence scores are calculated during feature extraction. The confidence score formula is introduced to reduce noise interference with the data; that is, the greater the noise, the lower the confidence score, while ensuring that all images or radar data can be used, avoiding omissions of diseases due to noise removal. The existence of confidence scores influences the weights calculated from the image or structural features. Or β is low, and +β=1 shifts the data's reliance towards relatively reliable apparent or structural data. Based on historical data, it learns the causal relationship between apparent defects (such as cracks) and structural defects (such as voids) (e.g., cracks may accelerate internal moisture penetration, leading to voids). This can be used for highway defect detection, timely identifying hidden internal defects on roads, reducing traffic accidents, and improving road efficiency. By analyzing and mining high-precision defect detection data, the distribution and evolution trends of road defects can be revealed. This provides relevant maintenance departments and decision-makers with scientific management basis, optimizes resource allocation and maintenance plans, reduces costs and increases efficiency, improves the service life and economic benefits of highways, supports preventive maintenance, and avoids the extensive nature of static assessment.

[0058] Furthermore, the image noise variance is obtained based on wavelet transform.

[0059] As described above, image noise variance refers to the intensity of random noise in a road image, analyzed and calculated using wavelet transform, a mathematical tool. This intensity is represented by variance. The larger the variance, the more severe the random interference in the image (such as sensor thermal noise, minor vibrations caused by uneven lighting, etc.), and the worse the image quality.

[0060] Furthermore, the image noise variance obtained based on wavelet transform further includes:

[0061] Choose a wavelet basis function;

[0062] A two-dimensional discrete wavelet transform is performed on the appearance image to decompose it into four sub-bands:

[0063] LL: Low frequency in both horizontal and vertical directions, containing the main approximate information of the appearance image;

[0064] LH: Horizontal low frequency, vertical high frequency, containing edge details in the vertical direction;

[0065] HL: Horizontal high frequencies, vertical low frequencies, including edge details in the horizontal direction;

[0066] HH: Both horizontal and vertical frequencies are high frequencies, containing noise and details in the diagonal direction;

[0067] Perform multi-level decomposition;

[0068] Noise figure extraction: Extract wavelet coefficients of the HH subband after the first level of decomposition;

[0069] Robust variance estimation:

[0070] The standard deviation is calculated using the absolute deviation of the median.

[0071] As described above, LL (low-low), LH (low-high), HL (high-low), and HH (high-high) are calculated by selecting the HH subband after the first-level decomposition. In this high-frequency subband, the vast majority of coefficients are contributed by noise, while the actual signal (image details) is minimal. Calculating the variance of these coefficients may be affected by a small amount of strong edge signals. Therefore, a more robust (more resistant to interference) statistical method is used: the median absolute deviation is used to calculate the standard deviation.

[0072] Furthermore, the standard deviation is calculated using the absolute deviation of the median, including:

[0073] ;in, For each wavelet coefficient in the HH subband; median This is the median of the absolute values ​​of all wavelet coefficients.

[0074] As can be seen from the above description, 0.6745 is an empirical constant.

[0075] Furthermore, the wavelet basis functions are db4 or sym8 in the Daubechies wavelet system.

[0076] As can be seen from the above description, wavelets of this type, such as db4 or sym8, are suitable for image processing.

[0077] Furthermore, the radar signal noise variance is obtained based on the S variation.

[0078] As described above, the GPR signal obtained by ground-penetrating radar, also known as ground-penetrating radar data, is a typical non-stationary signal whose frequency characteristics vary with time (i.e., detection depth). The S-transform combines the advantages of short-time Fourier transform (STFT) and wavelet transform, providing more accurate time-frequency localization characteristics, making it very suitable for analyzing this type of signal.

[0079] Furthermore, the radar signal noise variance obtained based on the S variation further includes:

[0080] Ground-penetrating radar data includes a one-dimensional GPR signal s(t);

[0081] S-transform: Performing an S-transform on a one-dimensional GPR signal s(t) yields a two-dimensional complex time-frequency matrix. ,in, f represents time, and 'f' represents depth; 'f' represents frequency.

[0082] Modulus of complex time-frequency matrix Indicates the signal in time and energy density at frequency f;

[0083] Noise region identification: Select a band where no effective signal is generated as the noise region;

[0084] Calculate the variance of the noise region: , where Var() is used to calculate the variance of this set.

[0085] As can be seen from the above description, in the time-frequency spectrum obtained by S-transform, the effective signal is represented by a high-frequency or low-frequency energy band with concentrated energy at a certain time point; random noise is represented by a "background" or "spurs" with relatively uniform distribution and low energy across the entire time-frequency plane. Therefore, selecting a band where no effective signal is generated as the noise region is for the convenience of comparison and calculation.

[0086] Furthermore, the extended Kalman filter algorithm is used to perform spatiotemporal alignment of the apparent image I and the ground penetrating radar data R.

[0087] Furthermore, when At that time, the apparent image I and the ground penetrating radar data R are output as independent events.

[0088] As can be seen from the above description, when This indicates that there is no correlation between the apparent image I and the ground-penetrating radar data R. The two are output independently and separated through manual analysis or other tools to avoid information distortion caused by improper fusion weights, thereby ensuring the reliability of the disease identification results.

[0089] Furthermore, the separately output visual image I and ground-penetrating radar data R are manually analyzed.

[0090] Furthermore, when If the value is less than 0.6, the current road area is considered unhealthy, and the maintenance recommendation for the current road area is to take immediate action.

[0091] when =0, and If the value is less than or equal to 0.7, the disease is considered a stable disease, and the maintenance recommendation for the current road area is to not take any action for the time being and to continue monitoring.

[0092] When -1 < When <0, and If the value is less than or equal to 0.7, the disease is considered to be a slowly developing disease, and the maintenance recommendation for the current road area is to pay attention to the trend and then take action.

[0093] when When ≤-1, and If the value is less than or equal to 0.7, the disease is considered to be a rapidly deteriorating disease, and the maintenance recommendation for the current road area is to prioritize its treatment. Example 1

[0094] A structural health assessment method based on multidimensional road detection data includes:

[0095] Simultaneously acquire the road's apparent image I and ground-penetrating radar data R, and use the extended Kalman filter algorithm for spatiotemporal alignment;

[0096] Apparent feature extraction: Using a pre-trained CNN to extract image features and obtain the apparent feature vector. , Apparent feature vector The disease features, including texture and shape, are extracted from the apparent image I using a CNN; and the confidence level of the apparent data is calculated. ; ;in, Image noise variance (calculated based on wavelet transform), apparent data confidence. Reflecting data quality, image noise variance refers to the intensity of random noise in a road image, analyzed and calculated using wavelet transform, a mathematical tool. This intensity is represented by variance. The larger the variance, the more severe the random interference in the image (such as sensor thermal noise, minor vibrations caused by uneven lighting, etc.), and the worse the image quality.

[0097] Wavelet transform includes:

[0098] Choose a suitable wavelet basis function for image processing, such as db4 or sym8 from the Daubechies wavelet system.

[0099] A two-dimensional discrete wavelet transform is performed on the appearance image to decompose it into four sub-bands:

[0100] LL (Low-Low): Low frequencies in both the horizontal and vertical directions, containing the main approximate information of the appearance image;

[0101] LH (Low-High): Horizontal low frequency, vertical high frequency, containing edge details in the vertical direction;

[0102] HL (High-Low): Horizontal high frequencies, vertical low frequencies, containing edge details in the horizontal direction;

[0103] HH (High-High): Both horizontal and vertical frequencies are high, containing noise and details in the diagonal direction;

[0104] Perform multi-level decomposition;

[0105] Noise figure extraction: Extract wavelet coefficients of the HH subband after the first level of decomposition; by selecting the HH subband after the first level of decomposition, most of the coefficients in this high-frequency subband are contributed by noise, while the real signal (image details) is very little.

[0106] Robust variance estimation:

[0107] The standard deviation is calculated using the absolute deviation of the median. However, calculating the variance of these coefficients can be affected by minor interference from strong edge signals. Therefore, a more robust (more resistant to interference) statistical method is used: the absolute deviation of the median to calculate the standard deviation. .

[0108] ;in, For each wavelet coefficient in the HH subband; median This is the median of the absolute values ​​of all wavelet coefficients. 0.6745 is an empirical constant.

[0109] Structural Feature Extraction: Using 3D CNN to process radar data to obtain structural feature vectors. , The data, including temporal spectrum and morphological features, was extracted from radar data R using a 3D CNN. The structural data confidence score was then calculated. ; ;in, This represents the radar signal noise variance. The confidence level of the structural data. It reflects the degree of electromagnetic interference.

[0110] The radar signal noise variance is obtained based on the S-transform. The GPR signal obtained by ground-penetrating radar, also known as ground-penetrating radar data, is a typical non-stationary signal whose frequency characteristics vary with time (i.e., detection depth). The S-transform combines the advantages of short-time Fourier transform (STFT) and wavelet transform, providing more accurate time-frequency localization characteristics, making it very suitable for analyzing this type of signal.

[0111] Ground-penetrating radar data includes a one-dimensional GPR signal s(t);

[0112] S-transform: Performing an S-transform on a one-dimensional GPR signal s(t) yields a two-dimensional complex time-frequency matrix. ,in, f represents time, and 'f' represents depth; 'f' represents frequency.

[0113] Modulus of complex time-frequency matrix Indicates the signal in time and energy density at frequency f;

[0114] Noise region identification: Select a band where no effective signal is generated as the noise region, i.e., a "quiet" time window without obvious reflected signals (a "quiet" time window without obvious reflected signals); in the time spectrum obtained by S-transform, the effective signal is represented by a high-frequency or low-frequency energy band with concentrated energy at a certain time point; random noise is represented by a "background" or "spurs" with relatively uniform distribution and low energy across the entire time-frequency plane; selecting a band where no effective signal is generated as the noise region is for the convenience of comparison and calculation.

[0115] Calculate the variance of the noise region: Where Var() is used to calculate the variance of this set. The S-transform is used to map the signal onto the time-frequency plane, thereby accurately separating the noise-dominant region in both time and frequency dimensions. Then, the degree of fluctuation (variance) of signal energy in this region is calculated to quantify the noise intensity.

[0116] Constructing a causal graph: Based on historical data, the causal relationship between apparent diseases and structural diseases can be learned through big data models (or the causal relationship coefficients between different apparent diseases and structural diseases can be manually labeled), and the causal graph can be used to guide the adjustment of weights; , ; where softmax() is a function; Corr() is the correlation coefficient between appearance and structural features, calculated using the Pearson correlation coefficient, reflecting the causal relationship of the disease; β is the attention weight for appearance features, and β is the weight for structural features; when At that time, the apparent image I and the ground-penetrating radar data R are output as independent events. The separately output apparent image I and ground-penetrating radar data R are then manually analyzed. The reason is that... At that time, either Either Image noise variance or As the radar noise variance approaches infinity, or if the noise becomes infinitely large, the apparent image I or the ground-penetrating radar data R is deemed invalid. This requires manual determination. The reason for not specifically removing noise from the apparent image I and the ground-penetrating radar data R is to ensure no omissions and to guarantee detection effectiveness. Even if the entire image or radar data is filled with noise, this is extremely rare. Therefore, in most cases, even with high noise levels, the data can still be used, unless... ,Right now and There are cases where there is no connection (e.g., no damage to the surface layer and no damage to the structural layer), which require manual identification and definition.

[0117] By combining weights and features, a road feature vector is output. , ;

[0118] Calculate the health index , And predict the evolution trend of the disease. , Where sigmoid() is a function, W is a weight matrix, and b is a bias term; W and b are obtained through training on historical data; The rate of change of the health index. This is a time interval (usually one year). The health index at time t-1;

[0119] Health Index The output is a health score for the current road area, ranging from 0 to 1 (closer to 1 indicates healthier, closer to 0 indicates unhealthier), combined with a prediction of disease evolution trends. Provide maintenance recommendations;

[0120] when If the value is less than 0.6, the current road area is considered unhealthy, and the maintenance recommendation for the current road area is to take immediate action.

[0121] when =0, and If the value is less than or equal to 0.7, the disease is considered a stable disease, and the maintenance recommendation for the current road area is to not take any action for the time being and to continue monitoring.

[0122] When -1 < When <0, and If the value is less than or equal to 0.7, the disease is considered to be a slowly developing disease, and the maintenance recommendation for the current road area is to pay attention to the trend and then take action.

[0123] when When ≤-1, and If the value is less than or equal to 0.7, the disease is considered to be a rapidly deteriorating disease, and the maintenance recommendation for the current road area is to prioritize its treatment.

[0124] If there are different maintenance recommendations for a certain road section, the one that is dealt with immediately takes priority over the other maintenance recommendations.

[0125] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A structural health assessment method based on multidimensional road detection data, characterized in that, include: The apparent image I of the collected road and the ground penetrating radar data R are spatiotemporally aligned; Extracting the appearance feature vector from the appearance image and apparent data confidence Extracting structural feature vectors from radar data and structural data confidence ; Apparent feature extraction: Using a pre-trained CNN to extract image features and obtain the apparent feature vector. And calculate the confidence level of the apparent data. ; ;in, The variance of image noise; Structural Feature Extraction: Using 3D CNN to process radar data to obtain structural feature vectors. And calculate the confidence level of the structured data. ; ;in, The variance of radar signal noise; Output road feature vector , ; As weight; Constructing a cause-effect graph: Based on historical data, learn the causal relationship between apparent diseases and structural diseases, and use the cause-effect graph to guide the adjustment of weights; , ;where, softmax() is a function; Corr() is the correlation coefficient between appearance and structural features, calculated using the Pearson correlation coefficient, reflecting the causal relationship of the disease; Calculate the health index , And predict the evolution trend of the disease. , Where sigmoid() is a function, W is the weight matrix, and b is the bias term; For time intervals; Health Index The output is the health score of the current road area, combined with the predicted trend of disease evolution. Provide maintenance recommendations; when If the value is less than 0.6, the current road area is considered unhealthy, and the maintenance recommendation for the current road area is to take immediate action. when =0, and If the value is ≤0.7, the disease is considered a stable disease, and the maintenance recommendation is to not treat it for the time being and to continue monitoring it. When -1 < When <0, and If the value is ≤0.7, the disease is considered to be a slowly developing disease, and the maintenance recommendation is to monitor the trend first, and then take action. when When ≤-1, and If the value is ≤0.7, the disease is identified as a rapidly deteriorating disease, and the maintenance recommendation is to prioritize treatment. when At that time, the apparent image and ground-penetrating radar data were output separately as independent events for manual analysis; If there are different maintenance recommendations for a certain road section, the one that is dealt with immediately takes priority over the other maintenance recommendations.

2. The structural health assessment method based on multidimensional road detection data according to claim 1, characterized in that, The image noise variance is obtained based on wavelet transform.

3. The structural health assessment method based on multidimensional road detection data according to claim 2, characterized in that, The image noise variance, obtained based on wavelet transform, further includes: Choose a wavelet basis function; A two-dimensional discrete wavelet transform is performed on the appearance image to decompose it into four sub-bands: LL: Low frequency in both horizontal and vertical directions, containing the main approximate information of the appearance image; LH: Horizontal low frequency, vertical high frequency, containing edge details in the vertical direction; HL: Horizontal high frequencies, vertical low frequencies, including edge details in the horizontal direction; HH: Both horizontal and vertical frequencies are high frequencies, containing noise and details in the diagonal direction; Perform multi-level decomposition; Noise figure extraction: Extract wavelet coefficients of the HH subband after the first level of decomposition; Robust variance estimation: The standard deviation is calculated using the absolute deviation of the median.

4. The structural health assessment method based on multidimensional road detection data according to claim 3, characterized in that, The standard deviation is calculated using the absolute deviation of the median, including: ;in, For each wavelet coefficient in the HH subband; median This is the median of the absolute values ​​of all wavelet coefficients.

5. The structural health assessment method based on multidimensional road detection data according to claim 4, characterized in that, The wavelet basis functions are db4 or sym8 in the Daubechies wavelet system.

6. The structural health assessment method based on multidimensional road detection data according to claim 1, characterized in that, The radar signal noise variance is obtained based on the S variation.

7. The structural health assessment method based on multidimensional road detection data according to claim 6, characterized in that, The radar signal noise variance, obtained based on the S-variance, further includes: Ground-penetrating radar data includes a one-dimensional GPR signal s(t); S-transform: Performing an S-transform on a one-dimensional GPR signal s(t) yields a two-dimensional complex time-frequency matrix. ,in, f represents time, and 'f' represents depth; 'f' represents frequency. Modulus of complex time-frequency matrix Indicates the signal in time and energy density at frequency f; Noise region identification: Select a band where no effective signal is generated as the noise region; Calculate the variance of the noise region: , where Var() is used to calculate the variance of this set.

8. The structural health assessment method based on multidimensional road detection data according to claim 1, characterized in that, The extended Kalman filter algorithm is used to perform spatiotemporal alignment of the appearance image and ground penetrating radar data.

Images